Merge branch 'shadps4-emu:main' into hybrid

src/core/address_space.h

@@ -19,8 +19,6 @@ enum class MemoryPermission : u32 {
 };
 DECLARE_ENUM_FLAG_OPERATORS(MemoryPermission)
 
-constexpr VAddr CODE_BASE_OFFSET = 0x100000000ULL;
-
 constexpr VAddr SYSTEM_MANAGED_MIN = 0x00000400000ULL;
 constexpr VAddr SYSTEM_MANAGED_MAX = 0x07FFFFBFFFULL;
 constexpr VAddr SYSTEM_RESERVED_MIN = 0x07FFFFC000ULL;

src/core/cpu_patches.cpp

@@ -464,9 +464,8 @@ static std::pair<bool, u64> TryPatch(u8* code, PatchModule* module) {
 
             if (needs_trampoline && instruction.length < 5) {
                 // Trampoline is needed but instruction is too short to patch.
-                // Return false and length to fall back to the illegal instruction handler,
+                // Return false and length to signal to AOT compilation that this instruction
-                // or to signal to AOT compilation that this instruction should be skipped and
+                // should be skipped and handled at runtime.
-                // handled at runtime.
                 return std::make_pair(false, instruction.length);
             }
 
@@ -512,136 +511,137 @@ static std::pair<bool, u64> TryPatch(u8* code, PatchModule* module) {
 
 #if defined(ARCH_X86_64)
 
+static bool Is4ByteExtrqOrInsertq(void* code_address) {
+    u8* bytes = (u8*)code_address;
+    if (bytes[0] == 0x66 && bytes[1] == 0x0F && bytes[2] == 0x79) {
+        return true; // extrq
+    } else if (bytes[0] == 0xF2 && bytes[1] == 0x0F && bytes[2] == 0x79) {
+        return true; // insertq
+    } else {
+        return false;
+    }
+}
+
 static bool TryExecuteIllegalInstruction(void* ctx, void* code_address) {
-    ZydisDecodedInstruction instruction;
+    // We need to decode the instruction to find out what it is. Normally we'd use a fully fleshed
-    ZydisDecodedOperand operands[ZYDIS_MAX_OPERAND_COUNT];
+    // out decoder like Zydis, however Zydis does a bunch of stuff that impact performance that we
-    const auto status =
+    // don't care about. We can get information about the instruction a lot faster by writing a mini
-        Common::Decoder::Instance()->decodeInstruction(instruction, operands, code_address);
+    // decoder here, since we know it is definitely an extrq or an insertq. If for some reason we
+    // need to interpret more instructions in the future (I don't see why we would), we can revert
+    // to using Zydis.
+    ZydisMnemonic mnemonic;
+    u8* bytes = (u8*)code_address;
+    if (bytes[0] == 0x66) {
+        mnemonic = ZYDIS_MNEMONIC_EXTRQ;
+    } else if (bytes[0] == 0xF2) {
+        mnemonic = ZYDIS_MNEMONIC_INSERTQ;
+    } else {
+        ZydisDecodedInstruction instruction;
+        ZydisDecodedOperand operands[ZYDIS_MAX_OPERAND_COUNT];
+        const auto status =
+            Common::Decoder::Instance()->decodeInstruction(instruction, operands, code_address);
+        LOG_ERROR(Core, "Unhandled illegal instruction at code address {}: {}",
+                  fmt::ptr(code_address),
+                  ZYAN_SUCCESS(status) ? ZydisMnemonicGetString(instruction.mnemonic)
+                                       : "Failed to decode");
+        return false;
+    }
 
-    switch (instruction.mnemonic) {
+    ASSERT(bytes[1] == 0x0F && bytes[2] == 0x79);
+
+    // Note: It's guaranteed that there's no REX prefix in these instructions checked by
+    // Is4ByteExtrqOrInsertq
+    u8 modrm = bytes[3];
+    u8 rm = modrm & 0b111;
+    u8 reg = (modrm >> 3) & 0b111;
+    u8 mod = (modrm >> 6) & 0b11;
+
+    ASSERT(mod == 0b11); // Any instruction we interpret here uses reg/reg addressing only
+
+    int dstIndex = reg;
+    int srcIndex = rm;
+
+    switch (mnemonic) {
     case ZYDIS_MNEMONIC_EXTRQ: {
-        bool immediateForm = operands[1].type == ZYDIS_OPERAND_TYPE_IMMEDIATE &&
+        const auto dst = Common::GetXmmPointer(ctx, dstIndex);
-                             operands[2].type == ZYDIS_OPERAND_TYPE_IMMEDIATE;
+        const auto src = Common::GetXmmPointer(ctx, srcIndex);
-        if (immediateForm) {
+
-            LOG_CRITICAL(Core, "EXTRQ immediate form should have been patched at code address: {}",
+        u64 lowQWordSrc;
-                         fmt::ptr(code_address));
+        memcpy(&lowQWordSrc, src, sizeof(lowQWordSrc));
-            return false;
+
+        u64 lowQWordDst;
+        memcpy(&lowQWordDst, dst, sizeof(lowQWordDst));
+
+        u64 length = lowQWordSrc & 0x3F;
+        u64 mask;
+        if (length == 0) {
+            length = 64; // for the check below
+            mask = 0xFFFF'FFFF'FFFF'FFFF;
         } else {
-            ASSERT_MSG(operands[0].type == ZYDIS_OPERAND_TYPE_REGISTER &&
+            mask = (1ULL << length) - 1;
-                           operands[1].type == ZYDIS_OPERAND_TYPE_REGISTER &&
-                           operands[0].reg.value >= ZYDIS_REGISTER_XMM0 &&
-                           operands[0].reg.value <= ZYDIS_REGISTER_XMM15 &&
-                           operands[1].reg.value >= ZYDIS_REGISTER_XMM0 &&
-                           operands[1].reg.value <= ZYDIS_REGISTER_XMM15,
-                       "Unexpected operand types for EXTRQ instruction");
-
-            const auto dstIndex = operands[0].reg.value - ZYDIS_REGISTER_XMM0;
-            const auto srcIndex = operands[1].reg.value - ZYDIS_REGISTER_XMM0;
-
-            const auto dst = Common::GetXmmPointer(ctx, dstIndex);
-            const auto src = Common::GetXmmPointer(ctx, srcIndex);
-
-            u64 lowQWordSrc;
-            memcpy(&lowQWordSrc, src, sizeof(lowQWordSrc));
-
-            u64 lowQWordDst;
-            memcpy(&lowQWordDst, dst, sizeof(lowQWordDst));
-
-            u64 length = lowQWordSrc & 0x3F;
-            u64 mask;
-            if (length == 0) {
-                length = 64; // for the check below
-                mask = 0xFFFF'FFFF'FFFF'FFFF;
-            } else {
-                mask = (1ULL << length) - 1;
-            }
-
-            u64 index = (lowQWordSrc >> 8) & 0x3F;
-            if (length + index > 64) {
-                // Undefined behavior if length + index is bigger than 64 according to the spec,
-                // we'll warn and continue execution.
-                LOG_TRACE(Core,
-                          "extrq at {} with length {} and index {} is bigger than 64, "
-                          "undefined behavior",
-                          fmt::ptr(code_address), length, index);
-            }
-
-            lowQWordDst >>= index;
-            lowQWordDst &= mask;
-
-            memcpy(dst, &lowQWordDst, sizeof(lowQWordDst));
-
-            Common::IncrementRip(ctx, instruction.length);
-
-            return true;
         }
-        break;
+
+        u64 index = (lowQWordSrc >> 8) & 0x3F;
+        if (length + index > 64) {
+            // Undefined behavior if length + index is bigger than 64 according to the spec,
+            // we'll warn and continue execution.
+            LOG_TRACE(Core,
+                      "extrq at {} with length {} and index {} is bigger than 64, "
+                      "undefined behavior",
+                      fmt::ptr(code_address), length, index);
+        }
+
+        lowQWordDst >>= index;
+        lowQWordDst &= mask;
+
+        memcpy(dst, &lowQWordDst, sizeof(lowQWordDst));
+
+        Common::IncrementRip(ctx, 4);
+
+        return true;
     }
     case ZYDIS_MNEMONIC_INSERTQ: {
-        bool immediateForm = operands[2].type == ZYDIS_OPERAND_TYPE_IMMEDIATE &&
+        const auto dst = Common::GetXmmPointer(ctx, dstIndex);
-                             operands[3].type == ZYDIS_OPERAND_TYPE_IMMEDIATE;
+        const auto src = Common::GetXmmPointer(ctx, srcIndex);
-        if (immediateForm) {
+
-            LOG_CRITICAL(Core,
+        u64 lowQWordSrc, highQWordSrc;
-                         "INSERTQ immediate form should have been patched at code address: {}",
+        memcpy(&lowQWordSrc, src, sizeof(lowQWordSrc));
-                         fmt::ptr(code_address));
+        memcpy(&highQWordSrc, (u8*)src + 8, sizeof(highQWordSrc));
-            return false;
+
+        u64 lowQWordDst;
+        memcpy(&lowQWordDst, dst, sizeof(lowQWordDst));
+
+        u64 length = highQWordSrc & 0x3F;
+        u64 mask;
+        if (length == 0) {
+            length = 64; // for the check below
+            mask = 0xFFFF'FFFF'FFFF'FFFF;
         } else {
-            ASSERT_MSG(operands[2].type == ZYDIS_OPERAND_TYPE_UNUSED &&
+            mask = (1ULL << length) - 1;
-                           operands[3].type == ZYDIS_OPERAND_TYPE_UNUSED,
-                       "operands 2 and 3 must be unused for register form.");
-
-            ASSERT_MSG(operands[0].type == ZYDIS_OPERAND_TYPE_REGISTER &&
-                           operands[1].type == ZYDIS_OPERAND_TYPE_REGISTER,
-                       "operands 0 and 1 must be registers.");
-
-            const auto dstIndex = operands[0].reg.value - ZYDIS_REGISTER_XMM0;
-            const auto srcIndex = operands[1].reg.value - ZYDIS_REGISTER_XMM0;
-
-            const auto dst = Common::GetXmmPointer(ctx, dstIndex);
-            const auto src = Common::GetXmmPointer(ctx, srcIndex);
-
-            u64 lowQWordSrc, highQWordSrc;
-            memcpy(&lowQWordSrc, src, sizeof(lowQWordSrc));
-            memcpy(&highQWordSrc, (u8*)src + 8, sizeof(highQWordSrc));
-
-            u64 lowQWordDst;
-            memcpy(&lowQWordDst, dst, sizeof(lowQWordDst));
-
-            u64 length = highQWordSrc & 0x3F;
-            u64 mask;
-            if (length == 0) {
-                length = 64; // for the check below
-                mask = 0xFFFF'FFFF'FFFF'FFFF;
-            } else {
-                mask = (1ULL << length) - 1;
-            }
-
-            u64 index = (highQWordSrc >> 8) & 0x3F;
-            if (length + index > 64) {
-                // Undefined behavior if length + index is bigger than 64 according to the spec,
-                // we'll warn and continue execution.
-                LOG_TRACE(Core,
-                          "insertq at {} with length {} and index {} is bigger than 64, "
-                          "undefined behavior",
-                          fmt::ptr(code_address), length, index);
-            }
-
-            lowQWordSrc &= mask;
-            lowQWordDst &= ~(mask << index);
-            lowQWordDst |= lowQWordSrc << index;
-
-            memcpy(dst, &lowQWordDst, sizeof(lowQWordDst));
-
-            Common::IncrementRip(ctx, instruction.length);
-
-            return true;
         }
-        break;
+
+        u64 index = (highQWordSrc >> 8) & 0x3F;
+        if (length + index > 64) {
+            // Undefined behavior if length + index is bigger than 64 according to the spec,
+            // we'll warn and continue execution.
+            LOG_TRACE(Core,
+                      "insertq at {} with length {} and index {} is bigger than 64, "
+                      "undefined behavior",
+                      fmt::ptr(code_address), length, index);
+        }
+
+        lowQWordSrc &= mask;
+        lowQWordDst &= ~(mask << index);
+        lowQWordDst |= lowQWordSrc << index;
+
+        memcpy(dst, &lowQWordDst, sizeof(lowQWordDst));
+
+        Common::IncrementRip(ctx, 4);
+
+        return true;
     }
     default: {
-        LOG_ERROR(Core, "Unhandled illegal instruction at code address {}: {}",
+        UNREACHABLE();
-                  fmt::ptr(code_address), ZydisMnemonicGetString(instruction.mnemonic));
-        return false;
     }
     }
 
@@ -695,9 +695,22 @@ static bool PatchesAccessViolationHandler(void* context, void* /* fault_address
 
 static bool PatchesIllegalInstructionHandler(void* context) {
     void* code_address = Common::GetRip(context);
-    if (!TryPatchJit(code_address)) {
+    if (Is4ByteExtrqOrInsertq(code_address)) {
+        // The instruction is not big enough for a relative jump, don't try to patch it and pass it
+        // to our illegal instruction interpreter directly
         return TryExecuteIllegalInstruction(context, code_address);
+    } else {
+        if (!TryPatchJit(code_address)) {
+            ZydisDecodedInstruction instruction;
+            ZydisDecodedOperand operands[ZYDIS_MAX_OPERAND_COUNT];
+            const auto status =
+                Common::Decoder::Instance()->decodeInstruction(instruction, operands, code_address);
+            LOG_ERROR(Core, "Failed to patch address {:x} -- mnemonic: {}", (u64)code_address,
+                      ZYAN_SUCCESS(status) ? ZydisMnemonicGetString(instruction.mnemonic)
+                                           : "Failed to decode");
+        }
     }
+
     return true;
 }
 

src/core/libraries/kernel/file_system.cpp

@@ -181,10 +181,6 @@ s32 PS4_SYSV_ABI open(const char* raw_path, s32 flags, u16 mode) {
             return -1;
         }
     } else {
-        // Start by opening as read-write so we can truncate regardless of flags.
-        // Since open starts by closing the file, this won't interfere with later open calls.
-        e = file->f.Open(file->m_host_name, Common::FS::FileAccessMode::ReadWrite);
-
         file->type = Core::FileSys::FileType::Regular;
 
         if (truncate && read_only) {
@@ -192,9 +188,14 @@ s32 PS4_SYSV_ABI open(const char* raw_path, s32 flags, u16 mode) {
             h->DeleteHandle(handle);
             *__Error() = POSIX_EROFS;
             return -1;
-        } else if (truncate && e == 0) {
+        } else if (truncate) {
-            // If the file was opened successfully and truncate was enabled, reduce size to 0
+            // Open the file as read-write so we can truncate regardless of flags.
-            file->f.SetSize(0);
+            // Since open starts by closing the file, this won't interfere with later open calls.
+            e = file->f.Open(file->m_host_name, Common::FS::FileAccessMode::ReadWrite);
+            if (e == 0) {
+                // If the file was opened successfully, reduce size to 0
+                file->f.SetSize(0);
+            }
         }
 
         if (read) {

src/core/libraries/kernel/threads/pthread.cpp

@@ -289,7 +289,12 @@ int PS4_SYSV_ABI posix_pthread_create_name_np(PthreadT* thread, const PthreadAtt
     /* Create thread */
     new_thread->native_thr = Core::NativeThread();
     int ret = new_thread->native_thr.Create(RunThread, new_thread, &new_thread->attr);
+
     ASSERT_MSG(ret == 0, "Failed to create thread with error {}", ret);
+
+    if (attr != nullptr && *attr != nullptr && (*attr)->cpuset != nullptr) {
+        new_thread->SetAffinity((*attr)->cpuset);
+    }
     if (ret) {
         *thread = nullptr;
     }
@@ -521,6 +526,69 @@ int PS4_SYSV_ABI posix_pthread_setcancelstate(PthreadCancelState state,
     return 0;
 }
 
+int Pthread::SetAffinity(const Cpuset* cpuset) {
+    const auto processor_count = std::thread::hardware_concurrency();
+    if (processor_count < 8) {
+        return 0;
+    }
+    if (cpuset == nullptr) {
+        return POSIX_EINVAL;
+    }
+
+    u64 mask = cpuset->bits;
+
+    uintptr_t handle = native_thr.GetHandle();
+    if (handle == 0) {
+        return POSIX_ESRCH;
+    }
+
+    // We don't use this currently because some games gets performance problems
+    // when applying affinity even on strong hardware
+    /*
+    #ifdef _WIN64
+        DWORD_PTR affinity_mask = static_cast<DWORD_PTR>(mask);
+        if (!SetThreadAffinityMask(reinterpret_cast<HANDLE>(handle), affinity_mask)) {
+            return POSIX_EINVAL;
+        }
+
+    #elif defined(__linux__)
+        cpu_set_t cpu_set;
+        CPU_ZERO(&cpu_set);
+
+        u64 mask = cpuset->bits;
+        for (int cpu = 0; cpu < std::min(64, CPU_SETSIZE); ++cpu) {
+            if (mask & (1ULL << cpu)) {
+                CPU_SET(cpu, &cpu_set);
+            }
+        }
+
+        int result =
+            pthread_setaffinity_np(static_cast<pthread_t>(handle), sizeof(cpu_set_t), &cpu_set);
+        if (result != 0) {
+            return POSIX_EINVAL;
+        }
+    #endif
+    */
+    return 0;
+}
+
+int PS4_SYSV_ABI posix_pthread_setaffinity_np(PthreadT thread, size_t cpusetsize,
+                                              const Cpuset* cpusetp) {
+    if (thread == nullptr || cpusetp == nullptr) {
+        return POSIX_EINVAL;
+    }
+    thread->attr.cpusetsize = cpusetsize;
+    return thread->SetAffinity(cpusetp);
+}
+
+int PS4_SYSV_ABI scePthreadSetaffinity(PthreadT thread, const Cpuset mask) {
+    int result = posix_pthread_setaffinity_np(thread, 0x10, &mask);
+    if (result != 0) {
+        return ErrnoToSceKernelError(result);
+    }
+    return 0;
+}
+
 void RegisterThread(Core::Loader::SymbolsResolver* sym) {
     // Posix
     LIB_FUNCTION("Z4QosVuAsA0", "libScePosix", 1, "libkernel", 1, 1, posix_pthread_once);
@@ -544,6 +612,7 @@ void RegisterThread(Core::Loader::SymbolsResolver* sym) {
     LIB_FUNCTION("Z4QosVuAsA0", "libkernel", 1, "libkernel", 1, 1, posix_pthread_once);
     LIB_FUNCTION("EotR8a3ASf4", "libkernel", 1, "libkernel", 1, 1, posix_pthread_self);
     LIB_FUNCTION("OxhIB8LB-PQ", "libkernel", 1, "libkernel", 1, 1, posix_pthread_create);
+    LIB_FUNCTION("5KWrg7-ZqvE", "libkernel", 1, "libkernel", 1, 1, posix_pthread_setaffinity_np);
 
     // Orbis
     LIB_FUNCTION("14bOACANTBo", "libkernel", 1, "libkernel", 1, 1, ORBIS(posix_pthread_once));
@@ -566,6 +635,7 @@ void RegisterThread(Core::Loader::SymbolsResolver* sym) {
     LIB_FUNCTION("W0Hpm2X0uPE", "libkernel", 1, "libkernel", 1, 1, ORBIS(posix_pthread_setprio));
     LIB_FUNCTION("rNhWz+lvOMU", "libkernel", 1, "libkernel", 1, 1, _sceKernelSetThreadDtors);
     LIB_FUNCTION("6XG4B33N09g", "libkernel", 1, "libkernel", 1, 1, sched_yield);
+    LIB_FUNCTION("bt3CTBKmGyI", "libkernel", 1, "libkernel", 1, 1, scePthreadSetaffinity)
 }
 
 } // namespace Libraries::Kernel

src/core/libraries/kernel/threads/pthread.h

@@ -332,6 +332,8 @@ struct Pthread {
             return true;
         }
     }
+
+    int SetAffinity(const Cpuset* cpuset);
 };
 using PthreadT = Pthread*;
 

src/core/libraries/save_data/dialog/savedatadialog_ui.cpp

@@ -49,13 +49,11 @@ void SaveDialogResult::CopyTo(OrbisSaveDataDialogResult& result) const {
     result.mode = this->mode;
     result.result = this->result;
     result.buttonId = this->button_id;
-    if (mode == SaveDataDialogMode::LIST || ElfInfo::Instance().FirmwareVer() >= ElfInfo::FW_45) {
+    if (result.dirName != nullptr) {
-        if (result.dirName != nullptr) {
+        result.dirName->data.FromString(this->dir_name);
-            result.dirName->data.FromString(this->dir_name);
+    }
-        }
+    if (result.param != nullptr && this->param.GetString(SaveParams::MAINTITLE).has_value()) {
-        if (result.param != nullptr && this->param.GetString(SaveParams::MAINTITLE).has_value()) {
+        result.param->FromSFO(this->param);
-            result.param->FromSFO(this->param);
-        }
     }
     result.userData = this->user_data;
 }
@@ -345,12 +343,15 @@ SaveDialogUi::SaveDialogUi(SaveDialogUi&& other) noexcept
     }
 }
 
-SaveDialogUi& SaveDialogUi::operator=(SaveDialogUi other) {
+SaveDialogUi& SaveDialogUi::operator=(SaveDialogUi&& other) noexcept {
     std::scoped_lock lock(draw_mutex, other.draw_mutex);
     using std::swap;
-    swap(state, other.state);
+    state = other.state;
-    swap(status, other.status);
+    other.state = nullptr;
-    swap(result, other.result);
+    status = other.status;
+    other.status = nullptr;
+    result = other.result;
+    other.result = nullptr;
     if (status && *status == Status::RUNNING) {
         first_render = true;
         AddLayer(this);

src/core/libraries/save_data/dialog/savedatadialog_ui.h

@@ -300,7 +300,8 @@ public:
     ~SaveDialogUi() override;
     SaveDialogUi(const SaveDialogUi& other) = delete;
     SaveDialogUi(SaveDialogUi&& other) noexcept;
-    SaveDialogUi& operator=(SaveDialogUi other);
+    SaveDialogUi& operator=(SaveDialogUi& other) = delete;
+    SaveDialogUi& operator=(SaveDialogUi&& other) noexcept;
 
     void Finish(ButtonId buttonId, CommonDialog::Result r = CommonDialog::Result::OK);
 

src/core/module.cpp

@@ -19,8 +19,7 @@ namespace Core {
 
 using EntryFunc = PS4_SYSV_ABI int (*)(size_t args, const void* argp, void* param);
 
-static u64 LoadOffset = CODE_BASE_OFFSET;
+static constexpr u64 ModuleLoadBase = 0x800000000;
-static constexpr u64 CODE_BASE_INCR = 0x010000000u;
 
 static u64 GetAlignedSize(const elf_program_header& phdr) {
     return (phdr.p_align != 0 ? (phdr.p_memsz + (phdr.p_align - 1)) & ~(phdr.p_align - 1)
@@ -84,7 +83,7 @@ static std::string StringToNid(std::string_view symbol) {
 }
 
 Module::Module(Core::MemoryManager* memory_, const std::filesystem::path& file_, u32& max_tls_index)
-    : memory{memory_}, file{file_}, name{file.stem().string()} {
+    : memory{memory_}, file{file_}, name{file.filename().string()} {
     elf.Open(file);
     if (elf.IsElfFile()) {
         LoadModuleToMemory(max_tls_index);
@@ -113,10 +112,8 @@ void Module::LoadModuleToMemory(u32& max_tls_index) {
 
     // Map module segments (and possible TLS trampolines)
     void** out_addr = reinterpret_cast<void**>(&base_virtual_addr);
-    memory->MapMemory(out_addr, memory->SystemReservedVirtualBase() + LoadOffset,
+    memory->MapMemory(out_addr, ModuleLoadBase, aligned_base_size + TrampolineSize,
-                      aligned_base_size + TrampolineSize, MemoryProt::CpuReadWrite,
+                      MemoryProt::CpuReadWrite, MemoryMapFlags::NoFlags, VMAType::Code, name, true);
-                      MemoryMapFlags::Fixed, VMAType::Code, name, true);
-    LoadOffset += CODE_BASE_INCR * (1 + aligned_base_size / CODE_BASE_INCR);
     LOG_INFO(Core_Linker, "Loading module {} to {}", name, fmt::ptr(*out_addr));
 
 #ifdef ARCH_X86_64
@@ -229,7 +226,7 @@ void Module::LoadModuleToMemory(u32& max_tls_index) {
     LOG_INFO(Core_Linker, "program entry addr ..........: {:#018x}", entry_addr);
 
     if (MemoryPatcher::g_eboot_address == 0) {
-        if (name == "eboot") {
+        if (name == "eboot.bin") {
             MemoryPatcher::g_eboot_address = base_virtual_addr;
             MemoryPatcher::g_eboot_image_size = base_size;
             MemoryPatcher::OnGameLoaded();

src/core/tls.h

@@ -5,6 +5,8 @@
 
 #include "common/types.h"
 
+void* memset(void* ptr, int value, size_t num);
+
 namespace Xbyak {
 class CodeGenerator;
 }
@@ -41,9 +43,18 @@ Tcb* GetTcbBase();
 /// Makes sure TLS is initialized for the thread before entering guest.
 void EnsureThreadInitialized();
 
+template <size_t size>
+__attribute__((optnone)) void ClearStack() {
+    volatile void* buf = alloca(size);
+    memset(const_cast<void*>(buf), 0, size);
+    buf = nullptr;
+}
+
 template <class ReturnType, class... FuncArgs, class... CallArgs>
 ReturnType ExecuteGuest(PS4_SYSV_ABI ReturnType (*func)(FuncArgs...), CallArgs&&... args) {
     EnsureThreadInitialized();
+    // clear stack to avoid trash from EnsureThreadInitialized
+    ClearStack<13_KB>();
     return func(std::forward<CallArgs>(args)...);
 }
 

src/emulator.cpp

@@ -10,7 +10,6 @@
 #include "common/logging/log.h"
 #ifdef ENABLE_QT_GUI
 #include <QtCore>
-#include "common/memory_patcher.h"
 #endif
 #include "common/assert.h"
 #ifdef ENABLE_DISCORD_RPC
@@ -20,6 +19,7 @@
 #include <WinSock2.h>
 #endif
 #include "common/elf_info.h"
+#include "common/memory_patcher.h"
 #include "common/ntapi.h"
 #include "common/path_util.h"
 #include "common/polyfill_thread.h"
@@ -54,27 +54,6 @@ Emulator::Emulator() {
     WSADATA wsaData;
     WSAStartup(versionWanted, &wsaData);
 #endif
-
-    // Create stdin/stdout/stderr
-    Common::Singleton<FileSys::HandleTable>::Instance()->CreateStdHandles();
-
-    // Defer until after logging is initialized.
-    memory = Core::Memory::Instance();
-    controller = Common::Singleton<Input::GameController>::Instance();
-    linker = Common::Singleton<Core::Linker>::Instance();
-
-    // Load renderdoc module.
-    VideoCore::LoadRenderDoc();
-
-    // Start the timer (Play Time)
-#ifdef ENABLE_QT_GUI
-    start_time = std::chrono::steady_clock::now();
-    const auto user_dir = Common::FS::GetUserPath(Common::FS::PathType::UserDir);
-    QString filePath = QString::fromStdString((user_dir / "play_time.txt").string());
-    QFile file(filePath);
-    ASSERT_MSG(file.open(QIODevice::ReadWrite | QIODevice::Text),
-               "Error opening or creating play_time.txt");
-#endif
 }
 
 Emulator::~Emulator() {
@@ -102,54 +81,89 @@ void Emulator::Run(const std::filesystem::path& file, const std::vector<std::str
     // Certain games may use /hostapp as well such as CUSA001100
     mnt->Mount(game_folder, "/hostapp", true);
 
-    auto& game_info = Common::ElfInfo::Instance();
+    const auto param_sfo_path = mnt->GetHostPath("/app0/sce_sys/param.sfo");
+    const auto param_sfo_exists = std::filesystem::exists(param_sfo_path);
 
-    // Loading param.sfo file if exists
+    // Load param.sfo details if it exists
     std::string id;
     std::string title;
     std::string app_version;
     u32 fw_version;
     Common::PSFAttributes psf_attributes{};
-
+    if (param_sfo_exists) {
-    const auto param_sfo_path = mnt->GetHostPath("/app0/sce_sys/param.sfo");
-    if (!std::filesystem::exists(param_sfo_path) || !Config::getSeparateLogFilesEnabled()) {
-        Common::Log::Initialize();
-        Common::Log::Start();
-    }
-
-    if (std::filesystem::exists(param_sfo_path)) {
         auto* param_sfo = Common::Singleton<PSF>::Instance();
-        const bool success = param_sfo->Open(param_sfo_path);
+        ASSERT_MSG(param_sfo->Open(param_sfo_path), "Failed to open param.sfo");
-        ASSERT_MSG(success, "Failed to open param.sfo");
+
         const auto content_id = param_sfo->GetString("CONTENT_ID");
         ASSERT_MSG(content_id.has_value(), "Failed to get CONTENT_ID");
+
         id = std::string(*content_id, 7, 9);
-
+        title = param_sfo->GetString("TITLE").value_or("Unknown title");
-        if (Config::getSeparateLogFilesEnabled()) {
+        fw_version = param_sfo->GetInteger("SYSTEM_VER").value_or(0x4700000);
-            Common::Log::Initialize(id + ".log");
+        app_version = param_sfo->GetString("APP_VER").value_or("Unknown version");
-            Common::Log::Start();
+        if (const auto raw_attributes = param_sfo->GetInteger("ATTRIBUTE")) {
+            psf_attributes.raw = *raw_attributes;
         }
-        LOG_INFO(Loader, "Starting shadps4 emulator v{} ", Common::g_version);
+    }
-        LOG_INFO(Loader, "Revision {}", Common::g_scm_rev);
-        LOG_INFO(Loader, "Branch {}", Common::g_scm_branch);
-        LOG_INFO(Loader, "Description {}", Common::g_scm_desc);
-        LOG_INFO(Loader, "Remote {}", Common::g_scm_remote_url);
 
-        LOG_INFO(Config, "General LogType: {}", Config::getLogType());
+    // Initialize logging as soon as possible
-        LOG_INFO(Config, "General isNeo: {}", Config::isNeoModeConsole());
+    if (!id.empty() && Config::getSeparateLogFilesEnabled()) {
-        LOG_INFO(Config, "GPU isNullGpu: {}", Config::nullGpu());
+        Common::Log::Initialize(id + ".log");
-        LOG_INFO(Config, "GPU shouldDumpShaders: {}", Config::dumpShaders());
+    } else {
-        LOG_INFO(Config, "GPU vblankDivider: {}", Config::vblankDiv());
+        Common::Log::Initialize();
-        LOG_INFO(Config, "Vulkan gpuId: {}", Config::getGpuId());
+    }
-        LOG_INFO(Config, "Vulkan vkValidation: {}", Config::vkValidationEnabled());
+    Common::Log::Start();
-        LOG_INFO(Config, "Vulkan vkValidationSync: {}", Config::vkValidationSyncEnabled());
-        LOG_INFO(Config, "Vulkan vkValidationGpu: {}", Config::vkValidationGpuEnabled());
-        LOG_INFO(Config, "Vulkan crashDiagnostics: {}", Config::getVkCrashDiagnosticEnabled());
-        LOG_INFO(Config, "Vulkan hostMarkers: {}", Config::getVkHostMarkersEnabled());
-        LOG_INFO(Config, "Vulkan guestMarkers: {}", Config::getVkGuestMarkersEnabled());
-        LOG_INFO(Config, "Vulkan rdocEnable: {}", Config::isRdocEnabled());
 
+    LOG_INFO(Loader, "Starting shadps4 emulator v{} ", Common::g_version);
+    LOG_INFO(Loader, "Revision {}", Common::g_scm_rev);
+    LOG_INFO(Loader, "Branch {}", Common::g_scm_branch);
+    LOG_INFO(Loader, "Description {}", Common::g_scm_desc);
+    LOG_INFO(Loader, "Remote {}", Common::g_scm_remote_url);
+
+    LOG_INFO(Config, "General LogType: {}", Config::getLogType());
+    LOG_INFO(Config, "General isNeo: {}", Config::isNeoModeConsole());
+    LOG_INFO(Config, "GPU isNullGpu: {}", Config::nullGpu());
+    LOG_INFO(Config, "GPU shouldDumpShaders: {}", Config::dumpShaders());
+    LOG_INFO(Config, "GPU vblankDivider: {}", Config::vblankDiv());
+    LOG_INFO(Config, "Vulkan gpuId: {}", Config::getGpuId());
+    LOG_INFO(Config, "Vulkan vkValidation: {}", Config::vkValidationEnabled());
+    LOG_INFO(Config, "Vulkan vkValidationSync: {}", Config::vkValidationSyncEnabled());
+    LOG_INFO(Config, "Vulkan vkValidationGpu: {}", Config::vkValidationGpuEnabled());
+    LOG_INFO(Config, "Vulkan crashDiagnostics: {}", Config::getVkCrashDiagnosticEnabled());
+    LOG_INFO(Config, "Vulkan hostMarkers: {}", Config::getVkHostMarkersEnabled());
+    LOG_INFO(Config, "Vulkan guestMarkers: {}", Config::getVkGuestMarkersEnabled());
+    LOG_INFO(Config, "Vulkan rdocEnable: {}", Config::isRdocEnabled());
+
+    if (param_sfo_exists) {
+        LOG_INFO(Loader, "Game id: {} Title: {}", id, title);
+        LOG_INFO(Loader, "Fw: {:#x} App Version: {}", fw_version, app_version);
+    }
+    if (!args.empty()) {
+        const auto argc = std::min<size_t>(args.size(), 32);
+        for (auto i = 0; i < argc; i++) {
+            LOG_INFO(Loader, "Game argument {}: {}", i, args[i]);
+        }
+        if (args.size() > 32) {
+            LOG_ERROR(Loader, "Too many game arguments, only passing the first 32");
+        }
+    }
+
+    // Create stdin/stdout/stderr
+    Common::Singleton<FileSys::HandleTable>::Instance()->CreateStdHandles();
+
+    // Initialize components
+    memory = Core::Memory::Instance();
+    controller = Common::Singleton<Input::GameController>::Instance();
+    linker = Common::Singleton<Core::Linker>::Instance();
+
+    // Load renderdoc module
+    VideoCore::LoadRenderDoc();
+
+    // Initialize patcher and trophies
+    if (!id.empty()) {
+        MemoryPatcher::g_game_serial = id;
         Libraries::NpTrophy::game_serial = id;
+
         const auto trophyDir =
             Common::FS::GetUserPath(Common::FS::PathType::MetaDataDir) / id / "TrophyFiles";
         if (!std::filesystem::exists(trophyDir)) {
@@ -158,41 +172,9 @@ void Emulator::Run(const std::filesystem::path& file, const std::vector<std::str
                 LOG_ERROR(Loader, "Couldn't extract trophies");
             }
         }
-#ifdef ENABLE_QT_GUI
-        MemoryPatcher::g_game_serial = id;
-
-        // Timer for 'Play Time'
-        QTimer* timer = new QTimer();
-        QObject::connect(timer, &QTimer::timeout, [this, id]() {
-            UpdatePlayTime(id);
-            start_time = std::chrono::steady_clock::now();
-        });
-        timer->start(60000); // 60000 ms = 1 minute
-#endif
-        title = param_sfo->GetString("TITLE").value_or("Unknown title");
-        LOG_INFO(Loader, "Game id: {} Title: {}", id, title);
-        fw_version = param_sfo->GetInteger("SYSTEM_VER").value_or(0x4700000);
-        app_version = param_sfo->GetString("APP_VER").value_or("Unknown version");
-        LOG_INFO(Loader, "Fw: {:#x} App Version: {}", fw_version, app_version);
-        if (const auto raw_attributes = param_sfo->GetInteger("ATTRIBUTE")) {
-            psf_attributes.raw = *raw_attributes;
-        }
-        if (!args.empty()) {
-            int argc = std::min<int>(args.size(), 32);
-            for (int i = 0; i < argc; i++) {
-                LOG_INFO(Loader, "Game argument {}: {}", i, args[i]);
-            }
-            if (args.size() > 32) {
-                LOG_ERROR(Loader, "Too many game arguments, only passing the first 32");
-            }
-        }
-    }
-
-    const auto pic1_path = mnt->GetHostPath("/app0/sce_sys/pic1.png");
-    if (std::filesystem::exists(pic1_path)) {
-        game_info.splash_path = pic1_path;
     }
 
+    auto& game_info = Common::ElfInfo::Instance();
     game_info.initialized = true;
     game_info.game_serial = id;
     game_info.title = title;
@@ -201,6 +183,11 @@ void Emulator::Run(const std::filesystem::path& file, const std::vector<std::str
     game_info.raw_firmware_ver = fw_version;
     game_info.psf_attributes = psf_attributes;
 
+    const auto pic1_path = mnt->GetHostPath("/app0/sce_sys/pic1.png");
+    if (std::filesystem::exists(pic1_path)) {
+        game_info.splash_path = pic1_path;
+    }
+
     std::string game_title = fmt::format("{} - {} <{}>", id, title, app_version);
     std::string window_title = "";
     if (Common::g_is_release) {
@@ -284,6 +271,25 @@ void Emulator::Run(const std::filesystem::path& file, const std::vector<std::str
     }
 #endif
 
+    // Start the timer (Play Time)
+#ifdef ENABLE_QT_GUI
+    if (!id.empty()) {
+        auto* timer = new QTimer();
+        QObject::connect(timer, &QTimer::timeout, [this, id]() {
+            UpdatePlayTime(id);
+            start_time = std::chrono::steady_clock::now();
+        });
+        timer->start(60000); // 60000 ms = 1 minute
+
+        start_time = std::chrono::steady_clock::now();
+        const auto user_dir = Common::FS::GetUserPath(Common::FS::PathType::UserDir);
+        QString filePath = QString::fromStdString((user_dir / "play_time.txt").string());
+        QFile file(filePath);
+        ASSERT_MSG(file.open(QIODevice::ReadWrite | QIODevice::Text),
+                   "Error opening or creating play_time.txt");
+    }
+#endif
+
     linker->Execute(args);
 
     window->InitTimers();

src/shader_recompiler/runtime_info.h

@@ -169,10 +169,10 @@ static constexpr u32 MaxColorBuffers = 8;
 
 struct PsColorBuffer {
     AmdGpu::NumberFormat num_format : 4;
-    AmdGpu::NumberConversion num_conversion : 2;
+    AmdGpu::NumberConversion num_conversion : 3;
     AmdGpu::Liverpool::ShaderExportFormat export_format : 4;
     u32 needs_unorm_fixup : 1;
-    u32 pad : 21;
+    u32 pad : 20;
     AmdGpu::CompMapping swizzle;
 
     auto operator<=>(const PsColorBuffer&) const noexcept = default;

src/video_core/buffer_cache/buffer_cache.cpp

@@ -269,8 +269,8 @@ void BufferCache::BindVertexBuffers(const Vulkan::GraphicsPipeline& pipeline) {
     if (instance.IsVertexInputDynamicState()) {
         cmdbuf.bindVertexBuffers(0, num_buffers, host_buffers.data(), host_offsets.data());
     } else {
-        cmdbuf.bindVertexBuffers2EXT(0, num_buffers, host_buffers.data(), host_offsets.data(),
+        cmdbuf.bindVertexBuffers2(0, num_buffers, host_buffers.data(), host_offsets.data(),
-                                     host_sizes.data(), host_strides.data());
+                                  host_sizes.data(), host_strides.data());
     }
 }
 

src/video_core/renderer_vulkan/vk_presenter.cpp

@@ -270,7 +270,25 @@ Frame* Presenter::PrepareLastFrame() {
     return frame;
 }
 
-Frame* Presenter::PrepareFrameInternal(VideoCore::ImageId image_id, bool is_eop) {
+static vk::Format GetFrameViewFormat(const Libraries::VideoOut::PixelFormat format) {
+    switch (format) {
+    case Libraries::VideoOut::PixelFormat::A8B8G8R8Srgb:
+        return vk::Format::eR8G8B8A8Srgb;
+    case Libraries::VideoOut::PixelFormat::A8R8G8B8Srgb:
+        return vk::Format::eB8G8R8A8Srgb;
+    case Libraries::VideoOut::PixelFormat::A2R10G10B10:
+    case Libraries::VideoOut::PixelFormat::A2R10G10B10Srgb:
+    case Libraries::VideoOut::PixelFormat::A2R10G10B10Bt2020Pq:
+        return vk::Format::eA2R10G10B10UnormPack32;
+    default:
+        break;
+    }
+    UNREACHABLE_MSG("Unknown format={}", static_cast<u32>(format));
+    return {};
+}
+
+Frame* Presenter::PrepareFrameInternal(VideoCore::ImageId image_id,
+                                       const Libraries::VideoOut::PixelFormat format, bool is_eop) {
     // Request a free presentation frame.
     Frame* frame = GetRenderFrame();
 
@@ -324,7 +342,7 @@ Frame* Presenter::PrepareFrameInternal(VideoCore::ImageId image_id, bool is_eop)
                       cmdbuf);
 
         VideoCore::ImageViewInfo info{};
-        info.format = image.info.pixel_format;
+        info.format = GetFrameViewFormat(format);
         // Exclude alpha from output frame to avoid blending with UI.
         info.mapping = vk::ComponentMapping{
             .r = vk::ComponentSwizzle::eIdentity,

src/video_core/renderer_vulkan/vk_presenter.h

@@ -70,11 +70,12 @@ public:
         auto desc = VideoCore::TextureCache::VideoOutDesc{attribute, cpu_address};
         const auto image_id = texture_cache.FindImage(desc);
         texture_cache.UpdateImage(image_id, is_eop ? nullptr : &flip_scheduler);
-        return PrepareFrameInternal(image_id, is_eop);
+        return PrepareFrameInternal(image_id, attribute.attrib.pixel_format, is_eop);
     }
 
     Frame* PrepareBlankFrame(bool is_eop) {
-        return PrepareFrameInternal(VideoCore::NULL_IMAGE_ID, is_eop);
+        return PrepareFrameInternal(VideoCore::NULL_IMAGE_ID,
+                                    Libraries::VideoOut::PixelFormat::Unknown, is_eop);
     }
 
     VideoCore::Image& RegisterVideoOutSurface(
@@ -119,7 +120,8 @@ public:
     }
 
 private:
-    Frame* PrepareFrameInternal(VideoCore::ImageId image_id, bool is_eop = true);
+    Frame* PrepareFrameInternal(VideoCore::ImageId image_id,
+                                Libraries::VideoOut::PixelFormat format, bool is_eop = true);
     Frame* GetRenderFrame();
 
     void SetExpectedGameSize(s32 width, s32 height);

src/video_core/texture_cache/image_info.cpp

@@ -16,14 +16,15 @@ using VideoOutFormat = Libraries::VideoOut::PixelFormat;
 
 static vk::Format ConvertPixelFormat(const VideoOutFormat format) {
     switch (format) {
-    case VideoOutFormat::A8R8G8B8Srgb:
-        return vk::Format::eB8G8R8A8Srgb;
     case VideoOutFormat::A8B8G8R8Srgb:
+    // Remaining formats are mapped to RGBA for internal consistency and changed to BGRA in the
+    // frame image view.
+    case VideoOutFormat::A8R8G8B8Srgb:
         return vk::Format::eR8G8B8A8Srgb;
     case VideoOutFormat::A2R10G10B10:
     case VideoOutFormat::A2R10G10B10Srgb:
     case VideoOutFormat::A2R10G10B10Bt2020Pq:
-        return vk::Format::eA2R10G10B10UnormPack32;
+        return vk::Format::eA2B10G10R10UnormPack32;
     default:
         break;
     }