Mac OS ARM Native reaches ELF entry

CMakeLists.txt

@@ -657,8 +657,11 @@ if (APPLE)
       target_link_libraries(shadps4 PRIVATE ${MOLTENVK})
   endif()
 
+  # Using x86_64 (ARM by default if not set)
+  if (CMAKE_OSX_ARCHITECTURES)
     # Reserve system-managed memory space.
     target_link_options(shadps4 PRIVATE -Wl,-no_pie,-no_fixup_chains,-no_huge,-pagezero_size,0x4000,-segaddr,TCB_SPACE,0x4000,-segaddr,GUEST_SYSTEM,0x400000,-image_base,0x20000000000)
+  endif()
 
   # Replacement for std::chrono::time_zone
   target_link_libraries(shadps4 PRIVATE date::date-tz)

externals/CMakeLists.txt

@@ -43,7 +43,15 @@ else()
 endif()
 
 if (NOT TARGET FFmpeg::ffmpeg)
+    if (APPLE)
+        if (CMAKE_OSX_ARCHITECTURES)
+            set(ARCHITECTURE "${CMAKE_OSX_ARCHITECTURES}")
+        else()
+            set(ARCHITECTURE "${CMAKE_SYSTEM_PROCESSOR}")
+        endif()
+    else()
         set(ARCHITECTURE "x86_64")
+    endif()
     add_subdirectory(ffmpeg-core)
     add_library(FFmpeg::ffmpeg ALIAS ffmpeg)
 endif()

src/common/assert.cpp

@@ -4,7 +4,11 @@
 #include "common/assert.h"
 #include "common/logging/backend.h"
 
+#ifdef __x86_64__
 #define Crash() __asm__ __volatile__("int $3")
+#elif __aarch64__
+#define Crash() ;
+#endif
 
 void assert_fail_impl() {
     Common::Log::Stop();

src/common/rdtsc.cpp

@@ -38,7 +38,10 @@ static u64 GetTimeNs() {
 #endif
 }
 
-u64 EstimateRDTSCFrequency() {
+u64 EstimateRDTSCFrequency() 
+{
+    #ifdef __x86_64__
+
     // Discard the first result measuring the rdtsc.
     FencedRDTSC();
     std::this_thread::sleep_for(std::chrono::milliseconds{1});
@@ -55,6 +58,12 @@ u64 EstimateRDTSCFrequency() {
     const u64 tsc_diff = tsc_end - tsc_start;
     const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, end_time - start_time);
     return RoundToNearest<100'000>(tsc_freq);
+
+    #else
+
+    return 0;
+
+    #endif
 }
 
 } // namespace Common

src/common/rdtsc.h

@@ -11,7 +11,10 @@
 
 namespace Common {
 
+#ifdef __x86_64__
+
 #ifdef _MSC_VER
+// MSVC
 __forceinline static u64 FencedRDTSC() {
     _mm_lfence();
     _ReadWriteBarrier();
@@ -21,6 +24,7 @@ __forceinline static u64 FencedRDTSC() {
     return result;
 }
 #else
+// Linux/Mac
 static inline u64 FencedRDTSC() {
     u64 eax;
     u64 edx;
@@ -32,6 +36,13 @@ static inline u64 FencedRDTSC() {
 }
 #endif
 
+#else
+// ARM
+static inline u64 FencedRDTSC() {
+    return 0;
+}
+#endif
+
 u64 EstimateRDTSCFrequency();
 
 } // namespace Common

src/core/address_space.cpp

@@ -17,9 +17,11 @@
 #endif
 
 #ifdef __APPLE__
+#ifdef __x86_64__
 // Reserve space for the system address space using a zerofill section.
 asm(".zerofill GUEST_SYSTEM,GUEST_SYSTEM,__guest_system,0xFBFC00000");
 #endif
+#endif
 
 namespace Core {
 
@@ -299,6 +301,7 @@ struct AddressSpace::Impl {
         constexpr int protection_flags = PROT_READ | PROT_WRITE;
         constexpr int base_map_flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE;
 #ifdef __APPLE__
+#ifdef __x86_64__
         // On ARM64 Macs, we run into limitations due to the commpage from 0xFC0000000 - 0xFFFFFFFFF
         // and the GPU carveout region from 0x1000000000 - 0x6FFFFFFFFF. We can allocate the system
         // managed region, as well as system reserved if reduced in size slightly, but we cannot map
@@ -313,6 +316,14 @@ struct AddressSpace::Impl {
         // Cannot guarantee enough space for these areas at the desired addresses, so not MAP_FIXED.
         user_base = reinterpret_cast<u8*>(mmap(reinterpret_cast<void*>(USER_MIN), user_size,
                                                protection_flags, base_map_flags, -1, 0));
+#else
+        const auto virtual_size = system_managed_size + system_reserved_size + user_size;
+        const auto virtual_base =
+            reinterpret_cast<u8*>(mmap(nullptr, virtual_size, protection_flags, base_map_flags | MAP_JIT, -1, 0));
+        system_managed_base = virtual_base;
+        system_reserved_base = virtual_base + (SYSTEM_RESERVED_MIN - SYSTEM_MANAGED_MIN);
+        user_base = virtual_base + (USER_MIN - SYSTEM_MANAGED_MIN);
+#endif
 #else
         const auto virtual_size = system_managed_size + system_reserved_size + user_size;
         const auto virtual_base =
@@ -382,13 +393,16 @@ struct AddressSpace::Impl {
     void* Map(VAddr virtual_addr, PAddr phys_addr, size_t size, PosixPageProtection prot,
               int fd = -1) {
         m_free_regions.subtract({virtual_addr, virtual_addr + size});
-        const int handle = phys_addr != -1 ? (fd == -1 ? backing_fd : fd) : -1;
+        if (phys_addr != -1) {
-        const off_t host_offset = phys_addr != -1 ? phys_addr : 0;
+            const int handle = fd == -1 ? backing_fd : fd;
-        const int flag = phys_addr != -1 ? MAP_SHARED : (MAP_ANONYMOUS | MAP_PRIVATE);
+            void* ret = mmap(reinterpret_cast<void*>(virtual_addr), size, prot, MAP_FIXED | MAP_SHARED,
-        void* ret = mmap(reinterpret_cast<void*>(virtual_addr), size, prot, MAP_FIXED | flag,
+                             handle, phys_addr);
-                         handle, host_offset);
             ASSERT_MSG(ret != MAP_FAILED, "mmap failed: {}", strerror(errno));
-        return ret;
+        } else {
+            int ret = mprotect(reinterpret_cast<void*>(virtual_addr), size, prot);
+            ASSERT_MSG(ret == 0, "mprotect failed: {}", strerror(errno));
+        }
+        return reinterpret_cast<void*>(virtual_addr);
     }
 
     void Unmap(VAddr virtual_addr, size_t size, bool) {
@@ -407,9 +421,8 @@ struct AddressSpace::Impl {
         m_free_regions.insert({start_address, end_address});
 
         // Return the adjusted pointers.
-        void* ret = mmap(reinterpret_cast<void*>(start_address), end_address - start_address,
+        int ret = mprotect(reinterpret_cast<void*>(start_address), end_address - start_address, PROT_NONE);
-                         PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0);
+        ASSERT_MSG(ret == 0, "mprotect failed: {}", strerror(errno));
-        ASSERT_MSG(ret != MAP_FAILED, "mmap failed: {}", strerror(errno));
     }
 
     void Protect(VAddr virtual_addr, size_t size, bool read, bool write, bool execute) {

src/core/cpu_patches.cpp

@@ -1,6 +1,9 @@
 // SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+// full file
+#ifdef __x86_64__
+
 #include <memory>
 #include <mutex>
 #include <Zydis/Zydis.h>
@@ -682,3 +685,5 @@ void PatchInstructions(u64 segment_addr, u64 segment_size, Xbyak::CodeGenerator&
 }
 
 } // namespace Core
+
+#endif // __x86_64__

src/core/cpu_patches.h

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#ifdef __x86_64__
 namespace Xbyak {
 class CodeGenerator;
 }
@@ -19,3 +20,4 @@ void CleanupThreadPatchStack();
 void PatchInstructions(u64 segment_addr, u64 segment_size, Xbyak::CodeGenerator& c);
 
 } // namespace Core
+#endif

src/core/libraries/kernel/thread_management.cpp

@@ -988,14 +988,22 @@ static void cleanup_thread(void* arg) {
             destructor(value);
         }
     }
+
+    #ifdef __x86_64__
     Core::SetTcbBase(nullptr);
+    #endif
+
     thread->is_almost_done = true;
 }
 
 static void* run_thread(void* arg) {
     auto* thread = static_cast<ScePthread>(arg);
     Common::SetCurrentThreadName(thread->name.c_str());
+
+    #ifdef __x86_64__
     Core::InitializeThreadPatchStack();
+    #endif
+    
     auto* linker = Common::Singleton<Core::Linker>::Instance();
     linker->InitTlsForThread(false);
     void* ret = nullptr;

src/core/linker.cpp

@@ -26,6 +26,7 @@ static PS4_SYSV_ABI void ProgramExitFunc() {
     fmt::print("exit function called\n");
 }
 
+#ifdef __x86_64__
 static void RunMainEntry(VAddr addr, EntryParams* params, ExitFunc exit_func) {
     // reinterpret_cast<entry_func_t>(addr)(params, exit_func); // can't be used, stack has to have
     // a specific layout
@@ -47,6 +48,13 @@ static void RunMainEntry(VAddr addr, EntryParams* params, ExitFunc exit_func) {
                  : "r"(addr), "r"(params), "r"(exit_func)
                  : "rax", "rsi", "rdi");
 }
+#elif __aarch64__
+static void RunMainEntry(VAddr addr, EntryParams* params, ExitFunc exit_func) {
+    printf("Arm Entry\n");
+    while(1) {}
+}
+#endif
+
 
 Linker::Linker() : memory{Memory::Instance()} {}
 
@@ -85,7 +93,11 @@ void Linker::Execute() {
 
     // Init primary thread.
     Common::SetCurrentThreadName("GAME_MainThread");
+
+    #ifdef __x86_64__
     InitializeThreadPatchStack();
+    #endif
+
     Libraries::Kernel::pthreadInitSelfMainThread();
     InitTlsForThread(true);
 
@@ -107,7 +119,9 @@ void Linker::Execute() {
         }
     }
 
+    #ifdef __x86_64__
     SetTcbBase(nullptr);
+    #endif
 }
 
 s32 Linker::LoadModule(const std::filesystem::path& elf_name, bool is_dynamic) {
@@ -292,7 +306,9 @@ bool Linker::Resolve(const std::string& name, Loader::SymbolType sym_type, Modul
     return false;
 }
 
-void* Linker::TlsGetAddr(u64 module_index, u64 offset) {
+void* Linker::TlsGetAddr(u64 module_index, u64 offset) 
+{
+#ifdef __x86_64__
     std::scoped_lock lk{mutex};
 
     DtvEntry* dtv_table = GetTcbBase()->tcb_dtv;
@@ -326,6 +342,9 @@ void* Linker::TlsGetAddr(u64 module_index, u64 offset) {
         addr = dest;
     }
     return addr + offset;
+#else
+    return 0;
+#endif
 }
 
 void Linker::InitTlsForThread(bool is_primary) {
@@ -395,8 +414,10 @@ void Linker::InitTlsForThread(bool is_primary) {
         tcb->tcb_dtv[module->tls.modid + 1].pointer = dest;
     }
 
+    #ifdef __x86_64__
     // Set pointer to FS base
     SetTcbBase(tcb);
+    #endif
 }
 
 void Linker::DebugDump() {

src/core/module.cpp

@@ -1,7 +1,10 @@
 // SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+#ifdef __x86_64__
 #include <xbyak/xbyak.h>
+#endif
+
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/logging/log.h"
@@ -93,9 +96,11 @@ void Module::LoadModuleToMemory(u32& max_tls_index) {
     LoadOffset += CODE_BASE_INCR * (1 + aligned_base_size / CODE_BASE_INCR);
     LOG_INFO(Core_Linker, "Loading module {} to {}", name, fmt::ptr(*out_addr));
 
+    #ifdef __x86_64__
     // Initialize trampoline generator.
     void* trampoline_addr = std::bit_cast<void*>(base_virtual_addr + aligned_base_size);
     Xbyak::CodeGenerator c(TrampolineSize, trampoline_addr);
+    #endif
 
     LOG_INFO(Core_Linker, "======== Load Module to Memory ========");
     LOG_INFO(Core_Linker, "base_virtual_addr ......: {:#018x}", base_virtual_addr);
@@ -105,7 +110,9 @@ void Module::LoadModuleToMemory(u32& max_tls_index) {
     const auto add_segment = [this](const elf_program_header& phdr, bool do_map = true) {
         const VAddr segment_addr = base_virtual_addr + phdr.p_vaddr;
         if (do_map) {
+            pthread_jit_write_protect_np(false);
             elf.LoadSegment(segment_addr, phdr.p_offset, phdr.p_filesz);
+            pthread_jit_write_protect_np(true);
         }
         auto& segment = info.segments[info.num_segments++];
         segment.address = segment_addr;
@@ -115,6 +122,9 @@ void Module::LoadModuleToMemory(u32& max_tls_index) {
 
     for (u16 i = 0; i < elf_header.e_phnum; i++) {
         const auto header_type = elf.ElfPheaderTypeStr(elf_pheader[i].p_type);
+
+        printf("%d/%d: %d\n", i, elf_header.e_phnum, elf_pheader[i].p_type);
+
         switch (elf_pheader[i].p_type) {
         case PT_LOAD:
         case PT_SCE_RELRO: {
@@ -134,8 +144,11 @@ void Module::LoadModuleToMemory(u32& max_tls_index) {
             LOG_INFO(Core_Linker, "segment_mode ..........: {}", segment_mode);
 
             add_segment(elf_pheader[i]);
-            if (elf_pheader[i].p_flags & PF_EXEC) {
+            if (elf_pheader[i].p_flags & PF_EXEC) 
+            {
+                #ifdef __x86_64__
                 PatchInstructions(segment_addr, segment_file_size, c);
+                #endif
             }
             break;
         }

src/core/tls.cpp

@@ -1,6 +1,8 @@
 // SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+#ifdef __x86_64__
+
 #include <mutex>
 #include "common/assert.h"
 #include "common/types.h"
@@ -147,3 +149,5 @@ Tcb* GetTcbBase() {
 #endif
 
 } // namespace Core
+
+#endif

src/core/tls.h

@@ -5,9 +5,11 @@
 
 #include "common/types.h"
 
+#ifdef __x86_64__
 namespace Xbyak {
 class CodeGenerator;
 }
+#endif
 
 namespace Core {
 

src/core/virtual_memory.cpp

@@ -127,6 +127,10 @@ bool memory_patch(u64 vaddr, u64 value) {
     MemoryMode old_mode{};
     // memory_protect(vaddr, 8, MemoryMode::ReadWrite, &old_mode);
 
+#if defined(__APPLE__) && defined(__aarch64__)
+    pthread_jit_write_protect_np(false);
+#endif
+
     auto* ptr = reinterpret_cast<uint64_t*>(vaddr);
 
     bool ret = (*ptr != value);
@@ -140,6 +144,10 @@ bool memory_patch(u64 vaddr, u64 value) {
         memory_flush(vaddr, 8);
     }
 
+#if defined(__APPLE__) && defined(__aarch64__)
+    pthread_jit_write_protect_np(true);
+#endif
+
     return ret;
 }
 static u64 AlignUp(u64 pos, u64 align) {

src/video_core/page_manager.cpp

@@ -194,11 +194,17 @@ struct PageManager::Impl {
     static void GuestFaultSignalHandler(int sig, siginfo_t* info, void* raw_context) {
         ucontext_t* ctx = reinterpret_cast<ucontext_t*>(raw_context);
         const VAddr address = reinterpret_cast<VAddr>(info->si_addr);
-#ifdef __APPLE__
+
+#ifdef __x86_64__
+    #ifdef __APPLE__
         const u32 err = ctx->uc_mcontext->__es.__err;
-#else
+    #else
         const greg_t err = ctx->uc_mcontext.gregs[REG_ERR];
+    #endif
+#else
+        u32 err = 0;
 #endif
+
         if (err & 0x2) {
             const VAddr addr_aligned = Common::AlignDown(address, PAGESIZE);
             rasterizer->InvalidateMemory(addr_aligned, PAGESIZE);