A64: Add fast path for Fcvtas_Gp/S/V, Fcvtau_Gp/S/V and Frinta_S/V in… (#3712)

* A64: Add fast path for Fcvtas_Gp/S/V, Fcvtau_Gp/S/V and Frinta_S/V instructions;

they use "Round to Nearest with Ties to Away" rounding mode not supported in x86.

All instructions involved have been tested locally in both release and debug modes, in both lowcq and highcq.

The titles Mario Strikers and Super Smash Bros. U. use these instructions intensively.

* Update Ptc.cs

* A32: Add fast path for Vcvta_RM, Vrinta_RM and Vrinta_V instructions aswell.

ARMeilleure/Instructions/InstEmitSimdArithmetic.cs

@@ -1617,18 +1617,32 @@ namespace ARMeilleure.Instructions
 
         public static void Frinta_S(ArmEmitterContext context)
         {
-            EmitScalarUnaryOpF(context, (op1) =>
+            if (Optimizations.UseSse41)
             {
-                return EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1);
-            });
+                EmitSse41ScalarRoundOpF(context, FPRoundingMode.ToNearestAway);
+            }
+            else
+            {
+                EmitScalarUnaryOpF(context, (op1) =>
+                {
+                    return EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1);
+                });
+            }
         }
 
         public static void Frinta_V(ArmEmitterContext context)
         {
-            EmitVectorUnaryOpF(context, (op1) =>
+            if (Optimizations.UseSse41)
             {
-                return EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1);
-            });
+                EmitSse41VectorRoundOpF(context, FPRoundingMode.ToNearestAway);
+            }
+            else
+            {
+                EmitVectorUnaryOpF(context, (op1) =>
+                {
+                    return EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1);
+                });
+            }
         }
 
         public static void Frinti_S(ArmEmitterContext context)
@@ -3516,9 +3530,18 @@ namespace ARMeilleure.Instructions
 
             Operand n = GetVec(op.Rn);
 
-            Intrinsic inst = (op.Size & 1) != 0 ? Intrinsic.X86Roundsd : Intrinsic.X86Roundss;
+            Operand res;
 
-            Operand res = context.AddIntrinsic(inst, n, Const(X86GetRoundControl(roundMode)));
+            if (roundMode != FPRoundingMode.ToNearestAway)
+            {
+                Intrinsic inst = (op.Size & 1) != 0 ? Intrinsic.X86Roundsd : Intrinsic.X86Roundss;
+
+                res = context.AddIntrinsic(inst, n, Const(X86GetRoundControl(roundMode)));
+            }
+            else
+            {
+                res = EmitSse41RoundToNearestWithTiesToAwayOpF(context, n, scalar: true);
+            }
 
             if ((op.Size & 1) != 0)
             {
@@ -3538,9 +3561,18 @@ namespace ARMeilleure.Instructions
 
             Operand n = GetVec(op.Rn);
 
-            Intrinsic inst = (op.Size & 1) != 0 ? Intrinsic.X86Roundpd : Intrinsic.X86Roundps;
+            Operand res;
+
+            if (roundMode != FPRoundingMode.ToNearestAway)
+            {
+                Intrinsic inst = (op.Size & 1) != 0 ? Intrinsic.X86Roundpd : Intrinsic.X86Roundps;
 
-            Operand res = context.AddIntrinsic(inst, n, Const(X86GetRoundControl(roundMode)));
+                res = context.AddIntrinsic(inst, n, Const(X86GetRoundControl(roundMode)));
+            }
+            else
+            {
+                res = EmitSse41RoundToNearestWithTiesToAwayOpF(context, n, scalar: false);
+            }
 
             if (op.RegisterSize == RegisterSize.Simd64)
             {

ARMeilleure/Instructions/InstEmitSimdCvt.cs

@@ -164,32 +164,74 @@ namespace ARMeilleure.Instructions
 
         public static void Fcvtas_Gp(ArmEmitterContext context)
         {
-            EmitFcvt_s_Gp(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1));
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvts_Gp(context, FPRoundingMode.ToNearestAway, isFixed: false);
+            }
+            else
+            {
+                EmitFcvt_s_Gp(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1));
+            }
         }
 
         public static void Fcvtas_S(ArmEmitterContext context)
         {
-            EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: true, scalar: true);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41FcvtsOpF(context, FPRoundingMode.ToNearestAway, scalar: true);
+            }
+            else
+            {
+                EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: true, scalar: true);
+            }
         }
 
         public static void Fcvtas_V(ArmEmitterContext context)
         {
-            EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: true, scalar: false);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41FcvtsOpF(context, FPRoundingMode.ToNearestAway, scalar: false);
+            }
+            else
+            {
+                EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: true, scalar: false);
+            }
         }
 
         public static void Fcvtau_Gp(ArmEmitterContext context)
         {
-            EmitFcvt_u_Gp(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1));
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvtu_Gp(context, FPRoundingMode.ToNearestAway, isFixed: false);
+            }
+            else
+            {
+                EmitFcvt_u_Gp(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1));
+            }
         }
 
         public static void Fcvtau_S(ArmEmitterContext context)
         {
-            EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: false, scalar: true);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41FcvtuOpF(context, FPRoundingMode.ToNearestAway, scalar: true);
+            }
+            else
+            {
+                EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: false, scalar: true);
+            }
         }
 
         public static void Fcvtau_V(ArmEmitterContext context)
         {
-            EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: false, scalar: false);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41FcvtuOpF(context, FPRoundingMode.ToNearestAway, scalar: false);
+            }
+            else
+            {
+                EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: false, scalar: false);
+            }
         }
 
         public static void Fcvtl_V(ArmEmitterContext context)
@@ -1223,7 +1265,14 @@ namespace ARMeilleure.Instructions
                     nRes = context.AddIntrinsic(Intrinsic.X86Mulps, nRes, fpScaledMask);
                 }
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar);
+                }
 
                 Operand nInt = context.AddIntrinsic(Intrinsic.X86Cvtps2dq, nRes);
 
@@ -1265,7 +1314,14 @@ namespace ARMeilleure.Instructions
                     nRes = context.AddIntrinsic(Intrinsic.X86Mulpd, nRes, fpScaledMask);
                 }
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar);
+                }
 
                 Operand nLong = EmitSse2CvtDoubleToInt64OpF(context, nRes, scalar);
 
@@ -1314,7 +1370,14 @@ namespace ARMeilleure.Instructions
                     nRes = context.AddIntrinsic(Intrinsic.X86Mulps, nRes, fpScaledMask);
                 }
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar);
+                }
 
                 Operand zero = context.VectorZero();
 
@@ -1369,7 +1432,14 @@ namespace ARMeilleure.Instructions
                     nRes = context.AddIntrinsic(Intrinsic.X86Mulpd, nRes, fpScaledMask);
                 }
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar);
+                }
 
                 Operand zero = context.VectorZero();
 
@@ -1424,7 +1494,14 @@ namespace ARMeilleure.Instructions
                     nRes = context.AddIntrinsic(Intrinsic.X86Mulss, nRes, fpScaledMask);
                 }
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }
 
                 Operand nIntOrLong = op.RegisterSize == RegisterSize.Int32
                     ? context.AddIntrinsicInt (Intrinsic.X86Cvtss2si, nRes)
@@ -1464,7 +1541,14 @@ namespace ARMeilleure.Instructions
                     nRes = context.AddIntrinsic(Intrinsic.X86Mulsd, nRes, fpScaledMask);
                 }
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }
 
                 Operand nIntOrLong = op.RegisterSize == RegisterSize.Int32
                     ? context.AddIntrinsicInt (Intrinsic.X86Cvtsd2si, nRes)
@@ -1512,7 +1596,14 @@ namespace ARMeilleure.Instructions
                     nRes = context.AddIntrinsic(Intrinsic.X86Mulss, nRes, fpScaledMask);
                 }
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }
 
                 Operand zero = context.VectorZero();
 
@@ -1567,7 +1658,14 @@ namespace ARMeilleure.Instructions
                     nRes = context.AddIntrinsic(Intrinsic.X86Mulsd, nRes, fpScaledMask);
                 }
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }
 
                 Operand zero = context.VectorZero();
 

ARMeilleure/Instructions/InstEmitSimdCvt32.cs

@@ -203,6 +203,9 @@ namespace ARMeilleure.Instructions
             FPRoundingMode roundMode;
             switch (rm)
             {
+                case 0b00:
+                    roundMode = FPRoundingMode.ToNearestAway;
+                    break;
                 case 0b01:
                     roundMode = FPRoundingMode.ToNearest;
                     break;
@@ -228,7 +231,7 @@ namespace ARMeilleure.Instructions
             bool unsigned = op.Opc == 0;
             int rm = op.Opc2 & 3;
 
-            if (Optimizations.UseSse41 && rm != 0b00)
+            if (Optimizations.UseSse41)
             {
                 EmitSse41ConvertInt32(context, RMToRoundMode(rm), !unsigned);
             }
@@ -267,15 +270,21 @@ namespace ARMeilleure.Instructions
 
             int rm = op.Opc2 & 3;
 
-            if (Optimizations.UseSse2 && rm != 0b00)
+            if (Optimizations.UseSse41)
             {
                 EmitScalarUnaryOpSimd32(context, (m) =>
                 {
-                    Intrinsic inst = (op.Size & 1) == 0 ? Intrinsic.X86Roundss : Intrinsic.X86Roundsd;
-
                     FPRoundingMode roundMode = RMToRoundMode(rm);
 
-                    return context.AddIntrinsic(inst, m, Const(X86GetRoundControl(roundMode)));
+                    if (roundMode != FPRoundingMode.ToNearestAway)
+                    {
+                        Intrinsic inst = (op.Size & 1) == 0 ? Intrinsic.X86Roundss : Intrinsic.X86Roundsd;
+                        return context.AddIntrinsic(inst, m, Const(X86GetRoundControl(roundMode)));
+                    }
+                    else
+                    {
+                        return EmitSse41RoundToNearestWithTiesToAwayOpF(context, m, scalar: true);
+                    }
                 });
             }
             else
@@ -305,7 +314,17 @@ namespace ARMeilleure.Instructions
         // VRINTA (vector).
         public static void Vrinta_V(ArmEmitterContext context)
         {
-            EmitVectorUnaryOpF32(context, (m) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, m));
+            if (Optimizations.UseSse41)
+            {
+                EmitVectorUnaryOpSimd32(context, (m) =>
+                {
+                    return EmitSse41RoundToNearestWithTiesToAwayOpF(context, m, scalar: false);
+                });
+            }
+            else
+            {
+                EmitVectorUnaryOpF32(context, (m) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, m));
+            }
         }
 
         // VRINTM (vector).
@@ -413,7 +432,14 @@ namespace ARMeilleure.Instructions
                 Operand nRes = context.AddIntrinsic(Intrinsic.X86Cmpss, n, n, Const((int)CmpCondition.OrderedQ));
                 nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, n);
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }
 
                 Operand zero = context.VectorZero();
 
@@ -464,7 +490,14 @@ namespace ARMeilleure.Instructions
                 Operand nRes = context.AddIntrinsic(Intrinsic.X86Cmpsd, n, n, Const((int)CmpCondition.OrderedQ));
                 nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, n);
 
-                nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }
 
                 Operand zero = context.VectorZero();
 

ARMeilleure/Instructions/InstEmitSimdHelper.cs

@@ -33,6 +33,14 @@ namespace ARMeilleure.Instructions
         };
 
         public static readonly long ZeroMask = 128L << 56 | 128L << 48 | 128L << 40 | 128L << 32 | 128L << 24 | 128L << 16 | 128L << 8 | 128L << 0;
+
+        public static ulong X86GetGf2p8LogicalShiftLeft(int shift)
+        {
+            ulong identity = (0b00000001UL << 56) | (0b00000010UL << 48) | (0b00000100UL << 40) | (0b00001000UL << 32) |
+                             (0b00010000UL << 24) | (0b00100000UL << 16) | (0b01000000UL <<  8) | (0b10000000UL <<  0);
+
+            return shift >= 0 ? identity >> (shift * 8) : identity << (-shift * 8);
+        }
 #endregion
 
 #region "X86 SSE Intrinsics"
@@ -243,19 +251,44 @@ namespace ARMeilleure.Instructions
             throw new ArgumentException($"Invalid rounding mode \"{roundMode}\".");
         }
 
-        public static ulong X86GetGf2p8LogicalShiftLeft(int shift)
+        public static Operand EmitSse41RoundToNearestWithTiesToAwayOpF(ArmEmitterContext context, Operand n, bool scalar)
         {
-            ulong identity = 
-                (0b00000001UL << 56) |
-                (0b00000010UL << 48) |
-                (0b00000100UL << 40) |
-                (0b00001000UL << 32) |
-                (0b00010000UL << 24) |
-                (0b00100000UL << 16) |
-                (0b01000000UL <<  8) |
-                (0b10000000UL <<  0);
+            Debug.Assert(n.Type == OperandType.V128);
 
-            return shift >= 0 ? identity >> (shift * 8) : identity << (-shift * 8);
+            Operand nCopy = context.Copy(n);
+
+            Operand rC = Const(X86GetRoundControl(FPRoundingMode.TowardsZero));
+
+            IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
+
+            if ((op.Size & 1) == 0)
+            {
+                Operand signMask = scalar ? X86GetScalar(context, int.MinValue) : X86GetAllElements(context, int.MinValue);
+                        signMask = context.AddIntrinsic(Intrinsic.X86Pand, signMask, nCopy);
+
+                // 0x3EFFFFFF == BitConverter.SingleToInt32Bits(0.5f) - 1
+                Operand valueMask = scalar ? X86GetScalar(context, 0x3EFFFFFF) : X86GetAllElements(context, 0x3EFFFFFF);
+                        valueMask = context.AddIntrinsic(Intrinsic.X86Por, valueMask, signMask);
+
+                nCopy = context.AddIntrinsic(scalar ? Intrinsic.X86Addss : Intrinsic.X86Addps, nCopy, valueMask);
+
+                nCopy = context.AddIntrinsic(scalar ? Intrinsic.X86Roundss : Intrinsic.X86Roundps, nCopy, rC);
+            }
+            else
+            {
+                Operand signMask = scalar ? X86GetScalar(context, long.MinValue) : X86GetAllElements(context, long.MinValue);
+                        signMask = context.AddIntrinsic(Intrinsic.X86Pand, signMask, nCopy);
+
+                // 0x3FDFFFFFFFFFFFFFL == BitConverter.DoubleToInt64Bits(0.5d) - 1L
+                Operand valueMask = scalar ? X86GetScalar(context, 0x3FDFFFFFFFFFFFFFL) : X86GetAllElements(context, 0x3FDFFFFFFFFFFFFFL);
+                        valueMask = context.AddIntrinsic(Intrinsic.X86Por, valueMask, signMask);
+
+                nCopy = context.AddIntrinsic(scalar ? Intrinsic.X86Addsd : Intrinsic.X86Addpd, nCopy, valueMask);
+
+                nCopy = context.AddIntrinsic(scalar ? Intrinsic.X86Roundsd : Intrinsic.X86Roundpd, nCopy, rC);
+            }
+
+            return nCopy;
         }
 
         public static Operand EmitCountSetBits8(ArmEmitterContext context, Operand op) // "size" is 8 (SIMD&FP Inst.).

ARMeilleure/State/FPRoundingMode.cs

@@ -2,9 +2,10 @@ namespace ARMeilleure.State
 {
     public enum FPRoundingMode
     {
-        ToNearest            = 0,
+        ToNearest            = 0, // With ties to even.
         TowardsPlusInfinity  = 1,
         TowardsMinusInfinity = 2,
-        TowardsZero          = 3
+        TowardsZero          = 3,
+        ToNearestAway        = 4 // With ties to away.
     }
 }

ARMeilleure/Translation/PTC/Ptc.cs

@@ -27,7 +27,7 @@ namespace ARMeilleure.Translation.PTC
         private const string OuterHeaderMagicString = "PTCohd\0\0";
         private const string InnerHeaderMagicString = "PTCihd\0\0";
 
-        private const uint InternalVersion = 3710; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 3713; //! To be incremented manually for each change to the ARMeilleure project.
 
         private const string ActualDir = "0";
         private const string BackupDir = "1";