Init GX2 to reduce screen corruption if no menu was shown

Bumps wiiu-env/devkitppc from 20240423 to 20240505.

---
updated-dependencies:
- dependency-name: wiiu-env/devkitppc
  dependency-type: direct:production
  update-type: version-update:semver-major
...

Signed-off-by: dependabot[bot] <support@github.com>

.github/dependabot.yml

@@ -0,0 +1,10 @@
+version: 2
+updates:
+  - package-ecosystem: "docker"
+    directory: "/"
+    schedule:
+      interval: "daily"
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "daily"

.github/workflows/ci.yml

@@ -9,7 +9,7 @@ jobs:
   clang-format:
     runs-on: ubuntu-22.04
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
     - name: clang-format
       run: |
         docker run --rm -v ${PWD}:/src ghcr.io/wiiu-env/clang-format:13.0.0-2 -r ./source --exclude ./source/elfio
@@ -17,7 +17,14 @@ jobs:
     runs-on: ubuntu-22.04
     needs: clang-format
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
+    - name: create version.h
+      run: |
+        git_hash=$(git rev-parse --short "$GITHUB_SHA")
+        cat <<EOF > ./source/version.h
+        #pragma once
+        #define ENVIRONMENT_LOADER_VERSION_EXTRA " (nightly-$git_hash)"
+        EOF
     - name: build binary
       run: |
         docker build . -t builder
@@ -42,7 +49,7 @@ jobs:
     - name: zip artifact
       run: zip -r ${{ env.REPOSITORY_NAME }}_${{ env.DATETIME }}.zip wiiu
     - name: Create Release
-      uses: "softprops/action-gh-release@v1"
+      uses: "softprops/action-gh-release@v2"
       with:
         tag_name: ${{ env.REPOSITORY_NAME }}-${{ env.DATETIME }}
         draft: false

.github/workflows/pr.yml

@@ -6,7 +6,7 @@ jobs:
   clang-format:
     runs-on: ubuntu-22.04
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
     - name: clang-format
       run: |
         docker run --rm -v ${PWD}:/src ghcr.io/wiiu-env/clang-format:13.0.0-2 -r ./source --exclude ./source/elfio
@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-22.04
     needs: clang-format
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - name: build binary with logging
         run: |
           docker build . -t builder
@@ -25,7 +25,14 @@ jobs:
     runs-on: ubuntu-22.04
     needs: clang-format
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
+    - name: create version.h
+      run: |
+        git_hash=$(git rev-parse --short "${{ github.event.pull_request.head.sha }}")
+        cat <<EOF > ./source/version.h
+        #pragma once
+        #define ENVIRONMENT_LOADER_VERSION_EXTRA " (nightly-$git_hash)"
+        EOF
     - name: build binary
       run: |
         docker build . -t builder

.gitignore

@@ -12,3 +12,4 @@ cmake-build-debug/
 *.txt
 build1/
 cmake-build-default/
+*.zip

Dockerfile

@@ -1,3 +1,3 @@
-FROM ghcr.io/wiiu-env/devkitppc:20230417
+FROM ghcr.io/wiiu-env/devkitppc:20240704
 
 WORKDIR project

source/ElfUtils.cpp

@@ -3,32 +3,39 @@
 #include <coreinit/cache.h>
 #include <coreinit/debug.h>
 #include <coreinit/dynload.h>
-#include <coreinit/memdefaultheap.h>
 
 #include "ElfUtils.h"
 #include "elfio/elfio.hpp"
 
-bool ElfUtils::doRelocation(const std::vector<std::unique_ptr<RelocationData>> &relocData, relocation_trampoline_entry_t *tramp_data, uint32_t tramp_length) {
+bool ElfUtils::doRelocation(const std::vector<RelocationData> &relocData, relocation_trampoline_entry_t *tramp_data, uint32_t tramp_length, std::map<std::string, OSDynLoad_Module> &usedRPls) {
     for (auto const &curReloc : relocData) {
-        std::string functionName   = curReloc->getName();
-        std::string rplName        = curReloc->getImportRPLInformation()->getRPLName();
-        int32_t isData             = curReloc->getImportRPLInformation()->isData();
+        std::string functionName   = curReloc.getName();
+        std::string rplName        = curReloc.getImportRPLInformation()->getRPLName();
+        int32_t isData             = curReloc.getImportRPLInformation()->isData();
         OSDynLoad_Module rplHandle = nullptr;
 
-
-        auto err = OSDynLoad_IsModuleLoaded(rplName.c_str(), &rplHandle);
-        if (err != OS_DYNLOAD_OK || rplHandle == nullptr) {
-            // only acquire if not already loaded.
-            OSDynLoad_Acquire(rplName.c_str(), &rplHandle);
+        if (!usedRPls.contains(rplName)) {
+            //DEBUG_FUNCTION_LINE_VERBOSE("Acquire %s", rplName.c_str());
+            // Always acquire to increase refcount and make sure it won't get unloaded while we're using it.
+            OSDynLoad_Error err = OSDynLoad_Acquire(rplName.c_str(), &rplHandle);
+            if (err != OS_DYNLOAD_OK) {
+                DEBUG_FUNCTION_LINE_ERR("Failed to acquire %s", rplName.c_str());
+                return false;
+            }
+            // Keep track RPLs we are using.
+            // They will be released on exit (See: AromaBaseModule)
+            usedRPls[rplName] = rplHandle;
+        } else {
+            //DEBUG_FUNCTION_LINE_VERBOSE("Use from usedRPLs cache! %s", rplName.c_str());
         }
+        rplHandle = usedRPls[rplName];
 
         uint32_t functionAddress = 0;
-        OSDynLoad_FindExport(rplHandle, (OSDynLoad_ExportType) isData, functionName.c_str(), (void **) &functionAddress);
-        if (functionAddress == 0) {
+        if ((OSDynLoad_FindExport(rplHandle, (OSDynLoad_ExportType) isData, functionName.c_str(), (void **) &functionAddress) != OS_DYNLOAD_OK) || functionAddress == 0) {
             DEBUG_FUNCTION_LINE_ERR("Failed to find export for %s %s %d", functionName.c_str(), rplName.c_str(), isData);
             return false;
         }
-        if (!ElfUtils::elfLinkOne(curReloc->getType(), curReloc->getOffset(), curReloc->getAddend(), (uint32_t) curReloc->getDestination(), functionAddress, tramp_data, tramp_length,
+        if (!ElfUtils::elfLinkOne(curReloc.getType(), curReloc.getOffset(), curReloc.getAddend(), (uint32_t) curReloc.getDestination(), functionAddress, tramp_data, tramp_length,
                                   RELOC_TYPE_IMPORT)) {
             DEBUG_FUNCTION_LINE_ERR("Relocation failed\n");
             return false;
@@ -152,7 +159,6 @@ bool ElfUtils::elfLinkOne(char type, size_t offset, int32_t addend, uint32_t des
                     freeSlot->trampoline[1] = 0x616B0000 | (((uint32_t) value) & 0x0000ffff);         // ori r11, r11, real_addr@l
                     freeSlot->trampoline[2] = 0x7D6903A6;                                             // mtctr   r11
                     freeSlot->trampoline[3] = 0x4E800420;                                             // bctr
-                    DCFlushRange((void *) freeSlot->trampoline, sizeof(freeSlot->trampoline));
                     ICInvalidateRange((unsigned char *) freeSlot->trampoline, sizeof(freeSlot->trampoline));
 
                     if (reloc_type == RELOC_TYPE_FIXED) {

source/ElfUtils.h

@@ -2,8 +2,10 @@
 
 #include "common/relocation_defines.h"
 #include "module/RelocationData.h"
+#include <coreinit/dynload.h>
 #include <cstdint>
 #include <cstdio>
+#include <map>
 #include <memory>
 #include <vector>
 
@@ -49,6 +51,5 @@ public:
     static bool elfLinkOne(char type, size_t offset, int32_t addend, uint32_t destination, uint32_t symbol_addr, relocation_trampoline_entry_t *trampolin_data, uint32_t trampolin_data_length,
                            RelocationType reloc_type);
 
-
-    static bool doRelocation(const std::vector<std::unique_ptr<RelocationData>> &relocData, relocation_trampoline_entry_t *tramp_data, uint32_t tramp_length);
+    static bool doRelocation(const std::vector<RelocationData> &relocData, relocation_trampoline_entry_t *tramp_data, uint32_t tramp_length, std::map<std::string, OSDynLoad_Module> &usedRPls);
 };

source/main.cpp

@@ -7,14 +7,17 @@
 #include <coreinit/filesystem_fsa.h>
 #include <coreinit/foreground.h>
 #include <coreinit/ios.h>
+#include <coreinit/savedframe.h>
 #include <coreinit/screen.h>
 #include <coreinit/title.h>
 #include <elfio/elfio.hpp>
 #include <fcntl.h>
+#include <gx2/display.h>
 #include <gx2/state.h>
 #include <malloc.h>
 #include <memory>
 #include <nn/act/client_cpp.h>
+#include <nsysccr/cdc.h>
 #include <proc_ui/procui.h>
 #include <sysapp/launch.h>
 #include <sysapp/title.h>
@@ -30,14 +33,18 @@
 #include "kernel.h"
 #include "module/ModuleDataFactory.h"
 #include "utils/DrawUtils.h"
+#include "utils/FileUtils.h"
+#include "utils/InputUtils.h"
+#include "utils/OnLeavingScope.h"
+#include "utils/PairUtils.h"
+#include "utils/utils.h"
+#include "utils/wiiu_zlib.hpp"
+#include "version.h"
 
-// clang-format off
-#define MEMORY_REGION_START         0x00A00000
-#define MEMORY_REGION_SIZE          0x00600000
-#define MEMORY_REGION_END           (MEMORY_REGION_START + MEMORY_REGION_SIZE)
+#define ENVIRONMENT_LOADER_VERSION "v0.3.2"
 
-#define AUTOBOOT_CONFIG_PATH        "fs:/vol/external01/wiiu/environments/default.cfg"
-// clang-format on
+#define MEMORY_REGION_START        0x00800000
+#define AUTOBOOT_CONFIG_PATH       "fs:/vol/external01/wiiu/environments/default.cfg"
 
 bool CheckRunning() {
     switch (ProcUIProcessMessages(true)) {
@@ -90,6 +97,8 @@ bool writeFileContent(const std::string &path, const std::string &content) {
 
 extern "C" void __fini();
 extern "C" void __init_wut_malloc();
+void LoadAndRunModule(std::string_view filepath, std::string_view environment_path);
+void ClearSavedFrameBuffers();
 
 int main(int argc, char **argv) {
     // We need to call __init_wut_malloc somewhere so wut_malloc will be used for the memory allocation.
@@ -108,28 +117,22 @@ int main(int argc, char **argv) {
 
     DEBUG_FUNCTION_LINE("Hello from EnvironmentLoader!");
 
-    char environmentPath[0x100];
-    memset(environmentPath, 0, sizeof(environmentPath));
-
-    auto handle = IOS_Open("/dev/mcp", IOS_OPEN_READ);
+    char environmentPathFromIOSU[0x100] = {};
+    auto handle                         = IOS_Open("/dev/mcp", IOS_OPEN_READ);
     if (handle >= 0) {
         int in = 0xF9; // IPC_CUSTOM_COPY_ENVIRONMENT_PATH
-        if (IOS_Ioctl(handle, 100, &in, sizeof(in), environmentPath, sizeof(environmentPath)) == IOS_ERROR_OK) {
-            DEBUG_FUNCTION_LINE("Boot into %s", environmentPath);
+        if (IOS_Ioctl(handle, 100, &in, sizeof(in), environmentPathFromIOSU, sizeof(environmentPathFromIOSU)) == IOS_ERROR_OK) {
+            DEBUG_FUNCTION_LINE("Boot into %s", environmentPathFromIOSU);
         }
 
         IOS_Close(handle);
     }
 
-    // We substract 0x100 to be safe.
-    uint32_t textSectionStart = textStart() - 0x100;
-
-    auto gModuleData = (module_information_t *) (textSectionStart - sizeof(module_information_t));
-
     bool noEnvironmentsFound = false;
+    bool shownMenu           = false;
 
-    std::string environment_path = std::string(environmentPath);
-    if (strncmp(environmentPath, "fs:/vol/external01/wiiu/environments/", strlen("fs:/vol/external01/wiiu/environments/")) != 0) {
+    std::string environmentPath = std::string(environmentPathFromIOSU);
+    if (!environmentPath.starts_with("fs:/vol/external01/wiiu/environments/")) { // If the environment path in IOSU is empty or unexpected, read config
         DirList environmentDirs("fs:/vol/external01/wiiu/environments/", nullptr, DirList::Dirs, 1);
 
         std::map<std::string, std::string> environmentPaths;
@@ -146,9 +149,9 @@ int main(int argc, char **argv) {
             for (auto const &[key, val] : environmentPaths) {
                 if (res.value() == key) {
                     DEBUG_FUNCTION_LINE("Found environment %s from config at index %d", res.value().c_str(), i);
-                    autobootIndex    = i;
-                    environment_path = val;
-                    forceMenu        = false;
+                    autobootIndex   = i;
+                    environmentPath = val;
+                    forceMenu       = false;
                     break;
                 }
                 i++;
@@ -157,29 +160,41 @@ int main(int argc, char **argv) {
             DEBUG_FUNCTION_LINE_ERR("No config found");
         }
 
-        VPADReadError err;
-        VPADStatus vpad_data;
-        VPADRead(VPAD_CHAN_0, &vpad_data, 1, &err);
+        InputUtils::Init();
 
-        uint32_t btn = 0;
-        if (err == VPAD_READ_SUCCESS) {
-            btn = vpad_data.hold | vpad_data.trigger;
-        }
+        InputUtils::InputData input = InputUtils::getControllerInput();
 
-        if (forceMenu || (btn & VPAD_BUTTON_X) == VPAD_BUTTON_X) {
+        if (forceMenu || ((input.trigger | input.hold) & VPAD_BUTTON_X) == VPAD_BUTTON_X) {
+            shownMenu = true;
             DEBUG_FUNCTION_LINE_VERBOSE("Open menu!");
-            environment_path = EnvironmentSelectionScreen(environmentPaths, autobootIndex);
+            environmentPath = EnvironmentSelectionScreen(environmentPaths, autobootIndex);
             if (environmentPaths.empty()) {
                 noEnvironmentsFound = true;
             } else {
-                DEBUG_FUNCTION_LINE_VERBOSE("Selected %s", environment_path.c_str());
+                DEBUG_FUNCTION_LINE_VERBOSE("Selected %s", environmentPath.c_str());
             }
+        } else {
         }
+        InputUtils::DeInit();
     }
+
+    if (!shownMenu) {
+        // Clear saved frame buffer to reduce screen corruption
+        ClearSavedFrameBuffers();
+
+        OSScreenInit();
+
+        // Call GX2Init to shut down OSScreen
+        GX2Init(nullptr);
+
+        GX2SetTVEnable(FALSE);
+        GX2SetDRCEnable(FALSE);
+    }
+
     RevertMainHook();
 
     if (!noEnvironmentsFound) {
-        DirList setupModules(environment_path + "/modules/setup", ".rpx", DirList::Files, 1);
+        DirList setupModules(environmentPath + "/modules/setup", ".rpx", DirList::Files, 1);
         setupModules.SortList();
 
         for (int i = 0; i < setupModules.GetFilecount(); i++) {
@@ -189,43 +204,9 @@ int main(int argc, char **argv) {
                 continue;
             }
 
-            // Some module may unmount the sd card on exit.
-            FSAInit();
-            auto client = FSAAddClient(nullptr);
-            if (client) {
-                FSAMount(client, "/dev/sdcard01", "/vol/external01", static_cast<FSAMountFlags>(0), nullptr, 0);
-                FSADelClient(client);
-            }
-
-            uint32_t destination_address_end = ((uint32_t) gModuleData) & 0xFFFF0000;
-            memset((void *) gModuleData, 0, sizeof(module_information_t));
-            DEBUG_FUNCTION_LINE("Trying to run %s.", setupModules.GetFilepath(i), destination_address_end, ((uint32_t) gModuleData) - MEMORY_REGION_START);
-            auto moduleData = ModuleDataFactory::load(setupModules.GetFilepath(i), destination_address_end, ((uint32_t) gModuleData) - MEMORY_REGION_START, gModuleData->trampolines,
-                                                      DYN_LINK_TRAMPOLIN_LIST_LENGTH);
-            if (!moduleData) {
-                DEBUG_FUNCTION_LINE_ERR("Failed to load %s", setupModules.GetFilepath(i));
-                OSFatal("EnvironmentLoader: Failed to load module");
-                continue;
-            }
-            DEBUG_FUNCTION_LINE("Loaded module data");
-            if (!ElfUtils::doRelocation(moduleData.value()->getRelocationDataList(), gModuleData->trampolines, DYN_LINK_TRAMPOLIN_LIST_LENGTH)) {
-                DEBUG_FUNCTION_LINE_ERR("Relocations failed");
-                OSFatal("EnvironmentLoader: Relocations failed");
-            } else {
-                DEBUG_FUNCTION_LINE("Relocation done");
-            }
-
-            DCFlushRange((void *) moduleData.value()->getStartAddress(), moduleData.value()->getEndAddress() - moduleData.value()->getStartAddress());
-            ICInvalidateRange((void *) moduleData.value()->getStartAddress(), moduleData.value()->getEndAddress() - moduleData.value()->getStartAddress());
-
-            DEBUG_FUNCTION_LINE("Calling entrypoint @%08X", moduleData.value()->getEntrypoint());
-            char *arr[1];
-            arr[0] = (char *) environment_path.c_str();
-            // clang-format off
-            ((int(*)(int, char **)) moduleData.value()->getEntrypoint())(1, arr);
-            // clang-format on
-            DEBUG_FUNCTION_LINE("Back from module");
+            LoadAndRunModule(setupModules.GetFilepath(i), environmentPath);
         }
+
     } else {
         DEBUG_FUNCTION_LINE("Return to Wii U Menu");
         ProcUIInit(OSSavesDone_ReadyToRelease);
@@ -258,17 +239,239 @@ int main(int argc, char **argv) {
     return 0;
 }
 
-#define COLOR_WHITE              Color(0xffffffff)
-#define COLOR_BLACK              Color(0, 0, 0, 255)
-#define COLOR_RED                Color(237, 28, 36, 255)
-#define COLOR_BACKGROUND         Color(0, 40, 100, 255)
-#define COLOR_TEXT               COLOR_WHITE
-#define COLOR_TEXT2              Color(0xB3ffffff)
-#define COLOR_AUTOBOOT           Color(0xaeea00ff)
-#define COLOR_BORDER             Color(204, 204, 204, 255)
-#define COLOR_BORDER_HIGHLIGHTED Color(0x3478e4ff)
+std::optional<HeapWrapper> GetHeapFromMappedMemory(uint32_t heapSize) {
+    void *(*MEMAllocFromDefaultHeapExForThreads)(uint32_t size, int align) = nullptr;
+    void (*MEMFreeToDefaultHeapForThreads)(void *ptr)                      = nullptr;
+
+    // Let's try to get the memalign and free functions from the memorymapping module.
+    OSDynLoad_Module module;
+    if (OSDynLoad_Acquire("homebrew_memorymapping", &module) != OS_DYNLOAD_OK) {
+        DEBUG_FUNCTION_LINE("Failed to acquire homebrew_memorymapping.");
+        return {};
+    }
+    /* Memory allocation functions */
+    uint32_t *allocPtr = nullptr, *freePtr = nullptr;
+    if (OSDynLoad_FindExport(module, OS_DYNLOAD_EXPORT_DATA, "MEMAllocFromMappedMemoryEx", reinterpret_cast<void **>(&allocPtr)) != OS_DYNLOAD_OK) {
+        DEBUG_FUNCTION_LINE("OSDynLoad_FindExport for MEMAllocFromDefaultHeapEx failed");
+        return {};
+    }
+    if (OSDynLoad_FindExport(module, OS_DYNLOAD_EXPORT_DATA, "MEMFreeToMappedMemory", reinterpret_cast<void **>(&freePtr)) != OS_DYNLOAD_OK) {
+        DEBUG_FUNCTION_LINE("OSDynLoad_FindExport for MEMFreeToDefaultHeap failed");
+        return {};
+    }
+
+    MEMAllocFromDefaultHeapExForThreads = (void *(*) (uint32_t, int) ) * allocPtr;
+    MEMFreeToDefaultHeapForThreads      = (void (*)(void *)) * freePtr;
+
+    if (!MEMAllocFromDefaultHeapExForThreads || !MEMFreeToDefaultHeapForThreads) {
+        DEBUG_FUNCTION_LINE_ERR("MEMAllocFromDefaultHeapExForThreads or MEMFreeToDefaultHeapForThreads is null");
+        // the mapped memory is not available (yet)
+        return {};
+    }
+
+    uint32_t size = heapSize;
+    auto ptr      = MEMAllocFromDefaultHeapExForThreads(size, 0x4);
+    if (!ptr) {
+        DEBUG_FUNCTION_LINE_ERR("Failed to alloc memory: %d bytes", size);
+        return {};
+    }
+
+    DEBUG_FUNCTION_LINE("Let's create a memory wrapper for 0x%08X, size: %d", ptr, size);
+    return HeapWrapper(MemoryWrapper(ptr, size, MEMFreeToDefaultHeapForThreads));
+}
+
+std::optional<HeapWrapper> GetHeapForModule(uint32_t heapSize) {
+    // If Aroma is already loaded, we can't use the region between MEMORY_REGION_START and MEMORY_REGION_END anymore because Aroma is using.
+    // So instead we check before loading a module if aromas memory mapping module is already usable. If yes, we use this to load the module instead
+    if (auto heapWrapper = GetHeapFromMappedMemory(heapSize)) {
+        return heapWrapper;
+    }
+
+    // If Aroma is not already loaded, we use the existing 0x00800000 - 0x01000000 memory region. This is where aroma is loaded to. Note: this region may be only mapped to the main core.
+    // The environment loader is loaded to the end of 0x00800000 - 0x01000000 memory region. With this helper we know the start of the .text section
+    uint32_t textSectionStart = textStart() - 0x100;
+
+    auto endOfUsableMemory = textSectionStart;
+    uint32_t startAddress  = ((uint32_t) endOfUsableMemory - heapSize) & 0xFFFF0000;
+    uint32_t size          = endOfUsableMemory - startAddress;
+
+    if (startAddress < MEMORY_REGION_START) {
+        DEBUG_FUNCTION_LINE_ERR("Not enough static memory");
+        return {};
+    }
+
+    DEBUG_FUNCTION_LINE("Let's create a memory wrapper for 0x%08X, size: %d", startAddress, size);
+    auto res = HeapWrapper(MemoryWrapper((void *) startAddress, size, /* we don't need to free this memory*/ nullptr));
+    if ((uint32_t) res.GetHeapHandle() != startAddress) {
+        OSFatal("EnvironmentLoader: Unexpected address");
+    }
+    return res;
+}
+
+void SetupKernelModule() {
+    void *(*KernelSetupDefaultSyscalls)() = nullptr;
+
+    OSDynLoad_Module module;
+    if (OSDynLoad_Acquire("homebrew_kernel", &module) != OS_DYNLOAD_OK) {
+        DEBUG_FUNCTION_LINE("Failed to acquire homebrew_kernel.");
+        return;
+    }
+
+    if (OSDynLoad_FindExport(module, OS_DYNLOAD_EXPORT_FUNC, "KernelSetupDefaultSyscalls", reinterpret_cast<void **>(&KernelSetupDefaultSyscalls)) != OS_DYNLOAD_OK) {
+        DEBUG_FUNCTION_LINE("OSDynLoad_FindExport for KernelSetupDefaultSyscalls failed");
+        OSFatal("EnvironmentLoader: KernelModule is missing the export\n"
+                "\"KernelSetupDefaultSyscalls\"... Please update Aroma!\n"
+                "\n"
+                "See https://wiiu.hacks.guide/ for more information.");
+        return;
+    }
+
+    if (!KernelSetupDefaultSyscalls) {
+        DEBUG_FUNCTION_LINE_WARN("KernelSetupDefaultSyscalls is null");
+        OSFatal("EnvironmentLoader: KernelModule is missing the export\n"
+                "\"KernelSetupDefaultSyscalls\"... Please update Aroma!\n"
+                "\n"
+                "See https://wiiu.hacks.guide/ for more information.");
+        return;
+    }
+
+    DEBUG_FUNCTION_LINE("Call KernelSetupDefaultSyscalls");
+    KernelSetupDefaultSyscalls();
+    OSDynLoad_Release(module);
+}
+
+void LoadAndRunModule(std::string_view filepath, std::string_view environment_path) {
+    // Some module may unmount the sd card on exit.
+    FSAInit();
+    auto client = FSAAddClient(nullptr);
+    if (client) {
+        FSAMount(client, "/dev/sdcard01", "/vol/external01", static_cast<FSAMountFlags>(0), nullptr, 0);
+        FSADelClient(client);
+    } else {
+        DEBUG_FUNCTION_LINE_ERR("Failed to add FSA client");
+    }
+
+    DEBUG_FUNCTION_LINE("Trying to load %s into memory", filepath.data());
+    uint8_t *buffer = nullptr;
+    uint32_t fsize  = 0;
+    if (LoadFileToMem(filepath.data(), &buffer, &fsize) < 0) {
+        DEBUG_FUNCTION_LINE_ERR("Failed to load file");
+        OSFatal("EnvironmentLoader: Failed to load file to memory");
+        return;
+    }
+
+    auto cleanupBuffer = onLeavingScope([buffer]() { free(buffer); });
+
+    ELFIO::elfio reader(new wiiu_zlib);
+    // Load ELF data
+    if (!reader.load(reinterpret_cast<const char *>(buffer), fsize)) {
+        DEBUG_FUNCTION_LINE_ERR("Can't parse .wms from buffer.");
+        OSFatal("Can't parse .wms from buffer.");
+        return;
+    }
+
+    uint32_t moduleSize = ModuleDataFactory::GetSizeOfModule(reader);
+    DEBUG_FUNCTION_LINE_VERBOSE("Module has size: %d", moduleSize);
+
+    uint32_t requiredHeapSize = moduleSize + sizeof(module_information_t) + 0x10000; // add another 0x10000 to be safe
+    DEBUG_FUNCTION_LINE_VERBOSE("Allocate %d bytes for heap (%.2f KiB)", requiredHeapSize, requiredHeapSize / 1024.0f);
+
+    if (auto heapWrapperOpt = GetHeapForModule(requiredHeapSize); heapWrapperOpt.has_value()) {
+        // Frees automatically, must not survive the heapWrapper.
+        auto moduleInfoOpt = heapWrapperOpt->Alloc(sizeof(module_information_t), 0x4);
+        if (!moduleInfoOpt) {
+            DEBUG_FUNCTION_LINE_ERR("Failed to alloc module information");
+            OSFatal("EnvironmentLoader: Failed to alloc module information");
+            return;
+        }
+
+        auto moduleInfo    = std::move(*moduleInfoOpt);
+        auto moduleInfoPtr = (module_information_t *) moduleInfo.data();
+        *moduleInfoPtr     = {};
+
+        // Frees automatically, must not survive the heapWrapper.
+        auto moduleData = ModuleDataFactory::load(reader, *heapWrapperOpt, moduleInfoPtr->trampolines, sizeof(moduleInfoPtr->trampolines) / sizeof(moduleInfoPtr->trampolines[0]));
+        if (!moduleData) {
+            DEBUG_FUNCTION_LINE_ERR("Failed to load %s", filepath);
+            OSFatal("EnvironmentLoader: Failed to load module");
+            return;
+        }
+
+        DEBUG_FUNCTION_LINE("Loaded module data");
+        std::map<std::string, OSDynLoad_Module> usedRPls;
+        if (!ElfUtils::doRelocation(moduleData.value()->getRelocationDataList(), moduleInfoPtr->trampolines, sizeof(moduleInfoPtr->trampolines) / sizeof(moduleInfoPtr->trampolines[0]), usedRPls)) {
+            DEBUG_FUNCTION_LINE_ERR("Relocations failed");
+            OSFatal("EnvironmentLoader: Relocations failed");
+        } else {
+            DEBUG_FUNCTION_LINE("Relocation done");
+        }
+
+        char *arr[4];
+        arr[0] = (char *) environment_path.data();
+        arr[1] = (char *) "EnvironmentLoader"; //
+        arr[2] = (char *) 0x02;                // Version
+        /*
+         * This is a hacky work around to tell Aromas Module Loader which memory region it can use safely. After using it, it's expected to expose new memory region via the
+         * custom rpl "homebrew_mappedmemory" (See: GetHeapFromMappedMemory). The returned memory is expected to be RWX for user and kernel.
+         * Once a custom memory allocator is provided, usable_mem_start and usable_mem_end are set to 0.
+         */
+        auto usable_mem_end = (uint32_t) heapWrapperOpt->GetHeapHandle();
+        if (heapWrapperOpt->IsAllocated()) { // Check if you use memory which is actually allocated. This means we can't give it to the module.
+            DEBUG_FUNCTION_LINE("Don't give the module a usable memory region because it will be loaded on a custom memory region.");
+            usable_mem_end = 0;
+        }
+        arr[3] = (char *) usable_mem_end; // End of usable memory
+
+        DEBUG_FUNCTION_LINE("Calling entrypoint @%08X with: \"%s\", \"%s\", %08X, %08X", moduleData.value()->getEntrypoint(), arr[0], arr[1], arr[2], arr[3]);
+        // clang-format off
+        ((int(*)(int, char **)) moduleData.value()->getEntrypoint())(sizeof(arr)/ sizeof(arr[0]), arr);
+        // clang-format on
+        DEBUG_FUNCTION_LINE("Back from module");
+
+        for (auto &rpl : usedRPls) {
+            DEBUG_FUNCTION_LINE_VERBOSE("Release %s", rpl.first.c_str());
+            OSDynLoad_Release(rpl.second);
+        }
+
+    } else {
+        DEBUG_FUNCTION_LINE_ERR("Failed to create heap");
+        OSFatal("EnvironmentLoader: Failed to create heap");
+    }
+
+    // module may override the syscalls used by the Aroma KernelModule. This (tries to) re-init(s) the KernelModule after a setup module has been run.
+    SetupKernelModule();
+}
+
+void ClearSavedFrameBuffers() {
+    // If GX2 is running make sure to shut it down and free all existing memory in the saved-frame area.
+    if (GX2GetMainCoreId() != -1) {
+        GX2SetTVEnable(FALSE);
+        GX2SetDRCEnable(FALSE);
+        GX2Shutdown();
+    }
+
+    __OSClearSavedFrame(OS_SAVED_FRAME_A, OS_SAVED_FRAME_SCREEN_TV);
+    __OSClearSavedFrame(OS_SAVED_FRAME_A, OS_SAVED_FRAME_SCREEN_DRC);
+    __OSClearSavedFrame(OS_SAVED_FRAME_B, OS_SAVED_FRAME_SCREEN_TV);
+    __OSClearSavedFrame(OS_SAVED_FRAME_B, OS_SAVED_FRAME_SCREEN_DRC);
+}
+
+void AbortQuickStartMenu() {
+    CCRCDCDrcState state = {};
+    CCRCDCSysGetDrcState(CCR_CDC_DESTINATION_DRC0, &state);
+    if (state.state == CCR_CDC_DRC_STATE_SUBACTIVE) {
+        state.state = CCR_CDC_DRC_STATE_ACTIVE;
+        CCRCDCSysSetDrcState(CCR_CDC_DESTINATION_DRC0, &state);
+    }
+}
 
 std::string EnvironmentSelectionScreen(const std::map<std::string, std::string> &payloads, int32_t autobootIndex) {
+    // Close quick start menu is selection screen is displayed
+    AbortQuickStartMenu();
+
+    // Clear saved frame buffer to reduce screen corruption
+    ClearSavedFrameBuffers();
+
     OSScreenInit();
 
     uint32_t tvBufferSize  = OSScreenGetBufferSizeEx(SCREEN_TV);
@@ -276,7 +479,7 @@ std::string EnvironmentSelectionScreen(const std::map<std::string, std::string>
 
     auto *screenBuffer = (uint8_t *) memalign(0x100, tvBufferSize + drcBufferSize);
     if (!screenBuffer) {
-        OSFatal("Fail to allocate screenBuffer");
+        OSFatal("EnvironmentLoader: Fail to allocate screenBuffer");
     }
     memset(screenBuffer, 0, tvBufferSize + drcBufferSize);
 
@@ -289,38 +492,38 @@ std::string EnvironmentSelectionScreen(const std::map<std::string, std::string>
     DrawUtils::initBuffers(screenBuffer, tvBufferSize, screenBuffer + tvBufferSize, drcBufferSize);
 
     if (!DrawUtils::initFont()) {
-        OSFatal("Failed to init font");
+        OSFatal("EnvironmentLoader: Failed to init font");
     }
 
     uint32_t selected = autobootIndex > 0 ? autobootIndex : 0;
     int autoBoot      = autobootIndex;
 
-    bool redraw = true;
-    while (true) {
-        VPADStatus vpad{};
-        VPADRead(VPAD_CHAN_0, &vpad, 1, nullptr);
-
-        if (vpad.trigger & VPAD_BUTTON_UP) {
-            if (selected > 0) {
-                selected--;
-                redraw = true;
+    {
+        PairMenu pairMenu;
+        while (true) {
+            if (pairMenu.ProcessPairScreen()) {
+                continue;
             }
-        } else if (vpad.trigger & VPAD_BUTTON_DOWN) {
-            if (selected < payloads.size() - 1) {
-                selected++;
-                redraw = true;
-            }
-        } else if (vpad.trigger & VPAD_BUTTON_A) {
-            break;
-        } else if (vpad.trigger & VPAD_BUTTON_X) {
-            autoBoot = -1;
-            redraw   = true;
-        } else if (vpad.trigger & VPAD_BUTTON_Y) {
-            autoBoot = selected;
-            redraw   = true;
-        }
 
-        if (redraw) {
+            InputUtils::InputData input = InputUtils::getControllerInput();
+
+            if (input.trigger & VPAD_BUTTON_UP) {
+                if (selected > 0) {
+                    selected--;
+                }
+            } else if (input.trigger & VPAD_BUTTON_DOWN) {
+                if (selected < payloads.size() - 1) {
+                    selected++;
+                }
+            } else if (input.trigger & VPAD_BUTTON_A) {
+                break;
+            } else if (input.trigger & (VPAD_BUTTON_X | VPAD_BUTTON_MINUS)) {
+                autoBoot = -1;
+            } else if (input.trigger & (VPAD_BUTTON_Y | VPAD_BUTTON_PLUS)) {
+                autoBoot = selected;
+            }
+
+
             DrawUtils::beginDraw();
             DrawUtils::clear(COLOR_BACKGROUND);
 
@@ -354,6 +557,8 @@ std::string EnvironmentSelectionScreen(const std::map<std::string, std::string>
             DrawUtils::setFontSize(24);
             DrawUtils::print(16, 6 + 24, "Environment Loader");
             DrawUtils::drawRectFilled(8, 8 + 24 + 4, SCREEN_WIDTH - 8 * 2, 3, COLOR_WHITE);
+            DrawUtils::setFontSize(16);
+            DrawUtils::print(SCREEN_WIDTH - 16, 6 + 24, ENVIRONMENT_LOADER_VERSION ENVIRONMENT_LOADER_VERSION_EXTRA, true);
 
             // draw bottom bar
             DrawUtils::drawRectFilled(8, SCREEN_HEIGHT - 24 - 8 - 4, SCREEN_WIDTH - 8 * 2, 3, COLOR_WHITE);
@@ -361,15 +566,13 @@ std::string EnvironmentSelectionScreen(const std::map<std::string, std::string>
             if (!payloads.empty()) {
                 DrawUtils::print(16, SCREEN_HEIGHT - 8, "\ue07d Navigate ");
                 DrawUtils::print(SCREEN_WIDTH - 16, SCREEN_HEIGHT - 8, "\ue000 Choose", true);
-                const char *autobootHints = "\ue002 Clear Default / \ue003 Select Default";
+                const char *autobootHints = "\ue002/\ue046 Clear Default / \ue003/\ue045 Select Default";
                 DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(autobootHints) / 2, SCREEN_HEIGHT - 8, autobootHints, true);
             } else {
                 DrawUtils::print(SCREEN_WIDTH - 20, SCREEN_HEIGHT - 8, "\ue000 Wii U Menu", true);
             }
 
             DrawUtils::endDraw();
-
-            redraw = false;
         }
     }
 
@@ -382,6 +585,9 @@ std::string EnvironmentSelectionScreen(const std::map<std::string, std::string>
     // Call GX2Init to shut down OSScreen
     GX2Init(nullptr);
 
+    GX2SetTVEnable(FALSE);
+    GX2SetDRCEnable(FALSE);
+
     free(screenBuffer);
 
     if (autoBoot != autobootIndex) {

source/module/ImportRPLInformation.h

@@ -39,7 +39,7 @@ public:
 
     [[nodiscard]] const char *getRPLName() const {
         if (name.max_size() < strlen("._import_") + 1) {
-            OSFatal("Invalid RPLName, is too short to be valid");
+            OSFatal("EnvironmentLoader: Invalid RPLName, is too short to be valid");
         }
         return name.c_str() + strlen("._import_");
     }

source/module/ModuleData.h

@@ -18,6 +18,7 @@
 #pragma once
 
 #include "RelocationData.h"
+#include "utils/MemoryUtils.h"
 #include <map>
 #include <set>
 #include <string>
@@ -33,11 +34,11 @@ public:
         this->entrypoint = address;
     }
 
-    void addRelocationData(std::unique_ptr<RelocationData> relocation_data) {
+    void addRelocationData(RelocationData relocation_data) {
         relocation_data_list.push_back(std::move(relocation_data));
     }
 
-    [[nodiscard]] const std::vector<std::unique_ptr<RelocationData>> &getRelocationDataList() const {
+    [[nodiscard]] const std::vector<RelocationData> &getRelocationDataList() const {
         return relocation_data_list;
     }
 
@@ -45,25 +46,16 @@ public:
         return entrypoint;
     }
 
-    void setStartAddress(uint32_t address) {
-        this->startAddress = address;
+    void setTextMemory(ExpHeapMemory &&memory) {
+        mTextMemory = std::move(memory);
     }
-
-    void setEndAddress(uint32_t address) {
-        this->endAddress = address;
-    }
-
-    [[nodiscard]] uint32_t getStartAddress() const {
-        return startAddress;
-    }
-
-    [[nodiscard]] uint32_t getEndAddress() const {
-        return endAddress;
+    void setDataMemory(ExpHeapMemory &&memory) {
+        mDataMemory = std::move(memory);
     }
 
 private:
-    std::vector<std::unique_ptr<RelocationData>> relocation_data_list;
-    uint32_t entrypoint   = 0;
-    uint32_t startAddress = 0;
-    uint32_t endAddress   = 0;
+    std::vector<RelocationData> relocation_data_list;
+    uint32_t entrypoint = 0;
+    ExpHeapMemory mTextMemory;
+    ExpHeapMemory mDataMemory;
 };

source/module/ModuleDataFactory.cpp

@@ -24,33 +24,32 @@
 #include <coreinit/cache.h>
 #include <map>
 #include <string>
-#include <vector>
+
+
+uint32_t ModuleDataFactory::GetSizeOfModule(const ELFIO::elfio &reader) {
+    uint32_t sec_num      = reader.sections.size();
+    uint32_t sizeOfModule = 0;
+    for (uint32_t i = 0; i < sec_num; ++i) {
+        ELFIO::section *psec = reader.sections[i];
+        if (psec->get_type() == 0x80000002 || psec->get_name() == ".wut_load_bounds") {
+            continue;
+        }
+
+        if ((psec->get_type() == ELFIO::SHT_PROGBITS || psec->get_type() == ELFIO::SHT_NOBITS) && (psec->get_flags() & ELFIO::SHF_ALLOC)) {
+            sizeOfModule += psec->get_size() + psec->get_addr_align();
+        }
+    }
+    return sizeOfModule;
+}
 
 std::optional<std::unique_ptr<ModuleData>>
-ModuleDataFactory::load(const std::string &path, uint32_t destination_address_end, uint32_t maximum_size, relocation_trampoline_entry_t *trampoline_data, uint32_t trampoline_data_length) {
-    ELFIO::elfio reader(new wiiu_zlib);
+ModuleDataFactory::load(const ELFIO::elfio &reader, const HeapWrapper &heapWrapper, relocation_trampoline_entry_t *trampoline_data, uint32_t trampoline_data_length) {
     auto moduleData = make_unique_nothrow<ModuleData>();
     if (!moduleData) {
         DEBUG_FUNCTION_LINE_ERR("Failed to allocate ModuleData");
         return {};
     }
 
-    uint8_t *buffer = nullptr;
-    uint32_t fsize  = 0;
-    if (LoadFileToMem(path.c_str(), &buffer, &fsize) < 0) {
-        DEBUG_FUNCTION_LINE_ERR("Failed to load file");
-        return {};
-    }
-
-    // Load ELF data
-    if (!reader.load(reinterpret_cast<char *>(buffer), fsize)) {
-        DEBUG_FUNCTION_LINE_ERR("Can't find or process %s", path.c_str());
-        free(buffer);
-        return {};
-    }
-
-    auto cleanupBuffer = onLeavingScope([buffer]() { free(buffer); });
-
     uint32_t sec_num = reader.sections.size();
 
     auto destinations = make_unique_nothrow<uint8_t *[]>(sec_num);
@@ -59,7 +58,9 @@ ModuleDataFactory::load(const std::string &path, uint32_t destination_address_en
         return {};
     }
 
-    uint32_t sizeOfModule = 0;
+    uint32_t text_size = 0;
+    uint32_t data_size = 0;
+
     for (uint32_t i = 0; i < sec_num; ++i) {
         ELFIO::section *psec = reader.sections[i];
         if (psec->get_type() == 0x80000002) {
@@ -67,25 +68,31 @@ ModuleDataFactory::load(const std::string &path, uint32_t destination_address_en
         }
 
         if ((psec->get_type() == ELFIO::SHT_PROGBITS || psec->get_type() == ELFIO::SHT_NOBITS) && (psec->get_flags() & ELFIO::SHF_ALLOC)) {
-            sizeOfModule += psec->get_size() + 1;
+            uint32_t sectionSize = psec->get_size();
+            auto address         = (uint32_t) psec->get_address();
+            if ((address >= 0x02000000) && address < 0x10000000) {
+                text_size += sectionSize + psec->get_addr_align();
+            } else if ((address >= 0x10000000) && address < 0xC0000000) {
+                data_size += sectionSize + psec->get_addr_align();
+            }
         }
     }
 
-    if (sizeOfModule > maximum_size) {
-        DEBUG_FUNCTION_LINE_ERR("Module is too big.");
-        return {};
+    auto text_dataOpt = heapWrapper.Alloc(text_size, 0x100);
+    if (!text_dataOpt) {
+        DEBUG_FUNCTION_LINE_ERR("Failed to alloc memory for the .text section (%d bytes)", text_size);
+        return std::nullopt;
     }
+    ExpHeapMemory text_data = std::move(*text_dataOpt);
 
-    uint32_t baseOffset   = (destination_address_end - sizeOfModule) & 0xFFFFFF00;
-    uint32_t startAddress = baseOffset;
+    auto data_dataOpt = heapWrapper.Alloc(data_size, 0x100);
+    if (!data_dataOpt) {
+        DEBUG_FUNCTION_LINE_ERR("Failed to alloc memory for the .data section (%d bytes)", data_size);
+        return std::nullopt;
+    }
+    ExpHeapMemory data_data = std::move(*data_dataOpt);
 
-    uint32_t offset_text = baseOffset;
-    uint32_t offset_data = offset_text;
-
-    uint32_t entrypoint = offset_text + (uint32_t) reader.get_entry() - 0x02000000;
-
-    uint32_t totalSize  = 0;
-    uint32_t endAddress = 0;
+    uint32_t entrypoint = (uint32_t) text_data.data() + (uint32_t) reader.get_entry() - 0x02000000;
 
     for (uint32_t i = 0; i < sec_num; ++i) {
         ELFIO::section *psec = reader.sections[i];
@@ -95,26 +102,34 @@ ModuleDataFactory::load(const std::string &path, uint32_t destination_address_en
 
         if ((psec->get_type() == ELFIO::SHT_PROGBITS || psec->get_type() == ELFIO::SHT_NOBITS) && (psec->get_flags() & ELFIO::SHF_ALLOC)) {
             uint32_t sectionSize = psec->get_size();
+            auto address         = (uint32_t) psec->get_address();
 
-            totalSize += sectionSize;
-            if (totalSize > maximum_size) {
-                DEBUG_FUNCTION_LINE_ERR("Couldn't load setup module because it's too big.");
-                return {};
-            }
+            uint32_t destination = address;
 
-            auto address = (uint32_t) psec->get_address();
-
-            destinations[psec->get_index()] = (uint8_t *) baseOffset;
-
-            uint32_t destination = baseOffset + address;
             if ((address >= 0x02000000) && address < 0x10000000) {
+                destination += (uint32_t) text_data.data();
                 destination -= 0x02000000;
-                destinations[psec->get_index()] -= 0x02000000;
-                baseOffset += sectionSize;
-                offset_data += sectionSize;
+                destinations[psec->get_index()] = (uint8_t *) text_data.data();
+
+                if (destination + sectionSize > (uint32_t) text_data.data() + text_size) {
+                    DEBUG_FUNCTION_LINE_ERR("Tried to overflow .text buffer. %08X > %08X", destination + sectionSize, (uint32_t) text_data.data() + text_data.size());
+                    OSFatal("EnvironmentLoader: Tried to overflow .text buffer");
+                } else if (destination < (uint32_t) text_data.data()) {
+                    DEBUG_FUNCTION_LINE_ERR("Tried to underflow .text buffer. %08X < %08X", destination, (uint32_t) text_data.data());
+                    OSFatal("EnvironmentLoader: Tried to underflow .text buffer");
+                }
             } else if ((address >= 0x10000000) && address < 0xC0000000) {
+                destination += (uint32_t) data_data.data();
                 destination -= 0x10000000;
-                destinations[psec->get_index()] -= 0x10000000;
+                destinations[psec->get_index()] = (uint8_t *) data_data.data();
+
+                if (destination + sectionSize > (uint32_t) data_data.data() + data_data.size()) {
+                    DEBUG_FUNCTION_LINE_ERR("Tried to overflow .data buffer. %08X > %08X", destination + sectionSize, (uint32_t) data_data.data() + data_data.size());
+                    OSFatal("EnvironmentLoader: Tried to overflow .data buffer");
+                } else if (destination < (uint32_t) data_data.data()) {
+                    DEBUG_FUNCTION_LINE_ERR("Tried to underflow .data buffer. %08X < %08X", destination, (uint32_t) data_data.data());
+                    OSFatal("EnvironmentLoader: Tried to underflow .data buffer");
+                }
             } else if (address >= 0xC0000000) {
                 DEBUG_FUNCTION_LINE_ERR("Loading section from 0xC0000000 is NOT supported");
                 return std::nullopt;
@@ -123,30 +138,22 @@ ModuleDataFactory::load(const std::string &path, uint32_t destination_address_en
                 return std::nullopt;
             }
 
-            const char *p = reader.sections[i]->get_data();
+            const char *p = psec->get_data();
 
-            if (destination + sectionSize > (uint32_t) destination_address_end) {
-                DEBUG_FUNCTION_LINE_ERR("Tried to overflow buffer. %08X > %08X", destination + sectionSize, destination_address_end);
-                OSFatal("EnvironmentLoader: Tried to overflow buffer");
+            uint32_t address_align = psec->get_addr_align();
+            if ((destination & (address_align - 1)) != 0) {
+                DEBUG_FUNCTION_LINE_WARN("Address not aligned: %08X %08X", destination, address_align);
+                OSFatal("EnvironmentLoader: Address not aligned");
             }
 
             if (psec->get_type() == ELFIO::SHT_NOBITS) {
-                DEBUG_FUNCTION_LINE("memset section %s %08X to 0 (%d bytes)", psec->get_name().c_str(), destination, sectionSize);
+                DEBUG_FUNCTION_LINE_VERBOSE("memset section %s %08X to 0 (%d bytes)", psec->get_name().c_str(), destination, sectionSize);
                 memset((void *) destination, 0, sectionSize);
             } else if (psec->get_type() == ELFIO::SHT_PROGBITS) {
-                DEBUG_FUNCTION_LINE("Copy section %s %08X -> %08X (%d bytes)", psec->get_name().c_str(), p, destination, sectionSize);
+                DEBUG_FUNCTION_LINE_VERBOSE("Copy section %s %08X -> %08X (%d bytes)", psec->get_name().c_str(), p, destination, sectionSize);
                 memcpy((void *) destination, p, sectionSize);
             }
-
-            //nextAddress = ROUNDUP(destination + sectionSize, 0x100);
-            if (psec->get_name() == ".bss" || psec->get_name() == ".sbss") {
-                DEBUG_FUNCTION_LINE("memset %s section. Location: %08X size: %08X", psec->get_name().c_str(), destination, sectionSize);
-                memset(reinterpret_cast<void *>(destination), 0, sectionSize);
-            }
-
-            if (endAddress < destination + sectionSize) {
-                endAddress = destination + sectionSize;
-            }
+            DEBUG_FUNCTION_LINE_VERBOSE("Saved %s section info. Location: %08X size: %08X", psec->get_name().c_str(), destination, sectionSize);
 
             DCFlushRange((void *) destination, sectionSize);
             ICInvalidateRange((void *) destination, sectionSize);
@@ -157,7 +164,7 @@ ModuleDataFactory::load(const std::string &path, uint32_t destination_address_en
         ELFIO::section *psec = reader.sections[i];
         if ((psec->get_type() == ELFIO::SHT_PROGBITS || psec->get_type() == ELFIO::SHT_NOBITS) && (psec->get_flags() & ELFIO::SHF_ALLOC)) {
             DEBUG_FUNCTION_LINE("Linking (%d)... %s", i, psec->get_name().c_str());
-            if (!linkSection(reader, psec->get_index(), (uint32_t) destinations[psec->get_index()], offset_text, offset_data, trampoline_data, trampoline_data_length)) {
+            if (!linkSection(reader, psec->get_index(), (uint32_t) destinations[psec->get_index()], (uint32_t) (text_data.data()), (uint32_t) (data_data.data()), trampoline_data, trampoline_data_length)) {
                 DEBUG_FUNCTION_LINE_ERR("elfLink failed");
                 return std::nullopt;
             }
@@ -165,18 +172,19 @@ ModuleDataFactory::load(const std::string &path, uint32_t destination_address_en
     }
     getImportRelocationData(moduleData, reader, destinations.get());
 
-    DCFlushRange((void *) baseOffset, totalSize);
-    ICInvalidateRange((void *) baseOffset, totalSize);
+    DCFlushRange((void *) data_data.data(), data_data.size());
+    ICInvalidateRange((void *) text_data.data(), text_data.size());
 
-    moduleData->setStartAddress(startAddress);
-    moduleData->setEndAddress(endAddress);
     moduleData->setEntrypoint(entrypoint);
+    moduleData->setTextMemory(std::move(text_data));
+    moduleData->setDataMemory(std::move(data_data));
+
     DEBUG_FUNCTION_LINE("Saved entrypoint as %08X", entrypoint);
 
     return moduleData;
 }
 
-bool ModuleDataFactory::getImportRelocationData(std::unique_ptr<ModuleData> &moduleData, ELFIO::elfio &reader, uint8_t **destinations) {
+bool ModuleDataFactory::getImportRelocationData(std::unique_ptr<ModuleData> &moduleData, const ELFIO::elfio &reader, uint8_t **destinations) {
     std::map<uint32_t, std::shared_ptr<ImportRPLInformation>> infoMap;
 
     uint32_t sec_num = reader.sections.size();
@@ -196,6 +204,7 @@ bool ModuleDataFactory::getImportRelocationData(std::unique_ptr<ModuleData> &mod
     for (uint32_t i = 0; i < sec_num; ++i) {
         ELFIO::section *psec = reader.sections[i];
         if (psec->get_type() == ELFIO::SHT_RELA || psec->get_type() == ELFIO::SHT_REL) {
+            DEBUG_FUNCTION_LINE_VERBOSE("Found relocation section %s", psec->get_name().c_str());
             ELFIO::relocation_section_accessor rel(reader, psec);
             for (uint32_t j = 0; j < (uint32_t) rel.get_entries_num(); ++j) {
                 ELFIO::Elf_Word symbol = 0;
@@ -232,29 +241,23 @@ bool ModuleDataFactory::getImportRelocationData(std::unique_ptr<ModuleData> &mod
                 uint32_t section_index = psec->get_info();
                 if (!infoMap.contains(sym_section_index)) {
                     DEBUG_FUNCTION_LINE_ERR("Relocation is referencing a unknown section. %d destination: %08X sym_name %s", section_index, destinations[section_index], sym_name.c_str());
-                    OSFatal("Relocation is referencing a unknown section.");
-                }
-
-                auto relocationData = make_unique_nothrow<RelocationData>(type,
-                                                                          offset - 0x02000000,
-                                                                          addend,
-                                                                          (void *) (destinations[section_index] + 0x02000000),
-                                                                          sym_name,
-                                                                          infoMap[sym_section_index]);
-
-                if (!relocationData) {
-                    DEBUG_FUNCTION_LINE_ERR("Failed to alloc relocation data");
+                    OSFatal("EnvironmentLoader: Relocation is referencing a unknown section.");
                     return false;
                 }
 
-                moduleData->addRelocationData(std::move(relocationData));
+                moduleData->addRelocationData(RelocationData(type,
+                                                             offset - 0x02000000,
+                                                             addend,
+                                                             (void *) (destinations[section_index]),
+                                                             sym_name,
+                                                             infoMap[sym_section_index]));
             }
         }
     }
     return true;
 }
 
-bool ModuleDataFactory::linkSection(ELFIO::elfio &reader, uint32_t section_index, uint32_t destination, uint32_t base_text, uint32_t base_data, relocation_trampoline_entry_t *trampoline_data,
+bool ModuleDataFactory::linkSection(const ELFIO::elfio &reader, uint32_t section_index, uint32_t destination, uint32_t base_text, uint32_t base_data, relocation_trampoline_entry_t *trampoline_data,
                                     uint32_t trampoline_data_length) {
     uint32_t sec_num = reader.sections.size();
 
@@ -307,17 +310,27 @@ bool ModuleDataFactory::linkSection(ELFIO::elfio &reader, uint32_t section_index
                     return false;
                 }
 
+                auto adjusted_offset = (uint32_t) offset;
+                if ((offset >= 0x02000000) && offset < 0x10000000) {
+                    adjusted_offset -= 0x02000000;
+                } else if ((adjusted_offset >= 0x10000000) && adjusted_offset < 0xC0000000) {
+                    adjusted_offset -= 0x10000000;
+                } else if (adjusted_offset >= 0xC0000000) {
+                    adjusted_offset -= 0xC0000000;
+                }
+
                 if (sym_section_index == ELFIO::SHN_ABS) {
                     //
                 } else if (sym_section_index > ELFIO::SHN_LORESERVE) {
                     DEBUG_FUNCTION_LINE_ERR("NOT IMPLEMENTED: %04X", sym_section_index);
                     return false;
                 }
-                if (!ElfUtils::elfLinkOne(type, offset, addend, destination, adjusted_sym_value, trampoline_data, trampoline_data_length, RELOC_TYPE_FIXED)) {
+                if (!ElfUtils::elfLinkOne(type, adjusted_offset, addend, destination, adjusted_sym_value, trampoline_data, trampoline_data_length, RELOC_TYPE_FIXED)) {
                     DEBUG_FUNCTION_LINE_ERR("Link failed");
                     return false;
                 }
             }
+            return true;
         }
     }
     return true;

source/module/ModuleDataFactory.h

@@ -20,17 +20,20 @@
 #include "../common/relocation_defines.h"
 #include "ModuleData.h"
 #include "elfio/elfio.hpp"
+#include "utils/MemoryUtils.h"
+#include "utils/utils.h"
 #include <map>
 #include <string>
 #include <vector>
 
 class ModuleDataFactory {
 public:
-    static std::optional<std::unique_ptr<ModuleData>>
-    load(const std::string &path, uint32_t destination_address_end, uint32_t maximum_size, relocation_trampoline_entry_t *trampoline_data, uint32_t trampoline_data_length);
+    static uint32_t GetSizeOfModule(const ELFIO::elfio &reader);
 
-    static bool linkSection(ELFIO::elfio &reader, uint32_t section_index, uint32_t destination, uint32_t base_text, uint32_t base_data, relocation_trampoline_entry_t *trampoline_data,
+    static std::optional<std::unique_ptr<ModuleData>> load(const ELFIO::elfio &reader, const HeapWrapper &heapWrapper, relocation_trampoline_entry_t *trampoline_data, uint32_t trampoline_data_length);
+
+    static bool linkSection(const ELFIO::elfio &reader, uint32_t section_index, uint32_t destination, uint32_t base_text, uint32_t base_data, relocation_trampoline_entry_t *trampoline_data,
                             uint32_t trampoline_data_length);
 
-    static bool getImportRelocationData(std::unique_ptr<ModuleData> &moduleData, ELFIO::elfio &reader, uint8_t **destinations);
+    static bool getImportRelocationData(std::unique_ptr<ModuleData> &moduleData, const ELFIO::elfio &reader, uint8_t **destinations);
 };

source/utils/DrawUtils.h

@@ -3,9 +3,19 @@
 #include "schrift.h"
 #include <cstdint>
 
+#define COLOR_WHITE              Color(0xffffffff)
+#define COLOR_BLACK              Color(0, 0, 0, 255)
+#define COLOR_RED                Color(237, 28, 36, 255)
+#define COLOR_BACKGROUND         Color(0, 40, 100, 255)
+#define COLOR_TEXT               COLOR_WHITE
+#define COLOR_TEXT2              Color(0xB3ffffff)
+#define COLOR_AUTOBOOT           Color(0xaeea00ff)
+#define COLOR_BORDER             Color(204, 204, 204, 255)
+#define COLOR_BORDER_HIGHLIGHTED Color(0x3478e4ff)
+
 // visible screen sizes
-#define SCREEN_WIDTH  854
-#define SCREEN_HEIGHT 480
+#define SCREEN_WIDTH             854
+#define SCREEN_HEIGHT            480
 
 union Color {
     explicit Color(uint32_t color) {

source/utils/InputUtils.cpp

@@ -0,0 +1,140 @@
+#include "InputUtils.h"
+#include <coreinit/thread.h>
+#include <padscore/kpad.h>
+#include <padscore/wpad.h>
+#include <vpad/input.h>
+
+uint32_t remapWiiMoteButtons(uint32_t buttons) {
+    uint32_t convButtons = 0;
+
+    if (buttons & WPAD_BUTTON_LEFT)
+        convButtons |= VPAD_BUTTON_LEFT;
+
+    if (buttons & WPAD_BUTTON_RIGHT)
+        convButtons |= VPAD_BUTTON_RIGHT;
+
+    if (buttons & WPAD_BUTTON_DOWN)
+        convButtons |= VPAD_BUTTON_DOWN;
+
+    if (buttons & WPAD_BUTTON_UP)
+        convButtons |= VPAD_BUTTON_UP;
+
+    if (buttons & WPAD_BUTTON_PLUS)
+        convButtons |= VPAD_BUTTON_PLUS;
+
+    if (buttons & WPAD_BUTTON_2)
+        convButtons |= VPAD_BUTTON_Y;
+
+    if (buttons & WPAD_BUTTON_1)
+        convButtons |= VPAD_BUTTON_X;
+
+    if (buttons & WPAD_BUTTON_B)
+        convButtons |= VPAD_BUTTON_B;
+
+    if (buttons & WPAD_BUTTON_A)
+        convButtons |= VPAD_BUTTON_A;
+
+    if (buttons & WPAD_BUTTON_MINUS)
+        convButtons |= VPAD_BUTTON_MINUS;
+
+    if (buttons & WPAD_BUTTON_HOME)
+        convButtons |= VPAD_BUTTON_HOME;
+
+    return convButtons;
+}
+
+uint32_t remapClassicButtons(uint32_t buttons) {
+    uint32_t convButtons = 0;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_LEFT)
+        convButtons |= VPAD_BUTTON_LEFT;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_RIGHT)
+        convButtons |= VPAD_BUTTON_RIGHT;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_DOWN)
+        convButtons |= VPAD_BUTTON_DOWN;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_UP)
+        convButtons |= VPAD_BUTTON_UP;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_PLUS)
+        convButtons |= VPAD_BUTTON_PLUS;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_X)
+        convButtons |= VPAD_BUTTON_X;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_Y)
+        convButtons |= VPAD_BUTTON_Y;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_B)
+        convButtons |= VPAD_BUTTON_B;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_A)
+        convButtons |= VPAD_BUTTON_A;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_MINUS)
+        convButtons |= VPAD_BUTTON_MINUS;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_HOME)
+        convButtons |= VPAD_BUTTON_HOME;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_ZR)
+        convButtons |= VPAD_BUTTON_ZR;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_ZL)
+        convButtons |= VPAD_BUTTON_ZL;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_R)
+        convButtons |= VPAD_BUTTON_R;
+
+    if (buttons & WPAD_CLASSIC_BUTTON_L)
+        convButtons |= VPAD_BUTTON_L;
+
+    return convButtons;
+}
+
+InputUtils::InputData InputUtils::getControllerInput() {
+    InputData inputData     = {};
+    VPADStatus vpadStatus   = {};
+    VPADReadError vpadError = VPAD_READ_UNINITIALIZED;
+    int maxAttempts         = 100;
+    do {
+        if (VPADRead(VPAD_CHAN_0, &vpadStatus, 1, &vpadError) > 0 && vpadError == VPAD_READ_SUCCESS) {
+            inputData.trigger = vpadStatus.trigger;
+            inputData.hold    = vpadStatus.hold;
+            inputData.release = vpadStatus.release;
+        } else {
+            OSSleepTicks(OSMillisecondsToTicks(1));
+        }
+    } while (--maxAttempts > 0 && vpadError == VPAD_READ_NO_SAMPLES);
+
+    KPADStatus kpadStatus = {};
+    KPADError kpadError   = KPAD_ERROR_UNINITIALIZED;
+    for (int32_t i = 0; i < 4; i++) {
+        if (KPADReadEx((KPADChan) i, &kpadStatus, 1, &kpadError) > 0) {
+            if (kpadError == KPAD_ERROR_OK && kpadStatus.extensionType != 0xFF) {
+                if (kpadStatus.extensionType == WPAD_EXT_CORE || kpadStatus.extensionType == WPAD_EXT_NUNCHUK) {
+                    inputData.trigger |= remapWiiMoteButtons(kpadStatus.trigger);
+                    inputData.hold |= remapWiiMoteButtons(kpadStatus.hold);
+                    inputData.release |= remapWiiMoteButtons(kpadStatus.release);
+                } else {
+                    inputData.trigger |= remapClassicButtons(kpadStatus.classic.trigger);
+                    inputData.hold |= remapClassicButtons(kpadStatus.classic.hold);
+                    inputData.release |= remapClassicButtons(kpadStatus.classic.release);
+                }
+            }
+        }
+    }
+
+    return inputData;
+}
+
+void InputUtils::Init() {
+    KPADInit();
+    WPADEnableURCC(1);
+}
+
+void InputUtils::DeInit() {
+    KPADShutdown();
+}

source/utils/InputUtils.h

@@ -0,0 +1,17 @@
+#pragma once
+#include <cstdint>
+#include <vpad/input.h>
+
+class InputUtils {
+public:
+    typedef struct InputData {
+        uint32_t trigger = 0;
+        uint32_t hold    = 0;
+        uint32_t release = 0;
+    } InputData;
+
+    static void Init();
+    static void DeInit();
+
+    static InputData getControllerInput();
+};

source/utils/MemoryUtils.h

@@ -0,0 +1,199 @@
+#pragma once
+#include "logger.h"
+#include <coreinit/memexpheap.h>
+#include <coreinit/memheap.h>
+#include <cstdint>
+#include <cstring>
+#include <optional>
+
+typedef void (*FreeMemoryFn)(void *);
+
+class MemoryWrapper {
+public:
+    MemoryWrapper(void *ptr, uint32_t size, FreeMemoryFn freeFn) : mPtr(ptr), mSize(size), mFreeFn(freeFn) {
+    }
+    ~MemoryWrapper() {
+        if (mPtr && mFreeFn) {
+            memset(mPtr, 0, mSize);
+            mFreeFn(mPtr);
+        }
+    }
+
+    MemoryWrapper(const MemoryWrapper &) = delete;
+    MemoryWrapper &operator=(const MemoryWrapper &) = delete;
+
+    MemoryWrapper(MemoryWrapper &&other) noexcept
+        : mPtr(other.mPtr), mSize(other.mSize), mFreeFn(other.mFreeFn) {
+        other.mPtr    = {};
+        other.mSize   = {};
+        other.mFreeFn = {};
+    }
+
+    MemoryWrapper &operator=(MemoryWrapper &&other) noexcept {
+        if (this != &other) {
+            mPtr          = other.mPtr;
+            mSize         = other.mSize;
+            mFreeFn       = other.mFreeFn;
+            other.mPtr    = {};
+            other.mSize   = 0;
+            other.mFreeFn = {};
+        }
+        return *this;
+    }
+
+    [[nodiscard]] void *data() const {
+        return mPtr;
+    }
+
+    [[nodiscard]] uint32_t size() const {
+        return mSize;
+    }
+
+    [[nodiscard]] bool IsAllocated() const {
+        return mFreeFn && mPtr && mSize > 0;
+    }
+
+private:
+    void *mPtr           = {};
+    uint32_t mSize       = 0;
+    FreeMemoryFn mFreeFn = {};
+};
+
+
+class ExpHeapMemory {
+
+public:
+    ExpHeapMemory(MEMHeapHandle heapHandle, void *data, uint32_t size) : mHeapHandle(heapHandle),
+                                                                         mData(data),
+                                                                         mSize(size) {
+    }
+    ExpHeapMemory() = default;
+
+    ~ExpHeapMemory() {
+        if (mData) {
+            MEMFreeToExpHeap(mHeapHandle, mData);
+        }
+        mData = nullptr;
+        mSize = 0;
+    }
+
+    // Delete the copy constructor and copy assignment operator
+    ExpHeapMemory(const ExpHeapMemory &) = delete;
+    ExpHeapMemory &operator=(const ExpHeapMemory &) = delete;
+
+    ExpHeapMemory(ExpHeapMemory &&other) noexcept
+        : mHeapHandle(other.mHeapHandle), mData(other.mData), mSize(other.mSize) {
+        other.mHeapHandle = {};
+        other.mData       = {};
+        other.mSize       = 0;
+    }
+
+    ExpHeapMemory &operator=(ExpHeapMemory &&other) noexcept {
+        if (this != &other) {
+            mHeapHandle       = other.mHeapHandle;
+            mData             = other.mData;
+            mSize             = other.mSize;
+            other.mHeapHandle = {};
+            other.mData       = {};
+            other.mSize       = 0;
+        }
+        return *this;
+    }
+
+    explicit operator bool() const {
+        return mData != nullptr;
+    }
+
+    explicit operator void *() const {
+        // Return the desired void* value
+        return mData;
+    }
+
+    [[nodiscard]] void *data() const {
+        return mData;
+    }
+
+    [[nodiscard]] std::size_t size() const {
+        return mSize;
+    }
+
+    static std::optional<ExpHeapMemory> Alloc(MEMHeapHandle heapHandle, uint32_t size, int32_t alignment) {
+        auto *ptr = MEMAllocFromExpHeapEx(heapHandle, size, alignment);
+        if (!ptr) {
+            return {};
+        }
+
+        return ExpHeapMemory(heapHandle, ptr, size);
+    }
+
+private:
+    MEMHeapHandle mHeapHandle{};
+    void *mData = nullptr;
+    uint32_t mSize{};
+};
+
+class HeapWrapper {
+public:
+    explicit HeapWrapper(MemoryWrapper &&memory) : mMemory(std::move(memory)) {
+        mHeapHandle = MEMCreateExpHeapEx(mMemory.data(), mMemory.size(), MEM_HEAP_FLAG_USE_LOCK);
+        if (mHeapHandle) {
+            mSize = mMemory.size();
+            mPtr  = mMemory.data();
+        }
+    }
+
+    ~HeapWrapper() {
+        if (mHeapHandle) {
+            MEMDestroyExpHeap(mHeapHandle);
+        }
+        if (mPtr) {
+            memset(mPtr, 0, mSize);
+        }
+    }
+
+    // Delete the copy constructor and copy assignment operator
+    HeapWrapper(const HeapWrapper &) = delete;
+    HeapWrapper &operator=(const HeapWrapper &) = delete;
+
+    HeapWrapper(HeapWrapper &&other) noexcept
+        : mMemory(std::move(other.mMemory)), mHeapHandle(other.mHeapHandle), mPtr(other.mPtr), mSize(other.mSize) {
+        other.mHeapHandle = {};
+        other.mPtr        = {};
+        other.mSize       = 0;
+    }
+
+    HeapWrapper &operator=(HeapWrapper &&other) noexcept {
+        if (this != &other) {
+            mMemory           = std::move(other.mMemory);
+            mHeapHandle       = other.mHeapHandle;
+            mPtr              = other.mPtr;
+            mSize             = other.mSize;
+            other.mHeapHandle = {};
+            other.mPtr        = {};
+            other.mSize       = 0;
+        }
+        return *this;
+    }
+
+    [[nodiscard]] MEMHeapHandle GetHeapHandle() const {
+        return mHeapHandle;
+    }
+
+    [[nodiscard]] uint32_t GetHeapSize() const {
+        return mSize;
+    }
+
+    [[nodiscard]] bool IsAllocated() const {
+        return mMemory.IsAllocated();
+    }
+
+    [[nodiscard]] std::optional<ExpHeapMemory> Alloc(uint32_t size, int align) const {
+        return ExpHeapMemory::Alloc(mHeapHandle, size, align);
+    }
+
+private:
+    MemoryWrapper mMemory;
+    MEMHeapHandle mHeapHandle = {};
+    void *mPtr                = {};
+    uint32_t mSize            = 0;
+};

source/utils/PairUtils.cpp

@@ -0,0 +1,239 @@
+#include "PairUtils.h"
+#include "DrawUtils.h"
+#include "InputUtils.h"
+#include "logger.h"
+#include "utils.h"
+#include <coreinit/cache.h>
+#include <coreinit/thread.h>
+#include <malloc.h>
+#include <nn/ccr/sys.h>
+#include <padscore/kpad.h>
+#include <padscore/wpad.h>
+#include <vpad/input.h>
+
+void PairMenu::drawPairKPADScreen() const {
+    DrawUtils::beginDraw();
+    DrawUtils::clear(COLOR_BACKGROUND);
+
+    DrawUtils::setFontColor(COLOR_TEXT);
+
+    DrawUtils::setFontSize(26);
+
+    std::string textLine1 = "Press the SYNC Button on the controller you want to pair.";
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(textLine1.c_str()) / 2, 40, textLine1.c_str(), true);
+
+
+    WPADExtensionType ext{};
+    for (int i = 0; i < 4; i++) {
+        bool isConnected     = WPADProbe((WPADChan) i, &ext) == 0;
+        std::string textLine = string_format("Slot %d: ", i + 1);
+        if (isConnected) {
+            textLine += ext == WPAD_EXT_PRO_CONTROLLER ? "Pro Controller" : "Wiimote";
+        } else {
+            textLine += "No controller";
+        }
+
+        DrawUtils::print(300, 140 + (i * 30), textLine.c_str());
+    }
+
+    DrawUtils::setFontSize(26);
+
+    std::string gamepadSyncText1 = "If you are pairing a Wii U GamePad, press the SYNC Button";
+    std::string gamepadSyncText2 = "on your Wii U console one more time";
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(gamepadSyncText1.c_str()) / 2, SCREEN_HEIGHT - 100, gamepadSyncText1.c_str(), true);
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(gamepadSyncText2.c_str()) / 2, SCREEN_HEIGHT - 70, gamepadSyncText2.c_str(), true);
+
+    DrawUtils::setFontSize(16);
+
+    const char *exitHints = "Press \ue001 to return";
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(exitHints) / 2, SCREEN_HEIGHT - 8, exitHints, true);
+
+    DrawUtils::endDraw();
+}
+
+void PairMenu::drawPairScreen() const {
+    DrawUtils::beginDraw();
+    DrawUtils::clear(COLOR_BACKGROUND);
+
+    DrawUtils::setFontColor(COLOR_TEXT);
+
+    // Convert the pin to symbols and set the text
+    static char pinSymbols[][4] = {
+            "\u2660",
+            "\u2665",
+            "\u2666",
+            "\u2663"};
+
+    uint32_t pincode = mGamePadPincode;
+
+    std::string pin = std::string(pinSymbols[(pincode / 1000) % 10]) +
+                      pinSymbols[(pincode / 100) % 10] +
+                      pinSymbols[(pincode / 10) % 10] +
+                      pinSymbols[pincode % 10];
+
+    std::string textLine1 = "Press the SYNC Button on the Wii U GamePad,";
+    std::string textLine2 = "and enter the four symbols shown below.";
+
+    DrawUtils::setFontSize(26);
+
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(textLine1.c_str()) / 2, 60, textLine1.c_str(), true);
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(textLine2.c_str()) / 2, 100, textLine2.c_str(), true);
+
+    DrawUtils::setFontSize(100);
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(pin.c_str()) / 2, (SCREEN_HEIGHT / 2) + 40, pin.c_str(), true);
+
+    DrawUtils::setFontSize(20);
+
+    std::string textLine3 = string_format("(%d seconds remaining) ", mGamePadSyncTimeout - (uint32_t) (OSTicksToSeconds(OSGetTime() - mSyncGamePadStartTime)));
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(textLine3.c_str()) / 2, SCREEN_HEIGHT - 80, textLine3.c_str(), true);
+
+    DrawUtils::setFontSize(26);
+
+    std::string textLine4 = "Press the SYNC Button on the Wii U console to exit.";
+    DrawUtils::print(SCREEN_WIDTH / 2 + DrawUtils::getTextWidth(textLine4.c_str()) / 2, SCREEN_HEIGHT - 40, textLine4.c_str(), true);
+
+    DrawUtils::endDraw();
+}
+
+PairMenu::PairMenu() {
+    CCRSysInit();
+
+    mState              = STATE_WAIT;
+    mGamePadSyncTimeout = 120;
+
+    // Initialize IM
+    mIMHandle = IM_Open();
+    if (mIMHandle < 0) {
+        DEBUG_FUNCTION_LINE_ERR("PairMenu: IM_Open failed");
+        OSFatal("EnvironmentLoader: PairMenu: IM_Open failed");
+    }
+    mIMRequest = (IMRequest *) memalign(0x40, sizeof(IMRequest));
+
+    // Allocate a separate request for IM_CancelGetEventNotify to avoid conflict with the pending IM_GetEventNotify request
+    mIMCancelRequest = (IMRequest *) memalign(0x40, sizeof(IMRequest));
+
+    if (!mIMRequest || !mIMCancelRequest) {
+        DEBUG_FUNCTION_LINE_ERR("Failed to allocate im request");
+        OSFatal("EnvironmentLoader: PairMenu: Failed to allocate im request");
+    }
+
+    mIMEventMask = IM_EVENT_SYNC;
+
+    // Notify about sync button events
+    IM_GetEventNotify(mIMHandle, mIMRequest, &mIMEventMask, PairMenu::SyncButtonCallback, this);
+}
+
+PairMenu::~PairMenu() {
+    // Close IM
+    IM_CancelGetEventNotify(mIMHandle, mIMCancelRequest, nullptr, nullptr);
+    IM_Close(mIMHandle);
+    if (mIMCancelRequest) {
+        free(mIMCancelRequest);
+        mIMCancelRequest = {};
+    }
+    if (mIMRequest) {
+        free(mIMRequest);
+        mIMRequest = {};
+    }
+
+    // Deinit CCRSys
+    CCRSysExit();
+}
+
+bool PairMenu::ProcessPairScreen() {
+    switch (mState) {
+        case STATE_SYNC_WPAD: {
+            // WPAD syncing stops after ~18 seconds, make sure to restart it.
+            if ((uint32_t) OSTicksToSeconds(OSGetTime() - mSyncWPADStartTime) >= 18) {
+                WPADStartSyncDevice();
+                mSyncWPADStartTime = OSGetTime();
+            }
+
+            InputUtils::InputData input = InputUtils::getControllerInput();
+
+            // Stop syncing when pressing A or B.
+            if (input.trigger & (VPAD_BUTTON_A | VPAD_BUTTON_B)) {
+                mState = STATE_WAIT;
+            }
+
+            break;
+        }
+        case STATE_SYNC_GAMEPAD: {
+            if (CCRSysGetPincode(&mGamePadPincode) != 0) {
+                DEBUG_FUNCTION_LINE_ERR("CCRSysGetPincode failed");
+                mState = STATE_WAIT;
+                break;
+            }
+
+            // Start pairing to slot 0
+            if (CCRSysStartPairing(0, mGamePadSyncTimeout) != 0) {
+                DEBUG_FUNCTION_LINE_ERR("CCRSysStartPairing failed.");
+                mState = STATE_WAIT;
+                break;
+            }
+
+            // Pairing has started, save start time
+            mSyncGamePadStartTime = OSGetTime();
+            mState                = STATE_PAIRING;
+
+            DEBUG_FUNCTION_LINE("Started GamePad syncing.");
+
+            break;
+        }
+        case STATE_PAIRING: {
+            // Get the current pairing state
+            CCRSysPairingState pairingState = CCRSysGetPairingState();
+            if (pairingState == CCR_SYS_PAIRING_TIMED_OUT) {
+                DEBUG_FUNCTION_LINE("GamePad SYNC timed out.");
+                // Pairing has timed out or was cancelled
+                CCRSysStopPairing();
+                mState = STATE_WAIT;
+            } else if (pairingState == CCR_SYS_PAIRING_FINISHED) {
+                DEBUG_FUNCTION_LINE("GamePad paired.");
+                mState = STATE_WAIT;
+            }
+            break;
+        }
+        case STATE_CANCEL: {
+            CCRSysStopPairing();
+            mState = STATE_WAIT;
+            break;
+        }
+        case STATE_WAIT:
+            break;
+    }
+    switch (mState) {
+        case STATE_WAIT: {
+            return false;
+        }
+        case STATE_SYNC_WPAD:
+            drawPairKPADScreen();
+            break;
+        case STATE_SYNC_GAMEPAD:
+        case STATE_PAIRING:
+        case STATE_CANCEL: {
+            drawPairScreen();
+            break;
+        }
+    }
+    return true;
+}
+
+
+void PairMenu::SyncButtonCallback(IOSError error, void *arg) {
+    auto *pairMenu = (PairMenu *) arg;
+
+    if (error == IOS_ERROR_OK && pairMenu && (pairMenu->mIMEventMask & IM_EVENT_SYNC)) {
+        if (pairMenu->mState == STATE_WAIT) {
+            pairMenu->mState = STATE_SYNC_WPAD;
+            // We need to restart the WPAD pairing every 18 seconds. For the timing we need to save the current time.
+            pairMenu->mSyncWPADStartTime = OSGetTime();
+        } else if (pairMenu->mState == STATE_SYNC_WPAD) {
+            pairMenu->mState = STATE_SYNC_GAMEPAD;
+        } else if (pairMenu->mState == STATE_SYNC_GAMEPAD || pairMenu->mState == STATE_PAIRING) {
+            pairMenu->mState = STATE_CANCEL;
+        }
+        OSMemoryBarrier();
+        IM_GetEventNotify(pairMenu->mIMHandle, pairMenu->mIMRequest, &pairMenu->mIMEventMask, PairMenu::SyncButtonCallback, pairMenu);
+    }
+}

source/utils/PairUtils.h

@@ -0,0 +1,43 @@
+#pragma once
+
+#include "logger.h"
+#include <coreinit/cache.h>
+#include <coreinit/im.h>
+#include <coreinit/ios.h>
+#include <coreinit/time.h>
+#include <malloc.h>
+#include <nn/ccr/sys.h>
+
+class PairMenu {
+public:
+    PairMenu();
+
+    ~PairMenu();
+
+    bool ProcessPairScreen();
+
+    static void SyncButtonCallback(IOSError error, void *arg);
+
+    void drawPairScreen() const;
+
+    void drawPairKPADScreen() const;
+
+private:
+    enum PairMenuState {
+        STATE_WAIT, // Wait for SYNC button press
+        STATE_SYNC_WPAD,
+        STATE_SYNC_GAMEPAD,
+        STATE_PAIRING,
+        STATE_CANCEL,
+    };
+
+    IOSHandle mIMHandle{};
+    IMRequest *mIMRequest{};
+    IMRequest *mIMCancelRequest{};
+    OSTime mSyncWPADStartTime    = 0;
+    OSTime mSyncGamePadStartTime = 0;
+    uint32_t mGamePadPincode     = 0;
+    PairMenuState mState         = STATE_WAIT;
+    uint32_t mGamePadSyncTimeout = 120;
+    IMEventMask mIMEventMask{};
+};

source/utils/logger.h

@@ -34,6 +34,7 @@ extern "C" {
 
 #define DEBUG_FUNCTION_LINE_WRITE(FMT, ARGS...) LOG(WHBLogWritef, FMT, ##ARGS)
 
+#define DEBUG_FUNCTION_LINE_WARN(FMT, ARGS...)  LOG_EX(WHBLogPrintf, "##WARN ## ", "", FMT, ##ARGS)
 #define DEBUG_FUNCTION_LINE_ERR(FMT, ARGS...)   LOG_EX(WHBLogPrintf, "##ERROR## ", "", FMT, ##ARGS)
 
 #else
@@ -44,6 +45,7 @@ extern "C" {
 
 #define DEBUG_FUNCTION_LINE_WRITE(FMT, ARGS...)   while (0)
 
+#define DEBUG_FUNCTION_LINE_WARN(FMT, ARGS...)    LOG_EX(OSReport, "##WARN ## ", "\n", FMT, ##ARGS)
 #define DEBUG_FUNCTION_LINE_ERR(FMT, ARGS...)     LOG_EX(OSReport, "##ERROR## ", "\n", FMT, ##ARGS)
 
 #endif

source/utils/utils.h

@@ -15,3 +15,16 @@ template<class T, class... Args>
 std::shared_ptr<T> make_shared_nothrow(Args &&...args) noexcept(noexcept(T(std::forward<Args>(args)...))) {
     return std::shared_ptr<T>(new (std::nothrow) T(std::forward<Args>(args)...));
 }
+
+template<typename... Args>
+std::string string_format(const std::string &format, Args... args) {
+    int size_s = std::snprintf(nullptr, 0, format.c_str(), args...) + 1; // Extra space for '\0'
+    auto size  = static_cast<size_t>(size_s);
+    auto buf   = std::make_unique<char[]>(size);
+    std::snprintf(buf.get(), size, format.c_str(), args...);
+    return std::string(buf.get(), buf.get() + size - 1); // We don't want the '\0' inside
+}
+
+// those work only in powers of 2
+#define ROUNDDOWN(val, align) ((val) & ~(align - 1))
+#define ROUNDUP(val, align)   ROUNDDOWN(((val) + (align - 1)), align)

source/version.h

@@ -0,0 +1,2 @@
+#pragma once
+#define ENVIRONMENT_LOADER_VERSION_EXTRA ""