3 files changed, 177 insertions, 13 deletions
diff --git a/ARMeilleure/Instructions/InstEmitSimdHelper.cs b/ARMeilleure/Instructions/InstEmitSimdHelper.cs
index 49c9e687..49c17560 100644
--- a/ARMeilleure/Instructions/InstEmitSimdHelper.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdHelper.cs
@@ -243,6 +243,21 @@ namespace ARMeilleure.Instructions
             throw new ArgumentException($"Invalid rounding mode \"{roundMode}\".");
         }
 
+        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);
+        }
+
         public static Operand EmitCountSetBits8(ArmEmitterContext context, Operand op) // "size" is 8 (SIMD&FP Inst.).
         {
             Debug.Assert(op.Type == OperandType.I32 || op.Type == OperandType.I64);
diff --git a/ARMeilleure/Instructions/InstEmitSimdLogical.cs b/ARMeilleure/Instructions/InstEmitSimdLogical.cs
index dbd1a1a0..624ae841 100644
--- a/ARMeilleure/Instructions/InstEmitSimdLogical.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdLogical.cs
@@ -336,20 +336,45 @@ namespace ARMeilleure.Instructions
         {
             OpCodeSimd op = (OpCodeSimd)context.CurrOp;
 
-            Operand res = context.VectorZero();
+            if (Optimizations.UseGfni)
+            {
+                const long bitMatrix =
+                    (0b10000000L << 56) |
+                    (0b01000000L << 48) |
+                    (0b00100000L << 40) |
+                    (0b00010000L << 32) |
+                    (0b00001000L << 24) |
+                    (0b00000100L << 16) |
+                    (0b00000010L <<  8) |
+                    (0b00000001L <<  0);
 
-            int elems = op.RegisterSize == RegisterSize.Simd128 ? 16 : 8;
+                Operand vBitMatrix = X86GetAllElements(context, bitMatrix);
 
-            for (int index = 0; index < elems; index++)
-            {
-                Operand ne = EmitVectorExtractZx(context, op.Rn, index, 0);
+                Operand res = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, GetVec(op.Rn), vBitMatrix, Const(0));
 
-                Operand de = EmitReverseBits8Op(context, ne);
+                if (op.RegisterSize == RegisterSize.Simd64)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
 
-                res = EmitVectorInsert(context, res, de, index, 0);
+                context.Copy(GetVec(op.Rd), res);
             }
+            else
+            {
+                Operand res = context.VectorZero();
+                int elems = op.RegisterSize == RegisterSize.Simd128 ? 16 : 8;
 
-            context.Copy(GetVec(op.Rd), res);
+                for (int index = 0; index < elems; index++)
+                {
+                    Operand ne = EmitVectorExtractZx(context, op.Rn, index, 0);
+
+                    Operand de = EmitReverseBits8Op(context, ne);
+
+                    res = EmitVectorInsert(context, res, de, index, 0);
+                }
+
+                context.Copy(GetVec(op.Rd), res);
+            }
         }
 
         private static Operand EmitReverseBits8Op(ArmEmitterContext context, Operand op)
diff --git a/ARMeilleure/Instructions/InstEmitSimdShift.cs b/ARMeilleure/Instructions/InstEmitSimdShift.cs
index 146aeafa..cf3b51bd 100644
--- a/ARMeilleure/Instructions/InstEmitSimdShift.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdShift.cs
@@ -88,8 +88,35 @@ namespace ARMeilleure.Instructions
             OpCodeSimdShImm op = (OpCodeSimdShImm)context.CurrOp;
 
             int shift = GetImmShl(op);
+            int eSize = 8 << op.Size;
 
-            if (Optimizations.UseSse2 && op.Size > 0)
+            if (shift >= eSize)
+            {
+                if ((op.RegisterSize == RegisterSize.Simd64))
+                {
+                    Operand res = context.VectorZeroUpper64(GetVec(op.Rd));
+
+                    context.Copy(GetVec(op.Rd), res);
+                }
+            }
+            else if (Optimizations.UseGfni && op.Size == 0)
+            {
+                Operand n = GetVec(op.Rn);
+
+                ulong bitMatrix = X86GetGf2p8LogicalShiftLeft(shift);
+
+                Operand vBitMatrix = X86GetElements(context, bitMatrix, bitMatrix);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, n, vBitMatrix, Const(0));
+
+                if (op.RegisterSize == RegisterSize.Simd64)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(GetVec(op.Rd), res);
+            }
+            else if (Optimizations.UseSse2 && op.Size > 0)
             {
                 Operand n = GetVec(op.Rn);
 
@@ -396,10 +423,40 @@ namespace ARMeilleure.Instructions
         {
             OpCodeSimdShImm op = (OpCodeSimdShImm)context.CurrOp;
 
-            if (Optimizations.UseSse2 && op.Size > 0 && op.Size < 3)
+            int shift = GetImmShr(op);
+
+            if (Optimizations.UseGfni && op.Size == 0)
             {
-                int shift = GetImmShr(op);
+                Operand n = GetVec(op.Rn);
 
+                ulong bitMatrix;
+
+                if (shift < 8)
+                {
+                    bitMatrix = X86GetGf2p8LogicalShiftLeft(-shift);
+
+                    // Extend sign-bit
+                    bitMatrix |= 0x8080808080808080UL >> (64 - shift * 8);
+                }
+                else
+                {
+                    // Replicate sign-bit into all bits
+                    bitMatrix = 0x8080808080808080UL;
+                }
+
+                Operand vBitMatrix = X86GetElements(context, bitMatrix, bitMatrix);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, n, vBitMatrix, Const(0));
+
+                if (op.RegisterSize == RegisterSize.Simd64)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(GetVec(op.Rd), res);
+            }
+            else if (Optimizations.UseSse2 && op.Size > 0 && op.Size < 3)
+            {
                 Operand n = GetVec(op.Rn);
 
                 Intrinsic sraInst = X86PsraInstruction[op.Size];
@@ -929,10 +986,44 @@ namespace ARMeilleure.Instructions
             OpCodeSimdShImm op = (OpCodeSimdShImm)context.CurrOp;
 
             int shift = GetImmShl(op);
+            int eSize = 8 << op.Size;
 
             ulong mask = shift != 0 ? ulong.MaxValue >> (64 - shift) : 0UL;
 
-            if (Optimizations.UseSse2 && op.Size > 0)
+            if (shift >= eSize)
+            {
+                if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
+                {
+                    Operand res = context.VectorZeroUpper64(GetVec(op.Rd));
+
+                    context.Copy(GetVec(op.Rd), res);
+                }
+            }
+            else if (Optimizations.UseGfni && op.Size == 0)
+            {
+                Operand d = GetVec(op.Rd);
+                Operand n = GetVec(op.Rn);
+
+                ulong bitMatrix = X86GetGf2p8LogicalShiftLeft(shift);
+
+                Operand vBitMatrix = X86GetElements(context, bitMatrix, bitMatrix);
+
+                Operand nShifted = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, n, vBitMatrix, Const(0));
+
+                Operand dMask = X86GetAllElements(context, (long)mask * _masks_SliSri[op.Size]);
+
+                Operand dMasked = context.AddIntrinsic(Intrinsic.X86Pand, d, dMask);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Por, nShifted, dMasked);
+
+                if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(d, res);
+            }
+            else if (Optimizations.UseSse2 && op.Size > 0)
             {
                 Operand d = GetVec(op.Rd);
                 Operand n = GetVec(op.Rn);
@@ -988,7 +1079,40 @@ namespace ARMeilleure.Instructions
 
             ulong mask = (ulong.MaxValue << (eSize - shift)) & (ulong.MaxValue >> (64 - eSize));
 
-            if (Optimizations.UseSse2 && op.Size > 0)
+            if (shift >= eSize)
+            {
+                if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
+                {
+                    Operand res = context.VectorZeroUpper64(GetVec(op.Rd));
+
+                    context.Copy(GetVec(op.Rd), res);
+                }
+            }
+            else if (Optimizations.UseGfni && op.Size == 0)
+            {
+                Operand d = GetVec(op.Rd);
+                Operand n = GetVec(op.Rn);
+
+                ulong bitMatrix = X86GetGf2p8LogicalShiftLeft(-shift);
+
+                Operand vBitMatrix = X86GetElements(context, bitMatrix, bitMatrix);
+
+                Operand nShifted = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, n, vBitMatrix, Const(0));
+
+                Operand dMask = X86GetAllElements(context, (long)mask * _masks_SliSri[op.Size]);
+
+                Operand dMasked = context.AddIntrinsic(Intrinsic.X86Pand, d, dMask);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Por, nShifted, dMasked);
+
+                if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(d, res);
+            }
+            else if (Optimizations.UseSse2 && op.Size > 0)
             {
                 Operand d = GetVec(op.Rd);
                 Operand n = GetVec(op.Rn);