early-access version 1759

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 1757.
+This is the source code for early-access 1759.
 
 ## Legal Notice
 

externals/dynarmic/src/dynarmic/backend/x64/a64_emit_x64.cpp

@@ -881,11 +881,11 @@ Xbyak::RegExp EmitVAddrLookup(BlockOfCode& code, A64EmitContext& ctx, size_t bit
 }
 
 Xbyak::RegExp EmitFastmemVAddr(BlockOfCode& code, A64EmitContext& ctx, Xbyak::Label& abort, Xbyak::Reg64 vaddr, bool& require_abort_handling) {
-    const size_t unused_top_bits = 64 - ctx.conf.page_table_address_space_bits;
+    const size_t unused_top_bits = 64 - ctx.conf.fastmem_address_space_bits;
 
     if (unused_top_bits == 0) {
         return r13 + vaddr;
-    } else if (ctx.conf.silently_mirror_page_table) {
+    } else if (ctx.conf.silently_mirror_fastmem) {
         Xbyak::Reg64 tmp = ctx.reg_alloc.ScratchGpr();
         if (unused_top_bits < 32) {
             code.mov(tmp, vaddr);
@@ -895,19 +895,19 @@ Xbyak::RegExp EmitFastmemVAddr(BlockOfCode& code, A64EmitContext& ctx, Xbyak::La
             code.mov(tmp.cvt32(), vaddr.cvt32());
         } else {
             code.mov(tmp.cvt32(), vaddr.cvt32());
-            code.and_(tmp, u32((1 << ctx.conf.page_table_address_space_bits) - 1));
+            code.and_(tmp, u32((1 << ctx.conf.fastmem_address_space_bits) - 1));
         }
         return r13 + tmp;
     } else {
-        if (ctx.conf.page_table_address_space_bits < 32) {
+        if (ctx.conf.fastmem_address_space_bits < 32) {
-            code.test(vaddr, u32(-(1 << ctx.conf.page_table_address_space_bits)));
+            code.test(vaddr, u32(-(1 << ctx.conf.fastmem_address_space_bits)));
             code.jnz(abort, code.T_NEAR);
             require_abort_handling = true;
         } else {
             // TODO: Consider having TEST as above but coalesce 64-bit constant in register allocator
             Xbyak::Reg64 tmp = ctx.reg_alloc.ScratchGpr();
             code.mov(tmp, vaddr);
-            code.shr(tmp, int(ctx.conf.page_table_address_space_bits));
+            code.shr(tmp, int(ctx.conf.fastmem_address_space_bits));
             code.jnz(abort, code.T_NEAR);
             require_abort_handling = true;
         }

externals/dynarmic/src/dynarmic/interface/A64/config.h

@@ -202,7 +202,7 @@ struct UserConfig {
     void** page_table = nullptr;
     /// Declares how many valid address bits are there in virtual addresses.
     /// Determines the size of page_table. Valid values are between 12 and 64 inclusive.
-    /// This is only used if page_table or fastmem_pointer is not nullptr.
+    /// This is only used if page_table is not nullptr.
     size_t page_table_address_space_bits = 36;
     /// Masks out the first N bits in host pointers from the page table.
     /// The intention behind this is to allow users of Dynarmic to pack attributes in the
@@ -213,7 +213,7 @@ struct UserConfig {
     /// page table. If true, Dynarmic will silently mirror page_table's address space. If
     /// false, accessing memory outside of page_table bounds will result in a call to the
     /// relevant memory callback.
-    /// This is only used if page_table or fastmem_pointer is not nullptr.
+    /// This is only used if page_table is not nullptr.
     bool silently_mirror_page_table = true;
     /// Determines if the pointer in the page_table shall be offseted locally or globally.
     /// 'false' will access page_table[addr >> bits][addr & mask]
@@ -243,6 +243,16 @@ struct UserConfig {
     /// Recompiled code will use the page_table if this is available, otherwise memory
     /// accesses will hit the memory callbacks.
     bool recompile_on_fastmem_failure = true;
+    /// Declares how many valid address bits are there in virtual addresses.
+    /// Determines the size of fastmem arena. Valid values are between 12 and 64 inclusive.
+    /// This is only used if fastmem_pointer is not nullptr.
+    size_t fastmem_address_space_bits = 36;
+    /// Determines what happens if the guest accesses an entry that is off the end of the
+    /// fastmem arena. If true, Dynarmic will silently mirror fastmem's address space. If
+    /// false, accessing memory outside of fastmem bounds will result in a call to the
+    /// relevant memory callback.
+    /// This is only used if fastmem_pointer is not nullptr.
+    bool silently_mirror_fastmem = true;
 
     /// This option relates to translation. Generally when we run into an unpredictable
     /// instruction the ExceptionRaised callback is called. If this is true, we define

externals/dynarmic/tests/A64/a64.cpp

@@ -852,9 +852,9 @@ TEST_CASE("A64: Memory access (fastmem)", "[a64]") {
     A64FastmemTestEnv env{backing_memory};
     Dynarmic::A64::UserConfig config{&env};
     config.fastmem_pointer = backing_memory;
-    config.page_table_address_space_bits = address_width;
+    config.fastmem_address_space_bits = address_width;
     config.recompile_on_fastmem_failure = false;
-    config.silently_mirror_page_table = true;
+    config.silently_mirror_fastmem = true;
     config.processor_id = 0;
 
     Dynarmic::A64::Jit jit{config};

src/common/CMakeLists.txt

@@ -130,6 +130,8 @@ add_library(common STATIC
     hash.h
     hex_util.cpp
     hex_util.h
+    host_memory.cpp
+    host_memory.h
     intrusive_red_black_tree.h
     logging/backend.cpp
     logging/backend.h

src/common/host_memory.cpp

@@ -0,0 +1,438 @@
+#ifdef __linux__
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+#include <fcntl.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#elif defined(_WIN32) // ^^^ Linux ^^^ vvv Windows vvv
+#ifdef _WIN32_WINNT
+#undef _WIN32_WINNT
+#endif
+#define _WIN32_WINNT 0x0A00 // Windows 10
+
+#include <Windows.h>
+
+#include <boost/icl/separate_interval_set.hpp>
+
+#include <iterator>
+#include <unordered_map>
+
+#pragma comment(lib, "mincore.lib")
+
+#endif // ^^^ Windows ^^^
+
+#include <mutex>
+
+#include "common/alignment.h"
+#include "common/assert.h"
+#include "common/host_memory.h"
+#include "common/logging/log.h"
+
+namespace Common {
+
+static constexpr size_t PageAlignment = 0x1000;
+static constexpr size_t HugePageSize = 0x200000;
+
+#ifdef _WIN32
+
+class HostMemory::Impl {
+public:
+    explicit Impl(size_t backing_size_, size_t virtual_size_)
+        : backing_size{backing_size_}, virtual_size{virtual_size_}, process{GetCurrentProcess()} {
+        // Allocate backing file map
+        backing_handle =
+            CreateFileMapping2(INVALID_HANDLE_VALUE, nullptr, FILE_MAP_WRITE | FILE_MAP_READ,
+                               PAGE_READWRITE, SEC_COMMIT, backing_size, nullptr, nullptr, 0);
+        if (!backing_handle) {
+            throw std::bad_alloc{};
+        }
+        // Allocate a virtual memory for the backing file map as placeholder
+        backing_base = static_cast<u8*>(VirtualAlloc2(process, nullptr, backing_size,
+                                                      MEM_RESERVE | MEM_RESERVE_PLACEHOLDER,
+                                                      PAGE_NOACCESS, nullptr, 0));
+        if (!backing_base) {
+            Release();
+            throw std::bad_alloc{};
+        }
+        // Map backing placeholder
+        void* const ret = MapViewOfFile3(backing_handle, process, backing_base, 0, backing_size,
+                                         MEM_REPLACE_PLACEHOLDER, PAGE_READWRITE, nullptr, 0);
+        if (ret != backing_base) {
+            Release();
+            throw std::bad_alloc{};
+        }
+        // Allocate virtual address placeholder
+        virtual_base = static_cast<u8*>(VirtualAlloc2(process, nullptr, virtual_size,
+                                                      MEM_RESERVE | MEM_RESERVE_PLACEHOLDER,
+                                                      PAGE_NOACCESS, nullptr, 0));
+        if (!virtual_base) {
+            Release();
+            throw std::bad_alloc{};
+        }
+    }
+
+    ~Impl() {
+        Release();
+    }
+
+    void Map(size_t virtual_offset, size_t host_offset, size_t length) {
+        std::unique_lock lock{placeholder_mutex};
+        if (!IsNiechePlaceholder(virtual_offset, length)) {
+            Split(virtual_offset, length);
+        }
+        ASSERT(placeholders.find({virtual_offset, virtual_offset + length}) == placeholders.end());
+        TrackPlaceholder(virtual_offset, host_offset, length);
+
+        MapView(virtual_offset, host_offset, length);
+    }
+
+    void Unmap(size_t virtual_offset, size_t length) {
+        std::lock_guard lock{placeholder_mutex};
+
+        // Unmap until there are no more placeholders
+        while (UnmapOnePlaceholder(virtual_offset, length)) {
+        }
+    }
+
+    void Protect(size_t virtual_offset, size_t length, bool read, bool write) {
+        DWORD new_flags{};
+        if (read && write) {
+            new_flags = PAGE_READWRITE;
+        } else if (read && !write) {
+            new_flags = PAGE_READONLY;
+        } else if (!read && !write) {
+            new_flags = PAGE_NOACCESS;
+        } else {
+            UNIMPLEMENTED_MSG("Protection flag combination read={} write={}", read, write);
+        }
+        const size_t virtual_end = virtual_offset + length;
+
+        std::lock_guard lock{placeholder_mutex};
+        auto [it, end] = placeholders.equal_range({virtual_offset, virtual_end});
+        while (it != end) {
+            const size_t offset = std::max(it->lower(), virtual_offset);
+            const size_t protect_length = std::min(it->upper(), virtual_end) - offset;
+            DWORD old_flags{};
+            if (!VirtualProtect(virtual_base + offset, protect_length, new_flags, &old_flags)) {
+                LOG_CRITICAL(HW_Memory, "Failed to change virtual memory protect rules");
+            }
+            ++it;
+        }
+    }
+
+    const size_t backing_size; ///< Size of the backing memory in bytes
+    const size_t virtual_size; ///< Size of the virtual address placeholder in bytes
+
+    u8* backing_base{};
+    u8* virtual_base{};
+
+private:
+    /// Release all resources in the object
+    void Release() {
+        if (!placeholders.empty()) {
+            for (const auto& placeholder : placeholders) {
+                if (!UnmapViewOfFile2(process, virtual_base + placeholder.lower(),
+                                      MEM_PRESERVE_PLACEHOLDER)) {
+                    LOG_CRITICAL(HW_Memory, "Failed to unmap virtual memory placeholder");
+                }
+            }
+            Coalesce(0, virtual_size);
+        }
+        if (virtual_base) {
+            if (!VirtualFree(virtual_base, 0, MEM_RELEASE)) {
+                LOG_CRITICAL(HW_Memory, "Failed to free virtual memory");
+            }
+        }
+        if (backing_base) {
+            if (!UnmapViewOfFile2(process, backing_base, MEM_PRESERVE_PLACEHOLDER)) {
+                LOG_CRITICAL(HW_Memory, "Failed to unmap backing memory placeholder");
+            }
+            if (!VirtualFreeEx(process, backing_base, 0, MEM_RELEASE)) {
+                LOG_CRITICAL(HW_Memory, "Failed to free backing memory");
+            }
+        }
+        if (!CloseHandle(backing_handle)) {
+            LOG_CRITICAL(HW_Memory, "Failed to free backing memory file handle");
+        }
+    }
+
+    /// Unmap one placeholder in the given range (partial unmaps are supported)
+    /// Return true when there are no more placeholders to unmap
+    bool UnmapOnePlaceholder(size_t virtual_offset, size_t length) {
+        const auto it = placeholders.find({virtual_offset, virtual_offset + length});
+        const auto begin = placeholders.begin();
+        const auto end = placeholders.end();
+        if (it == end) {
+            return false;
+        }
+        const size_t placeholder_begin = it->lower();
+        const size_t placeholder_end = it->upper();
+        const size_t unmap_begin = std::max(virtual_offset, placeholder_begin);
+        const size_t unmap_end = std::min(virtual_offset + length, placeholder_end);
+        ASSERT(unmap_begin >= placeholder_begin && unmap_begin < placeholder_end);
+        ASSERT(unmap_end <= placeholder_end && unmap_end > placeholder_begin);
+
+        const auto host_pointer_it = placeholder_host_pointers.find(placeholder_begin);
+        ASSERT(host_pointer_it != placeholder_host_pointers.end());
+        const size_t host_offset = host_pointer_it->second;
+
+        const bool split_left = unmap_begin > placeholder_begin;
+        const bool split_right = unmap_end < placeholder_end;
+
+        if (!UnmapViewOfFile2(process, virtual_base + placeholder_begin,
+                              MEM_PRESERVE_PLACEHOLDER)) {
+            LOG_CRITICAL(HW_Memory, "Failed to unmap placeholder");
+        }
+        // If we have to remap memory regions due to partial unmaps, we are in a data race as
+        // Windows doesn't support remapping memory without unmapping first. Avoid adding any extra
+        // logic within the panic region described below.
+
+        // Panic region, we are in a data race right now
+        if (split_left || split_right) {
+            Split(unmap_begin, unmap_end - unmap_begin);
+        }
+        if (split_left) {
+            MapView(placeholder_begin, host_offset, unmap_begin - placeholder_begin);
+        }
+        if (split_right) {
+            MapView(unmap_end, host_offset + unmap_end - placeholder_begin,
+                    placeholder_end - unmap_end);
+        }
+        // End panic region
+
+        size_t coalesce_begin = unmap_begin;
+        if (!split_left) {
+            // Try to coalesce pages to the left
+            coalesce_begin = it == begin ? 0 : std::prev(it)->upper();
+            if (coalesce_begin != placeholder_begin) {
+                Coalesce(coalesce_begin, unmap_end - coalesce_begin);
+            }
+        }
+        if (!split_right) {
+            // Try to coalesce pages to the right
+            const auto next = std::next(it);
+            const size_t next_begin = next == end ? virtual_size : next->lower();
+            if (placeholder_end != next_begin) {
+                // We can coalesce to the right
+                Coalesce(coalesce_begin, next_begin - coalesce_begin);
+            }
+        }
+        // Remove and reinsert placeholder trackers
+        UntrackPlaceholder(it);
+        if (split_left) {
+            TrackPlaceholder(placeholder_begin, host_offset, unmap_begin - placeholder_begin);
+        }
+        if (split_right) {
+            TrackPlaceholder(unmap_end, host_offset + unmap_end - placeholder_begin,
+                             placeholder_end - unmap_end);
+        }
+        return true;
+    }
+
+    void MapView(size_t virtual_offset, size_t host_offset, size_t length) {
+        if (!MapViewOfFile3(backing_handle, process, virtual_base + virtual_offset, host_offset,
+                            length, MEM_REPLACE_PLACEHOLDER, PAGE_READWRITE, nullptr, 0)) {
+            LOG_CRITICAL(HW_Memory, "Failed to map placeholder");
+        }
+    }
+
+    void Split(size_t virtual_offset, size_t length) {
+        if (!VirtualFreeEx(process, reinterpret_cast<LPVOID>(virtual_base + virtual_offset), length,
+                           MEM_RELEASE | MEM_PRESERVE_PLACEHOLDER)) {
+            LOG_CRITICAL(HW_Memory, "Failed to split placeholder");
+        }
+    }
+
+    void Coalesce(size_t virtual_offset, size_t length) {
+        if (!VirtualFreeEx(process, reinterpret_cast<LPVOID>(virtual_base + virtual_offset), length,
+                           MEM_RELEASE | MEM_COALESCE_PLACEHOLDERS)) {
+            LOG_CRITICAL(HW_Memory, "Failed to coalesce placeholders");
+        }
+    }
+
+    void TrackPlaceholder(size_t virtual_offset, size_t host_offset, size_t length) {
+        placeholders.insert({virtual_offset, virtual_offset + length});
+        placeholder_host_pointers.emplace(virtual_offset, host_offset);
+    }
+
+    void UntrackPlaceholder(boost::icl::separate_interval_set<size_t>::iterator it) {
+        placeholders.erase(it);
+        placeholder_host_pointers.erase(it->lower());
+    }
+
+    /// Return true when a given memory region is a "nieche" and the placeholders don't have to be
+    /// splitted.
+    bool IsNiechePlaceholder(size_t virtual_offset, size_t length) const {
+        const auto it = placeholders.upper_bound({virtual_offset, virtual_offset + length});
+        if (it != placeholders.end() && it->lower() == virtual_offset + length) {
+            const bool is_root = it == placeholders.begin() && virtual_offset == 0;
+            return is_root || std::prev(it)->upper() == virtual_offset;
+        }
+        return false;
+    }
+
+    HANDLE process{};        ///< Current process handle
+    HANDLE backing_handle{}; ///< File based backing memory
+
+    std::mutex placeholder_mutex;                                 ///< Mutex for placeholders
+    boost::icl::separate_interval_set<size_t> placeholders;       ///< Mapped placeholders
+    std::unordered_map<size_t, size_t> placeholder_host_pointers; ///< Placeholder backing offset
+};
+
+#elif defined(__linux__)
+
+class HostMemory::Impl {
+public:
+    explicit Impl(size_t backing_size_, size_t virtual_size_)
+        : backing_size{backing_size_}, virtual_size{virtual_size_} {
+
+        // Backing memory initialization
+        fd = memfd_create("HostMemory", 0);
+        if (fd == -1) {
+            LOG_CRITICAL(HW_Memory, "memfd_create failed: {}", strerror(errno));
+            Release();
+            throw std::bad_alloc{};
+        }
+
+        // Defined to extend the file with zeros
+        int ret = ftruncate(fd, backing_size);
+        if (ret != 0) {
+            LOG_CRITICAL(HW_Memory, "ftruncate failed with {}, are you out-of-memory?",
+                         strerror(errno));
+            Release();
+            throw std::bad_alloc{};
+        }
+
+        backing_base = static_cast<u8*>(
+            mmap(nullptr, backing_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0));
+        if (backing_base == MAP_FAILED) {
+            LOG_CRITICAL(HW_Memory, "mmap failed: {}", strerror(errno));
+            Release();
+            throw std::bad_alloc{};
+        }
+
+        // Virtual memory initialization
+        virtual_base = static_cast<u8*>(
+            mmap(nullptr, virtual_size, PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0));
+        if (virtual_base == MAP_FAILED) {
+            LOG_CRITICAL(HW_Memory, "mmap failed: {}", strerror(errno));
+            Release();
+            throw std::bad_alloc{};
+        }
+    }
+
+    ~Impl() {
+        Release();
+    }
+
+    void Map(size_t virtual_offset, size_t host_offset, size_t length) {
+
+        void* ret = mmap(virtual_base + virtual_offset, length, PROT_READ | PROT_WRITE,
+                         MAP_SHARED | MAP_FIXED, fd, host_offset);
+        ASSERT_MSG(ret != MAP_FAILED, "mmap failed: {}", strerror(errno));
+    }
+
+    void Unmap(size_t virtual_offset, size_t length) {
+        // The method name is wrong. We're still talking about the virtual range.
+        // We don't want to unmap, we want to reserve this memory.
+
+        void* ret = mmap(virtual_base + virtual_offset, length, PROT_NONE,
+                         MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0);
+        ASSERT_MSG(ret != MAP_FAILED, "mmap failed: {}", strerror(errno));
+    }
+
+    void Protect(size_t virtual_offset, size_t length, bool read, bool write) {
+        int flags = 0;
+        if (read)
+            flags |= PROT_READ;
+        if (write)
+            flags |= PROT_WRITE;
+        int ret = mprotect(virtual_base + virtual_offset, length, flags);
+        ASSERT_MSG(ret == 0, "mprotect failed: {}", strerror(errno));
+    }
+
+    const size_t backing_size; ///< Size of the backing memory in bytes
+    const size_t virtual_size; ///< Size of the virtual address placeholder in bytes
+
+    u8* backing_base{reinterpret_cast<u8*>(MAP_FAILED)};
+    u8* virtual_base{reinterpret_cast<u8*>(MAP_FAILED)};
+
+private:
+    /// Release all resources in the object
+    void Release() {
+        if (virtual_base != MAP_FAILED) {
+            int ret = munmap(virtual_base, virtual_size);
+            ASSERT_MSG(ret == 0, "munmap failed: {}", strerror(errno));
+        }
+
+        if (backing_base != MAP_FAILED) {
+            int ret = munmap(backing_base, backing_size);
+            ASSERT_MSG(ret == 0, "munmap failed: {}", strerror(errno));
+        }
+
+        if (fd != -1) {
+            int ret = close(fd);
+            ASSERT_MSG(ret == 0, "close failed: {}", strerror(errno));
+        }
+    }
+
+    int fd{-1}; // memfd file descriptor, -1 is the error value of memfd_create
+};
+
+#else
+
+#error Please implement the host memory for your platform
+
+#endif
+
+HostMemory::HostMemory(size_t backing_size_, size_t virtual_size_)
+    : backing_size(backing_size_),
+      virtual_size(virtual_size_), impl{std::make_unique<HostMemory::Impl>(
+                                       AlignUp(backing_size, PageAlignment),
+                                       AlignUp(virtual_size, PageAlignment) + 3 * HugePageSize)},
+      backing_base{impl->backing_base}, virtual_base{impl->virtual_base} {
+    virtual_base += 2 * HugePageSize - 1;
+    virtual_base -= reinterpret_cast<size_t>(virtual_base) & (HugePageSize - 1);
+    virtual_base_offset = virtual_base - impl->virtual_base;
+}
+
+HostMemory::~HostMemory() = default;
+
+HostMemory::HostMemory(HostMemory&&) noexcept = default;
+
+HostMemory& HostMemory::operator=(HostMemory&&) noexcept = default;
+
+void HostMemory::Map(size_t virtual_offset, size_t host_offset, size_t length) {
+    ASSERT(virtual_offset % PageAlignment == 0);
+    ASSERT(host_offset % PageAlignment == 0);
+    ASSERT(length % PageAlignment == 0);
+    ASSERT(virtual_offset + length <= virtual_size);
+    ASSERT(host_offset + length <= backing_size);
+    if (length == 0) {
+        return;
+    }
+    impl->Map(virtual_offset + virtual_base_offset, host_offset, length);
+}
+
+void HostMemory::Unmap(size_t virtual_offset, size_t length) {
+    ASSERT(virtual_offset % PageAlignment == 0);
+    ASSERT(length % PageAlignment == 0);
+    ASSERT(virtual_offset + length <= virtual_size);
+    if (length == 0) {
+        return;
+    }
+    impl->Unmap(virtual_offset + virtual_base_offset, length);
+}
+
+void HostMemory::Protect(size_t virtual_offset, size_t length, bool read, bool write) {
+    ASSERT(virtual_offset % PageAlignment == 0);
+    ASSERT(length % PageAlignment == 0);
+    ASSERT(virtual_offset + length <= virtual_size);
+    if (length == 0) {
+        return;
+    }
+    impl->Protect(virtual_offset + virtual_base_offset, length, read, write);
+}
+
+} // namespace Common

src/common/host_memory.h

@@ -0,0 +1,66 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include "common/common_types.h"
+
+namespace Common {
+
+/**
+ * A low level linear memory buffer, which supports multiple mappings
+ * Its purpose is to rebuild a given sparse memory layout, including mirrors.
+ */
+class HostMemory {
+public:
+    explicit HostMemory(size_t backing_size_, size_t virtual_size_);
+    ~HostMemory();
+
+    /**
+     * Copy constructors. They shall return a copy of the buffer without the mappings.
+     * TODO: Implement them with COW if needed.
+     */
+    HostMemory(HostMemory& other) = delete;
+    HostMemory& operator=(HostMemory& other) = delete;
+
+    /**
+     * Move constructors. They will move the buffer and the mappings to the new object.
+     */
+    HostMemory(HostMemory&& other) noexcept;
+    HostMemory& operator=(HostMemory&& other) noexcept;
+
+    void Map(size_t virtual_offset, size_t host_offset, size_t length);
+
+    void Unmap(size_t virtual_offset, size_t length);
+
+    void Protect(size_t virtual_offset, size_t length, bool read, bool write);
+
+    [[nodiscard]] u8* BackingBasePointer() noexcept {
+        return backing_base;
+    }
+    [[nodiscard]] const u8* BackingBasePointer() const noexcept {
+        return backing_base;
+    }
+
+    [[nodiscard]] u8* VirtualBasePointer() noexcept {
+        return virtual_base;
+    }
+    [[nodiscard]] const u8* VirtualBasePointer() const noexcept {
+        return virtual_base;
+    }
+
+private:
+    size_t backing_size{};
+    size_t virtual_size{};
+
+    // Low level handler for the platform dependent memory routines
+    class Impl;
+    std::unique_ptr<Impl> impl;
+    u8* backing_base{};
+    u8* virtual_base{};
+    size_t virtual_base_offset{};
+};
+
+} // namespace Common

src/common/page_table.h

@@ -111,6 +111,8 @@ struct PageTable {
     VirtualBuffer<u64> backing_addr;
 
     size_t current_address_space_width_in_bits;
+
+    u8* fastmem_arena;
 };
 
 } // namespace Common

src/common/settings.cpp

@@ -90,6 +90,13 @@ bool IsGPULevelHigh() {
            values.gpu_accuracy.GetValue() == GPUAccuracy::High;
 }
 
+bool IsFastmemEnabled() {
+    if (values.cpu_accuracy.GetValue() == CPUAccuracy::DebugMode) {
+        return values.cpuopt_fastmem;
+    }
+    return true;
+}
+
 float Volume() {
     if (values.audio_muted) {
         return 0.0f;
@@ -115,6 +122,7 @@ void RestoreGlobalState(bool is_powered_on) {
     values.cpuopt_unsafe_unfuse_fma.SetGlobal(true);
     values.cpuopt_unsafe_reduce_fp_error.SetGlobal(true);
     values.cpuopt_unsafe_inaccurate_nan.SetGlobal(true);
+    values.cpuopt_unsafe_fastmem_check.SetGlobal(true);
 
     // Renderer
     values.renderer_backend.SetGlobal(true);

src/common/settings.h

@@ -125,10 +125,12 @@ struct Values {
     bool cpuopt_const_prop;
     bool cpuopt_misc_ir;
     bool cpuopt_reduce_misalign_checks;
+    bool cpuopt_fastmem;
 
     Setting<bool> cpuopt_unsafe_unfuse_fma;
     Setting<bool> cpuopt_unsafe_reduce_fp_error;
     Setting<bool> cpuopt_unsafe_inaccurate_nan;
+    Setting<bool> cpuopt_unsafe_fastmem_check;
 
     // Renderer
     Setting<RendererBackend> renderer_backend;
@@ -249,6 +251,8 @@ void SetConfiguringGlobal(bool is_global);
 bool IsGPULevelExtreme();
 bool IsGPULevelHigh();
 
+bool IsFastmemEnabled();
+
 float Volume();
 
 std::string GetTimeZoneString();

src/core/arm/dynarmic/arm_dynarmic_32.cpp

@@ -128,6 +128,7 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable*
     if (page_table) {
         config.page_table = reinterpret_cast<std::array<std::uint8_t*, NUM_PAGE_TABLE_ENTRIES>*>(
             page_table->pointers.data());
+        config.fastmem_pointer = page_table->fastmem_arena;
     }
     config.absolute_offset_page_table = true;
     config.page_table_pointer_mask_bits = Common::PageTable::ATTRIBUTE_BITS;
@@ -143,7 +144,7 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable*
 
     // Code cache size
     config.code_cache_size = 512 * 1024 * 1024;
-    config.far_code_offset = 256 * 1024 * 1024;
+    config.far_code_offset = 400 * 1024 * 1024;
 
     // Safe optimizations
     if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::DebugMode) {
@@ -171,6 +172,9 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable*
         if (!Settings::values.cpuopt_reduce_misalign_checks) {
             config.only_detect_misalignment_via_page_table_on_page_boundary = false;
         }
+        if (!Settings::values.cpuopt_fastmem) {
+            config.fastmem_pointer = nullptr;
+        }
     }
 
     // Unsafe optimizations

src/core/arm/dynarmic/arm_dynarmic_64.cpp

@@ -160,6 +160,10 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
         config.absolute_offset_page_table = true;
         config.detect_misaligned_access_via_page_table = 16 | 32 | 64 | 128;
         config.only_detect_misalignment_via_page_table_on_page_boundary = true;
+
+        config.fastmem_pointer = page_table->fastmem_arena;
+        config.fastmem_address_space_bits = address_space_bits;
+        config.silently_mirror_fastmem = false;
     }
 
     // Multi-process state
@@ -181,7 +185,7 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
 
     // Code cache size
     config.code_cache_size = 512 * 1024 * 1024;
-    config.far_code_offset = 256 * 1024 * 1024;
+    config.far_code_offset = 400 * 1024 * 1024;
 
     // Safe optimizations
     if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::DebugMode) {
@@ -209,6 +213,9 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
         if (!Settings::values.cpuopt_reduce_misalign_checks) {
             config.only_detect_misalignment_via_page_table_on_page_boundary = false;
         }
+        if (!Settings::values.cpuopt_fastmem) {
+            config.fastmem_pointer = nullptr;
+        }
     }
 
     // Unsafe optimizations
@@ -223,6 +230,9 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
         if (Settings::values.cpuopt_unsafe_inaccurate_nan.GetValue()) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_InaccurateNaN;
         }
+        if (Settings::values.cpuopt_unsafe_fastmem_check.GetValue()) {
+            config.fastmem_address_space_bits = 64;
+        }
     }
 
     return std::make_shared<Dynarmic::A64::Jit>(config);

src/core/device_memory.cpp

@@ -6,7 +6,7 @@
 
 namespace Core {
 
-DeviceMemory::DeviceMemory() : buffer{DramMemoryMap::Size} {}
+DeviceMemory::DeviceMemory() : buffer{DramMemoryMap::Size, 1ULL << 39} {}
 DeviceMemory::~DeviceMemory() = default;
 
 } // namespace Core

src/core/device_memory.h

@@ -5,7 +5,7 @@
 #pragma once
 
 #include "common/common_types.h"
-#include "common/virtual_buffer.h"
+#include "common/host_memory.h"
 
 namespace Core {
 
@@ -21,27 +21,30 @@ enum : u64 {
 };
 }; // namespace DramMemoryMap
 
-class DeviceMemory : NonCopyable {
+class DeviceMemory {
 public:
     explicit DeviceMemory();
     ~DeviceMemory();
 
+    DeviceMemory& operator=(const DeviceMemory&) = delete;
+    DeviceMemory(const DeviceMemory&) = delete;
+
     template <typename T>
     PAddr GetPhysicalAddr(const T* ptr) const {
-        return (reinterpret_cast<uintptr_t>(ptr) - reinterpret_cast<uintptr_t>(buffer.data())) +
+        return (reinterpret_cast<uintptr_t>(ptr) -
+                reinterpret_cast<uintptr_t>(buffer.BackingBasePointer())) +
                DramMemoryMap::Base;
     }
 
     u8* GetPointer(PAddr addr) {
-        return buffer.data() + (addr - DramMemoryMap::Base);
+        return buffer.BackingBasePointer() + (addr - DramMemoryMap::Base);
     }
 
     const u8* GetPointer(PAddr addr) const {
-        return buffer.data() + (addr - DramMemoryMap::Base);
+        return buffer.BackingBasePointer() + (addr - DramMemoryMap::Base);
     }
 
-private:
+    Common::HostMemory buffer;
-    Common::VirtualBuffer<u8> buffer;
 };
 
 } // namespace Core

src/core/hle/kernel/hle_ipc.h

@@ -85,8 +85,8 @@ public:
      */
     void ClientDisconnected(KServerSession* session);
 
-    std::weak_ptr<ServiceThread> GetServiceThread() const {
+    std::shared_ptr<ServiceThread> GetServiceThread() const {
-        return service_thread;
+        return service_thread.lock();
     }
 
 protected:
@@ -152,7 +152,7 @@ public:
         session_handler = std::move(handler);
     }
 
-    std::weak_ptr<ServiceThread> GetServiceThread() const {
+    std::shared_ptr<ServiceThread> GetServiceThread() const {
         return session_handler->GetServiceThread();
     }
 

src/core/hle/kernel/k_server_session.cpp

@@ -119,9 +119,11 @@ ResultCode KServerSession::QueueSyncRequest(KThread* thread, Core::Memory::Memor
 
     context->PopulateFromIncomingCommandBuffer(kernel.CurrentProcess()->GetHandleTable(), cmd_buf);
 
-    if (auto strong_ptr = manager->GetServiceThread().lock()) {
+    if (auto strong_ptr = manager->GetServiceThread(); strong_ptr) {
         strong_ptr->QueueSyncRequest(*parent, std::move(context));
         return ResultSuccess;
+    } else {
+        ASSERT(false, "strong_ptr was nullptr!");
     }
 
     return ResultSuccess;

src/core/memory.cpp

@@ -12,6 +12,7 @@
 #include "common/common_types.h"
 #include "common/logging/log.h"
 #include "common/page_table.h"
+#include "common/settings.h"
 #include "common/swap.h"
 #include "core/arm/arm_interface.h"
 #include "core/core.h"
@@ -32,6 +33,7 @@ struct Memory::Impl {
 
     void SetCurrentPageTable(Kernel::KProcess& process, u32 core_id) {
         current_page_table = &process.PageTable().PageTableImpl();
+        current_page_table->fastmem_arena = system.DeviceMemory().buffer.VirtualBasePointer();
 
         const std::size_t address_space_width = process.PageTable().GetAddressSpaceWidth();
 
@@ -41,13 +43,23 @@ struct Memory::Impl {
     void MapMemoryRegion(Common::PageTable& page_table, VAddr base, u64 size, PAddr target) {
         ASSERT_MSG((size & PAGE_MASK) == 0, "non-page aligned size: {:016X}", size);
         ASSERT_MSG((base & PAGE_MASK) == 0, "non-page aligned base: {:016X}", base);
+        ASSERT_MSG(target >= DramMemoryMap::Base && target < DramMemoryMap::End,
+                   "Out of bounds target: {:016X}", target);
         MapPages(page_table, base / PAGE_SIZE, size / PAGE_SIZE, target, Common::PageType::Memory);
+
+        if (Settings::IsFastmemEnabled()) {
+            system.DeviceMemory().buffer.Map(base, target - DramMemoryMap::Base, size);
+        }
     }
 
     void UnmapRegion(Common::PageTable& page_table, VAddr base, u64 size) {
         ASSERT_MSG((size & PAGE_MASK) == 0, "non-page aligned size: {:016X}", size);
         ASSERT_MSG((base & PAGE_MASK) == 0, "non-page aligned base: {:016X}", base);
         MapPages(page_table, base / PAGE_SIZE, size / PAGE_SIZE, 0, Common::PageType::Unmapped);
+
+        if (Settings::IsFastmemEnabled()) {
+            system.DeviceMemory().buffer.Unmap(base, size);
+        }
     }
 
     bool IsValidVirtualAddress(const Kernel::KProcess& process, const VAddr vaddr) const {
@@ -466,6 +478,12 @@ struct Memory::Impl {
         if (vaddr == 0) {
             return;
         }
+
+        if (Settings::IsFastmemEnabled()) {
+            const bool is_read_enable = Settings::IsGPULevelHigh() || !cached;
+            system.DeviceMemory().buffer.Protect(vaddr, size, is_read_enable, !cached);
+        }
+
         // Iterate over a contiguous CPU address space, which corresponds to the specified GPU
         // address space, marking the region as un/cached. The region is marked un/cached at a
         // granularity of CPU pages, hence why we iterate on a CPU page basis (note: GPU page size

src/tests/CMakeLists.txt

@@ -2,6 +2,7 @@ add_executable(tests
     common/bit_field.cpp
     common/cityhash.cpp
     common/fibers.cpp
+    common/host_memory.cpp
     common/param_package.cpp
     common/ring_buffer.cpp
     core/core_timing.cpp

src/tests/common/host_memory.cpp

@@ -0,0 +1,183 @@
+// Copyright 2021 yuzu emulator team
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <catch2/catch.hpp>
+
+#include "common/host_memory.h"
+
+using Common::HostMemory;
+
+static constexpr size_t VIRTUAL_SIZE = 1ULL << 39;
+static constexpr size_t BACKING_SIZE = 4ULL * 1024 * 1024 * 1024;
+
+TEST_CASE("HostMemory: Initialize and deinitialize", "[common]") {
+    { HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE); }
+    { HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE); }
+}
+
+TEST_CASE("HostMemory: Simple map", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x5000, 0x8000, 0x1000);
+
+    volatile u8* const data = mem.VirtualBasePointer() + 0x5000;
+    data[0] = 50;
+    REQUIRE(data[0] == 50);
+}
+
+TEST_CASE("HostMemory: Simple mirror map", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x5000, 0x3000, 0x2000);
+    mem.Map(0x8000, 0x4000, 0x1000);
+
+    volatile u8* const mirror_a = mem.VirtualBasePointer() + 0x5000;
+    volatile u8* const mirror_b = mem.VirtualBasePointer() + 0x8000;
+    mirror_b[0] = 76;
+    REQUIRE(mirror_a[0x1000] == 76);
+}
+
+TEST_CASE("HostMemory: Simple unmap", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x5000, 0x3000, 0x2000);
+
+    volatile u8* const data = mem.VirtualBasePointer() + 0x5000;
+    data[75] = 50;
+    REQUIRE(data[75] == 50);
+
+    mem.Unmap(0x5000, 0x2000);
+}
+
+TEST_CASE("HostMemory: Simple unmap and remap", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x5000, 0x3000, 0x2000);
+
+    volatile u8* const data = mem.VirtualBasePointer() + 0x5000;
+    data[0] = 50;
+    REQUIRE(data[0] == 50);
+
+    mem.Unmap(0x5000, 0x2000);
+
+    mem.Map(0x5000, 0x3000, 0x2000);
+    REQUIRE(data[0] == 50);
+
+    mem.Map(0x7000, 0x2000, 0x5000);
+    REQUIRE(data[0x3000] == 50);
+}
+
+TEST_CASE("HostMemory: Nieche allocation", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x0000, 0, 0x20000);
+    mem.Unmap(0x0000, 0x4000);
+    mem.Map(0x1000, 0, 0x2000);
+    mem.Map(0x3000, 0, 0x1000);
+    mem.Map(0, 0, 0x1000);
+}
+
+TEST_CASE("HostMemory: Full unmap", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x8000, 0, 0x4000);
+    mem.Unmap(0x8000, 0x4000);
+    mem.Map(0x6000, 0, 0x16000);
+}
+
+TEST_CASE("HostMemory: Right out of bounds unmap", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x0000, 0, 0x4000);
+    mem.Unmap(0x2000, 0x4000);
+    mem.Map(0x2000, 0x80000, 0x4000);
+}
+
+TEST_CASE("HostMemory: Left out of bounds unmap", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x8000, 0, 0x4000);
+    mem.Unmap(0x6000, 0x4000);
+    mem.Map(0x8000, 0, 0x2000);
+}
+
+TEST_CASE("HostMemory: Multiple placeholder unmap", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x0000, 0, 0x4000);
+    mem.Map(0x4000, 0, 0x1b000);
+    mem.Unmap(0x3000, 0x1c000);
+    mem.Map(0x3000, 0, 0x20000);
+}
+
+TEST_CASE("HostMemory: Unmap between placeholders", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x0000, 0, 0x4000);
+    mem.Map(0x4000, 0, 0x4000);
+    mem.Unmap(0x2000, 0x4000);
+    mem.Map(0x2000, 0, 0x4000);
+}
+
+TEST_CASE("HostMemory: Unmap to origin", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x4000, 0, 0x4000);
+    mem.Map(0x8000, 0, 0x4000);
+    mem.Unmap(0x4000, 0x4000);
+    mem.Map(0, 0, 0x4000);
+    mem.Map(0x4000, 0, 0x4000);
+}
+
+TEST_CASE("HostMemory: Unmap to right", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x4000, 0, 0x4000);
+    mem.Map(0x8000, 0, 0x4000);
+    mem.Unmap(0x8000, 0x4000);
+    mem.Map(0x8000, 0, 0x4000);
+}
+
+TEST_CASE("HostMemory: Partial right unmap check bindings", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x4000, 0x10000, 0x4000);
+
+    volatile u8* const ptr = mem.VirtualBasePointer() + 0x4000;
+    ptr[0x1000] = 17;
+
+    mem.Unmap(0x6000, 0x2000);
+
+    REQUIRE(ptr[0x1000] == 17);
+}
+
+TEST_CASE("HostMemory: Partial left unmap check bindings", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x4000, 0x10000, 0x4000);
+
+    volatile u8* const ptr = mem.VirtualBasePointer() + 0x4000;
+    ptr[0x3000] = 19;
+    ptr[0x3fff] = 12;
+
+    mem.Unmap(0x4000, 0x2000);
+
+    REQUIRE(ptr[0x3000] == 19);
+    REQUIRE(ptr[0x3fff] == 12);
+}
+
+TEST_CASE("HostMemory: Partial middle unmap check bindings", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x4000, 0x10000, 0x4000);
+
+    volatile u8* const ptr = mem.VirtualBasePointer() + 0x4000;
+    ptr[0x0000] = 19;
+    ptr[0x3fff] = 12;
+
+    mem.Unmap(0x1000, 0x2000);
+
+    REQUIRE(ptr[0x0000] == 19);
+    REQUIRE(ptr[0x3fff] == 12);
+}
+
+TEST_CASE("HostMemory: Partial sparse middle unmap and check bindings", "[common]") {
+    HostMemory mem(BACKING_SIZE, VIRTUAL_SIZE);
+    mem.Map(0x4000, 0x10000, 0x2000);
+    mem.Map(0x6000, 0x20000, 0x2000);
+
+    volatile u8* const ptr = mem.VirtualBasePointer() + 0x4000;
+    ptr[0x0000] = 19;
+    ptr[0x3fff] = 12;
+
+    mem.Unmap(0x5000, 0x2000);
+
+    REQUIRE(ptr[0x0000] == 19);
+    REQUIRE(ptr[0x3fff] == 12);
+}

src/video_core/gpu_thread.cpp

@@ -99,25 +99,13 @@ void ThreadManager::FlushRegion(VAddr addr, u64 size) {
         PushCommand(FlushRegionCommand(addr, size));
         return;
     }
-
+    if (!Settings::IsGPULevelExtreme()) {
-    // Asynchronous GPU mode
+        return;
-    switch (Settings::values.gpu_accuracy.GetValue()) {
-    case Settings::GPUAccuracy::Normal:
-        PushCommand(FlushRegionCommand(addr, size));
-        break;
-    case Settings::GPUAccuracy::High:
-        // TODO(bunnei): Is this right? Preserving existing behavior for now
-        break;
-    case Settings::GPUAccuracy::Extreme: {
-        auto& gpu = system.GPU();
-        u64 fence = gpu.RequestFlush(addr, size);
-        PushCommand(GPUTickCommand(), true);
-        ASSERT(fence <= gpu.CurrentFlushRequestFence());
-        break;
-    }
-    default:
-        UNIMPLEMENTED_MSG("Unsupported gpu_accuracy {}", Settings::values.gpu_accuracy.GetValue());
     }
+    auto& gpu = system.GPU();
+    u64 fence = gpu.RequestFlush(addr, size);
+    PushCommand(GPUTickCommand(), true);
+    ASSERT(fence <= gpu.CurrentFlushRequestFence());
 }
 
 void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {

src/video_core/rasterizer_accelerated.cpp

@@ -2,6 +2,8 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <atomic>
+
 #include "common/assert.h"
 #include "common/common_types.h"
 #include "common/div_ceil.h"
@@ -10,35 +12,59 @@
 
 namespace VideoCore {
 
-RasterizerAccelerated::RasterizerAccelerated(Core::Memory::Memory& cpu_memory_)
+using namespace Core::Memory;
-    : cpu_memory{cpu_memory_} {}
+
+RasterizerAccelerated::RasterizerAccelerated(Memory& cpu_memory_) : cpu_memory{cpu_memory_} {}
 
 RasterizerAccelerated::~RasterizerAccelerated() = default;
 
 void RasterizerAccelerated::UpdatePagesCachedCount(VAddr addr, u64 size, int delta) {
-    const auto page_end = Common::DivCeil(addr + size, Core::Memory::PAGE_SIZE);
+    u64 uncache_begin = 0;
-    for (auto page = addr >> Core::Memory::PAGE_BITS; page != page_end; ++page) {
+    u64 cache_begin = 0;
-        auto& count = cached_pages.at(page >> 2).Count(page);
+    u64 uncache_bytes = 0;
+    u64 cache_bytes = 0;
+
+    std::atomic_thread_fence(std::memory_order_acquire);
+    const u64 page_end = Common::DivCeil(addr + size, PAGE_SIZE);
+    for (u64 page = addr >> PAGE_BITS; page != page_end; ++page) {
+        std::atomic_uint16_t& count = cached_pages.at(page >> 2).Count(page);
 
         if (delta > 0) {
-            ASSERT_MSG(count < UINT16_MAX, "Count may overflow!");
+            ASSERT_MSG(count.load(std::memory_order::relaxed) < UINT16_MAX, "Count may overflow!");
         } else if (delta < 0) {
-            ASSERT_MSG(count > 0, "Count may underflow!");
+            ASSERT_MSG(count.load(std::memory_order::relaxed) > 0, "Count may underflow!");
         } else {
-            ASSERT_MSG(true, "Delta must be non-zero!");
+            ASSERT_MSG(false, "Delta must be non-zero!");
         }
 
         // Adds or subtracts 1, as count is a unsigned 8-bit value
-        count += static_cast<u16>(delta);
+        count.fetch_add(static_cast<u16>(delta), std::memory_order_release);
 
         // Assume delta is either -1 or 1
-        if (count == 0) {
+        if (count.load(std::memory_order::relaxed) == 0) {
-            cpu_memory.RasterizerMarkRegionCached(page << Core::Memory::PAGE_BITS,
+            if (uncache_bytes == 0) {
-                                                  Core::Memory::PAGE_SIZE, false);
+                uncache_begin = page;
-        } else if (count == 1 && delta > 0) {
+            }
-            cpu_memory.RasterizerMarkRegionCached(page << Core::Memory::PAGE_BITS,
+            uncache_bytes += PAGE_SIZE;
-                                                  Core::Memory::PAGE_SIZE, true);
+        } else if (uncache_bytes > 0) {
+            cpu_memory.RasterizerMarkRegionCached(uncache_begin << PAGE_BITS, uncache_bytes, false);
+            uncache_bytes = 0;
         }
+        if (count.load(std::memory_order::relaxed) == 1 && delta > 0) {
+            if (cache_bytes == 0) {
+                cache_begin = page;
+            }
+            cache_bytes += PAGE_SIZE;
+        } else if (cache_bytes > 0) {
+            cpu_memory.RasterizerMarkRegionCached(cache_begin << PAGE_BITS, cache_bytes, true);
+            cache_bytes = 0;
+        }
+    }
+    if (uncache_bytes > 0) {
+        cpu_memory.RasterizerMarkRegionCached(uncache_begin << PAGE_BITS, uncache_bytes, false);
+    }
+    if (cache_bytes > 0) {
+        cpu_memory.RasterizerMarkRegionCached(cache_begin << PAGE_BITS, cache_bytes, true);
     }
 }
 

src/yuzu/configuration/config.cpp

@@ -756,6 +756,8 @@ void Config::ReadCpuValues() {
                       QStringLiteral("cpuopt_unsafe_reduce_fp_error"), true);
     ReadSettingGlobal(Settings::values.cpuopt_unsafe_inaccurate_nan,
                       QStringLiteral("cpuopt_unsafe_inaccurate_nan"), true);
+    ReadSettingGlobal(Settings::values.cpuopt_unsafe_fastmem_check,
+                      QStringLiteral("cpuopt_unsafe_fastmem_check"), true);
 
     if (global) {
         Settings::values.cpuopt_page_tables =
@@ -774,6 +776,8 @@ void Config::ReadCpuValues() {
             ReadSetting(QStringLiteral("cpuopt_misc_ir"), true).toBool();
         Settings::values.cpuopt_reduce_misalign_checks =
             ReadSetting(QStringLiteral("cpuopt_reduce_misalign_checks"), true).toBool();
+        Settings::values.cpuopt_fastmem =
+            ReadSetting(QStringLiteral("cpuopt_fastmem"), true).toBool();
     }
 
     qt_config->endGroup();
@@ -1332,6 +1336,8 @@ void Config::SaveCpuValues() {
                        Settings::values.cpuopt_unsafe_reduce_fp_error, true);
     WriteSettingGlobal(QStringLiteral("cpuopt_unsafe_inaccurate_nan"),
                        Settings::values.cpuopt_unsafe_inaccurate_nan, true);
+    WriteSettingGlobal(QStringLiteral("cpuopt_unsafe_fastmem_check"),
+                       Settings::values.cpuopt_unsafe_fastmem_check, true);
 
     if (global) {
         WriteSetting(QStringLiteral("cpuopt_page_tables"), Settings::values.cpuopt_page_tables,
@@ -1348,6 +1354,7 @@ void Config::SaveCpuValues() {
         WriteSetting(QStringLiteral("cpuopt_misc_ir"), Settings::values.cpuopt_misc_ir, true);
         WriteSetting(QStringLiteral("cpuopt_reduce_misalign_checks"),
                      Settings::values.cpuopt_reduce_misalign_checks, true);
+        WriteSetting(QStringLiteral("cpuopt_fastmem"), Settings::values.cpuopt_fastmem, true);
     }
 
     qt_config->endGroup();

src/yuzu/configuration/configure_cpu.cpp

@@ -35,12 +35,15 @@ void ConfigureCpu::SetConfiguration() {
     ui->cpuopt_unsafe_unfuse_fma->setEnabled(runtime_lock);
     ui->cpuopt_unsafe_reduce_fp_error->setEnabled(runtime_lock);
     ui->cpuopt_unsafe_inaccurate_nan->setEnabled(runtime_lock);
+    ui->cpuopt_unsafe_fastmem_check->setEnabled(runtime_lock);
 
     ui->cpuopt_unsafe_unfuse_fma->setChecked(Settings::values.cpuopt_unsafe_unfuse_fma.GetValue());
     ui->cpuopt_unsafe_reduce_fp_error->setChecked(
         Settings::values.cpuopt_unsafe_reduce_fp_error.GetValue());
     ui->cpuopt_unsafe_inaccurate_nan->setChecked(
         Settings::values.cpuopt_unsafe_inaccurate_nan.GetValue());
+    ui->cpuopt_unsafe_fastmem_check->setChecked(
+        Settings::values.cpuopt_unsafe_fastmem_check.GetValue());
 
     if (Settings::IsConfiguringGlobal()) {
         ui->accuracy->setCurrentIndex(static_cast<int>(Settings::values.cpu_accuracy.GetValue()));
@@ -84,6 +87,9 @@ void ConfigureCpu::ApplyConfiguration() {
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.cpuopt_unsafe_inaccurate_nan,
                                              ui->cpuopt_unsafe_inaccurate_nan,
                                              cpuopt_unsafe_inaccurate_nan);
+    ConfigurationShared::ApplyPerGameSetting(&Settings::values.cpuopt_unsafe_fastmem_check,
+                                             ui->cpuopt_unsafe_fastmem_check,
+                                             cpuopt_unsafe_fastmem_check);
 
     if (Settings::IsConfiguringGlobal()) {
         // Guard if during game and set to game-specific value
@@ -134,4 +140,7 @@ void ConfigureCpu::SetupPerGameUI() {
     ConfigurationShared::SetColoredTristate(ui->cpuopt_unsafe_inaccurate_nan,
                                             Settings::values.cpuopt_unsafe_inaccurate_nan,
                                             cpuopt_unsafe_inaccurate_nan);
+    ConfigurationShared::SetColoredTristate(ui->cpuopt_unsafe_fastmem_check,
+                                            Settings::values.cpuopt_unsafe_fastmem_check,
+                                            cpuopt_unsafe_fastmem_check);
 }

src/yuzu/configuration/configure_cpu.h

@@ -41,4 +41,5 @@ private:
     ConfigurationShared::CheckState cpuopt_unsafe_unfuse_fma;
     ConfigurationShared::CheckState cpuopt_unsafe_reduce_fp_error;
     ConfigurationShared::CheckState cpuopt_unsafe_inaccurate_nan;
+    ConfigurationShared::CheckState cpuopt_unsafe_fastmem_check;
 };

src/yuzu/configuration/configure_cpu.ui

@@ -123,6 +123,18 @@
           </property>
          </widget>
         </item>
+        <item>
+         <widget class="QCheckBox" name="cpuopt_unsafe_fastmem_check">
+          <property name="toolTip">
+           <string>
+            &lt;div&gt;This option improves speed by eliminating a safety check before every memory read/write in guest. Disabling it may allow a game to read/write the emulator's memory.&lt;/div&gt;
+           </string>
+          </property>
+          <property name="text">
+           <string>Disable adress space checks</string>
+          </property>
+         </widget>
+        </item>
        </layout>
       </widget>
      </item>

src/yuzu/configuration/configure_cpu_debug.cpp

@@ -39,6 +39,8 @@ void ConfigureCpuDebug::SetConfiguration() {
     ui->cpuopt_misc_ir->setChecked(Settings::values.cpuopt_misc_ir);
     ui->cpuopt_reduce_misalign_checks->setEnabled(runtime_lock);
     ui->cpuopt_reduce_misalign_checks->setChecked(Settings::values.cpuopt_reduce_misalign_checks);
+    ui->cpuopt_fastmem->setEnabled(runtime_lock);
+    ui->cpuopt_fastmem->setChecked(Settings::values.cpuopt_fastmem);
 }
 
 void ConfigureCpuDebug::ApplyConfiguration() {
@@ -50,6 +52,7 @@ void ConfigureCpuDebug::ApplyConfiguration() {
     Settings::values.cpuopt_const_prop = ui->cpuopt_const_prop->isChecked();
     Settings::values.cpuopt_misc_ir = ui->cpuopt_misc_ir->isChecked();
     Settings::values.cpuopt_reduce_misalign_checks = ui->cpuopt_reduce_misalign_checks->isChecked();
+    Settings::values.cpuopt_fastmem = ui->cpuopt_fastmem->isChecked();
 }
 
 void ConfigureCpuDebug::changeEvent(QEvent* event) {

src/yuzu/configuration/configure_cpu_debug.ui

@@ -139,6 +139,20 @@
           </property>
          </widget>
         </item>
+        <item>
+         <widget class="QCheckBox" name="cpuopt_fastmem">
+          <property name="text">
+           <string>Enable Host MMU Emulation</string>
+          </property>
+          <property name="toolTip">
+           <string>
+            &lt;div style="white-space: nowrap"&gt;This optimization speeds up memory accesses by the guest program.&lt;/div&gt;
+            &lt;div style="white-space: nowrap"&gt;Enabling it causes guest memory reads/writes to be done directly into memory and make use of Host's MMU.&lt;/div&gt;
+            &lt;div style="white-space: nowrap"&gt;Disabling this forces all memory accesses to use Software MMU Emulation.&lt;/div&gt;
+           </string>
+          </property>
+         </widget>
+        </item>
        </layout>
       </widget>
      </item>

src/yuzu_cmd/default_ini.h

@@ -150,6 +150,10 @@ cpuopt_misc_ir =
 # 0: Disabled, 1 (default): Enabled
 cpuopt_reduce_misalign_checks =
 
+# Enable Host MMU Emulation (faster guest memory access)
+# 0: Disabled, 1 (default): Enabled
+cpuopt_fastmem =
+
 [Renderer]
 # Which backend API to use.
 # 0 (default): OpenGL, 1: Vulkan