106 files changed, 2707 insertions, 1051 deletions

diff --git a/src/audio_core/CMakeLists.txt b/src/audio_core/CMakeLists.txt
index c4bad8cb0..869da5e83 100644
--- a/src/audio_core/CMakeLists.txt
+++ b/src/audio_core/CMakeLists.txt
@@ -2,13 +2,16 @@ set(SRCS
             audio_core.cpp
             codec.cpp
             hle/dsp.cpp
+            hle/filter.cpp
             hle/pipe.cpp
             )
 
 set(HEADERS
             audio_core.h
             codec.h
+            hle/common.h
             hle/dsp.h
+            hle/filter.h
             hle/pipe.h
             sink.h
             )
diff --git a/src/audio_core/hle/common.h b/src/audio_core/hle/common.h
new file mode 100644
index 000000000..37d441eb2
--- /dev/null
+++ b/src/audio_core/hle/common.h
@@ -0,0 +1,35 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+
+#include "audio_core/audio_core.h"
+
+#include "common/common_types.h"
+
+namespace DSP {
+namespace HLE {
+
+/// The final output to the speakers is stereo. Preprocessing output in Source is also stereo.
+using StereoFrame16 = std::array<std::array<s16, 2>, AudioCore::samples_per_frame>;
+
+/// The DSP is quadraphonic internally.
+using QuadFrame32   = std::array<std::array<s32, 4>, AudioCore::samples_per_frame>;
+
+/**
+ * This performs the filter operation defined by FilterT::ProcessSample on the frame in-place.
+ * FilterT::ProcessSample is called sequentially on the samples.
+ */
+template<typename FrameT, typename FilterT>
+void FilterFrame(FrameT& frame, FilterT& filter) {
+    std::transform(frame.begin(), frame.end(), frame.begin(), [&filter](const typename FrameT::value_type& sample) {
+        return filter.ProcessSample(sample);
+    });
+}
+
+} // namespace HLE
+} // namespace DSP
diff --git a/src/audio_core/hle/dsp.h b/src/audio_core/hle/dsp.h
index 376436c29..c15ef0b7a 100644
--- a/src/audio_core/hle/dsp.h
+++ b/src/audio_core/hle/dsp.h
@@ -126,8 +126,11 @@ struct SourceConfiguration {
         union {
             u32_le dirty_raw;
 
+            BitField<0, 1, u32_le> format_dirty;
+            BitField<1, 1, u32_le> mono_or_stereo_dirty;
             BitField<2, 1, u32_le> adpcm_coefficients_dirty;
             BitField<3, 1, u32_le> partial_embedded_buffer_dirty; ///< Tends to be set when a looped buffer is queued.
+            BitField<4, 1, u32_le> partial_reset_flag;
 
             BitField<16, 1, u32_le> enable_dirty;
             BitField<17, 1, u32_le> interpolation_dirty;
@@ -143,8 +146,7 @@ struct SourceConfiguration {
             BitField<27, 1, u32_le> gain_2_dirty;
             BitField<28, 1, u32_le> sync_dirty;
             BitField<29, 1, u32_le> reset_flag;
-
-            BitField<31, 1, u32_le> embedded_buffer_dirty;
+            BitField<30, 1, u32_le> embedded_buffer_dirty;
         };
 
         // Gain control
@@ -175,7 +177,8 @@ struct SourceConfiguration {
         /**
          * This is the simplest normalized first-order digital recursive filter.
          * The transfer function of this filter is:
-         *     H(z) = b0 / (1 + a1 z^-1)
+         *     H(z) = b0 / (1 - a1 z^-1)
+         * Note the feedbackward coefficient is negated.
          * Values are signed fixed point with 15 fractional bits.
          */
         struct SimpleFilter {
@@ -192,11 +195,11 @@ struct SourceConfiguration {
          * Values are signed fixed point with 14 fractional bits.
          */
         struct BiquadFilter {
-            s16_le b0;
-            s16_le b1;
-            s16_le b2;
-            s16_le a1;
             s16_le a2;
+            s16_le a1;
+            s16_le b2;
+            s16_le b1;
+            s16_le b0;
         };
 
         union {
diff --git a/src/audio_core/hle/filter.cpp b/src/audio_core/hle/filter.cpp
new file mode 100644
index 000000000..2c65ef026
--- /dev/null
+++ b/src/audio_core/hle/filter.cpp
@@ -0,0 +1,115 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <array>
+#include <cstddef>
+
+#include "audio_core/hle/common.h"
+#include "audio_core/hle/dsp.h"
+#include "audio_core/hle/filter.h"
+
+#include "common/common_types.h"
+#include "common/math_util.h"
+
+namespace DSP {
+namespace HLE {
+
+void SourceFilters::Reset() {
+    Enable(false, false);
+}
+
+void SourceFilters::Enable(bool simple, bool biquad) {
+    simple_filter_enabled = simple;
+    biquad_filter_enabled = biquad;
+
+    if (!simple)
+        simple_filter.Reset();
+    if (!biquad)
+        biquad_filter.Reset();
+}
+
+void SourceFilters::Configure(SourceConfiguration::Configuration::SimpleFilter config) {
+    simple_filter.Configure(config);
+}
+
+void SourceFilters::Configure(SourceConfiguration::Configuration::BiquadFilter config) {
+    biquad_filter.Configure(config);
+}
+
+void SourceFilters::ProcessFrame(StereoFrame16& frame) {
+    if (!simple_filter_enabled && !biquad_filter_enabled)
+        return;
+
+    if (simple_filter_enabled) {
+        FilterFrame(frame, simple_filter);
+    }
+
+    if (biquad_filter_enabled) {
+        FilterFrame(frame, biquad_filter);
+    }
+}
+
+// SimpleFilter
+
+void SourceFilters::SimpleFilter::Reset() {
+    y1.fill(0);
+    // Configure as passthrough.
+    a1 = 0;
+    b0 = 1 << 15;
+}
+
+void SourceFilters::SimpleFilter::Configure(SourceConfiguration::Configuration::SimpleFilter config) {
+    a1 = config.a1;
+    b0 = config.b0;
+}
+
+std::array<s16, 2> SourceFilters::SimpleFilter::ProcessSample(const std::array<s16, 2>& x0) {
+    std::array<s16, 2> y0;
+    for (size_t i = 0; i < 2; i++) {
+        const s32 tmp = (b0 * x0[i] + a1 * y1[i]) >> 15;
+        y0[i] = MathUtil::Clamp(tmp, -32768, 32767);
+    }
+
+    y1 = y0;
+
+    return y0;
+}
+
+// BiquadFilter
+
+void SourceFilters::BiquadFilter::Reset() {
+    x1.fill(0);
+    x2.fill(0);
+    y1.fill(0);
+    y2.fill(0);
+    // Configure as passthrough.
+    a1 = a2 = b1 = b2 = 0;
+    b0 = 1 << 14;
+}
+
+void SourceFilters::BiquadFilter::Configure(SourceConfiguration::Configuration::BiquadFilter config) {
+    a1 = config.a1;
+    a2 = config.a2;
+    b0 = config.b0;
+    b1 = config.b1;
+    b2 = config.b2;
+}
+
+std::array<s16, 2> SourceFilters::BiquadFilter::ProcessSample(const std::array<s16, 2>& x0) {
+    std::array<s16, 2> y0;
+    for (size_t i = 0; i < 2; i++) {
+        const s32 tmp = (b0 * x0[i] + b1 * x1[i] + b2 * x2[i] + a1 * y1[i] + a2 * y2[i]) >> 14;
+        y0[i] = MathUtil::Clamp(tmp, -32768, 32767);
+    }
+
+    x2 = x1;
+    x1 = x0;
+    y2 = y1;
+    y1 = y0;
+
+    return y0;
+}
+
+} // namespace HLE
+} // namespace DSP
diff --git a/src/audio_core/hle/filter.h b/src/audio_core/hle/filter.h
new file mode 100644
index 000000000..75738f600
--- /dev/null
+++ b/src/audio_core/hle/filter.h
@@ -0,0 +1,112 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <array>
+
+#include "audio_core/hle/common.h"
+#include "audio_core/hle/dsp.h"
+
+#include "common/common_types.h"
+
+namespace DSP {
+namespace HLE {
+
+/// Preprocessing filters. There is an independent set of filters for each Source.
+class SourceFilters final {
+    SourceFilters() { Reset(); }
+
+    /// Reset internal state.
+    void Reset();
+
+    /**
+     * Enable/Disable filters
+     * See also: SourceConfiguration::Configuration::simple_filter_enabled,
+     *           SourceConfiguration::Configuration::biquad_filter_enabled.
+     * @param simple If true, enables the simple filter. If false, disables it.
+     * @param simple If true, enables the biquad filter. If false, disables it.
+     */
+    void Enable(bool simple, bool biquad);
+
+    /**
+     * Configure simple filter.
+     * @param config Configuration from DSP shared memory.
+     */
+    void Configure(SourceConfiguration::Configuration::SimpleFilter config);
+
+    /**
+     * Configure biquad filter.
+     * @param config Configuration from DSP shared memory.
+     */
+    void Configure(SourceConfiguration::Configuration::BiquadFilter config);
+
+    /**
+     * Processes a frame in-place.
+     * @param frame Audio samples to process. Modified in-place.
+     */
+    void ProcessFrame(StereoFrame16& frame);
+
+private:
+    bool simple_filter_enabled;
+    bool biquad_filter_enabled;
+
+    struct SimpleFilter {
+        SimpleFilter() { Reset(); }
+
+        /// Resets internal state.
+        void Reset();
+
+        /**
+         * Configures this filter with application settings.
+         * @param config Configuration from DSP shared memory.
+         */
+        void Configure(SourceConfiguration::Configuration::SimpleFilter config);
+
+        /**
+         * Processes a single stereo PCM16 sample.
+         * @param x0 Input sample
+         * @return Output sample
+         */
+        std::array<s16, 2> ProcessSample(const std::array<s16, 2>& x0);
+
+    private:
+        // Configuration
+        s32 a1, b0;
+        // Internal state
+        std::array<s16, 2> y1;
+    } simple_filter;
+
+    struct BiquadFilter {
+        BiquadFilter() { Reset(); }
+
+        /// Resets internal state.
+        void Reset();
+
+        /**
+         * Configures this filter with application settings.
+         * @param config Configuration from DSP shared memory.
+         */
+        void Configure(SourceConfiguration::Configuration::BiquadFilter config);
+
+        /**
+         * Processes a single stereo PCM16 sample.
+         * @param x0 Input sample
+         * @return Output sample
+         */
+        std::array<s16, 2> ProcessSample(const std::array<s16, 2>& x0);
+
+    private:
+        // Configuration
+        s32 a1, a2, b0, b1, b2;
+        // Internal state
+        std::array<s16, 2> x1;
+        std::array<s16, 2> x2;
+        std::array<s16, 2> y1;
+        std::array<s16, 2> y2;
+    } biquad_filter;
+};
+
+} // namespace HLE
+} // namespace DSP
diff --git a/src/citra/citra.cpp b/src/citra/citra.cpp
index 40e40f192..d6ad13f69 100644
--- a/src/citra/citra.cpp
+++ b/src/citra/citra.cpp
@@ -5,6 +5,7 @@
 #include <string>
 #include <thread>
 #include <iostream>
+#include <memory>
 
 // This needs to be included before getopt.h because the latter #defines symbols used by it
 #include "common/microprofile.h"
@@ -19,7 +20,6 @@
 #include "common/logging/log.h"
 #include "common/logging/backend.h"
 #include "common/logging/filter.h"
-#include "common/make_unique.h"
 #include "common/scope_exit.h"
 
 #include "core/settings.h"
@@ -36,25 +36,43 @@
 
 static void PrintHelp()
 {
-    std::cout << "Usage: citra <filename>" << std::endl;
+    std::cout << "Usage: citra [options] <filename>" << std::endl;
+    std::cout << "--help, -h            Display this information" << std::endl;
+    std::cout << "--gdbport, -g number  Enable gdb stub on port number" << std::endl;
 }
 
 /// Application entry point
 int main(int argc, char **argv) {
+    Config config;
     int option_index = 0;
+    bool use_gdbstub = Settings::values.use_gdbstub;
+    u32 gdb_port = static_cast<u32>(Settings::values.gdbstub_port);
+    char *endarg;
     std::string boot_filename;
+
     static struct option long_options[] = {
         { "help", no_argument, 0, 'h' },
+        { "gdbport", required_argument, 0, 'g' },
         { 0, 0, 0, 0 }
     };
 
     while (optind < argc) {
-        char arg = getopt_long(argc, argv, ":h", long_options, &option_index);
+        char arg = getopt_long(argc, argv, ":hg:", long_options, &option_index);
         if (arg != -1) {
             switch (arg) {
             case 'h':
                 PrintHelp();
                 return 0;
+            case 'g':
+                errno = 0;
+                gdb_port = strtoul(optarg, &endarg, 0);
+                use_gdbstub = true;
+                if (endarg == optarg) errno = EINVAL;
+                if (errno != 0) {
+                    perror("--gdbport");
+                    exit(1);
+                }
+                break;
             }
         } else {
             boot_filename = argv[optind];
@@ -73,16 +91,14 @@ int main(int argc, char **argv) {
         return -1;
     }
 
-    Config config;
     log_filter.ParseFilterString(Settings::values.log_filter);
 
-    GDBStub::ToggleServer(Settings::values.use_gdbstub);
-    GDBStub::SetServerPort(static_cast<u32>(Settings::values.gdbstub_port));
-
-    std::unique_ptr<EmuWindow_SDL2> emu_window = Common::make_unique<EmuWindow_SDL2>();
+    // Apply the command line arguments
+    Settings::values.gdbstub_port = gdb_port;
+    Settings::values.use_gdbstub = use_gdbstub;
+    Settings::Apply();
 
-    VideoCore::g_hw_renderer_enabled = Settings::values.use_hw_renderer;
-    VideoCore::g_shader_jit_enabled = Settings::values.use_shader_jit;
+    std::unique_ptr<EmuWindow_SDL2> emu_window = std::make_unique<EmuWindow_SDL2>();
 
     System::Init(emu_window.get());
     SCOPE_EXIT({ System::Shutdown(); });
diff --git a/src/citra/config.cpp b/src/citra/config.cpp
index 9034b188e..6b6617352 100644
--- a/src/citra/config.cpp
+++ b/src/citra/config.cpp
@@ -2,6 +2,8 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <memory>
+
 #include <inih/cpp/INIReader.h>
 
 #include <SDL.h>
@@ -10,7 +12,6 @@
 
 #include "common/file_util.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 
 #include "core/settings.h"
 
@@ -19,7 +20,7 @@
 Config::Config() {
     // TODO: Don't hardcode the path; let the frontend decide where to put the config files.
     sdl2_config_loc = FileUtil::GetUserPath(D_CONFIG_IDX) + "sdl2-config.ini";
-    sdl2_config = Common::make_unique<INIReader>(sdl2_config_loc);
+    sdl2_config = std::make_unique<INIReader>(sdl2_config_loc);
 
     Reload();
 }
@@ -31,7 +32,7 @@ bool Config::LoadINI(const std::string& default_contents, bool retry) {
             LOG_WARNING(Config, "Failed to load %s. Creating file from defaults...", location);
             FileUtil::CreateFullPath(location);
             FileUtil::WriteStringToFile(true, default_contents, location);
-            sdl2_config = Common::make_unique<INIReader>(location); // Reopen file
+            sdl2_config = std::make_unique<INIReader>(location); // Reopen file
 
             return LoadINI(default_contents, false);
         }
diff --git a/src/citra_qt/CMakeLists.txt b/src/citra_qt/CMakeLists.txt
index 9b3eb2cd6..6660d9879 100644
--- a/src/citra_qt/CMakeLists.txt
+++ b/src/citra_qt/CMakeLists.txt
@@ -17,12 +17,16 @@ set(SRCS
             debugger/profiler.cpp
             debugger/ramview.cpp
             debugger/registers.cpp
-            game_list.cpp
             util/spinbox.cpp
             util/util.cpp
             bootmanager.cpp
+            configure_debug.cpp
+            configure_dialog.cpp
+            configure_general.cpp
+            game_list.cpp
             hotkeys.cpp
             main.cpp
+            ui_settings.cpp
             citra-qt.rc
             Info.plist
             )
@@ -44,12 +48,16 @@ set(HEADERS
             debugger/profiler.h
             debugger/ramview.h
             debugger/registers.h
-            game_list.h
             util/spinbox.h
             util/util.h
             bootmanager.h
+            configure_debug.h
+            configure_dialog.h
+            configure_general.h
+            game_list.h
             hotkeys.h
             main.h
+            ui_settings.h
             version.h
             )
 
@@ -59,6 +67,9 @@ set(UIS
             debugger/disassembler.ui
             debugger/profiler.ui
             debugger/registers.ui
+            configure.ui
+            configure_debug.ui
+            configure_general.ui
             hotkeys.ui
             main.ui
             )
diff --git a/src/citra_qt/config.cpp b/src/citra_qt/config.cpp
index d1a19ade9..e363be38a 100644
--- a/src/citra_qt/config.cpp
+++ b/src/citra_qt/config.cpp
@@ -7,16 +7,16 @@
 #include <QStringList>
 
 #include "citra_qt/config.h"
+#include "citra_qt/ui_settings.h"
 
 #include "common/file_util.h"
 #include "core/settings.h"
 
 Config::Config() {
-
     // TODO: Don't hardcode the path; let the frontend decide where to put the config files.
     qt_config_loc = FileUtil::GetUserPath(D_CONFIG_IDX) + "qt-config.ini";
     FileUtil::CreateFullPath(qt_config_loc);
-    qt_config = new QSettings(QString::fromLocal8Bit(qt_config_loc.c_str()), QSettings::IniFormat);
+    qt_config = new QSettings(QString::fromStdString(qt_config_loc), QSettings::IniFormat);
 
     Reload();
 }
@@ -67,6 +67,51 @@ void Config::ReadValues() {
     Settings::values.use_gdbstub = qt_config->value("use_gdbstub", false).toBool();
     Settings::values.gdbstub_port = qt_config->value("gdbstub_port", 24689).toInt();
     qt_config->endGroup();
+
+    qt_config->beginGroup("UI");
+
+    qt_config->beginGroup("UILayout");
+    UISettings::values.geometry = qt_config->value("geometry").toByteArray();
+    UISettings::values.state = qt_config->value("state").toByteArray();
+    UISettings::values.renderwindow_geometry = qt_config->value("geometryRenderWindow").toByteArray();
+    UISettings::values.gamelist_header_state = qt_config->value("gameListHeaderState").toByteArray();
+    UISettings::values.microprofile_geometry = qt_config->value("microProfileDialogGeometry").toByteArray();
+    UISettings::values.microprofile_visible = qt_config->value("microProfileDialogVisible", false).toBool();
+    qt_config->endGroup();
+
+    qt_config->beginGroup("Paths");
+    UISettings::values.roms_path = qt_config->value("romsPath").toString();
+    UISettings::values.symbols_path = qt_config->value("symbolsPath").toString();
+    UISettings::values.gamedir = qt_config->value("gameListRootDir", ".").toString();
+    UISettings::values.gamedir_deepscan = qt_config->value("gameListDeepScan", false).toBool();
+    UISettings::values.recent_files = qt_config->value("recentFiles").toStringList();
+    qt_config->endGroup();
+
+    qt_config->beginGroup("Shortcuts");
+    QStringList groups = qt_config->childGroups();
+    for (auto group : groups) {
+        qt_config->beginGroup(group);
+
+        QStringList hotkeys = qt_config->childGroups();
+        for (auto hotkey : hotkeys) {
+            qt_config->beginGroup(hotkey);
+            UISettings::values.shortcuts.emplace_back(
+                        UISettings::Shortcut(group + "/" + hotkey,
+                                             UISettings::ContextualShortcut(qt_config->value("KeySeq").toString(),
+                                                                           qt_config->value("Context").toInt())));
+            qt_config->endGroup();
+        }
+
+        qt_config->endGroup();
+    }
+    qt_config->endGroup();
+
+    UISettings::values.single_window_mode = qt_config->value("singleWindowMode", true).toBool();
+    UISettings::values.display_titlebar = qt_config->value("displayTitleBars", true).toBool();
+    UISettings::values.confirm_before_closing = qt_config->value("confirmClose",true).toBool();
+    UISettings::values.first_start = qt_config->value("firstStart", true).toBool();
+
+    qt_config->endGroup();
 }
 
 void Config::SaveValues() {
@@ -107,10 +152,44 @@ void Config::SaveValues() {
     qt_config->setValue("use_gdbstub", Settings::values.use_gdbstub);
     qt_config->setValue("gdbstub_port", Settings::values.gdbstub_port);
     qt_config->endGroup();
+
+    qt_config->beginGroup("UI");
+
+    qt_config->beginGroup("UILayout");
+    qt_config->setValue("geometry", UISettings::values.geometry);
+    qt_config->setValue("state", UISettings::values.state);
+    qt_config->setValue("geometryRenderWindow", UISettings::values.renderwindow_geometry);
+    qt_config->setValue("gameListHeaderState", UISettings::values.gamelist_header_state);
+    qt_config->setValue("microProfileDialogGeometry", UISettings::values.microprofile_geometry);
+    qt_config->setValue("microProfileDialogVisible", UISettings::values.microprofile_visible);
+    qt_config->endGroup();
+
+    qt_config->beginGroup("Paths");
+    qt_config->setValue("romsPath", UISettings::values.roms_path);
+    qt_config->setValue("symbolsPath", UISettings::values.symbols_path);
+    qt_config->setValue("gameListRootDir", UISettings::values.gamedir);
+    qt_config->setValue("gameListDeepScan", UISettings::values.gamedir_deepscan);
+    qt_config->setValue("recentFiles", UISettings::values.recent_files);
+    qt_config->endGroup();
+
+    qt_config->beginGroup("Shortcuts");
+    for (auto shortcut : UISettings::values.shortcuts ) {
+        qt_config->setValue(shortcut.first + "/KeySeq", shortcut.second.first);
+        qt_config->setValue(shortcut.first + "/Context", shortcut.second.second);
+    }
+    qt_config->endGroup();
+
+    qt_config->setValue("singleWindowMode", UISettings::values.single_window_mode);
+    qt_config->setValue("displayTitleBars", UISettings::values.display_titlebar);
+    qt_config->setValue("confirmClose", UISettings::values.confirm_before_closing);
+    qt_config->setValue("firstStart", UISettings::values.first_start);
+
+    qt_config->endGroup();
 }
 
 void Config::Reload() {
     ReadValues();
+    Settings::Apply();
 }
 
 void Config::Save() {
diff --git a/src/citra_qt/configure.ui b/src/citra_qt/configure.ui
new file mode 100644
index 000000000..6ae056ff9
--- /dev/null
+++ b/src/citra_qt/configure.ui
@@ -0,0 +1,97 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<ui version="4.0">
+ <class>ConfigureDialog</class>
+ <widget class="QDialog" name="ConfigureDialog">
+  <property name="geometry">
+   <rect>
+    <x>0</x>
+    <y>0</y>
+    <width>441</width>
+    <height>501</height>
+   </rect>
+  </property>
+  <property name="windowTitle">
+   <string>Citra Configuration</string>
+  </property>
+  <layout class="QVBoxLayout" name="verticalLayout">
+   <item>
+    <widget class="QTabWidget" name="tabWidget">
+     <property name="currentIndex">
+      <number>0</number>
+     </property>
+     <widget class="ConfigureGeneral" name="generalTab">
+      <attribute name="title">
+       <string>General</string>
+      </attribute>
+     </widget>
+     <widget class="QWidget" name="inputTab">
+      <attribute name="title">
+       <string>Input</string>
+      </attribute>
+     </widget>
+     <widget class="ConfigureDebug" name="debugTab">
+      <attribute name="title">
+       <string>Debug</string>
+      </attribute>
+     </widget>
+    </widget>
+   </item>
+   <item>
+    <widget class="QDialogButtonBox" name="buttonBox">
+     <property name="standardButtons">
+      <set>QDialogButtonBox::Cancel|QDialogButtonBox::Ok</set>
+     </property>
+    </widget>
+   </item>
+  </layout>
+ </widget>
+ <customwidgets>
+  <customwidget>
+   <class>ConfigureGeneral</class>
+   <extends>QWidget</extends>
+   <header>configure_general.h</header>
+   <container>1</container>
+  </customwidget>
+  <customwidget>
+   <class>ConfigureDebug</class>
+   <extends>QWidget</extends>
+   <header>configure_debug.h</header>
+   <container>1</container>
+  </customwidget>
+ </customwidgets>
+ <resources/>
+ <connections>
+  <connection>
+   <sender>buttonBox</sender>
+   <signal>accepted()</signal>
+   <receiver>ConfigureDialog</receiver>
+   <slot>accept()</slot>
+   <hints>
+    <hint type="sourcelabel">
+     <x>220</x>
+     <y>380</y>
+    </hint>
+    <hint type="destinationlabel">
+     <x>220</x>
+     <y>200</y>
+    </hint>
+   </hints>
+  </connection>
+  <connection>
+   <sender>buttonBox</sender>
+   <signal>rejected()</signal>
+   <receiver>ConfigureDialog</receiver>
+   <slot>reject()</slot>
+   <hints>
+    <hint type="sourcelabel">
+     <x>220</x>
+     <y>380</y>
+    </hint>
+    <hint type="destinationlabel">
+     <x>220</x>
+     <y>200</y>
+    </hint>
+   </hints>
+  </connection>
+ </connections>
+</ui>
diff --git a/src/citra_qt/configure_debug.cpp b/src/citra_qt/configure_debug.cpp
new file mode 100644
index 000000000..dc3d7b906
--- /dev/null
+++ b/src/citra_qt/configure_debug.cpp
@@ -0,0 +1,31 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "citra_qt/configure_debug.h"
+#include "ui_configure_debug.h"
+
+#include "core/settings.h"
+
+ConfigureDebug::ConfigureDebug(QWidget *parent) :
+    QWidget(parent),
+    ui(new Ui::ConfigureDebug)
+{
+    ui->setupUi(this);
+    this->setConfiguration();
+}
+
+ConfigureDebug::~ConfigureDebug() {
+}
+
+void ConfigureDebug::setConfiguration() {
+    ui->toogle_gdbstub->setChecked(Settings::values.use_gdbstub);
+    ui->gdbport_spinbox->setEnabled(Settings::values.use_gdbstub);
+    ui->gdbport_spinbox->setValue(Settings::values.gdbstub_port);
+}
+
+void ConfigureDebug::applyConfiguration() {
+    Settings::values.use_gdbstub = ui->toogle_gdbstub->isChecked();
+    Settings::values.gdbstub_port = ui->gdbport_spinbox->value();
+    Settings::Apply();
+}
diff --git a/src/citra_qt/configure_debug.h b/src/citra_qt/configure_debug.h
new file mode 100644
index 000000000..ab58ebbdc
--- /dev/null
+++ b/src/citra_qt/configure_debug.h
@@ -0,0 +1,29 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <QWidget>
+
+namespace Ui {
+class ConfigureDebug;
+}
+
+class ConfigureDebug : public QWidget
+{
+    Q_OBJECT
+
+public:
+    explicit ConfigureDebug(QWidget *parent = nullptr);
+    ~ConfigureDebug();
+
+    void applyConfiguration();
+
+private:
+    void setConfiguration();
+
+private:
+    std::unique_ptr<Ui::ConfigureDebug> ui;
+};
diff --git a/src/citra_qt/configure_debug.ui b/src/citra_qt/configure_debug.ui
new file mode 100644
index 000000000..3ba7f44da
--- /dev/null
+++ b/src/citra_qt/configure_debug.ui
@@ -0,0 +1,102 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<ui version="4.0">
+ <class>ConfigureDebug</class>
+ <widget class="QWidget" name="ConfigureDebug">
+  <property name="geometry">
+   <rect>
+    <x>0</x>
+    <y>0</y>
+    <width>400</width>
+    <height>300</height>
+   </rect>
+  </property>
+  <property name="windowTitle">
+   <string>Form</string>
+  </property>
+  <layout class="QVBoxLayout" name="verticalLayout">
+   <item>
+    <layout class="QVBoxLayout" name="verticalLayout_3">
+     <item>
+      <widget class="QGroupBox" name="groupBox">
+       <property name="title">
+        <string>GDB</string>
+       </property>
+       <layout class="QVBoxLayout" name="verticalLayout_2">
+        <item>
+         <layout class="QHBoxLayout" name="horizontalLayout_3">
+          <item>
+           <widget class="QCheckBox" name="toogle_gdbstub">
+            <property name="text">
+             <string>Enable GDB Stub</string>
+            </property>
+           </widget>
+          </item>
+          <item>
+           <spacer name="horizontalSpacer">
+            <property name="orientation">
+             <enum>Qt::Horizontal</enum>
+            </property>
+            <property name="sizeHint" stdset="0">
+             <size>
+              <width>40</width>
+              <height>20</height>
+             </size>
+            </property>
+           </spacer>
+          </item>
+          <item>
+           <widget class="QLabel" name="label">
+            <property name="text">
+             <string>Port:</string>
+            </property>
+           </widget>
+          </item>
+          <item>
+           <widget class="QSpinBox" name="gdbport_spinbox">
+            <property name="maximum">
+             <number>65536</number>
+            </property>
+           </widget>
+          </item>
+         </layout>
+        </item>
+       </layout>
+      </widget>
+     </item>
+    </layout>
+   </item>
+   <item>
+    <spacer name="verticalSpacer">
+     <property name="orientation">
+      <enum>Qt::Vertical</enum>
+     </property>
+     <property name="sizeHint" stdset="0">
+      <size>
+       <width>20</width>
+       <height>40</height>
+      </size>
+     </property>
+    </spacer>
+   </item>
+  </layout>
+ </widget>
+ <resources/>
+ <connections>
+  <connection>
+   <sender>toogle_gdbstub</sender>
+   <signal>toggled(bool)</signal>
+   <receiver>gdbport_spinbox</receiver>
+   <slot>setEnabled(bool)</slot>
+   <hints>
+    <hint type="sourcelabel">
+     <x>84</x>
+     <y>157</y>
+    </hint>
+    <hint type="destinationlabel">
+     <x>342</x>
+     <y>158</y>
+    </hint>
+   </hints>
+  </connection>
+ </connections>
+</ui>
diff --git a/src/citra_qt/configure_dialog.cpp b/src/citra_qt/configure_dialog.cpp
new file mode 100644
index 000000000..87c26c715
--- /dev/null
+++ b/src/citra_qt/configure_dialog.cpp
@@ -0,0 +1,29 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "citra_qt/config.h"
+#include "citra_qt/configure_dialog.h"
+#include "ui_configure.h"
+
+
+#include "core/settings.h"
+
+ConfigureDialog::ConfigureDialog(QWidget *parent) :
+    QDialog(parent),
+    ui(new Ui::ConfigureDialog)
+{
+    ui->setupUi(this);
+    this->setConfiguration();
+}
+
+ConfigureDialog::~ConfigureDialog() {
+}
+
+void ConfigureDialog::setConfiguration() {
+}
+
+void ConfigureDialog::applyConfiguration() {
+    ui->generalTab->applyConfiguration();
+    ui->debugTab->applyConfiguration();
+}
diff --git a/src/citra_qt/configure_dialog.h b/src/citra_qt/configure_dialog.h
new file mode 100644
index 000000000..89020eeb4
--- /dev/null
+++ b/src/citra_qt/configure_dialog.h
@@ -0,0 +1,29 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <QDialog>
+
+namespace Ui {
+class ConfigureDialog;
+}
+
+class ConfigureDialog : public QDialog
+{
+    Q_OBJECT
+
+public:
+    explicit ConfigureDialog(QWidget *parent = nullptr);
+    ~ConfigureDialog();
+
+    void applyConfiguration();
+
+private:
+    void setConfiguration();
+
+private:
+    std::unique_ptr<Ui::ConfigureDialog> ui;
+};
diff --git a/src/citra_qt/configure_general.cpp b/src/citra_qt/configure_general.cpp
new file mode 100644
index 000000000..a27d0d26c
--- /dev/null
+++ b/src/citra_qt/configure_general.cpp
@@ -0,0 +1,37 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "citra_qt/configure_general.h"
+#include "citra_qt/ui_settings.h"
+#include "ui_configure_general.h"
+
+#include "core/settings.h"
+
+ConfigureGeneral::ConfigureGeneral(QWidget *parent) :
+    QWidget(parent),
+    ui(new Ui::ConfigureGeneral)
+{
+    ui->setupUi(this);
+    this->setConfiguration();
+}
+
+ConfigureGeneral::~ConfigureGeneral() {
+}
+
+void ConfigureGeneral::setConfiguration() {
+    ui->toogle_deepscan->setChecked(UISettings::values.gamedir_deepscan);
+    ui->toogle_check_exit->setChecked(UISettings::values.confirm_before_closing);
+    ui->region_combobox->setCurrentIndex(Settings::values.region_value);
+    ui->toogle_hw_renderer->setChecked(Settings::values.use_hw_renderer);
+    ui->toogle_shader_jit->setChecked(Settings::values.use_shader_jit);
+}
+
+void ConfigureGeneral::applyConfiguration() {
+    UISettings::values.gamedir_deepscan = ui->toogle_deepscan->isChecked();
+    UISettings::values.confirm_before_closing = ui->toogle_check_exit->isChecked();
+    Settings::values.region_value = ui->region_combobox->currentIndex();
+    Settings::values.use_hw_renderer = ui->toogle_hw_renderer->isChecked();
+    Settings::values.use_shader_jit = ui->toogle_shader_jit->isChecked();
+    Settings::Apply();
+}
diff --git a/src/citra_qt/configure_general.h b/src/citra_qt/configure_general.h
new file mode 100644
index 000000000..a6c68e62d
--- /dev/null
+++ b/src/citra_qt/configure_general.h
@@ -0,0 +1,29 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <QWidget>
+
+namespace Ui {
+class ConfigureGeneral;
+}
+
+class ConfigureGeneral : public QWidget
+{
+    Q_OBJECT
+
+public:
+    explicit ConfigureGeneral(QWidget *parent = nullptr);
+    ~ConfigureGeneral();
+
+    void applyConfiguration();
+
+private:
+    void setConfiguration();
+
+private:
+    std::unique_ptr<Ui::ConfigureGeneral> ui;
+};
diff --git a/src/citra_qt/configure_general.ui b/src/citra_qt/configure_general.ui
new file mode 100644
index 000000000..47184c5c6
--- /dev/null
+++ b/src/citra_qt/configure_general.ui
@@ -0,0 +1,166 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<ui version="4.0">
+ <class>ConfigureGeneral</class>
+ <widget class="QWidget" name="ConfigureGeneral">
+  <property name="geometry">
+   <rect>
+    <x>0</x>
+    <y>0</y>
+    <width>300</width>
+    <height>377</height>
+   </rect>
+  </property>
+  <property name="windowTitle">
+   <string>Form</string>
+  </property>
+  <layout class="QHBoxLayout" name="horizontalLayout">
+   <item>
+    <layout class="QVBoxLayout" name="verticalLayout">
+     <item>
+      <widget class="QGroupBox" name="groupBox">
+       <property name="title">
+        <string>General</string>
+       </property>
+       <layout class="QHBoxLayout" name="horizontalLayout_3">
+        <item>
+         <layout class="QVBoxLayout" name="verticalLayout_2">
+          <item>
+           <widget class="QCheckBox" name="toogle_deepscan">
+            <property name="text">
+             <string>Recursive scan for game folder</string>
+            </property>
+           </widget>
+          </item>
+          <item>
+           <widget class="QCheckBox" name="toogle_check_exit">
+            <property name="text">
+             <string>Confirm exit while emulation is running</string>
+            </property>
+           </widget>
+          </item>
+         </layout>
+        </item>
+       </layout>
+      </widget>
+     </item>
+     <item>
+      <widget class="QGroupBox" name="groupBox_4">
+       <property name="title">
+        <string>Emulation</string>
+       </property>
+       <layout class="QHBoxLayout" name="horizontalLayout_5">
+        <item>
+         <layout class="QVBoxLayout" name="verticalLayout_6">
+          <item>
+           <layout class="QHBoxLayout" name="horizontalLayout_6">
+            <item>
+             <widget class="QLabel" name="label">
+              <property name="text">
+               <string>Region:</string>
+              </property>
+             </widget>
+            </item>
+            <item>
+             <widget class="QComboBox" name="region_combobox">
+              <item>
+               <property name="text">
+                <string notr="true">JPN</string>
+               </property>
+              </item>
+              <item>
+               <property name="text">
+                <string notr="true">USA</string>
+               </property>
+              </item>
+              <item>
+               <property name="text">
+                <string notr="true">EUR</string>
+               </property>
+              </item>
+              <item>
+               <property name="text">
+                <string notr="true">AUS</string>
+               </property>
+              </item>
+              <item>
+               <property name="text">
+                <string notr="true">CHN</string>
+               </property>
+              </item>
+              <item>
+               <property name="text">
+                <string notr="true">KOR</string>
+               </property>
+              </item>
+              <item>
+               <property name="text">
+                <string notr="true">TWN</string>
+               </property>
+              </item>
+             </widget>
+            </item>
+           </layout>
+          </item>
+         </layout>
+        </item>
+       </layout>
+      </widget>
+     </item>
+     <item>
+      <widget class="QGroupBox" name="groupBox_2">
+       <property name="title">
+        <string>Performance</string>
+       </property>
+       <layout class="QHBoxLayout" name="horizontalLayout_2">
+        <item>
+         <layout class="QVBoxLayout" name="verticalLayout_3">
+          <item>
+           <widget class="QCheckBox" name="toogle_hw_renderer">
+            <property name="text">
+             <string>Enable hardware renderer</string>
+            </property>
+           </widget>
+          </item>
+          <item>
+           <widget class="QCheckBox" name="toogle_shader_jit">
+            <property name="text">
+             <string>Enable shader JIT</string>
+            </property>
+           </widget>
+          </item>
+         </layout>
+        </item>
+       </layout>
+      </widget>
+     </item>
+     <item>
+      <widget class="QGroupBox" name="groupBox_3">
+       <property name="title">
+        <string>Hotkeys</string>
+       </property>
+       <layout class="QHBoxLayout" name="horizontalLayout_4">
+        <item>
+         <layout class="QVBoxLayout" name="verticalLayout_4">
+          <item>
+           <widget class="GHotkeysDialog" name="widget" native="true"/>
+          </item>
+         </layout>
+        </item>
+       </layout>
+      </widget>
+     </item>
+    </layout>
+   </item>
+  </layout>
+ </widget>
+ <customwidgets>
+  <customwidget>
+   <class>GHotkeysDialog</class>
+   <extends>QWidget</extends>
+   <header>hotkeys.h</header>
+   <container>1</container>
+  </customwidget>
+ </customwidgets>
+ <resources/>
+ <connections/>
+</ui>
diff --git a/src/citra_qt/game_list.cpp b/src/citra_qt/game_list.cpp
index a0b216b0a..d14532102 100644
--- a/src/citra_qt/game_list.cpp
+++ b/src/citra_qt/game_list.cpp
@@ -8,6 +8,7 @@
 
 #include "game_list.h"
 #include "game_list_p.h"
+#include "ui_settings.h"
 
 #include "core/loader/loader.h"
 
@@ -66,7 +67,7 @@ void GameList::ValidateEntry(const QModelIndex& item)
 
     if (file_path.isEmpty())
         return;
-    std::string std_file_path(file_path.toLocal8Bit());
+    std::string std_file_path(file_path.toStdString());
     if (!FileUtil::Exists(std_file_path) || FileUtil::IsDirectory(std_file_path))
         return;
     emit GameChosen(file_path);
@@ -100,19 +101,15 @@ void GameList::PopulateAsync(const QString& dir_path, bool deep_scan)
     current_worker = std::move(worker);
 }
 
-void GameList::SaveInterfaceLayout(QSettings& settings)
+void GameList::SaveInterfaceLayout()
 {
-    settings.beginGroup("UILayout");
-    settings.setValue("gameListHeaderState", tree_view->header()->saveState());
-    settings.endGroup();
+    UISettings::values.gamelist_header_state = tree_view->header()->saveState();
 }
 
-void GameList::LoadInterfaceLayout(QSettings& settings)
+void GameList::LoadInterfaceLayout()
 {
     auto header = tree_view->header();
-    settings.beginGroup("UILayout");
-    header->restoreState(settings.value("gameListHeaderState").toByteArray());
-    settings.endGroup();
+    header->restoreState(UISettings::values.gamelist_header_state);
 
     item_model->sort(header->sortIndicatorSection(), header->sortIndicatorOrder());
 }
@@ -148,7 +145,7 @@ void GameListWorker::AddFstEntriesToGameList(const std::string& dir_path, bool d
 
             emit EntryReady({
                 new GameListItem(QString::fromStdString(Loader::GetFileTypeString(filetype))),
-                new GameListItemPath(QString::fromLocal8Bit(physical_name.c_str())),
+                new GameListItemPath(QString::fromStdString(physical_name)),
                 new GameListItemSize(FileUtil::GetSize(physical_name)),
             });
         }
diff --git a/src/citra_qt/game_list.h b/src/citra_qt/game_list.h
index 0950d9622..48febdc60 100644
--- a/src/citra_qt/game_list.h
+++ b/src/citra_qt/game_list.h
@@ -31,8 +31,8 @@ public:
 
     void PopulateAsync(const QString& dir_path, bool deep_scan);
 
-    void SaveInterfaceLayout(QSettings& settings);
-    void LoadInterfaceLayout(QSettings& settings);
+    void SaveInterfaceLayout();
+    void LoadInterfaceLayout();
 
 public slots:
     void AddEntry(QList<QStandardItem*> entry_items);
diff --git a/src/citra_qt/hotkeys.cpp b/src/citra_qt/hotkeys.cpp
index ed6b12fc4..41f95c63d 100644
--- a/src/citra_qt/hotkeys.cpp
+++ b/src/citra_qt/hotkeys.cpp
@@ -4,11 +4,12 @@
 
 #include <map>
 
+#include <QtGlobal>
 #include <QKeySequence>
-#include <QSettings>
 #include <QShortcut>
 
 #include "citra_qt/hotkeys.h"
+#include "citra_qt/ui_settings.h"
 
 struct Hotkey
 {
@@ -24,54 +25,39 @@ typedef std::map<QString, HotkeyMap> HotkeyGroupMap;
 
 HotkeyGroupMap hotkey_groups;
 
-void SaveHotkeys(QSettings& settings)
+void SaveHotkeys()
 {
-    settings.beginGroup("Shortcuts");
-
+    UISettings::values.shortcuts.clear();
     for (auto group : hotkey_groups)
     {
-        settings.beginGroup(group.first);
         for (auto hotkey : group.second)
         {
-            settings.beginGroup(hotkey.first);
-            settings.setValue(QString("KeySeq"), hotkey.second.keyseq.toString());
-            settings.setValue(QString("Context"), hotkey.second.context);
-            settings.endGroup();
+            UISettings::values.shortcuts.emplace_back(
+                        UISettings::Shortcut(group.first + "/" + hotkey.first,
+                                             UISettings::ContextualShortcut(hotkey.second.keyseq.toString(),
+                                                                           hotkey.second.context)));
         }
-        settings.endGroup();
     }
-    settings.endGroup();
 }
 
-void LoadHotkeys(QSettings& settings)
+void LoadHotkeys()
 {
-    settings.beginGroup("Shortcuts");
-
     // Make sure NOT to use a reference here because it would become invalid once we call beginGroup()
-    QStringList groups = settings.childGroups();
-    for (auto group : groups)
+    for (auto shortcut : UISettings::values.shortcuts)
     {
-        settings.beginGroup(group);
+        QStringList cat = shortcut.first.split("/");
+        Q_ASSERT(cat.size() >= 2);
 
-        QStringList hotkeys = settings.childGroups();
-        for (auto hotkey : hotkeys)
+        // RegisterHotkey assigns default keybindings, so use old values as default parameters
+        Hotkey& hk = hotkey_groups[cat[0]][cat[1]];
+        if (!shortcut.second.first.isEmpty())
         {
-            settings.beginGroup(hotkey);
-
-            // RegisterHotkey assigns default keybindings, so use old values as default parameters
-            Hotkey& hk = hotkey_groups[group][hotkey];
-            hk.keyseq = QKeySequence::fromString(settings.value("KeySeq", hk.keyseq.toString()).toString());
-            hk.context = (Qt::ShortcutContext)settings.value("Context", hk.context).toInt();
-            if (hk.shortcut)
-                hk.shortcut->setKey(hk.keyseq);
-
-            settings.endGroup();
+            hk.keyseq = QKeySequence::fromString(shortcut.second.first);
+            hk.context = (Qt::ShortcutContext)shortcut.second.second;
         }
-
-        settings.endGroup();
+        if (hk.shortcut)
+            hk.shortcut->setKey(hk.keyseq);
     }
-
-    settings.endGroup();
 }
 
 void RegisterHotkey(const QString& group, const QString& action, const QKeySequence& default_keyseq, Qt::ShortcutContext default_context)
@@ -94,7 +80,7 @@ QShortcut* GetHotkey(const QString& group, const QString& action, QWidget* widge
 }
 
 
-GHotkeysDialog::GHotkeysDialog(QWidget* parent): QDialog(parent)
+GHotkeysDialog::GHotkeysDialog(QWidget* parent): QWidget(parent)
 {
     ui.setupUi(this);
 
diff --git a/src/citra_qt/hotkeys.h b/src/citra_qt/hotkeys.h
index 2fe635882..38aa5f012 100644
--- a/src/citra_qt/hotkeys.h
+++ b/src/citra_qt/hotkeys.h
@@ -2,6 +2,8 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#pragma once
+
 #include "ui_hotkeys.h"
 
 class QDialog;
@@ -33,16 +35,16 @@ QShortcut* GetHotkey(const QString& group, const QString& action, QWidget* widge
  *
  * @note Each hotkey group will be stored a settings group; For each hotkey inside that group, a settings group will be created to store the key sequence and the hotkey context.
  */
-void SaveHotkeys(QSettings& settings);
+void SaveHotkeys();
 
 /**
  * Loads hotkeys from the settings file.
  *
  * @note Yet unregistered hotkeys which are present in the settings will automatically be registered.
  */
-void LoadHotkeys(QSettings& settings);
+void LoadHotkeys();
 
-class GHotkeysDialog : public QDialog
+class GHotkeysDialog : public QWidget
 {
     Q_OBJECT
 
diff --git a/src/citra_qt/hotkeys.ui b/src/citra_qt/hotkeys.ui
index 38a9a14d1..050fe064e 100644
--- a/src/citra_qt/hotkeys.ui
+++ b/src/citra_qt/hotkeys.ui
@@ -1,7 +1,7 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <ui version="4.0">
  <class>hotkeys</class>
- <widget class="QDialog" name="hotkeys">
+ <widget class="QWidget" name="hotkeys">
   <property name="geometry">
    <rect>
     <x>0</x>
@@ -39,51 +39,8 @@
      </column>
     </widget>
    </item>
-   <item>
-    <widget class="QDialogButtonBox" name="buttonBox">
-     <property name="orientation">
-      <enum>Qt::Horizontal</enum>
-     </property>
-     <property name="standardButtons">
-      <set>QDialogButtonBox::Cancel|QDialogButtonBox::Ok|QDialogButtonBox::Reset</set>
-     </property>
-    </widget>
-   </item>
   </layout>
  </widget>
  <resources/>
- <connections>
-  <connection>
-   <sender>buttonBox</sender>
-   <signal>accepted()</signal>
-   <receiver>hotkeys</receiver>
-   <slot>accept()</slot>
-   <hints>
-    <hint type="sourcelabel">
-     <x>248</x>
-     <y>254</y>
-    </hint>
-    <hint type="destinationlabel">
-     <x>157</x>
-     <y>274</y>
-    </hint>
-   </hints>
-  </connection>
-  <connection>
-   <sender>buttonBox</sender>
-   <signal>rejected()</signal>
-   <receiver>hotkeys</receiver>
-   <slot>reject()</slot>
-   <hints>
-    <hint type="sourcelabel">
-     <x>316</x>
-     <y>260</y>
-    </hint>
-    <hint type="destinationlabel">
-     <x>286</x>
-     <y>274</y>
-    </hint>
-   </hints>
-  </connection>
- </connections>
+ <connections/>
 </ui>
diff --git a/src/citra_qt/main.cpp b/src/citra_qt/main.cpp
index 32cceaf7e..2ca1e51f6 100644
--- a/src/citra_qt/main.cpp
+++ b/src/citra_qt/main.cpp
@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include <clocale>
+#include <memory>
 #include <thread>
 
 #include <QDesktopWidget>
@@ -13,9 +14,11 @@
 
 #include "citra_qt/bootmanager.h"
 #include "citra_qt/config.h"
+#include "citra_qt/configure_dialog.h"
 #include "citra_qt/game_list.h"
 #include "citra_qt/hotkeys.h"
 #include "citra_qt/main.h"
+#include "citra_qt/ui_settings.h"
 
 // Debugger
 #include "citra_qt/debugger/callstack.h"
@@ -30,7 +33,6 @@
 #include "citra_qt/debugger/ramview.h"
 #include "citra_qt/debugger/registers.h"
 
-#include "common/make_unique.h"
 #include "common/microprofile.h"
 #include "common/platform.h"
 #include "common/scm_rev.h"
@@ -50,12 +52,10 @@
 
 #include "video_core/video_core.h"
 
-GMainWindow::GMainWindow() : emu_thread(nullptr)
+GMainWindow::GMainWindow() : config(new Config()), emu_thread(nullptr)
 {
     Pica::g_debug_context = Pica::DebugContext::Construct();
 
-    Config config;
-
     ui.setupUi(this);
     statusBar()->hide();
 
@@ -133,33 +133,18 @@ GMainWindow::GMainWindow() : emu_thread(nullptr)
     setGeometry(x, y, w, h);
 
     // Restore UI state
-    QSettings settings;
-
-    settings.beginGroup("UILayout");
-    restoreGeometry(settings.value("geometry").toByteArray());
-    restoreState(settings.value("state").toByteArray());
-    render_window->restoreGeometry(settings.value("geometryRenderWindow").toByteArray());
-    microProfileDialog->restoreGeometry(settings.value("microProfileDialogGeometry").toByteArray());
-    microProfileDialog->setVisible(settings.value("microProfileDialogVisible").toBool());
-    settings.endGroup();
-
-    game_list->LoadInterfaceLayout(settings);
-
-    ui.action_Use_Gdbstub->setChecked(Settings::values.use_gdbstub);
-    SetGdbstubEnabled(ui.action_Use_Gdbstub->isChecked());
-
-    GDBStub::SetServerPort(static_cast<u32>(Settings::values.gdbstub_port));
-
-    ui.action_Use_Hardware_Renderer->setChecked(Settings::values.use_hw_renderer);
-    SetHardwareRendererEnabled(ui.action_Use_Hardware_Renderer->isChecked());
+    restoreGeometry(UISettings::values.geometry);
+    restoreState(UISettings::values.state);
+    render_window->restoreGeometry(UISettings::values.renderwindow_geometry);
+    microProfileDialog->restoreGeometry(UISettings::values.microprofile_geometry);
+    microProfileDialog->setVisible(UISettings::values.microprofile_visible);
 
-    ui.action_Use_Shader_JIT->setChecked(Settings::values.use_shader_jit);
-    SetShaderJITEnabled(ui.action_Use_Shader_JIT->isChecked());
+    game_list->LoadInterfaceLayout();
 
-    ui.action_Single_Window_Mode->setChecked(settings.value("singleWindowMode", true).toBool());
+    ui.action_Single_Window_Mode->setChecked(UISettings::values.single_window_mode);
     ToggleWindowMode();
 
-    ui.actionDisplay_widget_title_bars->setChecked(settings.value("displayTitleBars", true).toBool());
+    ui.actionDisplay_widget_title_bars->setChecked(UISettings::values.display_titlebar);
     OnDisplayTitleBars(ui.actionDisplay_widget_title_bars->isChecked());
 
     // Prepare actions for recent files
@@ -172,21 +157,16 @@ GMainWindow::GMainWindow() : emu_thread(nullptr)
     }
     UpdateRecentFiles();
 
-    confirm_before_closing = settings.value("confirmClose", true).toBool();
-
     // Setup connections
-    connect(game_list, SIGNAL(GameChosen(QString)), this, SLOT(OnGameListLoadFile(QString)));
-    connect(ui.action_Load_File, SIGNAL(triggered()), this, SLOT(OnMenuLoadFile()));
+    connect(game_list, SIGNAL(GameChosen(QString)), this, SLOT(OnGameListLoadFile(QString)), Qt::DirectConnection);
+    connect(ui.action_Configure, SIGNAL(triggered()), this, SLOT(OnConfigure()));
+    connect(ui.action_Load_File, SIGNAL(triggered()), this, SLOT(OnMenuLoadFile()),Qt::DirectConnection);
     connect(ui.action_Load_Symbol_Map, SIGNAL(triggered()), this, SLOT(OnMenuLoadSymbolMap()));
     connect(ui.action_Select_Game_List_Root, SIGNAL(triggered()), this, SLOT(OnMenuSelectGameListRoot()));
     connect(ui.action_Start, SIGNAL(triggered()), this, SLOT(OnStartGame()));
     connect(ui.action_Pause, SIGNAL(triggered()), this, SLOT(OnPauseGame()));
     connect(ui.action_Stop, SIGNAL(triggered()), this, SLOT(OnStopGame()));
-    connect(ui.action_Use_Hardware_Renderer, SIGNAL(triggered(bool)), this, SLOT(SetHardwareRendererEnabled(bool)));
-    connect(ui.action_Use_Shader_JIT, SIGNAL(triggered(bool)), this, SLOT(SetShaderJITEnabled(bool)));
-    connect(ui.action_Use_Gdbstub, SIGNAL(triggered(bool)), this, SLOT(SetGdbstubEnabled(bool)));
     connect(ui.action_Single_Window_Mode, SIGNAL(triggered(bool)), this, SLOT(ToggleWindowMode()));
-    connect(ui.action_Hotkeys, SIGNAL(triggered()), this, SLOT(OnOpenHotkeysDialog()));
 
     connect(this, SIGNAL(EmulationStarting(EmuThread*)), disasmWidget, SLOT(OnEmulationStarting(EmuThread*)));
     connect(this, SIGNAL(EmulationStopping()), disasmWidget, SLOT(OnEmulationStopping()));
@@ -201,7 +181,7 @@ GMainWindow::GMainWindow() : emu_thread(nullptr)
     // Setup hotkeys
     RegisterHotkey("Main Window", "Load File", QKeySequence::Open);
     RegisterHotkey("Main Window", "Start Emulation");
-    LoadHotkeys(settings);
+    LoadHotkeys();
 
     connect(GetHotkey("Main Window", "Load File", this), SIGNAL(activated()), this, SLOT(OnMenuLoadFile()));
     connect(GetHotkey("Main Window", "Start Emulation", this), SIGNAL(activated()), this, SLOT(OnStartGame()));
@@ -211,7 +191,7 @@ GMainWindow::GMainWindow() : emu_thread(nullptr)
 
     show();
 
-    game_list->PopulateAsync(settings.value("gameListRootDir", ".").toString(), settings.value("gameListDeepScan", false).toBool());
+    game_list->PopulateAsync(UISettings::values.gamedir, UISettings::values.gamedir_deepscan);
 
     QStringList args = QApplication::arguments();
     if (args.length() >= 2) {
@@ -319,7 +299,7 @@ void GMainWindow::BootGame(const std::string& filename) {
         return;
 
     // Create and start the emulation thread
-    emu_thread = Common::make_unique<EmuThread>(render_window);
+    emu_thread = std::make_unique<EmuThread>(render_window);
     emit EmulationStarting(emu_thread.get());
     render_window->moveContext();
     emu_thread->start();
@@ -375,32 +355,24 @@ void GMainWindow::ShutdownGame() {
     emulation_running = false;
 }
 
-void GMainWindow::StoreRecentFile(const std::string& filename)
-{
-    QSettings settings;
-    QStringList recent_files = settings.value("recentFiles").toStringList();
-    recent_files.prepend(QString::fromStdString(filename));
-    recent_files.removeDuplicates();
-    while (recent_files.size() > max_recent_files_item) {
-        recent_files.removeLast();
+void GMainWindow::StoreRecentFile(const std::string& filename) {
+    UISettings::values.recent_files.prepend(QString::fromStdString(filename));
+    UISettings::values.recent_files.removeDuplicates();
+    while (UISettings::values.recent_files.size() > max_recent_files_item) {
+        UISettings::values.recent_files.removeLast();
     }
 
-    settings.setValue("recentFiles", recent_files);
-
     UpdateRecentFiles();
 }
 
 void GMainWindow::UpdateRecentFiles() {
-    QSettings settings;
-    QStringList recent_files = settings.value("recentFiles").toStringList();
-
-    unsigned int num_recent_files = std::min(recent_files.size(), static_cast<int>(max_recent_files_item));
+    unsigned int num_recent_files = std::min(UISettings::values.recent_files.size(), static_cast<int>(max_recent_files_item));
 
     for (unsigned int i = 0; i < num_recent_files; i++) {
-        QString text = QString("&%1. %2").arg(i + 1).arg(QFileInfo(recent_files[i]).fileName());
+        QString text = QString("&%1. %2").arg(i + 1).arg(QFileInfo(UISettings::values.recent_files[i]).fileName());
         actions_recent_files[i]->setText(text);
-        actions_recent_files[i]->setData(recent_files[i]);
-        actions_recent_files[i]->setToolTip(recent_files[i]);
+        actions_recent_files[i]->setData(UISettings::values.recent_files[i]);
+        actions_recent_files[i]->setToolTip(UISettings::values.recent_files[i]);
         actions_recent_files[i]->setVisible(true);
     }
 
@@ -417,40 +389,32 @@ void GMainWindow::UpdateRecentFiles() {
 }
 
 void GMainWindow::OnGameListLoadFile(QString game_path) {
-    BootGame(game_path.toLocal8Bit().data());
+    BootGame(game_path.toStdString());
 }
 
 void GMainWindow::OnMenuLoadFile() {
-    QSettings settings;
-    QString rom_path = settings.value("romsPath", QString()).toString();
-
-    QString filename = QFileDialog::getOpenFileName(this, tr("Load File"), rom_path, tr("3DS executable (*.3ds *.3dsx *.elf *.axf *.cci *.cxi)"));
+    QString filename = QFileDialog::getOpenFileName(this, tr("Load File"), UISettings::values.roms_path, tr("3DS executable (*.3ds *.3dsx *.elf *.axf *.cci *.cxi)"));
     if (!filename.isEmpty()) {
-        settings.setValue("romsPath", QFileInfo(filename).path());
+        UISettings::values.roms_path = QFileInfo(filename).path();
 
-        BootGame(filename.toLocal8Bit().data());
+        BootGame(filename.toStdString());
     }
 }
 
 void GMainWindow::OnMenuLoadSymbolMap() {
-    QSettings settings;
-    QString symbol_path = settings.value("symbolsPath", QString()).toString();
-
-    QString filename = QFileDialog::getOpenFileName(this, tr("Load Symbol Map"), symbol_path, tr("Symbol map (*)"));
+    QString filename = QFileDialog::getOpenFileName(this, tr("Load Symbol Map"), UISettings::values.symbols_path, tr("Symbol map (*)"));
     if (!filename.isEmpty()) {
-        settings.setValue("symbolsPath", QFileInfo(filename).path());
+        UISettings::values.symbols_path = QFileInfo(filename).path();
 
-        LoadSymbolMap(filename.toLocal8Bit().data());
+        LoadSymbolMap(filename.toStdString());
     }
 }
 
 void GMainWindow::OnMenuSelectGameListRoot() {
-    QSettings settings;
-
     QString dir_path = QFileDialog::getExistingDirectory(this, tr("Select Directory"));
     if (!dir_path.isEmpty()) {
-        settings.setValue("gameListRootDir", dir_path);
-        game_list->PopulateAsync(dir_path, settings.value("gameListDeepScan").toBool());
+        UISettings::values.gamedir = dir_path;
+        game_list->PopulateAsync(dir_path, UISettings::values.gamedir_deepscan);
     }
 }
 
@@ -461,15 +425,12 @@ void GMainWindow::OnMenuRecentFile() {
     QString filename = action->data().toString();
     QFileInfo file_info(filename);
     if (file_info.exists()) {
-        BootGame(filename.toLocal8Bit().data());
+        BootGame(filename.toStdString());
     } else {
         // Display an error message and remove the file from the list.
         QMessageBox::information(this, tr("File not found"), tr("File \"%1\" not found").arg(filename));
 
-        QSettings settings;
-        QStringList recent_files = settings.value("recentFiles").toStringList();
-        recent_files.removeOne(filename);
-        settings.setValue("recentFiles", recent_files);
+        UISettings::values.recent_files.removeOne(filename);
         UpdateRecentFiles();
     }
 }
@@ -496,31 +457,6 @@ void GMainWindow::OnStopGame() {
     ShutdownGame();
 }
 
-void GMainWindow::OnOpenHotkeysDialog() {
-    GHotkeysDialog dialog(this);
-    dialog.exec();
-}
-
-void GMainWindow::SetHardwareRendererEnabled(bool enabled) {
-    VideoCore::g_hw_renderer_enabled = enabled;
-
-    Config config;
-    Settings::values.use_hw_renderer = enabled;
-    config.Save();
-}
-
-void GMainWindow::SetGdbstubEnabled(bool enabled) {
-    GDBStub::ToggleServer(enabled);
-}
-
-void GMainWindow::SetShaderJITEnabled(bool enabled) {
-    VideoCore::g_shader_jit_enabled = enabled;
-
-    Config config;
-    Settings::values.use_shader_jit = enabled;
-    config.Save();
-}
-
 void GMainWindow::ToggleWindowMode() {
     if (ui.action_Single_Window_Mode->isChecked()) {
         // Render in the main window...
@@ -547,11 +483,17 @@ void GMainWindow::ToggleWindowMode() {
 }
 
 void GMainWindow::OnConfigure() {
-    //GControllerConfigDialog* dialog = new GControllerConfigDialog(controller_ports, this);
+    ConfigureDialog configureDialog(this);
+    auto result = configureDialog.exec();
+    if (result == QDialog::Accepted)
+    {
+        configureDialog.applyConfiguration();
+        config->Save();
+    }
 }
 
 bool GMainWindow::ConfirmClose() {
-    if (emu_thread == nullptr || !confirm_before_closing)
+    if (emu_thread == nullptr || !UISettings::values.confirm_before_closing)
         return true;
 
     auto answer = QMessageBox::question(this, tr("Citra"),
@@ -566,23 +508,18 @@ void GMainWindow::closeEvent(QCloseEvent* event) {
         return;
     }
 
-    // Save window layout
-    QSettings settings(QSettings::IniFormat, QSettings::UserScope, "Citra team", "Citra");
+    UISettings::values.geometry = saveGeometry();
+    UISettings::values.state = saveState();
+    UISettings::values.renderwindow_geometry = render_window->saveGeometry();
+    UISettings::values.microprofile_geometry = microProfileDialog->saveGeometry();
+    UISettings::values.microprofile_visible = microProfileDialog->isVisible();
 
-    settings.beginGroup("UILayout");
-    settings.setValue("geometry", saveGeometry());
-    settings.setValue("state", saveState());
-    settings.setValue("geometryRenderWindow", render_window->saveGeometry());
-    settings.setValue("microProfileDialogGeometry", microProfileDialog->saveGeometry());
-    settings.setValue("microProfileDialogVisible", microProfileDialog->isVisible());
-    settings.endGroup();
+    UISettings::values.single_window_mode = ui.action_Single_Window_Mode->isChecked();
+    UISettings::values.display_titlebar = ui.actionDisplay_widget_title_bars->isChecked();
+    UISettings::values.first_start = false;
 
-    settings.setValue("singleWindowMode", ui.action_Single_Window_Mode->isChecked());
-    settings.setValue("displayTitleBars", ui.actionDisplay_widget_title_bars->isChecked());
-    settings.setValue("firstStart", false);
-    settings.setValue("confirmClose", confirm_before_closing);
-    game_list->SaveInterfaceLayout(settings);
-    SaveHotkeys(settings);
+    game_list->SaveInterfaceLayout();
+    SaveHotkeys();
 
     // Shutdown session if the emu thread is active...
     if (emu_thread != nullptr)
@@ -607,7 +544,6 @@ int main(int argc, char* argv[]) {
     });
 
     // Init settings params
-    QSettings::setDefaultFormat(QSettings::IniFormat);
     QCoreApplication::setOrganizationName("Citra team");
     QCoreApplication::setApplicationName("Citra");
 
diff --git a/src/citra_qt/main.h b/src/citra_qt/main.h
index 6e4e56689..477db5c5c 100644
--- a/src/citra_qt/main.h
+++ b/src/citra_qt/main.h
@@ -10,6 +10,7 @@
 
 #include "ui_main.h"
 
+class Config;
 class GameList;
 class GImageInfo;
 class GRenderWindow;
@@ -104,12 +105,8 @@ private slots:
     /// Called whenever a user selects the "File->Select Game List Root" menu item
     void OnMenuSelectGameListRoot();
     void OnMenuRecentFile();
-    void OnOpenHotkeysDialog();
     void OnConfigure();
     void OnDisplayTitleBars(bool);
-    void SetHardwareRendererEnabled(bool);
-    void SetGdbstubEnabled(bool);
-    void SetShaderJITEnabled(bool);
     void ToggleWindowMode();
 
 private:
@@ -118,6 +115,8 @@ private:
     GRenderWindow* render_window;
     GameList* game_list;
 
+    std::unique_ptr<Config> config;
+
     // Whether emulation is currently running in Citra.
     bool emulation_running = false;
     std::unique_ptr<EmuThread> emu_thread;
@@ -131,7 +130,6 @@ private:
     GPUCommandListWidget* graphicsCommandsWidget;
 
     QAction* actions_recent_files[max_recent_files_item];
-    bool confirm_before_closing;
 };
 
 #endif // _CITRA_QT_MAIN_HXX_
diff --git a/src/citra_qt/main.ui b/src/citra_qt/main.ui
index 1e8a07cfb..441e0b81e 100644
--- a/src/citra_qt/main.ui
+++ b/src/citra_qt/main.ui
@@ -45,7 +45,7 @@
      <x>0</x>
      <y>0</y>
      <width>1081</width>
-     <height>22</height>
+     <height>19</height>
     </rect>
    </property>
    <widget class="QMenu" name="menu_File">
@@ -73,9 +73,6 @@
     <addaction name="action_Pause"/>
     <addaction name="action_Stop"/>
     <addaction name="separator"/>
-    <addaction name="action_Use_Hardware_Renderer"/>
-    <addaction name="action_Use_Shader_JIT"/>
-    <addaction name="action_Use_Gdbstub"/>
     <addaction name="action_Configure"/>
    </widget>
    <widget class="QMenu" name="menu_View">
@@ -84,7 +81,6 @@
     </property>
     <addaction name="action_Single_Window_Mode"/>
     <addaction name="actionDisplay_widget_title_bars"/>
-    <addaction name="action_Hotkeys"/>
    </widget>
    <widget class="QMenu" name="menu_Help">
     <property name="title">
@@ -150,35 +146,6 @@
     <string>Single Window Mode</string>
    </property>
   </action>
-  <action name="action_Hotkeys">
-   <property name="text">
-    <string>Configure &amp;Hotkeys ...</string>
-   </property>
-  </action>
-  <action name="action_Use_Hardware_Renderer">
-   <property name="checkable">
-    <bool>true</bool>
-   </property>
-   <property name="text">
-    <string>Use Hardware Renderer</string>
-   </property>
-  </action>
-  <action name="action_Use_Shader_JIT">
-   <property name="checkable">
-    <bool>true</bool>
-   </property>
-   <property name="text">
-    <string>Use Shader JIT</string>
-   </property>
-  </action>
-  <action name="action_Use_Gdbstub">
-    <property name="checkable">
-      <bool>true</bool>
-    </property>
-    <property name="text">
-      <string>Use Gdbstub</string>
-    </property>
-  </action>
   <action name="action_Configure">
    <property name="text">
     <string>Configure ...</string>
@@ -220,22 +187,6 @@
    </hints>
   </connection>
   <connection>
-   <sender>action_Configure</sender>
-   <signal>triggered()</signal>
-   <receiver>MainWindow</receiver>
-   <slot>OnConfigure()</slot>
-   <hints>
-    <hint type="sourcelabel">
-     <x>-1</x>
-     <y>-1</y>
-    </hint>
-    <hint type="destinationlabel">
-     <x>540</x>
-     <y>364</y>
-    </hint>
-   </hints>
-  </connection>
-  <connection>
    <sender>actionDisplay_widget_title_bars</sender>
    <signal>triggered(bool)</signal>
    <receiver>MainWindow</receiver>
diff --git a/src/citra_qt/ui_settings.cpp b/src/citra_qt/ui_settings.cpp
new file mode 100644
index 000000000..5f2215899
--- /dev/null
+++ b/src/citra_qt/ui_settings.cpp
@@ -0,0 +1,11 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "ui_settings.h"
+
+namespace UISettings {
+
+Values values = {};
+
+}
diff --git a/src/citra_qt/ui_settings.h b/src/citra_qt/ui_settings.h
new file mode 100644
index 000000000..62db4a73e
--- /dev/null
+++ b/src/citra_qt/ui_settings.h
@@ -0,0 +1,47 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <QByteArray>
+#include <QStringList>
+#include <QString>
+
+#include <vector>
+
+namespace UISettings {
+
+using ContextualShortcut = std::pair<QString, int> ;
+using Shortcut = std::pair<QString, ContextualShortcut>;
+
+struct Values {
+    QByteArray geometry;
+    QByteArray state;
+
+    QByteArray renderwindow_geometry;
+
+    QByteArray gamelist_header_state;
+
+    QByteArray microprofile_geometry;
+    bool microprofile_visible;
+
+    bool single_window_mode;
+    bool display_titlebar;
+
+    bool confirm_before_closing;
+    bool first_start;
+
+    QString roms_path;
+    QString symbols_path;
+    QString gamedir;
+    bool gamedir_deepscan;
+    QStringList recent_files;
+
+    // Shortcut name <Shortcut, context>
+    std::vector<Shortcut> shortcuts;
+};
+
+extern Values values;
+
+}
diff --git a/src/common/CMakeLists.txt b/src/common/CMakeLists.txt
index 1c9be718f..c839ce173 100644
--- a/src/common/CMakeLists.txt
+++ b/src/common/CMakeLists.txt
@@ -42,7 +42,6 @@ set(HEADERS
             logging/filter.h
             logging/log.h
             logging/backend.h
-            make_unique.h
             math_util.h
             memory_util.h
             microprofile.h
diff --git a/src/common/file_util.cpp b/src/common/file_util.cpp
index c3061479a..53700c865 100644
--- a/src/common/file_util.cpp
+++ b/src/common/file_util.cpp
@@ -85,7 +85,7 @@ bool Exists(const std::string &filename)
     StripTailDirSlashes(copy);
 
 #ifdef _WIN32
-    int result = _tstat64(Common::UTF8ToTStr(copy).c_str(), &file_info);
+    int result = _wstat64(Common::UTF8ToUTF16W(copy).c_str(), &file_info);
 #else
     int result = stat64(copy.c_str(), &file_info);
 #endif
@@ -102,7 +102,7 @@ bool IsDirectory(const std::string &filename)
     StripTailDirSlashes(copy);
 
 #ifdef _WIN32
-    int result = _tstat64(Common::UTF8ToTStr(copy).c_str(), &file_info);
+    int result = _wstat64(Common::UTF8ToUTF16W(copy).c_str(), &file_info);
 #else
     int result = stat64(copy.c_str(), &file_info);
 #endif
@@ -138,7 +138,7 @@ bool Delete(const std::string &filename)
     }
 
 #ifdef _WIN32
-    if (!DeleteFile(Common::UTF8ToTStr(filename).c_str()))
+    if (!DeleteFileW(Common::UTF8ToUTF16W(filename).c_str()))
     {
         LOG_ERROR(Common_Filesystem, "DeleteFile failed on %s: %s",
                  filename.c_str(), GetLastErrorMsg());
@@ -160,7 +160,7 @@ bool CreateDir(const std::string &path)
 {
     LOG_TRACE(Common_Filesystem, "directory %s", path.c_str());
 #ifdef _WIN32
-    if (::CreateDirectory(Common::UTF8ToTStr(path).c_str(), nullptr))
+    if (::CreateDirectoryW(Common::UTF8ToUTF16W(path).c_str(), nullptr))
         return true;
     DWORD error = GetLastError();
     if (error == ERROR_ALREADY_EXISTS)
@@ -241,7 +241,7 @@ bool DeleteDir(const std::string &filename)
     }
 
 #ifdef _WIN32
-    if (::RemoveDirectory(Common::UTF8ToTStr(filename).c_str()))
+    if (::RemoveDirectoryW(Common::UTF8ToUTF16W(filename).c_str()))
         return true;
 #else
     if (rmdir(filename.c_str()) == 0)
@@ -257,8 +257,13 @@ bool Rename(const std::string &srcFilename, const std::string &destFilename)
 {
     LOG_TRACE(Common_Filesystem, "%s --> %s",
             srcFilename.c_str(), destFilename.c_str());
+#ifdef _WIN32
+    if (_wrename(Common::UTF8ToUTF16W(srcFilename).c_str(), Common::UTF8ToUTF16W(destFilename).c_str()) == 0)
+        return true;
+#else
     if (rename(srcFilename.c_str(), destFilename.c_str()) == 0)
         return true;
+#endif
     LOG_ERROR(Common_Filesystem, "failed %s --> %s: %s",
               srcFilename.c_str(), destFilename.c_str(), GetLastErrorMsg());
     return false;
@@ -270,7 +275,7 @@ bool Copy(const std::string &srcFilename, const std::string &destFilename)
     LOG_TRACE(Common_Filesystem, "%s --> %s",
             srcFilename.c_str(), destFilename.c_str());
 #ifdef _WIN32
-    if (CopyFile(Common::UTF8ToTStr(srcFilename).c_str(), Common::UTF8ToTStr(destFilename).c_str(), FALSE))
+    if (CopyFileW(Common::UTF8ToUTF16W(srcFilename).c_str(), Common::UTF8ToUTF16W(destFilename).c_str(), FALSE))
         return true;
 
     LOG_ERROR(Common_Filesystem, "failed %s --> %s: %s",
@@ -358,7 +363,7 @@ u64 GetSize(const std::string &filename)
 
     struct stat64 buf;
 #ifdef _WIN32
-    if (_tstat64(Common::UTF8ToTStr(filename).c_str(), &buf) == 0)
+    if (_wstat64(Common::UTF8ToUTF16W(filename).c_str(), &buf) == 0)
 #else
     if (stat64(filename.c_str(), &buf) == 0)
 #endif
@@ -432,16 +437,16 @@ bool ForeachDirectoryEntry(unsigned* num_entries_out, const std::string &directo
 
 #ifdef _WIN32
     // Find the first file in the directory.
-    WIN32_FIND_DATA ffd;
+    WIN32_FIND_DATAW ffd;
 
-    HANDLE handle_find = FindFirstFile(Common::UTF8ToTStr(directory + "\\*").c_str(), &ffd);
+    HANDLE handle_find = FindFirstFileW(Common::UTF8ToUTF16W(directory + "\\*").c_str(), &ffd);
     if (handle_find == INVALID_HANDLE_VALUE) {
         FindClose(handle_find);
         return false;
     }
     // windows loop
     do {
-        const std::string virtual_name(Common::TStrToUTF8(ffd.cFileName));
+        const std::string virtual_name(Common::UTF16ToUTF8(ffd.cFileName));
 #else
     struct dirent dirent, *result = nullptr;
 
@@ -465,7 +470,7 @@ bool ForeachDirectoryEntry(unsigned* num_entries_out, const std::string &directo
         found_entries += ret_entries;
 
 #ifdef _WIN32
-    } while (FindNextFile(handle_find, &ffd) != 0);
+    } while (FindNextFileW(handle_find, &ffd) != 0);
     FindClose(handle_find);
 #else
     }
@@ -572,15 +577,23 @@ void CopyDir(const std::string &source_path, const std::string &dest_path)
 // Returns the current directory
 std::string GetCurrentDir()
 {
-    char *dir;
     // Get the current working directory (getcwd uses malloc)
+#ifdef _WIN32
+    wchar_t *dir;
+    if (!(dir = _wgetcwd(nullptr, 0))) {
+#else
+    char *dir;
     if (!(dir = getcwd(nullptr, 0))) {
-
+#endif
         LOG_ERROR(Common_Filesystem, "GetCurrentDirectory failed: %s",
                 GetLastErrorMsg());
         return nullptr;
     }
+#ifdef _WIN32
+    std::string strDir = Common::UTF16ToUTF8(dir);
+#else
     std::string strDir = dir;
+#endif
     free(dir);
     return strDir;
 }
@@ -588,7 +601,11 @@ std::string GetCurrentDir()
 // Sets the current directory to the given directory
 bool SetCurrentDir(const std::string &directory)
 {
+#ifdef _WIN32
+    return _wchdir(Common::UTF8ToUTF16W(directory).c_str()) == 0;
+#else
     return chdir(directory.c_str()) == 0;
+#endif
 }
 
 #if defined(__APPLE__)
@@ -613,9 +630,9 @@ std::string& GetExeDirectory()
     static std::string exe_path;
     if (exe_path.empty())
     {
-        TCHAR tchar_exe_path[2048];
-        GetModuleFileName(nullptr, tchar_exe_path, 2048);
-        exe_path = Common::TStrToUTF8(tchar_exe_path);
+        wchar_t wchar_exe_path[2048];
+        GetModuleFileNameW(nullptr, wchar_exe_path, 2048);
+        exe_path = Common::UTF16ToUTF8(wchar_exe_path);
         exe_path = exe_path.substr(0, exe_path.find_last_of('\\'));
     }
     return exe_path;
@@ -807,13 +824,12 @@ size_t WriteStringToFile(bool text_file, const std::string &str, const char *fil
 
 size_t ReadFileToString(bool text_file, const char *filename, std::string &str)
 {
-    FileUtil::IOFile file(filename, text_file ? "r" : "rb");
-    auto const f = file.GetHandle();
+    IOFile file(filename, text_file ? "r" : "rb");
 
-    if (!f)
+    if (!file)
         return false;
 
-    str.resize(static_cast<u32>(GetSize(f)));
+    str.resize(static_cast<u32>(file.GetSize()));
     return file.ReadArray(&str[0], str.size());
 }
 
@@ -860,15 +876,10 @@ void SplitFilename83(const std::string& filename, std::array<char, 9>& short_nam
 }
 
 IOFile::IOFile()
-    : m_file(nullptr), m_good(true)
-{}
-
-IOFile::IOFile(std::FILE* file)
-    : m_file(file), m_good(true)
-{}
+{
+}
 
 IOFile::IOFile(const std::string& filename, const char openmode[])
-    : m_file(nullptr), m_good(true)
 {
     Open(filename, openmode);
 }
@@ -879,7 +890,6 @@ IOFile::~IOFile()
 }
 
 IOFile::IOFile(IOFile&& other)
-    : m_file(nullptr), m_good(true)
 {
     Swap(other);
 }
@@ -900,7 +910,7 @@ bool IOFile::Open(const std::string& filename, const char openmode[])
 {
     Close();
 #ifdef _WIN32
-    _tfopen_s(&m_file, Common::UTF8ToTStr(filename).c_str(), Common::UTF8ToTStr(openmode).c_str());
+    _wfopen_s(&m_file, Common::UTF8ToUTF16W(filename).c_str(), Common::UTF8ToUTF16W(openmode).c_str());
 #else
     m_file = fopen(filename.c_str(), openmode);
 #endif
@@ -918,26 +928,12 @@ bool IOFile::Close()
     return m_good;
 }
 
-std::FILE* IOFile::ReleaseHandle()
-{
-    std::FILE* const ret = m_file;
-    m_file = nullptr;
-    return ret;
-}
-
-void IOFile::SetHandle(std::FILE* file)
-{
-    Close();
-    Clear();
-    m_file = file;
-}
-
-u64 IOFile::GetSize()
+u64 IOFile::GetSize() const
 {
     if (IsOpen())
         return FileUtil::GetSize(m_file);
-    else
-        return 0;
+
+    return 0;
 }
 
 bool IOFile::Seek(s64 off, int origin)
@@ -948,12 +944,12 @@ bool IOFile::Seek(s64 off, int origin)
     return m_good;
 }
 
-u64 IOFile::Tell()
+u64 IOFile::Tell() const
 {
     if (IsOpen())
         return ftello(m_file);
-    else
-        return -1;
+
+    return -1;
 }
 
 bool IOFile::Flush()
diff --git a/src/common/file_util.h b/src/common/file_util.h
index a85121aa6..b54a9fb72 100644
--- a/src/common/file_util.h
+++ b/src/common/file_util.h
@@ -14,6 +14,10 @@
 
 #include "common/common_types.h"
 
+#ifdef _MSC_VER
+#include "common/string_util.h"
+#endif
+
 // User directory indices for GetUserPath
 enum {
     D_USER_IDX,
@@ -172,7 +176,6 @@ class IOFile : public NonCopyable
 {
 public:
     IOFile();
-    IOFile(std::FILE* file);
     IOFile(const std::string& filename, const char openmode[]);
 
     ~IOFile();
@@ -188,6 +191,9 @@ public:
     template <typename T>
     size_t ReadArray(T* data, size_t length)
     {
+        static_assert(std::is_standard_layout<T>(), "Given array does not consist of standard layout objects");
+        static_assert(std::is_trivially_copyable<T>(), "Given array does not consist of trivially copyable objects");
+
         if (!IsOpen()) {
             m_good = false;
             return -1;
@@ -203,9 +209,8 @@ public:
     template <typename T>
     size_t WriteArray(const T* data, size_t length)
     {
-        static_assert(std::is_standard_layout<T>::value, "Given array does not consist of standard layout objects");
-        // TODO: gcc 4.8 does not support is_trivially_copyable, but we really should check for it here.
-        //static_assert(std::is_trivially_copyable<T>::value, "Given array does not consist of trivially copyable objects");
+        static_assert(std::is_standard_layout<T>(), "Given array does not consist of standard layout objects");
+        static_assert(std::is_trivially_copyable<T>(), "Given array does not consist of trivially copyable objects");
 
         if (!IsOpen()) {
             m_good = false;
@@ -235,32 +240,24 @@ public:
         return WriteArray(&object, 1);
     }
 
-    bool IsOpen() { return nullptr != m_file; }
+    bool IsOpen() const { return nullptr != m_file; }
 
     // m_good is set to false when a read, write or other function fails
-    bool IsGood() {    return m_good; }
-    operator void*() { return m_good ? m_file : nullptr; }
-
-    std::FILE* ReleaseHandle();
-
-    std::FILE* GetHandle() { return m_file; }
-
-    void SetHandle(std::FILE* file);
+    bool IsGood() const { return m_good; }
+    explicit operator bool() const { return IsGood(); }
 
     bool Seek(s64 off, int origin);
-    u64 Tell();
-    u64 GetSize();
+    u64 Tell() const;
+    u64 GetSize() const;
     bool Resize(u64 size);
     bool Flush();
 
     // clear error state
     void Clear() { m_good = true; std::clearerr(m_file); }
 
-    std::FILE* m_file;
-    bool m_good;
 private:
-    IOFile(IOFile&);
-    IOFile& operator=(IOFile& other);
+    std::FILE* m_file = nullptr;
+    bool m_good = true;
 };
 
 }  // namespace
diff --git a/src/common/logging/backend.cpp b/src/common/logging/backend.cpp
index 4c86151ab..3d39f94d5 100644
--- a/src/common/logging/backend.cpp
+++ b/src/common/logging/backend.cpp
@@ -34,6 +34,7 @@ namespace Log {
         SUB(Kernel, SVC) \
         CLS(Service) \
         SUB(Service, SRV) \
+        SUB(Service, FRD) \
         SUB(Service, FS) \
         SUB(Service, ERR) \
         SUB(Service, APT) \
@@ -46,8 +47,10 @@ namespace Log {
         SUB(Service, NIM) \
         SUB(Service, NWM) \
         SUB(Service, CAM) \
+        SUB(Service, CECD) \
         SUB(Service, CFG) \
         SUB(Service, DSP) \
+        SUB(Service, DLP) \
         SUB(Service, HID) \
         SUB(Service, SOC) \
         SUB(Service, IR) \
diff --git a/src/common/logging/log.h b/src/common/logging/log.h
index e4c39c308..521362317 100644
--- a/src/common/logging/log.h
+++ b/src/common/logging/log.h
@@ -49,6 +49,7 @@ enum class Class : ClassType {
     Service,                    ///< HLE implementation of system services. Each major service
                                 ///  should have its own subclass.
     Service_SRV,                ///< The SRV (Service Directory) implementation
+    Service_FRD,                ///< The FRD (Friends) service
     Service_FS,                 ///< The FS (Filesystem) service implementation
     Service_ERR,                ///< The ERR (Error) port implementation
     Service_APT,                ///< The APT (Applets) service
@@ -61,8 +62,10 @@ enum class Class : ClassType {
     Service_NIM,                ///< The NIM (Network interface manager) service
     Service_NWM,                ///< The NWM (Network wlan manager) service
     Service_CAM,                ///< The CAM (Camera) service
+    Service_CECD,               ///< The CECD (StreetPass) service
     Service_CFG,                ///< The CFG (Configuration) service
     Service_DSP,                ///< The DSP (DSP control) service
+    Service_DLP,                ///< The DLP (Download Play) service
     Service_HID,                ///< The HID (Human interface device) service
     Service_SOC,                ///< The SOC (Socket) service
     Service_IR,                 ///< The IR service
diff --git a/src/common/make_unique.h b/src/common/make_unique.h
deleted file mode 100644
index f6e7f017c..000000000
--- a/src/common/make_unique.h
+++ /dev/null
@@ -1,17 +0,0 @@
-// Copyright 2014 Citra Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#pragma once
-
-#include <algorithm>
-#include <memory>
-
-namespace Common {
-
-template <typename T, typename... Args>
-std::unique_ptr<T> make_unique(Args&&... args) {
-    return std::unique_ptr<T>(new T(std::forward<Args>(args)...));
-}
-
-} // namespace
diff --git a/src/common/string_util.cpp b/src/common/string_util.cpp
index 6d6fc591f..f0aa072db 100644
--- a/src/common/string_util.cpp
+++ b/src/common/string_util.cpp
@@ -320,27 +320,27 @@ std::u16string UTF8ToUTF16(const std::string& input)
 #endif
 }
 
-static std::string UTF16ToUTF8(const std::wstring& input)
+static std::wstring CPToUTF16(u32 code_page, const std::string& input)
 {
-    auto const size = WideCharToMultiByte(CP_UTF8, 0, input.data(), static_cast<int>(input.size()), nullptr, 0, nullptr, nullptr);
+    auto const size = MultiByteToWideChar(code_page, 0, input.data(), static_cast<int>(input.size()), nullptr, 0);
 
-    std::string output;
+    std::wstring output;
     output.resize(size);
 
-    if (size == 0 || size != WideCharToMultiByte(CP_UTF8, 0, input.data(), static_cast<int>(input.size()), &output[0], static_cast<int>(output.size()), nullptr, nullptr))
+    if (size == 0 || size != MultiByteToWideChar(code_page, 0, input.data(), static_cast<int>(input.size()), &output[0], static_cast<int>(output.size())))
         output.clear();
 
     return output;
 }
 
-static std::wstring CPToUTF16(u32 code_page, const std::string& input)
+std::string UTF16ToUTF8(const std::wstring& input)
 {
-    auto const size = MultiByteToWideChar(code_page, 0, input.data(), static_cast<int>(input.size()), nullptr, 0);
+    auto const size = WideCharToMultiByte(CP_UTF8, 0, input.data(), static_cast<int>(input.size()), nullptr, 0, nullptr, nullptr);
 
-    std::wstring output;
+    std::string output;
     output.resize(size);
 
-    if (size == 0 || size != MultiByteToWideChar(code_page, 0, input.data(), static_cast<int>(input.size()), &output[0], static_cast<int>(output.size())))
+    if (size == 0 || size != WideCharToMultiByte(CP_UTF8, 0, input.data(), static_cast<int>(input.size()), &output[0], static_cast<int>(output.size()), nullptr, nullptr))
         output.clear();
 
     return output;
diff --git a/src/common/string_util.h b/src/common/string_util.h
index c5c474c6f..89d9f133e 100644
--- a/src/common/string_util.h
+++ b/src/common/string_util.h
@@ -95,7 +95,7 @@ std::string CP1252ToUTF8(const std::string& str);
 std::string SHIFTJISToUTF8(const std::string& str);
 
 #ifdef _WIN32
-
+std::string UTF16ToUTF8(const std::wstring& input);
 std::wstring UTF8ToUTF16W(const std::string& str);
 
 #ifdef _UNICODE
diff --git a/src/common/thread.h b/src/common/thread.h
index 8255ee6d3..bbfa8befa 100644
--- a/src/common/thread.h
+++ b/src/common/thread.h
@@ -30,8 +30,7 @@
 #   endif
 #endif
 
-namespace Common
-{
+namespace Common {
 
 int CurrentThreadId();
 
@@ -43,55 +42,55 @@ public:
     Event() : is_set(false) {}
 
     void Set() {
-        std::lock_guard<std::mutex> lk(m_mutex);
+        std::lock_guard<std::mutex> lk(mutex);
         if (!is_set) {
             is_set = true;
-            m_condvar.notify_one();
+            condvar.notify_one();
         }
     }
 
     void Wait() {
-        std::unique_lock<std::mutex> lk(m_mutex);
-        m_condvar.wait(lk, [&]{ return is_set; });
+        std::unique_lock<std::mutex> lk(mutex);
+        condvar.wait(lk, [&]{ return is_set; });
         is_set = false;
     }
 
     void Reset() {
-        std::unique_lock<std::mutex> lk(m_mutex);
+        std::unique_lock<std::mutex> lk(mutex);
         // no other action required, since wait loops on the predicate and any lingering signal will get cleared on the first iteration
         is_set = false;
     }
 
 private:
     bool is_set;
-    std::condition_variable m_condvar;
-    std::mutex m_mutex;
+    std::condition_variable condvar;
+    std::mutex mutex;
 };
 
 class Barrier {
 public:
-    Barrier(size_t count) : m_count(count), m_waiting(0) {}
+    explicit Barrier(size_t count_) : count(count_), waiting(0), generation(0) {}
 
     /// Blocks until all "count" threads have called Sync()
     void Sync() {
-        std::unique_lock<std::mutex> lk(m_mutex);
+        std::unique_lock<std::mutex> lk(mutex);
+        const size_t current_generation = generation;
 
-        // TODO: broken when next round of Sync()s
-        // is entered before all waiting threads return from the notify_all
-
-        if (++m_waiting == m_count) {
-            m_waiting = 0;
-            m_condvar.notify_all();
+        if (++waiting == count) {
+            generation++;
+            waiting = 0;
+            condvar.notify_all();
         } else {
-            m_condvar.wait(lk, [&]{ return m_waiting == 0; });
+            condvar.wait(lk, [this, current_generation]{ return current_generation != generation; });
         }
     }
 
 private:
-    std::condition_variable m_condvar;
-    std::mutex m_mutex;
-    const size_t m_count;
-    size_t m_waiting;
+    std::condition_variable condvar;
+    std::mutex mutex;
+    const size_t count;
+    size_t waiting;
+    size_t generation; // Incremented once each time the barrier is used
 };
 
 void SleepCurrentThread(int ms);
@@ -100,8 +99,7 @@ void SwitchCurrentThread();    // On Linux, this is equal to sleep 1ms
 // Use this function during a spin-wait to make the current thread
 // relax while another thread is working. This may be more efficient
 // than using events because event functions use kernel calls.
-inline void YieldCPU()
-{
+inline void YieldCPU() {
     std::this_thread::yield();
 }
 
diff --git a/src/common/x64/emitter.cpp b/src/common/x64/emitter.cpp
index 1dcf2416c..5662f7f86 100644
--- a/src/common/x64/emitter.cpp
+++ b/src/common/x64/emitter.cpp
@@ -455,6 +455,18 @@ void XEmitter::CALL(const void* fnptr)
     Write32(u32(distance));
 }
 
+FixupBranch XEmitter::CALL()
+{
+    FixupBranch branch;
+    branch.type = 1;
+    branch.ptr = code + 5;
+
+    Write8(0xE8);
+    Write32(0);
+
+    return branch;
+}
+
 FixupBranch XEmitter::J(bool force5bytes)
 {
     FixupBranch branch;
@@ -531,6 +543,22 @@ void XEmitter::SetJumpTarget(const FixupBranch& branch)
     }
 }
 
+void XEmitter::SetJumpTarget(const FixupBranch& branch, const u8* target)
+{
+    if (branch.type == 0)
+    {
+        s64 distance = (s64)(target - branch.ptr);
+        ASSERT_MSG(distance >= -0x80 && distance < 0x80, "Jump target too far away, needs force5Bytes = true");
+        branch.ptr[-1] = (u8)(s8)distance;
+    }
+    else if (branch.type == 1)
+    {
+        s64 distance = (s64)(target - branch.ptr);
+        ASSERT_MSG(distance >= -0x80000000LL && distance < 0x80000000LL, "Jump target too far away, needs indirect register");
+        ((s32*)branch.ptr)[-1] = (s32)distance;
+    }
+}
+
 //Single byte opcodes
 //There is no PUSHAD/POPAD in 64-bit mode.
 void XEmitter::INT3() {Write8(0xCC);}
diff --git a/src/common/x64/emitter.h b/src/common/x64/emitter.h
index 7c6548fb5..a33724146 100644
--- a/src/common/x64/emitter.h
+++ b/src/common/x64/emitter.h
@@ -425,12 +425,14 @@ public:
 #undef CALL
 #endif
     void CALL(const void* fnptr);
+    FixupBranch CALL();
     void CALLptr(OpArg arg);
 
     FixupBranch J_CC(CCFlags conditionCode, bool force5bytes = false);
     void J_CC(CCFlags conditionCode, const u8* addr, bool force5Bytes = false);
 
     void SetJumpTarget(const FixupBranch& branch);
+    void SetJumpTarget(const FixupBranch& branch, const u8* target);
 
     void SETcc(CCFlags flag, OpArg dest);
     // Note: CMOV brings small if any benefit on current cpus.
diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
index 3473e2f5b..a8d891689 100644
--- a/src/core/CMakeLists.txt
+++ b/src/core/CMakeLists.txt
@@ -68,6 +68,7 @@ set(SRCS
             hle/service/cfg/cfg_s.cpp
             hle/service/cfg/cfg_u.cpp
             hle/service/csnd_snd.cpp
+            hle/service/dlp_srvr.cpp
             hle/service/dsp_dsp.cpp
             hle/service/err_f.cpp
             hle/service/frd/frd.cpp
@@ -200,6 +201,7 @@ set(HEADERS
             hle/service/cfg/cfg_s.h
             hle/service/cfg/cfg_u.h
             hle/service/csnd_snd.h
+            hle/service/dlp_srvr.h
             hle/service/dsp_dsp.h
             hle/service/err_f.h
             hle/service/frd/frd.h
diff --git a/src/core/arm/dyncom/arm_dyncom.cpp b/src/core/arm/dyncom/arm_dyncom.cpp
index f3be2c857..a3581132c 100644
--- a/src/core/arm/dyncom/arm_dyncom.cpp
+++ b/src/core/arm/dyncom/arm_dyncom.cpp
@@ -3,8 +3,7 @@
 // Refer to the license.txt file included.
 
 #include <cstring>
-
-#include "common/make_unique.h"
+#include <memory>
 
 #include "core/arm/skyeye_common/armstate.h"
 #include "core/arm/skyeye_common/armsupp.h"
@@ -18,7 +17,7 @@
 #include "core/core_timing.h"
 
 ARM_DynCom::ARM_DynCom(PrivilegeMode initial_mode) {
-    state = Common::make_unique<ARMul_State>(initial_mode);
+    state = std::make_unique<ARMul_State>(initial_mode);
 }
 
 ARM_DynCom::~ARM_DynCom() {
@@ -94,7 +93,7 @@ void ARM_DynCom::ResetContext(Core::ThreadContext& context, u32 stack_top, u32 e
     context.cpu_registers[0] = arg;
     context.pc = entry_point;
     context.sp = stack_top;
-    context.cpsr = 0x1F; // Usermode
+    context.cpsr = 0x1F | ((entry_point & 1) << 5); // Usermode and THUMB mode
 }
 
 void ARM_DynCom::SaveContext(Core::ThreadContext& ctx) {
diff --git a/src/core/arm/dyncom/arm_dyncom_interpreter.cpp b/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
index 5f8826034..647784208 100644
--- a/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
+++ b/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
@@ -36,7 +36,8 @@ enum {
     CALL            = (1 << 4),
     RET             = (1 << 5),
     END_OF_PAGE     = (1 << 6),
-    THUMB           = (1 << 7)
+    THUMB           = (1 << 7),
+    SINGLE_STEP     = (1 << 8)
 };
 
 #define RM    BITS(sht_oper, 0, 3)
@@ -3466,7 +3467,35 @@ enum {
 
 MICROPROFILE_DEFINE(DynCom_Decode, "DynCom", "Decode", MP_RGB(255, 64, 64));
 
-static int InterpreterTranslate(ARMul_State* cpu, int& bb_start, u32 addr) {
+static unsigned int InterpreterTranslateInstruction(const ARMul_State* cpu, const u32 phys_addr, ARM_INST_PTR& inst_base) {
+    unsigned int inst_size = 4;
+    unsigned int inst = Memory::Read32(phys_addr & 0xFFFFFFFC);
+
+    // If we are in Thumb mode, we'll translate one Thumb instruction to the corresponding ARM instruction
+    if (cpu->TFlag) {
+        u32 arm_inst;
+        ThumbDecodeStatus state = DecodeThumbInstruction(inst, phys_addr, &arm_inst, &inst_size, &inst_base);
+
+        // We have translated the Thumb branch instruction in the Thumb decoder
+        if (state == ThumbDecodeStatus::BRANCH) {
+            return inst_size;
+        }
+        inst = arm_inst;
+    }
+
+    int idx;
+    if (DecodeARMInstruction(inst, &idx) == ARMDecodeStatus::FAILURE) {
+        std::string disasm = ARM_Disasm::Disassemble(phys_addr, inst);
+        LOG_ERROR(Core_ARM11, "Decode failure.\tPC : [0x%x]\tInstruction : %s [%x]", phys_addr, disasm.c_str(), inst);
+        LOG_ERROR(Core_ARM11, "cpsr=0x%x, cpu->TFlag=%d, r15=0x%x", cpu->Cpsr, cpu->TFlag, cpu->Reg[15]);
+        CITRA_IGNORE_EXIT(-1);
+    }
+    inst_base = arm_instruction_trans[idx](inst, idx);
+
+    return inst_size;
+}
+
+static int InterpreterTranslateBlock(ARMul_State* cpu, int& bb_start, u32 addr) {
     Common::Profiling::ScopeTimer timer_decode(profile_decode);
     MICROPROFILE_SCOPE(DynCom_Decode);
 
@@ -3475,8 +3504,6 @@ static int InterpreterTranslate(ARMul_State* cpu, int& bb_start, u32 addr) {
     // Go on next, until terminal instruction
     // Save start addr of basicblock in CreamCache
     ARM_INST_PTR inst_base = nullptr;
-    unsigned int inst, inst_size = 4;
-    int idx;
     int ret = NON_BRANCH;
     int size = 0; // instruction size of basic block
     bb_start = top;
@@ -3485,30 +3512,10 @@ static int InterpreterTranslate(ARMul_State* cpu, int& bb_start, u32 addr) {
     u32 pc_start = cpu->Reg[15];
 
     while (ret == NON_BRANCH) {
-        inst = Memory::Read32(phys_addr & 0xFFFFFFFC);
+        unsigned int inst_size = InterpreterTranslateInstruction(cpu, phys_addr, inst_base);
 
         size++;
-        // If we are in Thumb mode, we'll translate one Thumb instruction to the corresponding ARM instruction
-        if (cpu->TFlag) {
-            u32 arm_inst;
-            ThumbDecodeStatus state = DecodeThumbInstruction(inst, phys_addr, &arm_inst, &inst_size, &inst_base);
-
-            // We have translated the Thumb branch instruction in the Thumb decoder
-            if (state == ThumbDecodeStatus::BRANCH) {
-                goto translated;
-            }
-            inst = arm_inst;
-        }
-
-        if (DecodeARMInstruction(inst, &idx) == ARMDecodeStatus::FAILURE) {
-            std::string disasm = ARM_Disasm::Disassemble(phys_addr, inst);
-            LOG_ERROR(Core_ARM11, "Decode failure.\tPC : [0x%x]\tInstruction : %s [%x]", phys_addr, disasm.c_str(), inst);
-            LOG_ERROR(Core_ARM11, "cpsr=0x%x, cpu->TFlag=%d, r15=0x%x", cpu->Cpsr, cpu->TFlag, cpu->Reg[15]);
-            CITRA_IGNORE_EXIT(-1);
-        }
-        inst_base = arm_instruction_trans[idx](inst, idx);
 
-translated:
         phys_addr += inst_size;
 
         if ((phys_addr & 0xfff) == 0) {
@@ -3522,6 +3529,27 @@ translated:
     return KEEP_GOING;
 }
 
+static int InterpreterTranslateSingle(ARMul_State* cpu, int& bb_start, u32 addr) {
+    Common::Profiling::ScopeTimer timer_decode(profile_decode);
+    MICROPROFILE_SCOPE(DynCom_Decode);
+
+    ARM_INST_PTR inst_base = nullptr;
+    bb_start = top;
+
+    u32 phys_addr = addr;
+    u32 pc_start = cpu->Reg[15];
+
+    InterpreterTranslateInstruction(cpu, phys_addr, inst_base);
+
+    if (inst_base->br == NON_BRANCH) {
+        inst_base->br = SINGLE_STEP;
+    }
+
+    cpu->instruction_cache[pc_start] = bb_start;
+
+    return KEEP_GOING;
+}
+
 static int clz(unsigned int x) {
     int n;
     if (x == 0) return (32);
@@ -3871,8 +3899,11 @@ unsigned InterpreterMainLoop(ARMul_State* cpu) {
         auto itr = cpu->instruction_cache.find(cpu->Reg[15]);
         if (itr != cpu->instruction_cache.end()) {
             ptr = itr->second;
+        } else if (cpu->NumInstrsToExecute != 1) {
+            if (InterpreterTranslateBlock(cpu, ptr, cpu->Reg[15]) == FETCH_EXCEPTION)
+                goto END;
         } else {
-            if (InterpreterTranslate(cpu, ptr, cpu->Reg[15]) == FETCH_EXCEPTION)
+            if (InterpreterTranslateSingle(cpu, ptr, cpu->Reg[15]) == FETCH_EXCEPTION)
                 goto END;
         }
 
@@ -3924,9 +3955,7 @@ unsigned InterpreterMainLoop(ARMul_State* cpu) {
         if (inst_base->cond == ConditionCode::AL || CondPassed(cpu, inst_base->cond)) {
             add_inst* const inst_cream = (add_inst*)inst_base->component;
 
-            u32 rn_val = RN;
-            if (inst_cream->Rn == 15)
-                rn_val += 2 * cpu->GetInstructionSize();
+            u32 rn_val = CHECK_READ_REG15_WA(cpu, inst_cream->Rn);
 
             bool carry;
             bool overflow;
@@ -4051,11 +4080,12 @@ unsigned InterpreterMainLoop(ARMul_State* cpu) {
         if ((inst_base->cond == ConditionCode::AL) || CondPassed(cpu, inst_base->cond)) {
             unsigned int inst = inst_cream->inst;
             if (BITS(inst, 20, 27) == 0x12 && BITS(inst, 4, 7) == 0x3) {
+                const u32 jump_address = cpu->Reg[inst_cream->val.Rm];
                 cpu->Reg[14] = (cpu->Reg[15] + cpu->GetInstructionSize());
                 if(cpu->TFlag)
                     cpu->Reg[14] |= 0x1;
-                cpu->Reg[15] = cpu->Reg[inst_cream->val.Rm] & 0xfffffffe;
-                cpu->TFlag = cpu->Reg[inst_cream->val.Rm] & 0x1;
+                cpu->Reg[15] = jump_address & 0xfffffffe;
+                cpu->TFlag = jump_address & 0x1;
             } else {
                 cpu->Reg[14] = (cpu->Reg[15] + cpu->GetInstructionSize());
                 cpu->TFlag = 0x1;
@@ -6136,9 +6166,7 @@ unsigned InterpreterMainLoop(ARMul_State* cpu) {
         if (inst_base->cond == ConditionCode::AL || CondPassed(cpu, inst_base->cond)) {
             sub_inst* const inst_cream = (sub_inst*)inst_base->component;
 
-            u32 rn_val = RN;
-            if (inst_cream->Rn == 15)
-                rn_val += 2 * cpu->GetInstructionSize();
+            u32 rn_val = CHECK_READ_REG15_WA(cpu, inst_cream->Rn);
 
             bool carry;
             bool overflow;
diff --git a/src/core/core.cpp b/src/core/core.cpp
index 84d6c392e..3bb843aab 100644
--- a/src/core/core.cpp
+++ b/src/core/core.cpp
@@ -4,7 +4,6 @@
 
 #include <memory>
 
-#include "common/make_unique.h"
 #include "common/logging/log.h"
 
 #include "core/core.h"
@@ -74,8 +73,8 @@ void Stop() {
 
 /// Initialize the core
 void Init() {
-    g_sys_core = Common::make_unique<ARM_DynCom>(USER32MODE);
-    g_app_core = Common::make_unique<ARM_DynCom>(USER32MODE);
+    g_sys_core = std::make_unique<ARM_DynCom>(USER32MODE);
+    g_app_core = std::make_unique<ARM_DynCom>(USER32MODE);
 
     LOG_DEBUG(Core, "Initialized OK");
 }
diff --git a/src/core/file_sys/archive_extsavedata.cpp b/src/core/file_sys/archive_extsavedata.cpp
index 961264fe5..1d9eaefcb 100644
--- a/src/core/file_sys/archive_extsavedata.cpp
+++ b/src/core/file_sys/archive_extsavedata.cpp
@@ -3,12 +3,12 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <memory>
 #include <vector>
 
 #include "common/common_types.h"
 #include "common/file_util.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 #include "common/string_util.h"
 
 #include "core/file_sys/archive_extsavedata.h"
@@ -84,7 +84,7 @@ ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_ExtSaveData::Open(cons
                               ErrorSummary::InvalidState, ErrorLevel::Status);
         }
     }
-    auto archive = Common::make_unique<DiskArchive>(fullpath);
+    auto archive = std::make_unique<DiskArchive>(fullpath);
     return MakeResult<std::unique_ptr<ArchiveBackend>>(std::move(archive));
 }
 
diff --git a/src/core/file_sys/archive_romfs.cpp b/src/core/file_sys/archive_romfs.cpp
index a9a29ebde..38828b546 100644
--- a/src/core/file_sys/archive_romfs.cpp
+++ b/src/core/file_sys/archive_romfs.cpp
@@ -7,7 +7,6 @@
 
 #include "common/common_types.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 
 #include "core/file_sys/archive_romfs.h"
 #include "core/file_sys/ivfc_archive.h"
@@ -25,7 +24,7 @@ ArchiveFactory_RomFS::ArchiveFactory_RomFS(Loader::AppLoader& app_loader) {
 }
 
 ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_RomFS::Open(const Path& path) {
-    auto archive = Common::make_unique<IVFCArchive>(romfs_file, data_offset, data_size);
+    auto archive = std::make_unique<IVFCArchive>(romfs_file, data_offset, data_size);
     return MakeResult<std::unique_ptr<ArchiveBackend>>(std::move(archive));
 }
 
diff --git a/src/core/file_sys/archive_savedata.cpp b/src/core/file_sys/archive_savedata.cpp
index fe020d21c..fd5711e14 100644
--- a/src/core/file_sys/archive_savedata.cpp
+++ b/src/core/file_sys/archive_savedata.cpp
@@ -3,11 +3,11 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <memory>
 
 #include "common/common_types.h"
 #include "common/file_util.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 #include "common/string_util.h"
 
 #include "core/file_sys/archive_savedata.h"
@@ -53,7 +53,7 @@ ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_SaveData::Open(const P
             ErrorSummary::InvalidState, ErrorLevel::Status);
     }
 
-    auto archive = Common::make_unique<DiskArchive>(std::move(concrete_mount_point));
+    auto archive = std::make_unique<DiskArchive>(std::move(concrete_mount_point));
     return MakeResult<std::unique_ptr<ArchiveBackend>>(std::move(archive));
 }
 
diff --git a/src/core/file_sys/archive_savedatacheck.cpp b/src/core/file_sys/archive_savedatacheck.cpp
index 3db11c500..9f65e5455 100644
--- a/src/core/file_sys/archive_savedatacheck.cpp
+++ b/src/core/file_sys/archive_savedatacheck.cpp
@@ -3,12 +3,12 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <memory>
 #include <vector>
 
 #include "common/common_types.h"
 #include "common/file_util.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 #include "common/string_util.h"
 
 #include "core/file_sys/archive_savedatacheck.h"
@@ -44,7 +44,7 @@ ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_SaveDataCheck::Open(co
     }
     auto size = file->GetSize();
 
-    auto archive = Common::make_unique<IVFCArchive>(file, 0, size);
+    auto archive = std::make_unique<IVFCArchive>(file, 0, size);
     return MakeResult<std::unique_ptr<ArchiveBackend>>(std::move(archive));
 }
 
diff --git a/src/core/file_sys/archive_sdmc.cpp b/src/core/file_sys/archive_sdmc.cpp
index 657221cbf..9b218af58 100644
--- a/src/core/file_sys/archive_sdmc.cpp
+++ b/src/core/file_sys/archive_sdmc.cpp
@@ -3,10 +3,10 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <memory>
 
 #include "common/file_util.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 
 #include "core/file_sys/archive_sdmc.h"
 #include "core/file_sys/disk_archive.h"
@@ -36,7 +36,7 @@ bool ArchiveFactory_SDMC::Initialize() {
 }
 
 ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_SDMC::Open(const Path& path) {
-    auto archive = Common::make_unique<DiskArchive>(sdmc_directory);
+    auto archive = std::make_unique<DiskArchive>(sdmc_directory);
     return MakeResult<std::unique_ptr<ArchiveBackend>>(std::move(archive));
 }
 
diff --git a/src/core/file_sys/archive_systemsavedata.cpp b/src/core/file_sys/archive_systemsavedata.cpp
index e1780de2f..1bcc228a1 100644
--- a/src/core/file_sys/archive_systemsavedata.cpp
+++ b/src/core/file_sys/archive_systemsavedata.cpp
@@ -3,11 +3,11 @@
 // Refer to the license.txt file included.
 
 #include <algorithm>
+#include <memory>
 #include <vector>
 
 #include "common/common_types.h"
 #include "common/file_util.h"
-#include "common/make_unique.h"
 #include "common/string_util.h"
 
 #include "core/file_sys/archive_systemsavedata.h"
@@ -59,7 +59,7 @@ ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_SystemSaveData::Open(c
         return ResultCode(ErrorDescription::FS_NotFormatted, ErrorModule::FS,
             ErrorSummary::InvalidState, ErrorLevel::Status);
     }
-    auto archive = Common::make_unique<DiskArchive>(fullpath);
+    auto archive = std::make_unique<DiskArchive>(fullpath);
     return MakeResult<std::unique_ptr<ArchiveBackend>>(std::move(archive));
 }
 
diff --git a/src/core/file_sys/disk_archive.cpp b/src/core/file_sys/disk_archive.cpp
index 8e4ea01c5..489cc96fb 100644
--- a/src/core/file_sys/disk_archive.cpp
+++ b/src/core/file_sys/disk_archive.cpp
@@ -4,11 +4,11 @@
 
 #include <algorithm>
 #include <cstdio>
+#include <memory>
 
 #include "common/common_types.h"
 #include "common/file_util.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 
 #include "core/file_sys/disk_archive.h"
 
@@ -19,7 +19,7 @@ namespace FileSys {
 
 ResultVal<std::unique_ptr<FileBackend>> DiskArchive::OpenFile(const Path& path, const Mode mode) const {
     LOG_DEBUG(Service_FS, "called path=%s mode=%01X", path.DebugStr().c_str(), mode.hex);
-    auto file = Common::make_unique<DiskFile>(*this, path, mode);
+    auto file = std::make_unique<DiskFile>(*this, path, mode);
     ResultCode result = file->Open();
     if (result.IsError())
         return result;
@@ -83,7 +83,7 @@ bool DiskArchive::RenameDirectory(const Path& src_path, const Path& dest_path) c
 
 std::unique_ptr<DirectoryBackend> DiskArchive::OpenDirectory(const Path& path) const {
     LOG_DEBUG(Service_FS, "called path=%s", path.DebugStr().c_str());
-    auto directory = Common::make_unique<DiskDirectory>(*this, path);
+    auto directory = std::make_unique<DiskDirectory>(*this, path);
     if (!directory->Open())
         return nullptr;
     return std::move(directory);
@@ -132,7 +132,7 @@ ResultCode DiskFile::Open() {
     // Open the file in binary mode, to avoid problems with CR/LF on Windows systems
     mode_string += "b";
 
-    file = Common::make_unique<FileUtil::IOFile>(path, mode_string.c_str());
+    file = std::make_unique<FileUtil::IOFile>(path, mode_string.c_str());
     if (file->IsOpen())
         return RESULT_SUCCESS;
     return ResultCode(ErrorDescription::FS_NotFound, ErrorModule::FS, ErrorSummary::NotFound, ErrorLevel::Status);
diff --git a/src/core/file_sys/ivfc_archive.cpp b/src/core/file_sys/ivfc_archive.cpp
index a8e9a72ef..c61791ef7 100644
--- a/src/core/file_sys/ivfc_archive.cpp
+++ b/src/core/file_sys/ivfc_archive.cpp
@@ -7,7 +7,6 @@
 
 #include "common/common_types.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 
 #include "core/file_sys/ivfc_archive.h"
 
@@ -21,7 +20,7 @@ std::string IVFCArchive::GetName() const {
 }
 
 ResultVal<std::unique_ptr<FileBackend>> IVFCArchive::OpenFile(const Path& path, const Mode mode) const {
-    return MakeResult<std::unique_ptr<FileBackend>>(Common::make_unique<IVFCFile>(romfs_file, data_offset, data_size));
+    return MakeResult<std::unique_ptr<FileBackend>>(std::make_unique<IVFCFile>(romfs_file, data_offset, data_size));
 }
 
 ResultCode IVFCArchive::DeleteFile(const Path& path) const {
@@ -58,7 +57,7 @@ bool IVFCArchive::RenameDirectory(const Path& src_path, const Path& dest_path) c
 }
 
 std::unique_ptr<DirectoryBackend> IVFCArchive::OpenDirectory(const Path& path) const {
-    return Common::make_unique<IVFCDirectory>();
+    return std::make_unique<IVFCDirectory>();
 }
 
 u64 IVFCArchive::GetFreeBytes() const {
diff --git a/src/core/gdbstub/gdbstub.cpp b/src/core/gdbstub/gdbstub.cpp
index 3a2445241..c1a7ec5bf 100644
--- a/src/core/gdbstub/gdbstub.cpp
+++ b/src/core/gdbstub/gdbstub.cpp
@@ -60,6 +60,59 @@ const u32 R15_REGISTER = 15;
 const u32 CPSR_REGISTER = 25;
 const u32 FPSCR_REGISTER = 58;
 
+// For sample XML files see the GDB source /gdb/features
+// GDB also wants the l character at the start
+// This XML defines what the registers are for this specific ARM device
+static const char* target_xml =
+R"(l<?xml version="1.0"?>
+<!DOCTYPE target SYSTEM "gdb-target.dtd">
+<target version="1.0">
+  <feature name="org.gnu.gdb.arm.core">
+    <reg name="r0" bitsize="32"/>
+    <reg name="r1" bitsize="32"/>
+    <reg name="r2" bitsize="32"/>
+    <reg name="r3" bitsize="32"/>
+    <reg name="r4" bitsize="32"/>
+    <reg name="r5" bitsize="32"/>
+    <reg name="r6" bitsize="32"/>
+    <reg name="r7" bitsize="32"/>
+    <reg name="r8" bitsize="32"/>
+    <reg name="r9" bitsize="32"/>
+    <reg name="r10" bitsize="32"/>
+    <reg name="r11" bitsize="32"/>
+    <reg name="r12" bitsize="32"/>
+    <reg name="sp" bitsize="32" type="data_ptr"/>
+    <reg name="lr" bitsize="32"/>
+    <reg name="pc" bitsize="32" type="code_ptr"/>
+
+    <!-- The CPSR is register 25, rather than register 16, because
+         the FPA registers historically were placed between the PC
+         and the CPSR in the "g" packet.  -->
+
+    <reg name="cpsr" bitsize="32" regnum="25"/>
+  </feature>
+  <feature name="org.gnu.gdb.arm.vfp">
+    <reg name="d0" bitsize="64" type="float"/>
+    <reg name="d1" bitsize="64" type="float"/>
+    <reg name="d2" bitsize="64" type="float"/>
+    <reg name="d3" bitsize="64" type="float"/>
+    <reg name="d4" bitsize="64" type="float"/>
+    <reg name="d5" bitsize="64" type="float"/>
+    <reg name="d6" bitsize="64" type="float"/>
+    <reg name="d7" bitsize="64" type="float"/>
+    <reg name="d8" bitsize="64" type="float"/>
+    <reg name="d9" bitsize="64" type="float"/>
+    <reg name="d10" bitsize="64" type="float"/>
+    <reg name="d11" bitsize="64" type="float"/>
+    <reg name="d12" bitsize="64" type="float"/>
+    <reg name="d13" bitsize="64" type="float"/>
+    <reg name="d14" bitsize="64" type="float"/>
+    <reg name="d15" bitsize="64" type="float"/>
+    <reg name="fpscr" bitsize="32" type="int" group="float"/>
+  </feature>
+</target>
+)";
+
 namespace GDBStub {
 
 static int gdbserver_socket = -1;
@@ -211,7 +264,7 @@ static u8 ReadByte() {
 }
 
 /// Calculate the checksum of the current command buffer.
-static u8 CalculateChecksum(u8 *buffer, u32 length) {
+static u8 CalculateChecksum(u8* buffer, u32 length) {
     return static_cast<u8>(std::accumulate(buffer, buffer + length, 0, std::plus<u8>()));
 }
 
@@ -353,8 +406,15 @@ static void SendReply(const char* reply) {
 static void HandleQuery() {
     LOG_DEBUG(Debug_GDBStub, "gdb: query '%s'\n", command_buffer + 1);
 
-    if (!strcmp(reinterpret_cast<const char*>(command_buffer + 1), "TStatus")) {
+    const char* query = reinterpret_cast<const char*>(command_buffer + 1);
+
+    if (strcmp(query, "TStatus") == 0 ) {
         SendReply("T0");
+    } else if (strncmp(query, "Supported:", strlen("Supported:")) == 0) {
+        // PacketSize needs to be large enough for target xml
+        SendReply("PacketSize=800;qXfer:features:read+");
+    } else if (strncmp(query, "Xfer:features:read:target.xml:", strlen("Xfer:features:read:target.xml:")) == 0) {
+        SendReply(target_xml);
     } else {
         SendReply("");
     }
@@ -491,29 +551,25 @@ static void ReadRegisters() {
     memset(buffer, 0, sizeof(buffer));
 
     u8* bufptr = buffer;
-    for (int i = 0, reg = 0; reg <= FPSCR_REGISTER; i++, reg++) {
-        if (reg <= R15_REGISTER) {
-            IntToGdbHex(bufptr + i * CHAR_BIT, Core::g_app_core->GetReg(reg));
-        } else if (reg == CPSR_REGISTER) {
-            IntToGdbHex(bufptr + i * CHAR_BIT, Core::g_app_core->GetCPSR());
-        } else if (reg == CPSR_REGISTER - 1) {
-            // Dummy FPA register, ignore
-            IntToGdbHex(bufptr + i * CHAR_BIT, 0);
-        } else if (reg < CPSR_REGISTER) {
-            // Dummy FPA registers, ignore
-            IntToGdbHex(bufptr + i * CHAR_BIT, 0);
-            IntToGdbHex(bufptr + (i + 1) * CHAR_BIT, 0);
-            IntToGdbHex(bufptr + (i + 2) * CHAR_BIT, 0);
-            i += 2;
-        } else if (reg > CPSR_REGISTER && reg < FPSCR_REGISTER) {
-            IntToGdbHex(bufptr + i * CHAR_BIT, Core::g_app_core->GetVFPReg(reg - CPSR_REGISTER - 1));
-            IntToGdbHex(bufptr + (i + 1) * CHAR_BIT, 0);
-            i++;
-        } else if (reg == FPSCR_REGISTER) {
-            IntToGdbHex(bufptr + i * CHAR_BIT, Core::g_app_core->GetVFPSystemReg(VFP_FPSCR));
-        }
+
+    for (int reg = 0; reg <= R15_REGISTER; reg++) {
+        IntToGdbHex(bufptr + reg * CHAR_BIT, Core::g_app_core->GetReg(reg));
     }
 
+    bufptr += (16 * CHAR_BIT);
+
+    IntToGdbHex(bufptr, Core::g_app_core->GetCPSR());
+
+    bufptr += CHAR_BIT;
+
+    for (int reg = 0; reg <= 31; reg++) {
+        IntToGdbHex(bufptr + reg * CHAR_BIT, Core::g_app_core->GetVFPReg(reg));
+    }
+
+    bufptr += (32 * CHAR_BIT);
+
+    IntToGdbHex(bufptr, Core::g_app_core->GetVFPSystemReg(VFP_FPSCR));
+
     SendReply(reinterpret_cast<char*>(buffer));
 }
 
@@ -885,6 +941,12 @@ void Init(u16 port) {
         LOG_ERROR(Debug_GDBStub, "Failed to create gdb socket");
     }
 
+    // Set socket to SO_REUSEADDR so it can always bind on the same port
+    int reuse_enabled = 1;
+    if (setsockopt(tmpsock, SOL_SOCKET, SO_REUSEADDR, (const char*)&reuse_enabled, sizeof(reuse_enabled)) < 0) {
+        LOG_ERROR(Debug_GDBStub, "Failed to set gdb socket option");
+    }
+
     const sockaddr* server_addr = reinterpret_cast<const sockaddr*>(&saddr_server);
     socklen_t server_addrlen = sizeof(saddr_server);
     if (bind(tmpsock, server_addr, server_addrlen) < 0) {
diff --git a/src/core/hle/config_mem.cpp b/src/core/hle/config_mem.cpp
index b1a72dc0c..ccd73cfcb 100644
--- a/src/core/hle/config_mem.cpp
+++ b/src/core/hle/config_mem.cpp
@@ -3,13 +3,6 @@
 // Refer to the license.txt file included.
 
 #include <cstring>
-
-#include "common/assert.h"
-#include "common/common_types.h"
-#include "common/common_funcs.h"
-
-#include "core/core.h"
-#include "core/memory.h"
 #include "core/hle/config_mem.h"
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
diff --git a/src/core/hle/hle.cpp b/src/core/hle/hle.cpp
index 331b1b22a..e545de3b5 100644
--- a/src/core/hle/hle.cpp
+++ b/src/core/hle/hle.cpp
@@ -8,8 +8,6 @@
 #include "core/arm/arm_interface.h"
 #include "core/core.h"
 #include "core/hle/hle.h"
-#include "core/hle/config_mem.h"
-#include "core/hle/shared_page.h"
 #include "core/hle/service/service.h"
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
diff --git a/src/core/hle/kernel/process.cpp b/src/core/hle/kernel/process.cpp
index 24b266eae..0546f6e16 100644
--- a/src/core/hle/kernel/process.cpp
+++ b/src/core/hle/kernel/process.cpp
@@ -2,10 +2,11 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <memory>
+
 #include "common/assert.h"
 #include "common/common_funcs.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 
 #include "core/hle/kernel/memory.h"
 #include "core/hle/kernel/process.h"
diff --git a/src/core/hle/result.h b/src/core/hle/result.h
index 0cb76ba1c..2d22652d9 100644
--- a/src/core/hle/result.h
+++ b/src/core/hle/result.h
@@ -24,6 +24,7 @@ enum class ErrorDescription : u32 {
     FS_InvalidOpenFlags = 230,
     FS_NotAFile = 250,
     FS_NotFormatted = 340, ///< This is used by the FS service when creating a SaveData archive
+    OutofRangeOrMisalignedAddress = 513, // TODO(purpasmart): Check if this name fits its actual usage
     FS_InvalidPath = 702,
     InvalidSection = 1000,
     TooLarge = 1001,
diff --git a/src/core/hle/service/am/am.cpp b/src/core/hle/service/am/am.cpp
index 06be9940e..9591522e5 100644
--- a/src/core/hle/service/am/am.cpp
+++ b/src/core/hle/service/am/am.cpp
@@ -2,6 +2,8 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <cinttypes>
+
 #include "common/logging/log.h"
 
 #include "core/hle/service/service.h"
@@ -9,30 +11,119 @@
 #include "core/hle/service/am/am_app.h"
 #include "core/hle/service/am/am_net.h"
 #include "core/hle/service/am/am_sys.h"
+#include "core/hle/service/am/am_u.h"
 
 namespace Service {
 namespace AM {
 
-void TitleIDListGetTotal(Service::Interface* self) {
+static std::array<u32, 3> am_content_count = { 0, 0, 0 };
+static std::array<u32, 3> am_titles_count = { 0, 0, 0 };
+static std::array<u32, 3> am_titles_list_count = { 0, 0, 0 };
+static u32 am_ticket_count = 0;
+static u32 am_ticket_list_count = 0;
+
+void GetTitleCount(Service::Interface* self) {
     u32* cmd_buff = Kernel::GetCommandBuffer();
+
     u32 media_type = cmd_buff[1] & 0xFF;
 
     cmd_buff[1] = RESULT_SUCCESS.raw;
-    cmd_buff[2] = 0;
+    cmd_buff[2] = am_titles_count[media_type];
+    LOG_WARNING(Service_AM, "(STUBBED) media_type=%u, title_count=0x%08x", media_type, am_titles_count[media_type]);
+}
+
+void FindContentInfos(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 media_type = cmd_buff[1] & 0xFF;
+    u64 title_id = (static_cast<u64>(cmd_buff[3]) << 32) | cmd_buff[2];
+    u32 content_ids_pointer = cmd_buff[6];
+    u32 content_info_pointer = cmd_buff[8];
+
+    am_content_count[media_type] = cmd_buff[4];
 
-    LOG_WARNING(Service_AM, "(STUBBED) media_type %u", media_type);
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    LOG_WARNING(Service_AM, "(STUBBED) media_type=%u, title_id=0x%016lx, content_cound=%u, content_ids_pointer=0x%08x, content_info_pointer=0x%08x",
+                media_type, title_id, am_content_count[media_type], content_ids_pointer, content_info_pointer);
 }
 
-void GetTitleIDList(Service::Interface* self) {
+void ListContentInfos(Service::Interface* self) {
     u32* cmd_buff = Kernel::GetCommandBuffer();
-    u32 num_titles = cmd_buff[1];
+
     u32 media_type = cmd_buff[2] & 0xFF;
-    u32 addr       = cmd_buff[4];
+    u64 title_id = (static_cast<u64>(cmd_buff[4]) << 32) | cmd_buff[3];
+    u32 start_index = cmd_buff[5];
+    u32 content_info_pointer = cmd_buff[7];
+
+    am_content_count[media_type] = cmd_buff[1];
 
     cmd_buff[1] = RESULT_SUCCESS.raw;
-    cmd_buff[2] = 0;
+    cmd_buff[2] = am_content_count[media_type];
+    LOG_WARNING(Service_AM, "(STUBBED) media_type=%u, content_count=%u, title_id=0x%016" PRIx64 ", start_index=0x%08x, content_info_pointer=0x%08X",
+                media_type, am_content_count[media_type], title_id, start_index, content_info_pointer);
+}
+
+void DeleteContents(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 media_type = cmd_buff[1] & 0xFF;
+    u64 title_id = (static_cast<u64>(cmd_buff[3]) << 32) | cmd_buff[2];
+    u32 content_ids_pointer = cmd_buff[6];
 
-    LOG_WARNING(Service_AM, "(STUBBED) Requested %u titles from media type %u. Address=0x%08X", num_titles, media_type, addr);
+    am_content_count[media_type] = cmd_buff[4];
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    LOG_WARNING(Service_AM, "(STUBBED) media_type=%u, title_id=0x%016" PRIx64 ", content_count=%u, content_ids_pointer=0x%08x",
+        media_type, title_id, am_content_count[media_type], content_ids_pointer);
+}
+
+void GetTitleList(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 media_type = cmd_buff[2] & 0xFF;
+    u32 title_ids_output_pointer = cmd_buff[4];
+
+    am_titles_list_count[media_type] = cmd_buff[1];
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    cmd_buff[2] = am_titles_list_count[media_type];
+    LOG_WARNING(Service_AM, "(STUBBED) media_type=%u, titles_list_count=0x%08X, title_ids_output_pointer=0x%08X",
+                media_type, am_titles_list_count[media_type], title_ids_output_pointer);
+}
+
+void GetTitleInfo(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 media_type = cmd_buff[1] & 0xFF;
+    u32 title_id_list_pointer = cmd_buff[4];
+    u32 title_list_pointer = cmd_buff[6];
+
+    am_titles_count[media_type] = cmd_buff[2];
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    LOG_WARNING(Service_AM, "(STUBBED) media_type=%u, total_titles=0x%08X, title_id_list_pointer=0x%08X, title_list_pointer=0x%08X",
+                media_type, am_titles_count[media_type], title_id_list_pointer, title_list_pointer);
+}
+
+void GetDataTitleInfos(Service::Interface* self) {
+    GetTitleInfo(self);
+
+    LOG_WARNING(Service_AM, "(STUBBED) called");
+}
+
+void ListDataTitleTicketInfos(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u64 title_id = (static_cast<u64>(cmd_buff[3]) << 32) | cmd_buff[2];
+    u32 start_index = cmd_buff[4];
+    u32 ticket_info_pointer = cmd_buff[6];
+
+    am_ticket_count = cmd_buff[1];
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    cmd_buff[2] = am_ticket_count;
+    LOG_WARNING(Service_AM, "(STUBBED) ticket_count=0x%08X, title_id=0x%016" PRIx64 ", start_index=0x%08X, ticket_info_pointer=0x%08X",
+                 am_ticket_count, title_id, start_index, ticket_info_pointer);
 }
 
 void GetNumContentInfos(Service::Interface* self) {
@@ -40,16 +131,47 @@ void GetNumContentInfos(Service::Interface* self) {
 
     cmd_buff[1] = RESULT_SUCCESS.raw;
     cmd_buff[2] = 1; // Number of content infos plus one
-
     LOG_WARNING(Service_AM, "(STUBBED) called");
 }
 
+void DeleteTicket(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u64 title_id = (static_cast<u64>(cmd_buff[2]) << 32) | cmd_buff[1];
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    LOG_WARNING(Service_AM, "(STUBBED) called title_id=0x%016" PRIx64 "",title_id);
+}
+
+void GetTicketCount(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    cmd_buff[2] = am_ticket_count;
+    LOG_WARNING(Service_AM, "(STUBBED) called ticket_count=0x%08x",am_ticket_count);
+}
+
+void GetTicketList(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 num_of_skip = cmd_buff[2];
+    u32 ticket_list_pointer = cmd_buff[4];
+
+    am_ticket_list_count = cmd_buff[1];
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    cmd_buff[2] = am_ticket_list_count;
+    LOG_WARNING(Service_AM, "(STUBBED) ticket_list_count=0x%08x, num_of_skip=0x%08x, ticket_list_pointer=0x%08x",
+                am_ticket_list_count, num_of_skip, ticket_list_pointer);
+}
+
 void Init() {
     using namespace Kernel;
 
     AddService(new AM_APP_Interface);
     AddService(new AM_NET_Interface);
     AddService(new AM_SYS_Interface);
+    AddService(new AM_U_Interface);
 }
 
 void Shutdown() {
diff --git a/src/core/hle/service/am/am.h b/src/core/hle/service/am/am.h
index 15e63bc7b..5676cdd5f 100644
--- a/src/core/hle/service/am/am.h
+++ b/src/core/hle/service/am/am.h
@@ -11,7 +11,7 @@ class Interface;
 namespace AM {
 
 /**
- * AM::TitleIDListGetTotal service function
+ * AM::GetTitleCount service function
  * Gets the number of installed titles in the requested media type
  *  Inputs:
  *      0 : Command header (0x00010040)
@@ -20,36 +20,140 @@ namespace AM {
  *      1 : Result, 0 on success, otherwise error code
  *      2 : The number of titles in the requested media type
  */
-void TitleIDListGetTotal(Service::Interface* self);
+void GetTitleCount(Service::Interface* self);
 
 /**
- * AM::GetTitleIDList service function
+ * AM::FindContentInfos service function
+ *  Inputs:
+ *      1 : MediaType
+ *    2-3 : u64, Title ID
+ *      4 : Content count
+ *      6 : Content IDs pointer
+ *      8 : Content Infos pointer
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ */
+void FindContentInfos(Service::Interface* self);
+
+/**
+ * AM::ListContentInfos service function
+ *  Inputs:
+ *      1 : Content count
+ *      2 : MediaType
+ *    3-4 : u64, Title ID
+ *      5 : Start Index
+ *      7 : Content Infos pointer
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ *      2 : Number of content infos returned
+ */
+void ListContentInfos(Service::Interface* self);
+
+/**
+ * AM::DeleteContents service function
+ *  Inputs:
+ *      1 : MediaType
+ *    2-3 : u64, Title ID
+ *      4 : Content count
+ *      6 : Content IDs pointer
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ */
+void DeleteContents(Service::Interface* self);
+
+/**
+ * AM::GetTitleList service function
  * Loads information about the desired number of titles from the desired media type into an array
  *  Inputs:
- *      0 : Command header (0x00020082)
- *      1 : The maximum number of titles to load
+ *      1 : Title count
  *      2 : Media type to load the titles from
- *      3 : Descriptor of the output buffer pointer
- *      4 : Address of the output buffer
+ *      4 : Title IDs output pointer
  *  Outputs:
  *      1 : Result, 0 on success, otherwise error code
  *      2 : The number of titles loaded from the requested media type
  */
-void GetTitleIDList(Service::Interface* self);
+void GetTitleList(Service::Interface* self);
+
+/**
+ * AM::GetTitleInfo service function
+ *  Inputs:
+ *      1 : u8 Mediatype
+ *      2 : Total titles
+ *      4 : TitleIDList pointer
+ *      6 : TitleList pointer
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ */
+void GetTitleInfo(Service::Interface* self);
+
+/**
+ * AM::GetDataTitleInfos service function
+ * Wrapper for AM::GetTitleInfo
+ *  Inputs:
+ *      1 : u8 Mediatype
+ *      2 : Total titles
+ *      4 : TitleIDList pointer
+ *      6 : TitleList pointer
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ */
+void GetDataTitleInfos(Service::Interface* self);
+
+/**
+ * AM::ListDataTitleTicketInfos service function
+ *  Inputs:
+ *      1 : Ticket count
+ *    2-3 : u64, Title ID
+ *      4 : Start Index?
+ *      5 : (TicketCount * 24) << 8 | 0x4
+ *      6 : Ticket Infos pointer
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ *      2 : Number of ticket infos returned
+ */
+void ListDataTitleTicketInfos(Service::Interface* self);
 
 /**
  * AM::GetNumContentInfos service function
  *  Inputs:
  *      0 : Command header (0x100100C0)
- *      1 : Unknown
- *      2 : Unknown
- *      3 : Unknown
+ *      1 : MediaType
+ *    2-3 : u64, Title ID
  *  Outputs:
  *      1 : Result, 0 on success, otherwise error code
  *      2 : Number of content infos plus one
  */
 void GetNumContentInfos(Service::Interface* self);
 
+/**
+ * AM::DeleteTicket service function
+ *  Inputs:
+ *    1-2 : u64, Title ID
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ */
+void DeleteTicket(Service::Interface* self);
+
+/**
+ * AM::GetTicketCount service function
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ *      2 : Total titles
+ */
+void GetTicketCount(Service::Interface* self);
+
+/**
+ * AM::GetTicketList service function
+ *  Inputs:
+ *      1 : Number of TicketList
+ *      2 : Number to skip
+ *      4 : TicketList pointer
+ *  Outputs:
+ *      1 : Result, 0 on success, otherwise error code
+ *      2 : Total TicketList
+ */
+void GetTicketList(Service::Interface* self);
+
 /// Initialize AM service
 void Init();
 
diff --git a/src/core/hle/service/am/am_app.cpp b/src/core/hle/service/am/am_app.cpp
index 16c76a1eb..d27b3defd 100644
--- a/src/core/hle/service/am/am_app.cpp
+++ b/src/core/hle/service/am/am_app.cpp
@@ -9,14 +9,14 @@ namespace Service {
 namespace AM {
 
 const Interface::FunctionInfo FunctionTable[] = {
-    {0x100100C0, GetNumContentInfos, "GetNumContentInfos"},
-    {0x10020104, nullptr, "FindContentInfos"},
-    {0x10030142, nullptr, "ListContentInfos"},
-    {0x10040102, nullptr, "DeleteContents"},
-    {0x10050084, nullptr, "GetDataTitleInfos"},
-    {0x10070102, nullptr, "ListDataTitleTicketInfos"},
-    {0x100900C0, nullptr, "IsDataTitleInUse"},
-    {0x100A0000, nullptr, "IsExternalTitleDatabaseInitialized"},
+    {0x100100C0, GetNumContentInfos,        "GetNumContentInfos"},
+    {0x10020104, FindContentInfos,          "FindContentInfos"},
+    {0x10030142, ListContentInfos,          "ListContentInfos"},
+    {0x10040102, DeleteContents,            "DeleteContents"},
+    {0x10050084, GetDataTitleInfos,         "GetDataTitleInfos"},
+    {0x10070102, ListDataTitleTicketInfos,  "ListDataTitleTicketInfos"},
+    {0x100900C0, nullptr,                   "IsDataTitleInUse"},
+    {0x100A0000, nullptr,                   "IsExternalTitleDatabaseInitialized"},
 };
 
 AM_APP_Interface::AM_APP_Interface() {
diff --git a/src/core/hle/service/am/am_net.cpp b/src/core/hle/service/am/am_net.cpp
index 065e04118..e75755245 100644
--- a/src/core/hle/service/am/am_net.cpp
+++ b/src/core/hle/service/am/am_net.cpp
@@ -9,16 +9,20 @@ namespace Service {
 namespace AM {
 
 const Interface::FunctionInfo FunctionTable[] = {
-    {0x00010040, TitleIDListGetTotal,   "TitleIDListGetTotal"},
-    {0x00020082, GetTitleIDList,        "GetTitleIDList"},
-    {0x00030084, nullptr,               "ListTitles"},
+    {0x00010040, GetTitleCount,         "GetTitleCount"},
+    {0x00020082, GetTitleList,          "GetTitleList"},
+    {0x00030084, GetTitleInfo,          "GetTitleInfo"},
     {0x000400C0, nullptr,               "DeleteApplicationTitle"},
     {0x000500C0, nullptr,               "GetTitleProductCode"},
-    {0x00080000, nullptr,               "TitleIDListGetTotal3"},
-    {0x00090082, nullptr,               "GetTitleIDList3"},
+    {0x000600C0, nullptr,               "GetTitleExtDataId"},
+    {0x00070080, DeleteTicket,          "DeleteTicket"},
+    {0x00080000, GetTicketCount,        "GetTicketCount"},
+    {0x00090082, GetTicketList,         "GetTicketList"},
     {0x000A0000, nullptr,               "GetDeviceID"},
-    {0x000D0084, nullptr,               "ListTitles2"},
-    {0x00140040, nullptr,               "FinishInstallToMedia"},
+    {0x000D0084, nullptr,               "GetPendingTitleInfo"},
+    {0x000E00C0, nullptr,               "DeletePendingTitle"},
+    {0x00140040, nullptr,               "FinalizePendingTitles"},
+    {0x00150040, nullptr,               "DeleteAllPendingTitles"},
     {0x00180080, nullptr,               "InitializeTitleDatabase"},
     {0x00190040, nullptr,               "ReloadDBS"},
     {0x001A00C0, nullptr,               "GetDSiWareExportSize"},
diff --git a/src/core/hle/service/am/am_sys.cpp b/src/core/hle/service/am/am_sys.cpp
index e38812297..8bad5e1c9 100644
--- a/src/core/hle/service/am/am_sys.cpp
+++ b/src/core/hle/service/am/am_sys.cpp
@@ -9,23 +9,27 @@ namespace Service {
 namespace AM {
 
 const Interface::FunctionInfo FunctionTable[] = {
-    {0x00010040, TitleIDListGetTotal,         "TitleIDListGetTotal"},
-    {0x00020082, GetTitleIDList,              "GetTitleIDList"},
-    {0x00030084, nullptr,                     "ListTitles"},
+    {0x00010040, GetTitleCount,               "GetTitleCount"},
+    {0x00020082, GetTitleList,                "GetTitleList"},
+    {0x00030084, GetTitleInfo,                "GetTitleInfo"},
     {0x000400C0, nullptr,                     "DeleteApplicationTitle"},
     {0x000500C0, nullptr,                     "GetTitleProductCode"},
-    {0x00080000, nullptr,                     "TitleIDListGetTotal3"},
-    {0x00090082, nullptr,                     "GetTitleIDList3"},
+    {0x000600C0, nullptr,                     "GetTitleExtDataId"},
+    {0x00070080, DeleteTicket,                "DeleteTicket"},
+    {0x00080000, GetTicketCount,              "GetTicketCount"},
+    {0x00090082, GetTicketList,               "GetTicketList"},
     {0x000A0000, nullptr,                     "GetDeviceID"},
-    {0x000D0084, nullptr,                     "ListTitles2"},
-    {0x00140040, nullptr,                     "FinishInstallToMedia"},
+    {0x000D0084, nullptr,                     "GetPendingTitleInfo"},
+    {0x000E00C0, nullptr,                     "DeletePendingTitle"},
+    {0x00140040, nullptr,                     "FinalizePendingTitles"},
+    {0x00150040, nullptr,                     "DeleteAllPendingTitles"},
     {0x00180080, nullptr,                     "InitializeTitleDatabase"},
     {0x00190040, nullptr,                     "ReloadDBS"},
     {0x001A00C0, nullptr,                     "GetDSiWareExportSize"},
     {0x001B0144, nullptr,                     "ExportDSiWare"},
     {0x001C0084, nullptr,                     "ImportDSiWare"},
-    {0x00230080, nullptr,                     "TitleIDListGetTotal2"},
-    {0x002400C2, nullptr,                     "GetTitleIDList2"}
+    {0x00230080, nullptr,                     "GetPendingTitleCount"},
+    {0x002400C2, nullptr,                     "GetPendingTitleList"}
 };
 
 AM_SYS_Interface::AM_SYS_Interface() {
diff --git a/src/core/hle/service/am/am_u.cpp b/src/core/hle/service/am/am_u.cpp
index c0392b754..d583dd9e6 100644
--- a/src/core/hle/service/am/am_u.cpp
+++ b/src/core/hle/service/am/am_u.cpp
@@ -9,16 +9,20 @@ namespace Service {
 namespace AM {
 
 const Interface::FunctionInfo FunctionTable[] = {
-    {0x00010040, TitleIDListGetTotal,         "TitleIDListGetTotal"},
-    {0x00020082, GetTitleIDList,              "GetTitleIDList"},
-    {0x00030084, nullptr,                     "ListTitles"},
+    {0x00010040, GetTitleCount,               "GetTitleCount"},
+    {0x00020082, GetTitleList,                "GetTitleList"},
+    {0x00030084, GetTitleInfo,                "GetTitleInfo"},
     {0x000400C0, nullptr,                     "DeleteApplicationTitle"},
     {0x000500C0, nullptr,                     "GetTitleProductCode"},
-    {0x00080000, nullptr,                     "TitleIDListGetTotal3"},
-    {0x00090082, nullptr,                     "GetTitleIDList3"},
+    {0x000600C0, nullptr,                     "GetTitleExtDataId"},
+    {0x00070080, DeleteTicket,                "DeleteTicket"},
+    {0x00080000, GetTicketCount,              "GetTicketCount"},
+    {0x00090082, GetTicketList,               "GetTicketList"},
     {0x000A0000, nullptr,                     "GetDeviceID"},
-    {0x000D0084, nullptr,                     "ListTitles2"},
-    {0x00140040, nullptr,                     "FinishInstallToMedia"},
+    {0x000D0084, nullptr,                     "GetPendingTitleInfo"},
+    {0x000E00C0, nullptr,                     "DeletePendingTitle"},
+    {0x00140040, nullptr,                     "FinalizePendingTitles"},
+    {0x00150040, nullptr,                     "DeleteAllPendingTitles"},
     {0x00180080, nullptr,                     "InitializeTitleDatabase"},
     {0x00190040, nullptr,                     "ReloadDBS"},
     {0x001A00C0, nullptr,                     "GetDSiWareExportSize"},
diff --git a/src/core/hle/service/apt/apt.cpp b/src/core/hle/service/apt/apt.cpp
index a49365287..6d72e8188 100644
--- a/src/core/hle/service/apt/apt.cpp
+++ b/src/core/hle/service/apt/apt.cpp
@@ -397,6 +397,23 @@ void GetAppletInfo(Service::Interface* self) {
     LOG_WARNING(Service_APT, "(stubbed) called appid=%u", app_id);
 }
 
+void GetStartupArgument(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+    u32 parameter_size = cmd_buff[1];
+    StartupArgumentType startup_argument_type = static_cast<StartupArgumentType>(cmd_buff[2]);
+
+    if (parameter_size >= 0x300) {
+        LOG_ERROR(Service_APT, "Parameter size is outside the valid range (capped to 0x300): parameter_size=0x%08x", parameter_size);
+        return;
+    }
+
+    LOG_WARNING(Service_APT,"(stubbed) called startup_argument_type=%u , parameter_size=0x%08x , parameter_value=0x%08x",
+                startup_argument_type, parameter_size, Memory::Read32(cmd_buff[41]));
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    cmd_buff[2] = (parameter_size > 0) ? 1 : 0;
+}
+
 void Init() {
     AddService(new APT_A_Interface);
     AddService(new APT_S_Interface);
diff --git a/src/core/hle/service/apt/apt.h b/src/core/hle/service/apt/apt.h
index 47a97c1a1..668b4a66f 100644
--- a/src/core/hle/service/apt/apt.h
+++ b/src/core/hle/service/apt/apt.h
@@ -67,6 +67,12 @@ enum class AppletId : u32 {
     Ed2                = 0x402,
 };
 
+enum class StartupArgumentType : u32 {
+    OtherApp   = 0,
+    Restart    = 1,
+    OtherMedia = 2,
+};
+
 /// Send a parameter to the currently-running application, which will read it via ReceiveParameter
 void SendParameter(const MessageParameter& parameter);
 
@@ -344,6 +350,17 @@ void PreloadLibraryApplet(Service::Interface* self);
  */
 void StartLibraryApplet(Service::Interface* self);
 
+/**
+ * APT::GetStartupArgument service function
+ *  Inputs:
+ *      1 : Parameter Size (capped to 0x300)
+ *      2 : StartupArgumentType
+ *  Outputs:
+ *      0 : Return header
+ *      1 : u8, Exists (0 = does not exist, 1 = exists)
+ */
+void GetStartupArgument(Service::Interface* self);
+
 /// Initialize the APT service
 void Init();
 
diff --git a/src/core/hle/service/apt/apt_a.cpp b/src/core/hle/service/apt/apt_a.cpp
index 0c6a77305..9ff47701a 100644
--- a/src/core/hle/service/apt/apt_a.cpp
+++ b/src/core/hle/service/apt/apt_a.cpp
@@ -13,9 +13,10 @@ const Interface::FunctionInfo FunctionTable[] = {
     {0x00020080, Initialize,                   "Initialize?"},
     {0x00030040, Enable,                       "Enable?"},
     {0x00040040, nullptr,                      "Finalize?"},
-    {0x00050040, nullptr,                      "GetAppletManInfo?"},
-    {0x00060040, nullptr,                      "GetAppletInfo?"},
+    {0x00050040, GetAppletManInfo,             "GetAppletManInfo"},
+    {0x00060040, GetAppletInfo,                "GetAppletInfo"},
     {0x00090040, IsRegistered,                 "IsRegistered"},
+    {0x000B0040, InquireNotification,          "InquireNotification"},
     {0x000C0104, SendParameter,                "SendParameter"},
     {0x000D0080, ReceiveParameter,             "ReceiveParameter"},
     {0x000E0080, GlanceParameter,              "GlanceParameter"},
@@ -24,9 +25,13 @@ const Interface::FunctionInfo FunctionTable[] = {
     {0x00180040, PrepareToStartLibraryApplet,  "PrepareToStartLibraryApplet"},
     {0x001E0084, StartLibraryApplet,           "StartLibraryApplet"},
     {0x003B0040, nullptr,                      "CancelLibraryApplet?"},
+    {0x003E0080, nullptr,                      "ReplySleepQuery"},
     {0x00430040, NotifyToWait,                 "NotifyToWait?"},
     {0x00440000, GetSharedFont,                "GetSharedFont?"},
     {0x004B00C2, AppletUtility,                "AppletUtility?"},
+    {0x004F0080, SetAppCpuTimeLimit,           "SetAppCpuTimeLimit"},
+    {0x00500040, GetAppCpuTimeLimit,           "GetAppCpuTimeLimit"},
+    {0x00510080, GetStartupArgument,           "GetStartupArgument"},
     {0x00550040, nullptr,                      "WriteInputToNsState?"},
 };
 
diff --git a/src/core/hle/service/apt/apt_s.cpp b/src/core/hle/service/apt/apt_s.cpp
index 7f6e81a63..ca54e593c 100644
--- a/src/core/hle/service/apt/apt_s.cpp
+++ b/src/core/hle/service/apt/apt_s.cpp
@@ -13,8 +13,8 @@ const Interface::FunctionInfo FunctionTable[] = {
     {0x00020080, Initialize,            "Initialize"},
     {0x00030040, Enable,                "Enable"},
     {0x00040040, nullptr,               "Finalize"},
-    {0x00050040, nullptr,               "GetAppletManInfo"},
-    {0x00060040, nullptr,               "GetAppletInfo"},
+    {0x00050040, GetAppletManInfo,      "GetAppletManInfo"},
+    {0x00060040, GetAppletInfo,         "GetAppletInfo"},
     {0x00070000, nullptr,               "GetLastSignaledAppletId"},
     {0x00080000, nullptr,               "CountRegisteredApplet"},
     {0x00090040, nullptr,               "IsRegistered"},
@@ -87,9 +87,9 @@ const Interface::FunctionInfo FunctionTable[] = {
     {0x004C0000, nullptr,               "SetFatalErrDispMode"},
     {0x004D0080, nullptr,               "GetAppletProgramInfo"},
     {0x004E0000, nullptr,               "HardwareResetAsync"},
-    {0x004F0080, nullptr,               "SetApplicationCpuTimeLimit"},
-    {0x00500040, nullptr,               "GetApplicationCpuTimeLimit"},
-    {0x00510080, nullptr,               "GetStartupArgument"},
+    {0x004F0080, SetAppCpuTimeLimit,    "SetAppCpuTimeLimit"},
+    {0x00500040, GetAppCpuTimeLimit,    "GetAppCpuTimeLimit"},
+    {0x00510080, GetStartupArgument,    "GetStartupArgument"},
     {0x00520104, nullptr,               "Wrap1"},
     {0x00530104, nullptr,               "Unwrap1"},
     {0x00580002, nullptr,               "GetProgramID"},
diff --git a/src/core/hle/service/apt/apt_u.cpp b/src/core/hle/service/apt/apt_u.cpp
index b13b51549..0e85c6d08 100644
--- a/src/core/hle/service/apt/apt_u.cpp
+++ b/src/core/hle/service/apt/apt_u.cpp
@@ -89,7 +89,7 @@ const Interface::FunctionInfo FunctionTable[] = {
     {0x004E0000, nullptr,                         "HardwareResetAsync"},
     {0x004F0080, SetAppCpuTimeLimit,              "SetAppCpuTimeLimit"},
     {0x00500040, GetAppCpuTimeLimit,              "GetAppCpuTimeLimit"},
-    {0x00510080, nullptr,                         "GetStartupArgument"},
+    {0x00510080, GetStartupArgument,              "GetStartupArgument"},
     {0x00520104, nullptr,                         "Wrap1"},
     {0x00530104, nullptr,                         "Unwrap1"},
     {0x00580002, nullptr,                         "GetProgramID"},
diff --git a/src/core/hle/service/cecd/cecd.cpp b/src/core/hle/service/cecd/cecd.cpp
index 6d79ce9b4..50c03495e 100644
--- a/src/core/hle/service/cecd/cecd.cpp
+++ b/src/core/hle/service/cecd/cecd.cpp
@@ -4,6 +4,7 @@
 
 #include "common/logging/log.h"
 
+#include "core/hle/kernel/event.h"
 #include "core/hle/service/service.h"
 #include "core/hle/service/cecd/cecd.h"
 #include "core/hle/service/cecd/cecd_s.h"
@@ -12,14 +13,47 @@
 namespace Service {
 namespace CECD {
 
-void Init() {
-    using namespace Kernel;
+static Kernel::SharedPtr<Kernel::Event> cecinfo_event;
+static Kernel::SharedPtr<Kernel::Event> change_state_event;
+
+void GetCecStateAbbreviated(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+    cmd_buff[2] = static_cast<u32>(CecStateAbbreviated::CEC_STATE_ABBREV_IDLE);
+
+    LOG_WARNING(Service_CECD, "(STUBBED) called");
+}
+
+void GetCecInfoEventHandle(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
 
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+    cmd_buff[3] = Kernel::g_handle_table.Create(cecinfo_event).MoveFrom(); // Event handle
+
+    LOG_WARNING(Service_CECD, "(STUBBED) called");
+}
+
+void GetChangeStateEventHandle(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+    cmd_buff[3] = Kernel::g_handle_table.Create(change_state_event).MoveFrom(); // Event handle
+
+    LOG_WARNING(Service_CECD, "(STUBBED) called");
+}
+
+void Init() {
     AddService(new CECD_S_Interface);
     AddService(new CECD_U_Interface);
+
+    cecinfo_event = Kernel::Event::Create(Kernel::ResetType::OneShot, "CECD_U::cecinfo_event");
+    change_state_event = Kernel::Event::Create(Kernel::ResetType::OneShot, "CECD_U::change_state_event");
 }
 
 void Shutdown() {
+    cecinfo_event = nullptr;
+    change_state_event = nullptr;
 }
 
 } // namespace CECD
diff --git a/src/core/hle/service/cecd/cecd.h b/src/core/hle/service/cecd/cecd.h
index 9e158521b..435611363 100644
--- a/src/core/hle/service/cecd/cecd.h
+++ b/src/core/hle/service/cecd/cecd.h
@@ -5,8 +5,49 @@
 #pragma once
 
 namespace Service {
+
+class Interface;
+
 namespace CECD {
 
+enum class CecStateAbbreviated {
+    CEC_STATE_ABBREV_IDLE      = 1, ///< Corresponds to CEC_STATE_IDLE
+    CEC_STATE_ABBREV_NOT_LOCAL = 2, ///< Corresponds to CEC_STATEs *FINISH*, *POST, and OVER_BOSS
+    CEC_STATE_ABBREV_SCANNING  = 3, ///< Corresponds to CEC_STATE_SCANNING
+    CEC_STATE_ABBREV_WLREADY   = 4, ///< Corresponds to CEC_STATE_WIRELESS_READY when some unknown bool is true
+    CEC_STATE_ABBREV_OTHER     = 5, ///< Corresponds to CEC_STATEs besides *FINISH*, *POST, and OVER_BOSS and those listed here
+};
+
+/**
+ * GetCecStateAbbreviated service function
+ *  Inputs:
+ *      0: 0x000E0000
+ *  Outputs:
+ *      1: ResultCode
+ *      2: CecStateAbbreviated
+ */
+void GetCecStateAbbreviated(Service::Interface* self);
+
+/**
+ * GetCecInfoEventHandle service function
+ *  Inputs:
+ *      0: 0x000F0000
+ *  Outputs:
+ *      1: ResultCode
+ *      3: Event Handle
+ */
+void GetCecInfoEventHandle(Service::Interface* self);
+
+/**
+ * GetChangeStateEventHandle service function
+ *  Inputs:
+ *      0: 0x00100000
+ *  Outputs:
+ *      1: ResultCode
+ *      3: Event Handle
+ */
+void GetChangeStateEventHandle(Service::Interface* self);
+
 /// Initialize CECD service(s)
 void Init();
 
diff --git a/src/core/hle/service/cecd/cecd_u.cpp b/src/core/hle/service/cecd/cecd_u.cpp
index 9b720a738..be6d4d8f6 100644
--- a/src/core/hle/service/cecd/cecd_u.cpp
+++ b/src/core/hle/service/cecd/cecd_u.cpp
@@ -2,13 +2,17 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include "core/hle/service/cecd/cecd.h"
 #include "core/hle/service/cecd/cecd_u.h"
 
 namespace Service {
 namespace CECD {
 
 static const Interface::FunctionInfo FunctionTable[] = {
-    { 0x00120104, nullptr, "ReadSavedData" },
+    {0x000E0000, GetCecStateAbbreviated,    "GetCecStateAbbreviated"},
+    {0x000F0000, GetCecInfoEventHandle,     "GetCecInfoEventHandle"},
+    {0x00100000, GetChangeStateEventHandle, "GetChangeStateEventHandle"},
+    {0x00120104, nullptr,                   "ReadSavedData"},
 };
 
 CECD_U_Interface::CECD_U_Interface() {
diff --git a/src/core/hle/service/cfg/cfg_i.cpp b/src/core/hle/service/cfg/cfg_i.cpp
index 0559a07b2..b18060f6d 100644
--- a/src/core/hle/service/cfg/cfg_i.cpp
+++ b/src/core/hle/service/cfg/cfg_i.cpp
@@ -9,6 +9,18 @@ namespace Service {
 namespace CFG {
 
 const Interface::FunctionInfo FunctionTable[] = {
+    // cfg common
+    {0x00010082, GetConfigInfoBlk2,                    "GetConfigInfoBlk2"},
+    {0x00020000, SecureInfoGetRegion,                  "SecureInfoGetRegion"},
+    {0x00030040, GenHashConsoleUnique,                 "GenHashConsoleUnique"},
+    {0x00040000, GetRegionCanadaUSA,                   "GetRegionCanadaUSA"},
+    {0x00050000, GetSystemModel,                       "GetSystemModel"},
+    {0x00060000, GetModelNintendo2DS,                  "GetModelNintendo2DS"},
+    {0x00070040, nullptr,                              "WriteToFirstByteCfgSavegame"},
+    {0x00080080, nullptr,                              "GoThroughTable"},
+    {0x00090040, GetCountryCodeString,                 "GetCountryCodeString"},
+    {0x000A0040, GetCountryCodeID,                     "GetCountryCodeID"},
+    // cfg:i
     {0x04010082, GetConfigInfoBlk8,                    "GetConfigInfoBlk8"},
     {0x04020082, nullptr,                              "SetConfigInfoBlk4"},
     {0x04030000, UpdateConfigNANDSavegame,             "UpdateConfigNANDSavegame"},
diff --git a/src/core/hle/service/cfg/cfg_s.cpp b/src/core/hle/service/cfg/cfg_s.cpp
index b03d290e5..e001f7687 100644
--- a/src/core/hle/service/cfg/cfg_s.cpp
+++ b/src/core/hle/service/cfg/cfg_s.cpp
@@ -9,10 +9,18 @@ namespace Service {
 namespace CFG {
 
 const Interface::FunctionInfo FunctionTable[] = {
+    // cfg common
     {0x00010082, GetConfigInfoBlk2,                    "GetConfigInfoBlk2"},
     {0x00020000, SecureInfoGetRegion,                  "SecureInfoGetRegion"},
     {0x00030040, GenHashConsoleUnique,                 "GenHashConsoleUnique"},
+    {0x00040000, GetRegionCanadaUSA,                   "GetRegionCanadaUSA"},
     {0x00050000, GetSystemModel,                       "GetSystemModel"},
+    {0x00060000, GetModelNintendo2DS,                  "GetModelNintendo2DS"},
+    {0x00070040, nullptr,                              "WriteToFirstByteCfgSavegame"},
+    {0x00080080, nullptr,                              "GoThroughTable"},
+    {0x00090040, GetCountryCodeString,                 "GetCountryCodeString"},
+    {0x000A0040, GetCountryCodeID,                     "GetCountryCodeID"},
+    // cfg:s
     {0x04010082, GetConfigInfoBlk8,                    "GetConfigInfoBlk8"},
     {0x04020082, nullptr,                              "SetConfigInfoBlk4"},
     {0x04030000, UpdateConfigNANDSavegame,             "UpdateConfigNANDSavegame"},
diff --git a/src/core/hle/service/cfg/cfg_u.cpp b/src/core/hle/service/cfg/cfg_u.cpp
index 89ae96c9e..606f7b2eb 100644
--- a/src/core/hle/service/cfg/cfg_u.cpp
+++ b/src/core/hle/service/cfg/cfg_u.cpp
@@ -9,6 +9,7 @@ namespace Service {
 namespace CFG {
 
 const Interface::FunctionInfo FunctionTable[] = {
+    // cfg common
     {0x00010082, GetConfigInfoBlk2,     "GetConfigInfoBlk2"},
     {0x00020000, SecureInfoGetRegion,   "SecureInfoGetRegion"},
     {0x00030040, GenHashConsoleUnique,  "GenHashConsoleUnique"},
diff --git a/src/core/hle/service/dlp_srvr.cpp b/src/core/hle/service/dlp_srvr.cpp
new file mode 100644
index 000000000..1f30188da
--- /dev/null
+++ b/src/core/hle/service/dlp_srvr.cpp
@@ -0,0 +1,36 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/logging/log.h"
+#include "core/hle/hle.h"
+#include "core/hle/service/dlp_srvr.h"
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Namespace DLP_SRVR
+
+namespace DLP_SRVR {
+
+static void unk_0x000E0040(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+    cmd_buff[2] = 0;
+
+    LOG_WARNING(Service_DLP, "(STUBBED) called");
+}
+
+const Interface::FunctionInfo FunctionTable[] = {
+    {0x00010183, nullptr,           "Initialize"},
+    {0x00020000, nullptr,           "Finalize"},
+    {0x000E0040, unk_0x000E0040,    "unk_0x000E0040"},
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Interface class
+
+Interface::Interface() {
+    Register(FunctionTable);
+}
+
+} // namespace
diff --git a/src/core/hle/service/dlp_srvr.h b/src/core/hle/service/dlp_srvr.h
new file mode 100644
index 000000000..d65d00814
--- /dev/null
+++ b/src/core/hle/service/dlp_srvr.h
@@ -0,0 +1,23 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "core/hle/service/service.h"
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Namespace DLP_SRVR
+
+namespace DLP_SRVR {
+
+class Interface : public Service::Interface {
+public:
+    Interface();
+
+    std::string GetPortName() const override {
+        return "dlp:SRVR";
+    }
+};
+
+} // namespace
diff --git a/src/core/hle/service/frd/frd.cpp b/src/core/hle/service/frd/frd.cpp
index c13ffd9d2..15d604bb6 100644
--- a/src/core/hle/service/frd/frd.cpp
+++ b/src/core/hle/service/frd/frd.cpp
@@ -2,6 +2,8 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include "common/string_util.h"
+
 #include "core/hle/service/service.h"
 #include "core/hle/service/frd/frd.h"
 #include "core/hle/service/frd/frd_a.h"
@@ -10,6 +12,95 @@
 namespace Service {
 namespace FRD {
 
+static FriendKey my_friend_key = {0, 0, 0ull};
+static MyPresence my_presence = {};
+
+void GetMyPresence(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 shifted_out_size = cmd_buff[64];
+    u32 my_presence_addr = cmd_buff[65];
+
+    ASSERT(shifted_out_size == ((sizeof(MyPresence) << 14) | 2));
+
+    Memory::WriteBlock(my_presence_addr, reinterpret_cast<const u8*>(&my_presence), sizeof(MyPresence));
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+
+    LOG_WARNING(Service_FRD, "(STUBBED) called");
+}
+
+void GetFriendKeyList(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 unknown = cmd_buff[1];
+    u32 frd_count = cmd_buff[2];
+    u32 frd_key_addr = cmd_buff[65];
+
+    FriendKey zero_key = {};
+    for (u32 i = 0; i < frd_count; ++i) {
+        Memory::WriteBlock(frd_key_addr + i * sizeof(FriendKey),
+                           reinterpret_cast<const u8*>(&zero_key), sizeof(FriendKey));
+    }
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+    cmd_buff[2] = 0; // 0 friends
+    LOG_WARNING(Service_FRD, "(STUBBED) called, unknown=%d, frd_count=%d, frd_key_addr=0x%08X",
+                unknown, frd_count, frd_key_addr);
+}
+
+void GetFriendProfile(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 count = cmd_buff[1];
+    u32 frd_key_addr = cmd_buff[3];
+    u32 profiles_addr = cmd_buff[65];
+
+    Profile zero_profile = {};
+    for (u32 i = 0; i < count; ++i) {
+        Memory::WriteBlock(profiles_addr + i * sizeof(Profile),
+            reinterpret_cast<const u8*>(&zero_profile), sizeof(Profile));
+    }
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+    LOG_WARNING(Service_FRD, "(STUBBED) called, count=%d, frd_key_addr=0x%08X, profiles_addr=0x%08X",
+                count, frd_key_addr, profiles_addr);
+}
+
+void GetFriendAttributeFlags(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    u32 count           = cmd_buff[1];
+    u32 frd_key_addr    = cmd_buff[3];
+    u32 attr_flags_addr = cmd_buff[65];
+
+    for (u32 i = 0; i < count; ++i) {
+        //TODO:(mailwl) figure out AttributeFlag size and zero all buffer. Assume 1 byte
+        Memory::Write8(attr_flags_addr + i, 0);
+    }
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+    LOG_WARNING(Service_FRD, "(STUBBED) called, count=%d, frd_key_addr=0x%08X, attr_flags_addr=0x%08X",
+                count, frd_key_addr, attr_flags_addr);
+}
+
+void GetMyFriendKey(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+    Memory::WriteBlock(cmd_buff[2], reinterpret_cast<const u8*>(&my_friend_key), sizeof(FriendKey));
+    LOG_WARNING(Service_FRD, "(STUBBED) called");
+}
+
+void GetMyScreenName(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+    // TODO: (mailwl) get the name from config
+    Common::UTF8ToUTF16("Citra").copy(reinterpret_cast<char16_t*>(&cmd_buff[2]), 11);
+    LOG_WARNING(Service_FRD, "(STUBBED) called");
+}
+
 void Init() {
     using namespace Kernel;
 
diff --git a/src/core/hle/service/frd/frd.h b/src/core/hle/service/frd/frd.h
index f9f88b444..c8283a7f3 100644
--- a/src/core/hle/service/frd/frd.h
+++ b/src/core/hle/service/frd/frd.h
@@ -4,9 +4,97 @@
 
 #pragma once
 
+#include "common/common_types.h"
+
 namespace Service {
+
+class Interface;
+
 namespace FRD {
 
+struct FriendKey {
+    u32 friend_id;
+    u32 unknown;
+    u64 friend_code;
+};
+
+struct MyPresence {
+    u8 unknown[0x12C];
+};
+
+struct Profile {
+    u8 region;
+    u8 country;
+    u8 area;
+    u8 language;
+    u32 unknown;
+};
+
+/**
+ * FRD::GetMyPresence service function
+ *  Inputs:
+ *      64 : sizeof (MyPresence) << 14 | 2
+ *      65 : Address of MyPresence structure
+ *  Outputs:
+ *      1 : Result of function, 0 on success, otherwise error code
+ */
+void GetMyPresence(Service::Interface* self);
+
+/**
+ * FRD::GetFriendKeyList service function
+ *  Inputs:
+ *      1 : Unknown
+ *      2 : Max friends count
+ *      65 : Address of FriendKey List
+ *  Outputs:
+ *      1 : Result of function, 0 on success, otherwise error code
+ *      2 : FriendKey count filled
+ */
+void GetFriendKeyList(Service::Interface* self);
+
+/**
+ * FRD::GetFriendProfile service function
+ *  Inputs:
+ *      1 : Friends count
+ *      2 : Friends count << 18 | 2
+ *      3 : Address of FriendKey List
+ *      64 : (count * sizeof (Profile)) << 10 | 2
+ *      65 : Address of Profiles List
+ *  Outputs:
+ *      1 : Result of function, 0 on success, otherwise error code
+ */
+void GetFriendProfile(Service::Interface* self);
+
+/**
+ * FRD::GetFriendAttributeFlags service function
+ *  Inputs:
+ *      1 : Friends count
+ *      2 : Friends count << 18 | 2
+ *      3 : Address of FriendKey List
+ *      65 : Address of AttributeFlags
+ *  Outputs:
+ *      1 : Result of function, 0 on success, otherwise error code
+ */
+void GetFriendAttributeFlags(Service::Interface* self);
+
+/**
+ * FRD::GetMyFriendKey service function
+ *  Inputs:
+ *      none
+ *  Outputs:
+ *      1 : Result of function, 0 on success, otherwise error code
+ *      2-5 : FriendKey
+ */
+void GetMyFriendKey(Service::Interface* self);
+
+/**
+ * FRD::GetMyScreenName service function
+ *  Outputs:
+ *      1 : Result of function, 0 on success, otherwise error code
+ *      2 : UTF16 encoded name (max 11 symbols)
+ */
+void GetMyScreenName(Service::Interface* self);
+
 /// Initialize FRD service(s)
 void Init();
 
diff --git a/src/core/hle/service/frd/frd_u.cpp b/src/core/hle/service/frd/frd_u.cpp
index 2c6885377..db8666416 100644
--- a/src/core/hle/service/frd/frd_u.cpp
+++ b/src/core/hle/service/frd/frd_u.cpp
@@ -2,65 +2,66 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include "core/hle/service/frd/frd.h"
 #include "core/hle/service/frd/frd_u.h"
 
 namespace Service {
 namespace FRD {
 
 const Interface::FunctionInfo FunctionTable[] = {
-    {0x00010000, nullptr,               "HasLoggedIn"},
-    {0x00020000, nullptr,               "IsOnline"},
-    {0x00030000, nullptr,               "Login"},
-    {0x00040000, nullptr,               "Logout"},
-    {0x00050000, nullptr,               "GetMyFriendKey"},
-    {0x00060000, nullptr,               "GetMyPreference"},
-    {0x00070000, nullptr,               "GetMyProfile"},
-    {0x00080000, nullptr,               "GetMyPresence"},
-    {0x00090000, nullptr,               "GetMyScreenName"},
-    {0x000A0000, nullptr,               "GetMyMii"},
-    {0x000B0000, nullptr,               "GetMyLocalAccountId"},
-    {0x000C0000, nullptr,               "GetMyPlayingGame"},
-    {0x000D0000, nullptr,               "GetMyFavoriteGame"},
-    {0x000E0000, nullptr,               "GetMyNcPrincipalId"},
-    {0x000F0000, nullptr,               "GetMyComment"},
-    {0x00100040, nullptr,               "GetMyPassword"},
-    {0x00110080, nullptr,               "GetFriendKeyList"},
-    {0x00120042, nullptr,               "GetFriendPresence"},
-    {0x00130142, nullptr,               "GetFriendScreenName"},
-    {0x00140044, nullptr,               "GetFriendMii"},
-    {0x00150042, nullptr,               "GetFriendProfile"},
-    {0x00160042, nullptr,               "GetFriendRelationship"},
-    {0x00170042, nullptr,               "GetFriendAttributeFlags"},
-    {0x00180044, nullptr,               "GetFriendPlayingGame"},
-    {0x00190042, nullptr,               "GetFriendFavoriteGame"},
-    {0x001A00C4, nullptr,               "GetFriendInfo"},
-    {0x001B0080, nullptr,               "IsIncludedInFriendList"},
-    {0x001C0042, nullptr,               "UnscrambleLocalFriendCode"},
-    {0x001D0002, nullptr,               "UpdateGameModeDescription"},
-    {0x001E02C2, nullptr,               "UpdateGameMode"},
-    {0x001F0042, nullptr,               "SendInvitation"},
-    {0x00200002, nullptr,               "AttachToEventNotification"},
-    {0x00210040, nullptr,               "SetNotificationMask"},
-    {0x00220040, nullptr,               "GetEventNotification"},
-    {0x00230000, nullptr,               "GetLastResponseResult"},
-    {0x00240040, nullptr,               "PrincipalIdToFriendCode"},
-    {0x00250080, nullptr,               "FriendCodeToPrincipalId"},
-    {0x00260080, nullptr,               "IsValidFriendCode"},
-    {0x00270040, nullptr,               "ResultToErrorCode"},
-    {0x00280244, nullptr,               "RequestGameAuthentication"},
-    {0x00290000, nullptr,               "GetGameAuthenticationData"},
-    {0x002A0204, nullptr,               "RequestServiceLocator"},
-    {0x002B0000, nullptr,               "GetServiceLocatorData"},
-    {0x002C0002, nullptr,               "DetectNatProperties"},
-    {0x002D0000, nullptr,               "GetNatProperties"},
-    {0x002E0000, nullptr,               "GetServerTimeInterval"},
-    {0x002F0040, nullptr,               "AllowHalfAwake"},
-    {0x00300000, nullptr,               "GetServerTypes"},
-    {0x00310082, nullptr,               "GetFriendComment"},
-    {0x00320042, nullptr,               "SetClientSdkVersion"},
-    {0x00330000, nullptr,               "GetMyApproachContext"},
-    {0x00340046, nullptr,               "AddFriendWithApproach"},
-    {0x00350082, nullptr,               "DecryptApproachContext"},
+    {0x00010000, nullptr,                 "HasLoggedIn"},
+    {0x00020000, nullptr,                 "IsOnline"},
+    {0x00030000, nullptr,                 "Login"},
+    {0x00040000, nullptr,                 "Logout"},
+    {0x00050000, GetMyFriendKey,          "GetMyFriendKey"},
+    {0x00060000, nullptr,                 "GetMyPreference"},
+    {0x00070000, nullptr,                 "GetMyProfile"},
+    {0x00080000, GetMyPresence,           "GetMyPresence"},
+    {0x00090000, GetMyScreenName,         "GetMyScreenName"},
+    {0x000A0000, nullptr,                 "GetMyMii"},
+    {0x000B0000, nullptr,                 "GetMyLocalAccountId"},
+    {0x000C0000, nullptr,                 "GetMyPlayingGame"},
+    {0x000D0000, nullptr,                 "GetMyFavoriteGame"},
+    {0x000E0000, nullptr,                 "GetMyNcPrincipalId"},
+    {0x000F0000, nullptr,                 "GetMyComment"},
+    {0x00100040, nullptr,                 "GetMyPassword"},
+    {0x00110080, GetFriendKeyList,        "GetFriendKeyList"},
+    {0x00120042, nullptr,                 "GetFriendPresence"},
+    {0x00130142, nullptr,                 "GetFriendScreenName"},
+    {0x00140044, nullptr,                 "GetFriendMii"},
+    {0x00150042, GetFriendProfile,        "GetFriendProfile"},
+    {0x00160042, nullptr,                 "GetFriendRelationship"},
+    {0x00170042, GetFriendAttributeFlags, "GetFriendAttributeFlags"},
+    {0x00180044, nullptr,                 "GetFriendPlayingGame"},
+    {0x00190042, nullptr,                 "GetFriendFavoriteGame"},
+    {0x001A00C4, nullptr,                 "GetFriendInfo"},
+    {0x001B0080, nullptr,                 "IsIncludedInFriendList"},
+    {0x001C0042, nullptr,                 "UnscrambleLocalFriendCode"},
+    {0x001D0002, nullptr,                 "UpdateGameModeDescription"},
+    {0x001E02C2, nullptr,                 "UpdateGameMode"},
+    {0x001F0042, nullptr,                 "SendInvitation"},
+    {0x00200002, nullptr,                 "AttachToEventNotification"},
+    {0x00210040, nullptr,                 "SetNotificationMask"},
+    {0x00220040, nullptr,                 "GetEventNotification"},
+    {0x00230000, nullptr,                 "GetLastResponseResult"},
+    {0x00240040, nullptr,                 "PrincipalIdToFriendCode"},
+    {0x00250080, nullptr,                 "FriendCodeToPrincipalId"},
+    {0x00260080, nullptr,                 "IsValidFriendCode"},
+    {0x00270040, nullptr,                 "ResultToErrorCode"},
+    {0x00280244, nullptr,                 "RequestGameAuthentication"},
+    {0x00290000, nullptr,                 "GetGameAuthenticationData"},
+    {0x002A0204, nullptr,                 "RequestServiceLocator"},
+    {0x002B0000, nullptr,                 "GetServiceLocatorData"},
+    {0x002C0002, nullptr,                 "DetectNatProperties"},
+    {0x002D0000, nullptr,                 "GetNatProperties"},
+    {0x002E0000, nullptr,                 "GetServerTimeInterval"},
+    {0x002F0040, nullptr,                 "AllowHalfAwake"},
+    {0x00300000, nullptr,                 "GetServerTypes"},
+    {0x00310082, nullptr,                 "GetFriendComment"},
+    {0x00320042, nullptr,                 "SetClientSdkVersion"},
+    {0x00330000, nullptr,                 "GetMyApproachContext"},
+    {0x00340046, nullptr,                 "AddFriendWithApproach"},
+    {0x00350082, nullptr,                 "DecryptApproachContext"},
 };
 
 FRD_U_Interface::FRD_U_Interface() {
diff --git a/src/core/hle/service/fs/archive.cpp b/src/core/hle/service/fs/archive.cpp
index 590697e76..e9588cb72 100644
--- a/src/core/hle/service/fs/archive.cpp
+++ b/src/core/hle/service/fs/archive.cpp
@@ -15,7 +15,6 @@
 #include "common/common_types.h"
 #include "common/file_util.h"
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 
 #include "core/file_sys/archive_backend.h"
 #include "core/file_sys/archive_extsavedata.h"
@@ -521,23 +520,23 @@ void ArchiveInit() {
 
     std::string sdmc_directory = FileUtil::GetUserPath(D_SDMC_IDX);
     std::string nand_directory = FileUtil::GetUserPath(D_NAND_IDX);
-    auto sdmc_factory = Common::make_unique<FileSys::ArchiveFactory_SDMC>(sdmc_directory);
+    auto sdmc_factory = std::make_unique<FileSys::ArchiveFactory_SDMC>(sdmc_directory);
     if (sdmc_factory->Initialize())
         RegisterArchiveType(std::move(sdmc_factory), ArchiveIdCode::SDMC);
     else
         LOG_ERROR(Service_FS, "Can't instantiate SDMC archive with path %s", sdmc_directory.c_str());
 
     // Create the SaveData archive
-    auto savedata_factory = Common::make_unique<FileSys::ArchiveFactory_SaveData>(sdmc_directory);
+    auto savedata_factory = std::make_unique<FileSys::ArchiveFactory_SaveData>(sdmc_directory);
     RegisterArchiveType(std::move(savedata_factory), ArchiveIdCode::SaveData);
 
-    auto extsavedata_factory = Common::make_unique<FileSys::ArchiveFactory_ExtSaveData>(sdmc_directory, false);
+    auto extsavedata_factory = std::make_unique<FileSys::ArchiveFactory_ExtSaveData>(sdmc_directory, false);
     if (extsavedata_factory->Initialize())
         RegisterArchiveType(std::move(extsavedata_factory), ArchiveIdCode::ExtSaveData);
     else
         LOG_ERROR(Service_FS, "Can't instantiate ExtSaveData archive with path %s", extsavedata_factory->GetMountPoint().c_str());
 
-    auto sharedextsavedata_factory = Common::make_unique<FileSys::ArchiveFactory_ExtSaveData>(nand_directory, true);
+    auto sharedextsavedata_factory = std::make_unique<FileSys::ArchiveFactory_ExtSaveData>(nand_directory, true);
     if (sharedextsavedata_factory->Initialize())
         RegisterArchiveType(std::move(sharedextsavedata_factory), ArchiveIdCode::SharedExtSaveData);
     else
@@ -545,10 +544,10 @@ void ArchiveInit() {
             sharedextsavedata_factory->GetMountPoint().c_str());
 
     // Create the SaveDataCheck archive, basically a small variation of the RomFS archive
-    auto savedatacheck_factory = Common::make_unique<FileSys::ArchiveFactory_SaveDataCheck>(nand_directory);
+    auto savedatacheck_factory = std::make_unique<FileSys::ArchiveFactory_SaveDataCheck>(nand_directory);
     RegisterArchiveType(std::move(savedatacheck_factory), ArchiveIdCode::SaveDataCheck);
 
-    auto systemsavedata_factory = Common::make_unique<FileSys::ArchiveFactory_SystemSaveData>(nand_directory);
+    auto systemsavedata_factory = std::make_unique<FileSys::ArchiveFactory_SystemSaveData>(nand_directory);
     RegisterArchiveType(std::move(systemsavedata_factory), ArchiveIdCode::SystemSaveData);
 }
 
diff --git a/src/core/hle/service/gsp_gpu.cpp b/src/core/hle/service/gsp_gpu.cpp
index 2ace2cade..0c655395e 100644
--- a/src/core/hle/service/gsp_gpu.cpp
+++ b/src/core/hle/service/gsp_gpu.cpp
@@ -31,6 +31,13 @@ const static u32 REGS_BEGIN = 0x1EB00000;
 
 namespace GSP_GPU {
 
+const ResultCode ERR_GSP_REGS_OUTOFRANGE_OR_MISALIGNED(ErrorDescription::OutofRangeOrMisalignedAddress, ErrorModule::GX,
+    ErrorSummary::InvalidArgument, ErrorLevel::Usage); // 0xE0E02A01
+const ResultCode ERR_GSP_REGS_MISALIGNED(ErrorDescription::MisalignedSize, ErrorModule::GX,
+    ErrorSummary::InvalidArgument, ErrorLevel::Usage); // 0xE0E02BF2
+const ResultCode ERR_GSP_REGS_INVALID_SIZE(ErrorDescription::InvalidSize, ErrorModule::GX,
+    ErrorSummary::InvalidArgument, ErrorLevel::Usage); // 0xE0E02BEC
+
 /// Event triggered when GSP interrupt has been signalled
 Kernel::SharedPtr<Kernel::Event> g_interrupt_event;
 /// GSP shared memoryings
@@ -59,47 +66,87 @@ static inline InterruptRelayQueue* GetInterruptRelayQueue(u32 thread_id) {
 }
 
 /**
- * Checks if the parameters in a register write call are valid and logs in the case that
- * they are not
- * @param base_address The first address in the sequence of registers that will be written
- * @param size_in_bytes The number of registers that will be written
- * @return true if the parameters are valid, false otherwise
+ * Writes sequential GSP GPU hardware registers using an array of source data
+ *
+ * @param base_address The address of the first register in the sequence
+ * @param size_in_bytes The number of registers to update (size of data)
+ * @param data A pointer to the source data
+ * @return RESULT_SUCCESS if the parameters are valid, error code otherwise
  */
-static bool CheckWriteParameters(u32 base_address, u32 size_in_bytes) {
-    // TODO: Return proper error codes
-    if (base_address + size_in_bytes >= 0x420000) {
-        LOG_ERROR(Service_GSP, "Write address out of range! (address=0x%08x, size=0x%08x)",
+static ResultCode WriteHWRegs(u32 base_address, u32 size_in_bytes, const u32* data) {
+    // This magic number is verified to be done by the gsp module
+    const u32 max_size_in_bytes = 0x80;
+
+    if (base_address & 3 || base_address >= 0x420000) {
+        LOG_ERROR(Service_GSP, "Write address was out of range or misaligned! (address=0x%08x, size=0x%08x)",
                   base_address, size_in_bytes);
-        return false;
-    }
+        return ERR_GSP_REGS_OUTOFRANGE_OR_MISALIGNED;
+    } else if (size_in_bytes <= max_size_in_bytes) {
+        if (size_in_bytes & 3) {
+            LOG_ERROR(Service_GSP, "Misaligned size 0x%08x", size_in_bytes);
+            return ERR_GSP_REGS_MISALIGNED;
+        } else {
+            while (size_in_bytes > 0) {
+                HW::Write<u32>(base_address + REGS_BEGIN, *data);
+
+                size_in_bytes -= 4;
+                ++data;
+                base_address += 4;
+            }
+            return RESULT_SUCCESS;
+        }
 
-    // size should be word-aligned
-    if ((size_in_bytes % 4) != 0) {
-        LOG_ERROR(Service_GSP, "Invalid size 0x%08x", size_in_bytes);
-        return false;
+    } else {
+        LOG_ERROR(Service_GSP, "Out of range size 0x%08x", size_in_bytes);
+        return ERR_GSP_REGS_INVALID_SIZE;
     }
-
-    return true;
 }
 
 /**
- * Writes sequential GSP GPU hardware registers using an array of source data
+ * Updates sequential GSP GPU hardware registers using parallel arrays of source data and masks.
+ * For each register, the value is updated only where the mask is high
  *
  * @param base_address The address of the first register in the sequence
  * @param size_in_bytes The number of registers to update (size of data)
- * @param data A pointer to the source data
+ * @param data A pointer to the source data to use for updates
+ * @param masks A pointer to the masks
+ * @return RESULT_SUCCESS if the parameters are valid, error code otherwise
  */
-static void WriteHWRegs(u32 base_address, u32 size_in_bytes, const u32* data) {
-    // TODO: Return proper error codes
-    if (!CheckWriteParameters(base_address, size_in_bytes))
-        return;
+static ResultCode WriteHWRegsWithMask(u32 base_address, u32 size_in_bytes, const u32* data, const u32* masks) {
+    // This magic number is verified to be done by the gsp module
+    const u32 max_size_in_bytes = 0x80;
 
-    while (size_in_bytes > 0) {
-        HW::Write<u32>(base_address + REGS_BEGIN, *data);
+    if (base_address & 3 || base_address >= 0x420000) {
+        LOG_ERROR(Service_GSP, "Write address was out of range or misaligned! (address=0x%08x, size=0x%08x)",
+                  base_address, size_in_bytes);
+        return ERR_GSP_REGS_OUTOFRANGE_OR_MISALIGNED;
+    } else if (size_in_bytes <= max_size_in_bytes) {
+        if (size_in_bytes & 3) {
+            LOG_ERROR(Service_GSP, "Misaligned size 0x%08x", size_in_bytes);
+            return ERR_GSP_REGS_MISALIGNED;
+        } else {
+            while (size_in_bytes > 0) {
+                const u32 reg_address = base_address + REGS_BEGIN;
+
+                u32 reg_value;
+                HW::Read<u32>(reg_value, reg_address);
+
+                // Update the current value of the register only for set mask bits
+                reg_value = (reg_value & ~*masks) | (*data | *masks);
+
+                HW::Write<u32>(reg_address, reg_value);
+
+                size_in_bytes -= 4;
+                ++data;
+                ++masks;
+                base_address += 4;
+            }
+            return RESULT_SUCCESS;
+        }
 
-        size_in_bytes -= 4;
-        ++data;
-        base_address += 4;
+    } else {
+        LOG_ERROR(Service_GSP, "Out of range size 0x%08x", size_in_bytes);
+        return ERR_GSP_REGS_INVALID_SIZE;
     }
 }
 
@@ -120,39 +167,7 @@ static void WriteHWRegs(Service::Interface* self) {
 
     u32* src = (u32*)Memory::GetPointer(cmd_buff[4]);
 
-    WriteHWRegs(reg_addr, size, src);
-}
-
-/**
- * Updates sequential GSP GPU hardware registers using parallel arrays of source data and masks.
- * For each register, the value is updated only where the mask is high
- *
- * @param base_address The address of the first register in the sequence
- * @param size_in_bytes The number of registers to update (size of data)
- * @param data A pointer to the source data to use for updates
- * @param masks A pointer to the masks
- */
-static void WriteHWRegsWithMask(u32 base_address, u32 size_in_bytes, const u32* data, const u32* masks) {
-    // TODO: Return proper error codes
-    if (!CheckWriteParameters(base_address, size_in_bytes))
-        return;
-
-    while (size_in_bytes > 0) {
-        const u32 reg_address = base_address + REGS_BEGIN;
-
-        u32 reg_value;
-        HW::Read<u32>(reg_value, reg_address);
-
-        // Update the current value of the register only for set mask bits
-        reg_value = (reg_value & ~*masks) | (*data | *masks);
-
-        HW::Write<u32>(reg_address, reg_value);
-
-        size_in_bytes -= 4;
-        ++data;
-        ++masks;
-        base_address += 4;
-    }
+    cmd_buff[1] = WriteHWRegs(reg_addr, size, src).raw;
 }
 
 /**
@@ -174,7 +189,7 @@ static void WriteHWRegsWithMask(Service::Interface* self) {
     u32* src_data = (u32*)Memory::GetPointer(cmd_buff[4]);
     u32* mask_data = (u32*)Memory::GetPointer(cmd_buff[6]);
 
-    WriteHWRegsWithMask(reg_addr, size, src_data, mask_data);
+    cmd_buff[1] = WriteHWRegsWithMask(reg_addr, size, src_data, mask_data).raw;
 }
 
 /// Read a GSP GPU hardware register
@@ -206,27 +221,27 @@ static void ReadHWRegs(Service::Interface* self) {
     }
 }
 
-void SetBufferSwap(u32 screen_id, const FrameBufferInfo& info) {
+ResultCode SetBufferSwap(u32 screen_id, const FrameBufferInfo& info) {
     u32 base_address = 0x400000;
     PAddr phys_address_left = Memory::VirtualToPhysicalAddress(info.address_left);
     PAddr phys_address_right = Memory::VirtualToPhysicalAddress(info.address_right);
     if (info.active_fb == 0) {
-        WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].address_left1)), 4,
-                &phys_address_left);
-        WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].address_right1)), 4,
-                &phys_address_right);
+        WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].address_left1)),
+                    4, &phys_address_left);
+        WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].address_right1)),
+                    4, &phys_address_right);
     } else {
-        WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].address_left2)), 4,
-                &phys_address_left);
-        WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].address_right2)), 4,
-                &phys_address_right);
+        WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].address_left2)),
+                    4, &phys_address_left);
+        WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].address_right2)),
+                    4, &phys_address_right);
     }
-    WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].stride)), 4,
-            &info.stride);
-    WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].color_format)), 4,
-            &info.format);
-    WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].active_fb)), 4,
-            &info.shown_fb);
+    WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].stride)),
+                4, &info.stride);
+    WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].color_format)),
+                4, &info.format);
+    WriteHWRegs(base_address + 4 * static_cast<u32>(GPU_REG_INDEX(framebuffer_config[screen_id].active_fb)),
+                4, &info.shown_fb);
 
     if (Pica::g_debug_context)
         Pica::g_debug_context->OnEvent(Pica::DebugContext::Event::BufferSwapped, nullptr);
@@ -234,6 +249,8 @@ void SetBufferSwap(u32 screen_id, const FrameBufferInfo& info) {
     if (screen_id == 0) {
         MicroProfileFlip();
     }
+
+    return RESULT_SUCCESS;
 }
 
 /**
@@ -251,9 +268,8 @@ static void SetBufferSwap(Service::Interface* self) {
     u32* cmd_buff = Kernel::GetCommandBuffer();
     u32 screen_id = cmd_buff[1];
     FrameBufferInfo* fb_info = (FrameBufferInfo*)&cmd_buff[2];
-    SetBufferSwap(screen_id, *fb_info);
 
-    cmd_buff[1] = 0; // No error
+    cmd_buff[1] = SetBufferSwap(screen_id, *fb_info).raw;
 }
 
 /**
@@ -286,6 +302,22 @@ static void FlushDataCache(Service::Interface* self) {
 }
 
 /**
+ * GSP_GPU::SetAxiConfigQoSMode service function
+ *  Inputs:
+ *      1 : Mode, unused in emulator
+ *  Outputs:
+ *      1 : Result of function, 0 on success, otherwise error code
+ */
+static void SetAxiConfigQoSMode(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+    u32 mode = cmd_buff[1];
+
+    cmd_buff[1] = RESULT_SUCCESS.raw; // No error
+
+    LOG_WARNING(Service_GSP, "(STUBBED) called mode=0x%08X", mode);
+}
+
+/**
  * GSP_GPU::RegisterInterruptRelayQueue service function
  *  Inputs:
  *      1 : "Flags" field, purpose is unknown
@@ -302,6 +334,12 @@ static void RegisterInterruptRelayQueue(Service::Interface* self) {
     g_interrupt_event = Kernel::g_handle_table.Get<Kernel::Event>(cmd_buff[3]);
     ASSERT_MSG((g_interrupt_event != nullptr), "handle is not valid!");
 
+    g_interrupt_event->name = "GSP_GPU::interrupt_event";
+
+    using Kernel::MemoryPermission;
+    g_shared_memory = Kernel::SharedMemory::Create(0x1000, MemoryPermission::ReadWrite,
+        MemoryPermission::ReadWrite, "GSPSharedMem");
+
     Handle shmem_handle = Kernel::g_handle_table.Create(g_shared_memory).MoveFrom();
 
     // This specific code is required for a successful initialization, rather than 0
@@ -314,6 +352,22 @@ static void RegisterInterruptRelayQueue(Service::Interface* self) {
 }
 
 /**
+ * GSP_GPU::UnregisterInterruptRelayQueue service function
+ *  Outputs:
+ *      1 : Result of function, 0 on success, otherwise error code
+ */
+static void UnregisterInterruptRelayQueue(Service::Interface* self) {
+    u32* cmd_buff = Kernel::GetCommandBuffer();
+
+    g_shared_memory = nullptr;
+    g_interrupt_event = nullptr;
+
+    cmd_buff[1] = RESULT_SUCCESS.raw;
+
+    LOG_WARNING(Service_GSP, "called");
+}
+
+/**
  * Signals that the specified interrupt type has occurred to userland code
  * @param interrupt_id ID of interrupt that is being signalled
  * @todo This should probably take a thread_id parameter and only signal this thread?
@@ -591,11 +645,11 @@ const Interface::FunctionInfo FunctionTable[] = {
     {0x000D0140, nullptr,                       "SetDisplayTransfer"},
     {0x000E0180, nullptr,                       "SetTextureCopy"},
     {0x000F0200, nullptr,                       "SetMemoryFill"},
-    {0x00100040, nullptr,                       "SetAxiConfigQoSMode"},
+    {0x00100040, SetAxiConfigQoSMode,           "SetAxiConfigQoSMode"},
     {0x00110040, nullptr,                       "SetPerfLogMode"},
     {0x00120000, nullptr,                       "GetPerfLog"},
     {0x00130042, RegisterInterruptRelayQueue,   "RegisterInterruptRelayQueue"},
-    {0x00140000, nullptr,                       "UnregisterInterruptRelayQueue"},
+    {0x00140000, UnregisterInterruptRelayQueue, "UnregisterInterruptRelayQueue"},
     {0x00150002, nullptr,                       "TryAcquireRight"},
     {0x00160042, nullptr,                       "AcquireRight"},
     {0x00170000, nullptr,                       "ReleaseRight"},
@@ -616,10 +670,7 @@ Interface::Interface() {
     Register(FunctionTable);
 
     g_interrupt_event = nullptr;
-
-    using Kernel::MemoryPermission;
-    g_shared_memory = Kernel::SharedMemory::Create(0x1000, MemoryPermission::ReadWrite,
-            MemoryPermission::ReadWrite, "GSPSharedMem");
+    g_shared_memory = nullptr;
 
     g_thread_id = 0;
 }
diff --git a/src/core/hle/service/gsp_gpu.h b/src/core/hle/service/gsp_gpu.h
index 0e2f7a21e..55a993bb8 100644
--- a/src/core/hle/service/gsp_gpu.h
+++ b/src/core/hle/service/gsp_gpu.h
@@ -194,7 +194,7 @@ public:
  */
 void SignalInterrupt(InterruptId interrupt_id);
 
-void SetBufferSwap(u32 screen_id, const FrameBufferInfo& info);
+ResultCode SetBufferSwap(u32 screen_id, const FrameBufferInfo& info);
 
 /**
  * Retrieves the framebuffer info stored in the GSP shared memory for the
diff --git a/src/core/hle/service/service.cpp b/src/core/hle/service/service.cpp
index 35b648409..0fe3a4d7a 100644
--- a/src/core/hle/service/service.cpp
+++ b/src/core/hle/service/service.cpp
@@ -9,6 +9,7 @@
 #include "core/hle/service/ac_u.h"
 #include "core/hle/service/act_u.h"
 #include "core/hle/service/csnd_snd.h"
+#include "core/hle/service/dlp_srvr.h"
 #include "core/hle/service/dsp_dsp.h"
 #include "core/hle/service/err_f.h"
 #include "core/hle/service/gsp_gpu.h"
@@ -70,9 +71,8 @@ ResultVal<bool> Interface::SyncRequest() {
         // TODO(bunnei): Hack - ignore error
         cmd_buff[1] = 0;
         return MakeResult<bool>(false);
-    } else {
-        LOG_TRACE(Service, "%s", MakeFunctionString(itr->second.name, GetPortName().c_str(), cmd_buff).c_str());
     }
+    LOG_TRACE(Service, "%s", MakeFunctionString(itr->second.name, GetPortName().c_str(), cmd_buff).c_str());
 
     itr->second.func(this);
 
@@ -121,6 +121,7 @@ void Init() {
     AddService(new AC_U::Interface);
     AddService(new ACT_U::Interface);
     AddService(new CSND_SND::Interface);
+    AddService(new DLP_SRVR::Interface);
     AddService(new DSP_DSP::Interface);
     AddService(new GSP_GPU::Interface);
     AddService(new GSP_LCD::Interface);
diff --git a/src/core/hle/shared_page.cpp b/src/core/hle/shared_page.cpp
index 50c5bc01b..2a1caeaac 100644
--- a/src/core/hle/shared_page.cpp
+++ b/src/core/hle/shared_page.cpp
@@ -16,6 +16,9 @@ void Init() {
     std::memset(&shared_page, 0, sizeof(shared_page));
 
     shared_page.running_hw = 0x1; // product
+
+    // Some games wait until this value becomes 0x1, before asking running_hw
+    shared_page.unknown_value = 0x1;
 }
 
 } // namespace
diff --git a/src/core/hle/shared_page.h b/src/core/hle/shared_page.h
index 379bb7b63..35a07c685 100644
--- a/src/core/hle/shared_page.h
+++ b/src/core/hle/shared_page.h
@@ -39,12 +39,14 @@ struct SharedPageDef {
     DateTime date_time_0;                // 20
     DateTime date_time_1;                // 40
     u8       wifi_macaddr[6];            // 60
-    u8       wifi_unknown1;              // 66
+    u8       wifi_link_level;            // 66
     u8       wifi_unknown2;              // 67
     INSERT_PADDING_BYTES(0x80 - 0x68);   // 68
     float_le sliderstate_3d;             // 80
     u8       ledstate_3d;                // 84
-    INSERT_PADDING_BYTES(0xA0 - 0x85);   // 85
+    INSERT_PADDING_BYTES(1);             // 85
+    u8       unknown_value;              // 86
+    INSERT_PADDING_BYTES(0xA0 - 0x87);   // 87
     u64_le   menu_title_id;              // A0
     u64_le   active_menu_title_id;       // A8
     INSERT_PADDING_BYTES(0x1000 - 0xB0); // B0
diff --git a/src/core/hw/y2r.cpp b/src/core/hw/y2r.cpp
index 48c45564f..083391e83 100644
--- a/src/core/hw/y2r.cpp
+++ b/src/core/hw/y2r.cpp
@@ -261,7 +261,7 @@ void PerformConversion(ConversionConfiguration& cvt) {
     ASSERT(cvt.block_alignment != BlockAlignment::Block8x8 || cvt.input_lines % 8 == 0);
     // Tiles per row
     size_t num_tiles = cvt.input_line_width / 8;
-    ASSERT(num_tiles < MAX_TILES);
+    ASSERT(num_tiles <= MAX_TILES);
 
     // Buffer used as a CDMA source/target.
     std::unique_ptr<u8[]> data_buffer(new u8[cvt.input_line_width * 8 * 4]);
diff --git a/src/core/loader/3dsx.cpp b/src/core/loader/3dsx.cpp
index 8eed6a50a..5fb3b9e2b 100644
--- a/src/core/loader/3dsx.cpp
+++ b/src/core/loader/3dsx.cpp
@@ -10,13 +10,9 @@
 #include "core/file_sys/archive_romfs.h"
 #include "core/hle/kernel/process.h"
 #include "core/hle/kernel/resource_limit.h"
-#include "core/hle/service/fs/archive.h"
-#include "core/loader/elf.h"
-#include "core/loader/ncch.h"
+#include "core/loader/3dsx.h"
 #include "core/memory.h"
 
-#include "3dsx.h"
-
 namespace Loader {
 
 /*
diff --git a/src/core/loader/loader.cpp b/src/core/loader/loader.cpp
index b1907cd55..886501c41 100644
--- a/src/core/loader/loader.cpp
+++ b/src/core/loader/loader.cpp
@@ -6,7 +6,6 @@
 #include <string>
 
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 #include "common/string_util.h"
 
 #include "core/file_sys/archive_romfs.h"
@@ -120,7 +119,7 @@ ResultStatus LoadFile(const std::string& filename) {
         AppLoader_THREEDSX app_loader(std::move(file), filename_filename, filename);
         // Load application and RomFS
         if (ResultStatus::Success == app_loader.Load()) {
-            Service::FS::RegisterArchiveType(Common::make_unique<FileSys::ArchiveFactory_RomFS>(app_loader), Service::FS::ArchiveIdCode::RomFS);
+            Service::FS::RegisterArchiveType(std::make_unique<FileSys::ArchiveFactory_RomFS>(app_loader), Service::FS::ArchiveIdCode::RomFS);
             return ResultStatus::Success;
         }
         break;
@@ -139,7 +138,7 @@ ResultStatus LoadFile(const std::string& filename) {
         // Load application and RomFS
         ResultStatus result = app_loader.Load();
         if (ResultStatus::Success == result) {
-            Service::FS::RegisterArchiveType(Common::make_unique<FileSys::ArchiveFactory_RomFS>(app_loader), Service::FS::ArchiveIdCode::RomFS);
+            Service::FS::RegisterArchiveType(std::make_unique<FileSys::ArchiveFactory_RomFS>(app_loader), Service::FS::ArchiveIdCode::RomFS);
         }
         return result;
     }
diff --git a/src/core/loader/ncch.cpp b/src/core/loader/ncch.cpp
index 93f21bec2..a4b47ef8c 100644
--- a/src/core/loader/ncch.cpp
+++ b/src/core/loader/ncch.cpp
@@ -7,7 +7,6 @@
 #include <memory>
 
 #include "common/logging/log.h"
-#include "common/make_unique.h"
 #include "common/string_util.h"
 #include "common/swap.h"
 
@@ -175,7 +174,7 @@ ResultStatus AppLoader_NCCH::LoadSectionExeFS(const char* name, std::vector<u8>&
         return ResultStatus::Error;
 
     LOG_DEBUG(Loader, "%d sections:", kMaxSections);
-    // Iterate through the ExeFs archive until we find the .code file...
+    // Iterate through the ExeFs archive until we find a section with the specified name...
     for (unsigned section_number = 0; section_number < kMaxSections; section_number++) {
         const auto& section = exefs_header.section[section_number];
 
@@ -187,7 +186,7 @@ ResultStatus AppLoader_NCCH::LoadSectionExeFS(const char* name, std::vector<u8>&
             s64 section_offset = (section.offset + exefs_offset + sizeof(ExeFs_Header) + ncch_offset);
             file.Seek(section_offset, SEEK_SET);
 
-            if (is_compressed) {
+            if (strcmp(section.name, ".code") == 0 && is_compressed) {
                 // Section is compressed, read compressed .code section...
                 std::unique_ptr<u8[]> temp_buffer;
                 try {
diff --git a/src/core/settings.cpp b/src/core/settings.cpp
index 8a14f75aa..1aa26fbd2 100644
--- a/src/core/settings.cpp
+++ b/src/core/settings.cpp
@@ -4,8 +4,22 @@
 
 #include "settings.h"
 
+#include "core/gdbstub/gdbstub.h"
+
+#include "video_core/video_core.h"
+
 namespace Settings {
 
 Values values = {};
 
+void Apply() {
+
+    GDBStub::SetServerPort(static_cast<u32>(values.gdbstub_port));
+    GDBStub::ToggleServer(values.use_gdbstub);
+
+    VideoCore::g_hw_renderer_enabled = values.use_hw_renderer;
+    VideoCore::g_shader_jit_enabled = values.use_shader_jit;
+
 }
+
+} // namespace
diff --git a/src/core/settings.h b/src/core/settings.h
index 97ddcdff9..4933a516d 100644
--- a/src/core/settings.h
+++ b/src/core/settings.h
@@ -67,4 +67,6 @@ struct Values {
     u16 gdbstub_port;
 } extern values;
 
+void Apply();
+
 }
diff --git a/src/video_core/command_processor.cpp b/src/video_core/command_processor.cpp
index 028b59348..3abe79c09 100644
--- a/src/video_core/command_processor.cpp
+++ b/src/video_core/command_processor.cpp
@@ -140,7 +140,7 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                         immediate_attribute_id = 0;
 
                         Shader::UnitState<false> shader_unit;
-                        Shader::Setup(shader_unit);
+                        Shader::Setup();
 
                         if (g_debug_context)
                             g_debug_context->OnEvent(DebugContext::Event::VertexLoaded, static_cast<void*>(&immediate_input));
@@ -249,10 +249,6 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
             const u16* index_address_16 = reinterpret_cast<const u16*>(index_address_8);
             bool index_u16 = index_info.format != 0;
 
-#if PICA_DUMP_GEOMETRY
-            DebugUtils::GeometryDumper geometry_dumper;
-            PrimitiveAssembler<DebugUtils::GeometryDumper::Vertex> dumping_primitive_assembler(regs.triangle_topology.Value());
-#endif
             PrimitiveAssembler<Shader::OutputVertex>& primitive_assembler = g_state.primitive_assembler;
 
             if (g_debug_context) {
@@ -304,7 +300,7 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
             vertex_cache_ids.fill(-1);
 
             Shader::UnitState<false> shader_unit;
-            Shader::Setup(shader_unit);
+            Shader::Setup();
 
             for (unsigned int index = 0; index < regs.num_vertices; ++index)
             {
@@ -388,17 +384,6 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                     if (g_debug_context)
                         g_debug_context->OnEvent(DebugContext::Event::VertexLoaded, (void*)&input);
 
-#if PICA_DUMP_GEOMETRY
-                    // NOTE: When dumping geometry, we simply assume that the first input attribute
-                    //       corresponds to the position for now.
-                    DebugUtils::GeometryDumper::Vertex dumped_vertex = {
-                        input.attr[0][0].ToFloat32(), input.attr[0][1].ToFloat32(), input.attr[0][2].ToFloat32()
-                    };
-                    using namespace std::placeholders;
-                    dumping_primitive_assembler.SubmitVertex(dumped_vertex,
-                                                             std::bind(&DebugUtils::GeometryDumper::AddTriangle,
-                                                                       &geometry_dumper, _1, _2, _3));
-#endif
                     // Send to vertex shader
                     output = Shader::Run(shader_unit, input, attribute_config.GetNumTotalAttributes());
 
@@ -424,10 +409,6 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                                                           range.second, range.first);
             }
 
-#if PICA_DUMP_GEOMETRY
-            geometry_dumper.Dump();
-#endif
-
             break;
         }
 
diff --git a/src/video_core/debug_utils/debug_utils.cpp b/src/video_core/debug_utils/debug_utils.cpp
index bac6d69c7..c3a9c9598 100644
--- a/src/video_core/debug_utils/debug_utils.cpp
+++ b/src/video_core/debug_utils/debug_utils.cpp
@@ -85,35 +85,6 @@ std::shared_ptr<DebugContext> g_debug_context; // TODO: Get rid of this global
 
 namespace DebugUtils {
 
-void GeometryDumper::AddTriangle(Vertex& v0, Vertex& v1, Vertex& v2) {
-    vertices.push_back(v0);
-    vertices.push_back(v1);
-    vertices.push_back(v2);
-
-    int num_vertices = (int)vertices.size();
-    faces.push_back({{ num_vertices-3, num_vertices-2, num_vertices-1 }});
-}
-
-void GeometryDumper::Dump() {
-    static int index = 0;
-    std::string filename = std::string("geometry_dump") + std::to_string(++index) + ".obj";
-
-    std::ofstream file(filename);
-
-    for (const auto& vertex : vertices) {
-        file << "v " << vertex.pos[0]
-             << " "  << vertex.pos[1]
-             << " "  << vertex.pos[2] << std::endl;
-    }
-
-    for (const Face& face : faces) {
-        file << "f " << 1+face.index[0]
-             << " "  << 1+face.index[1]
-             << " "  << 1+face.index[2] << std::endl;
-    }
-}
-
-
 void DumpShader(const std::string& filename, const Regs::ShaderConfig& config, const Shader::ShaderSetup& setup, const Regs::VSOutputAttributes* output_attributes)
 {
     struct StuffToWrite {
@@ -315,7 +286,7 @@ void StartPicaTracing()
     }
 
     std::lock_guard<std::mutex> lock(pica_trace_mutex);
-    pica_trace = std::unique_ptr<PicaTrace>(new PicaTrace);
+    pica_trace = std::make_unique<PicaTrace>();
 
     is_pica_tracing = true;
 }
@@ -615,6 +586,21 @@ TextureInfo TextureInfo::FromPicaRegister(const Regs::TextureConfig& config,
     return info;
 }
 
+#ifdef HAVE_PNG
+// Adapter functions to libpng to write/flush to File::IOFile instances.
+static void WriteIOFile(png_structp png_ptr, png_bytep data, png_size_t length) {
+    auto* fp = static_cast<FileUtil::IOFile*>(png_get_io_ptr(png_ptr));
+    if (!fp->WriteBytes(data, length))
+         png_error(png_ptr, "Failed to write to output PNG file.");
+}
+
+static void FlushIOFile(png_structp png_ptr) {
+    auto* fp = static_cast<FileUtil::IOFile*>(png_get_io_ptr(png_ptr));
+    if (!fp->Flush())
+         png_error(png_ptr, "Failed to flush to output PNG file.");
+}
+#endif
+
 void DumpTexture(const Pica::Regs::TextureConfig& texture_config, u8* data) {
 #ifndef HAVE_PNG
     return;
@@ -658,7 +644,7 @@ void DumpTexture(const Pica::Regs::TextureConfig& texture_config, u8* data) {
         goto finalise;
     }
 
-    png_init_io(png_ptr, fp.GetHandle());
+    png_set_write_fn(png_ptr, static_cast<void*>(&fp), WriteIOFile, FlushIOFile);
 
     // Write header (8 bit color depth)
     png_set_IHDR(png_ptr, info_ptr, texture_config.width, texture_config.height,
diff --git a/src/video_core/debug_utils/debug_utils.h b/src/video_core/debug_utils/debug_utils.h
index 795160a32..7df941619 100644
--- a/src/video_core/debug_utils/debug_utils.h
+++ b/src/video_core/debug_utils/debug_utils.h
@@ -158,30 +158,9 @@ extern std::shared_ptr<DebugContext> g_debug_context; // TODO: Get rid of this g
 
 namespace DebugUtils {
 
-#define PICA_DUMP_GEOMETRY 0
 #define PICA_DUMP_TEXTURES 0
 #define PICA_LOG_TEV 0
 
-// Simple utility class for dumping geometry data to an OBJ file
-class GeometryDumper {
-public:
-    struct Vertex {
-        std::array<float,3> pos;
-    };
-
-    void AddTriangle(Vertex& v0, Vertex& v1, Vertex& v2);
-
-    void Dump();
-
-private:
-    struct Face {
-        int index[3];
-    };
-
-    std::vector<Vertex> vertices;
-    std::vector<Face> faces;
-};
-
 void DumpShader(const std::string& filename, const Regs::ShaderConfig& config,
                 const Shader::ShaderSetup& setup, const Regs::VSOutputAttributes* output_attributes);
 
diff --git a/src/video_core/pica.h b/src/video_core/pica.h
index 16f9e4006..4552ff81c 100644
--- a/src/video_core/pica.h
+++ b/src/video_core/pica.h
@@ -578,7 +578,17 @@ struct Regs {
     }
 
     struct {
-        INSERT_PADDING_WORDS(0x6);
+        INSERT_PADDING_WORDS(0x3);
+
+        union {
+            BitField<0, 4, u32> allow_color_write; // 0 = disable, else enable
+        };
+
+        INSERT_PADDING_WORDS(0x1);
+
+        union {
+            BitField<0, 2, u32> allow_depth_stencil_write; // 0 = disable, else enable
+        };
 
         DepthFormat depth_format; // TODO: Should be a BitField!
         BitField<16, 3, ColorFormat> color_format;
diff --git a/src/video_core/primitive_assembly.cpp b/src/video_core/primitive_assembly.cpp
index 0061690f1..ff3e2b862 100644
--- a/src/video_core/primitive_assembly.cpp
+++ b/src/video_core/primitive_assembly.cpp
@@ -68,7 +68,5 @@ void PrimitiveAssembler<VertexType>::Reconfigure(Regs::TriangleTopology topology
 // explicitly instantiate use cases
 template
 struct PrimitiveAssembler<Shader::OutputVertex>;
-template
-struct PrimitiveAssembler<DebugUtils::GeometryDumper::Vertex>;
 
 } // namespace
diff --git a/src/video_core/rasterizer.cpp b/src/video_core/rasterizer.cpp
index fd02aa652..0434ad05a 100644
--- a/src/video_core/rasterizer.cpp
+++ b/src/video_core/rasterizer.cpp
@@ -809,7 +809,8 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
 
             auto UpdateStencil = [stencil_test, x, y, &old_stencil](Pica::Regs::StencilAction action) {
                 u8 new_stencil = PerformStencilAction(action, old_stencil, stencil_test.reference_value);
-                SetStencil(x >> 4, y >> 4, (new_stencil & stencil_test.write_mask) | (old_stencil & ~stencil_test.write_mask));
+                if (g_state.regs.framebuffer.allow_depth_stencil_write != 0)
+                    SetStencil(x >> 4, y >> 4, (new_stencil & stencil_test.write_mask) | (old_stencil & ~stencil_test.write_mask));
             };
 
             if (stencil_action_enable) {
@@ -909,7 +910,7 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
                 }
             }
 
-            if (output_merger.depth_write_enable)
+            if (regs.framebuffer.allow_depth_stencil_write != 0 && output_merger.depth_write_enable)
                 SetDepth(x >> 4, y >> 4, z);
 
             // The stencil depth_pass action is executed even if depth testing is disabled
@@ -922,92 +923,72 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
             if (output_merger.alphablend_enable) {
                 auto params = output_merger.alpha_blending;
 
-                auto LookupFactorRGB = [&](Regs::BlendFactor factor) -> Math::Vec3<u8> {
+                auto LookupFactor = [&](unsigned channel, Regs::BlendFactor factor) -> u8 {
+                    DEBUG_ASSERT(channel < 4);
+
+                    const Math::Vec4<u8> blend_const = {
+                        static_cast<u8>(output_merger.blend_const.r),
+                        static_cast<u8>(output_merger.blend_const.g),
+                        static_cast<u8>(output_merger.blend_const.b),
+                        static_cast<u8>(output_merger.blend_const.a)
+                    };
+
                     switch (factor) {
-                    case Regs::BlendFactor::Zero :
-                        return Math::Vec3<u8>(0, 0, 0);
+                    case Regs::BlendFactor::Zero:
+                        return 0;
 
-                    case Regs::BlendFactor::One :
-                        return Math::Vec3<u8>(255, 255, 255);
+                    case Regs::BlendFactor::One:
+                        return 255;
 
                     case Regs::BlendFactor::SourceColor:
-                        return combiner_output.rgb();
+                        return combiner_output[channel];
 
                     case Regs::BlendFactor::OneMinusSourceColor:
-                        return Math::Vec3<u8>(255 - combiner_output.r(), 255 - combiner_output.g(), 255 - combiner_output.b());
+                        return 255 - combiner_output[channel];
 
                     case Regs::BlendFactor::DestColor:
-                        return dest.rgb();
+                        return dest[channel];
 
                     case Regs::BlendFactor::OneMinusDestColor:
-                        return Math::Vec3<u8>(255 - dest.r(), 255 - dest.g(), 255 - dest.b());
+                        return 255 - dest[channel];
 
                     case Regs::BlendFactor::SourceAlpha:
-                        return Math::Vec3<u8>(combiner_output.a(), combiner_output.a(), combiner_output.a());
+                        return combiner_output.a();
 
                     case Regs::BlendFactor::OneMinusSourceAlpha:
-                        return Math::Vec3<u8>(255 - combiner_output.a(), 255 - combiner_output.a(), 255 - combiner_output.a());
+                        return 255 - combiner_output.a();
 
                     case Regs::BlendFactor::DestAlpha:
-                        return Math::Vec3<u8>(dest.a(), dest.a(), dest.a());
+                        return dest.a();
 
                     case Regs::BlendFactor::OneMinusDestAlpha:
-                        return Math::Vec3<u8>(255 - dest.a(), 255 - dest.a(), 255 - dest.a());
+                        return 255 - dest.a();
 
                     case Regs::BlendFactor::ConstantColor:
-                        return Math::Vec3<u8>(output_merger.blend_const.r, output_merger.blend_const.g, output_merger.blend_const.b);
+                        return blend_const[channel];
 
                     case Regs::BlendFactor::OneMinusConstantColor:
-                        return Math::Vec3<u8>(255 - output_merger.blend_const.r, 255 - output_merger.blend_const.g, 255 - output_merger.blend_const.b);
+                        return 255 - blend_const[channel];
 
                     case Regs::BlendFactor::ConstantAlpha:
-                        return Math::Vec3<u8>(output_merger.blend_const.a, output_merger.blend_const.a, output_merger.blend_const.a);
+                        return blend_const.a();
 
                     case Regs::BlendFactor::OneMinusConstantAlpha:
-                        return Math::Vec3<u8>(255 - output_merger.blend_const.a, 255 - output_merger.blend_const.a, 255 - output_merger.blend_const.a);
-
-                    default:
-                        LOG_CRITICAL(HW_GPU, "Unknown color blend factor %x", factor);
-                        UNIMPLEMENTED();
-                        break;
-                    }
-
-                    return {};
-                };
-
-                auto LookupFactorA = [&](Regs::BlendFactor factor) -> u8 {
-                    switch (factor) {
-                    case Regs::BlendFactor::Zero:
-                        return 0;
-
-                    case Regs::BlendFactor::One:
-                        return 255;
-
-                    case Regs::BlendFactor::SourceAlpha:
-                        return combiner_output.a();
-
-                    case Regs::BlendFactor::OneMinusSourceAlpha:
-                        return 255 - combiner_output.a();
+                        return 255 - blend_const.a();
 
-                    case Regs::BlendFactor::DestAlpha:
-                        return dest.a();
-
-                    case Regs::BlendFactor::OneMinusDestAlpha:
-                        return 255 - dest.a();
-
-                    case Regs::BlendFactor::ConstantAlpha:
-                        return output_merger.blend_const.a;
-
-                    case Regs::BlendFactor::OneMinusConstantAlpha:
-                        return 255 - output_merger.blend_const.a;
+                    case Regs::BlendFactor::SourceAlphaSaturate:
+                        // Returns 1.0 for the alpha channel
+                        if (channel == 3)
+                            return 255;
+                        return std::min(combiner_output.a(), static_cast<u8>(255 - dest.a()));
 
                     default:
-                        LOG_CRITICAL(HW_GPU, "Unknown alpha blend factor %x", factor);
+                        LOG_CRITICAL(HW_GPU, "Unknown blend factor %x", factor);
                         UNIMPLEMENTED();
                         break;
                     }
 
-                    return {};
+                    return combiner_output[channel];
                 };
 
                 static auto EvaluateBlendEquation = [](const Math::Vec4<u8>& src, const Math::Vec4<u8>& srcfactor,
@@ -1059,10 +1040,15 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
                                     MathUtil::Clamp(result.a(), 0, 255));
                 };
 
-                auto srcfactor = Math::MakeVec(LookupFactorRGB(params.factor_source_rgb),
-                                               LookupFactorA(params.factor_source_a));
-                auto dstfactor = Math::MakeVec(LookupFactorRGB(params.factor_dest_rgb),
-                                               LookupFactorA(params.factor_dest_a));
+                auto srcfactor = Math::MakeVec(LookupFactor(0, params.factor_source_rgb),
+                                               LookupFactor(1, params.factor_source_rgb),
+                                               LookupFactor(2, params.factor_source_rgb),
+                                               LookupFactor(3, params.factor_source_a));
+
+                auto dstfactor = Math::MakeVec(LookupFactor(0, params.factor_dest_rgb),
+                                               LookupFactor(1, params.factor_dest_rgb),
+                                               LookupFactor(2, params.factor_dest_rgb),
+                                               LookupFactor(3, params.factor_dest_a));
 
                 blend_output     = EvaluateBlendEquation(combiner_output, srcfactor, dest, dstfactor, params.blend_equation_rgb);
                 blend_output.a() = EvaluateBlendEquation(combiner_output, srcfactor, dest, dstfactor, params.blend_equation_a).a();
@@ -1133,7 +1119,8 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
                 output_merger.alpha_enable ? blend_output.a() : dest.a()
             };
 
-            DrawPixel(x >> 4, y >> 4, result);
+            if (regs.framebuffer.allow_color_write != 0)
+                DrawPixel(x >> 4, y >> 4, result);
         }
     }
 }
diff --git a/src/video_core/renderer_base.cpp b/src/video_core/renderer_base.cpp
index 6467ff723..101f84eb9 100644
--- a/src/video_core/renderer_base.cpp
+++ b/src/video_core/renderer_base.cpp
@@ -4,8 +4,6 @@
 
 #include <memory>
 
-#include "common/make_unique.h"
-
 #include "core/settings.h"
 
 #include "video_core/renderer_base.h"
@@ -19,9 +17,9 @@ void RendererBase::RefreshRasterizerSetting() {
         opengl_rasterizer_active = hw_renderer_enabled;
 
         if (hw_renderer_enabled) {
-            rasterizer = Common::make_unique<RasterizerOpenGL>();
+            rasterizer = std::make_unique<RasterizerOpenGL>();
         } else {
-            rasterizer = Common::make_unique<VideoCore::SWRasterizer>();
+            rasterizer = std::make_unique<VideoCore::SWRasterizer>();
         }
         rasterizer->InitObjects();
         rasterizer->Reset();
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 1fadcf5ae..6ca9f45e2 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -9,7 +9,6 @@
 
 #include "common/color.h"
 #include "common/file_util.h"
-#include "common/make_unique.h"
 #include "common/math_util.h"
 #include "common/microprofile.h"
 #include "common/profiler.h"
@@ -140,8 +139,9 @@ void RasterizerOpenGL::InitObjects() {
     }
     state.Apply();
 
-    ASSERT_MSG(glCheckFramebufferStatus(GL_FRAMEBUFFER) == GL_FRAMEBUFFER_COMPLETE,
-               "OpenGL rasterizer framebuffer setup failed, status %X", glCheckFramebufferStatus(GL_FRAMEBUFFER));
+    GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+    ASSERT_MSG(status == GL_FRAMEBUFFER_COMPLETE,
+               "OpenGL rasterizer framebuffer setup failed, status %X", status);
 }
 
 void RasterizerOpenGL::Reset() {
@@ -153,6 +153,9 @@ void RasterizerOpenGL::Reset() {
     SyncLogicOp();
     SyncStencilTest();
     SyncDepthTest();
+    SyncColorWriteMask();
+    SyncStencilWriteMask();
+    SyncDepthWriteMask();
 
     SetShader();
 
@@ -268,15 +271,36 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
         state.draw.shader_dirty = true;
         break;
 
-    // Stencil test
+    // Sync GL stencil test + stencil write mask
+    // (Pica stencil test function register also contains a stencil write mask)
     case PICA_REG_INDEX(output_merger.stencil_test.raw_func):
+        SyncStencilTest();
+        SyncStencilWriteMask();
+        break;
     case PICA_REG_INDEX(output_merger.stencil_test.raw_op):
+    case PICA_REG_INDEX(framebuffer.depth_format):
         SyncStencilTest();
         break;
 
-    // Depth test
+    // Sync GL depth test + depth and color write mask
+    // (Pica depth test function register also contains a depth and color write mask)
     case PICA_REG_INDEX(output_merger.depth_test_enable):
         SyncDepthTest();
+        SyncDepthWriteMask();
+        SyncColorWriteMask();
+        break;
+
+    // Sync GL depth and stencil write mask
+    // (This is a dedicated combined depth / stencil write-enable register)
+    case PICA_REG_INDEX(framebuffer.allow_depth_stencil_write):
+        SyncDepthWriteMask();
+        SyncStencilWriteMask();
+        break;
+
+    // Sync GL color write mask
+    // (This is a dedicated color write-enable register)
+    case PICA_REG_INDEX(framebuffer.allow_color_write):
+        SyncColorWriteMask();
         break;
 
     // Logic op
@@ -677,7 +701,7 @@ void RasterizerOpenGL::ReconfigureDepthTexture(DepthTextureInfo& texture, Pica::
 
 void RasterizerOpenGL::SetShader() {
     PicaShaderConfig config = PicaShaderConfig::CurrentConfig();
-    std::unique_ptr<PicaShader> shader = Common::make_unique<PicaShader>();
+    std::unique_ptr<PicaShader> shader = std::make_unique<PicaShader>();
 
     // Find (or generate) the GLSL shader for the current TEV state
     auto cached_shader = shader_cache.find(config);
@@ -808,6 +832,10 @@ void RasterizerOpenGL::SyncFramebuffer() {
 
         ReloadDepthBuffer();
     }
+
+    GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+    ASSERT_MSG(status == GL_FRAMEBUFFER_COMPLETE,
+               "OpenGL rasterizer framebuffer setup failed, status %X", status);
 }
 
 void RasterizerOpenGL::SyncCullMode() {
@@ -876,13 +904,39 @@ void RasterizerOpenGL::SyncLogicOp() {
     state.logic_op = PicaToGL::LogicOp(Pica::g_state.regs.output_merger.logic_op);
 }
 
+void RasterizerOpenGL::SyncColorWriteMask() {
+    const auto& regs = Pica::g_state.regs;
+
+    auto IsColorWriteEnabled = [&](u32 value) {
+        return (regs.framebuffer.allow_color_write != 0 && value != 0) ? GL_TRUE : GL_FALSE;
+    };
+
+    state.color_mask.red_enabled = IsColorWriteEnabled(regs.output_merger.red_enable);
+    state.color_mask.green_enabled = IsColorWriteEnabled(regs.output_merger.green_enable);
+    state.color_mask.blue_enabled = IsColorWriteEnabled(regs.output_merger.blue_enable);
+    state.color_mask.alpha_enabled = IsColorWriteEnabled(regs.output_merger.alpha_enable);
+}
+
+void RasterizerOpenGL::SyncStencilWriteMask() {
+    const auto& regs = Pica::g_state.regs;
+    state.stencil.write_mask = (regs.framebuffer.allow_depth_stencil_write != 0)
+                             ? static_cast<GLuint>(regs.output_merger.stencil_test.write_mask)
+                             : 0;
+}
+
+void RasterizerOpenGL::SyncDepthWriteMask() {
+    const auto& regs = Pica::g_state.regs;
+    state.depth.write_mask = (regs.framebuffer.allow_depth_stencil_write != 0 && regs.output_merger.depth_write_enable)
+                           ? GL_TRUE
+                           : GL_FALSE;
+}
+
 void RasterizerOpenGL::SyncStencilTest() {
     const auto& regs = Pica::g_state.regs;
     state.stencil.test_enabled = regs.output_merger.stencil_test.enable && regs.framebuffer.depth_format == Pica::Regs::DepthFormat::D24S8;
     state.stencil.test_func = PicaToGL::CompareFunc(regs.output_merger.stencil_test.func);
     state.stencil.test_ref = regs.output_merger.stencil_test.reference_value;
     state.stencil.test_mask = regs.output_merger.stencil_test.input_mask;
-    state.stencil.write_mask = regs.output_merger.stencil_test.write_mask;
     state.stencil.action_stencil_fail = PicaToGL::StencilOp(regs.output_merger.stencil_test.action_stencil_fail);
     state.stencil.action_depth_fail = PicaToGL::StencilOp(regs.output_merger.stencil_test.action_depth_fail);
     state.stencil.action_depth_pass = PicaToGL::StencilOp(regs.output_merger.stencil_test.action_depth_pass);
@@ -894,11 +948,6 @@ void RasterizerOpenGL::SyncDepthTest() {
                                regs.output_merger.depth_write_enable == 1;
     state.depth.test_func = regs.output_merger.depth_test_enable == 1 ?
                             PicaToGL::CompareFunc(regs.output_merger.depth_test_func) : GL_ALWAYS;
-    state.color_mask.red_enabled = regs.output_merger.red_enable;
-    state.color_mask.green_enabled = regs.output_merger.green_enable;
-    state.color_mask.blue_enabled = regs.output_merger.blue_enable;
-    state.color_mask.alpha_enabled = regs.output_merger.alpha_enable;
-    state.depth.write_mask = regs.output_merger.depth_write_enable ? GL_TRUE : GL_FALSE;
 }
 
 void RasterizerOpenGL::SyncCombinerColor() {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index fc85aa3ff..390349a0c 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -344,6 +344,15 @@ private:
     /// Syncs the logic op states to match the PICA register
     void SyncLogicOp();
 
+    /// Syncs the color write mask to match the PICA register state
+    void SyncColorWriteMask();
+
+    /// Syncs the stencil write mask to match the PICA register state
+    void SyncStencilWriteMask();
+
+    /// Syncs the depth write mask to match the PICA register state
+    void SyncDepthWriteMask();
+
     /// Syncs the stencil test states to match the PICA register
     void SyncStencilTest();
 
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index a9ad46fe0..1323c12e4 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -2,8 +2,9 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <memory>
+
 #include "common/hash.h"
-#include "common/make_unique.h"
 #include "common/math_util.h"
 #include "common/microprofile.h"
 #include "common/vector_math.h"
@@ -29,7 +30,7 @@ void RasterizerCacheOpenGL::LoadAndBindTexture(OpenGLState &state, unsigned text
     } else {
         MICROPROFILE_SCOPE(OpenGL_TextureUpload);
 
-        std::unique_ptr<CachedTexture> new_texture = Common::make_unique<CachedTexture>();
+        std::unique_ptr<CachedTexture> new_texture = std::make_unique<CachedTexture>();
 
         new_texture->texture.Create();
         state.texture_units[texture_unit].texture_2d = new_texture->texture.handle;
diff --git a/src/video_core/renderer_opengl/pica_to_gl.h b/src/video_core/renderer_opengl/pica_to_gl.h
index 3d6c4e9e5..fd3617d77 100644
--- a/src/video_core/renderer_opengl/pica_to_gl.h
+++ b/src/video_core/renderer_opengl/pica_to_gl.h
@@ -22,7 +22,7 @@ inline GLenum TextureFilterMode(Pica::Regs::TextureConfig::TextureFilter mode) {
     };
 
     // Range check table for input
-    if (mode >= ARRAY_SIZE(filter_mode_table)) {
+    if (static_cast<size_t>(mode) >= ARRAY_SIZE(filter_mode_table)) {
         LOG_CRITICAL(Render_OpenGL, "Unknown texture filtering mode %d", mode);
         UNREACHABLE();
 
@@ -51,7 +51,7 @@ inline GLenum WrapMode(Pica::Regs::TextureConfig::WrapMode mode) {
     };
 
     // Range check table for input
-    if (mode >= ARRAY_SIZE(wrap_mode_table)) {
+    if (static_cast<size_t>(mode) >= ARRAY_SIZE(wrap_mode_table)) {
         LOG_CRITICAL(Render_OpenGL, "Unknown texture wrap mode %d", mode);
         UNREACHABLE();
 
@@ -91,7 +91,7 @@ inline GLenum BlendFunc(Pica::Regs::BlendFactor factor) {
     };
 
     // Range check table for input
-    if ((unsigned)factor >= ARRAY_SIZE(blend_func_table)) {
+    if (static_cast<size_t>(factor) >= ARRAY_SIZE(blend_func_table)) {
         LOG_CRITICAL(Render_OpenGL, "Unknown blend factor %d", factor);
         UNREACHABLE();
 
@@ -122,7 +122,7 @@ inline GLenum LogicOp(Pica::Regs::LogicOp op) {
     };
 
     // Range check table for input
-    if ((unsigned)op >= ARRAY_SIZE(logic_op_table)) {
+    if (static_cast<size_t>(op) >= ARRAY_SIZE(logic_op_table)) {
         LOG_CRITICAL(Render_OpenGL, "Unknown logic op %d", op);
         UNREACHABLE();
 
@@ -145,7 +145,7 @@ inline GLenum CompareFunc(Pica::Regs::CompareFunc func) {
     };
 
     // Range check table for input
-    if ((unsigned)func >= ARRAY_SIZE(compare_func_table)) {
+    if (static_cast<size_t>(func) >= ARRAY_SIZE(compare_func_table)) {
         LOG_CRITICAL(Render_OpenGL, "Unknown compare function %d", func);
         UNREACHABLE();
 
@@ -168,7 +168,7 @@ inline GLenum StencilOp(Pica::Regs::StencilAction action) {
     };
 
     // Range check table for input
-    if ((unsigned)action >= ARRAY_SIZE(stencil_op_table)) {
+    if (static_cast<size_t>(action) >= ARRAY_SIZE(stencil_op_table)) {
         LOG_CRITICAL(Render_OpenGL, "Unknown stencil op %d", action);
         UNREACHABLE();
 
diff --git a/src/video_core/shader/shader.cpp b/src/video_core/shader/shader.cpp
index eb1db0778..75301accd 100644
--- a/src/video_core/shader/shader.cpp
+++ b/src/video_core/shader/shader.cpp
@@ -8,7 +8,6 @@
 #include <boost/range/algorithm/fill.hpp>
 
 #include "common/hash.h"
-#include "common/make_unique.h"
 #include "common/microprofile.h"
 #include "common/profiler.h"
 
@@ -29,36 +28,24 @@ namespace Pica {
 namespace Shader {
 
 #ifdef ARCHITECTURE_x86_64
-static std::unordered_map<u64, CompiledShader*> shader_map;
-static JitCompiler jit;
-static CompiledShader* jit_shader;
-
-static void ClearCache() {
-    shader_map.clear();
-    jit.Clear();
-    LOG_INFO(HW_GPU, "Shader JIT cache cleared");
-}
+static std::unordered_map<u64, std::unique_ptr<JitShader>> shader_map;
+static const JitShader* jit_shader;
 #endif // ARCHITECTURE_x86_64
 
-void Setup(UnitState<false>& state) {
+void Setup() {
 #ifdef ARCHITECTURE_x86_64
     if (VideoCore::g_shader_jit_enabled) {
         u64 cache_key = (Common::ComputeHash64(&g_state.vs.program_code, sizeof(g_state.vs.program_code)) ^
-            Common::ComputeHash64(&g_state.vs.swizzle_data, sizeof(g_state.vs.swizzle_data)) ^
-            g_state.regs.vs.main_offset);
+            Common::ComputeHash64(&g_state.vs.swizzle_data, sizeof(g_state.vs.swizzle_data)));
 
         auto iter = shader_map.find(cache_key);
         if (iter != shader_map.end()) {
-            jit_shader = iter->second;
+            jit_shader = iter->second.get();
         } else {
-            // Check if remaining JIT code space is enough for at least one more (massive) shader
-            if (jit.GetSpaceLeft() < jit_shader_size) {
-                // If not, clear the cache of all previously compiled shaders
-                ClearCache();
-            }
-
-            jit_shader = jit.Compile();
-            shader_map.emplace(cache_key, jit_shader);
+            auto shader = std::make_unique<JitShader>();
+            shader->Compile();
+            jit_shader = shader.get();
+            shader_map[cache_key] = std::move(shader);
         }
     }
 #endif // ARCHITECTURE_x86_64
@@ -66,7 +53,7 @@ void Setup(UnitState<false>& state) {
 
 void Shutdown() {
 #ifdef ARCHITECTURE_x86_64
-    ClearCache();
+    shader_map.clear();
 #endif // ARCHITECTURE_x86_64
 }
 
@@ -110,7 +97,7 @@ OutputVertex Run(UnitState<false>& state, const InputVertex& input, int num_attr
 
 #ifdef ARCHITECTURE_x86_64
     if (VideoCore::g_shader_jit_enabled)
-        jit_shader(&state.registers);
+        jit_shader->Run(&state.registers, g_state.regs.vs.main_offset);
     else
         RunInterpreter(state);
 #else
diff --git a/src/video_core/shader/shader.h b/src/video_core/shader/shader.h
index 7af8f1fa1..9c5bd97bd 100644
--- a/src/video_core/shader/shader.h
+++ b/src/video_core/shader/shader.h
@@ -339,9 +339,8 @@ struct UnitState {
 /**
  * Performs any shader unit setup that only needs to happen once per shader (as opposed to once per
  * vertex, which would happen within the `Run` function).
- * @param state Shader unit state, must be setup per shader and per shader unit
  */
-void Setup(UnitState<false>& state);
+void Setup();
 
 /// Performs any cleanup when the emulator is shutdown
 void Shutdown();
diff --git a/src/video_core/shader/shader_jit_x64.cpp b/src/video_core/shader/shader_jit_x64.cpp
index dffe051ef..b47d3beda 100644
--- a/src/video_core/shader/shader_jit_x64.cpp
+++ b/src/video_core/shader/shader_jit_x64.cpp
@@ -2,6 +2,7 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <algorithm>
 #include <smmintrin.h>
 
 #include "common/x64/abi.h"
@@ -19,73 +20,73 @@ namespace Shader {
 
 using namespace Gen;
 
-typedef void (JitCompiler::*JitFunction)(Instruction instr);
+typedef void (JitShader::*JitFunction)(Instruction instr);
 
 const JitFunction instr_table[64] = {
-    &JitCompiler::Compile_ADD,      // add
-    &JitCompiler::Compile_DP3,      // dp3
-    &JitCompiler::Compile_DP4,      // dp4
-    &JitCompiler::Compile_DPH,      // dph
+    &JitShader::Compile_ADD,        // add
+    &JitShader::Compile_DP3,        // dp3
+    &JitShader::Compile_DP4,        // dp4
+    &JitShader::Compile_DPH,        // dph
     nullptr,                        // unknown
-    &JitCompiler::Compile_EX2,      // ex2
-    &JitCompiler::Compile_LG2,      // lg2
+    &JitShader::Compile_EX2,        // ex2
+    &JitShader::Compile_LG2,        // lg2
     nullptr,                        // unknown
-    &JitCompiler::Compile_MUL,      // mul
-    &JitCompiler::Compile_SGE,      // sge
-    &JitCompiler::Compile_SLT,      // slt
-    &JitCompiler::Compile_FLR,      // flr
-    &JitCompiler::Compile_MAX,      // max
-    &JitCompiler::Compile_MIN,      // min
-    &JitCompiler::Compile_RCP,      // rcp
-    &JitCompiler::Compile_RSQ,      // rsq
+    &JitShader::Compile_MUL,        // mul
+    &JitShader::Compile_SGE,        // sge
+    &JitShader::Compile_SLT,        // slt
+    &JitShader::Compile_FLR,        // flr
+    &JitShader::Compile_MAX,        // max
+    &JitShader::Compile_MIN,        // min
+    &JitShader::Compile_RCP,        // rcp
+    &JitShader::Compile_RSQ,        // rsq
     nullptr,                        // unknown
     nullptr,                        // unknown
-    &JitCompiler::Compile_MOVA,     // mova
-    &JitCompiler::Compile_MOV,      // mov
+    &JitShader::Compile_MOVA,       // mova
+    &JitShader::Compile_MOV,        // mov
     nullptr,                        // unknown
     nullptr,                        // unknown
     nullptr,                        // unknown
     nullptr,                        // unknown
-    &JitCompiler::Compile_DPH,      // dphi
+    &JitShader::Compile_DPH,        // dphi
     nullptr,                        // unknown
-    &JitCompiler::Compile_SGE,      // sgei
-    &JitCompiler::Compile_SLT,      // slti
+    &JitShader::Compile_SGE,        // sgei
+    &JitShader::Compile_SLT,        // slti
     nullptr,                        // unknown
     nullptr,                        // unknown
     nullptr,                        // unknown
     nullptr,                        // unknown
     nullptr,                        // unknown
-    &JitCompiler::Compile_NOP,      // nop
-    &JitCompiler::Compile_END,      // end
+    &JitShader::Compile_NOP,        // nop
+    &JitShader::Compile_END,        // end
     nullptr,                        // break
-    &JitCompiler::Compile_CALL,     // call
-    &JitCompiler::Compile_CALLC,    // callc
-    &JitCompiler::Compile_CALLU,    // callu
-    &JitCompiler::Compile_IF,       // ifu
-    &JitCompiler::Compile_IF,       // ifc
-    &JitCompiler::Compile_LOOP,     // loop
+    &JitShader::Compile_CALL,       // call
+    &JitShader::Compile_CALLC,      // callc
+    &JitShader::Compile_CALLU,      // callu
+    &JitShader::Compile_IF,         // ifu
+    &JitShader::Compile_IF,         // ifc
+    &JitShader::Compile_LOOP,       // loop
     nullptr,                        // emit
     nullptr,                        // sete
-    &JitCompiler::Compile_JMP,      // jmpc
-    &JitCompiler::Compile_JMP,      // jmpu
-    &JitCompiler::Compile_CMP,      // cmp
-    &JitCompiler::Compile_CMP,      // cmp
-    &JitCompiler::Compile_MAD,      // madi
-    &JitCompiler::Compile_MAD,      // madi
-    &JitCompiler::Compile_MAD,      // madi
-    &JitCompiler::Compile_MAD,      // madi
-    &JitCompiler::Compile_MAD,      // madi
-    &JitCompiler::Compile_MAD,      // madi
-    &JitCompiler::Compile_MAD,      // madi
-    &JitCompiler::Compile_MAD,      // madi
-    &JitCompiler::Compile_MAD,      // mad
-    &JitCompiler::Compile_MAD,      // mad
-    &JitCompiler::Compile_MAD,      // mad
-    &JitCompiler::Compile_MAD,      // mad
-    &JitCompiler::Compile_MAD,      // mad
-    &JitCompiler::Compile_MAD,      // mad
-    &JitCompiler::Compile_MAD,      // mad
-    &JitCompiler::Compile_MAD,      // mad
+    &JitShader::Compile_JMP,        // jmpc
+    &JitShader::Compile_JMP,        // jmpu
+    &JitShader::Compile_CMP,        // cmp
+    &JitShader::Compile_CMP,        // cmp
+    &JitShader::Compile_MAD,        // madi
+    &JitShader::Compile_MAD,        // madi
+    &JitShader::Compile_MAD,        // madi
+    &JitShader::Compile_MAD,        // madi
+    &JitShader::Compile_MAD,        // madi
+    &JitShader::Compile_MAD,        // madi
+    &JitShader::Compile_MAD,        // madi
+    &JitShader::Compile_MAD,        // madi
+    &JitShader::Compile_MAD,        // mad
+    &JitShader::Compile_MAD,        // mad
+    &JitShader::Compile_MAD,        // mad
+    &JitShader::Compile_MAD,        // mad
+    &JitShader::Compile_MAD,        // mad
+    &JitShader::Compile_MAD,        // mad
+    &JitShader::Compile_MAD,        // mad
+    &JitShader::Compile_MAD,        // mad
 };
 
 // The following is used to alias some commonly used registers. Generally, RAX-RDX and XMM0-XMM3 can
@@ -138,13 +139,32 @@ static const u8 NO_SRC_REG_SWIZZLE = 0x1b;
 static const u8 NO_DEST_REG_MASK = 0xf;
 
 /**
+ * Get the vertex shader instruction for a given offset in the current shader program
+ * @param offset Offset in the current shader program of the instruction
+ * @return Instruction at the specified offset
+ */
+static Instruction GetVertexShaderInstruction(size_t offset) {
+    return { g_state.vs.program_code[offset] };
+}
+
+static void LogCritical(const char* msg) {
+    LOG_CRITICAL(HW_GPU, msg);
+}
+
+void JitShader::Compile_Assert(bool condition, const char* msg) {
+    if (!condition) {
+        ABI_CallFunctionP(reinterpret_cast<const void*>(LogCritical), const_cast<char*>(msg));
+    }
+}
+
+/**
  * Loads and swizzles a source register into the specified XMM register.
  * @param instr VS instruction, used for determining how to load the source register
  * @param src_num Number indicating which source register to load (1 = src1, 2 = src2, 3 = src3)
  * @param src_reg SourceRegister object corresponding to the source register to load
  * @param dest Destination XMM register to store the loaded, swizzled source register
  */
-void JitCompiler::Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg, X64Reg dest) {
+void JitShader::Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg, X64Reg dest) {
     X64Reg src_ptr;
     size_t src_offset;
 
@@ -216,7 +236,7 @@ void JitCompiler::Compile_SwizzleSrc(Instruction instr, unsigned src_num, Source
     }
 }
 
-void JitCompiler::Compile_DestEnable(Instruction instr,X64Reg src) {
+void JitShader::Compile_DestEnable(Instruction instr,X64Reg src) {
     DestRegister dest;
     unsigned operand_desc_id;
     if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
@@ -263,7 +283,7 @@ void JitCompiler::Compile_DestEnable(Instruction instr,X64Reg src) {
     }
 }
 
-void JitCompiler::Compile_SanitizedMul(Gen::X64Reg src1, Gen::X64Reg src2, Gen::X64Reg scratch) {
+void JitShader::Compile_SanitizedMul(Gen::X64Reg src1, Gen::X64Reg src2, Gen::X64Reg scratch) {
     MOVAPS(scratch, R(src1));
     CMPPS(scratch, R(src2), CMP_ORD);
 
@@ -276,7 +296,7 @@ void JitCompiler::Compile_SanitizedMul(Gen::X64Reg src1, Gen::X64Reg src2, Gen::
     ANDPS(src1, R(scratch));
 }
 
-void JitCompiler::Compile_EvaluateCondition(Instruction instr) {
+void JitShader::Compile_EvaluateCondition(Instruction instr) {
     // Note: NXOR is used below to check for equality
     switch (instr.flow_control.op) {
     case Instruction::FlowControlType::Or:
@@ -307,23 +327,23 @@ void JitCompiler::Compile_EvaluateCondition(Instruction instr) {
     }
 }
 
-void JitCompiler::Compile_UniformCondition(Instruction instr) {
+void JitShader::Compile_UniformCondition(Instruction instr) {
     int offset = offsetof(decltype(g_state.vs.uniforms), b) + (instr.flow_control.bool_uniform_id * sizeof(bool));
     CMP(sizeof(bool) * 8, MDisp(UNIFORMS, offset), Imm8(0));
 }
 
-BitSet32 JitCompiler::PersistentCallerSavedRegs() {
+BitSet32 JitShader::PersistentCallerSavedRegs() {
     return persistent_regs & ABI_ALL_CALLER_SAVED;
 }
 
-void JitCompiler::Compile_ADD(Instruction instr) {
+void JitShader::Compile_ADD(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
     ADDPS(SRC1, R(SRC2));
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_DP3(Instruction instr) {
+void JitShader::Compile_DP3(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
 
@@ -342,7 +362,7 @@ void JitCompiler::Compile_DP3(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_DP4(Instruction instr) {
+void JitShader::Compile_DP4(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
 
@@ -359,7 +379,7 @@ void JitCompiler::Compile_DP4(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_DPH(Instruction instr) {
+void JitShader::Compile_DPH(Instruction instr) {
     if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::DPHI) {
         Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
         Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
@@ -391,7 +411,7 @@ void JitCompiler::Compile_DPH(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_EX2(Instruction instr) {
+void JitShader::Compile_EX2(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     MOVSS(XMM0, R(SRC1));
 
@@ -404,7 +424,7 @@ void JitCompiler::Compile_EX2(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_LG2(Instruction instr) {
+void JitShader::Compile_LG2(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     MOVSS(XMM0, R(SRC1));
 
@@ -417,14 +437,14 @@ void JitCompiler::Compile_LG2(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_MUL(Instruction instr) {
+void JitShader::Compile_MUL(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
     Compile_SanitizedMul(SRC1, SRC2, SCRATCH);
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_SGE(Instruction instr) {
+void JitShader::Compile_SGE(Instruction instr) {
     if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::SGEI) {
         Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
         Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
@@ -439,7 +459,7 @@ void JitCompiler::Compile_SGE(Instruction instr) {
     Compile_DestEnable(instr, SRC2);
 }
 
-void JitCompiler::Compile_SLT(Instruction instr) {
+void JitShader::Compile_SLT(Instruction instr) {
     if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::SLTI) {
         Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
         Compile_SwizzleSrc(instr, 2, instr.common.src2i, SRC2);
@@ -454,7 +474,7 @@ void JitCompiler::Compile_SLT(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_FLR(Instruction instr) {
+void JitShader::Compile_FLR(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
 
     if (Common::GetCPUCaps().sse4_1) {
@@ -467,7 +487,7 @@ void JitCompiler::Compile_FLR(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_MAX(Instruction instr) {
+void JitShader::Compile_MAX(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
     // SSE semantics match PICA200 ones: In case of NaN, SRC2 is returned.
@@ -475,7 +495,7 @@ void JitCompiler::Compile_MAX(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_MIN(Instruction instr) {
+void JitShader::Compile_MIN(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
     // SSE semantics match PICA200 ones: In case of NaN, SRC2 is returned.
@@ -483,7 +503,7 @@ void JitCompiler::Compile_MIN(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_MOVA(Instruction instr) {
+void JitShader::Compile_MOVA(Instruction instr) {
     SwizzlePattern swiz = { g_state.vs.swizzle_data[instr.common.operand_desc_id] };
 
     if (!swiz.DestComponentEnabled(0) && !swiz.DestComponentEnabled(1)) {
@@ -528,12 +548,12 @@ void JitCompiler::Compile_MOVA(Instruction instr) {
     }
 }
 
-void JitCompiler::Compile_MOV(Instruction instr) {
+void JitShader::Compile_MOV(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_RCP(Instruction instr) {
+void JitShader::Compile_RCP(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
 
     // TODO(bunnei): RCPSS is a pretty rough approximation, this might cause problems if Pica
@@ -544,7 +564,7 @@ void JitCompiler::Compile_RCP(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_RSQ(Instruction instr) {
+void JitShader::Compile_RSQ(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
 
     // TODO(bunnei): RSQRTSS is a pretty rough approximation, this might cause problems if Pica
@@ -555,36 +575,41 @@ void JitCompiler::Compile_RSQ(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_NOP(Instruction instr) {
+void JitShader::Compile_NOP(Instruction instr) {
 }
 
-void JitCompiler::Compile_END(Instruction instr) {
+void JitShader::Compile_END(Instruction instr) {
     ABI_PopRegistersAndAdjustStack(ABI_ALL_CALLEE_SAVED, 8);
     RET();
 }
 
-void JitCompiler::Compile_CALL(Instruction instr) {
-    unsigned offset = instr.flow_control.dest_offset;
-    while (offset < (instr.flow_control.dest_offset + instr.flow_control.num_instructions)) {
-        Compile_NextInstr(&offset);
-    }
+void JitShader::Compile_CALL(Instruction instr) {
+    // Push offset of the return
+    PUSH(64, Imm32(instr.flow_control.dest_offset + instr.flow_control.num_instructions));
+
+    // Call the subroutine
+    FixupBranch b = CALL();
+    fixup_branches.push_back({ b, instr.flow_control.dest_offset });
+
+    // Skip over the return offset that's on the stack
+    ADD(64, R(RSP), Imm32(8));
 }
 
-void JitCompiler::Compile_CALLC(Instruction instr) {
+void JitShader::Compile_CALLC(Instruction instr) {
     Compile_EvaluateCondition(instr);
     FixupBranch b = J_CC(CC_Z, true);
     Compile_CALL(instr);
     SetJumpTarget(b);
 }
 
-void JitCompiler::Compile_CALLU(Instruction instr) {
+void JitShader::Compile_CALLU(Instruction instr) {
     Compile_UniformCondition(instr);
     FixupBranch b = J_CC(CC_Z, true);
     Compile_CALL(instr);
     SetJumpTarget(b);
 }
 
-void JitCompiler::Compile_CMP(Instruction instr) {
+void JitShader::Compile_CMP(Instruction instr) {
     using Op = Instruction::Common::CompareOpType::Op;
     Op op_x = instr.common.compare_op.x;
     Op op_y = instr.common.compare_op.y;
@@ -627,7 +652,7 @@ void JitCompiler::Compile_CMP(Instruction instr) {
     SHR(64, R(COND1), Imm8(63));
 }
 
-void JitCompiler::Compile_MAD(Instruction instr) {
+void JitShader::Compile_MAD(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.mad.src1, SRC1);
 
     if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
@@ -644,9 +669,8 @@ void JitCompiler::Compile_MAD(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
-void JitCompiler::Compile_IF(Instruction instr) {
-    ASSERT_MSG(instr.flow_control.dest_offset > *offset_ptr, "Backwards if-statements (%d -> %d) not supported",
-            *offset_ptr, instr.flow_control.dest_offset.Value());
+void JitShader::Compile_IF(Instruction instr) {
+    Compile_Assert(instr.flow_control.dest_offset >= program_counter, "Backwards if-statements not supported");
 
     // Evaluate the "IF" condition
     if (instr.opcode.Value() == OpCode::Id::IFU) {
@@ -676,10 +700,9 @@ void JitCompiler::Compile_IF(Instruction instr) {
     SetJumpTarget(b2);
 }
 
-void JitCompiler::Compile_LOOP(Instruction instr) {
-    ASSERT_MSG(instr.flow_control.dest_offset > *offset_ptr, "Backwards loops (%d -> %d) not supported",
-            *offset_ptr, instr.flow_control.dest_offset.Value());
-    ASSERT_MSG(!looping, "Nested loops not supported");
+void JitShader::Compile_LOOP(Instruction instr) {
+    Compile_Assert(instr.flow_control.dest_offset >= program_counter, "Backwards loops not supported");
+    Compile_Assert(!looping, "Nested loops not supported");
 
     looping = true;
 
@@ -705,10 +728,7 @@ void JitCompiler::Compile_LOOP(Instruction instr) {
     looping = false;
 }
 
-void JitCompiler::Compile_JMP(Instruction instr) {
-    ASSERT_MSG(instr.flow_control.dest_offset > *offset_ptr, "Backwards jumps (%d -> %d) not supported",
-            *offset_ptr, instr.flow_control.dest_offset.Value());
-
+void JitShader::Compile_JMP(Instruction instr) {
     if (instr.opcode.Value() == OpCode::Id::JMPC)
         Compile_EvaluateCondition(instr);
     else if (instr.opcode.Value() == OpCode::Id::JMPU)
@@ -718,30 +738,38 @@ void JitCompiler::Compile_JMP(Instruction instr) {
 
     bool inverted_condition = (instr.opcode.Value() == OpCode::Id::JMPU) &&
         (instr.flow_control.num_instructions & 1);
+
     FixupBranch b = J_CC(inverted_condition ? CC_Z : CC_NZ, true);
+    fixup_branches.push_back({ b, instr.flow_control.dest_offset });
+}
 
-    Compile_Block(instr.flow_control.dest_offset);
+void JitShader::Compile_Block(unsigned end) {
+    while (program_counter < end) {
+        Compile_NextInstr();
+    }
+}
+
+void JitShader::Compile_Return() {
+    // Peek return offset on the stack and check if we're at that offset
+    MOV(64, R(RAX), MDisp(RSP, 8));
+    CMP(32, R(RAX), Imm32(program_counter));
 
+    // If so, jump back to before CALL
+    FixupBranch b = J_CC(CC_NZ, true);
+    RET();
     SetJumpTarget(b);
 }
 
-void JitCompiler::Compile_Block(unsigned end) {
-    // Save current offset pointer
-    unsigned* prev_offset_ptr = offset_ptr;
-    unsigned offset = *prev_offset_ptr;
+void JitShader::Compile_NextInstr() {
+    if (std::binary_search(return_offsets.begin(), return_offsets.end(), program_counter)) {
+        Compile_Return();
+    }
 
-    while (offset < end)
-        Compile_NextInstr(&offset);
+    ASSERT_MSG(code_ptr[program_counter] == nullptr, "Tried to compile already compiled shader location!");
+    code_ptr[program_counter] = GetCodePtr();
 
-    // Restore current offset pointer
-    offset_ptr = prev_offset_ptr;
-    *offset_ptr = offset;
-}
+    Instruction instr = GetVertexShaderInstruction(program_counter++);
 
-void JitCompiler::Compile_NextInstr(unsigned* offset) {
-    offset_ptr = offset;
-
-    Instruction instr = *(Instruction*)&g_state.vs.program_code[(*offset_ptr)++];
     OpCode::Id opcode = instr.opcode.Value();
     auto instr_func = instr_table[static_cast<unsigned>(opcode)];
 
@@ -755,9 +783,35 @@ void JitCompiler::Compile_NextInstr(unsigned* offset) {
     }
 }
 
-CompiledShader* JitCompiler::Compile() {
-    const u8* start = GetCodePtr();
-    unsigned offset = g_state.regs.vs.main_offset;
+void JitShader::FindReturnOffsets() {
+    return_offsets.clear();
+
+    for (size_t offset = 0; offset < g_state.vs.program_code.size(); ++offset) {
+        Instruction instr = GetVertexShaderInstruction(offset);
+
+        switch (instr.opcode.Value()) {
+        case OpCode::Id::CALL:
+        case OpCode::Id::CALLC:
+        case OpCode::Id::CALLU:
+            return_offsets.push_back(instr.flow_control.dest_offset + instr.flow_control.num_instructions);
+            break;
+        }
+    }
+
+    // Sort for efficient binary search later
+    std::sort(return_offsets.begin(), return_offsets.end());
+}
+
+void JitShader::Compile() {
+    // Reset flow control state
+    program = (CompiledShader*)GetCodePtr();
+    program_counter = 0;
+    looping = false;
+    code_ptr.fill(nullptr);
+    fixup_branches.clear();
+
+    // Find all `CALL` instructions and identify return locations
+    FindReturnOffsets();
 
     // The stack pointer is 8 modulo 16 at the entry of a procedure
     ABI_PushRegistersAndAdjustStack(ABI_ALL_CALLEE_SAVED, 8);
@@ -780,21 +834,31 @@ CompiledShader* JitCompiler::Compile() {
     MOV(PTRBITS, R(RAX), ImmPtr(&neg));
     MOVAPS(NEGBIT, MatR(RAX));
 
-    looping = false;
+    // Jump to start of the shader program
+    JMPptr(R(ABI_PARAM2));
+
+    // Compile entire program
+    Compile_Block(static_cast<unsigned>(g_state.vs.program_code.size()));
 
-    while (offset < g_state.vs.program_code.size()) {
-        Compile_NextInstr(&offset);
+    // Set the target for any incomplete branches now that the entire shader program has been emitted
+    for (const auto& branch : fixup_branches) {
+        SetJumpTarget(branch.first, code_ptr[branch.second]);
     }
 
-    return (CompiledShader*)start;
-}
+    // Free memory that's no longer needed
+    return_offsets.clear();
+    return_offsets.shrink_to_fit();
+    fixup_branches.clear();
+    fixup_branches.shrink_to_fit();
+
+    uintptr_t size = reinterpret_cast<uintptr_t>(GetCodePtr()) - reinterpret_cast<uintptr_t>(program);
+    ASSERT_MSG(size <= MAX_SHADER_SIZE, "Compiled a shader that exceeds the allocated size!");
 
-JitCompiler::JitCompiler() {
-    AllocCodeSpace(jit_cache_size);
+    LOG_DEBUG(HW_GPU, "Compiled shader size=%d", size);
 }
 
-void JitCompiler::Clear() {
-    ClearCodeSpace();
+JitShader::JitShader() {
+    AllocCodeSpace(MAX_SHADER_SIZE);
 }
 
 } // namespace Shader
diff --git a/src/video_core/shader/shader_jit_x64.h b/src/video_core/shader/shader_jit_x64.h
index 5357c964b..cd6280ade 100644
--- a/src/video_core/shader/shader_jit_x64.h
+++ b/src/video_core/shader/shader_jit_x64.h
@@ -4,6 +4,9 @@
 
 #pragma once
 
+#include <utility>
+#include <vector>
+
 #include <nihstro/shader_bytecode.h>
 
 #include "common/x64/emitter.h"
@@ -19,24 +22,22 @@ namespace Pica {
 
 namespace Shader {
 
-/// Memory needed to be available to compile the next shader (otherwise, clear the cache)
-constexpr size_t jit_shader_size = 1024 * 512;
-/// Memory allocated for the JIT code space cache
-constexpr size_t jit_cache_size = 1024 * 1024 * 8;
-
-using CompiledShader = void(void* registers);
+/// Memory allocated for each compiled shader (64Kb)
+constexpr size_t MAX_SHADER_SIZE = 1024 * 64;
 
 /**
  * This class implements the shader JIT compiler. It recompiles a Pica shader program into x86_64
  * code that can be executed on the host machine directly.
  */
-class JitCompiler : public Gen::XCodeBlock {
+class JitShader : public Gen::XCodeBlock {
 public:
-    JitCompiler();
+    JitShader();
 
-    CompiledShader* Compile();
+    void Run(void* registers, unsigned offset) const {
+        program(registers, code_ptr[offset]);
+    }
 
-    void Clear();
+    void Compile();
 
     void Compile_ADD(Instruction instr);
     void Compile_DP3(Instruction instr);
@@ -66,8 +67,9 @@ public:
     void Compile_MAD(Instruction instr);
 
 private:
+
     void Compile_Block(unsigned end);
-    void Compile_NextInstr(unsigned* offset);
+    void Compile_NextInstr();
 
     void Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg, Gen::X64Reg dest);
     void Compile_DestEnable(Instruction instr, Gen::X64Reg dest);
@@ -81,13 +83,39 @@ private:
     void Compile_EvaluateCondition(Instruction instr);
     void Compile_UniformCondition(Instruction instr);
 
+    /**
+     * Emits the code to conditionally return from a subroutine envoked by the `CALL` instruction.
+     */
+    void Compile_Return();
+
     BitSet32 PersistentCallerSavedRegs();
 
-    /// Pointer to the variable that stores the current Pica code offset. Used to handle nested code blocks.
-    unsigned* offset_ptr = nullptr;
+    /**
+     * Assertion evaluated at compile-time, but only triggered if executed at runtime.
+     * @param msg Message to be logged if the assertion fails.
+     */
+    void Compile_Assert(bool condition, const char* msg);
+
+    /**
+     * Analyzes the entire shader program for `CALL` instructions before emitting any code,
+     * identifying the locations where a return needs to be inserted.
+     */
+    void FindReturnOffsets();
+
+    /// Mapping of Pica VS instructions to pointers in the emitted code
+    std::array<const u8*, 1024> code_ptr;
+
+    /// Offsets in code where a return needs to be inserted
+    std::vector<unsigned> return_offsets;
+
+    unsigned program_counter = 0;       ///< Offset of the next instruction to decode
+    bool looping = false;               ///< True if compiling a loop, used to check for nested loops
+
+    /// Branches that need to be fixed up once the entire shader program is compiled
+    std::vector<std::pair<Gen::FixupBranch, unsigned>> fixup_branches;
 
-    /// Set to true if currently in a loop, used to check for the existence of nested loops
-    bool looping = false;
+    using CompiledShader = void(void* registers, const u8* start_addr);
+    CompiledShader* program = nullptr;
 };
 
 } // Shader
diff --git a/src/video_core/video_core.cpp b/src/video_core/video_core.cpp
index ee5e50df1..256899c89 100644
--- a/src/video_core/video_core.cpp
+++ b/src/video_core/video_core.cpp
@@ -5,7 +5,6 @@
 #include <memory>
 
 #include "common/emu_window.h"
-#include "common/make_unique.h"
 #include "common/logging/log.h"
 
 #include "core/core.h"
@@ -32,7 +31,7 @@ bool Init(EmuWindow* emu_window) {
     Pica::Init();
 
     g_emu_window = emu_window;
-    g_renderer = Common::make_unique<RendererOpenGL>();
+    g_renderer = std::make_unique<RendererOpenGL>();
     g_renderer->SetWindow(g_emu_window);
     if (g_renderer->Init()) {
         LOG_DEBUG(Render, "initialized OK");