Combine guest memory and function table into one virtual allocation.

UnleashedRecomp/CMakeLists.txt

@@ -112,7 +112,6 @@ endif()
 
 set(SWA_CPU_CXX_SOURCES
     "cpu/guest_thread.cpp"
-    "cpu/code_cache.cpp"
 )
 
 set(SWA_GPU_CXX_SOURCES

UnleashedRecomp/api/SWA.inl

@@ -1,6 +1,5 @@
 #pragma once
 
-#include <cpu/guest_code.h>
 #include <cpu/guest_stack_var.h>
 #include <kernel/function.h>
 

UnleashedRecomp/apu/audio.cpp

@@ -3,7 +3,7 @@
 #include <bit>
 
 #include "audio.h"
-#include "cpu/code_cache.h"
+#include <kernel/memory.h>
 
 #define AUDIO_DRIVER_KEY (uint32_t)('DAUD')
 
@@ -21,7 +21,7 @@ uint32_t XAudioRegisterRenderDriverClient(be<uint32_t>* callback, be<uint32_t>*
 #endif
 
     *driver = AUDIO_DRIVER_KEY;
-    XAudioRegisterClient(KeFindHostFunction(*callback), callback[1]);
+    XAudioRegisterClient(g_memory.FindFunction(*callback), callback[1]);
     return 0;
 }
 

UnleashedRecomp/apu/driver/sdl2_driver.cpp

@@ -1,7 +1,5 @@
 #include "sdl2_driver.h"
-#include <cpu/code_cache.h>
 #include <cpu/guest_thread.h>
-#include <cpu/guest_code.h>
 #include <kernel/heap.h>
 
 static PPCFunc* g_clientCallback{};
@@ -50,7 +48,7 @@ static void AudioThread()
         if ((queuedAudioSize / callbackAudioSize) <= MAX_LATENCY)
         {
             ctx.ppcContext.r3.u32 = g_clientCallbackParam;
-            g_clientCallback(ctx.ppcContext, reinterpret_cast<uint8_t*>(g_memory.base));
+            g_clientCallback(ctx.ppcContext, g_memory.base);
         }
 
         auto now = std::chrono::steady_clock::now();

UnleashedRecomp/apu/driver/xaudio_driver.cpp

@@ -1,9 +1,7 @@
 #include <stdafx.h>
 #include "xaudio_driver.h"
 #include <xaudio2.h>
-#include <cpu/code_cache.h>
 #include <cpu/guest_thread.h>
-#include <cpu/guest_code.h>
 #include <cpu/ppc_context.h>
 #include <kernel/heap.h>
 
@@ -55,7 +53,7 @@ PPC_FUNC(DriverLoop)
         WaitForSingleObject(g_audioSemaphore, INFINITE);
 
         ctx.r3.u64 = g_clientCallbackParam;
-        GuestCode::Run((void*)g_clientCallback, &ctx);
+        g_clientCallback(ctx, g_memory.base);
     }
 }
 

UnleashedRecomp/cpu/code_cache.cpp

@@ -1,60 +0,0 @@
-#include <stdafx.h>
-#include "code_cache.h"
-#include "ppc_context.h"
-
-CodeCache::CodeCache()
-{
-#ifdef _WIN32
-    bucket = (char*)VirtualAlloc(nullptr, 0x200000000, MEM_RESERVE, PAGE_READWRITE);
-    assert(bucket != nullptr);
-#else
-    bucket = (char*)mmap(NULL, 0x200000000, PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
-    assert(bucket != (char*)MAP_FAILED);
-#endif
-}
-
-CodeCache::~CodeCache()
-{
-#ifdef _WIN32
-    VirtualFree(bucket, 0, MEM_RELEASE);
-#else
-    munmap(bucket, 0x200000000);
-#endif
-}
-
-void CodeCache::Init()
-{
-    for (size_t i = 0; PPCFuncMappings[i].guest != 0; i++)
-    {
-        if (PPCFuncMappings[i].host != nullptr)
-        {
-#ifdef _WIN32
-            VirtualAlloc(bucket + PPCFuncMappings[i].guest * 2, sizeof(void*), MEM_COMMIT, PAGE_READWRITE);
-#endif
-            *(void**)(bucket + PPCFuncMappings[i].guest * 2) = (void*)PPCFuncMappings[i].host;
-        }
-    }
-}
-
-void CodeCache::Insert(uint32_t guest, PPCFunc* host)
-{
-#ifdef _WIN32
-    VirtualAlloc(bucket + static_cast<uint64_t>(guest) * 2, sizeof(void*), MEM_COMMIT, PAGE_READWRITE);
-#endif
-    *reinterpret_cast<PPCFunc**>(bucket + static_cast<uint64_t>(guest) * 2) = host;
-}
-
-void* CodeCache::Find(uint32_t guest) const
-{
-    return *reinterpret_cast<void**>(bucket + static_cast<uint64_t>(guest) * 2);
-}
-
-SWA_API PPCFunc* KeFindHostFunction(uint32_t guest)
-{
-    return reinterpret_cast<PPCFunc*>(g_codeCache.Find(guest));
-}
-
-SWA_API void KeInsertHostFunction(uint32_t guest, PPCFunc* function)
-{
-    g_codeCache.Insert(guest, function);
-}

UnleashedRecomp/cpu/code_cache.h

@@ -1,19 +0,0 @@
-#pragma once
-
-struct CodeCache
-{
-    char* bucket{};
-
-    CodeCache();
-    ~CodeCache();
-
-    void Init();
-    void Insert(uint32_t guest, PPCFunc* host);
-
-    void* Find(uint32_t guest) const;
-};
-
-SWA_API PPCFunc* KeFindHostFunction(uint32_t guest);
-SWA_API void KeInsertHostFunction(uint32_t guest, PPCFunc* function);
-
-extern CodeCache g_codeCache;

UnleashedRecomp/cpu/guest_code.h

@@ -1,23 +0,0 @@
-#pragma once
-#include "ppc_context.h"
-#include <kernel/memory.h>
-
-struct GuestCode 
-{
-    inline static void Run(void* hostAddress, PPCContext* ctx, void* baseAddress)
-    {
-        ctx->fpscr.loadFromHost();
-        reinterpret_cast<PPCFunc*>(hostAddress)(*ctx, reinterpret_cast<uint8_t*>(baseAddress));
-    }
-
-    inline static void Run(void* hostAddress, PPCContext* ctx)
-    {
-        ctx->fpscr.loadFromHost();
-        reinterpret_cast<PPCFunc*>(hostAddress)(*ctx, reinterpret_cast<uint8_t*>(g_memory.base));
-    }
-
-    inline static void Run(void* hostAddress)
-    {
-        Run(hostAddress, GetPPCContext());
-    }
-};

UnleashedRecomp/cpu/guest_thread.cpp

@@ -3,8 +3,6 @@
 #include <kernel/memory.h>
 #include <kernel/heap.h>
 #include <kernel/function.h>
-#include "code_cache.h"
-#include "guest_code.h"
 #include "ppc_context.h"
 
 constexpr size_t PCR_SIZE = 0xAB0;
@@ -29,7 +27,6 @@ GuestThreadContext::GuestThreadContext(uint32_t cpuNumber)
     *(uint32_t*)(thread + PCR_SIZE + 0x10) = 0xFFFFFFFF; // that one TLS entry that felt quirky
     *(uint32_t*)(thread + PCR_SIZE + TLS_SIZE + 0x14C) = ByteSwap(GuestThread::GetCurrentThreadId()); // thread id
 
-    ppcContext.fn = (uint8_t*)g_codeCache.bucket;
     ppcContext.r1.u64 = g_memory.MapVirtual(thread + PCR_SIZE + TLS_SIZE + TEB_SIZE + STACK_SIZE); // stack pointer
     ppcContext.r13.u64 = g_memory.MapVirtual(thread);
     ppcContext.fpscr.loadFromHost();
@@ -78,7 +75,7 @@ uint32_t GuestThread::Start(const GuestThreadParams& params)
     GuestThreadContext ctx(cpuNumber);
     ctx.ppcContext.r3.u64 = params.value;
 
-    reinterpret_cast<PPCFunc*>(g_codeCache.Find(params.function))(ctx.ppcContext, reinterpret_cast<uint8_t*>(g_memory.base));
+    g_memory.FindFunction(params.function)(ctx.ppcContext, g_memory.base);
 
     return ctx.ppcContext.r3.u32;
 }

UnleashedRecomp/gpu/video.cpp

@@ -6,8 +6,6 @@
 
 #include <app.h>
 #include <bc_diff.h>
-#include <cpu/code_cache.h>
-#include <cpu/guest_code.h>
 #include <cpu/guest_thread.h>
 #include <decompressor.h>
 #include <kernel/function.h>
@@ -1682,7 +1680,7 @@ static uint32_t CreateDevice(uint32_t a1, uint32_t a2, uint32_t a3, uint32_t a4,
     memset(device, 0, sizeof(*device));
 
     uint32_t functionOffset = 0x443344; // D3D
-    g_codeCache.Insert(functionOffset, HostToGuestFunction<SetRenderStateUnimplemented>);
+    g_memory.InsertFunction(functionOffset, HostToGuestFunction<SetRenderStateUnimplemented>);
 
     for (size_t i = 0; i < std::size(device->setRenderStateFunctions); i++)
         device->setRenderStateFunctions[i] = functionOffset;
@@ -1690,7 +1688,7 @@ static uint32_t CreateDevice(uint32_t a1, uint32_t a2, uint32_t a3, uint32_t a4,
     for (auto& [state, function] : g_setRenderStateFunctions)
     {
         functionOffset += 4;
-        g_codeCache.Insert(functionOffset, function);
+        g_memory.InsertFunction(functionOffset, function);
         device->setRenderStateFunctions[state / 4] = functionOffset;
     }
 

UnleashedRecomp/kernel/function.h

@@ -314,19 +314,18 @@ T GuestToHostFunction(const TFunction& func, TArgs&&... argv)
     auto& currentCtx = *GetPPCContext();
 
     PPCContext newCtx; // NOTE: No need for zero initialization, has lots of unnecessary code generation.
-    newCtx.fn = currentCtx.fn;
     newCtx.r1 = currentCtx.r1;
     newCtx.r13 = currentCtx.r13;
     newCtx.fpscr = currentCtx.fpscr;
 
-    _translate_args_to_guest(newCtx, (uint8_t*)g_memory.base, args);
+    _translate_args_to_guest(newCtx, g_memory.base, args);
 
     SetPPCContext(newCtx);
 
     if constexpr (std::is_function_v<TFunction>)
-        func(newCtx, (uint8_t*)g_memory.base);
+        func(newCtx, g_memory.base);
     else
-        (*(PPCFunc**)(newCtx.fn + uint64_t(func) * 2))(newCtx, (uint8_t*)g_memory.base);
+        g_memory.FindFunction(func)(newCtx, g_memory.base);
 
     currentCtx.fpscr = newCtx.fpscr;
     SetPPCContext(currentCtx);

UnleashedRecomp/kernel/heap.cpp

@@ -8,10 +8,7 @@ constexpr size_t RESERVED_END = 0xA0000000;
 
 void Heap::Init()
 {
-    g_memory.Alloc(0x20000, RESERVED_BEGIN - 0x20000, MEM_COMMIT);
     heap = o1heapInit(g_memory.Translate(0x20000), RESERVED_BEGIN - 0x20000);
-
-    g_memory.Alloc(RESERVED_END, 0x100000000 - RESERVED_END, MEM_COMMIT);
     physicalHeap = o1heapInit(g_memory.Translate(RESERVED_END), 0x100000000 - RESERVED_END);
 }
 

UnleashedRecomp/kernel/memory.cpp

@@ -1,48 +1,30 @@
 #include <stdafx.h>
 #include "memory.h"
 
-Memory::Memory(void* address, size_t size) : size(size)
+Memory::Memory()
 {
 #ifdef _WIN32
-    base = (char*)VirtualAlloc(address, size, MEM_RESERVE, PAGE_READWRITE);
+    base = (uint8_t*)VirtualAlloc((void*)0x100000000ull, PPC_MEMORY_SIZE + PPC_FUNC_TABLE_SIZE, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE);
 
     if (base == nullptr)
-        base = (char*)VirtualAlloc(nullptr, size, MEM_RESERVE, PAGE_READWRITE);
+        base = (uint8_t*)VirtualAlloc(nullptr, PPC_MEMORY_SIZE + PPC_FUNC_TABLE_SIZE, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE);
-#else
-    base = (char*)mmap(address, size, PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
 
-    if (base == (char*)MAP_FAILED)
+    DWORD oldProtect;
-        base = (char*)mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
+    VirtualProtect(base, 4096, PAGE_NOACCESS, &oldProtect);
+#else
+    base = (uint8_t*)mmap((void*)0x100000000ull, PPC_MEMORY_SIZE + PPC_FUNC_TABLE_SIZE, PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
+
+    if (base == (uint8_t*)MAP_FAILED)
+        base = (uint8_t*)mmap(NULL, PPC_MEMORY_SIZE + PPC_FUNC_TABLE_SIZE, PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
 
     mprotect(base, 4096, PROT_NONE);
 #endif
-}
 
-void* Memory::Alloc(size_t offset, size_t size, uint32_t type)
+    for (size_t i = 0; PPCFuncMappings[i].guest != 0; i++)
-{
+    {
-#ifdef _WIN32
+        if (PPCFuncMappings[i].host != nullptr)
-    return VirtualAlloc(base + offset, size, type, PAGE_READWRITE);
+            InsertFunction(PPCFuncMappings[i].guest, PPCFuncMappings[i].host);
-#else
+    }
-    return base + offset;
-#endif
-}
-
-void* Memory::Commit(size_t offset, size_t size)
-{
-#ifdef _WIN32
-    return Alloc(offset, size, MEM_COMMIT);
-#else
-    return base + offset;
-#endif
-}
-
-void* Memory::Reserve(size_t offset, size_t size)
-{
-#ifdef _WIN32
-    return Alloc(offset, size, MEM_RESERVE);
-#else
-    return base + offset;
-#endif
 }
 
 void* MmGetHostAddress(uint32_t ptr)

UnleashedRecomp/kernel/memory.h

@@ -5,29 +5,21 @@
 #define MEM_RESERVE 0x00002000  
 #endif
 
-class Memory
+struct Memory
 {
-public:
+    uint8_t* base{};
-    char* base{};
-    size_t size{};
-    size_t guestBase{};
 
-    Memory(void* address, size_t size);
+    Memory();
-
-    void* Alloc(size_t offset, size_t size, uint32_t type);
-
-    void* Commit(size_t offset, size_t size);
-    void* Reserve(size_t offset, size_t size);
 
     bool IsInMemoryRange(const void* host) const noexcept
     {
-        return host >= base && host < (base + size);
+        return host >= base && host < (base + PPC_MEMORY_SIZE);
     }
 
     void* Translate(size_t offset) const noexcept
     {
         if (offset)
-            assert(offset < 0x100000000ull);
+            assert(offset < PPC_MEMORY_SIZE);
 
         return base + offset;
     }
@@ -37,7 +29,17 @@ public:
         if (host)
             assert(IsInMemoryRange(host));
 
-        return static_cast<uint32_t>(static_cast<const char*>(host) - base);
+        return static_cast<uint32_t>(static_cast<const uint8_t*>(host) - base);
+    }
+
+    PPCFunc* FindFunction(uint32_t guest) const noexcept
+    {
+        return *reinterpret_cast<PPCFunc**>(base + PPC_FUNC_TABLE_OFFSET + (uint64_t(guest) * 2));
+    }
+
+    void InsertFunction(uint32_t guest, PPCFunc* host)
+    {
+        *reinterpret_cast<PPCFunc**>(base + PPC_FUNC_TABLE_OFFSET + (uint64_t(guest) * 2)) = host;
     }
 };
 

UnleashedRecomp/main.cpp

@@ -1,5 +1,4 @@
 #include <stdafx.h>
-#include <cpu/code_cache.h>
 #include <cpu/guest_thread.h>
 #include <gpu/video.h>
 #include <kernel/function.h>
@@ -25,9 +24,8 @@
 const size_t XMAIOBegin = 0x7FEA0000;
 const size_t XMAIOEnd = XMAIOBegin + 0x0000FFFF;
 
-Memory g_memory{ reinterpret_cast<void*>(0x100000000), 0x100000000 };
+Memory g_memory;
 Heap g_userHeap;
-CodeCache g_codeCache;
 XDBFWrapper g_xdbfWrapper;
 std::unordered_map<uint16_t, GuestTexture*> g_xdbfTextureCache;
 
@@ -37,11 +35,7 @@ void HostStartup()
     CoInitializeEx(nullptr, COINIT_MULTITHREADED);
 #endif
 
-    g_memory.Alloc(0x10000, 0x1000, MEM_COMMIT);
     g_userHeap.Init();
-    g_codeCache.Init();
-
-    g_memory.Alloc(XMAIOBegin, 0xFFFF, MEM_COMMIT);
 
     hid::Init();
 }
@@ -109,8 +103,6 @@ uint32_t LdrLoadModule(const std::filesystem::path &path)
     auto* xex = reinterpret_cast<XEX_HEADER*>(loadResult.data());
     auto security = reinterpret_cast<XEX2_SECURITY_INFO*>((char*)xex + xex->AddressOfSecurityInfo);
 
-    g_memory.Alloc(security->ImageBase, security->SizeOfImage, MEM_COMMIT);
-
     auto format = Xex2FindOptionalHeader<XEX_FILE_FORMAT_INFO>(xex, XEX_HEADER_FILE_FORMAT_INFO);
     auto entry = *Xex2FindOptionalHeader<uint32_t>(xex, XEX_HEADER_ENTRY_POINT);
     ByteSwapInplace(entry);

UnleashedRecomp/patches/audio_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <user/config.h>
 #include <kernel/function.h>
 #include <os/media.h>

UnleashedRecomp/patches/camera_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <api/SWA.h>
 #include <ui/game_window.h>
 #include <user/config.h>

UnleashedRecomp/patches/fps_patches.cpp

@@ -1,5 +1,4 @@
-#include <cpu/code_cache.h>
+#include <kernel/memory.h>
-#include <cpu/guest_code.h>
 #include <api/SWA.h>
 #include <ui/game_window.h>
 #include <user/config.h>
@@ -87,13 +86,13 @@ bool LoadingUpdateMidAsmHook(PPCRegister& r31)
     double deltaTime = std::min(std::chrono::duration<double>(now - g_prev).count(), 1.0 / 15.0);
     g_prev = now;
 
-    uint8_t* base = reinterpret_cast<uint8_t*>(g_memory.base);
+    uint8_t* base = g_memory.base;
     uint32_t application = PPC_LOAD_U32(PPC_LOAD_U32(r31.u32 + 4));
     uint32_t update = PPC_LOAD_U32(PPC_LOAD_U32(application) + 20);
 
     g_ppcContext->r3.u32 = application;
     g_ppcContext->f1.f64 = deltaTime;
-    reinterpret_cast<PPCFunc*>(g_codeCache.Find(update))(*g_ppcContext, base);
+    g_memory.FindFunction(update)(*g_ppcContext, base);
 
     bool loading = PPC_LOAD_U8(0x83367A4C);
     if (loading)

UnleashedRecomp/patches/misc_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <api/SWA.h>
 #include <ui/game_window.h>
 #include <user/achievement_data.h>

UnleashedRecomp/patches/object_patches.cpp

@@ -1,5 +1,3 @@
-#include <cpu/guest_code.h>
-
 // CObjFlame::CObjFlame
 // A field is not zero initialized,
 // causing collisions to constantly get created

UnleashedRecomp/patches/player_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <api/SWA.h>
 #include <ui/game_window.h>
 #include <ui/window_events.h>

UnleashedRecomp/patches/resident_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <user/achievement_data.h>
 #include <user/config.h>
 #include <api/SWA.h>

UnleashedRecomp/patches/ui/CHudPause_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <kernel/function.h>
 #include <api/SWA.h>
 #include <ui/achievement_menu.h>

UnleashedRecomp/patches/ui/CTitleStateIntro_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <api/SWA.h>
 #include <locale/locale.h>
 #include <ui/fader.h>

UnleashedRecomp/patches/ui/CTitleStateMenu_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <api/SWA.h>
 #include <locale/locale.h>
 #include <ui/button_guide.h>

UnleashedRecomp/patches/video_patches.cpp

@@ -1,4 +1,3 @@
-#include <cpu/guest_code.h>
 #include <user/config.h>
 #include <api/SWA.h>
 #include <ui/game_window.h>

tools/PowerRecomp

@@ -1 +1 @@
-Subproject commit 4650dc69fb40d0e5857fe28c40e85eec7c97edf7
+Subproject commit de2840970ffc3161a4cb8743b10ddd4da93bdc9f