shader_ir: Unify blocks in decompiled shaders.

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -191,10 +191,12 @@ public:
 
         // TODO(Subv): Figure out the actual depth of the flow stack, for now it seems
         // unlikely that shaders will use 20 nested SSYs and PBKs.
-        constexpr u32 FLOW_STACK_SIZE = 20;
-        for (const auto stack : std::array{MetaStackClass::Ssy, MetaStackClass::Pbk}) {
-            code.AddLine("uint {}[{}];", FlowStackName(stack), FLOW_STACK_SIZE);
-            code.AddLine("uint {} = 0u;", FlowStackTopName(stack));
+        if (!ir.IsFlowStackDisabled()) {
+            constexpr u32 FLOW_STACK_SIZE = 20;
+            for (const auto stack : std::array{MetaStackClass::Ssy, MetaStackClass::Pbk}) {
+                code.AddLine("uint {}[{}];", FlowStackName(stack), FLOW_STACK_SIZE);
+                code.AddLine("uint {} = 0u;", FlowStackTopName(stack));
+            }
         }
 
         code.AddLine("while (true) {{");

src/video_core/shader/control_flow.cpp

@@ -1,5 +1,6 @@
 
 #include <list>
+#include <map>
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
@@ -104,28 +105,6 @@ struct BlockInfo {
     }
 };
 
-struct Stamp {
-    Stamp() = default;
-    Stamp(u32 address, u32 target) : address{address}, target{target} {}
-    u32 address{};
-    u32 target{};
-    bool operator==(const Stamp& sb) const {
-        return std::tie(address, target) == std::tie(sb.address, sb.target);
-    }
-    bool operator<(const Stamp& sb) const {
-        return address < sb.address;
-    }
-    bool operator>(const Stamp& sb) const {
-        return address > sb.address;
-    }
-    bool operator<=(const Stamp& sb) const {
-        return address <= sb.address;
-    }
-    bool operator>=(const Stamp& sb) const {
-        return address >= sb.address;
-    }
-};
-
 struct CFGRebuildState {
     explicit CFGRebuildState(const ProgramCode& program_code, const std::size_t program_size)
         : program_code{program_code}, program_size{program_size} {
@@ -144,8 +123,8 @@ struct CFGRebuildState {
     std::list<Query> queries{};
     std::unordered_map<u32, u32> registered{};
     std::unordered_set<u32> labels{};
-    std::set<Stamp> ssy_labels;
-    std::set<Stamp> pbk_labels;
+    std::map<u32, u32> ssy_labels;
+    std::map<u32, u32> pbk_labels;
     std::unordered_map<u32, BlockStack> stacks{};
     const ProgramCode& program_code;
     const std::size_t program_size;
@@ -393,7 +372,7 @@ bool TryInspectAddress(CFGRebuildState& state) {
     }
     case BlockCollision::Inside: {
         // This case is the tricky one:
-        // We need to Split the block in 2 sepprate blocks
+        // We need to Split the block in 2 sepparate blocks
         auto it = search_result.second;
         block_info = CreateBlockInfo(state, address, it->end);
         it->end = address - 1;
@@ -428,13 +407,11 @@ bool TryInspectAddress(CFGRebuildState& state) {
 }
 
 bool TryQuery(CFGRebuildState& state) {
-    auto gather_labels = ([](ControlStack& cc, std::set<Stamp> labels, BlockInfo& block) {
-        Stamp start{block.start, 0};
-        Stamp end{block.end, 0};
-        auto gather_start = labels.lower_bound(start);
-        auto gather_end = labels.upper_bound(end);
+    auto gather_labels = ([](ControlStack& cc, std::map<u32, u32>& labels, BlockInfo& block) {
+        auto gather_start = labels.lower_bound(block.start);
+        auto gather_end = labels.upper_bound(block.end);
         while (gather_start != gather_end) {
-            cc.Push(gather_start->target);
+            cc.Push(gather_start->second);
             gather_start++;
         }
     });
@@ -444,9 +421,13 @@ bool TryQuery(CFGRebuildState& state) {
     Query& q = state.queries.front();
     u32 block_index = state.registered[q.address];
     BlockInfo& block = state.block_info[block_index];
+    // If the block is visted, check if the stacks match, else gather the ssy/pbk
+    // labels into the current stack and look if the branch at the end of the block
+    // consumes a label. Schedule new queries accordingly
     if (block.visited) {
         BlockStack& stack = state.stacks[q.address];
-        bool all_okay = q.ssy_stack.Compare(stack.ssy_stack) && q.pbk_stack.Compare(stack.pbk_stack);
+        bool all_okay = (stack.ssy_stack.Size() == 0 || q.ssy_stack.Compare(stack.ssy_stack)) &&
+                        (stack.pbk_stack.Size() == 0 || q.pbk_stack.Compare(stack.pbk_stack));
         state.queries.pop_front();
         return all_okay;
     }
@@ -523,8 +504,10 @@ bool ScanFlow(const ProgramCode& program_code, u32 program_size, u32 start_addre
         result_out.blocks.push_back(new_block);
     }
     if (result_out.decompilable) {
+        result_out.labels = std::move(state.labels);
         return true;
     }
+    // If it's not decompilable, merge the unlabelled blocks together
     auto back = result_out.blocks.begin();
     auto next = std::next(back);
     while (next != result_out.blocks.end()) {

src/video_core/shader/control_flow.h

@@ -3,7 +3,7 @@
 #include <cstring>
 #include <list>
 #include <optional>
-#include <vector>
+#include <unordered_set>
 
 #include "video_core/engines/shader_bytecode.h"
 #include "video_core/shader/shader_ir.h"
@@ -48,6 +48,7 @@ struct ShaderCharacteristics {
     bool decompilable{};
     u32 start;
     u32 end;
+    std::unordered_set<u32> labels{};
 };
 
 bool ScanFlow(const ProgramCode& program_code, u32 program_size, u32 start_address,

src/video_core/shader/decode.cpp

@@ -38,32 +38,47 @@ constexpr bool IsSchedInstruction(u32 offset, u32 main_offset) {
 void ShaderIR::Decode() {
     std::memcpy(&header, program_code.data(), sizeof(Tegra::Shader::Header));
 
+    disable_flow_stack = false;
     ShaderCharacteristics shader_info{};
     bool can_proceed = ScanFlow(program_code, program_code.size(), main_offset, shader_info);
     if (can_proceed) {
         coverage_begin = shader_info.start;
         coverage_end = shader_info.end;
         if (shader_info.decompilable) {
+            disable_flow_stack = true;
+            auto insert_block = ([this](NodeBlock& nodes, u32 label) {
+                if (label == exit_branch) {
+                    return;
+                }
+                basic_blocks.insert({label, nodes});
+            });
             std::list<ShaderBlock>& blocks = shader_info.blocks;
+            NodeBlock current_block;
+            u32 current_label = exit_branch;
             for (auto& block : blocks) {
-                NodeBlock nodes;
+                if (shader_info.labels.count(block.start) != 0) {
+                    insert_block(current_block, current_label);
+                    current_block.clear();
+                    current_label = block.start;
+                }
                 if (!block.ignore_branch) {
-                    nodes = DecodeRange(block.start, block.end);
-                    InsertControlFlow(nodes, block);
+                    DecodeRangeInner(current_block, block.start, block.end);
+                    InsertControlFlow(current_block, block);
                 } else {
-                    nodes = DecodeRange(block.start, block.end + 1);
+                    DecodeRangeInner(current_block, block.start, block.end + 1);
                 }
-                basic_blocks.insert({block.start, nodes});
             }
+            insert_block(current_block, current_label);
             return;
         }
+        LOG_WARNING(HW_GPU, "Flow Stack Removing Failed! Falling back to old method");
         // we can't decompile it, fallback to standard method
         for (const auto& block : shader_info.blocks) {
             basic_blocks.insert({block.start, DecodeRange(block.start, block.end + 1)});
         }
         return;
     }
-    LOG_WARNING(HW_GPU, "Flow Analysis failed, falling back to brute force compiling");
+    LOG_WARNING(HW_GPU, "Flow Analysis Failed! Falling back to brute force compiling");
 
     // Now we need to deal with an undecompilable shader. We need to brute force
     // a shader that captures every position.
@@ -78,10 +93,14 @@ void ShaderIR::Decode() {
 
 NodeBlock ShaderIR::DecodeRange(u32 begin, u32 end) {
     NodeBlock basic_block;
+    DecodeRangeInner(basic_block, begin, end);
+    return basic_block;
+}
+
+void ShaderIR::DecodeRangeInner(NodeBlock& bb, u32 begin, u32 end) {
     for (u32 pc = begin; pc < (begin > end ? MAX_PROGRAM_LENGTH : end);) {
-        pc = DecodeInstr(basic_block, pc);
+        pc = DecodeInstr(bb, pc);
     }
-    return basic_block;
 }
 
 void ShaderIR::InsertControlFlow(NodeBlock& bb, const ShaderBlock& block) {

src/video_core/shader/decode/other.cpp

@@ -98,9 +98,10 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         } else {
             const u32 target = pc + 1;
             const Node op_a = GetConstBuffer(instr.cbuf36.index, instr.cbuf36.GetOffset());
-            const Node convert = SignedOperation(OperationCode::IArithmeticShiftRight,
-                                               true, PRECISE, op_a, Immediate(3));
-            const Node operand = Operation(OperationCode::IAdd, PRECISE, convert, Immediate(target));
+            const Node convert = SignedOperation(OperationCode::IArithmeticShiftRight, true,
+                                                 PRECISE, op_a, Immediate(3));
+            const Node operand =
+                Operation(OperationCode::IAdd, PRECISE, convert, Immediate(target));
             branch = Operation(OperationCode::BranchIndirect, convert);
         }
 
@@ -119,14 +120,14 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
             const Node index = GetRegister(instr.gpr8);
             const Node op_a =
                 GetConstBufferIndirect(instr.cbuf36.index, instr.cbuf36.GetOffset() + 0, index);
-            const Node convert = SignedOperation(OperationCode::IArithmeticShiftRight,
-                                               true, PRECISE, op_a, Immediate(3));
+            const Node convert = SignedOperation(OperationCode::IArithmeticShiftRight, true,
+                                                 PRECISE, op_a, Immediate(3));
             operand = Operation(OperationCode::IAdd, PRECISE, convert, Immediate(target));
         } else {
             const s32 target = pc + instr.brx.GetBranchExtend();
             const Node op_a = GetRegister(instr.gpr8);
-            const Node convert = SignedOperation(OperationCode::IArithmeticShiftRight,
-                                               true, PRECISE, op_a, Immediate(3));
+            const Node convert = SignedOperation(OperationCode::IArithmeticShiftRight, true,
+                                                 PRECISE, op_a, Immediate(3));
             operand = Operation(OperationCode::IAdd, PRECISE, convert, Immediate(target));
         }
         const Node branch = Operation(OperationCode::BranchIndirect, operand);
@@ -143,6 +144,10 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         UNIMPLEMENTED_IF_MSG(instr.bra.constant_buffer != 0,
                              "Constant buffer flow is not supported");
 
+        if (disable_flow_stack) {
+            break;
+        }
+
         // The SSY opcode tells the GPU where to re-converge divergent execution paths with SYNC.
         const u32 target = pc + instr.bra.GetBranchTarget();
         bb.push_back(
@@ -153,6 +158,10 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         UNIMPLEMENTED_IF_MSG(instr.bra.constant_buffer != 0,
                              "Constant buffer PBK is not supported");
 
+        if (disable_flow_stack) {
+            break;
+        }
+
         // PBK pushes to a stack the address where BRK will jump to.
         const u32 target = pc + instr.bra.GetBranchTarget();
         bb.push_back(
@@ -164,6 +173,10 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         UNIMPLEMENTED_IF_MSG(cc != Tegra::Shader::ConditionCode::T, "SYNC condition code used: {}",
                              static_cast<u32>(cc));
 
+        if (disable_flow_stack) {
+            break;
+        }
+
         // The SYNC opcode jumps to the address previously set by the SSY opcode
         bb.push_back(Operation(OperationCode::PopFlowStack, MetaStackClass::Ssy));
         break;
@@ -172,6 +185,9 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         const Tegra::Shader::ConditionCode cc = instr.flow_condition_code;
         UNIMPLEMENTED_IF_MSG(cc != Tegra::Shader::ConditionCode::T, "BRK condition code used: {}",
                              static_cast<u32>(cc));
+        if (disable_flow_stack) {
+            break;
+        }
 
         // The BRK opcode jumps to the address previously set by the PBK opcode
         bb.push_back(Operation(OperationCode::PopFlowStack, MetaStackClass::Pbk));

src/video_core/shader/node.h

@@ -148,12 +148,12 @@ enum class OperationCode {
 
     ImageStore, /// (MetaImage, float[N] coords) -> void
 
-    Branch,        /// (uint branch_target) -> void
-    BranchIndirect,/// (uint branch_target) -> void
-    PushFlowStack, /// (uint branch_target) -> void
-    PopFlowStack,  /// () -> void
-    Exit,          /// () -> void
-    Discard,       /// () -> void
+    Branch,         /// (uint branch_target) -> void
+    BranchIndirect, /// (uint branch_target) -> void
+    PushFlowStack,  /// (uint branch_target) -> void
+    PopFlowStack,   /// () -> void
+    Exit,           /// () -> void
+    Discard,        /// () -> void
 
     EmitVertex,   /// () -> void
     EndPrimitive, /// () -> void

src/video_core/shader/shader_ir.h

@@ -123,10 +123,15 @@ public:
         return header;
     }
 
+    bool IsFlowStackDisabled() const {
+        return disable_flow_stack;
+    }
+
 private:
     void Decode();
 
     NodeBlock DecodeRange(u32 begin, u32 end);
+    void DecodeRangeInner(NodeBlock& bb, u32 begin, u32 end);
     void InsertControlFlow(NodeBlock& bb, const ShaderBlock& block);
 
     /**
@@ -320,6 +325,7 @@ private:
     const ProgramCode& program_code;
     const u32 main_offset;
     const std::size_t program_size;
+    bool disable_flow_stack{};
 
     u32 coverage_begin{};
     u32 coverage_end{};