38 files changed, 394 insertions, 147 deletions

diff --git a/src/common/fs/file.h b/src/common/fs/file.h
index 209f9664b..50e270c5b 100644
--- a/src/common/fs/file.h
+++ b/src/common/fs/file.h
@@ -117,7 +117,7 @@ template <typename Path>
 }
 #endif
 
-class IOFile final : NonCopyable {
+class IOFile final {
 public:
     IOFile();
 
@@ -142,7 +142,10 @@ public:
                     FileType type = FileType::BinaryFile,
                     FileShareFlag flag = FileShareFlag::ShareReadOnly);
 
-    virtual ~IOFile();
+    ~IOFile();
+
+    IOFile(const IOFile&) = delete;
+    IOFile& operator=(const IOFile&) = delete;
 
     IOFile(IOFile&& other) noexcept;
     IOFile& operator=(IOFile&& other) noexcept;
@@ -441,8 +444,8 @@ public:
 
 private:
     std::filesystem::path file_path;
-    FileAccessMode file_access_mode;
-    FileType file_type;
+    FileAccessMode file_access_mode{};
+    FileType file_type{};
 
     std::FILE* file = nullptr;
 };
diff --git a/src/core/arm/arm_interface.h b/src/core/arm/arm_interface.h
index 9a0151736..689e3ceb5 100644
--- a/src/core/arm/arm_interface.h
+++ b/src/core/arm/arm_interface.h
@@ -65,9 +65,6 @@ public:
     /// Step CPU by one instruction
     virtual void Step() = 0;
 
-    /// Exits execution from a callback, the callback must rewind the stack
-    virtual void ExceptionalExit() = 0;
-
     /// Clear all instruction cache
     virtual void ClearInstructionCache() = 0;
 
@@ -159,8 +156,6 @@ public:
      */
     virtual void SetTPIDR_EL0(u64 value) = 0;
 
-    virtual void ChangeProcessorID(std::size_t new_core_id) = 0;
-
     virtual void SaveContext(ThreadContext32& ctx) = 0;
     virtual void SaveContext(ThreadContext64& ctx) = 0;
     virtual void LoadContext(const ThreadContext32& ctx) = 0;
diff --git a/src/core/arm/dynarmic/arm_dynarmic_32.cpp b/src/core/arm/dynarmic/arm_dynarmic_32.cpp
index 93d43e22e..e5b78210a 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_32.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_32.cpp
@@ -4,9 +4,9 @@
 
 #include <cinttypes>
 #include <memory>
-#include <dynarmic/A32/a32.h>
-#include <dynarmic/A32/config.h>
-#include <dynarmic/A32/context.h>
+#include <dynarmic/interface/A32/a32.h>
+#include <dynarmic/interface/A32/config.h>
+#include <dynarmic/interface/A32/context.h>
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "common/page_table.h"
@@ -78,7 +78,9 @@ public:
     }
 
     void CallSVC(u32 swi) override {
-        Kernel::Svc::Call(parent.system, swi);
+        parent.svc_called = true;
+        parent.svc_swi = swi;
+        parent.jit->HaltExecution();
     }
 
     void AddTicks(u64 ticks) override {
@@ -187,11 +189,17 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable*
 }
 
 void ARM_Dynarmic_32::Run() {
-    jit->Run();
-}
-
-void ARM_Dynarmic_32::ExceptionalExit() {
-    jit->ExceptionalExit();
+    while (true) {
+        jit->Run();
+        if (!svc_called) {
+            break;
+        }
+        svc_called = false;
+        Kernel::Svc::Call(system, svc_swi);
+        if (shutdown) {
+            break;
+        }
+    }
 }
 
 void ARM_Dynarmic_32::Step() {
@@ -255,10 +263,6 @@ void ARM_Dynarmic_32::SetTPIDR_EL0(u64 value) {
     cp15->uprw = static_cast<u32>(value);
 }
 
-void ARM_Dynarmic_32::ChangeProcessorID(std::size_t new_core_id) {
-    jit->ChangeProcessorID(new_core_id);
-}
-
 void ARM_Dynarmic_32::SaveContext(ThreadContext32& ctx) {
     Dynarmic::A32::Context context;
     jit->SaveContext(context);
@@ -279,6 +283,7 @@ void ARM_Dynarmic_32::LoadContext(const ThreadContext32& ctx) {
 
 void ARM_Dynarmic_32::PrepareReschedule() {
     jit->HaltExecution();
+    shutdown = true;
 }
 
 void ARM_Dynarmic_32::ClearInstructionCache() {
diff --git a/src/core/arm/dynarmic/arm_dynarmic_32.h b/src/core/arm/dynarmic/arm_dynarmic_32.h
index 42778c02c..063605b46 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_32.h
+++ b/src/core/arm/dynarmic/arm_dynarmic_32.h
@@ -7,9 +7,9 @@
 #include <memory>
 #include <unordered_map>
 
-#include <dynarmic/A32/a32.h>
-#include <dynarmic/A64/a64.h>
-#include <dynarmic/exclusive_monitor.h>
+#include <dynarmic/interface/A32/a32.h>
+#include <dynarmic/interface/A64/a64.h>
+#include <dynarmic/interface/exclusive_monitor.h>
 #include "common/common_types.h"
 #include "common/hash.h"
 #include "core/arm/arm_interface.h"
@@ -42,13 +42,11 @@ public:
     u32 GetPSTATE() const override;
     void SetPSTATE(u32 pstate) override;
     void Run() override;
-    void ExceptionalExit() override;
     void Step() override;
     VAddr GetTlsAddress() const override;
     void SetTlsAddress(VAddr address) override;
     void SetTPIDR_EL0(u64 value) override;
     u64 GetTPIDR_EL0() const override;
-    void ChangeProcessorID(std::size_t new_core_id) override;
 
     bool IsInThumbMode() const {
         return (GetPSTATE() & 0x20) != 0;
@@ -83,6 +81,12 @@ private:
     std::size_t core_index;
     DynarmicExclusiveMonitor& exclusive_monitor;
     std::shared_ptr<Dynarmic::A32::Jit> jit;
+
+    // SVC callback
+    u32 svc_swi{};
+    bool svc_called{};
+
+    bool shutdown{};
 };
 
 } // namespace Core
diff --git a/src/core/arm/dynarmic/arm_dynarmic_64.cpp b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
index 08fa85904..dd439f55e 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_64.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
@@ -4,8 +4,8 @@
 
 #include <cinttypes>
 #include <memory>
-#include <dynarmic/A64/a64.h>
-#include <dynarmic/A64/config.h>
+#include <dynarmic/interface/A64/a64.h>
+#include <dynarmic/interface/A64/config.h>
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "common/page_table.h"
@@ -102,7 +102,9 @@ public:
     }
 
     void CallSVC(u32 swi) override {
-        Kernel::Svc::Call(parent.system, swi);
+        parent.svc_called = true;
+        parent.svc_swi = swi;
+        parent.jit->HaltExecution();
     }
 
     void AddTicks(u64 ticks) override {
@@ -227,11 +229,17 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
 }
 
 void ARM_Dynarmic_64::Run() {
-    jit->Run();
-}
-
-void ARM_Dynarmic_64::ExceptionalExit() {
-    jit->ExceptionalExit();
+    while (true) {
+        jit->Run();
+        if (!svc_called) {
+            break;
+        }
+        svc_called = false;
+        Kernel::Svc::Call(system, svc_swi);
+        if (shutdown) {
+            break;
+        }
+    }
 }
 
 void ARM_Dynarmic_64::Step() {
@@ -296,10 +304,6 @@ void ARM_Dynarmic_64::SetTPIDR_EL0(u64 value) {
     cb->tpidr_el0 = value;
 }
 
-void ARM_Dynarmic_64::ChangeProcessorID(std::size_t new_core_id) {
-    jit->ChangeProcessorID(new_core_id);
-}
-
 void ARM_Dynarmic_64::SaveContext(ThreadContext64& ctx) {
     ctx.cpu_registers = jit->GetRegisters();
     ctx.sp = jit->GetSP();
@@ -324,6 +328,7 @@ void ARM_Dynarmic_64::LoadContext(const ThreadContext64& ctx) {
 
 void ARM_Dynarmic_64::PrepareReschedule() {
     jit->HaltExecution();
+    shutdown = true;
 }
 
 void ARM_Dynarmic_64::ClearInstructionCache() {
diff --git a/src/core/arm/dynarmic/arm_dynarmic_64.h b/src/core/arm/dynarmic/arm_dynarmic_64.h
index b81fbcc66..0c4e46c64 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_64.h
+++ b/src/core/arm/dynarmic/arm_dynarmic_64.h
@@ -7,7 +7,7 @@
 #include <memory>
 #include <unordered_map>
 
-#include <dynarmic/A64/a64.h>
+#include <dynarmic/interface/A64/a64.h>
 #include "common/common_types.h"
 #include "common/hash.h"
 #include "core/arm/arm_interface.h"
@@ -40,12 +40,10 @@ public:
     void SetPSTATE(u32 pstate) override;
     void Run() override;
     void Step() override;
-    void ExceptionalExit() override;
     VAddr GetTlsAddress() const override;
     void SetTlsAddress(VAddr address) override;
     void SetTPIDR_EL0(u64 value) override;
     u64 GetTPIDR_EL0() const override;
-    void ChangeProcessorID(std::size_t new_core_id) override;
 
     void SaveContext(ThreadContext32& ctx) override {}
     void SaveContext(ThreadContext64& ctx) override;
@@ -76,6 +74,12 @@ private:
     DynarmicExclusiveMonitor& exclusive_monitor;
 
     std::shared_ptr<Dynarmic::A64::Jit> jit;
+
+    // SVC callback
+    u32 svc_swi{};
+    bool svc_called{};
+
+    bool shutdown{};
 };
 
 } // namespace Core
diff --git a/src/core/arm/dynarmic/arm_dynarmic_cp15.h b/src/core/arm/dynarmic/arm_dynarmic_cp15.h
index 8597beddf..7c7ede79e 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_cp15.h
+++ b/src/core/arm/dynarmic/arm_dynarmic_cp15.h
@@ -7,7 +7,7 @@
 #include <memory>
 #include <optional>
 
-#include <dynarmic/A32/coprocessor.h>
+#include <dynarmic/interface/A32/coprocessor.h>
 #include "common/common_types.h"
 
 namespace Core {
diff --git a/src/core/arm/dynarmic/arm_exclusive_monitor.h b/src/core/arm/dynarmic/arm_exclusive_monitor.h
index f9f056a59..73d41f223 100644
--- a/src/core/arm/dynarmic/arm_exclusive_monitor.h
+++ b/src/core/arm/dynarmic/arm_exclusive_monitor.h
@@ -7,7 +7,7 @@
 #include <memory>
 #include <unordered_map>
 
-#include <dynarmic/exclusive_monitor.h>
+#include <dynarmic/interface/exclusive_monitor.h>
 
 #include "common/common_types.h"
 #include "core/arm/dynarmic/arm_dynarmic_32.h"
diff --git a/src/core/hle/kernel/k_client_port.h b/src/core/hle/kernel/k_client_port.h
index 8501156e8..f2fff3b01 100644
--- a/src/core/hle/kernel/k_client_port.h
+++ b/src/core/hle/kernel/k_client_port.h
@@ -22,7 +22,7 @@ class KClientPort final : public KSynchronizationObject {
 
 public:
     explicit KClientPort(KernelCore& kernel_);
-    virtual ~KClientPort() override;
+    ~KClientPort() override;
 
     void Initialize(KPort* parent_, s32 max_sessions_, std::string&& name_);
     void OnSessionFinalized();
@@ -49,8 +49,8 @@ public:
     bool IsServerClosed() const;
 
     // Overridden virtual functions.
-    virtual void Destroy() override;
-    virtual bool IsSignaled() const override;
+    void Destroy() override;
+    bool IsSignaled() const override;
 
     ResultCode CreateSession(KClientSession** out);
 
diff --git a/src/core/hle/kernel/k_client_session.h b/src/core/hle/kernel/k_client_session.h
index 720a8c243..b11d5b4e3 100644
--- a/src/core/hle/kernel/k_client_session.h
+++ b/src/core/hle/kernel/k_client_session.h
@@ -34,7 +34,7 @@ class KClientSession final
 
 public:
     explicit KClientSession(KernelCore& kernel_);
-    virtual ~KClientSession();
+    ~KClientSession() override;
 
     void Initialize(KSession* parent_, std::string&& name_) {
         // Set member variables.
@@ -42,7 +42,7 @@ public:
         name = std::move(name_);
     }
 
-    virtual void Destroy() override;
+    void Destroy() override;
     static void PostDestroy([[maybe_unused]] uintptr_t arg) {}
 
     KSession* GetParent() const {
diff --git a/src/core/hle/kernel/k_event.h b/src/core/hle/kernel/k_event.h
index 9a59ffb70..3d3ec99e2 100644
--- a/src/core/hle/kernel/k_event.h
+++ b/src/core/hle/kernel/k_event.h
@@ -20,23 +20,21 @@ class KEvent final : public KAutoObjectWithSlabHeapAndContainer<KEvent, KAutoObj
 
 public:
     explicit KEvent(KernelCore& kernel_);
-    virtual ~KEvent();
+    ~KEvent() override;
 
     void Initialize(std::string&& name);
 
-    virtual void Finalize() override;
+    void Finalize() override;
 
-    virtual bool IsInitialized() const override {
+    bool IsInitialized() const override {
         return initialized;
     }
 
-    virtual uintptr_t GetPostDestroyArgument() const override {
+    uintptr_t GetPostDestroyArgument() const override {
         return reinterpret_cast<uintptr_t>(owner);
     }
 
-    static void PostDestroy(uintptr_t arg);
-
-    virtual KProcess* GetOwner() const override {
+    KProcess* GetOwner() const override {
         return owner;
     }
 
@@ -48,6 +46,8 @@ public:
         return writable_event;
     }
 
+    static void PostDestroy(uintptr_t arg);
+
 private:
     KReadableEvent readable_event;
     KWritableEvent writable_event;
diff --git a/src/core/hle/kernel/k_port.h b/src/core/hle/kernel/k_port.h
index 960f1f3a3..4018ea2df 100644
--- a/src/core/hle/kernel/k_port.h
+++ b/src/core/hle/kernel/k_port.h
@@ -22,7 +22,7 @@ class KPort final : public KAutoObjectWithSlabHeapAndContainer<KPort, KAutoObjec
 
 public:
     explicit KPort(KernelCore& kernel_);
-    virtual ~KPort();
+    ~KPort() override;
 
     static void PostDestroy([[maybe_unused]] uintptr_t arg) {}
 
@@ -59,7 +59,6 @@ private:
         ServerClosed = 3,
     };
 
-private:
     KServerPort server;
     KClientPort client;
     State state{State::Invalid};
diff --git a/src/core/hle/kernel/k_process.h b/src/core/hle/kernel/k_process.h
index 123d71cd3..c0656b9af 100644
--- a/src/core/hle/kernel/k_process.h
+++ b/src/core/hle/kernel/k_process.h
@@ -331,19 +331,19 @@ public:
 
     void LoadModule(CodeSet code_set, VAddr base_addr);
 
-    virtual bool IsInitialized() const override {
+    bool IsInitialized() const override {
         return is_initialized;
     }
 
     static void PostDestroy([[maybe_unused]] uintptr_t arg) {}
 
-    virtual void Finalize();
+    void Finalize() override;
 
-    virtual u64 GetId() const override final {
+    u64 GetId() const override {
         return GetProcessID();
     }
 
-    virtual bool IsSignaled() const override;
+    bool IsSignaled() const override;
 
     void PinCurrentThread();
     void UnpinCurrentThread();
diff --git a/src/core/hle/kernel/k_readable_event.h b/src/core/hle/kernel/k_readable_event.h
index 33cd1dd3e..b2850ac7b 100644
--- a/src/core/hle/kernel/k_readable_event.h
+++ b/src/core/hle/kernel/k_readable_event.h
@@ -31,8 +31,8 @@ public:
         return parent;
     }
 
-    virtual bool IsSignaled() const override;
-    virtual void Destroy() override;
+    bool IsSignaled() const override;
+    void Destroy() override;
 
     ResultCode Signal();
     ResultCode Clear();
diff --git a/src/core/hle/kernel/k_resource_limit.h b/src/core/hle/kernel/k_resource_limit.h
index 0debbbb51..fab6005ff 100644
--- a/src/core/hle/kernel/k_resource_limit.h
+++ b/src/core/hle/kernel/k_resource_limit.h
@@ -37,10 +37,10 @@ class KResourceLimit final
 
 public:
     explicit KResourceLimit(KernelCore& kernel_);
-    virtual ~KResourceLimit();
+    ~KResourceLimit() override;
 
     void Initialize(const Core::Timing::CoreTiming* core_timing_);
-    virtual void Finalize() override;
+    void Finalize() override;
 
     s64 GetLimitValue(LimitableResource which) const;
     s64 GetCurrentValue(LimitableResource which) const;
diff --git a/src/core/hle/kernel/k_scheduler.cpp b/src/core/hle/kernel/k_scheduler.cpp
index 2f82fbcd6..6a7d80d03 100644
--- a/src/core/hle/kernel/k_scheduler.cpp
+++ b/src/core/hle/kernel/k_scheduler.cpp
@@ -659,7 +659,6 @@ void KScheduler::Unload(KThread* thread) {
 
     if (thread) {
         if (thread->IsCallingSvc()) {
-            system.ArmInterface(core_id).ExceptionalExit();
             thread->ClearIsCallingSvc();
         }
         if (!thread->IsTerminationRequested()) {
diff --git a/src/core/hle/kernel/k_server_port.h b/src/core/hle/kernel/k_server_port.h
index d1a757ec3..55481d63f 100644
--- a/src/core/hle/kernel/k_server_port.h
+++ b/src/core/hle/kernel/k_server_port.h
@@ -25,12 +25,9 @@ class SessionRequestHandler;
 class KServerPort final : public KSynchronizationObject {
     KERNEL_AUTOOBJECT_TRAITS(KServerPort, KSynchronizationObject);
 
-private:
-    using SessionList = boost::intrusive::list<KServerSession>;
-
 public:
     explicit KServerPort(KernelCore& kernel_);
-    virtual ~KServerPort() override;
+    ~KServerPort() override;
 
     void Initialize(KPort* parent_, std::string&& name_);
 
@@ -63,13 +60,14 @@ public:
     bool IsLight() const;
 
     // Overridden virtual functions.
-    virtual void Destroy() override;
-    virtual bool IsSignaled() const override;
+    void Destroy() override;
+    bool IsSignaled() const override;
 
 private:
+    using SessionList = boost::intrusive::list<KServerSession>;
+
     void CleanupSessions();
 
-private:
     SessionList session_list;
     SessionRequestHandlerPtr session_handler;
     KPort* parent{};
diff --git a/src/core/hle/kernel/k_server_session.h b/src/core/hle/kernel/k_server_session.h
index dd4de2904..27b757ad2 100644
--- a/src/core/hle/kernel/k_server_session.h
+++ b/src/core/hle/kernel/k_server_session.h
@@ -42,9 +42,9 @@ class KServerSession final : public KSynchronizationObject,
 
 public:
     explicit KServerSession(KernelCore& kernel_);
-    virtual ~KServerSession() override;
+    ~KServerSession() override;
 
-    virtual void Destroy() override;
+    void Destroy() override;
 
     void Initialize(KSession* parent_, std::string&& name_);
 
@@ -56,7 +56,7 @@ public:
         return parent;
     }
 
-    virtual bool IsSignaled() const override;
+    bool IsSignaled() const override;
 
     void OnClientClosed();
 
diff --git a/src/core/hle/kernel/k_session.h b/src/core/hle/kernel/k_session.h
index a981fd1f6..4ddd080d2 100644
--- a/src/core/hle/kernel/k_session.h
+++ b/src/core/hle/kernel/k_session.h
@@ -18,17 +18,17 @@ class KSession final : public KAutoObjectWithSlabHeapAndContainer<KSession, KAut
 
 public:
     explicit KSession(KernelCore& kernel_);
-    virtual ~KSession() override;
+    ~KSession() override;
 
     void Initialize(KClientPort* port_, const std::string& name_);
 
-    virtual void Finalize() override;
+    void Finalize() override;
 
-    virtual bool IsInitialized() const override {
+    bool IsInitialized() const override {
         return initialized;
     }
 
-    virtual uintptr_t GetPostDestroyArgument() const override {
+    uintptr_t GetPostDestroyArgument() const override {
         return reinterpret_cast<uintptr_t>(process);
     }
 
@@ -78,7 +78,6 @@ private:
         ServerClosed = 3,
     };
 
-private:
     void SetState(State state) {
         atomic_state = static_cast<u8>(state);
     }
@@ -87,7 +86,6 @@ private:
         return static_cast<State>(atomic_state.load(std::memory_order_relaxed));
     }
 
-private:
     KServerSession server;
     KClientSession client;
     std::atomic<std::underlying_type_t<State>> atomic_state{
diff --git a/src/core/hle/kernel/k_shared_memory.h b/src/core/hle/kernel/k_shared_memory.h
index 553a56327..e9815f90b 100644
--- a/src/core/hle/kernel/k_shared_memory.h
+++ b/src/core/hle/kernel/k_shared_memory.h
@@ -68,9 +68,9 @@ public:
         return device_memory->GetPointer(physical_address + offset);
     }
 
-    virtual void Finalize() override;
+    void Finalize() override;
 
-    virtual bool IsInitialized() const override {
+    bool IsInitialized() const override {
         return is_initialized;
     }
     static void PostDestroy([[maybe_unused]] uintptr_t arg) {}
diff --git a/src/core/hle/kernel/k_slab_heap.h b/src/core/hle/kernel/k_slab_heap.h
index 81d472a3e..0ad74b0a0 100644
--- a/src/core/hle/kernel/k_slab_heap.h
+++ b/src/core/hle/kernel/k_slab_heap.h
@@ -4,34 +4,213 @@
 
 #pragma once
 
+#include <atomic>
+
+#include "common/assert.h"
+#include "common/common_types.h"
+
 namespace Kernel {
 
 class KernelCore;
 
-/// This is a placeholder class to manage slab heaps for kernel objects. For now, we just allocate
-/// these with new/delete, but this can be re-implemented later to allocate these in emulated
-/// memory.
+namespace impl {
+
+class KSlabHeapImpl final : NonCopyable {
+public:
+    struct Node {
+        Node* next{};
+    };
+
+    constexpr KSlabHeapImpl() = default;
+
+    void Initialize(std::size_t size) {
+        ASSERT(head == nullptr);
+        obj_size = size;
+    }
+
+    constexpr std::size_t GetObjectSize() const {
+        return obj_size;
+    }
+
+    Node* GetHead() const {
+        return head;
+    }
+
+    void* Allocate() {
+        Node* ret = head.load();
+
+        do {
+            if (ret == nullptr) {
+                break;
+            }
+        } while (!head.compare_exchange_weak(ret, ret->next));
+
+        return ret;
+    }
+
+    void Free(void* obj) {
+        Node* node = static_cast<Node*>(obj);
+
+        Node* cur_head = head.load();
+        do {
+            node->next = cur_head;
+        } while (!head.compare_exchange_weak(cur_head, node));
+    }
+
+private:
+    std::atomic<Node*> head{};
+    std::size_t obj_size{};
+};
+
+} // namespace impl
+
+class KSlabHeapBase : NonCopyable {
+public:
+    constexpr KSlabHeapBase() = default;
+
+    constexpr bool Contains(uintptr_t addr) const {
+        return start <= addr && addr < end;
+    }
+
+    constexpr std::size_t GetSlabHeapSize() const {
+        return (end - start) / GetObjectSize();
+    }
+
+    constexpr std::size_t GetObjectSize() const {
+        return impl.GetObjectSize();
+    }
+
+    constexpr uintptr_t GetSlabHeapAddress() const {
+        return start;
+    }
+
+    std::size_t GetObjectIndexImpl(const void* obj) const {
+        return (reinterpret_cast<uintptr_t>(obj) - start) / GetObjectSize();
+    }
+
+    std::size_t GetPeakIndex() const {
+        return GetObjectIndexImpl(reinterpret_cast<const void*>(peak));
+    }
+
+    void* AllocateImpl() {
+        return impl.Allocate();
+    }
+
+    void FreeImpl(void* obj) {
+        // Don't allow freeing an object that wasn't allocated from this heap
+        ASSERT(Contains(reinterpret_cast<uintptr_t>(obj)));
+
+        impl.Free(obj);
+    }
+
+    void InitializeImpl(std::size_t obj_size, void* memory, std::size_t memory_size) {
+        // Ensure we don't initialize a slab using null memory
+        ASSERT(memory != nullptr);
+
+        // Initialize the base allocator
+        impl.Initialize(obj_size);
+
+        // Set our tracking variables
+        const std::size_t num_obj = (memory_size / obj_size);
+        start = reinterpret_cast<uintptr_t>(memory);
+        end = start + num_obj * obj_size;
+        peak = start;
+
+        // Free the objects
+        u8* cur = reinterpret_cast<u8*>(end);
+
+        for (std::size_t i{}; i < num_obj; i++) {
+            cur -= obj_size;
+            impl.Free(cur);
+        }
+    }
+
+private:
+    using Impl = impl::KSlabHeapImpl;
+
+    Impl impl;
+    uintptr_t peak{};
+    uintptr_t start{};
+    uintptr_t end{};
+};
 
 template <typename T>
-class KSlabHeap final : NonCopyable {
+class KSlabHeap final : public KSlabHeapBase {
 public:
-    KSlabHeap() = default;
+    enum class AllocationType {
+        Host,
+        Guest,
+    };
 
-    void Initialize([[maybe_unused]] void* memory, [[maybe_unused]] std::size_t memory_size) {
-        // Placeholder that should initialize the backing slab heap implementation.
+    explicit constexpr KSlabHeap(AllocationType allocation_type_ = AllocationType::Host)
+        : KSlabHeapBase(), allocation_type{allocation_type_} {}
+
+    void Initialize(void* memory, std::size_t memory_size) {
+        if (allocation_type == AllocationType::Guest) {
+            InitializeImpl(sizeof(T), memory, memory_size);
+        }
     }
 
     T* Allocate() {
-        return new T();
+        switch (allocation_type) {
+        case AllocationType::Host:
+            // Fallback for cases where we do not yet support allocating guest memory from the slab
+            // heap, such as for kernel memory regions.
+            return new T;
+
+        case AllocationType::Guest:
+            T* obj = static_cast<T*>(AllocateImpl());
+            if (obj != nullptr) {
+                new (obj) T();
+            }
+            return obj;
+        }
+
+        UNREACHABLE_MSG("Invalid AllocationType {}", allocation_type);
+        return nullptr;
     }
 
     T* AllocateWithKernel(KernelCore& kernel) {
-        return new T(kernel);
+        switch (allocation_type) {
+        case AllocationType::Host:
+            // Fallback for cases where we do not yet support allocating guest memory from the slab
+            // heap, such as for kernel memory regions.
+            return new T(kernel);
+
+        case AllocationType::Guest:
+            T* obj = static_cast<T*>(AllocateImpl());
+            if (obj != nullptr) {
+                new (obj) T(kernel);
+            }
+            return obj;
+        }
+
+        UNREACHABLE_MSG("Invalid AllocationType {}", allocation_type);
+        return nullptr;
     }
 
     void Free(T* obj) {
-        delete obj;
+        switch (allocation_type) {
+        case AllocationType::Host:
+            // Fallback for cases where we do not yet support allocating guest memory from the slab
+            // heap, such as for kernel memory regions.
+            delete obj;
+            return;
+
+        case AllocationType::Guest:
+            FreeImpl(obj);
+            return;
+        }
+
+        UNREACHABLE_MSG("Invalid AllocationType {}", allocation_type);
     }
+
+    constexpr std::size_t GetObjectIndex(const T* obj) const {
+        return GetObjectIndexImpl(obj);
+    }
+
+private:
+    const AllocationType allocation_type;
 };
 
 } // namespace Kernel
diff --git a/src/core/hle/kernel/k_synchronization_object.h b/src/core/hle/kernel/k_synchronization_object.h
index a41dd1220..3d4ce1fbc 100644
--- a/src/core/hle/kernel/k_synchronization_object.h
+++ b/src/core/hle/kernel/k_synchronization_object.h
@@ -29,7 +29,7 @@ public:
                                          KSynchronizationObject** objects, const s32 num_objects,
                                          s64 timeout);
 
-    virtual void Finalize() override;
+    void Finalize() override;
 
     [[nodiscard]] virtual bool IsSignaled() const = 0;
 
@@ -37,7 +37,7 @@ public:
 
 protected:
     explicit KSynchronizationObject(KernelCore& kernel);
-    virtual ~KSynchronizationObject();
+    ~KSynchronizationObject() override;
 
     virtual void OnFinalizeSynchronizationObject() {}
 
diff --git a/src/core/hle/kernel/k_thread.cpp b/src/core/hle/kernel/k_thread.cpp
index e3f08f256..3cf43d290 100644
--- a/src/core/hle/kernel/k_thread.cpp
+++ b/src/core/hle/kernel/k_thread.cpp
@@ -168,13 +168,13 @@ ResultCode KThread::Initialize(KThreadFunction func, uintptr_t arg, VAddr user_s
     std::memset(static_cast<void*>(std::addressof(GetStackParameters())), 0,
                 sizeof(StackParameters));
 
-    // Setup the TLS, if needed.
-    if (type == ThreadType::User) {
-        tls_address = owner->CreateTLSRegion();
-    }
-
     // Set parent, if relevant.
     if (owner != nullptr) {
+        // Setup the TLS, if needed.
+        if (type == ThreadType::User) {
+            tls_address = owner->CreateTLSRegion();
+        }
+
         parent = owner;
         parent->Open();
         parent->IncrementThreadCount();
diff --git a/src/core/hle/kernel/k_thread.h b/src/core/hle/kernel/k_thread.h
index 4abfc2b49..01eebb165 100644
--- a/src/core/hle/kernel/k_thread.h
+++ b/src/core/hle/kernel/k_thread.h
@@ -358,21 +358,21 @@ public:
         return termination_requested || GetRawState() == ThreadState::Terminated;
     }
 
-    [[nodiscard]] virtual u64 GetId() const override final {
+    [[nodiscard]] u64 GetId() const override {
         return this->GetThreadID();
     }
 
-    [[nodiscard]] virtual bool IsInitialized() const override {
+    [[nodiscard]] bool IsInitialized() const override {
         return initialized;
     }
 
-    [[nodiscard]] virtual uintptr_t GetPostDestroyArgument() const override {
+    [[nodiscard]] uintptr_t GetPostDestroyArgument() const override {
         return reinterpret_cast<uintptr_t>(parent) | (resource_limit_release_hint ? 1 : 0);
     }
 
-    virtual void Finalize() override;
+    void Finalize() override;
 
-    [[nodiscard]] virtual bool IsSignaled() const override;
+    [[nodiscard]] bool IsSignaled() const override;
 
     static void PostDestroy(uintptr_t arg);
 
diff --git a/src/core/hle/kernel/k_transfer_memory.h b/src/core/hle/kernel/k_transfer_memory.h
index c2d0f1eaf..31029a5c2 100644
--- a/src/core/hle/kernel/k_transfer_memory.h
+++ b/src/core/hle/kernel/k_transfer_memory.h
@@ -27,23 +27,23 @@ class KTransferMemory final
 
 public:
     explicit KTransferMemory(KernelCore& kernel_);
-    virtual ~KTransferMemory() override;
+    ~KTransferMemory() override;
 
     ResultCode Initialize(VAddr address_, std::size_t size_, Svc::MemoryPermission owner_perm_);
 
-    virtual void Finalize() override;
+    void Finalize() override;
 
-    virtual bool IsInitialized() const override {
+    bool IsInitialized() const override {
         return is_initialized;
     }
 
-    virtual uintptr_t GetPostDestroyArgument() const override {
+    uintptr_t GetPostDestroyArgument() const override {
         return reinterpret_cast<uintptr_t>(owner);
     }
 
     static void PostDestroy(uintptr_t arg);
 
-    KProcess* GetOwner() const {
+    KProcess* GetOwner() const override {
         return owner;
     }
 
diff --git a/src/core/hle/kernel/k_writable_event.h b/src/core/hle/kernel/k_writable_event.h
index 607b0eadb..858d982c4 100644
--- a/src/core/hle/kernel/k_writable_event.h
+++ b/src/core/hle/kernel/k_writable_event.h
@@ -21,7 +21,7 @@ public:
     explicit KWritableEvent(KernelCore& kernel_);
     ~KWritableEvent() override;
 
-    virtual void Destroy() override;
+    void Destroy() override;
 
     static void PostDestroy([[maybe_unused]] uintptr_t arg) {}
 
diff --git a/src/core/hle/kernel/kernel.cpp b/src/core/hle/kernel/kernel.cpp
index 8b55df82e..0ffb78d51 100644
--- a/src/core/hle/kernel/kernel.cpp
+++ b/src/core/hle/kernel/kernel.cpp
@@ -258,7 +258,7 @@ struct KernelCore::Impl {
             KAutoObject::Create(thread.get());
             ASSERT(KThread::InitializeDummyThread(thread.get()).IsSuccess());
             thread->SetName(fmt::format("DummyThread:{}", GetHostThreadId()));
-            return std::move(thread);
+            return thread;
         };
 
         thread_local auto thread = make_thread();
@@ -620,7 +620,8 @@ struct KernelCore::Impl {
 
     void InitializePageSlab() {
         // Allocate slab heaps
-        user_slab_heap_pages = std::make_unique<KSlabHeap<Page>>();
+        user_slab_heap_pages =
+            std::make_unique<KSlabHeap<Page>>(KSlabHeap<Page>::AllocationType::Guest);
 
         // TODO(ameerj): This should be derived, not hardcoded within the kernel
         constexpr u64 user_slab_heap_size{0x3de000};
diff --git a/src/core/hle/service/hid/hid.cpp b/src/core/hle/service/hid/hid.cpp
index 49c17fd14..df0fe1c8e 100644
--- a/src/core/hle/service/hid/hid.cpp
+++ b/src/core/hle/service/hid/hid.cpp
@@ -1770,7 +1770,7 @@ public:
             {232, nullptr, "GetIrSensorState"},
             {233, nullptr, "GetXcdHandleForNpadWithIrSensor"},
             {301, nullptr, "ActivateNpadSystem"},
-            {303, nullptr, "ApplyNpadSystemCommonPolicy"},
+            {303, &HidSys::ApplyNpadSystemCommonPolicy, "ApplyNpadSystemCommonPolicy"},
             {304, nullptr, "EnableAssigningSingleOnSlSrPress"},
             {305, nullptr, "DisableAssigningSingleOnSlSrPress"},
             {306, nullptr, "GetLastActiveNpad"},
@@ -1949,6 +1949,15 @@ public:
 
         RegisterHandlers(functions);
     }
+
+private:
+    void ApplyNpadSystemCommonPolicy(Kernel::HLERequestContext& ctx) {
+        // We already do this for homebrew so we can just stub it out
+        LOG_WARNING(Service_HID, "called");
+
+        IPC::ResponseBuilder rb{ctx, 2};
+        rb.Push(RESULT_SUCCESS);
+    }
 };
 
 class HidTmp final : public ServiceFramework<HidTmp> {
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index de971041f..9e6b87960 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -596,7 +596,7 @@ void BufferCache<P>::PopAsyncFlushes() {
             runtime.CopyBuffer(download_staging.buffer, slot_buffers[buffer_id], copies);
         }
         runtime.Finish();
-        for (const auto [copy, buffer_id] : downloads) {
+        for (const auto& [copy, buffer_id] : downloads) {
             const Buffer& buffer = slot_buffers[buffer_id];
             const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
             // Undo the modified offset
@@ -606,7 +606,7 @@ void BufferCache<P>::PopAsyncFlushes() {
         }
     } else {
         const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
-        for (const auto [copy, buffer_id] : downloads) {
+        for (const auto& [copy, buffer_id] : downloads) {
             Buffer& buffer = slot_buffers[buffer_id];
             buffer.ImmediateDownload(copy.src_offset, immediate_buffer.subspan(0, copy.size));
             const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
diff --git a/src/video_core/rasterizer_accelerated.cpp b/src/video_core/rasterizer_accelerated.cpp
index 62d84c0f8..6decd2546 100644
--- a/src/video_core/rasterizer_accelerated.cpp
+++ b/src/video_core/rasterizer_accelerated.cpp
@@ -18,10 +18,10 @@ RasterizerAccelerated::~RasterizerAccelerated() = default;
 void RasterizerAccelerated::UpdatePagesCachedCount(VAddr addr, u64 size, int delta) {
     const auto page_end = Common::DivCeil(addr + size, Core::Memory::PAGE_SIZE);
     for (auto page = addr >> Core::Memory::PAGE_BITS; page != page_end; ++page) {
-        auto& count = cached_pages.at(page >> 3).Count(page);
+        auto& count = cached_pages.at(page >> 2).Count(page);
 
         if (delta > 0) {
-            ASSERT_MSG(count < UINT8_MAX, "Count may overflow!");
+            ASSERT_MSG(count < UINT16_MAX, "Count may overflow!");
         } else if (delta < 0) {
             ASSERT_MSG(count > 0, "Count may underflow!");
         } else {
@@ -29,7 +29,7 @@ void RasterizerAccelerated::UpdatePagesCachedCount(VAddr addr, u64 size, int del
         }
 
         // Adds or subtracts 1, as count is a unsigned 8-bit value
-        count += static_cast<u8>(delta);
+        count += static_cast<u16>(delta);
 
         // Assume delta is either -1 or 1
         if (count == 0) {
diff --git a/src/video_core/rasterizer_accelerated.h b/src/video_core/rasterizer_accelerated.h
index 9227a4adc..ea879bfdd 100644
--- a/src/video_core/rasterizer_accelerated.h
+++ b/src/video_core/rasterizer_accelerated.h
@@ -29,20 +29,20 @@ private:
     public:
         CacheEntry() = default;
 
-        std::atomic_uint8_t& Count(std::size_t page) {
-            return values[page & 7];
+        std::atomic_uint16_t& Count(std::size_t page) {
+            return values[page & 3];
         }
 
-        const std::atomic_uint8_t& Count(std::size_t page) const {
-            return values[page & 7];
+        const std::atomic_uint16_t& Count(std::size_t page) const {
+            return values[page & 3];
         }
 
     private:
-        std::array<std::atomic_uint8_t, 8> values{};
+        std::array<std::atomic_uint16_t, 4> values{};
     };
     static_assert(sizeof(CacheEntry) == 8, "CacheEntry should be 8 bytes!");
 
-    std::array<CacheEntry, 0x800000> cached_pages;
+    std::array<CacheEntry, 0x1000000> cached_pages;
     Core::Memory::Memory& cpu_memory;
 };
 
diff --git a/src/yuzu/CMakeLists.txt b/src/yuzu/CMakeLists.txt
index cc0790e07..634fe66a5 100644
--- a/src/yuzu/CMakeLists.txt
+++ b/src/yuzu/CMakeLists.txt
@@ -4,6 +4,12 @@ set(CMAKE_AUTOUIC ON)
 set(CMAKE_INCLUDE_CURRENT_DIR ON)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${PROJECT_SOURCE_DIR}/CMakeModules)
 
+# Set the RPATH for Qt Libraries
+# This must be done before the `yuzu` target is created
+if (YUZU_USE_BUNDLED_QT AND (${CMAKE_SYSTEM_NAME} STREQUAL "Linux"))
+    set(CMAKE_BUILD_RPATH "${CMAKE_BINARY_DIR}/bin/lib/")
+endif()
+
 add_executable(yuzu
     Info.plist
     about_dialog.cpp
@@ -278,11 +284,14 @@ if(UNIX AND NOT APPLE)
     install(TARGETS yuzu RUNTIME DESTINATION "${CMAKE_INSTALL_PREFIX}/bin")
 endif()
 
-if (MSVC)
+if (YUZU_USE_BUNDLED_QT)
     include(CopyYuzuQt5Deps)
+    copy_yuzu_Qt5_deps(yuzu)
+endif()
+
+if (MSVC)
     include(CopyYuzuSDLDeps)
     include(CopyYuzuFFmpegDeps)
-    copy_yuzu_Qt5_deps(yuzu)
     copy_yuzu_SDL_deps(yuzu)
     copy_yuzu_FFmpeg_deps(yuzu)
 endif()
diff --git a/src/yuzu/configuration/configure_dialog.cpp b/src/yuzu/configuration/configure_dialog.cpp
index 3ad40d2b3..6028135c5 100644
--- a/src/yuzu/configuration/configure_dialog.cpp
+++ b/src/yuzu/configuration/configure_dialog.cpp
@@ -2,8 +2,11 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <QAbstractButton>
+#include <QDialogButtonBox>
 #include <QHash>
 #include <QListWidgetItem>
+#include <QPushButton>
 #include <QSignalBlocker>
 #include "common/settings.h"
 #include "core/core.h"
@@ -31,6 +34,12 @@ ConfigureDialog::ConfigureDialog(QWidget* parent, HotkeyRegistry& registry,
     connect(ui->selectorList, &QListWidget::itemSelectionChanged, this,
             &ConfigureDialog::UpdateVisibleTabs);
 
+    if (Core::System::GetInstance().IsPoweredOn()) {
+        QPushButton* apply_button = ui->buttonBox->addButton(QDialogButtonBox::Apply);
+        connect(apply_button, &QAbstractButton::clicked, this,
+                &ConfigureDialog::HandleApplyButtonClicked);
+    }
+
     adjustSize();
     ui->selectorList->setCurrentRow(0);
 }
@@ -80,6 +89,11 @@ void ConfigureDialog::RetranslateUI() {
     ui->tabWidget->setCurrentIndex(old_index);
 }
 
+void ConfigureDialog::HandleApplyButtonClicked() {
+    UISettings::values.configuration_applied = true;
+    ApplyConfiguration();
+}
+
 Q_DECLARE_METATYPE(QList<QWidget*>);
 
 void ConfigureDialog::PopulateSelectionList() {
diff --git a/src/yuzu/configuration/configure_dialog.h b/src/yuzu/configuration/configure_dialog.h
index 570c3b941..abe019635 100644
--- a/src/yuzu/configuration/configure_dialog.h
+++ b/src/yuzu/configuration/configure_dialog.h
@@ -35,9 +35,10 @@ signals:
 
 private:
     void changeEvent(QEvent* event) override;
-
     void RetranslateUI();
 
+    void HandleApplyButtonClicked();
+
     void SetConfiguration();
     void UpdateVisibleTabs();
     void PopulateSelectionList();
diff --git a/src/yuzu/configuration/configure_per_game.cpp b/src/yuzu/configuration/configure_per_game.cpp
index 3e13bd438..d89f1ad4b 100644
--- a/src/yuzu/configuration/configure_per_game.cpp
+++ b/src/yuzu/configuration/configure_per_game.cpp
@@ -6,9 +6,12 @@
 #include <memory>
 #include <utility>
 
+#include <QAbstractButton>
 #include <QCheckBox>
+#include <QDialogButtonBox>
 #include <QHeaderView>
 #include <QMenu>
+#include <QPushButton>
 #include <QStandardItemModel>
 #include <QString>
 #include <QTimer>
@@ -42,6 +45,12 @@ ConfigurePerGame::ConfigurePerGame(QWidget* parent, u64 title_id)
     scene = new QGraphicsScene;
     ui->icon_view->setScene(scene);
 
+    if (Core::System::GetInstance().IsPoweredOn()) {
+        QPushButton* apply_button = ui->buttonBox->addButton(QDialogButtonBox::Apply);
+        connect(apply_button, &QAbstractButton::clicked, this,
+                &ConfigurePerGame::HandleApplyButtonClicked);
+    }
+
     LoadConfiguration();
 }
 
@@ -74,6 +83,11 @@ void ConfigurePerGame::RetranslateUI() {
     ui->retranslateUi(this);
 }
 
+void ConfigurePerGame::HandleApplyButtonClicked() {
+    UISettings::values.configuration_applied = true;
+    ApplyConfiguration();
+}
+
 void ConfigurePerGame::LoadFromFile(FileSys::VirtualFile file) {
     this->file = std::move(file);
     LoadConfiguration();
diff --git a/src/yuzu/configuration/configure_per_game.h b/src/yuzu/configuration/configure_per_game.h
index 5f9a08cef..f6e6ab7c4 100644
--- a/src/yuzu/configuration/configure_per_game.h
+++ b/src/yuzu/configuration/configure_per_game.h
@@ -39,6 +39,8 @@ private:
     void changeEvent(QEvent* event) override;
     void RetranslateUI();
 
+    void HandleApplyButtonClicked();
+
     void LoadConfiguration();
 
     std::unique_ptr<Ui::ConfigurePerGame> ui;
diff --git a/src/yuzu/main.cpp b/src/yuzu/main.cpp
index 37ef62967..0f0e228b0 100644
--- a/src/yuzu/main.cpp
+++ b/src/yuzu/main.cpp
@@ -2587,12 +2587,12 @@ void GMainWindow::OnConfigure() {
             &GMainWindow::OnLanguageChanged);
 
     const auto result = configure_dialog.exec();
-    if (result != QDialog::Accepted) {
+    if (result != QDialog::Accepted && !UISettings::values.configuration_applied) {
         return;
+    } else if (result == QDialog::Accepted) {
+        configure_dialog.ApplyConfiguration();
+        controller_dialog->refreshConfiguration();
     }
-
-    configure_dialog.ApplyConfiguration();
-    controller_dialog->refreshConfiguration();
     InitializeHotkeys();
     if (UISettings::values.theme != old_theme) {
         UpdateUITheme();
@@ -2607,6 +2607,8 @@ void GMainWindow::OnConfigure() {
         game_list->PopulateAsync(UISettings::values.game_dirs);
     }
 
+    UISettings::values.configuration_applied = false;
+
     config->Save();
 
     if ((UISettings::values.hide_mouse || Settings::values.mouse_panning) && emulation_running) {
@@ -2636,23 +2638,27 @@ void GMainWindow::OpenPerGameConfiguration(u64 title_id, const std::string& file
     ConfigurePerGame dialog(this, title_id);
     dialog.LoadFromFile(v_file);
     const auto result = dialog.exec();
-    if (result == QDialog::Accepted) {
+
+    if (result != QDialog::Accepted && !UISettings::values.configuration_applied) {
+        Settings::RestoreGlobalState(system.IsPoweredOn());
+        return;
+    } else if (result == QDialog::Accepted) {
         dialog.ApplyConfiguration();
+    }
 
-        const auto reload = UISettings::values.is_game_list_reload_pending.exchange(false);
-        if (reload) {
-            game_list->PopulateAsync(UISettings::values.game_dirs);
-        }
+    const auto reload = UISettings::values.is_game_list_reload_pending.exchange(false);
+    if (reload) {
+        game_list->PopulateAsync(UISettings::values.game_dirs);
+    }
 
-        // Do not cause the global config to write local settings into the config file
-        const bool is_powered_on = system.IsPoweredOn();
-        Settings::RestoreGlobalState(is_powered_on);
+    // Do not cause the global config to write local settings into the config file
+    const bool is_powered_on = system.IsPoweredOn();
+    Settings::RestoreGlobalState(is_powered_on);
 
-        if (!is_powered_on) {
-            config->Save();
-        }
-    } else {
-        Settings::RestoreGlobalState(system.IsPoweredOn());
+    UISettings::values.configuration_applied = false;
+
+    if (!is_powered_on) {
+        config->Save();
     }
 }
 
diff --git a/src/yuzu/uisettings.h b/src/yuzu/uisettings.h
index 5ba00b8c8..49122ec32 100644
--- a/src/yuzu/uisettings.h
+++ b/src/yuzu/uisettings.h
@@ -95,6 +95,8 @@ struct Values {
     uint8_t row_2_text_id;
     std::atomic_bool is_game_list_reload_pending{false};
     bool cache_game_list;
+
+    bool configuration_applied;
 };
 
 extern Values values;