Merge branch 'shadps4-emu:main' into main

CMakeLists.txt

@@ -597,6 +597,8 @@ set(MISC_LIBS  src/core/libraries/screenshot/screenshot.cpp
                src/core/libraries/move/move.h
                src/core/libraries/ulobjmgr/ulobjmgr.cpp
                src/core/libraries/ulobjmgr/ulobjmgr.h
+               src/core/libraries/signin_dialog/signindialog.cpp
+               src/core/libraries/signin_dialog/signindialog.h
 )
 
 set(DEV_TOOLS src/core/devtools/layer.cpp

src/common/logging/filter.cpp

@@ -137,6 +137,7 @@ bool ParseFilterRule(Filter& instance, Iterator begin, Iterator end) {
     SUB(Lib, NpParty)                                                                              \
     SUB(Lib, Zlib)                                                                                 \
     SUB(Lib, Hmd)                                                                                  \
+    SUB(Lib, SigninDialog)                                                                         \
     CLS(Frontend)                                                                                  \
     CLS(Render)                                                                                    \
     SUB(Render, Vulkan)                                                                            \

src/common/logging/types.h

@@ -104,6 +104,7 @@ enum class Class : u8 {
     Lib_NpParty,           ///< The LibSceNpParty implementation
     Lib_Zlib,              ///< The LibSceZlib implementation.
     Lib_Hmd,               ///< The LibSceHmd implementation.
+    Lib_SigninDialog,      ///< The LibSigninDialog implementation.
     Frontend,              ///< Emulator UI
     Render,                ///< Video Core
     Render_Vulkan,         ///< Vulkan backend

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/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/libc_internal/printf.h

@@ -56,7 +56,6 @@
 #include <stdio.h>
 
 #include <cstdarg>
-#include <cstdbool>
 #include <cstddef>
 #include <cstdint>
 #include <cstring>

src/core/libraries/libs.cpp

@@ -45,6 +45,7 @@
 #include "core/libraries/save_data/savedata.h"
 #include "core/libraries/screenshot/screenshot.h"
 #include "core/libraries/share_play/shareplay.h"
+#include "core/libraries/signin_dialog/signindialog.h"
 #include "core/libraries/system/commondialog.h"
 #include "core/libraries/system/msgdialog.h"
 #include "core/libraries/system/posix.h"
@@ -120,6 +121,7 @@ void InitHLELibs(Core::Loader::SymbolsResolver* sym) {
     Libraries::Hmd::RegisterlibSceHmd(sym);
     Libraries::DiscMap::RegisterlibSceDiscMap(sym);
     Libraries::Ulobjmgr::RegisterlibSceUlobjmgr(sym);
+    Libraries::SigninDialog::RegisterlibSceSigninDialog(sym);
 }
 
 } // namespace Libraries

src/core/libraries/signin_dialog/signindialog.cpp

@@ -0,0 +1,64 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+// Generated By moduleGenerator
+#include "common/logging/log.h"
+#include "core/libraries/error_codes.h"
+#include "core/libraries/libs.h"
+#include "signindialog.h"
+
+namespace Libraries::SigninDialog {
+
+s32 PS4_SYSV_ABI sceSigninDialogInitialize() {
+    LOG_ERROR(Lib_SigninDialog, "(STUBBED) called");
+    return ORBIS_OK;
+}
+
+s32 PS4_SYSV_ABI sceSigninDialogOpen() {
+    LOG_ERROR(Lib_SigninDialog, "(STUBBED) called");
+    return ORBIS_OK;
+}
+
+Status PS4_SYSV_ABI sceSigninDialogGetStatus() {
+    LOG_ERROR(Lib_SigninDialog, "(STUBBED) called, return 'finished' status");
+    return Status::FINISHED;
+}
+
+Status PS4_SYSV_ABI sceSigninDialogUpdateStatus() {
+    LOG_ERROR(Lib_SigninDialog, "(STUBBED) called, return 'finished' status");
+    return Status::FINISHED;
+}
+
+s32 PS4_SYSV_ABI sceSigninDialogGetResult() {
+    LOG_ERROR(Lib_SigninDialog, "(STUBBED) called");
+    return ORBIS_OK;
+}
+
+s32 PS4_SYSV_ABI sceSigninDialogClose() {
+    LOG_ERROR(Lib_SigninDialog, "(STUBBED) called");
+    return ORBIS_OK;
+}
+
+s32 PS4_SYSV_ABI sceSigninDialogTerminate() {
+    LOG_ERROR(Lib_SigninDialog, "(STUBBED) called");
+    return ORBIS_OK;
+}
+
+void RegisterlibSceSigninDialog(Core::Loader::SymbolsResolver* sym) {
+    LIB_FUNCTION("mlYGfmqE3fQ", "libSceSigninDialog", 1, "libSceSigninDialog", 1, 1,
+                 sceSigninDialogInitialize);
+    LIB_FUNCTION("JlpJVoRWv7U", "libSceSigninDialog", 1, "libSceSigninDialog", 1, 1,
+                 sceSigninDialogOpen);
+    LIB_FUNCTION("2m077aeC+PA", "libSceSigninDialog", 1, "libSceSigninDialog", 1, 1,
+                 sceSigninDialogGetStatus);
+    LIB_FUNCTION("Bw31liTFT3A", "libSceSigninDialog", 1, "libSceSigninDialog", 1, 1,
+                 sceSigninDialogUpdateStatus);
+    LIB_FUNCTION("nqG7rqnYw1U", "libSceSigninDialog", 1, "libSceSigninDialog", 1, 1,
+                 sceSigninDialogGetResult);
+    LIB_FUNCTION("M3OkENHcyiU", "libSceSigninDialog", 1, "libSceSigninDialog", 1, 1,
+                 sceSigninDialogClose);
+    LIB_FUNCTION("LXlmS6PvJdU", "libSceSigninDialog", 1, "libSceSigninDialog", 1, 1,
+                 sceSigninDialogTerminate);
+};
+
+} // namespace Libraries::SigninDialog

src/core/libraries/signin_dialog/signindialog.h

@@ -0,0 +1,29 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+#include "common/types.h"
+
+namespace Core::Loader {
+class SymbolsResolver;
+}
+
+enum class Status : u32 {
+    NONE = 0,
+    INITIALIZED = 1,
+    RUNNING = 2,
+    FINISHED = 3,
+};
+
+namespace Libraries::SigninDialog {
+
+s32 PS4_SYSV_ABI sceSigninDialogInitialize();
+s32 PS4_SYSV_ABI sceSigninDialogOpen();
+Status PS4_SYSV_ABI sceSigninDialogGetStatus();
+Status PS4_SYSV_ABI sceSigninDialogUpdateStatus();
+s32 PS4_SYSV_ABI sceSigninDialogGetResult();
+s32 PS4_SYSV_ABI sceSigninDialogClose();
+s32 PS4_SYSV_ABI sceSigninDialogTerminate();
+
+void RegisterlibSceSigninDialog(Core::Loader::SymbolsResolver* sym);
+} // namespace Libraries::SigninDialog

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)...);
 }