about summary refs log tree commit diff
diff options
context:
space:
mode:
authorAbseil Team <absl-team@google.com>2018-02-21T16·32-0800
committerDerek Mauro <dmauro@google.com>2018-02-22T15·15-0500
commitdca2eb50f9f5ee378395ee020953dc20a2a84c97 (patch)
tree31c1476614cb761bc3f9547d3a3d0d37cf948cca
parent03c1513538584f4a04d666be5eb469e3979febba (diff)
Changes imported from Abseil "staging" branch:
  - db061dd2b795837e6154be0991077ca5a03ec471 Release the async-signal safe demangler as an internal im... by Derek Mauro <dmauro@google.com>
  - 310440eb33d946df79d26bf1edf795f6a0b466b9 Use static_cast to constrain nanoseconds to int range, av... by Abseil Team <absl-team@google.com>

GitOrigin-RevId: db061dd2b795837e6154be0991077ca5a03ec471
Change-Id: Ibc1bc1db355a48115451da5ce88d66a1f6e1a182
-rw-r--r--absl/base/internal/spinlock_wait.cc2
-rw-r--r--absl/debugging/BUILD.bazel24
-rw-r--r--absl/debugging/CMakeLists.txt38
-rw-r--r--absl/debugging/internal/demangle.cc1862
-rw-r--r--absl/debugging/internal/demangle.h67
-rw-r--r--absl/debugging/internal/demangle_test.cc191
6 files changed, 2182 insertions, 2 deletions
diff --git a/absl/base/internal/spinlock_wait.cc b/absl/base/internal/spinlock_wait.cc
index 8f951b66dc36..9f6e9911e109 100644
--- a/absl/base/internal/spinlock_wait.cc
+++ b/absl/base/internal/spinlock_wait.cc
@@ -72,7 +72,7 @@ int SpinLockSuggestedDelayNS(int loop) {
   // Mean is exponential in loop for first 32 iterations, then 8ms.
   // The futex path multiplies this by 16, since we expect explicit wakeups
   // almost always on that path.
-  return r >> (44 - (loop >> 3));
+  return static_cast<int>(r >> (44 - (loop >> 3)));
 }
 
 }  // namespace base_internal
diff --git a/absl/debugging/BUILD.bazel b/absl/debugging/BUILD.bazel
index d4cdf49223b3..20b455b836aa 100644
--- a/absl/debugging/BUILD.bazel
+++ b/absl/debugging/BUILD.bazel
@@ -64,10 +64,34 @@ cc_library(
     deps = [
         "//absl/base",
         "//absl/base:dynamic_annotations",
+    ],
+)
+
+cc_library(
+    name = "demangle_internal",
+    srcs = ["internal/demangle.cc"],
+    hdrs = ["internal/demangle.h"],
+    copts = ABSL_DEFAULT_COPTS,
+    deps = [
+        "//absl/base",
         "//absl/base:core_headers",
     ],
 )
 
+cc_test(
+    name = "demangle_test",
+    srcs = ["internal/demangle_test.cc"],
+    copts = ABSL_TEST_COPTS,
+    deps = [
+        ":demangle_internal",
+        ":stack_consumption",
+        "//absl/base",
+        "//absl/base:core_headers",
+        "//absl/memory",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
+
 cc_library(
     name = "leak_check",
     srcs = select({
diff --git a/absl/debugging/CMakeLists.txt b/absl/debugging/CMakeLists.txt
index 57efe80fd05e..758c00fe3a61 100644
--- a/absl/debugging/CMakeLists.txt
+++ b/absl/debugging/CMakeLists.txt
@@ -22,6 +22,7 @@ list(APPEND DEBUGGING_PUBLIC_HEADERS
 
 list(APPEND DEBUGGING_INTERNAL_HEADERS
   "internal/address_is_readable.h"
+  "internal/demangle.h"
   "internal/elf_mem_image.h"
   "internal/stacktrace_config.h"
   "internal/vdso_support.h"
@@ -31,6 +32,7 @@ list(APPEND DEBUGGING_INTERNAL_HEADERS
 list(APPEND STACKTRACE_SRC
   "stacktrace.cc"
   "internal/address_is_readable.cc"
+  "internal/demangle.cc"
   "internal/elf_mem_image.cc"
   "internal/vdso_support.cc"
   ${DEBUGGING_PUBLIC_HEADERS}
@@ -79,6 +81,41 @@ absl_header_library(
 ## TESTS
 #
 
+list(APPEND DEBUGGING_INTERNAL_TEST_HEADERS
+  "internal/stack_consumption.h"
+)
+
+list(APPEND STACK_CONSUMPTION_SRC
+  "internal/stack_consumption.cc"
+  ${DEBUGGING_INTERNAL_TEST_HEADERS}
+)
+
+absl_library(
+  TARGET
+    absl_stack_consumption
+  SOURCES
+    ${STACK_CONSUMPTION_SRC}
+)
+
+absl_test(
+  TARGET
+    absl_stack_consumption_test
+  SOURCES
+    ${STACK_CONSUMPTION_SRC}
+)
+
+list(APPEND DEMANGLE_TEST_SRC "demangle_test.cc")
+
+absl_test(
+  TARGET
+    demangle_test
+  SOURCES
+    ${DEMANGLE_TEST_SRC}
+  PUBLIC_LIBRARIES
+    absl_stacktrace absl_stack_consumption
+)
+
+
 # test leak_check_test
 list(APPEND LEAK_CHECK_TEST_SRC "leak_check_test.cc")
 
@@ -90,4 +127,3 @@ absl_test(
   PUBLIC_LIBRARIES
     absl_leak_check
 )
-
diff --git a/absl/debugging/internal/demangle.cc b/absl/debugging/internal/demangle.cc
new file mode 100644
index 000000000000..c9ca2f3bdb17
--- /dev/null
+++ b/absl/debugging/internal/demangle.cc
@@ -0,0 +1,1862 @@
+// 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
+//
+//      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.
+
+// For reference check out:
+// https://itanium-cxx-abi.github.io/cxx-abi/abi.html#mangling
+//
+// Note that we only have partial C++11 support yet.
+
+#include "absl/debugging/internal/demangle.h"
+
+#include <cstdint>
+#include <cstdio>
+#include <limits>
+
+namespace absl {
+namespace debugging_internal {
+
+typedef struct {
+  const char *abbrev;
+  const char *real_name;
+  // Number of arguments in <expression> context, or 0 if disallowed.
+  int arity;
+} AbbrevPair;
+
+// List of operators from Itanium C++ ABI.
+static const AbbrevPair kOperatorList[] = {
+    // New has special syntax (not currently supported).
+    {"nw", "new", 0},
+    {"na", "new[]", 0},
+
+    // Works except that the 'gs' prefix is not supported.
+    {"dl", "delete", 1},
+    {"da", "delete[]", 1},
+
+    {"ps", "+", 1},  // "positive"
+    {"ng", "-", 1},  // "negative"
+    {"ad", "&", 1},  // "address-of"
+    {"de", "*", 1},  // "dereference"
+    {"co", "~", 1},
+
+    {"pl", "+", 2},
+    {"mi", "-", 2},
+    {"ml", "*", 2},
+    {"dv", "/", 2},
+    {"rm", "%", 2},
+    {"an", "&", 2},
+    {"or", "|", 2},
+    {"eo", "^", 2},
+    {"aS", "=", 2},
+    {"pL", "+=", 2},
+    {"mI", "-=", 2},
+    {"mL", "*=", 2},
+    {"dV", "/=", 2},
+    {"rM", "%=", 2},
+    {"aN", "&=", 2},
+    {"oR", "|=", 2},
+    {"eO", "^=", 2},
+    {"ls", "<<", 2},
+    {"rs", ">>", 2},
+    {"lS", "<<=", 2},
+    {"rS", ">>=", 2},
+    {"eq", "==", 2},
+    {"ne", "!=", 2},
+    {"lt", "<", 2},
+    {"gt", ">", 2},
+    {"le", "<=", 2},
+    {"ge", ">=", 2},
+    {"nt", "!", 1},
+    {"aa", "&&", 2},
+    {"oo", "||", 2},
+    {"pp", "++", 1},
+    {"mm", "--", 1},
+    {"cm", ",", 2},
+    {"pm", "->*", 2},
+    {"pt", "->", 0},  // Special syntax
+    {"cl", "()", 0},  // Special syntax
+    {"ix", "[]", 2},
+    {"qu", "?", 3},
+    {"st", "sizeof", 0},  // Special syntax
+    {"sz", "sizeof", 1},  // Not a real operator name, but used in expressions.
+    {nullptr, nullptr, 0},
+};
+
+// List of builtin types from Itanium C++ ABI.
+static const AbbrevPair kBuiltinTypeList[] = {
+    {"v", "void", 0},
+    {"w", "wchar_t", 0},
+    {"b", "bool", 0},
+    {"c", "char", 0},
+    {"a", "signed char", 0},
+    {"h", "unsigned char", 0},
+    {"s", "short", 0},
+    {"t", "unsigned short", 0},
+    {"i", "int", 0},
+    {"j", "unsigned int", 0},
+    {"l", "long", 0},
+    {"m", "unsigned long", 0},
+    {"x", "long long", 0},
+    {"y", "unsigned long long", 0},
+    {"n", "__int128", 0},
+    {"o", "unsigned __int128", 0},
+    {"f", "float", 0},
+    {"d", "double", 0},
+    {"e", "long double", 0},
+    {"g", "__float128", 0},
+    {"z", "ellipsis", 0},
+    {nullptr, nullptr, 0},
+};
+
+// List of substitutions Itanium C++ ABI.
+static const AbbrevPair kSubstitutionList[] = {
+    {"St", "", 0},
+    {"Sa", "allocator", 0},
+    {"Sb", "basic_string", 0},
+    // std::basic_string<char, std::char_traits<char>,std::allocator<char> >
+    {"Ss", "string", 0},
+    // std::basic_istream<char, std::char_traits<char> >
+    {"Si", "istream", 0},
+    // std::basic_ostream<char, std::char_traits<char> >
+    {"So", "ostream", 0},
+    // std::basic_iostream<char, std::char_traits<char> >
+    {"Sd", "iostream", 0},
+    {nullptr, nullptr, 0},
+};
+
+// State needed for demangling.  This struct is copied in almost every stack
+// frame, so every byte counts.
+typedef struct {
+  int mangled_idx;                   // Cursor of mangled name.
+  int out_cur_idx;                   // Cursor of output std::string.
+  int prev_name_idx;                 // For constructors/destructors.
+  signed int prev_name_length : 16;  // For constructors/destructors.
+  signed int nest_level : 15;        // For nested names.
+  unsigned int append : 1;           // Append flag.
+  // Note: for some reason MSVC can't pack "bool append : 1" into the same int
+  // with the above two fields, so we use an int instead.  Amusingly it can pack
+  // "signed bool" as expected, but relying on that to continue to be a legal
+  // type seems ill-advised (as it's illegal in at least clang).
+} ParseState;
+
+static_assert(sizeof(ParseState) == 4 * sizeof(int),
+              "unexpected size of ParseState");
+
+// One-off state for demangling that's not subject to backtracking -- either
+// constant data, data that's intentionally immune to backtracking (steps), or
+// data that would never be changed by backtracking anyway (recursion_depth).
+//
+// Only one copy of this exists for each call to Demangle, so the size of this
+// struct is nearly inconsequential.
+typedef struct {
+  const char *mangled_begin;  // Beginning of input std::string.
+  char *out;                  // Beginning of output std::string.
+  int out_end_idx;            // One past last allowed output character.
+  int recursion_depth;        // For stack exhaustion prevention.
+  int steps;               // Cap how much work we'll do, regardless of depth.
+  ParseState parse_state;  // Backtrackable state copied for most frames.
+} State;
+
+namespace {
+// Prevent deep recursion / stack exhaustion.
+// Also prevent unbounded handling of complex inputs.
+class ComplexityGuard {
+ public:
+  explicit ComplexityGuard(State *state) : state_(state) {
+    ++state->recursion_depth;
+    ++state->steps;
+  }
+  ~ComplexityGuard() { --state_->recursion_depth; }
+
+  // 256 levels of recursion seems like a reasonable upper limit on depth.
+  // 128 is not enough to demagle synthetic tests from demangle_unittest.txt:
+  // "_ZaaZZZZ..." and "_ZaaZcvZcvZ..."
+  static constexpr int kRecursionDepthLimit = 256;
+
+  // We're trying to pick a charitable upper-limit on how many parse steps are
+  // necessary to handle something that a human could actually make use of.
+  // This is mostly in place as a bound on how much work we'll do if we are
+  // asked to demangle an mangled name from an untrusted source, so it should be
+  // much larger than the largest expected symbol, but much smaller than the
+  // amount of work we can do in, e.g., a second.
+  //
+  // Some real-world symbols from an arbitrary binary started failing between
+  // 2^12 and 2^13, so we multiply the latter by an extra factor of 16 to set
+  // the limit.
+  //
+  // Spending one second on 2^17 parse steps would require each step to take
+  // 7.6us, or ~30000 clock cycles, so it's safe to say this can be done in
+  // under a second.
+  static constexpr int kParseStepsLimit = 1 << 17;
+
+  bool IsTooComplex() const {
+    return state_->recursion_depth > kRecursionDepthLimit ||
+           state_->steps > kParseStepsLimit;
+  }
+
+ private:
+  State *state_;
+};
+}  // namespace
+
+// We don't use strlen() in libc since it's not guaranteed to be async
+// signal safe.
+static size_t StrLen(const char *str) {
+  size_t len = 0;
+  while (*str != '\0') {
+    ++str;
+    ++len;
+  }
+  return len;
+}
+
+// Returns true if "str" has at least "n" characters remaining.
+static bool AtLeastNumCharsRemaining(const char *str, int n) {
+  for (int i = 0; i < n; ++i) {
+    if (str[i] == '\0') {
+      return false;
+    }
+  }
+  return true;
+}
+
+// Returns true if "str" has "prefix" as a prefix.
+static bool StrPrefix(const char *str, const char *prefix) {
+  size_t i = 0;
+  while (str[i] != '\0' && prefix[i] != '\0' && str[i] == prefix[i]) {
+    ++i;
+  }
+  return prefix[i] == '\0';  // Consumed everything in "prefix".
+}
+
+static void InitState(State *state, const char *mangled, char *out,
+                      int out_size) {
+  state->mangled_begin = mangled;
+  state->out = out;
+  state->out_end_idx = out_size;
+  state->recursion_depth = 0;
+  state->steps = 0;
+
+  state->parse_state.mangled_idx = 0;
+  state->parse_state.out_cur_idx = 0;
+  state->parse_state.prev_name_idx = 0;
+  state->parse_state.prev_name_length = -1;
+  state->parse_state.nest_level = -1;
+  state->parse_state.append = true;
+}
+
+static inline const char *RemainingInput(State *state) {
+  return &state->mangled_begin[state->parse_state.mangled_idx];
+}
+
+// Returns true and advances "mangled_idx" if we find "one_char_token"
+// at "mangled_idx" position.  It is assumed that "one_char_token" does
+// not contain '\0'.
+static bool ParseOneCharToken(State *state, const char one_char_token) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (RemainingInput(state)[0] == one_char_token) {
+    ++state->parse_state.mangled_idx;
+    return true;
+  }
+  return false;
+}
+
+// Returns true and advances "mangled_cur" if we find "two_char_token"
+// at "mangled_cur" position.  It is assumed that "two_char_token" does
+// not contain '\0'.
+static bool ParseTwoCharToken(State *state, const char *two_char_token) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (RemainingInput(state)[0] == two_char_token[0] &&
+      RemainingInput(state)[1] == two_char_token[1]) {
+    state->parse_state.mangled_idx += 2;
+    return true;
+  }
+  return false;
+}
+
+// Returns true and advances "mangled_cur" if we find any character in
+// "char_class" at "mangled_cur" position.
+static bool ParseCharClass(State *state, const char *char_class) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (RemainingInput(state)[0] == '\0') {
+    return false;
+  }
+  const char *p = char_class;
+  for (; *p != '\0'; ++p) {
+    if (RemainingInput(state)[0] == *p) {
+      ++state->parse_state.mangled_idx;
+      return true;
+    }
+  }
+  return false;
+}
+
+static bool ParseDigit(State *state, int *digit) {
+  char c = RemainingInput(state)[0];
+  if (ParseCharClass(state, "0123456789")) {
+    if (digit != nullptr) {
+      *digit = c - '0';
+    }
+    return true;
+  }
+  return false;
+}
+
+// This function is used for handling an optional non-terminal.
+static bool Optional(bool /*status*/) { return true; }
+
+// This function is used for handling <non-terminal>+ syntax.
+typedef bool (*ParseFunc)(State *);
+static bool OneOrMore(ParseFunc parse_func, State *state) {
+  if (parse_func(state)) {
+    while (parse_func(state)) {
+    }
+    return true;
+  }
+  return false;
+}
+
+// This function is used for handling <non-terminal>* syntax. The function
+// always returns true and must be followed by a termination token or a
+// terminating sequence not handled by parse_func (e.g.
+// ParseOneCharToken(state, 'E')).
+static bool ZeroOrMore(ParseFunc parse_func, State *state) {
+  while (parse_func(state)) {
+  }
+  return true;
+}
+
+// Append "str" at "out_cur_idx".  If there is an overflow, out_cur_idx is
+// set to out_end_idx+1.  The output std::string is ensured to
+// always terminate with '\0' as long as there is no overflow.
+static void Append(State *state, const char *const str, const int length) {
+  for (int i = 0; i < length; ++i) {
+    if (state->parse_state.out_cur_idx + 1 <
+        state->out_end_idx) {  // +1 for '\0'
+      state->out[state->parse_state.out_cur_idx++] = str[i];
+    } else {
+      // signal overflow
+      state->parse_state.out_cur_idx = state->out_end_idx + 1;
+      break;
+    }
+  }
+  if (state->parse_state.out_cur_idx < state->out_end_idx) {
+    state->out[state->parse_state.out_cur_idx] =
+        '\0';  // Terminate it with '\0'
+  }
+}
+
+// We don't use equivalents in libc to avoid locale issues.
+static bool IsLower(char c) { return c >= 'a' && c <= 'z'; }
+
+static bool IsAlpha(char c) {
+  return (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z');
+}
+
+static bool IsDigit(char c) { return c >= '0' && c <= '9'; }
+
+// Returns true if "str" is a function clone suffix.  These suffixes are used
+// by GCC 4.5.x and later versions (and our locally-modified version of GCC
+// 4.4.x) to indicate functions which have been cloned during optimization.
+// We treat any sequence (.<alpha>+.<digit>+)+ as a function clone suffix.
+static bool IsFunctionCloneSuffix(const char *str) {
+  size_t i = 0;
+  while (str[i] != '\0') {
+    // Consume a single .<alpha>+.<digit>+ sequence.
+    if (str[i] != '.' || !IsAlpha(str[i + 1])) {
+      return false;
+    }
+    i += 2;
+    while (IsAlpha(str[i])) {
+      ++i;
+    }
+    if (str[i] != '.' || !IsDigit(str[i + 1])) {
+      return false;
+    }
+    i += 2;
+    while (IsDigit(str[i])) {
+      ++i;
+    }
+  }
+  return true;  // Consumed everything in "str".
+}
+
+static bool EndsWith(State *state, const char chr) {
+  return state->parse_state.out_cur_idx > 0 &&
+         chr == state->out[state->parse_state.out_cur_idx - 1];
+}
+
+// Append "str" with some tweaks, iff "append" state is true.
+static void MaybeAppendWithLength(State *state, const char *const str,
+                                  const int length) {
+  if (state->parse_state.append && length > 0) {
+    // Append a space if the output buffer ends with '<' and "str"
+    // starts with '<' to avoid <<<.
+    if (str[0] == '<' && EndsWith(state, '<')) {
+      Append(state, " ", 1);
+    }
+    // Remember the last identifier name for ctors/dtors.
+    if (IsAlpha(str[0]) || str[0] == '_') {
+      state->parse_state.prev_name_idx = state->parse_state.out_cur_idx;
+      state->parse_state.prev_name_length = length;
+    }
+    Append(state, str, length);
+  }
+}
+
+// Appends a positive decimal number to the output if appending is enabled.
+static bool MaybeAppendDecimal(State *state, unsigned int val) {
+  // Max {32-64}-bit unsigned int is 20 digits.
+  constexpr size_t kMaxLength = 20;
+  char buf[kMaxLength];
+
+  // We can't use itoa or sprintf as neither is specified to be
+  // async-signal-safe.
+  if (state->parse_state.append) {
+    // We can't have a one-before-the-beginning pointer, so instead start with
+    // one-past-the-end and manipulate one character before the pointer.
+    char *p = &buf[kMaxLength];
+    do {  // val=0 is the only input that should write a leading zero digit.
+      *--p = (val % 10) + '0';
+      val /= 10;
+    } while (p > buf && val != 0);
+
+    // 'p' landed on the last character we set.  How convenient.
+    Append(state, p, kMaxLength - (p - buf));
+  }
+
+  return true;
+}
+
+// A convenient wrapper around MaybeAppendWithLength().
+// Returns true so that it can be placed in "if" conditions.
+static bool MaybeAppend(State *state, const char *const str) {
+  if (state->parse_state.append) {
+    int length = StrLen(str);
+    MaybeAppendWithLength(state, str, length);
+  }
+  return true;
+}
+
+// This function is used for handling nested names.
+static bool EnterNestedName(State *state) {
+  state->parse_state.nest_level = 0;
+  return true;
+}
+
+// This function is used for handling nested names.
+static bool LeaveNestedName(State *state, int16_t prev_value) {
+  state->parse_state.nest_level = prev_value;
+  return true;
+}
+
+// Disable the append mode not to print function parameters, etc.
+static bool DisableAppend(State *state) {
+  state->parse_state.append = false;
+  return true;
+}
+
+// Restore the append mode to the previous state.
+static bool RestoreAppend(State *state, bool prev_value) {
+  state->parse_state.append = prev_value;
+  return true;
+}
+
+// Increase the nest level for nested names.
+static void MaybeIncreaseNestLevel(State *state) {
+  if (state->parse_state.nest_level > -1) {
+    ++state->parse_state.nest_level;
+  }
+}
+
+// Appends :: for nested names if necessary.
+static void MaybeAppendSeparator(State *state) {
+  if (state->parse_state.nest_level >= 1) {
+    MaybeAppend(state, "::");
+  }
+}
+
+// Cancel the last separator if necessary.
+static void MaybeCancelLastSeparator(State *state) {
+  if (state->parse_state.nest_level >= 1 && state->parse_state.append &&
+      state->parse_state.out_cur_idx >= 2) {
+    state->parse_state.out_cur_idx -= 2;
+    state->out[state->parse_state.out_cur_idx] = '\0';
+  }
+}
+
+// Returns true if the identifier of the given length pointed to by
+// "mangled_cur" is anonymous namespace.
+static bool IdentifierIsAnonymousNamespace(State *state, int length) {
+  // Returns true if "anon_prefix" is a proper prefix of "mangled_cur".
+  static const char anon_prefix[] = "_GLOBAL__N_";
+  return (length > static_cast<int>(sizeof(anon_prefix) - 1) &&
+          StrPrefix(RemainingInput(state), anon_prefix));
+}
+
+// Forward declarations of our parsing functions.
+static bool ParseMangledName(State *state);
+static bool ParseEncoding(State *state);
+static bool ParseName(State *state);
+static bool ParseUnscopedName(State *state);
+static bool ParseNestedName(State *state);
+static bool ParsePrefix(State *state);
+static bool ParseUnqualifiedName(State *state);
+static bool ParseSourceName(State *state);
+static bool ParseLocalSourceName(State *state);
+static bool ParseUnnamedTypeName(State *state);
+static bool ParseNumber(State *state, int *number_out);
+static bool ParseFloatNumber(State *state);
+static bool ParseSeqId(State *state);
+static bool ParseIdentifier(State *state, int length);
+static bool ParseOperatorName(State *state, int *arity);
+static bool ParseSpecialName(State *state);
+static bool ParseCallOffset(State *state);
+static bool ParseNVOffset(State *state);
+static bool ParseVOffset(State *state);
+static bool ParseCtorDtorName(State *state);
+static bool ParseDecltype(State *state);
+static bool ParseType(State *state);
+static bool ParseCVQualifiers(State *state);
+static bool ParseBuiltinType(State *state);
+static bool ParseFunctionType(State *state);
+static bool ParseBareFunctionType(State *state);
+static bool ParseClassEnumType(State *state);
+static bool ParseArrayType(State *state);
+static bool ParsePointerToMemberType(State *state);
+static bool ParseTemplateParam(State *state);
+static bool ParseTemplateTemplateParam(State *state);
+static bool ParseTemplateArgs(State *state);
+static bool ParseTemplateArg(State *state);
+static bool ParseBaseUnresolvedName(State *state);
+static bool ParseUnresolvedName(State *state);
+static bool ParseExpression(State *state);
+static bool ParseExprPrimary(State *state);
+static bool ParseExprCastValue(State *state);
+static bool ParseLocalName(State *state);
+static bool ParseLocalNameSuffix(State *state);
+static bool ParseDiscriminator(State *state);
+static bool ParseSubstitution(State *state, bool accept_std);
+
+// Implementation note: the following code is a straightforward
+// translation of the Itanium C++ ABI defined in BNF with a couple of
+// exceptions.
+//
+// - Support GNU extensions not defined in the Itanium C++ ABI
+// - <prefix> and <template-prefix> are combined to avoid infinite loop
+// - Reorder patterns to shorten the code
+// - Reorder patterns to give greedier functions precedence
+//   We'll mark "Less greedy than" for these cases in the code
+//
+// Each parsing function changes the parse state and returns true on
+// success, or returns false and doesn't change the parse state (note:
+// the parse-steps counter increases regardless of success or failure).
+// To ensure that the parse state isn't changed in the latter case, we
+// save the original state before we call multiple parsing functions
+// consecutively with &&, and restore it if unsuccessful.  See
+// ParseEncoding() as an example of this convention.  We follow the
+// convention throughout the code.
+//
+// Originally we tried to do demangling without following the full ABI
+// syntax but it turned out we needed to follow the full syntax to
+// parse complicated cases like nested template arguments.  Note that
+// implementing a full-fledged demangler isn't trivial (libiberty's
+// cp-demangle.c has +4300 lines).
+//
+// Note that (foo) in <(foo) ...> is a modifier to be ignored.
+//
+// Reference:
+// - Itanium C++ ABI
+//   <https://mentorembedded.github.io/cxx-abi/abi.html#mangling>
+
+// <mangled-name> ::= _Z <encoding>
+static bool ParseMangledName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  return ParseTwoCharToken(state, "_Z") && ParseEncoding(state);
+}
+
+// <encoding> ::= <(function) name> <bare-function-type>
+//            ::= <(data) name>
+//            ::= <special-name>
+static bool ParseEncoding(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  // Implementing the first two productions together as <name>
+  // [<bare-function-type>] avoids exponential blowup of backtracking.
+  //
+  // Since Optional(...) can't fail, there's no need to copy the state for
+  // backtracking.
+  if (ParseName(state) && Optional(ParseBareFunctionType(state))) {
+    return true;
+  }
+
+  if (ParseSpecialName(state)) {
+    return true;
+  }
+  return false;
+}
+
+// <name> ::= <nested-name>
+//        ::= <unscoped-template-name> <template-args>
+//        ::= <unscoped-name>
+//        ::= <local-name>
+static bool ParseName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (ParseNestedName(state) || ParseLocalName(state)) {
+    return true;
+  }
+
+  // We reorganize the productions to avoid re-parsing unscoped names.
+  // - Inline <unscoped-template-name> productions:
+  //   <name> ::= <substitution> <template-args>
+  //          ::= <unscoped-name> <template-args>
+  //          ::= <unscoped-name>
+  // - Merge the two productions that start with unscoped-name:
+  //   <name> ::= <unscoped-name> [<template-args>]
+
+  ParseState copy = state->parse_state;
+  // "std<...>" isn't a valid name.
+  if (ParseSubstitution(state, /*accept_std=*/false) &&
+      ParseTemplateArgs(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Note there's no need to restore state after this since only the first
+  // subparser can fail.
+  return ParseUnscopedName(state) && Optional(ParseTemplateArgs(state));
+}
+
+// <unscoped-name> ::= <unqualified-name>
+//                 ::= St <unqualified-name>
+static bool ParseUnscopedName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (ParseUnqualifiedName(state)) {
+    return true;
+  }
+
+  ParseState copy = state->parse_state;
+  if (ParseTwoCharToken(state, "St") && MaybeAppend(state, "std::") &&
+      ParseUnqualifiedName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <ref-qualifer> ::= R // lvalue method reference qualifier
+//                ::= O // rvalue method reference qualifier
+static inline bool ParseRefQualifier(State *state) {
+  return ParseCharClass(state, "OR");
+}
+
+// <nested-name> ::= N [<CV-qualifiers>] [<ref-qualifier>] <prefix>
+//                   <unqualified-name> E
+//               ::= N [<CV-qualifiers>] [<ref-qualifier>] <template-prefix>
+//                   <template-args> E
+static bool ParseNestedName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'N') && EnterNestedName(state) &&
+      Optional(ParseCVQualifiers(state)) &&
+      Optional(ParseRefQualifier(state)) && ParsePrefix(state) &&
+      LeaveNestedName(state, copy.nest_level) &&
+      ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// This part is tricky.  If we literally translate them to code, we'll
+// end up infinite loop.  Hence we merge them to avoid the case.
+//
+// <prefix> ::= <prefix> <unqualified-name>
+//          ::= <template-prefix> <template-args>
+//          ::= <template-param>
+//          ::= <substitution>
+//          ::= # empty
+// <template-prefix> ::= <prefix> <(template) unqualified-name>
+//                   ::= <template-param>
+//                   ::= <substitution>
+static bool ParsePrefix(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  bool has_something = false;
+  while (true) {
+    MaybeAppendSeparator(state);
+    if (ParseTemplateParam(state) ||
+        ParseSubstitution(state, /*accept_std=*/true) ||
+        ParseUnscopedName(state) ||
+        (ParseOneCharToken(state, 'M') && ParseUnnamedTypeName(state))) {
+      has_something = true;
+      MaybeIncreaseNestLevel(state);
+      continue;
+    }
+    MaybeCancelLastSeparator(state);
+    if (has_something && ParseTemplateArgs(state)) {
+      return ParsePrefix(state);
+    } else {
+      break;
+    }
+  }
+  return true;
+}
+
+// <unqualified-name> ::= <operator-name>
+//                    ::= <ctor-dtor-name>
+//                    ::= <source-name>
+//                    ::= <local-source-name> // GCC extension; see below.
+//                    ::= <unnamed-type-name>
+static bool ParseUnqualifiedName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  return (ParseOperatorName(state, nullptr) || ParseCtorDtorName(state) ||
+          ParseSourceName(state) || ParseLocalSourceName(state) ||
+          ParseUnnamedTypeName(state));
+}
+
+// <source-name> ::= <positive length number> <identifier>
+static bool ParseSourceName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  int length = -1;
+  if (ParseNumber(state, &length) && ParseIdentifier(state, length)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <local-source-name> ::= L <source-name> [<discriminator>]
+//
+// References:
+//   http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31775
+//   http://gcc.gnu.org/viewcvs?view=rev&revision=124467
+static bool ParseLocalSourceName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'L') && ParseSourceName(state) &&
+      Optional(ParseDiscriminator(state))) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <unnamed-type-name> ::= Ut [<(nonnegative) number>] _
+//                     ::= <closure-type-name>
+// <closure-type-name> ::= Ul <lambda-sig> E [<(nonnegative) number>] _
+// <lambda-sig>        ::= <(parameter) type>+
+static bool ParseUnnamedTypeName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  // Type's 1-based index n is encoded as { "", n == 1; itoa(n-2), otherwise }.
+  // Optionally parse the encoded value into 'which' and add 2 to get the index.
+  int which = -1;
+
+  // Unnamed type local to function or class.
+  if (ParseTwoCharToken(state, "Ut") && Optional(ParseNumber(state, &which)) &&
+      which <= std::numeric_limits<int>::max() - 2 &&  // Don't overflow.
+      ParseOneCharToken(state, '_')) {
+    MaybeAppend(state, "{unnamed type#");
+    MaybeAppendDecimal(state, 2 + which);
+    MaybeAppend(state, "}");
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Closure type.
+  which = -1;
+  if (ParseTwoCharToken(state, "Ul") && DisableAppend(state) &&
+      OneOrMore(ParseType, state) && RestoreAppend(state, copy.append) &&
+      ParseOneCharToken(state, 'E') && Optional(ParseNumber(state, &which)) &&
+      which <= std::numeric_limits<int>::max() - 2 &&  // Don't overflow.
+      ParseOneCharToken(state, '_')) {
+    MaybeAppend(state, "{lambda()#");
+    MaybeAppendDecimal(state, 2 + which);
+    MaybeAppend(state, "}");
+    return true;
+  }
+  state->parse_state = copy;
+
+  return false;
+}
+
+// <number> ::= [n] <non-negative decimal integer>
+// If "number_out" is non-null, then *number_out is set to the value of the
+// parsed number on success.
+static bool ParseNumber(State *state, int *number_out) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  bool negative = false;
+  if (ParseOneCharToken(state, 'n')) {
+    negative = true;
+  }
+  const char *p = RemainingInput(state);
+  uint64_t number = 0;
+  for (; *p != '\0'; ++p) {
+    if (IsDigit(*p)) {
+      number = number * 10 + (*p - '0');
+    } else {
+      break;
+    }
+  }
+  // Apply the sign with uint64_t arithmetic so overflows aren't UB.  Gives
+  // "incorrect" results for out-of-range inputs, but negative values only
+  // appear for literals, which aren't printed.
+  if (negative) {
+    number = ~number + 1;
+  }
+  if (p != RemainingInput(state)) {  // Conversion succeeded.
+    state->parse_state.mangled_idx += p - RemainingInput(state);
+    if (number_out != nullptr) {
+      // Note: possibly truncate "number".
+      *number_out = number;
+    }
+    return true;
+  }
+  return false;
+}
+
+// Floating-point literals are encoded using a fixed-length lowercase
+// hexadecimal std::string.
+static bool ParseFloatNumber(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  const char *p = RemainingInput(state);
+  for (; *p != '\0'; ++p) {
+    if (!IsDigit(*p) && !(*p >= 'a' && *p <= 'f')) {
+      break;
+    }
+  }
+  if (p != RemainingInput(state)) {  // Conversion succeeded.
+    state->parse_state.mangled_idx += p - RemainingInput(state);
+    return true;
+  }
+  return false;
+}
+
+// The <seq-id> is a sequence number in base 36,
+// using digits and upper case letters
+static bool ParseSeqId(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  const char *p = RemainingInput(state);
+  for (; *p != '\0'; ++p) {
+    if (!IsDigit(*p) && !(*p >= 'A' && *p <= 'Z')) {
+      break;
+    }
+  }
+  if (p != RemainingInput(state)) {  // Conversion succeeded.
+    state->parse_state.mangled_idx += p - RemainingInput(state);
+    return true;
+  }
+  return false;
+}
+
+// <identifier> ::= <unqualified source code identifier> (of given length)
+static bool ParseIdentifier(State *state, int length) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (length < 0 || !AtLeastNumCharsRemaining(RemainingInput(state), length)) {
+    return false;
+  }
+  if (IdentifierIsAnonymousNamespace(state, length)) {
+    MaybeAppend(state, "(anonymous namespace)");
+  } else {
+    MaybeAppendWithLength(state, RemainingInput(state), length);
+  }
+  state->parse_state.mangled_idx += length;
+  return true;
+}
+
+// <operator-name> ::= nw, and other two letters cases
+//                 ::= cv <type>  # (cast)
+//                 ::= v  <digit> <source-name> # vendor extended operator
+static bool ParseOperatorName(State *state, int *arity) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (!AtLeastNumCharsRemaining(RemainingInput(state), 2)) {
+    return false;
+  }
+  // First check with "cv" (cast) case.
+  ParseState copy = state->parse_state;
+  if (ParseTwoCharToken(state, "cv") && MaybeAppend(state, "operator ") &&
+      EnterNestedName(state) && ParseType(state) &&
+      LeaveNestedName(state, copy.nest_level)) {
+    if (arity != nullptr) {
+      *arity = 1;
+    }
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Then vendor extended operators.
+  if (ParseOneCharToken(state, 'v') && ParseDigit(state, arity) &&
+      ParseSourceName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Other operator names should start with a lower alphabet followed
+  // by a lower/upper alphabet.
+  if (!(IsLower(RemainingInput(state)[0]) &&
+        IsAlpha(RemainingInput(state)[1]))) {
+    return false;
+  }
+  // We may want to perform a binary search if we really need speed.
+  const AbbrevPair *p;
+  for (p = kOperatorList; p->abbrev != nullptr; ++p) {
+    if (RemainingInput(state)[0] == p->abbrev[0] &&
+        RemainingInput(state)[1] == p->abbrev[1]) {
+      if (arity != nullptr) {
+        *arity = p->arity;
+      }
+      MaybeAppend(state, "operator");
+      if (IsLower(*p->real_name)) {  // new, delete, etc.
+        MaybeAppend(state, " ");
+      }
+      MaybeAppend(state, p->real_name);
+      state->parse_state.mangled_idx += 2;
+      return true;
+    }
+  }
+  return false;
+}
+
+// <special-name> ::= TV <type>
+//                ::= TT <type>
+//                ::= TI <type>
+//                ::= TS <type>
+//                ::= Tc <call-offset> <call-offset> <(base) encoding>
+//                ::= GV <(object) name>
+//                ::= T <call-offset> <(base) encoding>
+// G++ extensions:
+//                ::= TC <type> <(offset) number> _ <(base) type>
+//                ::= TF <type>
+//                ::= TJ <type>
+//                ::= GR <name>
+//                ::= GA <encoding>
+//                ::= Th <call-offset> <(base) encoding>
+//                ::= Tv <call-offset> <(base) encoding>
+//
+// Note: we don't care much about them since they don't appear in
+// stack traces.  The are special data.
+static bool ParseSpecialName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'T') && ParseCharClass(state, "VTIS") &&
+      ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseTwoCharToken(state, "Tc") && ParseCallOffset(state) &&
+      ParseCallOffset(state) && ParseEncoding(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseTwoCharToken(state, "GV") && ParseName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'T') && ParseCallOffset(state) &&
+      ParseEncoding(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // G++ extensions
+  if (ParseTwoCharToken(state, "TC") && ParseType(state) &&
+      ParseNumber(state, nullptr) && ParseOneCharToken(state, '_') &&
+      DisableAppend(state) && ParseType(state)) {
+    RestoreAppend(state, copy.append);
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'T') && ParseCharClass(state, "FJ") &&
+      ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseTwoCharToken(state, "GR") && ParseName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseTwoCharToken(state, "GA") && ParseEncoding(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'T') && ParseCharClass(state, "hv") &&
+      ParseCallOffset(state) && ParseEncoding(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <call-offset> ::= h <nv-offset> _
+//               ::= v <v-offset> _
+static bool ParseCallOffset(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'h') && ParseNVOffset(state) &&
+      ParseOneCharToken(state, '_')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'v') && ParseVOffset(state) &&
+      ParseOneCharToken(state, '_')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  return false;
+}
+
+// <nv-offset> ::= <(offset) number>
+static bool ParseNVOffset(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  return ParseNumber(state, nullptr);
+}
+
+// <v-offset>  ::= <(offset) number> _ <(virtual offset) number>
+static bool ParseVOffset(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseNumber(state, nullptr) && ParseOneCharToken(state, '_') &&
+      ParseNumber(state, nullptr)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <ctor-dtor-name> ::= C1 | C2 | C3
+//                  ::= D0 | D1 | D2
+// # GCC extensions: "unified" constructor/destructor.  See
+// # https://github.com/gcc-mirror/gcc/blob/7ad17b583c3643bd4557f29b8391ca7ef08391f5/gcc/cp/mangle.c#L1847
+//                  ::= C4 | D4
+static bool ParseCtorDtorName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'C') && ParseCharClass(state, "1234")) {
+    const char *const prev_name = state->out + state->parse_state.prev_name_idx;
+    MaybeAppendWithLength(state, prev_name,
+                          state->parse_state.prev_name_length);
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'D') && ParseCharClass(state, "0124")) {
+    const char *const prev_name = state->out + state->parse_state.prev_name_idx;
+    MaybeAppend(state, "~");
+    MaybeAppendWithLength(state, prev_name,
+                          state->parse_state.prev_name_length);
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <decltype> ::= Dt <expression> E  # decltype of an id-expression or class
+//                                   # member access (C++0x)
+//            ::= DT <expression> E  # decltype of an expression (C++0x)
+static bool ParseDecltype(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'D') && ParseCharClass(state, "tT") &&
+      ParseExpression(state) && ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  return false;
+}
+
+// <type> ::= <CV-qualifiers> <type>
+//        ::= P <type>   # pointer-to
+//        ::= R <type>   # reference-to
+//        ::= O <type>   # rvalue reference-to (C++0x)
+//        ::= C <type>   # complex pair (C 2000)
+//        ::= G <type>   # imaginary (C 2000)
+//        ::= U <source-name> <type>  # vendor extended type qualifier
+//        ::= <builtin-type>
+//        ::= <function-type>
+//        ::= <class-enum-type>  # note: just an alias for <name>
+//        ::= <array-type>
+//        ::= <pointer-to-member-type>
+//        ::= <template-template-param> <template-args>
+//        ::= <template-param>
+//        ::= <decltype>
+//        ::= <substitution>
+//        ::= Dp <type>          # pack expansion of (C++0x)
+//
+static bool ParseType(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+
+  // We should check CV-qualifers, and PRGC things first.
+  //
+  // CV-qualifiers overlap with some operator names, but an operator name is not
+  // valid as a type.  To avoid an ambiguity that can lead to exponential time
+  // complexity, refuse to backtrack the CV-qualifiers.
+  //
+  // _Z4aoeuIrMvvE
+  //  => _Z 4aoeuI        rM  v     v   E
+  //         aoeu<operator%=, void, void>
+  //  => _Z 4aoeuI r Mv v              E
+  //         aoeu<void void::* restrict>
+  //
+  // By consuming the CV-qualifiers first, the former parse is disabled.
+  if (ParseCVQualifiers(state)) {
+    const bool result = ParseType(state);
+    if (!result) state->parse_state = copy;
+    return result;
+  }
+  state->parse_state = copy;
+
+  // Similarly, these tag characters can overlap with other <name>s resulting in
+  // two different parse prefixes that land on <template-args> in the same
+  // place, such as "C3r1xI...".  So, disable the "ctor-name = C3" parse by
+  // refusing to backtrack the tag characters.
+  if (ParseCharClass(state, "OPRCG")) {
+    const bool result = ParseType(state);
+    if (!result) state->parse_state = copy;
+    return result;
+  }
+  state->parse_state = copy;
+
+  if (ParseTwoCharToken(state, "Dp") && ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'U') && ParseSourceName(state) &&
+      ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseBuiltinType(state) || ParseFunctionType(state) ||
+      ParseClassEnumType(state) || ParseArrayType(state) ||
+      ParsePointerToMemberType(state) || ParseDecltype(state) ||
+      // "std" on its own isn't a type.
+      ParseSubstitution(state, /*accept_std=*/false)) {
+    return true;
+  }
+
+  if (ParseTemplateTemplateParam(state) && ParseTemplateArgs(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Less greedy than <template-template-param> <template-args>.
+  if (ParseTemplateParam(state)) {
+    return true;
+  }
+
+  return false;
+}
+
+// <CV-qualifiers> ::= [r] [V] [K]
+// We don't allow empty <CV-qualifiers> to avoid infinite loop in
+// ParseType().
+static bool ParseCVQualifiers(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  int num_cv_qualifiers = 0;
+  num_cv_qualifiers += ParseOneCharToken(state, 'r');
+  num_cv_qualifiers += ParseOneCharToken(state, 'V');
+  num_cv_qualifiers += ParseOneCharToken(state, 'K');
+  return num_cv_qualifiers > 0;
+}
+
+// <builtin-type> ::= v, etc.
+//                ::= u <source-name>
+static bool ParseBuiltinType(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  const AbbrevPair *p;
+  for (p = kBuiltinTypeList; p->abbrev != nullptr; ++p) {
+    if (RemainingInput(state)[0] == p->abbrev[0]) {
+      MaybeAppend(state, p->real_name);
+      ++state->parse_state.mangled_idx;
+      return true;
+    }
+  }
+
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'u') && ParseSourceName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <function-type> ::= F [Y] <bare-function-type> E
+static bool ParseFunctionType(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'F') &&
+      Optional(ParseOneCharToken(state, 'Y')) && ParseBareFunctionType(state) &&
+      ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <bare-function-type> ::= <(signature) type>+
+static bool ParseBareFunctionType(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  DisableAppend(state);
+  if (OneOrMore(ParseType, state)) {
+    RestoreAppend(state, copy.append);
+    MaybeAppend(state, "()");
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <class-enum-type> ::= <name>
+static bool ParseClassEnumType(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  return ParseName(state);
+}
+
+// <array-type> ::= A <(positive dimension) number> _ <(element) type>
+//              ::= A [<(dimension) expression>] _ <(element) type>
+static bool ParseArrayType(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'A') && ParseNumber(state, nullptr) &&
+      ParseOneCharToken(state, '_') && ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'A') && Optional(ParseExpression(state)) &&
+      ParseOneCharToken(state, '_') && ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <pointer-to-member-type> ::= M <(class) type> <(member) type>
+static bool ParsePointerToMemberType(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'M') && ParseType(state) && ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <template-param> ::= T_
+//                  ::= T <parameter-2 non-negative number> _
+static bool ParseTemplateParam(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (ParseTwoCharToken(state, "T_")) {
+    MaybeAppend(state, "?");  // We don't support template substitutions.
+    return true;
+  }
+
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'T') && ParseNumber(state, nullptr) &&
+      ParseOneCharToken(state, '_')) {
+    MaybeAppend(state, "?");  // We don't support template substitutions.
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <template-template-param> ::= <template-param>
+//                           ::= <substitution>
+static bool ParseTemplateTemplateParam(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  return (ParseTemplateParam(state) ||
+          // "std" on its own isn't a template.
+          ParseSubstitution(state, /*accept_std=*/false));
+}
+
+// <template-args> ::= I <template-arg>+ E
+static bool ParseTemplateArgs(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  DisableAppend(state);
+  if (ParseOneCharToken(state, 'I') && OneOrMore(ParseTemplateArg, state) &&
+      ParseOneCharToken(state, 'E')) {
+    RestoreAppend(state, copy.append);
+    MaybeAppend(state, "<>");
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <template-arg>  ::= <type>
+//                 ::= <expr-primary>
+//                 ::= J <template-arg>* E        # argument pack
+//                 ::= X <expression> E
+static bool ParseTemplateArg(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'J') && ZeroOrMore(ParseTemplateArg, state) &&
+      ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // There can be significant overlap between the following leading to
+  // exponential backtracking:
+  //
+  //   <expr-primary> ::= L <type> <expr-cast-value> E
+  //                 e.g. L 2xxIvE 1                 E
+  //   <type>         ==> <local-source-name> <template-args>
+  //                 e.g. L 2xx               IvE
+  //
+  // This means parsing an entire <type> twice, and <type> can contain
+  // <template-arg>, so this can generate exponential backtracking.  There is
+  // only overlap when the remaining input starts with "L <source-name>", so
+  // parse all cases that can start this way jointly to share the common prefix.
+  //
+  // We have:
+  //
+  //   <template-arg> ::= <type>
+  //                  ::= <expr-primary>
+  //
+  // First, drop all the productions of <type> that must start with something
+  // other than 'L'.  All that's left is <class-enum-type>; inline it.
+  //
+  //   <type> ::= <nested-name> # starts with 'N'
+  //          ::= <unscoped-name>
+  //          ::= <unscoped-template-name> <template-args>
+  //          ::= <local-name> # starts with 'Z'
+  //
+  // Drop and inline again:
+  //
+  //   <type> ::= <unscoped-name>
+  //          ::= <unscoped-name> <template-args>
+  //          ::= <substitution> <template-args> # starts with 'S'
+  //
+  // Merge the first two, inline <unscoped-name>, drop last:
+  //
+  //   <type> ::= <unqualified-name> [<template-args>]
+  //          ::= St <unqualified-name> [<template-args>] # starts with 'S'
+  //
+  // Drop and inline:
+  //
+  //   <type> ::= <operator-name> [<template-args>] # starts with lowercase
+  //          ::= <ctor-dtor-name> [<template-args>] # starts with 'C' or 'D'
+  //          ::= <source-name> [<template-args>] # starts with digit
+  //          ::= <local-source-name> [<template-args>]
+  //          ::= <unnamed-type-name> [<template-args>] # starts with 'U'
+  //
+  // One more time:
+  //
+  //   <type> ::= L <source-name> [<template-args>]
+  //
+  // Likewise with <expr-primary>:
+  //
+  //   <expr-primary> ::= L <type> <expr-cast-value> E
+  //                  ::= LZ <encoding> E # cannot overlap; drop
+  //                  ::= L <mangled_name> E # cannot overlap; drop
+  //
+  // By similar reasoning as shown above, the only <type>s starting with
+  // <source-name> are "<source-name> [<template-args>]".  Inline this.
+  //
+  //   <expr-primary> ::= L <source-name> [<template-args>] <expr-cast-value> E
+  //
+  // Now inline both of these into <template-arg>:
+  //
+  //   <template-arg> ::= L <source-name> [<template-args>]
+  //                  ::= L <source-name> [<template-args>] <expr-cast-value> E
+  //
+  // Merge them and we're done:
+  //   <template-arg>
+  //     ::= L <source-name> [<template-args>] [<expr-cast-value> E]
+  if (ParseLocalSourceName(state) && Optional(ParseTemplateArgs(state))) {
+    copy = state->parse_state;
+    if (ParseExprCastValue(state) && ParseOneCharToken(state, 'E')) {
+      return true;
+    }
+    state->parse_state = copy;
+    return true;
+  }
+
+  // Now that the overlapping cases can't reach this code, we can safely call
+  // both of these.
+  if (ParseType(state) || ParseExprPrimary(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'X') && ParseExpression(state) &&
+      ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <unresolved-type> ::= <template-param> [<template-args>]
+//                   ::= <decltype>
+//                   ::= <substitution>
+static inline bool ParseUnresolvedType(State *state) {
+  // No ComplexityGuard because we don't copy the state in this stack frame.
+  return (ParseTemplateParam(state) && Optional(ParseTemplateArgs(state))) ||
+         ParseDecltype(state) || ParseSubstitution(state, /*accept_std=*/false);
+}
+
+// <simple-id> ::= <source-name> [<template-args>]
+static inline bool ParseSimpleId(State *state) {
+  // No ComplexityGuard because we don't copy the state in this stack frame.
+
+  // Note: <simple-id> cannot be followed by a parameter pack; see comment in
+  // ParseUnresolvedType.
+  return ParseSourceName(state) && Optional(ParseTemplateArgs(state));
+}
+
+// <base-unresolved-name> ::= <source-name> [<template-args>]
+//                        ::= on <operator-name> [<template-args>]
+//                        ::= dn <destructor-name>
+static bool ParseBaseUnresolvedName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+
+  if (ParseSimpleId(state)) {
+    return true;
+  }
+
+  ParseState copy = state->parse_state;
+  if (ParseTwoCharToken(state, "on") && ParseOperatorName(state, nullptr) &&
+      Optional(ParseTemplateArgs(state))) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseTwoCharToken(state, "dn") &&
+      (ParseUnresolvedType(state) || ParseSimpleId(state))) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  return false;
+}
+
+// <unresolved-name> ::= [gs] <base-unresolved-name>
+//                   ::= sr <unresolved-type> <base-unresolved-name>
+//                   ::= srN <unresolved-type> <unresolved-qualifier-level>+ E
+//                         <base-unresolved-name>
+//                   ::= [gs] sr <unresolved-qualifier-level>+ E
+//                         <base-unresolved-name>
+static bool ParseUnresolvedName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+
+  ParseState copy = state->parse_state;
+  if (Optional(ParseTwoCharToken(state, "gs")) &&
+      ParseBaseUnresolvedName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseTwoCharToken(state, "sr") && ParseUnresolvedType(state) &&
+      ParseBaseUnresolvedName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseTwoCharToken(state, "sr") && ParseOneCharToken(state, 'N') &&
+      ParseUnresolvedType(state) &&
+      OneOrMore(/* <unresolved-qualifier-level> ::= */ ParseSimpleId, state) &&
+      ParseOneCharToken(state, 'E') && ParseBaseUnresolvedName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (Optional(ParseTwoCharToken(state, "gs")) &&
+      ParseTwoCharToken(state, "sr") &&
+      OneOrMore(/* <unresolved-qualifier-level> ::= */ ParseSimpleId, state) &&
+      ParseOneCharToken(state, 'E') && ParseBaseUnresolvedName(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  return false;
+}
+
+// <expression> ::= <1-ary operator-name> <expression>
+//              ::= <2-ary operator-name> <expression> <expression>
+//              ::= <3-ary operator-name> <expression> <expression> <expression>
+//              ::= cl <expression>+ E
+//              ::= cv <type> <expression>      # type (expression)
+//              ::= cv <type> _ <expression>* E # type (expr-list)
+//              ::= st <type>
+//              ::= <template-param>
+//              ::= <function-param>
+//              ::= <expr-primary>
+//              ::= dt <expression> <unresolved-name> # expr.name
+//              ::= pt <expression> <unresolved-name> # expr->name
+//              ::= sp <expression>         # argument pack expansion
+//              ::= sr <type> <unqualified-name> <template-args>
+//              ::= sr <type> <unqualified-name>
+// <function-param> ::= fp <(top-level) CV-qualifiers> _
+//                  ::= fp <(top-level) CV-qualifiers> <number> _
+//                  ::= fL <number> p <(top-level) CV-qualifiers> _
+//                  ::= fL <number> p <(top-level) CV-qualifiers> <number> _
+static bool ParseExpression(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (ParseTemplateParam(state) || ParseExprPrimary(state)) {
+    return true;
+  }
+
+  // Object/function call expression.
+  ParseState copy = state->parse_state;
+  if (ParseTwoCharToken(state, "cl") && OneOrMore(ParseExpression, state) &&
+      ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Function-param expression (level 0).
+  if (ParseTwoCharToken(state, "fp") && Optional(ParseCVQualifiers(state)) &&
+      Optional(ParseNumber(state, nullptr)) && ParseOneCharToken(state, '_')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Function-param expression (level 1+).
+  if (ParseTwoCharToken(state, "fL") && Optional(ParseNumber(state, nullptr)) &&
+      ParseOneCharToken(state, 'p') && Optional(ParseCVQualifiers(state)) &&
+      Optional(ParseNumber(state, nullptr)) && ParseOneCharToken(state, '_')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Parse the conversion expressions jointly to avoid re-parsing the <type> in
+  // their common prefix.  Parsed as:
+  // <expression> ::= cv <type> <conversion-args>
+  // <conversion-args> ::= _ <expression>* E
+  //                   ::= <expression>
+  //
+  // Also don't try ParseOperatorName after seeing "cv", since ParseOperatorName
+  // also needs to accept "cv <type>" in other contexts.
+  if (ParseTwoCharToken(state, "cv")) {
+    if (ParseType(state)) {
+      ParseState copy2 = state->parse_state;
+      if (ParseOneCharToken(state, '_') && ZeroOrMore(ParseExpression, state) &&
+          ParseOneCharToken(state, 'E')) {
+        return true;
+      }
+      state->parse_state = copy2;
+      if (ParseExpression(state)) {
+        return true;
+      }
+    }
+  } else {
+    // Parse unary, binary, and ternary operator expressions jointly, taking
+    // care not to re-parse subexpressions repeatedly. Parse like:
+    //   <expression> ::= <operator-name> <expression>
+    //                    [<one-to-two-expressions>]
+    //   <one-to-two-expressions> ::= <expression> [<expression>]
+    int arity = -1;
+    if (ParseOperatorName(state, &arity) &&
+        arity > 0 &&  // 0 arity => disabled.
+        (arity < 3 || ParseExpression(state)) &&
+        (arity < 2 || ParseExpression(state)) &&
+        (arity < 1 || ParseExpression(state))) {
+      return true;
+    }
+  }
+  state->parse_state = copy;
+
+  // sizeof type
+  if (ParseTwoCharToken(state, "st") && ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Object and pointer member access expressions.
+  if ((ParseTwoCharToken(state, "dt") || ParseTwoCharToken(state, "pt")) &&
+      ParseExpression(state) && ParseType(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Parameter pack expansion
+  if (ParseTwoCharToken(state, "sp") && ParseExpression(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  return ParseUnresolvedName(state);
+}
+
+// <expr-primary> ::= L <type> <(value) number> E
+//                ::= L <type> <(value) float> E
+//                ::= L <mangled-name> E
+//                // A bug in g++'s C++ ABI version 2 (-fabi-version=2).
+//                ::= LZ <encoding> E
+//
+// Warning, subtle: the "bug" LZ production above is ambiguous with the first
+// production where <type> starts with <local-name>, which can lead to
+// exponential backtracking in two scenarios:
+//
+// - When whatever follows the E in the <local-name> in the first production is
+//   not a name, we backtrack the whole <encoding> and re-parse the whole thing.
+//
+// - When whatever follows the <local-name> in the first production is not a
+//   number and this <expr-primary> may be followed by a name, we backtrack the
+//   <name> and re-parse it.
+//
+// Moreover this ambiguity isn't always resolved -- for example, the following
+// has two different parses:
+//
+//   _ZaaILZ4aoeuE1x1EvE
+//   => operator&&<aoeu, x, E, void>
+//   => operator&&<(aoeu::x)(1), void>
+//
+// To resolve this, we just do what GCC's demangler does, and refuse to parse
+// casts to <local-name> types.
+static bool ParseExprPrimary(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+
+  // The "LZ" special case: if we see LZ, we commit to accept "LZ <encoding> E"
+  // or fail, no backtracking.
+  if (ParseTwoCharToken(state, "LZ")) {
+    if (ParseEncoding(state) && ParseOneCharToken(state, 'E')) {
+      return true;
+    }
+
+    state->parse_state = copy;
+    return false;
+  }
+
+  // The merged cast production.
+  if (ParseOneCharToken(state, 'L') && ParseType(state) &&
+      ParseExprCastValue(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseOneCharToken(state, 'L') && ParseMangledName(state) &&
+      ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  return false;
+}
+
+// <number> or <float>, followed by 'E', as described above ParseExprPrimary.
+static bool ParseExprCastValue(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  // We have to be able to backtrack after accepting a number because we could
+  // have e.g. "7fffE", which will accept "7" as a number but then fail to find
+  // the 'E'.
+  ParseState copy = state->parse_state;
+  if (ParseNumber(state, nullptr) && ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  if (ParseFloatNumber(state) && ParseOneCharToken(state, 'E')) {
+    return true;
+  }
+  state->parse_state = copy;
+
+  return false;
+}
+
+// <local-name> ::= Z <(function) encoding> E <(entity) name> [<discriminator>]
+//              ::= Z <(function) encoding> E s [<discriminator>]
+//
+// Parsing a common prefix of these two productions together avoids an
+// exponential blowup of backtracking.  Parse like:
+//   <local-name> := Z <encoding> E <local-name-suffix>
+//   <local-name-suffix> ::= s [<discriminator>]
+//                       ::= <name> [<discriminator>]
+
+static bool ParseLocalNameSuffix(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+
+  if (MaybeAppend(state, "::") && ParseName(state) &&
+      Optional(ParseDiscriminator(state))) {
+    return true;
+  }
+
+  // Since we're not going to overwrite the above "::" by re-parsing the
+  // <encoding> (whose trailing '\0' byte was in the byte now holding the
+  // first ':'), we have to rollback the "::" if the <name> parse failed.
+  if (state->parse_state.append) {
+    state->out[state->parse_state.out_cur_idx - 2] = '\0';
+  }
+
+  return ParseOneCharToken(state, 's') && Optional(ParseDiscriminator(state));
+}
+
+static bool ParseLocalName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'Z') && ParseEncoding(state) &&
+      ParseOneCharToken(state, 'E') && ParseLocalNameSuffix(state)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <discriminator> := _ <(non-negative) number>
+static bool ParseDiscriminator(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, '_') && ParseNumber(state, nullptr)) {
+    return true;
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// <substitution> ::= S_
+//                ::= S <seq-id> _
+//                ::= St, etc.
+//
+// "St" is special in that it's not valid as a standalone name, and it *is*
+// allowed to precede a name without being wrapped in "N...E".  This means that
+// if we accept it on its own, we can accept "St1a" and try to parse
+// template-args, then fail and backtrack, accept "St" on its own, then "1a" as
+// an unqualified name and re-parse the same template-args.  To block this
+// exponential backtracking, we disable it with 'accept_std=false' in
+// problematic contexts.
+static bool ParseSubstitution(State *state, bool accept_std) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (ParseTwoCharToken(state, "S_")) {
+    MaybeAppend(state, "?");  // We don't support substitutions.
+    return true;
+  }
+
+  ParseState copy = state->parse_state;
+  if (ParseOneCharToken(state, 'S') && ParseSeqId(state) &&
+      ParseOneCharToken(state, '_')) {
+    MaybeAppend(state, "?");  // We don't support substitutions.
+    return true;
+  }
+  state->parse_state = copy;
+
+  // Expand abbreviations like "St" => "std".
+  if (ParseOneCharToken(state, 'S')) {
+    const AbbrevPair *p;
+    for (p = kSubstitutionList; p->abbrev != nullptr; ++p) {
+      if (RemainingInput(state)[0] == p->abbrev[1] &&
+          (accept_std || p->abbrev[1] != 't')) {
+        MaybeAppend(state, "std");
+        if (p->real_name[0] != '\0') {
+          MaybeAppend(state, "::");
+          MaybeAppend(state, p->real_name);
+        }
+        ++state->parse_state.mangled_idx;
+        return true;
+      }
+    }
+  }
+  state->parse_state = copy;
+  return false;
+}
+
+// Parse <mangled-name>, optionally followed by either a function-clone suffix
+// or version suffix.  Returns true only if all of "mangled_cur" was consumed.
+static bool ParseTopLevelMangledName(State *state) {
+  ComplexityGuard guard(state);
+  if (guard.IsTooComplex()) return false;
+  if (ParseMangledName(state)) {
+    if (RemainingInput(state)[0] != '\0') {
+      // Drop trailing function clone suffix, if any.
+      if (IsFunctionCloneSuffix(RemainingInput(state))) {
+        return true;
+      }
+      // Append trailing version suffix if any.
+      // ex. _Z3foo@@GLIBCXX_3.4
+      if (RemainingInput(state)[0] == '@') {
+        MaybeAppend(state, RemainingInput(state));
+        return true;
+      }
+      return false;  // Unconsumed suffix.
+    }
+    return true;
+  }
+  return false;
+}
+
+static bool Overflowed(const State *state) {
+  return state->parse_state.out_cur_idx >= state->out_end_idx;
+}
+
+// The demangler entry point.
+bool Demangle(const char *mangled, char *out, int out_size) {
+  State state;
+  InitState(&state, mangled, out, out_size);
+  return ParseTopLevelMangledName(&state) && !Overflowed(&state);
+}
+
+}  // namespace debugging_internal
+}  // namespace absl
diff --git a/absl/debugging/internal/demangle.h b/absl/debugging/internal/demangle.h
new file mode 100644
index 000000000000..2e75564ed35b
--- /dev/null
+++ b/absl/debugging/internal/demangle.h
@@ -0,0 +1,67 @@
+// 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
+//
+//      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.
+
+// An async-signal-safe and thread-safe demangler for Itanium C++ ABI
+// (aka G++ V3 ABI).
+//
+// The demangler is implemented to be used in async signal handlers to
+// symbolize stack traces.  We cannot use libstdc++'s
+// abi::__cxa_demangle() in such signal handlers since it's not async
+// signal safe (it uses malloc() internally).
+//
+// Note that this demangler doesn't support full demangling.  More
+// specifically, it doesn't print types of function parameters and
+// types of template arguments.  It just skips them.  However, it's
+// still very useful to extract basic information such as class,
+// function, constructor, destructor, and operator names.
+//
+// See the implementation note in demangle.cc if you are interested.
+//
+// Example:
+//
+// | Mangled Name  | The Demangler | abi::__cxa_demangle()
+// |---------------|---------------|-----------------------
+// | _Z1fv         | f()           | f()
+// | _Z1fi         | f()           | f(int)
+// | _Z3foo3bar    | foo()         | foo(bar)
+// | _Z1fIiEvi     | f<>()         | void f<int>(int)
+// | _ZN1N1fE      | N::f          | N::f
+// | _ZN3Foo3BarEv | Foo::Bar()    | Foo::Bar()
+// | _Zrm1XS_"     | operator%()   | operator%(X, X)
+// | _ZN3FooC1Ev   | Foo::Foo()    | Foo::Foo()
+// | _Z1fSs        | f()           | f(std::basic_string<char,
+// |               |               |   std::char_traits<char>,
+// |               |               |   std::allocator<char> >)
+//
+// See the unit test for more examples.
+//
+// Note: we might want to write demanglers for ABIs other than Itanium
+// C++ ABI in the future.
+//
+
+#ifndef ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
+#define ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
+
+namespace absl {
+namespace debugging_internal {
+
+// Demangle `mangled`.  On success, return true and write the
+// demangled symbol name to `out`.  Otherwise, return false.
+// `out` is modified even if demangling is unsuccessful.
+bool Demangle(const char *mangled, char *out, int out_size);
+
+}  // namespace debugging_internal
+}  // namespace absl
+
+#endif  // ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
diff --git a/absl/debugging/internal/demangle_test.cc b/absl/debugging/internal/demangle_test.cc
new file mode 100644
index 000000000000..b9d9008f0077
--- /dev/null
+++ b/absl/debugging/internal/demangle_test.cc
@@ -0,0 +1,191 @@
+// 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
+//
+//      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 "absl/debugging/internal/demangle.h"
+
+#include <cstdlib>
+#include <string>
+
+#include "gtest/gtest.h"
+#include "absl/base/internal/raw_logging.h"
+#include "absl/debugging/internal/stack_consumption.h"
+#include "absl/memory/memory.h"
+
+namespace absl {
+namespace debugging_internal {
+namespace {
+
+// A wrapper function for Demangle() to make the unit test simple.
+static const char *DemangleIt(const char * const mangled) {
+  static char demangled[4096];
+  if (Demangle(mangled, demangled, sizeof(demangled))) {
+    return demangled;
+  } else {
+    return mangled;
+  }
+}
+
+// Test corner cases of bounary conditions.
+TEST(Demangle, CornerCases) {
+  char tmp[10];
+  EXPECT_TRUE(Demangle("_Z6foobarv", tmp, sizeof(tmp)));
+  // sizeof("foobar()") == 9
+  EXPECT_STREQ("foobar()", tmp);
+  EXPECT_TRUE(Demangle("_Z6foobarv", tmp, 9));
+  EXPECT_STREQ("foobar()", tmp);
+  EXPECT_FALSE(Demangle("_Z6foobarv", tmp, 8));  // Not enough.
+  EXPECT_FALSE(Demangle("_Z6foobarv", tmp, 1));
+  EXPECT_FALSE(Demangle("_Z6foobarv", tmp, 0));
+  EXPECT_FALSE(Demangle("_Z6foobarv", nullptr, 0));  // Should not cause SEGV.
+  EXPECT_FALSE(Demangle("_Z1000000", tmp, 9));
+}
+
+// Test handling of functions suffixed with .clone.N, which is used
+// by GCC 4.5.x (and our locally-modified version of GCC 4.4.x), and
+// .constprop.N and .isra.N, which are used by GCC 4.6.x.  These
+// suffixes are used to indicate functions which have been cloned
+// during optimization.  We ignore these suffixes.
+TEST(Demangle, Clones) {
+  char tmp[20];
+  EXPECT_TRUE(Demangle("_ZL3Foov", tmp, sizeof(tmp)));
+  EXPECT_STREQ("Foo()", tmp);
+  EXPECT_TRUE(Demangle("_ZL3Foov.clone.3", tmp, sizeof(tmp)));
+  EXPECT_STREQ("Foo()", tmp);
+  EXPECT_TRUE(Demangle("_ZL3Foov.constprop.80", tmp, sizeof(tmp)));
+  EXPECT_STREQ("Foo()", tmp);
+  EXPECT_TRUE(Demangle("_ZL3Foov.isra.18", tmp, sizeof(tmp)));
+  EXPECT_STREQ("Foo()", tmp);
+  EXPECT_TRUE(Demangle("_ZL3Foov.isra.2.constprop.18", tmp, sizeof(tmp)));
+  EXPECT_STREQ("Foo()", tmp);
+  // Invalid (truncated), should not demangle.
+  EXPECT_FALSE(Demangle("_ZL3Foov.clo", tmp, sizeof(tmp)));
+  // Invalid (.clone. not followed by number), should not demangle.
+  EXPECT_FALSE(Demangle("_ZL3Foov.clone.", tmp, sizeof(tmp)));
+  // Invalid (.clone. followed by non-number), should not demangle.
+  EXPECT_FALSE(Demangle("_ZL3Foov.clone.foo", tmp, sizeof(tmp)));
+  // Invalid (.constprop. not followed by number), should not demangle.
+  EXPECT_FALSE(Demangle("_ZL3Foov.isra.2.constprop.", tmp, sizeof(tmp)));
+}
+
+// Tests that verify that Demangle footprint is within some limit.
+// They are not to be run under sanitizers as the sanitizers increase
+// stack consumption by about 4x.
+#if defined(ABSL_INTERNAL_HAVE_DEBUGGING_STACK_CONSUMPTION) && \
+    !ADDRESS_SANITIZER && !MEMORY_SANITIZER && !THREAD_SANITIZER
+
+static const char *g_mangled;
+static char g_demangle_buffer[4096];
+static char *g_demangle_result;
+
+static void DemangleSignalHandler(int signo) {
+  if (Demangle(g_mangled, g_demangle_buffer, sizeof(g_demangle_buffer))) {
+    g_demangle_result = g_demangle_buffer;
+  } else {
+    g_demangle_result = nullptr;
+  }
+}
+
+// Call Demangle and figure out the stack footprint of this call.
+static const char *DemangleStackConsumption(const char *mangled,
+                                            int *stack_consumed) {
+  g_mangled = mangled;
+  *stack_consumed = GetSignalHandlerStackConsumption(DemangleSignalHandler);
+  ABSL_RAW_LOG(INFO, "Stack consumption of Demangle: %d", *stack_consumed);
+  return g_demangle_result;
+}
+
+// Demangle stack consumption should be within 8kB for simple mangled names
+// with some level of nesting. With alternate signal stack we have 64K,
+// but some signal handlers run on thread stack, and could have arbitrarily
+// little space left (so we don't want to make this number too large).
+const int kStackConsumptionUpperLimit = 8192;
+
+// Returns a mangled name nested to the given depth.
+static std::string NestedMangledName(int depth) {
+  std::string mangled_name = "_Z1a";
+  if (depth > 0) {
+    mangled_name += "IXL";
+    mangled_name += NestedMangledName(depth - 1);
+    mangled_name += "EEE";
+  }
+  return mangled_name;
+}
+
+TEST(Demangle, DemangleStackConsumption) {
+  // Measure stack consumption of Demangle for nested mangled names of varying
+  // depth.  Since Demangle is implemented as a recursive descent parser,
+  // stack consumption will grow as the nesting depth increases.  By measuring
+  // the stack consumption for increasing depths, we can see the growing
+  // impact of any stack-saving changes made to the code for Demangle.
+  int stack_consumed = 0;
+
+  const char *demangled =
+      DemangleStackConsumption("_Z6foobarv", &stack_consumed);
+  EXPECT_STREQ("foobar()", demangled);
+  EXPECT_GT(stack_consumed, 0);
+  EXPECT_LT(stack_consumed, kStackConsumptionUpperLimit);
+
+  const std::string nested_mangled_name0 = NestedMangledName(0);
+  demangled = DemangleStackConsumption(nested_mangled_name0.c_str(),
+                                       &stack_consumed);
+  EXPECT_STREQ("a", demangled);
+  EXPECT_GT(stack_consumed, 0);
+  EXPECT_LT(stack_consumed, kStackConsumptionUpperLimit);
+
+  const std::string nested_mangled_name1 = NestedMangledName(1);
+  demangled = DemangleStackConsumption(nested_mangled_name1.c_str(),
+                                       &stack_consumed);
+  EXPECT_STREQ("a<>", demangled);
+  EXPECT_GT(stack_consumed, 0);
+  EXPECT_LT(stack_consumed, kStackConsumptionUpperLimit);
+
+  const std::string nested_mangled_name2 = NestedMangledName(2);
+  demangled = DemangleStackConsumption(nested_mangled_name2.c_str(),
+                                       &stack_consumed);
+  EXPECT_STREQ("a<>", demangled);
+  EXPECT_GT(stack_consumed, 0);
+  EXPECT_LT(stack_consumed, kStackConsumptionUpperLimit);
+
+  const std::string nested_mangled_name3 = NestedMangledName(3);
+  demangled = DemangleStackConsumption(nested_mangled_name3.c_str(),
+                                       &stack_consumed);
+  EXPECT_STREQ("a<>", demangled);
+  EXPECT_GT(stack_consumed, 0);
+  EXPECT_LT(stack_consumed, kStackConsumptionUpperLimit);
+}
+
+#endif  // Stack consumption tests
+
+static void TestOnInput(const char* input) {
+  static const int kOutSize = 1048576;
+  auto out = absl::make_unique<char[]>(kOutSize);
+  Demangle(input, out.get(), kOutSize);
+}
+
+TEST(DemangleRegression, NegativeLength) {
+  TestOnInput("_ZZn4");
+}
+TEST(DemangleRegression, DeeplyNestedArrayType) {
+  const int depth = 100000;
+  std::string data = "_ZStI";
+  data.reserve(data.size() + 3 * depth + 1);
+  for (int i = 0; i < depth; i++) {
+    data += "A1_";
+  }
+  TestOnInput(data.c_str());
+}
+
+}  // namespace
+}  // namespace debugging_internal
+}  // namespace absl