Shader Cahe: Fix Phi Nodes on GLASM.

yuzu_cmd, yuzu qt: Use SDL to disable the screen saver

src/CMakeLists.txt

@@ -32,6 +32,7 @@ if (MSVC)
     # /Zc:externConstexpr - Allow extern constexpr variables to have external linkage, like the standard mandates
     # /Zc:inline          - Let codegen omit inline functions in object files
     # /Zc:throwingNew     - Let codegen assume `operator new` (without std::nothrow) will never return null
+    # /GT                 - Supports fiber safety for data allocated using static thread-local storage
     add_compile_options(
         /MP
         /Zi
@@ -44,6 +45,7 @@ if (MSVC)
         /Zc:externConstexpr
         /Zc:inline
         /Zc:throwingNew
+        /GT
 
         # External headers diagnostics
         /experimental:external  # Enables the external headers options. This option isn't required in Visual Studio 2019 version 16.10 and later
@@ -69,6 +71,10 @@ if (MSVC)
         /we5038 # data member 'member1' will be initialized after data member 'member2'
     )
 
+    if (ARCHITECTURE_x86_64)
+        add_compile_options(/QIntel-jcc-erratum)
+    endif()
+
     # /GS- - No stack buffer overflow checks
     add_compile_options("$<$<CONFIG:Release>:/GS->")
 

src/core/file_sys/control_metadata.cpp

@@ -9,7 +9,7 @@
 
 namespace FileSys {
 
-const std::array<const char*, 15> LANGUAGE_NAMES{{
+const std::array<const char*, 16> LANGUAGE_NAMES{{
     "AmericanEnglish",
     "BritishEnglish",
     "Japanese",
@@ -25,6 +25,7 @@ const std::array<const char*, 15> LANGUAGE_NAMES{{
     "Korean",
     "Taiwanese",
     "Chinese",
+    "BrazilianPortuguese",
 }};
 
 std::string LanguageEntry::GetApplicationName() const {

src/core/file_sys/control_metadata.h

@@ -88,11 +88,12 @@ enum class Language : u8 {
     Korean = 12,
     Taiwanese = 13,
     Chinese = 14,
+    BrazilianPortuguese = 15,
 
     Default = 255,
 };
 
-extern const std::array<const char*, 15> LANGUAGE_NAMES;
+extern const std::array<const char*, 16> LANGUAGE_NAMES;
 
 // A class representing the format used by NX metadata files, typically named Control.nacp.
 // These store application name, dev name, title id, and other miscellaneous data.

src/core/hle/result.h

@@ -206,7 +206,7 @@ public:
         return result;
     }
 
-    ResultVal(const ResultVal& o) : result_code(o.result_code) {
+    ResultVal(const ResultVal& o) noexcept : result_code(o.result_code) {
         if (!o.empty()) {
             new (&object) T(o.object);
         }
@@ -224,7 +224,7 @@ public:
         }
     }
 
-    ResultVal& operator=(const ResultVal& o) {
+    ResultVal& operator=(const ResultVal& o) noexcept {
         if (this == &o) {
             return *this;
         }
@@ -244,6 +244,26 @@ public:
         return *this;
     }
 
+    ResultVal& operator=(ResultVal&& o) noexcept {
+        if (this == &o) {
+            return *this;
+        }
+        if (!empty()) {
+            if (!o.empty()) {
+                object = std::move(o.object);
+            } else {
+                object.~T();
+            }
+        } else {
+            if (!o.empty()) {
+                new (&object) T(std::move(o.object));
+            }
+        }
+        result_code = o.result_code;
+
+        return *this;
+    }
+
     /**
      * Replaces the current result with a new constructed result value in-place. The code must not
      * be an error code.
@@ -329,8 +349,8 @@ template <typename T, typename... Args>
  * copy or move constructing.
  */
 template <typename Arg>
-[[nodiscard]] ResultVal<std::remove_reference_t<Arg>> MakeResult(Arg&& arg) {
-    return ResultVal<std::remove_reference_t<Arg>>::WithCode(ResultSuccess, std::forward<Arg>(arg));
+[[nodiscard]] ResultVal<std::remove_cvref_t<Arg>> MakeResult(Arg&& arg) {
+    return ResultVal<std::remove_cvref_t<Arg>>::WithCode(ResultSuccess, std::forward<Arg>(arg));
 }
 
 /**

src/core/hle/service/ns/language.cpp

@@ -277,6 +277,25 @@ constexpr ApplicationLanguagePriorityList priority_list_simplified_chinese = {{
     ApplicationLanguage::Korean,
 }};
 
+constexpr ApplicationLanguagePriorityList priority_list_brazilian_portuguese = {{
+    ApplicationLanguage::BrazilianPortuguese,
+    ApplicationLanguage::Portuguese,
+    ApplicationLanguage::LatinAmericanSpanish,
+    ApplicationLanguage::AmericanEnglish,
+    ApplicationLanguage::BritishEnglish,
+    ApplicationLanguage::Japanese,
+    ApplicationLanguage::French,
+    ApplicationLanguage::German,
+    ApplicationLanguage::Spanish,
+    ApplicationLanguage::Italian,
+    ApplicationLanguage::Dutch,
+    ApplicationLanguage::CanadianFrench,
+    ApplicationLanguage::Russian,
+    ApplicationLanguage::Korean,
+    ApplicationLanguage::SimplifiedChinese,
+    ApplicationLanguage::TraditionalChinese,
+}};
+
 const ApplicationLanguagePriorityList* GetApplicationLanguagePriorityList(
     const ApplicationLanguage lang) {
     switch (lang) {
@@ -310,6 +329,8 @@ const ApplicationLanguagePriorityList* GetApplicationLanguagePriorityList(
         return &priority_list_traditional_chinese;
     case ApplicationLanguage::SimplifiedChinese:
         return &priority_list_simplified_chinese;
+    case ApplicationLanguage::BrazilianPortuguese:
+        return &priority_list_brazilian_portuguese;
     default:
         return nullptr;
     }
@@ -339,7 +360,6 @@ std::optional<ApplicationLanguage> ConvertToApplicationLanguage(
     case Set::LanguageCode::FR_CA:
         return ApplicationLanguage::CanadianFrench;
     case Set::LanguageCode::PT:
-    case Set::LanguageCode::PT_BR:
         return ApplicationLanguage::Portuguese;
     case Set::LanguageCode::RU:
         return ApplicationLanguage::Russian;
@@ -351,6 +371,8 @@ std::optional<ApplicationLanguage> ConvertToApplicationLanguage(
     case Set::LanguageCode::ZH_CN:
     case Set::LanguageCode::ZH_HANS:
         return ApplicationLanguage::SimplifiedChinese;
+    case Set::LanguageCode::PT_BR:
+        return ApplicationLanguage::BrazilianPortuguese;
     default:
         return std::nullopt;
     }
@@ -388,6 +410,8 @@ std::optional<Set::LanguageCode> ConvertToLanguageCode(const ApplicationLanguage
         return Set::LanguageCode::ZH_HANT;
     case ApplicationLanguage::SimplifiedChinese:
         return Set::LanguageCode::ZH_HANS;
+    case ApplicationLanguage::BrazilianPortuguese:
+        return Set::LanguageCode::PT_BR;
     default:
         return std::nullopt;
     }

src/core/hle/service/ns/language.h

@@ -30,6 +30,7 @@ enum class ApplicationLanguage : u8 {
     Korean,
     TraditionalChinese,
     SimplifiedChinese,
+    BrazilianPortuguese,
     Count
 };
 using ApplicationLanguagePriorityList =

src/shader_recompiler/backend/glasm/emit_glasm_not_implemented.cpp

@@ -18,7 +18,7 @@ namespace Shader::Backend::GLASM {
 #define NotImplemented() throw NotImplementedException("GLASM instruction {}", __LINE__)
 
 static void DefinePhi(EmitContext& ctx, IR::Inst& phi) {
-    switch (phi.Arg(0).Type()) {
+    switch (phi.Type()) {
     case IR::Type::U1:
     case IR::Type::U32:
     case IR::Type::F32:

src/shader_recompiler/backend/glsl/emit_glsl_special.cpp

@@ -68,7 +68,7 @@ void EmitPhi(EmitContext& ctx, IR::Inst& phi) {
     }
     if (!phi.Definition<Id>().is_valid) {
         // The phi node wasn't forward defined
-        ctx.var_alloc.PhiDefine(phi, phi.Arg(0).Type());
+        ctx.var_alloc.PhiDefine(phi, phi.Type());
     }
 }
 
@@ -80,7 +80,7 @@ void EmitReference(EmitContext& ctx, const IR::Value& value) {
 
 void EmitPhiMove(EmitContext& ctx, const IR::Value& phi_value, const IR::Value& value) {
     IR::Inst& phi{*phi_value.InstRecursive()};
-    const auto phi_type{phi.Arg(0).Type()};
+    const auto phi_type{phi.Type()};
     if (!phi.Definition<Id>().is_valid) {
         // The phi node wasn't forward defined
         ctx.var_alloc.PhiDefine(phi, phi_type);

src/shader_recompiler/frontend/ir/basic_block.h

@@ -152,6 +152,17 @@ public:
         return instructions.crend();
     }
 
+    // Set the order of the block, it can be set pre order, the user decides
+    void SetOrder(u32 new_order) {
+        order = new_order;
+    }
+
+    // Get the order of the block.
+    // The higher, the closer is the block to the end.
+    [[nodiscard]] u32 GetOrder() const {
+        return order;
+    }
+
 private:
     /// Memory pool for instruction list
     ObjectPool<Inst>* inst_pool;
@@ -171,6 +182,9 @@ private:
 
     /// Intrusively stored host definition of this block.
     u32 definition{};
+
+    /// Order of the block.
+    u32 order{};
 };
 
 using BlockList = std::vector<Block*>;

src/shader_recompiler/frontend/ir/microinstruction.cpp

@@ -6,6 +6,7 @@
 #include <memory>
 
 #include "shader_recompiler/exception.h"
+#include "shader_recompiler/frontend/ir/basic_block.h"
 #include "shader_recompiler/frontend/ir/type.h"
 #include "shader_recompiler/frontend/ir/value.h"
 
@@ -253,6 +254,10 @@ Inst* Inst::GetAssociatedPseudoOperation(IR::Opcode opcode) {
 }
 
 IR::Type Inst::Type() const {
+    if (op == IR::Opcode::Phi) {
+        // The type of a phi node is stored in its flags
+        return Flags<IR::Type>();
+    }
     return TypeOf(op);
 }
 
@@ -291,6 +296,16 @@ void Inst::AddPhiOperand(Block* predecessor, const Value& value) {
     phi_args.emplace_back(predecessor, value);
 }
 
+void Inst::OrderPhiArgs() {
+    if (op != Opcode::Phi) {
+        throw LogicError("{} is not a Phi instruction", op);
+    }
+    std::sort(phi_args.begin(), phi_args.end(),
+              [](const std::pair<Block*, Value>& a, const std::pair<Block*, Value>& b) {
+                  return a.first->GetOrder() < b.first->GetOrder();
+              });
+}
+
 void Inst::Invalidate() {
     ClearArgs();
     ReplaceOpcode(Opcode::Void);

src/shader_recompiler/frontend/ir/value.h

@@ -182,6 +182,9 @@ public:
     /// Add phi operand to a phi instruction.
     void AddPhiOperand(Block* predecessor, const Value& value);
 
+    /// Orders the Phi arguments from farthest away to nearest.
+    void OrderPhiArgs();
+
     void Invalidate();
     void ClearArgs();
 

src/shader_recompiler/frontend/maxwell/translate_program.cpp

@@ -27,9 +27,11 @@ IR::BlockList GenerateBlocks(const IR::AbstractSyntaxList& syntax_list) {
     }
     IR::BlockList blocks;
     blocks.reserve(num_syntax_blocks);
+    u32 order_index{};
     for (const auto& node : syntax_list) {
         if (node.type == IR::AbstractSyntaxNode::Type::Block) {
             blocks.push_back(node.data.block);
+            blocks.back()->SetOrder(order_index++);
         }
     }
     return blocks;

src/shader_recompiler/ir_opt/ssa_rewrite_pass.cpp

@@ -14,6 +14,7 @@
 //      https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6
 //
 
+#include <deque>
 #include <span>
 #include <variant>
 #include <vector>
@@ -370,6 +371,26 @@ void VisitBlock(Pass& pass, IR::Block* block) {
     }
     pass.SealBlock(block);
 }
+
+IR::Type GetConcreteType(IR::Inst* inst) {
+    std::deque<IR::Inst*> queue;
+    queue.push_back(inst);
+    while (!queue.empty()) {
+        IR::Inst* current = queue.front();
+        queue.pop_front();
+        const size_t num_args{current->NumArgs()};
+        for (size_t i = 0; i < num_args; ++i) {
+            const auto set_type = current->Arg(i).Type();
+            if (set_type != IR::Type::Opaque) {
+                return set_type;
+            }
+            if (!current->Arg(i).IsImmediate()) {
+                queue.push_back(current->Arg(i).Inst());
+            }
+        }
+    }
+    return IR::Type::Opaque;
+}
 } // Anonymous namespace
 
 void SsaRewritePass(IR::Program& program) {
@@ -378,6 +399,16 @@ void SsaRewritePass(IR::Program& program) {
     for (auto block = program.post_order_blocks.rbegin(); block != end; ++block) {
         VisitBlock(pass, *block);
     }
+    for (auto block = program.post_order_blocks.rbegin(); block != end; ++block) {
+        for (IR::Inst& inst : (*block)->Instructions()) {
+            if (inst.GetOpcode() == IR::Opcode::Phi) {
+                if (inst.Type() == IR::Type::Opaque) {
+                    inst.SetFlags(GetConcreteType(&inst));
+                }
+                inst.OrderPhiArgs();
+            }
+        }
+    }
 }
 
 } // namespace Shader::Optimization

src/video_core/renderer_opengl/gl_device.cpp

@@ -181,6 +181,21 @@ Device::Device() {
         LOG_ERROR(Render_OpenGL, "Assembly shaders enabled but not supported");
         shader_backend = Settings::ShaderBackend::GLSL;
     }
+
+    if (shader_backend == Settings::ShaderBackend::GLSL && is_nvidia &&
+        !Settings::values.renderer_debug) {
+        const std::string_view driver_version = version.substr(13);
+        const int version_major =
+            std::atoi(driver_version.substr(0, driver_version.find(".")).data());
+
+        if (version_major >= 495) {
+            LOG_WARNING(Render_OpenGL, "NVIDIA drivers 495 and later causes significant problems "
+                                       "with yuzu. Forcing GLASM as a mitigation.");
+            shader_backend = Settings::ShaderBackend::GLASM;
+            use_assembly_shaders = true;
+        }
+    }
+
     // Blocks AMD and Intel OpenGL drivers on Windows from using asynchronous shader compilation.
     use_asynchronous_shaders = Settings::values.use_asynchronous_shaders.GetValue() &&
                                !(is_amd || (is_intel && !is_linux));

src/yuzu/CMakeLists.txt

@@ -299,6 +299,11 @@ if (YUZU_USE_BUNDLED_QT)
     copy_yuzu_Qt5_deps(yuzu)
 endif()
 
+if (ENABLE_SDL2)
+    target_link_libraries(yuzu PRIVATE SDL2)
+    target_compile_definitions(yuzu PRIVATE HAVE_SDL2)
+endif()
+
 if (MSVC)
     include(CopyYuzuSDLDeps)
     include(CopyYuzuFFmpegDeps)

src/yuzu/configuration/configure_dialog.cpp

@@ -81,8 +81,11 @@ ConfigureDialog::ConfigureDialog(QWidget* parent, HotkeyRegistry& registry,
     SetConfiguration();
     PopulateSelectionList();
 
-    connect(ui->tabWidget, &QTabWidget::currentChanged, this,
-            [this]() { debug_tab_tab->SetCurrentIndex(0); });
+    connect(ui->tabWidget, &QTabWidget::currentChanged, this, [this](int index) {
+        if (index != -1) {
+            debug_tab_tab->SetCurrentIndex(0);
+        }
+    });
     connect(ui_tab.get(), &ConfigureUi::LanguageChanged, this, &ConfigureDialog::OnLanguageChanged);
     connect(ui->selectorList, &QListWidget::itemSelectionChanged, this,
             &ConfigureDialog::UpdateVisibleTabs);

src/yuzu/configuration/configure_per_game.cpp

@@ -66,7 +66,7 @@ ConfigurePerGame::ConfigurePerGame(QWidget* parent, u64 title_id, const std::str
     ui->tabWidget->addTab(system_tab.get(), tr("System"));
     ui->tabWidget->addTab(cpu_tab.get(), tr("CPU"));
     ui->tabWidget->addTab(graphics_tab.get(), tr("Graphics"));
-    ui->tabWidget->addTab(graphics_advanced_tab.get(), tr("GraphicsAdvanced"));
+    ui->tabWidget->addTab(graphics_advanced_tab.get(), tr("Adv. Graphics"));
     ui->tabWidget->addTab(audio_tab.get(), tr("Audio"));
 
     setFocusPolicy(Qt::ClickFocus);

src/yuzu/configuration/configure_per_game.ui

@@ -2,14 +2,6 @@
 <ui version="4.0">
  <class>ConfigurePerGame</class>
  <widget class="QDialog" name="ConfigurePerGame">
-  <property name="geometry">
-   <rect>
-    <x>0</x>
-    <y>0</y>
-    <width>900</width>
-    <height>630</height>
-   </rect>
-  </property>
   <property name="minimumSize">
    <size>
     <width>900</width>

src/yuzu/main.cpp

@@ -66,6 +66,10 @@ static FileSys::VirtualFile VfsDirectoryCreateFileWrapper(const FileSys::Virtual
 #include <QUrl>
 #include <QtConcurrent/QtConcurrent>
 
+#ifdef HAVE_SDL2
+#include <SDL.h> // For SDL ScreenSaver functions
+#endif
+
 #include <fmt/format.h>
 #include "common/detached_tasks.h"
 #include "common/fs/fs.h"
@@ -287,6 +291,14 @@ GMainWindow::GMainWindow()
 
     ui->action_Fullscreen->setChecked(false);
 
+#if defined(HAVE_SDL2) && !defined(_WIN32)
+    SDL_InitSubSystem(SDL_INIT_VIDEO);
+    // SDL disables the screen saver by default, and setting the hint
+    // SDL_HINT_VIDEO_ALLOW_SCREENSAVER doesn't seem to work, so we just enable the screen saver
+    // for now.
+    SDL_EnableScreenSaver();
+#endif
+
     QStringList args = QApplication::arguments();
 
     if (args.size() < 2) {
@@ -357,8 +369,9 @@ GMainWindow::GMainWindow()
 
 GMainWindow::~GMainWindow() {
     // will get automatically deleted otherwise
-    if (render_window->parent() == nullptr)
+    if (render_window->parent() == nullptr) {
         delete render_window;
+    }
 }
 
 void GMainWindow::RegisterMetaTypes() {
@@ -1223,12 +1236,16 @@ void GMainWindow::OnDisplayTitleBars(bool show) {
 void GMainWindow::PreventOSSleep() {
 #ifdef _WIN32
     SetThreadExecutionState(ES_CONTINUOUS | ES_SYSTEM_REQUIRED | ES_DISPLAY_REQUIRED);
+#elif defined(HAVE_SDL2)
+    SDL_DisableScreenSaver();
 #endif
 }
 
 void GMainWindow::AllowOSSleep() {
 #ifdef _WIN32
     SetThreadExecutionState(ES_CONTINUOUS);
+#elif defined(HAVE_SDL2)
+    SDL_EnableScreenSaver();
 #endif
 }