Remove all the calls to StaticCast methods (#605)

ChocolArm64/Instructions/InstEmitMemoryHelper.cs

@@ -200,7 +200,7 @@ namespace ChocolArm64.Instructions
 
             switch (size)
             {
-                case 2: context.EmitCall(typeof(Sse),  nameof(Sse.LoadScalarVector128));  break;
+                case 2: context.EmitCall(typeof(Sse), nameof(Sse.LoadScalarVector128));  break;
 
                 case 3:
                 {
@@ -208,12 +208,10 @@ namespace ChocolArm64.Instructions
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.LoadScalarVector128), types));
 
-                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleToSingle));
-
                     break;
                 }
 
-                case 4: context.EmitCall(typeof(Sse),  nameof(Sse.LoadAlignedVector128)); break;
+                case 4: context.EmitCall(typeof(Sse), nameof(Sse.LoadAlignedVector128)); break;
 
                 throw new InvalidOperationException($"Invalid vector load size of {1 << size} bytes.");
             }
@@ -283,18 +281,9 @@ namespace ChocolArm64.Instructions
 
             switch (size)
             {
-                case 2: context.EmitCall(typeof(Sse), nameof(Sse.StoreScalar)); break;
-
-                case 3:
-                {
-                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToDouble));
-
-                    context.EmitCall(typeof(Sse2), nameof(Sse2.StoreScalar));
-
-                    break;
-                }
-
-                case 4: context.EmitCall(typeof(Sse), nameof(Sse.StoreAligned)); break;
+                case 2: context.EmitCall(typeof(Sse),  nameof(Sse.StoreScalar));  break;
+                case 3: context.EmitCall(typeof(Sse2), nameof(Sse2.StoreScalar)); break;
+                case 4: context.EmitCall(typeof(Sse),  nameof(Sse.StoreAligned)); break;
 
                 default: throw new InvalidOperationException($"Invalid vector store size of {1 << size} bytes.");
             }

ChocolArm64/Instructions/InstEmitSimdArithmetic.cs

@@ -194,8 +194,7 @@ namespace ChocolArm64.Instructions
                     context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SubtractScalar), typesSubAndNot));
-
-                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.AndNot), typesSubAndNot));
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.AndNot),         typesSubAndNot));
 
                     context.EmitStvec(op.Rd);
 
@@ -209,14 +208,13 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(-0d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), typesSsv));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), typesSubAndNot));
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot),         typesSubAndNot));
 
-                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesSubAndNot));
-
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
 
                     EmitVectorZeroUpper(context, op.Rd);
                 }
@@ -252,8 +250,7 @@ namespace ChocolArm64.Instructions
                     context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Subtract), typesSubAndNot));
-
-                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.AndNot), typesSubAndNot));
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.AndNot),   typesSubAndNot));
 
                     context.EmitStvec(op.Rd);
 
@@ -270,14 +267,13 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(-0d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesSubAndNot));
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot),   typesSubAndNot));
 
-                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesSubAndNot));
-
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -321,11 +317,11 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(-0d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), typesSsv));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAndNot));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
 
                     EmitVectorZeroUpper(context, op.Rd);
                 }
@@ -374,11 +370,11 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(-0d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAndNot));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -445,12 +441,12 @@ namespace ChocolArm64.Instructions
                 {
                     Type[] typesAddH = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
                     context.Emit(OpCodes.Dup);
 
                     context.EmitCall(typeof(Sse3).GetMethod(nameof(Sse3.HorizontalAdd), typesAddH));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
 
                     EmitVectorZeroUpper(context, op.Rd);
                 }
@@ -536,14 +532,14 @@ namespace ChocolArm64.Instructions
                 {
                     Type[] typesMulAdd = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                    EmitLdvecWithCastToDouble(context, op.Ra);
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Ra);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), typesMulAdd));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AddScalar),      typesMulAdd));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
 
                     EmitVectorZeroUpper(context, op.Rd);
                 }
@@ -718,14 +714,14 @@ namespace ChocolArm64.Instructions
                 {
                     Type[] typesMulAdd = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                    EmitLdvecWithCastToDouble(context, op.Rd);
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rd);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesMulAdd));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add),      typesMulAdd));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -751,18 +747,14 @@ namespace ChocolArm64.Instructions
                     Type[] typesMulAdd = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     context.EmitLdvec(op.Rd);
-
                     context.EmitLdvec(op.Rn);
-
                     context.EmitLdvec(op.Rm);
                     context.Emit(OpCodes.Dup);
 
                     context.EmitLdc_I4(op.Index | op.Index << 2 | op.Index << 4 | op.Index << 6);
-                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle), typesSfl));
-
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle),  typesSfl));
                     context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesMulAdd));
-
-                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Add), typesMulAdd));
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Add),      typesMulAdd));
 
                     context.EmitStvec(op.Rd);
 
@@ -776,21 +768,17 @@ namespace ChocolArm64.Instructions
                     Type[] typesSfl    = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>), typeof(byte) };
                     Type[] typesMulAdd = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                    EmitLdvecWithCastToDouble(context, op.Rd);
-
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rd);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
                     context.Emit(OpCodes.Dup);
 
                     context.EmitLdc_I4(op.Index | op.Index << 1);
-                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Shuffle), typesSfl));
-
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Shuffle),  typesSfl));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesMulAdd));
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add),      typesMulAdd));
 
-                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesMulAdd));
-
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -841,14 +829,14 @@ namespace ChocolArm64.Instructions
                 {
                     Type[] typesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                    EmitLdvecWithCastToDouble(context, op.Rd);
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rd);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesMulSub));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesMulSub));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -874,17 +862,13 @@ namespace ChocolArm64.Instructions
                     Type[] typesMulSub = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     context.EmitLdvec(op.Rd);
-
                     context.EmitLdvec(op.Rn);
-
                     context.EmitLdvec(op.Rm);
                     context.Emit(OpCodes.Dup);
 
                     context.EmitLdc_I4(op.Index | op.Index << 2 | op.Index << 4 | op.Index << 6);
-                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle), typesSfl));
-
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle),  typesSfl));
                     context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesMulSub));
-
                     context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Subtract), typesMulSub));
 
                     context.EmitStvec(op.Rd);
@@ -899,21 +883,17 @@ namespace ChocolArm64.Instructions
                     Type[] typesSfl    = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>), typeof(byte) };
                     Type[] typesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                    EmitLdvecWithCastToDouble(context, op.Rd);
-
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rd);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
                     context.Emit(OpCodes.Dup);
 
                     context.EmitLdc_I4(op.Index | op.Index << 1);
-                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Shuffle), typesSfl));
-
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Shuffle),  typesSfl));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesMulSub));
-
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesMulSub));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -950,14 +930,14 @@ namespace ChocolArm64.Instructions
                 {
                     Type[] typesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                    EmitLdvecWithCastToDouble(context, op.Ra);
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Ra);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), typesMulSub));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), typesMulSub));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
 
                     EmitVectorZeroUpper(context, op.Rd);
                 }
@@ -1020,13 +1000,11 @@ namespace ChocolArm64.Instructions
                     Type[] typesMul = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
 
                     context.EmitLdvec(op.Rn);
-
                     context.EmitLdvec(op.Rm);
                     context.Emit(OpCodes.Dup);
 
                     context.EmitLdc_I4(op.Index | op.Index << 2 | op.Index << 4 | op.Index << 6);
-                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle), typesSfl));
-
+                    context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Shuffle),  typesSfl));
                     context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesMul));
 
                     context.EmitStvec(op.Rd);
@@ -1041,17 +1019,15 @@ namespace ChocolArm64.Instructions
                     Type[] typesSfl = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>), typeof(byte) };
                     Type[] typesMul = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
                     context.Emit(OpCodes.Dup);
 
                     context.EmitLdc_I4(op.Index | op.Index << 1);
-                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Shuffle), typesSfl));
-
+                    context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Shuffle),  typesSfl));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesMul));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -1125,11 +1101,11 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(-0d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), typesSsv));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesXor));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
 
                     EmitVectorZeroUpper(context, op.Rd);
                 }
@@ -1175,11 +1151,11 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(-0d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesXor));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -1242,8 +1218,7 @@ namespace ChocolArm64.Instructions
 
             int sizeF = op.Size & 1;
 
-            if (Optimizations.FastFP && Optimizations.UseSse
-                                     && sizeF == 0)
+            if (Optimizations.FastFP && Optimizations.UseSse && sizeF == 0)
             {
                 EmitScalarSseOrSse2OpF(context, nameof(Sse.ReciprocalScalar));
             }
@@ -1262,8 +1237,7 @@ namespace ChocolArm64.Instructions
 
             int sizeF = op.Size & 1;
 
-            if (Optimizations.FastFP && Optimizations.UseSse
-                                     && sizeF == 0)
+            if (Optimizations.FastFP && Optimizations.UseSse && sizeF == 0)
             {
                 EmitVectorSseOrSse2OpF(context, nameof(Sse.Reciprocal));
             }
@@ -1310,13 +1284,13 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(2d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), typesSsv));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), typesMulSub));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), typesMulSub));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
 
                     EmitVectorZeroUpper(context, op.Rd);
                 }
@@ -1367,13 +1341,13 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(2d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesMulSub));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesMulSub));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -1579,8 +1553,7 @@ namespace ChocolArm64.Instructions
 
             int sizeF = op.Size & 1;
 
-            if (Optimizations.FastFP && Optimizations.UseSse
-                                     && sizeF == 0)
+            if (Optimizations.FastFP && Optimizations.UseSse && sizeF == 0)
             {
                 EmitScalarSseOrSse2OpF(context, nameof(Sse.ReciprocalSqrtScalar));
             }
@@ -1599,8 +1572,7 @@ namespace ChocolArm64.Instructions
 
             int sizeF = op.Size & 1;
 
-            if (Optimizations.FastFP && Optimizations.UseSse
-                                     && sizeF == 0)
+            if (Optimizations.FastFP && Optimizations.UseSse && sizeF == 0)
             {
                 EmitVectorSseOrSse2OpF(context, nameof(Sse.ReciprocalSqrt));
             }
@@ -1654,14 +1626,14 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(3d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), typesSsv));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), typesMulSub));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SubtractScalar), typesMulSub));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), typesMulSub));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
 
                     EmitVectorZeroUpper(context, op.Rd);
                 }
@@ -1719,14 +1691,14 @@ namespace ChocolArm64.Instructions
                     context.EmitLdc_R8(3d);
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
-                    EmitLdvecWithCastToDouble(context, op.Rm);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesMulSub));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesMulSub));
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), typesMulSub));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -1864,11 +1836,11 @@ namespace ChocolArm64.Instructions
 
                 VectorHelper.EmitCall(context, namesSzv[op.Size]);
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesSub));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -1953,14 +1925,14 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
 
                 context.EmitCall(typeof(Sse41).GetMethod(namesCvt[op.Size], typesCvt));
 
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -1969,7 +1941,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -1999,9 +1971,8 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size + 1);
-
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2010,7 +1981,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -2027,12 +1998,12 @@ namespace ChocolArm64.Instructions
                 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.EmitLdvec(op.Rn);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp();
 
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rm);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp2();
@@ -2046,10 +2017,9 @@ namespace ChocolArm64.Instructions
 
                 context.EmitLdc_I4(1);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), typesSra));
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add),                  typesAndXorAdd));
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAndXorAdd));
-
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2083,23 +2053,21 @@ namespace ChocolArm64.Instructions
 
                 context.EmitStvectmp();
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
                 context.EmitLdvectmp();
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAddSub));
 
                 context.Emit(OpCodes.Dup);
 
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rm);
                 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.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);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2128,12 +2096,12 @@ namespace ChocolArm64.Instructions
 
                 Type typeSse = op.Size == 1 ? typeof(Sse2) : typeof(Sse41);
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitCall(typeSse.GetMethod(nameof(Sse2.Max), typesMax));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2169,12 +2137,12 @@ namespace ChocolArm64.Instructions
 
                 Type typeSse = op.Size == 1 ? typeof(Sse2) : typeof(Sse41);
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitCall(typeSse.GetMethod(nameof(Sse2.Min), typesMin));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2219,16 +2187,15 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithSignedCast(context, op.Rd, op.Size + 1);
-
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rn);
 
                 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.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2239,7 +2206,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesMulAdd));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -2279,16 +2246,15 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithSignedCast(context, op.Rd, op.Size + 1);
-
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rn);
 
                 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.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2299,7 +2265,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesMulSub));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -2426,20 +2392,19 @@ namespace ChocolArm64.Instructions
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp();
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
                 context.EmitLdvectmp();
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesSubAdd));
 
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rm);
                 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));
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Average), typesAvg));
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add),     typesSubAdd));
-
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2478,14 +2443,14 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
 
                 context.EmitCall(typeof(Sse41).GetMethod(namesCvt[op.Size], typesCvt));
 
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2494,7 +2459,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesSub));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -2519,9 +2484,8 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size + 1);
-
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2530,7 +2494,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesSub));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -2632,14 +2596,14 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 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.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2648,7 +2612,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -2697,9 +2661,8 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size + 1);
-
-                EmitLdvecWithUnsignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2708,7 +2671,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -2725,12 +2688,12 @@ namespace ChocolArm64.Instructions
                 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.EmitLdvec(op.Rn);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp();
 
-                EmitLdvecWithUnsignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rm);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp2();
@@ -2744,10 +2707,9 @@ namespace ChocolArm64.Instructions
 
                 context.EmitLdc_I4(1);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrl));
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add),               typesAndXorAdd));
 
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAndXorAdd));
-
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2774,16 +2736,15 @@ namespace ChocolArm64.Instructions
             {
                 Type[] typesAvgSub = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], VectorUIntTypesPerSizeLog2[op.Size] };
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
                 context.Emit(OpCodes.Dup);
 
-                EmitLdvecWithUnsignedCast(context, op.Rm, op.Size);
-
-                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Average), typesAvgSub));
+                context.EmitLdvec(op.Rm);
 
+                context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Average),  typesAvgSub));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesAvgSub));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2812,12 +2773,12 @@ namespace ChocolArm64.Instructions
 
                 Type typeSse = op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
-                EmitLdvecWithUnsignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitCall(typeSse.GetMethod(nameof(Sse2.Max), typesMax));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2853,12 +2814,12 @@ namespace ChocolArm64.Instructions
 
                 Type typeSse = op.Size == 0 ? typeof(Sse2) : typeof(Sse41);
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
-                EmitLdvecWithUnsignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitCall(typeSse.GetMethod(nameof(Sse2.Min), typesMin));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -2903,16 +2864,15 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithUnsignedCast(context, op.Rd, op.Size + 1);
-
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rn);
 
                 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.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2923,7 +2883,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesMulAdd));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -2963,16 +2923,15 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithUnsignedCast(context, op.Rd, op.Size + 1);
-
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rn);
 
                 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.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -2983,7 +2942,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesMulSub));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -3052,12 +3011,12 @@ namespace ChocolArm64.Instructions
             {
                 Type[] typesAvg = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], VectorUIntTypesPerSizeLog2[op.Size] };
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
-                EmitLdvecWithUnsignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Average), typesAvg));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -3106,14 +3065,14 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 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.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -3122,7 +3081,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesSub));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -3147,9 +3106,8 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size + 1);
-
-                EmitLdvecWithUnsignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSrl));
@@ -3158,7 +3116,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Subtract), typesSub));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {

ChocolArm64/Instructions/InstEmitSimdCmp.cs

@@ -382,7 +382,7 @@ namespace ChocolArm64.Instructions
                     ILLabel lblNaN = new ILLabel();
                     ILLabel lblEnd = new ILLabel();
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
 
                     context.Emit(OpCodes.Dup);
                     context.EmitStvectmp();
@@ -393,7 +393,7 @@ namespace ChocolArm64.Instructions
                     }
                     else
                     {
-                        EmitLdvecWithCastToDouble(context, op.Rm);
+                        context.EmitLdvec(op.Rm);
                     }
 
                     context.Emit(OpCodes.Dup);
@@ -656,12 +656,12 @@ namespace ChocolArm64.Instructions
 
                 if (!isLeOrLt)
                 {
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
                 }
 
                 if (op is OpCodeSimdReg64 binOp)
                 {
-                    EmitLdvecWithCastToDouble(context, binOp.Rm);
+                    context.EmitLdvec(binOp.Rm);
                 }
                 else
                 {
@@ -670,12 +670,12 @@ namespace ChocolArm64.Instructions
 
                 if (isLeOrLt)
                 {
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
                 }
 
                 context.EmitCall(typeof(Sse2).GetMethod(name, types));
 
-                EmitStvecWithCastFromDouble(context, op.Rd);
+                context.EmitStvec(op.Rd);
 
                 if (scalar)
                 {

ChocolArm64/Instructions/InstEmitSimdCvt.cs

@@ -23,7 +23,7 @@ namespace ChocolArm64.Instructions
                     //Double -> Single.
                     VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleZero));
 
-                    EmitLdvecWithCastToDouble(context, op.Rn);
+                    context.EmitLdvec(op.Rn);
 
                     Type[] types = new Type[] { typeof(Vector128<float>), typeof(Vector128<double>) };
 
@@ -42,7 +42,7 @@ namespace ChocolArm64.Instructions
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertScalarToVector128Double), types));
 
-                    EmitStvecWithCastFromDouble(context, op.Rd);
+                    context.EmitStvec(op.Rd);
                 }
                 else
                 {
@@ -91,7 +91,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Double), typesCvt));
 
-                EmitStvecWithCastFromDouble(context, op.Rd);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -154,7 +154,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
 
-                EmitLdvecWithCastToDouble(context, op.Rn);
+                context.EmitLdvec(op.Rn);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
                 context.Emit(OpCodes.Dup);
 
@@ -332,7 +332,7 @@ namespace ChocolArm64.Instructions
             {
                 Type[] typesCvt = new Type[] { typeof(Vector128<int>) };
 
-                EmitLdvecWithSignedCast(context, op.Rn, 2);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
 

ChocolArm64/Instructions/InstEmitSimdHelper.cs

@@ -86,13 +86,13 @@ namespace ChocolArm64.Instructions
         {
             OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
 
-            EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+            context.EmitLdvec(op.Rn);
 
             Type baseType = VectorIntTypesPerSizeLog2[op.Size];
 
             if (op is OpCodeSimdReg64 binOp)
             {
-                EmitLdvecWithSignedCast(context, binOp.Rm, op.Size);
+                context.EmitLdvec(binOp.Rm);
 
                 context.EmitCall(type.GetMethod(name, new Type[] { baseType, baseType }));
             }
@@ -101,7 +101,7 @@ namespace ChocolArm64.Instructions
                 context.EmitCall(type.GetMethod(name, new Type[] { baseType }));
             }
 
-            EmitStvecWithSignedCast(context, op.Rd, op.Size);
+            context.EmitStvec(op.Rd);
 
             if (op.RegisterSize == RegisterSize.Simd64)
             {
@@ -109,80 +109,6 @@ namespace ChocolArm64.Instructions
             }
         }
 
-        public static void EmitLdvecWithSignedCast(ILEmitterCtx context, int reg, int size)
-        {
-            context.EmitLdvec(reg);
-
-            switch (size)
-            {
-                case 0: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToSByte)); break;
-                case 1: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToInt16)); break;
-                case 2: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToInt32)); break;
-                case 3: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToInt64)); break;
-
-                default: throw new ArgumentOutOfRangeException(nameof(size));
-            }
-        }
-
-        public static void EmitLdvecWithCastToDouble(ILEmitterCtx context, int reg)
-        {
-            context.EmitLdvec(reg);
-
-            VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToDouble));
-        }
-
-        public static void EmitStvecWithCastFromDouble(ILEmitterCtx context, int reg)
-        {
-            VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleToSingle));
-
-            context.EmitStvec(reg);
-        }
-
-        public static void EmitLdvecWithUnsignedCast(ILEmitterCtx context, int reg, int size)
-        {
-            context.EmitLdvec(reg);
-
-            switch (size)
-            {
-                case 0: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToByte));   break;
-                case 1: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToUInt16)); break;
-                case 2: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToUInt32)); break;
-                case 3: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToUInt64)); break;
-
-                default: throw new ArgumentOutOfRangeException(nameof(size));
-            }
-        }
-
-        public static void EmitStvecWithSignedCast(ILEmitterCtx context, int reg, int size)
-        {
-            switch (size)
-            {
-                case 0: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSByteToSingle)); break;
-                case 1: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorInt16ToSingle)); break;
-                case 2: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorInt32ToSingle)); break;
-                case 3: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorInt64ToSingle)); break;
-
-                default: throw new ArgumentOutOfRangeException(nameof(size));
-            }
-
-            context.EmitStvec(reg);
-        }
-
-        public static void EmitStvecWithUnsignedCast(ILEmitterCtx context, int reg, int size)
-        {
-            switch (size)
-            {
-                case 0: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorByteToSingle));   break;
-                case 1: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorUInt16ToSingle)); break;
-                case 2: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorUInt32ToSingle)); break;
-                case 3: VectorHelper.EmitCall(context, nameof(VectorHelper.VectorUInt64ToSingle)); break;
-
-                default: throw new ArgumentOutOfRangeException(nameof(size));
-            }
-
-            context.EmitStvec(reg);
-        }
-
         public static void EmitScalarSseOrSse2OpF(ILEmitterCtx context, string name)
         {
             EmitSseOrSse2OpF(context, name, true);
@@ -199,17 +125,7 @@ namespace ChocolArm64.Instructions
 
             int sizeF = op.Size & 1;
 
-            void Ldvec(int reg)
-            {
-                context.EmitLdvec(reg);
-
-                if (sizeF == 1)
-                {
-                    VectorHelper.EmitCall(context, nameof(VectorHelper.VectorSingleToDouble));
-                }
-            }
-
-            Ldvec(op.Rn);
+            context.EmitLdvec(op.Rn);
 
             Type type;
             Type baseType;
@@ -227,7 +143,7 @@ namespace ChocolArm64.Instructions
 
             if (op is OpCodeSimdReg64 binOp)
             {
-                Ldvec(binOp.Rm);
+                context.EmitLdvec(binOp.Rm);
 
                 context.EmitCall(type.GetMethod(name, new Type[] { baseType, baseType }));
             }
@@ -236,11 +152,6 @@ namespace ChocolArm64.Instructions
                 context.EmitCall(type.GetMethod(name, new Type[] { baseType }));
             }
 
-            if (sizeF == 1)
-            {
-                VectorHelper.EmitCall(context, nameof(VectorHelper.VectorDoubleToSingle));
-            }
-
             context.EmitStvec(op.Rd);
 
             if (scalar)
@@ -1014,12 +925,12 @@ namespace ChocolArm64.Instructions
             {
                 Type[] types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
 
-                EmitLdvecWithCastToDouble(context, op.Rn);
+                context.EmitLdvec(op.Rn);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp();
 
-                EmitLdvecWithCastToDouble(context, op.Rm);
+                context.EmitLdvec(op.Rm);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp2();
@@ -1033,7 +944,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(name, types));
 
-                EmitStvecWithCastFromDouble(context, op.Rd);
+                context.EmitStvec(op.Rd);
             }
         }
 
@@ -1277,13 +1188,9 @@ namespace ChocolArm64.Instructions
         }
 
         // TSrc (16bit, 32bit, 64bit; signed, unsigned) > TDst (8bit, 16bit, 32bit; signed, unsigned).
-        public static void EmitSatQ(
-            ILEmitterCtx context,
-            int  sizeDst,
-            bool signedSrc,
-            bool signedDst)
+        public static void EmitSatQ(ILEmitterCtx context, int sizeDst, bool signedSrc, bool signedDst)
         {
-            if (sizeDst > 2)
+            if ((uint)sizeDst > 2)
             {
                 throw new ArgumentOutOfRangeException(nameof(sizeDst));
             }

ChocolArm64/Instructions/InstEmitSimdLogical.cs

@@ -32,12 +32,12 @@ namespace ChocolArm64.Instructions
 
                 Type[] typesAndNot = new Type[] { typeof(Vector128<byte>), typeof(Vector128<byte>) };
 
-                EmitLdvecWithUnsignedCast(context, op.Rm, 0);
-                EmitLdvecWithUnsignedCast(context, op.Rn, 0);
+                context.EmitLdvec(op.Rm);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAndNot));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, 0);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -83,16 +83,16 @@ namespace ChocolArm64.Instructions
 
                 string nameAndNot = notRm ? nameof(Sse2.AndNot) : nameof(Sse2.And);
 
-                EmitLdvecWithUnsignedCast(context, op.Rd, 0);
-                EmitLdvecWithUnsignedCast(context, op.Rm, 0);
-                EmitLdvecWithUnsignedCast(context, op.Rn, 0);
-                EmitLdvecWithUnsignedCast(context, op.Rd, 0);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rm);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rd);
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesXorAndNot));
                 context.EmitCall(typeof(Sse2).GetMethod(nameAndNot,       typesXorAndNot));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesXorAndNot));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, 0);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -141,20 +141,20 @@ namespace ChocolArm64.Instructions
 
                 Type[] typesXorAnd = new Type[] { typeof(Vector128<byte>), typeof(Vector128<byte>) };
 
-                EmitLdvecWithUnsignedCast(context, op.Rm, 0);
+                context.EmitLdvec(op.Rm);
                 context.Emit(OpCodes.Dup);
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, 0);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesXorAnd));
 
-                EmitLdvecWithUnsignedCast(context, op.Rd, 0);
+                context.EmitLdvec(op.Rd);
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.And), typesXorAnd));
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Xor), typesXorAnd));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, 0);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -199,14 +199,14 @@ namespace ChocolArm64.Instructions
                 Type[] typesSav    = new Type[] { typeof(byte) };
                 Type[] typesAndNot = new Type[] { typeof(Vector128<byte>), typeof(Vector128<byte>) };
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, 0);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(byte.MaxValue);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAndNot));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, 0);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -228,8 +228,8 @@ namespace ChocolArm64.Instructions
                 Type[] typesSav      = new Type[] { typeof(byte) };
                 Type[] typesAndNotOr = new Type[] { typeof(Vector128<byte>), typeof(Vector128<byte>) };
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, 0);
-                EmitLdvecWithUnsignedCast(context, op.Rm, 0);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4(byte.MaxValue);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
@@ -237,7 +237,7 @@ namespace ChocolArm64.Instructions
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AndNot), typesAndNotOr));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or),     typesAndNotOr));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, 0);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -305,7 +305,7 @@ namespace ChocolArm64.Instructions
                 Type[] typesSve = new Type[] { typeof(long), typeof(long) };
                 Type[] typesSfl = new Type[] { typeof(Vector128<sbyte>), typeof(Vector128<sbyte>) };
 
-                EmitLdvecWithSignedCast(context, op.Rn, 0); // value
+                context.EmitLdvec(op.Rn); // value
 
                 context.EmitLdc_I8(14L << 56 | 15L << 48 | 12L << 40 | 13L << 32 | 10L << 24 | 11L << 16 | 08L << 8 | 09L << 0); // maskE1
                 context.EmitLdc_I8(06L << 56 | 07L << 48 | 04L << 40 | 05L << 32 | 02L << 24 | 03L << 16 | 00L << 8 | 01L << 0); // maskE0
@@ -314,7 +314,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesSfl));
 
-                EmitStvecWithSignedCast(context, op.Rd, 0);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -336,7 +336,7 @@ namespace ChocolArm64.Instructions
                 Type[] typesSve = new Type[] { typeof(long), typeof(long) };
                 Type[] typesSfl = new Type[] { typeof(Vector128<sbyte>), typeof(Vector128<sbyte>) };
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size); // value
+                context.EmitLdvec(op.Rn); // value
 
                 if (op.Size == 0)
                 {
@@ -353,7 +353,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesSfl));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -375,7 +375,7 @@ namespace ChocolArm64.Instructions
                 Type[] typesSve = new Type[] { typeof(long), typeof(long) };
                 Type[] typesSfl = new Type[] { typeof(Vector128<sbyte>), typeof(Vector128<sbyte>) };
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size); // value
+                context.EmitLdvec(op.Rn); // value
 
                 if (op.Size == 0)
                 {
@@ -397,7 +397,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), typesSfl));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {

ChocolArm64/Instructions/InstEmitSimdMove.cs

@@ -59,7 +59,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -108,7 +108,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -138,7 +138,7 @@ namespace ChocolArm64.Instructions
                 Type[] typesShs = new Type[] { typeof(Vector128<byte>), typeof(byte) };
                 Type[] typesOr  = new Type[] { typeof(Vector128<byte>), typeof(Vector128<byte>) };
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, 0);
+                context.EmitLdvec(op.Rn);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -150,7 +150,7 @@ namespace ChocolArm64.Instructions
                 context.EmitLdc_I4(op.Imm4);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesShs));
 
-                EmitLdvecWithUnsignedCast(context, op.Rm, 0);
+                context.EmitLdvec(op.Rm);
 
                 context.EmitLdc_I4((op.RegisterSize == RegisterSize.Simd64 ? 8 : 16) - op.Imm4);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical128BitLane), typesShs));
@@ -164,7 +164,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Or), typesOr));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, 0);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -418,7 +418,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MoveLowToHigh)));
 
-                EmitLdvecWithSignedCast(context, op.Rn, 0); // value
+                context.EmitLdvec(op.Rn); // value
 
                 context.EmitLdc_I8(_masksE0_TrnUzpXtn[op.Size]); // mask
                 context.Emit(OpCodes.Dup); // mask
@@ -492,7 +492,7 @@ namespace ChocolArm64.Instructions
                     ? nameof(Sse2.UnpackLow)
                     : nameof(Sse2.UnpackHigh);
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size); // value
+                context.EmitLdvec(op.Rn); // value
 
                 if (op.Size < 3)
                 {
@@ -504,7 +504,7 @@ namespace ChocolArm64.Instructions
                     context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), GetTypesSflUpk(0)));
                 }
 
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size); // value
+                context.EmitLdvec(op.Rm); // value
 
                 if (op.Size < 3)
                 {
@@ -518,7 +518,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(op.Size)));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -560,7 +560,7 @@ namespace ChocolArm64.Instructions
 
                 if (op.RegisterSize == RegisterSize.Simd128)
                 {
-                    EmitLdvecWithSignedCast(context, op.Rn, op.Size); // value
+                    context.EmitLdvec(op.Rn); // value
 
                     if (op.Size < 3)
                     {
@@ -572,7 +572,7 @@ namespace ChocolArm64.Instructions
                         context.EmitCall(typeof(Ssse3).GetMethod(nameof(Ssse3.Shuffle), GetTypesSflUpk(0)));
                     }
 
-                    EmitLdvecWithSignedCast(context, op.Rm, op.Size); // value
+                    context.EmitLdvec(op.Rm); // value
 
                     if (op.Size < 3)
                     {
@@ -586,12 +586,12 @@ namespace ChocolArm64.Instructions
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(3)));
 
-                    EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                    context.EmitStvec(op.Rd);
                 }
                 else
                 {
-                    EmitLdvecWithSignedCast(context, op.Rn, op.Size);
-                    EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                    context.EmitLdvec(op.Rn);
+                    context.EmitLdvec(op.Rm);
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.UnpackLow), GetTypesSflUpk(op.Size))); // value
 
@@ -609,7 +609,7 @@ namespace ChocolArm64.Instructions
 
                     context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(3)));
 
-                    EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                    context.EmitStvec(op.Rd);
                 }
             }
             else
@@ -648,8 +648,8 @@ namespace ChocolArm64.Instructions
                     ? nameof(Sse2.UnpackLow)
                     : nameof(Sse2.UnpackHigh);
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
-                EmitLdvecWithSignedCast(context, op.Rm, op.Size);
+                context.EmitLdvec(op.Rn);
+                context.EmitLdvec(op.Rm);
 
                 if (op.RegisterSize == RegisterSize.Simd128)
                 {
@@ -663,7 +663,7 @@ namespace ChocolArm64.Instructions
                     context.EmitCall(typeof(Sse2).GetMethod(nameUpk, GetTypesSflUpk(3)));
                 }
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
             }
             else
             {

ChocolArm64/Instructions/InstEmitSimdShift.cs

@@ -42,12 +42,12 @@ namespace ChocolArm64.Instructions
             {
                 Type[] typesSll = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], typeof(byte) };
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(shift);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesSll));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -82,7 +82,7 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSll));
@@ -92,7 +92,7 @@ namespace ChocolArm64.Instructions
                 context.EmitLdc_I4(shift);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesSll));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -280,7 +280,7 @@ namespace ChocolArm64.Instructions
                 int shift = GetImmShr(op);
                 int eSize = 8 << op.Size;
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp();
@@ -298,7 +298,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -329,8 +329,8 @@ namespace ChocolArm64.Instructions
                 int shift = GetImmShr(op);
                 int eSize = 8 << op.Size;
 
-                EmitLdvecWithSignedCast(context, op.Rd, op.Size);
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rn);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp();
@@ -349,7 +349,7 @@ namespace ChocolArm64.Instructions
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -405,7 +405,7 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSll));
@@ -415,7 +415,7 @@ namespace ChocolArm64.Instructions
                 context.EmitLdc_I4(shift);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesSll));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -437,12 +437,12 @@ namespace ChocolArm64.Instructions
             {
                 Type[] typesSra = new Type[] { VectorIntTypesPerSizeLog2[op.Size], typeof(byte) };
 
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(GetImmShr(op));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), typesSra));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -470,15 +470,15 @@ namespace ChocolArm64.Instructions
                 Type[] typesSra = new Type[] { VectorIntTypesPerSizeLog2[op.Size], typeof(byte) };
                 Type[] typesAdd = new Type[] { VectorIntTypesPerSizeLog2[op.Size], VectorIntTypesPerSizeLog2[op.Size] };
 
-                EmitLdvecWithSignedCast(context, op.Rd, op.Size);
-                EmitLdvecWithSignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(GetImmShr(op));
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightArithmetic), typesSra));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithSignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -610,7 +610,7 @@ namespace ChocolArm64.Instructions
                 int shift = GetImmShr(op);
                 int eSize = 8 << op.Size;
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp();
@@ -628,7 +628,7 @@ namespace ChocolArm64.Instructions
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -658,8 +658,8 @@ namespace ChocolArm64.Instructions
                 int shift = GetImmShr(op);
                 int eSize = 8 << op.Size;
 
-                EmitLdvecWithUnsignedCast(context, op.Rd, op.Size);
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rn);
 
                 context.Emit(OpCodes.Dup);
                 context.EmitStvectmp();
@@ -678,7 +678,7 @@ namespace ChocolArm64.Instructions
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -734,7 +734,7 @@ namespace ChocolArm64.Instructions
 
                 int numBytes = op.RegisterSize == RegisterSize.Simd128 ? 8 : 0;
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(numBytes);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical128BitLane), typesSll));
@@ -744,7 +744,7 @@ namespace ChocolArm64.Instructions
                 context.EmitLdc_I4(shift);
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesSll));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size + 1);
+                context.EmitStvec(op.Rd);
             }
             else
             {
@@ -765,12 +765,12 @@ namespace ChocolArm64.Instructions
             {
                 Type[] typesSrl = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], typeof(byte) };
 
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(GetImmShr(op));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrl));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {
@@ -797,15 +797,15 @@ namespace ChocolArm64.Instructions
                 Type[] typesSrl = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], typeof(byte) };
                 Type[] typesAdd = new Type[] { VectorUIntTypesPerSizeLog2[op.Size], VectorUIntTypesPerSizeLog2[op.Size] };
 
-                EmitLdvecWithUnsignedCast(context, op.Rd, op.Size);
-                EmitLdvecWithUnsignedCast(context, op.Rn, op.Size);
+                context.EmitLdvec(op.Rd);
+                context.EmitLdvec(op.Rn);
 
                 context.EmitLdc_I4(GetImmShr(op));
 
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrl));
                 context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Add), typesAdd));
 
-                EmitStvecWithUnsignedCast(context, op.Rd, op.Size);
+                context.EmitStvec(op.Rd);
 
                 if (op.RegisterSize == RegisterSize.Simd64)
                 {

ChocolArm64/Instructions/VectorHelper.cs

@@ -565,203 +565,5 @@ namespace ChocolArm64.Instructions
 
             throw new PlatformNotSupportedException();
         }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<sbyte> VectorSingleToSByte(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, sbyte>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<short> VectorSingleToInt16(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, short>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<int> VectorSingleToInt32(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, int>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<long> VectorSingleToInt64(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, long>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<byte> VectorSingleToByte(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, byte>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<ushort> VectorSingleToUInt16(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, ushort>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<uint> VectorSingleToUInt32(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, uint>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<ulong> VectorSingleToUInt64(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, ulong>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<double> VectorSingleToDouble(Vector128<float> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<float, double>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorSByteToSingle(Vector128<sbyte> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<sbyte, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorInt16ToSingle(Vector128<short> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<short, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorInt32ToSingle(Vector128<int> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<int, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorInt64ToSingle(Vector128<long> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<long, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorByteToSingle(Vector128<byte> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<byte, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorUInt16ToSingle(Vector128<ushort> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<ushort, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorUInt32ToSingle(Vector128<uint> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<uint, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorUInt64ToSingle(Vector128<ulong> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<ulong, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static Vector128<float> VectorDoubleToSingle(Vector128<double> vector)
-        {
-            if (Sse.IsSupported)
-            {
-                return Sse.StaticCast<double, float>(vector);
-            }
-
-            throw new PlatformNotSupportedException();
-        }
     }
 }