Implement SHF (funnel shift) shader instruction (#2702)

* Implement SHF shader instruction

* Shader cache version bump

* Better name

Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

@@ -40,7 +40,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <summary>
         /// Version of the codegen (to be changed when codegen or guest format change).
         /// </summary>
-        private const ulong ShaderCodeGenVersion = 2697;
+        private const ulong ShaderCodeGenVersion = 2702;
 
         // Progress reporting helpers
         private volatile int _shaderCount;

Ryujinx.Graphics.Shader/Instructions/InstEmit.cs

@@ -418,34 +418,6 @@ namespace Ryujinx.Graphics.Shader.Instructions
             context.Config.GpuAccessor.Log("Shader instruction Setlmembase is not implemented.");
         }
 
-        public static void ShfLR(EmitterContext context)
-        {
-            InstShfLR op = context.GetOp<InstShfLR>();
-
-            context.Config.GpuAccessor.Log("Shader instruction ShfLR is not implemented.");
-        }
-
-        public static void ShfRR(EmitterContext context)
-        {
-            InstShfRR op = context.GetOp<InstShfRR>();
-
-            context.Config.GpuAccessor.Log("Shader instruction ShfRR is not implemented.");
-        }
-
-        public static void ShfLI(EmitterContext context)
-        {
-            InstShfLI op = context.GetOp<InstShfLI>();
-
-            context.Config.GpuAccessor.Log("Shader instruction ShfLI is not implemented.");
-        }
-
-        public static void ShfRI(EmitterContext context)
-        {
-            InstShfRI op = context.GetOp<InstShfRI>();
-
-            context.Config.GpuAccessor.Log("Shader instruction ShfRI is not implemented.");
-        }
-
         public static void St(EmitterContext context)
         {
             InstSt op = context.GetOp<InstSt>();

Ryujinx.Graphics.Shader/Instructions/InstEmitShift.cs

@@ -9,6 +9,50 @@ namespace Ryujinx.Graphics.Shader.Instructions
 {
     static partial class InstEmit
     {
+        public static void ShfLR(EmitterContext context)
+        {
+            InstShfLR op = context.GetOp<InstShfLR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitShf(context, op.MaxShift, srcA, srcB, srcC, op.Dest, op.M, left: true, op.WriteCC);
+        }
+
+        public static void ShfRR(EmitterContext context)
+        {
+            InstShfRR op = context.GetOp<InstShfRR>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = GetSrcReg(context, op.SrcB);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitShf(context, op.MaxShift, srcA, srcB, srcC, op.Dest, op.M, left: false, op.WriteCC);
+        }
+
+        public static void ShfLI(EmitterContext context)
+        {
+            InstShfLI op = context.GetOp<InstShfLI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = Const(op.Imm6);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitShf(context, op.MaxShift, srcA, srcB, srcC, op.Dest, op.M, left: true, op.WriteCC);
+        }
+
+        public static void ShfRI(EmitterContext context)
+        {
+            InstShfRI op = context.GetOp<InstShfRI>();
+
+            var srcA = GetSrcReg(context, op.SrcA);
+            var srcB = Const(op.Imm6);
+            var srcC = GetSrcReg(context, op.SrcC);
+
+            EmitShf(context, op.MaxShift, srcA, srcB, srcC, op.Dest, op.M, left: false, op.WriteCC);
+        }
+
         public static void ShlR(EmitterContext context)
         {
             InstShlR op = context.GetOp<InstShlR>();
@@ -60,6 +104,79 @@ namespace Ryujinx.Graphics.Shader.Instructions
             EmitShr(context, srcA, srcB, op.Dest, op.M, op.Brev, op.Signed);
         }
 
+        private static void EmitShf(
+            EmitterContext context,
+            MaxShift maxShift,
+            Operand srcA,
+            Operand srcB,
+            Operand srcC,
+            int rd,
+            bool mask,
+            bool left,
+            bool writeCC)
+        {
+            bool isLongShift = maxShift == MaxShift.U64 || maxShift == MaxShift.S64;
+            bool signedShift = maxShift == MaxShift.S64;
+            int maxShiftConst = isLongShift ? 64 : 32;
+
+            if (mask)
+            {
+                srcB = context.BitwiseAnd(srcB, Const(maxShiftConst - 1));
+            }
+
+            Operand res;
+
+            if (left)
+            {
+                // res = (C << B) | (A >> (32 - B))
+                res = context.ShiftLeft(srcC, srcB);
+                res = context.BitwiseOr(res, context.ShiftRightU32(srcA, context.ISubtract(Const(32), srcB)));
+
+                if (isLongShift)
+                {
+                    // res = B >= 32 ? A << (B - 32) : res
+                    Operand lowerShift = context.ShiftLeft(srcA, context.ISubtract(srcB, Const(32)));
+
+                    Operand shiftGreaterThan31 = context.ICompareGreaterOrEqualUnsigned(srcB, Const(32));
+                    res = context.ConditionalSelect(shiftGreaterThan31, lowerShift, res);
+                }
+            }
+            else
+            {
+                // res = (A >> B) | (C << (32 - B))
+                res = context.ShiftRightU32(srcA, srcB);
+                res = context.BitwiseOr(res, context.ShiftLeft(srcC, context.ISubtract(Const(32), srcB)));
+
+                if (isLongShift)
+                {
+                    // res = B >= 32 ? C >> (B - 32) : res
+                    Operand upperShift = signedShift
+                        ? context.ShiftRightS32(srcC, context.ISubtract(srcB, Const(32)))
+                        : context.ShiftRightU32(srcC, context.ISubtract(srcB, Const(32)));
+
+                    Operand shiftGreaterThan31 = context.ICompareGreaterOrEqualUnsigned(srcB, Const(32));
+                    res = context.ConditionalSelect(shiftGreaterThan31, upperShift, res);
+                }
+            }
+
+            if (!mask)
+            {
+                // Clamped shift value.
+                Operand isLessThanMax = context.ICompareLessUnsigned(srcB, Const(maxShiftConst));
+
+                res = context.ConditionalSelect(isLessThanMax, res, Const(0));
+            }
+
+            context.Copy(GetDest(rd), res);
+
+            if (writeCC)
+            {
+                InstEmitAluHelper.SetZnFlags(context, res, writeCC);
+            }
+
+            // TODO: X.
+        }
+
         private static void EmitShl(EmitterContext context, Operand srcA, Operand srcB, int rd, bool mask)
         {
             if (mask)