Add 9+7 fast/slow FP inst. impls.; add 14 FP Tests. (#437)

* Update CpuTest.cs

* Delete CpuTestSimdCmp.cs

Obsolete.

* Update CpuTestSimdArithmetic.cs

Superseded.

* Update CpuTestSimd.cs

* Update CpuTestSimdReg.cs

* Update AInstEmitSimdArithmetic.cs

* Update AInstEmitSimdHelper.cs

* Update ASoftFloat.cs

* Nit.

* Update AOpCodeTable.cs

* Update AOptimizations.cs

* Update AInstEmitSimdArithmetic.cs

* Update ASoftFloat.cs

* Update CpuTest.cs

* Update CpuTestSimd.cs

* Update CpuTestSimdReg.cs

* Update AOpCodeTable.cs

* Update AInstEmitSimdArithmetic.cs

* Update ASoftFloat.cs

* Update CpuTestSimdReg.cs

* Update AOpCodeTable.cs

* Update AInstEmitSimdArithmetic.cs

* Update ASoftFloat.cs

* Update CpuTestSimd.cs

* Update CpuTestSimdReg.cs

AOpCodeTable.cs

@@ -301,6 +301,8 @@ namespace ChocolArm64
             SetA64("010111111<<xxxxx1001x0xxxxxxxxxx", AInstEmit.Fmul_Se,       typeof(AOpCodeSimdRegElemF));
             SetA64("0>1011100<1xxxxx110111xxxxxxxxxx", AInstEmit.Fmul_V,        typeof(AOpCodeSimdReg));
             SetA64("0x0011111<<xxxxx1001x0xxxxxxxxxx", AInstEmit.Fmul_Ve,       typeof(AOpCodeSimdRegElemF));
+            SetA64("010111100x1xxxxx110111xxxxxxxxxx", AInstEmit.Fmulx_S,       typeof(AOpCodeSimdReg));
+            SetA64("0>0011100<1xxxxx110111xxxxxxxxxx", AInstEmit.Fmulx_V,       typeof(AOpCodeSimdReg));
             SetA64("000111100x100001010000xxxxxxxxxx", AInstEmit.Fneg_S,        typeof(AOpCodeSimd));
             SetA64("0>1011101<100000111110xxxxxxxxxx", AInstEmit.Fneg_V,        typeof(AOpCodeSimd));
             SetA64("000111110x1xxxxx0xxxxxxxxxxxxxxx", AInstEmit.Fnmadd_S,      typeof(AOpCodeSimdReg));
@@ -310,6 +312,7 @@ namespace ChocolArm64
             SetA64("0>0011101<100001110110xxxxxxxxxx", AInstEmit.Frecpe_V,      typeof(AOpCodeSimd));
             SetA64("010111100x1xxxxx111111xxxxxxxxxx", AInstEmit.Frecps_S,      typeof(AOpCodeSimdReg));
             SetA64("0>0011100<1xxxxx111111xxxxxxxxxx", AInstEmit.Frecps_V,      typeof(AOpCodeSimdReg));
+            SetA64("010111101x100001111110xxxxxxxxxx", AInstEmit.Frecpx_S,      typeof(AOpCodeSimd));
             SetA64("000111100x100110010000xxxxxxxxxx", AInstEmit.Frinta_S,      typeof(AOpCodeSimd));
             SetA64("0>1011100<100001100010xxxxxxxxxx", AInstEmit.Frinta_V,      typeof(AOpCodeSimd));
             SetA64("000111100x100111110000xxxxxxxxxx", AInstEmit.Frinti_S,      typeof(AOpCodeSimd));
@@ -327,6 +330,7 @@ namespace ChocolArm64
             SetA64("010111101x1xxxxx111111xxxxxxxxxx", AInstEmit.Frsqrts_S,     typeof(AOpCodeSimdReg));
             SetA64("0>0011101<1xxxxx111111xxxxxxxxxx", AInstEmit.Frsqrts_V,     typeof(AOpCodeSimdReg));
             SetA64("000111100x100001110000xxxxxxxxxx", AInstEmit.Fsqrt_S,       typeof(AOpCodeSimd));
+            SetA64("0>1011101<100001111110xxxxxxxxxx", AInstEmit.Fsqrt_V,       typeof(AOpCodeSimd));
             SetA64("000111100x1xxxxx001110xxxxxxxxxx", AInstEmit.Fsub_S,        typeof(AOpCodeSimdReg));
             SetA64("0>0011101<1xxxxx110101xxxxxxxxxx", AInstEmit.Fsub_V,        typeof(AOpCodeSimdReg));
             SetA64("01001110000xxxxx000111xxxxxxxxxx", AInstEmit.Ins_Gp,        typeof(AOpCodeSimdIns));

AOptimizations.cs

@@ -2,6 +2,8 @@ using System.Runtime.Intrinsics.X86;
 
 public static class AOptimizations
 {
+    internal static bool FastFP = true;
+
     private static bool UseAllSseIfAvailable = true;
 
     private static bool UseSseIfAvailable   = true;

Instruction/AInstEmitSimdArithmetic.cs

@@ -174,25 +174,33 @@ namespace ChocolArm64.Instruction
 
         public static void Fadd_S(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse && AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
                 EmitScalarSseOrSse2OpF(Context, nameof(Sse.AddScalar));
             }
             else
             {
-                EmitScalarBinaryOpF(Context, () => Context.Emit(OpCodes.Add));
+                EmitScalarBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPAdd));
+                });
             }
         }
 
         public static void Fadd_V(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse && AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
                 EmitVectorSseOrSse2OpF(Context, nameof(Sse.Add));
             }
             else
             {
-                EmitVectorBinaryOpF(Context, () => Context.Emit(OpCodes.Add));
+                EmitVectorBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPAdd));
+                });
             }
         }
 
@@ -217,42 +225,50 @@ namespace ChocolArm64.Instruction
 
         public static void Fdiv_S(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse && AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
                 EmitScalarSseOrSse2OpF(Context, nameof(Sse.DivideScalar));
             }
             else
             {
-                EmitScalarBinaryOpF(Context, () => Context.Emit(OpCodes.Div));
+                EmitScalarBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPDiv));
+                });
             }
         }
 
         public static void Fdiv_V(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse && AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
                 EmitVectorSseOrSse2OpF(Context, nameof(Sse.Divide));
             }
             else
             {
-                EmitVectorBinaryOpF(Context, () => Context.Emit(OpCodes.Div));
+                EmitVectorBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPDiv));
+                });
             }
         }
 
         public static void Fmadd_S(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse2)
             {
                 AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
 
                 if (Op.Size == 0)
                 {
+                    Type[] Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
+
                     Context.EmitLdvec(Op.Ra);
                     Context.EmitLdvec(Op.Rn);
                     Context.EmitLdvec(Op.Rm);
 
-                    Type[] Types = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
-
                     Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), Types));
                     Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.AddScalar),      Types));
 
@@ -262,12 +278,12 @@ namespace ChocolArm64.Instruction
                 }
                 else /* if (Op.Size == 1) */
                 {
+                    Type[] Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
+
                     EmitLdvecWithCastToDouble(Context, Op.Ra);
                     EmitLdvecWithCastToDouble(Context, Op.Rn);
                     EmitLdvecWithCastToDouble(Context, Op.Rm);
 
-                    Type[] Types = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
-
                     Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), Types));
                     Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.AddScalar),      Types));
 
@@ -280,33 +296,48 @@ namespace ChocolArm64.Instruction
             {
                 EmitScalarTernaryRaOpF(Context, () =>
                 {
-                    Context.Emit(OpCodes.Mul);
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMulAdd));
-                    Context.Emit(OpCodes.Add);
                 });
             }
         }
 
         public static void Fmax_S(AILEmitterCtx Context)
         {
-            EmitScalarBinaryOpF(Context, () =>
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.Max));
+                EmitScalarSseOrSse2OpF(Context, nameof(Sse.MaxScalar));
-            });
+            }
+            else
+            {
+                EmitScalarBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMax));
+                });
+            }
         }
 
         public static void Fmax_V(AILEmitterCtx Context)
         {
-            EmitVectorBinaryOpF(Context, () =>
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.Max));
+                EmitVectorSseOrSse2OpF(Context, nameof(Sse.Max));
-            });
+            }
+            else
+            {
+                EmitVectorBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMax));
+                });
+            }
         }
 
         public static void Fmaxnm_S(AILEmitterCtx Context)
         {
             EmitScalarBinaryOpF(Context, () =>
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.MaxNum));
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMaxNum));
             });
         }
 
@@ -314,36 +345,55 @@ namespace ChocolArm64.Instruction
         {
             EmitVectorBinaryOpF(Context, () =>
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.MaxNum));
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMaxNum));
             });
         }
 
         public static void Fmaxp_V(AILEmitterCtx Context)
         {
-            EmitVectorPairwiseOpF(Context, () => EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.Max)));
+            EmitVectorPairwiseOpF(Context, () =>
+            {
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMax));
+            });
         }
 
         public static void Fmin_S(AILEmitterCtx Context)
         {
-            EmitScalarBinaryOpF(Context, () =>
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.Min));
+                EmitScalarSseOrSse2OpF(Context, nameof(Sse.MinScalar));
-            });
+            }
+            else
+            {
+                EmitScalarBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMin));
+                });
+            }
         }
 
         public static void Fmin_V(AILEmitterCtx Context)
         {
-            EmitVectorBinaryOpF(Context, () =>
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.Min));
+                EmitVectorSseOrSse2OpF(Context, nameof(Sse.Min));
-            });
+            }
+            else
+            {
+                EmitVectorBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMin));
+                });
+            }
         }
 
         public static void Fminnm_S(AILEmitterCtx Context)
         {
             EmitScalarBinaryOpF(Context, () =>
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.MinNum));
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMinNum));
             });
         }
 
@@ -351,13 +401,16 @@ namespace ChocolArm64.Instruction
         {
             EmitVectorBinaryOpF(Context, () =>
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.MinNum));
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMinNum));
             });
         }
 
         public static void Fminp_V(AILEmitterCtx Context)
         {
-            EmitVectorPairwiseOpF(Context, () => EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.Min)));
+            EmitVectorPairwiseOpF(Context, () =>
+            {
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMin));
+            });
         }
 
         public static void Fmla_Se(AILEmitterCtx Context)
@@ -407,22 +460,63 @@ namespace ChocolArm64.Instruction
 
         public static void Fmsub_S(AILEmitterCtx Context)
         {
-            EmitScalarTernaryRaOpF(Context, () =>
+            if (AOptimizations.FastFP && AOptimizations.UseSse2)
             {
-                Context.Emit(OpCodes.Mul);
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-                Context.Emit(OpCodes.Sub);
+
-            });
+                if (Op.Size == 0)
+                {
+                    Type[] Types = 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.SubtractScalar), Types));
+
+                    Context.EmitStvec(Op.Rd);
+
+                    EmitVectorZero32_128(Context, Op.Rd);
+                }
+                else /* if (Op.Size == 1) */
+                {
+                    Type[] Types = 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.SubtractScalar), Types));
+
+                    EmitStvecWithCastFromDouble(Context, Op.Rd);
+
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
+            {
+                EmitScalarTernaryRaOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMulSub));
+                });
+            }
         }
 
         public static void Fmul_S(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse && AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
                 EmitScalarSseOrSse2OpF(Context, nameof(Sse.MultiplyScalar));
             }
             else
             {
-                EmitScalarBinaryOpF(Context, () => Context.Emit(OpCodes.Mul));
+                EmitScalarBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMul));
+                });
             }
         }
 
@@ -433,13 +527,17 @@ namespace ChocolArm64.Instruction
 
         public static void Fmul_V(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse && AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
                 EmitVectorSseOrSse2OpF(Context, nameof(Sse.Multiply));
             }
             else
             {
-                EmitVectorBinaryOpF(Context, () => Context.Emit(OpCodes.Mul));
+                EmitVectorBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMul));
+                });
             }
         }
 
@@ -448,6 +546,22 @@ namespace ChocolArm64.Instruction
             EmitVectorBinaryOpByElemF(Context, () => Context.Emit(OpCodes.Mul));
         }
 
+        public static void Fmulx_S(AILEmitterCtx Context)
+        {
+            EmitScalarBinaryOpF(Context, () =>
+            {
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMulX));
+            });
+        }
+
+        public static void Fmulx_V(AILEmitterCtx Context)
+        {
+            EmitVectorBinaryOpF(Context, () =>
+            {
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPMulX));
+            });
+        }
+
         public static void Fneg_S(AILEmitterCtx Context)
         {
             EmitScalarUnaryOpF(Context, () => Context.Emit(OpCodes.Neg));
@@ -524,17 +638,122 @@ namespace ChocolArm64.Instruction
 
         public static void Frecps_S(AILEmitterCtx Context)
         {
-            EmitScalarBinaryOpF(Context, () =>
+            if (AOptimizations.FastFP && AOptimizations.UseSse2)
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.RecipStep));
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-            });
+
+                int SizeF = Op.Size & 1;
+
+                if (SizeF == 0)
+                {
+                    Type[] Types = new Type[] { typeof(float) };
+
+                    Context.EmitLdc_R4(2f);
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), Types));
+
+                    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.SubtractScalar), Types));
+
+                    Context.EmitStvec(Op.Rd);
+
+                    EmitVectorZero32_128(Context, Op.Rd);
+                }
+                else /* if (SizeF == 1) */
+                {
+                    Type[] Types = new Type[] { typeof(double) };
+
+                    Context.EmitLdc_R8(2d);
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), Types));
+
+                    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.SubtractScalar), Types));
+
+                    EmitStvecWithCastFromDouble(Context, Op.Rd);
+
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
+            {
+                EmitScalarBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPRecipStepFused));
+                });
+            }
         }
 
         public static void Frecps_V(AILEmitterCtx Context)
         {
-            EmitVectorBinaryOpF(Context, () =>
+            if (AOptimizations.FastFP && AOptimizations.UseSse2)
             {
-                EmitBinarySoftFloatCall(Context, nameof(ASoftFloat.RecipStep));
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+                int SizeF = Op.Size & 1;
+
+                if (SizeF == 0)
+                {
+                    Type[] Types = new Type[] { typeof(float) };
+
+                    Context.EmitLdc_R4(2f);
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), Types));
+
+                    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.Subtract), Types));
+
+                    Context.EmitStvec(Op.Rd);
+
+                    if (Op.RegisterSize == ARegisterSize.SIMD64)
+                    {
+                        EmitVectorZeroUpper(Context, Op.Rd);
+                    }
+                }
+                else /* if (SizeF == 1) */
+                {
+                    Type[] Types = new Type[] { typeof(double) };
+
+                    Context.EmitLdc_R8(2d);
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), Types));
+
+                    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.Subtract), Types));
+
+                    EmitStvecWithCastFromDouble(Context, Op.Rd);
+                }
+            }
+            else
+            {
+                EmitVectorBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPRecipStepFused));
+                });
+            }
+        }
+
+        public static void Frecpx_S(AILEmitterCtx Context)
+        {
+            EmitScalarUnaryOpF(Context, () =>
+            {
+                EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPRecpX));
             });
         }
 
@@ -728,97 +947,194 @@ namespace ChocolArm64.Instruction
 
         public static void Frsqrts_S(AILEmitterCtx Context)
         {
-            EmitFrsqrts(Context, 0, Scalar: true);
+            if (AOptimizations.FastFP && AOptimizations.UseSse2)
+            {
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
+
+                int SizeF = Op.Size & 1;
+
+                if (SizeF == 0)
+                {
+                    Type[] Types = new Type[] { typeof(float) };
+
+                    Context.EmitLdc_R4(0.5f);
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), Types));
+
+                    Context.EmitLdc_R4(3f);
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetScalarVector128), Types));
+
+                    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.SubtractScalar), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.MultiplyScalar), Types));
+
+                    Context.EmitStvec(Op.Rd);
+
+                    EmitVectorZero32_128(Context, Op.Rd);
+                }
+                else /* if (SizeF == 1) */
+                {
+                    Type[] Types = new Type[] { typeof(double) };
+
+                    Context.EmitLdc_R8(0.5d);
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), Types));
+
+                    Context.EmitLdc_R8(3d);
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetScalarVector128), Types));
+
+                    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.SubtractScalar), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.MultiplyScalar), Types));
+
+                    EmitStvecWithCastFromDouble(Context, Op.Rd);
+
+                    EmitVectorZeroUpper(Context, Op.Rd);
+                }
+            }
+            else
+            {
+                EmitScalarBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPRSqrtStepFused));
+                });
+            }
         }
 
         public static void Frsqrts_V(AILEmitterCtx Context)
         {
-            AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
+            if (AOptimizations.FastFP && AOptimizations.UseSse2)
-
-            int SizeF = Op.Size & 1;
-
-            int Bytes = Op.GetBitsCount() >> 3;
-
-            for (int Index = 0; Index < Bytes >> SizeF + 2; Index++)
             {
-                EmitFrsqrts(Context, Index, Scalar: false);
+                AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-            }
 
-            if (Op.RegisterSize == ARegisterSize.SIMD64)
+                int SizeF = Op.Size & 1;
+
+                if (SizeF == 0)
+                {
+                    Type[] Types = new Type[] { typeof(float) };
+
+                    Context.EmitLdc_R4(0.5f);
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), Types));
+
+                    Context.EmitLdc_R4(3f);
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.SetAllVector128), Types));
+
+                    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.Subtract), Types));
+                    Context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), Types));
+
+                    Context.EmitStvec(Op.Rd);
+
+                    if (Op.RegisterSize == ARegisterSize.SIMD64)
+                    {
+                        EmitVectorZeroUpper(Context, Op.Rd);
+                    }
+                }
+                else /* if (SizeF == 1) */
+                {
+                    Type[] Types = new Type[] { typeof(double) };
+
+                    Context.EmitLdc_R8(0.5d);
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), Types));
+
+                    Context.EmitLdc_R8(3d);
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), Types));
+
+                    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.Subtract), Types));
+                    Context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.Multiply), Types));
+
+                    EmitStvecWithCastFromDouble(Context, Op.Rd);
+                }
+            }
+            else
             {
-                EmitVectorZeroUpper(Context, Op.Rd);
+                EmitVectorBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPRSqrtStepFused));
+                });
             }
         }
 
-        private static void EmitFrsqrts(AILEmitterCtx Context, int Index, bool Scalar)
-        {
-            AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
-
-            int SizeF = Op.Size & 1;
-
-            if (SizeF == 0)
-            {
-                Context.EmitLdc_R4(3);
-            }
-            else /* if (SizeF == 1) */
-            {
-                Context.EmitLdc_R8(3);
-            }
-
-            EmitVectorExtractF(Context, Op.Rn, Index, SizeF);
-            EmitVectorExtractF(Context, Op.Rm, Index, SizeF);
-
-            Context.Emit(OpCodes.Mul);
-            Context.Emit(OpCodes.Sub);
-
-            if (SizeF == 0)
-            {
-                Context.EmitLdc_R4(0.5f);
-            }
-            else /* if (SizeF == 1) */
-            {
-                Context.EmitLdc_R8(0.5);
-            }
-
-            Context.Emit(OpCodes.Mul);
-
-            if (Scalar)
-            {
-                EmitVectorZeroAll(Context, Op.Rd);
-            }
-
-            EmitVectorInsertF(Context, Op.Rd, Index, SizeF);
-        }
-
         public static void Fsqrt_S(AILEmitterCtx Context)
         {
-            EmitScalarUnaryOpF(Context, () =>
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
-                EmitUnaryMathCall(Context, nameof(Math.Sqrt));
+                EmitScalarSseOrSse2OpF(Context, nameof(Sse.SqrtScalar));
-            });
+            }
+            else
+            {
+                EmitScalarUnaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPSqrt));
+                });
+            }
+        }
+
+        public static void Fsqrt_V(AILEmitterCtx Context)
+        {
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
+            {
+                EmitVectorSseOrSse2OpF(Context, nameof(Sse.Sqrt));
+            }
+            else
+            {
+                EmitVectorUnaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPSqrt));
+                });
+            }
         }
 
         public static void Fsub_S(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse && AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
                 EmitScalarSseOrSse2OpF(Context, nameof(Sse.SubtractScalar));
             }
             else
             {
-                EmitScalarBinaryOpF(Context, () => Context.Emit(OpCodes.Sub));
+                EmitScalarBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPSub));
+                });
             }
         }
 
         public static void Fsub_V(AILEmitterCtx Context)
         {
-            if (AOptimizations.UseSse && AOptimizations.UseSse2)
+            if (AOptimizations.FastFP && AOptimizations.UseSse
+                                      && AOptimizations.UseSse2)
             {
                 EmitVectorSseOrSse2OpF(Context, nameof(Sse.Subtract));
             }
             else
             {
-                EmitVectorBinaryOpF(Context, () => Context.Emit(OpCodes.Sub));
+                EmitVectorBinaryOpF(Context, () =>
+                {
+                    EmitSoftFloatCall(Context, nameof(ASoftFloat_32.FPSub));
+                });
             }
         }
 
@@ -1170,7 +1486,6 @@ namespace ChocolArm64.Instruction
             EmitVectorTernaryOpZx(Context, () =>
             {
                 Context.Emit(OpCodes.Sub);
-
                 EmitAbs(Context);
 
                 Context.Emit(OpCodes.Add);
@@ -1182,7 +1497,6 @@ namespace ChocolArm64.Instruction
             EmitVectorWidenRnRmTernaryOpZx(Context, () =>
             {
                 Context.Emit(OpCodes.Sub);
-
                 EmitAbs(Context);
 
                 Context.Emit(OpCodes.Add);
@@ -1194,7 +1508,6 @@ namespace ChocolArm64.Instruction
             EmitVectorBinaryOpZx(Context, () =>
             {
                 Context.Emit(OpCodes.Sub);
-
                 EmitAbs(Context);
             });
         }
@@ -1204,7 +1517,6 @@ namespace ChocolArm64.Instruction
             EmitVectorWidenRnRmBinaryOpZx(Context, () =>
             {
                 Context.Emit(OpCodes.Sub);
-
                 EmitAbs(Context);
             });
         }

Instruction/AInstEmitSimdHelper.cs

@@ -306,25 +306,19 @@ namespace ChocolArm64.Instruction
 
             int SizeF = Op.Size & 1;
 
-            Context.EmitLdc_I4((int)RoundMode);
-
             MethodInfo MthdInfo;
 
-            Type[] Types = new Type[] { null, typeof(MidpointRounding) };
-
-            Types[0] = SizeF == 0
-                ? typeof(float)
-                : typeof(double);
-
             if (SizeF == 0)
             {
-                MthdInfo = typeof(MathF).GetMethod(nameof(MathF.Round), Types);
+                MthdInfo = typeof(MathF).GetMethod(nameof(MathF.Round), new Type[] { typeof(float), typeof(MidpointRounding) });
             }
             else /* if (SizeF == 1) */
             {
-                MthdInfo = typeof(Math).GetMethod(nameof(Math.Round), Types);
+                MthdInfo = typeof(Math).GetMethod(nameof(Math.Round), new Type[] { typeof(double), typeof(MidpointRounding) });
             }
 
+            Context.EmitLdc_I4((int)RoundMode);
+
             Context.EmitCall(MthdInfo);
         }
 
@@ -348,24 +342,17 @@ namespace ChocolArm64.Instruction
             Context.EmitCall(MthdInfo);
         }
 
-        public static void EmitBinarySoftFloatCall(AILEmitterCtx Context, string Name)
+        public static void EmitSoftFloatCall(AILEmitterCtx Context, string Name)
         {
             IAOpCodeSimd Op = (IAOpCodeSimd)Context.CurrOp;
 
-            int SizeF = Op.Size & 1;
+            Type Type = (Op.Size & 1) == 0
+                ? typeof(ASoftFloat_32)
+                : typeof(ASoftFloat_64);
 
-            MethodInfo MthdInfo;
+            Context.EmitLdarg(ATranslatedSub.StateArgIdx);
 
-            if (SizeF == 0)
+            Context.EmitCall(Type, Name);
-            {
-                MthdInfo = typeof(ASoftFloat).GetMethod(Name, new Type[] { typeof(float), typeof(float) });
-            }
-            else /* if (SizeF == 1) */
-            {
-                MthdInfo = typeof(ASoftFloat).GetMethod(Name, new Type[] { typeof(double), typeof(double) });
-            }
-
-            Context.EmitCall(MthdInfo);
         }
 
         public static void EmitScalarBinaryOpByElemF(AILEmitterCtx Context, Action Emit)