shader: SSA and dominance

src/shader_recompiler/CMakeLists.txt

@@ -11,6 +11,8 @@ add_executable(shader_recompiler
     frontend/ir/condition.h
     frontend/ir/flow_test.cpp
     frontend/ir/flow_test.h
+    frontend/ir/function.cpp
+    frontend/ir/function.h
     frontend/ir/ir_emitter.cpp
     frontend/ir/ir_emitter.h
     frontend/ir/microinstruction.cpp
@@ -51,6 +53,7 @@ add_executable(shader_recompiler
     ir_opt/get_set_elimination_pass.cpp
     ir_opt/identity_removal_pass.cpp
     ir_opt/passes.h
+    ir_opt/ssa_rewrite_pass.cpp
     ir_opt/verification_pass.cpp
     main.cpp
 )

src/shader_recompiler/frontend/ir/basic_block.cpp

@@ -37,6 +37,10 @@ Block::iterator Block::PrependNewInst(iterator insertion_point, Opcode op,
     return result_it;
 }
 
+void Block::AddImmediatePredecessor(IR::Block* immediate_predecessor) {
+    imm_predecessors.push_back(immediate_predecessor);
+}
+
 u32 Block::LocationBegin() const noexcept {
     return location_begin;
 }
@@ -53,6 +57,18 @@ const Block::InstructionList& Block::Instructions() const noexcept {
     return instructions;
 }
 
+std::span<IR::Block* const> Block::ImmediatePredecessors() const noexcept {
+    return imm_predecessors;
+}
+
+static std::string BlockToIndex(const std::map<const Block*, size_t>& block_to_index,
+                                Block* block) {
+    if (const auto it{block_to_index.find(block)}; it != block_to_index.end()) {
+        return fmt::format("{{Block ${}}}", it->second);
+    }
+    return fmt::format("$<unknown block {:016x}>", reinterpret_cast<u64>(block));
+}
+
 static std::string ArgToIndex(const std::map<const Block*, size_t>& block_to_index,
                               const std::map<const Inst*, size_t>& inst_to_index,
                               const Value& arg) {
@@ -60,10 +76,7 @@ static std::string ArgToIndex(const std::map<const Block*, size_t>& block_to_ind
         return "<null>";
     }
     if (arg.IsLabel()) {
-        if (const auto it{block_to_index.find(arg.Label())}; it != block_to_index.end()) {
+        return BlockToIndex(block_to_index, arg.Label());
-            return fmt::format("{{Block ${}}}", it->second);
-        }
-        return fmt::format("$<unknown block {:016x}>", reinterpret_cast<u64>(arg.Label()));
     }
     if (!arg.IsImmediate()) {
         if (const auto it{inst_to_index.find(arg.Inst())}; it != inst_to_index.end()) {
@@ -115,16 +128,26 @@ std::string DumpBlock(const Block& block, const std::map<const Block*, size_t>&
         } else {
             ret += fmt::format("         {}", op); // '%00000 = ' -> 1 + 5 + 3 = 9 spaces
         }
-        const size_t arg_count{NumArgsOf(op)};
+        if (op == Opcode::Phi) {
-        for (size_t arg_index = 0; arg_index < arg_count; ++arg_index) {
+            size_t val_index{0};
-            const Value arg{inst.Arg(arg_index)};
+            for (const auto& [phi_block, phi_val] : inst.PhiOperands()) {
-            ret += arg_index != 0 ? ", " : " ";
+                ret += val_index != 0 ? ", " : " ";
-            ret += ArgToIndex(block_to_index, inst_to_index, arg);
+                ret += fmt::format("[ {}, {} ]", ArgToIndex(block_to_index, inst_to_index, phi_val),
+                                   BlockToIndex(block_to_index, phi_block));
+                ++val_index;
+            }
+        } else {
+            const size_t arg_count{NumArgsOf(op)};
+            for (size_t arg_index = 0; arg_index < arg_count; ++arg_index) {
+                const Value arg{inst.Arg(arg_index)};
+                ret += arg_index != 0 ? ", " : " ";
+                ret += ArgToIndex(block_to_index, inst_to_index, arg);
 
-            const Type actual_type{arg.Type()};
+                const Type actual_type{arg.Type()};
-            const Type expected_type{ArgTypeOf(op, arg_index)};
+                const Type expected_type{ArgTypeOf(op, arg_index)};
-            if (!AreTypesCompatible(actual_type, expected_type)) {
+                if (!AreTypesCompatible(actual_type, expected_type)) {
-                ret += fmt::format("<type error: {} != {}>", actual_type, expected_type);
+                    ret += fmt::format("<type error: {} != {}>", actual_type, expected_type);
+                }
             }
         }
         if (TypeOf(op) != Type::Void) {

src/shader_recompiler/frontend/ir/basic_block.h

@@ -6,6 +6,8 @@
 
 #include <initializer_list>
 #include <map>
+#include <span>
+#include <vector>
 
 #include <boost/intrusive/list.hpp>
 #include <boost/pool/pool_alloc.hpp>
@@ -36,7 +38,11 @@ public:
     void AppendNewInst(Opcode op, std::initializer_list<Value> args);
 
     /// Prepends a new instruction to this basic block before the insertion point.
-    iterator PrependNewInst(iterator insertion_point, Opcode op, std::initializer_list<Value> args);
+    iterator PrependNewInst(iterator insertion_point, Opcode op,
+                            std::initializer_list<Value> args = {});
+
+    /// Adds a new immediate predecessor to the basic block.
+    void AddImmediatePredecessor(IR::Block* immediate_predecessor);
 
     /// Gets the starting location of this basic block.
     [[nodiscard]] u32 LocationBegin() const noexcept;
@@ -44,9 +50,12 @@ public:
     [[nodiscard]] u32 LocationEnd() const noexcept;
 
     /// Gets a mutable reference to the instruction list for this basic block.
-    InstructionList& Instructions() noexcept;
+    [[nodiscard]] InstructionList& Instructions() noexcept;
     /// Gets an immutable reference to the instruction list for this basic block.
-    const InstructionList& Instructions() const noexcept;
+    [[nodiscard]] const InstructionList& Instructions() const noexcept;
+
+    /// Gets an immutable span to the immediate predecessors.
+    [[nodiscard]] std::span<IR::Block* const> ImmediatePredecessors() const noexcept;
 
     [[nodiscard]] bool empty() const {
         return instructions.empty();
@@ -115,13 +124,16 @@ private:
     /// End location of this block
     u32 location_end;
 
-    /// List of instructions in this block.
+    /// List of instructions in this block
     InstructionList instructions;
 
     /// Memory pool for instruction list
     boost::fast_pool_allocator<Inst, boost::default_user_allocator_malloc_free,
                                boost::details::pool::null_mutex>
         instruction_alloc_pool;
+
+    /// Block immediate predecessors
+    std::vector<IR::Block*> imm_predecessors;
 };
 
 [[nodiscard]] std::string DumpBlock(const Block& block);

src/shader_recompiler/frontend/ir/function.cpp

@@ -0,0 +1,5 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "shader_recompiler/frontend/ir/function.h"

src/shader_recompiler/frontend/ir/function.h

@@ -0,0 +1,25 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <vector>
+
+#include "shader_recompiler/frontend/ir/basic_block.h"
+
+namespace Shader::IR {
+
+struct Function {
+    struct InplaceDelete {
+        void operator()(IR::Block* block) const noexcept {
+            std::destroy_at(block);
+        }
+    };
+    using UniqueBlock = std::unique_ptr<IR::Block, InplaceDelete>;
+
+    std::vector<UniqueBlock> blocks;
+};
+
+} // namespace Shader::IR

src/shader_recompiler/frontend/ir/microinstruction.cpp

@@ -30,6 +30,11 @@ static void RemovePseudoInstruction(IR::Inst*& inst, IR::Opcode expected_opcode)
 
 bool Inst::MayHaveSideEffects() const noexcept {
     switch (op) {
+    case Opcode::Branch:
+    case Opcode::BranchConditional:
+    case Opcode::Exit:
+    case Opcode::Return:
+    case Opcode::Unreachable:
     case Opcode::SetAttribute:
     case Opcode::SetAttributeIndexed:
     case Opcode::WriteGlobalU8:
@@ -113,6 +118,17 @@ void Inst::SetArg(size_t index, Value value) {
     args[index] = value;
 }
 
+std::span<const std::pair<Block*, Value>> Inst::PhiOperands() const noexcept {
+    return phi_operands;
+}
+
+void Inst::AddPhiOperand(Block* predecessor, const Value& value) {
+    if (!value.IsImmediate()) {
+        Use(value);
+    }
+    phi_operands.emplace_back(predecessor, value);
+}
+
 void Inst::Invalidate() {
     ClearArgs();
     op = Opcode::Void;
@@ -125,6 +141,12 @@ void Inst::ClearArgs() {
         }
         value = {};
     }
+    for (auto& [phi_block, phi_op] : phi_operands) {
+        if (!phi_op.IsImmediate()) {
+            UndoUse(phi_op);
+        }
+    }
+    phi_operands.clear();
 }
 
 void Inst::ReplaceUsesWith(Value replacement) {

src/shader_recompiler/frontend/ir/microinstruction.h

@@ -5,6 +5,8 @@
 #pragma once
 
 #include <array>
+#include <span>
+#include <vector>
 
 #include <boost/intrusive/list.hpp>
 
@@ -15,6 +17,8 @@
 
 namespace Shader::IR {
 
+class Block;
+
 constexpr size_t MAX_ARG_COUNT = 4;
 
 class Inst : public boost::intrusive::list_base_hook<> {
@@ -59,6 +63,11 @@ public:
     /// Set the value of a given argument index.
     void SetArg(size_t index, Value value);
 
+    /// Get an immutable span to the phi operands.
+    [[nodiscard]] std::span<const std::pair<Block*, Value>> PhiOperands() const noexcept;
+    /// Add phi operand to a phi instruction.
+    void AddPhiOperand(Block* predecessor, const Value& value);
+
     void Invalidate();
     void ClearArgs();
 
@@ -76,6 +85,7 @@ private:
     Inst* carry_inst{};
     Inst* overflow_inst{};
     Inst* zsco_inst{};
+    std::vector<std::pair<Block*, Value>> phi_operands;
     u64 flags{};
 };
 

src/shader_recompiler/frontend/ir/opcode.inc

@@ -5,6 +5,7 @@
 //     opcode name,                                         return type,    arg1 type,      arg2 type,      arg3 type,      arg4 type,      ...
 OPCODE(Void,                                                Void,                                                                           )
 OPCODE(Identity,                                            Opaque,         Opaque,                                                         )
+OPCODE(Phi,                                                 Opaque,         /*todo*/                                                        )
 
 // Control flow
 OPCODE(Branch,                                              Void,           Label,                                                          )
@@ -35,6 +36,13 @@ OPCODE(SetSFlag,                                            Void,           U1,
 OPCODE(SetCFlag,                                            Void,           U1,                                                             )
 OPCODE(SetOFlag,                                            Void,           U1,                                                             )
 
+// Undefined
+OPCODE(Undef1,                                              U1,                                                                             )
+OPCODE(Undef8,                                              U8,                                                                             )
+OPCODE(Undef16,                                             U16,                                                                            )
+OPCODE(Undef32,                                             U32,                                                                            )
+OPCODE(Undef64,                                             U64,                                                                            )
+
 // Memory operations
 OPCODE(WriteGlobalU8,                                       Void,           U64,            U32,                                            )
 OPCODE(WriteGlobalS8,                                       Void,           U64,            U32,                                            )

src/shader_recompiler/frontend/ir/pred.h

@@ -10,6 +10,13 @@ namespace Shader::IR {
 
 enum class Pred { P0, P1, P2, P3, P4, P5, P6, PT };
 
+constexpr size_t NUM_USER_PREDS = 6;
+constexpr size_t NUM_PREDS = 7;
+
+[[nodiscard]] constexpr size_t PredIndex(Pred pred) noexcept {
+    return static_cast<size_t>(pred);
+}
+
 } // namespace Shader::IR
 
 template <>

src/shader_recompiler/frontend/ir/reg.h

@@ -271,6 +271,9 @@ enum class Reg : u64 {
 };
 static_assert(static_cast<int>(Reg::RZ) == 255);
 
+constexpr size_t NUM_USER_REGS = 255;
+constexpr size_t NUM_REGS = 256;
+
 [[nodiscard]] constexpr Reg operator+(Reg reg, int num) {
     if (reg == Reg::RZ) {
         // Adding or subtracting registers from RZ yields RZ
@@ -290,8 +293,12 @@ static_assert(static_cast<int>(Reg::RZ) == 255);
     return reg + (-num);
 }
 
+[[nodiscard]] constexpr size_t RegIndex(Reg reg) noexcept {
+    return static_cast<size_t>(reg);
+}
+
 [[nodiscard]] constexpr bool IsAligned(Reg reg, size_t align) {
-    return (static_cast<size_t>(reg) / align) * align == static_cast<size_t>(reg);
+    return (RegIndex(reg) / align) * align == RegIndex(reg);
 }
 
 } // namespace Shader::IR

src/shader_recompiler/frontend/ir/value.cpp

@@ -115,6 +115,43 @@ u64 Value::U64() const {
     return imm_u64;
 }
 
+bool Value::operator==(const Value& other) const {
+    if (type != other.type) {
+        return false;
+    }
+    switch (type) {
+    case Type::Void:
+        return true;
+    case Type::Opaque:
+        return inst == other.inst;
+    case Type::Label:
+        return label == other.label;
+    case Type::Reg:
+        return reg == other.reg;
+    case Type::Pred:
+        return pred == other.pred;
+    case Type::Attribute:
+        return attribute == other.attribute;
+    case Type::U1:
+        return imm_u1 == other.imm_u1;
+    case Type::U8:
+        return imm_u8 == other.imm_u8;
+    case Type::U16:
+        return imm_u16 == other.imm_u16;
+    case Type::U32:
+        return imm_u32 == other.imm_u32;
+    case Type::U64:
+        return imm_u64 == other.imm_u64;
+    case Type::ZSCO:
+        throw NotImplementedException("ZSCO comparison");
+    }
+    throw LogicError("Invalid type {}", type);
+}
+
+bool Value::operator!=(const Value& other) const {
+    return !operator==(other);
+}
+
 void Value::ValidateAccess(IR::Type expected) const {
     if (type != expected) {
         throw LogicError("Reading {} out of {}", expected, type);

src/shader_recompiler/frontend/ir/value.h

@@ -48,6 +48,9 @@ public:
     [[nodiscard]] u32 U32() const;
     [[nodiscard]] u64 U64() const;
 
+    [[nodiscard]] bool operator==(const Value& other) const;
+    [[nodiscard]] bool operator!=(const Value& other) const;
+
 private:
     void ValidateAccess(IR::Type expected) const;
 

src/shader_recompiler/frontend/maxwell/control_flow.cpp

@@ -36,6 +36,7 @@ static std::array<Block, 2> Split(Block&& block, Location pc, BlockId new_id) {
             .cond{true},
             .branch_true{new_id},
             .branch_false{UNREACHABLE_BLOCK_ID},
+            .imm_predecessors{},
         },
         Block{
             .begin{pc},
@@ -46,6 +47,7 @@ static std::array<Block, 2> Split(Block&& block, Location pc, BlockId new_id) {
             .cond{block.cond},
             .branch_true{block.branch_true},
             .branch_false{block.branch_false},
+            .imm_predecessors{},
         },
     };
 }
@@ -108,7 +110,7 @@ static bool HasFlowTest(Opcode opcode) {
     }
 }
 
-static std::string Name(const Block& block) {
+static std::string NameOf(const Block& block) {
     if (block.begin.IsVirtual()) {
         return fmt::format("\"Virtual {}\"", block.id);
     } else {
@@ -154,13 +156,127 @@ bool Block::Contains(Location pc) const noexcept {
 }
 
 Function::Function(Location start_address)
-    : entrypoint{start_address}, labels{Label{
+    : entrypoint{start_address}, labels{{
                                      .address{start_address},
                                      .block_id{0},
                                      .stack{},
                                  }} {}
 
+void Function::BuildBlocksMap() {
+    const size_t num_blocks{NumBlocks()};
+    blocks_map.resize(num_blocks);
+    for (size_t block_index = 0; block_index < num_blocks; ++block_index) {
+        Block& block{blocks_data[block_index]};
+        blocks_map[block.id] = &block;
+    }
+}
+
+void Function::BuildImmediatePredecessors() {
+    for (const Block& block : blocks_data) {
+        if (block.branch_true != UNREACHABLE_BLOCK_ID) {
+            blocks_map[block.branch_true]->imm_predecessors.push_back(block.id);
+        }
+        if (block.branch_false != UNREACHABLE_BLOCK_ID) {
+            blocks_map[block.branch_false]->imm_predecessors.push_back(block.id);
+        }
+    }
+}
+
+void Function::BuildPostOrder() {
+    boost::container::small_vector<BlockId, 0x110> block_stack;
+    post_order_map.resize(NumBlocks());
+
+    Block& first_block{blocks_data[blocks.front()]};
+    first_block.post_order_visited = true;
+    block_stack.push_back(first_block.id);
+
+    const auto visit_branch = [&](BlockId block_id, BlockId branch_id) {
+        if (branch_id == UNREACHABLE_BLOCK_ID) {
+            return false;
+        }
+        if (blocks_map[branch_id]->post_order_visited) {
+            return false;
+        }
+        blocks_map[branch_id]->post_order_visited = true;
+
+        // Calling push_back twice is faster than insert on msvc
+        block_stack.push_back(block_id);
+        block_stack.push_back(branch_id);
+        return true;
+    };
+    while (!block_stack.empty()) {
+        const Block* const block{blocks_map[block_stack.back()]};
+        block_stack.pop_back();
+
+        if (!visit_branch(block->id, block->branch_true) &&
+            !visit_branch(block->id, block->branch_false)) {
+            post_order_map[block->id] = static_cast<u32>(post_order_blocks.size());
+            post_order_blocks.push_back(block->id);
+        }
+    }
+}
+
+void Function::BuildImmediateDominators() {
+    auto transform_block_id{std::views::transform([this](BlockId id) { return blocks_map[id]; })};
+    auto reverse_order_but_first{std::views::reverse | std::views::drop(1) | transform_block_id};
+    auto has_idom{std::views::filter([](Block* block) { return block->imm_dominator; })};
+    auto intersect{[this](Block* finger1, Block* finger2) {
+        while (finger1 != finger2) {
+            while (post_order_map[finger1->id] < post_order_map[finger2->id]) {
+                finger1 = finger1->imm_dominator;
+            }
+            while (post_order_map[finger2->id] < post_order_map[finger1->id]) {
+                finger2 = finger2->imm_dominator;
+            }
+        }
+        return finger1;
+    }};
+    for (Block& block : blocks_data) {
+        block.imm_dominator = nullptr;
+    }
+    Block* const start_block{&blocks_data[blocks.front()]};
+    start_block->imm_dominator = start_block;
+
+    bool changed{true};
+    while (changed) {
+        changed = false;
+        for (Block* const block : post_order_blocks | reverse_order_but_first) {
+            Block* new_idom{};
+            for (Block* predecessor : block->imm_predecessors | transform_block_id | has_idom) {
+                new_idom = new_idom ? intersect(predecessor, new_idom) : predecessor;
+            }
+            changed |= block->imm_dominator != new_idom;
+            block->imm_dominator = new_idom;
+        }
+    }
+}
+
+void Function::BuildDominanceFrontier() {
+    auto transform_block_id{std::views::transform([this](BlockId id) { return blocks_map[id]; })};
+    auto has_enough_predecessors{[](Block& block) { return block.imm_predecessors.size() >= 2; }};
+    for (Block& block : blocks_data | std::views::filter(has_enough_predecessors)) {
+        for (Block* current : block.imm_predecessors | transform_block_id) {
+            while (current != block.imm_dominator) {
+                current->dominance_frontiers.push_back(current->id);
+                current = current->imm_dominator;
+            }
+        }
+    }
+}
+
 CFG::CFG(Environment& env_, Location start_address) : env{env_} {
+    VisitFunctions(start_address);
+
+    for (Function& function : functions) {
+        function.BuildBlocksMap();
+        function.BuildImmediatePredecessors();
+        function.BuildPostOrder();
+        function.BuildImmediateDominators();
+        function.BuildDominanceFrontier();
+    }
+}
+
+void CFG::VisitFunctions(Location start_address) {
     functions.emplace_back(start_address);
     for (FunctionId function_id = 0; function_id < functions.size(); ++function_id) {
         while (!functions[function_id].labels.empty()) {
@@ -202,6 +318,7 @@ void CFG::AnalyzeLabel(FunctionId function_id, Label& label) {
         .cond{true},
         .branch_true{UNREACHABLE_BLOCK_ID},
         .branch_false{UNREACHABLE_BLOCK_ID},
+        .imm_predecessors{},
     };
     // Analyze instructions until it reaches an already visited block or there's a branch
     bool is_branch{false};
@@ -310,7 +427,7 @@ CFG::AnalysisState CFG::AnalyzeInst(Block& block, FunctionId function_id, Locati
         // Technically CAL pushes into PRET, but that's implicit in the function call for us
         // Insert the function into the list if it doesn't exist
         if (std::ranges::find(functions, cal_pc, &Function::entrypoint) == functions.end()) {
-            functions.push_back(cal_pc);
+            functions.emplace_back(cal_pc);
         }
         // Handle CAL like a regular instruction
         break;
@@ -352,6 +469,7 @@ void CFG::AnalyzeCondInst(Block& block, FunctionId function_id, Location pc,
         .cond{cond},
         .branch_true{conditional_block_id},
         .branch_false{UNREACHABLE_BLOCK_ID},
+        .imm_predecessors{},
     })};
     // Set the end properties of the conditional instruction and give it a new identity
     Block& conditional_block{block};
@@ -465,14 +583,14 @@ std::string CFG::Dot() const {
         dot += fmt::format("\t\tnode [style=filled];\n");
         for (const u32 block_index : function.blocks) {
             const Block& block{function.blocks_data[block_index]};
-            const std::string name{Name(block)};
+            const std::string name{NameOf(block)};
             const auto add_branch = [&](BlockId branch_id, bool add_label) {
                 const auto it{std::ranges::find(function.blocks_data, branch_id, &Block::id)};
                 dot += fmt::format("\t\t{}->", name);
                 if (it == function.blocks_data.end()) {
                     dot += fmt::format("\"Unknown label {}\"", branch_id);
                 } else {
-                    dot += Name(*it);
+                    dot += NameOf(*it);
                 };
                 if (add_label && block.cond != true && block.cond != false) {
                     dot += fmt::format(" [label=\"{}\"]", block.cond);
@@ -520,7 +638,7 @@ std::string CFG::Dot() const {
         if (functions.front().blocks.empty()) {
             dot += "Start;\n";
         } else {
-            dot += fmt::format("\tStart -> {};\n", Name(functions.front().blocks_data.front()));
+            dot += fmt::format("\tStart -> {};\n", NameOf(functions.front().blocks_data.front()));
         }
         dot += fmt::format("\tStart [shape=diamond];\n");
     }

src/shader_recompiler/frontend/maxwell/control_flow.h

@@ -70,6 +70,12 @@ struct Block {
     IR::Condition cond;
     BlockId branch_true;
     BlockId branch_false;
+    boost::container::small_vector<BlockId, 4> imm_predecessors;
+    boost::container::small_vector<BlockId, 8> dominance_frontiers;
+    union {
+        bool post_order_visited{false};
+        Block* imm_dominator;
+    };
 };
 
 struct Label {
@@ -81,11 +87,30 @@ struct Label {
 struct Function {
     Function(Location start_address);
 
+    void BuildBlocksMap();
+
+    void BuildImmediatePredecessors();
+
+    void BuildPostOrder();
+
+    void BuildImmediateDominators();
+
+    void BuildDominanceFrontier();
+
+    [[nodiscard]] size_t NumBlocks() const noexcept {
+        return static_cast<size_t>(current_block_id) + 1;
+    }
+
     Location entrypoint;
     BlockId current_block_id{0};
     boost::container::small_vector<Label, 16> labels;
     boost::container::small_vector<u32, 0x130> blocks;
     boost::container::small_vector<Block, 0x130> blocks_data;
+    // Translates from BlockId to block index
+    boost::container::small_vector<Block*, 0x130> blocks_map;
+
+    boost::container::small_vector<u32, 0x130> post_order_blocks;
+    boost::container::small_vector<BlockId, 0x130> post_order_map;
 };
 
 class CFG {
@@ -97,6 +122,12 @@ class CFG {
 public:
     explicit CFG(Environment& env, Location start_address);
 
+    CFG& operator=(const CFG&) = delete;
+    CFG(const CFG&) = delete;
+
+    CFG& operator=(CFG&&) = delete;
+    CFG(CFG&&) = delete;
+
     [[nodiscard]] std::string Dot() const;
 
     [[nodiscard]] std::span<const Function> Functions() const noexcept {
@@ -104,20 +135,22 @@ public:
     }
 
 private:
+    void VisitFunctions(Location start_address);
+
     void AnalyzeLabel(FunctionId function_id, Label& label);
 
     /// Inspect already visited blocks.
     /// Return true when the block has already been visited
-    [[nodiscard]] bool InspectVisitedBlocks(FunctionId function_id, const Label& label);
+    bool InspectVisitedBlocks(FunctionId function_id, const Label& label);
 
-    [[nodiscard]] AnalysisState AnalyzeInst(Block& block, FunctionId function_id, Location pc);
+    AnalysisState AnalyzeInst(Block& block, FunctionId function_id, Location pc);
 
     void AnalyzeCondInst(Block& block, FunctionId function_id, Location pc, EndClass insn_end_class,
                          IR::Condition cond);
 
     /// Return true when the branch instruction is confirmed to be a branch
-    [[nodiscard]] bool AnalyzeBranch(Block& block, FunctionId function_id, Location pc,
+    bool AnalyzeBranch(Block& block, FunctionId function_id, Location pc, Instruction inst,
-                                     Instruction inst, Opcode opcode);
+                       Opcode opcode);
 
     void AnalyzeBRA(Block& block, FunctionId function_id, Location pc, Instruction inst,
                     bool is_absolute);
@@ -126,8 +159,7 @@ private:
     AnalysisState AnalyzeEXIT(Block& block, FunctionId function_id, Location pc, Instruction inst);
 
     /// Return the branch target block id
-    [[nodiscard]] BlockId AddLabel(const Block& block, Stack stack, Location pc,
+    BlockId AddLabel(const Block& block, Stack stack, Location pc, FunctionId function_id);
-                                   FunctionId function_id);
 
     Environment& env;
     boost::container::small_vector<Function, 1> functions;

src/shader_recompiler/frontend/maxwell/program.cpp

@@ -8,40 +8,53 @@
 #include "shader_recompiler/frontend/maxwell/program.h"
 #include "shader_recompiler/frontend/maxwell/termination_code.h"
 #include "shader_recompiler/frontend/maxwell/translate/translate.h"
+#include "shader_recompiler/ir_opt/passes.h"
 
 namespace Shader::Maxwell {
+namespace {
+void TranslateCode(Environment& env, const Flow::Function& cfg_function, IR::Function& function,
+                   std::span<IR::Block*> block_map, IR::Block* block_memory) {
+    const size_t num_blocks{cfg_function.blocks.size()};
+    function.blocks.reserve(num_blocks);
 
-Program::Function::~Function() {
+    for (const Flow::BlockId block_id : cfg_function.blocks) {
-    std::ranges::for_each(blocks, &std::destroy_at<IR::Block>);
+        const Flow::Block& flow_block{cfg_function.blocks_data[block_id]};
+
+        function.blocks.emplace_back(std::construct_at(block_memory, Translate(env, flow_block)));
+        block_map[flow_block.id] = function.blocks.back().get();
+        ++block_memory;
+    }
 }
 
-Program::Program(Environment& env, const Flow::CFG& cfg) {
+void EmitTerminationInsts(const Flow::Function& cfg_function,
+                          std::span<IR::Block* const> block_map) {
+    for (const Flow::BlockId block_id : cfg_function.blocks) {
+        const Flow::Block& flow_block{cfg_function.blocks_data[block_id]};
+        EmitTerminationCode(flow_block, block_map);
+    }
+}
+
+void TranslateFunction(Environment& env, const Flow::Function& cfg_function, IR::Function& function,
+                       IR::Block* block_memory) {
     std::vector<IR::Block*> block_map;
+    block_map.resize(cfg_function.blocks_data.size());
+
+    TranslateCode(env, cfg_function, function, block_map, block_memory);
+    EmitTerminationInsts(cfg_function, block_map);
+}
+} // Anonymous namespace
+
+Program::Program(Environment& env, const Flow::CFG& cfg) {
     functions.reserve(cfg.Functions().size());
-
     for (const Flow::Function& cfg_function : cfg.Functions()) {
-        Function& function{functions.emplace_back()};
+        TranslateFunction(env, cfg_function, functions.emplace_back(),
-
+                          block_alloc_pool.allocate(cfg_function.blocks.size()));
-        const size_t num_blocks{cfg_function.blocks.size()};
+    }
-        IR::Block* block_memory{block_alloc_pool.allocate(num_blocks)};
+    std::ranges::for_each(functions, Optimization::SsaRewritePass);
-        function.blocks.reserve(num_blocks);
+    for (IR::Function& function : functions) {
-
+        Optimization::Invoke(Optimization::DeadCodeEliminationPass, function);
-        block_map.resize(cfg_function.blocks_data.size());
+        Optimization::Invoke(Optimization::IdentityRemovalPass, function);
-
+        // Optimization::Invoke(Optimization::VerificationPass, function);
-        // Visit the instructions of all blocks
-        for (const Flow::BlockId block_id : cfg_function.blocks) {
-            const Flow::Block& flow_block{cfg_function.blocks_data[block_id]};
-
-            IR::Block* const block{std::construct_at(block_memory, Translate(env, flow_block))};
-            ++block_memory;
-            function.blocks.push_back(block);
-            block_map[flow_block.id] = block;
-        }
-        // Now that all blocks are defined, emit the termination instructions
-        for (const Flow::BlockId block_id : cfg_function.blocks) {
-            const Flow::Block& flow_block{cfg_function.blocks_data[block_id]};
-            EmitTerminationCode(flow_block, block_map);
-        }
     }
 }
 
@@ -50,16 +63,16 @@ std::string DumpProgram(const Program& program) {
     std::map<const IR::Inst*, size_t> inst_to_index;
     std::map<const IR::Block*, size_t> block_to_index;
 
-    for (const Program::Function& function : program.functions) {
+    for (const IR::Function& function : program.functions) {
-        for (const IR::Block* const block : function.blocks) {
+        for (const auto& block : function.blocks) {
-            block_to_index.emplace(block, index);
+            block_to_index.emplace(block.get(), index);
             ++index;
         }
     }
     std::string ret;
-    for (const Program::Function& function : program.functions) {
+    for (const IR::Function& function : program.functions) {
         ret += fmt::format("Function\n");
-        for (const IR::Block* const block : function.blocks) {
+        for (const auto& block : function.blocks) {
             ret += IR::DumpBlock(*block, block_to_index, inst_to_index, index) + '\n';
         }
     }

src/shader_recompiler/frontend/maxwell/program.h

@@ -4,13 +4,16 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <vector>
 
+#include <boost/container/small_vector.hpp>
 #include <boost/pool/pool_alloc.hpp>
 
 #include "shader_recompiler/environment.h"
 #include "shader_recompiler/frontend/ir/basic_block.h"
+#include "shader_recompiler/frontend/ir/function.h"
 #include "shader_recompiler/frontend/maxwell/control_flow.h"
 
 namespace Shader::Maxwell {
@@ -22,16 +25,10 @@ public:
     explicit Program(Environment& env, const Flow::CFG& cfg);
 
 private:
-    struct Function {
-        ~Function();
-
-        std::vector<IR::Block*> blocks;
-    };
-
     boost::pool_allocator<IR::Block, boost::default_user_allocator_new_delete,
                           boost::details::pool::null_mutex>
         block_alloc_pool;
-    std::vector<Function> functions;
+    boost::container::small_vector<IR::Function, 1> functions;
 };
 
 [[nodiscard]] std::string DumpProgram(const Program& program);

src/shader_recompiler/frontend/maxwell/termination_code.cpp

@@ -47,12 +47,19 @@ static IR::U1 GetCond(IR::Condition cond, IR::IREmitter& ir) {
 
 static void EmitBranch(const Flow::Block& flow_block, std::span<IR::Block* const> block_map,
                        IR::IREmitter& ir) {
+    const auto add_immediate_predecessor = [&](Flow::BlockId label) {
+        block_map[label]->AddImmediatePredecessor(&ir.block);
+    };
     if (flow_block.cond == true) {
+        add_immediate_predecessor(flow_block.branch_true);
         return ir.Branch(block_map[flow_block.branch_true]);
     }
     if (flow_block.cond == false) {
+        add_immediate_predecessor(flow_block.branch_false);
         return ir.Branch(block_map[flow_block.branch_false]);
     }
+    add_immediate_predecessor(flow_block.branch_true);
+    add_immediate_predecessor(flow_block.branch_false);
     return ir.BranchConditional(GetCond(flow_block.cond, ir), block_map[flow_block.branch_true],
                                 block_map[flow_block.branch_false]);
 }

src/shader_recompiler/frontend/maxwell/termination_code.h

@@ -11,6 +11,7 @@
 
 namespace Shader::Maxwell {
 
+/// Emit termination instructions and collect immediate predecessors
 void EmitTerminationCode(const Flow::Block& flow_block, std::span<IR::Block* const> block_map);
 
 } // namespace Shader::Maxwell

src/shader_recompiler/frontend/maxwell/translate/impl/impl.h

@@ -208,7 +208,7 @@ public:
     void P2R_reg(u64 insn);
     void P2R_cbuf(u64 insn);
     void P2R_imm(u64 insn);
-    void PBK(u64 insn);
+    void PBK();
     void PCNT(u64 insn);
     void PEXIT(u64 insn);
     void PIXLD(u64 insn);
@@ -252,7 +252,7 @@ public:
     void SHR_reg(u64 insn);
     void SHR_cbuf(u64 insn);
     void SHR_imm(u64 insn);
-    void SSY(u64 insn);
+    void SSY();
     void ST(u64 insn);
     void STG(u64 insn);
     void STL(u64 insn);

src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp

@@ -762,7 +762,7 @@ void TranslatorVisitor::P2R_imm(u64) {
     ThrowNotImplemented(Opcode::P2R_imm);
 }
 
-void TranslatorVisitor::PBK(u64) {
+void TranslatorVisitor::PBK() {
     // PBK is a no-op
 }
 
@@ -938,8 +938,8 @@ void TranslatorVisitor::SHR_imm(u64) {
     ThrowNotImplemented(Opcode::SHR_imm);
 }
 
-void TranslatorVisitor::SSY(u64) {
+void TranslatorVisitor::SSY() {
-    ThrowNotImplemented(Opcode::SSY);
+    // SSY is a no-op
 }
 
 void TranslatorVisitor::ST(u64) {

src/shader_recompiler/ir_opt/identity_removal_pass.cpp

@@ -28,7 +28,6 @@ void IdentityRemovalPass(IR::Block& block) {
             ++inst;
         }
     }
-
     for (IR::Inst* const inst : to_invalidate) {
         inst->Invalidate();
     }

src/shader_recompiler/ir_opt/passes.h

@@ -5,12 +5,21 @@
 #pragma once
 
 #include "shader_recompiler/frontend/ir/basic_block.h"
+#include "shader_recompiler/frontend/ir/function.h"
 
 namespace Shader::Optimization {
 
+template <typename Func>
+void Invoke(Func&& func, IR::Function& function) {
+    for (const auto& block : function.blocks) {
+        func(*block);
+    }
+}
+
 void DeadCodeEliminationPass(IR::Block& block);
 void GetSetElimination(IR::Block& block);
 void IdentityRemovalPass(IR::Block& block);
+void SsaRewritePass(IR::Function& function);
 void VerificationPass(const IR::Block& block);
 
 } // namespace Shader::Optimization

src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp

@@ -0,0 +1,155 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+// This file implements the SSA rewriting algorithm proposed in
+//
+//      Simple and Efficient Construction of Static Single Assignment Form.
+//      Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013)
+//      In: Jhala R., De Bosschere K. (eds)
+//      Compiler Construction. CC 2013.
+//      Lecture Notes in Computer Science, vol 7791.
+//      Springer, Berlin, Heidelberg
+//
+//      https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6
+//
+
+#include <map>
+
+#include <boost/container/flat_map.hpp>
+
+#include "shader_recompiler/frontend/ir/basic_block.h"
+#include "shader_recompiler/frontend/ir/function.h"
+#include "shader_recompiler/frontend/ir/microinstruction.h"
+#include "shader_recompiler/frontend/ir/opcode.h"
+#include "shader_recompiler/frontend/ir/pred.h"
+#include "shader_recompiler/frontend/ir/reg.h"
+#include "shader_recompiler/ir_opt/passes.h"
+
+namespace Shader::Optimization {
+namespace {
+using ValueMap = boost::container::flat_map<IR::Block*, IR::Value, std::less<IR::Block*>>;
+
+struct DefTable {
+    [[nodiscard]] ValueMap& operator[](IR::Reg variable) noexcept {
+        return regs[IR::RegIndex(variable)];
+    }
+
+    [[nodiscard]] ValueMap& operator[](IR::Pred variable) noexcept {
+        return preds[IR::PredIndex(variable)];
+    }
+
+    std::array<ValueMap, IR::NUM_USER_REGS> regs;
+    std::array<ValueMap, IR::NUM_USER_PREDS> preds;
+};
+
+IR::Opcode UndefOpcode(IR::Reg) noexcept {
+    return IR::Opcode::Undef32;
+}
+
+IR::Opcode UndefOpcode(IR::Pred) noexcept {
+    return IR::Opcode::Undef1;
+}
+
+[[nodiscard]] bool IsPhi(const IR::Inst& inst) noexcept {
+    return inst.Opcode() == IR::Opcode::Phi;
+}
+
+class Pass {
+public:
+    void WriteVariable(auto variable, IR::Block* block, const IR::Value& value) {
+        current_def[variable].insert_or_assign(block, value);
+    }
+
+    IR::Value ReadVariable(auto variable, IR::Block* block) {
+        auto& def{current_def[variable]};
+        if (const auto it{def.find(block)}; it != def.end()) {
+            return it->second;
+        }
+        return ReadVariableRecursive(variable, block);
+    }
+
+private:
+    IR::Value ReadVariableRecursive(auto variable, IR::Block* block) {
+        IR::Value val;
+        if (const std::span preds{block->ImmediatePredecessors()}; preds.size() == 1) {
+            val = ReadVariable(variable, preds.front());
+        } else {
+            // Break potential cycles with operandless phi
+            val = IR::Value{&*block->PrependNewInst(block->begin(), IR::Opcode::Phi)};
+            WriteVariable(variable, block, val);
+            val = AddPhiOperands(variable, val, block);
+        }
+        WriteVariable(variable, block, val);
+        return val;
+    }
+
+    IR::Value AddPhiOperands(auto variable, const IR::Value& phi, IR::Block* block) {
+        for (IR::Block* const pred : block->ImmediatePredecessors()) {
+            phi.Inst()->AddPhiOperand(pred, ReadVariable(variable, pred));
+        }
+        return TryRemoveTrivialPhi(phi, block, UndefOpcode(variable));
+    }
+
+    IR::Value TryRemoveTrivialPhi(const IR::Value& phi, IR::Block* block, IR::Opcode undef_opcode) {
+        IR::Value same;
+        for (const auto& pair : phi.Inst()->PhiOperands()) {
+            const IR::Value& op{pair.second};
+            if (op == same || op == phi) {
+                // Unique value or self-reference
+                continue;
+            }
+            if (!same.IsEmpty()) {
+                // The phi merges at least two values: not trivial
+                return phi;
+            }
+            same = op;
+        }
+        if (same.IsEmpty()) {
+            // The phi is unreachable or in the start block
+            const auto first_not_phi{std::ranges::find_if_not(block->Instructions(), IsPhi)};
+            same = IR::Value{&*block->PrependNewInst(first_not_phi, undef_opcode)};
+        }
+        // Reroute all uses of phi to same and remove phi
+        phi.Inst()->ReplaceUsesWith(same);
+        // TODO: Try to recursively remove all phi users, which might have become trivial
+        return same;
+    }
+
+    DefTable current_def;
+};
+} // Anonymous namespace
+
+void SsaRewritePass(IR::Function& function) {
+    Pass pass;
+    for (const auto& block : function.blocks) {
+        for (IR::Inst& inst : block->Instructions()) {
+            switch (inst.Opcode()) {
+            case IR::Opcode::SetRegister:
+                if (const IR::Reg reg{inst.Arg(0).Reg()}; reg != IR::Reg::RZ) {
+                    pass.WriteVariable(reg, block.get(), inst.Arg(1));
+                }
+                break;
+            case IR::Opcode::SetPred:
+                if (const IR::Pred pred{inst.Arg(0).Pred()}; pred != IR::Pred::PT) {
+                    pass.WriteVariable(pred, block.get(), inst.Arg(1));
+                }
+                break;
+            case IR::Opcode::GetRegister:
+                if (const IR::Reg reg{inst.Arg(0).Reg()}; reg != IR::Reg::RZ) {
+                    inst.ReplaceUsesWith(pass.ReadVariable(reg, block.get()));
+                }
+                break;
+            case IR::Opcode::GetPred:
+                if (const IR::Pred pred{inst.Arg(0).Pred()}; pred != IR::Pred::PT) {
+                    inst.ReplaceUsesWith(pass.ReadVariable(pred, block.get()));
+                }
+                break;
+            default:
+                break;
+            }
+        }
+    }
+}
+
+} // namespace Shader::Optimization

src/shader_recompiler/main.cpp

@@ -35,12 +35,12 @@ void RunDatabase() {
     ForEachFile("D:\\Shaders\\Database", [&](const std::filesystem::path& path) {
         map.emplace_back(std::make_unique<FileEnvironment>(path.string().c_str()));
     });
-    for (int i = 0; i < 1; ++i) {
+    for (int i = 0; i < 300; ++i) {
         for (auto& env : map) {
             // fmt::print(stdout, "Decoding {}\n", path.string());
             const Location start_address{0};
             auto cfg{std::make_unique<Flow::CFG>(*env, start_address)};
-            // fmt::print(stdout, "{}\n", cfg.Dot());
+            // fmt::print(stdout, "{}\n", cfg->Dot());
             // IR::Program program{env, cfg};
             // Optimize(program);
             // const std::string code{EmitGLASM(program)};