Improve shader global memory to storage pass (#2200)

* Improve shader global memory to storage pass

* Formatting and more comments

* Shader cache version bump

Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

@@ -35,7 +35,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <summary>
         /// Version of the codegen (to be changed when codegen or guest format change).
         /// </summary>
-        private const ulong ShaderCodeGenVersion = 2163;
+        private const ulong ShaderCodeGenVersion = 2200;
 
         // Progress reporting helpers
         private volatile int _shaderCount;

Ryujinx.Graphics.Shader/Translation/Optimizations/BindlessElimination.cs

@@ -5,66 +5,6 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
 {
     class BindlessElimination
     {
-        private static Operation FindBranchSource(BasicBlock block)
-        {
-            foreach (BasicBlock sourceBlock in block.Predecessors)
-            {
-                if (sourceBlock.Operations.Count > 0)
-                {
-                    Operation lastOp = sourceBlock.Operations.Last.Value as Operation;
-
-                    if (lastOp != null &&
-                        ((sourceBlock.Next == block && lastOp.Inst == Instruction.BranchIfFalse) ||
-                        (sourceBlock.Branch == block && lastOp.Inst == Instruction.BranchIfTrue)))
-                    {
-                        return lastOp;
-                    }
-                }
-            }
-
-            return null;
-        }
-
-        private static bool BlockConditionsMatch(BasicBlock currentBlock, BasicBlock queryBlock)
-        {
-            // Check if all the conditions for the query block are satisfied by the current block.
-            // Just checks the top-most conditional for now.
-
-            Operation currentBranch = FindBranchSource(currentBlock);
-            Operation queryBranch = FindBranchSource(queryBlock);
-
-            Operand currentCondition = currentBranch?.GetSource(0);
-            Operand queryCondition = queryBranch?.GetSource(0);
-
-            // The condition should be the same operand instance.
-
-            return currentBranch != null && queryBranch != null &&
-                   currentBranch.Inst == queryBranch.Inst &&
-                   currentCondition == queryCondition;
-        }
-
-        private static Operand FindLastOperation(Operand source, BasicBlock block)
-        {
-            if (source.AsgOp is PhiNode phiNode)
-            {
-                // This source can have a different value depending on a previous branch.
-                // Ensure that conditions met for that branch are also met for the current one.
-                // Prefer the latest sources for the phi node.
-
-                for (int i = phiNode.SourcesCount - 1; i >= 0; i--)
-                {
-                    BasicBlock phiBlock = phiNode.GetBlock(i);
-
-                    if (BlockConditionsMatch(block, phiBlock))
-                    {
-                        return phiNode.GetSource(i);
-                    }
-                }
-            }
-
-            return source;
-        }
-
         public static void RunPass(BasicBlock block, ShaderConfig config)
         {
             // We can turn a bindless into regular access by recognizing the pattern
@@ -89,7 +29,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                     texOp.Inst == Instruction.TextureSample ||
                     texOp.Inst == Instruction.TextureSize)
                 {
-                    Operand bindlessHandle = FindLastOperation(texOp.GetSource(0), block);
+                    Operand bindlessHandle = Utils.FindLastOperation(texOp.GetSource(0), block);
 
                     if (bindlessHandle.Type == OperandType.ConstantBuffer)
                     {
@@ -107,8 +47,8 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                         continue;
                     }
 
-                    Operand src0 = FindLastOperation(handleCombineOp.GetSource(0), block);
-                    Operand src1 = FindLastOperation(handleCombineOp.GetSource(1), block);
+                    Operand src0 = Utils.FindLastOperation(handleCombineOp.GetSource(0), block);
+                    Operand src1 = Utils.FindLastOperation(handleCombineOp.GetSource(1), block);
 
                     if (src0.Type != OperandType.ConstantBuffer ||
                         src1.Type != OperandType.ConstantBuffer || src0.GetCbufSlot() != src1.GetCbufSlot())
@@ -120,7 +60,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                 }
                 else if (texOp.Inst == Instruction.ImageLoad || texOp.Inst == Instruction.ImageStore)
                 {
-                    Operand src0 = FindLastOperation(texOp.GetSource(0), block);
+                    Operand src0 = Utils.FindLastOperation(texOp.GetSource(0), block);
 
                     if (src0.Type == OperandType.ConstantBuffer)
                     {

Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs

@@ -25,32 +25,29 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                 {
                     Operand source = operation.GetSource(0);
 
-                    if (source.AsgOp is Operation asgOperation)
+                    int storageIndex = SearchForStorageBase(block, source, sbStart, sbEnd);
+
+                    if (storageIndex >= 0)
+                    {
+                        // Storage buffers are implemented using global memory access.
+                        // If we know from where the base address of the access is loaded,
+                        // we can guess which storage buffer it is accessing.
+                        // We can then replace the global memory access with a storage
+                        // buffer access.
+                        node = ReplaceGlobalWithStorage(node, config, storageIndex);
+                    }
+                    else if (config.Stage == ShaderStage.Compute && operation.Inst == Instruction.LoadGlobal)
                     {
-                        int storageIndex = SearchForStorageBase(asgOperation, sbStart, sbEnd);
+                        // Here we effectively try to replace a LDG instruction with LDC.
+                        // The hardware only supports a limited amount of constant buffers
+                        // so NVN "emulates" more constant buffers using global memory access.
+                        // Here we try to replace the global access back to a constant buffer
+                        // load.
+                        storageIndex = SearchForStorageBase(block, source, UbeBaseOffset, UbeBaseOffset + UbeDescsSize);
 
                         if (storageIndex >= 0)
                         {
-                            // Storage buffers are implemented using global memory access.
-                            // If we know from where the base address of the access is loaded,
-                            // we can guess which storage buffer it is accessing.
-                            // We can then replace the global memory access with a storage
-                            // buffer access.
-                            node = ReplaceGlobalWithStorage(node, config, storageIndex);
-                        }
-                        else if (config.Stage == ShaderStage.Compute && operation.Inst == Instruction.LoadGlobal)
-                        {
-                            // Here we effectively try to replace a LDG instruction with LDC.
-                            // The hardware only supports a limited amount of constant buffers
-                            // so NVN "emulates" more constant buffers using global memory access.
-                            // Here we try to replace the global access back to a constant buffer
-                            // load.
-                            storageIndex = SearchForStorageBase(asgOperation, UbeBaseOffset, UbeBaseOffset + UbeDescsSize);
-
-                            if (storageIndex >= 0)
-                            {
-                                node = ReplaceLdgWithLdc(node, config, storageIndex);
-                            }
+                            node = ReplaceLdgWithLdc(node, config, storageIndex);
                         }
                     }
                 }
@@ -184,35 +181,70 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
             return node;
         }
 
-        private static int SearchForStorageBase(Operation operation, int sbStart, int sbEnd)
+        private static int SearchForStorageBase(BasicBlock block, Operand globalAddress, int sbStart, int sbEnd)
         {
-            Queue<Operation> assignments = new Queue<Operation>();
+            globalAddress = Utils.FindLastOperation(globalAddress, block);
 
-            assignments.Enqueue(operation);
+            if (globalAddress.Type == OperandType.ConstantBuffer)
+            {
+                return GetStorageIndex(globalAddress, sbStart, sbEnd);
+            }
+
+            Operation operation = globalAddress.AsgOp as Operation;
 
-            while (assignments.TryDequeue(out operation))
+            if (operation == null || operation.Inst != Instruction.Add)
             {
-                for (int index = 0; index < operation.SourcesCount; index++)
+                return -1;
+            }
+
+            Operand src1 = operation.GetSource(0);
+            Operand src2 = operation.GetSource(1);
+
+            if ((src1.Type == OperandType.LocalVariable && src2.Type == OperandType.Constant) ||
+                (src2.Type == OperandType.LocalVariable && src1.Type == OperandType.Constant))
+            {
+                if (src1.Type == OperandType.LocalVariable)
                 {
-                    Operand source = operation.GetSource(index);
+                    operation = Utils.FindLastOperation(src1, block).AsgOp as Operation;
+                }
+                else
+                {
+                    operation = Utils.FindLastOperation(src2, block).AsgOp as Operation;
+                }
 
-                    if (source.Type == OperandType.ConstantBuffer)
-                    {
-                        int slot   = source.GetCbufSlot();
-                        int offset = source.GetCbufOffset();
+                if (operation == null || operation.Inst != Instruction.Add)
+                {
+                    return -1;
+                }
+            }
 
-                        if (slot == 0 && offset >= sbStart && offset < sbEnd)
-                        {
-                            int storageIndex = (offset - sbStart) / StorageDescSize;
+            for (int index = 0; index < operation.SourcesCount; index++)
+            {
+                Operand source = operation.GetSource(index);
 
-                            return storageIndex;
-                        }
-                    }
+                int storageIndex = GetStorageIndex(source, sbStart, sbEnd);
 
-                    if (source.AsgOp is Operation asgOperation)
-                    {
-                        assignments.Enqueue(asgOperation);
-                    }
+                if (storageIndex != -1)
+                {
+                    return storageIndex;
+                }
+            }
+
+            return -1;
+        }
+
+        private static int GetStorageIndex(Operand operand, int sbStart, int sbEnd)
+        {
+            if (operand.Type == OperandType.ConstantBuffer)
+            {
+                int slot   = operand.GetCbufSlot();
+                int offset = operand.GetCbufOffset();
+
+                if (slot == 0 && offset >= sbStart && offset < sbEnd)
+                {
+                    int storageIndex = (offset - sbStart) / StorageDescSize;
+
+                    return storageIndex;
                 }
             }
 

Ryujinx.Graphics.Shader/Translation/Optimizations/Optimizer.cs

@@ -10,11 +10,22 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
     {
         public static void RunPass(BasicBlock[] blocks, ShaderConfig config)
         {
+            RunOptimizationPasses(blocks);
+
+            // Those passes are looking for specific patterns and only needs to run once.
             for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
             {
                 GlobalToStorage.RunPass(blocks[blkIndex], config);
+                BindlessToIndexed.RunPass(blocks[blkIndex]);
+                BindlessElimination.RunPass(blocks[blkIndex], config);
             }
 
+            // Run optimizations one last time to remove any code that is now optimizable after above passes.
+            RunOptimizationPasses(blocks);
+        }
+
+        private static void RunOptimizationPasses(BasicBlock[] blocks)
+        {
             bool modified;
 
             do
@@ -85,27 +96,6 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                 }
             }
             while (modified);
-
-            for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
-            {
-                BindlessToIndexed.RunPass(blocks[blkIndex]);
-                BindlessElimination.RunPass(blocks[blkIndex], config);
-
-                // Try to eliminate any operations that are now unused.
-                LinkedListNode<INode> node = blocks[blkIndex].Operations.First;
-
-                while (node != null)
-                {
-                    LinkedListNode<INode> nextNode = node.Next;
-
-                    if (IsUnused(node.Value))
-                    {
-                        RemoveNode(blocks[blkIndex], node);
-                    }
-
-                    node = nextNode;
-                }
-            }
         }
 
         private static void PropagateCopy(Operation copyOp)

Ryujinx.Graphics.Shader/Translation/Optimizations/Utils.cs

@@ -0,0 +1,67 @@
+using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+
+namespace Ryujinx.Graphics.Shader.Translation.Optimizations
+{
+    static class Utils
+    {
+        private static Operation FindBranchSource(BasicBlock block)
+        {
+            foreach (BasicBlock sourceBlock in block.Predecessors)
+            {
+                if (sourceBlock.Operations.Count > 0)
+                {
+                    Operation lastOp = sourceBlock.Operations.Last.Value as Operation;
+
+                    if (lastOp != null &&
+                        ((sourceBlock.Next == block && lastOp.Inst == Instruction.BranchIfFalse) ||
+                        (sourceBlock.Branch == block && lastOp.Inst == Instruction.BranchIfTrue)))
+                    {
+                        return lastOp;
+                    }
+                }
+            }
+
+            return null;
+        }
+
+        private static bool BlockConditionsMatch(BasicBlock currentBlock, BasicBlock queryBlock)
+        {
+            // Check if all the conditions for the query block are satisfied by the current block.
+            // Just checks the top-most conditional for now.
+
+            Operation currentBranch = FindBranchSource(currentBlock);
+            Operation queryBranch = FindBranchSource(queryBlock);
+
+            Operand currentCondition = currentBranch?.GetSource(0);
+            Operand queryCondition = queryBranch?.GetSource(0);
+
+            // The condition should be the same operand instance.
+
+            return currentBranch != null && queryBranch != null &&
+                   currentBranch.Inst == queryBranch.Inst &&
+                   currentCondition == queryCondition;
+        }
+
+        public static Operand FindLastOperation(Operand source, BasicBlock block)
+        {
+            if (source.AsgOp is PhiNode phiNode)
+            {
+                // This source can have a different value depending on a previous branch.
+                // Ensure that conditions met for that branch are also met for the current one.
+                // Prefer the latest sources for the phi node.
+
+                for (int i = phiNode.SourcesCount - 1; i >= 0; i--)
+                {
+                    BasicBlock phiBlock = phiNode.GetBlock(i);
+
+                    if (BlockConditionsMatch(block, phiBlock))
+                    {
+                        return phiNode.GetSource(i);
+                    }
+                }
+            }
+
+            return source;
+        }
+    }
+}