re2 1.20.0 → 1.20.2

package/vendor/abseil-cpp/absl/debugging/symbolize_elf.inc

@@ -0,0 +1,1637 @@
+// Copyright 2018 The Abseil Authors.
+//
+// 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
+//
+//      https://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.
+
+// This library provides Symbolize() function that symbolizes program
+// counters to their corresponding symbol names on linux platforms.
+// This library has a minimal implementation of an ELF symbol table
+// reader (i.e. it doesn't depend on libelf, etc.).
+//
+// The algorithm used in Symbolize() is as follows.
+//
+//   1. Go through a list of maps in /proc/self/maps and find the map
+//   containing the program counter.
+//
+//   2. Open the mapped file and find a regular symbol table inside.
+//   Iterate over symbols in the symbol table and look for the symbol
+//   containing the program counter.  If such a symbol is found,
+//   obtain the symbol name, and demangle the symbol if possible.
+//   If the symbol isn't found in the regular symbol table (binary is
+//   stripped), try the same thing with a dynamic symbol table.
+//
+// Note that Symbolize() is originally implemented to be used in
+// signal handlers, hence it doesn't use malloc() and other unsafe
+// operations.  It should be both thread-safe and async-signal-safe.
+//
+// Implementation note:
+//
+// We don't use heaps but only use stacks.  We want to reduce the
+// stack consumption so that the symbolizer can run on small stacks.
+//
+// Here are some numbers collected with GCC 4.1.0 on x86:
+// - sizeof(Elf32_Sym)  = 16
+// - sizeof(Elf32_Shdr) = 40
+// - sizeof(Elf64_Sym)  = 24
+// - sizeof(Elf64_Shdr) = 64
+//
+// This implementation is intended to be async-signal-safe but uses some
+// functions which are not guaranteed to be so, such as memchr() and
+// memmove().  We assume they are async-signal-safe.
+
+#include <dlfcn.h>
+#include <elf.h>
+#include <fcntl.h>
+#include <link.h>  // For ElfW() macro.
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <cerrno>
+#include <cinttypes>
+#include <climits>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+#include "absl/base/casts.h"
+#include "absl/base/dynamic_annotations.h"
+#include "absl/base/internal/low_level_alloc.h"
+#include "absl/base/internal/raw_logging.h"
+#include "absl/base/internal/spinlock.h"
+#include "absl/base/port.h"
+#include "absl/debugging/internal/demangle.h"
+#include "absl/debugging/internal/vdso_support.h"
+#include "absl/strings/string_view.h"
+
+#if defined(__FreeBSD__) && !defined(ElfW)
+#define ElfW(x) __ElfN(x)
+#endif
+
+namespace absl {
+ABSL_NAMESPACE_BEGIN
+
+// Value of argv[0]. Used by MaybeInitializeObjFile().
+static char *argv0_value = nullptr;
+
+void InitializeSymbolizer(const char *argv0) {
+#ifdef ABSL_HAVE_VDSO_SUPPORT
+  // We need to make sure VDSOSupport::Init() is called before any setuid or
+  // chroot calls, so InitializeSymbolizer() should be called very early in the
+  // life of a program.
+  absl::debugging_internal::VDSOSupport::Init();
+#endif
+  if (argv0_value != nullptr) {
+    free(argv0_value);
+    argv0_value = nullptr;
+  }
+  if (argv0 != nullptr && argv0[0] != '\0') {
+    argv0_value = strdup(argv0);
+  }
+}
+
+namespace debugging_internal {
+namespace {
+
+// Re-runs fn until it doesn't cause EINTR.
+#define NO_INTR(fn) \
+  do {              \
+  } while ((fn) < 0 && errno == EINTR)
+
+// On Linux, ELF_ST_* are defined in <linux/elf.h>.  To make this portable
+// we define our own ELF_ST_BIND and ELF_ST_TYPE if not available.
+#ifndef ELF_ST_BIND
+#define ELF_ST_BIND(info) (((unsigned char)(info)) >> 4)
+#endif
+
+#ifndef ELF_ST_TYPE
+#define ELF_ST_TYPE(info) (((unsigned char)(info)) & 0xF)
+#endif
+
+// Some platforms use a special .opd section to store function pointers.
+const char kOpdSectionName[] = ".opd";
+
+#if (defined(__powerpc__) && !(_CALL_ELF > 1)) || defined(__ia64)
+// Use opd section for function descriptors on these platforms, the function
+// address is the first word of the descriptor.
+enum { kPlatformUsesOPDSections = 1 };
+#else  // not PPC or IA64
+enum { kPlatformUsesOPDSections = 0 };
+#endif
+
+// This works for PowerPC & IA64 only.  A function descriptor consist of two
+// pointers and the first one is the function's entry.
+const size_t kFunctionDescriptorSize = sizeof(void *) * 2;
+
+const int kMaxDecorators = 10;  // Seems like a reasonable upper limit.
+
+struct InstalledSymbolDecorator {
+  SymbolDecorator fn;
+  void *arg;
+  int ticket;
+};
+
+int g_num_decorators;
+InstalledSymbolDecorator g_decorators[kMaxDecorators];
+
+struct FileMappingHint {
+  const void *start;
+  const void *end;
+  uint64_t offset;
+  const char *filename;
+};
+
+// Protects g_decorators.
+// We are using SpinLock and not a Mutex here, because we may be called
+// from inside Mutex::Lock itself, and it prohibits recursive calls.
+// This happens in e.g. base/stacktrace_syscall_unittest.
+// Moreover, we are using only TryLock(), if the decorator list
+// is being modified (is busy), we skip all decorators, and possibly
+// loose some info. Sorry, that's the best we could do.
+ABSL_CONST_INIT absl::base_internal::SpinLock g_decorators_mu(
+    absl::kConstInit, absl::base_internal::SCHEDULE_KERNEL_ONLY);
+
+const int kMaxFileMappingHints = 8;
+int g_num_file_mapping_hints;
+FileMappingHint g_file_mapping_hints[kMaxFileMappingHints];
+// Protects g_file_mapping_hints.
+ABSL_CONST_INIT absl::base_internal::SpinLock g_file_mapping_mu(
+    absl::kConstInit, absl::base_internal::SCHEDULE_KERNEL_ONLY);
+
+// Async-signal-safe function to zero a buffer.
+// memset() is not guaranteed to be async-signal-safe.
+static void SafeMemZero(void* p, size_t size) {
+  unsigned char *c = static_cast<unsigned char *>(p);
+  while (size--) {
+    *c++ = 0;
+  }
+}
+
+struct ObjFile {
+  ObjFile()
+      : filename(nullptr),
+        start_addr(nullptr),
+        end_addr(nullptr),
+        offset(0),
+        fd(-1),
+        elf_type(-1) {
+    SafeMemZero(&elf_header, sizeof(elf_header));
+    SafeMemZero(&phdr[0], sizeof(phdr));
+  }
+
+  char *filename;
+  const void *start_addr;
+  const void *end_addr;
+  uint64_t offset;
+
+  // The following fields are initialized on the first access to the
+  // object file.
+  int fd;
+  int elf_type;
+  ElfW(Ehdr) elf_header;
+
+  // PT_LOAD program header describing executable code.
+  // Normally we expect just one, but SWIFT binaries have two.
+  // CUDA binaries have 3 (see cr/473913254 description).
+  std::array<ElfW(Phdr), 4> phdr;
+};
+
+// Build 4-way associative cache for symbols. Within each cache line, symbols
+// are replaced in LRU order.
+enum {
+  ASSOCIATIVITY = 4,
+};
+struct SymbolCacheLine {
+  const void *pc[ASSOCIATIVITY];
+  char *name[ASSOCIATIVITY];
+
+  // age[i] is incremented when a line is accessed. it's reset to zero if the
+  // i'th entry is read.
+  uint32_t age[ASSOCIATIVITY];
+};
+
+// ---------------------------------------------------------------
+// An async-signal-safe arena for LowLevelAlloc
+static std::atomic<base_internal::LowLevelAlloc::Arena *> g_sig_safe_arena;
+
+static base_internal::LowLevelAlloc::Arena *SigSafeArena() {
+  return g_sig_safe_arena.load(std::memory_order_acquire);
+}
+
+static void InitSigSafeArena() {
+  if (SigSafeArena() == nullptr) {
+    base_internal::LowLevelAlloc::Arena *new_arena =
+        base_internal::LowLevelAlloc::NewArena(
+            base_internal::LowLevelAlloc::kAsyncSignalSafe);
+    base_internal::LowLevelAlloc::Arena *old_value = nullptr;
+    if (!g_sig_safe_arena.compare_exchange_strong(old_value, new_arena,
+                                                  std::memory_order_release,
+                                                  std::memory_order_relaxed)) {
+      // We lost a race to allocate an arena; deallocate.
+      base_internal::LowLevelAlloc::DeleteArena(new_arena);
+    }
+  }
+}
+
+// ---------------------------------------------------------------
+// An AddrMap is a vector of ObjFile, using SigSafeArena() for allocation.
+
+class AddrMap {
+ public:
+  AddrMap() : size_(0), allocated_(0), obj_(nullptr) {}
+  ~AddrMap() { base_internal::LowLevelAlloc::Free(obj_); }
+  size_t Size() const { return size_; }
+  ObjFile *At(size_t i) { return &obj_[i]; }
+  ObjFile *Add();
+  void Clear();
+
+ private:
+  size_t size_;       // count of valid elements (<= allocated_)
+  size_t allocated_;  // count of allocated elements
+  ObjFile *obj_;      // array of allocated_ elements
+  AddrMap(const AddrMap &) = delete;
+  AddrMap &operator=(const AddrMap &) = delete;
+};
+
+void AddrMap::Clear() {
+  for (size_t i = 0; i != size_; i++) {
+    At(i)->~ObjFile();
+  }
+  size_ = 0;
+}
+
+ObjFile *AddrMap::Add() {
+  if (size_ == allocated_) {
+    size_t new_allocated = allocated_ * 2 + 50;
+    ObjFile *new_obj_ =
+        static_cast<ObjFile *>(base_internal::LowLevelAlloc::AllocWithArena(
+            new_allocated * sizeof(*new_obj_), SigSafeArena()));
+    if (obj_) {
+      memcpy(new_obj_, obj_, allocated_ * sizeof(*new_obj_));
+      base_internal::LowLevelAlloc::Free(obj_);
+    }
+    obj_ = new_obj_;
+    allocated_ = new_allocated;
+  }
+  return new (&obj_[size_++]) ObjFile;
+}
+
+// ---------------------------------------------------------------
+
+enum FindSymbolResult { SYMBOL_NOT_FOUND = 1, SYMBOL_TRUNCATED, SYMBOL_FOUND };
+
+class Symbolizer {
+ public:
+  Symbolizer();
+  ~Symbolizer();
+  const char *GetSymbol(const void *const pc);
+
+ private:
+  char *CopyString(const char *s) {
+    size_t len = strlen(s);
+    char *dst = static_cast<char *>(
+        base_internal::LowLevelAlloc::AllocWithArena(len + 1, SigSafeArena()));
+    ABSL_RAW_CHECK(dst != nullptr, "out of memory");
+    memcpy(dst, s, len + 1);
+    return dst;
+  }
+  ObjFile *FindObjFile(const void *const start,
+                       size_t size) ABSL_ATTRIBUTE_NOINLINE;
+  static bool RegisterObjFile(const char *filename,
+                              const void *const start_addr,
+                              const void *const end_addr, uint64_t offset,
+                              void *arg);
+  SymbolCacheLine *GetCacheLine(const void *const pc);
+  const char *FindSymbolInCache(const void *const pc);
+  const char *InsertSymbolInCache(const void *const pc, const char *name);
+  void AgeSymbols(SymbolCacheLine *line);
+  void ClearAddrMap();
+  FindSymbolResult GetSymbolFromObjectFile(const ObjFile &obj,
+                                           const void *const pc,
+                                           const ptrdiff_t relocation,
+                                           char *out, size_t out_size,
+                                           char *tmp_buf, size_t tmp_buf_size);
+  const char *GetUncachedSymbol(const void *pc);
+
+  enum {
+    SYMBOL_BUF_SIZE = 3072,
+    TMP_BUF_SIZE = 1024,
+    SYMBOL_CACHE_LINES = 128,
+  };
+
+  AddrMap addr_map_;
+
+  bool ok_;
+  bool addr_map_read_;
+
+  char symbol_buf_[SYMBOL_BUF_SIZE];
+
+  // tmp_buf_ will be used to store arrays of ElfW(Shdr) and ElfW(Sym)
+  // so we ensure that tmp_buf_ is properly aligned to store either.
+  alignas(16) char tmp_buf_[TMP_BUF_SIZE];
+  static_assert(alignof(ElfW(Shdr)) <= 16,
+                "alignment of tmp buf too small for Shdr");
+  static_assert(alignof(ElfW(Sym)) <= 16,
+                "alignment of tmp buf too small for Sym");
+
+  SymbolCacheLine symbol_cache_[SYMBOL_CACHE_LINES];
+};
+
+static std::atomic<Symbolizer *> g_cached_symbolizer;
+
+}  // namespace
+
+static size_t SymbolizerSize() {
+#if defined(__wasm__) || defined(__asmjs__)
+  auto pagesize = static_cast<size_t>(getpagesize());
+#else
+  auto pagesize = static_cast<size_t>(sysconf(_SC_PAGESIZE));
+#endif
+  return ((sizeof(Symbolizer) - 1) / pagesize + 1) * pagesize;
+}
+
+// Return (and set null) g_cached_symbolized_state if it is not null.
+// Otherwise return a new symbolizer.
+static Symbolizer *AllocateSymbolizer() {
+  InitSigSafeArena();
+  Symbolizer *symbolizer =
+      g_cached_symbolizer.exchange(nullptr, std::memory_order_acquire);
+  if (symbolizer != nullptr) {
+    return symbolizer;
+  }
+  return new (base_internal::LowLevelAlloc::AllocWithArena(
+      SymbolizerSize(), SigSafeArena())) Symbolizer();
+}
+
+// Set g_cached_symbolize_state to s if it is null, otherwise
+// delete s.
+static void FreeSymbolizer(Symbolizer *s) {
+  Symbolizer *old_cached_symbolizer = nullptr;
+  if (!g_cached_symbolizer.compare_exchange_strong(old_cached_symbolizer, s,
+                                                   std::memory_order_release,
+                                                   std::memory_order_relaxed)) {
+    s->~Symbolizer();
+    base_internal::LowLevelAlloc::Free(s);
+  }
+}
+
+Symbolizer::Symbolizer() : ok_(true), addr_map_read_(false) {
+  for (SymbolCacheLine &symbol_cache_line : symbol_cache_) {
+    for (size_t j = 0; j < ABSL_ARRAYSIZE(symbol_cache_line.name); ++j) {
+      symbol_cache_line.pc[j] = nullptr;
+      symbol_cache_line.name[j] = nullptr;
+      symbol_cache_line.age[j] = 0;
+    }
+  }
+}
+
+Symbolizer::~Symbolizer() {
+  for (SymbolCacheLine &symbol_cache_line : symbol_cache_) {
+    for (char *s : symbol_cache_line.name) {
+      base_internal::LowLevelAlloc::Free(s);
+    }
+  }
+  ClearAddrMap();
+}
+
+// We don't use assert() since it's not guaranteed to be
+// async-signal-safe.  Instead we define a minimal assertion
+// macro. So far, we don't need pretty printing for __FILE__, etc.
+#define SAFE_ASSERT(expr) ((expr) ? static_cast<void>(0) : abort())
+
+// Read up to "count" bytes from file descriptor "fd" into the buffer
+// starting at "buf" while handling short reads and EINTR.  On
+// success, return the number of bytes read.  Otherwise, return -1.
+static ssize_t ReadPersistent(int fd, void *buf, size_t count) {
+  SAFE_ASSERT(fd >= 0);
+  SAFE_ASSERT(count <= SSIZE_MAX);
+  char *buf0 = reinterpret_cast<char *>(buf);
+  size_t num_bytes = 0;
+  while (num_bytes < count) {
+    ssize_t len;
+    NO_INTR(len = read(fd, buf0 + num_bytes, count - num_bytes));
+    if (len < 0) {  // There was an error other than EINTR.
+      ABSL_RAW_LOG(WARNING, "read failed: errno=%d", errno);
+      return -1;
+    }
+    if (len == 0) {  // Reached EOF.
+      break;
+    }
+    num_bytes += static_cast<size_t>(len);
+  }
+  SAFE_ASSERT(num_bytes <= count);
+  return static_cast<ssize_t>(num_bytes);
+}
+
+// Read up to "count" bytes from "offset" in the file pointed by file
+// descriptor "fd" into the buffer starting at "buf".  On success,
+// return the number of bytes read.  Otherwise, return -1.
+static ssize_t ReadFromOffset(const int fd, void *buf, const size_t count,
+                              const off_t offset) {
+  off_t off = lseek(fd, offset, SEEK_SET);
+  if (off == (off_t)-1) {
+    ABSL_RAW_LOG(WARNING, "lseek(%d, %jd, SEEK_SET) failed: errno=%d", fd,
+                 static_cast<intmax_t>(offset), errno);
+    return -1;
+  }
+  return ReadPersistent(fd, buf, count);
+}
+
+// Try reading exactly "count" bytes from "offset" bytes in a file
+// pointed by "fd" into the buffer starting at "buf" while handling
+// short reads and EINTR.  On success, return true. Otherwise, return
+// false.
+static bool ReadFromOffsetExact(const int fd, void *buf, const size_t count,
+                                const off_t offset) {
+  ssize_t len = ReadFromOffset(fd, buf, count, offset);
+  return len >= 0 && static_cast<size_t>(len) == count;
+}
+
+// Returns elf_header.e_type if the file pointed by fd is an ELF binary.
+static int FileGetElfType(const int fd) {
+  ElfW(Ehdr) elf_header;
+  if (!ReadFromOffsetExact(fd, &elf_header, sizeof(elf_header), 0)) {
+    return -1;
+  }
+  if (memcmp(elf_header.e_ident, ELFMAG, SELFMAG) != 0) {
+    return -1;
+  }
+  return elf_header.e_type;
+}
+
+// Read the section headers in the given ELF binary, and if a section
+// of the specified type is found, set the output to this section header
+// and return true.  Otherwise, return false.
+// To keep stack consumption low, we would like this function to not get
+// inlined.
+static ABSL_ATTRIBUTE_NOINLINE bool GetSectionHeaderByType(
+    const int fd, ElfW(Half) sh_num, const off_t sh_offset, ElfW(Word) type,
+    ElfW(Shdr) * out, char *tmp_buf, size_t tmp_buf_size) {
+  ElfW(Shdr) *buf = reinterpret_cast<ElfW(Shdr) *>(tmp_buf);
+  const size_t buf_entries = tmp_buf_size / sizeof(buf[0]);
+  const size_t buf_bytes = buf_entries * sizeof(buf[0]);
+
+  for (size_t i = 0; static_cast<int>(i) < sh_num;) {
+    const size_t num_bytes_left =
+        (static_cast<size_t>(sh_num) - i) * sizeof(buf[0]);
+    const size_t num_bytes_to_read =
+        (buf_bytes > num_bytes_left) ? num_bytes_left : buf_bytes;
+    const off_t offset = sh_offset + static_cast<off_t>(i * sizeof(buf[0]));
+    const ssize_t len = ReadFromOffset(fd, buf, num_bytes_to_read, offset);
+    if (len < 0) {
+      ABSL_RAW_LOG(
+          WARNING,
+          "Reading %zu bytes from offset %ju returned %zd which is negative.",
+          num_bytes_to_read, static_cast<intmax_t>(offset), len);
+      return false;
+    }
+    if (static_cast<size_t>(len) % sizeof(buf[0]) != 0) {
+      ABSL_RAW_LOG(
+          WARNING,
+          "Reading %zu bytes from offset %jd returned %zd which is not a "
+          "multiple of %zu.",
+          num_bytes_to_read, static_cast<intmax_t>(offset), len,
+          sizeof(buf[0]));
+      return false;
+    }
+    const size_t num_headers_in_buf = static_cast<size_t>(len) / sizeof(buf[0]);
+    SAFE_ASSERT(num_headers_in_buf <= buf_entries);
+    for (size_t j = 0; j < num_headers_in_buf; ++j) {
+      if (buf[j].sh_type == type) {
+        *out = buf[j];
+        return true;
+      }
+    }
+    i += num_headers_in_buf;
+  }
+  return false;
+}
+
+// There is no particular reason to limit section name to 63 characters,
+// but there has (as yet) been no need for anything longer either.
+const int kMaxSectionNameLen = 64;
+
+bool ForEachSection(int fd,
+                    const std::function<bool(absl::string_view name,
+                                             const ElfW(Shdr) &)> &callback) {
+  ElfW(Ehdr) elf_header;
+  if (!ReadFromOffsetExact(fd, &elf_header, sizeof(elf_header), 0)) {
+    return false;
+  }
+
+  ElfW(Shdr) shstrtab;
+  off_t shstrtab_offset = static_cast<off_t>(elf_header.e_shoff) +
+                          elf_header.e_shentsize * elf_header.e_shstrndx;
+  if (!ReadFromOffsetExact(fd, &shstrtab, sizeof(shstrtab), shstrtab_offset)) {
+    return false;
+  }
+
+  for (int i = 0; i < elf_header.e_shnum; ++i) {
+    ElfW(Shdr) out;
+    off_t section_header_offset =
+        static_cast<off_t>(elf_header.e_shoff) + elf_header.e_shentsize * i;
+    if (!ReadFromOffsetExact(fd, &out, sizeof(out), section_header_offset)) {
+      return false;
+    }
+    off_t name_offset = static_cast<off_t>(shstrtab.sh_offset) + out.sh_name;
+    char header_name[kMaxSectionNameLen];
+    ssize_t n_read =
+        ReadFromOffset(fd, &header_name, kMaxSectionNameLen, name_offset);
+    if (n_read < 0) {
+      return false;
+    } else if (n_read > kMaxSectionNameLen) {
+      // Long read?
+      return false;
+    }
+
+    absl::string_view name(header_name,
+                           strnlen(header_name, static_cast<size_t>(n_read)));
+    if (!callback(name, out)) {
+      break;
+    }
+  }
+  return true;
+}
+
+// name_len should include terminating '\0'.
+bool GetSectionHeaderByName(int fd, const char *name, size_t name_len,
+                            ElfW(Shdr) * out) {
+  char header_name[kMaxSectionNameLen];
+  if (sizeof(header_name) < name_len) {
+    ABSL_RAW_LOG(WARNING,
+                 "Section name '%s' is too long (%zu); "
+                 "section will not be found (even if present).",
+                 name, name_len);
+    // No point in even trying.
+    return false;
+  }
+
+  ElfW(Ehdr) elf_header;
+  if (!ReadFromOffsetExact(fd, &elf_header, sizeof(elf_header), 0)) {
+    return false;
+  }
+
+  ElfW(Shdr) shstrtab;
+  off_t shstrtab_offset = static_cast<off_t>(elf_header.e_shoff) +
+                          elf_header.e_shentsize * elf_header.e_shstrndx;
+  if (!ReadFromOffsetExact(fd, &shstrtab, sizeof(shstrtab), shstrtab_offset)) {
+    return false;
+  }
+
+  for (int i = 0; i < elf_header.e_shnum; ++i) {
+    off_t section_header_offset =
+        static_cast<off_t>(elf_header.e_shoff) + elf_header.e_shentsize * i;
+    if (!ReadFromOffsetExact(fd, out, sizeof(*out), section_header_offset)) {
+      return false;
+    }
+    off_t name_offset = static_cast<off_t>(shstrtab.sh_offset) + out->sh_name;
+    ssize_t n_read = ReadFromOffset(fd, &header_name, name_len, name_offset);
+    if (n_read < 0) {
+      return false;
+    } else if (static_cast<size_t>(n_read) != name_len) {
+      // Short read -- name could be at end of file.
+      continue;
+    }
+    if (memcmp(header_name, name, name_len) == 0) {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Compare symbols at in the same address.
+// Return true if we should pick symbol1.
+static bool ShouldPickFirstSymbol(const ElfW(Sym) & symbol1,
+                                  const ElfW(Sym) & symbol2) {
+  // If one of the symbols is weak and the other is not, pick the one
+  // this is not a weak symbol.
+  char bind1 = ELF_ST_BIND(symbol1.st_info);
+  char bind2 = ELF_ST_BIND(symbol1.st_info);
+  if (bind1 == STB_WEAK && bind2 != STB_WEAK) return false;
+  if (bind2 == STB_WEAK && bind1 != STB_WEAK) return true;
+
+  // If one of the symbols has zero size and the other is not, pick the
+  // one that has non-zero size.
+  if (symbol1.st_size != 0 && symbol2.st_size == 0) {
+    return true;
+  }
+  if (symbol1.st_size == 0 && symbol2.st_size != 0) {
+    return false;
+  }
+
+  // If one of the symbols has no type and the other is not, pick the
+  // one that has a type.
+  char type1 = ELF_ST_TYPE(symbol1.st_info);
+  char type2 = ELF_ST_TYPE(symbol1.st_info);
+  if (type1 != STT_NOTYPE && type2 == STT_NOTYPE) {
+    return true;
+  }
+  if (type1 == STT_NOTYPE && type2 != STT_NOTYPE) {
+    return false;
+  }
+
+  // Pick the first one, if we still cannot decide.
+  return true;
+}
+
+// Return true if an address is inside a section.
+static bool InSection(const void *address, const ElfW(Shdr) * section) {
+  const char *start = reinterpret_cast<const char *>(section->sh_addr);
+  size_t size = static_cast<size_t>(section->sh_size);
+  return start <= address && address < (start + size);
+}
+
+static const char *ComputeOffset(const char *base, ptrdiff_t offset) {
+  // Note: cast to intptr_t to avoid undefined behavior when base evaluates to
+  // zero and offset is non-zero.
+  return reinterpret_cast<const char *>(reinterpret_cast<intptr_t>(base) +
+                                        offset);
+}
+
+// Read a symbol table and look for the symbol containing the
+// pc. Iterate over symbols in a symbol table and look for the symbol
+// containing "pc".  If the symbol is found, and its name fits in
+// out_size, the name is written into out and SYMBOL_FOUND is returned.
+// If the name does not fit, truncated name is written into out,
+// and SYMBOL_TRUNCATED is returned. Out is NUL-terminated.
+// If the symbol is not found, SYMBOL_NOT_FOUND is returned;
+// To keep stack consumption low, we would like this function to not get
+// inlined.
+static ABSL_ATTRIBUTE_NOINLINE FindSymbolResult FindSymbol(
+    const void *const pc, const int fd, char *out, size_t out_size,
+    ptrdiff_t relocation, const ElfW(Shdr) * strtab, const ElfW(Shdr) * symtab,
+    const ElfW(Shdr) * opd, char *tmp_buf, size_t tmp_buf_size) {
+  if (symtab == nullptr) {
+    return SYMBOL_NOT_FOUND;
+  }
+
+  // Read multiple symbols at once to save read() calls.
+  ElfW(Sym) *buf = reinterpret_cast<ElfW(Sym) *>(tmp_buf);
+  const size_t buf_entries = tmp_buf_size / sizeof(buf[0]);
+
+  const size_t num_symbols = symtab->sh_size / symtab->sh_entsize;
+
+  // On platforms using an .opd section (PowerPC & IA64), a function symbol
+  // has the address of a function descriptor, which contains the real
+  // starting address.  However, we do not always want to use the real
+  // starting address because we sometimes want to symbolize a function
+  // pointer into the .opd section, e.g. FindSymbol(&foo,...).
+  const bool pc_in_opd =
+      kPlatformUsesOPDSections && opd != nullptr && InSection(pc, opd);
+  const bool deref_function_descriptor_pointer =
+      kPlatformUsesOPDSections && opd != nullptr && !pc_in_opd;
+
+  ElfW(Sym) best_match;
+  SafeMemZero(&best_match, sizeof(best_match));
+  bool found_match = false;
+  for (size_t i = 0; i < num_symbols;) {
+    off_t offset =
+        static_cast<off_t>(symtab->sh_offset + i * symtab->sh_entsize);
+    const size_t num_remaining_symbols = num_symbols - i;
+    const size_t entries_in_chunk =
+        std::min(num_remaining_symbols, buf_entries);
+    const size_t bytes_in_chunk = entries_in_chunk * sizeof(buf[0]);
+    const ssize_t len = ReadFromOffset(fd, buf, bytes_in_chunk, offset);
+    SAFE_ASSERT(len >= 0);
+    SAFE_ASSERT(static_cast<size_t>(len) % sizeof(buf[0]) == 0);
+    const size_t num_symbols_in_buf = static_cast<size_t>(len) / sizeof(buf[0]);
+    SAFE_ASSERT(num_symbols_in_buf <= entries_in_chunk);
+    for (size_t j = 0; j < num_symbols_in_buf; ++j) {
+      const ElfW(Sym) &symbol = buf[j];
+
+      // For a DSO, a symbol address is relocated by the loading address.
+      // We keep the original address for opd redirection below.
+      const char *const original_start_address =
+          reinterpret_cast<const char *>(symbol.st_value);
+      const char *start_address =
+          ComputeOffset(original_start_address, relocation);
+
+#ifdef __arm__
+      // ARM functions are always aligned to multiples of two bytes; the
+      // lowest-order bit in start_address is ignored by the CPU and indicates
+      // whether the function contains ARM (0) or Thumb (1) code. We don't care
+      // about what encoding is being used; we just want the real start address
+      // of the function.
+      start_address = reinterpret_cast<const char *>(
+          reinterpret_cast<uintptr_t>(start_address) & ~1u);
+#endif
+
+      if (deref_function_descriptor_pointer &&
+          InSection(original_start_address, opd)) {
+        // The opd section is mapped into memory.  Just dereference
+        // start_address to get the first double word, which points to the
+        // function entry.
+        start_address = *reinterpret_cast<const char *const *>(start_address);
+      }
+
+      // If pc is inside the .opd section, it points to a function descriptor.
+      const size_t size = pc_in_opd ? kFunctionDescriptorSize : symbol.st_size;
+      const void *const end_address =
+          ComputeOffset(start_address, static_cast<ptrdiff_t>(size));
+      if (symbol.st_value != 0 &&  // Skip null value symbols.
+          symbol.st_shndx != 0 &&  // Skip undefined symbols.
+#ifdef STT_TLS
+          ELF_ST_TYPE(symbol.st_info) != STT_TLS &&  // Skip thread-local data.
+#endif                                               // STT_TLS
+          ((start_address <= pc && pc < end_address) ||
+           (start_address == pc && pc == end_address))) {
+        if (!found_match || ShouldPickFirstSymbol(symbol, best_match)) {
+          found_match = true;
+          best_match = symbol;
+        }
+      }
+    }
+    i += num_symbols_in_buf;
+  }
+
+  if (found_match) {
+    const off_t off =
+        static_cast<off_t>(strtab->sh_offset) + best_match.st_name;
+    const ssize_t n_read = ReadFromOffset(fd, out, out_size, off);
+    if (n_read <= 0) {
+      // This should never happen.
+      ABSL_RAW_LOG(WARNING,
+                   "Unable to read from fd %d at offset %lld: n_read = %zd", fd,
+                   static_cast<long long>(off), n_read);
+      return SYMBOL_NOT_FOUND;
+    }
+    ABSL_RAW_CHECK(static_cast<size_t>(n_read) <= out_size,
+                   "ReadFromOffset read too much data.");
+
+    // strtab->sh_offset points into .strtab-like section that contains
+    // NUL-terminated strings: '\0foo\0barbaz\0...".
+    //
+    // sh_offset+st_name points to the start of symbol name, but we don't know
+    // how long the symbol is, so we try to read as much as we have space for,
+    // and usually over-read (i.e. there is a NUL somewhere before n_read).
+    if (memchr(out, '\0', static_cast<size_t>(n_read)) == nullptr) {
+      // Either out_size was too small (n_read == out_size and no NUL), or
+      // we tried to read past the EOF (n_read < out_size) and .strtab is
+      // corrupt (missing terminating NUL; should never happen for valid ELF).
+      out[n_read - 1] = '\0';
+      return SYMBOL_TRUNCATED;
+    }
+    return SYMBOL_FOUND;
+  }
+
+  return SYMBOL_NOT_FOUND;
+}
+
+// Get the symbol name of "pc" from the file pointed by "fd".  Process
+// both regular and dynamic symbol tables if necessary.
+// See FindSymbol() comment for description of return value.
+FindSymbolResult Symbolizer::GetSymbolFromObjectFile(
+    const ObjFile &obj, const void *const pc, const ptrdiff_t relocation,
+    char *out, size_t out_size, char *tmp_buf, size_t tmp_buf_size) {
+  ElfW(Shdr) symtab;
+  ElfW(Shdr) strtab;
+  ElfW(Shdr) opd;
+  ElfW(Shdr) *opd_ptr = nullptr;
+
+  // On platforms using an .opd sections for function descriptor, read
+  // the section header.  The .opd section is in data segment and should be
+  // loaded but we check that it is mapped just to be extra careful.
+  if (kPlatformUsesOPDSections) {
+    if (GetSectionHeaderByName(obj.fd, kOpdSectionName,
+                               sizeof(kOpdSectionName) - 1, &opd) &&
+        FindObjFile(reinterpret_cast<const char *>(opd.sh_addr) + relocation,
+                    opd.sh_size) != nullptr) {
+      opd_ptr = &opd;
+    } else {
+      return SYMBOL_NOT_FOUND;
+    }
+  }
+
+  // Consult a regular symbol table, then fall back to the dynamic symbol table.
+  for (const auto symbol_table_type : {SHT_SYMTAB, SHT_DYNSYM}) {
+    if (!GetSectionHeaderByType(obj.fd, obj.elf_header.e_shnum,
+                                static_cast<off_t>(obj.elf_header.e_shoff),
+                                static_cast<ElfW(Word)>(symbol_table_type),
+                                &symtab, tmp_buf, tmp_buf_size)) {
+      continue;
+    }
+    if (!ReadFromOffsetExact(
+            obj.fd, &strtab, sizeof(strtab),
+            static_cast<off_t>(obj.elf_header.e_shoff +
+                               symtab.sh_link * sizeof(symtab)))) {
+      continue;
+    }
+    const FindSymbolResult rc =
+        FindSymbol(pc, obj.fd, out, out_size, relocation, &strtab, &symtab,
+                   opd_ptr, tmp_buf, tmp_buf_size);
+    if (rc != SYMBOL_NOT_FOUND) {
+      return rc;
+    }
+  }
+
+  return SYMBOL_NOT_FOUND;
+}
+
+namespace {
+// Thin wrapper around a file descriptor so that the file descriptor
+// gets closed for sure.
+class FileDescriptor {
+ public:
+  explicit FileDescriptor(int fd) : fd_(fd) {}
+  FileDescriptor(const FileDescriptor &) = delete;
+  FileDescriptor &operator=(const FileDescriptor &) = delete;
+
+  ~FileDescriptor() {
+    if (fd_ >= 0) {
+      close(fd_);
+    }
+  }
+
+  int get() const { return fd_; }
+
+ private:
+  const int fd_;
+};
+
+// Helper class for reading lines from file.
+//
+// Note: we don't use ProcMapsIterator since the object is big (it has
+// a 5k array member) and uses async-unsafe functions such as sscanf()
+// and snprintf().
+class LineReader {
+ public:
+  explicit LineReader(int fd, char *buf, size_t buf_len)
+      : fd_(fd),
+        buf_len_(buf_len),
+        buf_(buf),
+        bol_(buf),
+        eol_(buf),
+        eod_(buf) {}
+
+  LineReader(const LineReader &) = delete;
+  LineReader &operator=(const LineReader &) = delete;
+
+  // Read '\n'-terminated line from file.  On success, modify "bol"
+  // and "eol", then return true.  Otherwise, return false.
+  //
+  // Note: if the last line doesn't end with '\n', the line will be
+  // dropped.  It's an intentional behavior to make the code simple.
+  bool ReadLine(const char **bol, const char **eol) {
+    if (BufferIsEmpty()) {  // First time.
+      const ssize_t num_bytes = ReadPersistent(fd_, buf_, buf_len_);
+      if (num_bytes <= 0) {  // EOF or error.
+        return false;
+      }
+      eod_ = buf_ + num_bytes;
+      bol_ = buf_;
+    } else {
+      bol_ = eol_ + 1;            // Advance to the next line in the buffer.
+      SAFE_ASSERT(bol_ <= eod_);  // "bol_" can point to "eod_".
+      if (!HasCompleteLine()) {
+        const auto incomplete_line_length = static_cast<size_t>(eod_ - bol_);
+        // Move the trailing incomplete line to the beginning.
+        memmove(buf_, bol_, incomplete_line_length);
+        // Read text from file and append it.
+        char *const append_pos = buf_ + incomplete_line_length;
+        const size_t capacity_left = buf_len_ - incomplete_line_length;
+        const ssize_t num_bytes =
+            ReadPersistent(fd_, append_pos, capacity_left);
+        if (num_bytes <= 0) {  // EOF or error.
+          return false;
+        }
+        eod_ = append_pos + num_bytes;
+        bol_ = buf_;
+      }
+    }
+    eol_ = FindLineFeed();
+    if (eol_ == nullptr) {  // '\n' not found.  Malformed line.
+      return false;
+    }
+    *eol_ = '\0';  // Replace '\n' with '\0'.
+
+    *bol = bol_;
+    *eol = eol_;
+    return true;
+  }
+
+ private:
+  char *FindLineFeed() const {
+    return reinterpret_cast<char *>(
+        memchr(bol_, '\n', static_cast<size_t>(eod_ - bol_)));
+  }
+
+  bool BufferIsEmpty() const { return buf_ == eod_; }
+
+  bool HasCompleteLine() const {
+    return !BufferIsEmpty() && FindLineFeed() != nullptr;
+  }
+
+  const int fd_;
+  const size_t buf_len_;
+  char *const buf_;
+  char *bol_;
+  char *eol_;
+  const char *eod_;  // End of data in "buf_".
+};
+}  // namespace
+
+// Place the hex number read from "start" into "*hex".  The pointer to
+// the first non-hex character or "end" is returned.
+static const char *GetHex(const char *start, const char *end,
+                          uint64_t *const value) {
+  uint64_t hex = 0;
+  const char *p;
+  for (p = start; p < end; ++p) {
+    int ch = *p;
+    if ((ch >= '0' && ch <= '9') || (ch >= 'A' && ch <= 'F') ||
+        (ch >= 'a' && ch <= 'f')) {
+      hex = (hex << 4) |
+            static_cast<uint64_t>(ch < 'A' ? ch - '0' : (ch & 0xF) + 9);
+    } else {  // Encountered the first non-hex character.
+      break;
+    }
+  }
+  SAFE_ASSERT(p <= end);
+  *value = hex;
+  return p;
+}
+
+static const char *GetHex(const char *start, const char *end,
+                          const void **const addr) {
+  uint64_t hex = 0;
+  const char *p = GetHex(start, end, &hex);
+  *addr = reinterpret_cast<void *>(hex);
+  return p;
+}
+
+// Normally we are only interested in "r?x" maps.
+// On the PowerPC, function pointers point to descriptors in the .opd
+// section.  The descriptors themselves are not executable code, so
+// we need to relax the check below to "r??".
+static bool ShouldUseMapping(const char *const flags) {
+  return flags[0] == 'r' && (kPlatformUsesOPDSections || flags[2] == 'x');
+}
+
+// Read /proc/self/maps and run "callback" for each mmapped file found.  If
+// "callback" returns false, stop scanning and return true. Else continue
+// scanning /proc/self/maps. Return true if no parse error is found.
+static ABSL_ATTRIBUTE_NOINLINE bool ReadAddrMap(
+    bool (*callback)(const char *filename, const void *const start_addr,
+                     const void *const end_addr, uint64_t offset, void *arg),
+    void *arg, void *tmp_buf, size_t tmp_buf_size) {
+  // Use /proc/self/task/<pid>/maps instead of /proc/self/maps. The latter
+  // requires kernel to stop all threads, and is significantly slower when there
+  // are 1000s of threads.
+  char maps_path[80];
+  snprintf(maps_path, sizeof(maps_path), "/proc/self/task/%d/maps", getpid());
+
+  int maps_fd;
+  NO_INTR(maps_fd = open(maps_path, O_RDONLY));
+  FileDescriptor wrapped_maps_fd(maps_fd);
+  if (wrapped_maps_fd.get() < 0) {
+    ABSL_RAW_LOG(WARNING, "%s: errno=%d", maps_path, errno);
+    return false;
+  }
+
+  // Iterate over maps and look for the map containing the pc.  Then
+  // look into the symbol tables inside.
+  LineReader reader(wrapped_maps_fd.get(), static_cast<char *>(tmp_buf),
+                    tmp_buf_size);
+  while (true) {
+    const char *cursor;
+    const char *eol;
+    if (!reader.ReadLine(&cursor, &eol)) {  // EOF or malformed line.
+      break;
+    }
+
+    const char *line = cursor;
+    const void *start_address;
+    // Start parsing line in /proc/self/maps.  Here is an example:
+    //
+    // 08048000-0804c000 r-xp 00000000 08:01 2142121    /bin/cat
+    //
+    // We want start address (08048000), end address (0804c000), flags
+    // (r-xp) and file name (/bin/cat).
+
+    // Read start address.
+    cursor = GetHex(cursor, eol, &start_address);
+    if (cursor == eol || *cursor != '-') {
+      ABSL_RAW_LOG(WARNING, "Corrupt /proc/self/maps line: %s", line);
+      return false;
+    }
+    ++cursor;  // Skip '-'.
+
+    // Read end address.
+    const void *end_address;
+    cursor = GetHex(cursor, eol, &end_address);
+    if (cursor == eol || *cursor != ' ') {
+      ABSL_RAW_LOG(WARNING, "Corrupt /proc/self/maps line: %s", line);
+      return false;
+    }
+    ++cursor;  // Skip ' '.
+
+    // Read flags.  Skip flags until we encounter a space or eol.
+    const char *const flags_start = cursor;
+    while (cursor < eol && *cursor != ' ') {
+      ++cursor;
+    }
+    // We expect at least four letters for flags (ex. "r-xp").
+    if (cursor == eol || cursor < flags_start + 4) {
+      ABSL_RAW_LOG(WARNING, "Corrupt /proc/self/maps: %s", line);
+      return false;
+    }
+
+    // Check flags.
+    if (!ShouldUseMapping(flags_start)) {
+      continue;  // We skip this map.
+    }
+    ++cursor;  // Skip ' '.
+
+    // Read file offset.
+    uint64_t offset;
+    cursor = GetHex(cursor, eol, &offset);
+    ++cursor;  // Skip ' '.
+
+    // Skip to file name.  "cursor" now points to dev.  We need to skip at least
+    // two spaces for dev and inode.
+    int num_spaces = 0;
+    while (cursor < eol) {
+      if (*cursor == ' ') {
+        ++num_spaces;
+      } else if (num_spaces >= 2) {
+        // The first non-space character after  skipping two spaces
+        // is the beginning of the file name.
+        break;
+      }
+      ++cursor;
+    }
+
+    // Check whether this entry corresponds to our hint table for the true
+    // filename.
+    bool hinted =
+        GetFileMappingHint(&start_address, &end_address, &offset, &cursor);
+    if (!hinted && (cursor == eol || cursor[0] == '[')) {
+      // not an object file, typically [vdso] or [vsyscall]
+      continue;
+    }
+    if (!callback(cursor, start_address, end_address, offset, arg)) break;
+  }
+  return true;
+}
+
+// Find the objfile mapped in address region containing [addr, addr + len).
+ObjFile *Symbolizer::FindObjFile(const void *const addr, size_t len) {
+  for (int i = 0; i < 2; ++i) {
+    if (!ok_) return nullptr;
+
+    // Read /proc/self/maps if necessary
+    if (!addr_map_read_) {
+      addr_map_read_ = true;
+      if (!ReadAddrMap(RegisterObjFile, this, tmp_buf_, TMP_BUF_SIZE)) {
+        ok_ = false;
+        return nullptr;
+      }
+    }
+
+    size_t lo = 0;
+    size_t hi = addr_map_.Size();
+    while (lo < hi) {
+      size_t mid = (lo + hi) / 2;
+      if (addr < addr_map_.At(mid)->end_addr) {
+        hi = mid;
+      } else {
+        lo = mid + 1;
+      }
+    }
+    if (lo != addr_map_.Size()) {
+      ObjFile *obj = addr_map_.At(lo);
+      SAFE_ASSERT(obj->end_addr > addr);
+      if (addr >= obj->start_addr &&
+          reinterpret_cast<const char *>(addr) + len <= obj->end_addr)
+        return obj;
+    }
+
+    // The address mapping may have changed since it was last read.  Retry.
+    ClearAddrMap();
+  }
+  return nullptr;
+}
+
+void Symbolizer::ClearAddrMap() {
+  for (size_t i = 0; i != addr_map_.Size(); i++) {
+    ObjFile *o = addr_map_.At(i);
+    base_internal::LowLevelAlloc::Free(o->filename);
+    if (o->fd >= 0) {
+      close(o->fd);
+    }
+  }
+  addr_map_.Clear();
+  addr_map_read_ = false;
+}
+
+// Callback for ReadAddrMap to register objfiles in an in-memory table.
+bool Symbolizer::RegisterObjFile(const char *filename,
+                                 const void *const start_addr,
+                                 const void *const end_addr, uint64_t offset,
+                                 void *arg) {
+  Symbolizer *impl = static_cast<Symbolizer *>(arg);
+
+  // Files are supposed to be added in the increasing address order.  Make
+  // sure that's the case.
+  size_t addr_map_size = impl->addr_map_.Size();
+  if (addr_map_size != 0) {
+    ObjFile *old = impl->addr_map_.At(addr_map_size - 1);
+    if (old->end_addr > end_addr) {
+      ABSL_RAW_LOG(ERROR,
+                   "Unsorted addr map entry: 0x%" PRIxPTR ": %s <-> 0x%" PRIxPTR
+                   ": %s",
+                   reinterpret_cast<uintptr_t>(end_addr), filename,
+                   reinterpret_cast<uintptr_t>(old->end_addr), old->filename);
+      return true;
+    } else if (old->end_addr == end_addr) {
+      // The same entry appears twice. This sometimes happens for [vdso].
+      if (old->start_addr != start_addr ||
+          strcmp(old->filename, filename) != 0) {
+        ABSL_RAW_LOG(ERROR,
+                     "Duplicate addr 0x%" PRIxPTR ": %s <-> 0x%" PRIxPTR ": %s",
+                     reinterpret_cast<uintptr_t>(end_addr), filename,
+                     reinterpret_cast<uintptr_t>(old->end_addr), old->filename);
+      }
+      return true;
+    } else if (old->end_addr == start_addr &&
+               reinterpret_cast<uintptr_t>(old->start_addr) - old->offset ==
+                   reinterpret_cast<uintptr_t>(start_addr) - offset &&
+               strcmp(old->filename, filename) == 0) {
+      // Two contiguous map entries that span a contiguous region of the file,
+      // perhaps because some part of the file was mlock()ed. Combine them.
+      old->end_addr = end_addr;
+      return true;
+    }
+  }
+  ObjFile *obj = impl->addr_map_.Add();
+  obj->filename = impl->CopyString(filename);
+  obj->start_addr = start_addr;
+  obj->end_addr = end_addr;
+  obj->offset = offset;
+  obj->elf_type = -1;  // filled on demand
+  obj->fd = -1;        // opened on demand
+  return true;
+}
+
+// This function wraps the Demangle function to provide an interface
+// where the input symbol is demangled in-place.
+// To keep stack consumption low, we would like this function to not
+// get inlined.
+static ABSL_ATTRIBUTE_NOINLINE void DemangleInplace(char *out, size_t out_size,
+                                                    char *tmp_buf,
+                                                    size_t tmp_buf_size) {
+  if (Demangle(out, tmp_buf, tmp_buf_size)) {
+    // Demangling succeeded. Copy to out if the space allows.
+    size_t len = strlen(tmp_buf);
+    if (len + 1 <= out_size) {  // +1 for '\0'.
+      SAFE_ASSERT(len < tmp_buf_size);
+      memmove(out, tmp_buf, len + 1);
+    }
+  }
+}
+
+SymbolCacheLine *Symbolizer::GetCacheLine(const void *const pc) {
+  uintptr_t pc0 = reinterpret_cast<uintptr_t>(pc);
+  pc0 >>= 3;  // drop the low 3 bits
+
+  // Shuffle bits.
+  pc0 ^= (pc0 >> 6) ^ (pc0 >> 12) ^ (pc0 >> 18);
+  return &symbol_cache_[pc0 % SYMBOL_CACHE_LINES];
+}
+
+void Symbolizer::AgeSymbols(SymbolCacheLine *line) {
+  for (uint32_t &age : line->age) {
+    ++age;
+  }
+}
+
+const char *Symbolizer::FindSymbolInCache(const void *const pc) {
+  if (pc == nullptr) return nullptr;
+
+  SymbolCacheLine *line = GetCacheLine(pc);
+  for (size_t i = 0; i < ABSL_ARRAYSIZE(line->pc); ++i) {
+    if (line->pc[i] == pc) {
+      AgeSymbols(line);
+      line->age[i] = 0;
+      return line->name[i];
+    }
+  }
+  return nullptr;
+}
+
+const char *Symbolizer::InsertSymbolInCache(const void *const pc,
+                                            const char *name) {
+  SAFE_ASSERT(pc != nullptr);
+
+  SymbolCacheLine *line = GetCacheLine(pc);
+  uint32_t max_age = 0;
+  size_t oldest_index = 0;
+  bool found_oldest_index = false;
+  for (size_t i = 0; i < ABSL_ARRAYSIZE(line->pc); ++i) {
+    if (line->pc[i] == nullptr) {
+      AgeSymbols(line);
+      line->pc[i] = pc;
+      line->name[i] = CopyString(name);
+      line->age[i] = 0;
+      return line->name[i];
+    }
+    if (line->age[i] >= max_age) {
+      max_age = line->age[i];
+      oldest_index = i;
+      found_oldest_index = true;
+    }
+  }
+
+  AgeSymbols(line);
+  ABSL_RAW_CHECK(found_oldest_index, "Corrupt cache");
+  base_internal::LowLevelAlloc::Free(line->name[oldest_index]);
+  line->pc[oldest_index] = pc;
+  line->name[oldest_index] = CopyString(name);
+  line->age[oldest_index] = 0;
+  return line->name[oldest_index];
+}
+
+static void MaybeOpenFdFromSelfExe(ObjFile *obj) {
+  if (memcmp(obj->start_addr, ELFMAG, SELFMAG) != 0) {
+    return;
+  }
+  int fd = open("/proc/self/exe", O_RDONLY);
+  if (fd == -1) {
+    return;
+  }
+  // Verify that contents of /proc/self/exe matches in-memory image of
+  // the binary. This can fail if the "deleted" binary is in fact not
+  // the main executable, or for binaries that have the first PT_LOAD
+  // segment smaller than 4K. We do it in four steps so that the
+  // buffer is smaller and we don't consume too much stack space.
+  const char *mem = reinterpret_cast<const char *>(obj->start_addr);
+  for (int i = 0; i < 4; ++i) {
+    char buf[1024];
+    ssize_t n = read(fd, buf, sizeof(buf));
+    if (n != sizeof(buf) || memcmp(buf, mem, sizeof(buf)) != 0) {
+      close(fd);
+      return;
+    }
+    mem += sizeof(buf);
+  }
+  obj->fd = fd;
+}
+
+static bool MaybeInitializeObjFile(ObjFile *obj) {
+  if (obj->fd < 0) {
+    obj->fd = open(obj->filename, O_RDONLY);
+
+    if (obj->fd < 0) {
+      // Getting /proc/self/exe here means that we were hinted.
+      if (strcmp(obj->filename, "/proc/self/exe") == 0) {
+        // /proc/self/exe may be inaccessible (due to setuid, etc.), so try
+        // accessing the binary via argv0.
+        if (argv0_value != nullptr) {
+          obj->fd = open(argv0_value, O_RDONLY);
+        }
+      } else {
+        MaybeOpenFdFromSelfExe(obj);
+      }
+    }
+
+    if (obj->fd < 0) {
+      ABSL_RAW_LOG(WARNING, "%s: open failed: errno=%d", obj->filename, errno);
+      return false;
+    }
+    obj->elf_type = FileGetElfType(obj->fd);
+    if (obj->elf_type < 0) {
+      ABSL_RAW_LOG(WARNING, "%s: wrong elf type: %d", obj->filename,
+                   obj->elf_type);
+      return false;
+    }
+
+    if (!ReadFromOffsetExact(obj->fd, &obj->elf_header, sizeof(obj->elf_header),
+                             0)) {
+      ABSL_RAW_LOG(WARNING, "%s: failed to read elf header", obj->filename);
+      return false;
+    }
+    const int phnum = obj->elf_header.e_phnum;
+    const int phentsize = obj->elf_header.e_phentsize;
+    auto phoff = static_cast<off_t>(obj->elf_header.e_phoff);
+    size_t num_executable_load_segments = 0;
+    for (int j = 0; j < phnum; j++) {
+      ElfW(Phdr) phdr;
+      if (!ReadFromOffsetExact(obj->fd, &phdr, sizeof(phdr), phoff)) {
+        ABSL_RAW_LOG(WARNING, "%s: failed to read program header %d",
+                     obj->filename, j);
+        return false;
+      }
+      phoff += phentsize;
+      constexpr int rx = PF_X | PF_R;
+      if (phdr.p_type != PT_LOAD || (phdr.p_flags & rx) != rx) {
+        // Not a LOAD segment, or not executable code.
+        continue;
+      }
+      if (num_executable_load_segments < obj->phdr.size()) {
+        memcpy(&obj->phdr[num_executable_load_segments++], &phdr, sizeof(phdr));
+      } else {
+        ABSL_RAW_LOG(
+            WARNING, "%s: too many executable LOAD segments: %zu >= %zu",
+            obj->filename, num_executable_load_segments, obj->phdr.size());
+        break;
+      }
+    }
+    if (num_executable_load_segments == 0) {
+      // This object has no "r-x" LOAD segments. That's unexpected.
+      ABSL_RAW_LOG(WARNING, "%s: no executable LOAD segments", obj->filename);
+      return false;
+    }
+  }
+  return true;
+}
+
+// The implementation of our symbolization routine.  If it
+// successfully finds the symbol containing "pc" and obtains the
+// symbol name, returns pointer to that symbol. Otherwise, returns nullptr.
+// If any symbol decorators have been installed via InstallSymbolDecorator(),
+// they are called here as well.
+// To keep stack consumption low, we would like this function to not
+// get inlined.
+const char *Symbolizer::GetUncachedSymbol(const void *pc) {
+  ObjFile *const obj = FindObjFile(pc, 1);
+  ptrdiff_t relocation = 0;
+  int fd = -1;
+  if (obj != nullptr) {
+    if (MaybeInitializeObjFile(obj)) {
+      const size_t start_addr = reinterpret_cast<size_t>(obj->start_addr);
+      if (obj->elf_type == ET_DYN && start_addr >= obj->offset) {
+        // This object was relocated.
+        //
+        // For obj->offset > 0, adjust the relocation since a mapping at offset
+        // X in the file will have a start address of [true relocation]+X.
+        relocation = static_cast<ptrdiff_t>(start_addr - obj->offset);
+
+        // Note: some binaries have multiple "rx" LOAD segments. We must
+        // find the right one.
+        ElfW(Phdr) *phdr = nullptr;
+        for (size_t j = 0; j < obj->phdr.size(); j++) {
+          ElfW(Phdr) &p = obj->phdr[j];
+          if (p.p_type != PT_LOAD) {
+            // We only expect PT_LOADs. This must be PT_NULL that we didn't
+            // write over (i.e. we exhausted all interesting PT_LOADs).
+            ABSL_RAW_CHECK(p.p_type == PT_NULL, "unexpected p_type");
+            break;
+          }
+          if (pc < reinterpret_cast<void *>(start_addr + p.p_memsz)) {
+            phdr = &p;
+            break;
+          }
+        }
+        if (phdr == nullptr) {
+          // That's unexpected. Hope for the best.
+          ABSL_RAW_LOG(
+              WARNING,
+              "%s: unable to find LOAD segment for pc: %p, start_addr: %zx",
+              obj->filename, pc, start_addr);
+        } else {
+          // Adjust relocation in case phdr.p_vaddr != 0.
+          // This happens for binaries linked with `lld --rosegment`, and for
+          // binaries linked with BFD `ld -z separate-code`.
+          relocation -= phdr->p_vaddr - phdr->p_offset;
+        }
+      }
+
+      fd = obj->fd;
+      if (GetSymbolFromObjectFile(*obj, pc, relocation, symbol_buf_,
+                                  sizeof(symbol_buf_), tmp_buf_,
+                                  sizeof(tmp_buf_)) == SYMBOL_FOUND) {
+        // Only try to demangle the symbol name if it fit into symbol_buf_.
+        DemangleInplace(symbol_buf_, sizeof(symbol_buf_), tmp_buf_,
+                        sizeof(tmp_buf_));
+      }
+    }
+  } else {
+#if ABSL_HAVE_VDSO_SUPPORT
+    VDSOSupport vdso;
+    if (vdso.IsPresent()) {
+      VDSOSupport::SymbolInfo symbol_info;
+      if (vdso.LookupSymbolByAddress(pc, &symbol_info)) {
+        // All VDSO symbols are known to be short.
+        size_t len = strlen(symbol_info.name);
+        ABSL_RAW_CHECK(len + 1 < sizeof(symbol_buf_),
+                       "VDSO symbol unexpectedly long");
+        memcpy(symbol_buf_, symbol_info.name, len + 1);
+      }
+    }
+#endif
+  }
+
+  if (g_decorators_mu.TryLock()) {
+    if (g_num_decorators > 0) {
+      SymbolDecoratorArgs decorator_args = {
+          pc,       relocation,       fd,     symbol_buf_, sizeof(symbol_buf_),
+          tmp_buf_, sizeof(tmp_buf_), nullptr};
+      for (int i = 0; i < g_num_decorators; ++i) {
+        decorator_args.arg = g_decorators[i].arg;
+        g_decorators[i].fn(&decorator_args);
+      }
+    }
+    g_decorators_mu.Unlock();
+  }
+  if (symbol_buf_[0] == '\0') {
+    return nullptr;
+  }
+  symbol_buf_[sizeof(symbol_buf_) - 1] = '\0';  // Paranoia.
+  return InsertSymbolInCache(pc, symbol_buf_);
+}
+
+const char *Symbolizer::GetSymbol(const void *pc) {
+  const char *entry = FindSymbolInCache(pc);
+  if (entry != nullptr) {
+    return entry;
+  }
+  symbol_buf_[0] = '\0';
+
+#ifdef __hppa__
+  {
+    // In some contexts (e.g., return addresses), PA-RISC uses the lowest two
+    // bits of the address to indicate the privilege level. Clear those bits
+    // before trying to symbolize.
+    const auto pc_bits = reinterpret_cast<uintptr_t>(pc);
+    const auto address = pc_bits & ~0x3;
+    entry = GetUncachedSymbol(reinterpret_cast<const void *>(address));
+    if (entry != nullptr) {
+      return entry;
+    }
+
+    // In some contexts, PA-RISC also uses bit 1 of the address to indicate that
+    // this is a cross-DSO function pointer. Such function pointers actually
+    // point to a procedure label, a struct whose first 32-bit (pointer) element
+    // actually points to the function text. With no symbol found for this
+    // address so far, try interpreting it as a cross-DSO function pointer and
+    // see how that goes.
+    if (pc_bits & 0x2) {
+      return GetUncachedSymbol(*reinterpret_cast<const void *const *>(address));
+    }
+
+    return nullptr;
+  }
+#else
+  return GetUncachedSymbol(pc);
+#endif
+}
+
+bool RemoveAllSymbolDecorators(void) {
+  if (!g_decorators_mu.TryLock()) {
+    // Someone else is using decorators. Get out.
+    return false;
+  }
+  g_num_decorators = 0;
+  g_decorators_mu.Unlock();
+  return true;
+}
+
+bool RemoveSymbolDecorator(int ticket) {
+  if (!g_decorators_mu.TryLock()) {
+    // Someone else is using decorators. Get out.
+    return false;
+  }
+  for (int i = 0; i < g_num_decorators; ++i) {
+    if (g_decorators[i].ticket == ticket) {
+      while (i < g_num_decorators - 1) {
+        g_decorators[i] = g_decorators[i + 1];
+        ++i;
+      }
+      g_num_decorators = i;
+      break;
+    }
+  }
+  g_decorators_mu.Unlock();
+  return true;  // Decorator is known to be removed.
+}
+
+int InstallSymbolDecorator(SymbolDecorator decorator, void *arg) {
+  static int ticket = 0;
+
+  if (!g_decorators_mu.TryLock()) {
+    // Someone else is using decorators. Get out.
+    return -2;
+  }
+  int ret = ticket;
+  if (g_num_decorators >= kMaxDecorators) {
+    ret = -1;
+  } else {
+    g_decorators[g_num_decorators] = {decorator, arg, ticket++};
+    ++g_num_decorators;
+  }
+  g_decorators_mu.Unlock();
+  return ret;
+}
+
+bool RegisterFileMappingHint(const void *start, const void *end, uint64_t offset,
+                             const char *filename) {
+  SAFE_ASSERT(start <= end);
+  SAFE_ASSERT(filename != nullptr);
+
+  InitSigSafeArena();
+
+  if (!g_file_mapping_mu.TryLock()) {
+    return false;
+  }
+
+  bool ret = true;
+  if (g_num_file_mapping_hints >= kMaxFileMappingHints) {
+    ret = false;
+  } else {
+    // TODO(ckennelly): Move this into a string copy routine.
+    size_t len = strlen(filename);
+    char *dst = static_cast<char *>(
+        base_internal::LowLevelAlloc::AllocWithArena(len + 1, SigSafeArena()));
+    ABSL_RAW_CHECK(dst != nullptr, "out of memory");
+    memcpy(dst, filename, len + 1);
+
+    auto &hint = g_file_mapping_hints[g_num_file_mapping_hints++];
+    hint.start = start;
+    hint.end = end;
+    hint.offset = offset;
+    hint.filename = dst;
+  }
+
+  g_file_mapping_mu.Unlock();
+  return ret;
+}
+
+bool GetFileMappingHint(const void **start, const void **end, uint64_t *offset,
+                        const char **filename) {
+  if (!g_file_mapping_mu.TryLock()) {
+    return false;
+  }
+  bool found = false;
+  for (int i = 0; i < g_num_file_mapping_hints; i++) {
+    if (g_file_mapping_hints[i].start <= *start &&
+        *end <= g_file_mapping_hints[i].end) {
+      // We assume that the start_address for the mapping is the base
+      // address of the ELF section, but when [start_address,end_address) is
+      // not strictly equal to [hint.start, hint.end), that assumption is
+      // invalid.
+      //
+      // This uses the hint's start address (even though hint.start is not
+      // necessarily equal to start_address) to ensure the correct
+      // relocation is computed later.
+      *start = g_file_mapping_hints[i].start;
+      *end = g_file_mapping_hints[i].end;
+      *offset = g_file_mapping_hints[i].offset;
+      *filename = g_file_mapping_hints[i].filename;
+      found = true;
+      break;
+    }
+  }
+  g_file_mapping_mu.Unlock();
+  return found;
+}
+
+}  // namespace debugging_internal
+
+bool Symbolize(const void *pc, char *out, int out_size) {
+  // Symbolization is very slow under tsan.
+  ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN();
+  SAFE_ASSERT(out_size >= 0);
+  debugging_internal::Symbolizer *s = debugging_internal::AllocateSymbolizer();
+  const char *name = s->GetSymbol(pc);
+  bool ok = false;
+  if (name != nullptr && out_size > 0) {
+    strncpy(out, name, static_cast<size_t>(out_size));
+    ok = true;
+    if (out[static_cast<size_t>(out_size) - 1] != '\0') {
+      // strncpy() does not '\0' terminate when it truncates.  Do so, with
+      // trailing ellipsis.
+      static constexpr char kEllipsis[] = "...";
+      size_t ellipsis_size =
+          std::min(strlen(kEllipsis), static_cast<size_t>(out_size) - 1);
+      memcpy(out + static_cast<size_t>(out_size) - ellipsis_size - 1, kEllipsis,
+             ellipsis_size);
+      out[static_cast<size_t>(out_size) - 1] = '\0';
+    }
+  }
+  debugging_internal::FreeSymbolizer(s);
+  ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_END();
+  return ok;
+}
+
+ABSL_NAMESPACE_END
+}  // namespace absl
+
+extern "C" bool AbslInternalGetFileMappingHint(const void **start,
+                                               const void **end, uint64_t *offset,
+                                               const char **filename) {
+  return absl::debugging_internal::GetFileMappingHint(start, end, offset,
+                                                      filename);
+}