Quelle  runtime.cc

/*
 * Copyright (C) 2011 The Android Open Source Project
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include "runtime.h"

#include <optional>
#include <utility>

#ifdef __linux__
#include <sys/prctl.h>
#endif

#include <fcntl.h>
#include <signal.h>
#include <sys/mount.h>
#include <sys/syscall.h>

#if defined(__APPLE__)
#include <crt_externs.h>  // for _NSGetEnviron
#endif

#include <android-base/properties.h>
#include <android-base/strings.h>
#include <string.h>

#include <cstdio>
#include <cstdlib>
#include <limits>
#include <thread>
#include <unordered_set>
#include <vector>

#include "arch/arm/registers_arm.h"
#include "arch/arm64/registers_arm64.h"
#include "arch/context.h"
#include "arch/instruction_set_features.h"
#include "arch/x86/registers_x86.h"
#include "arch/x86_64/registers_x86_64.h"
#include "art_field-inl.h"
#include "art_method-inl.h"
#include "asm_support.h"
#include "assume_value_signatures.h"
#include "base/aborting.h"
#include "base/arena_allocator.h"
#include "base/atomic.h"
#include "base/calloc_arena_pool.h"
#include "base/dumpable.h"
#include "base/file_utils.h"
#include "base/flags.h"
#include "base/mem_map_arena_pool.h"
#include "base/memory_tool.h"
#include "base/mutex.h"
#include "base/os.h"
#include "base/pointer_size.h"
#include "base/quasi_atomic.h"
#include "base/sdk_version.h"
#include "base/stl_util.h"
#include "base/systrace.h"
#include "base/unix_file/fd_file.h"
#include "base/utils.h"
#include "class_linker-inl.h"
#include "class_root-inl.h"
#include "com_android_art_rw_flags.h"
#include "compiler_callbacks.h"
#include "debugger.h"
#include "dex/art_dex_file_loader.h"
#include "dex/dex_file_loader.h"
#include "entrypoints/entrypoint_utils-inl.h"
#include "entrypoints/runtime_asm_entrypoints.h"
#include "experimental_flags.h"
#include "fault_handler.h"
#include "gc/accounting/card_table-inl.h"
#include "gc/heap.h"
#include "gc/scoped_gc_critical_section.h"
#include "gc/space/image_space.h"
#include "gc/space/space-inl.h"
#include "gc/system_weak.h"
#include "gc/task_processor.h"
#include "handle_scope-inl.h"
#include "hidden_api.h"
#include "instrumentation.h"
#include "intern_table-inl.h"
#include "interpreter/interpreter.h"
#include "jit/jit.h"
#include "jit/jit_code_cache.h"
#include "jit/profile_saver.h"
#include "jni/java_vm_ext.h"
#include "jni/jni_id_manager.h"
#include "jni_id_type.h"
#include "linear_alloc.h"
#include "memory_representation.h"
#include "metrics/statsd.h"
#include "mirror/array.h"
#include "mirror/class-alloc-inl.h"
#include "mirror/class-inl.h"
#include "mirror/class_ext.h"
#include "mirror/class_loader-inl.h"
#include "mirror/emulated_stack_frame.h"
#include "mirror/field.h"
#include "mirror/method.h"
#include "mirror/method_handle_impl.h"
#include "mirror/method_handles_lookup.h"
#include "mirror/method_type.h"
#include "mirror/stack_trace_element.h"
#include "mirror/throwable.h"
#include "mirror/var_handle.h"
#include "monitor.h"
#include "native/dalvik_system_BaseDexClassLoader.h"
#include "native/dalvik_system_DexFile.h"
#include "native/dalvik_system_VMDebug.h"
#include "native/dalvik_system_VMRuntime.h"
#include "native/dalvik_system_VMStack.h"
#include "native/dalvik_system_ZygoteHooks.h"
#include "native/java_lang_Class.h"
#include "native/java_lang_Object.h"
#include "native/java_lang_StackStreamFactory.h"
#include "native/java_lang_String.h"
#include "native/java_lang_StringFactory.h"
#include "native/java_lang_System.h"
#include "native/java_lang_Thread.h"
#include "native/java_lang_Throwable.h"
#include "native/java_lang_VMClassLoader.h"
#include "native/java_lang_invoke_MethodHandle.h"
#include "native/java_lang_invoke_MethodHandleImpl.h"
#include "native/java_lang_ref_FinalizerReference.h"
#include "native/java_lang_ref_Reference.h"
#include "native/java_lang_reflect_Array.h"
#include "native/java_lang_reflect_Constructor.h"
#include "native/java_lang_reflect_Executable.h"
#include "native/java_lang_reflect_Field.h"
#include "native/java_lang_reflect_Method.h"
#include "native/java_lang_reflect_Parameter.h"
#include "native/java_lang_reflect_Proxy.h"
#include "native/jdk_internal_misc_Unsafe.h"
#include "native/jdk_internal_vm_Continuation.h"
#include "native/libcore_io_Memory.h"
#include "native/libcore_util_CharsetUtils.h"
#include "native/org_apache_harmony_dalvik_ddmc_DdmServer.h"
#include "native/org_apache_harmony_dalvik_ddmc_DdmVmInternal.h"
#include "native/sun_misc_Unsafe.h"
#include "native_bridge_art_interface.h"
#include "native_stack_dump.h"
#include "nativehelper/scoped_local_ref.h"
#include "nterp_helpers.h"
#include "oat/elf_file.h"
#include "oat/image-inl.h"
#include "oat/oat.h"
#include "oat/oat_file_manager.h"
#include "oat/oat_quick_method_header.h"
#include "object_callbacks.h"
#include "odr_statslog/odr_statslog.h"
#include "parsed_options.h"
#include "quick/quick_method_frame_info.h"
#include "reflection.h"
#include "runtime_callbacks.h"
#include "runtime_common.h"
#include "runtime_image.h"
#include "runtime_intrinsics.h"
#include "runtime_options.h"
#include "scoped_thread_state_change-inl.h"
#include "sigchain.h"
#include "signal_catcher.h"
#include "signal_set.h"
#include "thread.h"
#include "thread_list.h"
#include "ti/agent.h"
#include "trace.h"
#include "vdex_file.h"
#include "verifier/class_verifier.h"
#include "well_known_classes-inl.h"

#ifdef ART_TARGET_ANDROID
#include <android/api-level.h>
#include <android/set_abort_message.h>
#include <linux/magic.h>
#include <sys/vfs.h>

#include "com_android_apex.h"

namespace apex = com::android::apex;

#endif
#ifdef ART_USE_SIMULATOR
#include "code_simulator_container.h"
#endif

// Static asserts to check the values of generated assembly-support macros.
#define ASM_DEFINE(NAME, EXPR) static_assert((NAME) == (EXPR), "Unexpected value of " #NAME);
#include "asm_defines.def"
#undef ASM_DEFINE

namespace art HIDDEN {

// If a signal isn't handled properly, enable a handler that attempts to dump the Java stack.
static constexpr bool kEnableJavaStackTraceHandler = false;
// Tuned by compiling GmsCore under perf and measuring time spent in DescriptorEquals for class
// linking.
static constexpr double kLowMemoryMinLoadFactor = 0.5;
static constexpr double kLowMemoryMaxLoadFactor = 0.8;
static constexpr double kNormalMinLoadFactor = 0.4;
static constexpr double kNormalMaxLoadFactor = 0.7;

#ifdef ART_PAGE_SIZE_AGNOSTIC
// Declare the constant as ALWAYS_HIDDEN to ensure it isn't visible from outside libart.so.
const size_t PageSize::value_ ALWAYS_HIDDEN = GetPageSizeSlow();
PageSize gPageSize ALWAYS_HIDDEN;
#endif

Runtime* Runtime::instance_ = nullptr;

struct TraceConfig {
  Trace::TraceMode trace_mode;
  TraceOutputMode trace_output_mode;
  std::string trace_file;
  size_t trace_file_size;
  TraceClockSource clock_source;
};

extern bool ShouldUseGenerationalGC();

namespace {

#ifdef __APPLE__
inline char** GetEnviron() {
  // When Google Test is built as a framework on MacOS X, the environ variable
  // is unavailable. Apple's documentation (man environ) recommends using
  // _NSGetEnviron() instead.
  return *_NSGetEnviron();
}
#else
// Some POSIX platforms expect you to declare environ. extern "C" makes
// it reside in the global namespace.
EXPORT extern "C" char** environ;
inline char** GetEnviron() { return environ; }
#endif

void CheckConstants() {
  CHECK_EQ(mirror::Array::kFirstElementOffset, mirror::Array::FirstElementOffset());
}

// Helper method do determine if the given location can safely assume a larger readahead window.
// For now, we only assume this if 1) the location is in /data, and 2) /data is f2fs. This query
// is cheap enough to invoke before each madvise, avoiding extra syscalls and allocations
bool LocationSupportsLargeReadahead([[maybe_unused]] std::string_view location) {
#if defined(ART_TARGET_ANDROID)
  static const char* kDataDir = [] {
    const char* data_dir = getenv("ANDROID_DATA");
    return (data_dir != nullptr) ? data_dir : "/data";
  }();

  static const bool kDataIsF2fs = [] {
    struct statfs buf;
    // Note that the stat call may fail in sandboxed processes. As this query is purely for
    // potential optimizations, treat that failure as benign.
    return statfs(kDataDir, &buf) == 0 && buf.f_type == F2FS_SUPER_MAGIC;
  }();

  if (!kDataIsF2fs) {
    return false;
  }

  std::string_view data_dir(kDataDir);

  // Normalize the ending to simplify root equivalence checks.
  if (UNLIKELY(data_dir.ends_with('/'))) {
    data_dir.remove_suffix(1);
  }

  // Perform a fast boundary-safe lexical check.
  if (!location.starts_with(data_dir)) {
    return false;
  }

  // Check if identical or if the next character is a separator, avoiding matches of
  // `/datafoo/bar against `/data`.
  return location.length() == data_dir.length() || location[data_dir.length()] == '/';
#else
  return false;
#endif  // defined(ART_TARGET_ANDROID)
}

}  // namespace

Runtime::Runtime()
    : resolution_method_(nullptr),
      imt_conflict_method_(nullptr),
      imt_unimplemented_method_(nullptr),
      instruction_set_(InstructionSet::kNone),
      compiler_callbacks_(nullptr),
      is_zygote_(false),
      is_primary_zygote_(false),
      is_system_server_(false),
      must_relocate_(false),
      is_concurrent_gc_enabled_(true),
      is_explicit_gc_disabled_(false),
      is_eagerly_release_explicit_gc_disabled_(false),
      image_dex2oat_enabled_(true),
      default_stack_size_(0),
      heap_(nullptr),
      max_spins_before_thin_lock_inflation_(Monitor::kDefaultMaxSpinsBeforeThinLockInflation),
      monitor_list_(nullptr),
      monitor_pool_(nullptr),
      thread_list_(nullptr),
      intern_table_(nullptr),
      class_linker_(nullptr),
      signal_catcher_(nullptr),
      java_vm_(nullptr),
      thread_pool_ref_count_(0u),
      fault_message_(nullptr),
      threads_being_born_(0),
      shutdown_cond_(new ConditionVariable("Runtime shutdown", *Locks::runtime_shutdown_lock_)),
      shutting_down_(false),
      shutting_down_started_(false),
      started_(false),
      finished_starting_(false),
      vfprintf_(nullptr),
      exit_(nullptr),
      abort_(nullptr),
      stats_enabled_(false),
      is_running_on_memory_tool_(kRunningOnMemoryTool),
      instrumentation_(new instrumentation::Instrumentation()),
      main_thread_group_(nullptr),
      system_thread_group_(nullptr),
      system_class_loader_(nullptr),
      dump_gc_performance_on_shutdown_(false),
      active_transaction_(false),
      verify_(verifier::VerifyMode::kNone),
      target_sdk_version_(static_cast<uint32_t>(SdkVersion::kUnset)),
      sdk_version_(
          android::base::GetUintProperty(
              "ro.build.version.sdk", static_cast<uint32_t>(SdkVersion::kUnset))),
      compat_framework_(),
      implicit_null_checks_(false),
      implicit_so_checks_(false),
      implicit_suspend_checks_(false),
      no_sig_chain_(false),
      force_native_bridge_(false),
      is_native_bridge_loaded_(false),
      is_native_debuggable_(false),
      async_exceptions_thrown_(false),
      non_standard_exits_enabled_(false),
      runtime_debug_state_(RuntimeDebugState::kNonJavaDebuggable),
      monitor_timeout_enable_(false),
      monitor_timeout_ns_(0),
      zygote_max_failed_boots_(0),
      experimental_flags_(ExperimentalFlags::kNone),
      oat_file_manager_(nullptr),
      is_low_memory_mode_(false),
      madvise_willneed_total_dex_size_(0),
      madvise_willneed_odex_filesize_(0),
      madvise_willneed_art_filesize_(0),
      safe_mode_(false),
      hidden_api_policy_(hiddenapi::EnforcementPolicy::kDisabled),
      core_platform_api_policy_(hiddenapi::EnforcementPolicy::kDisabled),
      test_api_policy_(hiddenapi::EnforcementPolicy::kDisabled),
      dedupe_hidden_api_warnings_(true),
      hidden_api_access_event_log_rate_(0),
      dump_native_stack_on_sig_quit_(true),
      // Initially assume we perceive jank in case the process state is never updated.
      process_state_(kProcessStateJankPerceptible),
      zygote_no_threads_(false),
      verifier_logging_threshold_ms_(100),
      verifier_missing_kthrow_fatal_(false),
      perfetto_hprof_enabled_(false),
      perfetto_javaheapprof_enabled_(false),
      out_of_memory_error_hook_(nullptr) {
  static_assert(Runtime::kCalleeSaveSize ==
                    static_cast<uint32_t>(CalleeSaveType::kLastCalleeSaveType), "Unexpected size");
  CheckConstants();

  std::fill(callee_save_methods_, callee_save_methods_ + arraysize(callee_save_methods_), 0u);
  interpreter::CheckInterpreterAsmConstants();
  callbacks_.reset(new RuntimeCallbacks());
  for (size_t i = 0; i <= static_cast<size_t>(DeoptimizationKind::kLast); ++i) {
    deoptimization_counts_[i] = 0u;
  }
}

Runtime::~Runtime() {
  ScopedTrace trace("Runtime shutdown");
  if (is_native_bridge_loaded_) {
    UnloadNativeBridge();
  }

  Thread* self = Thread::Current();
  const bool attach_shutdown_thread = self == nullptr;
  if (attach_shutdown_thread) {
    // We can only create a peer if the runtime is actually started. This is only not true during
    // some tests. If there is extreme memory pressure the allocation of the thread peer can fail.
    // In this case we will just try again without allocating a peer so that shutdown can continue.
    // Very few things are actually capable of distinguishing between the peer & peerless states so
    // this should be fine.
    // Running callbacks is prone to deadlocks in libjdwp tests that need an event handler lock to
    // process any event. We also need to enter a GCCriticalSection when processing certain events
    // (for ex: removing the last breakpoint). These two restrictions together make the tear down
    // of the jdwp tests deadlock prone if we fail to finish Thread::Attach callback.
    // (TODO:b/251163712) Remove this once we update deopt manager to not use GCCriticalSection.
    bool thread_attached = AttachCurrentThread("Shutdown thread",
                                               /* as_daemon= */ false,
                                               GetSystemThreadGroup(),
                                               /* create_peer= */ IsStarted(),
                                               /* should_run_callbacks= */ false);
    if (UNLIKELY(!thread_attached)) {
      LOG(WARNING) << "Failed to attach shutdown thread. Trying again without a peer.";
      CHECK(AttachCurrentThread("Shutdown thread (no java peer)",
                                /* as_daemon= */   false,
                                /* thread_group=*/ nullptr,
                                /* create_peer= */ false));
    }
    self = Thread::Current();
  } else {
    LOG(WARNING) << "Current thread not detached in Runtime shutdown";
  }

  if (dump_gc_performance_on_shutdown_) {
    heap_->CalculatePreGcWeightedAllocatedBytes();
    uint64_t process_cpu_end_time = ProcessCpuNanoTime();
    ScopedLogSeverity sls(LogSeverity::INFO);
    // This can't be called from the Heap destructor below because it
    // could call RosAlloc::InspectAll() which needs the thread_list
    // to be still alive.
    heap_->DumpGcPerformanceInfo(LOG_STREAM(INFO));

    uint64_t process_cpu_time = process_cpu_end_time - heap_->GetProcessCpuStartTime();
    uint64_t gc_cpu_time = heap_->GetTotalGcCpuTime();
    float ratio = static_cast<float>(gc_cpu_time) / process_cpu_time;
    LOG_STREAM(INFO) << "GC CPU time " << PrettyDuration(gc_cpu_time)
        << " out of process CPU time " << PrettyDuration(process_cpu_time)
        << " (" << ratio << ")"
        << "\n";
    double pre_gc_weighted_allocated_bytes =
        heap_->GetPreGcWeightedAllocatedBytes() / process_cpu_time;
    // Here we don't use process_cpu_time for normalization, because VM shutdown is not a real
    // GC. Both numerator and denominator take into account until the end of the last GC,
    // instead of the whole process life time like pre_gc_weighted_allocated_bytes.
    double post_gc_weighted_allocated_bytes =
        heap_->GetPostGcWeightedAllocatedBytes() /
          (heap_->GetPostGCLastProcessCpuTime() - heap_->GetProcessCpuStartTime());

    LOG_STREAM(INFO) << "Average bytes allocated at GC start, weighted by CPU time between GCs: "
        << static_cast<uint64_t>(pre_gc_weighted_allocated_bytes)
        << " (" <<  PrettySize(pre_gc_weighted_allocated_bytes)  << ")";
    LOG_STREAM(INFO) << "Average bytes allocated at GC end, weighted by CPU time between GCs: "
        << static_cast<uint64_t>(post_gc_weighted_allocated_bytes)
        << " (" <<  PrettySize(post_gc_weighted_allocated_bytes)  << ")"
        << "\n";
  }

  // Wait for the workers of thread pools to be created since there can't be any
  // threads attaching during shutdown.
  WaitForThreadPoolWorkersToStart();
  if (jit_ != nullptr) {
    jit_->WaitForWorkersToBeCreated();
    // Stop the profile saver thread before marking the runtime as shutting down.
    // The saver will try to dump the profiles before being sopped and that
    // requires holding the mutator lock.
    jit_->StopProfileSaver();
    // Delete thread pool before the thread list since we don't want to wait forever on the
    // JIT compiler threads. Also this should be run before marking the runtime
    // as shutting down as some tasks may require mutator access.
    jit_->DeleteThreadPool();
  }
  if (oat_file_manager_ != nullptr) {
    oat_file_manager_->WaitForWorkersToBeCreated();
  }
  // Disable GC before deleting the thread-pool and shutting down runtime as it
  // restricts attaching new threads.
  heap_->DisableGCForShutdown();
  heap_->WaitForWorkersToBeCreated();
  // Make sure to let the GC complete if it is running.
  heap_->WaitForGcToComplete(gc::kGcCauseBackground, self);

  // Shutdown any trace before SetShuttingDown. Trace uses thread pool workers to flush entries
  // and we want to make sure they are fully created. Threads cannot attach while shutting down.
  Trace::Shutdown();

  {
    ScopedTrace trace2("Wait for shutdown cond");
    MutexLock mu(self, *Locks::runtime_shutdown_lock_);
    shutting_down_started_ = true;
    while (threads_being_born_ > 0) {
      shutdown_cond_->Wait(self);
    }
    SetShuttingDown();
  }
  // Shutdown and wait for the daemons.
  CHECK(self != nullptr);
  if (IsFinishedStarting()) {
    ScopedTrace trace2("Waiting for Daemons");
    self->ClearException();
    ScopedObjectAccess soa(self);
    WellKnownClasses::java_lang_Daemons_stop->InvokeStatic<'V'>(self);
  }

  // Report death. Clients may require a working thread, still, so do it before GC completes and
  // all non-daemon threads are done.
  {
    ScopedObjectAccess soa(self);
    callbacks_->NextRuntimePhase(RuntimePhaseCallback::RuntimePhase::kDeath);
  }

  // Delete thread pools before detaching the current thread in case tasks
  // getting deleted need to have access to Thread::Current.
  heap_->DeleteThreadPool();
  if (oat_file_manager_ != nullptr) {
    oat_file_manager_->DeleteThreadPool();
  }
  DeleteThreadPool();
  CHECK(thread_pool_ == nullptr);

  if (attach_shutdown_thread) {
    DetachCurrentThread(/* should_run_callbacks= */ false);
    self = nullptr;
  }

  // Make sure our internal threads are dead before we start tearing down things they're using.
  GetRuntimeCallbacks()->StopDebugger();
  // Deletion ordering is tricky. Null out everything we've deleted.
  delete signal_catcher_;
  signal_catcher_ = nullptr;

  // Shutdown metrics reporting.
  metrics_reporter_.reset();

  // Make sure all other non-daemon threads have terminated, and all daemon threads are suspended.
  // Also wait for daemon threads to quiesce, so that in addition to being "suspended", they
  // no longer access monitor and thread list data structures. We leak user daemon threads
  // themselves, since we have no mechanism for shutting them down.
  {
    ScopedTrace trace2("Delete thread list");
    thread_list_->ShutDown();
  }

  // TODO Maybe do some locking.
  for (auto& agent : agents_) {
    agent->Unload();
  }

  // TODO Maybe do some locking
  for (auto& plugin : plugins_) {
    plugin.Unload();
  }

  // Finally delete the thread list.
  // Thread_list_ can be accessed by "suspended" threads, e.g. in InflateThinLocked.
  // We assume that by this point, we've waited long enough for things to quiesce.
  delete thread_list_;
  thread_list_ = nullptr;

  // Delete the JIT after thread list to ensure that there is no remaining threads which could be
  // accessing the instrumentation when we delete it.
  if (jit_ != nullptr) {
    VLOG(jit) << "Deleting jit";
    jit_.reset(nullptr);
    jit_code_cache_.reset(nullptr);
  }

  // Shutdown the fault manager if it was initialized.
  fault_manager.Shutdown();

  ScopedTrace trace2("Delete state");
  delete monitor_list_;
  monitor_list_ = nullptr;
  delete monitor_pool_;
  monitor_pool_ = nullptr;
  delete class_linker_;
  class_linker_ = nullptr;
  delete small_lrt_allocator_;
  small_lrt_allocator_ = nullptr;
  delete heap_;
  heap_ = nullptr;
  delete intern_table_;
  intern_table_ = nullptr;
  delete oat_file_manager_;
  oat_file_manager_ = nullptr;
  Thread::Shutdown();

  // Destroy allocators before shutting down the MemMap because they may use it.
  java_vm_.reset();
  linear_alloc_.reset();
  delete ReleaseStartupLinearAlloc();
  linear_alloc_arena_pool_.reset();
  arena_pool_.reset();
  jit_arena_pool_.reset();
  protected_fault_page_.Reset();
  assume_value_field_signatures_.clear();
  MemMap::Shutdown();

  // TODO: acquire a static mutex on Runtime to avoid racing.
  CHECK(instance_ == nullptr || instance_ == this);
  instance_ = nullptr;

  // Well-known classes must be deleted or it is impossible to successfully start another Runtime
  // instance. We rely on a small initialization order issue in Runtime::Start() that requires
  // elements of WellKnownClasses to be null, see b/65500943.
  WellKnownClasses::Clear();

#ifdef ART_PAGE_SIZE_AGNOSTIC
  // This is added to ensure no test is able to access gPageSize prior to initializing Runtime just
  // because a Runtime instance was created (and subsequently destroyed) by another test.
  gPageSize.DisallowAccess();
#endif
}

struct AbortState {
  void Dump(std::ostream& os) const {
    if (gAborting > 1) {
      os << "Runtime aborting --- recursively, so no thread-specific detail!\n";
      DumpRecursiveAbort(os);
      return;
    }
    gAborting++;
    os << "Runtime aborting...\n";
    if (Runtime::Current() == nullptr) {
      os << "(Runtime does not yet exist!)\n";
      DumpNativeStack(os, GetTid(), "  native: ", nullptr);
      return;
    }
    Thread* self = Thread::Current();

    // Dump all threads first and then the aborting thread. While this is counter the logical flow,
    // it improves the chance of relevant data surviving in the Android logs.

    DumpAllThreads(os, self);

    if (self == nullptr) {
      os << "(Aborting thread was not attached to runtime!)\n";
      DumpNativeStack(os, GetTid(), "  native: ", nullptr);
    } else {
      os << "Aborting thread:\n";
      if (Locks::mutator_lock_->IsExclusiveHeld(self) || Locks::mutator_lock_->IsSharedHeld(self)) {
        DumpThread(os, self);
      } else {
        if (Locks::mutator_lock_->SharedTryLock(self)) {
          DumpThread(os, self);
          Locks::mutator_lock_->SharedUnlock(self);
        }
      }
    }
  }

  // No thread-safety analysis as we do explicitly test for holding the mutator lock.
  void DumpThread(std::ostream& os, Thread* self) const NO_THREAD_SAFETY_ANALYSIS {
    DCHECK(Locks::mutator_lock_->IsExclusiveHeld(self) || Locks::mutator_lock_->IsSharedHeld(self));
    self->Dump(os);
    if (self->IsExceptionPending()) {
      mirror::Throwable* exception = self->GetException();
      os << "Pending exception " << exception->Dump();
    }
  }

  void DumpAllThreads(std::ostream& os, Thread* self) const {
    Runtime* runtime = Runtime::Current();
    if (runtime != nullptr) {
      ThreadList* thread_list = runtime->GetThreadList();
      if (thread_list != nullptr) {
        // Dump requires ThreadListLock and ThreadSuspendCountLock to not be held (they will be
        // grabbed).
        // TODO(b/134167395): Change Dump to work with the locks held, and have a loop with timeout
        //                    acquiring the locks.
        bool tll_already_held = Locks::thread_list_lock_->IsExclusiveHeld(self);
        bool tscl_already_held = Locks::thread_suspend_count_lock_->IsExclusiveHeld(self);
        if (tll_already_held || tscl_already_held) {
          os << "Skipping all-threads dump as locks are held:"
             << (tll_already_held ? "" : " thread_list_lock")
             << (tscl_already_held ? "" : " thread_suspend_count_lock")
             << "\n";
          return;
        }
        bool ml_already_exlusively_held = Locks::mutator_lock_->IsExclusiveHeld(self);
        if (ml_already_exlusively_held) {
          os << "Skipping all-threads dump as mutator lock is exclusively held.";
          return;
        }
        bool ml_already_held = Locks::mutator_lock_->IsSharedHeld(self);
        if (!ml_already_held) {
          os << "Dumping all threads without mutator lock held\n";
        }
        os << "All threads:\n";
        thread_list->Dump(os);
      }
    }
  }

  // For recursive aborts.
  void DumpRecursiveAbort(std::ostream& os) const NO_THREAD_SAFETY_ANALYSIS {
    // The only thing we'll attempt is dumping the native stack of the current thread. We will only
    // try this if we haven't exceeded an arbitrary amount of recursions, to recover and actually
    // die.
    // Note: as we're using a global counter for the recursive abort detection, there is a potential
    //       race here and it is not OK to just print when the counter is "2" (one from
    //       Runtime::Abort(), one from previous Dump() call). Use a number that seems large enough.
    static constexpr size_t kOnlyPrintWhenRecursionLessThan = 100u;
    if (gAborting < kOnlyPrintWhenRecursionLessThan) {
      gAborting++;
      DumpNativeStack(os, GetTid());
    }
  }
};

void Runtime::SetAbortMessage(const char* msg) {
  auto old_value = gAborting.fetch_add(1);  // set before taking any locks

  // Only set the first abort message.
  if (old_value == 0) {
#ifdef ART_TARGET_ANDROID
    android_set_abort_message(msg);
#endif
    // Set the runtime fault message in case our unexpected-signal code will run.
    Runtime* current = Runtime::Current();
    if (current != nullptr) {
      current->SetFaultMessage(msg);
    }
  }
}

void Runtime::Abort(const char* msg) {
  SetAbortMessage(msg);

  // May be coming from an unattached thread.
  if (Thread::Current() == nullptr) {
    Runtime* current = Runtime::Current();
    if (current != nullptr && current->IsStarted() && !current->IsShuttingDownUnsafe()) {
      // We do not flag this to the unexpected-signal handler so that that may dump the stack.
      abort();
      UNREACHABLE();
    }
  }

  {
    // Ensure that we don't have multiple threads trying to abort at once,
    // which would result in significantly worse diagnostics.
    ScopedThreadStateChange tsc(Thread::Current(), ThreadState::kNativeForAbort);
    Locks::abort_lock_->ExclusiveLock(Thread::Current());
  }

  // Get any pending output out of the way.
  fflush(nullptr);

  // Many people have difficulty distinguish aborts from crashes,
  // so be explicit.
  // Note: use cerr on the host to print log lines immediately, so we get at least some output
  //       in case of recursive aborts. We lose annotation with the source file and line number
  //       here, which is a minor issue. The same is significantly more complicated on device,
  //       which is why we ignore the issue there.
  AbortState state;
  if (kIsTargetBuild) {
    LOG(FATAL_WITHOUT_ABORT) << Dumpable<AbortState>(state);
  } else {
    std::cerr << Dumpable<AbortState>(state);
  }

  // Sometimes we dump long messages, and the Android abort message only retains the first line.
  // In those cases, just log the message again, to avoid logcat limits.
  if (msg != nullptr && strchr(msg, '\n') != nullptr) {
    LOG(FATAL_WITHOUT_ABORT) << msg;
  }

  FlagRuntimeAbort();

  // Call the abort hook if we have one.
  if (Runtime::Current() != nullptr && Runtime::Current()->abort_ != nullptr) {
    LOG(FATAL_WITHOUT_ABORT) << "Calling abort hook...";
    Runtime::Current()->abort_();
    // notreached
    LOG(FATAL_WITHOUT_ABORT) << "Unexpectedly returned from abort hook!";
  }

  abort();
  // notreached
}

/**
 * Update entrypoints of methods before the first fork. This
 * helps sharing pages where ArtMethods are allocated between the zygote and
 * forked apps.
 */
class UpdateMethodsPreFirstForkVisitor : public ClassVisitor {
 public:
  explicit UpdateMethodsPreFirstForkVisitor(ClassLinker* class_linker)
      : class_linker_(class_linker),
        can_use_nterp_(interpreter::CanRuntimeUseNterp()) {}

  bool operator()(ObjPtr<mirror::Class> klass) override REQUIRES_SHARED(Locks::mutator_lock_) {
    bool is_initialized = klass->IsVisiblyInitialized();
    for (ArtMethod& method : klass->GetDeclaredMethods(kRuntimePointerSize)) {
      if (!is_initialized && method.NeedsClinitCheckBeforeCall() && can_use_nterp_) {
        const void* existing = method.GetEntryPointFromQuickCompiledCode();
        if (class_linker_->IsQuickResolutionStub(existing) && CanMethodUseNterp(&method)) {
          method.SetEntryPointFromQuickCompiledCode(interpreter::GetNterpWithClinitEntryPoint());
        }
      }
    }
    return true;
  }

 private:
  ClassLinker* const class_linker_;
  const bool can_use_nterp_;

  DISALLOW_COPY_AND_ASSIGN(UpdateMethodsPreFirstForkVisitor);
};

// Wait until the kernel thinks we are single-threaded again.
static void WaitUntilSingleThreaded() {
#if defined(__linux__)
  // Read num_threads field from /proc/self/stat, avoiding higher-level IO libraries that may
  // break atomicity of the read.
  static constexpr size_t kNumTries = 2000;
  static constexpr size_t kNumThreadsIndex = 20;
  static constexpr size_t BUF_SIZE = 500;
  static constexpr size_t BUF_PRINT_SIZE = 150;  // Only log this much on failure to limit length.
  static_assert(BUF_SIZE > BUF_PRINT_SIZE);
  char buf[BUF_SIZE];
  size_t bytes_read = 0;
  uint64_t millis = 0;
  for (size_t tries = 0; tries < kNumTries; ++tries) {
    bytes_read = GetOsThreadStat(getpid(), buf, BUF_SIZE);
    CHECK_NE(bytes_read, 0u);
    size_t pos = 0;
    while (pos < bytes_read && buf[pos++] != ')') {}
    ++pos;
    // We're now positioned at the beginning of the third field. Don't count blanks embedded in
    // second (command) field.
    size_t blanks_seen = 2;
    while (pos < bytes_read && blanks_seen < kNumThreadsIndex - 1) {
      if (buf[pos++] == ' ') {
        ++blanks_seen;
      }
    }
    CHECK(pos < bytes_read - 2);
    // pos is first character of num_threads field.
    CHECK_EQ(buf[pos + 1], ' ');  // We never have more than single-digit threads here.
    if (buf[pos] == '1') {
      return;  //  num_threads == 1; success.
    }
    if (millis == 0) {
      millis = MilliTime();
    }
    usleep(tries < 10 ? 1000 : 2000);
  }
  buf[std::min(BUF_PRINT_SIZE, bytes_read)] = '\0';  // Truncate buf before printing.
  LOG(ERROR) << "Not single threaded: bytes_read = " << bytes_read << " stat contents = \"" << buf
             << "...\"";
  LOG(ERROR) << "Other threads' abbreviated stats: " << GetOtherThreadOsStats();
  bytes_read = GetOsThreadStat(getpid(), buf, BUF_PRINT_SIZE);
  CHECK_NE(bytes_read, 0u);
  LOG(ERROR) << "After re-read: bytes_read = " << bytes_read << " stat contents = \"" << buf
             << "...\"";
  LOG(FATAL) << "Failed to reach single-threaded state: wait_time = " << MilliTime() - millis;
#else  // Not Linux; shouldn't matter, but this has a high probability of working slowly.
  usleep(20'000);
#endif
}

void Runtime::PreZygoteFork() {
  if (GetJit() != nullptr) {
    GetJit()->PreZygoteFork();
  }
  // All other threads have already been joined, but they may not have finished
  // removing themselves from the thread list. Wait until the other threads have completely
  // finished, and are no longer in the thread list.
  // TODO: Since the threads Unregister() themselves before exiting, the first wait should be
  // unnecessary. But since we're reading from a /proc entry that's concurrently changing, for
  // now we play this as safe as possible.
  ThreadList* tl = GetThreadList();
  {
    Thread* self = Thread::Current();
    MutexLock mu(self, *Locks::thread_list_lock_);
    tl->WaitForUnregisterToComplete(self);
    if (kIsDebugBuild) {
      auto list = tl->GetList();
      if (list.size() != 1) {
        for (Thread* t : list) {
          std::string name;
          t->GetThreadName(name);
          LOG(ERROR) << "Remaining pre-fork thread: " << name;
        }
      }
    }
    CHECK_EQ(tl->Size(), 1u);
    // And then wait until the kernel thinks the threads are gone.
    WaitUntilSingleThreaded();
  }

  if (!heap_->HasZygoteSpace()) {
    Thread* self = Thread::Current();
    // This is the first fork. Update ArtMethods in the boot classpath now to
    // avoid having forked apps dirty the memory.

    // Ensure we call FixupStaticTrampolines on all methods that are
    // initialized.
    class_linker_->MakeInitializedClassesVisiblyInitialized(self, /*wait=*/ true);

    ScopedObjectAccess soa(self);
    UpdateMethodsPreFirstForkVisitor visitor(class_linker_);
    class_linker_->VisitClasses(&visitor);
  }
  heap_->PreZygoteFork();
  PreZygoteForkNativeBridge();
}

void Runtime::PostZygoteFork() {
  jit::Jit* jit = GetJit();
  if (jit != nullptr) {
    jit->PostZygoteFork();
    // Ensure that the threads in the JIT pool have been created with the right
    // priority.
    if (kIsDebugBuild && jit->GetThreadPool() != nullptr) {
      jit->GetThreadPool()->CheckPthreadPriority(
          IsZygote() ? jit->GetZygoteThreadPoolPthreadPriority()
                     : jit->GetThreadPoolPthreadPriority());
    }
  }
  // Reset all stats.
  ResetStats(0xFFFFFFFF);
}

void Runtime::CallExitHook(jint status) {
  if (exit_ != nullptr) {
    ScopedThreadStateChange tsc(Thread::Current(), ThreadState::kNative);
    exit_(status);
    LOG(WARNING) << "Exit hook returned instead of exiting!";
  }
}

void Runtime::SweepSystemWeaks(IsMarkedVisitor* visitor) {
  // Userfaultfd compaction updates weak intern-table page-by-page via
  // LinearAlloc.
  bool in_uffd_compaction = GetHeap()->IsPerformingUffdCompaction();
  if (!in_uffd_compaction) {
    GetInternTable()->SweepInternTableWeaks(visitor);
  }
  GetMonitorList()->SweepMonitorList(visitor);
  GetJavaVM()->SweepJniWeakGlobals(visitor);
  GetHeap()->SweepAllocationRecords(visitor);
  // Sweep JIT tables only if the GC is moving as in other cases the entries are
  // not updated.
  if (GetJit() != nullptr && GetHeap()->IsMovingGc()) {
    auto* gc = static_cast<gc::collector::GarbageCollector*>(visitor);
    if (in_uffd_compaction) {
      gc->GetTimings()->StartTiming("SweepJitCodeCache");
    }
    // Visit JIT literal tables. Objects in these tables are classes and strings
    // and only classes can be affected by class unloading. The strings always
    // stay alive as they are strongly interned.
    // TODO: Move this closer to CleanupClassLoaders, to avoid blocking weak accesses
    // from mutators. See b/32167580.
    GetJit()->GetCodeCache()->SweepRootTables(visitor);
    if (in_uffd_compaction) {
      gc->GetTimings()->EndTiming();
    }
  }

  // All other generic system-weak holders.
  for (gc::AbstractSystemWeakHolder* holder : system_weak_holders_) {
    holder->Sweep(visitor);
  }
}

bool Runtime::ParseOptions(const RuntimeOptions& raw_options,
                           bool ignore_unrecognized,
                           RuntimeArgumentMap* runtime_options) {
  Locks::Init();
  InitLogging(/* argv= */ nullptr, Abort);  // Calls Locks::Init() as a side effect.
  bool parsed = ParsedOptions::Parse(raw_options, ignore_unrecognized, runtime_options);
  if (!parsed) {
    LOG(ERROR) << "Failed to parse options";
    return false;
  }
  return true;
}

// Callback to check whether it is safe to call Abort (e.g., to use a call to
// LOG(FATAL)).  It is only safe to call Abort if the runtime has been created,
// properly initialized, and has not shut down.
static bool IsSafeToCallAbort() NO_THREAD_SAFETY_ANALYSIS {
  Runtime* runtime = Runtime::Current();
  return runtime != nullptr && runtime->IsStarted() && !runtime->IsShuttingDownLocked();
}

void Runtime::AddGeneratedCodeRange(const void* start, size_t size) {
  if (HandlesSignalsInCompiledCode()) {
    fault_manager.AddGeneratedCodeRange(start, size);
  }
}

void Runtime::RemoveGeneratedCodeRange(const void* start, size_t size) {
  if (HandlesSignalsInCompiledCode()) {
    fault_manager.RemoveGeneratedCodeRange(start, size);
  }
}

bool Runtime::Create(RuntimeArgumentMap&& runtime_options) {
  // TODO: acquire a static mutex on Runtime to avoid racing.
  if (Runtime::instance_ != nullptr) {
    return false;
  }
  instance_ = new Runtime;
  Locks::SetClientCallback(IsSafeToCallAbort);
  if (!instance_->Init(std::move(runtime_options))) {
    // TODO: Currently deleting the instance will abort the runtime on destruction. Now This will
    // leak memory, instead. Fix the destructor. b/19100793.
    // delete instance_;
    instance_ = nullptr;
    return false;
  }
  return true;
}

bool Runtime::Create(const RuntimeOptions& raw_options, bool ignore_unrecognized) {
  RuntimeArgumentMap runtime_options;
  return ParseOptions(raw_options, ignore_unrecognized, &runtime_options) &&
      Create(std::move(runtime_options));
}

static jobject CreateSystemClassLoader(Runtime* runtime) {
  if (runtime->IsAotCompiler() && !runtime->GetCompilerCallbacks()->IsBootImage()) {
    return nullptr;
  }

  ScopedObjectAccess soa(Thread::Current());
  ClassLinker* cl = runtime->GetClassLinker();
  auto pointer_size = cl->GetImagePointerSize();

  ObjPtr<mirror::Class> class_loader_class = GetClassRoot<mirror::ClassLoader>(cl);
  DCHECK(class_loader_class->IsInitialized());  // Class roots have been initialized.

  ArtMethod* getSystemClassLoader = class_loader_class->FindClassMethod(
      "getSystemClassLoader", "()Ljava/lang/ClassLoader;", pointer_size);
  CHECK(getSystemClassLoader != nullptr);
  CHECK(getSystemClassLoader->IsStatic());

  ObjPtr<mirror::Object> system_class_loader = getSystemClassLoader->InvokeStatic<'L'>(soa.Self());
  CHECK(system_class_loader != nullptr)
      << (soa.Self()->IsExceptionPending() ? soa.Self()->GetException()->Dump() : "<null>");

  ScopedAssertNoThreadSuspension sants(__FUNCTION__);
  jobject g_system_class_loader =
      runtime->GetJavaVM()->AddGlobalRef(soa.Self(), system_class_loader);
  soa.Self()->SetClassLoaderOverride(g_system_class_loader);

  ObjPtr<mirror::Class> thread_class = WellKnownClasses::java_lang_Thread.Get();
  ArtField* contextClassLoader =
      thread_class->FindDeclaredInstanceField("contextClassLoader", "Ljava/lang/ClassLoader;");
  CHECK(contextClassLoader != nullptr);

  // We can't run in a transaction yet.
  contextClassLoader->SetObject<false>(soa.Self()->GetPeer(), system_class_loader);

  return g_system_class_loader;
}

std::string Runtime::GetCompilerExecutable() const {
  if (!compiler_executable_.empty()) {
    return compiler_executable_;
  }
  std::string compiler_executable = GetArtBinDir() + "/dex2oat";
  if (kIsDebugBuild) {
    compiler_executable += 'd';
  }
  if (kIsTargetBuild) {
    compiler_executable += Is64BitInstructionSet(kRuntimeISA) ? "64" : "32";
  }
  return compiler_executable;
}

void Runtime::RunRootClinits(Thread* self) {
  class_linker_->RunRootClinits(self);

  GcRoot<mirror::Throwable>* exceptions[] = {
      &pre_allocated_OutOfMemoryError_when_throwing_exception_,
      // &pre_allocated_OutOfMemoryError_when_throwing_oome_,             // Same class as above.
      // &pre_allocated_OutOfMemoryError_when_handling_stack_overflow_,   // Same class as above.
      &pre_allocated_NoClassDefFoundError_,
  };
  for (GcRoot<mirror::Throwable>* exception : exceptions) {
    StackHandleScope<1> hs(self);
    Handle<mirror::Class> klass = hs.NewHandle<mirror::Class>(exception->Read()->GetClass());
    class_linker_->EnsureInitialized(self, klass, true, true);
    self->AssertNoPendingException();
  }
}

bool Runtime::Start() {
  VLOG(startup) << "Runtime::Start entering";

  CHECK(!no_sig_chain_) << "A started runtime should have sig chain enabled";

  // If a debug host build, disable ptrace restriction for debugging and test timeout thread dump.
  // Only 64-bit as prctl() may fail in 32 bit userspace on a 64-bit kernel.
#if defined(__linux__) && !defined(ART_TARGET_ANDROID) && defined(__x86_64__)
  if (kIsDebugBuild) {
    if (prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY) != 0) {
      PLOG(WARNING) << "Failed setting PR_SET_PTRACER to PR_SET_PTRACER_ANY";
    }
  }
#endif

  // Restore main thread state to kNative as expected by native code.
  Thread* self = Thread::Current();

  started_ = true;

  // Before running any clinit, set up the native methods provided by the runtime itself.
  RegisterRuntimeNativeMethods(self->GetJniEnv());

  class_linker_->RunEarlyRootClinits(self);
  InitializeIntrinsics();

  self->TransitionFromRunnableToSuspended(ThreadState::kNative);

  // InitNativeMethods needs to be after started_ so that the classes
  // it touches will have methods linked to the oat file if necessary.
  {
    ScopedTrace trace2("InitNativeMethods");
    InitNativeMethods();
  }

  // InitializeCorePlatformApiPrivateFields() needs to be called after well known class
  // initializtion in InitNativeMethods().
  art::hiddenapi::InitializeCorePlatformApiPrivateFields();

  // Initialize well known thread group values that may be accessed threads while attaching.
  InitThreadGroups(self);

  Thread::FinishStartup();

  // Create the JIT either if we have to use JIT compilation or save profiling info. This is
  // done after FinishStartup as the JIT pool needs Java thread peers, which require the main
  // ThreadGroup to exist.
  //
  // TODO(calin): We use the JIT class as a proxy for JIT compilation and for
  // recoding profiles. Maybe we should consider changing the name to be more clear it's
  // not only about compiling. b/28295073.
  if (jit_options_->UseJitCompilation() || jit_options_->GetSaveProfilingInfo()) {
    CreateJit();
#ifdef ADDRESS_SANITIZER
    // (b/238730394): In older implementations of sanitizer + glibc there is a race between
    // pthread_create and dlopen that could cause a deadlock. pthread_create interceptor in ASAN
    // uses dl_pthread_iterator with a callback that could request a dl_load_lock via call to
    // __tls_get_addr [1]. dl_pthread_iterate would already hold dl_load_lock so this could cause a
    // deadlock. __tls_get_addr needs a dl_load_lock only when there is a dlopen happening in
    // parallel. As a workaround we wait for the pthread_create (i.e JIT thread pool creation) to
    // finish before going to the next phase. Creating a system class loader could need a dlopen so
    // we wait here till threads are initialized.
    // [1] https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/sanitizer_common/sanitizer_linux_libcdep.cpp#L408
    // See this for more context: https://reviews.llvm.org/D98926
    // TODO(b/238730394): Revisit this workaround once we migrate to musl libc.
    if (jit_ != nullptr) {
      jit_->GetThreadPool()->WaitForWorkersToBeCreated();
    }
#endif
  }

  // Send the start phase event. We have to wait till here as this is when the main thread peer
  // has just been generated, important root clinits have been run and JNI is completely functional.
  {
    ScopedObjectAccess soa(self);
    callbacks_->NextRuntimePhase(RuntimePhaseCallback::RuntimePhase::kStart);
  }

  system_class_loader_ = CreateSystemClassLoader(this);

  if (!is_zygote_) {
    if (is_native_bridge_loaded_) {
      PreInitializeNativeBridge(".");
    }
    NativeBridgeAction action = force_native_bridge_
        ? NativeBridgeAction::kInitialize
        : NativeBridgeAction::kUnload;
    InitNonZygoteOrPostFork(self->GetJniEnv(),
                            /* is_system_server= */ false,
                            /* is_child_zygote= */ false,
                            action,
                            GetInstructionSetString(kRuntimeISA));
  }

  {
    ScopedObjectAccess soa(self);
    StartDaemonThreads();
    self->GetJniEnv()->AssertLocalsEmpty();

    // Send the initialized phase event. Send it after starting the Daemon threads so that agents
    // cannot delay the daemon threads from starting forever.
    callbacks_->NextRuntimePhase(RuntimePhaseCallback::RuntimePhase::kInit);
    self->GetJniEnv()->AssertLocalsEmpty();
  }

  VLOG(startup) << "Runtime::Start exiting";
  finished_starting_ = true;

  if (trace_config_.get() != nullptr && trace_config_->trace_file != "") {
    ScopedThreadStateChange tsc(self, ThreadState::kWaitingForMethodTracingStart);
    int flags = 0;
    if (trace_config_->clock_source == TraceClockSource::kDual) {
      flags = Trace::TraceFlag::kTraceClockSourceWallClock |
              Trace::TraceFlag::kTraceClockSourceThreadCpu;
    } else if (trace_config_->clock_source == TraceClockSource::kWall) {
      flags = Trace::TraceFlag::kTraceClockSourceWallClock;
    } else if (TraceClockSource::kThreadCpu == trace_config_->clock_source) {
      flags = Trace::TraceFlag::kTraceClockSourceThreadCpu;
    } else {
      LOG(ERROR) << "Unexpected clock source";
    }
    Trace::Start(trace_config_->trace_file.c_str(),
                 static_cast<int>(trace_config_->trace_file_size),
                 flags,
                 trace_config_->trace_output_mode,
                 trace_config_->trace_mode,
                 0);
  }

  // In case we have a profile path passed as a command line argument,
  // register the current class path for profiling now. Note that we cannot do
  // this before we create the JIT and having it here is the most convenient way.
  // This is used when testing profiles with dalvikvm command as there is no
  // framework to register the dex files for profiling.
  if (jit_.get() != nullptr && jit_options_->GetSaveProfilingInfo() &&
      !jit_options_->GetProfileSaverOptions().GetProfilePath().empty()) {
    std::vector<std::string> dex_filenames;
    Split(class_path_string_, ':', &dex_filenames);

    // We pass "" as the package name because at this point we don't know it. It could be the
    // Zygote or it could be a dalvikvm cmd line execution. The package name will be re-set during
    // post-fork or during RegisterAppInfo.
    //
    // Also, it's ok to pass "" to the ref profile filename. It indicates we don't have
    // a reference profile.
    RegisterAppInfo(
        /*package_name=*/ "",
        dex_filenames,
        jit_options_->GetProfileSaverOptions().GetProfilePath(),
        /*ref_profile_filename=*/ "",
        kVMRuntimePrimaryApk);
  }
  // Add a concurrent-gc task after runtime has started if we are in continuous-gc mode.
  if (heap_->InContinuousGCMode()) {
    heap_->RequestConcurrentGC(self, gc::kGcCauseBackground, false, heap_->GetCurrentGcNum());
  }
  return true;
}

void Runtime::EndThreadBirth() REQUIRES(Locks::runtime_shutdown_lock_) {
  DCHECK_GT(threads_being_born_, 0U);
  threads_being_born_--;
  if (shutting_down_started_ && threads_being_born_ == 0) {
    shutdown_cond_->Broadcast(Thread::Current());
  }
}

void Runtime::InitNonZygoteOrPostFork(
    JNIEnv* env,
    bool is_system_server,
    // This is true when we are initializing a child-zygote. It requires
    // native bridge initialization to be able to run guest native code in
    // doPreload().
    bool is_child_zygote,
    NativeBridgeAction action,
    const char* isa,
    bool profile_system_server) {
  if (is_native_bridge_loaded_) {
    switch (action) {
      case NativeBridgeAction::kUnload:
        UnloadNativeBridge();
        is_native_bridge_loaded_ = false;
        break;
      case NativeBridgeAction::kInitialize:
        InitializeNativeBridge(env, isa);
        break;
    }
  }

  if (is_child_zygote) {
    // If creating a child-zygote we only initialize native bridge. The rest of
    // runtime post-fork logic would spin up threads for Binder and JDWP.
    // Instead, the Java side of the child process will call a static main in a
    // class specified by the parent.
    return;
  }

  DCHECK(!IsZygote());

  if (is_system_server) {
    // Register the system server code paths.
    // TODO: Ideally this should be done by the VMRuntime#RegisterAppInfo. However, right now
    // the method is only called when we set up the profile. It should be called all the time
    // (simillar to the apps). Once that's done this manual registration can be removed.
    const char* system_server_classpath = getenv("SYSTEMSERVERCLASSPATH");
    if (system_server_classpath == nullptr || (strlen(system_server_classpath) == 0)) {
      LOG(WARNING) << "System server class path not set";
    } else {
      std::vector<std::string> jars = android::base::Split(system_server_classpath, ":");
      app_info_.RegisterAppInfo("android",
                                jars,
                                /*profile_output_filename=*/ "",
                                /*ref_profile_filename=*/ "",
                                AppInfo::CodeType::kPrimaryApk);
    }

    // Set the system server package name to "android".
    // This is used to tell the difference between samples provided by system server
    // and samples generated by other apps when processing boot image profiles.
    SetProcessPackageName("android");
    if (profile_system_server) {
      jit_options_->SetWaitForJitNotificationsToSaveProfile(false);
      VLOG(profiler) << "Enabling system server profiles";
    }
  }

  // Create the thread pool for loading app images.
  // Avoid creating the runtime thread pool for system server since it will not be used and would
  // waste memory.
  if (!is_system_server &&
      android::base::GetBoolProperty("dalvik.vm.parallel-image-loading", false)) {
    ScopedTrace timing("CreateThreadPool");
    constexpr size_t kStackSize = 64 * KB;
    constexpr size_t kMaxRuntimeWorkers = 4u;
    const size_t num_workers =
        std::min(static_cast<size_t>(std::thread::hardware_concurrency()), kMaxRuntimeWorkers);
    MutexLock mu(Thread::Current(), *Locks::runtime_thread_pool_lock_);
    CHECK(thread_pool_ == nullptr);
    thread_pool_.reset(
        ThreadPool::Create("Runtime", num_workers, /*create_peers=*/false, kStackSize));
    thread_pool_->StartWorkers(Thread::Current());
  }

  // Reset the gc performance data and metrics at zygote fork so that the events from
  // before fork aren't attributed to an app.
  heap_->ResetGcPerformanceInfo();
  GetMetrics()->Reset();

  if (AreMetricsInitialized()) {
    // Now that we know if we are an app or system server, reload the metrics reporter config
    // in case there are any difference.
    metrics::ReportingConfig metrics_config =
        metrics::ReportingConfig::FromFlags(is_system_server);

    metrics_reporter_->ReloadConfig(metrics_config);

    metrics::SessionData session_data{metrics::SessionData::CreateDefault()};
    // Start the session id from 1 to avoid clashes with the default value.
    // (better for debugability)
    session_data.session_id = GetRandomNumber<int64_t>(1, std::numeric_limits<int64_t>::max());
    // TODO: set session_data.compilation_reason and session_data.compiler_filter
    metrics_reporter_->MaybeStartBackgroundThread(session_data);
    // Also notify about any updates to the app info.
    metrics_reporter_->NotifyAppInfoUpdated(&app_info_);
  }

  StartSignalCatcher();

  ScopedObjectAccess soa(Thread::Current());
  if (IsPerfettoHprofEnabled() &&
      (Dbg::IsJdwpAllowed() || IsProfileable() || IsProfileableFromShell() || IsJavaDebuggable() ||
       Runtime::Current()->IsSystemServer())) {
    std::string err;
    ScopedTrace tr("perfetto_hprof init.");
    ScopedThreadSuspension sts(Thread::Current(), ThreadState::kNative);
    if (!EnsurePerfettoPlugin(&err)) {
      LOG(WARNING) << "Failed to load perfetto_hprof: " << err;
    }
  }
  if (IsPerfettoJavaHeapStackProfEnabled() &&
      (Dbg::IsJdwpAllowed() || IsProfileable() || IsProfileableFromShell() || IsJavaDebuggable() ||
       Runtime::Current()->IsSystemServer())) {
    // Marker used for dev tracing similar to above markers.
    ScopedTrace tr("perfetto_javaheapprof init.");
  }
  if (Runtime::Current()->IsSystemServer()) {
    std::string err;
    ScopedTrace tr("odrefresh and device stats logging");
    ScopedThreadSuspension sts(Thread::Current(), ThreadState::kNative);
    // Report stats if available. This should be moved into ART Services when they are ready.
    if (!odrefresh::UploadStatsIfAvailable(&err)) {
      LOG(WARNING) << "Failed to upload odrefresh metrics: " << err;
    }
  }

  if (LIKELY(automatically_set_jni_ids_indirection_) && CanSetJniIdType()) {
    if (IsJavaDebuggable()) {
      SetJniIdType(JniIdType::kIndices);
    } else {
      SetJniIdType(JniIdType::kPointer);
    }
  }
  ATraceIntegerValue(
      "profilebootclasspath",
      static_cast<int>(jit_options_->GetProfileSaverOptions().GetProfileBootClassPath()));
  // Start the JDWP thread. If the command-line debugger flags specified "suspend=y",
  // this will pause the runtime (in the internal debugger implementation), so we probably want
  // this to come last.
  GetRuntimeCallbacks()->StartDebugger();
}

void Runtime::StartSignalCatcher() {
  if (!is_zygote_) {
    signal_catcher_ = new SignalCatcher();
  }
}

bool Runtime::IsShuttingDown(Thread* self) {
  MutexLock mu(self, *Locks::runtime_shutdown_lock_);
  return IsShuttingDownLocked();
}

void Runtime::StartDaemonThreads() {
  ScopedTrace trace(__FUNCTION__);
  VLOG(startup) << "Runtime::StartDaemonThreads entering";

  Thread* self = Thread::Current();

  DCHECK_EQ(self->GetState(), ThreadState::kRunnable);

  WellKnownClasses::java_lang_Daemons_start->InvokeStatic<'V'>(self);
  if (UNLIKELY(self->IsExceptionPending())) {
    LOG(FATAL) << "Error starting java.lang.Daemons: " << self->GetException()->Dump();
  }

  VLOG(startup) << "Runtime::StartDaemonThreads exiting";
}

static size_t OpenBootDexFiles(ArrayRef<const std::string> dex_filenames,
                               ArrayRef<const std::string> dex_locations,
                               ArrayRef<File> dex_files,
                               std::vector<std::unique_ptr<const DexFile>>* out_dex_files) {
  DCHECK(out_dex_files != nullptr) << "OpenDexFiles: out-param is nullptr";
  size_t failure_count = 0;
  for (size_t i = 0; i < dex_filenames.size(); i++) {
    const char* dex_filename = dex_filenames[i].c_str();
    const char* dex_location = dex_locations[i].c_str();
    File noFile;
    File* file = i < dex_files.size() ? &dex_files[i] : &noFile;
    static constexpr bool kVerifyChecksum = true;
    std::string error_msg;
    if (!OS::FileExists(dex_filename) && file->IsValid()) {
      LOG(WARNING) << "Skipping non-existent dex file '" << dex_filename << "'";
      continue;
    }
    bool verify = Runtime::Current()->IsVerificationEnabled();
    ArtDexFileLoader dex_file_loader(dex_filename, file, dex_location);
    if (!dex_file_loader.Open(verify, kVerifyChecksum, &error_msg, out_dex_files)) {
      LOG(WARNING) << "Failed to open .dex from file '" << dex_filename << "' / fd " << file->Fd()
                   << ": " << error_msg;
      ++failure_count;
    }
    if (file->IsValid()) {
      bool close_ok = file->Close();
      DCHECK(close_ok) << dex_filename;
    }
  }
  return failure_count;
}

void Runtime::SetSentinel(ObjPtr<mirror::Object> sentinel) {
  CHECK(sentinel_.Read() == nullptr);
  CHECK(sentinel != nullptr);
  CHECK(!heap_->IsMovableObject(sentinel));
  sentinel_ = GcRoot<mirror::Object>(sentinel);
}

GcRoot<mirror::Object> Runtime::GetSentinel() {
  return sentinel_;
}

static inline void CreatePreAllocatedException(Thread* self,
                                               Runtime* runtime,
                                               GcRoot<mirror::Throwable>* exception,
                                               const char* exception_class_descriptor,
                                               const char* msg)
    REQUIRES_SHARED(Locks::mutator_lock_) {
  DCHECK_EQ(self, Thread::Current());
  ClassLinker* class_linker = runtime->GetClassLinker();
  // Allocate an object without initializing the class to allow non-trivial Throwable.<clinit>().
  ObjPtr<mirror::Class> klass = class_linker->FindSystemClass(self, exception_class_descriptor);
  CHECK(klass != nullptr);
  gc::AllocatorType allocator_type = runtime->GetHeap()->GetCurrentAllocator();
  ObjPtr<mirror::Throwable> exception_object = ObjPtr<mirror::Throwable>::DownCast(
      klass->Alloc(self, allocator_type));
  CHECK(exception_object != nullptr);
  *exception = GcRoot<mirror::Throwable>(exception_object);
  // Initialize the "detailMessage" field.
  ObjPtr<mirror::String> message = mirror::String::AllocFromModifiedUtf8(self, msg);
  CHECK(message != nullptr);
  ObjPtr<mirror::Class> throwable = GetClassRoot<mirror::Throwable>(class_linker);
  ArtField* detailMessageField =
      throwable->FindDeclaredInstanceField("detailMessage", "Ljava/lang/String;");
  CHECK(detailMessageField != nullptr);
  detailMessageField->SetObject</* kTransactionActive= */ false>(exception->Read(), message);
}

inline void Runtime::CreatePreAllocatedExceptions(Thread* self) {
  // Pre-allocate an OutOfMemoryError for the case when we fail to
  // allocate the exception to be thrown.
  CreatePreAllocatedException(self,
                              this,
                              &pre_allocated_OutOfMemoryError_when_throwing_exception_,
                              "Ljava/lang/OutOfMemoryError;",
                              "OutOfMemoryError thrown while trying to throw an exception; "
                              "no stack trace available");
  // Pre-allocate an OutOfMemoryError for the double-OOME case.
  CreatePreAllocatedException(self,
                              this,
                              &pre_allocated_OutOfMemoryError_when_throwing_oome_,
                              "Ljava/lang/OutOfMemoryError;",
                              "OutOfMemoryError thrown while trying to throw OutOfMemoryError; "
                              "no stack trace available");
  // Pre-allocate an OutOfMemoryError for the case when we fail to
  // allocate while handling a stack overflow.
  CreatePreAllocatedException(self,
                              this,
                              &pre_allocated_OutOfMemoryError_when_handling_stack_overflow_,
                              "Ljava/lang/OutOfMemoryError;",
                              "OutOfMemoryError thrown while trying to handle a stack overflow; "
                              "no stack trace available");
  // Pre-allocate a NoClassDefFoundError for the common case of failing to find a system class
  // ahead of checking the application's class loader.
  CreatePreAllocatedException(self,
                              this,
                              &pre_allocated_NoClassDefFoundError_,
                              "Ljava/lang/NoClassDefFoundError;",
                              "Class not found using the boot class loader; "
                              "no stack trace available");
}

std::string Runtime::GetApexVersions(ArrayRef<const std::string> boot_class_path_locations) {
  std::vector<std::string_view> bcp_apexes;
  for (std::string_view jar : boot_class_path_locations) {
    std::string_view apex = ApexNameFromLocation(jar);
    if (!apex.empty()) {
      bcp_apexes.push_back(apex);
    }
  }
  static const char* kApexFileName = "/apex/apex-info-list.xml";
  // Start with empty markers.
  std::string empty_apex_versions(bcp_apexes.size(), '/');
  // When running on host or chroot, we just use empty markers.
  if (!kIsTargetBuild || !OS::FileExists(kApexFileName)) {
    return empty_apex_versions;
  }
#ifdef ART_TARGET_ANDROID
  if (access(kApexFileName, R_OK) != 0) {
    PLOG(WARNING) << "Failed to read " << kApexFileName;
    return empty_apex_versions;
  }
  auto info_list = apex::readApexInfoList(kApexFileName);
  if (!info_list.has_value()) {
    LOG(WARNING) << "Failed to parse " << kApexFileName;
    return empty_apex_versions;
  }

  std::string result;
  std::map<std::string_view, const apex::ApexInfo*> apex_infos;
  for (const apex::ApexInfo& info : info_list->getApexInfo()) {
    if (info.getIsActive()) {
      apex_infos.emplace(info.getModuleName(), &info);
    }
  }
  for (const std::string_view& str : bcp_apexes) {
    auto info = apex_infos.find(str);
    if (info == apex_infos.end() || info->second->getIsFactory()) {
      result += '/';
    } else {
      // In case lastUpdateMillis field is populated in apex-info-list.xml, we prefer to use it as
      // version scheme.
      // If the field is missing we fallback to the version code of the APEX. This only happens
      // when apexd cannot stat the APEX, which should be very rare in practice because we would
      // probably have crashed already due to apexd failing to mount the APEX.
      uint64_t version = info->second->hasLastUpdateMillis()
          ? info->second->getLastUpdateMillis()
          : info->second->getVersionCode();
      android::base::StringAppendF(&result, "/%" PRIu64, version);
    }
  }
  return result;
#else
  return empty_apex_versions;  // Not an Android build.
#endif
}

void Runtime::InitializeApexVersions() {
  apex_versions_ =
      GetApexVersions(ArrayRef<const std::string>(Runtime::Current()->GetBootClassPathLocations()));
}

std::optional<AssumeValueSignature> Runtime::LookupAssumeValueSignature(ArtField* field) const {
  auto it = assume_value_field_signatures_.find(field);
  return it != assume_value_field_signatures_.end() ? std::optional(*it->second) : std::nullopt;
}

void Runtime::ReloadAllFlags(const std::string& caller) {
  FlagBase::ReloadAllFlags(caller);
}

static std::vector<File> FileFdsToFileObjects(std::vector<int>&& fds) {
  std::vector<File> files;
  files.reserve(fds.size());
  for (int fd : fds) {
    files.push_back(File(fd, /*check_usage=*/false));
  }
  return files;
}

inline static uint64_t GetThreadSuspendTimeout(const RuntimeArgumentMap* runtime_options) {
  auto suspend_timeout_opt = runtime_options->GetOptional(RuntimeArgumentMap::ThreadSuspendTimeout);
  return suspend_timeout_opt.has_value() ?
             suspend_timeout_opt.value().GetNanoseconds() :
             ThreadList::kDefaultThreadSuspendTimeout *
                 android::base::GetIntProperty("ro.hw_timeout_multiplier", 1);
}

bool Runtime::Init(RuntimeArgumentMap&& runtime_options_in) {
  // (b/30160149): protect subprocesses from modifications to LD_LIBRARY_PATH, etc.
  // Take a snapshot of the environment at the time the runtime was created, for use by Exec, etc.
  env_snapshot_.TakeSnapshot();

#ifdef ART_PAGE_SIZE_AGNOSTIC
  gPageSize.AllowAccess();
#endif

  using Opt = RuntimeArgumentMap;
  Opt runtime_options(std::move(runtime_options_in));
  ScopedTrace trace(__FUNCTION__);
  CHECK_EQ(static_cast<size_t>(sysconf(_SC_PAGE_SIZE)), gPageSize);

  // Reload all the flags value (from system properties and device configs).
  ReloadAllFlags(__FUNCTION__);

  deny_art_apex_data_files_ = runtime_options.Exists(Opt::DenyArtApexDataFiles);
  if (deny_art_apex_data_files_) {
    // We will run slower without those files if the system has taken an ART APEX update.
    LOG(WARNING) << "ART APEX data files are untrusted.";
  }

  // Early override for logging output.
  if (runtime_options.Exists(Opt::UseStderrLogger)) {
    android::base::SetLogger(android::base::StderrLogger);
  }

  MemMap::Init();

  verifier_missing_kthrow_fatal_ = runtime_options.GetOrDefault(Opt::VerifierMissingKThrowFatal);
  force_java_zygote_fork_loop_ = runtime_options.GetOrDefault(Opt::ForceJavaZygoteForkLoop);
  perfetto_hprof_enabled_ = runtime_options.GetOrDefault(Opt::PerfettoHprof);
  perfetto_javaheapprof_enabled_ = runtime_options.GetOrDefault(Opt::PerfettoJavaHeapStackProf);

  // Try to reserve a dedicated fault page. This is allocated for clobbered registers and sentinels.
  // If we cannot reserve it, log a warning.
  // Note: We allocate this first to have a good chance of grabbing the page. The address (0xebad..)
  //       is out-of-the-way enough that it should not collide with boot image mapping.
  // Note: Don't request an error message. That will lead to a maps dump in the case of failure,
  //       leading to logspam.
  {
    const uintptr_t sentinel_addr =
        RoundDown(static_cast<uintptr_t>(Context::kBadGprBase), gPageSize);
    protected_fault_page_ = MemMap::MapAnonymous("Sentinel fault page",
                                                 reinterpret_cast<uint8_t*>(sentinel_addr),
                                                 gPageSize,
                                                 PROT_NONE,
                                                 /*low_4gb=*/ true,
                                                 /*reuse=*/ false,
                                                 /*reservation=*/ nullptr,
                                                 /*error_msg=*/ nullptr);
    if (!protected_fault_page_.IsValid()) {
      LOG(WARNING) << "Could not reserve sentinel fault page";
    } else if (reinterpret_cast<uintptr_t>(protected_fault_page_.Begin()) != sentinel_addr) {
      LOG(WARNING) << "Could not reserve sentinel fault page at the right address.";
      protected_fault_page_.Reset();
    }
  }

  VLOG(startup) << "Runtime::Init -verbose:startup enabled";

  oat_file_manager_ = new OatFileManager();

  jni_id_manager_.reset(new jni::JniIdManager());

  Thread::SetSensitiveThreadHook(runtime_options.GetOrDefault(Opt::HookIsSensitiveThread));
  Monitor::Init(runtime_options.GetOrDefault(Opt::LockProfThreshold),
                runtime_options.GetOrDefault(Opt::StackDumpLockProfThreshold));

  image_locations_ = runtime_options.ReleaseOrDefault(Opt::Image);

  SetInstructionSet(runtime_options.GetOrDefault(Opt::ImageInstructionSet));
  boot_class_path_ = runtime_options.ReleaseOrDefault(Opt::BootClassPath);
  boot_class_path_locations_ = runtime_options.ReleaseOrDefault(Opt::BootClassPathLocations);
  DCHECK(boot_class_path_locations_.empty() ||
         boot_class_path_locations_.size() == boot_class_path_.size());
  if (boot_class_path_.empty()) {
    LOG(ERROR) << "Boot classpath is empty";
    return false;
  }

  boot_class_path_files_ =
      FileFdsToFileObjects(runtime_options.ReleaseOrDefault(Opt::BootClassPathFds));
  if (!boot_class_path_files_.empty() && boot_class_path_files_.size() != boot_class_path_.size()) {
    LOG(ERROR) << "Number of FDs specified in -Xbootclasspathfds must match the number of JARs in "
               << "-Xbootclasspath.";
    return false;
  }

  boot_class_path_image_files_ =
      FileFdsToFileObjects(runtime_options.ReleaseOrDefault(Opt::BootClassPathImageFds));
  boot_class_path_vdex_files_ =
      FileFdsToFileObjects(runtime_options.ReleaseOrDefault(Opt::BootClassPathVdexFds));
  boot_class_path_oat_files_ =
      FileFdsToFileObjects(runtime_options.ReleaseOrDefault(Opt::BootClassPathOatFds));
  CHECK(boot_class_path_image_files_.empty() ||
        boot_class_path_image_files_.size() == boot_class_path_.size());
  CHECK(boot_class_path_vdex_files_.empty() ||
        boot_class_path_vdex_files_.size() == boot_class_path_.size());
  CHECK(boot_class_path_oat_files_.empty() ||
        boot_class_path_oat_files_.size() == boot_class_path_.size());

  class_path_string_ = runtime_options.ReleaseOrDefault(Opt::ClassPath);
  properties_ = runtime_options.ReleaseOrDefault(Opt::PropertiesList);

  compiler_callbacks_ = runtime_options.GetOrDefault(Opt::CompilerCallbacksPtr);
  must_relocate_ = runtime_options.GetOrDefault(Opt::Relocate);
  is_zygote_ = runtime_options.Exists(Opt::Zygote);
  is_primary_zygote_ = runtime_options.Exists(Opt::PrimaryZygote);
  is_explicit_gc_disabled_ = runtime_options.Exists(Opt::DisableExplicitGC);
  is_eagerly_release_explicit_gc_disabled_ =
      runtime_options.Exists(Opt::DisableEagerlyReleaseExplicitGC);
  image_dex2oat_enabled_ = runtime_options.GetOrDefault(Opt::ImageDex2Oat);
  dump_native_stack_on_sig_quit_ = runtime_options.GetOrDefault(Opt::DumpNativeStackOnSigQuit);
  allow_in_memory_compilation_ = runtime_options.Exists(Opt::AllowInMemoryCompilation);

  if (is_zygote_ || runtime_options.Exists(Opt::OnlyUseTrustedOatFiles)) {
    oat_file_manager_->SetOnlyUseTrustedOatFiles();
  }

  vfprintf_ = runtime_options.GetOrDefault(Opt::HookVfprintf);
  exit_ = runtime_options.GetOrDefault(Opt::HookExit);
  abort_ = runtime_options.GetOrDefault(Opt::HookAbort);

  default_stack_size_ = runtime_options.GetOrDefault(Opt::StackSize);

  compiler_executable_ = runtime_options.ReleaseOrDefault(Opt::Compiler);
  compiler_options_ = runtime_options.ReleaseOrDefault(Opt::CompilerOptions);
  for (const std::string& option : Runtime::Current()->GetCompilerOptions()) {
    if (option == "--debuggable") {
      SetRuntimeDebugState(RuntimeDebugState::kJavaDebuggableAtInit);
      break;
    }
  }
  image_compiler_options_ = runtime_options.ReleaseOrDefault(Opt::ImageCompilerOptions);

  finalizer_timeout_ms_ = runtime_options.GetOrDefault(Opt::FinalizerTimeoutMs);
  max_spins_before_thin_lock_inflation_ =
      runtime_options.GetOrDefault(Opt::MaxSpinsBeforeThinLockInflation);

  monitor_list_ = new MonitorList;
  monitor_pool_ = MonitorPool::Create();
  thread_list_ = new ThreadList(GetThreadSuspendTimeout(&runtime_options));
  intern_table_ = new InternTable;

  monitor_timeout_enable_ = runtime_options.GetOrDefault(Opt::MonitorTimeoutEnable);
  int monitor_timeout_ms = runtime_options.GetOrDefault(Opt::MonitorTimeout);
  if (monitor_timeout_ms < Monitor::kMonitorTimeoutMinMs) {
    LOG(WARNING) << "Monitor timeout too short: Increasing";
    monitor_timeout_ms = Monitor::kMonitorTimeoutMinMs;
  }
  if (monitor_timeout_ms >= Monitor::kMonitorTimeoutMaxMs) {
    LOG(WARNING) << "Monitor timeout too long: Decreasing";
    monitor_timeout_ms = Monitor::kMonitorTimeoutMaxMs - 1;
  }
  monitor_timeout_ns_ = MsToNs(monitor_timeout_ms);

  verify_ = runtime_options.GetOrDefault(Opt::Verify);

  target_sdk_version_ = runtime_options.GetOrDefault(Opt::TargetSdkVersion);

  // Set hidden API enforcement policy. The checks are disabled by default and
  // we only enable them if:
  // (a) runtime was started with a command line flag that enables the checks, or
  // (b) Zygote forked a new process that is not exempt (see ZygoteHooks).
  hidden_api_policy_ = runtime_options.GetOrDefault(Opt::HiddenApiPolicy);
  DCHECK_IMPLIES(is_zygote_, hidden_api_policy_ == hiddenapi::EnforcementPolicy::kDisabled);

  // Set core platform API enforcement policy. Always enabled if the
  // hiddenapi_platform_enforcement flag is set, otherwise the checks are
  // disabled by default and can be enabled with a command line flag.
  // AndroidRuntime will pass the flag if a system property is set.
  {
    bool always_enable = false;
#ifdef ART_TARGET_ANDROID
    if (hiddenapi::EnableHiddenapiPlatformEnforcement()) {
      always_enable = true;
    }
#endif
    const char* reason;
    if (always_enable) {
      core_platform_api_policy_ = hiddenapi::EnforcementPolicy::kEnabled;
      reason = "from the hiddenapi_platform_enforcement flag and the device API level";
    } else {
      core_platform_api_policy_ = runtime_options.GetOrDefault(Opt::CorePlatformApiPolicy);
      reason = "by runtime option";
    }
    if (core_platform_api_policy_ != hiddenapi::EnforcementPolicy::kDisabled) {
      LOG(INFO) << "Core platform API "
                << (core_platform_api_policy_ == hiddenapi::EnforcementPolicy::kEnabled
                        ? "enforcement"
                        : "reporting")
                << " enabled " << reason;
    }
  }

  // Dex2Oat's Runtime does not need the signal chain or the fault handler
  // and it passes the `NoSigChain` option to `Runtime` to indicate this.
  no_sig_chain_ = runtime_options.Exists(Opt::NoSigChain);
  force_native_bridge_ = runtime_options.Exists(Opt::ForceNativeBridge);

  Split(runtime_options.GetOrDefault(Opt::CpuAbiList), ',', &cpu_abilist_);

  fingerprint_ = runtime_options.ReleaseOrDefault(Opt::Fingerprint);

  if (runtime_options.GetOrDefault(Opt::Interpret)) {
    GetInstrumentation()->ForceInterpretOnly();
  }

  zygote_max_failed_boots_ = runtime_options.GetOrDefault(Opt::ZygoteMaxFailedBoots);
  experimental_flags_ = runtime_options.GetOrDefault(Opt::Experimental);
  is_low_memory_mode_ = runtime_options.Exists(Opt::LowMemoryMode);
  madvise_willneed_total_dex_size_ = runtime_options.GetOrDefault(Opt::MadviseWillNeedVdexFileSize);
  madvise_willneed_odex_filesize_ = runtime_options.GetOrDefault(Opt::MadviseWillNeedOdexFileSize);
  madvise_willneed_art_filesize_ = runtime_options.GetOrDefault(Opt::MadviseWillNeedArtFileSize);

  jni_ids_indirection_ = runtime_options.GetOrDefault(Opt::OpaqueJniIds);
  automatically_set_jni_ids_indirection_ =
      runtime_options.GetOrDefault(Opt::AutoPromoteOpaqueJniIds);

  plugins_ = runtime_options.ReleaseOrDefault(Opt::Plugins);
  agent_specs_ = runtime_options.ReleaseOrDefault(Opt::AgentPath);
  // TODO Add back in -agentlib
  // for (auto lib : runtime_options.ReleaseOrDefault(Opt::AgentLib)) {
  //   agents_.push_back(lib);
  // }

  if (InstructionSetFeatures::IsRuntimeDetectionSupported()) {
    runtime_instruction_set_features_ = InstructionSetFeatures::FromRuntimeDetection();
  }

  float foreground_heap_growth_multiplier;
  if (is_low_memory_mode_ && !runtime_options.Exists(Opt::ForegroundHeapGrowthMultiplier)) {
    // If low memory mode, use 1.0 as the multiplier by default.
    foreground_heap_growth_multiplier = 1.0f;
  } else {
    // Extra added to the default heap growth multiplier for concurrent GC
    // compaction algorithms. This is done for historical reasons.
    // TODO: remove when we revisit heap configurations.
    foreground_heap_growth_multiplier =
        runtime_options.GetOrDefault(Opt::ForegroundHeapGrowthMultiplier) + 1.0f;
  }
  XGcOption xgc_option = runtime_options.GetOrDefault(Opt::GcOption);

  // Generational CC collection is currently only compatible with Baker read barriers.
  bool use_generational_gc = (kUseBakerReadBarrier || gUseUserfaultfd) &&
                             xgc_option.generational_gc && ShouldUseGenerationalGC();

  // Cache the apex versions.
  InitializeApexVersions();

  BackgroundGcOption background_gc =
      gUseReadBarrier ? BackgroundGcOption(gc::kCollectorTypeCCBackground) :
                        (gUseUserfaultfd ? BackgroundGcOption(gc::kCollectorTypeCMCBackground) :
                                           runtime_options.GetOrDefault(Opt::BackgroundGc));

  bool enable_time_based_gc_trigger =
      runtime_options.GetOrDefault(Opt::EnableTimeBasedGcTrigger) &&
      !GetBoolProperty(
          "persist.device_config.runtime_native_boot.force_disable_time_based_gc_trigger", false);

#ifdef ART_USE_SIMULATOR
  if (IsSimulatorMode()) {
    instruction_set_ = kRuntimeQuickCodeISA;
    simulator_container_.reset(new CodeSimulatorContainer(kRuntimeQuickCodeISA));
  }
#endif

  heap_ = new gc::Heap(runtime_options.GetOrDefault(Opt::MemoryInitialSize),
                       runtime_options.GetOrDefault(Opt::HeapGrowthLimit),
                       runtime_options.GetOrDefault(Opt::HeapMinFree),
                       runtime_options.GetOrDefault(Opt::HeapMaxFree),
                       runtime_options.GetOrDefault(Opt::HeapTargetUtilization),
                       enable_time_based_gc_trigger,
                       runtime_options.GetOrDefault(Opt::HeapMemoryGcCostFactor),
                       foreground_heap_growth_multiplier,
                       runtime_options.GetOrDefault(Opt::StopForNativeAllocs),
                       runtime_options.GetOrDefault(Opt::MemoryMaximumSize),
                       runtime_options.GetOrDefault(Opt::NonMovingSpaceCapacity),
                       GetBootClassPath(),
                       GetBootClassPathLocations(),
                       GetBootClassPathFiles(),
                       GetBootClassPathImageFiles(),
                       GetBootClassPathVdexFiles(),
                       GetBootClassPathOatFiles(),
                       image_locations_,
                       instruction_set_,
                       // Override the collector type to CC if the read barrier config.
                       gUseReadBarrier ? gc::kCollectorTypeCC : xgc_option.collector_type_,
                       background_gc,
                       runtime_options.GetOrDefault(Opt::LargeObjectSpace),
                       runtime_options.GetOrDefault(Opt::LargeObjectThreshold),
                       runtime_options.GetOrDefault(Opt::ParallelGCThreads),
                       runtime_options.GetOrDefault(Opt::ConcGCThreads),
                       runtime_options.Exists(Opt::LowMemoryMode),
                       runtime_options.GetOrDefault(Opt::LongPauseLogThreshold),
                       runtime_options.GetOrDefault(Opt::LongGCLogThreshold),
                       runtime_options.Exists(Opt::IgnoreMaxFootprint),
                       runtime_options.GetOrDefault(Opt::AlwaysLogExplicitGcs),
                       runtime_options.GetOrDefault(Opt::UseTLAB),
                       xgc_option.verify_pre_gc_heap_,
                       xgc_option.verify_pre_sweeping_heap_,
                       xgc_option.verify_post_gc_heap_,
                       xgc_option.verify_pre_gc_rosalloc_,
                       xgc_option.verify_pre_sweeping_rosalloc_,
                       xgc_option.verify_post_gc_rosalloc_,
                       xgc_option.gcstress_,
                       xgc_option.continuous_gc_,
                       xgc_option.measure_,
                       runtime_options.GetOrDefault(Opt::EnableHSpaceCompactForOOM),
                       use_generational_gc,
                       runtime_options.GetOrDefault(Opt::HSpaceCompactForOOMMinIntervalsMs),
                       runtime_options.Exists(Opt::DumpRegionInfoBeforeGC),
                       runtime_options.Exists(Opt::DumpRegionInfoAfterGC));

  dump_gc_performance_on_shutdown_ = runtime_options.Exists(Opt::DumpGCPerformanceOnShutdown);

  bool has_explicit_jdwp_options = runtime_options.Get(Opt::JdwpOptions) != nullptr;
  jdwp_options_ = runtime_options.GetOrDefault(Opt::JdwpOptions);
  jdwp_provider_ = CanonicalizeJdwpProvider(runtime_options.GetOrDefault(Opt::JdwpProvider),
                                            IsJavaDebuggable());
  switch (jdwp_provider_) {
    case JdwpProvider::kNone: {
      VLOG(jdwp) << "Disabling all JDWP support.";
      if (!jdwp_options_.empty()) {
        bool has_transport = jdwp_options_.find("transport") != std::string::npos;
        std::string adb_connection_args =
            std::string("  -XjdwpProvider:adbconnection -XjdwpOptions:") + jdwp_options_;
        if (has_explicit_jdwp_options) {
          LOG(WARNING) << "Jdwp options given when jdwp is disabled! You probably want to enable "
                      << "jdwp with one of:" << std::endl
                      << "  -Xplugin:libopenjdkjvmti" << (kIsDebugBuild ? "d" : "") << ".so "
                      << "-agentpath:libjdwp.so=" << jdwp_options_ << std::endl
                      << (has_transport ? "" : adb_connection_args);
        }
      }
      break;
    }
    case JdwpProvider::kAdbConnection: {
      constexpr const char* plugin_name = kIsDebugBuild ? "libadbconnectiond.so"
                                                        : "libadbconnection.so";
      plugins_.push_back(Plugin::Create(plugin_name));
      break;
    }
    case JdwpProvider::kUnset: {
      LOG(FATAL) << "Illegal jdwp provider " << jdwp_provider_ << " was not filtered out!";
    }
  }
  callbacks_->AddThreadLifecycleCallback(Dbg::GetThreadLifecycleCallback());

  jit_options_.reset(jit::JitOptions::CreateFromRuntimeArguments(runtime_options));
  if (IsAotCompiler()) {
    // If we are already the compiler at this point, we must be dex2oat. Don't create the jit in
    // this case.
    // If runtime_options doesn't have UseJIT set to true then CreateFromRuntimeArguments returns
    // null and we don't create the jit.
    jit_options_->SetUseJitCompilation(false);
    jit_options_->SetSaveProfilingInfo(false);
  }

  // Use MemMap arena pool for jit, malloc otherwise. Malloc arenas are faster to allocate but
  // can't be trimmed as easily.
  const bool use_malloc = IsAotCompiler();
  if (use_malloc) {
    arena_pool_.reset(new CallocArenaPool());
    jit_arena_pool_.reset(new CallocArenaPool());
  } else {
    arena_pool_.reset(new MemMapArenaPool(/* low_4gb= */ false));
    jit_arena_pool_.reset(new MemMapArenaPool(/* low_4gb= */ false, "CompilerMetadata"));
  }

  // For 64 bit compilers, it needs to be in low 4GB in the case where we are cross compiling for a
  // 32 bit target. In this case, we have 32 bit pointers in the dex cache arrays which can't hold
  // when we have 64 bit ArtMethod pointers.
  const bool low_4gb = IsAotCompiler() && Is64BitInstructionSet(kRuntimeISA);
  if (gUseUserfaultfd) {
    linear_alloc_arena_pool_.reset(new GcVisitedArenaPool(low_4gb, IsZygote()));
  } else if (low_4gb) {
    linear_alloc_arena_pool_.reset(new MemMapArenaPool(low_4gb));
  }
  linear_alloc_.reset(CreateLinearAlloc());
  startup_linear_alloc_.store(CreateLinearAlloc(), std::memory_order_relaxed);

  small_lrt_allocator_ = new jni::SmallLrtAllocator();

  BlockSignals();
  InitPlatformSignalHandlers();

  // Change the implicit checks flags based on runtime architecture.
  switch (kRuntimeQuickCodeISA) {
    case InstructionSet::kArm64:
      implicit_suspend_checks_ = true;
      FALLTHROUGH_INTENDED;
    case InstructionSet::kArm:
    case InstructionSet::kThumb2:
    case InstructionSet::kRiscv64:
    case InstructionSet::kX86:
    case InstructionSet::kX86_64:
      implicit_null_checks_ = true;
      // Historical note: Installing stack protection was not playing well with Valgrind.
      implicit_so_checks_ = true;
      break;
    default:
      // Keep the defaults.
      break;
  }

#ifdef ART_USE_RESTRICTED_MODE
  // TODO(Simulator): support implicit suspend checks.
  implicit_suspend_checks_ = false;
#endif  // ART_USE_RESTRICTED_MODE

  fault_manager.Init(!no_sig_chain_);
  if (!no_sig_chain_) {
    if (HandlesSignalsInCompiledCode()) {
      // These need to be in a specific order.  The null point check handler must be
      // after the suspend check and stack overflow check handlers.
      //
      // Note: The manager will delete the handlers on Shutdown().
      if (implicit_suspend_checks_) {
        fault_manager.AddHandler(new SuspensionHandler(),
                                 SuspensionHandler::IsGeneratedCodeHandler());
      }

      if (implicit_so_checks_) {
        fault_manager.AddHandler(new StackOverflowHandler(),
                                 StackOverflowHandler::IsGeneratedCodeHandler());
      }

      if (implicit_null_checks_) {
#ifdef ART_USE_SIMULATOR
        fault_manager.AddHandler(new NullPointerHandlerSimulator(),
                                 NullPointerHandlerSimulator::IsGeneratedCodeHandler());
#else
        fault_manager.AddHandler(new NullPointerHandler(),
                                 NullPointerHandler::IsGeneratedCodeHandler());
#endif
      }

      if (kEnableJavaStackTraceHandler) {
        fault_manager.AddHandler(new JavaStackTraceHandler(&fault_manager),
                                 JavaStackTraceHandler::IsGeneratedCodeHandler());
      }

      if (interpreter::CanRuntimeUseNterp()) {
        // Nterp code can use signal handling just like the compiled managed code.
        OatQuickMethodHeader* nterp_header = OatQuickMethodHeader::NterpMethodHeader;
        fault_manager.AddGeneratedCodeRange(nterp_header->GetCode(), nterp_header->GetCodeSize());
      }
    }
  }

  verifier_logging_threshold_ms_ = runtime_options.GetOrDefault(Opt::VerifierLoggingThreshold);

  std::string error_msg;
  java_vm_ = JavaVMExt::Create(this, runtime_options, &error_msg);
  if (java_vm_.get() == nullptr) {
    LOG(ERROR) << "Could not initialize JavaVMExt: " << error_msg;
    return false;
  }

  // Add the JniEnv handler.
  // TODO Refactor this stuff.
  java_vm_->AddEnvironmentHook(JNIEnvExt::GetEnvHandler);

  Thread::Startup();

  // ClassLinker needs an attached thread, but we can't fully attach a thread without creating
  // objects. We can't supply a thread group yet; it will be fixed later. Since we are the main
  // thread, we do not get a java peer.
  Thread* self = Thread::Attach("main", false, nullptr, false, /* should_run_callbacks= */ true);
  CHECK_EQ(self->GetThreadId(), ThreadList::kMainThreadId);
  CHECK(self != nullptr);

  self->SetIsRuntimeThread(IsAotCompiler());

  // Set us to runnable so tools using a runtime can allocate and GC by default
  self->TransitionFromSuspendedToRunnable();

  // Now we're attached, we can take the heap locks and validate the heap.
  GetHeap()->EnableObjectValidation();

  CHECK_GE(GetHeap()->GetContinuousSpaces().size(), 1U);

  if (UNLIKELY(IsAotCompiler())) {
    class_linker_ = compiler_callbacks_->CreateAotClassLinker(intern_table_);
  } else {
    class_linker_ = new ClassLinker(
        intern_table_,
        runtime_options.GetOrDefault(Opt::FastClassNotFoundException));
  }
  if (GetHeap()->HasBootImageSpace()) {
    bool result = class_linker_->InitFromBootImage(&error_msg);
    if (!result) {
      LOG(ERROR) << "Could not initialize from image: " << error_msg;
      return false;
    }
    if (kIsDebugBuild) {
      for (auto image_space : GetHeap()->GetBootImageSpaces()) {
        image_space->VerifyImageAllocations();
      }
    }
    {
      ScopedTrace trace2("AddImageStringsToTable");
      for (gc::space::ImageSpace* image_space : heap_->GetBootImageSpaces()) {
        GetInternTable()->AddImageStringsToTable(image_space, VoidFunctor());
      }
    }

    const size_t total_components = gc::space::ImageSpace::GetNumberOfComponents(
        ArrayRef<gc::space::ImageSpace* const>(heap_->GetBootImageSpaces()));
    if (total_components != GetBootClassPath().size()) {
      // The boot image did not contain all boot class path components. Load the rest.
      CHECK_LT(total_components, GetBootClassPath().size());
      size_t start = total_components;
      DCHECK_LT(start, GetBootClassPath().size());
      std::vector<std::unique_ptr<const DexFile>> extra_boot_class_path;
      if (runtime_options.Exists(Opt::BootClassPathDexList)) {
        extra_boot_class_path.swap(*runtime_options.GetOrDefault(Opt::BootClassPathDexList));
      } else {
        ArrayRef<File> bcp_files = start < GetBootClassPathFiles().size() ?
                                       ArrayRef<File>(GetBootClassPathFiles()).SubArray(start) :
                                       ArrayRef<File>();
        OpenBootDexFiles(ArrayRef<const std::string>(GetBootClassPath()).SubArray(start),
                         ArrayRef<const std::string>(GetBootClassPathLocations()).SubArray(start),
                         bcp_files,
                         &extra_boot_class_path);
      }
      class_linker_->AddExtraBootDexFiles(self, std::move(extra_boot_class_path));
    }
    if (IsJavaDebuggable() || jit_options_->GetProfileSaverOptions().GetProfileBootClassPath()) {
      // Deoptimize the boot image if debuggable  as the code may have been compiled non-debuggable.
      // Also deoptimize if we are profiling the boot class path.
      ScopedThreadSuspension sts(self, ThreadState::kNative);
      ScopedSuspendAll ssa(__FUNCTION__);
      DeoptimizeBootImage();
    }
  } else {
    std::vector<std::unique_ptr<const DexFile>> boot_class_path;
    if (runtime_options.Exists(Opt::BootClassPathDexList)) {
      boot_class_path.swap(*runtime_options.GetOrDefault(Opt::BootClassPathDexList));
    } else {
      OpenBootDexFiles(ArrayRef<const std::string>(GetBootClassPath()),
                       ArrayRef<const std::string>(GetBootClassPathLocations()),
                       ArrayRef<File>(GetBootClassPathFiles()),
                       &boot_class_path);
    }
    if (!class_linker_->InitWithoutImage(std::move(boot_class_path), &error_msg)) {
      LOG(ERROR) << "Could not initialize without image: " << error_msg;
      return false;
    }

    // TODO: Should we move the following to InitWithoutImage?
    SetInstructionSet(instruction_set_);
    for (uint32_t i = 0; i < kCalleeSaveSize; i++) {
      CalleeSaveType type = CalleeSaveType(i);
      if (!HasCalleeSaveMethod(type)) {
        SetCalleeSaveMethod(CreateCalleeSaveMethod(), type);
      }
    }
  }

  // Now that the boot image space is set, cache the boot classpath checksums,
  // to be used when validating oat files.
  ArrayRef<gc::space::ImageSpace* const> image_spaces(GetHeap()->GetBootImageSpaces());
  ArrayRef<const DexFile* const> bcp_dex_files(GetClassLinker()->GetBootClassPath());
  boot_class_path_checksums_ = gc::space::ImageSpace::GetBootClassPathChecksums(image_spaces,
                                                                                bcp_dex_files);

  CHECK(class_linker_ != nullptr);

  if (runtime_options.Exists(Opt::MethodTrace)) {
    trace_config_.reset(new TraceConfig());
    trace_config_->trace_file = runtime_options.ReleaseOrDefault(Opt::MethodTraceFile);
    trace_config_->trace_file_size = runtime_options.ReleaseOrDefault(Opt::MethodTraceFileSize);
    trace_config_->trace_mode = Trace::TraceMode::kMethodTracing;
    trace_config_->trace_output_mode = runtime_options.Exists(Opt::MethodTraceStreaming) ?
                                           TraceOutputMode::kStreaming :
                                           TraceOutputMode::kFile;
    trace_config_->clock_source = runtime_options.GetOrDefault(Opt::MethodTraceClock);
  }

  if (GetHeap()->HasBootImageSpace()) {
    const ImageHeader& image_header = GetHeap()->GetBootImageSpaces()[0]->GetImageHeader();
    ObjPtr<mirror::ObjectArray<mirror::Object>> boot_image_live_objects =
        ObjPtr<mirror::ObjectArray<mirror::Object>>::DownCast(
            image_header.GetImageRoot(ImageHeader::kBootImageLiveObjects));
    pre_allocated_OutOfMemoryError_when_throwing_exception_ = GcRoot<mirror::Throwable>(
        boot_image_live_objects->Get(ImageHeader::kOomeWhenThrowingException)->AsThrowable());
    DCHECK(pre_allocated_OutOfMemoryError_when_throwing_exception_.Read()->GetClass()
               ->DescriptorEquals("Ljava/lang/OutOfMemoryError;"));
    pre_allocated_OutOfMemoryError_when_throwing_oome_ = GcRoot<mirror::Throwable>(
        boot_image_live_objects->Get(ImageHeader::kOomeWhenThrowingOome)->AsThrowable());
    DCHECK(pre_allocated_OutOfMemoryError_when_throwing_oome_.Read()->GetClass()
               ->DescriptorEquals("Ljava/lang/OutOfMemoryError;"));
    pre_allocated_OutOfMemoryError_when_handling_stack_overflow_ = GcRoot<mirror::Throwable>(
        boot_image_live_objects->Get(ImageHeader::kOomeWhenHandlingStackOverflow)->AsThrowable());
    DCHECK(pre_allocated_OutOfMemoryError_when_handling_stack_overflow_.Read()->GetClass()
               ->DescriptorEquals("Ljava/lang/OutOfMemoryError;"));
    pre_allocated_NoClassDefFoundError_ = GcRoot<mirror::Throwable>(
        boot_image_live_objects->Get(ImageHeader::kNoClassDefFoundError)->AsThrowable());
    DCHECK(pre_allocated_NoClassDefFoundError_.Read()->GetClass()
               ->DescriptorEquals("Ljava/lang/NoClassDefFoundError;"));
  } else {
    CreatePreAllocatedExceptions(self);
  }

  // Class-roots are setup, we can now finish initializing the JniIdManager.
  GetJniIdManager()->Init(self);

  // Initialize metrics only for the Zygote process or
  // if explicitly enabled via command line argument.
  if (IsZygote() || gFlags.MetricsForceEnable.GetValue()) {
    LOG(INFO) << "Initializing ART runtime metrics";
    InitMetrics();
  }

  // Runtime initialization is largely done now.
  // We load plugins first since that can modify the runtime state slightly.
  // Load all plugins
  {
    // The init method of plugins expect the state of the thread to be non runnable.
    ScopedThreadSuspension sts(self, ThreadState::kNative);
    for (auto& plugin : plugins_) {
      std::string err;
      if (!plugin.Load(&err)) {
        LOG(FATAL) << plugin << " failed to load: " << err;
      }
    }
  }

  // Look for a native bridge.
  //
  // The intended flow here is, in the case of a running system:
  //
  // Runtime::Init() (zygote):
  //   LoadNativeBridge -> dlopen from cmd line parameter.
  //  |
  //  V
  // Runtime::Start() (zygote):
  //   No-op wrt native bridge.
  //  |
  //  | start app
  //  V
  // DidForkFromZygote(action)
  //   action = kUnload -> dlclose native bridge.
  //   action = kInitialize -> initialize library
  //
  //
  // The intended flow here is, in the case of a simple dalvikvm call:
  //
  // Runtime::Init():
  //   LoadNativeBridge -> dlopen from cmd line parameter.
  //  |
  //  V
  // Runtime::Start():
  //   DidForkFromZygote(kInitialize) -> try to initialize any native bridge given.
  //   No-op wrt native bridge.
  {
    std::string native_bridge_file_name = runtime_options.ReleaseOrDefault(Opt::NativeBridge);
    is_native_bridge_loaded_ = LoadNativeBridge(native_bridge_file_name);
  }

  // Startup agents
  // TODO Maybe we should start a new thread to run these on. Investigate RI behavior more.
  for (auto& agent_spec : agent_specs_) {
    // TODO Check err
    int res = 0;
    std::string err = "";
    ti::LoadError error;
    std::unique_ptr<ti::Agent> agent = agent_spec.Load(&res, &error, &err);

    if (agent != nullptr) {
      agents_.push_back(std::move(agent));
      continue;
    }

    switch (error) {
      case ti::LoadError::kInitializationError:
        LOG(FATAL) << "Unable to initialize agent!";
        UNREACHABLE();

      case ti::LoadError::kLoadingError:
        LOG(ERROR) << "Unable to load an agent: " << err;
        continue;

      case ti::LoadError::kNoError:
        break;
    }
    LOG(FATAL) << "Unreachable";
    UNREACHABLE();
  }

  Locks::mutator_lock_->AssertSharedHeld(self);
  callbacks_->NextRuntimePhase(RuntimePhaseCallback::RuntimePhase::kInitialAgents);

  if (IsZygote() && IsPerfettoHprofEnabled()) {
    constexpr const char* plugin_name = kIsDebugBuild ?
        "libperfetto_hprofd.so" : "libperfetto_hprof.so";
    // Load eagerly in Zygote to improve app startup times. This will make
    // subsequent dlopens for the library no-ops.
    dlopen(plugin_name, RTLD_NOW | RTLD_LOCAL);
  }

  for (const auto& signature : AssumeValueSignatures::kSignatures) {
    ArtField* field = signature.LookupField();
    if (field != nullptr) {
      assume_value_field_signatures_.insert_or_assign(field, &signature);
    } else {
      // Don't treat this as an error; assumed values are purely an optimization, and we reserve
      // the right to selectively ignore/deprecate optimizations for certain fields.
      VLOG(compiler) << "Failed to find field corresponding to " << signature.AsKey();
    }
  }

  VLOG(startup) << "Runtime::Init exiting";

  return true;
}

void Runtime::InitMetrics() {
  metrics::ReportingConfig metrics_config = metrics::ReportingConfig::FromFlags();
  metrics_reporter_ = metrics::MetricsReporter::Create(metrics_config, this);
}

void Runtime::RequestMetricsReport(bool synchronous) {
  if (AreMetricsInitialized()) {
    metrics_reporter_->RequestMetricsReport(synchronous);
  }
}

bool Runtime::EnsurePluginLoaded(const char* plugin_name, std::string* error_msg) {
  // Is the plugin already loaded?
  for (const Plugin& p : plugins_) {
    if (p.GetLibrary() == plugin_name) {
      return true;
    }
  }
  Plugin new_plugin = Plugin::Create(plugin_name);

  if (!new_plugin.Load(error_msg)) {
    return false;
  }
  plugins_.push_back(std::move(new_plugin));
  return true;
}

bool Runtime::EnsurePerfettoPlugin(std::string* error_msg) {
  constexpr const char* plugin_name = kIsDebugBuild ?
    "libperfetto_hprofd.so" : "libperfetto_hprof.so";
  return EnsurePluginLoaded(plugin_name, error_msg);
}

static bool EnsureJvmtiPlugin(Runtime* runtime,
                              std::string* error_msg) {
  // TODO Rename Dbg::IsJdwpAllowed is IsDebuggingAllowed.
  DCHECK(Dbg::IsJdwpAllowed() || !runtime->IsJavaDebuggable())
      << "Being debuggable requires that jdwp (i.e. debugging) is allowed.";
  // Is the process debuggable? Otherwise, do not attempt to load the plugin unless we are
  // specifically allowed.
  if (!Dbg::IsJdwpAllowed()) {
    *error_msg = "Process is not allowed to load openjdkjvmti plugin. Process must be debuggable";
    return false;
  }

  constexpr const char* plugin_name = kIsDebugBuild ? "libopenjdkjvmtid.so" : "libopenjdkjvmti.so";
  return runtime->EnsurePluginLoaded(plugin_name, error_msg);
}

// Attach a new agent and add it to the list of runtime agents
//
// TODO: once we decide on the threading model for agents,
//   revisit this and make sure we're doing this on the right thread
//   (and we synchronize access to any shared data structures like "agents_")
//
void Runtime::AttachAgent(JNIEnv* env, const std::string& agent_arg, jobject class_loader) {
  std::string error_msg;
  if (!EnsureJvmtiPlugin(this, &error_msg)) {
    LOG(WARNING) << "Could not load plugin: " << error_msg;
    ScopedObjectAccess soa(Thread::Current());
    ThrowIOException("%s", error_msg.c_str());
    return;
  }

  ti::AgentSpec agent_spec(agent_arg);

  int res = 0;
  ti::LoadError error;
  std::unique_ptr<ti::Agent> agent = agent_spec.Attach(env, class_loader, &res, &error, &error_msg);

  if (agent != nullptr) {
    agents_.push_back(std::move(agent));
  } else {
    LOG(WARNING) << "Agent attach failed (result=" << error << ") : " << error_msg;
    ScopedObjectAccess soa(Thread::Current());
    ThrowIOException("%s", error_msg.c_str());
  }
}

void Runtime::InitNativeMethods() {
  VLOG(startup) << "Runtime::InitNativeMethods entering";
  Thread* self = Thread::Current();
  JNIEnv* env = self->GetJniEnv();

  // Must be in the kNative state for calling native methods (JNI_OnLoad code).
  CHECK_EQ(self->GetState(), ThreadState::kNative);

  // Then set up libjavacore / libopenjdk / libicu_jni ,which are just
  // a regular JNI libraries with a regular JNI_OnLoad. Most JNI libraries can
  // just use System.loadLibrary, but libcore can't because it's the library
  // that implements System.loadLibrary!
  //
  // By setting calling class to java.lang.Object, the caller location for these
  // JNI libs is core-oj.jar in the ART APEX, and hence they are loaded from the
  // com_android_art linker namespace.
  jclass java_lang_Object;
  {
    // Use global JNI reference to keep the local references empty. If we allocated a
    // local reference here, the `PushLocalFrame(128)` that these internal libraries do
    // in their `JNI_OnLoad()` would reserve a lot of unnecessary space due to rounding.
    ScopedObjectAccess soa(self);
    java_lang_Object = reinterpret_cast<jclass>(
        GetJavaVM()->AddGlobalRef(self, GetClassRoot<mirror::Object>(GetClassLinker())));
  }

  // libicu_jni has to be initialized before libopenjdk{d} due to runtime dependency from
  // libopenjdk{d} to Icu4cMetadata native methods in libicu_jni. See http://b/143888405
  {
    std::string error_msg;
    if (!java_vm_->LoadNativeLibrary(
          env, "libicu_jni.so", nullptr, java_lang_Object, &error_msg)) {
      LOG(FATAL) << "LoadNativeLibrary failed for \"libicu_jni.so\": " << error_msg;
    }
  }
  {
    std::string error_msg;
    if (!java_vm_->LoadNativeLibrary(
          env, "libjavacore.so", nullptr, java_lang_Object, &error_msg)) {
      LOG(FATAL) << "LoadNativeLibrary failed for \"libjavacore.so\": " << error_msg;
    }
  }
  {
    constexpr const char* kOpenJdkLibrary = kIsDebugBuild
                                                ? "libopenjdkd.so"
                                                : "libopenjdk.so";
    std::string error_msg;
    if (!java_vm_->LoadNativeLibrary(
          env, kOpenJdkLibrary, nullptr, java_lang_Object, &error_msg)) {
      LOG(FATAL) << "LoadNativeLibrary failed for \"" << kOpenJdkLibrary << "\": " << error_msg;
    }
  }
  env->DeleteGlobalRef(java_lang_Object);

  // Initialize well known classes that may invoke runtime native methods.
  WellKnownClasses::LateInit(env);

  VLOG(startup) << "Runtime::InitNativeMethods exiting";
}

void Runtime::ReclaimArenaPoolMemory() {
  arena_pool_->LockReclaimMemory();
}

void Runtime::InitThreadGroups(Thread* self) {
  ScopedObjectAccess soa(self);
  ArtField* main_thread_group_field = WellKnownClasses::java_lang_ThreadGroup_mainThreadGroup;
  ArtField* system_thread_group_field = WellKnownClasses::java_lang_ThreadGroup_systemThreadGroup;
  // Note: This is running before `ClassLinker::RunRootClinits()`, so we cannot rely on
  // `ThreadGroup` and `Thread` being initialized.
  // TODO: Clean up initialization order after all well-known methods are converted to `ArtMethod*`
  // (and therefore the `WellKnownClasses::Init()` shall not initialize any classes).
  StackHandleScope<2u> hs(self);
  Handle<mirror::Class> thread_group_class =
      hs.NewHandle(main_thread_group_field->GetDeclaringClass());
  bool initialized = GetClassLinker()->EnsureInitialized(
      self, thread_group_class, /*can_init_fields=*/ true, /*can_init_parents=*/ true);
  CHECK(initialized);
  Handle<mirror::Class> thread_class = hs.NewHandle(WellKnownClasses::java_lang_Thread.Get());
  initialized = GetClassLinker()->EnsureInitialized(
      self, thread_class, /*can_init_fields=*/ true, /*can_init_parents=*/ true);
  CHECK(initialized);
  main_thread_group_ =
      soa.Vm()->AddGlobalRef(self, main_thread_group_field->GetObject(thread_group_class.Get()));
  CHECK_IMPLIES(main_thread_group_ == nullptr, IsAotCompiler());
  system_thread_group_ =
      soa.Vm()->AddGlobalRef(self, system_thread_group_field->GetObject(thread_group_class.Get()));
  CHECK_IMPLIES(system_thread_group_ == nullptr, IsAotCompiler());
}

jobject Runtime::GetMainThreadGroup() const {
  CHECK_IMPLIES(main_thread_group_ == nullptr, IsAotCompiler());
  return main_thread_group_;
}

jobject Runtime::GetSystemThreadGroup() const {
  CHECK_IMPLIES(system_thread_group_ == nullptr, IsAotCompiler());
  return system_thread_group_;
}

jobject Runtime::GetSystemClassLoader() const {
  CHECK_IMPLIES(system_class_loader_ == nullptr, IsAotCompiler());
  return system_class_loader_;
}

void Runtime::RegisterRuntimeNativeMethods(JNIEnv* env) {
  register_dalvik_system_DexFile(env);
  register_dalvik_system_BaseDexClassLoader(env);
  register_dalvik_system_VMDebug(env);
  real_register_dalvik_system_VMRuntime(env);
  register_dalvik_system_VMStack(env);
  register_dalvik_system_ZygoteHooks(env);
  register_java_lang_Class(env);
  register_java_lang_Object(env);
  register_java_lang_invoke_MethodHandle(env);
  register_java_lang_invoke_MethodHandleImpl(env);
  register_java_lang_ref_FinalizerReference(env);
  register_java_lang_reflect_Array(env);
  register_java_lang_reflect_Constructor(env);
  register_java_lang_reflect_Executable(env);
  register_java_lang_reflect_Field(env);
  register_java_lang_reflect_Method(env);
  register_java_lang_reflect_Parameter(env);
  register_java_lang_reflect_Proxy(env);
  register_java_lang_ref_Reference(env);
  register_java_lang_StackStreamFactory(env);
  register_java_lang_String(env);
  register_java_lang_StringFactory(env);
  register_java_lang_System(env);
  register_java_lang_Thread(env);
  register_java_lang_Throwable(env);
  register_java_lang_VMClassLoader(env);
  register_jdk_internal_misc_Unsafe(env);
  register_jdk_internal_vm_Continuation(env);
  register_libcore_io_Memory(env);
  register_libcore_util_CharsetUtils(env);
  register_org_apache_harmony_dalvik_ddmc_DdmServer(env);
  register_org_apache_harmony_dalvik_ddmc_DdmVmInternal(env);
  register_sun_misc_Unsafe(env);
}

std::ostream& operator<<(std::ostream& os, const DeoptimizationKind& kind) {
  os << GetDeoptimizationKindName(kind);
  return os;
}

void Runtime::DumpDeoptimizations(std::ostream& os) {
  for (size_t i = 0; i <= static_cast<size_t>(DeoptimizationKind::kLast); ++i) {
    if (deoptimization_counts_[i] != 0) {
      os << "Number of "
         << GetDeoptimizationKindName(static_cast<DeoptimizationKind>(i))
         << " deoptimizations: "
         << deoptimization_counts_[i]
         << "\n";
    }
  }
}

std::optional<uint64_t> Runtime::SigQuitNanoTime() const {
  return signal_catcher_ != nullptr ? signal_catcher_->SigQuitNanoTime() : std::nullopt;
}

void Runtime::DumpForSigQuit(std::ostream& os) {
  // Print backtraces first since they are important do diagnose ANRs,
  // and ANRs can often be trimmed to limit upload size.
  thread_list_->DumpForSigQuit(os);
  GetClassLinker()->DumpForSigQuit(os);
  GetInternTable()->DumpForSigQuit(os);
  GetJavaVM()->DumpForSigQuit(os);
  GetHeap()->DumpForSigQuit(os);
  oat_file_manager_->DumpForSigQuit(os);
  if (GetJit() != nullptr) {
    GetJit()->DumpForSigQuit(os);
  } else {
    os << "Running non JIT\n";
  }
  DumpDeoptimizations(os);
  TrackedAllocators::Dump(os);
  GetMetrics()->DumpForSigQuit(os);
  os << "\n";

  BaseMutex::DumpAll(os);

  // Inform anyone else who is interested in SigQuit.
  {
    ScopedObjectAccess soa(Thread::Current());
    callbacks_->SigQuit();
  }
}

void Runtime::DumpLockHolders(std::ostream& os) {
  pid_t mutator_lock_owner = Locks::mutator_lock_->GetExclusiveOwnerTid();
  pid_t thread_list_lock_owner = GetThreadList()->GetLockOwner();
  pid_t classes_lock_owner = GetClassLinker()->GetClassesLockOwner();
  pid_t dex_lock_owner = GetClassLinker()->GetDexLockOwner();
  if ((mutator_lock_owner | thread_list_lock_owner | classes_lock_owner | dex_lock_owner) != 0) {
    os << "Mutator lock exclusive owner tid: " << mutator_lock_owner << "\n"
       << "ThreadList lock owner tid: " << thread_list_lock_owner << "\n"
       << "ClassLinker classes lock owner tid: " << classes_lock_owner << "\n"
       << "ClassLinker dex lock owner tid: " << dex_lock_owner << "\n";
  }
}

void Runtime::SetStatsEnabled(bool new_state) {
  Thread* self = Thread::Current();
  MutexLock mu(self, *Locks::instrument_entrypoints_lock_);
  if (new_state == true) {
    GetStats()->Clear(~0);
    // TODO: wouldn't it make more sense to clear _all_ threads' stats?
    self->GetStats()->Clear(~0);
    if (stats_enabled_ != new_state) {
      GetInstrumentation()->InstrumentQuickAllocEntryPointsLocked();
    }
  } else if (stats_enabled_ != new_state) {
    GetInstrumentation()->UninstrumentQuickAllocEntryPointsLocked();
  }
  stats_enabled_ = new_state;
}

void Runtime::ResetStats(int kinds) {
  GetStats()->Clear(kinds & 0xffff);
  // TODO: wouldn't it make more sense to clear _all_ threads' stats?
  Thread::Current()->GetStats()->Clear(kinds >> 16);
}

uint64_t Runtime::GetStat(int kind) {
  RuntimeStats* stats;
  if (kind < (1<<16)) {
    stats = GetStats();
  } else {
    stats = Thread::Current()->GetStats();
    kind >>= 16;
  }
  switch (kind) {
  case KIND_ALLOCATED_OBJECTS:
    return stats->allocated_objects;
  case KIND_ALLOCATED_BYTES:
    return stats->allocated_bytes;
  case KIND_FREED_OBJECTS:
    return stats->freed_objects;
  case KIND_FREED_BYTES:
    return stats->freed_bytes;
  case KIND_GC_INVOCATIONS:
    return stats->gc_for_alloc_count;
  case KIND_CLASS_INIT_COUNT:
    return stats->class_init_count;
  case KIND_CLASS_INIT_TIME:
    return stats->class_init_time_ns;
  case KIND_EXT_ALLOCATED_OBJECTS:
  case KIND_EXT_ALLOCATED_BYTES:
  case KIND_EXT_FREED_OBJECTS:
  case KIND_EXT_FREED_BYTES:
    return 0;  // backward compatibility
  default:
    LOG(FATAL) << "Unknown statistic " << kind;
    UNREACHABLE();
  }
}

void Runtime::BlockSignals() {
  SignalSet signals;
  signals.Add(SIGPIPE);
  // SIGQUIT is used to dump the runtime's state (including stack traces).
  signals.Add(SIGQUIT);
  // SIGUSR1 is used to initiate a GC.
  signals.Add(SIGUSR1);
  signals.Block();
}

bool Runtime::AttachCurrentThread(const char* thread_name, bool as_daemon, jobject thread_group,
                                  bool create_peer, bool should_run_callbacks) {
  ScopedTrace trace(__FUNCTION__);
  Thread* self = Thread::Attach(thread_name,
                                as_daemon,
                                thread_group,
                                create_peer,
                                should_run_callbacks);
  // Run ThreadGroup.add to notify the group that this thread is now started.
  if (self != nullptr && create_peer && !IsAotCompiler()) {
    ScopedObjectAccess soa(self);
    self->NotifyThreadGroup(soa, thread_group);
  }
  return self != nullptr;
}

void Runtime::DetachCurrentThread(bool should_run_callbacks) {
  ScopedTrace trace(__FUNCTION__);
  Thread* self = Thread::Current();
  if (self == nullptr) {
    LOG(FATAL) << "attempting to detach thread that is not attached";
  }
  if (self->HasManagedStack()) {
    LOG(FATAL) << *Thread::Current() << " attempting to detach while still running code";
  }
  thread_list_->Unregister(self, should_run_callbacks);
}

mirror::Throwable* Runtime::GetPreAllocatedOutOfMemoryErrorWhenThrowingException() {
  mirror::Throwable* oome = pre_allocated_OutOfMemoryError_when_throwing_exception_.Read();
  if (oome == nullptr) {
    LOG(ERROR) << "Failed to return pre-allocated OOME-when-throwing-exception";
  }
  return oome;
}

mirror::Throwable* Runtime::GetPreAllocatedOutOfMemoryErrorWhenThrowingOOME() {
  mirror::Throwable* oome = pre_allocated_OutOfMemoryError_when_throwing_oome_.Read();
  if (oome == nullptr) {
    LOG(ERROR) << "Failed to return pre-allocated OOME-when-throwing-OOME";
  }
  return oome;
}

mirror::Throwable* Runtime::GetPreAllocatedOutOfMemoryErrorWhenHandlingStackOverflow() {
  mirror::Throwable* oome = pre_allocated_OutOfMemoryError_when_handling_stack_overflow_.Read();
  if (oome == nullptr) {
    LOG(ERROR) << "Failed to return pre-allocated OOME-when-handling-stack-overflow";
  }
  return oome;
}

mirror::Throwable* Runtime::GetPreAllocatedNoClassDefFoundError() {
  mirror::Throwable* ncdfe = pre_allocated_NoClassDefFoundError_.Read();
  if (ncdfe == nullptr) {
    LOG(ERROR) << "Failed to return pre-allocated NoClassDefFoundError";
  }
  return ncdfe;
}

void Runtime::VisitConstantRoots(RootVisitor* visitor) {
  // Visiting the roots of these ArtMethods is not currently required since all the GcRoots are
  // null.
  BufferedRootVisitor<16> buffered_visitor(visitor, RootInfo(kRootVMInternal));
  const PointerSize pointer_size = GetClassLinker()->GetImagePointerSize();
  if (HasResolutionMethod()) {
    resolution_method_->VisitRoots(buffered_visitor, pointer_size);
  }
  if (HasImtConflictMethod()) {
    imt_conflict_method_->VisitRoots(buffered_visitor, pointer_size);
  }
  if (imt_unimplemented_method_ != nullptr) {
    imt_unimplemented_method_->VisitRoots(buffered_visitor, pointer_size);
  }
  for (uint32_t i = 0; i < kCalleeSaveSize; ++i) {
    auto* m = reinterpret_cast<ArtMethod*>(callee_save_methods_[i]);
    if (m != nullptr) {
      m->VisitRoots(buffered_visitor, pointer_size);
    }
  }
}

void Runtime::VisitConcurrentRoots(RootVisitor* visitor, VisitRootFlags flags) {
  // Userfaultfd compaction updates intern-tables and class-tables page-by-page
  // via LinearAlloc. So don't visit them here.
  if (GetHeap()->IsPerformingUffdCompaction()) {
    class_linker_->VisitRoots(visitor, flags, /*visit_class_roots=*/false);
  } else {
    intern_table_->VisitRoots(visitor, flags);
    class_linker_->VisitRoots(visitor, flags, /*visit_class_roots=*/true);
  }
  jni_id_manager_->VisitRoots(visitor);
  heap_->VisitAllocationRecords(visitor);
  if (jit_ != nullptr) {
    jit_->VisitRoots(visitor);
  }
  if ((flags & kVisitRootFlagNewRoots) == 0) {
    // Guaranteed to have no new roots in the constant roots.
    VisitConstantRoots(visitor);
  }
}

void Runtime::VisitNonThreadRoots(RootVisitor* visitor) {
  java_vm_->VisitRoots(visitor);
  sentinel_.VisitRootIfNonNull(visitor, RootInfo(kRootVMInternal));
  pre_allocated_OutOfMemoryError_when_throwing_exception_
      .VisitRootIfNonNull(visitor, RootInfo(kRootVMInternal));
  pre_allocated_OutOfMemoryError_when_throwing_oome_
      .VisitRootIfNonNull(visitor, RootInfo(kRootVMInternal));
  pre_allocated_OutOfMemoryError_when_handling_stack_overflow_
      .VisitRootIfNonNull(visitor, RootInfo(kRootVMInternal));
  pre_allocated_NoClassDefFoundError_.VisitRootIfNonNull(visitor, RootInfo(kRootVMInternal));
  VisitImageRoots(visitor);
  class_linker_->VisitTransactionRoots(visitor);
}

void Runtime::VisitNonConcurrentRoots(RootVisitor* visitor, VisitRootFlags flags) {
  VisitThreadRoots(visitor, flags);
  VisitNonThreadRoots(visitor);
}

void Runtime::VisitThreadRoots(RootVisitor* visitor, VisitRootFlags flags) {
  thread_list_->VisitRoots(visitor, flags);
}

void Runtime::VisitRoots(RootVisitor* visitor, VisitRootFlags flags) {
  VisitNonConcurrentRoots(visitor, flags);
  VisitConcurrentRoots(visitor, flags);
}

void Runtime::VisitReflectiveTargets(ReflectiveValueVisitor *visitor) {
  thread_list_->VisitReflectiveTargets(visitor);
  heap_->VisitReflectiveTargets(visitor);
  jni_id_manager_->VisitReflectiveTargets(visitor);
  callbacks_->VisitReflectiveTargets(visitor);
}

void Runtime::VisitImageRoots(RootVisitor* visitor) {
  // We only confirm that image roots are unchanged.
  if (kIsDebugBuild) {
    for (auto* space : GetHeap()->GetContinuousSpaces()) {
      if (space->IsImageSpace()) {
        auto* image_space = space->AsImageSpace();
        const auto& image_header = image_space->GetImageHeader();
        for (int32_t i = 0, size = image_header.GetImageRoots()->GetLength(); i != size; ++i) {
          mirror::Object* obj =
              image_header.GetImageRoot(static_cast<ImageHeader::ImageRoot>(i)).Ptr();
          if (obj != nullptr) {
            mirror::Object* after_obj = obj;
            visitor->VisitRoot(&after_obj, RootInfo(kRootStickyClass));
            CHECK_EQ(after_obj, obj);
          }
        }
      }
    }
  }
}

static ArtMethod* CreateRuntimeMethod(ClassLinker* class_linker, LinearAlloc* linear_alloc)
    REQUIRES_SHARED(Locks::mutator_lock_) {
  const PointerSize image_pointer_size = class_linker->GetImagePointerSize();
  const size_t method_alignment = ArtMethod::Alignment(image_pointer_size);
  const size_t method_size = ArtMethod::Size(image_pointer_size);
  LengthPrefixedArray<ArtMethod>* method_array = class_linker->AllocArtMethodArray(
      Thread::Current(),
      linear_alloc,
      1);
  ArtMethod* method = &method_array->At(0, method_size, method_alignment);
  CHECK(method != nullptr);
  method->SetDexMethodIndex(dex::kDexNoIndex);
  CHECK(method->IsRuntimeMethod());
  return method;
}

ArtMethod* Runtime::CreateImtConflictMethod(LinearAlloc* linear_alloc) {
  ClassLinker* const class_linker = GetClassLinker();
  ArtMethod* method = CreateRuntimeMethod(class_linker, linear_alloc);
  // When compiling, the code pointer will get set later when the image is loaded.
  const PointerSize pointer_size = GetInstructionSetPointerSize(instruction_set_);
  if (IsAotCompiler()) {
    method->SetEntryPointFromQuickCompiledCodePtrSize(nullptr, pointer_size);
  } else {
    method->SetEntryPointFromQuickCompiledCode(GetQuickImtConflictStub());
  }
  // Create empty conflict table.
  method->SetImtConflictTable(class_linker->CreateImtConflictTable(/*count=*/0u, linear_alloc),
                              pointer_size);
  return method;
}

void Runtime::SetImtConflictMethod(ArtMethod* method) {
  CHECK(method != nullptr);
  CHECK(method->IsRuntimeMethod());
  imt_conflict_method_ = method;
}

ArtMethod* Runtime::CreateResolutionMethod() {
  auto* method = CreateRuntimeMethod(GetClassLinker(), GetLinearAlloc());
  // When compiling, the code pointer will get set later when the image is loaded.
  if (IsAotCompiler()) {
    PointerSize pointer_size = GetInstructionSetPointerSize(instruction_set_);
    method->SetEntryPointFromQuickCompiledCodePtrSize(nullptr, pointer_size);
    method->SetEntryPointFromJniPtrSize(nullptr, pointer_size);
  } else {
    method->SetEntryPointFromQuickCompiledCode(GetQuickResolutionStub());
    method->SetEntryPointFromJni(GetJniDlsymLookupCriticalStub());
  }
  return method;
}

ArtMethod* Runtime::CreateCalleeSaveMethod() {
  auto* method = CreateRuntimeMethod(GetClassLinker(), GetLinearAlloc());
  PointerSize pointer_size = GetInstructionSetPointerSize(instruction_set_);
  method->SetEntryPointFromQuickCompiledCodePtrSize(nullptr, pointer_size);
  DCHECK_NE(instruction_set_, InstructionSet::kNone);
  DCHECK(method->IsRuntimeMethod());
  return method;
}

void Runtime::DisallowNewSystemWeaks() {
  CHECK(!gUseReadBarrier);
  monitor_list_->DisallowNewMonitors();
  intern_table_->ChangeWeakRootState(gc::kWeakRootStateNoReadsOrWrites);
  java_vm_->DisallowNewWeakGlobals();
  heap_->DisallowNewAllocationRecords();
  if (GetJit() != nullptr) {
    GetJit()->GetCodeCache()->DisallowInlineCacheAccess();
  }

  // All other generic system-weak holders.
  for (gc::AbstractSystemWeakHolder* holder : system_weak_holders_) {
    holder->Disallow();
  }
}

void Runtime::AllowNewSystemWeaks() {
  CHECK(!gUseReadBarrier);
  monitor_list_->AllowNewMonitors();
  intern_table_->ChangeWeakRootState(gc::kWeakRootStateNormal);  // TODO: Do this in the sweeping.
  java_vm_->AllowNewWeakGlobals();
  heap_->AllowNewAllocationRecords();
  if (GetJit() != nullptr) {
    GetJit()->GetCodeCache()->AllowInlineCacheAccess();
  }

  // All other generic system-weak holders.
  for (gc::AbstractSystemWeakHolder* holder : system_weak_holders_) {
    holder->Allow();
  }
}

void Runtime::BroadcastForNewSystemWeaks(bool broadcast_for_checkpoint) {
  // This is used for the read barrier case that uses the thread-local
  // Thread::GetWeakRefAccessEnabled() flag and the checkpoint while weak ref access is disabled
  // (see ThreadList::RunCheckpoint).
  monitor_list_->BroadcastForNewMonitors();
  intern_table_->BroadcastForNewInterns();
  java_vm_->BroadcastForNewWeakGlobals();
  heap_->BroadcastForNewAllocationRecords();
  if (GetJit() != nullptr) {
    GetJit()->GetCodeCache()->BroadcastForInlineCacheAccess();
  }

  // All other generic system-weak holders.
  for (gc::AbstractSystemWeakHolder* holder : system_weak_holders_) {
    holder->Broadcast(broadcast_for_checkpoint);
  }
}

void Runtime::SetInstructionSet(InstructionSet instruction_set) {
  instruction_set_ = instruction_set;
  switch (instruction_set) {
    case InstructionSet::kThumb2:
      // kThumb2 is the same as kArm, use the canonical value.
      instruction_set_ = InstructionSet::kArm;
      break;
    case InstructionSet::kArm:
    case InstructionSet::kArm64:
    case InstructionSet::kRiscv64:
    case InstructionSet::kX86:
    case InstructionSet::kX86_64:
      break;
    default:
      UNIMPLEMENTED(FATAL) << instruction_set_;
      UNREACHABLE();
  }
}

void Runtime::ClearInstructionSet() {
  instruction_set_ = InstructionSet::kNone;
}

void Runtime::SetCalleeSaveMethod(ArtMethod* method, CalleeSaveType type) {
  DCHECK_LT(static_cast<uint32_t>(type), kCalleeSaveSize);
  CHECK(method != nullptr);
  callee_save_methods_[static_cast<size_t>(type)] = reinterpret_cast<uintptr_t>(method);
}

void Runtime::ClearCalleeSaveMethods() {
  for (size_t i = 0; i < kCalleeSaveSize; ++i) {
    callee_save_methods_[i] = reinterpret_cast<uintptr_t>(nullptr);
  }
}

void Runtime::RegisterAppInfo(const std::string& package_name,
                              const std::vector<std::string>& code_paths,
                              const std::string& profile_output_filename,
                              const std::string& ref_profile_filename,
                              int32_t code_type) {
  AppInfo::CodeType internal_code_type = AppInfo::FromVMRuntimeConstants(code_type);
  app_info_.RegisterAppInfo(
      package_name,
      code_paths,
      profile_output_filename,
      ref_profile_filename,
      internal_code_type);

  if (AreMetricsInitialized()) {
    metrics_reporter_->NotifyAppInfoUpdated(&app_info_);
  }

  if (jit_ == nullptr) {
    // We are not JITing. Nothing to do.
    return;
  }

  VLOG(profiler) << "Register app with " << profile_output_filename
      << " " << android::base::Join(code_paths, ':');
  VLOG(profiler) << "Reference profile is: " << ref_profile_filename;

  if (profile_output_filename.empty()) {
    LOG(WARNING) << "JIT profile information will not be recorded: profile filename is empty.";
    return;
  }
  if (code_paths.empty()) {
    LOG(WARNING) << "JIT profile information will not be recorded: code paths is empty.";
    return;
  }

  jit_->RegisterAppInfo(internal_code_type, app_info_.GetCompilerFilter(code_paths[0]));

  // Framework calls this method for all split APKs. Ignore the calls for the ones with no dex code
  // so that we don't unnecessarily create profiles for them or write bootclasspath profiling info
  // to those profiles.
  bool has_code = false;
  for (const std::string& path : code_paths) {
    std::string error_msg;
    std::optional<uint32_t> checksum;
    std::vector<std::string> dex_locations;
    DexFileLoader loader(path);
    if (!loader.GetMultiDexChecksum(&checksum, &error_msg)) {
      LOG(WARNING) << error_msg;
      continue;
    }
    if (checksum.has_value()) {
      has_code = true;
      break;
    }
  }
  if (!has_code) {
    VLOG(profiler) << ART_FORMAT(
        "JIT profile information will not be recorded: no dex code in '{}'.",
        android::base::Join(code_paths, ','));
    return;
  }

  jit_->StartProfileSaver(profile_output_filename,
                          code_paths,
                          ref_profile_filename,
                          AppInfo::FromVMRuntimeConstants(code_type));
}

void Runtime::SetFaultMessage(const std::string& message) {
  std::string* new_msg = new std::string(message);
  std::string* cur_msg = fault_message_.exchange(new_msg);
  delete cur_msg;
}

std::string Runtime::GetFaultMessage() {
  // Retrieve the message. Temporarily replace with null so that SetFaultMessage will not delete
  // the string in parallel.
  std::string* cur_msg = fault_message_.exchange(nullptr);

  // Make a copy of the string.
  std::string ret = cur_msg == nullptr ? "" : *cur_msg;

  // Put the message back if it hasn't been updated.
  std::string* null_str = nullptr;
  if (!fault_message_.compare_exchange_strong(null_str, cur_msg)) {
    // Already replaced.
    delete cur_msg;
  }

  return ret;
}

void Runtime::AddCurrentRuntimeFeaturesAsDex2OatArguments(std::vector<std::string>* argv)
    const {
  if (GetInstrumentation()->InterpretOnly()) {
    argv->push_back("--compiler-filter=verify");
  }

  // Make the dex2oat instruction set match that of the launching runtime. If we have multiple
  // architecture support, dex2oat may be compiled as a different instruction-set than that
  // currently being executed.
  std::string instruction_set("--instruction-set=");
  // The dex2oat instruction set should match the runtime's target ISA.
  instruction_set += GetInstructionSetString(kRuntimeQuickCodeISA);
  argv->push_back(instruction_set);

  if (InstructionSetFeatures::IsRuntimeDetectionSupported()) {
    argv->push_back("--instruction-set-features=runtime");
  } else {
    std::unique_ptr<const InstructionSetFeatures> features(
        InstructionSetFeatures::FromCppDefines());
    std::string feature_string("--instruction-set-features=");
    feature_string += features->GetFeatureString();
    argv->push_back(feature_string);
  }
}

void Runtime::CreateJit() {
  DCHECK(jit_code_cache_ == nullptr);
  DCHECK(jit_ == nullptr);
  if (kIsDebugBuild && GetInstrumentation()->IsForcedInterpretOnly()) {
    DCHECK(!jit_options_->UseJitCompilation());
  }

  if (!jit_options_->UseJitCompilation() && !jit_options_->GetSaveProfilingInfo()) {
    return;
  }

  if (IsSafeMode()) {
    LOG(INFO) << "Not creating JIT because of SafeMode.";
    return;
  }

  std::string error_msg;
  bool profiling_only = !jit_options_->UseJitCompilation();
  jit_code_cache_.reset(jit::JitCodeCache::Create(profiling_only,
                                                  /*rwx_memory_allowed=*/ true,
                                                  IsZygote(),
                                                  &error_msg));
  if (jit_code_cache_.get() == nullptr) {
    LOG(WARNING) << "Failed to create JIT Code Cache: " << error_msg;
    return;
  }

  jit_ = jit::Jit::Create(jit_code_cache_.get(), jit_options_.get());
  jit_->CreateThreadPool();
}

bool Runtime::CanRelocate() const {
  return !IsAotCompiler();
}

bool Runtime::IsCompilingBootImage() const {
  return IsCompiler() && compiler_callbacks_->IsBootImage();
}

void Runtime::SetResolutionMethod(ArtMethod* method) {
  CHECK(method != nullptr);
  CHECK(method->IsRuntimeMethod()) << method;
  resolution_method_ = method;
}

void Runtime::SetImtUnimplementedMethod(ArtMethod* method) {
  CHECK(method != nullptr);
  CHECK(method->IsRuntimeMethod());
  imt_unimplemented_method_ = method;
}

void Runtime::FixupConflictTables() {
  // We can only do this after the class linker is created.
  const PointerSize pointer_size = GetClassLinker()->GetImagePointerSize();
  if (imt_unimplemented_method_->GetImtConflictTable(pointer_size) == nullptr) {
    imt_unimplemented_method_->SetImtConflictTable(
        ClassLinker::CreateImtConflictTable(/*count=*/0u, GetLinearAlloc(), pointer_size),
        pointer_size);
  }
  if (imt_conflict_method_->GetImtConflictTable(pointer_size) == nullptr) {
    imt_conflict_method_->SetImtConflictTable(
          ClassLinker::CreateImtConflictTable(/*count=*/0u, GetLinearAlloc(), pointer_size),
          pointer_size);
  }
}

void Runtime::DisableVerifier() {
  verify_ = verifier::VerifyMode::kNone;
}

bool Runtime::IsVerificationEnabled() const {
  return verify_ == verifier::VerifyMode::kEnable ||
      verify_ == verifier::VerifyMode::kSoftFail;
}

bool Runtime::IsVerificationSoftFail() const {
  return verify_ == verifier::VerifyMode::kSoftFail;
}

bool Runtime::IsAsyncDeoptimizeable(ArtMethod* method, uintptr_t code) const {
  if (OatQuickMethodHeader::NterpMethodHeader != nullptr) {
    if (OatQuickMethodHeader::NterpMethodHeader->Contains(code)) {
      return true;
    }
  }

  // We only support async deopt (ie the compiled code is not explicitly asking for
  // deopt, but something else like the debugger) in debuggable JIT code.
  // We could look at the oat file where `code` is being defined,
  // and check whether it's been compiled debuggable, but we decided to
  // only rely on the JIT for debuggable apps.
  // The JIT-zygote is not debuggable so we need to be sure to exclude code from the non-private
  // region as well.
  if (GetJit() != nullptr &&
      GetJit()->GetCodeCache()->PrivateRegionContainsPc(reinterpret_cast<const void*>(code))) {
    // If the code is JITed code then check if it was compiled as debuggable.
    const OatQuickMethodHeader* header = method->GetOatQuickMethodHeader(code);
    return CodeInfo::IsDebuggable(header->GetOptimizedCodeInfoPtr());
  }

  return false;
}


LinearAlloc* Runtime::CreateLinearAlloc() {
  ArenaPool* pool = linear_alloc_arena_pool_.get();
  return pool != nullptr
      ? new LinearAlloc(pool, gUseUserfaultfd)
      : new LinearAlloc(arena_pool_.get(), /*track_allocs=*/ false);
}

class Runtime::SetupLinearAllocForZygoteFork : public AllocatorVisitor {
 public:
  explicit SetupLinearAllocForZygoteFork(Thread* self) : self_(self) {}

  bool Visit(LinearAlloc* alloc) override {
    alloc->SetupForPostZygoteFork(self_);
    return true;
  }

 private:
  Thread* self_;
};

void Runtime::SetupLinearAllocForPostZygoteFork(Thread* self) {
  if (gUseUserfaultfd) {
    // Setup all the linear-allocs out there for post-zygote fork. This will
    // basically force the arena allocator to ask for a new arena for the next
    // allocation. All arenas allocated from now on will be in the userfaultfd
    // visited space.
    if (GetLinearAlloc() != nullptr) {
      GetLinearAlloc()->SetupForPostZygoteFork(self);
    }
    if (GetStartupLinearAlloc() != nullptr) {
      GetStartupLinearAlloc()->SetupForPostZygoteFork(self);
    }
    {
      Locks::mutator_lock_->AssertNotHeld(self);
      ReaderMutexLock mu2(self, *Locks::mutator_lock_);
      ReaderMutexLock mu3(self, *Locks::classlinker_classes_lock_);
      SetupLinearAllocForZygoteFork visitor(self);
      GetClassLinker()->VisitAllocators(&visitor);
    }
    static_cast<GcVisitedArenaPool*>(GetLinearAllocArenaPool())->SetupPostZygoteMode();
  }
}

double Runtime::GetHashTableMinLoadFactor() const {
  return is_low_memory_mode_ ? kLowMemoryMinLoadFactor : kNormalMinLoadFactor;
}

double Runtime::GetHashTableMaxLoadFactor() const {
  return is_low_memory_mode_ ? kLowMemoryMaxLoadFactor : kNormalMaxLoadFactor;
}

void Runtime::UpdateProcessState(ProcessState process_state) {
  ProcessState old_process_state = process_state_;
  process_state_ = process_state;
  was_ever_jank_perceptible_ |= InJankPerceptibleProcessState();
  GetHeap()->UpdateProcessState(old_process_state, process_state);

  // When the application switches to the foreground, lock contention on classlinker_classes_lock_
  // and priority inversion occasionally occur. Delay profile saving on hot/warm startup can reduce
  // the occurrence of the problem. This feature depends on whether the
  // jank_perceptible_narrow flag is enabled, which is only implemented and enabled on Android 16
  // and above.On older Android versions, the process is always in kProcessStateJankPerceptible as
  // long as it is not cached, so the profile saving delay is not applicable.
  if (IsSdkVersionSetAndAtLeast(sdk_version_, SdkVersion::kB) &&
      process_state == kProcessStateJankPerceptible) {
    ProfileSaver::NotifyDelayProfileSaving();
  }
}

void Runtime::RegisterSensitiveThread() const {
  Thread::SetJitSensitiveThread();
}

// Returns true if JIT compilations are enabled. GetJit() will be not null in this case.
bool Runtime::UseJitCompilation() const {
  return (jit_ != nullptr) && jit_->UseJitCompilation();
}

void Runtime::EnvSnapshot::TakeSnapshot() {
  char** env = GetEnviron();
  for (size_t i = 0; env[i] != nullptr; ++i) {
    name_value_pairs_.emplace_back(new std::string(env[i]));
  }
  // The strings in name_value_pairs_ retain ownership of the c_str, but we assign pointers
  // for quick use by GetSnapshot.  This avoids allocation and copying cost at Exec.
  c_env_vector_.reset(new char*[name_value_pairs_.size() + 1]);
  for (size_t i = 0; env[i] != nullptr; ++i) {
    c_env_vector_[i] = const_cast<char*>(name_value_pairs_[i]->c_str());
  }
  c_env_vector_[name_value_pairs_.size()] = nullptr;
}

char** Runtime::EnvSnapshot::GetSnapshot() const {
  return c_env_vector_.get();
}

void Runtime::AddSystemWeakHolder(gc::AbstractSystemWeakHolder* holder) {
  gc::ScopedGCCriticalSection gcs(Thread::Current(),
                                  gc::kGcCauseAddRemoveSystemWeakHolder,
                                  gc::kCollectorTypeAddRemoveSystemWeakHolder);
  // Note: The ScopedGCCriticalSection also ensures that the rest of the function is in
  //       a critical section.
  system_weak_holders_.push_back(holder);
}

void Runtime::RemoveSystemWeakHolder(gc::AbstractSystemWeakHolder* holder) {
  gc::ScopedGCCriticalSection gcs(Thread::Current(),
                                  gc::kGcCauseAddRemoveSystemWeakHolder,
                                  gc::kCollectorTypeAddRemoveSystemWeakHolder);
  auto it = std::find(system_weak_holders_.begin(), system_weak_holders_.end(), holder);
  if (it != system_weak_holders_.end()) {
    system_weak_holders_.erase(it);
  }
}

RuntimeCallbacks* Runtime::GetRuntimeCallbacks() {
  return callbacks_.get();
}

// Used to update boot image to not use AOT code. This is used when transitioning the runtime to
// java debuggable. This visitor re-initializes the entry points without using AOT code. This also
// disables shared hotness counters so the necessary methods can be JITed more efficiently.
class DeoptimizeBootImageClassVisitor : public ClassVisitor {
 public:
  explicit DeoptimizeBootImageClassVisitor(instrumentation::Instrumentation* instrumentation)
      : instrumentation_(instrumentation) {}

  bool operator()(ObjPtr<mirror::Class> klass) override REQUIRES(Locks::mutator_lock_) {
    DCHECK(Locks::mutator_lock_->IsExclusiveHeld(Thread::Current()));
    auto pointer_size = Runtime::Current()->GetClassLinker()->GetImagePointerSize();
    for (auto& m : klass->GetMethods(pointer_size)) {
      const void* code = m.GetEntryPointFromQuickCompiledCode();
      if (!m.IsInvokable()) {
        continue;
      }
      // For java debuggable runtimes we also deoptimize native methods. For other cases (boot
      // image profiling) we don't need to deoptimize native methods. If this changes also
      // update Instrumentation::CanUseAotCode.
      bool deoptimize_native_methods = Runtime::Current()->IsJavaDebuggable();
      if (Runtime::Current()->GetHeap()->IsInBootImageOatFile(code) &&
          (!m.IsNative() || deoptimize_native_methods) &&
          !m.IsProxyMethod()) {
        instrumentation_->ReinitializeMethodsCode(&m);
      }

      if (Runtime::Current()->GetJit() != nullptr &&
          Runtime::Current()->GetJit()->GetCodeCache()->IsInZygoteExecSpace(code) &&
          (!m.IsNative() || deoptimize_native_methods)) {
        DCHECK(!m.IsProxyMethod());
        instrumentation_->ReinitializeMethodsCode(&m);
      }

      if (m.IsPreCompiled()) {
        // Precompilation is incompatible with debuggable, so clear the flag
        // and update the entrypoint in case it has been compiled.
        m.ClearPreCompiled();
        instrumentation_->ReinitializeMethodsCode(&m);
      }

      // Clear MemorySharedAccessFlags so the boot class methods can be JITed better.
      m.ClearMemorySharedMethod();
    }
    return true;
  }

 private:
  instrumentation::Instrumentation* const instrumentation_;
};

void Runtime::SetRuntimeDebugState(RuntimeDebugState state) {
  if (state != RuntimeDebugState::kJavaDebuggableAtInit) {
    // We never change the state if we started as a debuggable runtime.
    DCHECK(runtime_debug_state_ != RuntimeDebugState::kJavaDebuggableAtInit);
  }
  runtime_debug_state_ = state;
}

void Runtime::DeoptimizeBootImage() {
  // If we've already started and we are setting this runtime to debuggable,
  // we patch entry points of methods in boot image to interpreter bridge, as
  // boot image code may be AOT compiled as not debuggable.
  DeoptimizeBootImageClassVisitor visitor(GetInstrumentation());
  GetClassLinker()->VisitClasses(&visitor);
  jit::Jit* jit = GetJit();
  if (jit != nullptr) {
    // Code previously compiled may not be compiled debuggable.
    jit->GetCodeCache()->TransitionToDebuggable();
  }
}

Runtime::ScopedThreadPoolUsage::ScopedThreadPoolUsage()
    : thread_pool_(Runtime::Current()->AcquireThreadPool()) {}

Runtime::ScopedThreadPoolUsage::~ScopedThreadPoolUsage() {
  Runtime::Current()->ReleaseThreadPool();
}

bool Runtime::DeleteThreadPool() {
  // Make sure workers are started to prevent thread shutdown errors.
  WaitForThreadPoolWorkersToStart();
  std::unique_ptr<ThreadPool> thread_pool;
  {
    MutexLock mu(Thread::Current(), *Locks::runtime_thread_pool_lock_);
    if (thread_pool_ref_count_ == 0) {
      thread_pool = std::move(thread_pool_);
    }
  }
  return thread_pool != nullptr;
}

ThreadPool* Runtime::AcquireThreadPool() {
  MutexLock mu(Thread::Current(), *Locks::runtime_thread_pool_lock_);
  ++thread_pool_ref_count_;
  return thread_pool_.get();
}

void Runtime::ReleaseThreadPool() {
  MutexLock mu(Thread::Current(), *Locks::runtime_thread_pool_lock_);
  CHECK_GT(thread_pool_ref_count_, 0u);
  --thread_pool_ref_count_;
}

void Runtime::WaitForThreadPoolWorkersToStart() {
  // Need to make sure workers are created before deleting the pool.
  ScopedThreadPoolUsage stpu;
  if (stpu.GetThreadPool() != nullptr) {
    stpu.GetThreadPool()->WaitForWorkersToBeCreated();
  }
}

void Runtime::ResetStartupCompleted() {
  startup_completed_.store(false, std::memory_order_seq_cst);
}

bool Runtime::NotifyStartupCompleted() {
  DCHECK(!IsZygote());
  bool expected = false;
  if (!startup_completed_.compare_exchange_strong(expected, true, std::memory_order_seq_cst)) {
    // Right now NotifyStartupCompleted will be called up to twice, once from profiler and up to
    // once externally. For this reason there are no asserts.
    return false;
  }

  VLOG(startup) << app_info_;

  ProfileSaver::NotifyStartupCompleted();

  if (AreMetricsInitialized()) {
    metrics_reporter_->NotifyStartupCompleted();
  }
  return true;
}

void Runtime::NotifyDexFileLoaded() {
  if (AreMetricsInitialized()) {
    metrics_reporter_->NotifyAppInfoUpdated(&app_info_);
  }
}

bool Runtime::GetStartupCompleted() const {
  return startup_completed_.load(std::memory_order_seq_cst);
}

void Runtime::SetSignalHookDebuggable(bool value) {
  SkipAddSignalHandler(value);
}

void Runtime::SetJniIdType(JniIdType t) {
  CHECK(CanSetJniIdType()) << "Not allowed to change id type!";
  if (t == GetJniIdType()) {
    return;
  }
  jni_ids_indirection_ = t;
  JNIEnvExt::ResetFunctionTable();
  WellKnownClasses::HandleJniIdTypeChange(Thread::Current()->GetJniEnv());
}

bool Runtime::IsSystemServerProfiled() const {
  return IsSystemServer() && jit_options_->GetSaveProfilingInfo();
}

bool Runtime::GetOatFilesExecutable() const {
  return !IsAotCompiler() && !IsSystemServerProfiled();
}

size_t Runtime::MadviseFileForRange(size_t madvise_size_limit_bytes,
                                    size_t map_size_bytes,
                                    const uint8_t* map_begin,
                                    const uint8_t* map_end,
                                    const std::string& file_name,
                                    int optional_fd) {
  // TODO(b/359932564): Fix map_size_bytes adjustment to account for map_begin alignment.
  map_begin = AlignDown(map_begin, gPageSize);
  map_size_bytes = RoundUp(map_size_bytes, gPageSize);

  size_t madvised_bytes = 0;
  size_t target_size_bytes = std::min<size_t>(map_size_bytes, madvise_size_limit_bytes);
  if (target_size_bytes > 0) {
    // Based on requested size (target_size_bytes)
    const uint8_t* target_pos = map_begin + target_size_bytes;

    // Clamp endOfFile if its past map_end
    if (target_pos > map_end) {
      target_pos = map_end;
    }

    // Apply madvise(WILLNEED) to the mapped range up to target_pos, in ideal transfer sized chunks.
    // Note: madvise(WILLNEED) will prefetch max(fd readahead size, optimal block size for device)
    // per call, hence the need for chunks. 128KB is a sensible default.
    static constexpr size_t kDefaultIoTransferSizeBytes = 128 * KB;
    static constexpr size_t kFadviseSafeIoTransferSizeBytes = 256 * KB;
    static constexpr size_t kFadviseLargeIoTransferSizeBytes = 16 * MB;
    static constexpr size_t kFadviseThresholdBytes = kFadviseSafeIoTransferSizeBytes;

    size_t io_transfer_size_bytes = kDefaultIoTransferSizeBytes;

    // fadvise req 1: The target size is sufficiently large *and* we have a valid backing FD.
    // This unlocks larger readahead and improves IO, but minimizes extra syscalls for small reads.
    const bool should_fadvise =
        com::android::art::rw::flags::madvise_optimized_readahead() &&
        optional_fd >= 0 &&
        target_size_bytes > kFadviseThresholdBytes;
    if (should_fadvise) {
      // fadvise req 2: File is in f2fs-backed /data partition for larger readahead support.
      // Otherwise, fall back to a more conservative but universally supported window.
      if (LocationSupportsLargeReadahead(file_name)) {
        io_transfer_size_bytes = kFadviseLargeIoTransferSizeBytes;
      } else {
        io_transfer_size_bytes = kFadviseSafeIoTransferSizeBytes;
      }
      // Note: We temporarily hint the entire file, resetting after madvise completes.
      posix_fadvise(optional_fd, 0, 0, POSIX_FADV_SEQUENTIAL);
    }

    size_t chunks = (target_pos - map_begin + io_transfer_size_bytes - 1) / io_transfer_size_bytes;
    SCOPED_TRACE << "madvising " << file_name
                 << " size=" << target_size_bytes
                 << " chunks=" << chunks;

    for (const uint8_t* madvise_start = map_begin;
         madvise_start < target_pos;
         madvise_start += io_transfer_size_bytes) {
      void* madvise_addr = const_cast<void*>(reinterpret_cast<const void*>(madvise_start));
      size_t madvise_length =
          std::min(io_transfer_size_bytes, static_cast<size_t>(target_pos - madvise_start));
      int status = madvise(madvise_addr, madvise_length, MADV_WILLNEED);
      // In case of error we stop madvising rest of the file
      if (status < 0) {
        LOG(ERROR) << "Failed to madvise file " << file_name
                   << " for size:" << map_size_bytes
                   << ": " << strerror(errno);
        break;
      }
      madvised_bytes += madvise_length;
    }
    if (should_fadvise) {
      // Restore the default file-backed readahead behavior.
      posix_fadvise(optional_fd, 0, 0, POSIX_FADV_NORMAL);
    }
  }

  DCHECK_LE(madvised_bytes, madvise_size_limit_bytes)
      << "Madvise should not have advised more than the requested size.";
  return madvised_bytes;
}

bool Runtime::ShouldMadviseForAppStartup(const char* dex_location) {
  // Short-circuit the madvise optimization for background processes. This
  // avoids IO and memory contention with foreground processes, particularly
  // those involving app startup.
  // Note: We can only safely short-circuit the madvise on T+, as it requires
  // the framework to always immediately notify ART of process states.
  const bool accurate_process_state_at_startup =
      IsSdkVersionSetAndAtLeast(sdk_version_, SdkVersion::kT);
  if (accurate_process_state_at_startup && !InJankPerceptibleProcessState()) {
    return false;
  }

  if (!app_info_.HasRegisteredAppInfo()) {
    // Conservatively madvise everything until app registration is complete and we can definitively
    // distinguish between primary and secondary dex artifacts.
    return true;
  }

  // Only madvise primary/split dex artifacts to reduce unnecessary faulting of unowned code that
  // may not be on the critical path.
  const AppInfo::CodeType code_type = app_info_.GetRegisteredCodeType(dex_location);
  return code_type == AppInfo::CodeType::kPrimaryApk || code_type == AppInfo::CodeType::kSplitApk;
}

// Return whether a boot image has a profile. This means we'll need to pre-JIT
// methods in that profile for performance.
bool Runtime::HasImageWithProfile() const {
  for (gc::space::ImageSpace* space : GetHeap()->GetBootImageSpaces()) {
    if (!space->GetProfileFiles().empty()) {
      return true;
    }
  }
  return false;
}

void Runtime::AppendToBootClassPath(const std::string& filename, const std::string& location) {
  DCHECK(!DexFileLoader::IsMultiDexLocation(filename));
  boot_class_path_.push_back(filename);
  if (!boot_class_path_locations_.empty()) {
    DCHECK(!DexFileLoader::IsMultiDexLocation(location));
    boot_class_path_locations_.push_back(location);
  }
}

void Runtime::AppendToBootClassPath(
    const std::string& filename,
    const std::string& location,
    const std::vector<std::unique_ptr<const art::DexFile>>& dex_files) {
  AppendToBootClassPath(filename, location);
  ScopedObjectAccess soa(Thread::Current());
  for (const std::unique_ptr<const art::DexFile>& dex_file : dex_files) {
    // The first element must not be at a multi-dex location, while other elements must be.
    DCHECK_NE(DexFileLoader::IsMultiDexLocation(dex_file->GetLocation()),
              dex_file.get() == dex_files.begin()->get());
    GetClassLinker()->AppendToBootClassPath(Thread::Current(), dex_file.get());
  }
}

void Runtime::AppendToBootClassPath(const std::string& filename,
                                    const std::string& location,
                                    const std::vector<const art::DexFile*>& dex_files) {
  AppendToBootClassPath(filename, location);
  ScopedObjectAccess soa(Thread::Current());
  for (const art::DexFile* dex_file : dex_files) {
    // The first element must not be at a multi-dex location, while other elements must be.
    DCHECK_NE(DexFileLoader::IsMultiDexLocation(dex_file->GetLocation()),
              dex_file == *dex_files.begin());
    GetClassLinker()->AppendToBootClassPath(Thread::Current(), dex_file);
  }
}

void Runtime::AppendToBootClassPath(
    const std::string& filename,
    const std::string& location,
    const std::vector<std::pair<const art::DexFile*, ObjPtr<mirror::DexCache>>>&
        dex_files_and_cache) {
  AppendToBootClassPath(filename, location);
  ScopedObjectAccess soa(Thread::Current());
  for (const auto& [dex_file, dex_cache] : dex_files_and_cache) {
    // The first element must not be at a multi-dex location, while other elements must be.
    DCHECK_NE(DexFileLoader::IsMultiDexLocation(dex_file->GetLocation()),
              dex_file == dex_files_and_cache.begin()->first);
    GetClassLinker()->AppendToBootClassPath(dex_file, dex_cache);
  }
}

void Runtime::AddExtraBootDexFiles(const std::string& filename,
                                   const std::string& location,
                                   std::vector<std::unique_ptr<const art::DexFile>>&& dex_files) {
  AppendToBootClassPath(filename, location);
  ScopedObjectAccess soa(Thread::Current());
  if (kIsDebugBuild) {
    for (const std::unique_ptr<const art::DexFile>& dex_file : dex_files) {
      // The first element must not be at a multi-dex location, while other elements must be.
      DCHECK_NE(DexFileLoader::IsMultiDexLocation(dex_file->GetLocation()),
                dex_file.get() == dex_files.begin()->get());
    }
  }
  GetClassLinker()->AddExtraBootDexFiles(Thread::Current(), std::move(dex_files));
}

void Runtime::DCheckNoTransactionCheckAllowed() {
  if (kIsDebugBuild) {
    Thread* self = Thread::Current();
    if (self != nullptr) {
      self->AssertNoTransactionCheckAllowed();
    }
  }
}

}  // namespace art