early-access version 1592

src/video_core/shader/control_flow.cpp

@@ -7,6 +7,7 @@
 #include <set>
 #include <stack>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 #include "common/assert.h"
@@ -26,17 +27,29 @@ using Tegra::Shader::OpCode;
 
 constexpr s32 unassigned_branch = -2;
 
+enum class JumpLabel : u32 {
+    SSYClass = 0,
+    PBKClass = 1,
+};
+
+struct JumpItem {
+    JumpLabel type;
+    u32 address;
+
+    bool operator==(const JumpItem& other) const {
+        return std::tie(type, address) == std::tie(other.type, other.address);
+    }
+};
+
 struct Query {
     u32 address{};
-    std::stack<u32> ssy_stack{};
-    std::stack<u32> pbk_stack{};
+    std::stack<JumpItem> stack{};
 };
 
 struct BlockStack {
     BlockStack() = default;
-    explicit BlockStack(const Query& q) : ssy_stack{q.ssy_stack}, pbk_stack{q.pbk_stack} {}
-    std::stack<u32> ssy_stack{};
-    std::stack<u32> pbk_stack{};
+    explicit BlockStack(const Query& q) : stack{q.stack} {}
+    std::stack<JumpItem> stack{};
 };
 
 template <typename T, typename... Args>
@@ -65,20 +78,36 @@ struct BlockInfo {
     }
 };
 
-struct CFGRebuildState {
-    explicit CFGRebuildState(const ProgramCode& program_code_, u32 start_, Registry& registry_)
-        : program_code{program_code_}, registry{registry_}, start{start_} {}
+struct ProgramControl {
+    std::unordered_set<u32> found_functions{};
+    std::list<u32> pending_functions{};
 
+    void RegisterFunction(u32 address) {
+        if (found_functions.count(address) != 0) {
+            return;
+        }
+        found_functions.insert(address);
+        pending_functions.emplace_back(address);
+    }
+};
+
+struct CFGRebuildState {
+    explicit CFGRebuildState(ProgramControl& control_, const ProgramCode& program_code_, u32 start_,
+                             u32 base_start_, Registry& registry_)
+        : control{control_}, program_code{program_code_}, registry{registry_}, start{start_},
+          base_start{base_start_} {}
+
+    ProgramControl& control;
     const ProgramCode& program_code;
     Registry& registry;
     u32 start{};
+    u32 base_start{};
     std::vector<BlockInfo> block_info;
     std::list<u32> inspect_queries;
     std::list<Query> queries;
     std::unordered_map<u32, u32> registered;
     std::set<u32> labels;
-    std::map<u32, u32> ssy_labels;
-    std::map<u32, u32> pbk_labels;
+    std::map<u32, JumpItem> jump_labels;
     std::unordered_map<u32, BlockStack> stacks;
     ASTManager* manager{};
 };
@@ -153,7 +182,7 @@ template <typename Result, typename TestCallable, typename PackCallable>
 std::optional<Result> TrackInstruction(const CFGRebuildState& state, u32& pos, TestCallable test,
                                        PackCallable pack) {
     for (; pos >= state.start; --pos) {
-        if (IsSchedInstruction(pos, state.start)) {
+        if (IsSchedInstruction(pos, state.base_start)) {
             continue;
         }
         const Instruction instr = state.program_code[pos];
@@ -262,7 +291,7 @@ std::pair<ParseResult, ParseInfo> ParseCode(CFGRebuildState& state, u32 address)
             single_branch.ignore = true;
             break;
         }
-        if (IsSchedInstruction(offset, state.start)) {
+        if (IsSchedInstruction(offset, state.base_start)) {
             offset++;
             continue;
         }
@@ -274,6 +303,7 @@ std::pair<ParseResult, ParseInfo> ParseCode(CFGRebuildState& state, u32 address)
         }
 
         switch (opcode->get().GetId()) {
+        case OpCode::Id::RET:
         case OpCode::Id::EXIT: {
             const auto pred_index = static_cast<u32>(instr.pred.pred_index);
             single_branch.condition.predicate = GetPredicate(pred_index, instr.negate_pred != 0);
@@ -411,13 +441,20 @@ std::pair<ParseResult, ParseInfo> ParseCode(CFGRebuildState& state, u32 address)
         case OpCode::Id::SSY: {
             const u32 target = offset + instr.bra.GetBranchTarget();
             insert_label(state, target);
-            state.ssy_labels.emplace(offset, target);
+            JumpItem it = {JumpLabel::SSYClass, target};
+            state.jump_labels.emplace(offset, it);
             break;
         }
         case OpCode::Id::PBK: {
             const u32 target = offset + instr.bra.GetBranchTarget();
             insert_label(state, target);
-            state.pbk_labels.emplace(offset, target);
+            JumpItem it = {JumpLabel::PBKClass, target};
+            state.jump_labels.emplace(offset, it);
+            break;
+        }
+        case OpCode::Id::CAL: {
+            const u32 target = offset + instr.bra.GetBranchTarget();
+            state.control.RegisterFunction(target);
             break;
         }
         case OpCode::Id::BRX: {
@@ -513,7 +550,7 @@ bool TryInspectAddress(CFGRebuildState& state) {
 }
 
 bool TryQuery(CFGRebuildState& state) {
-    const auto gather_labels = [](std::stack<u32>& cc, std::map<u32, u32>& labels,
+    const auto gather_labels = [](std::stack<JumpItem>& cc, std::map<u32, JumpItem>& labels,
                                   BlockInfo& block) {
         auto gather_start = labels.lower_bound(block.start);
         const auto gather_end = labels.upper_bound(block.end);
@@ -522,6 +559,19 @@ bool TryQuery(CFGRebuildState& state) {
             ++gather_start;
         }
     };
+    const auto pop_labels = [](JumpLabel type, SingleBranch* branch, Query& query) -> bool {
+        while (!query.stack.empty() && query.stack.top().type != type) {
+            query.stack.pop();
+        }
+        if (query.stack.empty()) {
+            return false;
+        }
+        if (branch->address == unassigned_branch) {
+            branch->address = query.stack.top().address;
+        }
+        query.stack.pop();
+        return true;
+    };
     if (state.queries.empty()) {
         return false;
     }
@@ -534,8 +584,7 @@ bool TryQuery(CFGRebuildState& state) {
     // consumes a label. Schedule new queries accordingly
     if (block.visited) {
         BlockStack& stack = state.stacks[q.address];
-        const bool all_okay = (stack.ssy_stack.empty() || q.ssy_stack == stack.ssy_stack) &&
-                              (stack.pbk_stack.empty() || q.pbk_stack == stack.pbk_stack);
+        const bool all_okay = (stack.stack.empty() || q.stack == stack.stack);
         state.queries.pop_front();
         return all_okay;
     }
@@ -544,8 +593,7 @@ bool TryQuery(CFGRebuildState& state) {
 
     Query q2(q);
     state.queries.pop_front();
-    gather_labels(q2.ssy_stack, state.ssy_labels, block);
-    gather_labels(q2.pbk_stack, state.pbk_labels, block);
+    gather_labels(q2.stack, state.jump_labels, block);
     if (std::holds_alternative<SingleBranch>(*block.branch)) {
         auto* branch = std::get_if<SingleBranch>(block.branch.get());
         if (!branch->condition.IsUnconditional()) {
@@ -555,16 +603,10 @@ bool TryQuery(CFGRebuildState& state) {
 
         auto& conditional_query = state.queries.emplace_back(q2);
         if (branch->is_sync) {
-            if (branch->address == unassigned_branch) {
-                branch->address = conditional_query.ssy_stack.top();
-            }
-            conditional_query.ssy_stack.pop();
+            pop_labels(JumpLabel::SSYClass, branch, conditional_query);
         }
         if (branch->is_brk) {
-            if (branch->address == unassigned_branch) {
-                branch->address = conditional_query.pbk_stack.top();
-            }
-            conditional_query.pbk_stack.pop();
+            pop_labels(JumpLabel::PBKClass, branch, conditional_query);
         }
         conditional_query.address = branch->address;
         return true;
@@ -646,25 +688,23 @@ void DecompileShader(CFGRebuildState& state) {
     state.manager->Decompile();
 }
 
-} // Anonymous namespace
-
-std::unique_ptr<ShaderCharacteristics> ScanFlow(const ProgramCode& program_code, u32 start_address,
-                                                const CompilerSettings& settings,
-                                                Registry& registry) {
-    auto result_out = std::make_unique<ShaderCharacteristics>();
+ShaderFunction ScanFunction(ProgramControl& control, const ProgramCode& program_code,
+                            u32 start_address, u32 base_start, const CompilerSettings& settings,
+                            Registry& registry) {
+    ShaderFunction result_out{};
     if (settings.depth == CompileDepth::BruteForce) {
-        result_out->settings.depth = CompileDepth::BruteForce;
+        result_out.settings.depth = CompileDepth::BruteForce;
         return result_out;
     }
 
-    CFGRebuildState state{program_code, start_address, registry};
+    CFGRebuildState state{control, program_code, start_address, base_start, registry};
     // Inspect Code and generate blocks
     state.labels.clear();
     state.labels.emplace(start_address);
     state.inspect_queries.push_back(state.start);
     while (!state.inspect_queries.empty()) {
         if (!TryInspectAddress(state)) {
-            result_out->settings.depth = CompileDepth::BruteForce;
+            result_out.settings.depth = CompileDepth::BruteForce;
             return result_out;
         }
     }
@@ -675,7 +715,7 @@ std::unique_ptr<ShaderCharacteristics> ScanFlow(const ProgramCode& program_code,
 
     if (settings.depth != CompileDepth::FlowStack) {
         // Decompile Stacks
-        state.queries.push_back(Query{state.start, {}, {}});
+        state.queries.push_back(Query{state.start, {}});
         decompiled = true;
         while (!state.queries.empty()) {
             if (!TryQuery(state)) {
@@ -705,19 +745,18 @@ std::unique_ptr<ShaderCharacteristics> ScanFlow(const ProgramCode& program_code,
             state.manager->ShowCurrentState("Of Shader");
             state.manager->Clear();
         } else {
-            auto characteristics = std::make_unique<ShaderCharacteristics>();
-            characteristics->start = start_address;
-            characteristics->settings.depth = settings.depth;
-            characteristics->manager = std::move(manager);
-            characteristics->end = state.block_info.back().end + 1;
-            return characteristics;
+            result_out.start = start_address;
+            result_out.settings.depth = settings.depth;
+            result_out.manager = std::move(manager);
+            result_out.end = state.block_info.back().end + 1;
+            return result_out;
         }
     }
 
-    result_out->start = start_address;
-    result_out->settings.depth =
+    result_out.start = start_address;
+    result_out.settings.depth =
         use_flow_stack ? CompileDepth::FlowStack : CompileDepth::NoFlowStack;
-    result_out->blocks.clear();
+    result_out.blocks.clear();
     for (auto& block : state.block_info) {
         ShaderBlock new_block{};
         new_block.start = block.start;
@@ -726,20 +765,20 @@ std::unique_ptr<ShaderCharacteristics> ScanFlow(const ProgramCode& program_code,
         if (!new_block.ignore_branch) {
             new_block.branch = block.branch;
         }
-        result_out->end = std::max(result_out->end, block.end);
-        result_out->blocks.push_back(new_block);
+        result_out.end = std::max(result_out.end, block.end);
+        result_out.blocks.push_back(new_block);
     }
     if (!use_flow_stack) {
-        result_out->labels = std::move(state.labels);
+        result_out.labels = std::move(state.labels);
         return result_out;
     }
 
-    auto back = result_out->blocks.begin();
+    auto back = result_out.blocks.begin();
     auto next = std::next(back);
-    while (next != result_out->blocks.end()) {
+    while (next != result_out.blocks.end()) {
         if (!state.labels.contains(next->start) && next->start == back->end + 1) {
             back->end = next->end;
-            next = result_out->blocks.erase(next);
+            next = result_out.blocks.erase(next);
             continue;
         }
         back = next;
@@ -748,4 +787,22 @@ std::unique_ptr<ShaderCharacteristics> ScanFlow(const ProgramCode& program_code,
 
     return result_out;
 }
+
+} // Anonymous namespace
+
+std::unique_ptr<ShaderProgram> ScanFlow(const ProgramCode& program_code, u32 start_address,
+                                        const CompilerSettings& settings, Registry& registry) {
+    ProgramControl control{};
+    auto result_out = std::make_unique<ShaderProgram>();
+    result_out->main =
+        ScanFunction(control, program_code, start_address, start_address, settings, registry);
+    while (!control.pending_functions.empty()) {
+        u32 address = control.pending_functions.front();
+        auto fun = ScanFunction(control, program_code, address, start_address, settings, registry);
+        result_out->subfunctions.emplace(address, std::move(fun));
+        control.pending_functions.pop_front();
+    }
+    return result_out;
+}
+
 } // namespace VideoCommon::Shader

src/video_core/shader/control_flow.h

@@ -5,6 +5,7 @@
 #pragma once
 
 #include <list>
+#include <map>
 #include <optional>
 #include <set>
 #include <variant>
@@ -101,7 +102,7 @@ struct ShaderBlock {
     }
 };
 
-struct ShaderCharacteristics {
+struct ShaderFunction {
     std::list<ShaderBlock> blocks{};
     std::set<u32> labels{};
     u32 start{};
@@ -110,8 +111,12 @@ struct ShaderCharacteristics {
     CompilerSettings settings{};
 };
 
-std::unique_ptr<ShaderCharacteristics> ScanFlow(const ProgramCode& program_code, u32 start_address,
-                                                const CompilerSettings& settings,
-                                                Registry& registry);
+struct ShaderProgram {
+    ShaderFunction main;
+    std::map<u32, ShaderFunction> subfunctions;
+};
+
+std::unique_ptr<ShaderProgram> ScanFlow(const ProgramCode& program_code, u32 start_address,
+                                        const CompilerSettings& settings, Registry& registry);
 
 } // namespace VideoCommon::Shader

src/video_core/shader/decode.cpp

@@ -64,9 +64,52 @@ std::optional<u32> TryDeduceSamplerSize(const SamplerEntry& sampler_to_deduce,
 
 } // Anonymous namespace
 
+class ExprDecoder {
+public:
+    explicit ExprDecoder(ShaderIR& ir_) : ir(ir_) {}
+
+    void operator()(const ExprAnd& expr) {
+        Visit(expr.operand1);
+        Visit(expr.operand2);
+    }
+
+    void operator()(const ExprOr& expr) {
+        Visit(expr.operand1);
+        Visit(expr.operand2);
+    }
+
+    void operator()(const ExprNot& expr) {
+        Visit(expr.operand1);
+    }
+
+    void operator()(const ExprPredicate& expr) {
+        const auto pred = static_cast<Tegra::Shader::Pred>(expr.predicate);
+        if (pred != Pred::UnusedIndex && pred != Pred::NeverExecute) {
+            ir.used_predicates.insert(pred);
+        }
+    }
+
+    void operator()(const ExprCondCode& expr) {}
+
+    void operator()(const ExprVar& expr) {}
+
+    void operator()(const ExprBoolean& expr) {}
+
+    void operator()(const ExprGprEqual& expr) {
+        ir.used_registers.insert(expr.gpr);
+    }
+
+    void Visit(const Expr& node) {
+        return std::visit(*this, *node);
+    }
+
+private:
+    ShaderIR& ir;
+};
+
 class ASTDecoder {
 public:
-    explicit ASTDecoder(ShaderIR& ir_) : ir(ir_) {}
+    explicit ASTDecoder(ShaderIR& ir_) : ir(ir_), decoder(ir_) {}
 
     void operator()(ASTProgram& ast) {
         ASTNode current = ast.nodes.GetFirst();
@@ -77,6 +120,7 @@ public:
     }
 
     void operator()(ASTIfThen& ast) {
+        decoder.Visit(ast.condition);
         ASTNode current = ast.nodes.GetFirst();
         while (current) {
             Visit(current);
@@ -96,13 +140,18 @@ public:
 
     void operator()(ASTBlockDecoded& ast) {}
 
-    void operator()(ASTVarSet& ast) {}
+    void operator()(ASTVarSet& ast) {
+        decoder.Visit(ast.condition);
+    }
 
     void operator()(ASTLabel& ast) {}
 
-    void operator()(ASTGoto& ast) {}
+    void operator()(ASTGoto& ast) {
+        decoder.Visit(ast.condition);
+    }
 
     void operator()(ASTDoWhile& ast) {
+        decoder.Visit(ast.condition);
         ASTNode current = ast.nodes.GetFirst();
         while (current) {
             Visit(current);
@@ -110,9 +159,13 @@ public:
         }
     }
 
-    void operator()(ASTReturn& ast) {}
+    void operator()(ASTReturn& ast) {
+        decoder.Visit(ast.condition);
+    }
 
-    void operator()(ASTBreak& ast) {}
+    void operator()(ASTBreak& ast) {
+        decoder.Visit(ast.condition);
+    }
 
     void Visit(ASTNode& node) {
         std::visit(*this, *node->GetInnerData());
@@ -125,77 +178,113 @@ public:
 
 private:
     ShaderIR& ir;
+    ExprDecoder decoder;
 };
 
 void ShaderIR::Decode() {
+    const auto decode_function = ([this](ShaderFunction& shader_info) {
+        coverage_end = std::max<u32>(0, shader_info.end);
+        switch (shader_info.settings.depth) {
+        case CompileDepth::FlowStack: {
+            for (const auto& block : shader_info.blocks) {
+                basic_blocks.insert({block.start, DecodeRange(block.start, block.end + 1)});
+            }
+            break;
+        }
+        case CompileDepth::NoFlowStack: {
+            disable_flow_stack = true;
+            const auto insert_block = [this](NodeBlock& nodes, u32 label) {
+                if (label == static_cast<u32>(exit_branch)) {
+                    return;
+                }
+                basic_blocks.insert({label, nodes});
+            };
+            const auto& blocks = shader_info.blocks;
+            NodeBlock current_block;
+            u32 current_label = static_cast<u32>(exit_branch);
+            for (const auto& block : blocks) {
+                if (shader_info.labels.contains(block.start)) {
+                    insert_block(current_block, current_label);
+                    current_block.clear();
+                    current_label = block.start;
+                }
+                if (!block.ignore_branch) {
+                    DecodeRangeInner(current_block, block.start, block.end);
+                    InsertControlFlow(current_block, block);
+                } else {
+                    DecodeRangeInner(current_block, block.start, block.end + 1);
+                }
+            }
+            insert_block(current_block, current_label);
+            break;
+        }
+        case CompileDepth::DecompileBackwards:
+        case CompileDepth::FullDecompile: {
+            program_manager = std::move(shader_info.manager);
+            disable_flow_stack = true;
+            decompiled = true;
+            ASTDecoder decoder{*this};
+            ASTNode program = program_manager.GetProgram();
+            decoder.Visit(program);
+            break;
+        }
+        default:
+            LOG_CRITICAL(HW_GPU, "Unknown decompilation mode!");
+            [[fallthrough]];
+        case CompileDepth::BruteForce: {
+            const auto shader_end = static_cast<u32>(program_code.size());
+            coverage_begin = main_offset;
+            coverage_end = shader_end;
+            for (u32 label = main_offset; label < shader_end; ++label) {
+                basic_blocks.insert({label, DecodeRange(label, label + 1)});
+            }
+            break;
+        }
+        }
+        if (settings.depth != shader_info.settings.depth) {
+            LOG_WARNING(
+                HW_GPU,
+                "Decompiling to this setting \"{}\" failed, downgrading to this setting \"{}\"",
+                CompileDepthAsString(settings.depth),
+                CompileDepthAsString(shader_info.settings.depth));
+        }
+    });
+    const auto gen_function =
+        ([this](ShaderFunction& shader_info, u32 id) -> std::shared_ptr<ShaderFunctionIR> {
+            std::shared_ptr<ShaderFunctionIR> result;
+            if (decompiled) {
+                result = std::make_shared<ShaderFunctionIR>(std::move(program_manager), id,
+                                                            shader_info.start, shader_info.end);
+            } else {
+                result =
+                    std::make_shared<ShaderFunctionIR>(std::move(basic_blocks), disable_flow_stack,
+                                                       id, shader_info.start, shader_info.end);
+            }
+            decompiled = false;
+            disable_flow_stack = false;
+            basic_blocks.clear();
+            program_manager.Clear();
+            return result;
+        });
     std::memcpy(&header, program_code.data(), sizeof(Tegra::Shader::Header));
 
     decompiled = false;
     auto info = ScanFlow(program_code, main_offset, settings, registry);
-    auto& shader_info = *info;
-    coverage_begin = shader_info.start;
-    coverage_end = shader_info.end;
-    switch (shader_info.settings.depth) {
-    case CompileDepth::FlowStack: {
-        for (const auto& block : shader_info.blocks) {
-            basic_blocks.insert({block.start, DecodeRange(block.start, block.end + 1)});
-        }
-        break;
+    u32 id_start = 1;
+    for (auto& pair : info->subfunctions) {
+        func_map.emplace(pair.first, id_start);
+        id_start++;
     }
-    case CompileDepth::NoFlowStack: {
-        disable_flow_stack = true;
-        const auto insert_block = [this](NodeBlock& nodes, u32 label) {
-            if (label == static_cast<u32>(exit_branch)) {
-                return;
-            }
-            basic_blocks.insert({label, nodes});
-        };
-        const auto& blocks = shader_info.blocks;
-        NodeBlock current_block;
-        u32 current_label = static_cast<u32>(exit_branch);
-        for (const auto& block : blocks) {
-            if (shader_info.labels.contains(block.start)) {
-                insert_block(current_block, current_label);
-                current_block.clear();
-                current_label = block.start;
-            }
-            if (!block.ignore_branch) {
-                DecodeRangeInner(current_block, block.start, block.end);
-                InsertControlFlow(current_block, block);
-            } else {
-                DecodeRangeInner(current_block, block.start, block.end + 1);
-            }
-        }
-        insert_block(current_block, current_label);
-        break;
-    }
-    case CompileDepth::DecompileBackwards:
-    case CompileDepth::FullDecompile: {
-        program_manager = std::move(shader_info.manager);
-        disable_flow_stack = true;
-        decompiled = true;
-        ASTDecoder decoder{*this};
-        ASTNode program = GetASTProgram();
-        decoder.Visit(program);
-        break;
-    }
-    default:
-        LOG_CRITICAL(HW_GPU, "Unknown decompilation mode!");
-        [[fallthrough]];
-    case CompileDepth::BruteForce: {
-        const auto shader_end = static_cast<u32>(program_code.size());
-        coverage_begin = main_offset;
-        coverage_end = shader_end;
-        for (u32 label = main_offset; label < shader_end; ++label) {
-            basic_blocks.insert({label, DecodeRange(label, label + 1)});
-        }
-        break;
-    }
-    }
-    if (settings.depth != shader_info.settings.depth) {
-        LOG_WARNING(
-            HW_GPU, "Decompiling to this setting \"{}\" failed, downgrading to this setting \"{}\"",
-            CompileDepthAsString(settings.depth), CompileDepthAsString(shader_info.settings.depth));
+    coverage_begin = info->main.start;
+    coverage_end = 0;
+    decode_function(info->main);
+    main_function = gen_function(info->main, 0);
+    subfunctions.resize(info->subfunctions.size());
+    for (auto& pair : info->subfunctions) {
+        auto& func_info = pair.second;
+        decode_function(func_info);
+        u32 id = func_map[pair.first];
+        subfunctions[id - 1] = gen_function(func_info, id);
     }
 }
 

src/video_core/shader/decode/other.cpp

@@ -33,6 +33,7 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         // With the previous preconditions, this instruction is a no-operation.
         break;
     }
+    case OpCode::Id::RET:
     case OpCode::Id::EXIT: {
         const ConditionCode cc = instr.flow_condition_code;
         UNIMPLEMENTED_IF_MSG(cc != ConditionCode::T, "EXIT condition code used: {}", cc);
@@ -312,6 +313,16 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         LOG_DEBUG(HW_GPU, "DEPBAR instruction is stubbed");
         break;
     }
+    case OpCode::Id::CAL: {
+        const u32 target = pc + instr.bra.GetBranchTarget();
+        const auto it = func_map.find(target);
+        if (it == func_map.end()) {
+            UNREACHABLE();
+            break;
+        }
+        bb.push_back(FunctionCall(it->second));
+        break;
+    }
     default:
         UNIMPLEMENTED_MSG("Unhandled instruction: {}", opcode->get().GetName());
     }

src/video_core/shader/decode/texture.cpp

@@ -339,8 +339,6 @@ u32 ShaderIR::DecodeTexture(NodeBlock& bb, u32 pc) {
         const TextureType texture_type{instr.tlds.GetTextureType()};
         const bool is_array{instr.tlds.IsArrayTexture()};
 
-        UNIMPLEMENTED_IF_MSG(instr.tlds.UsesMiscMode(TextureMiscMode::AOFFI),
-                             "AOFFI is not implemented");
         UNIMPLEMENTED_IF_MSG(instr.tlds.UsesMiscMode(TextureMiscMode::MZ), "MZ is not implemented");
 
         const Node4 components = GetTldsCode(instr, texture_type, is_array);
@@ -822,7 +820,7 @@ Node4 ShaderIR::GetTldsCode(Instruction instr, TextureType texture_type, bool is
     for (std::size_t i = 0; i < type_coord_count; ++i) {
         const bool last = (i == (type_coord_count - 1)) && (type_coord_count > 1);
         coords.push_back(
-            GetRegister(last && !aoffi_enabled ? last_coord_register : coord_register + i));
+            GetRegister(last && !aoffi_enabled ? last_coord_register : (coord_register + i)));
     }
 
     const Node array = is_array ? GetRegister(array_register) : nullptr;

src/video_core/shader/node.h

@@ -267,10 +267,11 @@ class PatchNode;
 class SmemNode;
 class GmemNode;
 class CommentNode;
+class FunctionCallNode;
 
 using NodeData = std::variant<OperationNode, ConditionalNode, GprNode, CustomVarNode, ImmediateNode,
                               InternalFlagNode, PredicateNode, AbufNode, PatchNode, CbufNode,
-                              LmemNode, SmemNode, GmemNode, CommentNode>;
+                              LmemNode, SmemNode, GmemNode, FunctionCallNode, CommentNode>;
 using Node = std::shared_ptr<NodeData>;
 using Node4 = std::array<Node, 4>;
 using NodeBlock = std::vector<Node>;
@@ -494,6 +495,18 @@ private:
     std::vector<Node> code; ///< Code to execute
 };
 
+class FunctionCallNode final : public AmendNode {
+public:
+    explicit FunctionCallNode(u32 func_id_) : func_id{func_id_} {}
+
+    [[nodiscard]] u32 GetFuncId() const {
+        return func_id;
+    }
+
+private:
+    u32 func_id; ///< Id of the function to call
+};
+
 /// A general purpose register
 class GprNode final {
 public:

src/video_core/shader/node_helper.cpp

@@ -19,6 +19,11 @@ Node Comment(std::string text) {
     return MakeNode<CommentNode>(std::move(text));
 }
 
+/// Creates a function call
+Node FunctionCall(u32 func_id) {
+    return MakeNode<FunctionCallNode>(func_id);
+}
+
 Node Immediate(u32 value) {
     return MakeNode<ImmediateNode>(value);
 }

src/video_core/shader/node_helper.h

@@ -27,6 +27,9 @@ Node Conditional(Node condition, std::vector<Node> code);
 /// Creates a commentary node
 Node Comment(std::string text);
 
+/// Creates a function call
+Node FunctionCall(u32 func_id);
+
 /// Creates an u32 immediate
 Node Immediate(u32 value);
 

src/video_core/shader/shader_ir.h

@@ -26,7 +26,7 @@ namespace VideoCommon::Shader {
 
 struct ShaderBlock;
 
-constexpr u32 MAX_PROGRAM_LENGTH = 0x1000;
+constexpr u32 MAX_PROGRAM_LENGTH = 0x2000;
 
 struct ConstBuffer {
     constexpr explicit ConstBuffer(u32 max_offset_, bool is_indirect_)
@@ -64,16 +64,68 @@ struct GlobalMemoryUsage {
     bool is_written{};
 };
 
+class ShaderFunctionIR final {
+public:
+    explicit ShaderFunctionIR(std::map<u32, NodeBlock>&& basic_blocks_, bool disable_flow_stack_,
+                              u32 id_, u32 coverage_begin_, u32 coverage_end_)
+        : basic_blocks{std::move(basic_blocks_)}, decompiled{false},
+          disable_flow_stack{disable_flow_stack_}, id{id_}, coverage_begin{coverage_begin_},
+          coverage_end{coverage_end_} {}
+    explicit ShaderFunctionIR(ASTManager&& program_manager_, u32 id_, u32 coverage_begin_,
+                              u32 coverage_end_)
+        : program_manager{std::move(program_manager_)}, decompiled{true}, disable_flow_stack{true},
+          id{id_}, coverage_begin{coverage_begin_}, coverage_end{coverage_end_} {}
+
+    const std::map<u32, NodeBlock>& GetBasicBlocks() const {
+        return basic_blocks;
+    }
+
+    [[nodiscard]] bool IsFlowStackDisabled() const {
+        return disable_flow_stack;
+    }
+
+    [[nodiscard]] bool IsDecompiled() const {
+        return decompiled;
+    }
+
+    const ASTManager& GetASTManager() const {
+        return program_manager;
+    }
+
+    [[nodiscard]] ASTNode GetASTProgram() const {
+        return program_manager.GetProgram();
+    }
+
+    [[nodiscard]] u32 GetASTNumVariables() const {
+        return program_manager.GetVariables();
+    }
+
+    [[nodiscard]] bool IsMain() const {
+        return id == 0;
+    }
+
+    [[nodiscard]] u32 GetId() const {
+        return id;
+    }
+
+private:
+    std::map<u32, NodeBlock> basic_blocks;
+    ASTManager program_manager{true, true};
+
+    bool decompiled{};
+    bool disable_flow_stack{};
+    u32 id{};
+
+    u32 coverage_begin{};
+    u32 coverage_end{};
+};
+
 class ShaderIR final {
 public:
     explicit ShaderIR(const ProgramCode& program_code_, u32 main_offset_,
                       CompilerSettings settings_, Registry& registry_);
     ~ShaderIR();
 
-    const std::map<u32, NodeBlock>& GetBasicBlocks() const {
-        return basic_blocks;
-    }
-
     const std::set<u32>& GetRegisters() const {
         return used_registers;
     }
@@ -155,26 +207,6 @@ public:
         return header;
     }
 
-    bool IsFlowStackDisabled() const {
-        return disable_flow_stack;
-    }
-
-    bool IsDecompiled() const {
-        return decompiled;
-    }
-
-    const ASTManager& GetASTManager() const {
-        return program_manager;
-    }
-
-    ASTNode GetASTProgram() const {
-        return program_manager.GetProgram();
-    }
-
-    u32 GetASTNumVariables() const {
-        return program_manager.GetVariables();
-    }
-
     u32 ConvertAddressToNvidiaSpace(u32 address) const {
         return (address - main_offset) * static_cast<u32>(sizeof(Tegra::Shader::Instruction));
     }
@@ -190,7 +222,16 @@ public:
         return num_custom_variables;
     }
 
+    std::shared_ptr<ShaderFunctionIR> GetMainFunction() const {
+        return main_function;
+    }
+
+    const std::vector<std::shared_ptr<ShaderFunctionIR>>& GetSubFunctions() const {
+        return subfunctions;
+    }
+
 private:
+    friend class ExprDecoder;
     friend class ASTDecoder;
 
     struct SamplerInfo {
@@ -453,6 +494,10 @@ private:
     std::vector<Node> amend_code;
     u32 num_custom_variables{};
 
+    std::shared_ptr<ShaderFunctionIR> main_function;
+    std::vector<std::shared_ptr<ShaderFunctionIR>> subfunctions;
+    std::unordered_map<u32, u32> func_map;
+
     std::set<u32> used_registers;
     std::set<Tegra::Shader::Pred> used_predicates;
     std::set<Tegra::Shader::Attribute::Index> used_input_attributes;