Merge pull request #4009 from ogniK5377/macro-jit-prod

video_core: Implement Macro JIT

src/core/settings.h

@@ -474,6 +474,7 @@ struct Values {
     bool reporting_services;
     bool quest_flag;
     bool disable_cpu_opt;
+    bool disable_macro_jit;
 
     // BCAT
     std::string bcat_backend;

src/video_core/CMakeLists.txt

@@ -25,6 +25,12 @@ add_library(video_core STATIC
     engines/shader_bytecode.h
     engines/shader_header.h
     engines/shader_type.h
+    macro/macro.cpp
+    macro/macro.h
+    macro/macro_interpreter.cpp
+    macro/macro_interpreter.h
+    macro/macro_jit_x64.cpp
+    macro/macro_jit_x64.h
     fence_manager.h
     gpu.cpp
     gpu.h
@@ -36,8 +42,6 @@ add_library(video_core STATIC
     gpu_thread.h
     guest_driver.cpp
     guest_driver.h
-    macro_interpreter.cpp
-    macro_interpreter.h
     memory_manager.cpp
     memory_manager.h
     morton.cpp

src/video_core/engines/maxwell_3d.cpp

@@ -25,9 +25,8 @@ constexpr u32 MacroRegistersStart = 0xE00;
 Maxwell3D::Maxwell3D(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
                      MemoryManager& memory_manager)
     : system{system}, rasterizer{rasterizer}, memory_manager{memory_manager},
-      macro_interpreter{*this}, upload_state{memory_manager, regs.upload} {
+      macro_engine{GetMacroEngine(*this)}, upload_state{memory_manager, regs.upload} {
     dirty.flags.flip();
-
     InitializeRegisterDefaults();
 }
 
@@ -120,7 +119,7 @@ void Maxwell3D::InitializeRegisterDefaults() {
     mme_inline[MAXWELL3D_REG_INDEX(index_array.count)] = true;
 }
 
-void Maxwell3D::CallMacroMethod(u32 method, std::size_t num_parameters, const u32* parameters) {
+void Maxwell3D::CallMacroMethod(u32 method, const std::vector<u32>& parameters) {
     // Reset the current macro.
     executing_macro = 0;
 
@@ -129,7 +128,7 @@ void Maxwell3D::CallMacroMethod(u32 method, std::size_t num_parameters, const u3
         ((method - MacroRegistersStart) >> 1) % static_cast<u32>(macro_positions.size());
 
     // Execute the current macro.
-    macro_interpreter.Execute(macro_positions[entry], num_parameters, parameters);
+    macro_engine->Execute(macro_positions[entry], parameters);
     if (mme_draw.current_mode != MMEDrawMode::Undefined) {
         FlushMMEInlineDraw();
     }
@@ -165,7 +164,7 @@ void Maxwell3D::CallMethod(u32 method, u32 method_argument, bool is_last_call) {
 
         // Call the macro when there are no more parameters in the command buffer
         if (is_last_call) {
-            CallMacroMethod(executing_macro, macro_params.size(), macro_params.data());
+            CallMacroMethod(executing_macro, macro_params);
             macro_params.clear();
         }
         return;
@@ -201,7 +200,7 @@ void Maxwell3D::CallMethod(u32 method, u32 method_argument, bool is_last_call) {
         break;
     }
     case MAXWELL3D_REG_INDEX(macros.data): {
-        ProcessMacroUpload(arg);
+        macro_engine->AddCode(regs.macros.upload_address, arg);
         break;
     }
     case MAXWELL3D_REG_INDEX(macros.bind): {
@@ -310,7 +309,7 @@ void Maxwell3D::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
 
         // Call the macro when there are no more parameters in the command buffer
         if (amount == methods_pending) {
-            CallMacroMethod(executing_macro, macro_params.size(), macro_params.data());
+            CallMacroMethod(executing_macro, macro_params);
             macro_params.clear();
         }
         return;
@@ -424,9 +423,7 @@ void Maxwell3D::FlushMMEInlineDraw() {
 }
 
 void Maxwell3D::ProcessMacroUpload(u32 data) {
-    ASSERT_MSG(regs.macros.upload_address < macro_memory.size(),
+    macro_engine->AddCode(regs.macros.upload_address++, data);
-               "upload_address exceeded macro_memory size!");
-    macro_memory[regs.macros.upload_address++] = data;
 }
 
 void Maxwell3D::ProcessMacroBind(u32 data) {

src/video_core/engines/maxwell_3d.h

@@ -23,7 +23,7 @@
 #include "video_core/engines/engine_upload.h"
 #include "video_core/engines/shader_type.h"
 #include "video_core/gpu.h"
-#include "video_core/macro_interpreter.h"
+#include "video_core/macro/macro.h"
 #include "video_core/textures/texture.h"
 
 namespace Core {
@@ -1411,15 +1411,6 @@ public:
 
     const VideoCore::GuestDriverProfile& AccessGuestDriverProfile() const override;
 
-    /// Memory for macro code - it's undetermined how big this is, however 1MB is much larger than
-    /// we've seen used.
-    using MacroMemory = std::array<u32, 0x40000>;
-
-    /// Gets a reference to macro memory.
-    const MacroMemory& GetMacroMemory() const {
-        return macro_memory;
-    }
-
     bool ShouldExecute() const {
         return execute_on;
     }
@@ -1468,16 +1459,13 @@ private:
 
     std::array<bool, Regs::NUM_REGS> mme_inline{};
 
-    /// Memory for macro code
-    MacroMemory macro_memory;
-
     /// Macro method that is currently being executed / being fed parameters.
     u32 executing_macro = 0;
     /// Parameters that have been submitted to the macro call so far.
     std::vector<u32> macro_params;
 
     /// Interpreter for the macro codes uploaded to the GPU.
-    MacroInterpreter macro_interpreter;
+    std::unique_ptr<MacroEngine> macro_engine;
 
     static constexpr u32 null_cb_data = 0xFFFFFFFF;
     struct {
@@ -1506,7 +1494,7 @@ private:
      * @param num_parameters Number of arguments
      * @param parameters Arguments to the method call
      */
-    void CallMacroMethod(u32 method, std::size_t num_parameters, const u32* parameters);
+    void CallMacroMethod(u32 method, const std::vector<u32>& parameters);
 
     /// Handles writes to the macro uploading register.
     void ProcessMacroUpload(u32 data);

src/video_core/macro/macro.cpp

@@ -0,0 +1,45 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/assert.h"
+#include "common/logging/log.h"
+#include "core/settings.h"
+#include "video_core/macro/macro.h"
+#include "video_core/macro/macro_interpreter.h"
+#include "video_core/macro/macro_jit_x64.h"
+
+namespace Tegra {
+
+void MacroEngine::AddCode(u32 method, u32 data) {
+    uploaded_macro_code[method].push_back(data);
+}
+
+void MacroEngine::Execute(u32 method, const std::vector<u32>& parameters) {
+    auto compiled_macro = macro_cache.find(method);
+    if (compiled_macro != macro_cache.end()) {
+        compiled_macro->second->Execute(parameters, method);
+    } else {
+        // Macro not compiled, check if it's uploaded and if so, compile it
+        auto macro_code = uploaded_macro_code.find(method);
+        if (macro_code == uploaded_macro_code.end()) {
+            UNREACHABLE_MSG("Macro 0x{0:x} was not uploaded", method);
+            return;
+        }
+        macro_cache[method] = Compile(macro_code->second);
+        macro_cache[method]->Execute(parameters, method);
+    }
+}
+
+std::unique_ptr<MacroEngine> GetMacroEngine(Engines::Maxwell3D& maxwell3d) {
+    if (Settings::values.disable_macro_jit) {
+        return std::make_unique<MacroInterpreter>(maxwell3d);
+    }
+#ifdef ARCHITECTURE_x86_64
+    return std::make_unique<MacroJITx64>(maxwell3d);
+#else
+    return std::make_unique<MacroInterpreter>(maxwell3d);
+#endif
+}
+
+} // namespace Tegra

src/video_core/macro/macro.h

@@ -0,0 +1,128 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <unordered_map>
+#include <vector>
+#include "common/bit_field.h"
+#include "common/common_types.h"
+
+namespace Tegra {
+namespace Engines {
+class Maxwell3D;
+}
+namespace Macro {
+constexpr std::size_t NUM_MACRO_REGISTERS = 8;
+enum class Operation : u32 {
+    ALU = 0,
+    AddImmediate = 1,
+    ExtractInsert = 2,
+    ExtractShiftLeftImmediate = 3,
+    ExtractShiftLeftRegister = 4,
+    Read = 5,
+    Unused = 6, // This operation doesn't seem to be a valid encoding.
+    Branch = 7,
+};
+
+enum class ALUOperation : u32 {
+    Add = 0,
+    AddWithCarry = 1,
+    Subtract = 2,
+    SubtractWithBorrow = 3,
+    // Operations 4-7 don't seem to be valid encodings.
+    Xor = 8,
+    Or = 9,
+    And = 10,
+    AndNot = 11,
+    Nand = 12
+};
+
+enum class ResultOperation : u32 {
+    IgnoreAndFetch = 0,
+    Move = 1,
+    MoveAndSetMethod = 2,
+    FetchAndSend = 3,
+    MoveAndSend = 4,
+    FetchAndSetMethod = 5,
+    MoveAndSetMethodFetchAndSend = 6,
+    MoveAndSetMethodSend = 7
+};
+
+enum class BranchCondition : u32 {
+    Zero = 0,
+    NotZero = 1,
+};
+
+union Opcode {
+    u32 raw;
+    BitField<0, 3, Operation> operation;
+    BitField<4, 3, ResultOperation> result_operation;
+    BitField<4, 1, BranchCondition> branch_condition;
+    // If set on a branch, then the branch doesn't have a delay slot.
+    BitField<5, 1, u32> branch_annul;
+    BitField<7, 1, u32> is_exit;
+    BitField<8, 3, u32> dst;
+    BitField<11, 3, u32> src_a;
+    BitField<14, 3, u32> src_b;
+    // The signed immediate overlaps the second source operand and the alu operation.
+    BitField<14, 18, s32> immediate;
+
+    BitField<17, 5, ALUOperation> alu_operation;
+
+    // Bitfield instructions data
+    BitField<17, 5, u32> bf_src_bit;
+    BitField<22, 5, u32> bf_size;
+    BitField<27, 5, u32> bf_dst_bit;
+
+    u32 GetBitfieldMask() const {
+        return (1 << bf_size) - 1;
+    }
+
+    s32 GetBranchTarget() const {
+        return static_cast<s32>(immediate * sizeof(u32));
+    }
+};
+
+union MethodAddress {
+    u32 raw;
+    BitField<0, 12, u32> address;
+    BitField<12, 6, u32> increment;
+};
+
+} // namespace Macro
+
+class CachedMacro {
+public:
+    virtual ~CachedMacro() = default;
+    /**
+     * Executes the macro code with the specified input parameters.
+     * @param code The macro byte code to execute
+     * @param parameters The parameters of the macro
+     */
+    virtual void Execute(const std::vector<u32>& parameters, u32 method) = 0;
+};
+
+class MacroEngine {
+public:
+    virtual ~MacroEngine() = default;
+
+    // Store the uploaded macro code to compile them when they're called.
+    void AddCode(u32 method, u32 data);
+
+    // Compiles the macro if its not in the cache, and executes the compiled macro
+    void Execute(u32 method, const std::vector<u32>& parameters);
+
+protected:
+    virtual std::unique_ptr<CachedMacro> Compile(const std::vector<u32>& code) = 0;
+
+private:
+    std::unordered_map<u32, std::unique_ptr<CachedMacro>> macro_cache;
+    std::unordered_map<u32, std::vector<u32>> uploaded_macro_code;
+};
+
+std::unique_ptr<MacroEngine> GetMacroEngine(Engines::Maxwell3D& maxwell3d);
+
+} // namespace Tegra

src/video_core/macro/macro_interpreter.cpp

@@ -1,4 +1,4 @@
-// Copyright 2018 yuzu Emulator Project
+// Copyright 2020 yuzu Emulator Project
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
@@ -6,109 +6,46 @@
 #include "common/logging/log.h"
 #include "common/microprofile.h"
 #include "video_core/engines/maxwell_3d.h"
-#include "video_core/macro_interpreter.h"
+#include "video_core/macro/macro_interpreter.h"
 
 MICROPROFILE_DEFINE(MacroInterp, "GPU", "Execute macro interpreter", MP_RGB(128, 128, 192));
 
 namespace Tegra {
-namespace {
-enum class Operation : u32 {
-    ALU = 0,
-    AddImmediate = 1,
-    ExtractInsert = 2,
-    ExtractShiftLeftImmediate = 3,
-    ExtractShiftLeftRegister = 4,
-    Read = 5,
-    Unused = 6, // This operation doesn't seem to be a valid encoding.
-    Branch = 7,
-};
-} // Anonymous namespace
-
-enum class MacroInterpreter::ALUOperation : u32 {
-    Add = 0,
-    AddWithCarry = 1,
-    Subtract = 2,
-    SubtractWithBorrow = 3,
-    // Operations 4-7 don't seem to be valid encodings.
-    Xor = 8,
-    Or = 9,
-    And = 10,
-    AndNot = 11,
-    Nand = 12
-};
-
-enum class MacroInterpreter::ResultOperation : u32 {
-    IgnoreAndFetch = 0,
-    Move = 1,
-    MoveAndSetMethod = 2,
-    FetchAndSend = 3,
-    MoveAndSend = 4,
-    FetchAndSetMethod = 5,
-    MoveAndSetMethodFetchAndSend = 6,
-    MoveAndSetMethodSend = 7
-};
-
-enum class MacroInterpreter::BranchCondition : u32 {
-    Zero = 0,
-    NotZero = 1,
-};
-
-union MacroInterpreter::Opcode {
-    u32 raw;
-    BitField<0, 3, Operation> operation;
-    BitField<4, 3, ResultOperation> result_operation;
-    BitField<4, 1, BranchCondition> branch_condition;
-    // If set on a branch, then the branch doesn't have a delay slot.
-    BitField<5, 1, u32> branch_annul;
-    BitField<7, 1, u32> is_exit;
-    BitField<8, 3, u32> dst;
-    BitField<11, 3, u32> src_a;
-    BitField<14, 3, u32> src_b;
-    // The signed immediate overlaps the second source operand and the alu operation.
-    BitField<14, 18, s32> immediate;
-
-    BitField<17, 5, ALUOperation> alu_operation;
-
-    // Bitfield instructions data
-    BitField<17, 5, u32> bf_src_bit;
-    BitField<22, 5, u32> bf_size;
-    BitField<27, 5, u32> bf_dst_bit;
-
-    u32 GetBitfieldMask() const {
-        return (1 << bf_size) - 1;
-    }
-
-    s32 GetBranchTarget() const {
-        return static_cast<s32>(immediate * sizeof(u32));
-    }
-};
-
 MacroInterpreter::MacroInterpreter(Engines::Maxwell3D& maxwell3d) : maxwell3d(maxwell3d) {}
 
-void MacroInterpreter::Execute(u32 offset, std::size_t num_parameters, const u32* parameters) {
+std::unique_ptr<CachedMacro> MacroInterpreter::Compile(const std::vector<u32>& code) {
+    return std::make_unique<MacroInterpreterImpl>(maxwell3d, code);
+}
+
+MacroInterpreterImpl::MacroInterpreterImpl(Engines::Maxwell3D& maxwell3d,
+                                           const std::vector<u32>& code)
+    : maxwell3d(maxwell3d), code(code) {}
+
+void MacroInterpreterImpl::Execute(const std::vector<u32>& parameters, u32 method) {
     MICROPROFILE_SCOPE(MacroInterp);
     Reset();
 
     registers[1] = parameters[0];
+    num_parameters = parameters.size();
 
     if (num_parameters > parameters_capacity) {
         parameters_capacity = num_parameters;
         this->parameters = std::make_unique<u32[]>(num_parameters);
     }
-    std::memcpy(this->parameters.get(), parameters, num_parameters * sizeof(u32));
+    std::memcpy(this->parameters.get(), parameters.data(), num_parameters * sizeof(u32));
     this->num_parameters = num_parameters;
 
     // Execute the code until we hit an exit condition.
     bool keep_executing = true;
     while (keep_executing) {
-        keep_executing = Step(offset, false);
+        keep_executing = Step(false);
     }
 
     // Assert the the macro used all the input parameters
     ASSERT(next_parameter_index == num_parameters);
 }
 
-void MacroInterpreter::Reset() {
+void MacroInterpreterImpl::Reset() {
     registers = {};
     pc = 0;
     delayed_pc = {};
@@ -120,10 +57,10 @@ void MacroInterpreter::Reset() {
     carry_flag = false;
 }
 
-bool MacroInterpreter::Step(u32 offset, bool is_delay_slot) {
+bool MacroInterpreterImpl::Step(bool is_delay_slot) {
     u32 base_address = pc;
 
-    Opcode opcode = GetOpcode(offset);
+    Macro::Opcode opcode = GetOpcode();
     pc += 4;
 
     // Update the program counter if we were delayed
@@ -134,18 +71,18 @@ bool MacroInterpreter::Step(u32 offset, bool is_delay_slot) {
     }
 
     switch (opcode.operation) {
-    case Operation::ALU: {
+    case Macro::Operation::ALU: {
         u32 result = GetALUResult(opcode.alu_operation, GetRegister(opcode.src_a),
                                   GetRegister(opcode.src_b));
         ProcessResult(opcode.result_operation, opcode.dst, result);
         break;
     }
-    case Operation::AddImmediate: {
+    case Macro::Operation::AddImmediate: {
         ProcessResult(opcode.result_operation, opcode.dst,
                       GetRegister(opcode.src_a) + opcode.immediate);
         break;
     }
-    case Operation::ExtractInsert: {
+    case Macro::Operation::ExtractInsert: {
         u32 dst = GetRegister(opcode.src_a);
         u32 src = GetRegister(opcode.src_b);
 
@@ -155,7 +92,7 @@ bool MacroInterpreter::Step(u32 offset, bool is_delay_slot) {
         ProcessResult(opcode.result_operation, opcode.dst, dst);
         break;
     }
-    case Operation::ExtractShiftLeftImmediate: {
+    case Macro::Operation::ExtractShiftLeftImmediate: {
         u32 dst = GetRegister(opcode.src_a);
         u32 src = GetRegister(opcode.src_b);
 
@@ -164,7 +101,7 @@ bool MacroInterpreter::Step(u32 offset, bool is_delay_slot) {
         ProcessResult(opcode.result_operation, opcode.dst, result);
         break;
     }
-    case Operation::ExtractShiftLeftRegister: {
+    case Macro::Operation::ExtractShiftLeftRegister: {
         u32 dst = GetRegister(opcode.src_a);
         u32 src = GetRegister(opcode.src_b);
 
@@ -173,12 +110,12 @@ bool MacroInterpreter::Step(u32 offset, bool is_delay_slot) {
         ProcessResult(opcode.result_operation, opcode.dst, result);
         break;
     }
-    case Operation::Read: {
+    case Macro::Operation::Read: {
         u32 result = Read(GetRegister(opcode.src_a) + opcode.immediate);
         ProcessResult(opcode.result_operation, opcode.dst, result);
         break;
     }
-    case Operation::Branch: {
+    case Macro::Operation::Branch: {
         ASSERT_MSG(!is_delay_slot, "Executing a branch in a delay slot is not valid");
         u32 value = GetRegister(opcode.src_a);
         bool taken = EvaluateBranchCondition(opcode.branch_condition, value);
@@ -191,7 +128,7 @@ bool MacroInterpreter::Step(u32 offset, bool is_delay_slot) {
 
             delayed_pc = base_address + opcode.GetBranchTarget();
             // Execute one more instruction due to the delay slot.
-            return Step(offset, true);
+            return Step(true);
         }
         break;
     }
@@ -204,51 +141,44 @@ bool MacroInterpreter::Step(u32 offset, bool is_delay_slot) {
     // cause an exit if it's executed inside a delay slot.
     if (opcode.is_exit && !is_delay_slot) {
         // Exit has a delay slot, execute the next instruction
-        Step(offset, true);
+        Step(true);
         return false;
     }
 
     return true;
 }
 
-MacroInterpreter::Opcode MacroInterpreter::GetOpcode(u32 offset) const {
+u32 MacroInterpreterImpl::GetALUResult(Macro::ALUOperation operation, u32 src_a, u32 src_b) {
-    const auto& macro_memory{maxwell3d.GetMacroMemory()};
-    ASSERT((pc % sizeof(u32)) == 0);
-    ASSERT((pc + offset) < macro_memory.size() * sizeof(u32));
-    return {macro_memory[offset + pc / sizeof(u32)]};
-}
-
-u32 MacroInterpreter::GetALUResult(ALUOperation operation, u32 src_a, u32 src_b) {
     switch (operation) {
-    case ALUOperation::Add: {
+    case Macro::ALUOperation::Add: {
         const u64 result{static_cast<u64>(src_a) + src_b};
         carry_flag = result > 0xffffffff;
         return static_cast<u32>(result);
     }
-    case ALUOperation::AddWithCarry: {
+    case Macro::ALUOperation::AddWithCarry: {
         const u64 result{static_cast<u64>(src_a) + src_b + (carry_flag ? 1ULL : 0ULL)};
         carry_flag = result > 0xffffffff;
         return static_cast<u32>(result);
     }
-    case ALUOperation::Subtract: {
+    case Macro::ALUOperation::Subtract: {
         const u64 result{static_cast<u64>(src_a) - src_b};
         carry_flag = result < 0x100000000;
         return static_cast<u32>(result);
     }
-    case ALUOperation::SubtractWithBorrow: {
+    case Macro::ALUOperation::SubtractWithBorrow: {
         const u64 result{static_cast<u64>(src_a) - src_b - (carry_flag ? 0ULL : 1ULL)};
         carry_flag = result < 0x100000000;
         return static_cast<u32>(result);
     }
-    case ALUOperation::Xor:
+    case Macro::ALUOperation::Xor:
         return src_a ^ src_b;
-    case ALUOperation::Or:
+    case Macro::ALUOperation::Or:
         return src_a | src_b;
-    case ALUOperation::And:
+    case Macro::ALUOperation::And:
         return src_a & src_b;
-    case ALUOperation::AndNot:
+    case Macro::ALUOperation::AndNot:
         return src_a & ~src_b;
-    case ALUOperation::Nand:
+    case Macro::ALUOperation::Nand:
         return ~(src_a & src_b);
 
     default:
@@ -257,43 +187,43 @@ u32 MacroInterpreter::GetALUResult(ALUOperation operation, u32 src_a, u32 src_b)
     }
 }
 
-void MacroInterpreter::ProcessResult(ResultOperation operation, u32 reg, u32 result) {
+void MacroInterpreterImpl::ProcessResult(Macro::ResultOperation operation, u32 reg, u32 result) {
     switch (operation) {
-    case ResultOperation::IgnoreAndFetch:
+    case Macro::ResultOperation::IgnoreAndFetch:
         // Fetch parameter and ignore result.
         SetRegister(reg, FetchParameter());
         break;
-    case ResultOperation::Move:
+    case Macro::ResultOperation::Move:
         // Move result.
         SetRegister(reg, result);
         break;
-    case ResultOperation::MoveAndSetMethod:
+    case Macro::ResultOperation::MoveAndSetMethod:
         // Move result and use as Method Address.
         SetRegister(reg, result);
         SetMethodAddress(result);
         break;
-    case ResultOperation::FetchAndSend:
+    case Macro::ResultOperation::FetchAndSend:
         // Fetch parameter and send result.
         SetRegister(reg, FetchParameter());
         Send(result);
         break;
-    case ResultOperation::MoveAndSend:
+    case Macro::ResultOperation::MoveAndSend:
         // Move and send result.
         SetRegister(reg, result);
         Send(result);
         break;
-    case ResultOperation::FetchAndSetMethod:
+    case Macro::ResultOperation::FetchAndSetMethod:
         // Fetch parameter and use result as Method Address.
         SetRegister(reg, FetchParameter());
         SetMethodAddress(result);
         break;
-    case ResultOperation::MoveAndSetMethodFetchAndSend:
+    case Macro::ResultOperation::MoveAndSetMethodFetchAndSend:
         // Move result and use as Method Address, then fetch and send parameter.
         SetRegister(reg, result);
         SetMethodAddress(result);
         Send(FetchParameter());
         break;
-    case ResultOperation::MoveAndSetMethodSend:
+    case Macro::ResultOperation::MoveAndSetMethodSend:
         // Move result and use as Method Address, then send bits 12:17 of result.
         SetRegister(reg, result);
         SetMethodAddress(result);
@@ -304,16 +234,28 @@ void MacroInterpreter::ProcessResult(ResultOperation operation, u32 reg, u32 res
     }
 }
 
-u32 MacroInterpreter::FetchParameter() {
+bool MacroInterpreterImpl::EvaluateBranchCondition(Macro::BranchCondition cond, u32 value) const {
-    ASSERT(next_parameter_index < num_parameters);
+    switch (cond) {
-    return parameters[next_parameter_index++];
+    case Macro::BranchCondition::Zero:
+        return value == 0;
+    case Macro::BranchCondition::NotZero:
+        return value != 0;
+    }
+    UNREACHABLE();
+    return true;
 }
 
-u32 MacroInterpreter::GetRegister(u32 register_id) const {
+Macro::Opcode MacroInterpreterImpl::GetOpcode() const {
+    ASSERT((pc % sizeof(u32)) == 0);
+    ASSERT(pc < code.size() * sizeof(u32));
+    return {code[pc / sizeof(u32)]};
+}
+
+u32 MacroInterpreterImpl::GetRegister(u32 register_id) const {
     return registers.at(register_id);
 }
 
-void MacroInterpreter::SetRegister(u32 register_id, u32 value) {
+void MacroInterpreterImpl::SetRegister(u32 register_id, u32 value) {
     // Register 0 is hardwired as the zero register.
     // Ensure no writes to it actually occur.
     if (register_id == 0) {
@@ -323,30 +265,24 @@ void MacroInterpreter::SetRegister(u32 register_id, u32 value) {
     registers.at(register_id) = value;
 }
 
-void MacroInterpreter::SetMethodAddress(u32 address) {
+void MacroInterpreterImpl::SetMethodAddress(u32 address) {
     method_address.raw = address;
 }
 
-void MacroInterpreter::Send(u32 value) {
+void MacroInterpreterImpl::Send(u32 value) {
     maxwell3d.CallMethodFromMME(method_address.address, value);
     // Increment the method address by the method increment.
     method_address.address.Assign(method_address.address.Value() +
                                   method_address.increment.Value());
 }
 
-u32 MacroInterpreter::Read(u32 method) const {
+u32 MacroInterpreterImpl::Read(u32 method) const {
     return maxwell3d.GetRegisterValue(method);
 }
 
-bool MacroInterpreter::EvaluateBranchCondition(BranchCondition cond, u32 value) const {
+u32 MacroInterpreterImpl::FetchParameter() {
-    switch (cond) {
+    ASSERT(next_parameter_index < num_parameters);
-    case BranchCondition::Zero:
+    return parameters[next_parameter_index++];
-        return value == 0;
-    case BranchCondition::NotZero:
-        return value != 0;
-    }
-    UNREACHABLE();
-    return true;
 }
 
 } // namespace Tegra

src/video_core/macro/macro_interpreter.h

@@ -1,44 +1,37 @@
-// Copyright 2018 yuzu Emulator Project
+// Copyright 2020 yuzu Emulator Project
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
 #pragma once
-
 #include <array>
 #include <optional>
-
+#include <vector>
 #include "common/bit_field.h"
 #include "common/common_types.h"
+#include "video_core/macro/macro.h"
 
 namespace Tegra {
 namespace Engines {
 class Maxwell3D;
 }
 
-class MacroInterpreter final {
+class MacroInterpreter final : public MacroEngine {
 public:
     explicit MacroInterpreter(Engines::Maxwell3D& maxwell3d);
 
-    /**
+protected:
-     * Executes the macro code with the specified input parameters.
+    std::unique_ptr<CachedMacro> Compile(const std::vector<u32>& code) override;
-     * @param offset Offset to start execution at.
-     * @param parameters The parameters of the macro.
-     */
-    void Execute(u32 offset, std::size_t num_parameters, const u32* parameters);
 
 private:
-    enum class ALUOperation : u32;
+    Engines::Maxwell3D& maxwell3d;
-    enum class BranchCondition : u32;
+};
-    enum class ResultOperation : u32;
 
-    union Opcode;
+class MacroInterpreterImpl : public CachedMacro {
-
+public:
-    union MethodAddress {
+    MacroInterpreterImpl(Engines::Maxwell3D& maxwell3d, const std::vector<u32>& code);
-        u32 raw;
+    void Execute(const std::vector<u32>& parameters, u32 method) override;
-        BitField<0, 12, u32> address;
-        BitField<12, 6, u32> increment;
-    };
 
+private:
     /// Resets the execution engine state, zeroing registers, etc.
     void Reset();
 
@@ -49,20 +42,20 @@ private:
      * @param is_delay_slot Whether the current step is being executed due to a delay slot in a
      * previous instruction.
      */
-    bool Step(u32 offset, bool is_delay_slot);
+    bool Step(bool is_delay_slot);
 
     /// Calculates the result of an ALU operation. src_a OP src_b;
-    u32 GetALUResult(ALUOperation operation, u32 src_a, u32 src_b);
+    u32 GetALUResult(Macro::ALUOperation operation, u32 src_a, u32 src_b);
 
     /// Performs the result operation on the input result and stores it in the specified register
     /// (if necessary).
-    void ProcessResult(ResultOperation operation, u32 reg, u32 result);
+    void ProcessResult(Macro::ResultOperation operation, u32 reg, u32 result);
 
     /// Evaluates the branch condition and returns whether the branch should be taken or not.
-    bool EvaluateBranchCondition(BranchCondition cond, u32 value) const;
+    bool EvaluateBranchCondition(Macro::BranchCondition cond, u32 value) const;
 
     /// Reads an opcode at the current program counter location.
-    Opcode GetOpcode(u32 offset) const;
+    Macro::Opcode GetOpcode() const;
 
     /// Returns the specified register's value. Register 0 is hardcoded to always return 0.
     u32 GetRegister(u32 register_id) const;
@@ -89,13 +82,11 @@ private:
     /// Program counter to execute at after the delay slot is executed.
     std::optional<u32> delayed_pc;
 
-    static constexpr std::size_t NumMacroRegisters = 8;
-
     /// General purpose macro registers.
-    std::array<u32, NumMacroRegisters> registers = {};
+    std::array<u32, Macro::NUM_MACRO_REGISTERS> registers = {};
 
     /// Method address to use for the next Send instruction.
-    MethodAddress method_address = {};
+    Macro::MethodAddress method_address = {};
 
     /// Input parameters of the current macro.
     std::unique_ptr<u32[]> parameters;
@@ -105,5 +96,7 @@ private:
     u32 next_parameter_index = 0;
 
     bool carry_flag = false;
+    const std::vector<u32>& code;
 };
+
 } // namespace Tegra

src/video_core/macro/macro_jit_x64.cpp

@@ -0,0 +1,640 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/assert.h"
+#include "common/logging/log.h"
+#include "common/microprofile.h"
+#include "common/x64/xbyak_util.h"
+#include "video_core/engines/maxwell_3d.h"
+#include "video_core/macro/macro_interpreter.h"
+#include "video_core/macro/macro_jit_x64.h"
+
+MICROPROFILE_DEFINE(MacroJitCompile, "GPU", "Compile macro JIT", MP_RGB(173, 255, 47));
+MICROPROFILE_DEFINE(MacroJitExecute, "GPU", "Execute macro JIT", MP_RGB(255, 255, 0));
+
+namespace Tegra {
+static const Xbyak::Reg64 PARAMETERS = Xbyak::util::r9;
+static const Xbyak::Reg64 REGISTERS = Xbyak::util::r10;
+static const Xbyak::Reg64 STATE = Xbyak::util::r11;
+static const Xbyak::Reg64 NEXT_PARAMETER = Xbyak::util::r12;
+static const Xbyak::Reg32 RESULT = Xbyak::util::r13d;
+static const Xbyak::Reg64 RESULT_64 = Xbyak::util::r13;
+static const Xbyak::Reg32 METHOD_ADDRESS = Xbyak::util::r14d;
+static const Xbyak::Reg64 METHOD_ADDRESS_64 = Xbyak::util::r14;
+static const Xbyak::Reg64 BRANCH_HOLDER = Xbyak::util::r15;
+
+static const std::bitset<32> PERSISTENT_REGISTERS = Common::X64::BuildRegSet({
+    PARAMETERS,
+    REGISTERS,
+    STATE,
+    NEXT_PARAMETER,
+    RESULT,
+    METHOD_ADDRESS,
+    BRANCH_HOLDER,
+});
+
+MacroJITx64::MacroJITx64(Engines::Maxwell3D& maxwell3d) : maxwell3d(maxwell3d) {}
+
+std::unique_ptr<CachedMacro> MacroJITx64::Compile(const std::vector<u32>& code) {
+    return std::make_unique<MacroJITx64Impl>(maxwell3d, code);
+}
+
+MacroJITx64Impl::MacroJITx64Impl(Engines::Maxwell3D& maxwell3d, const std::vector<u32>& code)
+    : Xbyak::CodeGenerator(MAX_CODE_SIZE), code(code), maxwell3d(maxwell3d) {
+    Compile();
+}
+
+MacroJITx64Impl::~MacroJITx64Impl() = default;
+
+void MacroJITx64Impl::Execute(const std::vector<u32>& parameters, u32 method) {
+    MICROPROFILE_SCOPE(MacroJitExecute);
+    ASSERT_OR_EXECUTE(program != nullptr, { return; });
+    JITState state{};
+    state.maxwell3d = &maxwell3d;
+    state.registers = {};
+    state.parameters = parameters.data();
+    program(&state);
+}
+
+void MacroJITx64Impl::Compile_ALU(Macro::Opcode opcode) {
+    const bool is_a_zero = opcode.src_a == 0;
+    const bool is_b_zero = opcode.src_b == 0;
+    const bool valid_operation = !is_a_zero && !is_b_zero;
+    const bool is_move_operation = !is_a_zero && is_b_zero;
+    const bool has_zero_register = is_a_zero || is_b_zero;
+
+    Xbyak::Reg64 src_a;
+    Xbyak::Reg32 src_b;
+
+    if (!optimizer.zero_reg_skip) {
+        src_a = Compile_GetRegister(opcode.src_a, RESULT_64);
+        src_b = Compile_GetRegister(opcode.src_b, ebx);
+    } else {
+        if (!is_a_zero) {
+            src_a = Compile_GetRegister(opcode.src_a, RESULT_64);
+        }
+        if (!is_b_zero) {
+            src_b = Compile_GetRegister(opcode.src_b, ebx);
+        }
+    }
+    Xbyak::Label skip_carry{};
+
+    bool has_emitted = false;
+
+    switch (opcode.alu_operation) {
+    case Macro::ALUOperation::Add:
+        if (optimizer.zero_reg_skip) {
+            if (valid_operation) {
+                add(src_a, src_b);
+            }
+        } else {
+            add(src_a, src_b);
+        }
+
+        if (!optimizer.can_skip_carry) {
+            setc(byte[STATE + offsetof(JITState, carry_flag)]);
+        }
+        break;
+    case Macro::ALUOperation::AddWithCarry:
+        bt(dword[STATE + offsetof(JITState, carry_flag)], 0);
+        adc(src_a, src_b);
+        setc(byte[STATE + offsetof(JITState, carry_flag)]);
+        break;
+    case Macro::ALUOperation::Subtract:
+        if (optimizer.zero_reg_skip) {
+            if (valid_operation) {
+                sub(src_a, src_b);
+                has_emitted = true;
+            }
+        } else {
+            sub(src_a, src_b);
+            has_emitted = true;
+        }
+        if (!optimizer.can_skip_carry && has_emitted) {
+            setc(byte[STATE + offsetof(JITState, carry_flag)]);
+        }
+        break;
+    case Macro::ALUOperation::SubtractWithBorrow:
+        bt(dword[STATE + offsetof(JITState, carry_flag)], 0);
+        sbb(src_a, src_b);
+        setc(byte[STATE + offsetof(JITState, carry_flag)]);
+        break;
+    case Macro::ALUOperation::Xor:
+        if (optimizer.zero_reg_skip) {
+            if (valid_operation) {
+                xor_(src_a, src_b);
+            }
+        } else {
+            xor_(src_a, src_b);
+        }
+        break;
+    case Macro::ALUOperation::Or:
+        if (optimizer.zero_reg_skip) {
+            if (valid_operation) {
+                or_(src_a, src_b);
+            }
+        } else {
+            or_(src_a, src_b);
+        }
+        break;
+    case Macro::ALUOperation::And:
+        if (optimizer.zero_reg_skip) {
+            if (!has_zero_register) {
+                and_(src_a, src_b);
+            }
+        } else {
+            and_(src_a, src_b);
+        }
+        break;
+    case Macro::ALUOperation::AndNot:
+        if (optimizer.zero_reg_skip) {
+            if (!is_a_zero) {
+                not_(src_b);
+                and_(src_a, src_b);
+            }
+        } else {
+            not_(src_b);
+            and_(src_a, src_b);
+        }
+        break;
+    case Macro::ALUOperation::Nand:
+        if (optimizer.zero_reg_skip) {
+            if (!is_a_zero) {
+                and_(src_a, src_b);
+                not_(src_a);
+            }
+        } else {
+            and_(src_a, src_b);
+            not_(src_a);
+        }
+        break;
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented ALU operation {}",
+                          static_cast<std::size_t>(opcode.alu_operation.Value()));
+        break;
+    }
+    Compile_ProcessResult(opcode.result_operation, opcode.dst);
+}
+
+void MacroJITx64Impl::Compile_AddImmediate(Macro::Opcode opcode) {
+    if (optimizer.skip_dummy_addimmediate) {
+        // Games tend to use this as an exit instruction placeholder. It's to encode an instruction
+        // without doing anything. In our case we can just not emit anything.
+        if (opcode.result_operation == Macro::ResultOperation::Move && opcode.dst == 0) {
+            return;
+        }
+    }
+    // Check for redundant moves
+    if (optimizer.optimize_for_method_move &&
+        opcode.result_operation == Macro::ResultOperation::MoveAndSetMethod) {
+        if (next_opcode.has_value()) {
+            const auto next = *next_opcode;
+            if (next.result_operation == Macro::ResultOperation::MoveAndSetMethod) {
+                return;
+            }
+        }
+    }
+    if (optimizer.zero_reg_skip && opcode.src_a == 0) {
+        if (opcode.immediate == 0) {
+            xor_(RESULT, RESULT);
+        } else {
+            mov(RESULT, opcode.immediate);
+        }
+    } else {
+        auto result = Compile_GetRegister(opcode.src_a, RESULT);
+        if (opcode.immediate > 2) {
+            add(result, opcode.immediate);
+        } else if (opcode.immediate == 1) {
+            inc(result);
+        } else if (opcode.immediate < 0) {
+            sub(result, opcode.immediate * -1);
+        }
+    }
+    Compile_ProcessResult(opcode.result_operation, opcode.dst);
+}
+
+void MacroJITx64Impl::Compile_ExtractInsert(Macro::Opcode opcode) {
+    auto dst = Compile_GetRegister(opcode.src_a, RESULT);
+    auto src = Compile_GetRegister(opcode.src_b, eax);
+
+    if (opcode.bf_src_bit != 0 && opcode.bf_src_bit != 31) {
+        shr(src, opcode.bf_src_bit);
+    } else if (opcode.bf_src_bit == 31) {
+        xor_(src, src);
+    }
+    // Don't bother masking the whole register since we're using a 32 bit register
+    if (opcode.bf_size != 31 && opcode.bf_size != 0) {
+        and_(src, opcode.GetBitfieldMask());
+    } else if (opcode.bf_size == 0) {
+        xor_(src, src);
+    }
+    if (opcode.bf_dst_bit != 31 && opcode.bf_dst_bit != 0) {
+        shl(src, opcode.bf_dst_bit);
+    } else if (opcode.bf_dst_bit == 31) {
+        xor_(src, src);
+    }
+
+    const u32 mask = ~(opcode.GetBitfieldMask() << opcode.bf_dst_bit);
+    if (mask != 0xffffffff) {
+        and_(dst, mask);
+    }
+    or_(dst, src);
+    Compile_ProcessResult(opcode.result_operation, opcode.dst);
+}
+
+void MacroJITx64Impl::Compile_ExtractShiftLeftImmediate(Macro::Opcode opcode) {
+    auto dst = Compile_GetRegister(opcode.src_a, eax);
+    auto src = Compile_GetRegister(opcode.src_b, RESULT);
+
+    shr(src, al);
+    if (opcode.bf_size != 0 && opcode.bf_size != 31) {
+        and_(src, opcode.GetBitfieldMask());
+    } else if (opcode.bf_size == 0) {
+        xor_(src, src);
+    }
+
+    if (opcode.bf_dst_bit != 0 && opcode.bf_dst_bit != 31) {
+        shl(src, opcode.bf_dst_bit);
+    } else if (opcode.bf_dst_bit == 31) {
+        xor_(src, src);
+    }
+    Compile_ProcessResult(opcode.result_operation, opcode.dst);
+}
+
+void MacroJITx64Impl::Compile_ExtractShiftLeftRegister(Macro::Opcode opcode) {
+    auto dst = Compile_GetRegister(opcode.src_a, eax);
+    auto src = Compile_GetRegister(opcode.src_b, RESULT);
+
+    if (opcode.bf_src_bit != 0) {
+        shr(src, opcode.bf_src_bit);
+    }
+
+    if (opcode.bf_size != 31) {
+        and_(src, opcode.GetBitfieldMask());
+    }
+    shl(src, al);
+    Compile_ProcessResult(opcode.result_operation, opcode.dst);
+}
+
+static u32 Read(Engines::Maxwell3D* maxwell3d, u32 method) {
+    return maxwell3d->GetRegisterValue(method);
+}
+
+static void Send(Engines::Maxwell3D* maxwell3d, Macro::MethodAddress method_address, u32 value) {
+    maxwell3d->CallMethodFromMME(method_address.address, value);
+}
+
+void MacroJITx64Impl::Compile_Read(Macro::Opcode opcode) {
+    if (optimizer.zero_reg_skip && opcode.src_a == 0) {
+        if (opcode.immediate == 0) {
+            xor_(RESULT, RESULT);
+        } else {
+            mov(RESULT, opcode.immediate);
+        }
+    } else {
+        auto result = Compile_GetRegister(opcode.src_a, RESULT);
+        if (opcode.immediate > 2) {
+            add(result, opcode.immediate);
+        } else if (opcode.immediate == 1) {
+            inc(result);
+        } else if (opcode.immediate < 0) {
+            sub(result, opcode.immediate * -1);
+        }
+    }
+    Common::X64::ABI_PushRegistersAndAdjustStackGPS(*this, PersistentCallerSavedRegs(), 0);
+    mov(Common::X64::ABI_PARAM1, qword[STATE]);
+    mov(Common::X64::ABI_PARAM2, RESULT);
+    Common::X64::CallFarFunction(*this, &Read);
+    Common::X64::ABI_PopRegistersAndAdjustStackGPS(*this, PersistentCallerSavedRegs(), 0);
+    mov(RESULT, Common::X64::ABI_RETURN.cvt32());
+    Compile_ProcessResult(opcode.result_operation, opcode.dst);
+}
+
+void Tegra::MacroJITx64Impl::Compile_Send(Xbyak::Reg32 value) {
+    Common::X64::ABI_PushRegistersAndAdjustStackGPS(*this, PersistentCallerSavedRegs(), 0);
+    mov(Common::X64::ABI_PARAM1, qword[STATE]);
+    mov(Common::X64::ABI_PARAM2, METHOD_ADDRESS);
+    mov(Common::X64::ABI_PARAM3, value);
+    Common::X64::CallFarFunction(*this, &Send);
+    Common::X64::ABI_PopRegistersAndAdjustStackGPS(*this, PersistentCallerSavedRegs(), 0);
+
+    Xbyak::Label dont_process{};
+    // Get increment
+    test(METHOD_ADDRESS, 0x3f000);
+    // If zero, method address doesn't update
+    je(dont_process);
+
+    mov(ecx, METHOD_ADDRESS);
+    and_(METHOD_ADDRESS, 0xfff);
+    shr(ecx, 12);
+    and_(ecx, 0x3f);
+    lea(eax, ptr[rcx + METHOD_ADDRESS_64]);
+    sal(ecx, 12);
+    or_(eax, ecx);
+
+    mov(METHOD_ADDRESS, eax);
+
+    L(dont_process);
+}
+
+void Tegra::MacroJITx64Impl::Compile_Branch(Macro::Opcode opcode) {
+    ASSERT_MSG(!is_delay_slot, "Executing a branch in a delay slot is not valid");
+    const s32 jump_address =
+        static_cast<s32>(pc) + static_cast<s32>(opcode.GetBranchTarget() / sizeof(s32));
+
+    Xbyak::Label end;
+    auto value = Compile_GetRegister(opcode.src_a, eax);
+    test(value, value);
+    if (optimizer.has_delayed_pc) {
+        switch (opcode.branch_condition) {
+        case Macro::BranchCondition::Zero:
+            jne(end, T_NEAR);
+            break;
+        case Macro::BranchCondition::NotZero:
+            je(end, T_NEAR);
+            break;
+        }
+
+        if (opcode.branch_annul) {
+            xor_(BRANCH_HOLDER, BRANCH_HOLDER);
+            jmp(labels[jump_address], T_NEAR);
+        } else {
+            Xbyak::Label handle_post_exit{};
+            Xbyak::Label skip{};
+            jmp(skip, T_NEAR);
+            if (opcode.is_exit) {
+                L(handle_post_exit);
+                // Execute 1 instruction
+                mov(BRANCH_HOLDER, end_of_code);
+                // Jump to next instruction to skip delay slot check
+                jmp(labels[jump_address], T_NEAR);
+            } else {
+                L(handle_post_exit);
+                xor_(BRANCH_HOLDER, BRANCH_HOLDER);
+                jmp(labels[jump_address], T_NEAR);
+            }
+            L(skip);
+            mov(BRANCH_HOLDER, handle_post_exit);
+            jmp(delay_skip[pc], T_NEAR);
+        }
+    } else {
+        switch (opcode.branch_condition) {
+        case Macro::BranchCondition::Zero:
+            je(labels[jump_address], T_NEAR);
+            break;
+        case Macro::BranchCondition::NotZero:
+            jne(labels[jump_address], T_NEAR);
+            break;
+        }
+    }
+
+    L(end);
+}
+
+void Tegra::MacroJITx64Impl::Optimizer_ScanFlags() {
+    optimizer.can_skip_carry = true;
+    optimizer.has_delayed_pc = false;
+    for (auto raw_op : code) {
+        Macro::Opcode op{};
+        op.raw = raw_op;
+
+        if (op.operation == Macro::Operation::ALU) {
+            // Scan for any ALU operations which actually use the carry flag, if they don't exist in
+            // our current code we can skip emitting the carry flag handling operations
+            if (op.alu_operation == Macro::ALUOperation::AddWithCarry ||
+                op.alu_operation == Macro::ALUOperation::SubtractWithBorrow) {
+                optimizer.can_skip_carry = false;
+            }
+        }
+
+        if (op.operation == Macro::Operation::Branch) {
+            if (!op.branch_annul) {
+                optimizer.has_delayed_pc = true;
+            }
+        }
+    }
+}
+
+void MacroJITx64Impl::Compile() {
+    MICROPROFILE_SCOPE(MacroJitCompile);
+    bool keep_executing = true;
+    labels.fill(Xbyak::Label());
+
+    Common::X64::ABI_PushRegistersAndAdjustStackGPS(*this, Common::X64::ABI_ALL_CALLEE_SAVED, 8);
+    // JIT state
+    mov(STATE, Common::X64::ABI_PARAM1);
+    mov(PARAMETERS, qword[Common::X64::ABI_PARAM1 +
+                          static_cast<Xbyak::uint32>(offsetof(JITState, parameters))]);
+    mov(REGISTERS, Common::X64::ABI_PARAM1);
+    add(REGISTERS, static_cast<Xbyak::uint32>(offsetof(JITState, registers)));
+    xor_(RESULT, RESULT);
+    xor_(METHOD_ADDRESS, METHOD_ADDRESS);
+    xor_(NEXT_PARAMETER, NEXT_PARAMETER);
+    xor_(BRANCH_HOLDER, BRANCH_HOLDER);
+
+    mov(dword[REGISTERS + 4], Compile_FetchParameter());
+
+    // Track get register for zero registers and mark it as no-op
+    optimizer.zero_reg_skip = true;
+
+    // AddImmediate tends to be used as a NOP instruction, if we detect this we can
+    // completely skip the entire code path and no emit anything
+    optimizer.skip_dummy_addimmediate = true;
+
+    // SMO tends to emit a lot of unnecessary method moves, we can mitigate this by only emitting
+    // one if our register isn't "dirty"
+    optimizer.optimize_for_method_move = true;
+
+    // Check to see if we can skip emitting certain instructions
+    Optimizer_ScanFlags();
+
+    const u32 op_count = static_cast<u32>(code.size());
+    for (u32 i = 0; i < op_count; i++) {
+        if (i < op_count - 1) {
+            pc = i + 1;
+            next_opcode = GetOpCode();
+        } else {
+            next_opcode = {};
+        }
+        pc = i;
+        Compile_NextInstruction();
+    }
+
+    L(end_of_code);
+
+    Common::X64::ABI_PopRegistersAndAdjustStackGPS(*this, Common::X64::ABI_ALL_CALLEE_SAVED, 8);
+    ret();
+    ready();
+    program = getCode<ProgramType>();
+}
+
+bool MacroJITx64Impl::Compile_NextInstruction() {
+    const auto opcode = GetOpCode();
+    if (labels[pc].getAddress()) {
+        return false;
+    }
+
+    L(labels[pc]);
+
+    switch (opcode.operation) {
+    case Macro::Operation::ALU:
+        Compile_ALU(opcode);
+        break;
+    case Macro::Operation::AddImmediate:
+        Compile_AddImmediate(opcode);
+        break;
+    case Macro::Operation::ExtractInsert:
+        Compile_ExtractInsert(opcode);
+        break;
+    case Macro::Operation::ExtractShiftLeftImmediate:
+        Compile_ExtractShiftLeftImmediate(opcode);
+        break;
+    case Macro::Operation::ExtractShiftLeftRegister:
+        Compile_ExtractShiftLeftRegister(opcode);
+        break;
+    case Macro::Operation::Read:
+        Compile_Read(opcode);
+        break;
+    case Macro::Operation::Branch:
+        Compile_Branch(opcode);
+        break;
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented opcode {}", opcode.operation.Value());
+        break;
+    }
+
+    if (optimizer.has_delayed_pc) {
+        if (opcode.is_exit) {
+            mov(rax, end_of_code);
+            test(BRANCH_HOLDER, BRANCH_HOLDER);
+            cmove(BRANCH_HOLDER, rax);
+            // Jump to next instruction to skip delay slot check
+            je(labels[pc + 1], T_NEAR);
+        } else {
+            // TODO(ogniK): Optimize delay slot branching
+            Xbyak::Label no_delay_slot{};
+            test(BRANCH_HOLDER, BRANCH_HOLDER);
+            je(no_delay_slot, T_NEAR);
+            mov(rax, BRANCH_HOLDER);
+            xor_(BRANCH_HOLDER, BRANCH_HOLDER);
+            jmp(rax);
+            L(no_delay_slot);
+        }
+        L(delay_skip[pc]);
+        if (opcode.is_exit) {
+            return false;
+        }
+    } else {
+        test(BRANCH_HOLDER, BRANCH_HOLDER);
+        jne(end_of_code, T_NEAR);
+        if (opcode.is_exit) {
+            inc(BRANCH_HOLDER);
+            return false;
+        }
+    }
+    return true;
+}
+
+Xbyak::Reg32 Tegra::MacroJITx64Impl::Compile_FetchParameter() {
+    mov(eax, dword[PARAMETERS + NEXT_PARAMETER * sizeof(u32)]);
+    inc(NEXT_PARAMETER);
+    return eax;
+}
+
+Xbyak::Reg32 MacroJITx64Impl::Compile_GetRegister(u32 index, Xbyak::Reg32 dst) {
+    if (index == 0) {
+        // Register 0 is always zero
+        xor_(dst, dst);
+    } else {
+        mov(dst, dword[REGISTERS + index * sizeof(u32)]);
+    }
+
+    return dst;
+}
+
+Xbyak::Reg64 Tegra::MacroJITx64Impl::Compile_GetRegister(u32 index, Xbyak::Reg64 dst) {
+    if (index == 0) {
+        // Register 0 is always zero
+        xor_(dst, dst);
+    } else {
+        mov(dst, dword[REGISTERS + index * sizeof(u32)]);
+    }
+
+    return dst;
+}
+
+void Tegra::MacroJITx64Impl::Compile_WriteCarry(Xbyak::Reg64 dst) {
+    Xbyak::Label zero{}, end{};
+    xor_(ecx, ecx);
+    shr(dst, 32);
+    setne(cl);
+    mov(dword[STATE + offsetof(JITState, carry_flag)], ecx);
+}
+
+void MacroJITx64Impl::Compile_ProcessResult(Macro::ResultOperation operation, u32 reg) {
+    auto SetRegister = [=](u32 reg, Xbyak::Reg32 result) {
+        // Register 0 is supposed to always return 0. NOP is implemented as a store to the zero
+        // register.
+        if (reg == 0) {
+            return;
+        }
+        mov(dword[REGISTERS + reg * sizeof(u32)], result);
+    };
+    auto SetMethodAddress = [=](Xbyak::Reg32 reg) { mov(METHOD_ADDRESS, reg); };
+
+    switch (operation) {
+    case Macro::ResultOperation::IgnoreAndFetch:
+        SetRegister(reg, Compile_FetchParameter());
+        break;
+    case Macro::ResultOperation::Move:
+        SetRegister(reg, RESULT);
+        break;
+    case Macro::ResultOperation::MoveAndSetMethod:
+        SetRegister(reg, RESULT);
+        SetMethodAddress(RESULT);
+        break;
+    case Macro::ResultOperation::FetchAndSend:
+        // Fetch parameter and send result.
+        SetRegister(reg, Compile_FetchParameter());
+        Compile_Send(RESULT);
+        break;
+    case Macro::ResultOperation::MoveAndSend:
+        // Move and send result.
+        SetRegister(reg, RESULT);
+        Compile_Send(RESULT);
+        break;
+    case Macro::ResultOperation::FetchAndSetMethod:
+        // Fetch parameter and use result as Method Address.
+        SetRegister(reg, Compile_FetchParameter());
+        SetMethodAddress(RESULT);
+        break;
+    case Macro::ResultOperation::MoveAndSetMethodFetchAndSend:
+        // Move result and use as Method Address, then fetch and send parameter.
+        SetRegister(reg, RESULT);
+        SetMethodAddress(RESULT);
+        Compile_Send(Compile_FetchParameter());
+        break;
+    case Macro::ResultOperation::MoveAndSetMethodSend:
+        // Move result and use as Method Address, then send bits 12:17 of result.
+        SetRegister(reg, RESULT);
+        SetMethodAddress(RESULT);
+        shr(RESULT, 12);
+        and_(RESULT, 0b111111);
+        Compile_Send(RESULT);
+        break;
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented macro operation {}", static_cast<std::size_t>(operation));
+    }
+}
+
+Macro::Opcode MacroJITx64Impl::GetOpCode() const {
+    ASSERT(pc < code.size());
+    return {code[pc]};
+}
+
+std::bitset<32> MacroJITx64Impl::PersistentCallerSavedRegs() const {
+    return PERSISTENT_REGISTERS & Common::X64::ABI_ALL_CALLER_SAVED;
+}
+
+} // namespace Tegra

src/video_core/macro/macro_jit_x64.h

@@ -0,0 +1,100 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <array>
+#include <bitset>
+#include <xbyak.h>
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "common/x64/xbyak_abi.h"
+#include "video_core/macro/macro.h"
+
+namespace Tegra {
+
+namespace Engines {
+class Maxwell3D;
+}
+
+/// MAX_CODE_SIZE is arbitrarily chosen based on current booting games
+constexpr size_t MAX_CODE_SIZE = 0x10000;
+
+class MacroJITx64 final : public MacroEngine {
+public:
+    explicit MacroJITx64(Engines::Maxwell3D& maxwell3d);
+
+protected:
+    std::unique_ptr<CachedMacro> Compile(const std::vector<u32>& code) override;
+
+private:
+    Engines::Maxwell3D& maxwell3d;
+};
+
+class MacroJITx64Impl : public Xbyak::CodeGenerator, public CachedMacro {
+public:
+    MacroJITx64Impl(Engines::Maxwell3D& maxwell3d, const std::vector<u32>& code);
+    ~MacroJITx64Impl();
+
+    void Execute(const std::vector<u32>& parameters, u32 method) override;
+
+    void Compile_ALU(Macro::Opcode opcode);
+    void Compile_AddImmediate(Macro::Opcode opcode);
+    void Compile_ExtractInsert(Macro::Opcode opcode);
+    void Compile_ExtractShiftLeftImmediate(Macro::Opcode opcode);
+    void Compile_ExtractShiftLeftRegister(Macro::Opcode opcode);
+    void Compile_Read(Macro::Opcode opcode);
+    void Compile_Branch(Macro::Opcode opcode);
+
+private:
+    void Optimizer_ScanFlags();
+
+    void Compile();
+    bool Compile_NextInstruction();
+
+    Xbyak::Reg32 Compile_FetchParameter();
+    Xbyak::Reg32 Compile_GetRegister(u32 index, Xbyak::Reg32 dst);
+    Xbyak::Reg64 Compile_GetRegister(u32 index, Xbyak::Reg64 dst);
+    void Compile_WriteCarry(Xbyak::Reg64 dst);
+
+    void Compile_ProcessResult(Macro::ResultOperation operation, u32 reg);
+    void Compile_Send(Xbyak::Reg32 value);
+
+    Macro::Opcode GetOpCode() const;
+    std::bitset<32> PersistentCallerSavedRegs() const;
+
+    struct JITState {
+        Engines::Maxwell3D* maxwell3d{};
+        std::array<u32, Macro::NUM_MACRO_REGISTERS> registers{};
+        const u32* parameters{};
+        u32 carry_flag{};
+    };
+    static_assert(offsetof(JITState, maxwell3d) == 0, "Maxwell3D is not at 0x0");
+    using ProgramType = void (*)(JITState*);
+
+    struct OptimizerState {
+        bool can_skip_carry{};
+        bool has_delayed_pc{};
+        bool zero_reg_skip{};
+        bool skip_dummy_addimmediate{};
+        bool optimize_for_method_move{};
+    };
+    OptimizerState optimizer{};
+
+    std::optional<Macro::Opcode> next_opcode{};
+    ProgramType program{nullptr};
+
+    std::array<Xbyak::Label, MAX_CODE_SIZE> labels{};
+    std::array<Xbyak::Label, MAX_CODE_SIZE> delay_skip{};
+    Xbyak::Label end_of_code{};
+
+    bool is_delay_slot{};
+    u32 pc{};
+    std::optional<u32> delayed_pc;
+
+    const std::vector<u32>& code;
+    Engines::Maxwell3D& maxwell3d;
+};
+
+} // namespace Tegra

src/yuzu/configuration/config.cpp

@@ -533,6 +533,8 @@ void Config::ReadDebuggingValues() {
     Settings::values.quest_flag = ReadSetting(QStringLiteral("quest_flag"), false).toBool();
     Settings::values.disable_cpu_opt =
         ReadSetting(QStringLiteral("disable_cpu_opt"), false).toBool();
+    Settings::values.disable_macro_jit =
+        ReadSetting(QStringLiteral("disable_macro_jit"), false).toBool();
 
     qt_config->endGroup();
 }
@@ -1011,6 +1013,7 @@ void Config::SaveDebuggingValues() {
     WriteSetting(QStringLiteral("dump_nso"), Settings::values.dump_nso, false);
     WriteSetting(QStringLiteral("quest_flag"), Settings::values.quest_flag, false);
     WriteSetting(QStringLiteral("disable_cpu_opt"), Settings::values.disable_cpu_opt, false);
+    WriteSetting(QStringLiteral("disable_macro_jit"), Settings::values.disable_macro_jit, false);
 
     qt_config->endGroup();
 }

src/yuzu/configuration/configure_debug.cpp

@@ -39,6 +39,8 @@ void ConfigureDebug::SetConfiguration() {
     ui->disable_cpu_opt->setChecked(Settings::values.disable_cpu_opt);
     ui->enable_graphics_debugging->setEnabled(!Core::System::GetInstance().IsPoweredOn());
     ui->enable_graphics_debugging->setChecked(Settings::values.renderer_debug);
+    ui->disable_macro_jit->setEnabled(!Core::System::GetInstance().IsPoweredOn());
+    ui->disable_macro_jit->setChecked(Settings::values.disable_macro_jit);
 }
 
 void ConfigureDebug::ApplyConfiguration() {
@@ -51,6 +53,7 @@ void ConfigureDebug::ApplyConfiguration() {
     Settings::values.quest_flag = ui->quest_flag->isChecked();
     Settings::values.disable_cpu_opt = ui->disable_cpu_opt->isChecked();
     Settings::values.renderer_debug = ui->enable_graphics_debugging->isChecked();
+    Settings::values.disable_macro_jit = ui->disable_macro_jit->isChecked();
     Debugger::ToggleConsole();
     Log::Filter filter;
     filter.ParseFilterString(Settings::values.log_filter);

src/yuzu/configuration/configure_debug.ui

@@ -148,6 +148,19 @@
         </property>
        </widget>
       </item>
+      <item>
+       <widget class="QCheckBox" name="disable_macro_jit">
+        <property name="enabled">
+         <bool>true</bool>
+        </property>
+        <property name="whatsThis">
+         <string>When checked, it disables the macro Just In Time compiler. Enabled this makes games run slower</string>
+        </property>
+        <property name="text">
+         <string>Disable Macro JIT</string>
+        </property>
+       </widget>
+      </item>
      </layout>
     </widget>
    </item>

src/yuzu_cmd/config.cpp

@@ -432,6 +432,8 @@ void Config::ReadValues() {
     Settings::values.quest_flag = sdl2_config->GetBoolean("Debugging", "quest_flag", false);
     Settings::values.disable_cpu_opt =
         sdl2_config->GetBoolean("Debugging", "disable_cpu_opt", false);
+    Settings::values.disable_macro_jit =
+        sdl2_config->GetBoolean("Debugging", "disable_macro_jit", false);
 
     const auto title_list = sdl2_config->Get("AddOns", "title_ids", "");
     std::stringstream ss(title_list);

src/yuzu_cmd/default_ini.h

@@ -291,6 +291,8 @@ quest_flag =
 # Determines whether or not JIT CPU optimizations are enabled
 # false: Optimizations Enabled, true: Optimizations Disabled
 disable_cpu_opt =
+# Enables/Disables the macro JIT compiler
+disable_macro_jit=false
 
 [WebService]
 # Whether or not to enable telemetry