Adding upstream version 110.0.1.upstream/110.0.1upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

100 files changed, 20795 insertions, 0 deletions

diff --git a/memory/app.mozbuild b/memory/app.mozbuild
new file mode 100644
index 0000000000..d4f9c357df
--- /dev/null
+++ b/memory/app.mozbuild
@@ -0,0 +1,27 @@
+# -*- Mode: python; c-basic-offset: 4; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+# Indirectly necessary for chromium's lock.h, included from LogAlloc.cpp
+EXPORTS.mozilla += [
+    '/mozglue/misc/Printf.h',
+    '/xpcom/base/Logging.h',
+]
+
+DIRS += [
+    '/memory',
+    '/mfbt',
+]
+
+EXPORTS.mozilla += [
+    "/mozglue/misc/IntegerPrintfMacros.h",
+    "/mozglue/misc/Sprintf.h",
+]
+
+if CONFIG["OS_ARCH"] == "WINNT":
+    EXPORTS.mozilla += [
+        "/mozglue/misc/StackWalk_windows.h",
+        "/mozglue/misc/StackWalkThread.h",
+    ]
\ No newline at end of file
diff --git a/memory/build/Makefile.in b/memory/build/Makefile.in
new file mode 100644
index 0000000000..c41ec8ea7d
--- /dev/null
+++ b/memory/build/Makefile.in
@@ -0,0 +1,9 @@
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+# Workaround for alignment problems in gcov code. See
+# https://bugzilla.mozilla.org/show_bug.cgi?id=1413570#c2.
+ifeq ($(CPU_ARCH),x86)
+PROFILE_GEN_CFLAGS += -mno-sse2
+endif
diff --git a/memory/build/Mutex.cpp b/memory/build/Mutex.cpp
new file mode 100644
index 0000000000..8bc69635ef
--- /dev/null
+++ b/memory/build/Mutex.cpp
@@ -0,0 +1,21 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "Mutex.h"
+
+#if defined(XP_DARWIN)
+
+// static
+bool Mutex::SpinInKernelSpace() {
+  if (__builtin_available(macOS 10.15, *)) {
+    return true;
+  }
+
+  return false;
+}
+
+// static
+bool Mutex::gSpinInKernelSpace = SpinInKernelSpace();
+
+#endif  // defined(XP_DARWIN)
diff --git a/memory/build/Mutex.h b/memory/build/Mutex.h
new file mode 100644
index 0000000000..d165028b14
--- /dev/null
+++ b/memory/build/Mutex.h
@@ -0,0 +1,197 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef Mutex_h
+#define Mutex_h
+
+#if defined(XP_WIN)
+#  include <windows.h>
+#elif defined(XP_DARWIN)
+#  include "mozilla/Assertions.h"
+#  include <os/lock.h>
+#else
+#  include <pthread.h>
+#endif
+#include "mozilla/Attributes.h"
+#include "mozilla/ThreadSafety.h"
+
+#if defined(XP_DARWIN)
+// For information about the following undocumented flags and functions see
+// https://github.com/apple/darwin-xnu/blob/main/bsd/sys/ulock.h and
+// https://github.com/apple/darwin-libplatform/blob/main/private/os/lock_private.h
+#  define OS_UNFAIR_LOCK_DATA_SYNCHRONIZATION (0x00010000)
+#  define OS_UNFAIR_LOCK_ADAPTIVE_SPIN (0x00040000)
+
+extern "C" {
+
+typedef uint32_t os_unfair_lock_options_t;
+OS_UNFAIR_LOCK_AVAILABILITY
+OS_EXPORT OS_NOTHROW OS_NONNULL_ALL void os_unfair_lock_lock_with_options(
+    os_unfair_lock_t lock, os_unfair_lock_options_t options);
+}
+#endif  // defined(XP_DARWIN)
+
+// Mutexes based on spinlocks.  We can't use normal pthread spinlocks in all
+// places, because they require malloc()ed memory, which causes bootstrapping
+// issues in some cases.  We also can't use constructors, because for statics,
+// they would fire after the first use of malloc, resetting the locks.
+struct MOZ_CAPABILITY("mutex") Mutex {
+#if defined(XP_WIN)
+  CRITICAL_SECTION mMutex;
+#elif defined(XP_DARWIN)
+  os_unfair_lock mMutex;
+#else
+  pthread_mutex_t mMutex;
+#endif
+
+  // Initializes a mutex. Returns whether initialization succeeded.
+  inline bool Init() {
+#if defined(XP_WIN)
+    if (!InitializeCriticalSectionAndSpinCount(&mMutex, 5000)) {
+      return false;
+    }
+#elif defined(XP_DARWIN)
+    mMutex = OS_UNFAIR_LOCK_INIT;
+#elif defined(XP_LINUX) && !defined(ANDROID)
+    pthread_mutexattr_t attr;
+    if (pthread_mutexattr_init(&attr) != 0) {
+      return false;
+    }
+    pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_ADAPTIVE_NP);
+    if (pthread_mutex_init(&mMutex, &attr) != 0) {
+      pthread_mutexattr_destroy(&attr);
+      return false;
+    }
+    pthread_mutexattr_destroy(&attr);
+#else
+    if (pthread_mutex_init(&mMutex, nullptr) != 0) {
+      return false;
+    }
+#endif
+    return true;
+  }
+
+  inline void Lock() MOZ_CAPABILITY_ACQUIRE() {
+#if defined(XP_WIN)
+    EnterCriticalSection(&mMutex);
+#elif defined(XP_DARWIN)
+    // We rely on a non-public function to improve performance here.
+    // The OS_UNFAIR_LOCK_DATA_SYNCHRONIZATION flag informs the kernel that
+    // the calling thread is able to make progress even in absence of actions
+    // from other threads and the OS_UNFAIR_LOCK_ADAPTIVE_SPIN one causes the
+    // kernel to spin on a contested lock if the owning thread is running on
+    // the same physical core (presumably only on x86 CPUs given that ARM
+    // macs don't have cores capable of SMT). On versions of macOS older than
+    // 10.15 the latter is not available and we spin in userspace instead.
+    if (Mutex::gSpinInKernelSpace) {
+      os_unfair_lock_lock_with_options(
+          &mMutex,
+          OS_UNFAIR_LOCK_DATA_SYNCHRONIZATION | OS_UNFAIR_LOCK_ADAPTIVE_SPIN);
+    } else {
+#  if defined(__x86_64__)
+      // On older versions of macOS (10.14 and older) the
+      // `OS_UNFAIR_LOCK_ADAPTIVE_SPIN` flag is not supported by the kernel,
+      // we spin in user-space instead like `OSSpinLock` does:
+      // https://github.com/apple/darwin-libplatform/blob/215b09856ab5765b7462a91be7076183076600df/src/os/lock.c#L183-L198
+      // Note that `OSSpinLock` uses 1000 iterations on x86-64:
+      // https://github.com/apple/darwin-libplatform/blob/215b09856ab5765b7462a91be7076183076600df/src/os/lock.c#L93
+      // ...but we only use 100 like it does on ARM:
+      // https://github.com/apple/darwin-libplatform/blob/215b09856ab5765b7462a91be7076183076600df/src/os/lock.c#L90
+      // We choose this value because it yields the same results in our
+      // benchmarks but is less likely to have detrimental effects caused by
+      // excessive spinning.
+      uint32_t retries = 100;
+
+      do {
+        if (os_unfair_lock_trylock(&mMutex)) {
+          return;
+        }
+
+        __asm__ __volatile__("pause");
+      } while (retries--);
+
+      os_unfair_lock_lock_with_options(&mMutex,
+                                       OS_UNFAIR_LOCK_DATA_SYNCHRONIZATION);
+#  else
+      MOZ_CRASH("User-space spin-locks should never be used on ARM");
+#  endif  // defined(__x86_64__)
+    }
+#else
+    pthread_mutex_lock(&mMutex);
+#endif
+  }
+
+  inline void Unlock() MOZ_CAPABILITY_RELEASE() {
+#if defined(XP_WIN)
+    LeaveCriticalSection(&mMutex);
+#elif defined(XP_DARWIN)
+    os_unfair_lock_unlock(&mMutex);
+#else
+    pthread_mutex_unlock(&mMutex);
+#endif
+  }
+
+#if defined(XP_DARWIN)
+  static bool SpinInKernelSpace();
+  static bool gSpinInKernelSpace;
+#endif  // XP_DARWIN
+};
+
+// Mutex that can be used for static initialization.
+// On Windows, CRITICAL_SECTION requires a function call to be initialized,
+// but for the initialization lock, a static initializer calling the
+// function would be called too late. We need no-function-call
+// initialization, which SRWLock provides.
+// Ideally, we'd use the same type of locks everywhere, but SRWLocks
+// everywhere incur a performance penalty. See bug 1418389.
+#if defined(XP_WIN)
+struct MOZ_CAPABILITY("mutex") StaticMutex {
+  SRWLOCK mMutex;
+
+  inline void Lock() MOZ_CAPABILITY_ACQUIRE() {
+    AcquireSRWLockExclusive(&mMutex);
+  }
+
+  inline void Unlock() MOZ_CAPABILITY_RELEASE() {
+    ReleaseSRWLockExclusive(&mMutex);
+  }
+};
+
+// Normally, we'd use a constexpr constructor, but MSVC likes to create
+// static initializers anyways.
+#  define STATIC_MUTEX_INIT SRWLOCK_INIT
+
+#else
+typedef Mutex StaticMutex;
+
+#  if defined(XP_DARWIN)
+#    define STATIC_MUTEX_INIT OS_UNFAIR_LOCK_INIT
+#  elif defined(XP_LINUX) && !defined(ANDROID)
+#    define STATIC_MUTEX_INIT PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP
+#  else
+#    define STATIC_MUTEX_INIT PTHREAD_MUTEX_INITIALIZER
+#  endif
+
+#endif
+
+template <typename T>
+struct MOZ_SCOPED_CAPABILITY MOZ_RAII AutoLock {
+  explicit AutoLock(T& aMutex) MOZ_CAPABILITY_ACQUIRE(aMutex) : mMutex(aMutex) {
+    mMutex.Lock();
+  }
+
+  ~AutoLock() MOZ_CAPABILITY_RELEASE() { mMutex.Unlock(); }
+
+  AutoLock(const AutoLock&) = delete;
+  AutoLock(AutoLock&&) = delete;
+
+ private:
+  T& mMutex;
+};
+
+using MutexAutoLock = AutoLock<Mutex>;
+
+#endif
diff --git a/memory/build/Utils.h b/memory/build/Utils.h
new file mode 100644
index 0000000000..6032bf23d0
--- /dev/null
+++ b/memory/build/Utils.h
@@ -0,0 +1,136 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef Utils_h
+#define Utils_h
+
+#include <type_traits>
+
+#include "mozilla/CheckedInt.h"
+#include "mozilla/TemplateLib.h"
+
+// Helper for log2 of powers of 2 at compile time.
+template <size_t N>
+struct Log2 : mozilla::tl::CeilingLog2<N> {
+  using mozilla::tl::CeilingLog2<N>::value;
+  static_assert(1ULL << value == N, "Number is not a power of 2");
+};
+#define LOG2(N) Log2<N>::value
+
+enum class Order {
+  eLess = -1,
+  eEqual = 0,
+  eGreater = 1,
+};
+
+// Compare two integers. Returns whether the first integer is Less,
+// Equal or Greater than the second integer.
+template <typename T>
+Order CompareInt(T aValue1, T aValue2) {
+  static_assert(std::is_integral_v<T>, "Type must be integral");
+  if (aValue1 < aValue2) {
+    return Order::eLess;
+  }
+  if (aValue1 > aValue2) {
+    return Order::eGreater;
+  }
+  return Order::eEqual;
+}
+
+// Compare two addresses. Returns whether the first address is Less,
+// Equal or Greater than the second address.
+template <typename T>
+Order CompareAddr(T* aAddr1, T* aAddr2) {
+  return CompareInt(uintptr_t(aAddr1), uintptr_t(aAddr2));
+}
+
+// User-defined literals to make constants more legible
+constexpr size_t operator"" _KiB(unsigned long long int aNum) {
+  return size_t(aNum) * 1024;
+}
+
+constexpr size_t operator"" _KiB(long double aNum) {
+  return size_t(aNum * 1024);
+}
+
+constexpr size_t operator"" _MiB(unsigned long long int aNum) {
+  return size_t(aNum) * 1024_KiB;
+}
+
+constexpr size_t operator"" _MiB(long double aNum) {
+  return size_t(aNum * 1024_KiB);
+}
+
+constexpr double operator""_percent(long double aPercent) {
+  return double(aPercent) / 100;
+}
+
+// Helper for (fast) comparison of fractions without involving divisions or
+// floats.
+class Fraction {
+ public:
+  explicit constexpr Fraction(size_t aNumerator, size_t aDenominator)
+      : mNumerator(aNumerator), mDenominator(aDenominator) {}
+
+  MOZ_IMPLICIT constexpr Fraction(long double aValue)
+      // We use an arbitrary power of two as denominator that provides enough
+      // precision for our use case.
+      : mNumerator(aValue * 4096), mDenominator(4096) {}
+
+  inline bool operator<(const Fraction& aOther) const {
+#ifndef MOZ_DEBUG
+    // We are comparing A / B < C / D, with all A, B, C and D being positive
+    // numbers. Multiplying both sides with B * D, we have:
+    // (A * B * D) / B < (C * B * D) / D, which can then be simplified as
+    // A * D < C * B. When can thus compare our fractions without actually
+    // doing any division.
+    // This however assumes the multiplied quantities are small enough not
+    // to overflow the multiplication. We use CheckedInt on debug builds
+    // to enforce the assumption.
+    return mNumerator * aOther.mDenominator < aOther.mNumerator * mDenominator;
+#else
+    mozilla::CheckedInt<size_t> numerator(mNumerator);
+    mozilla::CheckedInt<size_t> denominator(mDenominator);
+    // value() asserts when the multiplication overflowed.
+    size_t lhs = (numerator * aOther.mDenominator).value();
+    size_t rhs = (aOther.mNumerator * denominator).value();
+    return lhs < rhs;
+#endif
+  }
+
+  inline bool operator>(const Fraction& aOther) const { return aOther < *this; }
+
+  inline bool operator>=(const Fraction& aOther) const {
+    return !(*this < aOther);
+  }
+
+  inline bool operator<=(const Fraction& aOther) const {
+    return !(*this > aOther);
+  }
+
+  inline bool operator==(const Fraction& aOther) const {
+#ifndef MOZ_DEBUG
+    // Same logic as operator<
+    return mNumerator * aOther.mDenominator == aOther.mNumerator * mDenominator;
+#else
+    mozilla::CheckedInt<size_t> numerator(mNumerator);
+    mozilla::CheckedInt<size_t> denominator(mDenominator);
+    size_t lhs = (numerator * aOther.mDenominator).value();
+    size_t rhs = (aOther.mNumerator * denominator).value();
+    return lhs == rhs;
+#endif
+  }
+
+  inline bool operator!=(const Fraction& aOther) const {
+    return !(*this == aOther);
+  }
+
+ private:
+  size_t mNumerator;
+  size_t mDenominator;
+};
+
+#endif
diff --git a/memory/build/fallback.cpp b/memory/build/fallback.cpp
new file mode 100644
index 0000000000..6f1791bed9
--- /dev/null
+++ b/memory/build/fallback.cpp
@@ -0,0 +1,40 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "mozmemory.h"
+#include "mozjemalloc.h"
+#include <stdlib.h>
+
+#ifndef HAVE_MEMALIGN
+MOZ_MEMORY_API void* memalign(size_t aAlignment, size_t aSize) {
+#  ifdef XP_WIN
+  return _aligned_malloc(aSize, aAlignment);
+#  else
+  void* ret;
+  if (posix_memalign(&ret, aAlignment, aSize) != 0) {
+    return nullptr;
+  }
+  return ret;
+#  endif
+}
+#endif
+
+struct SystemMalloc {
+#define MALLOC_DECL(name, return_type, ...)                                \
+  static inline return_type name(ARGS_HELPER(TYPED_ARGS, ##__VA_ARGS__)) { \
+    return ::name(ARGS_HELPER(ARGS, ##__VA_ARGS__));                       \
+  }
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#include "malloc_decls.h"
+};
+
+#define MALLOC_DECL(name, return_type, ...)                                   \
+  MOZ_JEMALLOC_API return_type name(ARGS_HELPER(TYPED_ARGS, ##__VA_ARGS__)) { \
+    return DummyArenaAllocator<SystemMalloc>::name(                           \
+        ARGS_HELPER(ARGS, ##__VA_ARGS__));                                    \
+  }
+#define MALLOC_FUNCS MALLOC_FUNCS_ARENA
+#include "malloc_decls.h"
diff --git a/memory/build/malloc_decls.h b/memory/build/malloc_decls.h
new file mode 100644
index 0000000000..49a7437afc
--- /dev/null
+++ b/memory/build/malloc_decls.h
@@ -0,0 +1,147 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// Helper header to declare all the supported malloc functions.
+// MALLOC_DECL arguments are:
+//   - function name
+//   - return type
+//   - argument types
+
+#ifndef malloc_decls_h
+#  define malloc_decls_h
+
+#  include "mozjemalloc_types.h"
+
+#  define MALLOC_FUNCS_MALLOC_BASE 1
+#  define MALLOC_FUNCS_MALLOC_EXTRA 2
+#  define MALLOC_FUNCS_MALLOC \
+    (MALLOC_FUNCS_MALLOC_BASE | MALLOC_FUNCS_MALLOC_EXTRA)
+#  define MALLOC_FUNCS_JEMALLOC 4
+#  define MALLOC_FUNCS_ARENA_BASE 8
+#  define MALLOC_FUNCS_ARENA_ALLOC 16
+#  define MALLOC_FUNCS_ARENA \
+    (MALLOC_FUNCS_ARENA_BASE | MALLOC_FUNCS_ARENA_ALLOC)
+#  define MALLOC_FUNCS_ALL \
+    (MALLOC_FUNCS_MALLOC | MALLOC_FUNCS_JEMALLOC | MALLOC_FUNCS_ARENA)
+
+#endif  // malloc_decls_h
+
+#ifndef MALLOC_FUNCS
+#  define MALLOC_FUNCS MALLOC_FUNCS_ALL
+#endif
+
+#ifdef MALLOC_DECL
+// NOTHROW_MALLOC_DECL is intended for functions where the standard library
+// declares the functions in question as `throw()`.  Not all platforms
+// consistent declare certain functions as `throw()`, though.
+
+// Bionic and OS X don't seem to care about `throw()`ness.
+#  if defined(ANDROID) || defined(XP_DARWIN)
+#    undef NOTHROW_MALLOC_DECL
+#    define NOTHROW_MALLOC_DECL MALLOC_DECL
+// Some places don't care about the distinction.
+#  elif !defined(NOTHROW_MALLOC_DECL)
+#    define NOTHROW_MALLOC_DECL MALLOC_DECL
+#  endif
+
+#  if MALLOC_FUNCS & MALLOC_FUNCS_MALLOC_BASE
+MALLOC_DECL(malloc, void*, size_t)
+MALLOC_DECL(calloc, void*, size_t, size_t)
+MALLOC_DECL(realloc, void*, void*, size_t)
+NOTHROW_MALLOC_DECL(free, void, void*)
+NOTHROW_MALLOC_DECL(memalign, void*, size_t, size_t)
+#  endif
+#  if MALLOC_FUNCS & MALLOC_FUNCS_MALLOC_EXTRA
+NOTHROW_MALLOC_DECL(posix_memalign, int, void**, size_t, size_t)
+NOTHROW_MALLOC_DECL(aligned_alloc, void*, size_t, size_t)
+NOTHROW_MALLOC_DECL(valloc, void*, size_t)
+NOTHROW_MALLOC_DECL(malloc_usable_size, size_t, usable_ptr_t)
+MALLOC_DECL(malloc_good_size, size_t, size_t)
+#  endif
+
+#  if MALLOC_FUNCS & MALLOC_FUNCS_JEMALLOC
+// The 2nd argument points to an optional array exactly
+// jemalloc_stats_num_bins() long to be filled in (if non-null).
+MALLOC_DECL(jemalloc_stats_internal, void, jemalloc_stats_t*,
+            jemalloc_bin_stats_t*)
+
+// Return the size of the jemalloc_bin_stats_t array.
+MALLOC_DECL(jemalloc_stats_num_bins, size_t)
+
+// On some operating systems (Mac), we use madvise(MADV_FREE) to hand pages
+// back to the operating system.  On Mac, the operating system doesn't take
+// this memory back immediately; instead, the OS takes it back only when the
+// machine is running out of physical memory.
+//
+// This is great from the standpoint of efficiency, but it makes measuring our
+// actual RSS difficult, because pages which we've MADV_FREE'd shouldn't count
+// against our RSS.
+//
+// This function explicitly purges any MADV_FREE'd pages from physical memory,
+// causing our reported RSS match the amount of memory we're actually using.
+//
+// Note that this call is expensive in two ways.  First, it may be slow to
+// execute, because it may make a number of slow syscalls to free memory.  This
+// function holds the big jemalloc locks, so basically all threads are blocked
+// while this function runs.
+//
+// This function is also expensive in that the next time we go to access a page
+// which we've just explicitly decommitted, the operating system has to attach
+// to it a physical page!  If we hadn't run this function, the OS would have
+// less work to do.
+//
+// If MALLOC_DOUBLE_PURGE is not defined, this function does nothing.
+MALLOC_DECL(jemalloc_purge_freed_pages, void)
+
+// Free all unused dirty pages in all arenas. Calling this function will slow
+// down subsequent allocations so it is recommended to use it only when
+// memory needs to be reclaimed at all costs (see bug 805855). This function
+// provides functionality similar to mallctl("arenas.purge") in jemalloc 3.
+MALLOC_DECL(jemalloc_free_dirty_pages, void)
+
+// Opt in or out of a thread local arena (bool argument is whether to opt-in
+// (true) or out (false)).
+MALLOC_DECL(jemalloc_thread_local_arena, void, bool)
+
+// Provide information about any allocation enclosing the given address.
+MALLOC_DECL(jemalloc_ptr_info, void, const void*, jemalloc_ptr_info_t*)
+#  endif
+
+#  if MALLOC_FUNCS & MALLOC_FUNCS_ARENA_BASE
+
+// Creates a separate arena, and returns its id, valid to use with moz_arena_*
+// functions. A helper is provided in mozmemory.h that doesn't take any
+// arena_params_t: moz_create_arena.
+MALLOC_DECL(moz_create_arena_with_params, arena_id_t, arena_params_t*)
+
+// Dispose of the given arena. Subsequent uses of the arena will crash.
+// Passing an invalid id (inexistent or already disposed) to this function
+// will crash. The arena must be empty prior to calling this function.
+MALLOC_DECL(moz_dispose_arena, void, arena_id_t)
+#  endif
+
+#  if MALLOC_FUNCS & MALLOC_FUNCS_ARENA_ALLOC
+// Same as the functions without the moz_arena_ prefix, but using arenas
+// created with moz_create_arena.
+// The contract, even if not enforced at runtime in some configurations,
+// is that moz_arena_realloc and moz_arena_free will crash if the given
+// arena doesn't own the given pointer. All functions will crash if the
+// arena id is invalid.
+// Although discouraged, plain realloc and free can still be used on
+// pointers allocated with these functions. Realloc will properly keep
+// new pointers in the same arena as the original.
+MALLOC_DECL(moz_arena_malloc, void*, arena_id_t, size_t)
+MALLOC_DECL(moz_arena_calloc, void*, arena_id_t, size_t, size_t)
+MALLOC_DECL(moz_arena_realloc, void*, arena_id_t, void*, size_t)
+MALLOC_DECL(moz_arena_free, void, arena_id_t, void*)
+MALLOC_DECL(moz_arena_memalign, void*, arena_id_t, size_t, size_t)
+#  endif
+
+#endif  // MALLOC_DECL
+
+#undef NOTHROW_MALLOC_DECL
+#undef MALLOC_DECL
+#undef MALLOC_FUNCS
diff --git a/memory/build/moz.build b/memory/build/moz.build
new file mode 100644
index 0000000000..70a2864585
--- /dev/null
+++ b/memory/build/moz.build
@@ -0,0 +1,62 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+EXPORTS += [
+    "malloc_decls.h",
+    "mozjemalloc_types.h",
+    "mozmemory.h",
+    "mozmemory_wrap.h",
+]
+
+LIBRARY_DEFINES["MOZ_HAS_MOZGLUE"] = True
+DEFINES["MOZ_MEMORY_IMPL"] = True
+
+if CONFIG["MOZ_REPLACE_MALLOC"]:
+    EXPORTS += [
+        "replace_malloc.h",
+        "replace_malloc_bridge.h",
+    ]
+
+if CONFIG["MOZ_PHC"]:
+    DEFINES["MOZ_PHC"] = True
+
+if CONFIG["MOZ_MEMORY"]:
+    UNIFIED_SOURCES += [
+        "mozjemalloc.cpp",
+        "mozmemory_wrap.cpp",
+    ]
+else:
+    UNIFIED_SOURCES += [
+        "fallback.cpp",
+    ]
+
+if CONFIG["OS_TARGET"] == "Darwin" and (
+    CONFIG["MOZ_REPLACE_MALLOC"] or CONFIG["MOZ_MEMORY"]
+):
+    SOURCES += [
+        "Mutex.cpp",
+        "zone.c",
+    ]
+
+Library("memory")
+
+if CONFIG["OS_TARGET"] == "Android" and CONFIG["CC_TYPE"] == "clang":
+    CXXFLAGS += [
+        "-Wno-tautological-pointer-compare",
+    ]
+
+if CONFIG["MOZ_BUILD_APP"] != "memory":
+    FINAL_LIBRARY = "mozglue"
+
+if CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+    DEFINES["MOZ_REPLACE_MALLOC_STATIC"] = True
+
+DisableStlWrapping()
+
+if CONFIG["CC_TYPE"] == "clang-cl":
+    AllowCompilerWarnings()  # workaround for bug 1090497
+
+REQUIRES_UNIFIED_BUILD = True
diff --git a/memory/build/mozjemalloc.cpp b/memory/build/mozjemalloc.cpp
new file mode 100644
index 0000000000..b8439b90de
--- /dev/null
+++ b/memory/build/mozjemalloc.cpp
@@ -0,0 +1,5190 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// Portions of this file were originally under the following license:
+//
+// Copyright (C) 2006-2008 Jason Evans <jasone@FreeBSD.org>.
+// All rights reserved.
+// Copyright (C) 2007-2017 Mozilla Foundation.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+// 1. Redistributions of source code must retain the above copyright
+//    notice(s), this list of conditions and the following disclaimer as
+//    the first lines of this file unmodified other than the possible
+//    addition of one or more copyright notices.
+// 2. Redistributions in binary form must reproduce the above copyright
+//    notice(s), this list of conditions and the following disclaimer in
+//    the documentation and/or other materials provided with the
+//    distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE
+// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
+// BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+// OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
+// EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// *****************************************************************************
+//
+// This allocator implementation is designed to provide scalable performance
+// for multi-threaded programs on multi-processor systems.  The following
+// features are included for this purpose:
+//
+//   + Multiple arenas are used if there are multiple CPUs, which reduces lock
+//     contention and cache sloshing.
+//
+//   + Cache line sharing between arenas is avoided for internal data
+//     structures.
+//
+//   + Memory is managed in chunks and runs (chunks can be split into runs),
+//     rather than as individual pages.  This provides a constant-time
+//     mechanism for associating allocations with particular arenas.
+//
+// Allocation requests are rounded up to the nearest size class, and no record
+// of the original request size is maintained.  Allocations are broken into
+// categories according to size class.  Assuming runtime defaults, the size
+// classes in each category are as follows (for x86, x86_64 and Apple Silicon):
+//
+//   |=========================================================|
+//   | Category | Subcategory    |     x86 |  x86_64 | Mac ARM |
+//   |---------------------------+---------+---------+---------|
+//   | Word size                 |  32 bit |  64 bit |  64 bit |
+//   | Page size                 |    4 Kb |    4 Kb |   16 Kb |
+//   |=========================================================|
+//   | Small    | Tiny           |    4/-w |      -w |       - |
+//   |          |                |       8 |    8/-w |       8 |
+//   |          |----------------+---------|---------|---------|
+//   |          | Quantum-spaced |      16 |      16 |      16 |
+//   |          |                |      32 |      32 |      32 |
+//   |          |                |      48 |      48 |      48 |
+//   |          |                |     ... |     ... |     ... |
+//   |          |                |     480 |     480 |     480 |
+//   |          |                |     496 |     496 |     496 |
+//   |          |----------------+---------|---------|---------|
+//   |          | Quantum-wide-  |     512 |     512 |     512 |
+//   |          | spaced         |     768 |     768 |     768 |
+//   |          |                |     ... |     ... |     ... |
+//   |          |                |    3584 |    3584 |    3584 |
+//   |          |                |    3840 |    3840 |    3840 |
+//   |          |----------------+---------|---------|---------|
+//   |          | Sub-page       |       - |       - |    4096 |
+//   |          |                |       - |       - |    8 kB |
+//   |=========================================================|
+//   | Large                     |    4 kB |    4 kB |       - |
+//   |                           |    8 kB |    8 kB |       - |
+//   |                           |   12 kB |   12 kB |       - |
+//   |                           |   16 kB |   16 kB |   16 kB |
+//   |                           |     ... |     ... |       - |
+//   |                           |   32 kB |   32 kB |   32 kB |
+//   |                           |     ... |     ... |     ... |
+//   |                           | 1008 kB | 1008 kB | 1008 kB |
+//   |                           | 1012 kB | 1012 kB |       - |
+//   |                           | 1016 kB | 1016 kB |       - |
+//   |                           | 1020 kB | 1020 kB |       - |
+//   |=========================================================|
+//   | Huge                      |    1 MB |    1 MB |    1 MB |
+//   |                           |    2 MB |    2 MB |    2 MB |
+//   |                           |    3 MB |    3 MB |    3 MB |
+//   |                           |     ... |     ... |     ... |
+//   |=========================================================|
+//
+// Legend:
+//   n:    Size class exists for this platform.
+//   n/-w: This size class doesn't exist on Windows (see kMinTinyClass).
+//   -:    This size class doesn't exist for this platform.
+//   ...:  Size classes follow a pattern here.
+//
+// NOTE: Due to Mozilla bug 691003, we cannot reserve less than one word for an
+// allocation on Linux or Mac.  So on 32-bit *nix, the smallest bucket size is
+// 4 bytes, and on 64-bit, the smallest bucket size is 8 bytes.
+//
+// A different mechanism is used for each category:
+//
+//   Small : Each size class is segregated into its own set of runs.  Each run
+//           maintains a bitmap of which regions are free/allocated.
+//
+//   Large : Each allocation is backed by a dedicated run.  Metadata are stored
+//           in the associated arena chunk header maps.
+//
+//   Huge : Each allocation is backed by a dedicated contiguous set of chunks.
+//          Metadata are stored in a separate red-black tree.
+//
+// *****************************************************************************
+
+#include "mozmemory_wrap.h"
+#include "mozjemalloc.h"
+#include "mozjemalloc_types.h"
+
+#include <cstring>
+#include <cerrno>
+#ifdef XP_WIN
+#  include <io.h>
+#  include <windows.h>
+#else
+#  include <sys/mman.h>
+#  include <unistd.h>
+#endif
+#ifdef XP_DARWIN
+#  include <libkern/OSAtomic.h>
+#  include <mach/mach_init.h>
+#  include <mach/vm_map.h>
+#endif
+
+#include "mozilla/Atomics.h"
+#include "mozilla/Alignment.h"
+#include "mozilla/ArrayUtils.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/CheckedInt.h"
+#include "mozilla/DoublyLinkedList.h"
+#include "mozilla/HelperMacros.h"
+#include "mozilla/Likely.h"
+#include "mozilla/MathAlgorithms.h"
+#include "mozilla/RandomNum.h"
+#include "mozilla/Sprintf.h"
+// Note: MozTaggedAnonymousMmap() could call an LD_PRELOADed mmap
+// instead of the one defined here; use only MozTagAnonymousMemory().
+#include "mozilla/TaggedAnonymousMemory.h"
+#include "mozilla/ThreadLocal.h"
+#include "mozilla/UniquePtr.h"
+#include "mozilla/Unused.h"
+#include "mozilla/XorShift128PlusRNG.h"
+#include "mozilla/fallible.h"
+#include "rb.h"
+#include "Mutex.h"
+#include "Utils.h"
+
+// For GetGeckoProcessType(), when it's used.
+#if defined(XP_WIN) && !defined(JS_STANDALONE)
+#  include "mozilla/ProcessType.h"
+#endif
+
+using namespace mozilla;
+
+// On Linux, we use madvise(MADV_DONTNEED) to release memory back to the
+// operating system.  If we release 1MB of live pages with MADV_DONTNEED, our
+// RSS will decrease by 1MB (almost) immediately.
+//
+// On Mac, we use madvise(MADV_FREE).  Unlike MADV_DONTNEED on Linux, MADV_FREE
+// on Mac doesn't cause the OS to release the specified pages immediately; the
+// OS keeps them in our process until the machine comes under memory pressure.
+//
+// It's therefore difficult to measure the process's RSS on Mac, since, in the
+// absence of memory pressure, the contribution from the heap to RSS will not
+// decrease due to our madvise calls.
+//
+// We therefore define MALLOC_DOUBLE_PURGE on Mac.  This causes jemalloc to
+// track which pages have been MADV_FREE'd.  You can then call
+// jemalloc_purge_freed_pages(), which will force the OS to release those
+// MADV_FREE'd pages, making the process's RSS reflect its true memory usage.
+//
+// The jemalloc_purge_freed_pages definition in memory/build/mozmemory.h needs
+// to be adjusted if MALLOC_DOUBLE_PURGE is ever enabled on Linux.
+
+#ifdef XP_DARWIN
+#  define MALLOC_DOUBLE_PURGE
+#endif
+
+#ifdef XP_WIN
+#  define MALLOC_DECOMMIT
+#endif
+
+// When MALLOC_STATIC_PAGESIZE is defined, the page size is fixed at
+// compile-time for better performance, as opposed to determined at
+// runtime. Some platforms can have different page sizes at runtime
+// depending on kernel configuration, so they are opted out by default.
+// Debug builds are opted out too, for test coverage.
+#ifndef MOZ_DEBUG
+#  if !defined(__ia64__) && !defined(__sparc__) && !defined(__mips__) &&       \
+      !defined(__aarch64__) && !defined(__powerpc__) && !defined(XP_MACOSX) && \
+      !defined(__loongarch__)
+#    define MALLOC_STATIC_PAGESIZE 1
+#  endif
+#endif
+
+#ifdef XP_WIN
+#  define STDERR_FILENO 2
+
+// Implement getenv without using malloc.
+static char mozillaMallocOptionsBuf[64];
+
+#  define getenv xgetenv
+static char* getenv(const char* name) {
+  if (GetEnvironmentVariableA(name, mozillaMallocOptionsBuf,
+                              sizeof(mozillaMallocOptionsBuf)) > 0) {
+    return mozillaMallocOptionsBuf;
+  }
+
+  return nullptr;
+}
+#endif
+
+#ifndef XP_WIN
+// Newer Linux systems support MADV_FREE, but we're not supporting
+// that properly. bug #1406304.
+#  if defined(XP_LINUX) && defined(MADV_FREE)
+#    undef MADV_FREE
+#  endif
+#  ifndef MADV_FREE
+#    define MADV_FREE MADV_DONTNEED
+#  endif
+#endif
+
+// Some tools, such as /dev/dsp wrappers, LD_PRELOAD libraries that
+// happen to override mmap() and call dlsym() from their overridden
+// mmap(). The problem is that dlsym() calls malloc(), and this ends
+// up in a dead lock in jemalloc.
+// On these systems, we prefer to directly use the system call.
+// We do that for Linux systems and kfreebsd with GNU userland.
+// Note sanity checks are not done (alignment of offset, ...) because
+// the uses of mmap are pretty limited, in jemalloc.
+//
+// On Alpha, glibc has a bug that prevents syscall() to work for system
+// calls with 6 arguments.
+#if (defined(XP_LINUX) && !defined(__alpha__)) || \
+    (defined(__FreeBSD_kernel__) && defined(__GLIBC__))
+#  include <sys/syscall.h>
+#  if defined(SYS_mmap) || defined(SYS_mmap2)
+static inline void* _mmap(void* addr, size_t length, int prot, int flags,
+                          int fd, off_t offset) {
+// S390 only passes one argument to the mmap system call, which is a
+// pointer to a structure containing the arguments.
+#    ifdef __s390__
+  struct {
+    void* addr;
+    size_t length;
+    long prot;
+    long flags;
+    long fd;
+    off_t offset;
+  } args = {addr, length, prot, flags, fd, offset};
+  return (void*)syscall(SYS_mmap, &args);
+#    else
+#      if defined(ANDROID) && defined(__aarch64__) && defined(SYS_mmap2)
+// Android NDK defines SYS_mmap2 for AArch64 despite it not supporting mmap2.
+#        undef SYS_mmap2
+#      endif
+#      ifdef SYS_mmap2
+  return (void*)syscall(SYS_mmap2, addr, length, prot, flags, fd, offset >> 12);
+#      else
+  return (void*)syscall(SYS_mmap, addr, length, prot, flags, fd, offset);
+#      endif
+#    endif
+}
+#    define mmap _mmap
+#    define munmap(a, l) syscall(SYS_munmap, a, l)
+#  endif
+#endif
+
+// ***************************************************************************
+// Structures for chunk headers for chunks used for non-huge allocations.
+
+struct arena_t;
+
+// Each element of the chunk map corresponds to one page within the chunk.
+struct arena_chunk_map_t {
+  // Linkage for run trees.  There are two disjoint uses:
+  //
+  // 1) arena_t's tree or available runs.
+  // 2) arena_run_t conceptually uses this linkage for in-use non-full
+  //    runs, rather than directly embedding linkage.
+  RedBlackTreeNode<arena_chunk_map_t> link;
+
+  // Run address (or size) and various flags are stored together.  The bit
+  // layout looks like (assuming 32-bit system):
+  //
+  //   ???????? ???????? ????---- -mckdzla
+  //
+  // ? : Unallocated: Run address for first/last pages, unset for internal
+  //                  pages.
+  //     Small: Run address.
+  //     Large: Run size for first page, unset for trailing pages.
+  // - : Unused.
+  // m : MADV_FREE/MADV_DONTNEED'ed?
+  // c : decommitted?
+  // k : key?
+  // d : dirty?
+  // z : zeroed?
+  // l : large?
+  // a : allocated?
+  //
+  // Following are example bit patterns for the three types of runs.
+  //
+  // r : run address
+  // s : run size
+  // x : don't care
+  // - : 0
+  // [cdzla] : bit set
+  //
+  //   Unallocated:
+  //     ssssssss ssssssss ssss---- --c-----
+  //     xxxxxxxx xxxxxxxx xxxx---- ----d---
+  //     ssssssss ssssssss ssss---- -----z--
+  //
+  //   Small:
+  //     rrrrrrrr rrrrrrrr rrrr---- -------a
+  //     rrrrrrrr rrrrrrrr rrrr---- -------a
+  //     rrrrrrrr rrrrrrrr rrrr---- -------a
+  //
+  //   Large:
+  //     ssssssss ssssssss ssss---- ------la
+  //     -------- -------- -------- ------la
+  //     -------- -------- -------- ------la
+  size_t bits;
+
+// Note that CHUNK_MAP_DECOMMITTED's meaning varies depending on whether
+// MALLOC_DECOMMIT and MALLOC_DOUBLE_PURGE are defined.
+//
+// If MALLOC_DECOMMIT is defined, a page which is CHUNK_MAP_DECOMMITTED must be
+// re-committed with pages_commit() before it may be touched.  If
+// MALLOC_DECOMMIT is defined, MALLOC_DOUBLE_PURGE may not be defined.
+//
+// If neither MALLOC_DECOMMIT nor MALLOC_DOUBLE_PURGE is defined, pages which
+// are madvised (with either MADV_DONTNEED or MADV_FREE) are marked with
+// CHUNK_MAP_MADVISED.
+//
+// Otherwise, if MALLOC_DECOMMIT is not defined and MALLOC_DOUBLE_PURGE is
+// defined, then a page which is madvised is marked as CHUNK_MAP_MADVISED.
+// When it's finally freed with jemalloc_purge_freed_pages, the page is marked
+// as CHUNK_MAP_DECOMMITTED.
+#define CHUNK_MAP_MADVISED ((size_t)0x40U)
+#define CHUNK_MAP_DECOMMITTED ((size_t)0x20U)
+#define CHUNK_MAP_MADVISED_OR_DECOMMITTED \
+  (CHUNK_MAP_MADVISED | CHUNK_MAP_DECOMMITTED)
+#define CHUNK_MAP_KEY ((size_t)0x10U)
+#define CHUNK_MAP_DIRTY ((size_t)0x08U)
+#define CHUNK_MAP_ZEROED ((size_t)0x04U)
+#define CHUNK_MAP_LARGE ((size_t)0x02U)
+#define CHUNK_MAP_ALLOCATED ((size_t)0x01U)
+};
+
+// Arena chunk header.
+struct arena_chunk_t {
+  // Arena that owns the chunk.
+  arena_t* arena;
+
+  // Linkage for the arena's tree of dirty chunks.
+  RedBlackTreeNode<arena_chunk_t> link_dirty;
+
+#ifdef MALLOC_DOUBLE_PURGE
+  // If we're double-purging, we maintain a linked list of chunks which
+  // have pages which have been madvise(MADV_FREE)'d but not explicitly
+  // purged.
+  //
+  // We're currently lazy and don't remove a chunk from this list when
+  // all its madvised pages are recommitted.
+  DoublyLinkedListElement<arena_chunk_t> chunks_madvised_elem;
+#endif
+
+  // Number of dirty pages.
+  size_t ndirty;
+
+  // Map of pages within chunk that keeps track of free/large/small.
+  arena_chunk_map_t map[1];  // Dynamically sized.
+};
+
+// ***************************************************************************
+// Constants defining allocator size classes and behavior.
+
+// Maximum size of L1 cache line.  This is used to avoid cache line aliasing,
+// so over-estimates are okay (up to a point), but under-estimates will
+// negatively affect performance.
+static const size_t kCacheLineSize = 64;
+
+// Our size classes are inclusive ranges of memory sizes.  By describing the
+// minimums and how memory is allocated in each range the maximums can be
+// calculated.
+
+// Smallest size class to support.  On Windows the smallest allocation size
+// must be 8 bytes on 32-bit, 16 bytes on 64-bit.  On Linux and Mac, even
+// malloc(1) must reserve a word's worth of memory (see Mozilla bug 691003).
+#ifdef XP_WIN
+static const size_t kMinTinyClass = sizeof(void*) * 2;
+#else
+static const size_t kMinTinyClass = sizeof(void*);
+#endif
+
+// Maximum tiny size class.
+static const size_t kMaxTinyClass = 8;
+
+// Smallest quantum-spaced size classes. It could actually also be labelled a
+// tiny allocation, and is spaced as such from the largest tiny size class.
+// Tiny classes being powers of 2, this is twice as large as the largest of
+// them.
+static const size_t kMinQuantumClass = kMaxTinyClass * 2;
+static const size_t kMinQuantumWideClass = 512;
+static const size_t kMinSubPageClass = 4_KiB;
+
+// Amount (quantum) separating quantum-spaced size classes.
+static const size_t kQuantum = 16;
+static const size_t kQuantumMask = kQuantum - 1;
+static const size_t kQuantumWide = 256;
+static const size_t kQuantumWideMask = kQuantumWide - 1;
+
+static const size_t kMaxQuantumClass = kMinQuantumWideClass - kQuantum;
+static const size_t kMaxQuantumWideClass = kMinSubPageClass - kQuantumWide;
+
+// We can optimise some divisions to shifts if these are powers of two.
+static_assert(mozilla::IsPowerOfTwo(kQuantum),
+              "kQuantum is not a power of two");
+static_assert(mozilla::IsPowerOfTwo(kQuantumWide),
+              "kQuantumWide is not a power of two");
+
+static_assert(kMaxQuantumClass % kQuantum == 0,
+              "kMaxQuantumClass is not a multiple of kQuantum");
+static_assert(kMaxQuantumWideClass % kQuantumWide == 0,
+              "kMaxQuantumWideClass is not a multiple of kQuantumWide");
+static_assert(kQuantum < kQuantumWide,
+              "kQuantum must be smaller than kQuantumWide");
+static_assert(mozilla::IsPowerOfTwo(kMinSubPageClass),
+              "kMinSubPageClass is not a power of two");
+
+// Number of (2^n)-spaced tiny classes.
+static const size_t kNumTinyClasses =
+    LOG2(kMaxTinyClass) - LOG2(kMinTinyClass) + 1;
+
+// Number of quantum-spaced classes.  We add kQuantum(Max) before subtracting to
+// avoid underflow when a class is empty (Max<Min).
+static const size_t kNumQuantumClasses =
+    (kMaxQuantumClass + kQuantum - kMinQuantumClass) / kQuantum;
+static const size_t kNumQuantumWideClasses =
+    (kMaxQuantumWideClass + kQuantumWide - kMinQuantumWideClass) / kQuantumWide;
+
+// Size and alignment of memory chunks that are allocated by the OS's virtual
+// memory system.
+static const size_t kChunkSize = 1_MiB;
+static const size_t kChunkSizeMask = kChunkSize - 1;
+
+#ifdef MALLOC_STATIC_PAGESIZE
+// VM page size. It must divide the runtime CPU page size or the code
+// will abort.
+// Platform specific page size conditions copied from js/public/HeapAPI.h
+#  if defined(__powerpc64__)
+static const size_t gPageSize = 64_KiB;
+#  elif defined(__loongarch64)
+static const size_t gPageSize = 16_KiB;
+#  else
+static const size_t gPageSize = 4_KiB;
+#  endif
+static const size_t gRealPageSize = gPageSize;
+
+#else
+// When MALLOC_OPTIONS contains one or several `P`s, the page size used
+// across the allocator is multiplied by 2 for each `P`, but we also keep
+// the real page size for code paths that need it. gPageSize is thus a
+// power of two greater or equal to gRealPageSize.
+static size_t gRealPageSize;
+static size_t gPageSize;
+#endif
+
+#ifdef MALLOC_STATIC_PAGESIZE
+#  define DECLARE_GLOBAL(type, name)
+#  define DEFINE_GLOBALS
+#  define END_GLOBALS
+#  define DEFINE_GLOBAL(type) static const type
+#  define GLOBAL_LOG2 LOG2
+#  define GLOBAL_ASSERT_HELPER1(x) static_assert(x, #x)
+#  define GLOBAL_ASSERT_HELPER2(x, y) static_assert(x, y)
+#  define GLOBAL_ASSERT(...)                                               \
+    MACRO_CALL(                                                            \
+        MOZ_PASTE_PREFIX_AND_ARG_COUNT(GLOBAL_ASSERT_HELPER, __VA_ARGS__), \
+        (__VA_ARGS__))
+#  define GLOBAL_CONSTEXPR constexpr
+#else
+#  define DECLARE_GLOBAL(type, name) static type name;
+#  define DEFINE_GLOBALS static void DefineGlobals() {
+#  define END_GLOBALS }
+#  define DEFINE_GLOBAL(type)
+#  define GLOBAL_LOG2 FloorLog2
+#  define GLOBAL_ASSERT MOZ_RELEASE_ASSERT
+#  define GLOBAL_CONSTEXPR
+#endif
+
+DECLARE_GLOBAL(size_t, gMaxSubPageClass)
+DECLARE_GLOBAL(uint8_t, gNumSubPageClasses)
+DECLARE_GLOBAL(uint8_t, gPageSize2Pow)
+DECLARE_GLOBAL(size_t, gPageSizeMask)
+DECLARE_GLOBAL(size_t, gChunkNumPages)
+DECLARE_GLOBAL(size_t, gChunkHeaderNumPages)
+DECLARE_GLOBAL(size_t, gMaxLargeClass)
+
+DEFINE_GLOBALS
+
+// Largest sub-page size class, or zero if there are none
+DEFINE_GLOBAL(size_t)
+gMaxSubPageClass = gPageSize / 2 >= kMinSubPageClass ? gPageSize / 2 : 0;
+
+// Max size class for bins.
+#define gMaxBinClass \
+  (gMaxSubPageClass ? gMaxSubPageClass : kMaxQuantumWideClass)
+
+// Number of sub-page bins.
+DEFINE_GLOBAL(uint8_t)
+gNumSubPageClasses = []() GLOBAL_CONSTEXPR -> uint8_t {
+  if GLOBAL_CONSTEXPR (gMaxSubPageClass != 0) {
+    return FloorLog2(gMaxSubPageClass) - LOG2(kMinSubPageClass) + 1;
+  }
+  return 0;
+}();
+
+DEFINE_GLOBAL(uint8_t) gPageSize2Pow = GLOBAL_LOG2(gPageSize);
+DEFINE_GLOBAL(size_t) gPageSizeMask = gPageSize - 1;
+
+// Number of pages in a chunk.
+DEFINE_GLOBAL(size_t) gChunkNumPages = kChunkSize >> gPageSize2Pow;
+
+// Number of pages necessary for a chunk header plus a guard page.
+DEFINE_GLOBAL(size_t)
+gChunkHeaderNumPages =
+    1 + (((sizeof(arena_chunk_t) +
+           sizeof(arena_chunk_map_t) * (gChunkNumPages - 1) + gPageSizeMask) &
+          ~gPageSizeMask) >>
+         gPageSize2Pow);
+
+// One chunk, minus the header, minus a guard page
+DEFINE_GLOBAL(size_t)
+gMaxLargeClass =
+    kChunkSize - gPageSize - (gChunkHeaderNumPages << gPageSize2Pow);
+
+// Various sanity checks that regard configuration.
+GLOBAL_ASSERT(1ULL << gPageSize2Pow == gPageSize,
+              "Page size is not a power of two");
+GLOBAL_ASSERT(kQuantum >= sizeof(void*));
+GLOBAL_ASSERT(kQuantum <= kQuantumWide);
+GLOBAL_ASSERT(!kNumQuantumWideClasses ||
+              kQuantumWide <= (kMinSubPageClass - kMaxQuantumClass));
+
+GLOBAL_ASSERT(kQuantumWide <= kMaxQuantumClass);
+
+GLOBAL_ASSERT(gMaxSubPageClass >= kMinSubPageClass || gMaxSubPageClass == 0);
+GLOBAL_ASSERT(gMaxLargeClass >= gMaxSubPageClass);
+GLOBAL_ASSERT(kChunkSize >= gPageSize);
+GLOBAL_ASSERT(kQuantum * 4 <= kChunkSize);
+
+END_GLOBALS
+
+// Recycle at most 128 MiB of chunks. This means we retain at most
+// 6.25% of the process address space on a 32-bit OS for later use.
+static const size_t gRecycleLimit = 128_MiB;
+
+// The current amount of recycled bytes, updated atomically.
+static Atomic<size_t, ReleaseAcquire> gRecycledSize;
+
+// Maximum number of dirty pages per arena.
+#define DIRTY_MAX_DEFAULT (1U << 8)
+
+static size_t opt_dirty_max = DIRTY_MAX_DEFAULT;
+
+// Return the smallest chunk multiple that is >= s.
+#define CHUNK_CEILING(s) (((s) + kChunkSizeMask) & ~kChunkSizeMask)
+
+// Return the smallest cacheline multiple that is >= s.
+#define CACHELINE_CEILING(s) \
+  (((s) + (kCacheLineSize - 1)) & ~(kCacheLineSize - 1))
+
+// Return the smallest quantum multiple that is >= a.
+#define QUANTUM_CEILING(a) (((a) + (kQuantumMask)) & ~(kQuantumMask))
+#define QUANTUM_WIDE_CEILING(a) \
+  (((a) + (kQuantumWideMask)) & ~(kQuantumWideMask))
+
+// Return the smallest sub page-size  that is >= a.
+#define SUBPAGE_CEILING(a) (RoundUpPow2(a))
+
+// Return the smallest pagesize multiple that is >= s.
+#define PAGE_CEILING(s) (((s) + gPageSizeMask) & ~gPageSizeMask)
+
+// Number of all the small-allocated classes
+#define NUM_SMALL_CLASSES                                          \
+  (kNumTinyClasses + kNumQuantumClasses + kNumQuantumWideClasses + \
+   gNumSubPageClasses)
+
+// ***************************************************************************
+// MALLOC_DECOMMIT and MALLOC_DOUBLE_PURGE are mutually exclusive.
+#if defined(MALLOC_DECOMMIT) && defined(MALLOC_DOUBLE_PURGE)
+#  error MALLOC_DECOMMIT and MALLOC_DOUBLE_PURGE are mutually exclusive.
+#endif
+
+static void* base_alloc(size_t aSize);
+
+// Set to true once the allocator has been initialized.
+#if defined(_MSC_VER) && !defined(__clang__)
+// MSVC may create a static initializer for an Atomic<bool>, which may actually
+// run after `malloc_init` has been called once, which triggers multiple
+// initializations.
+// We work around the problem by not using an Atomic<bool> at all. There is a
+// theoretical problem with using `malloc_initialized` non-atomically, but
+// practically, this is only true if `malloc_init` is never called before
+// threads are created.
+static bool malloc_initialized;
+#else
+static Atomic<bool, SequentiallyConsistent> malloc_initialized;
+#endif
+
+static StaticMutex gInitLock MOZ_UNANNOTATED = {STATIC_MUTEX_INIT};
+
+// ***************************************************************************
+// Statistics data structures.
+
+struct arena_stats_t {
+  // Number of bytes currently mapped.
+  size_t mapped;
+
+  // Current number of committed pages.
+  size_t committed;
+
+  // Per-size-category statistics.
+  size_t allocated_small;
+
+  size_t allocated_large;
+};
+
+// ***************************************************************************
+// Extent data structures.
+
+enum ChunkType {
+  UNKNOWN_CHUNK,
+  ZEROED_CHUNK,    // chunk only contains zeroes.
+  ARENA_CHUNK,     // used to back arena runs created by arena_t::AllocRun.
+  HUGE_CHUNK,      // used to back huge allocations (e.g. arena_t::MallocHuge).
+  RECYCLED_CHUNK,  // chunk has been stored for future use by chunk_recycle.
+};
+
+// Tree of extents.
+struct extent_node_t {
+  union {
+    // Linkage for the size/address-ordered tree for chunk recycling.
+    RedBlackTreeNode<extent_node_t> mLinkBySize;
+    // Arena id for huge allocations. It's meant to match mArena->mId,
+    // which only holds true when the arena hasn't been disposed of.
+    arena_id_t mArenaId;
+  };
+
+  // Linkage for the address-ordered tree.
+  RedBlackTreeNode<extent_node_t> mLinkByAddr;
+
+  // Pointer to the extent that this tree node is responsible for.
+  void* mAddr;
+
+  // Total region size.
+  size_t mSize;
+
+  union {
+    // What type of chunk is there; used for chunk recycling.
+    ChunkType mChunkType;
+
+    // A pointer to the associated arena, for huge allocations.
+    arena_t* mArena;
+  };
+};
+
+struct ExtentTreeSzTrait {
+  static RedBlackTreeNode<extent_node_t>& GetTreeNode(extent_node_t* aThis) {
+    return aThis->mLinkBySize;
+  }
+
+  static inline Order Compare(extent_node_t* aNode, extent_node_t* aOther) {
+    Order ret = CompareInt(aNode->mSize, aOther->mSize);
+    return (ret != Order::eEqual) ? ret
+                                  : CompareAddr(aNode->mAddr, aOther->mAddr);
+  }
+};
+
+struct ExtentTreeTrait {
+  static RedBlackTreeNode<extent_node_t>& GetTreeNode(extent_node_t* aThis) {
+    return aThis->mLinkByAddr;
+  }
+
+  static inline Order Compare(extent_node_t* aNode, extent_node_t* aOther) {
+    return CompareAddr(aNode->mAddr, aOther->mAddr);
+  }
+};
+
+struct ExtentTreeBoundsTrait : public ExtentTreeTrait {
+  static inline Order Compare(extent_node_t* aKey, extent_node_t* aNode) {
+    uintptr_t key_addr = reinterpret_cast<uintptr_t>(aKey->mAddr);
+    uintptr_t node_addr = reinterpret_cast<uintptr_t>(aNode->mAddr);
+    size_t node_size = aNode->mSize;
+
+    // Is aKey within aNode?
+    if (node_addr <= key_addr && key_addr < node_addr + node_size) {
+      return Order::eEqual;
+    }
+
+    return CompareAddr(aKey->mAddr, aNode->mAddr);
+  }
+};
+
+// Describe size classes to which allocations are rounded up to.
+// TODO: add large and huge types when the arena allocation code
+// changes in a way that allows it to be beneficial.
+class SizeClass {
+ public:
+  enum ClassType {
+    Tiny,
+    Quantum,
+    QuantumWide,
+    SubPage,
+    Large,
+  };
+
+  explicit inline SizeClass(size_t aSize) {
+    if (aSize <= kMaxTinyClass) {
+      mType = Tiny;
+      mSize = std::max(RoundUpPow2(aSize), kMinTinyClass);
+    } else if (aSize <= kMaxQuantumClass) {
+      mType = Quantum;
+      mSize = QUANTUM_CEILING(aSize);
+    } else if (aSize <= kMaxQuantumWideClass) {
+      mType = QuantumWide;
+      mSize = QUANTUM_WIDE_CEILING(aSize);
+    } else if (aSize <= gMaxSubPageClass) {
+      mType = SubPage;
+      mSize = SUBPAGE_CEILING(aSize);
+    } else if (aSize <= gMaxLargeClass) {
+      mType = Large;
+      mSize = PAGE_CEILING(aSize);
+    } else {
+      MOZ_MAKE_COMPILER_ASSUME_IS_UNREACHABLE("Invalid size");
+    }
+  }
+
+  SizeClass& operator=(const SizeClass& aOther) = default;
+
+  bool operator==(const SizeClass& aOther) { return aOther.mSize == mSize; }
+
+  size_t Size() { return mSize; }
+
+  ClassType Type() { return mType; }
+
+  SizeClass Next() { return SizeClass(mSize + 1); }
+
+ private:
+  ClassType mType;
+  size_t mSize;
+};
+
+// Fast division
+//
+// During deallocation we want to divide by the size class.  This class
+// provides a routine and sets up a constant as follows.
+//
+// To divide by a number D that is not a power of two we multiply by (2^17 /
+// D) and then right shift by 17 positions.
+//
+//   X / D
+//
+// becomes
+//
+//   (X * m) >> p
+//
+// Where m is calculated during the FastDivisor constructor similarly to:
+//
+//   m = 2^p / D
+//
+template <typename T>
+class FastDivisor {
+ private:
+  // The shift amount (p) is chosen to minimise the size of m while
+  // working for divisors up to 65536 in steps of 16.  I arrived at 17
+  // experimentally.  I wanted a low number to minimise the range of m
+  // so it can fit in a uint16_t, 16 didn't work but 17 worked perfectly.
+  //
+  // We'd need to increase this if we allocated memory on smaller boundaries
+  // than 16.
+  static const unsigned p = 17;
+
+  // We can fit the inverted divisor in 16 bits, but we template it here for
+  // convenience.
+  T m;
+
+ public:
+  // Needed so mBins can be constructed.
+  FastDivisor() : m(0) {}
+
+  FastDivisor(unsigned div, unsigned max) {
+    MOZ_ASSERT(div <= max);
+
+    // divide_inv_shift is large enough.
+    MOZ_ASSERT((1U << p) >= div);
+
+    // The calculation here for m is formula 26 from Section
+    // 10-9 "Unsigned Division by Divisors >= 1" in
+    // Henry S. Warren, Jr.'s Hacker's Delight, 2nd Ed.
+    unsigned m_ = ((1U << p) + div - 1 - (((1U << p) - 1) % div)) / div;
+
+    // Make sure that max * m does not overflow.
+    MOZ_DIAGNOSTIC_ASSERT(max < UINT_MAX / m_);
+
+    MOZ_ASSERT(m_ <= std::numeric_limits<T>::max());
+    m = static_cast<T>(m_);
+
+    // Initialisation made m non-zero.
+    MOZ_ASSERT(m);
+
+    // Test that all the divisions in the range we expected would work.
+#ifdef MOZ_DEBUG
+    for (unsigned num = 0; num < max; num += div) {
+      MOZ_ASSERT(num / div == divide(num));
+    }
+#endif
+  }
+
+  // Note that this always occurs in uint32_t regardless of m's type.  If m is
+  // a uint16_t it will be zero-extended before the multiplication.  We also use
+  // uint32_t rather than something that could possibly be larger because it is
+  // most-likely the cheapest multiplication.
+  inline uint32_t divide(uint32_t num) const {
+    // Check that m was initialised.
+    MOZ_ASSERT(m);
+    return (num * m) >> p;
+  }
+};
+
+template <typename T>
+unsigned inline operator/(unsigned num, FastDivisor<T> divisor) {
+  return divisor.divide(num);
+}
+
+// ***************************************************************************
+// Radix tree data structures.
+//
+// The number of bits passed to the template is the number of significant bits
+// in an address to do a radix lookup with.
+//
+// An address is looked up by splitting it in kBitsPerLevel bit chunks, except
+// the most significant bits, where the bit chunk is kBitsAtLevel1 which can be
+// different if Bits is not a multiple of kBitsPerLevel.
+//
+// With e.g. sizeof(void*)=4, Bits=16 and kBitsPerLevel=8, an address is split
+// like the following:
+// 0x12345678 -> mRoot[0x12][0x34]
+template <size_t Bits>
+class AddressRadixTree {
+// Size of each radix tree node (as a power of 2).
+// This impacts tree depth.
+#ifdef HAVE_64BIT_BUILD
+  static const size_t kNodeSize = kCacheLineSize;
+#else
+  static const size_t kNodeSize = 16_KiB;
+#endif
+  static const size_t kBitsPerLevel = LOG2(kNodeSize) - LOG2(sizeof(void*));
+  static const size_t kBitsAtLevel1 =
+      (Bits % kBitsPerLevel) ? Bits % kBitsPerLevel : kBitsPerLevel;
+  static const size_t kHeight = (Bits + kBitsPerLevel - 1) / kBitsPerLevel;
+  static_assert(kBitsAtLevel1 + (kHeight - 1) * kBitsPerLevel == Bits,
+                "AddressRadixTree parameters don't work out");
+
+  Mutex mLock MOZ_UNANNOTATED;
+  void** mRoot;
+
+ public:
+  bool Init();
+
+  inline void* Get(void* aAddr);
+
+  // Returns whether the value was properly set.
+  inline bool Set(void* aAddr, void* aValue);
+
+  inline bool Unset(void* aAddr) { return Set(aAddr, nullptr); }
+
+ private:
+  inline void** GetSlot(void* aAddr, bool aCreate = false);
+};
+
+// ***************************************************************************
+// Arena data structures.
+
+struct arena_bin_t;
+
+struct ArenaChunkMapLink {
+  static RedBlackTreeNode<arena_chunk_map_t>& GetTreeNode(
+      arena_chunk_map_t* aThis) {
+    return aThis->link;
+  }
+};
+
+struct ArenaRunTreeTrait : public ArenaChunkMapLink {
+  static inline Order Compare(arena_chunk_map_t* aNode,
+                              arena_chunk_map_t* aOther) {
+    MOZ_ASSERT(aNode);
+    MOZ_ASSERT(aOther);
+    return CompareAddr(aNode, aOther);
+  }
+};
+
+struct ArenaAvailTreeTrait : public ArenaChunkMapLink {
+  static inline Order Compare(arena_chunk_map_t* aNode,
+                              arena_chunk_map_t* aOther) {
+    size_t size1 = aNode->bits & ~gPageSizeMask;
+    size_t size2 = aOther->bits & ~gPageSizeMask;
+    Order ret = CompareInt(size1, size2);
+    return (ret != Order::eEqual)
+               ? ret
+               : CompareAddr((aNode->bits & CHUNK_MAP_KEY) ? nullptr : aNode,
+                             aOther);
+  }
+};
+
+struct ArenaDirtyChunkTrait {
+  static RedBlackTreeNode<arena_chunk_t>& GetTreeNode(arena_chunk_t* aThis) {
+    return aThis->link_dirty;
+  }
+
+  static inline Order Compare(arena_chunk_t* aNode, arena_chunk_t* aOther) {
+    MOZ_ASSERT(aNode);
+    MOZ_ASSERT(aOther);
+    return CompareAddr(aNode, aOther);
+  }
+};
+
+#ifdef MALLOC_DOUBLE_PURGE
+namespace mozilla {
+
+template <>
+struct GetDoublyLinkedListElement<arena_chunk_t> {
+  static DoublyLinkedListElement<arena_chunk_t>& Get(arena_chunk_t* aThis) {
+    return aThis->chunks_madvised_elem;
+  }
+};
+}  // namespace mozilla
+#endif
+
+struct arena_run_t {
+#if defined(MOZ_DIAGNOSTIC_ASSERT_ENABLED)
+  uint32_t mMagic;
+#  define ARENA_RUN_MAGIC 0x384adf93
+
+  // On 64-bit platforms, having the arena_bin_t pointer following
+  // the mMagic field means there's padding between both fields, making
+  // the run header larger than necessary.
+  // But when MOZ_DIAGNOSTIC_ASSERT_ENABLED is not set, starting the
+  // header with this field followed by the arena_bin_t pointer yields
+  // the same padding. We do want the mMagic field to appear first, so
+  // depending whether MOZ_DIAGNOSTIC_ASSERT_ENABLED is set or not, we
+  // move some field to avoid padding.
+
+  // Number of free regions in run.
+  unsigned mNumFree;
+#endif
+
+  // Bin this run is associated with.
+  arena_bin_t* mBin;
+
+  // Index of first element that might have a free region.
+  unsigned mRegionsMinElement;
+
+#if !defined(MOZ_DIAGNOSTIC_ASSERT_ENABLED)
+  // Number of free regions in run.
+  unsigned mNumFree;
+#endif
+
+  // Bitmask of in-use regions (0: in use, 1: free).
+  unsigned mRegionsMask[1];  // Dynamically sized.
+};
+
+struct arena_bin_t {
+  // Current run being used to service allocations of this bin's size
+  // class.
+  arena_run_t* mCurrentRun;
+
+  // Tree of non-full runs.  This tree is used when looking for an
+  // existing run when mCurrentRun is no longer usable.  We choose the
+  // non-full run that is lowest in memory; this policy tends to keep
+  // objects packed well, and it can also help reduce the number of
+  // almost-empty chunks.
+  RedBlackTree<arena_chunk_map_t, ArenaRunTreeTrait> mNonFullRuns;
+
+  // Bin's size class.
+  size_t mSizeClass;
+
+  // Total number of regions in a run for this bin's size class.
+  uint32_t mRunNumRegions;
+
+  // Number of elements in a run's mRegionsMask for this bin's size class.
+  uint32_t mRunNumRegionsMask;
+
+  // Offset of first region in a run for this bin's size class.
+  uint32_t mRunFirstRegionOffset;
+
+  // Current number of runs in this bin, full or otherwise.
+  uint32_t mNumRuns;
+
+  // A constant for fast division by size class.  This value is 16 bits wide so
+  // it is placed last.
+  FastDivisor<uint16_t> mSizeDivisor;
+
+  // Total number of pages in a run for this bin's size class.
+  uint8_t mRunSizePages;
+
+  // Amount of overhead runs are allowed to have.
+  static constexpr double kRunOverhead = 1.6_percent;
+  static constexpr double kRunRelaxedOverhead = 2.4_percent;
+
+  // Initialize a bin for the given size class.
+  // The generated run sizes, for a page size of 4 KiB, are:
+  //   size|run       size|run       size|run       size|run
+  //  class|size     class|size     class|size     class|size
+  //     4   4 KiB      8   4 KiB     16   4 KiB     32   4 KiB
+  //    48   4 KiB     64   4 KiB     80   4 KiB     96   4 KiB
+  //   112   4 KiB    128   8 KiB    144   4 KiB    160   8 KiB
+  //   176   4 KiB    192   4 KiB    208   8 KiB    224   4 KiB
+  //   240   8 KiB    256  16 KiB    272   8 KiB    288   4 KiB
+  //   304  12 KiB    320  12 KiB    336   4 KiB    352   8 KiB
+  //   368   4 KiB    384   8 KiB    400  20 KiB    416  16 KiB
+  //   432  12 KiB    448   4 KiB    464  16 KiB    480   8 KiB
+  //   496  20 KiB    512  32 KiB    768  16 KiB   1024  64 KiB
+  //  1280  24 KiB   1536  32 KiB   1792  16 KiB   2048 128 KiB
+  //  2304  16 KiB   2560  48 KiB   2816  36 KiB   3072  64 KiB
+  //  3328  36 KiB   3584  32 KiB   3840  64 KiB
+  inline void Init(SizeClass aSizeClass);
+};
+
+// We try to keep the above structure aligned with common cache lines sizes,
+// often that's 64 bytes on x86 and ARM, we don't make assumptions for other
+// architectures.
+#if defined(__x86_64__) || defined(__aarch64__)
+// On 64bit platforms this structure is often 48 bytes
+// long, which means every other array element will be properly aligned.
+static_assert(sizeof(arena_bin_t) == 48);
+#elif defined(__x86__) || defined(__arm__)
+static_assert(sizeof(arena_bin_t) == 32);
+#endif
+
+struct arena_t {
+#if defined(MOZ_DIAGNOSTIC_ASSERT_ENABLED)
+  uint32_t mMagic;
+#  define ARENA_MAGIC 0x947d3d24
+#endif
+
+  // Linkage for the tree of arenas by id.
+  RedBlackTreeNode<arena_t> mLink;
+
+  // Arena id, that we keep away from the beginning of the struct so that
+  // free list pointers in TypedBaseAlloc<arena_t> don't overflow in it,
+  // and it keeps the value it had after the destructor.
+  arena_id_t mId;
+
+  // All operations on this arena require that lock be locked.
+  Mutex mLock MOZ_UNANNOTATED;
+
+  arena_stats_t mStats;
+
+ private:
+  // Tree of dirty-page-containing chunks this arena manages.
+  RedBlackTree<arena_chunk_t, ArenaDirtyChunkTrait> mChunksDirty;
+
+#ifdef MALLOC_DOUBLE_PURGE
+  // Head of a linked list of MADV_FREE'd-page-containing chunks this
+  // arena manages.
+  DoublyLinkedList<arena_chunk_t> mChunksMAdvised;
+#endif
+
+  // In order to avoid rapid chunk allocation/deallocation when an arena
+  // oscillates right on the cusp of needing a new chunk, cache the most
+  // recently freed chunk.  The spare is left in the arena's chunk trees
+  // until it is deleted.
+  //
+  // There is one spare chunk per arena, rather than one spare total, in
+  // order to avoid interactions between multiple threads that could make
+  // a single spare inadequate.
+  arena_chunk_t* mSpare;
+
+  // A per-arena opt-in to randomize the offset of small allocations
+  bool mRandomizeSmallAllocations;
+
+  // Whether this is a private arena. Multiple public arenas are just a
+  // performance optimization and not a safety feature.
+  //
+  // Since, for example, we don't want thread-local arenas to grow too much, we
+  // use the default arena for bigger allocations. We use this member to allow
+  // realloc() to switch out of our arena if needed (which is not allowed for
+  // private arenas for security).
+  bool mIsPrivate;
+
+  // A pseudorandom number generator. Initially null, it gets initialized
+  // on first use to avoid recursive malloc initialization (e.g. on OSX
+  // arc4random allocates memory).
+  mozilla::non_crypto::XorShift128PlusRNG* mPRNG;
+
+ public:
+  // Current count of pages within unused runs that are potentially
+  // dirty, and for which madvise(... MADV_FREE) has not been called.  By
+  // tracking this, we can institute a limit on how much dirty unused
+  // memory is mapped for each arena.
+  size_t mNumDirty;
+
+  // Maximum value allowed for mNumDirty.
+  size_t mMaxDirty;
+
+ private:
+  // Size/address-ordered tree of this arena's available runs.  This tree
+  // is used for first-best-fit run allocation.
+  RedBlackTree<arena_chunk_map_t, ArenaAvailTreeTrait> mRunsAvail;
+
+ public:
+  // mBins is used to store rings of free regions of the following sizes,
+  // assuming a 16-byte quantum, 4kB pagesize, and default MALLOC_OPTIONS.
+  //
+  //   mBins[i] | size |
+  //   --------+------+
+  //        0  |    2 |
+  //        1  |    4 |
+  //        2  |    8 |
+  //   --------+------+
+  //        3  |   16 |
+  //        4  |   32 |
+  //        5  |   48 |
+  //        6  |   64 |
+  //           :      :
+  //           :      :
+  //       33  |  496 |
+  //       34  |  512 |
+  //   --------+------+
+  //       35  |  768 |
+  //       36  | 1024 |
+  //           :      :
+  //           :      :
+  //       46  | 3584 |
+  //       47  | 3840 |
+  //   --------+------+
+  arena_bin_t mBins[1];  // Dynamically sized.
+
+  explicit arena_t(arena_params_t* aParams, bool aIsPrivate);
+  ~arena_t();
+
+ private:
+  void InitChunk(arena_chunk_t* aChunk, bool aZeroed);
+
+  void DeallocChunk(arena_chunk_t* aChunk);
+
+  arena_run_t* AllocRun(size_t aSize, bool aLarge, bool aZero);
+
+  void DallocRun(arena_run_t* aRun, bool aDirty);
+
+  [[nodiscard]] bool SplitRun(arena_run_t* aRun, size_t aSize, bool aLarge,
+                              bool aZero);
+
+  void TrimRunHead(arena_chunk_t* aChunk, arena_run_t* aRun, size_t aOldSize,
+                   size_t aNewSize);
+
+  void TrimRunTail(arena_chunk_t* aChunk, arena_run_t* aRun, size_t aOldSize,
+                   size_t aNewSize, bool dirty);
+
+  arena_run_t* GetNonFullBinRun(arena_bin_t* aBin);
+
+  inline uint8_t FindFreeBitInMask(uint32_t aMask, uint32_t& aRng);
+
+  inline void* ArenaRunRegAlloc(arena_run_t* aRun, arena_bin_t* aBin);
+
+  inline void* MallocSmall(size_t aSize, bool aZero);
+
+  void* MallocLarge(size_t aSize, bool aZero);
+
+  void* MallocHuge(size_t aSize, bool aZero);
+
+  void* PallocLarge(size_t aAlignment, size_t aSize, size_t aAllocSize);
+
+  void* PallocHuge(size_t aSize, size_t aAlignment, bool aZero);
+
+  void RallocShrinkLarge(arena_chunk_t* aChunk, void* aPtr, size_t aSize,
+                         size_t aOldSize);
+
+  bool RallocGrowLarge(arena_chunk_t* aChunk, void* aPtr, size_t aSize,
+                       size_t aOldSize);
+
+  void* RallocSmallOrLarge(void* aPtr, size_t aSize, size_t aOldSize);
+
+  void* RallocHuge(void* aPtr, size_t aSize, size_t aOldSize);
+
+ public:
+  inline void* Malloc(size_t aSize, bool aZero);
+
+  void* Palloc(size_t aAlignment, size_t aSize);
+
+  inline void DallocSmall(arena_chunk_t* aChunk, void* aPtr,
+                          arena_chunk_map_t* aMapElm);
+
+  void DallocLarge(arena_chunk_t* aChunk, void* aPtr);
+
+  void* Ralloc(void* aPtr, size_t aSize, size_t aOldSize);
+
+  void Purge(bool aAll);
+
+  void HardPurge();
+
+  void* operator new(size_t aCount) = delete;
+
+  void* operator new(size_t aCount, const fallible_t&) noexcept;
+
+  void operator delete(void*);
+};
+
+struct ArenaTreeTrait {
+  static RedBlackTreeNode<arena_t>& GetTreeNode(arena_t* aThis) {
+    return aThis->mLink;
+  }
+
+  static inline Order Compare(arena_t* aNode, arena_t* aOther) {
+    MOZ_ASSERT(aNode);
+    MOZ_ASSERT(aOther);
+    return CompareInt(aNode->mId, aOther->mId);
+  }
+};
+
+// Bookkeeping for all the arenas used by the allocator.
+// Arenas are separated in two categories:
+// - "private" arenas, used through the moz_arena_* API
+// - all the other arenas: the default arena, and thread-local arenas,
+//   used by the standard API.
+class ArenaCollection {
+ public:
+  bool Init() {
+    mArenas.Init();
+    mPrivateArenas.Init();
+    arena_params_t params;
+    // The main arena allows more dirty pages than the default for other arenas.
+    params.mMaxDirty = opt_dirty_max;
+    mDefaultArena =
+        mLock.Init() ? CreateArena(/* aIsPrivate = */ false, &params) : nullptr;
+    return bool(mDefaultArena);
+  }
+
+  inline arena_t* GetById(arena_id_t aArenaId, bool aIsPrivate);
+
+  arena_t* CreateArena(bool aIsPrivate, arena_params_t* aParams);
+
+  void DisposeArena(arena_t* aArena) {
+    MutexAutoLock lock(mLock);
+    MOZ_RELEASE_ASSERT(mPrivateArenas.Search(aArena),
+                       "Can only dispose of private arenas");
+    mPrivateArenas.Remove(aArena);
+    delete aArena;
+  }
+
+  using Tree = RedBlackTree<arena_t, ArenaTreeTrait>;
+
+  struct Iterator : Tree::Iterator {
+    explicit Iterator(Tree* aTree, Tree* aSecondTree)
+        : Tree::Iterator(aTree), mNextTree(aSecondTree) {}
+
+    Item<Iterator> begin() {
+      return Item<Iterator>(this, *Tree::Iterator::begin());
+    }
+
+    Item<Iterator> end() { return Item<Iterator>(this, nullptr); }
+
+    arena_t* Next() {
+      arena_t* result = Tree::Iterator::Next();
+      if (!result && mNextTree) {
+        new (this) Iterator(mNextTree, nullptr);
+        result = *Tree::Iterator::begin();
+      }
+      return result;
+    }
+
+   private:
+    Tree* mNextTree;
+  };
+
+  Iterator iter() { return Iterator(&mArenas, &mPrivateArenas); }
+
+  inline arena_t* GetDefault() { return mDefaultArena; }
+
+  Mutex mLock MOZ_UNANNOTATED;
+
+ private:
+  inline arena_t* GetByIdInternal(arena_id_t aArenaId, bool aIsPrivate);
+
+  arena_t* mDefaultArena;
+  arena_id_t mLastPublicArenaId;
+  Tree mArenas;
+  Tree mPrivateArenas;
+};
+
+static ArenaCollection gArenas;
+
+// ******
+// Chunks.
+static AddressRadixTree<(sizeof(void*) << 3) - LOG2(kChunkSize)> gChunkRTree;
+
+// Protects chunk-related data structures.
+static Mutex chunks_mtx;
+
+// Trees of chunks that were previously allocated (trees differ only in node
+// ordering).  These are used when allocating chunks, in an attempt to re-use
+// address space.  Depending on function, different tree orderings are needed,
+// which is why there are two trees with the same contents.
+static RedBlackTree<extent_node_t, ExtentTreeSzTrait> gChunksBySize
+    MOZ_GUARDED_BY(chunks_mtx);
+static RedBlackTree<extent_node_t, ExtentTreeTrait> gChunksByAddress
+    MOZ_GUARDED_BY(chunks_mtx);
+
+// Protects huge allocation-related data structures.
+static Mutex huge_mtx;
+
+// Tree of chunks that are stand-alone huge allocations.
+static RedBlackTree<extent_node_t, ExtentTreeTrait> huge
+    MOZ_GUARDED_BY(huge_mtx);
+
+// Huge allocation statistics.
+static size_t huge_allocated MOZ_GUARDED_BY(huge_mtx);
+static size_t huge_mapped MOZ_GUARDED_BY(huge_mtx);
+
+// **************************
+// base (internal allocation).
+
+static Mutex base_mtx;
+
+// Current pages that are being used for internal memory allocations.  These
+// pages are carved up in cacheline-size quanta, so that there is no chance of
+// false cache line sharing.
+static void* base_pages MOZ_GUARDED_BY(base_mtx);
+static void* base_next_addr MOZ_GUARDED_BY(base_mtx);
+static void* base_next_decommitted MOZ_GUARDED_BY(base_mtx);
+// Address immediately past base_pages.
+static void* base_past_addr MOZ_GUARDED_BY(base_mtx);
+static size_t base_mapped MOZ_GUARDED_BY(base_mtx);
+static size_t base_committed MOZ_GUARDED_BY(base_mtx);
+
+// ******
+// Arenas.
+
+// The arena associated with the current thread (per
+// jemalloc_thread_local_arena) On OSX, __thread/thread_local circles back
+// calling malloc to allocate storage on first access on each thread, which
+// leads to an infinite loop, but pthread-based TLS somehow doesn't have this
+// problem.
+#if !defined(XP_DARWIN)
+static MOZ_THREAD_LOCAL(arena_t*) thread_arena;
+#else
+static detail::ThreadLocal<arena_t*, detail::ThreadLocalKeyStorage>
+    thread_arena;
+#endif
+
+// *****************************
+// Runtime configuration options.
+
+const uint8_t kAllocJunk = 0xe4;
+const uint8_t kAllocPoison = 0xe5;
+
+#ifdef MOZ_DEBUG
+static bool opt_junk = true;
+static bool opt_zero = false;
+#else
+static const bool opt_junk = false;
+static const bool opt_zero = false;
+#endif
+static bool opt_randomize_small = true;
+
+// ***************************************************************************
+// Begin forward declarations.
+
+static void* chunk_alloc(size_t aSize, size_t aAlignment, bool aBase,
+                         bool* aZeroed = nullptr);
+static void chunk_dealloc(void* aChunk, size_t aSize, ChunkType aType);
+static void chunk_ensure_zero(void* aPtr, size_t aSize, bool aZeroed);
+static void huge_dalloc(void* aPtr, arena_t* aArena);
+static bool malloc_init_hard();
+
+#ifndef XP_WIN
+#  ifdef XP_DARWIN
+#    define FORK_HOOK extern "C"
+#  else
+#    define FORK_HOOK static
+#  endif
+FORK_HOOK void _malloc_prefork(void);
+FORK_HOOK void _malloc_postfork_parent(void);
+FORK_HOOK void _malloc_postfork_child(void);
+#endif
+
+// End forward declarations.
+// ***************************************************************************
+
+// FreeBSD's pthreads implementation calls malloc(3), so the malloc
+// implementation has to take pains to avoid infinite recursion during
+// initialization.
+// Returns whether the allocator was successfully initialized.
+static inline bool malloc_init() {
+  if (malloc_initialized == false) {
+    return malloc_init_hard();
+  }
+
+  return true;
+}
+
+static void _malloc_message(const char* p) {
+#if !defined(XP_WIN)
+#  define _write write
+#endif
+  // Pretend to check _write() errors to suppress gcc warnings about
+  // warn_unused_result annotations in some versions of glibc headers.
+  if (_write(STDERR_FILENO, p, (unsigned int)strlen(p)) < 0) {
+    return;
+  }
+}
+
+template <typename... Args>
+static void _malloc_message(const char* p, Args... args) {
+  _malloc_message(p);
+  _malloc_message(args...);
+}
+
+#ifdef ANDROID
+// Android's pthread.h does not declare pthread_atfork() until SDK 21.
+extern "C" MOZ_EXPORT int pthread_atfork(void (*)(void), void (*)(void),
+                                         void (*)(void));
+#endif
+
+// ***************************************************************************
+// Begin Utility functions/macros.
+
+// Return the chunk address for allocation address a.
+static inline arena_chunk_t* GetChunkForPtr(const void* aPtr) {
+  return (arena_chunk_t*)(uintptr_t(aPtr) & ~kChunkSizeMask);
+}
+
+// Return the chunk offset of address a.
+static inline size_t GetChunkOffsetForPtr(const void* aPtr) {
+  return (size_t)(uintptr_t(aPtr) & kChunkSizeMask);
+}
+
+static inline const char* _getprogname(void) { return "<jemalloc>"; }
+
+// Fill the given range of memory with zeroes or junk depending on opt_junk and
+// opt_zero. Callers can force filling with zeroes through the aForceZero
+// argument.
+static inline void ApplyZeroOrJunk(void* aPtr, size_t aSize) {
+  if (opt_junk) {
+    memset(aPtr, kAllocJunk, aSize);
+  } else if (opt_zero) {
+    memset(aPtr, 0, aSize);
+  }
+}
+
+// On Windows, delay crashing on OOM.
+#ifdef XP_WIN
+
+// Implementation of VirtualAlloc wrapper (bug 1716727).
+namespace MozAllocRetries {
+
+// Maximum retry count on OOM.
+constexpr size_t kMaxAttempts = 10;
+// Minimum delay time between retries. (The actual delay time may be larger. See
+// Microsoft's documentation for ::Sleep() for details.)
+constexpr size_t kDelayMs = 50;
+
+struct StallSpecs {
+  size_t maxAttempts;
+  size_t delayMs;
+};
+
+static constexpr StallSpecs maxStall = {.maxAttempts = kMaxAttempts,
+                                        .delayMs = kDelayMs};
+
+static inline StallSpecs GetStallSpecs() {
+#  if defined(JS_STANDALONE)
+  // GetGeckoProcessType() isn't available in this configuration. (SpiderMonkey
+  // on Windows mostly skips this in favor of directly calling ::VirtualAlloc(),
+  // though, so it's probably not going to matter whether we stall here or not.)
+  return maxStall;
+#  else
+  switch (GetGeckoProcessType()) {
+    // For the main process, stall for the maximum permissible time period. (The
+    // main process is the most important one to keep alive.)
+    case GeckoProcessType::GeckoProcessType_Default:
+      return maxStall;
+
+    // For all other process types, stall for at most half as long.
+    default:
+      return {.maxAttempts = kMaxAttempts / 2, .delayMs = kDelayMs};
+  }
+#  endif
+}
+
+// Drop-in wrapper around VirtualAlloc. When out of memory, may attempt to stall
+// and retry rather than returning immediately, in hopes that the page file is
+// about to be expanded by Windows.
+//
+// Ref: https://docs.microsoft.com/en-us/troubleshoot/windows-client/performance/slow-page-file-growth-memory-allocation-errors
+[[nodiscard]] void* MozVirtualAlloc(LPVOID lpAddress, SIZE_T dwSize,
+                                    DWORD flAllocationType, DWORD flProtect) {
+  constexpr auto IsOOMError = [] {
+    switch (::GetLastError()) {
+      // This is the usual error result from VirtualAlloc for OOM.
+      case ERROR_COMMITMENT_LIMIT:
+      // Although rare, this has also been observed in low-memory situations.
+      // (Presumably this means Windows can't allocate enough kernel-side space
+      // for its own internal representation of the process's virtual address
+      // space.)
+      case ERROR_NOT_ENOUGH_MEMORY:
+        return true;
+    }
+    return false;
+  };
+
+  {
+    void* ptr = ::VirtualAlloc(lpAddress, dwSize, flAllocationType, flProtect);
+    if (MOZ_LIKELY(ptr)) return ptr;
+
+    // We can't do anything for errors other than OOM...
+    if (!IsOOMError()) return nullptr;
+    // ... or if this wasn't a request to commit memory in the first place.
+    // (This function has no strategy for resolving MEM_RESERVE failures.)
+    if (!(flAllocationType & MEM_COMMIT)) return nullptr;
+  }
+
+  // Retry as many times as desired (possibly zero).
+  const StallSpecs stallSpecs = GetStallSpecs();
+
+  for (size_t i = 0; i < stallSpecs.maxAttempts; ++i) {
+    ::Sleep(stallSpecs.delayMs);
+    void* ptr = ::VirtualAlloc(lpAddress, dwSize, flAllocationType, flProtect);
+
+    if (ptr) {
+      // The OOM status has been handled, and should not be reported to
+      // telemetry.
+      if (IsOOMError()) {
+        ::SetLastError(0);
+      }
+      return ptr;
+    }
+
+    // Failure for some reason other than OOM.
+    if (!IsOOMError()) {
+      return nullptr;
+    }
+  }
+
+  // Ah, well. We tried.
+  return nullptr;
+}
+}  // namespace MozAllocRetries
+
+using MozAllocRetries::MozVirtualAlloc;
+#endif  // XP_WIN
+
+// ***************************************************************************
+
+static inline void pages_decommit(void* aAddr, size_t aSize) {
+#ifdef XP_WIN
+  // The region starting at addr may have been allocated in multiple calls
+  // to VirtualAlloc and recycled, so decommitting the entire region in one
+  // go may not be valid. However, since we allocate at least a chunk at a
+  // time, we may touch any region in chunksized increments.
+  size_t pages_size = std::min(aSize, kChunkSize - GetChunkOffsetForPtr(aAddr));
+  while (aSize > 0) {
+    // This will cause Access Violation on read and write and thus act as a
+    // guard page or region as well.
+    if (!VirtualFree(aAddr, pages_size, MEM_DECOMMIT)) {
+      MOZ_CRASH();
+    }
+    aAddr = (void*)((uintptr_t)aAddr + pages_size);
+    aSize -= pages_size;
+    pages_size = std::min(aSize, kChunkSize);
+  }
+#else
+  if (mmap(aAddr, aSize, PROT_NONE, MAP_FIXED | MAP_PRIVATE | MAP_ANON, -1,
+           0) == MAP_FAILED) {
+    // We'd like to report the OOM for our tooling, but we can't allocate
+    // memory at this point, so avoid the use of printf.
+    const char out_of_mappings[] =
+        "[unhandlable oom] Failed to mmap, likely no more mappings "
+        "available " __FILE__ " : " MOZ_STRINGIFY(__LINE__);
+    if (errno == ENOMEM) {
+#  ifndef ANDROID
+      fputs(out_of_mappings, stderr);
+      fflush(stderr);
+#  endif
+      MOZ_CRASH_ANNOTATE(out_of_mappings);
+    }
+    MOZ_REALLY_CRASH(__LINE__);
+  }
+  MozTagAnonymousMemory(aAddr, aSize, "jemalloc-decommitted");
+#endif
+}
+
+// Commit pages. Returns whether pages were committed.
+[[nodiscard]] static inline bool pages_commit(void* aAddr, size_t aSize) {
+#ifdef XP_WIN
+  // The region starting at addr may have been allocated in multiple calls
+  // to VirtualAlloc and recycled, so committing the entire region in one
+  // go may not be valid. However, since we allocate at least a chunk at a
+  // time, we may touch any region in chunksized increments.
+  size_t pages_size = std::min(aSize, kChunkSize - GetChunkOffsetForPtr(aAddr));
+  while (aSize > 0) {
+    if (!MozVirtualAlloc(aAddr, pages_size, MEM_COMMIT, PAGE_READWRITE)) {
+      return false;
+    }
+    aAddr = (void*)((uintptr_t)aAddr + pages_size);
+    aSize -= pages_size;
+    pages_size = std::min(aSize, kChunkSize);
+  }
+#else
+  if (mmap(aAddr, aSize, PROT_READ | PROT_WRITE,
+           MAP_FIXED | MAP_PRIVATE | MAP_ANON, -1, 0) == MAP_FAILED) {
+    return false;
+  }
+  MozTagAnonymousMemory(aAddr, aSize, "jemalloc");
+#endif
+  return true;
+}
+
+static bool base_pages_alloc(size_t minsize) MOZ_REQUIRES(base_mtx) {
+  size_t csize;
+  size_t pminsize;
+
+  MOZ_ASSERT(minsize != 0);
+  csize = CHUNK_CEILING(minsize);
+  base_pages = chunk_alloc(csize, kChunkSize, true);
+  if (!base_pages) {
+    return true;
+  }
+  base_next_addr = base_pages;
+  base_past_addr = (void*)((uintptr_t)base_pages + csize);
+  // Leave enough pages for minsize committed, since otherwise they would
+  // have to be immediately recommitted.
+  pminsize = PAGE_CEILING(minsize);
+  base_next_decommitted = (void*)((uintptr_t)base_pages + pminsize);
+  if (pminsize < csize) {
+    pages_decommit(base_next_decommitted, csize - pminsize);
+  }
+  base_mapped += csize;
+  base_committed += pminsize;
+
+  return false;
+}
+
+static void* base_alloc(size_t aSize) {
+  void* ret;
+  size_t csize;
+
+  // Round size up to nearest multiple of the cacheline size.
+  csize = CACHELINE_CEILING(aSize);
+
+  MutexAutoLock lock(base_mtx);
+  // Make sure there's enough space for the allocation.
+  if ((uintptr_t)base_next_addr + csize > (uintptr_t)base_past_addr) {
+    if (base_pages_alloc(csize)) {
+      return nullptr;
+    }
+  }
+  // Allocate.
+  ret = base_next_addr;
+  base_next_addr = (void*)((uintptr_t)base_next_addr + csize);
+  // Make sure enough pages are committed for the new allocation.
+  if ((uintptr_t)base_next_addr > (uintptr_t)base_next_decommitted) {
+    void* pbase_next_addr = (void*)(PAGE_CEILING((uintptr_t)base_next_addr));
+
+    if (!pages_commit(
+            base_next_decommitted,
+            (uintptr_t)pbase_next_addr - (uintptr_t)base_next_decommitted)) {
+      return nullptr;
+    }
+
+    base_committed +=
+        (uintptr_t)pbase_next_addr - (uintptr_t)base_next_decommitted;
+    base_next_decommitted = pbase_next_addr;
+  }
+
+  return ret;
+}
+
+static void* base_calloc(size_t aNumber, size_t aSize) {
+  void* ret = base_alloc(aNumber * aSize);
+  if (ret) {
+    memset(ret, 0, aNumber * aSize);
+  }
+  return ret;
+}
+
+// A specialization of the base allocator with a free list.
+template <typename T>
+struct TypedBaseAlloc {
+  static T* sFirstFree;
+
+  static size_t size_of() { return sizeof(T); }
+
+  static T* alloc() {
+    T* ret;
+
+    base_mtx.Lock();
+    if (sFirstFree) {
+      ret = sFirstFree;
+      sFirstFree = *(T**)ret;
+      base_mtx.Unlock();
+    } else {
+      base_mtx.Unlock();
+      ret = (T*)base_alloc(size_of());
+    }
+
+    return ret;
+  }
+
+  static void dealloc(T* aNode) {
+    MutexAutoLock lock(base_mtx);
+    *(T**)aNode = sFirstFree;
+    sFirstFree = aNode;
+  }
+};
+
+using ExtentAlloc = TypedBaseAlloc<extent_node_t>;
+
+template <>
+extent_node_t* ExtentAlloc::sFirstFree = nullptr;
+
+template <>
+arena_t* TypedBaseAlloc<arena_t>::sFirstFree = nullptr;
+
+template <>
+size_t TypedBaseAlloc<arena_t>::size_of() {
+  // Allocate enough space for trailing bins.
+  return sizeof(arena_t) + (sizeof(arena_bin_t) * (NUM_SMALL_CLASSES - 1));
+}
+
+template <typename T>
+struct BaseAllocFreePolicy {
+  void operator()(T* aPtr) { TypedBaseAlloc<T>::dealloc(aPtr); }
+};
+
+using UniqueBaseNode =
+    UniquePtr<extent_node_t, BaseAllocFreePolicy<extent_node_t>>;
+
+// End Utility functions/macros.
+// ***************************************************************************
+// Begin chunk management functions.
+
+#ifdef XP_WIN
+
+static void* pages_map(void* aAddr, size_t aSize) {
+  void* ret = nullptr;
+  ret = MozVirtualAlloc(aAddr, aSize, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE);
+  return ret;
+}
+
+static void pages_unmap(void* aAddr, size_t aSize) {
+  if (VirtualFree(aAddr, 0, MEM_RELEASE) == 0) {
+    _malloc_message(_getprogname(), ": (malloc) Error in VirtualFree()\n");
+  }
+}
+#else
+
+static void pages_unmap(void* aAddr, size_t aSize) {
+  if (munmap(aAddr, aSize) == -1) {
+    char buf[64];
+
+    if (strerror_r(errno, buf, sizeof(buf)) == 0) {
+      _malloc_message(_getprogname(), ": (malloc) Error in munmap(): ", buf,
+                      "\n");
+    }
+  }
+}
+
+static void* pages_map(void* aAddr, size_t aSize) {
+  void* ret;
+#  if defined(__ia64__) || \
+      (defined(__sparc__) && defined(__arch64__) && defined(__linux__))
+  // The JS engine assumes that all allocated pointers have their high 17 bits
+  // clear, which ia64's mmap doesn't support directly. However, we can emulate
+  // it by passing mmap an "addr" parameter with those bits clear. The mmap will
+  // return that address, or the nearest available memory above that address,
+  // providing a near-guarantee that those bits are clear. If they are not, we
+  // return nullptr below to indicate out-of-memory.
+  //
+  // The addr is chosen as 0x0000070000000000, which still allows about 120TB of
+  // virtual address space.
+  //
+  // See Bug 589735 for more information.
+  bool check_placement = true;
+  if (!aAddr) {
+    aAddr = (void*)0x0000070000000000;
+    check_placement = false;
+  }
+#  endif
+
+#  if defined(__sparc__) && defined(__arch64__) && defined(__linux__)
+  const uintptr_t start = 0x0000070000000000ULL;
+  const uintptr_t end = 0x0000800000000000ULL;
+
+  // Copied from js/src/gc/Memory.cpp and adapted for this source
+  uintptr_t hint;
+  void* region = MAP_FAILED;
+  for (hint = start; region == MAP_FAILED && hint + aSize <= end;
+       hint += kChunkSize) {
+    region = mmap((void*)hint, aSize, PROT_READ | PROT_WRITE,
+                  MAP_PRIVATE | MAP_ANON, -1, 0);
+    if (region != MAP_FAILED) {
+      if (((size_t)region + (aSize - 1)) & 0xffff800000000000) {
+        if (munmap(region, aSize)) {
+          MOZ_ASSERT(errno == ENOMEM);
+        }
+        region = MAP_FAILED;
+      }
+    }
+  }
+  ret = region;
+#  else
+  // We don't use MAP_FIXED here, because it can cause the *replacement*
+  // of existing mappings, and we only want to create new mappings.
+  ret =
+      mmap(aAddr, aSize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0);
+  MOZ_ASSERT(ret);
+#  endif
+  if (ret == MAP_FAILED) {
+    ret = nullptr;
+  }
+#  if defined(__ia64__) || \
+      (defined(__sparc__) && defined(__arch64__) && defined(__linux__))
+  // If the allocated memory doesn't have its upper 17 bits clear, consider it
+  // as out of memory.
+  else if ((long long)ret & 0xffff800000000000) {
+    munmap(ret, aSize);
+    ret = nullptr;
+  }
+  // If the caller requested a specific memory location, verify that's what mmap
+  // returned.
+  else if (check_placement && ret != aAddr) {
+#  else
+  else if (aAddr && ret != aAddr) {
+#  endif
+    // We succeeded in mapping memory, but not in the right place.
+    pages_unmap(ret, aSize);
+    ret = nullptr;
+  }
+  if (ret) {
+    MozTagAnonymousMemory(ret, aSize, "jemalloc");
+  }
+
+#  if defined(__ia64__) || \
+      (defined(__sparc__) && defined(__arch64__) && defined(__linux__))
+  MOZ_ASSERT(!ret || (!check_placement && ret) ||
+             (check_placement && ret == aAddr));
+#  else
+  MOZ_ASSERT(!ret || (!aAddr && ret != aAddr) || (aAddr && ret == aAddr));
+#  endif
+  return ret;
+}
+#endif
+
+#ifdef XP_DARWIN
+#  define VM_COPY_MIN kChunkSize
+static inline void pages_copy(void* dest, const void* src, size_t n) {
+  MOZ_ASSERT((void*)((uintptr_t)dest & ~gPageSizeMask) == dest);
+  MOZ_ASSERT(n >= VM_COPY_MIN);
+  MOZ_ASSERT((void*)((uintptr_t)src & ~gPageSizeMask) == src);
+
+  kern_return_t r = vm_copy(mach_task_self(), (vm_address_t)src, (vm_size_t)n,
+                            (vm_address_t)dest);
+  if (r != KERN_SUCCESS) {
+    MOZ_CRASH("vm_copy() failed");
+  }
+}
+
+#endif
+
+template <size_t Bits>
+bool AddressRadixTree<Bits>::Init() {
+  mLock.Init();
+  mRoot = (void**)base_calloc(1 << kBitsAtLevel1, sizeof(void*));
+  return mRoot;
+}
+
+template <size_t Bits>
+void** AddressRadixTree<Bits>::GetSlot(void* aKey, bool aCreate) {
+  uintptr_t key = reinterpret_cast<uintptr_t>(aKey);
+  uintptr_t subkey;
+  unsigned i, lshift, height, bits;
+  void** node;
+  void** child;
+
+  for (i = lshift = 0, height = kHeight, node = mRoot; i < height - 1;
+       i++, lshift += bits, node = child) {
+    bits = i ? kBitsPerLevel : kBitsAtLevel1;
+    subkey = (key << lshift) >> ((sizeof(void*) << 3) - bits);
+    child = (void**)node[subkey];
+    if (!child && aCreate) {
+      child = (void**)base_calloc(1 << kBitsPerLevel, sizeof(void*));
+      if (child) {
+        node[subkey] = child;
+      }
+    }
+    if (!child) {
+      return nullptr;
+    }
+  }
+
+  // node is a leaf, so it contains values rather than node
+  // pointers.
+  bits = i ? kBitsPerLevel : kBitsAtLevel1;
+  subkey = (key << lshift) >> ((sizeof(void*) << 3) - bits);
+  return &node[subkey];
+}
+
+template <size_t Bits>
+void* AddressRadixTree<Bits>::Get(void* aKey) {
+  void* ret = nullptr;
+
+  void** slot = GetSlot(aKey);
+
+  if (slot) {
+    ret = *slot;
+  }
+#ifdef MOZ_DEBUG
+  MutexAutoLock lock(mLock);
+
+  // Suppose that it were possible for a jemalloc-allocated chunk to be
+  // munmap()ped, followed by a different allocator in another thread re-using
+  // overlapping virtual memory, all without invalidating the cached rtree
+  // value.  The result would be a false positive (the rtree would claim that
+  // jemalloc owns memory that it had actually discarded).  I don't think this
+  // scenario is possible, but the following assertion is a prudent sanity
+  // check.
+  if (!slot) {
+    // In case a slot has been created in the meantime.
+    slot = GetSlot(aKey);
+  }
+  if (slot) {
+    // The MutexAutoLock above should act as a memory barrier, forcing
+    // the compiler to emit a new read instruction for *slot.
+    MOZ_ASSERT(ret == *slot);
+  } else {
+    MOZ_ASSERT(ret == nullptr);
+  }
+#endif
+  return ret;
+}
+
+template <size_t Bits>
+bool AddressRadixTree<Bits>::Set(void* aKey, void* aValue) {
+  MutexAutoLock lock(mLock);
+  void** slot = GetSlot(aKey, /* aCreate = */ true);
+  if (slot) {
+    *slot = aValue;
+  }
+  return slot;
+}
+
+// pages_trim, chunk_alloc_mmap_slow and chunk_alloc_mmap were cherry-picked
+// from upstream jemalloc 3.4.1 to fix Mozilla bug 956501.
+
+// Return the offset between a and the nearest aligned address at or below a.
+#define ALIGNMENT_ADDR2OFFSET(a, alignment) \
+  ((size_t)((uintptr_t)(a) & ((alignment)-1)))
+
+// Return the smallest alignment multiple that is >= s.
+#define ALIGNMENT_CEILING(s, alignment) \
+  (((s) + ((alignment)-1)) & (~((alignment)-1)))
+
+static void* pages_trim(void* addr, size_t alloc_size, size_t leadsize,
+                        size_t size) {
+  void* ret = (void*)((uintptr_t)addr + leadsize);
+
+  MOZ_ASSERT(alloc_size >= leadsize + size);
+#ifdef XP_WIN
+  {
+    void* new_addr;
+
+    pages_unmap(addr, alloc_size);
+    new_addr = pages_map(ret, size);
+    if (new_addr == ret) {
+      return ret;
+    }
+    if (new_addr) {
+      pages_unmap(new_addr, size);
+    }
+    return nullptr;
+  }
+#else
+  {
+    size_t trailsize = alloc_size - leadsize - size;
+
+    if (leadsize != 0) {
+      pages_unmap(addr, leadsize);
+    }
+    if (trailsize != 0) {
+      pages_unmap((void*)((uintptr_t)ret + size), trailsize);
+    }
+    return ret;
+  }
+#endif
+}
+
+static void* chunk_alloc_mmap_slow(size_t size, size_t alignment) {
+  void *ret, *pages;
+  size_t alloc_size, leadsize;
+
+  alloc_size = size + alignment - gRealPageSize;
+  // Beware size_t wrap-around.
+  if (alloc_size < size) {
+    return nullptr;
+  }
+  do {
+    pages = pages_map(nullptr, alloc_size);
+    if (!pages) {
+      return nullptr;
+    }
+    leadsize =
+        ALIGNMENT_CEILING((uintptr_t)pages, alignment) - (uintptr_t)pages;
+    ret = pages_trim(pages, alloc_size, leadsize, size);
+  } while (!ret);
+
+  MOZ_ASSERT(ret);
+  return ret;
+}
+
+static void* chunk_alloc_mmap(size_t size, size_t alignment) {
+  void* ret;
+  size_t offset;
+
+  // Ideally, there would be a way to specify alignment to mmap() (like
+  // NetBSD has), but in the absence of such a feature, we have to work
+  // hard to efficiently create aligned mappings. The reliable, but
+  // slow method is to create a mapping that is over-sized, then trim the
+  // excess. However, that always results in one or two calls to
+  // pages_unmap().
+  //
+  // Optimistically try mapping precisely the right amount before falling
+  // back to the slow method, with the expectation that the optimistic
+  // approach works most of the time.
+  ret = pages_map(nullptr, size);
+  if (!ret) {
+    return nullptr;
+  }
+  offset = ALIGNMENT_ADDR2OFFSET(ret, alignment);
+  if (offset != 0) {
+    pages_unmap(ret, size);
+    return chunk_alloc_mmap_slow(size, alignment);
+  }
+
+  MOZ_ASSERT(ret);
+  return ret;
+}
+
+// Purge and release the pages in the chunk of length `length` at `addr` to
+// the OS.
+// Returns whether the pages are guaranteed to be full of zeroes when the
+// function returns.
+// The force_zero argument explicitly requests that the memory is guaranteed
+// to be full of zeroes when the function returns.
+static bool pages_purge(void* addr, size_t length, bool force_zero) {
+  pages_decommit(addr, length);
+  return true;
+}
+
+static void* chunk_recycle(size_t aSize, size_t aAlignment, bool* aZeroed) {
+  extent_node_t key;
+
+  size_t alloc_size = aSize + aAlignment - kChunkSize;
+  // Beware size_t wrap-around.
+  if (alloc_size < aSize) {
+    return nullptr;
+  }
+  key.mAddr = nullptr;
+  key.mSize = alloc_size;
+  chunks_mtx.Lock();
+  extent_node_t* node = gChunksBySize.SearchOrNext(&key);
+  if (!node) {
+    chunks_mtx.Unlock();
+    return nullptr;
+  }
+  size_t leadsize = ALIGNMENT_CEILING((uintptr_t)node->mAddr, aAlignment) -
+                    (uintptr_t)node->mAddr;
+  MOZ_ASSERT(node->mSize >= leadsize + aSize);
+  size_t trailsize = node->mSize - leadsize - aSize;
+  void* ret = (void*)((uintptr_t)node->mAddr + leadsize);
+  ChunkType chunk_type = node->mChunkType;
+  if (aZeroed) {
+    *aZeroed = (chunk_type == ZEROED_CHUNK);
+  }
+  // Remove node from the tree.
+  gChunksBySize.Remove(node);
+  gChunksByAddress.Remove(node);
+  if (leadsize != 0) {
+    // Insert the leading space as a smaller chunk.
+    node->mSize = leadsize;
+    gChunksBySize.Insert(node);
+    gChunksByAddress.Insert(node);
+    node = nullptr;
+  }
+  if (trailsize != 0) {
+    // Insert the trailing space as a smaller chunk.
+    if (!node) {
+      // An additional node is required, but
+      // TypedBaseAlloc::alloc() can cause a new base chunk to be
+      // allocated.  Drop chunks_mtx in order to avoid
+      // deadlock, and if node allocation fails, deallocate
+      // the result before returning an error.
+      chunks_mtx.Unlock();
+      node = ExtentAlloc::alloc();
+      if (!node) {
+        chunk_dealloc(ret, aSize, chunk_type);
+        return nullptr;
+      }
+      chunks_mtx.Lock();
+    }
+    node->mAddr = (void*)((uintptr_t)(ret) + aSize);
+    node->mSize = trailsize;
+    node->mChunkType = chunk_type;
+    gChunksBySize.Insert(node);
+    gChunksByAddress.Insert(node);
+    node = nullptr;
+  }
+
+  gRecycledSize -= aSize;
+
+  chunks_mtx.Unlock();
+
+  if (node) {
+    ExtentAlloc::dealloc(node);
+  }
+  if (!pages_commit(ret, aSize)) {
+    return nullptr;
+  }
+  // pages_commit is guaranteed to zero the chunk.
+  if (aZeroed) {
+    *aZeroed = true;
+  }
+
+  return ret;
+}
+
+#ifdef XP_WIN
+// On Windows, calls to VirtualAlloc and VirtualFree must be matched, making it
+// awkward to recycle allocations of varying sizes. Therefore we only allow
+// recycling when the size equals the chunksize, unless deallocation is entirely
+// disabled.
+#  define CAN_RECYCLE(size) ((size) == kChunkSize)
+#else
+#  define CAN_RECYCLE(size) true
+#endif
+
+// Allocates `size` bytes of system memory aligned for `alignment`.
+// `base` indicates whether the memory will be used for the base allocator
+// (e.g. base_alloc).
+// `zeroed` is an outvalue that returns whether the allocated memory is
+// guaranteed to be full of zeroes. It can be omitted when the caller doesn't
+// care about the result.
+static void* chunk_alloc(size_t aSize, size_t aAlignment, bool aBase,
+                         bool* aZeroed) {
+  void* ret = nullptr;
+
+  MOZ_ASSERT(aSize != 0);
+  MOZ_ASSERT((aSize & kChunkSizeMask) == 0);
+  MOZ_ASSERT(aAlignment != 0);
+  MOZ_ASSERT((aAlignment & kChunkSizeMask) == 0);
+
+  // Base allocations can't be fulfilled by recycling because of
+  // possible deadlock or infinite recursion.
+  if (CAN_RECYCLE(aSize) && !aBase) {
+    ret = chunk_recycle(aSize, aAlignment, aZeroed);
+  }
+  if (!ret) {
+    ret = chunk_alloc_mmap(aSize, aAlignment);
+    if (aZeroed) {
+      *aZeroed = true;
+    }
+  }
+  if (ret && !aBase) {
+    if (!gChunkRTree.Set(ret, ret)) {
+      chunk_dealloc(ret, aSize, UNKNOWN_CHUNK);
+      return nullptr;
+    }
+  }
+
+  MOZ_ASSERT(GetChunkOffsetForPtr(ret) == 0);
+  return ret;
+}
+
+static void chunk_ensure_zero(void* aPtr, size_t aSize, bool aZeroed) {
+  if (aZeroed == false) {
+    memset(aPtr, 0, aSize);
+  }
+#ifdef MOZ_DEBUG
+  else {
+    size_t i;
+    size_t* p = (size_t*)(uintptr_t)aPtr;
+
+    for (i = 0; i < aSize / sizeof(size_t); i++) {
+      MOZ_ASSERT(p[i] == 0);
+    }
+  }
+#endif
+}
+
+static void chunk_record(void* aChunk, size_t aSize, ChunkType aType) {
+  extent_node_t key;
+
+  if (aType != ZEROED_CHUNK) {
+    if (pages_purge(aChunk, aSize, aType == HUGE_CHUNK)) {
+      aType = ZEROED_CHUNK;
+    }
+  }
+
+  // Allocate a node before acquiring chunks_mtx even though it might not
+  // be needed, because TypedBaseAlloc::alloc() may cause a new base chunk to
+  // be allocated, which could cause deadlock if chunks_mtx were already
+  // held.
+  UniqueBaseNode xnode(ExtentAlloc::alloc());
+  // Use xprev to implement conditional deferred deallocation of prev.
+  UniqueBaseNode xprev;
+
+  // RAII deallocates xnode and xprev defined above after unlocking
+  // in order to avoid potential dead-locks
+  MutexAutoLock lock(chunks_mtx);
+  key.mAddr = (void*)((uintptr_t)aChunk + aSize);
+  extent_node_t* node = gChunksByAddress.SearchOrNext(&key);
+  // Try to coalesce forward.
+  if (node && node->mAddr == key.mAddr) {
+    // Coalesce chunk with the following address range.  This does
+    // not change the position within gChunksByAddress, so only
+    // remove/insert from/into gChunksBySize.
+    gChunksBySize.Remove(node);
+    node->mAddr = aChunk;
+    node->mSize += aSize;
+    if (node->mChunkType != aType) {
+      node->mChunkType = RECYCLED_CHUNK;
+    }
+    gChunksBySize.Insert(node);
+  } else {
+    // Coalescing forward failed, so insert a new node.
+    if (!xnode) {
+      // TypedBaseAlloc::alloc() failed, which is an exceedingly
+      // unlikely failure.  Leak chunk; its pages have
+      // already been purged, so this is only a virtual
+      // memory leak.
+      return;
+    }
+    node = xnode.release();
+    node->mAddr = aChunk;
+    node->mSize = aSize;
+    node->mChunkType = aType;
+    gChunksByAddress.Insert(node);
+    gChunksBySize.Insert(node);
+  }
+
+  // Try to coalesce backward.
+  extent_node_t* prev = gChunksByAddress.Prev(node);
+  if (prev && (void*)((uintptr_t)prev->mAddr + prev->mSize) == aChunk) {
+    // Coalesce chunk with the previous address range.  This does
+    // not change the position within gChunksByAddress, so only
+    // remove/insert node from/into gChunksBySize.
+    gChunksBySize.Remove(prev);
+    gChunksByAddress.Remove(prev);
+
+    gChunksBySize.Remove(node);
+    node->mAddr = prev->mAddr;
+    node->mSize += prev->mSize;
+    if (node->mChunkType != prev->mChunkType) {
+      node->mChunkType = RECYCLED_CHUNK;
+    }
+    gChunksBySize.Insert(node);
+
+    xprev.reset(prev);
+  }
+
+  gRecycledSize += aSize;
+}
+
+static void chunk_dealloc(void* aChunk, size_t aSize, ChunkType aType) {
+  MOZ_ASSERT(aChunk);
+  MOZ_ASSERT(GetChunkOffsetForPtr(aChunk) == 0);
+  MOZ_ASSERT(aSize != 0);
+  MOZ_ASSERT((aSize & kChunkSizeMask) == 0);
+
+  gChunkRTree.Unset(aChunk);
+
+  if (CAN_RECYCLE(aSize)) {
+    size_t recycled_so_far = gRecycledSize;
+    // In case some race condition put us above the limit.
+    if (recycled_so_far < gRecycleLimit) {
+      size_t recycle_remaining = gRecycleLimit - recycled_so_far;
+      size_t to_recycle;
+      if (aSize > recycle_remaining) {
+        to_recycle = recycle_remaining;
+        // Drop pages that would overflow the recycle limit
+        pages_trim(aChunk, aSize, 0, to_recycle);
+      } else {
+        to_recycle = aSize;
+      }
+      chunk_record(aChunk, to_recycle, aType);
+      return;
+    }
+  }
+
+  pages_unmap(aChunk, aSize);
+}
+
+#undef CAN_RECYCLE
+
+// End chunk management functions.
+// ***************************************************************************
+// Begin arena.
+
+static inline arena_t* thread_local_arena(bool enabled) {
+  arena_t* arena;
+
+  if (enabled) {
+    // The arena will essentially be leaked if this function is
+    // called with `false`, but it doesn't matter at the moment.
+    // because in practice nothing actually calls this function
+    // with `false`, except maybe at shutdown.
+    arena =
+        gArenas.CreateArena(/* aIsPrivate = */ false, /* aParams = */ nullptr);
+  } else {
+    arena = gArenas.GetDefault();
+  }
+  thread_arena.set(arena);
+  return arena;
+}
+
+template <>
+inline void MozJemalloc::jemalloc_thread_local_arena(bool aEnabled) {
+  if (malloc_init()) {
+    thread_local_arena(aEnabled);
+  }
+}
+
+// Choose an arena based on a per-thread value.
+static inline arena_t* choose_arena(size_t size) {
+  arena_t* ret = nullptr;
+
+  // We can only use TLS if this is a PIC library, since for the static
+  // library version, libc's malloc is used by TLS allocation, which
+  // introduces a bootstrapping issue.
+
+  if (size > kMaxQuantumClass) {
+    // Force the default arena for larger allocations.
+    ret = gArenas.GetDefault();
+  } else {
+    // Check TLS to see if our thread has requested a pinned arena.
+    ret = thread_arena.get();
+    if (!ret) {
+      // Nothing in TLS. Pin this thread to the default arena.
+      ret = thread_local_arena(false);
+    }
+  }
+
+  MOZ_DIAGNOSTIC_ASSERT(ret);
+  return ret;
+}
+
+inline uint8_t arena_t::FindFreeBitInMask(uint32_t aMask, uint32_t& aRng) {
+  if (mPRNG != nullptr) {
+    if (aRng == UINT_MAX) {
+      aRng = mPRNG->next() % 32;
+    }
+    uint8_t bitIndex;
+    // RotateRight asserts when provided bad input.
+    aMask = aRng ? RotateRight(aMask, aRng)
+                 : aMask;  // Rotate the mask a random number of slots
+    bitIndex = CountTrailingZeroes32(aMask);
+    return (bitIndex + aRng) % 32;
+  }
+  return CountTrailingZeroes32(aMask);
+}
+
+inline void* arena_t::ArenaRunRegAlloc(arena_run_t* aRun, arena_bin_t* aBin) {
+  void* ret;
+  unsigned i, mask, bit, regind;
+  uint32_t rndPos = UINT_MAX;
+
+  MOZ_DIAGNOSTIC_ASSERT(aRun->mMagic == ARENA_RUN_MAGIC);
+  MOZ_ASSERT(aRun->mRegionsMinElement < aBin->mRunNumRegionsMask);
+
+  // Move the first check outside the loop, so that aRun->mRegionsMinElement can
+  // be updated unconditionally, without the possibility of updating it
+  // multiple times.
+  i = aRun->mRegionsMinElement;
+  mask = aRun->mRegionsMask[i];
+  if (mask != 0) {
+    bit = FindFreeBitInMask(mask, rndPos);
+
+    regind = ((i << (LOG2(sizeof(int)) + 3)) + bit);
+    MOZ_ASSERT(regind < aBin->mRunNumRegions);
+    ret = (void*)(((uintptr_t)aRun) + aBin->mRunFirstRegionOffset +
+                  (aBin->mSizeClass * regind));
+
+    // Clear bit.
+    mask ^= (1U << bit);
+    aRun->mRegionsMask[i] = mask;
+
+    return ret;
+  }
+
+  for (i++; i < aBin->mRunNumRegionsMask; i++) {
+    mask = aRun->mRegionsMask[i];
+    if (mask != 0) {
+      bit = FindFreeBitInMask(mask, rndPos);
+
+      regind = ((i << (LOG2(sizeof(int)) + 3)) + bit);
+      MOZ_ASSERT(regind < aBin->mRunNumRegions);
+      ret = (void*)(((uintptr_t)aRun) + aBin->mRunFirstRegionOffset +
+                    (aBin->mSizeClass * regind));
+
+      // Clear bit.
+      mask ^= (1U << bit);
+      aRun->mRegionsMask[i] = mask;
+
+      // Make a note that nothing before this element
+      // contains a free region.
+      aRun->mRegionsMinElement = i;  // Low payoff: + (mask == 0);
+
+      return ret;
+    }
+  }
+  // Not reached.
+  MOZ_DIAGNOSTIC_ASSERT(0);
+  return nullptr;
+}
+
+static inline void arena_run_reg_dalloc(arena_run_t* run, arena_bin_t* bin,
+                                        void* ptr, size_t size) {
+  uint32_t diff, regind;
+  unsigned elm, bit;
+
+  MOZ_DIAGNOSTIC_ASSERT(run->mMagic == ARENA_RUN_MAGIC);
+
+  // Avoid doing division with a variable divisor if possible.  Using
+  // actual division here can reduce allocator throughput by over 20%!
+  diff =
+      (uint32_t)((uintptr_t)ptr - (uintptr_t)run - bin->mRunFirstRegionOffset);
+
+  MOZ_ASSERT(diff <=
+             (static_cast<unsigned>(bin->mRunSizePages) << gPageSize2Pow));
+  regind = diff / bin->mSizeDivisor;
+
+  MOZ_DIAGNOSTIC_ASSERT(diff == regind * size);
+  MOZ_DIAGNOSTIC_ASSERT(regind < bin->mRunNumRegions);
+
+  elm = regind >> (LOG2(sizeof(int)) + 3);
+  if (elm < run->mRegionsMinElement) {
+    run->mRegionsMinElement = elm;
+  }
+  bit = regind - (elm << (LOG2(sizeof(int)) + 3));
+  MOZ_RELEASE_ASSERT((run->mRegionsMask[elm] & (1U << bit)) == 0,
+                     "Double-free?");
+  run->mRegionsMask[elm] |= (1U << bit);
+}
+
+bool arena_t::SplitRun(arena_run_t* aRun, size_t aSize, bool aLarge,
+                       bool aZero) {
+  arena_chunk_t* chunk;
+  size_t old_ndirty, run_ind, total_pages, need_pages, rem_pages, i;
+
+  chunk = GetChunkForPtr(aRun);
+  old_ndirty = chunk->ndirty;
+  run_ind = (unsigned)((uintptr_t(aRun) - uintptr_t(chunk)) >> gPageSize2Pow);
+  total_pages = (chunk->map[run_ind].bits & ~gPageSizeMask) >> gPageSize2Pow;
+  need_pages = (aSize >> gPageSize2Pow);
+  MOZ_ASSERT(need_pages > 0);
+  MOZ_ASSERT(need_pages <= total_pages);
+  rem_pages = total_pages - need_pages;
+
+  for (i = 0; i < need_pages; i++) {
+    // Commit decommitted pages if necessary.  If a decommitted
+    // page is encountered, commit all needed adjacent decommitted
+    // pages in one operation, in order to reduce system call
+    // overhead.
+    if (chunk->map[run_ind + i].bits & CHUNK_MAP_MADVISED_OR_DECOMMITTED) {
+      size_t j;
+
+      // Advance i+j to just past the index of the last page
+      // to commit.  Clear CHUNK_MAP_DECOMMITTED and
+      // CHUNK_MAP_MADVISED along the way.
+      for (j = 0; i + j < need_pages && (chunk->map[run_ind + i + j].bits &
+                                         CHUNK_MAP_MADVISED_OR_DECOMMITTED);
+           j++) {
+        // DECOMMITTED and MADVISED are mutually exclusive.
+        MOZ_ASSERT(!(chunk->map[run_ind + i + j].bits & CHUNK_MAP_DECOMMITTED &&
+                     chunk->map[run_ind + i + j].bits & CHUNK_MAP_MADVISED));
+
+        chunk->map[run_ind + i + j].bits &= ~CHUNK_MAP_MADVISED_OR_DECOMMITTED;
+      }
+
+#ifdef MALLOC_DECOMMIT
+      bool committed = pages_commit(
+          (void*)(uintptr_t(chunk) + ((run_ind + i) << gPageSize2Pow)),
+          j << gPageSize2Pow);
+      // pages_commit zeroes pages, so mark them as such if it succeeded.
+      // That's checked further below to avoid manually zeroing the pages.
+      for (size_t k = 0; k < j; k++) {
+        chunk->map[run_ind + i + k].bits |=
+            committed ? CHUNK_MAP_ZEROED : CHUNK_MAP_DECOMMITTED;
+      }
+      if (!committed) {
+        return false;
+      }
+#endif
+
+      mStats.committed += j;
+    }
+  }
+
+  mRunsAvail.Remove(&chunk->map[run_ind]);
+
+  // Keep track of trailing unused pages for later use.
+  if (rem_pages > 0) {
+    chunk->map[run_ind + need_pages].bits =
+        (rem_pages << gPageSize2Pow) |
+        (chunk->map[run_ind + need_pages].bits & gPageSizeMask);
+    chunk->map[run_ind + total_pages - 1].bits =
+        (rem_pages << gPageSize2Pow) |
+        (chunk->map[run_ind + total_pages - 1].bits & gPageSizeMask);
+    mRunsAvail.Insert(&chunk->map[run_ind + need_pages]);
+  }
+
+  for (i = 0; i < need_pages; i++) {
+    // Zero if necessary.
+    if (aZero) {
+      if ((chunk->map[run_ind + i].bits & CHUNK_MAP_ZEROED) == 0) {
+        memset((void*)(uintptr_t(chunk) + ((run_ind + i) << gPageSize2Pow)), 0,
+               gPageSize);
+        // CHUNK_MAP_ZEROED is cleared below.
+      }
+    }
+
+    // Update dirty page accounting.
+    if (chunk->map[run_ind + i].bits & CHUNK_MAP_DIRTY) {
+      chunk->ndirty--;
+      mNumDirty--;
+      // CHUNK_MAP_DIRTY is cleared below.
+    }
+
+    // Initialize the chunk map.
+    if (aLarge) {
+      chunk->map[run_ind + i].bits = CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED;
+    } else {
+      chunk->map[run_ind + i].bits = size_t(aRun) | CHUNK_MAP_ALLOCATED;
+    }
+  }
+
+  // Set the run size only in the first element for large runs.  This is
+  // primarily a debugging aid, since the lack of size info for trailing
+  // pages only matters if the application tries to operate on an
+  // interior pointer.
+  if (aLarge) {
+    chunk->map[run_ind].bits |= aSize;
+  }
+
+  if (chunk->ndirty == 0 && old_ndirty > 0) {
+    mChunksDirty.Remove(chunk);
+  }
+  return true;
+}
+
+void arena_t::InitChunk(arena_chunk_t* aChunk, bool aZeroed) {
+  size_t i;
+  // WARNING: The following relies on !aZeroed meaning "used to be an arena
+  // chunk".
+  // When the chunk we're initializating as an arena chunk is zeroed, we
+  // mark all runs are decommitted and zeroed.
+  // When it is not, which we can assume means it's a recycled arena chunk,
+  // all it can contain is an arena chunk header (which we're overwriting),
+  // and zeroed or poisoned memory (because a recycled arena chunk will
+  // have been emptied before being recycled). In that case, we can get
+  // away with reusing the chunk as-is, marking all runs as madvised.
+
+  size_t flags =
+      aZeroed ? CHUNK_MAP_DECOMMITTED | CHUNK_MAP_ZEROED : CHUNK_MAP_MADVISED;
+
+  mStats.mapped += kChunkSize;
+
+  aChunk->arena = this;
+
+  // Claim that no pages are in use, since the header is merely overhead.
+  aChunk->ndirty = 0;
+
+  // Initialize the map to contain one maximal free untouched run.
+  arena_run_t* run = (arena_run_t*)(uintptr_t(aChunk) +
+                                    (gChunkHeaderNumPages << gPageSize2Pow));
+
+  // Clear the bits for the real header pages.
+  for (i = 0; i < gChunkHeaderNumPages - 1; i++) {
+    aChunk->map[i].bits = 0;
+  }
+  // Mark the leading guard page (last header page) as decommitted.
+  aChunk->map[i++].bits = CHUNK_MAP_DECOMMITTED;
+
+  // Mark the area usable for runs as available, note size at start and end
+  aChunk->map[i++].bits = gMaxLargeClass | flags;
+  for (; i < gChunkNumPages - 2; i++) {
+    aChunk->map[i].bits = flags;
+  }
+  aChunk->map[gChunkNumPages - 2].bits = gMaxLargeClass | flags;
+
+  // Mark the trailing guard page as decommitted.
+  aChunk->map[gChunkNumPages - 1].bits = CHUNK_MAP_DECOMMITTED;
+
+#ifdef MALLOC_DECOMMIT
+  // Start out decommitted, in order to force a closer correspondence
+  // between dirty pages and committed untouched pages. This includes
+  // leading and trailing guard pages.
+  pages_decommit((void*)(uintptr_t(run) - gPageSize),
+                 gMaxLargeClass + 2 * gPageSize);
+#else
+  // Decommit the last header page (=leading page) as a guard.
+  pages_decommit((void*)(uintptr_t(run) - gPageSize), gPageSize);
+  // Decommit the last page as a guard.
+  pages_decommit((void*)(uintptr_t(aChunk) + kChunkSize - gPageSize),
+                 gPageSize);
+#endif
+
+  mStats.committed += gChunkHeaderNumPages;
+
+  // Insert the run into the tree of available runs.
+  mRunsAvail.Insert(&aChunk->map[gChunkHeaderNumPages]);
+
+#ifdef MALLOC_DOUBLE_PURGE
+  new (&aChunk->chunks_madvised_elem) DoublyLinkedListElement<arena_chunk_t>();
+#endif
+}
+
+void arena_t::DeallocChunk(arena_chunk_t* aChunk) {
+  if (mSpare) {
+    if (mSpare->ndirty > 0) {
+      aChunk->arena->mChunksDirty.Remove(mSpare);
+      mNumDirty -= mSpare->ndirty;
+      mStats.committed -= mSpare->ndirty;
+    }
+
+#ifdef MALLOC_DOUBLE_PURGE
+    if (mChunksMAdvised.ElementProbablyInList(mSpare)) {
+      mChunksMAdvised.remove(mSpare);
+    }
+#endif
+
+    chunk_dealloc((void*)mSpare, kChunkSize, ARENA_CHUNK);
+    mStats.mapped -= kChunkSize;
+    mStats.committed -= gChunkHeaderNumPages;
+  }
+
+  // Remove run from the tree of available runs, so that the arena does not use
+  // it. Dirty page flushing only uses the tree of dirty chunks, so leaving this
+  // chunk in the chunks_* trees is sufficient for that purpose.
+  mRunsAvail.Remove(&aChunk->map[gChunkHeaderNumPages]);
+
+  mSpare = aChunk;
+}
+
+arena_run_t* arena_t::AllocRun(size_t aSize, bool aLarge, bool aZero) {
+  arena_run_t* run;
+  arena_chunk_map_t* mapelm;
+  arena_chunk_map_t key;
+
+  MOZ_ASSERT(aSize <= gMaxLargeClass);
+  MOZ_ASSERT((aSize & gPageSizeMask) == 0);
+
+  // Search the arena's chunks for the lowest best fit.
+  key.bits = aSize | CHUNK_MAP_KEY;
+  mapelm = mRunsAvail.SearchOrNext(&key);
+  if (mapelm) {
+    arena_chunk_t* chunk = GetChunkForPtr(mapelm);
+    size_t pageind =
+        (uintptr_t(mapelm) - uintptr_t(chunk->map)) / sizeof(arena_chunk_map_t);
+
+    run = (arena_run_t*)(uintptr_t(chunk) + (pageind << gPageSize2Pow));
+  } else if (mSpare) {
+    // Use the spare.
+    arena_chunk_t* chunk = mSpare;
+    mSpare = nullptr;
+    run = (arena_run_t*)(uintptr_t(chunk) +
+                         (gChunkHeaderNumPages << gPageSize2Pow));
+    // Insert the run into the tree of available runs.
+    mRunsAvail.Insert(&chunk->map[gChunkHeaderNumPages]);
+  } else {
+    // No usable runs.  Create a new chunk from which to allocate
+    // the run.
+    bool zeroed;
+    arena_chunk_t* chunk =
+        (arena_chunk_t*)chunk_alloc(kChunkSize, kChunkSize, false, &zeroed);
+    if (!chunk) {
+      return nullptr;
+    }
+
+    InitChunk(chunk, zeroed);
+    run = (arena_run_t*)(uintptr_t(chunk) +
+                         (gChunkHeaderNumPages << gPageSize2Pow));
+  }
+  // Update page map.
+  return SplitRun(run, aSize, aLarge, aZero) ? run : nullptr;
+}
+
+void arena_t::Purge(bool aAll) {
+  arena_chunk_t* chunk;
+  size_t i, npages;
+  // If all is set purge all dirty pages.
+  size_t dirty_max = aAll ? 1 : mMaxDirty;
+#ifdef MOZ_DEBUG
+  size_t ndirty = 0;
+  for (auto chunk : mChunksDirty.iter()) {
+    ndirty += chunk->ndirty;
+  }
+  MOZ_ASSERT(ndirty == mNumDirty);
+#endif
+  MOZ_DIAGNOSTIC_ASSERT(aAll || (mNumDirty > mMaxDirty));
+
+  // Iterate downward through chunks until enough dirty memory has been
+  // purged.  Terminate as soon as possible in order to minimize the
+  // number of system calls, even if a chunk has only been partially
+  // purged.
+  while (mNumDirty > (dirty_max >> 1)) {
+#ifdef MALLOC_DOUBLE_PURGE
+    bool madvised = false;
+#endif
+    chunk = mChunksDirty.Last();
+    MOZ_DIAGNOSTIC_ASSERT(chunk);
+    // Last page is DECOMMITTED as a guard page.
+    MOZ_ASSERT((chunk->map[gChunkNumPages - 1].bits & CHUNK_MAP_DECOMMITTED) !=
+               0);
+    for (i = gChunkNumPages - 2; chunk->ndirty > 0; i--) {
+      MOZ_DIAGNOSTIC_ASSERT(i >= gChunkHeaderNumPages);
+
+      if (chunk->map[i].bits & CHUNK_MAP_DIRTY) {
+#ifdef MALLOC_DECOMMIT
+        const size_t free_operation = CHUNK_MAP_DECOMMITTED;
+#else
+        const size_t free_operation = CHUNK_MAP_MADVISED;
+#endif
+        MOZ_ASSERT((chunk->map[i].bits & CHUNK_MAP_MADVISED_OR_DECOMMITTED) ==
+                   0);
+        chunk->map[i].bits ^= free_operation | CHUNK_MAP_DIRTY;
+        // Find adjacent dirty run(s).
+        for (npages = 1; i > gChunkHeaderNumPages &&
+                         (chunk->map[i - 1].bits & CHUNK_MAP_DIRTY);
+             npages++) {
+          i--;
+          MOZ_ASSERT((chunk->map[i].bits & CHUNK_MAP_MADVISED_OR_DECOMMITTED) ==
+                     0);
+          chunk->map[i].bits ^= free_operation | CHUNK_MAP_DIRTY;
+        }
+        chunk->ndirty -= npages;
+        mNumDirty -= npages;
+
+#ifdef MALLOC_DECOMMIT
+        pages_decommit((void*)(uintptr_t(chunk) + (i << gPageSize2Pow)),
+                       (npages << gPageSize2Pow));
+#endif
+        mStats.committed -= npages;
+
+#ifndef MALLOC_DECOMMIT
+#  ifdef XP_SOLARIS
+        posix_madvise((void*)(uintptr_t(chunk) + (i << gPageSize2Pow)),
+                      (npages << gPageSize2Pow), MADV_FREE);
+#  else
+        madvise((void*)(uintptr_t(chunk) + (i << gPageSize2Pow)),
+                (npages << gPageSize2Pow), MADV_FREE);
+#  endif
+#  ifdef MALLOC_DOUBLE_PURGE
+        madvised = true;
+#  endif
+#endif
+        if (mNumDirty <= (dirty_max >> 1)) {
+          break;
+        }
+      }
+    }
+
+    if (chunk->ndirty == 0) {
+      mChunksDirty.Remove(chunk);
+    }
+#ifdef MALLOC_DOUBLE_PURGE
+    if (madvised) {
+      // The chunk might already be in the list, but this
+      // makes sure it's at the front.
+      if (mChunksMAdvised.ElementProbablyInList(chunk)) {
+        mChunksMAdvised.remove(chunk);
+      }
+      mChunksMAdvised.pushFront(chunk);
+    }
+#endif
+  }
+}
+
+void arena_t::DallocRun(arena_run_t* aRun, bool aDirty) {
+  arena_chunk_t* chunk;
+  size_t size, run_ind, run_pages;
+
+  chunk = GetChunkForPtr(aRun);
+  run_ind = (size_t)((uintptr_t(aRun) - uintptr_t(chunk)) >> gPageSize2Pow);
+  MOZ_DIAGNOSTIC_ASSERT(run_ind >= gChunkHeaderNumPages);
+  MOZ_RELEASE_ASSERT(run_ind < gChunkNumPages - 1);
+  if ((chunk->map[run_ind].bits & CHUNK_MAP_LARGE) != 0) {
+    size = chunk->map[run_ind].bits & ~gPageSizeMask;
+    run_pages = (size >> gPageSize2Pow);
+  } else {
+    run_pages = aRun->mBin->mRunSizePages;
+    size = run_pages << gPageSize2Pow;
+  }
+
+  // Mark pages as unallocated in the chunk map.
+  if (aDirty) {
+    size_t i;
+
+    for (i = 0; i < run_pages; i++) {
+      MOZ_DIAGNOSTIC_ASSERT((chunk->map[run_ind + i].bits & CHUNK_MAP_DIRTY) ==
+                            0);
+      chunk->map[run_ind + i].bits = CHUNK_MAP_DIRTY;
+    }
+
+    if (chunk->ndirty == 0) {
+      mChunksDirty.Insert(chunk);
+    }
+    chunk->ndirty += run_pages;
+    mNumDirty += run_pages;
+  } else {
+    size_t i;
+
+    for (i = 0; i < run_pages; i++) {
+      chunk->map[run_ind + i].bits &= ~(CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED);
+    }
+  }
+  chunk->map[run_ind].bits = size | (chunk->map[run_ind].bits & gPageSizeMask);
+  chunk->map[run_ind + run_pages - 1].bits =
+      size | (chunk->map[run_ind + run_pages - 1].bits & gPageSizeMask);
+
+  // Try to coalesce forward.
+  if (run_ind + run_pages < gChunkNumPages - 1 &&
+      (chunk->map[run_ind + run_pages].bits & CHUNK_MAP_ALLOCATED) == 0) {
+    size_t nrun_size = chunk->map[run_ind + run_pages].bits & ~gPageSizeMask;
+
+    // Remove successor from tree of available runs; the coalesced run is
+    // inserted later.
+    mRunsAvail.Remove(&chunk->map[run_ind + run_pages]);
+
+    size += nrun_size;
+    run_pages = size >> gPageSize2Pow;
+
+    MOZ_DIAGNOSTIC_ASSERT((chunk->map[run_ind + run_pages - 1].bits &
+                           ~gPageSizeMask) == nrun_size);
+    chunk->map[run_ind].bits =
+        size | (chunk->map[run_ind].bits & gPageSizeMask);
+    chunk->map[run_ind + run_pages - 1].bits =
+        size | (chunk->map[run_ind + run_pages - 1].bits & gPageSizeMask);
+  }
+
+  // Try to coalesce backward.
+  if (run_ind > gChunkHeaderNumPages &&
+      (chunk->map[run_ind - 1].bits & CHUNK_MAP_ALLOCATED) == 0) {
+    size_t prun_size = chunk->map[run_ind - 1].bits & ~gPageSizeMask;
+
+    run_ind -= prun_size >> gPageSize2Pow;
+
+    // Remove predecessor from tree of available runs; the coalesced run is
+    // inserted later.
+    mRunsAvail.Remove(&chunk->map[run_ind]);
+
+    size += prun_size;
+    run_pages = size >> gPageSize2Pow;
+
+    MOZ_DIAGNOSTIC_ASSERT((chunk->map[run_ind].bits & ~gPageSizeMask) ==
+                          prun_size);
+    chunk->map[run_ind].bits =
+        size | (chunk->map[run_ind].bits & gPageSizeMask);
+    chunk->map[run_ind + run_pages - 1].bits =
+        size | (chunk->map[run_ind + run_pages - 1].bits & gPageSizeMask);
+  }
+
+  // Insert into tree of available runs, now that coalescing is complete.
+  mRunsAvail.Insert(&chunk->map[run_ind]);
+
+  // Deallocate chunk if it is now completely unused.
+  if ((chunk->map[gChunkHeaderNumPages].bits &
+       (~gPageSizeMask | CHUNK_MAP_ALLOCATED)) == gMaxLargeClass) {
+    DeallocChunk(chunk);
+  }
+
+  // Enforce mMaxDirty.
+  if (mNumDirty > mMaxDirty) {
+    Purge(false);
+  }
+}
+
+void arena_t::TrimRunHead(arena_chunk_t* aChunk, arena_run_t* aRun,
+                          size_t aOldSize, size_t aNewSize) {
+  size_t pageind = (uintptr_t(aRun) - uintptr_t(aChunk)) >> gPageSize2Pow;
+  size_t head_npages = (aOldSize - aNewSize) >> gPageSize2Pow;
+
+  MOZ_ASSERT(aOldSize > aNewSize);
+
+  // Update the chunk map so that arena_t::RunDalloc() can treat the
+  // leading run as separately allocated.
+  aChunk->map[pageind].bits =
+      (aOldSize - aNewSize) | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED;
+  aChunk->map[pageind + head_npages].bits =
+      aNewSize | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED;
+
+  DallocRun(aRun, false);
+}
+
+void arena_t::TrimRunTail(arena_chunk_t* aChunk, arena_run_t* aRun,
+                          size_t aOldSize, size_t aNewSize, bool aDirty) {
+  size_t pageind = (uintptr_t(aRun) - uintptr_t(aChunk)) >> gPageSize2Pow;
+  size_t npages = aNewSize >> gPageSize2Pow;
+
+  MOZ_ASSERT(aOldSize > aNewSize);
+
+  // Update the chunk map so that arena_t::RunDalloc() can treat the
+  // trailing run as separately allocated.
+  aChunk->map[pageind].bits = aNewSize | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED;
+  aChunk->map[pageind + npages].bits =
+      (aOldSize - aNewSize) | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED;
+
+  DallocRun((arena_run_t*)(uintptr_t(aRun) + aNewSize), aDirty);
+}
+
+arena_run_t* arena_t::GetNonFullBinRun(arena_bin_t* aBin) {
+  arena_chunk_map_t* mapelm;
+  arena_run_t* run;
+  unsigned i, remainder;
+
+  // Look for a usable run.
+  mapelm = aBin->mNonFullRuns.First();
+  if (mapelm) {
+    // run is guaranteed to have available space.
+    aBin->mNonFullRuns.Remove(mapelm);
+    run = (arena_run_t*)(mapelm->bits & ~gPageSizeMask);
+    return run;
+  }
+  // No existing runs have any space available.
+
+  // Allocate a new run.
+  run = AllocRun(static_cast<size_t>(aBin->mRunSizePages) << gPageSize2Pow,
+                 false, false);
+  if (!run) {
+    return nullptr;
+  }
+  // Don't initialize if a race in arena_t::RunAlloc() allowed an existing
+  // run to become usable.
+  if (run == aBin->mCurrentRun) {
+    return run;
+  }
+
+  // Initialize run internals.
+  run->mBin = aBin;
+
+  for (i = 0; i < aBin->mRunNumRegionsMask - 1; i++) {
+    run->mRegionsMask[i] = UINT_MAX;
+  }
+  remainder = aBin->mRunNumRegions & ((1U << (LOG2(sizeof(int)) + 3)) - 1);
+  if (remainder == 0) {
+    run->mRegionsMask[i] = UINT_MAX;
+  } else {
+    // The last element has spare bits that need to be unset.
+    run->mRegionsMask[i] =
+        (UINT_MAX >> ((1U << (LOG2(sizeof(int)) + 3)) - remainder));
+  }
+
+  run->mRegionsMinElement = 0;
+
+  run->mNumFree = aBin->mRunNumRegions;
+#if defined(MOZ_DIAGNOSTIC_ASSERT_ENABLED)
+  run->mMagic = ARENA_RUN_MAGIC;
+#endif
+
+  aBin->mNumRuns++;
+  return run;
+}
+
+void arena_bin_t::Init(SizeClass aSizeClass) {
+  size_t try_run_size;
+  unsigned try_nregs, try_mask_nelms, try_reg0_offset;
+  // Size of the run header, excluding mRegionsMask.
+  static const size_t kFixedHeaderSize = offsetof(arena_run_t, mRegionsMask);
+
+  MOZ_ASSERT(aSizeClass.Size() <= gMaxBinClass);
+
+  try_run_size = gPageSize;
+
+  mCurrentRun = nullptr;
+  mNonFullRuns.Init();
+  mSizeClass = aSizeClass.Size();
+  mNumRuns = 0;
+
+  // Run size expansion loop.
+  while (true) {
+    try_nregs = ((try_run_size - kFixedHeaderSize) / mSizeClass) +
+                1;  // Counter-act try_nregs-- in loop.
+
+    // The do..while loop iteratively reduces the number of regions until
+    // the run header and the regions no longer overlap.  A closed formula
+    // would be quite messy, since there is an interdependency between the
+    // header's mask length and the number of regions.
+    do {
+      try_nregs--;
+      try_mask_nelms =
+          (try_nregs >> (LOG2(sizeof(int)) + 3)) +
+          ((try_nregs & ((1U << (LOG2(sizeof(int)) + 3)) - 1)) ? 1 : 0);
+      try_reg0_offset = try_run_size - (try_nregs * mSizeClass);
+    } while (kFixedHeaderSize + (sizeof(unsigned) * try_mask_nelms) >
+             try_reg0_offset);
+
+    // Try to keep the run overhead below kRunOverhead.
+    if (Fraction(try_reg0_offset, try_run_size) <= kRunOverhead) {
+      break;
+    }
+
+    // If the overhead is larger than the size class, it means the size class
+    // is small and doesn't align very well with the header. It's desirable to
+    // have smaller run sizes for them, so relax the overhead requirement.
+    if (try_reg0_offset > mSizeClass) {
+      if (Fraction(try_reg0_offset, try_run_size) <= kRunRelaxedOverhead) {
+        break;
+      }
+    }
+
+    // The run header includes one bit per region of the given size. For sizes
+    // small enough, the number of regions is large enough that growing the run
+    // size barely moves the needle for the overhead because of all those bits.
+    // For example, for a size of 8 bytes, adding 4KiB to the run size adds
+    // close to 512 bits to the header, which is 64 bytes.
+    // With such overhead, there is no way to get to the wanted overhead above,
+    // so we give up if the required size for mRegionsMask more than doubles the
+    // size of the run header.
+    if (try_mask_nelms * sizeof(unsigned) >= kFixedHeaderSize) {
+      break;
+    }
+
+    // If next iteration is going to be larger than the largest possible large
+    // size class, then we didn't find a setup where the overhead is small
+    // enough, and we can't do better than the current settings, so just use
+    // that.
+    if (try_run_size + gPageSize > gMaxLargeClass) {
+      break;
+    }
+
+    // Try more aggressive settings.
+    try_run_size += gPageSize;
+  }
+
+  MOZ_ASSERT(kFixedHeaderSize + (sizeof(unsigned) * try_mask_nelms) <=
+             try_reg0_offset);
+  MOZ_ASSERT((try_mask_nelms << (LOG2(sizeof(int)) + 3)) >= try_nregs);
+
+  // Copy final settings.
+  MOZ_ASSERT((try_run_size >> gPageSize2Pow) <= UINT8_MAX);
+  mRunSizePages = static_cast<uint8_t>(try_run_size >> gPageSize2Pow);
+  mRunNumRegions = try_nregs;
+  mRunNumRegionsMask = try_mask_nelms;
+  mRunFirstRegionOffset = try_reg0_offset;
+  mSizeDivisor = FastDivisor<uint16_t>(aSizeClass.Size(), try_run_size);
+}
+
+void* arena_t::MallocSmall(size_t aSize, bool aZero) {
+  void* ret;
+  arena_bin_t* bin;
+  arena_run_t* run;
+  SizeClass sizeClass(aSize);
+  aSize = sizeClass.Size();
+
+  switch (sizeClass.Type()) {
+    case SizeClass::Tiny:
+      bin = &mBins[FloorLog2(aSize / kMinTinyClass)];
+      break;
+    case SizeClass::Quantum:
+      // Although we divide 2 things by kQuantum, the compiler will
+      // reduce `kMinQuantumClass / kQuantum` and `kNumTinyClasses` to a
+      // single constant.
+      bin = &mBins[kNumTinyClasses + (aSize / kQuantum) -
+                   (kMinQuantumClass / kQuantum)];
+      break;
+    case SizeClass::QuantumWide:
+      bin =
+          &mBins[kNumTinyClasses + kNumQuantumClasses + (aSize / kQuantumWide) -
+                 (kMinQuantumWideClass / kQuantumWide)];
+      break;
+    case SizeClass::SubPage:
+      bin =
+          &mBins[kNumTinyClasses + kNumQuantumClasses + kNumQuantumWideClasses +
+                 (FloorLog2(aSize) - LOG2(kMinSubPageClass))];
+      break;
+    default:
+      MOZ_MAKE_COMPILER_ASSUME_IS_UNREACHABLE("Unexpected size class type");
+  }
+  MOZ_DIAGNOSTIC_ASSERT(aSize == bin->mSizeClass);
+
+  {
+    // Before we lock, we determine if we need to randomize the allocation
+    // because if we do, we need to create the PRNG which might require
+    // allocating memory (arc4random on OSX for example) and we need to
+    // avoid the deadlock
+    if (MOZ_UNLIKELY(mRandomizeSmallAllocations && mPRNG == nullptr)) {
+      // This is frustrating. Because the code backing RandomUint64 (arc4random
+      // for example) may allocate memory, and because
+      // mRandomizeSmallAllocations is true and we haven't yet initilized mPRNG,
+      // we would re-enter this same case and cause a deadlock inside e.g.
+      // arc4random.  So we temporarily disable mRandomizeSmallAllocations to
+      // skip this case and then re-enable it
+      mRandomizeSmallAllocations = false;
+      mozilla::Maybe<uint64_t> prngState1 = mozilla::RandomUint64();
+      mozilla::Maybe<uint64_t> prngState2 = mozilla::RandomUint64();
+      void* backing =
+          base_alloc(sizeof(mozilla::non_crypto::XorShift128PlusRNG));
+      mPRNG = new (backing) mozilla::non_crypto::XorShift128PlusRNG(
+          prngState1.valueOr(0), prngState2.valueOr(0));
+      mRandomizeSmallAllocations = true;
+    }
+    MOZ_ASSERT(!mRandomizeSmallAllocations || mPRNG);
+
+    MutexAutoLock lock(mLock);
+    run = bin->mCurrentRun;
+    if (MOZ_UNLIKELY(!run || run->mNumFree == 0)) {
+      run = bin->mCurrentRun = GetNonFullBinRun(bin);
+    }
+    if (MOZ_UNLIKELY(!run)) {
+      return nullptr;
+    }
+    MOZ_DIAGNOSTIC_ASSERT(run->mMagic == ARENA_RUN_MAGIC);
+    MOZ_DIAGNOSTIC_ASSERT(run->mNumFree > 0);
+    ret = ArenaRunRegAlloc(run, bin);
+    MOZ_DIAGNOSTIC_ASSERT(ret);
+    run->mNumFree--;
+    if (!ret) {
+      return nullptr;
+    }
+
+    mStats.allocated_small += aSize;
+  }
+
+  if (!aZero) {
+    ApplyZeroOrJunk(ret, aSize);
+  } else {
+    memset(ret, 0, aSize);
+  }
+
+  return ret;
+}
+
+void* arena_t::MallocLarge(size_t aSize, bool aZero) {
+  void* ret;
+
+  // Large allocation.
+  aSize = PAGE_CEILING(aSize);
+
+  {
+    MutexAutoLock lock(mLock);
+    ret = AllocRun(aSize, true, aZero);
+    if (!ret) {
+      return nullptr;
+    }
+    mStats.allocated_large += aSize;
+  }
+
+  if (!aZero) {
+    ApplyZeroOrJunk(ret, aSize);
+  }
+
+  return ret;
+}
+
+void* arena_t::Malloc(size_t aSize, bool aZero) {
+  MOZ_DIAGNOSTIC_ASSERT(mMagic == ARENA_MAGIC);
+  MOZ_ASSERT(aSize != 0);
+
+  if (aSize <= gMaxBinClass) {
+    return MallocSmall(aSize, aZero);
+  }
+  if (aSize <= gMaxLargeClass) {
+    return MallocLarge(aSize, aZero);
+  }
+  return MallocHuge(aSize, aZero);
+}
+
+// Only handles large allocations that require more than page alignment.
+void* arena_t::PallocLarge(size_t aAlignment, size_t aSize, size_t aAllocSize) {
+  void* ret;
+  size_t offset;
+  arena_chunk_t* chunk;
+
+  MOZ_ASSERT((aSize & gPageSizeMask) == 0);
+  MOZ_ASSERT((aAlignment & gPageSizeMask) == 0);
+
+  {
+    MutexAutoLock lock(mLock);
+    ret = AllocRun(aAllocSize, true, false);
+    if (!ret) {
+      return nullptr;
+    }
+
+    chunk = GetChunkForPtr(ret);
+
+    offset = uintptr_t(ret) & (aAlignment - 1);
+    MOZ_ASSERT((offset & gPageSizeMask) == 0);
+    MOZ_ASSERT(offset < aAllocSize);
+    if (offset == 0) {
+      TrimRunTail(chunk, (arena_run_t*)ret, aAllocSize, aSize, false);
+    } else {
+      size_t leadsize, trailsize;
+
+      leadsize = aAlignment - offset;
+      if (leadsize > 0) {
+        TrimRunHead(chunk, (arena_run_t*)ret, aAllocSize,
+                    aAllocSize - leadsize);
+        ret = (void*)(uintptr_t(ret) + leadsize);
+      }
+
+      trailsize = aAllocSize - leadsize - aSize;
+      if (trailsize != 0) {
+        // Trim trailing space.
+        MOZ_ASSERT(trailsize < aAllocSize);
+        TrimRunTail(chunk, (arena_run_t*)ret, aSize + trailsize, aSize, false);
+      }
+    }
+
+    mStats.allocated_large += aSize;
+  }
+
+  ApplyZeroOrJunk(ret, aSize);
+  return ret;
+}
+
+void* arena_t::Palloc(size_t aAlignment, size_t aSize) {
+  void* ret;
+  size_t ceil_size;
+
+  // Round size up to the nearest multiple of alignment.
+  //
+  // This done, we can take advantage of the fact that for each small
+  // size class, every object is aligned at the smallest power of two
+  // that is non-zero in the base two representation of the size.  For
+  // example:
+  //
+  //   Size |   Base 2 | Minimum alignment
+  //   -----+----------+------------------
+  //     96 |  1100000 |  32
+  //    144 | 10100000 |  32
+  //    192 | 11000000 |  64
+  //
+  // Depending on runtime settings, it is possible that arena_malloc()
+  // will further round up to a power of two, but that never causes
+  // correctness issues.
+  ceil_size = ALIGNMENT_CEILING(aSize, aAlignment);
+
+  // (ceil_size < aSize) protects against the combination of maximal
+  // alignment and size greater than maximal alignment.
+  if (ceil_size < aSize) {
+    // size_t overflow.
+    return nullptr;
+  }
+
+  if (ceil_size <= gPageSize ||
+      (aAlignment <= gPageSize && ceil_size <= gMaxLargeClass)) {
+    ret = Malloc(ceil_size, false);
+  } else {
+    size_t run_size;
+
+    // We can't achieve sub-page alignment, so round up alignment
+    // permanently; it makes later calculations simpler.
+    aAlignment = PAGE_CEILING(aAlignment);
+    ceil_size = PAGE_CEILING(aSize);
+
+    // (ceil_size < aSize) protects against very large sizes within
+    // pagesize of SIZE_T_MAX.
+    //
+    // (ceil_size + aAlignment < ceil_size) protects against the
+    // combination of maximal alignment and ceil_size large enough
+    // to cause overflow.  This is similar to the first overflow
+    // check above, but it needs to be repeated due to the new
+    // ceil_size value, which may now be *equal* to maximal
+    // alignment, whereas before we only detected overflow if the
+    // original size was *greater* than maximal alignment.
+    if (ceil_size < aSize || ceil_size + aAlignment < ceil_size) {
+      // size_t overflow.
+      return nullptr;
+    }
+
+    // Calculate the size of the over-size run that arena_palloc()
+    // would need to allocate in order to guarantee the alignment.
+    if (ceil_size >= aAlignment) {
+      run_size = ceil_size + aAlignment - gPageSize;
+    } else {
+      // It is possible that (aAlignment << 1) will cause
+      // overflow, but it doesn't matter because we also
+      // subtract pagesize, which in the case of overflow
+      // leaves us with a very large run_size.  That causes
+      // the first conditional below to fail, which means
+      // that the bogus run_size value never gets used for
+      // anything important.
+      run_size = (aAlignment << 1) - gPageSize;
+    }
+
+    if (run_size <= gMaxLargeClass) {
+      ret = PallocLarge(aAlignment, ceil_size, run_size);
+    } else if (aAlignment <= kChunkSize) {
+      ret = MallocHuge(ceil_size, false);
+    } else {
+      ret = PallocHuge(ceil_size, aAlignment, false);
+    }
+  }
+
+  MOZ_ASSERT((uintptr_t(ret) & (aAlignment - 1)) == 0);
+  return ret;
+}
+
+class AllocInfo {
+ public:
+  template <bool Validate = false>
+  static inline AllocInfo Get(const void* aPtr) {
+    // If the allocator is not initialized, the pointer can't belong to it.
+    if (Validate && malloc_initialized == false) {
+      return AllocInfo();
+    }
+
+    auto chunk = GetChunkForPtr(aPtr);
+    if (Validate) {
+      if (!chunk || !gChunkRTree.Get(chunk)) {
+        return AllocInfo();
+      }
+    }
+
+    if (chunk != aPtr) {
+      MOZ_DIAGNOSTIC_ASSERT(chunk->arena->mMagic == ARENA_MAGIC);
+
+      size_t pageind = (((uintptr_t)aPtr - (uintptr_t)chunk) >> gPageSize2Pow);
+      size_t mapbits = chunk->map[pageind].bits;
+      MOZ_DIAGNOSTIC_ASSERT((mapbits & CHUNK_MAP_ALLOCATED) != 0);
+
+      size_t size;
+      if ((mapbits & CHUNK_MAP_LARGE) == 0) {
+        arena_run_t* run = (arena_run_t*)(mapbits & ~gPageSizeMask);
+        MOZ_DIAGNOSTIC_ASSERT(run->mMagic == ARENA_RUN_MAGIC);
+        size = run->mBin->mSizeClass;
+      } else {
+        size = mapbits & ~gPageSizeMask;
+        MOZ_DIAGNOSTIC_ASSERT(size != 0);
+      }
+
+      return AllocInfo(size, chunk);
+    }
+
+    extent_node_t key;
+
+    // Huge allocation
+    key.mAddr = chunk;
+    MutexAutoLock lock(huge_mtx);
+    extent_node_t* node = huge.Search(&key);
+    if (Validate && !node) {
+      return AllocInfo();
+    }
+    return AllocInfo(node->mSize, node);
+  }
+
+  // Validate ptr before assuming that it points to an allocation.  Currently,
+  // the following validation is performed:
+  //
+  // + Check that ptr is not nullptr.
+  //
+  // + Check that ptr lies within a mapped chunk.
+  static inline AllocInfo GetValidated(const void* aPtr) {
+    return Get<true>(aPtr);
+  }
+
+  AllocInfo() : mSize(0), mChunk(nullptr) {}
+
+  explicit AllocInfo(size_t aSize, arena_chunk_t* aChunk)
+      : mSize(aSize), mChunk(aChunk) {
+    MOZ_ASSERT(mSize <= gMaxLargeClass);
+  }
+
+  explicit AllocInfo(size_t aSize, extent_node_t* aNode)
+      : mSize(aSize), mNode(aNode) {
+    MOZ_ASSERT(mSize > gMaxLargeClass);
+  }
+
+  size_t Size() { return mSize; }
+
+  arena_t* Arena() {
+    if (mSize <= gMaxLargeClass) {
+      return mChunk->arena;
+    }
+    // Best effort detection that we're not trying to access an already
+    // disposed arena. In the case of a disposed arena, the memory location
+    // pointed by mNode->mArena is either free (but still a valid memory
+    // region, per TypedBaseAlloc<arena_t>), in which case its id was reset,
+    // or has been reallocated for a new region, and its id is very likely
+    // different (per randomness). In both cases, the id is unlikely to
+    // match what it was for the disposed arena.
+    MOZ_RELEASE_ASSERT(mNode->mArenaId == mNode->mArena->mId);
+    return mNode->mArena;
+  }
+
+ private:
+  size_t mSize;
+  union {
+    // Pointer to the chunk associated with the allocation for small
+    // and large allocations.
+    arena_chunk_t* mChunk;
+
+    // Pointer to the extent node for huge allocations.
+    extent_node_t* mNode;
+  };
+};
+
+template <>
+inline void MozJemalloc::jemalloc_ptr_info(const void* aPtr,
+                                           jemalloc_ptr_info_t* aInfo) {
+  arena_chunk_t* chunk = GetChunkForPtr(aPtr);
+
+  // Is the pointer null, or within one chunk's size of null?
+  // Alternatively, if the allocator is not initialized yet, the pointer
+  // can't be known.
+  if (!chunk || !malloc_initialized) {
+    *aInfo = {TagUnknown, nullptr, 0, 0};
+    return;
+  }
+
+  // Look for huge allocations before looking for |chunk| in gChunkRTree.
+  // This is necessary because |chunk| won't be in gChunkRTree if it's
+  // the second or subsequent chunk in a huge allocation.
+  extent_node_t* node;
+  extent_node_t key;
+  {
+    MutexAutoLock lock(huge_mtx);
+    key.mAddr = const_cast<void*>(aPtr);
+    node =
+        reinterpret_cast<RedBlackTree<extent_node_t, ExtentTreeBoundsTrait>*>(
+            &huge)
+            ->Search(&key);
+    if (node) {
+      *aInfo = {TagLiveAlloc, node->mAddr, node->mSize, node->mArena->mId};
+      return;
+    }
+  }
+
+  // It's not a huge allocation. Check if we have a known chunk.
+  if (!gChunkRTree.Get(chunk)) {
+    *aInfo = {TagUnknown, nullptr, 0, 0};
+    return;
+  }
+
+  MOZ_DIAGNOSTIC_ASSERT(chunk->arena->mMagic == ARENA_MAGIC);
+
+  // Get the page number within the chunk.
+  size_t pageind = (((uintptr_t)aPtr - (uintptr_t)chunk) >> gPageSize2Pow);
+  if (pageind < gChunkHeaderNumPages) {
+    // Within the chunk header.
+    *aInfo = {TagUnknown, nullptr, 0, 0};
+    return;
+  }
+
+  size_t mapbits = chunk->map[pageind].bits;
+
+  if (!(mapbits & CHUNK_MAP_ALLOCATED)) {
+    void* pageaddr = (void*)(uintptr_t(aPtr) & ~gPageSizeMask);
+    *aInfo = {TagFreedPage, pageaddr, gPageSize, chunk->arena->mId};
+    return;
+  }
+
+  if (mapbits & CHUNK_MAP_LARGE) {
+    // It's a large allocation. Only the first page of a large
+    // allocation contains its size, so if the address is not in
+    // the first page, scan back to find the allocation size.
+    size_t size;
+    while (true) {
+      size = mapbits & ~gPageSizeMask;
+      if (size != 0) {
+        break;
+      }
+
+      // The following two return paths shouldn't occur in
+      // practice unless there is heap corruption.
+      pageind--;
+      MOZ_DIAGNOSTIC_ASSERT(pageind >= gChunkHeaderNumPages);
+      if (pageind < gChunkHeaderNumPages) {
+        *aInfo = {TagUnknown, nullptr, 0, 0};
+        return;
+      }
+
+      mapbits = chunk->map[pageind].bits;
+      MOZ_DIAGNOSTIC_ASSERT(mapbits & CHUNK_MAP_LARGE);
+      if (!(mapbits & CHUNK_MAP_LARGE)) {
+        *aInfo = {TagUnknown, nullptr, 0, 0};
+        return;
+      }
+    }
+
+    void* addr = ((char*)chunk) + (pageind << gPageSize2Pow);
+    *aInfo = {TagLiveAlloc, addr, size, chunk->arena->mId};
+    return;
+  }
+
+  // It must be a small allocation.
+  auto run = (arena_run_t*)(mapbits & ~gPageSizeMask);
+  MOZ_DIAGNOSTIC_ASSERT(run->mMagic == ARENA_RUN_MAGIC);
+
+  // The allocation size is stored in the run metadata.
+  size_t size = run->mBin->mSizeClass;
+
+  // Address of the first possible pointer in the run after its headers.
+  uintptr_t reg0_addr = (uintptr_t)run + run->mBin->mRunFirstRegionOffset;
+  if (aPtr < (void*)reg0_addr) {
+    // In the run header.
+    *aInfo = {TagUnknown, nullptr, 0, 0};
+    return;
+  }
+
+  // Position in the run.
+  unsigned regind = ((uintptr_t)aPtr - reg0_addr) / size;
+
+  // Pointer to the allocation's base address.
+  void* addr = (void*)(reg0_addr + regind * size);
+
+  // Check if the allocation has been freed.
+  unsigned elm = regind >> (LOG2(sizeof(int)) + 3);
+  unsigned bit = regind - (elm << (LOG2(sizeof(int)) + 3));
+  PtrInfoTag tag =
+      ((run->mRegionsMask[elm] & (1U << bit))) ? TagFreedAlloc : TagLiveAlloc;
+
+  *aInfo = {tag, addr, size, chunk->arena->mId};
+}
+
+namespace Debug {
+// Helper for debuggers. We don't want it to be inlined and optimized out.
+MOZ_NEVER_INLINE jemalloc_ptr_info_t* jemalloc_ptr_info(const void* aPtr) {
+  static jemalloc_ptr_info_t info;
+  MozJemalloc::jemalloc_ptr_info(aPtr, &info);
+  return &info;
+}
+}  // namespace Debug
+
+void arena_t::DallocSmall(arena_chunk_t* aChunk, void* aPtr,
+                          arena_chunk_map_t* aMapElm) {
+  arena_run_t* run;
+  arena_bin_t* bin;
+  size_t size;
+
+  run = (arena_run_t*)(aMapElm->bits & ~gPageSizeMask);
+  MOZ_DIAGNOSTIC_ASSERT(run->mMagic == ARENA_RUN_MAGIC);
+  bin = run->mBin;
+  size = bin->mSizeClass;
+  MOZ_DIAGNOSTIC_ASSERT(uintptr_t(aPtr) >=
+                        uintptr_t(run) + bin->mRunFirstRegionOffset);
+
+  memset(aPtr, kAllocPoison, size);
+
+  arena_run_reg_dalloc(run, bin, aPtr, size);
+  run->mNumFree++;
+
+  if (run->mNumFree == bin->mRunNumRegions) {
+    // Deallocate run.
+    if (run == bin->mCurrentRun) {
+      bin->mCurrentRun = nullptr;
+    } else if (bin->mRunNumRegions != 1) {
+      size_t run_pageind =
+          (uintptr_t(run) - uintptr_t(aChunk)) >> gPageSize2Pow;
+      arena_chunk_map_t* run_mapelm = &aChunk->map[run_pageind];
+
+      // This block's conditional is necessary because if the
+      // run only contains one region, then it never gets
+      // inserted into the non-full runs tree.
+      MOZ_DIAGNOSTIC_ASSERT(bin->mNonFullRuns.Search(run_mapelm) == run_mapelm);
+      bin->mNonFullRuns.Remove(run_mapelm);
+    }
+#if defined(MOZ_DIAGNOSTIC_ASSERT_ENABLED)
+    run->mMagic = 0;
+#endif
+    DallocRun(run, true);
+    bin->mNumRuns--;
+  } else if (run->mNumFree == 1 && run != bin->mCurrentRun) {
+    // Make sure that bin->mCurrentRun always refers to the lowest
+    // non-full run, if one exists.
+    if (!bin->mCurrentRun) {
+      bin->mCurrentRun = run;
+    } else if (uintptr_t(run) < uintptr_t(bin->mCurrentRun)) {
+      // Switch mCurrentRun.
+      if (bin->mCurrentRun->mNumFree > 0) {
+        arena_chunk_t* runcur_chunk = GetChunkForPtr(bin->mCurrentRun);
+        size_t runcur_pageind =
+            (uintptr_t(bin->mCurrentRun) - uintptr_t(runcur_chunk)) >>
+            gPageSize2Pow;
+        arena_chunk_map_t* runcur_mapelm = &runcur_chunk->map[runcur_pageind];
+
+        // Insert runcur.
+        MOZ_DIAGNOSTIC_ASSERT(!bin->mNonFullRuns.Search(runcur_mapelm));
+        bin->mNonFullRuns.Insert(runcur_mapelm);
+      }
+      bin->mCurrentRun = run;
+    } else {
+      size_t run_pageind =
+          (uintptr_t(run) - uintptr_t(aChunk)) >> gPageSize2Pow;
+      arena_chunk_map_t* run_mapelm = &aChunk->map[run_pageind];
+
+      MOZ_DIAGNOSTIC_ASSERT(bin->mNonFullRuns.Search(run_mapelm) == nullptr);
+      bin->mNonFullRuns.Insert(run_mapelm);
+    }
+  }
+  mStats.allocated_small -= size;
+}
+
+void arena_t::DallocLarge(arena_chunk_t* aChunk, void* aPtr) {
+  MOZ_DIAGNOSTIC_ASSERT((uintptr_t(aPtr) & gPageSizeMask) == 0);
+  size_t pageind = (uintptr_t(aPtr) - uintptr_t(aChunk)) >> gPageSize2Pow;
+  size_t size = aChunk->map[pageind].bits & ~gPageSizeMask;
+
+  memset(aPtr, kAllocPoison, size);
+  mStats.allocated_large -= size;
+
+  DallocRun((arena_run_t*)aPtr, true);
+}
+
+static inline void arena_dalloc(void* aPtr, size_t aOffset, arena_t* aArena) {
+  MOZ_ASSERT(aPtr);
+  MOZ_ASSERT(aOffset != 0);
+  MOZ_ASSERT(GetChunkOffsetForPtr(aPtr) == aOffset);
+
+  auto chunk = (arena_chunk_t*)((uintptr_t)aPtr - aOffset);
+  auto arena = chunk->arena;
+  MOZ_ASSERT(arena);
+  MOZ_DIAGNOSTIC_ASSERT(arena->mMagic == ARENA_MAGIC);
+  MOZ_RELEASE_ASSERT(!aArena || arena == aArena);
+
+  MutexAutoLock lock(arena->mLock);
+  size_t pageind = aOffset >> gPageSize2Pow;
+  arena_chunk_map_t* mapelm = &chunk->map[pageind];
+  MOZ_RELEASE_ASSERT((mapelm->bits & CHUNK_MAP_DECOMMITTED) == 0,
+                     "Freeing in decommitted page.");
+  MOZ_RELEASE_ASSERT((mapelm->bits & CHUNK_MAP_ALLOCATED) != 0, "Double-free?");
+  if ((mapelm->bits & CHUNK_MAP_LARGE) == 0) {
+    // Small allocation.
+    arena->DallocSmall(chunk, aPtr, mapelm);
+  } else {
+    // Large allocation.
+    arena->DallocLarge(chunk, aPtr);
+  }
+}
+
+static inline void idalloc(void* ptr, arena_t* aArena) {
+  size_t offset;
+
+  MOZ_ASSERT(ptr);
+
+  offset = GetChunkOffsetForPtr(ptr);
+  if (offset != 0) {
+    arena_dalloc(ptr, offset, aArena);
+  } else {
+    huge_dalloc(ptr, aArena);
+  }
+}
+
+void arena_t::RallocShrinkLarge(arena_chunk_t* aChunk, void* aPtr, size_t aSize,
+                                size_t aOldSize) {
+  MOZ_ASSERT(aSize < aOldSize);
+
+  // Shrink the run, and make trailing pages available for other
+  // allocations.
+  MutexAutoLock lock(mLock);
+  TrimRunTail(aChunk, (arena_run_t*)aPtr, aOldSize, aSize, true);
+  mStats.allocated_large -= aOldSize - aSize;
+}
+
+// Returns whether reallocation was successful.
+bool arena_t::RallocGrowLarge(arena_chunk_t* aChunk, void* aPtr, size_t aSize,
+                              size_t aOldSize) {
+  size_t pageind = (uintptr_t(aPtr) - uintptr_t(aChunk)) >> gPageSize2Pow;
+  size_t npages = aOldSize >> gPageSize2Pow;
+
+  MutexAutoLock lock(mLock);
+  MOZ_DIAGNOSTIC_ASSERT(aOldSize ==
+                        (aChunk->map[pageind].bits & ~gPageSizeMask));
+
+  // Try to extend the run.
+  MOZ_ASSERT(aSize > aOldSize);
+  if (pageind + npages < gChunkNumPages - 1 &&
+      (aChunk->map[pageind + npages].bits & CHUNK_MAP_ALLOCATED) == 0 &&
+      (aChunk->map[pageind + npages].bits & ~gPageSizeMask) >=
+          aSize - aOldSize) {
+    // The next run is available and sufficiently large.  Split the
+    // following run, then merge the first part with the existing
+    // allocation.
+    if (!SplitRun((arena_run_t*)(uintptr_t(aChunk) +
+                                 ((pageind + npages) << gPageSize2Pow)),
+                  aSize - aOldSize, true, false)) {
+      return false;
+    }
+
+    aChunk->map[pageind].bits = aSize | CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED;
+    aChunk->map[pageind + npages].bits = CHUNK_MAP_LARGE | CHUNK_MAP_ALLOCATED;
+
+    mStats.allocated_large += aSize - aOldSize;
+    return true;
+  }
+
+  return false;
+}
+
+void* arena_t::RallocSmallOrLarge(void* aPtr, size_t aSize, size_t aOldSize) {
+  void* ret;
+  size_t copysize;
+  SizeClass sizeClass(aSize);
+
+  // Try to avoid moving the allocation.
+  if (aOldSize <= gMaxLargeClass && sizeClass.Size() == aOldSize) {
+    if (aSize < aOldSize) {
+      memset((void*)(uintptr_t(aPtr) + aSize), kAllocPoison, aOldSize - aSize);
+    }
+    return aPtr;
+  }
+  if (sizeClass.Type() == SizeClass::Large && aOldSize > gMaxBinClass &&
+      aOldSize <= gMaxLargeClass) {
+    arena_chunk_t* chunk = GetChunkForPtr(aPtr);
+    if (sizeClass.Size() < aOldSize) {
+      // Fill before shrinking in order to avoid a race.
+      memset((void*)((uintptr_t)aPtr + aSize), kAllocPoison, aOldSize - aSize);
+      RallocShrinkLarge(chunk, aPtr, sizeClass.Size(), aOldSize);
+      return aPtr;
+    }
+    if (RallocGrowLarge(chunk, aPtr, sizeClass.Size(), aOldSize)) {
+      ApplyZeroOrJunk((void*)((uintptr_t)aPtr + aOldSize), aSize - aOldSize);
+      return aPtr;
+    }
+  }
+
+  // If we get here, then aSize and aOldSize are different enough that we
+  // need to move the object.  In that case, fall back to allocating new
+  // space and copying. Allow non-private arenas to switch arenas.
+  ret = (mIsPrivate ? this : choose_arena(aSize))->Malloc(aSize, false);
+  if (!ret) {
+    return nullptr;
+  }
+
+  // Junk/zero-filling were already done by arena_t::Malloc().
+  copysize = (aSize < aOldSize) ? aSize : aOldSize;
+#ifdef VM_COPY_MIN
+  if (copysize >= VM_COPY_MIN) {
+    pages_copy(ret, aPtr, copysize);
+  } else
+#endif
+  {
+    memcpy(ret, aPtr, copysize);
+  }
+  idalloc(aPtr, this);
+  return ret;
+}
+
+void* arena_t::Ralloc(void* aPtr, size_t aSize, size_t aOldSize) {
+  MOZ_DIAGNOSTIC_ASSERT(mMagic == ARENA_MAGIC);
+  MOZ_ASSERT(aPtr);
+  MOZ_ASSERT(aSize != 0);
+
+  return (aSize <= gMaxLargeClass) ? RallocSmallOrLarge(aPtr, aSize, aOldSize)
+                                   : RallocHuge(aPtr, aSize, aOldSize);
+}
+
+void* arena_t::operator new(size_t aCount, const fallible_t&) noexcept {
+  MOZ_ASSERT(aCount == sizeof(arena_t));
+  return TypedBaseAlloc<arena_t>::alloc();
+}
+
+void arena_t::operator delete(void* aPtr) {
+  TypedBaseAlloc<arena_t>::dealloc((arena_t*)aPtr);
+}
+
+arena_t::arena_t(arena_params_t* aParams, bool aIsPrivate) {
+  unsigned i;
+
+  MOZ_RELEASE_ASSERT(mLock.Init());
+
+  memset(&mLink, 0, sizeof(mLink));
+  memset(&mStats, 0, sizeof(arena_stats_t));
+  mId = 0;
+
+  // Initialize chunks.
+  mChunksDirty.Init();
+#ifdef MALLOC_DOUBLE_PURGE
+  new (&mChunksMAdvised) DoublyLinkedList<arena_chunk_t>();
+#endif
+  mSpare = nullptr;
+
+  mRandomizeSmallAllocations = opt_randomize_small;
+  if (aParams) {
+    uint32_t flags = aParams->mFlags & ARENA_FLAG_RANDOMIZE_SMALL_MASK;
+    switch (flags) {
+      case ARENA_FLAG_RANDOMIZE_SMALL_ENABLED:
+        mRandomizeSmallAllocations = true;
+        break;
+      case ARENA_FLAG_RANDOMIZE_SMALL_DISABLED:
+        mRandomizeSmallAllocations = false;
+        break;
+      case ARENA_FLAG_RANDOMIZE_SMALL_DEFAULT:
+      default:
+        break;
+    }
+  }
+  mPRNG = nullptr;
+
+  mIsPrivate = aIsPrivate;
+
+  mNumDirty = 0;
+  // The default maximum amount of dirty pages allowed on arenas is a fraction
+  // of opt_dirty_max.
+  mMaxDirty = (aParams && aParams->mMaxDirty) ? aParams->mMaxDirty
+                                              : (opt_dirty_max / 8);
+
+  mRunsAvail.Init();
+
+  // Initialize bins.
+  SizeClass sizeClass(1);
+
+  for (i = 0;; i++) {
+    arena_bin_t& bin = mBins[i];
+    bin.Init(sizeClass);
+
+    // SizeClass doesn't want sizes larger than gMaxBinClass for now.
+    if (sizeClass.Size() == gMaxBinClass) {
+      break;
+    }
+    sizeClass = sizeClass.Next();
+  }
+  MOZ_ASSERT(i == NUM_SMALL_CLASSES - 1);
+
+#if defined(MOZ_DIAGNOSTIC_ASSERT_ENABLED)
+  mMagic = ARENA_MAGIC;
+#endif
+}
+
+arena_t::~arena_t() {
+  size_t i;
+  MutexAutoLock lock(mLock);
+  MOZ_RELEASE_ASSERT(!mLink.Left() && !mLink.Right(),
+                     "Arena is still registered");
+  MOZ_RELEASE_ASSERT(!mStats.allocated_small && !mStats.allocated_large,
+                     "Arena is not empty");
+  if (mSpare) {
+    chunk_dealloc(mSpare, kChunkSize, ARENA_CHUNK);
+  }
+  for (i = 0; i < NUM_SMALL_CLASSES; i++) {
+    MOZ_RELEASE_ASSERT(!mBins[i].mNonFullRuns.First(), "Bin is not empty");
+  }
+#ifdef MOZ_DEBUG
+  {
+    MutexAutoLock lock(huge_mtx);
+    // This is an expensive check, so we only do it on debug builds.
+    for (auto node : huge.iter()) {
+      MOZ_RELEASE_ASSERT(node->mArenaId != mId, "Arena has huge allocations");
+    }
+  }
+#endif
+  mId = 0;
+}
+
+arena_t* ArenaCollection::CreateArena(bool aIsPrivate,
+                                      arena_params_t* aParams) {
+  arena_t* ret = new (fallible) arena_t(aParams, aIsPrivate);
+  if (!ret) {
+    // Only reached if there is an OOM error.
+
+    // OOM here is quite inconvenient to propagate, since dealing with it
+    // would require a check for failure in the fast path.  Instead, punt
+    // by using the first arena.
+    // In practice, this is an extremely unlikely failure.
+    _malloc_message(_getprogname(), ": (malloc) Error initializing arena\n");
+
+    return mDefaultArena;
+  }
+
+  MutexAutoLock lock(mLock);
+
+  // For public arenas, it's fine to just use incrementing arena id
+  if (!aIsPrivate) {
+    ret->mId = mLastPublicArenaId++;
+    mArenas.Insert(ret);
+    return ret;
+  }
+
+  // For private arenas, generate a cryptographically-secure random id for the
+  // new arena. If an attacker manages to get control of the process, this
+  // should make it more difficult for them to "guess" the ID of a memory
+  // arena, stopping them from getting data they may want
+
+  while (true) {
+    mozilla::Maybe<uint64_t> maybeRandomId = mozilla::RandomUint64();
+    MOZ_RELEASE_ASSERT(maybeRandomId.isSome());
+
+    // Avoid 0 as an arena Id. We use 0 for disposed arenas.
+    if (!maybeRandomId.value()) {
+      continue;
+    }
+
+    // Keep looping until we ensure that the random number we just generated
+    // isn't already in use by another active arena
+    arena_t* existingArena =
+        GetByIdInternal(maybeRandomId.value(), true /*aIsPrivate*/);
+
+    if (!existingArena) {
+      ret->mId = static_cast<arena_id_t>(maybeRandomId.value());
+      mPrivateArenas.Insert(ret);
+      return ret;
+    }
+  }
+}
+
+// End arena.
+// ***************************************************************************
+// Begin general internal functions.
+
+void* arena_t::MallocHuge(size_t aSize, bool aZero) {
+  return PallocHuge(aSize, kChunkSize, aZero);
+}
+
+void* arena_t::PallocHuge(size_t aSize, size_t aAlignment, bool aZero) {
+  void* ret;
+  size_t csize;
+  size_t psize;
+  extent_node_t* node;
+  bool zeroed;
+
+  // We're going to configure guard pages in the region between the
+  // page-aligned size and the chunk-aligned size, so if those are the same
+  // then we need to force that region into existence.
+  csize = CHUNK_CEILING(aSize + gPageSize);
+  if (csize < aSize) {
+    // size is large enough to cause size_t wrap-around.
+    return nullptr;
+  }
+
+  // Allocate an extent node with which to track the chunk.
+  node = ExtentAlloc::alloc();
+  if (!node) {
+    return nullptr;
+  }
+
+  // Allocate one or more contiguous chunks for this request.
+  ret = chunk_alloc(csize, aAlignment, false, &zeroed);
+  if (!ret) {
+    ExtentAlloc::dealloc(node);
+    return nullptr;
+  }
+  psize = PAGE_CEILING(aSize);
+  if (aZero) {
+    // We will decommit anything past psize so there is no need to zero
+    // further.
+    chunk_ensure_zero(ret, psize, zeroed);
+  }
+
+  // Insert node into huge.
+  node->mAddr = ret;
+  node->mSize = psize;
+  node->mArena = this;
+  node->mArenaId = mId;
+
+  {
+    MutexAutoLock lock(huge_mtx);
+    huge.Insert(node);
+
+    // Although we allocated space for csize bytes, we indicate that we've
+    // allocated only psize bytes.
+    //
+    // If DECOMMIT is defined, this is a reasonable thing to do, since
+    // we'll explicitly decommit the bytes in excess of psize.
+    //
+    // If DECOMMIT is not defined, then we're relying on the OS to be lazy
+    // about how it allocates physical pages to mappings.  If we never
+    // touch the pages in excess of psize, the OS won't allocate a physical
+    // page, and we won't use more than psize bytes of physical memory.
+    //
+    // A correct program will only touch memory in excess of how much it
+    // requested if it first calls malloc_usable_size and finds out how
+    // much space it has to play with.  But because we set node->mSize =
+    // psize above, malloc_usable_size will return psize, not csize, and
+    // the program will (hopefully) never touch bytes in excess of psize.
+    // Thus those bytes won't take up space in physical memory, and we can
+    // reasonably claim we never "allocated" them in the first place.
+    huge_allocated += psize;
+    huge_mapped += csize;
+  }
+
+  pages_decommit((void*)((uintptr_t)ret + psize), csize - psize);
+
+  if (!aZero) {
+    ApplyZeroOrJunk(ret, psize);
+  }
+
+  return ret;
+}
+
+void* arena_t::RallocHuge(void* aPtr, size_t aSize, size_t aOldSize) {
+  void* ret;
+  size_t copysize;
+
+  // Avoid moving the allocation if the size class would not change.
+  if (aOldSize > gMaxLargeClass &&
+      CHUNK_CEILING(aSize + gPageSize) == CHUNK_CEILING(aOldSize + gPageSize)) {
+    size_t psize = PAGE_CEILING(aSize);
+    if (aSize < aOldSize) {
+      memset((void*)((uintptr_t)aPtr + aSize), kAllocPoison, aOldSize - aSize);
+    }
+    if (psize < aOldSize) {
+      extent_node_t key;
+
+      pages_decommit((void*)((uintptr_t)aPtr + psize), aOldSize - psize);
+
+      // Update recorded size.
+      MutexAutoLock lock(huge_mtx);
+      key.mAddr = const_cast<void*>(aPtr);
+      extent_node_t* node = huge.Search(&key);
+      MOZ_ASSERT(node);
+      MOZ_ASSERT(node->mSize == aOldSize);
+      MOZ_RELEASE_ASSERT(node->mArena == this);
+      huge_allocated -= aOldSize - psize;
+      // No need to change huge_mapped, because we didn't (un)map anything.
+      node->mSize = psize;
+    } else if (psize > aOldSize) {
+      if (!pages_commit((void*)((uintptr_t)aPtr + aOldSize),
+                        psize - aOldSize)) {
+        return nullptr;
+      }
+
+      // We need to update the recorded size if the size increased,
+      // so malloc_usable_size doesn't return a value smaller than
+      // what was requested via realloc().
+      extent_node_t key;
+      MutexAutoLock lock(huge_mtx);
+      key.mAddr = const_cast<void*>(aPtr);
+      extent_node_t* node = huge.Search(&key);
+      MOZ_ASSERT(node);
+      MOZ_ASSERT(node->mSize == aOldSize);
+      MOZ_RELEASE_ASSERT(node->mArena == this);
+      huge_allocated += psize - aOldSize;
+      // No need to change huge_mapped, because we didn't
+      // (un)map anything.
+      node->mSize = psize;
+    }
+
+    if (aSize > aOldSize) {
+      ApplyZeroOrJunk((void*)((uintptr_t)aPtr + aOldSize), aSize - aOldSize);
+    }
+    return aPtr;
+  }
+
+  // If we get here, then aSize and aOldSize are different enough that we
+  // need to use a different size class.  In that case, fall back to allocating
+  // new space and copying. Allow non-private arenas to switch arenas.
+  ret = (mIsPrivate ? this : choose_arena(aSize))->MallocHuge(aSize, false);
+  if (!ret) {
+    return nullptr;
+  }
+
+  copysize = (aSize < aOldSize) ? aSize : aOldSize;
+#ifdef VM_COPY_MIN
+  if (copysize >= VM_COPY_MIN) {
+    pages_copy(ret, aPtr, copysize);
+  } else
+#endif
+  {
+    memcpy(ret, aPtr, copysize);
+  }
+  idalloc(aPtr, this);
+  return ret;
+}
+
+static void huge_dalloc(void* aPtr, arena_t* aArena) {
+  extent_node_t* node;
+  size_t mapped = 0;
+  {
+    extent_node_t key;
+    MutexAutoLock lock(huge_mtx);
+
+    // Extract from tree of huge allocations.
+    key.mAddr = aPtr;
+    node = huge.Search(&key);
+    MOZ_RELEASE_ASSERT(node, "Double-free?");
+    MOZ_ASSERT(node->mAddr == aPtr);
+    MOZ_RELEASE_ASSERT(!aArena || node->mArena == aArena);
+    // See AllocInfo::Arena.
+    MOZ_RELEASE_ASSERT(node->mArenaId == node->mArena->mId);
+    huge.Remove(node);
+
+    mapped = CHUNK_CEILING(node->mSize + gPageSize);
+    huge_allocated -= node->mSize;
+    huge_mapped -= mapped;
+  }
+
+  // Unmap chunk.
+  chunk_dealloc(node->mAddr, mapped, HUGE_CHUNK);
+
+  ExtentAlloc::dealloc(node);
+}
+
+static size_t GetKernelPageSize() {
+  static size_t kernel_page_size = ([]() {
+#ifdef XP_WIN
+    SYSTEM_INFO info;
+    GetSystemInfo(&info);
+    return info.dwPageSize;
+#else
+    long result = sysconf(_SC_PAGESIZE);
+    MOZ_ASSERT(result != -1);
+    return result;
+#endif
+  })();
+  return kernel_page_size;
+}
+
+// Returns whether the allocator was successfully initialized.
+static bool malloc_init_hard() {
+  unsigned i;
+  const char* opts;
+
+  AutoLock<StaticMutex> lock(gInitLock);
+
+  if (malloc_initialized) {
+    // Another thread initialized the allocator before this one
+    // acquired gInitLock.
+    return true;
+  }
+
+  if (!thread_arena.init()) {
+    return true;
+  }
+
+  // Get page size and number of CPUs
+  const size_t result = GetKernelPageSize();
+  // We assume that the page size is a power of 2.
+  MOZ_ASSERT(((result - 1) & result) == 0);
+#ifdef MALLOC_STATIC_PAGESIZE
+  if (gPageSize % result) {
+    _malloc_message(
+        _getprogname(),
+        "Compile-time page size does not divide the runtime one.\n");
+    MOZ_CRASH();
+  }
+#else
+  gRealPageSize = gPageSize = result;
+#endif
+
+  // Get runtime configuration.
+  if ((opts = getenv("MALLOC_OPTIONS"))) {
+    for (i = 0; opts[i] != '\0'; i++) {
+      unsigned j, nreps;
+      bool nseen;
+
+      // Parse repetition count, if any.
+      for (nreps = 0, nseen = false;; i++, nseen = true) {
+        switch (opts[i]) {
+          case '0':
+          case '1':
+          case '2':
+          case '3':
+          case '4':
+          case '5':
+          case '6':
+          case '7':
+          case '8':
+          case '9':
+            nreps *= 10;
+            nreps += opts[i] - '0';
+            break;
+          default:
+            goto MALLOC_OUT;
+        }
+      }
+    MALLOC_OUT:
+      if (nseen == false) {
+        nreps = 1;
+      }
+
+      for (j = 0; j < nreps; j++) {
+        switch (opts[i]) {
+          case 'f':
+            opt_dirty_max >>= 1;
+            break;
+          case 'F':
+            if (opt_dirty_max == 0) {
+              opt_dirty_max = 1;
+            } else if ((opt_dirty_max << 1) != 0) {
+              opt_dirty_max <<= 1;
+            }
+            break;
+#ifdef MOZ_DEBUG
+          case 'j':
+            opt_junk = false;
+            break;
+          case 'J':
+            opt_junk = true;
+            break;
+          case 'z':
+            opt_zero = false;
+            break;
+          case 'Z':
+            opt_zero = true;
+            break;
+#  ifndef MALLOC_STATIC_PAGESIZE
+          case 'P':
+            if (gPageSize < 64_KiB) {
+              gPageSize <<= 1;
+            }
+            break;
+#  endif
+#endif
+          case 'r':
+            opt_randomize_small = false;
+            break;
+          case 'R':
+            opt_randomize_small = true;
+            break;
+          default: {
+            char cbuf[2];
+
+            cbuf[0] = opts[i];
+            cbuf[1] = '\0';
+            _malloc_message(_getprogname(),
+                            ": (malloc) Unsupported character "
+                            "in malloc options: '",
+                            cbuf, "'\n");
+          }
+        }
+      }
+    }
+  }
+
+#ifndef MALLOC_STATIC_PAGESIZE
+  DefineGlobals();
+#endif
+  gRecycledSize = 0;
+
+  // Initialize chunks data.
+  chunks_mtx.Init();
+  MOZ_PUSH_IGNORE_THREAD_SAFETY
+  gChunksBySize.Init();
+  gChunksByAddress.Init();
+  MOZ_POP_THREAD_SAFETY
+
+  // Initialize huge allocation data.
+  huge_mtx.Init();
+  MOZ_PUSH_IGNORE_THREAD_SAFETY
+  huge.Init();
+  huge_allocated = 0;
+  huge_mapped = 0;
+  MOZ_POP_THREAD_SAFETY
+
+  // Initialize base allocation data structures.
+  base_mtx.Init();
+  MOZ_PUSH_IGNORE_THREAD_SAFETY
+  base_mapped = 0;
+  base_committed = 0;
+  MOZ_POP_THREAD_SAFETY
+
+  // Initialize arenas collection here.
+  if (!gArenas.Init()) {
+    return false;
+  }
+
+  // Assign the default arena to the initial thread.
+  thread_arena.set(gArenas.GetDefault());
+
+  if (!gChunkRTree.Init()) {
+    return false;
+  }
+
+  malloc_initialized = true;
+
+  // Dummy call so that the function is not removed by dead-code elimination
+  Debug::jemalloc_ptr_info(nullptr);
+
+#if !defined(XP_WIN) && !defined(XP_DARWIN)
+  // Prevent potential deadlock on malloc locks after fork.
+  pthread_atfork(_malloc_prefork, _malloc_postfork_parent,
+                 _malloc_postfork_child);
+#endif
+
+  return true;
+}
+
+// End general internal functions.
+// ***************************************************************************
+// Begin malloc(3)-compatible functions.
+
+// The BaseAllocator class is a helper class that implements the base allocator
+// functions (malloc, calloc, realloc, free, memalign) for a given arena,
+// or an appropriately chosen arena (per choose_arena()) when none is given.
+struct BaseAllocator {
+#define MALLOC_DECL(name, return_type, ...) \
+  inline return_type name(__VA_ARGS__);
+
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#include "malloc_decls.h"
+
+  explicit BaseAllocator(arena_t* aArena) : mArena(aArena) {}
+
+ private:
+  arena_t* mArena;
+};
+
+#define MALLOC_DECL(name, return_type, ...)                  \
+  template <>                                                \
+  inline return_type MozJemalloc::name(                      \
+      ARGS_HELPER(TYPED_ARGS, ##__VA_ARGS__)) {              \
+    BaseAllocator allocator(nullptr);                        \
+    return allocator.name(ARGS_HELPER(ARGS, ##__VA_ARGS__)); \
+  }
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#include "malloc_decls.h"
+
+inline void* BaseAllocator::malloc(size_t aSize) {
+  void* ret;
+  arena_t* arena;
+
+  if (!malloc_init()) {
+    ret = nullptr;
+    goto RETURN;
+  }
+
+  if (aSize == 0) {
+    aSize = 1;
+  }
+  arena = mArena ? mArena : choose_arena(aSize);
+  ret = arena->Malloc(aSize, /* aZero = */ false);
+
+RETURN:
+  if (!ret) {
+    errno = ENOMEM;
+  }
+
+  return ret;
+}
+
+inline void* BaseAllocator::memalign(size_t aAlignment, size_t aSize) {
+  MOZ_ASSERT(((aAlignment - 1) & aAlignment) == 0);
+
+  if (!malloc_init()) {
+    return nullptr;
+  }
+
+  if (aSize == 0) {
+    aSize = 1;
+  }
+
+  aAlignment = aAlignment < sizeof(void*) ? sizeof(void*) : aAlignment;
+  arena_t* arena = mArena ? mArena : choose_arena(aSize);
+  return arena->Palloc(aAlignment, aSize);
+}
+
+inline void* BaseAllocator::calloc(size_t aNum, size_t aSize) {
+  void* ret;
+
+  if (malloc_init()) {
+    CheckedInt<size_t> checkedSize = CheckedInt<size_t>(aNum) * aSize;
+    if (checkedSize.isValid()) {
+      size_t allocSize = checkedSize.value();
+      if (allocSize == 0) {
+        allocSize = 1;
+      }
+      arena_t* arena = mArena ? mArena : choose_arena(allocSize);
+      ret = arena->Malloc(allocSize, /* aZero = */ true);
+    } else {
+      ret = nullptr;
+    }
+  } else {
+    ret = nullptr;
+  }
+
+  if (!ret) {
+    errno = ENOMEM;
+  }
+
+  return ret;
+}
+
+inline void* BaseAllocator::realloc(void* aPtr, size_t aSize) {
+  void* ret;
+
+  if (aSize == 0) {
+    aSize = 1;
+  }
+
+  if (aPtr) {
+    MOZ_RELEASE_ASSERT(malloc_initialized);
+
+    auto info = AllocInfo::Get(aPtr);
+    auto arena = info.Arena();
+    MOZ_RELEASE_ASSERT(!mArena || arena == mArena);
+    ret = arena->Ralloc(aPtr, aSize, info.Size());
+  } else {
+    if (!malloc_init()) {
+      ret = nullptr;
+    } else {
+      arena_t* arena = mArena ? mArena : choose_arena(aSize);
+      ret = arena->Malloc(aSize, /* aZero = */ false);
+    }
+  }
+
+  if (!ret) {
+    errno = ENOMEM;
+  }
+  return ret;
+}
+
+inline void BaseAllocator::free(void* aPtr) {
+  size_t offset;
+
+  // A version of idalloc that checks for nullptr pointer.
+  offset = GetChunkOffsetForPtr(aPtr);
+  if (offset != 0) {
+    MOZ_RELEASE_ASSERT(malloc_initialized);
+    arena_dalloc(aPtr, offset, mArena);
+  } else if (aPtr) {
+    MOZ_RELEASE_ASSERT(malloc_initialized);
+    huge_dalloc(aPtr, mArena);
+  }
+}
+
+template <void* (*memalign)(size_t, size_t)>
+struct AlignedAllocator {
+  static inline int posix_memalign(void** aMemPtr, size_t aAlignment,
+                                   size_t aSize) {
+    void* result;
+
+    // alignment must be a power of two and a multiple of sizeof(void*)
+    if (((aAlignment - 1) & aAlignment) != 0 || aAlignment < sizeof(void*)) {
+      return EINVAL;
+    }
+
+    // The 0-->1 size promotion is done in the memalign() call below
+    result = memalign(aAlignment, aSize);
+
+    if (!result) {
+      return ENOMEM;
+    }
+
+    *aMemPtr = result;
+    return 0;
+  }
+
+  static inline void* aligned_alloc(size_t aAlignment, size_t aSize) {
+    if (aSize % aAlignment) {
+      return nullptr;
+    }
+    return memalign(aAlignment, aSize);
+  }
+
+  static inline void* valloc(size_t aSize) {
+    return memalign(GetKernelPageSize(), aSize);
+  }
+};
+
+template <>
+inline int MozJemalloc::posix_memalign(void** aMemPtr, size_t aAlignment,
+                                       size_t aSize) {
+  return AlignedAllocator<memalign>::posix_memalign(aMemPtr, aAlignment, aSize);
+}
+
+template <>
+inline void* MozJemalloc::aligned_alloc(size_t aAlignment, size_t aSize) {
+  return AlignedAllocator<memalign>::aligned_alloc(aAlignment, aSize);
+}
+
+template <>
+inline void* MozJemalloc::valloc(size_t aSize) {
+  return AlignedAllocator<memalign>::valloc(aSize);
+}
+
+// End malloc(3)-compatible functions.
+// ***************************************************************************
+// Begin non-standard functions.
+
+// This was added by Mozilla for use by SQLite.
+template <>
+inline size_t MozJemalloc::malloc_good_size(size_t aSize) {
+  if (aSize <= gMaxLargeClass) {
+    // Small or large
+    aSize = SizeClass(aSize).Size();
+  } else {
+    // Huge.  We use PAGE_CEILING to get psize, instead of using
+    // CHUNK_CEILING to get csize.  This ensures that this
+    // malloc_usable_size(malloc(n)) always matches
+    // malloc_good_size(n).
+    aSize = PAGE_CEILING(aSize);
+  }
+  return aSize;
+}
+
+template <>
+inline size_t MozJemalloc::malloc_usable_size(usable_ptr_t aPtr) {
+  return AllocInfo::GetValidated(aPtr).Size();
+}
+
+template <>
+inline void MozJemalloc::jemalloc_stats_internal(
+    jemalloc_stats_t* aStats, jemalloc_bin_stats_t* aBinStats) {
+  size_t non_arena_mapped, chunk_header_size;
+
+  if (!aStats) {
+    return;
+  }
+  if (!malloc_init()) {
+    memset(aStats, 0, sizeof(*aStats));
+    return;
+  }
+  if (aBinStats) {
+    memset(aBinStats, 0, sizeof(jemalloc_bin_stats_t) * NUM_SMALL_CLASSES);
+  }
+
+  // Gather runtime settings.
+  aStats->opt_junk = opt_junk;
+  aStats->opt_zero = opt_zero;
+  aStats->quantum = kQuantum;
+  aStats->quantum_max = kMaxQuantumClass;
+  aStats->quantum_wide = kQuantumWide;
+  aStats->quantum_wide_max = kMaxQuantumWideClass;
+  aStats->subpage_max = gMaxSubPageClass;
+  aStats->large_max = gMaxLargeClass;
+  aStats->chunksize = kChunkSize;
+  aStats->page_size = gPageSize;
+  aStats->dirty_max = opt_dirty_max;
+
+  // Gather current memory usage statistics.
+  aStats->narenas = 0;
+  aStats->mapped = 0;
+  aStats->allocated = 0;
+  aStats->waste = 0;
+  aStats->page_cache = 0;
+  aStats->bookkeeping = 0;
+  aStats->bin_unused = 0;
+
+  non_arena_mapped = 0;
+
+  // Get huge mapped/allocated.
+  {
+    MutexAutoLock lock(huge_mtx);
+    non_arena_mapped += huge_mapped;
+    aStats->allocated += huge_allocated;
+    MOZ_ASSERT(huge_mapped >= huge_allocated);
+  }
+
+  // Get base mapped/allocated.
+  {
+    MutexAutoLock lock(base_mtx);
+    non_arena_mapped += base_mapped;
+    aStats->bookkeeping += base_committed;
+    MOZ_ASSERT(base_mapped >= base_committed);
+  }
+
+  gArenas.mLock.Lock();
+  // Iterate over arenas.
+  for (auto arena : gArenas.iter()) {
+    size_t arena_mapped, arena_allocated, arena_committed, arena_dirty, j,
+        arena_unused, arena_headers;
+
+    arena_headers = 0;
+    arena_unused = 0;
+
+    {
+      MutexAutoLock lock(arena->mLock);
+
+      arena_mapped = arena->mStats.mapped;
+
+      // "committed" counts dirty and allocated memory.
+      arena_committed = arena->mStats.committed << gPageSize2Pow;
+
+      arena_allocated =
+          arena->mStats.allocated_small + arena->mStats.allocated_large;
+
+      arena_dirty = arena->mNumDirty << gPageSize2Pow;
+
+      for (j = 0; j < NUM_SMALL_CLASSES; j++) {
+        arena_bin_t* bin = &arena->mBins[j];
+        size_t bin_unused = 0;
+        size_t num_non_full_runs = 0;
+
+        for (auto mapelm : bin->mNonFullRuns.iter()) {
+          arena_run_t* run = (arena_run_t*)(mapelm->bits & ~gPageSizeMask);
+          bin_unused += run->mNumFree * bin->mSizeClass;
+          num_non_full_runs++;
+        }
+
+        if (bin->mCurrentRun) {
+          bin_unused += bin->mCurrentRun->mNumFree * bin->mSizeClass;
+          num_non_full_runs++;
+        }
+
+        arena_unused += bin_unused;
+        arena_headers += bin->mNumRuns * bin->mRunFirstRegionOffset;
+        if (aBinStats) {
+          aBinStats[j].size = bin->mSizeClass;
+          aBinStats[j].num_non_full_runs += num_non_full_runs;
+          aBinStats[j].num_runs += bin->mNumRuns;
+          aBinStats[j].bytes_unused += bin_unused;
+          size_t bytes_per_run = static_cast<size_t>(bin->mRunSizePages)
+                                 << gPageSize2Pow;
+          aBinStats[j].bytes_total +=
+              bin->mNumRuns * (bytes_per_run - bin->mRunFirstRegionOffset);
+          aBinStats[j].bytes_per_run = bytes_per_run;
+        }
+      }
+    }
+
+    MOZ_ASSERT(arena_mapped >= arena_committed);
+    MOZ_ASSERT(arena_committed >= arena_allocated + arena_dirty);
+
+    aStats->mapped += arena_mapped;
+    aStats->allocated += arena_allocated;
+    aStats->page_cache += arena_dirty;
+    // "waste" is committed memory that is neither dirty nor
+    // allocated.  If you change this definition please update
+    // memory/replace/logalloc/replay/Replay.cpp's jemalloc_stats calculation of
+    // committed.
+    aStats->waste += arena_committed - arena_allocated - arena_dirty -
+                     arena_unused - arena_headers;
+    aStats->bin_unused += arena_unused;
+    aStats->bookkeeping += arena_headers;
+    aStats->narenas++;
+  }
+  gArenas.mLock.Unlock();
+
+  // Account for arena chunk headers in bookkeeping rather than waste.
+  chunk_header_size =
+      ((aStats->mapped / aStats->chunksize) * gChunkHeaderNumPages)
+      << gPageSize2Pow;
+
+  aStats->mapped += non_arena_mapped;
+  aStats->bookkeeping += chunk_header_size;
+  aStats->waste -= chunk_header_size;
+
+  MOZ_ASSERT(aStats->mapped >= aStats->allocated + aStats->waste +
+                                   aStats->page_cache + aStats->bookkeeping);
+}
+
+template <>
+inline size_t MozJemalloc::jemalloc_stats_num_bins() {
+  return NUM_SMALL_CLASSES;
+}
+
+#ifdef MALLOC_DOUBLE_PURGE
+
+// Explicitly remove all of this chunk's MADV_FREE'd pages from memory.
+static void hard_purge_chunk(arena_chunk_t* aChunk) {
+  // See similar logic in arena_t::Purge().
+  for (size_t i = gChunkHeaderNumPages; i < gChunkNumPages; i++) {
+    // Find all adjacent pages with CHUNK_MAP_MADVISED set.
+    size_t npages;
+    for (npages = 0; aChunk->map[i + npages].bits & CHUNK_MAP_MADVISED &&
+                     i + npages < gChunkNumPages;
+         npages++) {
+      // Turn off the chunk's MADV_FREED bit and turn on its
+      // DECOMMITTED bit.
+      MOZ_DIAGNOSTIC_ASSERT(
+          !(aChunk->map[i + npages].bits & CHUNK_MAP_DECOMMITTED));
+      aChunk->map[i + npages].bits ^= CHUNK_MAP_MADVISED_OR_DECOMMITTED;
+    }
+
+    // We could use mincore to find out which pages are actually
+    // present, but it's not clear that's better.
+    if (npages > 0) {
+      pages_decommit(((char*)aChunk) + (i << gPageSize2Pow),
+                     npages << gPageSize2Pow);
+      Unused << pages_commit(((char*)aChunk) + (i << gPageSize2Pow),
+                             npages << gPageSize2Pow);
+    }
+    i += npages;
+  }
+}
+
+// Explicitly remove all of this arena's MADV_FREE'd pages from memory.
+void arena_t::HardPurge() {
+  MutexAutoLock lock(mLock);
+
+  while (!mChunksMAdvised.isEmpty()) {
+    arena_chunk_t* chunk = mChunksMAdvised.popFront();
+    hard_purge_chunk(chunk);
+  }
+}
+
+template <>
+inline void MozJemalloc::jemalloc_purge_freed_pages() {
+  if (malloc_initialized) {
+    MutexAutoLock lock(gArenas.mLock);
+    for (auto arena : gArenas.iter()) {
+      arena->HardPurge();
+    }
+  }
+}
+
+#else  // !defined MALLOC_DOUBLE_PURGE
+
+template <>
+inline void MozJemalloc::jemalloc_purge_freed_pages() {
+  // Do nothing.
+}
+
+#endif  // defined MALLOC_DOUBLE_PURGE
+
+template <>
+inline void MozJemalloc::jemalloc_free_dirty_pages(void) {
+  if (malloc_initialized) {
+    MutexAutoLock lock(gArenas.mLock);
+    for (auto arena : gArenas.iter()) {
+      MutexAutoLock arena_lock(arena->mLock);
+      arena->Purge(true);
+    }
+  }
+}
+
+inline arena_t* ArenaCollection::GetByIdInternal(arena_id_t aArenaId,
+                                                 bool aIsPrivate) {
+  // Use AlignedStorage2 to avoid running the arena_t constructor, while
+  // we only need it as a placeholder for mId.
+  mozilla::AlignedStorage2<arena_t> key;
+  key.addr()->mId = aArenaId;
+  return (aIsPrivate ? mPrivateArenas : mArenas).Search(key.addr());
+}
+
+inline arena_t* ArenaCollection::GetById(arena_id_t aArenaId, bool aIsPrivate) {
+  if (!malloc_initialized) {
+    return nullptr;
+  }
+
+  MutexAutoLock lock(mLock);
+  arena_t* result = GetByIdInternal(aArenaId, aIsPrivate);
+  MOZ_RELEASE_ASSERT(result);
+  return result;
+}
+
+template <>
+inline arena_id_t MozJemalloc::moz_create_arena_with_params(
+    arena_params_t* aParams) {
+  if (malloc_init()) {
+    arena_t* arena = gArenas.CreateArena(/* IsPrivate = */ true, aParams);
+    return arena->mId;
+  }
+  return 0;
+}
+
+template <>
+inline void MozJemalloc::moz_dispose_arena(arena_id_t aArenaId) {
+  arena_t* arena = gArenas.GetById(aArenaId, /* IsPrivate = */ true);
+  MOZ_RELEASE_ASSERT(arena);
+  gArenas.DisposeArena(arena);
+}
+
+#define MALLOC_DECL(name, return_type, ...)                          \
+  template <>                                                        \
+  inline return_type MozJemalloc::moz_arena_##name(                  \
+      arena_id_t aArenaId, ARGS_HELPER(TYPED_ARGS, ##__VA_ARGS__)) { \
+    BaseAllocator allocator(                                         \
+        gArenas.GetById(aArenaId, /* IsPrivate = */ true));          \
+    return allocator.name(ARGS_HELPER(ARGS, ##__VA_ARGS__));         \
+  }
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#include "malloc_decls.h"
+
+// End non-standard functions.
+// ***************************************************************************
+#ifndef XP_WIN
+// Begin library-private functions, used by threading libraries for protection
+// of malloc during fork().  These functions are only called if the program is
+// running in threaded mode, so there is no need to check whether the program
+// is threaded here.
+FORK_HOOK
+void _malloc_prefork(void) MOZ_NO_THREAD_SAFETY_ANALYSIS {
+  // Acquire all mutexes in a safe order.
+  gArenas.mLock.Lock();
+
+  for (auto arena : gArenas.iter()) {
+    arena->mLock.Lock();
+  }
+
+  base_mtx.Lock();
+
+  huge_mtx.Lock();
+}
+
+FORK_HOOK
+void _malloc_postfork_parent(void) MOZ_NO_THREAD_SAFETY_ANALYSIS {
+  // Release all mutexes, now that fork() has completed.
+  huge_mtx.Unlock();
+
+  base_mtx.Unlock();
+
+  for (auto arena : gArenas.iter()) {
+    arena->mLock.Unlock();
+  }
+
+  gArenas.mLock.Unlock();
+}
+
+FORK_HOOK
+void _malloc_postfork_child(void) {
+  // Reinitialize all mutexes, now that fork() has completed.
+  huge_mtx.Init();
+
+  base_mtx.Init();
+
+  for (auto arena : gArenas.iter()) {
+    arena->mLock.Init();
+  }
+
+  gArenas.mLock.Init();
+}
+#endif  // XP_WIN
+
+// End library-private functions.
+// ***************************************************************************
+#ifdef MOZ_REPLACE_MALLOC
+// Windows doesn't come with weak imports as they are possible with
+// LD_PRELOAD or DYLD_INSERT_LIBRARIES on Linux/OSX. On this platform,
+// the replacement functions are defined as variable pointers to the
+// function resolved with GetProcAddress() instead of weak definitions
+// of functions. On Android, the same needs to happen as well, because
+// the Android linker doesn't handle weak linking with non LD_PRELOADed
+// libraries, but LD_PRELOADing is not very convenient on Android, with
+// the zygote.
+#  ifdef XP_DARWIN
+#    define MOZ_REPLACE_WEAK __attribute__((weak_import))
+#  elif defined(XP_WIN) || defined(ANDROID)
+#    define MOZ_DYNAMIC_REPLACE_INIT
+#    define replace_init replace_init_decl
+#  elif defined(__GNUC__)
+#    define MOZ_REPLACE_WEAK __attribute__((weak))
+#  endif
+
+#  include "replace_malloc.h"
+
+#  define MALLOC_DECL(name, return_type, ...) MozJemalloc::name,
+
+// The default malloc table, i.e. plain allocations. It never changes. It's
+// used by init(), and not used after that.
+static const malloc_table_t gDefaultMallocTable = {
+#  include "malloc_decls.h"
+};
+
+// The malloc table installed by init(). It never changes from that point
+// onward. It will be the same as gDefaultMallocTable if no replace-malloc tool
+// is enabled at startup.
+static malloc_table_t gOriginalMallocTable = {
+#  include "malloc_decls.h"
+};
+
+// The malloc table installed by jemalloc_replace_dynamic(). (Read the
+// comments above that function for more details.)
+static malloc_table_t gDynamicMallocTable = {
+#  include "malloc_decls.h"
+};
+
+// This briefly points to gDefaultMallocTable at startup. After that, it points
+// to either gOriginalMallocTable or gDynamicMallocTable. It's atomic to avoid
+// races when switching between tables.
+static Atomic<malloc_table_t const*, mozilla::MemoryOrdering::Relaxed>
+    gMallocTablePtr;
+
+#  ifdef MOZ_DYNAMIC_REPLACE_INIT
+#    undef replace_init
+typedef decltype(replace_init_decl) replace_init_impl_t;
+static replace_init_impl_t* replace_init = nullptr;
+#  endif
+
+#  ifdef XP_WIN
+typedef HMODULE replace_malloc_handle_t;
+
+static replace_malloc_handle_t replace_malloc_handle() {
+  wchar_t replace_malloc_lib[1024];
+  if (GetEnvironmentVariableW(L"MOZ_REPLACE_MALLOC_LIB", replace_malloc_lib,
+                              ArrayLength(replace_malloc_lib)) > 0) {
+    return LoadLibraryW(replace_malloc_lib);
+  }
+  return nullptr;
+}
+
+#    define REPLACE_MALLOC_GET_INIT_FUNC(handle) \
+      (replace_init_impl_t*)GetProcAddress(handle, "replace_init")
+
+#  elif defined(ANDROID)
+#    include <dlfcn.h>
+
+typedef void* replace_malloc_handle_t;
+
+static replace_malloc_handle_t replace_malloc_handle() {
+  const char* replace_malloc_lib = getenv("MOZ_REPLACE_MALLOC_LIB");
+  if (replace_malloc_lib && *replace_malloc_lib) {
+    return dlopen(replace_malloc_lib, RTLD_LAZY);
+  }
+  return nullptr;
+}
+
+#    define REPLACE_MALLOC_GET_INIT_FUNC(handle) \
+      (replace_init_impl_t*)dlsym(handle, "replace_init")
+
+#  endif
+
+static void replace_malloc_init_funcs(malloc_table_t*);
+
+#  ifdef MOZ_REPLACE_MALLOC_STATIC
+extern "C" void logalloc_init(malloc_table_t*, ReplaceMallocBridge**);
+
+extern "C" void dmd_init(malloc_table_t*, ReplaceMallocBridge**);
+
+extern "C" void phc_init(malloc_table_t*, ReplaceMallocBridge**);
+#  endif
+
+bool Equals(const malloc_table_t& aTable1, const malloc_table_t& aTable2) {
+  return memcmp(&aTable1, &aTable2, sizeof(malloc_table_t)) == 0;
+}
+
+// Below is the malloc implementation overriding jemalloc and calling the
+// replacement functions if they exist.
+static ReplaceMallocBridge* gReplaceMallocBridge = nullptr;
+static void init() {
+  malloc_table_t tempTable = gDefaultMallocTable;
+
+#  ifdef MOZ_DYNAMIC_REPLACE_INIT
+  replace_malloc_handle_t handle = replace_malloc_handle();
+  if (handle) {
+    replace_init = REPLACE_MALLOC_GET_INIT_FUNC(handle);
+  }
+#  endif
+
+  // Set this *before* calling replace_init, otherwise if replace_init calls
+  // malloc() we'll get an infinite loop.
+  gMallocTablePtr = &gDefaultMallocTable;
+
+  // Pass in the default allocator table so replace functions can copy and use
+  // it for their allocations. The replace_init() function should modify the
+  // table if it wants to be active, otherwise leave it unmodified.
+  if (replace_init) {
+    replace_init(&tempTable, &gReplaceMallocBridge);
+  }
+#  ifdef MOZ_REPLACE_MALLOC_STATIC
+  if (Equals(tempTable, gDefaultMallocTable)) {
+    logalloc_init(&tempTable, &gReplaceMallocBridge);
+  }
+#    ifdef MOZ_DMD
+  if (Equals(tempTable, gDefaultMallocTable)) {
+    dmd_init(&tempTable, &gReplaceMallocBridge);
+  }
+#    endif
+#    ifdef MOZ_PHC
+  if (Equals(tempTable, gDefaultMallocTable)) {
+    phc_init(&tempTable, &gReplaceMallocBridge);
+  }
+#    endif
+#  endif
+  if (!Equals(tempTable, gDefaultMallocTable)) {
+    replace_malloc_init_funcs(&tempTable);
+  }
+  gOriginalMallocTable = tempTable;
+  gMallocTablePtr = &gOriginalMallocTable;
+}
+
+// WARNING WARNING WARNING: this function should be used with extreme care. It
+// is not as general-purpose as it looks. It is currently used by
+// tools/profiler/core/memory_hooks.cpp for counting allocations and probably
+// should not be used for any other purpose.
+//
+// This function allows the original malloc table to be temporarily replaced by
+// a different malloc table. Or, if the argument is nullptr, it switches back to
+// the original malloc table.
+//
+// Limitations:
+//
+// - It is not threadsafe. If multiple threads pass it the same
+//   `replace_init_func` at the same time, there will be data races writing to
+//   the malloc_table_t within that function.
+//
+// - Only one replacement can be installed. No nesting is allowed.
+//
+// - The new malloc table must be able to free allocations made by the original
+//   malloc table, and upon removal the original malloc table must be able to
+//   free allocations made by the new malloc table. This means the new malloc
+//   table can only do simple things like recording extra information, while
+//   delegating actual allocation/free operations to the original malloc table.
+//
+MOZ_JEMALLOC_API void jemalloc_replace_dynamic(
+    jemalloc_init_func replace_init_func) {
+  if (replace_init_func) {
+    malloc_table_t tempTable = gOriginalMallocTable;
+    (*replace_init_func)(&tempTable, &gReplaceMallocBridge);
+    if (!Equals(tempTable, gOriginalMallocTable)) {
+      replace_malloc_init_funcs(&tempTable);
+
+      // Temporarily switch back to the original malloc table. In the
+      // (supported) non-nested case, this is a no-op. But just in case this is
+      // a (unsupported) nested call, it makes the overwriting of
+      // gDynamicMallocTable less racy, because ongoing calls to malloc() and
+      // friends won't go through gDynamicMallocTable.
+      gMallocTablePtr = &gOriginalMallocTable;
+
+      gDynamicMallocTable = tempTable;
+      gMallocTablePtr = &gDynamicMallocTable;
+      // We assume that dynamic replaces don't occur close enough for a
+      // thread to still have old copies of the table pointer when the 2nd
+      // replace occurs.
+    }
+  } else {
+    // Switch back to the original malloc table.
+    gMallocTablePtr = &gOriginalMallocTable;
+  }
+}
+
+#  define MALLOC_DECL(name, return_type, ...)                           \
+    template <>                                                         \
+    inline return_type ReplaceMalloc::name(                             \
+        ARGS_HELPER(TYPED_ARGS, ##__VA_ARGS__)) {                       \
+      if (MOZ_UNLIKELY(!gMallocTablePtr)) {                             \
+        init();                                                         \
+      }                                                                 \
+      return (*gMallocTablePtr).name(ARGS_HELPER(ARGS, ##__VA_ARGS__)); \
+    }
+#  include "malloc_decls.h"
+
+MOZ_JEMALLOC_API struct ReplaceMallocBridge* get_bridge(void) {
+  if (MOZ_UNLIKELY(!gMallocTablePtr)) {
+    init();
+  }
+  return gReplaceMallocBridge;
+}
+
+// posix_memalign, aligned_alloc, memalign and valloc all implement some kind
+// of aligned memory allocation. For convenience, a replace-malloc library can
+// skip defining replace_posix_memalign, replace_aligned_alloc and
+// replace_valloc, and default implementations will be automatically derived
+// from replace_memalign.
+static void replace_malloc_init_funcs(malloc_table_t* table) {
+  if (table->posix_memalign == MozJemalloc::posix_memalign &&
+      table->memalign != MozJemalloc::memalign) {
+    table->posix_memalign =
+        AlignedAllocator<ReplaceMalloc::memalign>::posix_memalign;
+  }
+  if (table->aligned_alloc == MozJemalloc::aligned_alloc &&
+      table->memalign != MozJemalloc::memalign) {
+    table->aligned_alloc =
+        AlignedAllocator<ReplaceMalloc::memalign>::aligned_alloc;
+  }
+  if (table->valloc == MozJemalloc::valloc &&
+      table->memalign != MozJemalloc::memalign) {
+    table->valloc = AlignedAllocator<ReplaceMalloc::memalign>::valloc;
+  }
+  if (table->moz_create_arena_with_params ==
+          MozJemalloc::moz_create_arena_with_params &&
+      table->malloc != MozJemalloc::malloc) {
+#  define MALLOC_DECL(name, ...) \
+    table->name = DummyArenaAllocator<ReplaceMalloc>::name;
+#  define MALLOC_FUNCS MALLOC_FUNCS_ARENA_BASE
+#  include "malloc_decls.h"
+  }
+  if (table->moz_arena_malloc == MozJemalloc::moz_arena_malloc &&
+      table->malloc != MozJemalloc::malloc) {
+#  define MALLOC_DECL(name, ...) \
+    table->name = DummyArenaAllocator<ReplaceMalloc>::name;
+#  define MALLOC_FUNCS MALLOC_FUNCS_ARENA_ALLOC
+#  include "malloc_decls.h"
+  }
+}
+
+#endif  // MOZ_REPLACE_MALLOC
+// ***************************************************************************
+// Definition of all the _impl functions
+// GENERIC_MALLOC_DECL2_MINGW is only used for the MinGW build, and aliases
+// the malloc funcs (e.g. malloc) to the je_ versions. It does not generate
+// aliases for the other functions (jemalloc and arena functions).
+//
+// We do need aliases for the other mozglue.def-redirected functions though,
+// these are done at the bottom of mozmemory_wrap.cpp
+#define GENERIC_MALLOC_DECL2_MINGW(name, name_impl, return_type, ...) \
+  return_type name(ARGS_HELPER(TYPED_ARGS, ##__VA_ARGS__))            \
+      __attribute__((alias(MOZ_STRINGIFY(name_impl))));
+
+#define GENERIC_MALLOC_DECL2(attributes, name, name_impl, return_type, ...)  \
+  return_type name_impl(ARGS_HELPER(TYPED_ARGS, ##__VA_ARGS__)) attributes { \
+    return DefaultMalloc::name(ARGS_HELPER(ARGS, ##__VA_ARGS__));            \
+  }
+
+#ifndef __MINGW32__
+#  define GENERIC_MALLOC_DECL(attributes, name, return_type, ...)    \
+    GENERIC_MALLOC_DECL2(attributes, name, name##_impl, return_type, \
+                         ##__VA_ARGS__)
+#else
+#  define GENERIC_MALLOC_DECL(attributes, name, return_type, ...)    \
+    GENERIC_MALLOC_DECL2(attributes, name, name##_impl, return_type, \
+                         ##__VA_ARGS__)                              \
+    GENERIC_MALLOC_DECL2_MINGW(name, name##_impl, return_type, ##__VA_ARGS__)
+#endif
+
+#define NOTHROW_MALLOC_DECL(...) \
+  MOZ_MEMORY_API MACRO_CALL(GENERIC_MALLOC_DECL, (noexcept(true), __VA_ARGS__))
+#define MALLOC_DECL(...) \
+  MOZ_MEMORY_API MACRO_CALL(GENERIC_MALLOC_DECL, (, __VA_ARGS__))
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC
+#include "malloc_decls.h"
+
+#undef GENERIC_MALLOC_DECL
+#define GENERIC_MALLOC_DECL(attributes, name, return_type, ...) \
+  GENERIC_MALLOC_DECL2(attributes, name, name, return_type, ##__VA_ARGS__)
+
+#define MALLOC_DECL(...) \
+  MOZ_JEMALLOC_API MACRO_CALL(GENERIC_MALLOC_DECL, (, __VA_ARGS__))
+#define MALLOC_FUNCS (MALLOC_FUNCS_JEMALLOC | MALLOC_FUNCS_ARENA)
+#include "malloc_decls.h"
+// ***************************************************************************
+
+#ifdef HAVE_DLOPEN
+#  include <dlfcn.h>
+#endif
+
+#if defined(__GLIBC__) && !defined(__UCLIBC__)
+// glibc provides the RTLD_DEEPBIND flag for dlopen which can make it possible
+// to inconsistently reference libc's malloc(3)-compatible functions
+// (bug 493541).
+//
+// These definitions interpose hooks in glibc.  The functions are actually
+// passed an extra argument for the caller return address, which will be
+// ignored.
+
+extern "C" {
+MOZ_EXPORT void (*__free_hook)(void*) = free_impl;
+MOZ_EXPORT void* (*__malloc_hook)(size_t) = malloc_impl;
+MOZ_EXPORT void* (*__realloc_hook)(void*, size_t) = realloc_impl;
+MOZ_EXPORT void* (*__memalign_hook)(size_t, size_t) = memalign_impl;
+}
+
+#elif defined(RTLD_DEEPBIND)
+// XXX On systems that support RTLD_GROUP or DF_1_GROUP, do their
+// implementations permit similar inconsistencies?  Should STV_SINGLETON
+// visibility be used for interposition where available?
+#  error \
+      "Interposing malloc is unsafe on this system without libc malloc hooks."
+#endif
+
+#ifdef XP_WIN
+MOZ_EXPORT void* _recalloc(void* aPtr, size_t aCount, size_t aSize) {
+  size_t oldsize = aPtr ? AllocInfo::Get(aPtr).Size() : 0;
+  CheckedInt<size_t> checkedSize = CheckedInt<size_t>(aCount) * aSize;
+
+  if (!checkedSize.isValid()) {
+    return nullptr;
+  }
+
+  size_t newsize = checkedSize.value();
+
+  // In order for all trailing bytes to be zeroed, the caller needs to
+  // use calloc(), followed by recalloc().  However, the current calloc()
+  // implementation only zeros the bytes requested, so if recalloc() is
+  // to work 100% correctly, calloc() will need to change to zero
+  // trailing bytes.
+  aPtr = DefaultMalloc::realloc(aPtr, newsize);
+  if (aPtr && oldsize < newsize) {
+    memset((void*)((uintptr_t)aPtr + oldsize), 0, newsize - oldsize);
+  }
+
+  return aPtr;
+}
+
+// This impl of _expand doesn't ever actually expand or shrink blocks: it
+// simply replies that you may continue using a shrunk block.
+MOZ_EXPORT void* _expand(void* aPtr, size_t newsize) {
+  if (AllocInfo::Get(aPtr).Size() >= newsize) {
+    return aPtr;
+  }
+
+  return nullptr;
+}
+
+MOZ_EXPORT size_t _msize(void* aPtr) {
+  return DefaultMalloc::malloc_usable_size(aPtr);
+}
+#endif
diff --git a/memory/build/mozjemalloc.h b/memory/build/mozjemalloc.h
new file mode 100644
index 0000000000..60700756f3
--- /dev/null
+++ b/memory/build/mozjemalloc.h
@@ -0,0 +1,77 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozjemalloc_h
+#define mozjemalloc_h
+
+#include "mozjemalloc_types.h"
+#include "mozilla/MacroArgs.h"
+
+// Macro helpers
+
+#define MACRO_CALL(a, b) a b
+// Can't use macros recursively, so we need another one doing the same as above.
+#define MACRO_CALL2(a, b) a b
+
+#define ARGS_HELPER(name, ...)                                     \
+  MACRO_CALL2(MOZ_PASTE_PREFIX_AND_ARG_COUNT(name, ##__VA_ARGS__), \
+              (__VA_ARGS__))
+#define TYPED_ARGS0()
+#define TYPED_ARGS1(t1) t1 arg1
+#define TYPED_ARGS2(t1, t2) TYPED_ARGS1(t1), t2 arg2
+#define TYPED_ARGS3(t1, t2, t3) TYPED_ARGS2(t1, t2), t3 arg3
+
+#define ARGS0()
+#define ARGS1(t1) arg1
+#define ARGS2(t1, t2) ARGS1(t1), arg2
+#define ARGS3(t1, t2, t3) ARGS2(t1, t2), arg3
+
+#ifdef MOZ_MEMORY
+
+// Generic interface exposing the whole public allocator API
+// This facilitates the implementation of things like replace-malloc.
+// Note: compilers are expected to be able to optimize out `this`.
+template <typename T>
+struct Allocator : public T {
+#  define MALLOC_DECL(name, return_type, ...) \
+    static return_type name(__VA_ARGS__);
+#  include "malloc_decls.h"
+};
+
+// The MozJemalloc allocator
+struct MozJemallocBase {};
+typedef Allocator<MozJemallocBase> MozJemalloc;
+
+#  ifdef MOZ_REPLACE_MALLOC
+// The replace-malloc allocator
+struct ReplaceMallocBase {};
+typedef Allocator<ReplaceMallocBase> ReplaceMalloc;
+
+typedef ReplaceMalloc DefaultMalloc;
+#  else
+typedef MozJemalloc DefaultMalloc;
+#  endif
+
+#endif  // MOZ_MEMORY
+
+// Dummy implementation of the moz_arena_* API, falling back to a given
+// implementation of the base allocator.
+template <typename T>
+struct DummyArenaAllocator {
+  static arena_id_t moz_create_arena_with_params(arena_params_t*) { return 0; }
+
+  static void moz_dispose_arena(arena_id_t) {}
+
+#define MALLOC_DECL(name, return_type, ...)                 \
+  static return_type moz_arena_##name(                      \
+      arena_id_t, ARGS_HELPER(TYPED_ARGS, ##__VA_ARGS__)) { \
+    return T::name(ARGS_HELPER(ARGS, ##__VA_ARGS__));       \
+  }
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#include "malloc_decls.h"
+};
+
+#endif
diff --git a/memory/build/mozjemalloc_types.h b/memory/build/mozjemalloc_types.h
new file mode 100644
index 0000000000..80d5493f4a
--- /dev/null
+++ b/memory/build/mozjemalloc_types.h
@@ -0,0 +1,181 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// Portions of this file were originally under the following license:
+//
+// Copyright (C) 2006-2008 Jason Evans <jasone@FreeBSD.org>.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+// 1. Redistributions of source code must retain the above copyright
+//    notice(s), this list of conditions and the following disclaimer as
+//    the first lines of this file unmodified other than the possible
+//    addition of one or more copyright notices.
+// 2. Redistributions in binary form must reproduce the above copyright
+//    notice(s), this list of conditions and the following disclaimer in
+//    the documentation and/or other materials provided with the
+//    distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE
+// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
+// BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+// OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
+// EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#ifndef _JEMALLOC_TYPES_H_
+#define _JEMALLOC_TYPES_H_
+
+#include <stdint.h>
+
+// grab size_t
+#ifdef _MSC_VER
+#  include <crtdefs.h>
+#else
+#  include <stddef.h>
+#endif
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef MALLOC_USABLE_SIZE_CONST_PTR
+#  define MALLOC_USABLE_SIZE_CONST_PTR const
+#endif
+
+typedef MALLOC_USABLE_SIZE_CONST_PTR void* usable_ptr_t;
+
+typedef size_t arena_id_t;
+
+#define ARENA_FLAG_RANDOMIZE_SMALL_MASK 0x3
+#define ARENA_FLAG_RANDOMIZE_SMALL_DEFAULT 0
+#define ARENA_FLAG_RANDOMIZE_SMALL_ENABLED 1
+#define ARENA_FLAG_RANDOMIZE_SMALL_DISABLED 2
+
+typedef struct arena_params_s {
+  size_t mMaxDirty;
+  uint32_t mFlags;
+
+#ifdef __cplusplus
+  arena_params_s() : mMaxDirty(0), mFlags(0) {}
+#endif
+} arena_params_t;
+
+// jemalloc_stats() is not a stable interface.  When using jemalloc_stats_t, be
+// sure that the compiled results of jemalloc.c are in sync with this header
+// file.
+typedef struct {
+  // Run-time configuration settings.
+  bool opt_junk;            // Fill allocated memory with kAllocJunk?
+  bool opt_zero;            // Fill allocated memory with 0x0?
+  size_t narenas;           // Number of arenas.
+  size_t quantum;           // Allocation quantum.
+  size_t quantum_max;       // Max quantum-spaced allocation size.
+  size_t quantum_wide;      // Allocation quantum (QuantuWide).
+  size_t quantum_wide_max;  // Max quantum-wide-spaced allocation size.
+  size_t subpage_max;       // Max subpage allocation size.
+  size_t large_max;         // Max sub-chunksize allocation size.
+  size_t chunksize;         // Size of each virtual memory mapping.
+  size_t page_size;         // Size of pages.
+  size_t dirty_max;         // Max dirty pages per arena.
+
+  // Current memory usage statistics.
+  size_t mapped;       // Bytes mapped (not necessarily committed).
+  size_t allocated;    // Bytes allocated (committed, in use by application).
+  size_t waste;        // Bytes committed, not in use by the
+                       // application, and not intentionally left
+                       // unused (i.e., not dirty).
+  size_t page_cache;   // Committed, unused pages kept around as a
+                       // cache.  (jemalloc calls these "dirty".)
+  size_t bookkeeping;  // Committed bytes used internally by the
+                       // allocator.
+  size_t bin_unused;   // Bytes committed to a bin but currently unused.
+} jemalloc_stats_t;
+
+typedef struct {
+  size_t size;               // The size of objects in this bin, zero if this
+                             // bin stats array entry is unused (no more bins).
+  size_t num_non_full_runs;  // The number of non-full runs
+  size_t num_runs;           // The number of runs in this bin
+  size_t bytes_unused;       // The unallocated bytes across all these bins
+  size_t bytes_total;        // The total storage area for runs in this bin,
+  size_t bytes_per_run;      // The number of bytes per run, including headers.
+} jemalloc_bin_stats_t;
+
+enum PtrInfoTag {
+  // The pointer is not currently known to the allocator.
+  // 'addr', 'size', and 'arenaId' are always 0.
+  TagUnknown,
+
+  // The pointer is within a live allocation.
+  // 'addr', 'size', and 'arenaId' describe the allocation.
+  TagLiveAlloc,
+
+  // The pointer is within a small freed allocation.
+  // 'addr', 'size', and 'arenaId' describe the allocation.
+  TagFreedAlloc,
+
+  // The pointer is within a freed page. Details about the original
+  // allocation, including its size, are not available.
+  // 'addr', 'size', and 'arenaId' describe the page.
+  TagFreedPage,
+};
+
+// The information in jemalloc_ptr_info_t could be represented in a variety of
+// ways. The chosen representation has the following properties.
+// - The number of fields is minimized.
+// - The 'tag' field unambiguously defines the meaning of the subsequent fields.
+// Helper functions are used to group together related categories of tags.
+typedef struct jemalloc_ptr_info_s {
+  enum PtrInfoTag tag;
+  void* addr;   // meaning depends on tag; see above
+  size_t size;  // meaning depends on tag; see above
+
+#ifdef MOZ_DEBUG
+  arena_id_t arenaId;  // meaning depends on tag; see above
+#endif
+
+#ifdef __cplusplus
+  jemalloc_ptr_info_s() = default;
+  jemalloc_ptr_info_s(enum PtrInfoTag aTag, void* aAddr, size_t aSize,
+                      arena_id_t aArenaId)
+      : tag(aTag),
+        addr(aAddr),
+        size(aSize)
+#  ifdef MOZ_DEBUG
+        ,
+        arenaId(aArenaId)
+#  endif
+  {
+  }
+#endif
+} jemalloc_ptr_info_t;
+
+static inline bool jemalloc_ptr_is_live(jemalloc_ptr_info_t* info) {
+  return info->tag == TagLiveAlloc;
+}
+
+static inline bool jemalloc_ptr_is_freed(jemalloc_ptr_info_t* info) {
+  return info->tag == TagFreedAlloc || info->tag == TagFreedPage;
+}
+
+static inline bool jemalloc_ptr_is_freed_page(jemalloc_ptr_info_t* info) {
+  return info->tag == TagFreedPage;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // _JEMALLOC_TYPES_H_
diff --git a/memory/build/mozmemory.h b/memory/build/mozmemory.h
new file mode 100644
index 0000000000..e6ed1a352e
--- /dev/null
+++ b/memory/build/mozmemory.h
@@ -0,0 +1,79 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozmemory_h
+#define mozmemory_h
+
+// This header is meant to be used when the following functions are
+// necessary:
+//   - malloc_good_size (used to be called je_malloc_usable_in_advance)
+//   - jemalloc_stats
+//   - jemalloc_stats_num_bins
+//   - jemalloc_purge_freed_pages
+//   - jemalloc_free_dirty_pages
+//   - jemalloc_thread_local_arena
+//   - jemalloc_ptr_info
+
+#ifdef MALLOC_H
+#  include MALLOC_H
+#endif
+#include "mozmemory_wrap.h"
+#include "mozilla/Attributes.h"
+#include "mozilla/Types.h"
+#include "mozjemalloc_types.h"
+#include "stdbool.h"
+
+#ifdef MOZ_MEMORY
+// On OSX, malloc/malloc.h contains the declaration for malloc_good_size,
+// which will call back in jemalloc, through the zone allocator so just use it.
+#  ifndef XP_DARWIN
+MOZ_MEMORY_API size_t malloc_good_size_impl(size_t size);
+
+// Note: the MOZ_GLUE_IN_PROGRAM ifdef below is there to avoid -Werror turning
+// the protective if into errors. MOZ_GLUE_IN_PROGRAM is what triggers MFBT_API
+// to use weak imports.
+static inline size_t _malloc_good_size(size_t size) {
+#    if defined(MOZ_GLUE_IN_PROGRAM) && !defined(IMPL_MFBT)
+  if (!malloc_good_size) return size;
+#    endif
+  return malloc_good_size_impl(size);
+}
+
+#    define malloc_good_size _malloc_good_size
+#  endif
+
+#  define MALLOC_DECL(name, return_type, ...) \
+    MOZ_JEMALLOC_API return_type name(__VA_ARGS__);
+#  define MALLOC_FUNCS MALLOC_FUNCS_JEMALLOC
+#  include "malloc_decls.h"
+
+#  ifdef __cplusplus
+static inline void jemalloc_stats(jemalloc_stats_t* aStats,
+                                  jemalloc_bin_stats_t* aBinStats = nullptr) {
+  jemalloc_stats_internal(aStats, aBinStats);
+}
+#  else
+static inline void jemalloc_stats(jemalloc_stats_t* aStats) {
+  jemalloc_stats_internal(aStats, NULL);
+}
+#  endif
+
+#endif  // MOZ_MEMORY
+
+#define NOTHROW_MALLOC_DECL(name, return_type, ...) \
+  MOZ_JEMALLOC_API return_type name(__VA_ARGS__) noexcept(true);
+#define MALLOC_DECL(name, return_type, ...) \
+  MOZ_JEMALLOC_API return_type name(__VA_ARGS__);
+#define MALLOC_FUNCS MALLOC_FUNCS_ARENA
+#include "malloc_decls.h"
+
+#ifdef __cplusplus
+#  define moz_create_arena() moz_create_arena_with_params(nullptr)
+#else
+#  define moz_create_arena() moz_create_arena_with_params(NULL)
+#endif
+
+#endif  // mozmemory_h
diff --git a/memory/build/mozmemory_wrap.cpp b/memory/build/mozmemory_wrap.cpp
new file mode 100644
index 0000000000..3420fd9693
--- /dev/null
+++ b/memory/build/mozmemory_wrap.cpp
@@ -0,0 +1,145 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <string.h>
+#include "mozmemory_wrap.h"
+#include "mozilla/Types.h"
+
+// Declare malloc implementation functions with the right return and
+// argument types.
+#define NOTHROW_MALLOC_DECL(name, return_type, ...) \
+  MOZ_MEMORY_API return_type name##_impl(__VA_ARGS__) noexcept(true);
+#define MALLOC_DECL(name, return_type, ...) \
+  MOZ_MEMORY_API return_type name##_impl(__VA_ARGS__);
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC
+#include "malloc_decls.h"
+
+// strndup and strdup may be defined as macros in string.h, which would
+// clash with the definitions below.
+#undef strndup
+#undef strdup
+
+MOZ_MEMORY_API char* strndup_impl(const char* src, size_t len) {
+  char* dst = (char*)malloc_impl(len + 1);
+  if (dst) {
+    strncpy(dst, src, len);
+    dst[len] = '\0';
+  }
+  return dst;
+}
+
+MOZ_MEMORY_API char* strdup_impl(const char* src) {
+  size_t len = strlen(src);
+  return strndup_impl(src, len);
+}
+
+#ifdef ANDROID
+#  include <stdarg.h>
+#  include <stdio.h>
+
+MOZ_MEMORY_API int vasprintf_impl(char** str, const char* fmt, va_list ap) {
+  char *ptr, *_ptr;
+  int ret;
+
+  if (str == NULL || fmt == NULL) {
+    return -1;
+  }
+
+  ptr = (char*)malloc_impl(128);
+  if (ptr == NULL) {
+    *str = NULL;
+    return -1;
+  }
+
+  ret = vsnprintf(ptr, 128, fmt, ap);
+  if (ret < 0) {
+    free_impl(ptr);
+    *str = NULL;
+    return -1;
+  }
+
+  _ptr = reinterpret_cast<char*>(realloc_impl(ptr, ret + 1));
+  if (_ptr == NULL) {
+    free_impl(ptr);
+    *str = NULL;
+    return -1;
+  }
+
+  *str = _ptr;
+
+  return ret;
+}
+
+MOZ_MEMORY_API int asprintf_impl(char** str, const char* fmt, ...) {
+  int ret;
+  va_list ap;
+  va_start(ap, fmt);
+
+  ret = vasprintf_impl(str, fmt, ap);
+
+  va_end(ap);
+
+  return ret;
+}
+#endif
+
+#ifdef XP_WIN
+#  include <wchar.h>
+
+// We also need to provide our own impl of wcsdup so that we don't ask
+// the CRT for memory from its heap (which will then be unfreeable).
+MOZ_MEMORY_API wchar_t* wcsdup_impl(const wchar_t* src) {
+  size_t len = wcslen(src);
+  wchar_t* dst = (wchar_t*)malloc_impl((len + 1) * sizeof(wchar_t));
+  if (dst) wcsncpy(dst, src, len + 1);
+  return dst;
+}
+
+MOZ_MEMORY_API void* _aligned_malloc_impl(size_t size, size_t alignment) {
+  return memalign_impl(alignment, size);
+}
+
+#  ifdef __MINGW32__
+MOZ_BEGIN_EXTERN_C
+// As in mozjemalloc.cpp, we generate aliases for functions
+// redirected in mozglue.def
+void* _aligned_malloc(size_t size, size_t alignment)
+    __attribute__((alias(MOZ_STRINGIFY(_aligned_malloc_impl))));
+void _aligned_free(void* aPtr) __attribute__((alias(MOZ_STRINGIFY(free_impl))));
+
+char* strndup(const char* src, size_t len)
+    __attribute__((alias(MOZ_STRINGIFY(strdup_impl))));
+char* strdup(const char* src)
+    __attribute__((alias(MOZ_STRINGIFY(strdup_impl))));
+char* _strdup(const char* src)
+    __attribute__((alias(MOZ_STRINGIFY(strdup_impl))));
+wchar_t* wcsdup(const wchar_t* src)
+    __attribute__((alias(MOZ_STRINGIFY(wcsdup_impl))));
+wchar_t* _wcsdup(const wchar_t* src)
+    __attribute__((alias(MOZ_STRINGIFY(wcsdup_impl))));
+
+// jemalloc has _aligned_malloc, and friends. libc++.a contains
+// references to __imp__aligned_malloc (and friends) because it
+// is declared dllimport in the headers.
+//
+// The linker sees jemalloc's _aligned_malloc symbol in our objects,
+// but then libc++.a comes along and needs __imp__aligned_malloc, which
+// pulls in those parts of libucrt.a (or libmsvcrt.a in practice),
+// which define both __imp__aligned_malloc and _aligned_malloc, and
+// this causes a conflict.  (And repeat for each of the symbols defined
+// here.)
+//
+// The fix is to define not only an _aligned_malloc symbol (via an
+// alias), but also define the __imp__aligned_malloc pointer to it.
+// This prevents those parts of libucrt from being pulled in and causing
+// conflicts.
+// This is done with __MINGW_IMP_SYMBOL to handle x86/x64 differences.
+void (*__MINGW_IMP_SYMBOL(_aligned_free))(void*) = _aligned_free;
+void* (*__MINGW_IMP_SYMBOL(_aligned_malloc))(size_t, size_t) = _aligned_malloc;
+char* (*__MINGW_IMP_SYMBOL(_strdup))(const char* src) = _strdup;
+MOZ_END_EXTERN_C
+#  endif
+#endif  // XP_WIN
diff --git a/memory/build/mozmemory_wrap.h b/memory/build/mozmemory_wrap.h
new file mode 100644
index 0000000000..92d0a1176a
--- /dev/null
+++ b/memory/build/mozmemory_wrap.h
@@ -0,0 +1,166 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozmemory_wrap_h
+#define mozmemory_wrap_h
+
+// This header contains #defines which tweak the names of various memory
+// allocation functions.
+//
+// There are several types of functions related to memory allocation
+// that are meant to be used publicly by the Gecko codebase:
+//
+// - malloc implementation functions:
+//   - malloc
+//   - posix_memalign
+//   - aligned_alloc
+//   - calloc
+//   - realloc
+//   - free
+//   - memalign
+//   - valloc
+//   - malloc_usable_size
+//   - malloc_good_size
+//   Some of these functions are specific to some systems, but for
+//   convenience, they are treated as being cross-platform, and available
+//   as such.
+//
+// - duplication functions:
+//   - strndup
+//   - strdup
+//   - wcsdup (Windows only)
+//
+// - jemalloc specific functions:
+//   - jemalloc_stats
+//   - jemalloc_stats_num_bins
+//   - jemalloc_purge_freed_pages
+//   - jemalloc_free_dirty_pages
+//   - jemalloc_thread_local_arena
+//   - jemalloc_ptr_info
+//   (these functions are native to mozjemalloc)
+//
+// These functions are all exported as part of libmozglue (see
+// $(topsrcdir)/mozglue/build/Makefile.in), with a few implementation
+// peculiarities:
+//
+// - On Windows, the malloc implementation functions are all prefixed with
+//   "je_", the duplication functions are prefixed with "wrap_", and jemalloc
+//   specific functions are left unprefixed. All these functions are however
+//   aliased when exporting them, such that the resulting mozglue.dll exports
+//   them unprefixed (see $(topsrcdir)/mozglue/build/mozglue.def.in). The
+//   prefixed malloc implementation and duplication functions are not
+//   exported.
+//
+// - On MacOSX, the system libc has a zone allocator, which allows us to
+//   hook custom malloc implementation functions without exporting them.
+//   However, since we want things in Firefox to skip the system zone
+//   allocator, the malloc implementation functions are all exported
+//   unprefixed, as well as duplication functions.
+//   Jemalloc-specific functions are also left unprefixed.
+//
+// - On Android all functions are left unprefixed.
+//
+// - On other systems (mostly Linux), all functions are left unprefixed.
+//
+// On all platforms, C++ allocation functions are also exported.
+//
+// Proper exporting of the various functions is done with the MOZ_MEMORY_API
+// and MOZ_JEMALLOC_API macros. MOZ_MEMORY_API is meant to be used for malloc
+// implementation and duplication functions, while MOZ_JEMALLOC_API is
+// dedicated to jemalloc specific functions.
+//
+//
+// All these functions are meant to be called with no prefix from Gecko code.
+// In most cases, this is because that's how they are available at runtime.
+// However, on Android, this relies on faulty.lib (the custom dynamic linker)
+// resolving mozglue symbols before libc symbols, which is guaranteed by the
+// way faulty.lib works (it respects the DT_NEEDED order, and libc always
+// appears after mozglue ; which we double check when building anyways)
+//
+//
+// Within libmozglue (when MOZ_MEMORY_IMPL is defined), all the functions
+// should be suffixed with "_impl" both for declarations and use.
+// That is, the implementation declaration for e.g. strdup would look like:
+//   char* strdup_impl(const char *)
+// That implementation would call malloc by using "malloc_impl".
+
+#if defined(MOZ_MEMORY_IMPL) && !defined(IMPL_MFBT)
+#  ifdef MFBT_API  // mozilla/Types.h was already included
+#    error mozmemory_wrap.h has to be included before mozilla/Types.h when MOZ_MEMORY_IMPL is set and IMPL_MFBT is not.
+#  endif
+#  define IMPL_MFBT
+#endif
+
+#include "mozilla/Types.h"
+
+#ifndef MOZ_EXTERN_C
+#  ifdef __cplusplus
+#    define MOZ_EXTERN_C extern "C"
+#  else
+#    define MOZ_EXTERN_C
+#  endif
+#endif
+
+#ifdef MOZ_MEMORY_IMPL
+#  define MOZ_JEMALLOC_API MOZ_EXTERN_C MFBT_API
+#  if defined(XP_WIN)
+#    define mozmem_malloc_impl(a) je_##a
+#  else
+#    define MOZ_MEMORY_API MOZ_EXTERN_C MFBT_API
+#  endif
+#endif
+#ifdef XP_WIN
+#  define mozmem_dup_impl(a) wrap_##a
+#endif
+
+#if !defined(MOZ_MEMORY_IMPL)
+#  define MOZ_MEMORY_API MOZ_EXTERN_C MFBT_API
+#  define MOZ_JEMALLOC_API MOZ_EXTERN_C MFBT_API
+#endif
+
+#ifndef MOZ_MEMORY_API
+#  define MOZ_MEMORY_API MOZ_EXTERN_C
+#endif
+#ifndef MOZ_JEMALLOC_API
+#  define MOZ_JEMALLOC_API MOZ_EXTERN_C
+#endif
+
+#ifndef mozmem_malloc_impl
+#  define mozmem_malloc_impl(a) a
+#endif
+#ifndef mozmem_dup_impl
+#  define mozmem_dup_impl(a) a
+#endif
+
+// Malloc implementation functions
+#define malloc_impl mozmem_malloc_impl(malloc)
+#define posix_memalign_impl mozmem_malloc_impl(posix_memalign)
+#define aligned_alloc_impl mozmem_malloc_impl(aligned_alloc)
+#define calloc_impl mozmem_malloc_impl(calloc)
+#define realloc_impl mozmem_malloc_impl(realloc)
+#define free_impl mozmem_malloc_impl(free)
+#define memalign_impl mozmem_malloc_impl(memalign)
+#define valloc_impl mozmem_malloc_impl(valloc)
+#define malloc_usable_size_impl mozmem_malloc_impl(malloc_usable_size)
+#define malloc_good_size_impl mozmem_malloc_impl(malloc_good_size)
+
+// Duplication functions
+#define strndup_impl mozmem_dup_impl(strndup)
+#define strdup_impl mozmem_dup_impl(strdup)
+#ifdef XP_WIN
+#  define wcsdup_impl mozmem_dup_impl(wcsdup)
+#  define _aligned_malloc_impl mozmem_dup_impl(_aligned_malloc)
+#endif
+
+// String functions
+#ifdef ANDROID
+// Bug 801571 and Bug 879668, libstagefright uses vasprintf, causing malloc()/
+// free() to be mismatched between bionic and mozglue implementation.
+#  define vasprintf_impl mozmem_dup_impl(vasprintf)
+#  define asprintf_impl mozmem_dup_impl(asprintf)
+#endif
+
+#endif  // mozmemory_wrap_h
diff --git a/memory/build/rb.h b/memory/build/rb.h
new file mode 100644
index 0000000000..418d206911
--- /dev/null
+++ b/memory/build/rb.h
@@ -0,0 +1,741 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// Portions of this file were originally under the following license:
+//
+// Copyright (C) 2008 Jason Evans <jasone@FreeBSD.org>.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+// 1. Redistributions of source code must retain the above copyright
+//    notice(s), this list of conditions and the following disclaimer
+//    unmodified other than the allowable addition of one or more
+//    copyright notices.
+// 2. Redistributions in binary form must reproduce the above copyright
+//    notice(s), this list of conditions and the following disclaimer in
+//    the documentation and/or other materials provided with the
+//    distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE
+// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
+// BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+// OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
+// EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// ****************************************************************************
+//
+// C++ template implementation of left-leaning red-black trees.
+//
+// All operations are done non-recursively.  Parent pointers are not used, and
+// color bits are stored in the least significant bit of right-child pointers,
+// thus making node linkage as compact as is possible for red-black trees.
+//
+// The RedBlackTree template expects two type arguments: the type of the nodes,
+// containing a RedBlackTreeNode, and a trait providing two methods:
+//  - a GetTreeNode method that returns a reference to the RedBlackTreeNode
+//    corresponding to a given node with the following signature:
+//      static RedBlackTreeNode<T>& GetTreeNode(T*)
+//  - a Compare function with the following signature:
+//      static Order Compare(T* aNode, T* aOther)
+//                              ^^^^^
+//                           or aKey
+//
+// Interpretation of comparision function return values:
+//
+//   Order::eLess: aNode <  aOther
+//   Order::eEqual: aNode == aOther
+//   Order::eGreater: aNode >  aOther
+//
+// In all cases, the aNode or aKey argument is the first argument to the
+// comparison function, which makes it possible to write comparison functions
+// that treat the first argument specially.
+//
+// ***************************************************************************
+
+#ifndef RB_H_
+#define RB_H_
+
+#include "mozilla/Alignment.h"
+#include "mozilla/Assertions.h"
+#include "Utils.h"
+
+enum NodeColor {
+  Black = 0,
+  Red = 1,
+};
+
+// Node structure.
+template <typename T>
+class RedBlackTreeNode {
+  T* mLeft;
+  // The lowest bit is the color
+  T* mRightAndColor;
+
+ public:
+  T* Left() { return mLeft; }
+
+  void SetLeft(T* aValue) { mLeft = aValue; }
+
+  T* Right() {
+    return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(mRightAndColor) &
+                                uintptr_t(~1));
+  }
+
+  void SetRight(T* aValue) {
+    mRightAndColor = reinterpret_cast<T*>(
+        (reinterpret_cast<uintptr_t>(aValue) & uintptr_t(~1)) | Color());
+  }
+
+  NodeColor Color() {
+    return static_cast<NodeColor>(reinterpret_cast<uintptr_t>(mRightAndColor) &
+                                  1);
+  }
+
+  bool IsBlack() { return Color() == NodeColor::Black; }
+
+  bool IsRed() { return Color() == NodeColor::Red; }
+
+  void SetColor(NodeColor aColor) {
+    mRightAndColor = reinterpret_cast<T*>(
+        (reinterpret_cast<uintptr_t>(mRightAndColor) & uintptr_t(~1)) | aColor);
+  }
+};
+
+// Tree structure.
+template <typename T, typename Trait>
+class RedBlackTree {
+ public:
+  void Init() { mRoot = nullptr; }
+
+  T* First(T* aStart = nullptr) { return First(TreeNode(aStart)).Get(); }
+
+  T* Last(T* aStart = nullptr) { return Last(TreeNode(aStart)).Get(); }
+
+  T* Next(T* aNode) { return Next(TreeNode(aNode)).Get(); }
+
+  T* Prev(T* aNode) { return Prev(TreeNode(aNode)).Get(); }
+
+  T* Search(T* aKey) { return Search(TreeNode(aKey)).Get(); }
+
+  // Find a match if it exists. Otherwise, find the next greater node, if one
+  // exists.
+  T* SearchOrNext(T* aKey) { return SearchOrNext(TreeNode(aKey)).Get(); }
+
+  void Insert(T* aNode) { Insert(TreeNode(aNode)); }
+
+  void Remove(T* aNode) { Remove(TreeNode(aNode)); }
+
+  // Helper class to avoid having all the tree traversal code further below
+  // have to use Trait::GetTreeNode and do manual null pointer checks, adding
+  // visual noise. Practically speaking TreeNode(nullptr) acts as a virtual
+  // sentinel, that loops back to itself for Left() and Right() and is always
+  // black.
+  class TreeNode {
+   public:
+    constexpr TreeNode() : mNode(nullptr) {}
+
+    MOZ_IMPLICIT TreeNode(T* aNode) : mNode(aNode) {}
+
+    TreeNode& operator=(TreeNode aOther) {
+      mNode = aOther.mNode;
+      return *this;
+    }
+
+    TreeNode Left() {
+      return TreeNode(mNode ? Trait::GetTreeNode(mNode).Left() : nullptr);
+    }
+
+    void SetLeft(TreeNode aNode) {
+      MOZ_RELEASE_ASSERT(mNode);
+      Trait::GetTreeNode(mNode).SetLeft(aNode.mNode);
+    }
+
+    TreeNode Right() {
+      return TreeNode(mNode ? Trait::GetTreeNode(mNode).Right() : nullptr);
+    }
+
+    void SetRight(TreeNode aNode) {
+      MOZ_RELEASE_ASSERT(mNode);
+      Trait::GetTreeNode(mNode).SetRight(aNode.mNode);
+    }
+
+    NodeColor Color() {
+      return mNode ? Trait::GetTreeNode(mNode).Color() : NodeColor::Black;
+    }
+
+    bool IsRed() { return Color() == NodeColor::Red; }
+
+    bool IsBlack() { return Color() == NodeColor::Black; }
+
+    void SetColor(NodeColor aColor) {
+      MOZ_RELEASE_ASSERT(mNode);
+      Trait::GetTreeNode(mNode).SetColor(aColor);
+    }
+
+    T* Get() { return mNode; }
+
+    MOZ_IMPLICIT operator bool() { return !!mNode; }
+
+    bool operator==(TreeNode& aOther) { return mNode == aOther.mNode; }
+
+   private:
+    T* mNode;
+  };
+
+ private:
+  // Ideally we'd use a TreeNode for mRoot, but we need RedBlackTree to stay
+  // a POD type to avoid a static initializer for gArenas.
+  T* mRoot;
+
+  TreeNode First(TreeNode aStart) {
+    TreeNode ret;
+    for (ret = aStart ? aStart : mRoot; ret.Left(); ret = ret.Left()) {
+    }
+    return ret;
+  }
+
+  TreeNode Last(TreeNode aStart) {
+    TreeNode ret;
+    for (ret = aStart ? aStart : mRoot; ret.Right(); ret = ret.Right()) {
+    }
+    return ret;
+  }
+
+  TreeNode Next(TreeNode aNode) {
+    TreeNode ret;
+    if (aNode.Right()) {
+      ret = First(aNode.Right());
+    } else {
+      TreeNode rbp_n_t = mRoot;
+      MOZ_ASSERT(rbp_n_t);
+      ret = nullptr;
+      while (true) {
+        Order rbp_n_cmp = Trait::Compare(aNode.Get(), rbp_n_t.Get());
+        if (rbp_n_cmp == Order::eLess) {
+          ret = rbp_n_t;
+          rbp_n_t = rbp_n_t.Left();
+        } else if (rbp_n_cmp == Order::eGreater) {
+          rbp_n_t = rbp_n_t.Right();
+        } else {
+          break;
+        }
+        MOZ_ASSERT(rbp_n_t);
+      }
+    }
+    return ret;
+  }
+
+  TreeNode Prev(TreeNode aNode) {
+    TreeNode ret;
+    if (aNode.Left()) {
+      ret = Last(aNode.Left());
+    } else {
+      TreeNode rbp_p_t = mRoot;
+      MOZ_ASSERT(rbp_p_t);
+      ret = nullptr;
+      while (true) {
+        Order rbp_p_cmp = Trait::Compare(aNode.Get(), rbp_p_t.Get());
+        if (rbp_p_cmp == Order::eLess) {
+          rbp_p_t = rbp_p_t.Left();
+        } else if (rbp_p_cmp == Order::eGreater) {
+          ret = rbp_p_t;
+          rbp_p_t = rbp_p_t.Right();
+        } else {
+          break;
+        }
+        MOZ_ASSERT(rbp_p_t);
+      }
+    }
+    return ret;
+  }
+
+  TreeNode Search(TreeNode aKey) {
+    TreeNode ret = mRoot;
+    Order rbp_se_cmp;
+    while (ret && (rbp_se_cmp = Trait::Compare(aKey.Get(), ret.Get())) !=
+                      Order::eEqual) {
+      if (rbp_se_cmp == Order::eLess) {
+        ret = ret.Left();
+      } else {
+        ret = ret.Right();
+      }
+    }
+    return ret;
+  }
+
+  TreeNode SearchOrNext(TreeNode aKey) {
+    TreeNode ret = nullptr;
+    TreeNode rbp_ns_t = mRoot;
+    while (rbp_ns_t) {
+      Order rbp_ns_cmp = Trait::Compare(aKey.Get(), rbp_ns_t.Get());
+      if (rbp_ns_cmp == Order::eLess) {
+        ret = rbp_ns_t;
+        rbp_ns_t = rbp_ns_t.Left();
+      } else if (rbp_ns_cmp == Order::eGreater) {
+        rbp_ns_t = rbp_ns_t.Right();
+      } else {
+        ret = rbp_ns_t;
+        break;
+      }
+    }
+    return ret;
+  }
+
+  void Insert(TreeNode aNode) {
+    // rbp_i_s is only used as a placeholder for its RedBlackTreeNode. Use
+    // AlignedStorage2 to avoid running the TreeNode base class constructor.
+    mozilla::AlignedStorage2<T> rbp_i_s;
+    TreeNode rbp_i_g, rbp_i_p, rbp_i_c, rbp_i_t, rbp_i_u;
+    Order rbp_i_cmp = Order::eEqual;
+    rbp_i_g = nullptr;
+    rbp_i_p = rbp_i_s.addr();
+    rbp_i_p.SetLeft(mRoot);
+    rbp_i_p.SetRight(nullptr);
+    rbp_i_p.SetColor(NodeColor::Black);
+    rbp_i_c = mRoot;
+    // Iteratively search down the tree for the insertion point,
+    // splitting 4-nodes as they are encountered. At the end of each
+    // iteration, rbp_i_g->rbp_i_p->rbp_i_c is a 3-level path down
+    // the tree, assuming a sufficiently deep tree.
+    while (rbp_i_c) {
+      rbp_i_t = rbp_i_c.Left();
+      rbp_i_u = rbp_i_t.Left();
+      if (rbp_i_t.IsRed() && rbp_i_u.IsRed()) {
+        // rbp_i_c is the top of a logical 4-node, so split it.
+        // This iteration does not move down the tree, due to the
+        // disruptiveness of node splitting.
+        //
+        // Rotate right.
+        rbp_i_t = RotateRight(rbp_i_c);
+        // Pass red links up one level.
+        rbp_i_u = rbp_i_t.Left();
+        rbp_i_u.SetColor(NodeColor::Black);
+        if (rbp_i_p.Left() == rbp_i_c) {
+          rbp_i_p.SetLeft(rbp_i_t);
+          rbp_i_c = rbp_i_t;
+        } else {
+          // rbp_i_c was the right child of rbp_i_p, so rotate
+          // left in order to maintain the left-leaning invariant.
+          MOZ_ASSERT(rbp_i_p.Right() == rbp_i_c);
+          rbp_i_p.SetRight(rbp_i_t);
+          rbp_i_u = LeanLeft(rbp_i_p);
+          if (rbp_i_g.Left() == rbp_i_p) {
+            rbp_i_g.SetLeft(rbp_i_u);
+          } else {
+            MOZ_ASSERT(rbp_i_g.Right() == rbp_i_p);
+            rbp_i_g.SetRight(rbp_i_u);
+          }
+          rbp_i_p = rbp_i_u;
+          rbp_i_cmp = Trait::Compare(aNode.Get(), rbp_i_p.Get());
+          if (rbp_i_cmp == Order::eLess) {
+            rbp_i_c = rbp_i_p.Left();
+          } else {
+            MOZ_ASSERT(rbp_i_cmp == Order::eGreater);
+            rbp_i_c = rbp_i_p.Right();
+          }
+          continue;
+        }
+      }
+      rbp_i_g = rbp_i_p;
+      rbp_i_p = rbp_i_c;
+      rbp_i_cmp = Trait::Compare(aNode.Get(), rbp_i_c.Get());
+      if (rbp_i_cmp == Order::eLess) {
+        rbp_i_c = rbp_i_c.Left();
+      } else {
+        MOZ_ASSERT(rbp_i_cmp == Order::eGreater);
+        rbp_i_c = rbp_i_c.Right();
+      }
+    }
+    // rbp_i_p now refers to the node under which to insert.
+    aNode.SetLeft(nullptr);
+    aNode.SetRight(nullptr);
+    aNode.SetColor(NodeColor::Red);
+    if (rbp_i_cmp == Order::eGreater) {
+      rbp_i_p.SetRight(aNode);
+      rbp_i_t = LeanLeft(rbp_i_p);
+      if (rbp_i_g.Left() == rbp_i_p) {
+        rbp_i_g.SetLeft(rbp_i_t);
+      } else if (rbp_i_g.Right() == rbp_i_p) {
+        rbp_i_g.SetRight(rbp_i_t);
+      }
+    } else {
+      rbp_i_p.SetLeft(aNode);
+    }
+    // Update the root and make sure that it is black.
+    TreeNode root = TreeNode(rbp_i_s.addr()).Left();
+    root.SetColor(NodeColor::Black);
+    mRoot = root.Get();
+  }
+
+  void Remove(TreeNode aNode) {
+    // rbp_r_s is only used as a placeholder for its RedBlackTreeNode. Use
+    // AlignedStorage2 to avoid running the TreeNode base class constructor.
+    mozilla::AlignedStorage2<T> rbp_r_s;
+    TreeNode rbp_r_p, rbp_r_c, rbp_r_xp, rbp_r_t, rbp_r_u;
+    Order rbp_r_cmp;
+    rbp_r_p = TreeNode(rbp_r_s.addr());
+    rbp_r_p.SetLeft(mRoot);
+    rbp_r_p.SetRight(nullptr);
+    rbp_r_p.SetColor(NodeColor::Black);
+    rbp_r_c = mRoot;
+    rbp_r_xp = nullptr;
+    // Iterate down the tree, but always transform 2-nodes to 3- or
+    // 4-nodes in order to maintain the invariant that the current
+    // node is not a 2-node. This allows simple deletion once a leaf
+    // is reached. Handle the root specially though, since there may
+    // be no way to convert it from a 2-node to a 3-node.
+    rbp_r_cmp = Trait::Compare(aNode.Get(), rbp_r_c.Get());
+    if (rbp_r_cmp == Order::eLess) {
+      rbp_r_t = rbp_r_c.Left();
+      rbp_r_u = rbp_r_t.Left();
+      if (rbp_r_t.IsBlack() && rbp_r_u.IsBlack()) {
+        // Apply standard transform to prepare for left move.
+        rbp_r_t = MoveRedLeft(rbp_r_c);
+        rbp_r_t.SetColor(NodeColor::Black);
+        rbp_r_p.SetLeft(rbp_r_t);
+        rbp_r_c = rbp_r_t;
+      } else {
+        // Move left.
+        rbp_r_p = rbp_r_c;
+        rbp_r_c = rbp_r_c.Left();
+      }
+    } else {
+      if (rbp_r_cmp == Order::eEqual) {
+        MOZ_ASSERT(aNode == rbp_r_c);
+        if (!rbp_r_c.Right()) {
+          // Delete root node (which is also a leaf node).
+          if (rbp_r_c.Left()) {
+            rbp_r_t = LeanRight(rbp_r_c);
+            rbp_r_t.SetRight(nullptr);
+          } else {
+            rbp_r_t = nullptr;
+          }
+          rbp_r_p.SetLeft(rbp_r_t);
+        } else {
+          // This is the node we want to delete, but we will
+          // instead swap it with its successor and delete the
+          // successor. Record enough information to do the
+          // swap later. rbp_r_xp is the aNode's parent.
+          rbp_r_xp = rbp_r_p;
+          rbp_r_cmp = Order::eGreater;  // Note that deletion is incomplete.
+        }
+      }
+      if (rbp_r_cmp == Order::eGreater) {
+        if (rbp_r_c.Right().Left().IsBlack()) {
+          rbp_r_t = rbp_r_c.Left();
+          if (rbp_r_t.IsRed()) {
+            // Standard transform.
+            rbp_r_t = MoveRedRight(rbp_r_c);
+          } else {
+            // Root-specific transform.
+            rbp_r_c.SetColor(NodeColor::Red);
+            rbp_r_u = rbp_r_t.Left();
+            if (rbp_r_u.IsRed()) {
+              rbp_r_u.SetColor(NodeColor::Black);
+              rbp_r_t = RotateRight(rbp_r_c);
+              rbp_r_u = RotateLeft(rbp_r_c);
+              rbp_r_t.SetRight(rbp_r_u);
+            } else {
+              rbp_r_t.SetColor(NodeColor::Red);
+              rbp_r_t = RotateLeft(rbp_r_c);
+            }
+          }
+          rbp_r_p.SetLeft(rbp_r_t);
+          rbp_r_c = rbp_r_t;
+        } else {
+          // Move right.
+          rbp_r_p = rbp_r_c;
+          rbp_r_c = rbp_r_c.Right();
+        }
+      }
+    }
+    if (rbp_r_cmp != Order::eEqual) {
+      while (true) {
+        MOZ_ASSERT(rbp_r_p);
+        rbp_r_cmp = Trait::Compare(aNode.Get(), rbp_r_c.Get());
+        if (rbp_r_cmp == Order::eLess) {
+          rbp_r_t = rbp_r_c.Left();
+          if (!rbp_r_t) {
+            // rbp_r_c now refers to the successor node to
+            // relocate, and rbp_r_xp/aNode refer to the
+            // context for the relocation.
+            if (rbp_r_xp.Left() == aNode) {
+              rbp_r_xp.SetLeft(rbp_r_c);
+            } else {
+              MOZ_ASSERT(rbp_r_xp.Right() == (aNode));
+              rbp_r_xp.SetRight(rbp_r_c);
+            }
+            rbp_r_c.SetLeft(aNode.Left());
+            rbp_r_c.SetRight(aNode.Right());
+            rbp_r_c.SetColor(aNode.Color());
+            if (rbp_r_p.Left() == rbp_r_c) {
+              rbp_r_p.SetLeft(nullptr);
+            } else {
+              MOZ_ASSERT(rbp_r_p.Right() == rbp_r_c);
+              rbp_r_p.SetRight(nullptr);
+            }
+            break;
+          }
+          rbp_r_u = rbp_r_t.Left();
+          if (rbp_r_t.IsBlack() && rbp_r_u.IsBlack()) {
+            rbp_r_t = MoveRedLeft(rbp_r_c);
+            if (rbp_r_p.Left() == rbp_r_c) {
+              rbp_r_p.SetLeft(rbp_r_t);
+            } else {
+              rbp_r_p.SetRight(rbp_r_t);
+            }
+            rbp_r_c = rbp_r_t;
+          } else {
+            rbp_r_p = rbp_r_c;
+            rbp_r_c = rbp_r_c.Left();
+          }
+        } else {
+          // Check whether to delete this node (it has to be
+          // the correct node and a leaf node).
+          if (rbp_r_cmp == Order::eEqual) {
+            MOZ_ASSERT(aNode == rbp_r_c);
+            if (!rbp_r_c.Right()) {
+              // Delete leaf node.
+              if (rbp_r_c.Left()) {
+                rbp_r_t = LeanRight(rbp_r_c);
+                rbp_r_t.SetRight(nullptr);
+              } else {
+                rbp_r_t = nullptr;
+              }
+              if (rbp_r_p.Left() == rbp_r_c) {
+                rbp_r_p.SetLeft(rbp_r_t);
+              } else {
+                rbp_r_p.SetRight(rbp_r_t);
+              }
+              break;
+            }
+            // This is the node we want to delete, but we
+            // will instead swap it with its successor
+            // and delete the successor. Record enough
+            // information to do the swap later.
+            // rbp_r_xp is aNode's parent.
+            rbp_r_xp = rbp_r_p;
+          }
+          rbp_r_t = rbp_r_c.Right();
+          rbp_r_u = rbp_r_t.Left();
+          if (rbp_r_u.IsBlack()) {
+            rbp_r_t = MoveRedRight(rbp_r_c);
+            if (rbp_r_p.Left() == rbp_r_c) {
+              rbp_r_p.SetLeft(rbp_r_t);
+            } else {
+              rbp_r_p.SetRight(rbp_r_t);
+            }
+            rbp_r_c = rbp_r_t;
+          } else {
+            rbp_r_p = rbp_r_c;
+            rbp_r_c = rbp_r_c.Right();
+          }
+        }
+      }
+    }
+    // Update root.
+    mRoot = TreeNode(rbp_r_s.addr()).Left().Get();
+    aNode.SetLeft(nullptr);
+    aNode.SetRight(nullptr);
+    aNode.SetColor(NodeColor::Black);
+  }
+
+  TreeNode RotateLeft(TreeNode aNode) {
+    TreeNode node = aNode.Right();
+    aNode.SetRight(node.Left());
+    node.SetLeft(aNode);
+    return node;
+  }
+
+  TreeNode RotateRight(TreeNode aNode) {
+    TreeNode node = aNode.Left();
+    aNode.SetLeft(node.Right());
+    node.SetRight(aNode);
+    return node;
+  }
+
+  TreeNode LeanLeft(TreeNode aNode) {
+    TreeNode node = RotateLeft(aNode);
+    NodeColor color = aNode.Color();
+    node.SetColor(color);
+    aNode.SetColor(NodeColor::Red);
+    return node;
+  }
+
+  TreeNode LeanRight(TreeNode aNode) {
+    TreeNode node = RotateRight(aNode);
+    NodeColor color = aNode.Color();
+    node.SetColor(color);
+    aNode.SetColor(NodeColor::Red);
+    return node;
+  }
+
+  TreeNode MoveRedLeft(TreeNode aNode) {
+    TreeNode node;
+    TreeNode rbp_mrl_t, rbp_mrl_u;
+    rbp_mrl_t = aNode.Left();
+    rbp_mrl_t.SetColor(NodeColor::Red);
+    rbp_mrl_t = aNode.Right();
+    rbp_mrl_u = rbp_mrl_t.Left();
+    if (rbp_mrl_u.IsRed()) {
+      rbp_mrl_u = RotateRight(rbp_mrl_t);
+      aNode.SetRight(rbp_mrl_u);
+      node = RotateLeft(aNode);
+      rbp_mrl_t = aNode.Right();
+      if (rbp_mrl_t.IsRed()) {
+        rbp_mrl_t.SetColor(NodeColor::Black);
+        aNode.SetColor(NodeColor::Red);
+        rbp_mrl_t = RotateLeft(aNode);
+        node.SetLeft(rbp_mrl_t);
+      } else {
+        aNode.SetColor(NodeColor::Black);
+      }
+    } else {
+      aNode.SetColor(NodeColor::Red);
+      node = RotateLeft(aNode);
+    }
+    return node;
+  }
+
+  TreeNode MoveRedRight(TreeNode aNode) {
+    TreeNode node;
+    TreeNode rbp_mrr_t;
+    rbp_mrr_t = aNode.Left();
+    if (rbp_mrr_t.IsRed()) {
+      TreeNode rbp_mrr_u, rbp_mrr_v;
+      rbp_mrr_u = rbp_mrr_t.Right();
+      rbp_mrr_v = rbp_mrr_u.Left();
+      if (rbp_mrr_v.IsRed()) {
+        rbp_mrr_u.SetColor(aNode.Color());
+        rbp_mrr_v.SetColor(NodeColor::Black);
+        rbp_mrr_u = RotateLeft(rbp_mrr_t);
+        aNode.SetLeft(rbp_mrr_u);
+        node = RotateRight(aNode);
+        rbp_mrr_t = RotateLeft(aNode);
+        node.SetRight(rbp_mrr_t);
+      } else {
+        rbp_mrr_t.SetColor(aNode.Color());
+        rbp_mrr_u.SetColor(NodeColor::Red);
+        node = RotateRight(aNode);
+        rbp_mrr_t = RotateLeft(aNode);
+        node.SetRight(rbp_mrr_t);
+      }
+      aNode.SetColor(NodeColor::Red);
+    } else {
+      rbp_mrr_t.SetColor(NodeColor::Red);
+      rbp_mrr_t = rbp_mrr_t.Left();
+      if (rbp_mrr_t.IsRed()) {
+        rbp_mrr_t.SetColor(NodeColor::Black);
+        node = RotateRight(aNode);
+        rbp_mrr_t = RotateLeft(aNode);
+        node.SetRight(rbp_mrr_t);
+      } else {
+        node = RotateLeft(aNode);
+      }
+    }
+    return node;
+  }
+
+  // The iterator simulates recursion via an array of pointers that store the
+  // current path.  This is critical to performance, since a series of calls to
+  // rb_{next,prev}() would require time proportional to (n lg n), whereas this
+  // implementation only requires time proportional to (n).
+  //
+  // Since the iterator caches a path down the tree, any tree modification may
+  // cause the cached path to become invalid. Don't modify the tree during an
+  // iteration.
+
+  // Size the path arrays such that they are always large enough, even if a
+  // tree consumes all of memory.  Since each node must contain a minimum of
+  // two pointers, there can never be more nodes than:
+  //
+  //   1 << ((sizeof(void*)<<3) - (log2(sizeof(void*))+1))
+  //
+  // Since the depth of a tree is limited to 3*lg(#nodes), the maximum depth
+  // is:
+  //
+  //   (3 * ((sizeof(void*)<<3) - (log2(sizeof(void*))+1)))
+  //
+  // This works out to a maximum depth of 87 and 180 for 32- and 64-bit
+  // systems, respectively (approximately 348 and 1440 bytes, respectively).
+ public:
+  class Iterator {
+    TreeNode mPath[3 * ((sizeof(void*) << 3) - (LOG2(sizeof(void*)) + 1))];
+    unsigned mDepth;
+
+   public:
+    explicit Iterator(RedBlackTree<T, Trait>* aTree) : mDepth(0) {
+      // Initialize the path to contain the left spine.
+      if (aTree->mRoot) {
+        TreeNode node;
+        mPath[mDepth++] = aTree->mRoot;
+        while ((node = mPath[mDepth - 1].Left())) {
+          mPath[mDepth++] = node;
+        }
+      }
+    }
+
+    template <typename Iterator>
+    class Item {
+      Iterator* mIterator;
+      T* mItem;
+
+     public:
+      Item(Iterator* aIterator, T* aItem)
+          : mIterator(aIterator), mItem(aItem) {}
+
+      bool operator!=(const Item& aOther) const {
+        return (mIterator != aOther.mIterator) || (mItem != aOther.mItem);
+      }
+
+      T* operator*() const { return mItem; }
+
+      const Item& operator++() {
+        mItem = mIterator->Next();
+        return *this;
+      }
+    };
+
+    Item<Iterator> begin() {
+      return Item<Iterator>(this,
+                            mDepth > 0 ? mPath[mDepth - 1].Get() : nullptr);
+    }
+
+    Item<Iterator> end() { return Item<Iterator>(this, nullptr); }
+
+    T* Next() {
+      TreeNode node;
+      if ((node = mPath[mDepth - 1].Right())) {
+        // The successor is the left-most node in the right subtree.
+        mPath[mDepth++] = node;
+        while ((node = mPath[mDepth - 1].Left())) {
+          mPath[mDepth++] = node;
+        }
+      } else {
+        // The successor is above the current node.  Unwind until a
+        // left-leaning edge is removed from the path, of the path is empty.
+        for (mDepth--; mDepth > 0; mDepth--) {
+          if (mPath[mDepth - 1].Left() == mPath[mDepth]) {
+            break;
+          }
+        }
+      }
+      return mDepth > 0 ? mPath[mDepth - 1].Get() : nullptr;
+    }
+  };
+
+  Iterator iter() { return Iterator(this); }
+};
+
+#endif  // RB_H_
diff --git a/memory/build/replace_malloc.h b/memory/build/replace_malloc.h
new file mode 100644
index 0000000000..9ea3493f99
--- /dev/null
+++ b/memory/build/replace_malloc.h
@@ -0,0 +1,113 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef replace_malloc_h
+#define replace_malloc_h
+
+// The replace_malloc facility allows an external library to replace or
+// supplement the jemalloc implementation.
+//
+// The external library may be hooked by setting one of the following
+// environment variables to the library path:
+//   - LD_PRELOAD on Linux,
+//   - DYLD_INSERT_LIBRARIES on OSX,
+//   - MOZ_REPLACE_MALLOC_LIB on Windows and Android.
+//
+// An initialization function is called before any malloc replacement
+// function, and has the following declaration:
+//
+//   void replace_init(malloc_table_t*, ReplaceMallocBridge**)
+//
+// The malloc_table_t pointer given to that function is a table containing
+// pointers to the original allocator implementation, so that replacement
+// functions can call them back if they need to. The initialization function
+// needs to alter that table to replace the function it wants to replace.
+// If it needs the original implementation, it thus needs a copy of the
+// original table.
+//
+// The ReplaceMallocBridge* pointer is an outparam that allows the
+// replace_init function to return a pointer to its ReplaceMallocBridge
+// (see replace_malloc_bridge.h).
+//
+// The functions to be implemented in the external library are of the form:
+//
+//   void* replace_malloc(size_t size)
+//   {
+//     // Fiddle with the size if necessary.
+//     // orig->malloc doesn't have to be called if the external library
+//     // provides its own allocator, but in this case it will have to
+//     // implement all functions.
+//     void *ptr = orig->malloc(size);
+//     // Do whatever you want with the ptr.
+//     return ptr;
+//   }
+//
+// where "orig" is a pointer to a copy of the table replace_init got.
+//
+// See malloc_decls.h for a list of functions that can be replaced this
+// way. The implementations are all in the form:
+//   return_type replace_name(arguments [,...])
+//
+// They don't all need to be provided.
+//
+// Building a replace-malloc library is like rocket science. It can end up
+// with things blowing up, especially when trying to use complex types, and
+// even more especially when these types come from XPCOM or other parts of the
+// Mozilla codebase.
+// It is recommended to add the following to a replace-malloc implementation's
+// moz.build:
+//   DISABLE_STL_WRAPPING = True # Avoid STL wrapping
+//
+// If your replace-malloc implementation lives under memory/replace, these
+// are taken care of by memory/replace/defs.mk.
+
+#ifdef replace_malloc_bridge_h
+#  error Do not include replace_malloc_bridge.h before replace_malloc.h. \
+  In fact, you only need the latter.
+#endif
+
+#define REPLACE_MALLOC_IMPL
+
+#include "replace_malloc_bridge.h"
+
+// Implementing a replace-malloc library is incompatible with using mozalloc.
+#define MOZ_NO_MOZALLOC 1
+
+#include "mozilla/MacroArgs.h"
+#include "mozilla/Types.h"
+
+MOZ_BEGIN_EXTERN_C
+
+// MOZ_REPLACE_WEAK is only defined in mozjemalloc.cpp. Normally including
+// this header will add function definitions.
+#ifndef MOZ_REPLACE_WEAK
+#  define MOZ_REPLACE_WEAK
+#endif
+
+// When building a replace-malloc library for static linking, we want
+// each to have a different name for their "public" functions.
+// The build system defines MOZ_REPLACE_MALLOC_PREFIX in that case.
+#ifdef MOZ_REPLACE_MALLOC_PREFIX
+#  define replace_init MOZ_CONCAT(MOZ_REPLACE_MALLOC_PREFIX, _init)
+#  define MOZ_REPLACE_PUBLIC
+#else
+#  define MOZ_REPLACE_PUBLIC MOZ_EXPORT
+#endif
+
+struct ReplaceMallocBridge;
+typedef void (*jemalloc_init_func)(malloc_table_t*,
+                                   struct ReplaceMallocBridge**);
+
+// Replace-malloc library initialization function. See top of this file
+MOZ_REPLACE_PUBLIC void replace_init(
+    malloc_table_t*, struct ReplaceMallocBridge**) MOZ_REPLACE_WEAK;
+
+// ensure this is visible and libxul/etc reference it with a weak ref
+MFBT_API void jemalloc_replace_dynamic(jemalloc_init_func);
+
+MOZ_END_EXTERN_C
+
+#endif  // replace_malloc_h
diff --git a/memory/build/replace_malloc_bridge.h b/memory/build/replace_malloc_bridge.h
new file mode 100644
index 0000000000..358d9f25c4
--- /dev/null
+++ b/memory/build/replace_malloc_bridge.h
@@ -0,0 +1,257 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef replace_malloc_bridge_h
+#define replace_malloc_bridge_h
+
+// The replace-malloc bridge allows bidirectional method calls between
+// a program and the replace-malloc library that has been loaded for it.
+// In Firefox, this is used to allow method calls between code in libxul
+// and code in the replace-malloc library, without libxul needing to link
+// against that library or vice-versa.
+//
+// Subsystems can add methods for their own need. Replace-malloc libraries
+// can decide to implement those methods or not.
+//
+// Replace-malloc libraries can provide such a bridge by implementing
+// a ReplaceMallocBridge-derived class, and a replace_get_bridge function
+// returning an instance of that class. The default methods in
+// ReplaceMallocBridge are expected to return values that callers would
+// understand as "the bridge doesn't implement this method", so that a
+// replace-malloc library doesn't have to implement all methods.
+//
+// The ReplaceMallocBridge class contains definitions for methods for
+// all replace-malloc libraries. Each library picks the methods it wants
+// to reply to in its ReplaceMallocBridge-derived class instance.
+// All methods of ReplaceMallocBridge must be virtual. Similarly,
+// anything passed as an argument to those methods must be plain data, or
+// an instance of a class with only virtual methods.
+//
+// Binary compatibility is expected to be maintained, such that a newer
+// Firefox can be used with an old replace-malloc library, or an old
+// Firefox can be used with a newer replace-malloc library. As such, only
+// new virtual methods should be added to ReplaceMallocBridge, and
+// each change should have a corresponding bump of the mVersion value.
+// At the same time, each virtual method should have a corresponding
+// wrapper calling the virtual method on the instance from
+// ReplaceMallocBridge::Get(), giving it the version the virtual method
+// was added.
+//
+// Parts that are not relevant to the replace-malloc library end of the
+// bridge are hidden when REPLACE_MALLOC_IMPL is not defined, which is
+// the case when including replace_malloc.h.
+
+struct ReplaceMallocBridge;
+
+#include "mozilla/Types.h"
+
+MOZ_BEGIN_EXTERN_C
+
+#ifndef REPLACE_MALLOC_IMPL
+// Returns the replace-malloc bridge if there is one to be returned.
+MFBT_API ReplaceMallocBridge* get_bridge();
+#endif
+
+// Table of malloc functions.
+//   e.g. void* (*malloc)(size_t), etc.
+
+#define MALLOC_DECL(name, return_type, ...) \
+  typedef return_type(name##_impl_t)(__VA_ARGS__);
+
+#include "malloc_decls.h"
+
+#define MALLOC_DECL(name, return_type, ...) name##_impl_t* name;
+
+typedef struct {
+#include "malloc_decls.h"
+} malloc_table_t;
+
+MOZ_END_EXTERN_C
+
+#ifdef __cplusplus
+
+// Table of malloc hook functions.
+// Those functions are called with the arguments and results of malloc
+// functions after they are called.
+//   e.g. void* (*malloc_hook)(void*, size_t), etc.
+// They can either return the result they're given, or alter it before
+// returning it.
+// The hooks corresponding to functions, like free(void*), that return no
+// value, don't take an extra argument.
+// The table must at least contain a pointer for malloc_hook and free_hook
+// functions. They will be used as fallback if no pointer is given for
+// other allocation functions, like calloc_hook.
+namespace mozilla {
+namespace detail {
+template <typename R, typename... Args>
+struct AllocHookType {
+  using Type = R (*)(R, Args...);
+};
+
+template <typename... Args>
+struct AllocHookType<void, Args...> {
+  using Type = void (*)(Args...);
+};
+
+}  // namespace detail
+}  // namespace mozilla
+
+#  define MALLOC_DECL(name, return_type, ...)                                 \
+    typename mozilla::detail::AllocHookType<return_type, ##__VA_ARGS__>::Type \
+        name##_hook;
+
+typedef struct {
+#  include "malloc_decls.h"
+  // Like free_hook, but called before realloc_hook. free_hook is called
+  // instead of not given.
+  void (*realloc_hook_before)(void* aPtr);
+} malloc_hook_table_t;
+
+namespace mozilla {
+namespace dmd {
+struct DMDFuncs;
+}  // namespace dmd
+
+namespace phc {
+class AddrInfo;
+}  // namespace phc
+
+// Callbacks to register debug file handles for Poison IO interpose.
+// See Mozilla(|Un)RegisterDebugHandle in xpcom/build/PoisonIOInterposer.h
+struct DebugFdRegistry {
+  virtual void RegisterHandle(intptr_t aFd);
+
+  virtual void UnRegisterHandle(intptr_t aFd);
+};
+
+}  // namespace mozilla
+
+struct ReplaceMallocBridge {
+  ReplaceMallocBridge() : mVersion(4) {}
+
+  // This method was added in version 1 of the bridge.
+  virtual mozilla::dmd::DMDFuncs* GetDMDFuncs() { return nullptr; }
+
+  // Send a DebugFdRegistry instance to the replace-malloc library so that
+  // it can register/unregister file descriptors whenever needed. The
+  // instance is valid until the process dies.
+  // This method was added in version 2 of the bridge.
+  virtual void InitDebugFd(mozilla::DebugFdRegistry&) {}
+
+  // Register a list of malloc functions and hook functions to the
+  // replace-malloc library so that it can choose to dispatch to them
+  // when needed. The details of what is dispatched when is left to the
+  // replace-malloc library.
+  // Passing a nullptr for either table will unregister a previously
+  // registered table under the same name.
+  // Returns nullptr if registration failed.
+  // If registration succeeded, a table of "pure" malloc functions is
+  // returned. Those "pure" malloc functions won't call hooks.
+  // /!\ Do not rely on registration/unregistration to be instantaneous.
+  // Functions from a previously registered table may still be called for
+  // a brief time after RegisterHook returns.
+  // This method was added in version 3 of the bridge.
+  virtual const malloc_table_t* RegisterHook(
+      const char* aName, const malloc_table_t* aTable,
+      const malloc_hook_table_t* aHookTable) {
+    return nullptr;
+  }
+
+  // If this is a PHC-handled address, return true, and if an AddrInfo is
+  // provided, fill in all of its fields. Otherwise, return false and leave
+  // AddrInfo unchanged.
+  // This method was added in version 4 of the bridge.
+  virtual bool IsPHCAllocation(const void*, mozilla::phc::AddrInfo*) {
+    return false;
+  }
+
+  // Disable PHC allocations on the current thread. Only useful for tests. Note
+  // that PHC deallocations will still occur as needed.
+  // This method was added in version 4 of the bridge.
+  virtual void DisablePHCOnCurrentThread() {}
+
+  // Re-enable PHC allocations on the current thread. Only useful for tests.
+  // This method was added in version 4 of the bridge.
+  virtual void ReenablePHCOnCurrentThread() {}
+
+  // Test whether PHC allocations are enabled on the current thread. Only
+  // useful for tests.
+  // This method was added in version 4 of the bridge.
+  virtual bool IsPHCEnabledOnCurrentThread() { return false; }
+
+#  ifndef REPLACE_MALLOC_IMPL
+  // Returns the replace-malloc bridge if its version is at least the
+  // requested one.
+  static ReplaceMallocBridge* Get(int aMinimumVersion) {
+    static ReplaceMallocBridge* sSingleton = get_bridge();
+    return (sSingleton && sSingleton->mVersion >= aMinimumVersion) ? sSingleton
+                                                                   : nullptr;
+  }
+#  endif
+
+ protected:
+  const int mVersion;
+};
+
+#  ifndef REPLACE_MALLOC_IMPL
+// Class containing wrappers for calls to ReplaceMallocBridge methods.
+// Those wrappers need to be static methods in a class because compilers
+// complain about unused static global functions, and linkers complain
+// about multiple definitions of non-static global functions.
+// Using a separate class from ReplaceMallocBridge allows the function
+// names to be identical.
+struct ReplaceMalloc {
+  // Don't call this method from performance critical code. Use
+  // mozilla::dmd::DMDFuncs::Get() instead, it has less overhead.
+  static mozilla::dmd::DMDFuncs* GetDMDFuncs() {
+    auto singleton = ReplaceMallocBridge::Get(/* minimumVersion */ 1);
+    return singleton ? singleton->GetDMDFuncs() : nullptr;
+  }
+
+  static void InitDebugFd(mozilla::DebugFdRegistry& aRegistry) {
+    auto singleton = ReplaceMallocBridge::Get(/* minimumVersion */ 2);
+    if (singleton) {
+      singleton->InitDebugFd(aRegistry);
+    }
+  }
+
+  static const malloc_table_t* RegisterHook(
+      const char* aName, const malloc_table_t* aTable,
+      const malloc_hook_table_t* aHookTable) {
+    auto singleton = ReplaceMallocBridge::Get(/* minimumVersion */ 3);
+    return singleton ? singleton->RegisterHook(aName, aTable, aHookTable)
+                     : nullptr;
+  }
+
+  static bool IsPHCAllocation(const void* aPtr, mozilla::phc::AddrInfo* aOut) {
+    auto singleton = ReplaceMallocBridge::Get(/* minimumVersion */ 4);
+    return singleton ? singleton->IsPHCAllocation(aPtr, aOut) : false;
+  }
+
+  static void DisablePHCOnCurrentThread() {
+    auto singleton = ReplaceMallocBridge::Get(/* minimumVersion */ 4);
+    if (singleton) {
+      singleton->DisablePHCOnCurrentThread();
+    }
+  }
+
+  static void ReenablePHCOnCurrentThread() {
+    auto singleton = ReplaceMallocBridge::Get(/* minimumVersion */ 4);
+    if (singleton) {
+      singleton->ReenablePHCOnCurrentThread();
+    }
+  }
+
+  static bool IsPHCEnabledOnCurrentThread() {
+    auto singleton = ReplaceMallocBridge::Get(/* minimumVersion */ 4);
+    return singleton ? singleton->IsPHCEnabledOnCurrentThread() : false;
+  }
+};
+#  endif
+
+#endif  // __cplusplus
+
+#endif  // replace_malloc_bridge_h
diff --git a/memory/build/zone.c b/memory/build/zone.c
new file mode 100644
index 0000000000..7311ccf27b
--- /dev/null
+++ b/memory/build/zone.c
@@ -0,0 +1,377 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "mozmemory_wrap.h"
+
+#include <stdlib.h>
+#include <mach/mach_types.h>
+#include "mozilla/Assertions.h"
+
+// Malloc implementation functions are MOZ_MEMORY_API, and jemalloc
+// specific functions MOZ_JEMALLOC_API; see mozmemory_wrap.h
+
+#define MALLOC_DECL(name, return_type, ...) \
+  MOZ_MEMORY_API return_type name##_impl(__VA_ARGS__);
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC
+#include "malloc_decls.h"
+
+#define MALLOC_DECL(name, return_type, ...) \
+  MOZ_JEMALLOC_API return_type name##_impl(__VA_ARGS__);
+#define MALLOC_FUNCS MALLOC_FUNCS_JEMALLOC
+#include "malloc_decls.h"
+
+// Definitions of the following structs in malloc/malloc.h might be too old
+// for the built binary to run on newer versions of OSX. So use the newest
+// possible version of those structs.
+
+typedef struct _malloc_zone_t {
+  void* reserved1;
+  void* reserved2;
+  size_t (*size)(struct _malloc_zone_t*, const void*);
+  void* (*malloc)(struct _malloc_zone_t*, size_t);
+  void* (*calloc)(struct _malloc_zone_t*, size_t, size_t);
+  void* (*valloc)(struct _malloc_zone_t*, size_t);
+  void (*free)(struct _malloc_zone_t*, void*);
+  void* (*realloc)(struct _malloc_zone_t*, void*, size_t);
+  void (*destroy)(struct _malloc_zone_t*);
+  const char* zone_name;
+  unsigned (*batch_malloc)(struct _malloc_zone_t*, size_t, void**, unsigned);
+  void (*batch_free)(struct _malloc_zone_t*, void**, unsigned);
+  struct malloc_introspection_t* introspect;
+  unsigned version;
+  void* (*memalign)(struct _malloc_zone_t*, size_t, size_t);
+  void (*free_definite_size)(struct _malloc_zone_t*, void*, size_t);
+  size_t (*pressure_relief)(struct _malloc_zone_t*, size_t);
+} malloc_zone_t;
+
+typedef struct {
+  vm_address_t address;
+  vm_size_t size;
+} vm_range_t;
+
+typedef struct malloc_statistics_t {
+  unsigned blocks_in_use;
+  size_t size_in_use;
+  size_t max_size_in_use;
+  size_t size_allocated;
+} malloc_statistics_t;
+
+typedef kern_return_t memory_reader_t(task_t, vm_address_t, vm_size_t, void**);
+
+typedef void vm_range_recorder_t(task_t, void*, unsigned type, vm_range_t*,
+                                 unsigned);
+
+typedef struct malloc_introspection_t {
+  kern_return_t (*enumerator)(task_t, void*, unsigned, vm_address_t,
+                              memory_reader_t, vm_range_recorder_t);
+  size_t (*good_size)(malloc_zone_t*, size_t);
+  boolean_t (*check)(malloc_zone_t*);
+  void (*print)(malloc_zone_t*, boolean_t);
+  void (*log)(malloc_zone_t*, void*);
+  void (*force_lock)(malloc_zone_t*);
+  void (*force_unlock)(malloc_zone_t*);
+  void (*statistics)(malloc_zone_t*, malloc_statistics_t*);
+  boolean_t (*zone_locked)(malloc_zone_t*);
+  boolean_t (*enable_discharge_checking)(malloc_zone_t*);
+  boolean_t (*disable_discharge_checking)(malloc_zone_t*);
+  void (*discharge)(malloc_zone_t*, void*);
+#ifdef __BLOCKS__
+  void (*enumerate_discharged_pointers)(malloc_zone_t*, void (^)(void*, void*));
+#else
+  void* enumerate_unavailable_without_blocks;
+#endif
+  void (*reinit_lock)(malloc_zone_t*);
+} malloc_introspection_t;
+
+extern kern_return_t malloc_get_all_zones(task_t, memory_reader_t,
+                                          vm_address_t**, unsigned*);
+
+extern malloc_zone_t* malloc_default_zone(void);
+
+extern void malloc_zone_register(malloc_zone_t* zone);
+
+extern void malloc_zone_unregister(malloc_zone_t* zone);
+
+extern malloc_zone_t* malloc_default_purgeable_zone(void);
+
+extern malloc_zone_t* malloc_zone_from_ptr(const void* ptr);
+
+extern void malloc_zone_free(malloc_zone_t* zone, void* ptr);
+
+extern void* malloc_zone_realloc(malloc_zone_t* zone, void* ptr, size_t size);
+
+// The following is a OSX zone allocator implementation.
+// /!\ WARNING. It assumes the underlying malloc implementation's
+// malloc_usable_size returns 0 when the given pointer is not owned by
+// the allocator. Sadly, OSX does call zone_size with pointers not
+// owned by the allocator.
+
+static size_t zone_size(malloc_zone_t* zone, const void* ptr) {
+  return malloc_usable_size_impl(ptr);
+}
+
+static void* zone_malloc(malloc_zone_t* zone, size_t size) {
+  return malloc_impl(size);
+}
+
+static void* zone_calloc(malloc_zone_t* zone, size_t num, size_t size) {
+  return calloc_impl(num, size);
+}
+
+static void* zone_realloc(malloc_zone_t* zone, void* ptr, size_t size) {
+  if (malloc_usable_size_impl(ptr)) return realloc_impl(ptr, size);
+
+  // Sometimes, system libraries call malloc_zone_* functions with the wrong
+  // zone (e.g. CoreFoundation does). In that case, we need to find the real
+  // one. We can't call libSystem's realloc directly because we're exporting
+  // realloc from libmozglue and we'd pick that one, so we manually find the
+  // right zone and realloc with it.
+  malloc_zone_t* real_zone = malloc_zone_from_ptr(ptr);
+  // The system allocator crashes voluntarily by default when a pointer can't
+  // be traced back to a zone. Do the same.
+  MOZ_RELEASE_ASSERT(real_zone);
+  MOZ_RELEASE_ASSERT(real_zone != zone);
+  return malloc_zone_realloc(real_zone, ptr, size);
+}
+
+static void other_zone_free(malloc_zone_t* original_zone, void* ptr) {
+  // Sometimes, system libraries call malloc_zone_* functions with the wrong
+  // zone (e.g. CoreFoundation does). In that case, we need to find the real
+  // one. We can't call libSystem's free directly because we're exporting
+  // free from libmozglue and we'd pick that one, so we manually find the
+  // right zone and free with it.
+  if (!ptr) {
+    return;
+  }
+  malloc_zone_t* zone = malloc_zone_from_ptr(ptr);
+  // The system allocator crashes voluntarily by default when a pointer can't
+  // be traced back to a zone. Do the same.
+  MOZ_RELEASE_ASSERT(zone);
+  MOZ_RELEASE_ASSERT(zone != original_zone);
+  return malloc_zone_free(zone, ptr);
+}
+
+static void zone_free(malloc_zone_t* zone, void* ptr) {
+  if (malloc_usable_size_impl(ptr)) {
+    free_impl(ptr);
+    return;
+  }
+  other_zone_free(zone, ptr);
+}
+
+static void zone_free_definite_size(malloc_zone_t* zone, void* ptr,
+                                    size_t size) {
+  size_t current_size = malloc_usable_size_impl(ptr);
+  if (current_size) {
+    MOZ_ASSERT(current_size == size);
+    free_impl(ptr);
+    return;
+  }
+  other_zone_free(zone, ptr);
+}
+
+static void* zone_memalign(malloc_zone_t* zone, size_t alignment, size_t size) {
+  void* ptr;
+  if (posix_memalign_impl(&ptr, alignment, size) == 0) return ptr;
+  return NULL;
+}
+
+static void* zone_valloc(malloc_zone_t* zone, size_t size) {
+  return valloc_impl(size);
+}
+
+static void zone_destroy(malloc_zone_t* zone) {
+  // This function should never be called.
+  MOZ_CRASH();
+}
+
+static unsigned zone_batch_malloc(malloc_zone_t* zone, size_t size,
+                                  void** results, unsigned num_requested) {
+  unsigned i;
+
+  for (i = 0; i < num_requested; i++) {
+    results[i] = malloc_impl(size);
+    if (!results[i]) break;
+  }
+
+  return i;
+}
+
+static void zone_batch_free(malloc_zone_t* zone, void** to_be_freed,
+                            unsigned num_to_be_freed) {
+  unsigned i;
+
+  for (i = 0; i < num_to_be_freed; i++) {
+    zone_free(zone, to_be_freed[i]);
+    to_be_freed[i] = NULL;
+  }
+}
+
+static size_t zone_pressure_relief(malloc_zone_t* zone, size_t goal) {
+  return 0;
+}
+
+static size_t zone_good_size(malloc_zone_t* zone, size_t size) {
+  return malloc_good_size_impl(size);
+}
+
+static kern_return_t zone_enumerator(task_t task, void* data,
+                                     unsigned type_mask,
+                                     vm_address_t zone_address,
+                                     memory_reader_t reader,
+                                     vm_range_recorder_t recorder) {
+  return KERN_SUCCESS;
+}
+
+static boolean_t zone_check(malloc_zone_t* zone) { return true; }
+
+static void zone_print(malloc_zone_t* zone, boolean_t verbose) {}
+
+static void zone_log(malloc_zone_t* zone, void* address) {}
+
+extern void _malloc_prefork(void);
+extern void _malloc_postfork_child(void);
+
+static void zone_force_lock(malloc_zone_t* zone) {
+  // /!\ This calls into mozjemalloc. It works because we're linked in the
+  // same library.
+  _malloc_prefork();
+}
+
+static void zone_force_unlock(malloc_zone_t* zone) {
+  // /!\ This calls into mozjemalloc. It works because we're linked in the
+  // same library.
+  _malloc_postfork_child();
+}
+
+static void zone_statistics(malloc_zone_t* zone, malloc_statistics_t* stats) {
+  // We make no effort to actually fill the values
+  stats->blocks_in_use = 0;
+  stats->size_in_use = 0;
+  stats->max_size_in_use = 0;
+  stats->size_allocated = 0;
+}
+
+static boolean_t zone_locked(malloc_zone_t* zone) {
+  // Pretend no lock is being held
+  return false;
+}
+
+static void zone_reinit_lock(malloc_zone_t* zone) {
+  // As of OSX 10.12, this function is only used when force_unlock would
+  // be used if the zone version were < 9. So just use force_unlock.
+  zone_force_unlock(zone);
+}
+
+static malloc_zone_t zone;
+static struct malloc_introspection_t zone_introspect;
+
+static malloc_zone_t* get_default_zone() {
+  malloc_zone_t** zones = NULL;
+  unsigned int num_zones = 0;
+
+  // On OSX 10.12, malloc_default_zone returns a special zone that is not
+  // present in the list of registered zones. That zone uses a "lite zone"
+  // if one is present (apparently enabled when malloc stack logging is
+  // enabled), or the first registered zone otherwise. In practice this
+  // means unless malloc stack logging is enabled, the first registered
+  // zone is the default.
+  // So get the list of zones to get the first one, instead of relying on
+  // malloc_default_zone.
+  if (KERN_SUCCESS !=
+      malloc_get_all_zones(0, NULL, (vm_address_t**)&zones, &num_zones)) {
+    // Reset the value in case the failure happened after it was set.
+    num_zones = 0;
+  }
+  if (num_zones) {
+    return zones[0];
+  }
+  return malloc_default_zone();
+}
+
+__attribute__((constructor)) static void register_zone(void) {
+  malloc_zone_t* default_zone = get_default_zone();
+
+  zone.size = zone_size;
+  zone.malloc = zone_malloc;
+  zone.calloc = zone_calloc;
+  zone.valloc = zone_valloc;
+  zone.free = zone_free;
+  zone.realloc = zone_realloc;
+  zone.destroy = zone_destroy;
+#ifdef MOZ_REPLACE_MALLOC
+  zone.zone_name = "replace_malloc_zone";
+#else
+  zone.zone_name = "jemalloc_zone";
+#endif
+  zone.batch_malloc = zone_batch_malloc;
+  zone.batch_free = zone_batch_free;
+  zone.introspect = &zone_introspect;
+  zone.version = 9;
+  zone.memalign = zone_memalign;
+  zone.free_definite_size = zone_free_definite_size;
+  zone.pressure_relief = zone_pressure_relief;
+  zone_introspect.enumerator = zone_enumerator;
+  zone_introspect.good_size = zone_good_size;
+  zone_introspect.check = zone_check;
+  zone_introspect.print = zone_print;
+  zone_introspect.log = zone_log;
+  zone_introspect.force_lock = zone_force_lock;
+  zone_introspect.force_unlock = zone_force_unlock;
+  zone_introspect.statistics = zone_statistics;
+  zone_introspect.zone_locked = zone_locked;
+  zone_introspect.enable_discharge_checking = NULL;
+  zone_introspect.disable_discharge_checking = NULL;
+  zone_introspect.discharge = NULL;
+#ifdef __BLOCKS__
+  zone_introspect.enumerate_discharged_pointers = NULL;
+#else
+  zone_introspect.enumerate_unavailable_without_blocks = NULL;
+#endif
+  zone_introspect.reinit_lock = zone_reinit_lock;
+
+  // The default purgeable zone is created lazily by OSX's libc.  It uses
+  // the default zone when it is created for "small" allocations
+  // (< 15 KiB), but assumes the default zone is a scalable_zone.  This
+  // obviously fails when the default zone is the jemalloc zone, so
+  // malloc_default_purgeable_zone is called beforehand so that the
+  // default purgeable zone is created when the default zone is still
+  // a scalable_zone.
+  malloc_zone_t* purgeable_zone = malloc_default_purgeable_zone();
+
+  // There is a problem related to the above with the system nano zone, which
+  // is hard to work around from here, and that is instead worked around by
+  // disabling the nano zone through an environment variable
+  // (MallocNanoZone=0). In Firefox, we do that through
+  // browser/app/macbuild/Contents/Info.plist.in.
+
+  // Register the custom zone.  At this point it won't be the default.
+  malloc_zone_register(&zone);
+
+  do {
+    // Unregister and reregister the default zone.  On OSX >= 10.6,
+    // unregistering takes the last registered zone and places it at the
+    // location of the specified zone.  Unregistering the default zone thus
+    // makes the last registered one the default.  On OSX < 10.6,
+    // unregistering shifts all registered zones.  The first registered zone
+    // then becomes the default.
+    malloc_zone_unregister(default_zone);
+    malloc_zone_register(default_zone);
+
+    // On OSX 10.6, having the default purgeable zone appear before the default
+    // zone makes some things crash because it thinks it owns the default
+    // zone allocated pointers. We thus unregister/re-register it in order to
+    // ensure it's always after the default zone. On OSX < 10.6, as
+    // unregistering shifts registered zones, this simply removes the purgeable
+    // zone from the list and adds it back at the end, after the default zone.
+    // On OSX >= 10.6, unregistering replaces the purgeable zone with the last
+    // registered zone above, i.e the default zone. Registering it again then
+    // puts it at the end, obviously after the default zone.
+    malloc_zone_unregister(purgeable_zone);
+    malloc_zone_register(purgeable_zone);
+    default_zone = get_default_zone();
+  } while (default_zone != &zone);
+}
diff --git a/memory/gtest/TestJemalloc.cpp b/memory/gtest/TestJemalloc.cpp
new file mode 100644
index 0000000000..7f3b3f9cf3
--- /dev/null
+++ b/memory/gtest/TestJemalloc.cpp
@@ -0,0 +1,756 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "mozilla/mozalloc.h"
+#include "mozilla/UniquePtr.h"
+#include "mozilla/Unused.h"
+#include "mozilla/Vector.h"
+#include "mozilla/gtest/MozHelpers.h"
+#include "mozmemory.h"
+#include "nsCOMPtr.h"
+#include "Utils.h"
+
+#include "gtest/gtest.h"
+
+#ifdef MOZ_PHC
+#  include "replace_malloc_bridge.h"
+#endif
+
+using namespace mozilla;
+
+class AutoDisablePHCOnCurrentThread {
+ public:
+  AutoDisablePHCOnCurrentThread() {
+#ifdef MOZ_PHC
+    ReplaceMalloc::DisablePHCOnCurrentThread();
+#endif
+  }
+
+  ~AutoDisablePHCOnCurrentThread() {
+#ifdef MOZ_PHC
+    ReplaceMalloc::ReenablePHCOnCurrentThread();
+#endif
+  }
+};
+
+static inline void TestOne(size_t size) {
+  size_t req = size;
+  size_t adv = malloc_good_size(req);
+  char* p = (char*)malloc(req);
+  size_t usable = moz_malloc_usable_size(p);
+  // NB: Using EXPECT here so that we still free the memory on failure.
+  EXPECT_EQ(adv, usable) << "malloc_good_size(" << req << ") --> " << adv
+                         << "; "
+                            "malloc_usable_size("
+                         << req << ") --> " << usable;
+  free(p);
+}
+
+static inline void TestThree(size_t size) {
+  ASSERT_NO_FATAL_FAILURE(TestOne(size - 1));
+  ASSERT_NO_FATAL_FAILURE(TestOne(size));
+  ASSERT_NO_FATAL_FAILURE(TestOne(size + 1));
+}
+
+TEST(Jemalloc, UsableSizeInAdvance)
+{
+  /*
+   * Test every size up to a certain point, then (N-1, N, N+1) triplets for a
+   * various sizes beyond that.
+   */
+
+  for (size_t n = 0; n < 16_KiB; n++) ASSERT_NO_FATAL_FAILURE(TestOne(n));
+
+  for (size_t n = 16_KiB; n < 1_MiB; n += 4_KiB)
+    ASSERT_NO_FATAL_FAILURE(TestThree(n));
+
+  for (size_t n = 1_MiB; n < 8_MiB; n += 128_KiB)
+    ASSERT_NO_FATAL_FAILURE(TestThree(n));
+}
+
+static int gStaticVar;
+
+bool InfoEq(jemalloc_ptr_info_t& aInfo, PtrInfoTag aTag, void* aAddr,
+            size_t aSize, arena_id_t arenaId) {
+  return aInfo.tag == aTag && aInfo.addr == aAddr && aInfo.size == aSize
+#ifdef MOZ_DEBUG
+         && aInfo.arenaId == arenaId
+#endif
+      ;
+}
+
+bool InfoEqFreedPage(jemalloc_ptr_info_t& aInfo, void* aAddr, size_t aPageSize,
+                     arena_id_t arenaId) {
+  size_t pageSizeMask = aPageSize - 1;
+
+  return jemalloc_ptr_is_freed_page(&aInfo) &&
+         aInfo.addr == (void*)(uintptr_t(aAddr) & ~pageSizeMask) &&
+         aInfo.size == aPageSize
+#ifdef MOZ_DEBUG
+         && aInfo.arenaId == arenaId
+#endif
+      ;
+}
+
+TEST(Jemalloc, PtrInfo)
+{
+  arena_id_t arenaId = moz_create_arena();
+  ASSERT_TRUE(arenaId != 0);
+
+  jemalloc_stats_t stats;
+  jemalloc_stats(&stats);
+
+  jemalloc_ptr_info_t info;
+  Vector<char*> small, large, huge;
+
+  // For small (less than half the page size) allocations, test every position
+  // within many possible sizes.
+  size_t small_max =
+      stats.subpage_max ? stats.subpage_max : stats.quantum_wide_max;
+  for (size_t n = 0; n <= small_max; n += 8) {
+    auto p = (char*)moz_arena_malloc(arenaId, n);
+    size_t usable = moz_malloc_size_of(p);
+    ASSERT_TRUE(small.append(p));
+    for (size_t j = 0; j < usable; j++) {
+      jemalloc_ptr_info(&p[j], &info);
+      ASSERT_TRUE(InfoEq(info, TagLiveAlloc, p, usable, arenaId));
+    }
+  }
+
+  // Similar for large (small_max + 1 KiB .. 1MiB - 8KiB) allocations.
+  for (size_t n = small_max + 1_KiB; n <= stats.large_max; n += 1_KiB) {
+    auto p = (char*)moz_arena_malloc(arenaId, n);
+    size_t usable = moz_malloc_size_of(p);
+    ASSERT_TRUE(large.append(p));
+    for (size_t j = 0; j < usable; j += 347) {
+      jemalloc_ptr_info(&p[j], &info);
+      ASSERT_TRUE(InfoEq(info, TagLiveAlloc, p, usable, arenaId));
+    }
+  }
+
+  // Similar for huge (> 1MiB - 8KiB) allocations.
+  for (size_t n = stats.chunksize; n <= 10_MiB; n += 512_KiB) {
+    auto p = (char*)moz_arena_malloc(arenaId, n);
+    size_t usable = moz_malloc_size_of(p);
+    ASSERT_TRUE(huge.append(p));
+    for (size_t j = 0; j < usable; j += 567) {
+      jemalloc_ptr_info(&p[j], &info);
+      ASSERT_TRUE(InfoEq(info, TagLiveAlloc, p, usable, arenaId));
+    }
+  }
+
+  // The following loops check freed allocations. We step through the vectors
+  // using prime-sized steps, which gives full coverage of the arrays while
+  // avoiding deallocating in the same order we allocated.
+  size_t len;
+
+  // Free the small allocations and recheck them.
+  int isFreedAlloc = 0, isFreedPage = 0;
+  len = small.length();
+  for (size_t i = 0, j = 0; i < len; i++, j = (j + 19) % len) {
+    char* p = small[j];
+    size_t usable = moz_malloc_size_of(p);
+    free(p);
+    for (size_t k = 0; k < usable; k++) {
+      jemalloc_ptr_info(&p[k], &info);
+      // There are two valid outcomes here.
+      if (InfoEq(info, TagFreedAlloc, p, usable, arenaId)) {
+        isFreedAlloc++;
+      } else if (InfoEqFreedPage(info, &p[k], stats.page_size, arenaId)) {
+        isFreedPage++;
+      } else {
+        ASSERT_TRUE(false);
+      }
+    }
+  }
+  // There should be both FreedAlloc and FreedPage results, but a lot more of
+  // the former.
+  ASSERT_TRUE(isFreedAlloc != 0);
+  ASSERT_TRUE(isFreedPage != 0);
+  ASSERT_TRUE(isFreedAlloc / isFreedPage > 8);
+
+  // Free the large allocations and recheck them.
+  len = large.length();
+  for (size_t i = 0, j = 0; i < len; i++, j = (j + 31) % len) {
+    char* p = large[j];
+    size_t usable = moz_malloc_size_of(p);
+    free(p);
+    for (size_t k = 0; k < usable; k += 357) {
+      jemalloc_ptr_info(&p[k], &info);
+      ASSERT_TRUE(InfoEqFreedPage(info, &p[k], stats.page_size, arenaId));
+    }
+  }
+
+  // Free the huge allocations and recheck them.
+  len = huge.length();
+  for (size_t i = 0, j = 0; i < len; i++, j = (j + 7) % len) {
+    char* p = huge[j];
+    size_t usable = moz_malloc_size_of(p);
+    free(p);
+    for (size_t k = 0; k < usable; k += 587) {
+      jemalloc_ptr_info(&p[k], &info);
+      ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+    }
+  }
+
+  // Null ptr.
+  jemalloc_ptr_info(nullptr, &info);
+  ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+
+  // Near-null ptr.
+  jemalloc_ptr_info((void*)0x123, &info);
+  ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+
+  // Maximum address.
+  jemalloc_ptr_info((void*)uintptr_t(-1), &info);
+  ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+
+  // Stack memory.
+  int stackVar;
+  jemalloc_ptr_info(&stackVar, &info);
+  ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+
+  // Code memory.
+  jemalloc_ptr_info((const void*)&jemalloc_ptr_info, &info);
+  ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+
+  // Static memory.
+  jemalloc_ptr_info(&gStaticVar, &info);
+  ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+
+  // Chunk header.
+  UniquePtr<int> p = MakeUnique<int>();
+  size_t chunksizeMask = stats.chunksize - 1;
+  char* chunk = (char*)(uintptr_t(p.get()) & ~chunksizeMask);
+  size_t chunkHeaderSize = stats.chunksize - stats.large_max - stats.page_size;
+  for (size_t i = 0; i < chunkHeaderSize; i += 64) {
+    jemalloc_ptr_info(&chunk[i], &info);
+    ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+  }
+
+  // Run header.
+  size_t page_sizeMask = stats.page_size - 1;
+  char* run = (char*)(uintptr_t(p.get()) & ~page_sizeMask);
+  for (size_t i = 0; i < 4 * sizeof(void*); i++) {
+    jemalloc_ptr_info(&run[i], &info);
+    ASSERT_TRUE(InfoEq(info, TagUnknown, nullptr, 0U, 0U));
+  }
+
+  // Entire chunk. It's impossible to check what is put into |info| for all of
+  // these addresses; this is more about checking that we don't crash.
+  for (size_t i = 0; i < stats.chunksize; i += 256) {
+    jemalloc_ptr_info(&chunk[i], &info);
+  }
+
+  moz_dispose_arena(arenaId);
+}
+
+size_t sSizes[] = {1,      42,      79,      918,     1.4_KiB,
+                   73_KiB, 129_KiB, 1.1_MiB, 2.6_MiB, 5.1_MiB};
+
+TEST(Jemalloc, Arenas)
+{
+  arena_id_t arena = moz_create_arena();
+  ASSERT_TRUE(arena != 0);
+  void* ptr = moz_arena_malloc(arena, 42);
+  ASSERT_TRUE(ptr != nullptr);
+  ptr = moz_arena_realloc(arena, ptr, 64);
+  ASSERT_TRUE(ptr != nullptr);
+  moz_arena_free(arena, ptr);
+  ptr = moz_arena_calloc(arena, 24, 2);
+  // For convenience, free can be used to free arena pointers.
+  free(ptr);
+  moz_dispose_arena(arena);
+
+  // Avoid death tests adding some unnecessary (long) delays.
+  SAVE_GDB_SLEEP_LOCAL();
+
+  // Can't use an arena after it's disposed.
+  // ASSERT_DEATH_WRAP(moz_arena_malloc(arena, 80), "");
+
+  // Arena id 0 can't be used to somehow get to the main arena.
+  ASSERT_DEATH_WRAP(moz_arena_malloc(0, 80), "");
+
+  arena = moz_create_arena();
+  arena_id_t arena2 = moz_create_arena();
+  // Ensure arena2 is used to prevent OSX errors:
+  (void)arena2;
+
+  // For convenience, realloc can also be used to reallocate arena pointers.
+  // The result should be in the same arena. Test various size class
+  // transitions.
+  for (size_t from_size : sSizes) {
+    SCOPED_TRACE(testing::Message() << "from_size = " << from_size);
+    for (size_t to_size : sSizes) {
+      SCOPED_TRACE(testing::Message() << "to_size = " << to_size);
+      ptr = moz_arena_malloc(arena, from_size);
+      ptr = realloc(ptr, to_size);
+      // Freeing with the wrong arena should crash.
+      ASSERT_DEATH_WRAP(moz_arena_free(arena2, ptr), "");
+      // Likewise for moz_arena_realloc.
+      ASSERT_DEATH_WRAP(moz_arena_realloc(arena2, ptr, from_size), "");
+      // The following will crash if it's not in the right arena.
+      moz_arena_free(arena, ptr);
+    }
+  }
+
+  moz_dispose_arena(arena2);
+  moz_dispose_arena(arena);
+
+  RESTORE_GDB_SLEEP_LOCAL();
+}
+
+// Check that a buffer aPtr is entirely filled with a given character from
+// aOffset to aSize. For faster comparison, the caller is required to fill a
+// reference buffer with the wanted character, and give the size of that
+// reference buffer.
+static void bulk_compare(char* aPtr, size_t aOffset, size_t aSize,
+                         char* aReference, size_t aReferenceSize) {
+  for (size_t i = aOffset; i < aSize; i += aReferenceSize) {
+    size_t length = std::min(aSize - i, aReferenceSize);
+    if (memcmp(aPtr + i, aReference, length)) {
+      // We got a mismatch, we now want to report more precisely where.
+      for (size_t j = i; j < i + length; j++) {
+        ASSERT_EQ(aPtr[j], *aReference);
+      }
+    }
+  }
+}
+
+// A range iterator for size classes between two given values.
+class SizeClassesBetween {
+ public:
+  SizeClassesBetween(size_t aStart, size_t aEnd) : mStart(aStart), mEnd(aEnd) {}
+
+  class Iterator {
+   public:
+    explicit Iterator(size_t aValue) : mValue(malloc_good_size(aValue)) {}
+
+    operator size_t() const { return mValue; }
+    size_t operator*() const { return mValue; }
+    Iterator& operator++() {
+      mValue = malloc_good_size(mValue + 1);
+      return *this;
+    }
+
+   private:
+    size_t mValue;
+  };
+
+  Iterator begin() { return Iterator(mStart); }
+  Iterator end() { return Iterator(mEnd); }
+
+ private:
+  size_t mStart, mEnd;
+};
+
+#define ALIGNMENT_CEILING(s, alignment) \
+  (((s) + ((alignment)-1)) & (~((alignment)-1)))
+
+#define ALIGNMENT_FLOOR(s, alignment) ((s) & (~((alignment)-1)))
+
+static bool IsSameRoundedHugeClass(size_t aSize1, size_t aSize2,
+                                   jemalloc_stats_t& aStats) {
+  return (aSize1 > aStats.large_max && aSize2 > aStats.large_max &&
+          ALIGNMENT_CEILING(aSize1 + aStats.page_size, aStats.chunksize) ==
+              ALIGNMENT_CEILING(aSize2 + aStats.page_size, aStats.chunksize));
+}
+
+static bool CanReallocInPlace(size_t aFromSize, size_t aToSize,
+                              jemalloc_stats_t& aStats) {
+  // PHC allocations must be disabled because PHC reallocs differently to
+  // mozjemalloc.
+#ifdef MOZ_PHC
+  MOZ_RELEASE_ASSERT(!ReplaceMalloc::IsPHCEnabledOnCurrentThread());
+#endif
+
+  if (aFromSize == malloc_good_size(aToSize)) {
+    // Same size class: in-place.
+    return true;
+  }
+  if (aFromSize >= aStats.page_size && aFromSize <= aStats.large_max &&
+      aToSize >= aStats.page_size && aToSize <= aStats.large_max) {
+    // Any large class to any large class: in-place when there is space to.
+    return true;
+  }
+  if (IsSameRoundedHugeClass(aFromSize, aToSize, aStats)) {
+    // Huge sizes that round up to the same multiple of the chunk size:
+    // in-place.
+    return true;
+  }
+  return false;
+}
+
+TEST(Jemalloc, InPlace)
+{
+  // Disable PHC allocations for this test, because CanReallocInPlace() isn't
+  // valid for PHC allocations.
+  AutoDisablePHCOnCurrentThread disable;
+
+  jemalloc_stats_t stats;
+  jemalloc_stats(&stats);
+
+  // Using a separate arena, which is always emptied after an iteration, ensures
+  // that in-place reallocation happens in all cases it can happen. This test is
+  // intended for developers to notice they may have to adapt other tests if
+  // they change the conditions for in-place reallocation.
+  arena_id_t arena = moz_create_arena();
+
+  for (size_t from_size : SizeClassesBetween(1, 2 * stats.chunksize)) {
+    SCOPED_TRACE(testing::Message() << "from_size = " << from_size);
+    for (size_t to_size : sSizes) {
+      SCOPED_TRACE(testing::Message() << "to_size = " << to_size);
+      char* ptr = (char*)moz_arena_malloc(arena, from_size);
+      char* ptr2 = (char*)moz_arena_realloc(arena, ptr, to_size);
+      if (CanReallocInPlace(from_size, to_size, stats)) {
+        EXPECT_EQ(ptr, ptr2);
+      } else {
+        EXPECT_NE(ptr, ptr2);
+      }
+      moz_arena_free(arena, ptr2);
+    }
+  }
+
+  moz_dispose_arena(arena);
+}
+
+// Bug 1474254: disable this test for windows ccov builds because it leads to
+// timeout.
+#if !defined(XP_WIN) || !defined(MOZ_CODE_COVERAGE)
+TEST(Jemalloc, JunkPoison)
+{
+  // Disable PHC allocations for this test, because CanReallocInPlace() isn't
+  // valid for PHC allocations, and the testing UAFs aren't valid.
+  AutoDisablePHCOnCurrentThread disable;
+
+  jemalloc_stats_t stats;
+  jemalloc_stats(&stats);
+
+  // Avoid death tests adding some unnecessary (long) delays.
+  SAVE_GDB_SLEEP_LOCAL();
+
+  // Create buffers in a separate arena, for faster comparisons with
+  // bulk_compare.
+  arena_id_t buf_arena = moz_create_arena();
+  char* junk_buf = (char*)moz_arena_malloc(buf_arena, stats.page_size);
+  // Depending on its configuration, the allocator will either fill the
+  // requested allocation with the junk byte (0xe4) or with zeroes, or do
+  // nothing, in which case, since we're allocating in a fresh arena,
+  // we'll be getting zeroes.
+  char junk = stats.opt_junk ? '\xe4' : '\0';
+  for (size_t i = 0; i < stats.page_size; i++) {
+    ASSERT_EQ(junk_buf[i], junk);
+  }
+
+  char* poison_buf = (char*)moz_arena_malloc(buf_arena, stats.page_size);
+  memset(poison_buf, 0xe5, stats.page_size);
+
+  static const char fill = 0x42;
+  char* fill_buf = (char*)moz_arena_malloc(buf_arena, stats.page_size);
+  memset(fill_buf, fill, stats.page_size);
+
+  arena_params_t params;
+  // Allow as many dirty pages in the arena as possible, so that purge never
+  // happens in it. Purge breaks some of the tests below randomly depending on
+  // what other things happen on other threads.
+  params.mMaxDirty = size_t(-1);
+  arena_id_t arena = moz_create_arena_with_params(&params);
+
+  // Allocating should junk the buffer, and freeing should poison the buffer.
+  for (size_t size : sSizes) {
+    if (size <= stats.large_max) {
+      SCOPED_TRACE(testing::Message() << "size = " << size);
+      char* buf = (char*)moz_arena_malloc(arena, size);
+      size_t allocated = moz_malloc_usable_size(buf);
+      if (stats.opt_junk || stats.opt_zero) {
+        ASSERT_NO_FATAL_FAILURE(
+            bulk_compare(buf, 0, allocated, junk_buf, stats.page_size));
+      }
+      moz_arena_free(arena, buf);
+      // We purposefully do a use-after-free here, to check that the data was
+      // poisoned.
+      ASSERT_NO_FATAL_FAILURE(
+          bulk_compare(buf, 0, allocated, poison_buf, stats.page_size));
+    }
+  }
+
+  // Shrinking in the same size class should be in place and poison between the
+  // new allocation size and the old one.
+  size_t prev = 0;
+  for (size_t size : SizeClassesBetween(1, 2 * stats.chunksize)) {
+    SCOPED_TRACE(testing::Message() << "size = " << size);
+    SCOPED_TRACE(testing::Message() << "prev = " << prev);
+    char* ptr = (char*)moz_arena_malloc(arena, size);
+    memset(ptr, fill, moz_malloc_usable_size(ptr));
+    char* ptr2 = (char*)moz_arena_realloc(arena, ptr, prev + 1);
+    ASSERT_EQ(ptr, ptr2);
+    ASSERT_NO_FATAL_FAILURE(
+        bulk_compare(ptr, 0, prev + 1, fill_buf, stats.page_size));
+    ASSERT_NO_FATAL_FAILURE(
+        bulk_compare(ptr, prev + 1, size, poison_buf, stats.page_size));
+    moz_arena_free(arena, ptr);
+    prev = size;
+  }
+
+  // In-place realloc should junk the new bytes when growing and poison the old
+  // bytes when shrinking.
+  for (size_t from_size : SizeClassesBetween(1, 2 * stats.chunksize)) {
+    SCOPED_TRACE(testing::Message() << "from_size = " << from_size);
+    for (size_t to_size : sSizes) {
+      SCOPED_TRACE(testing::Message() << "to_size = " << to_size);
+      if (CanReallocInPlace(from_size, to_size, stats)) {
+        char* ptr = (char*)moz_arena_malloc(arena, from_size);
+        memset(ptr, fill, moz_malloc_usable_size(ptr));
+        char* ptr2 = (char*)moz_arena_realloc(arena, ptr, to_size);
+        ASSERT_EQ(ptr, ptr2);
+        // Shrinking allocation
+        if (from_size >= to_size) {
+          ASSERT_NO_FATAL_FAILURE(
+              bulk_compare(ptr, 0, to_size, fill_buf, stats.page_size));
+          // Huge allocations have guards and will crash when accessing
+          // beyond the valid range.
+          if (to_size > stats.large_max) {
+            size_t page_limit = ALIGNMENT_CEILING(to_size, stats.page_size);
+            ASSERT_NO_FATAL_FAILURE(bulk_compare(ptr, to_size, page_limit,
+                                                 poison_buf, stats.page_size));
+            ASSERT_DEATH_WRAP(ptr[page_limit] = 0, "");
+          } else {
+            ASSERT_NO_FATAL_FAILURE(bulk_compare(ptr, to_size, from_size,
+                                                 poison_buf, stats.page_size));
+          }
+        } else {
+          // Enlarging allocation
+          ASSERT_NO_FATAL_FAILURE(
+              bulk_compare(ptr, 0, from_size, fill_buf, stats.page_size));
+          if (stats.opt_junk || stats.opt_zero) {
+            ASSERT_NO_FATAL_FAILURE(bulk_compare(ptr, from_size, to_size,
+                                                 junk_buf, stats.page_size));
+          }
+          // Huge allocation, so should have a guard page following
+          if (to_size > stats.large_max) {
+            ASSERT_DEATH_WRAP(
+                ptr[ALIGNMENT_CEILING(to_size, stats.page_size)] = 0, "");
+          }
+        }
+        moz_arena_free(arena, ptr2);
+      }
+    }
+  }
+
+  // Growing to a different size class should poison the old allocation,
+  // preserve the original bytes, and junk the new bytes in the new allocation.
+  for (size_t from_size : SizeClassesBetween(1, 2 * stats.chunksize)) {
+    SCOPED_TRACE(testing::Message() << "from_size = " << from_size);
+    for (size_t to_size : sSizes) {
+      if (from_size < to_size && malloc_good_size(to_size) != from_size &&
+          !IsSameRoundedHugeClass(from_size, to_size, stats)) {
+        SCOPED_TRACE(testing::Message() << "to_size = " << to_size);
+        char* ptr = (char*)moz_arena_malloc(arena, from_size);
+        memset(ptr, fill, moz_malloc_usable_size(ptr));
+        // Avoid in-place realloc by allocating a buffer, expecting it to be
+        // right after the buffer we just received. Buffers smaller than the
+        // page size and exactly or larger than the size of the largest large
+        // size class can't be reallocated in-place.
+        char* avoid_inplace = nullptr;
+        if (from_size >= stats.page_size && from_size < stats.large_max) {
+          avoid_inplace = (char*)moz_arena_malloc(arena, stats.page_size);
+          ASSERT_EQ(ptr + from_size, avoid_inplace);
+        }
+        char* ptr2 = (char*)moz_arena_realloc(arena, ptr, to_size);
+        ASSERT_NE(ptr, ptr2);
+        if (from_size <= stats.large_max) {
+          ASSERT_NO_FATAL_FAILURE(
+              bulk_compare(ptr, 0, from_size, poison_buf, stats.page_size));
+        }
+        ASSERT_NO_FATAL_FAILURE(
+            bulk_compare(ptr2, 0, from_size, fill_buf, stats.page_size));
+        if (stats.opt_junk || stats.opt_zero) {
+          size_t rounded_to_size = malloc_good_size(to_size);
+          ASSERT_NE(to_size, rounded_to_size);
+          ASSERT_NO_FATAL_FAILURE(bulk_compare(ptr2, from_size, rounded_to_size,
+                                               junk_buf, stats.page_size));
+        }
+        moz_arena_free(arena, ptr2);
+        moz_arena_free(arena, avoid_inplace);
+      }
+    }
+  }
+
+  // Shrinking to a different size class should poison the old allocation,
+  // preserve the original bytes, and junk the extra bytes in the new
+  // allocation.
+  for (size_t from_size : SizeClassesBetween(1, 2 * stats.chunksize)) {
+    SCOPED_TRACE(testing::Message() << "from_size = " << from_size);
+    for (size_t to_size : sSizes) {
+      if (from_size > to_size &&
+          !CanReallocInPlace(from_size, to_size, stats)) {
+        SCOPED_TRACE(testing::Message() << "to_size = " << to_size);
+        char* ptr = (char*)moz_arena_malloc(arena, from_size);
+        memset(ptr, fill, from_size);
+        char* ptr2 = (char*)moz_arena_realloc(arena, ptr, to_size);
+        ASSERT_NE(ptr, ptr2);
+        if (from_size <= stats.large_max) {
+          ASSERT_NO_FATAL_FAILURE(
+              bulk_compare(ptr, 0, from_size, poison_buf, stats.page_size));
+        }
+        ASSERT_NO_FATAL_FAILURE(
+            bulk_compare(ptr2, 0, to_size, fill_buf, stats.page_size));
+        if (stats.opt_junk || stats.opt_zero) {
+          size_t rounded_to_size = malloc_good_size(to_size);
+          ASSERT_NE(to_size, rounded_to_size);
+          ASSERT_NO_FATAL_FAILURE(bulk_compare(ptr2, from_size, rounded_to_size,
+                                               junk_buf, stats.page_size));
+        }
+        moz_arena_free(arena, ptr2);
+      }
+    }
+  }
+
+  moz_dispose_arena(arena);
+
+  moz_arena_free(buf_arena, poison_buf);
+  moz_arena_free(buf_arena, junk_buf);
+  moz_arena_free(buf_arena, fill_buf);
+  moz_dispose_arena(buf_arena);
+
+  RESTORE_GDB_SLEEP_LOCAL();
+}
+#endif  // !defined(XP_WIN) || !defined(MOZ_CODE_COVERAGE)
+
+TEST(Jemalloc, TrailingGuard)
+{
+  // Disable PHC allocations for this test, because even a single PHC
+  // allocation occurring can throw it off.
+  AutoDisablePHCOnCurrentThread disable;
+
+  jemalloc_stats_t stats;
+  jemalloc_stats(&stats);
+
+  // Avoid death tests adding some unnecessary (long) delays.
+  SAVE_GDB_SLEEP_LOCAL();
+
+  arena_id_t arena = moz_create_arena();
+  ASSERT_TRUE(arena != 0);
+
+  // Do enough large allocations to fill a chunk, and then one additional one,
+  // and check that the guard page is still present after the one-but-last
+  // allocation, i.e. that we didn't allocate the guard.
+  Vector<void*> ptr_list;
+  for (size_t cnt = 0; cnt < stats.large_max / stats.page_size; cnt++) {
+    void* ptr = moz_arena_malloc(arena, stats.page_size);
+    ASSERT_TRUE(ptr != nullptr);
+    ASSERT_TRUE(ptr_list.append(ptr));
+  }
+
+  void* last_ptr_in_chunk = ptr_list[ptr_list.length() - 1];
+  void* extra_ptr = moz_arena_malloc(arena, stats.page_size);
+  void* guard_page = (void*)ALIGNMENT_CEILING(
+      (uintptr_t)last_ptr_in_chunk + stats.page_size, stats.page_size);
+  jemalloc_ptr_info_t info;
+  jemalloc_ptr_info(guard_page, &info);
+  ASSERT_TRUE(jemalloc_ptr_is_freed_page(&info));
+
+  ASSERT_DEATH_WRAP(*(char*)guard_page = 0, "");
+
+  for (void* ptr : ptr_list) {
+    moz_arena_free(arena, ptr);
+  }
+  moz_arena_free(arena, extra_ptr);
+
+  moz_dispose_arena(arena);
+
+  RESTORE_GDB_SLEEP_LOCAL();
+}
+
+TEST(Jemalloc, LeadingGuard)
+{
+  // Disable PHC allocations for this test, because even a single PHC
+  // allocation occurring can throw it off.
+  AutoDisablePHCOnCurrentThread disable;
+
+  jemalloc_stats_t stats;
+  jemalloc_stats(&stats);
+
+  // Avoid death tests adding some unnecessary (long) delays.
+  SAVE_GDB_SLEEP_LOCAL();
+
+  arena_id_t arena = moz_create_arena();
+  ASSERT_TRUE(arena != 0);
+
+  // Do a simple normal allocation, but force all the allocation space
+  // in the chunk to be used up. This allows us to check that we get
+  // the safe area right in the logic that follows (all memory will be
+  // committed and initialized), and it forces this pointer to the start
+  // of the zone to sit at the very start of the usable chunk area.
+  void* ptr = moz_arena_malloc(arena, stats.large_max);
+  ASSERT_TRUE(ptr != nullptr);
+  // If ptr is chunk-aligned, the above allocation went wrong.
+  void* chunk_start = (void*)ALIGNMENT_FLOOR((uintptr_t)ptr, stats.chunksize);
+  ASSERT_NE((uintptr_t)ptr, (uintptr_t)chunk_start);
+  // If ptr is 1 page after the chunk start (so right after the header),
+  // we must have missed adding the guard page.
+  ASSERT_NE((uintptr_t)ptr, (uintptr_t)chunk_start + stats.page_size);
+  // The actual start depends on the amount of metadata versus the page
+  // size, so we can't check equality without pulling in too many
+  // implementation details.
+
+  // Guard page should be right before data area
+  void* guard_page = (void*)(((uintptr_t)ptr) - sizeof(void*));
+  jemalloc_ptr_info_t info;
+  jemalloc_ptr_info(guard_page, &info);
+  ASSERT_TRUE(info.tag == TagUnknown);
+  ASSERT_DEATH_WRAP(*(char*)guard_page = 0, "");
+
+  moz_arena_free(arena, ptr);
+  moz_dispose_arena(arena);
+
+  RESTORE_GDB_SLEEP_LOCAL();
+}
+
+TEST(Jemalloc, DisposeArena)
+{
+  jemalloc_stats_t stats;
+  jemalloc_stats(&stats);
+
+  // Avoid death tests adding some unnecessary (long) delays.
+  SAVE_GDB_SLEEP_LOCAL();
+
+  arena_id_t arena = moz_create_arena();
+  void* ptr = moz_arena_malloc(arena, 42);
+  // Disposing of the arena when it's not empty is a MOZ_CRASH-worthy error.
+  ASSERT_DEATH_WRAP(moz_dispose_arena(arena), "");
+  moz_arena_free(arena, ptr);
+  moz_dispose_arena(arena);
+
+  arena = moz_create_arena();
+  ptr = moz_arena_malloc(arena, stats.page_size * 2);
+  // Disposing of the arena when it's not empty is a MOZ_CRASH-worthy error.
+  ASSERT_DEATH_WRAP(moz_dispose_arena(arena), "");
+  moz_arena_free(arena, ptr);
+  moz_dispose_arena(arena);
+
+  arena = moz_create_arena();
+  ptr = moz_arena_malloc(arena, stats.chunksize * 2);
+#ifdef MOZ_DEBUG
+  // On debug builds, we do the expensive check that arenas are empty.
+  ASSERT_DEATH_WRAP(moz_dispose_arena(arena), "");
+  moz_arena_free(arena, ptr);
+  moz_dispose_arena(arena);
+#else
+  // Currently, the allocator can't trivially check whether the arena is empty
+  // of huge allocations, so disposing of it works.
+  moz_dispose_arena(arena);
+  // But trying to free a pointer that belongs to it will MOZ_CRASH.
+  ASSERT_DEATH_WRAP(free(ptr), "");
+  // Likewise for realloc
+  ASSERT_DEATH_WRAP(ptr = realloc(ptr, stats.chunksize * 3), "");
+#endif
+
+  // Using the arena after it's been disposed of is MOZ_CRASH-worthy.
+  ASSERT_DEATH_WRAP(moz_arena_malloc(arena, 42), "");
+
+  RESTORE_GDB_SLEEP_LOCAL();
+}
diff --git a/memory/gtest/moz.build b/memory/gtest/moz.build
new file mode 100644
index 0000000000..f2a3e3f709
--- /dev/null
+++ b/memory/gtest/moz.build
@@ -0,0 +1,22 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+# skip the test on windows10-aarch64 due to perma-crash - bug 1544961
+if CONFIG["OS_TARGET"] != "Android" and not (
+    CONFIG["OS_TARGET"] == "WINNT" and CONFIG["CPU_ARCH"] == "aarch64"
+):
+    UNIFIED_SOURCES += [
+        "TestJemalloc.cpp",
+    ]
+
+    if CONFIG["MOZ_PHC"]:
+        DEFINES["MOZ_PHC"] = True
+
+FINAL_LIBRARY = "xul-gtest"
+
+LOCAL_INCLUDES += [
+    "../build",
+]
diff --git a/memory/moz.build b/memory/moz.build
new file mode 100644
index 0000000000..e1da146a3c
--- /dev/null
+++ b/memory/moz.build
@@ -0,0 +1,31 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+with Files("**"):
+    BUG_COMPONENT = ("Core", "Memory Allocator")
+
+DIRS += [
+    "build",
+]
+
+if CONFIG["MOZ_BUILD_APP"] == "memory":
+    DIRS += [
+        "mozjemalloc_info",
+    ]
+else:
+    # For now, don't build mozalloc when building with --enable-project=memory
+    DIRS += [
+        "mozalloc",
+    ]
+
+if CONFIG["MOZ_WIDGET_TOOLKIT"]:
+    DIRS += ["volatile"]
+
+# NB: gtest dir is included in toolkit/toolkit.build due to its dependency
+# on libxul.
+
+if CONFIG["MOZ_REPLACE_MALLOC"]:
+    DIRS += ["replace"]
diff --git a/memory/moz.configure b/memory/moz.configure
new file mode 100644
index 0000000000..9745b253cd
--- /dev/null
+++ b/memory/moz.configure
@@ -0,0 +1,7 @@
+# -*- Mode: python; c-basic-offset: 4; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+imply_option("--enable-jemalloc", True)
diff --git a/memory/mozalloc/cxxalloc.cpp b/memory/mozalloc/cxxalloc.cpp
new file mode 100644
index 0000000000..41f419fe2d
--- /dev/null
+++ b/memory/mozalloc/cxxalloc.cpp
@@ -0,0 +1,26 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#define MOZ_MEMORY_IMPL
+#include "mozmemory_wrap.h"
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC
+// See mozmemory_wrap.h for more details. Files that are part of libmozglue,
+// need to use _impl suffixes, which is becoming cumbersome. We'll have to use
+// something like a malloc.h wrapper and allow the use of the functions without
+// a _impl suffix. In the meanwhile, this is enough to get by for C++ code.
+#define MALLOC_DECL(name, return_type, ...) \
+  MOZ_MEMORY_API return_type name##_impl(__VA_ARGS__);
+#include "malloc_decls.h"
+
+#include "mozilla/Attributes.h"
+
+extern "C" MFBT_API void* moz_xmalloc(size_t size) MOZ_INFALLIBLE_ALLOCATOR;
+
+namespace std {
+struct nothrow_t;
+}
+
+#define MOZALLOC_EXPORT_NEW MFBT_API
+
+#include "mozilla/cxxalloc.h"
diff --git a/memory/mozalloc/cxxalloc.h b/memory/mozalloc/cxxalloc.h
new file mode 100644
index 0000000000..c6fb4bb1dc
--- /dev/null
+++ b/memory/mozalloc/cxxalloc.h
@@ -0,0 +1,82 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozilla_cxxalloc_h
+#define mozilla_cxxalloc_h
+
+/*
+ * We implement the default operators new/delete as part of
+ * libmozalloc, replacing their definitions in libstdc++.  The
+ * operator new* definitions in libmozalloc will never return a NULL
+ * pointer.
+ *
+ * Each operator new immediately below returns a pointer to memory
+ * that can be delete'd by any of
+ *
+ *   (1) the matching infallible operator delete immediately below
+ *   (2) the matching system |operator delete(void*, std::nothrow)|
+ *   (3) the matching system |operator delete(void*) noexcept(false)|
+ *
+ * NB: these are declared |noexcept(false)|, though they will never
+ * throw that exception.  This declaration is consistent with the rule
+ * that |::operator new() noexcept(false)| will never return NULL.
+ *
+ * NB: mozilla::fallible can be used instead of std::nothrow.
+ */
+
+#ifndef MOZALLOC_EXPORT_NEW
+#  define MOZALLOC_EXPORT_NEW MFBT_API
+#endif
+
+MOZALLOC_EXPORT_NEW void* operator new(size_t size) noexcept(false) {
+  return moz_xmalloc(size);
+}
+
+MOZALLOC_EXPORT_NEW void* operator new(size_t size,
+                                       const std::nothrow_t&) noexcept(true) {
+  return malloc_impl(size);
+}
+
+MOZALLOC_EXPORT_NEW void* operator new[](size_t size) noexcept(false) {
+  return moz_xmalloc(size);
+}
+
+MOZALLOC_EXPORT_NEW void* operator new[](size_t size,
+                                         const std::nothrow_t&) noexcept(true) {
+  return malloc_impl(size);
+}
+
+MOZALLOC_EXPORT_NEW void operator delete(void* ptr) noexcept(true) {
+  return free_impl(ptr);
+}
+
+MOZALLOC_EXPORT_NEW void operator delete(void* ptr,
+                                         const std::nothrow_t&) noexcept(true) {
+  return free_impl(ptr);
+}
+
+MOZALLOC_EXPORT_NEW void operator delete[](void* ptr) noexcept(true) {
+  return free_impl(ptr);
+}
+
+MOZALLOC_EXPORT_NEW void operator delete[](
+    void* ptr, const std::nothrow_t&) noexcept(true) {
+  return free_impl(ptr);
+}
+
+#if defined(XP_WIN)
+// We provide the global sized delete overloads unconditionally because the
+// MSVC runtime headers do, despite compiling with /Zc:sizedDealloc-
+MOZALLOC_EXPORT_NEW void operator delete(void* ptr,
+                                         size_t /*size*/) noexcept(true) {
+  return free_impl(ptr);
+}
+
+MOZALLOC_EXPORT_NEW void operator delete[](void* ptr,
+                                           size_t /*size*/) noexcept(true) {
+  return free_impl(ptr);
+}
+#endif
+
+#endif /* mozilla_cxxalloc_h */
diff --git a/memory/mozalloc/moz.build b/memory/mozalloc/moz.build
new file mode 100644
index 0000000000..2d13c6cf14
--- /dev/null
+++ b/memory/mozalloc/moz.build
@@ -0,0 +1,59 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+NoVisibilityFlags()
+
+EXPORTS.mozilla += [
+    "cxxalloc.h",
+    "mozalloc.h",
+    "mozalloc_abort.h",
+    "mozalloc_oom.h",
+]
+
+if CONFIG["WRAP_STL_INCLUDES"]:
+    if CONFIG["CC_TYPE"] in ("clang", "gcc"):
+        EXPORTS.mozilla += ["throw_gcc.h"]
+    elif CONFIG["CC_TYPE"] == "clang-cl":
+        DEFINES["_HAS_EXCEPTIONS"] = 0
+        SOURCES += [
+            "msvc_raise_wrappers.cpp",
+        ]
+
+if CONFIG["OS_TARGET"] == "WINNT":
+    # Keep this file separate to avoid #include'ing windows.h everywhere.
+    SOURCES += [
+        "winheap.cpp",
+    ]
+
+UNIFIED_SOURCES += [
+    "mozalloc.cpp",
+    "mozalloc_abort.cpp",
+    "mozalloc_oom.cpp",
+]
+
+if CONFIG["MOZ_MEMORY"]:
+    # In MinGW, we don't want to actually export these functions out of the library
+    # as the functions in libc++ correctly forward to jemalloc and exporting them
+    # produces duplicate symbol errors.
+    if not (CONFIG["CC_TYPE"] == "clang" and CONFIG["OS_TARGET"] == "WINNT"):
+        SOURCES += [
+            "cxxalloc.cpp",
+        ]
+
+FINAL_LIBRARY = "mozglue"
+
+# The strndup declaration in string.h is in an ifdef __USE_GNU section
+DEFINES["_GNU_SOURCE"] = True
+
+DisableStlWrapping()
+
+LOCAL_INCLUDES += [
+    "!/xpcom",
+    "/memory/build",
+]
+
+DIST_INSTALL = True
+
+REQUIRES_UNIFIED_BUILD = True
diff --git a/memory/mozalloc/mozalloc.cpp b/memory/mozalloc/mozalloc.cpp
new file mode 100644
index 0000000000..7e38e3722d
--- /dev/null
+++ b/memory/mozalloc/mozalloc.cpp
@@ -0,0 +1,157 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <stddef.h>  // for size_t
+
+#if defined(MALLOC_H)
+#  include MALLOC_H  // for memalign, malloc_size, malloc_us
+#endif               // if defined(MALLOC_H)
+
+#if !defined(MOZ_MEMORY)
+// When jemalloc is disabled, or when building the static runtime variant,
+// we need not to use the suffixes.
+
+#  include <stdlib.h>  // for malloc, free
+#  if defined(XP_UNIX)
+#    include <unistd.h>
+#  endif  // if defined(XP_UNIX)
+
+#  define malloc_impl malloc
+#  define calloc_impl calloc
+#  define realloc_impl realloc
+#  define free_impl free
+#  define memalign_impl memalign
+#  define malloc_usable_size_impl malloc_usable_size
+#  define strdup_impl strdup
+#  define strndup_impl strndup
+
+#endif
+
+#include <errno.h>
+#include <new>  // for std::bad_alloc
+#include <cstring>
+
+#include <sys/types.h>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/CheckedInt.h"
+#include "mozilla/Likely.h"
+#include "mozilla/mozalloc.h"
+#include "mozilla/mozalloc_oom.h"  // for mozalloc_handle_oom
+
+#if defined(MOZ_MEMORY)
+MOZ_MEMORY_API char* strdup_impl(const char*);
+MOZ_MEMORY_API char* strndup_impl(const char*, size_t);
+#endif
+
+void* moz_xmalloc(size_t size) {
+  void* ptr = malloc_impl(size);
+  if (MOZ_UNLIKELY(!ptr && size)) {
+    mozalloc_handle_oom(size);
+    return moz_xmalloc(size);
+  }
+  return ptr;
+}
+
+void* moz_xcalloc(size_t nmemb, size_t size) {
+  void* ptr = calloc_impl(nmemb, size);
+  if (MOZ_UNLIKELY(!ptr && nmemb && size)) {
+    mozilla::CheckedInt<size_t> totalSize =
+        mozilla::CheckedInt<size_t>(nmemb) * size;
+    mozalloc_handle_oom(totalSize.isValid() ? totalSize.value() : SIZE_MAX);
+    return moz_xcalloc(nmemb, size);
+  }
+  return ptr;
+}
+
+void* moz_xrealloc(void* ptr, size_t size) {
+  void* newptr = realloc_impl(ptr, size);
+  if (MOZ_UNLIKELY(!newptr && size)) {
+    mozalloc_handle_oom(size);
+    return moz_xrealloc(ptr, size);
+  }
+  return newptr;
+}
+
+char* moz_xstrdup(const char* str) {
+  char* dup = strdup_impl(str);
+  if (MOZ_UNLIKELY(!dup)) {
+    mozalloc_handle_oom(0);
+    return moz_xstrdup(str);
+  }
+  return dup;
+}
+
+#if defined(HAVE_STRNDUP)
+char* moz_xstrndup(const char* str, size_t strsize) {
+  char* dup = strndup_impl(str, strsize);
+  if (MOZ_UNLIKELY(!dup)) {
+    mozalloc_handle_oom(strsize);
+    return moz_xstrndup(str, strsize);
+  }
+  return dup;
+}
+#endif  // if defined(HAVE_STRNDUP)
+
+void* moz_xmemdup(const void* ptr, size_t size) {
+  void* newPtr = moz_xmalloc(size);
+  memcpy(newPtr, ptr, size);
+  return newPtr;
+}
+
+#ifndef HAVE_MEMALIGN
+// We always have a definition of memalign, but system headers don't
+// necessarily come with a declaration.
+extern "C" void* memalign(size_t, size_t);
+#endif
+
+void* moz_xmemalign(size_t boundary, size_t size) {
+  void* ptr = memalign_impl(boundary, size);
+  if (MOZ_UNLIKELY(!ptr && EINVAL != errno)) {
+    mozalloc_handle_oom(size);
+    return moz_xmemalign(boundary, size);
+  }
+  // non-NULL ptr or errno == EINVAL
+  return ptr;
+}
+
+size_t moz_malloc_usable_size(void* ptr) {
+  if (!ptr) return 0;
+
+#if defined(XP_DARWIN)
+  return malloc_size(ptr);
+#elif defined(HAVE_MALLOC_USABLE_SIZE) || defined(MOZ_MEMORY)
+  return malloc_usable_size_impl(ptr);
+#elif defined(XP_WIN)
+  return _msize(ptr);
+#else
+  return 0;
+#endif
+}
+
+size_t moz_malloc_size_of(const void* ptr) {
+  return moz_malloc_usable_size((void*)ptr);
+}
+
+#if defined(MOZ_MEMORY)
+#  include "mozjemalloc_types.h"
+// mozmemory.h declares jemalloc_ptr_info(), but including that header in this
+// file is complicated. So we just redeclare it here instead, and include
+// mozjemalloc_types.h for jemalloc_ptr_info_t.
+MOZ_JEMALLOC_API void jemalloc_ptr_info(const void* ptr,
+                                        jemalloc_ptr_info_t* info);
+#endif
+
+size_t moz_malloc_enclosing_size_of(const void* ptr) {
+#if defined(MOZ_MEMORY)
+  jemalloc_ptr_info_t info;
+  jemalloc_ptr_info(ptr, &info);
+  return jemalloc_ptr_is_live(&info) ? info.size : 0;
+#else
+  return 0;
+#endif
+}
diff --git a/memory/mozalloc/mozalloc.h b/memory/mozalloc/mozalloc.h
new file mode 100644
index 0000000000..549f92eded
--- /dev/null
+++ b/memory/mozalloc/mozalloc.h
@@ -0,0 +1,186 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozilla_mozalloc_h
+#define mozilla_mozalloc_h
+
+/*
+ * https://bugzilla.mozilla.org/show_bug.cgi?id=427099
+ */
+
+#if defined(__cplusplus)
+#  include <new>
+// Since libstdc++ 6, including the C headers (e.g. stdlib.h) instead of the
+// corresponding C++ header (e.g. cstdlib) can cause confusion in C++ code
+// using things defined there. Specifically, with stdlib.h, the use of abs()
+// in gfx/graphite2/src/inc/UtfCodec.h somehow ends up picking the wrong abs()
+#  include <cstdlib>
+#else
+#  include <stdlib.h>
+#endif
+
+#if defined(MOZ_MEMORY) && defined(IMPL_MFBT)
+#  define MOZ_MEMORY_IMPL
+#  include "mozmemory_wrap.h"
+#  define MALLOC_FUNCS MALLOC_FUNCS_MALLOC
+// See mozmemory_wrap.h for more details. Files that are part of libmozglue,
+// need to use _impl suffixes, which is becoming cumbersome. We'll have to use
+// something like a malloc.h wrapper and allow the use of the functions without
+// a _impl suffix. In the meanwhile, this is enough to get by for C++ code.
+#  define NOTHROW_MALLOC_DECL(name, return_type, ...) \
+    MOZ_MEMORY_API return_type name##_impl(__VA_ARGS__) noexcept(true);
+#  define MALLOC_DECL(name, return_type, ...) \
+    MOZ_MEMORY_API return_type name##_impl(__VA_ARGS__);
+#  include "malloc_decls.h"
+#endif
+
+#if defined(__cplusplus)
+#  include "mozilla/fallible.h"
+#  include "mozilla/mozalloc_abort.h"
+#  include "mozilla/TemplateLib.h"
+#endif
+#include "mozilla/Attributes.h"
+#include "mozilla/Types.h"
+
+MOZ_BEGIN_EXTERN_C
+
+/*
+ * We need to use malloc_impl and free_impl in this file when they are
+ * defined, because of how mozglue.dll is linked on Windows, where using
+ * malloc/free would end up using the symbols from the MSVCRT instead of
+ * ours.
+ */
+#ifndef free_impl
+#  define free_impl free
+#  define free_impl_
+#endif
+#ifndef malloc_impl
+#  define malloc_impl malloc
+#  define malloc_impl_
+#endif
+
+/*
+ * Each declaration below is analogous to a "standard" allocation
+ * function, except that the out-of-memory handling is made explicit.
+ * The |moz_x| versions will never return a NULL pointer; if memory
+ * is exhausted, they abort.  The |moz_| versions may return NULL
+ * pointers if memory is exhausted: their return value must be checked.
+ *
+ * All these allocation functions are *guaranteed* to return a pointer
+ * to memory allocated in such a way that that memory can be freed by
+ * passing that pointer to |free()|.
+ */
+
+MFBT_API void* moz_xmalloc(size_t size) MOZ_INFALLIBLE_ALLOCATOR;
+
+MFBT_API void* moz_xcalloc(size_t nmemb, size_t size) MOZ_INFALLIBLE_ALLOCATOR;
+
+MFBT_API void* moz_xrealloc(void* ptr, size_t size) MOZ_INFALLIBLE_ALLOCATOR;
+
+MFBT_API char* moz_xstrdup(const char* str) MOZ_INFALLIBLE_ALLOCATOR;
+
+#if defined(HAVE_STRNDUP)
+MFBT_API char* moz_xstrndup(const char* str,
+                            size_t strsize) MOZ_INFALLIBLE_ALLOCATOR;
+#endif /* if defined(HAVE_STRNDUP) */
+
+MFBT_API void* moz_xmemdup(const void* ptr,
+                           size_t size) MOZ_INFALLIBLE_ALLOCATOR;
+
+MFBT_API void* moz_xmemalign(size_t boundary,
+                             size_t size) MOZ_INFALLIBLE_ALLOCATOR;
+
+MFBT_API size_t moz_malloc_usable_size(void* ptr);
+
+MFBT_API size_t moz_malloc_size_of(const void* ptr);
+
+/*
+ * Like moz_malloc_size_of(), but works reliably with interior pointers, i.e.
+ * pointers into the middle of a live allocation.
+ */
+MFBT_API size_t moz_malloc_enclosing_size_of(const void* ptr);
+
+MOZ_END_EXTERN_C
+
+#ifdef __cplusplus
+
+/* NB: This is defined just to silence vacuous warnings about symbol
+ * visibility on OS X/gcc. These symbols are force-inline and not
+ * exported. */
+#  if defined(XP_MACOSX)
+#    define MOZALLOC_EXPORT_NEW MFBT_API MOZ_ALWAYS_INLINE_EVEN_DEBUG
+#  else
+#    define MOZALLOC_EXPORT_NEW MOZ_ALWAYS_INLINE_EVEN_DEBUG
+#  endif
+
+#  include "mozilla/cxxalloc.h"
+
+/*
+ * This policy is identical to MallocAllocPolicy, except it uses
+ * moz_xmalloc/moz_xcalloc/moz_xrealloc instead of
+ * malloc/calloc/realloc.
+ */
+class InfallibleAllocPolicy {
+ public:
+  template <typename T>
+  T* maybe_pod_malloc(size_t aNumElems) {
+    return pod_malloc<T>(aNumElems);
+  }
+
+  template <typename T>
+  T* maybe_pod_calloc(size_t aNumElems) {
+    return pod_calloc<T>(aNumElems);
+  }
+
+  template <typename T>
+  T* maybe_pod_realloc(T* aPtr, size_t aOldSize, size_t aNewSize) {
+    return pod_realloc<T>(aPtr, aOldSize, aNewSize);
+  }
+
+  template <typename T>
+  T* pod_malloc(size_t aNumElems) {
+    if (aNumElems & mozilla::tl::MulOverflowMask<sizeof(T)>::value) {
+      reportAllocOverflow();
+    }
+    return static_cast<T*>(moz_xmalloc(aNumElems * sizeof(T)));
+  }
+
+  template <typename T>
+  T* pod_calloc(size_t aNumElems) {
+    return static_cast<T*>(moz_xcalloc(aNumElems, sizeof(T)));
+  }
+
+  template <typename T>
+  T* pod_realloc(T* aPtr, size_t aOldSize, size_t aNewSize) {
+    if (aNewSize & mozilla::tl::MulOverflowMask<sizeof(T)>::value) {
+      reportAllocOverflow();
+    }
+    return static_cast<T*>(moz_xrealloc(aPtr, aNewSize * sizeof(T)));
+  }
+
+  template <typename T>
+  void free_(T* aPtr, size_t aNumElems = 0) {
+    free_impl(aPtr);
+  }
+
+  void reportAllocOverflow() const { mozalloc_abort("alloc overflow"); }
+
+  bool checkSimulatedOOM() const { return true; }
+};
+
+#endif /* ifdef __cplusplus */
+
+#ifdef malloc_impl_
+#  undef malloc_impl_
+#  undef malloc_impl
+#endif
+#ifdef free_impl_
+#  undef free_impl_
+#  undef free_impl
+#endif
+
+#endif /* ifndef mozilla_mozalloc_h */
diff --git a/memory/mozalloc/mozalloc_abort.cpp b/memory/mozalloc/mozalloc_abort.cpp
new file mode 100644
index 0000000000..3cfc92533e
--- /dev/null
+++ b/memory/mozalloc/mozalloc_abort.cpp
@@ -0,0 +1,96 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "mozilla/mozalloc_abort.h"
+
+#ifdef ANDROID
+#  include <android/log.h>
+#endif
+#ifdef MOZ_WIDGET_ANDROID
+#  include "APKOpen.h"
+#  include "dlfcn.h"
+#endif
+#include <stdio.h>
+#include <string.h>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/Sprintf.h"
+
+void mozalloc_abort(const char* const msg) {
+#ifndef ANDROID
+  fputs(msg, stderr);
+  fputs("\n", stderr);
+#else
+  __android_log_print(ANDROID_LOG_ERROR, "Gecko", "mozalloc_abort: %s", msg);
+#endif
+
+#ifdef MOZ_WIDGET_ANDROID
+  abortThroughJava(msg);
+#endif
+
+  MOZ_CRASH_UNSAFE(msg);
+}
+
+#ifdef MOZ_WIDGET_ANDROID
+template <size_t N>
+void fillAbortMessage(char (&msg)[N], uintptr_t retAddress) {
+  /*
+   * On Android, we often don't have reliable backtrace when crashing inside
+   * abort(). Therefore, we try to find out who is calling abort() and add
+   * that to the message.
+   */
+  Dl_info info = {};
+  dladdr(reinterpret_cast<void*>(retAddress), &info);
+
+  const char* const module = info.dli_fname ? info.dli_fname : "";
+  const char* const base_module = strrchr(module, '/');
+  const void* const module_offset =
+      reinterpret_cast<void*>(retAddress - uintptr_t(info.dli_fbase));
+  const char* const sym = info.dli_sname ? info.dli_sname : "";
+
+  SprintfLiteral(msg, "abort() called from %s:%p (%s)",
+                 base_module ? base_module + 1 : module, module_offset, sym);
+}
+#endif
+
+#if defined(XP_UNIX) && !defined(MOZ_ASAN) && !defined(MOZ_TSAN) && \
+    !defined(LIBFUZZER)
+// Define abort() here, so that it is used instead of the system abort(). This
+// lets us control the behavior when aborting, in order to get better results
+// on *NIX platforms. See mozalloc_abort for details.
+//
+// For AddressSanitizer, we must not redefine system abort because the ASan
+// option "abort_on_error=1" calls abort() and therefore causes the following
+// call chain with our redefined abort:
+//
+// ASan -> abort() -> moz_abort() -> MOZ_CRASH() -> Segmentation fault
+//
+// That segmentation fault will be interpreted as another bug by ASan and as a
+// result, ASan will just exit(1) instead of aborting.
+//
+// The same applies to ThreadSanitizer when run with "halt_on_error=1" in
+// combination with "abort_on_error=1".
+//
+// When building with libFuzzer, it pulls in the UndefinedBehaviorSanitizer
+// runtime which also requires the same workaround as with ASan or TSan.
+extern "C" void abort(void) {
+#  ifdef MOZ_WIDGET_ANDROID
+  char msg[64] = {};
+  fillAbortMessage(msg, uintptr_t(__builtin_return_address(0)));
+#  else
+  const char* const msg = "Redirecting call to abort() to mozalloc_abort\n";
+#  endif
+
+  mozalloc_abort(msg);
+
+  // We won't reach here because mozalloc_abort() is MOZ_NORETURN. But that
+  // annotation isn't used on ARM (see mozalloc_abort.h for why) so we add a
+  // unreachable marker here to avoid a "'noreturn' function does return"
+  // warning.
+  MOZ_ASSUME_UNREACHABLE_MARKER();
+}
+#endif
diff --git a/memory/mozalloc/mozalloc_abort.h b/memory/mozalloc/mozalloc_abort.h
new file mode 100644
index 0000000000..b9ff92a18e
--- /dev/null
+++ b/memory/mozalloc/mozalloc_abort.h
@@ -0,0 +1,28 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozilla_mozalloc_abort_h
+#define mozilla_mozalloc_abort_h
+
+#include "mozilla/Attributes.h"
+#include "mozilla/Types.h"
+
+/**
+ * Terminate this process in such a way that breakpad is triggered, if
+ * at all possible.
+ *
+ * Note: MOZ_NORETURN seems to break crash stacks on ARM, so we don't
+ * use that annotation there.
+ */
+extern "C" MFBT_API
+#if !defined(__arm__)
+    MOZ_NORETURN
+#endif
+    void
+    mozalloc_abort(const char* const msg);
+
+#endif /* ifndef mozilla_mozalloc_abort_h */
diff --git a/memory/mozalloc/mozalloc_oom.cpp b/memory/mozalloc/mozalloc_oom.cpp
new file mode 100644
index 0000000000..efe5dab4a2
--- /dev/null
+++ b/memory/mozalloc/mozalloc_oom.cpp
@@ -0,0 +1,52 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "mozilla/mozalloc_abort.h"
+#include "mozilla/mozalloc_oom.h"
+#include "mozilla/Assertions.h"
+
+#define OOM_MSG_LEADER "out of memory: 0x"
+#define OOM_MSG_DIGITS "0000000000000000"  // large enough for 2^64
+#define OOM_MSG_TRAILER " bytes requested"
+#define OOM_MSG_FIRST_DIGIT_OFFSET sizeof(OOM_MSG_LEADER) - 1
+#define OOM_MSG_LAST_DIGIT_OFFSET \
+  sizeof(OOM_MSG_LEADER) + sizeof(OOM_MSG_DIGITS) - 3
+
+MFBT_DATA size_t gOOMAllocationSize = 0;
+
+static const char* hex = "0123456789ABCDEF";
+
+void mozalloc_handle_oom(size_t size) {
+  char oomMsg[] = OOM_MSG_LEADER OOM_MSG_DIGITS OOM_MSG_TRAILER;
+  size_t i;
+
+  // NB: this is handle_oom() stage 1, which simply aborts on OOM.
+  // we might proceed to a stage 2 in which an attempt is made to
+  // reclaim memory
+  // Warning: when stage 2 is done by, for example, notifying
+  // "memory-pressure" synchronously, please audit all
+  // nsExpirationTrackers and ensure that the actions they take
+  // on memory-pressure notifications (via NotifyExpired) are safe.
+  // Note that Document::SelectorCache::NotifyExpired is _known_
+  // to not be safe: it will delete the selector it's caching,
+  // which might be in use at the time under querySelector or
+  // querySelectorAll.
+
+  gOOMAllocationSize = size;
+
+  static_assert(OOM_MSG_FIRST_DIGIT_OFFSET > 0,
+                "Loop below will never terminate (i can't go below 0)");
+
+  // Insert size into the diagnostic message using only primitive operations
+  for (i = OOM_MSG_LAST_DIGIT_OFFSET; size && i >= OOM_MSG_FIRST_DIGIT_OFFSET;
+       i--) {
+    oomMsg[i] = hex[size % 16];
+    size /= 16;
+  }
+
+  mozalloc_abort(oomMsg);
+}
diff --git a/memory/mozalloc/mozalloc_oom.h b/memory/mozalloc/mozalloc_oom.h
new file mode 100644
index 0000000000..d66d013aaa
--- /dev/null
+++ b/memory/mozalloc/mozalloc_oom.h
@@ -0,0 +1,24 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozilla_mozalloc_oom_h
+#define mozilla_mozalloc_oom_h
+
+#include "mozalloc.h"
+
+/**
+ * Called when memory is critically low.  Returns iff it was able to
+ * remedy the critical memory situation; if not, it will abort().
+ */
+MFBT_API void mozalloc_handle_oom(size_t requestedSize);
+
+extern MFBT_DATA size_t gOOMAllocationSize;
+
+/* TODO: functions to query system memory usage and register
+ * critical-memory handlers. */
+
+#endif /* ifndef mozilla_mozalloc_oom_h */
diff --git a/memory/mozalloc/msvc_raise_wrappers.cpp b/memory/mozalloc/msvc_raise_wrappers.cpp
new file mode 100644
index 0000000000..6eb8bdbe31
--- /dev/null
+++ b/memory/mozalloc/msvc_raise_wrappers.cpp
@@ -0,0 +1,17 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <exception>
+#include "mozalloc_abort.h"
+
+static void __cdecl RaiseHandler(const std::exception& e) {
+  mozalloc_abort(e.what());
+}
+
+static struct StaticScopeStruct final {
+  StaticScopeStruct() { std::exception::_Set_raise_handler(RaiseHandler); }
+} StaticScopeInvoke;
diff --git a/memory/mozalloc/throw_gcc.h b/memory/mozalloc/throw_gcc.h
new file mode 100644
index 0000000000..aeeb770ba7
--- /dev/null
+++ b/memory/mozalloc/throw_gcc.h
@@ -0,0 +1,148 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozilla_throw_gcc_h
+#define mozilla_throw_gcc_h
+
+#include "mozilla/Attributes.h"
+
+#include <stdio.h>   // snprintf
+#include <string.h>  // strerror
+
+// For gcc, we define these inline to abort so that we're absolutely
+// certain that (i) no exceptions are thrown from Gecko; (ii) these
+// errors are always terminal and caught by breakpad.
+
+#include "mozilla/mozalloc_abort.h"
+
+// libc++ 4.0.0 and higher use C++11 [[noreturn]] attributes for the functions
+// below, and since clang does not allow mixing __attribute__((noreturn)) and
+// [[noreturn]], we have to explicitly use the latter here.  See bug 1329520.
+#if defined(__clang__)
+#  if __has_feature(cxx_attributes) && defined(_LIBCPP_VERSION) && \
+      _LIBCPP_VERSION >= 4000
+#    define MOZ_THROW_NORETURN [[noreturn]]
+#  endif
+#endif
+#ifndef MOZ_THROW_NORETURN
+#  define MOZ_THROW_NORETURN MOZ_NORETURN
+#endif
+
+// MinGW doesn't appropriately inline these functions in debug builds,
+// so we need to do some extra coercion for it to do so. Bug 1332747
+#ifdef __MINGW32__
+#  define MOZ_THROW_INLINE MOZ_ALWAYS_INLINE_EVEN_DEBUG
+#  define MOZ_THROW_EXPORT
+#else
+#  define MOZ_THROW_INLINE MOZ_ALWAYS_INLINE
+#  define MOZ_THROW_EXPORT MOZ_EXPORT
+#endif
+
+namespace std {
+
+// NB: user code is not supposed to touch the std:: namespace.  We're
+// doing this after careful review because we want to define our own
+// exception throwing semantics.  Don't try this at home!
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_bad_exception(
+    void) {
+  mozalloc_abort("fatal: STL threw bad_exception");
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_bad_alloc(
+    void) {
+  mozalloc_abort("fatal: STL threw bad_alloc");
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_bad_cast(
+    void) {
+  mozalloc_abort("fatal: STL threw bad_cast");
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_bad_typeid(
+    void) {
+  mozalloc_abort("fatal: STL threw bad_typeid");
+}
+
+// used by <functional>
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void
+__throw_bad_function_call(void) {
+  mozalloc_abort("fatal: STL threw bad_function_call");
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_logic_error(
+    const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_domain_error(
+    const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void
+__throw_invalid_argument(const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_length_error(
+    const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_out_of_range(
+    const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_runtime_error(
+    const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_range_error(
+    const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void
+__throw_overflow_error(const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void
+__throw_underflow_error(const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_ios_failure(
+    const char* msg) {
+  mozalloc_abort(msg);
+}
+
+MOZ_THROW_NORETURN MOZ_THROW_EXPORT MOZ_THROW_INLINE void __throw_system_error(
+    int err) {
+  char error[128];
+  snprintf(error, sizeof(error) - 1, "fatal: STL threw system_error: %s (%d)",
+           strerror(err), err);
+  mozalloc_abort(error);
+}
+
+MOZ_THROW_NORETURN MOZ_EXPORT MOZ_ALWAYS_INLINE void __throw_regex_error(
+    int err) {
+  char error[128];
+  snprintf(error, sizeof(error) - 1, "fatal: STL threw regex_error: %s (%d)",
+           strerror(err), err);
+  mozalloc_abort(error);
+}
+
+}  // namespace std
+
+#undef MOZ_THROW_NORETURN
+#undef MOZ_THROW_INLINE
+
+#endif  // mozilla_throw_gcc_h
diff --git a/memory/mozalloc/winheap.cpp b/memory/mozalloc/winheap.cpp
new file mode 100644
index 0000000000..1d2e1e5599
--- /dev/null
+++ b/memory/mozalloc/winheap.cpp
@@ -0,0 +1,55 @@
+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: sw=2 ts=4 et :
+ */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "mozilla/mozalloc.h"
+#include <windows.h>
+
+#if !defined(MOZ_MEMORY)
+#  include <malloc.h>
+#  define malloc_impl malloc
+#  define calloc_impl calloc
+#  define realloc_impl realloc
+#  define free_impl free
+#endif
+
+// Warning: C4273: 'HeapAlloc': inconsistent dll linkage
+// The Windows headers define HeapAlloc as dllimport, but we define it as
+// dllexport, which is a voluntary inconsistency.
+#pragma warning(disable : 4273)
+
+MFBT_API
+LPVOID WINAPI HeapAlloc(_In_ HANDLE hHeap, _In_ DWORD dwFlags,
+                        _In_ SIZE_T dwBytes) {
+  if (dwFlags & HEAP_ZERO_MEMORY) {
+    return calloc_impl(1, dwBytes);
+  }
+  return malloc_impl(dwBytes);
+}
+
+MFBT_API
+LPVOID WINAPI HeapReAlloc(_In_ HANDLE hHeap, _In_ DWORD dwFlags,
+                          _In_ LPVOID lpMem, _In_ SIZE_T dwBytes) {
+  // The HeapReAlloc contract is that failures preserve the existing
+  // allocation. We can't try to realloc in-place without possibly
+  // freeing the original allocation, breaking the contract.
+  // We also can't guarantee we zero all the memory from the end of
+  // the original allocation to the end of the new one because of the
+  // difference between the originally requested size and what
+  // malloc_usable_size would return us.
+  // So for both cases, just tell the caller we can't do what they
+  // requested.
+  if (dwFlags & (HEAP_REALLOC_IN_PLACE_ONLY | HEAP_ZERO_MEMORY)) {
+    return NULL;
+  }
+  return realloc_impl(lpMem, dwBytes);
+}
+
+MFBT_API
+BOOL WINAPI HeapFree(_In_ HANDLE hHeap, _In_ DWORD dwFlags, _In_ LPVOID lpMem) {
+  free_impl(lpMem);
+  return true;
+}
diff --git a/memory/mozjemalloc_info/MozjemallocInfo.cpp b/memory/mozjemalloc_info/MozjemallocInfo.cpp
new file mode 100644
index 0000000000..0795be8047
--- /dev/null
+++ b/memory/mozjemalloc_info/MozjemallocInfo.cpp
@@ -0,0 +1,46 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "mozmemory.h"
+
+/*
+ * Print the configured size classes which we can then use to update
+ * documentation.
+ */
+int main() {
+  jemalloc_stats_t stats;
+
+  const size_t num_bins = jemalloc_stats_num_bins();
+  const size_t MAX_NUM_BINS = 100;
+  if (num_bins > MAX_NUM_BINS) {
+    fprintf(stderr, "Exceeded maximum number of jemalloc stats bins");
+    return 1;
+  }
+  jemalloc_bin_stats_t bin_stats[MAX_NUM_BINS] = {{0}};
+  jemalloc_stats(&stats, bin_stats);
+
+  printf("Page size:    %5zu\n", stats.page_size);
+  printf("Chunk size:   %5zuKiB\n", stats.chunksize / 1024);
+
+  printf("Quantum:      %5zu\n", stats.quantum);
+  printf("Quantum max:  %5zu\n", stats.quantum_max);
+  printf("Sub-page max: %5zu\n", stats.page_size / 2);
+  printf("Large max:    %5zuKiB\n", stats.large_max / 1024);
+
+  printf("\nBin stats:\n");
+  for (unsigned i = 0; i < num_bins; i++) {
+    auto& bin = bin_stats[i];
+    if (bin.size) {
+      printf("  Bin %5zu has runs of %3zuKiB\n", bin.size,
+             bin.bytes_per_run / 1024);
+    }
+  }
+
+  return EXIT_SUCCESS;
+}
diff --git a/memory/mozjemalloc_info/moz.build b/memory/mozjemalloc_info/moz.build
new file mode 100644
index 0000000000..e7932cb87a
--- /dev/null
+++ b/memory/mozjemalloc_info/moz.build
@@ -0,0 +1,41 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+Program("mozjemalloc-info")
+
+SOURCES += [
+    "/mfbt/Assertions.cpp",
+    "/mfbt/RandomNum.cpp",
+    "/mozglue/misc/StackWalk.cpp",
+    "MozjemallocInfo.cpp",
+]
+
+# Link replace-malloc and the default allocator.
+USE_LIBS += [
+    "memory",
+]
+
+# The memory library defines this, so it's needed here too.
+DEFINES["IMPL_MFBT"] = True
+
+if CONFIG["MOZ_NEEDS_LIBATOMIC"]:
+    OS_LIBS += ["atomic"]
+
+UNIFIED_SOURCES += [
+    "/mfbt/double-conversion/double-conversion/bignum-dtoa.cc",
+    "/mfbt/double-conversion/double-conversion/bignum.cc",
+    "/mfbt/double-conversion/double-conversion/cached-powers.cc",
+    "/mfbt/double-conversion/double-conversion/double-to-string.cc",
+    "/mfbt/double-conversion/double-conversion/fast-dtoa.cc",
+    "/mfbt/double-conversion/double-conversion/fixed-dtoa.cc",
+    "/mozglue/misc/Printf.cpp",
+]
+
+DisableStlWrapping()
+
+include("/mozglue/build/replace_malloc.mozbuild")
+
+REQUIRES_UNIFIED_BUILD = True
diff --git a/memory/replace/dmd/DMD.cpp b/memory/replace/dmd/DMD.cpp
new file mode 100644
index 0000000000..70705d7f2b
--- /dev/null
+++ b/memory/replace/dmd/DMD.cpp
@@ -0,0 +1,1888 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <ctype.h>
+#include <errno.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#if !defined(MOZ_PROFILING)
+#  error "DMD requires MOZ_PROFILING"
+#endif
+
+#ifdef XP_WIN
+#  include <windows.h>
+#  include <process.h>
+#else
+#  include <pthread.h>
+#  include <sys/types.h>
+#  include <unistd.h>
+#endif
+
+#ifdef ANDROID
+#  include <android/log.h>
+#endif
+
+#include "nscore.h"
+
+#include "mozilla/Assertions.h"
+#include "mozilla/FastBernoulliTrial.h"
+#include "mozilla/HashFunctions.h"
+#include "mozilla/HashTable.h"
+#include "mozilla/IntegerPrintfMacros.h"
+#include "mozilla/JSONWriter.h"
+#include "mozilla/Likely.h"
+#include "mozilla/MemoryReporting.h"
+#include "mozilla/PodOperations.h"
+#include "mozilla/StackWalk.h"
+#include "mozilla/ThreadLocal.h"
+
+// CodeAddressService is defined entirely in the header, so this does not make
+// DMD depend on XPCOM's object file.
+#include "CodeAddressService.h"
+
+// replace_malloc.h needs to be included before replace_malloc_bridge.h,
+// which DMD.h includes, so DMD.h needs to be included after replace_malloc.h.
+#include "replace_malloc.h"
+#include "DMD.h"
+
+namespace mozilla {
+namespace dmd {
+
+class DMDBridge : public ReplaceMallocBridge {
+  virtual DMDFuncs* GetDMDFuncs() override;
+};
+
+static DMDBridge* gDMDBridge;
+static DMDFuncs gDMDFuncs;
+
+DMDFuncs* DMDBridge::GetDMDFuncs() { return &gDMDFuncs; }
+
+MOZ_FORMAT_PRINTF(1, 2)
+inline void StatusMsg(const char* aFmt, ...) {
+  va_list ap;
+  va_start(ap, aFmt);
+  gDMDFuncs.StatusMsg(aFmt, ap);
+  va_end(ap);
+}
+
+//---------------------------------------------------------------------------
+// Utilities
+//---------------------------------------------------------------------------
+
+#ifndef DISALLOW_COPY_AND_ASSIGN
+#  define DISALLOW_COPY_AND_ASSIGN(T) \
+    T(const T&);                      \
+    void operator=(const T&)
+#endif
+
+static malloc_table_t gMallocTable;
+
+// This provides infallible allocations (they abort on OOM).  We use it for all
+// of DMD's own allocations, which fall into the following three cases.
+//
+// - Direct allocations (the easy case).
+//
+// - Indirect allocations in mozilla::{Vector,HashSet,HashMap} -- this class
+//   serves as their AllocPolicy.
+//
+// - Other indirect allocations (e.g. MozStackWalk) -- see the comments on
+//   Thread::mBlockIntercepts and in replace_malloc for how these work.
+//
+// It would be nice if we could use the InfallibleAllocPolicy from mozalloc,
+// but DMD cannot use mozalloc.
+//
+class InfallibleAllocPolicy {
+  static void ExitOnFailure(const void* aP);
+
+ public:
+  template <typename T>
+  static T* maybe_pod_malloc(size_t aNumElems) {
+    if (aNumElems & mozilla::tl::MulOverflowMask<sizeof(T)>::value)
+      return nullptr;
+    return (T*)gMallocTable.malloc(aNumElems * sizeof(T));
+  }
+
+  template <typename T>
+  static T* maybe_pod_calloc(size_t aNumElems) {
+    return (T*)gMallocTable.calloc(aNumElems, sizeof(T));
+  }
+
+  template <typename T>
+  static T* maybe_pod_realloc(T* aPtr, size_t aOldSize, size_t aNewSize) {
+    if (aNewSize & mozilla::tl::MulOverflowMask<sizeof(T)>::value)
+      return nullptr;
+    return (T*)gMallocTable.realloc(aPtr, aNewSize * sizeof(T));
+  }
+
+  static void* malloc_(size_t aSize) {
+    void* p = gMallocTable.malloc(aSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  template <typename T>
+  static T* pod_malloc(size_t aNumElems) {
+    T* p = maybe_pod_malloc<T>(aNumElems);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  static void* calloc_(size_t aCount, size_t aSize) {
+    void* p = gMallocTable.calloc(aCount, aSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  template <typename T>
+  static T* pod_calloc(size_t aNumElems) {
+    T* p = maybe_pod_calloc<T>(aNumElems);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  static void* realloc_(void* aPtr, size_t aNewSize) {
+    void* p = gMallocTable.realloc(aPtr, aNewSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  template <typename T>
+  static T* pod_realloc(T* aPtr, size_t aOldSize, size_t aNewSize) {
+    T* p = maybe_pod_realloc(aPtr, aOldSize, aNewSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  static void* memalign_(size_t aAlignment, size_t aSize) {
+    void* p = gMallocTable.memalign(aAlignment, aSize);
+    ExitOnFailure(p);
+    return p;
+  }
+
+  template <typename T>
+  static void free_(T* aPtr, size_t aSize = 0) {
+    gMallocTable.free(aPtr);
+  }
+
+  static char* strdup_(const char* aStr) {
+    char* s = (char*)InfallibleAllocPolicy::malloc_(strlen(aStr) + 1);
+    strcpy(s, aStr);
+    return s;
+  }
+
+  template <class T>
+  static T* new_() {
+    void* mem = malloc_(sizeof(T));
+    return new (mem) T;
+  }
+
+  template <class T, typename P1>
+  static T* new_(const P1& aP1) {
+    void* mem = malloc_(sizeof(T));
+    return new (mem) T(aP1);
+  }
+
+  template <class T>
+  static void delete_(T* aPtr) {
+    if (aPtr) {
+      aPtr->~T();
+      InfallibleAllocPolicy::free_(aPtr);
+    }
+  }
+
+  static void reportAllocOverflow() { ExitOnFailure(nullptr); }
+  bool checkSimulatedOOM() const { return true; }
+};
+
+// This is only needed because of the |const void*| vs |void*| arg mismatch.
+static size_t MallocSizeOf(const void* aPtr) {
+  return gMallocTable.malloc_usable_size(const_cast<void*>(aPtr));
+}
+
+void DMDFuncs::StatusMsg(const char* aFmt, va_list aAp) {
+#ifdef ANDROID
+  __android_log_vprint(ANDROID_LOG_INFO, "DMD", aFmt, aAp);
+#else
+  // The +64 is easily enough for the "DMD[<pid>] " prefix and the NUL.
+  size_t size = strlen(aFmt) + 64;
+  char* fmt = (char*)InfallibleAllocPolicy::malloc_(size);
+  snprintf(fmt, size, "DMD[%d] %s", getpid(), aFmt);
+  vfprintf(stderr, fmt, aAp);
+  InfallibleAllocPolicy::free_(fmt);
+#endif
+}
+
+/* static */
+void InfallibleAllocPolicy::ExitOnFailure(const void* aP) {
+  if (!aP) {
+    MOZ_CRASH("DMD out of memory; aborting");
+  }
+}
+
+static double Percent(size_t part, size_t whole) {
+  return (whole == 0) ? 0 : 100 * (double)part / whole;
+}
+
+// Commifies the number.
+static char* Show(size_t n, char* buf, size_t buflen) {
+  int nc = 0, i = 0, lasti = buflen - 2;
+  buf[lasti + 1] = '\0';
+  if (n == 0) {
+    buf[lasti - i] = '0';
+    i++;
+  } else {
+    while (n > 0) {
+      if (((i - nc) % 3) == 0 && i != 0) {
+        buf[lasti - i] = ',';
+        i++;
+        nc++;
+      }
+      buf[lasti - i] = static_cast<char>((n % 10) + '0');
+      i++;
+      n /= 10;
+    }
+  }
+  int firstCharIndex = lasti - i + 1;
+
+  MOZ_ASSERT(firstCharIndex >= 0);
+  return &buf[firstCharIndex];
+}
+
+//---------------------------------------------------------------------------
+// Options (Part 1)
+//---------------------------------------------------------------------------
+
+class Options {
+  template <typename T>
+  struct NumOption {
+    const T mDefault;
+    const T mMax;
+    T mActual;
+    NumOption(T aDefault, T aMax)
+        : mDefault(aDefault), mMax(aMax), mActual(aDefault) {}
+  };
+
+  // DMD has several modes. These modes affect what data is recorded and
+  // written to the output file, and the written data affects the
+  // post-processing that dmd.py can do.
+  //
+  // Users specify the mode as soon as DMD starts. This leads to minimal memory
+  // usage and log file size. It has the disadvantage that is inflexible -- if
+  // you want to change modes you have to re-run DMD. But in practice changing
+  // modes seems to be rare, so it's not much of a problem.
+  //
+  // An alternative possibility would be to always record and output *all* the
+  // information needed for all modes. This would let you choose the mode when
+  // running dmd.py, and so you could do multiple kinds of profiling on a
+  // single DMD run. But if you are only interested in one of the simpler
+  // modes, you'd pay the price of (a) increased memory usage and (b) *very*
+  // large log files.
+  //
+  // Finally, another alternative possibility would be to do mode selection
+  // partly at DMD startup or recording, and then partly in dmd.py. This would
+  // give some extra flexibility at moderate memory and file size cost. But
+  // certain mode pairs wouldn't work, which would be confusing.
+  //
+  enum class Mode {
+    // For each live block, this mode outputs: size (usable and slop) and
+    // (possibly) and allocation stack. This mode is good for live heap
+    // profiling.
+    Live,
+
+    // Like "Live", but for each live block it also outputs: zero or more
+    // report stacks. This mode is good for identifying where memory reporters
+    // should be added. This is the default mode.
+    DarkMatter,
+
+    // Like "Live", but also outputs the same data for dead blocks. This mode
+    // does cumulative heap profiling, which is good for identifying where large
+    // amounts of short-lived allocations ("heap churn") occur.
+    Cumulative,
+
+    // Like "Live", but this mode also outputs for each live block the address
+    // of the block and the values contained in the blocks. This mode is useful
+    // for investigating leaks, by helping to figure out which blocks refer to
+    // other blocks. This mode force-enables full stacks coverage.
+    Scan
+  };
+
+  // With full stacks, every heap block gets a stack trace recorded for it.
+  // This is complete but slow.
+  //
+  // With partial stacks, not all heap blocks will get a stack trace recorded.
+  // A Bernoulli trial (see mfbt/FastBernoulliTrial.h for details) is performed
+  // for each heap block to decide if it gets one. Because bigger heap blocks
+  // are more likely to get a stack trace, even though most heap *blocks* won't
+  // get a stack trace, most heap *bytes* will.
+  enum class Stacks { Full, Partial };
+
+  char* mDMDEnvVar;  // a saved copy, for later printing
+
+  Mode mMode;
+  Stacks mStacks;
+  bool mShowDumpStats;
+
+  void BadArg(const char* aArg);
+  static const char* ValueIfMatch(const char* aArg, const char* aOptionName);
+  static bool GetLong(const char* aArg, const char* aOptionName, long aMin,
+                      long aMax, long* aValue);
+  static bool GetBool(const char* aArg, const char* aOptionName, bool* aValue);
+
+ public:
+  explicit Options(const char* aDMDEnvVar);
+
+  bool IsLiveMode() const { return mMode == Mode::Live; }
+  bool IsDarkMatterMode() const { return mMode == Mode::DarkMatter; }
+  bool IsCumulativeMode() const { return mMode == Mode::Cumulative; }
+  bool IsScanMode() const { return mMode == Mode::Scan; }
+
+  const char* ModeString() const;
+
+  const char* DMDEnvVar() const { return mDMDEnvVar; }
+
+  bool DoFullStacks() const { return mStacks == Stacks::Full; }
+  size_t ShowDumpStats() const { return mShowDumpStats; }
+};
+
+static Options* gOptions;
+
+//---------------------------------------------------------------------------
+// The global lock
+//---------------------------------------------------------------------------
+
+// MutexBase implements the platform-specific parts of a mutex.
+
+#ifdef XP_WIN
+
+class MutexBase {
+  CRITICAL_SECTION mCS;
+
+  DISALLOW_COPY_AND_ASSIGN(MutexBase);
+
+ public:
+  MutexBase() { InitializeCriticalSection(&mCS); }
+  ~MutexBase() { DeleteCriticalSection(&mCS); }
+
+  void Lock() { EnterCriticalSection(&mCS); }
+  void Unlock() { LeaveCriticalSection(&mCS); }
+};
+
+#else
+
+class MutexBase {
+  pthread_mutex_t mMutex;
+
+  MutexBase(const MutexBase&) = delete;
+
+  const MutexBase& operator=(const MutexBase&) = delete;
+
+ public:
+  MutexBase() { pthread_mutex_init(&mMutex, nullptr); }
+
+  void Lock() { pthread_mutex_lock(&mMutex); }
+  void Unlock() { pthread_mutex_unlock(&mMutex); }
+};
+
+#endif
+
+class Mutex : private MutexBase {
+  bool mIsLocked;
+
+  Mutex(const Mutex&) = delete;
+
+  const Mutex& operator=(const Mutex&) = delete;
+
+ public:
+  Mutex() : mIsLocked(false) {}
+
+  void Lock() {
+    MutexBase::Lock();
+    MOZ_ASSERT(!mIsLocked);
+    mIsLocked = true;
+  }
+
+  void Unlock() {
+    MOZ_ASSERT(mIsLocked);
+    mIsLocked = false;
+    MutexBase::Unlock();
+  }
+
+  bool IsLocked() { return mIsLocked; }
+};
+
+// This lock must be held while manipulating global state such as
+// gStackTraceTable, gLiveBlockTable, gDeadBlockTable. Note that gOptions is
+// *not* protected by this lock because it is only written to by Options(),
+// which is only invoked at start-up and in ResetEverything(), which is only
+// used by SmokeDMD.cpp.
+static Mutex* gStateLock = nullptr;
+
+class AutoLockState {
+  AutoLockState(const AutoLockState&) = delete;
+
+  const AutoLockState& operator=(const AutoLockState&) = delete;
+
+ public:
+  AutoLockState() { gStateLock->Lock(); }
+  ~AutoLockState() { gStateLock->Unlock(); }
+};
+
+class AutoUnlockState {
+  AutoUnlockState(const AutoUnlockState&) = delete;
+
+  const AutoUnlockState& operator=(const AutoUnlockState&) = delete;
+
+ public:
+  AutoUnlockState() { gStateLock->Unlock(); }
+  ~AutoUnlockState() { gStateLock->Lock(); }
+};
+
+//---------------------------------------------------------------------------
+// Per-thread blocking of intercepts
+//---------------------------------------------------------------------------
+
+// On MacOS, the first __thread/thread_local access calls malloc, which leads
+// to an infinite loop. So we use pthread-based TLS instead, which somehow
+// doesn't have this problem.
+#if !defined(XP_DARWIN)
+#  define DMD_THREAD_LOCAL(T) MOZ_THREAD_LOCAL(T)
+#else
+#  define DMD_THREAD_LOCAL(T) \
+    detail::ThreadLocal<T, detail::ThreadLocalKeyStorage>
+#endif
+
+class Thread {
+  // Required for allocation via InfallibleAllocPolicy::new_.
+  friend class InfallibleAllocPolicy;
+
+  // When true, this blocks intercepts, which allows malloc interception
+  // functions to themselves call malloc.  (Nb: for direct calls to malloc we
+  // can just use InfallibleAllocPolicy::{malloc_,new_}, but we sometimes
+  // indirectly call vanilla malloc via functions like MozStackWalk.)
+  bool mBlockIntercepts;
+
+  Thread() : mBlockIntercepts(false) {}
+
+  Thread(const Thread&) = delete;
+
+  const Thread& operator=(const Thread&) = delete;
+
+  static DMD_THREAD_LOCAL(Thread*) tlsThread;
+
+ public:
+  static void Init() {
+    if (!tlsThread.init()) {
+      MOZ_CRASH();
+    }
+  }
+
+  static Thread* Fetch() {
+    Thread* t = tlsThread.get();
+    if (MOZ_UNLIKELY(!t)) {
+      // This memory is never freed, even if the thread dies. It's a leak, but
+      // only a tiny one.
+      t = InfallibleAllocPolicy::new_<Thread>();
+      tlsThread.set(t);
+    }
+
+    return t;
+  }
+
+  bool BlockIntercepts() {
+    MOZ_ASSERT(!mBlockIntercepts);
+    return mBlockIntercepts = true;
+  }
+
+  bool UnblockIntercepts() {
+    MOZ_ASSERT(mBlockIntercepts);
+    return mBlockIntercepts = false;
+  }
+
+  bool InterceptsAreBlocked() const { return mBlockIntercepts; }
+};
+
+DMD_THREAD_LOCAL(Thread*) Thread::tlsThread;
+
+// An object of this class must be created (on the stack) before running any
+// code that might allocate.
+class AutoBlockIntercepts {
+  Thread* const mT;
+
+  AutoBlockIntercepts(const AutoBlockIntercepts&) = delete;
+
+  const AutoBlockIntercepts& operator=(const AutoBlockIntercepts&) = delete;
+
+ public:
+  explicit AutoBlockIntercepts(Thread* aT) : mT(aT) { mT->BlockIntercepts(); }
+  ~AutoBlockIntercepts() {
+    MOZ_ASSERT(mT->InterceptsAreBlocked());
+    mT->UnblockIntercepts();
+  }
+};
+
+//---------------------------------------------------------------------------
+// Location service
+//---------------------------------------------------------------------------
+
+struct DescribeCodeAddressLock {
+  static void Unlock() { gStateLock->Unlock(); }
+  static void Lock() { gStateLock->Lock(); }
+  static bool IsLocked() { return gStateLock->IsLocked(); }
+};
+
+typedef CodeAddressService<InfallibleAllocPolicy, DescribeCodeAddressLock>
+    CodeAddressService;
+
+//---------------------------------------------------------------------------
+// Stack traces
+//---------------------------------------------------------------------------
+
+class StackTrace {
+ public:
+  static const uint32_t MaxFrames = 24;
+
+ private:
+  uint32_t mLength;             // The number of PCs.
+  const void* mPcs[MaxFrames];  // The PCs themselves.
+
+ public:
+  StackTrace() : mLength(0) {}
+  StackTrace(const StackTrace& aOther) : mLength(aOther.mLength) {
+    PodCopy(mPcs, aOther.mPcs, mLength);
+  }
+
+  uint32_t Length() const { return mLength; }
+  const void* Pc(uint32_t i) const {
+    MOZ_ASSERT(i < mLength);
+    return mPcs[i];
+  }
+
+  uint32_t Size() const { return mLength * sizeof(mPcs[0]); }
+
+  // The stack trace returned by this function is interned in gStackTraceTable,
+  // and so is immortal and unmovable.
+  static const StackTrace* Get(Thread* aT);
+
+  // Hash policy.
+
+  typedef StackTrace* Lookup;
+
+  static mozilla::HashNumber hash(const StackTrace* const& aSt) {
+    return mozilla::HashBytes(aSt->mPcs, aSt->Size());
+  }
+
+  static bool match(const StackTrace* const& aA, const StackTrace* const& aB) {
+    return aA->mLength == aB->mLength &&
+           memcmp(aA->mPcs, aB->mPcs, aA->Size()) == 0;
+  }
+
+ private:
+  static void StackWalkCallback(uint32_t aFrameNumber, void* aPc, void* aSp,
+                                void* aClosure) {
+    StackTrace* st = (StackTrace*)aClosure;
+    MOZ_ASSERT(st->mLength < MaxFrames);
+    st->mPcs[st->mLength] = aPc;
+    st->mLength++;
+    MOZ_ASSERT(st->mLength == aFrameNumber);
+  }
+};
+
+typedef mozilla::HashSet<StackTrace*, StackTrace, InfallibleAllocPolicy>
+    StackTraceTable;
+static StackTraceTable* gStackTraceTable = nullptr;
+
+typedef mozilla::HashSet<const StackTrace*,
+                         mozilla::DefaultHasher<const StackTrace*>,
+                         InfallibleAllocPolicy>
+    StackTraceSet;
+
+typedef mozilla::HashSet<const void*, mozilla::DefaultHasher<const void*>,
+                         InfallibleAllocPolicy>
+    PointerSet;
+typedef mozilla::HashMap<const void*, uint32_t,
+                         mozilla::DefaultHasher<const void*>,
+                         InfallibleAllocPolicy>
+    PointerIdMap;
+
+// We won't GC the stack trace table until it this many elements.
+static uint32_t gGCStackTraceTableWhenSizeExceeds = 4 * 1024;
+
+/* static */ const StackTrace* StackTrace::Get(Thread* aT) {
+  MOZ_ASSERT(gStateLock->IsLocked());
+  MOZ_ASSERT(aT->InterceptsAreBlocked());
+
+  // On Windows, MozStackWalk can acquire a lock from the shared library
+  // loader.  Another thread might call malloc while holding that lock (when
+  // loading a shared library).  So we can't be in gStateLock during the call
+  // to MozStackWalk.  For details, see
+  // https://bugzilla.mozilla.org/show_bug.cgi?id=374829#c8
+  // On Linux, something similar can happen;  see bug 824340.
+  // So let's just release it on all platforms.
+  StackTrace tmp;
+  {
+    AutoUnlockState unlock;
+    // In each of the following cases, skipFrames is chosen so that the
+    // first frame in each stack trace is a replace_* function (or as close as
+    // possible, given the vagaries of inlining on different platforms).
+#if defined(XP_WIN) && defined(_M_IX86)
+    // This avoids MozStackWalk(), which causes unusably slow startup on Win32
+    // when it is called during static initialization (see bug 1241684).
+    //
+    // This code is cribbed from the Gecko Profiler, which also uses
+    // FramePointerStackWalk() on Win32: Registers::SyncPopulate() for the
+    // frame pointer, and GetStackTop() for the stack end.
+    CONTEXT context;
+    RtlCaptureContext(&context);
+    void** fp = reinterpret_cast<void**>(context.Ebp);
+
+    PNT_TIB pTib = reinterpret_cast<PNT_TIB>(NtCurrentTeb());
+    void* stackEnd = static_cast<void*>(pTib->StackBase);
+    FramePointerStackWalk(StackWalkCallback, MaxFrames, &tmp, fp, stackEnd);
+#elif defined(XP_MACOSX)
+    // This avoids MozStackWalk(), which has become unusably slow on Mac due to
+    // changes in libunwind.
+    //
+    // This code is cribbed from the Gecko Profiler, which also uses
+    // FramePointerStackWalk() on Mac: Registers::SyncPopulate() for the frame
+    // pointer, and GetStackTop() for the stack end.
+#  pragma GCC diagnostic push
+#  pragma GCC diagnostic ignored "-Wframe-address"
+    void** fp = reinterpret_cast<void**>(__builtin_frame_address(1));
+#  pragma GCC diagnostic pop
+    void* stackEnd = pthread_get_stackaddr_np(pthread_self());
+    FramePointerStackWalk(StackWalkCallback, MaxFrames, &tmp, fp, stackEnd);
+#else
+    MozStackWalk(StackWalkCallback, nullptr, MaxFrames, &tmp);
+#endif
+  }
+
+  StackTraceTable::AddPtr p = gStackTraceTable->lookupForAdd(&tmp);
+  if (!p) {
+    StackTrace* stnew = InfallibleAllocPolicy::new_<StackTrace>(tmp);
+    MOZ_ALWAYS_TRUE(gStackTraceTable->add(p, stnew));
+  }
+  return *p;
+}
+
+//---------------------------------------------------------------------------
+// Heap blocks
+//---------------------------------------------------------------------------
+
+// This class combines a 2-byte-aligned pointer (i.e. one whose bottom bit
+// is zero) with a 1-bit tag.
+//
+// |T| is the pointer type, e.g. |int*|, not the pointed-to type.  This makes
+// is easier to have const pointers, e.g. |TaggedPtr<const int*>|.
+template <typename T>
+class TaggedPtr {
+  union {
+    T mPtr;
+    uintptr_t mUint;
+  };
+
+  static const uintptr_t kTagMask = uintptr_t(0x1);
+  static const uintptr_t kPtrMask = ~kTagMask;
+
+  static bool IsTwoByteAligned(T aPtr) {
+    return (uintptr_t(aPtr) & kTagMask) == 0;
+  }
+
+ public:
+  TaggedPtr() : mPtr(nullptr) {}
+
+  TaggedPtr(T aPtr, bool aBool) : mPtr(aPtr) {
+    MOZ_ASSERT(IsTwoByteAligned(aPtr));
+    uintptr_t tag = uintptr_t(aBool);
+    MOZ_ASSERT(tag <= kTagMask);
+    mUint |= (tag & kTagMask);
+  }
+
+  void Set(T aPtr, bool aBool) {
+    MOZ_ASSERT(IsTwoByteAligned(aPtr));
+    mPtr = aPtr;
+    uintptr_t tag = uintptr_t(aBool);
+    MOZ_ASSERT(tag <= kTagMask);
+    mUint |= (tag & kTagMask);
+  }
+
+  T Ptr() const { return reinterpret_cast<T>(mUint & kPtrMask); }
+
+  bool Tag() const { return bool(mUint & kTagMask); }
+};
+
+// A live heap block. Stores both basic data and data about reports, if we're
+// in DarkMatter mode.
+class LiveBlock {
+  const void* mPtr;
+  const size_t mReqSize;  // size requested
+
+  // The stack trace where this block was allocated, or nullptr if we didn't
+  // record one.
+  const StackTrace* const mAllocStackTrace;
+
+  // This array has two elements because we record at most two reports of a
+  // block.
+  // - Ptr: |mReportStackTrace| - stack trace where this block was reported.
+  //   nullptr if not reported.
+  // - Tag bit 0: |mReportedOnAlloc| - was the block reported immediately on
+  //   allocation?  If so, DMD must not clear the report at the end of
+  //   Analyze(). Only relevant if |mReportStackTrace| is non-nullptr.
+  //
+  // |mPtr| is used as the key in LiveBlockTable, so it's ok for this member
+  // to be |mutable|.
+  //
+  // Only used in DarkMatter mode.
+  mutable TaggedPtr<const StackTrace*> mReportStackTrace_mReportedOnAlloc[2];
+
+ public:
+  LiveBlock(const void* aPtr, size_t aReqSize,
+            const StackTrace* aAllocStackTrace)
+      : mPtr(aPtr),
+        mReqSize(aReqSize),
+        mAllocStackTrace(aAllocStackTrace),
+        mReportStackTrace_mReportedOnAlloc()  // all fields get zeroed
+  {}
+
+  const void* Address() const { return mPtr; }
+
+  size_t ReqSize() const { return mReqSize; }
+
+  size_t SlopSize() const { return MallocSizeOf(mPtr) - mReqSize; }
+
+  const StackTrace* AllocStackTrace() const { return mAllocStackTrace; }
+
+  const StackTrace* ReportStackTrace1() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    return mReportStackTrace_mReportedOnAlloc[0].Ptr();
+  }
+
+  const StackTrace* ReportStackTrace2() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    return mReportStackTrace_mReportedOnAlloc[1].Ptr();
+  }
+
+  bool ReportedOnAlloc1() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    return mReportStackTrace_mReportedOnAlloc[0].Tag();
+  }
+
+  bool ReportedOnAlloc2() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    return mReportStackTrace_mReportedOnAlloc[1].Tag();
+  }
+
+  void AddStackTracesToTable(StackTraceSet& aStackTraces) const {
+    if (AllocStackTrace()) {
+      MOZ_ALWAYS_TRUE(aStackTraces.put(AllocStackTrace()));
+    }
+    if (gOptions->IsDarkMatterMode()) {
+      if (ReportStackTrace1()) {
+        MOZ_ALWAYS_TRUE(aStackTraces.put(ReportStackTrace1()));
+      }
+      if (ReportStackTrace2()) {
+        MOZ_ALWAYS_TRUE(aStackTraces.put(ReportStackTrace2()));
+      }
+    }
+  }
+
+  uint32_t NumReports() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    if (ReportStackTrace2()) {
+      MOZ_ASSERT(ReportStackTrace1());
+      return 2;
+    }
+    if (ReportStackTrace1()) {
+      return 1;
+    }
+    return 0;
+  }
+
+  // This is |const| thanks to the |mutable| fields above.
+  void Report(Thread* aT, bool aReportedOnAlloc) const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    // We don't bother recording reports after the 2nd one.
+    uint32_t numReports = NumReports();
+    if (numReports < 2) {
+      mReportStackTrace_mReportedOnAlloc[numReports].Set(StackTrace::Get(aT),
+                                                         aReportedOnAlloc);
+    }
+  }
+
+  void UnreportIfNotReportedOnAlloc() const {
+    MOZ_ASSERT(gOptions->IsDarkMatterMode());
+    if (!ReportedOnAlloc1() && !ReportedOnAlloc2()) {
+      mReportStackTrace_mReportedOnAlloc[0].Set(nullptr, 0);
+      mReportStackTrace_mReportedOnAlloc[1].Set(nullptr, 0);
+
+    } else if (!ReportedOnAlloc1() && ReportedOnAlloc2()) {
+      // Shift the 2nd report down to the 1st one.
+      mReportStackTrace_mReportedOnAlloc[0] =
+          mReportStackTrace_mReportedOnAlloc[1];
+      mReportStackTrace_mReportedOnAlloc[1].Set(nullptr, 0);
+
+    } else if (ReportedOnAlloc1() && !ReportedOnAlloc2()) {
+      mReportStackTrace_mReportedOnAlloc[1].Set(nullptr, 0);
+    }
+  }
+
+  // Hash policy.
+
+  typedef const void* Lookup;
+
+  static mozilla::HashNumber hash(const void* const& aPtr) {
+    return mozilla::HashGeneric(aPtr);
+  }
+
+  static bool match(const LiveBlock& aB, const void* const& aPtr) {
+    return aB.mPtr == aPtr;
+  }
+};
+
+// A table of live blocks where the lookup key is the block address.
+typedef mozilla::HashSet<LiveBlock, LiveBlock, InfallibleAllocPolicy>
+    LiveBlockTable;
+static LiveBlockTable* gLiveBlockTable = nullptr;
+
+class AggregatedLiveBlockHashPolicy {
+ public:
+  typedef const LiveBlock* const Lookup;
+
+  static mozilla::HashNumber hash(const LiveBlock* const& aB) {
+    return gOptions->IsDarkMatterMode()
+               ? mozilla::HashGeneric(
+                     aB->ReqSize(), aB->SlopSize(), aB->AllocStackTrace(),
+                     aB->ReportedOnAlloc1(), aB->ReportedOnAlloc2())
+               : mozilla::HashGeneric(aB->ReqSize(), aB->SlopSize(),
+                                      aB->AllocStackTrace());
+  }
+
+  static bool match(const LiveBlock* const& aA, const LiveBlock* const& aB) {
+    return gOptions->IsDarkMatterMode()
+               ? aA->ReqSize() == aB->ReqSize() &&
+                     aA->SlopSize() == aB->SlopSize() &&
+                     aA->AllocStackTrace() == aB->AllocStackTrace() &&
+                     aA->ReportStackTrace1() == aB->ReportStackTrace1() &&
+                     aA->ReportStackTrace2() == aB->ReportStackTrace2()
+               : aA->ReqSize() == aB->ReqSize() &&
+                     aA->SlopSize() == aB->SlopSize() &&
+                     aA->AllocStackTrace() == aB->AllocStackTrace();
+  }
+};
+
+// A table of live blocks where the lookup key is everything but the block
+// address. For aggregating similar live blocks at output time.
+typedef mozilla::HashMap<const LiveBlock*, size_t,
+                         AggregatedLiveBlockHashPolicy, InfallibleAllocPolicy>
+    AggregatedLiveBlockTable;
+
+// A freed heap block.
+class DeadBlock {
+  const size_t mReqSize;   // size requested
+  const size_t mSlopSize;  // slop above size requested
+
+  // The stack trace where this block was allocated.
+  const StackTrace* const mAllocStackTrace;
+
+ public:
+  DeadBlock() : mReqSize(0), mSlopSize(0), mAllocStackTrace(nullptr) {}
+
+  explicit DeadBlock(const LiveBlock& aLb)
+      : mReqSize(aLb.ReqSize()),
+        mSlopSize(aLb.SlopSize()),
+        mAllocStackTrace(aLb.AllocStackTrace()) {}
+
+  ~DeadBlock() {}
+
+  size_t ReqSize() const { return mReqSize; }
+  size_t SlopSize() const { return mSlopSize; }
+
+  const StackTrace* AllocStackTrace() const { return mAllocStackTrace; }
+
+  void AddStackTracesToTable(StackTraceSet& aStackTraces) const {
+    if (AllocStackTrace()) {
+      MOZ_ALWAYS_TRUE(aStackTraces.put(AllocStackTrace()));
+    }
+  }
+
+  // Hash policy.
+
+  typedef DeadBlock Lookup;
+
+  static mozilla::HashNumber hash(const DeadBlock& aB) {
+    return mozilla::HashGeneric(aB.ReqSize(), aB.SlopSize(),
+                                aB.AllocStackTrace());
+  }
+
+  static bool match(const DeadBlock& aA, const DeadBlock& aB) {
+    return aA.ReqSize() == aB.ReqSize() && aA.SlopSize() == aB.SlopSize() &&
+           aA.AllocStackTrace() == aB.AllocStackTrace();
+  }
+};
+
+// For each unique DeadBlock value we store a count of how many actual dead
+// blocks have that value.
+typedef mozilla::HashMap<DeadBlock, size_t, DeadBlock, InfallibleAllocPolicy>
+    DeadBlockTable;
+static DeadBlockTable* gDeadBlockTable = nullptr;
+
+// Add the dead block to the dead block table, if that's appropriate.
+void MaybeAddToDeadBlockTable(const DeadBlock& aDb) {
+  if (gOptions->IsCumulativeMode() && aDb.AllocStackTrace()) {
+    AutoLockState lock;
+    if (DeadBlockTable::AddPtr p = gDeadBlockTable->lookupForAdd(aDb)) {
+      p->value() += 1;
+    } else {
+      MOZ_ALWAYS_TRUE(gDeadBlockTable->add(p, aDb, 1));
+    }
+  }
+}
+
+// Add a pointer to each live stack trace into the given StackTraceSet.  (A
+// stack trace is live if it's used by one of the live blocks.)
+static void GatherUsedStackTraces(StackTraceSet& aStackTraces) {
+  MOZ_ASSERT(gStateLock->IsLocked());
+  MOZ_ASSERT(Thread::Fetch()->InterceptsAreBlocked());
+
+  aStackTraces.clear();
+  MOZ_ALWAYS_TRUE(aStackTraces.reserve(512));
+
+  for (auto iter = gLiveBlockTable->iter(); !iter.done(); iter.next()) {
+    iter.get().AddStackTracesToTable(aStackTraces);
+  }
+
+  for (auto iter = gDeadBlockTable->iter(); !iter.done(); iter.next()) {
+    iter.get().key().AddStackTracesToTable(aStackTraces);
+  }
+}
+
+// Delete stack traces that we aren't using, and compact our hashtable.
+static void GCStackTraces() {
+  MOZ_ASSERT(gStateLock->IsLocked());
+  MOZ_ASSERT(Thread::Fetch()->InterceptsAreBlocked());
+
+  StackTraceSet usedStackTraces;
+  GatherUsedStackTraces(usedStackTraces);
+
+  // Delete all unused stack traces from gStackTraceTable.  The ModIterator
+  // destructor will automatically rehash and compact the table.
+  for (auto iter = gStackTraceTable->modIter(); !iter.done(); iter.next()) {
+    StackTrace* const& st = iter.get();
+    if (!usedStackTraces.has(st)) {
+      iter.remove();
+      InfallibleAllocPolicy::delete_(st);
+    }
+  }
+
+  // Schedule a GC when we have twice as many stack traces as we had right after
+  // this GC finished.
+  gGCStackTraceTableWhenSizeExceeds = 2 * gStackTraceTable->count();
+}
+
+//---------------------------------------------------------------------------
+// malloc/free callbacks
+//---------------------------------------------------------------------------
+
+static FastBernoulliTrial* gBernoulli;
+
+// In testing, a probability of 0.003 resulted in ~25% of heap blocks getting
+// a stack trace and ~80% of heap bytes getting a stack trace. (This is
+// possible because big heap blocks are more likely to get a stack trace.)
+//
+// We deliberately choose not to give the user control over this probability
+// (other than effectively setting it to 1 via --stacks=full) because it's
+// quite inscrutable and generally the user just wants "faster and imprecise"
+// or "slower and precise".
+//
+// The random number seeds are arbitrary and were obtained from random.org. If
+// you change them you'll need to change the tests as well, because their
+// expected output is based on the particular sequence of trial results that we
+// get with these seeds.
+static void ResetBernoulli() {
+  new (gBernoulli)
+      FastBernoulliTrial(0.003, 0x8e26eeee166bc8ca, 0x56820f304a9c9ae0);
+}
+
+static void AllocCallback(void* aPtr, size_t aReqSize, Thread* aT) {
+  if (!aPtr) {
+    return;
+  }
+
+  AutoLockState lock;
+  AutoBlockIntercepts block(aT);
+
+  size_t actualSize = gMallocTable.malloc_usable_size(aPtr);
+
+  // We may or may not record the allocation stack trace, depending on the
+  // options and the outcome of a Bernoulli trial.
+  bool getTrace = gOptions->DoFullStacks() || gBernoulli->trial(actualSize);
+  LiveBlock b(aPtr, aReqSize, getTrace ? StackTrace::Get(aT) : nullptr);
+  LiveBlockTable::AddPtr p = gLiveBlockTable->lookupForAdd(aPtr);
+  if (!p) {
+    // Most common case: there wasn't a record already.
+    MOZ_ALWAYS_TRUE(gLiveBlockTable->add(p, b));
+  } else {
+    // Edge-case: there was a record for the same address. We'll assume the
+    // allocator is not giving out a pointer to an existing allocation, so
+    // this means the previously recorded allocation was freed while we were
+    // blocking interceptions. This can happen while processing the data in
+    // e.g. AnalyzeImpl.
+    if (gOptions->IsCumulativeMode()) {
+      // Copy it out so it can be added to the dead block list later.
+      DeadBlock db(*p);
+      MaybeAddToDeadBlockTable(db);
+    }
+    gLiveBlockTable->remove(p);
+    MOZ_ALWAYS_TRUE(gLiveBlockTable->putNew(aPtr, b));
+  }
+}
+
+static void FreeCallback(void* aPtr, Thread* aT, DeadBlock* aDeadBlock) {
+  if (!aPtr) {
+    return;
+  }
+
+  AutoLockState lock;
+  AutoBlockIntercepts block(aT);
+
+  if (LiveBlockTable::Ptr lb = gLiveBlockTable->lookup(aPtr)) {
+    if (gOptions->IsCumulativeMode()) {
+      // Copy it out so it can be added to the dead block list later.
+      new (aDeadBlock) DeadBlock(*lb);
+    }
+    gLiveBlockTable->remove(lb);
+  } else {
+    // We have no record of the block. It must be a bogus pointer, or one that
+    // DMD wasn't able to see allocated. This should be extremely rare.
+  }
+
+  if (gStackTraceTable->count() > gGCStackTraceTableWhenSizeExceeds) {
+    GCStackTraces();
+  }
+}
+
+//---------------------------------------------------------------------------
+// malloc/free interception
+//---------------------------------------------------------------------------
+
+static bool Init(malloc_table_t* aMallocTable);
+
+}  // namespace dmd
+}  // namespace mozilla
+
+static void* replace_malloc(size_t aSize) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    // Intercepts are blocked, which means this must be a call to malloc
+    // triggered indirectly by DMD (e.g. via MozStackWalk).  Be infallible.
+    return InfallibleAllocPolicy::malloc_(aSize);
+  }
+
+  // This must be a call to malloc from outside DMD.  Intercept it.
+  void* ptr = gMallocTable.malloc(aSize);
+  AllocCallback(ptr, aSize, t);
+  return ptr;
+}
+
+static void* replace_calloc(size_t aCount, size_t aSize) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    return InfallibleAllocPolicy::calloc_(aCount, aSize);
+  }
+
+  // |aCount * aSize| could overflow, but if that happens then
+  // |gMallocTable.calloc()| will return nullptr and |AllocCallback()| will
+  // return immediately without using the overflowed value.
+  void* ptr = gMallocTable.calloc(aCount, aSize);
+  AllocCallback(ptr, aCount * aSize, t);
+  return ptr;
+}
+
+static void* replace_realloc(void* aOldPtr, size_t aSize) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    return InfallibleAllocPolicy::realloc_(aOldPtr, aSize);
+  }
+
+  // If |aOldPtr| is nullptr, the call is equivalent to |malloc(aSize)|.
+  if (!aOldPtr) {
+    return replace_malloc(aSize);
+  }
+
+  // Be very careful here!  Must remove the block from the table before doing
+  // the realloc to avoid races, just like in replace_free().
+  // Nb: This does an unnecessary hashtable remove+add if the block doesn't
+  // move, but doing better isn't worth the effort.
+  DeadBlock db;
+  FreeCallback(aOldPtr, t, &db);
+  void* ptr = gMallocTable.realloc(aOldPtr, aSize);
+  if (ptr) {
+    AllocCallback(ptr, aSize, t);
+    MaybeAddToDeadBlockTable(db);
+  } else {
+    // If realloc fails, we undo the prior operations by re-inserting the old
+    // pointer into the live block table. We don't have to do anything with the
+    // dead block list because the dead block hasn't yet been inserted. The
+    // block will end up looking like it was allocated for the first time here,
+    // which is untrue, and the slop bytes will be zero, which may be untrue.
+    // But this case is rare and doing better isn't worth the effort.
+    AllocCallback(aOldPtr, gMallocTable.malloc_usable_size(aOldPtr), t);
+  }
+  return ptr;
+}
+
+static void* replace_memalign(size_t aAlignment, size_t aSize) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    return InfallibleAllocPolicy::memalign_(aAlignment, aSize);
+  }
+
+  void* ptr = gMallocTable.memalign(aAlignment, aSize);
+  AllocCallback(ptr, aSize, t);
+  return ptr;
+}
+
+static void replace_free(void* aPtr) {
+  using namespace mozilla::dmd;
+
+  Thread* t = Thread::Fetch();
+  if (t->InterceptsAreBlocked()) {
+    return InfallibleAllocPolicy::free_(aPtr);
+  }
+
+  // Do the actual free after updating the table.  Otherwise, another thread
+  // could call malloc and get the freed block and update the table, and then
+  // our update here would remove the newly-malloc'd block.
+  DeadBlock db;
+  FreeCallback(aPtr, t, &db);
+  MaybeAddToDeadBlockTable(db);
+  gMallocTable.free(aPtr);
+}
+
+void replace_init(malloc_table_t* aMallocTable, ReplaceMallocBridge** aBridge) {
+  if (mozilla::dmd::Init(aMallocTable)) {
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#define MALLOC_DECL(name, ...) aMallocTable->name = replace_##name;
+#include "malloc_decls.h"
+    *aBridge = mozilla::dmd::gDMDBridge;
+  }
+}
+
+namespace mozilla {
+namespace dmd {
+
+//---------------------------------------------------------------------------
+// Options (Part 2)
+//---------------------------------------------------------------------------
+
+// Given an |aOptionName| like "foo", succeed if |aArg| has the form "foo=blah"
+// (where "blah" is non-empty) and return the pointer to "blah".  |aArg| can
+// have leading space chars (but not other whitespace).
+const char* Options::ValueIfMatch(const char* aArg, const char* aOptionName) {
+  MOZ_ASSERT(!isspace(*aArg));  // any leading whitespace should not remain
+  size_t optionLen = strlen(aOptionName);
+  if (strncmp(aArg, aOptionName, optionLen) == 0 && aArg[optionLen] == '=' &&
+      aArg[optionLen + 1]) {
+    return aArg + optionLen + 1;
+  }
+  return nullptr;
+}
+
+// Extracts a |long| value for an option from an argument.  It must be within
+// the range |aMin..aMax| (inclusive).
+bool Options::GetLong(const char* aArg, const char* aOptionName, long aMin,
+                      long aMax, long* aValue) {
+  if (const char* optionValue = ValueIfMatch(aArg, aOptionName)) {
+    char* endPtr;
+    *aValue = strtol(optionValue, &endPtr, /* base */ 10);
+    if (!*endPtr && aMin <= *aValue && *aValue <= aMax && *aValue != LONG_MIN &&
+        *aValue != LONG_MAX) {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Extracts a |bool| value for an option -- encoded as "yes" or "no" -- from an
+// argument.
+bool Options::GetBool(const char* aArg, const char* aOptionName, bool* aValue) {
+  if (const char* optionValue = ValueIfMatch(aArg, aOptionName)) {
+    if (strcmp(optionValue, "yes") == 0) {
+      *aValue = true;
+      return true;
+    }
+    if (strcmp(optionValue, "no") == 0) {
+      *aValue = false;
+      return true;
+    }
+  }
+  return false;
+}
+
+Options::Options(const char* aDMDEnvVar)
+    : mDMDEnvVar(aDMDEnvVar ? InfallibleAllocPolicy::strdup_(aDMDEnvVar)
+                            : nullptr),
+      mMode(Mode::DarkMatter),
+      mStacks(Stacks::Partial),
+      mShowDumpStats(false) {
+  char* e = mDMDEnvVar;
+  if (e && strcmp(e, "1") != 0) {
+    bool isEnd = false;
+    while (!isEnd) {
+      // Consume leading whitespace.
+      while (isspace(*e)) {
+        e++;
+      }
+
+      // Save the start of the arg.
+      const char* arg = e;
+
+      // Find the first char after the arg, and temporarily change it to '\0'
+      // to isolate the arg.
+      while (!isspace(*e) && *e != '\0') {
+        e++;
+      }
+      char replacedChar = *e;
+      isEnd = replacedChar == '\0';
+      *e = '\0';
+
+      // Handle arg
+      bool myBool;
+      if (strcmp(arg, "--mode=live") == 0) {
+        mMode = Mode::Live;
+      } else if (strcmp(arg, "--mode=dark-matter") == 0) {
+        mMode = Mode::DarkMatter;
+      } else if (strcmp(arg, "--mode=cumulative") == 0) {
+        mMode = Mode::Cumulative;
+      } else if (strcmp(arg, "--mode=scan") == 0) {
+        mMode = Mode::Scan;
+
+      } else if (strcmp(arg, "--stacks=full") == 0) {
+        mStacks = Stacks::Full;
+      } else if (strcmp(arg, "--stacks=partial") == 0) {
+        mStacks = Stacks::Partial;
+
+      } else if (GetBool(arg, "--show-dump-stats", &myBool)) {
+        mShowDumpStats = myBool;
+
+      } else if (strcmp(arg, "") == 0) {
+        // This can only happen if there is trailing whitespace.  Ignore.
+        MOZ_ASSERT(isEnd);
+
+      } else {
+        BadArg(arg);
+      }
+
+      // Undo the temporary isolation.
+      *e = replacedChar;
+    }
+  }
+
+  if (mMode == Mode::Scan) {
+    mStacks = Stacks::Full;
+  }
+}
+
+void Options::BadArg(const char* aArg) {
+  StatusMsg("\n");
+  StatusMsg("Bad entry in the $DMD environment variable: '%s'.\n", aArg);
+  StatusMsg("See the output of |mach help run| for the allowed options.\n");
+  exit(1);
+}
+
+const char* Options::ModeString() const {
+  switch (mMode) {
+    case Mode::Live:
+      return "live";
+    case Mode::DarkMatter:
+      return "dark-matter";
+    case Mode::Cumulative:
+      return "cumulative";
+    case Mode::Scan:
+      return "scan";
+    default:
+      MOZ_ASSERT(false);
+      return "(unknown DMD mode)";
+  }
+}
+
+//---------------------------------------------------------------------------
+// DMD start-up
+//---------------------------------------------------------------------------
+
+#ifndef XP_WIN
+static void prefork() {
+  if (gStateLock) {
+    gStateLock->Lock();
+  }
+}
+
+static void postfork() {
+  if (gStateLock) {
+    gStateLock->Unlock();
+  }
+}
+#endif
+
+// WARNING: this function runs *very* early -- before all static initializers
+// have run.  For this reason, non-scalar globals such as gStateLock and
+// gStackTraceTable are allocated dynamically (so we can guarantee their
+// construction in this function) rather than statically.
+static bool Init(malloc_table_t* aMallocTable) {
+  // DMD is controlled by the |DMD| environment variable.
+  const char* e = getenv("DMD");
+
+  if (!e) {
+    return false;
+  }
+  // Initialize the function table first, because StatusMsg uses
+  // InfallibleAllocPolicy::malloc_, which uses it.
+  gMallocTable = *aMallocTable;
+
+  StatusMsg("$DMD = '%s'\n", e);
+
+  gDMDBridge = InfallibleAllocPolicy::new_<DMDBridge>();
+
+#ifndef XP_WIN
+  // Avoid deadlocks when forking by acquiring our state lock prior to forking
+  // and releasing it after forking. See |LogAlloc|'s |replace_init| for
+  // in-depth details.
+  //
+  // Note: This must run after attempting an allocation so as to give the
+  // system malloc a chance to insert its own atfork handler.
+  pthread_atfork(prefork, postfork, postfork);
+#endif
+  // Parse $DMD env var.
+  gOptions = InfallibleAllocPolicy::new_<Options>(e);
+
+  gStateLock = InfallibleAllocPolicy::new_<Mutex>();
+
+  gBernoulli = (FastBernoulliTrial*)InfallibleAllocPolicy::malloc_(
+      sizeof(FastBernoulliTrial));
+  ResetBernoulli();
+
+  Thread::Init();
+
+  {
+    AutoLockState lock;
+
+    gStackTraceTable = InfallibleAllocPolicy::new_<StackTraceTable>(8192);
+    gLiveBlockTable = InfallibleAllocPolicy::new_<LiveBlockTable>(8192);
+
+    // Create this even if the mode isn't Cumulative (albeit with a small
+    // size), in case the mode is changed later on (as is done by SmokeDMD.cpp,
+    // for example).
+    size_t tableSize = gOptions->IsCumulativeMode() ? 8192 : 4;
+    gDeadBlockTable = InfallibleAllocPolicy::new_<DeadBlockTable>(tableSize);
+  }
+
+  return true;
+}
+
+//---------------------------------------------------------------------------
+// Block reporting and unreporting
+//---------------------------------------------------------------------------
+
+static void ReportHelper(const void* aPtr, bool aReportedOnAlloc) {
+  if (!gOptions->IsDarkMatterMode() || !aPtr) {
+    return;
+  }
+
+  Thread* t = Thread::Fetch();
+
+  AutoBlockIntercepts block(t);
+  AutoLockState lock;
+
+  if (LiveBlockTable::Ptr p = gLiveBlockTable->lookup(aPtr)) {
+    p->Report(t, aReportedOnAlloc);
+  } else {
+    // We have no record of the block. It must be a bogus pointer. This should
+    // be extremely rare because Report() is almost always called in
+    // conjunction with a malloc_size_of-style function. Print a message so
+    // that we get some feedback.
+    StatusMsg("Unknown pointer %p\n", aPtr);
+  }
+}
+
+void DMDFuncs::Report(const void* aPtr) {
+  ReportHelper(aPtr, /* onAlloc */ false);
+}
+
+void DMDFuncs::ReportOnAlloc(const void* aPtr) {
+  ReportHelper(aPtr, /* onAlloc */ true);
+}
+
+//---------------------------------------------------------------------------
+// DMD output
+//---------------------------------------------------------------------------
+
+// The version number of the output format. Increment this if you make
+// backwards-incompatible changes to the format. See DMD.h for the version
+// history.
+static const int kOutputVersionNumber = 5;
+
+// Note that, unlike most SizeOf* functions, this function does not take a
+// |mozilla::MallocSizeOf| argument.  That's because those arguments are
+// primarily to aid DMD track heap blocks... but DMD deliberately doesn't track
+// heap blocks it allocated for itself!
+//
+// SizeOfInternal should be called while you're holding the state lock and
+// while intercepts are blocked; SizeOf acquires the lock and blocks
+// intercepts.
+
+static void SizeOfInternal(Sizes* aSizes) {
+  MOZ_ASSERT(gStateLock->IsLocked());
+  MOZ_ASSERT(Thread::Fetch()->InterceptsAreBlocked());
+
+  aSizes->Clear();
+
+  StackTraceSet usedStackTraces;
+  GatherUsedStackTraces(usedStackTraces);
+
+  for (auto iter = gStackTraceTable->iter(); !iter.done(); iter.next()) {
+    StackTrace* const& st = iter.get();
+
+    if (usedStackTraces.has(st)) {
+      aSizes->mStackTracesUsed += MallocSizeOf(st);
+    } else {
+      aSizes->mStackTracesUnused += MallocSizeOf(st);
+    }
+  }
+
+  aSizes->mStackTraceTable =
+      gStackTraceTable->shallowSizeOfIncludingThis(MallocSizeOf);
+
+  aSizes->mLiveBlockTable =
+      gLiveBlockTable->shallowSizeOfIncludingThis(MallocSizeOf);
+
+  aSizes->mDeadBlockTable =
+      gDeadBlockTable->shallowSizeOfIncludingThis(MallocSizeOf);
+}
+
+void DMDFuncs::SizeOf(Sizes* aSizes) {
+  aSizes->Clear();
+
+  AutoBlockIntercepts block(Thread::Fetch());
+  AutoLockState lock;
+  SizeOfInternal(aSizes);
+}
+
+void DMDFuncs::ClearReports() {
+  if (!gOptions->IsDarkMatterMode()) {
+    return;
+  }
+
+  AutoLockState lock;
+
+  // Unreport all blocks that were marked reported by a memory reporter.  This
+  // excludes those that were reported on allocation, because they need to keep
+  // their reported marking.
+  for (auto iter = gLiveBlockTable->iter(); !iter.done(); iter.next()) {
+    iter.get().UnreportIfNotReportedOnAlloc();
+  }
+}
+
+class ToIdStringConverter final {
+ public:
+  ToIdStringConverter() : mIdMap(512), mNextId(0) {}
+
+  // Converts a pointer to a unique ID. Reuses the existing ID for the pointer
+  // if it's been seen before.
+  const char* ToIdString(const void* aPtr) {
+    uint32_t id;
+    PointerIdMap::AddPtr p = mIdMap.lookupForAdd(aPtr);
+    if (!p) {
+      id = mNextId++;
+      MOZ_ALWAYS_TRUE(mIdMap.add(p, aPtr, id));
+    } else {
+      id = p->value();
+    }
+    return Base32(id);
+  }
+
+  size_t sizeOfExcludingThis(mozilla::MallocSizeOf aMallocSizeOf) const {
+    return mIdMap.shallowSizeOfExcludingThis(aMallocSizeOf);
+  }
+
+ private:
+  // This function converts an integer to base-32. We use base-32 values for
+  // indexing into the traceTable and the frameTable, for the following reasons.
+  //
+  // - Base-32 gives more compact indices than base-16.
+  //
+  // - 32 is a power-of-two, which makes the necessary div/mod calculations
+  //   fast.
+  //
+  // - We can (and do) choose non-numeric digits for base-32. When
+  //   inspecting/debugging the JSON output, non-numeric indices are easier to
+  //   search for than numeric indices.
+  //
+  char* Base32(uint32_t aN) {
+    static const char digits[] = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdef";
+
+    char* b = mIdBuf + kIdBufLen - 1;
+    *b = '\0';
+    do {
+      b--;
+      if (b == mIdBuf) {
+        MOZ_CRASH("Base32 buffer too small");
+      }
+      *b = digits[aN % 32];
+      aN /= 32;
+    } while (aN);
+
+    return b;
+  }
+
+  PointerIdMap mIdMap;
+  uint32_t mNextId;
+
+  // |mIdBuf| must have space for at least eight chars, which is the space
+  // needed to hold 'Dffffff' (including the terminating null char), which is
+  // the base-32 representation of 0xffffffff.
+  static const size_t kIdBufLen = 16;
+  char mIdBuf[kIdBufLen];
+};
+
+// Helper class for converting a pointer value to a string.
+class ToStringConverter {
+ public:
+  const char* ToPtrString(const void* aPtr) {
+    snprintf(kPtrBuf, sizeof(kPtrBuf) - 1, "%" PRIxPTR, (uintptr_t)aPtr);
+    return kPtrBuf;
+  }
+
+ private:
+  char kPtrBuf[32];
+};
+
+static void WriteBlockContents(JSONWriter& aWriter, const LiveBlock& aBlock) {
+  size_t numWords = aBlock.ReqSize() / sizeof(uintptr_t*);
+  if (numWords == 0) {
+    return;
+  }
+
+  aWriter.StartArrayProperty("contents", aWriter.SingleLineStyle);
+  {
+    const uintptr_t** block = (const uintptr_t**)aBlock.Address();
+    ToStringConverter sc;
+    for (size_t i = 0; i < numWords; ++i) {
+      aWriter.StringElement(MakeStringSpan(sc.ToPtrString(block[i])));
+    }
+  }
+  aWriter.EndArray();
+}
+
+static void AnalyzeImpl(UniquePtr<JSONWriteFunc> aWriter) {
+  // Some blocks may have been allocated while creating |aWriter|. Those blocks
+  // will be freed at the end of this function when |write| is destroyed. The
+  // allocations will have occurred while intercepts were not blocked, so the
+  // frees better be as well, otherwise we'll get assertion failures.
+  // Therefore, this declaration must precede the AutoBlockIntercepts
+  // declaration, to ensure that |write| is destroyed *after* intercepts are
+  // unblocked.
+  JSONWriter writer(std::move(aWriter));
+
+  AutoBlockIntercepts block(Thread::Fetch());
+  AutoLockState lock;
+
+  // Allocate this on the heap instead of the stack because it's fairly large.
+  auto locService = InfallibleAllocPolicy::new_<CodeAddressService>();
+
+  StackTraceSet usedStackTraces(512);
+  PointerSet usedPcs(512);
+
+  size_t iscSize;
+
+  static int analysisCount = 1;
+  StatusMsg("Dump %d {\n", analysisCount++);
+
+  writer.Start();
+  {
+    writer.IntProperty("version", kOutputVersionNumber);
+
+    writer.StartObjectProperty("invocation");
+    {
+      const char* var = gOptions->DMDEnvVar();
+      if (var) {
+        writer.StringProperty("dmdEnvVar", MakeStringSpan(var));
+      } else {
+        writer.NullProperty("dmdEnvVar");
+      }
+
+      writer.StringProperty("mode", MakeStringSpan(gOptions->ModeString()));
+    }
+    writer.EndObject();
+
+    StatusMsg("  Constructing the heap block list...\n");
+
+    ToIdStringConverter isc;
+    ToStringConverter sc;
+
+    writer.StartArrayProperty("blockList");
+    {
+      // Lambda that writes out a live block.
+      auto writeLiveBlock = [&](const LiveBlock& aB, size_t aNum) {
+        aB.AddStackTracesToTable(usedStackTraces);
+
+        MOZ_ASSERT_IF(gOptions->IsScanMode(), aNum == 1);
+
+        writer.StartObjectElement(writer.SingleLineStyle);
+        {
+          if (gOptions->IsScanMode()) {
+            writer.StringProperty("addr",
+                                  MakeStringSpan(sc.ToPtrString(aB.Address())));
+            WriteBlockContents(writer, aB);
+          }
+          writer.IntProperty("req", aB.ReqSize());
+          if (aB.SlopSize() > 0) {
+            writer.IntProperty("slop", aB.SlopSize());
+          }
+
+          if (aB.AllocStackTrace()) {
+            writer.StringProperty(
+                "alloc", MakeStringSpan(isc.ToIdString(aB.AllocStackTrace())));
+          }
+
+          if (gOptions->IsDarkMatterMode() && aB.NumReports() > 0) {
+            writer.StartArrayProperty("reps");
+            {
+              if (aB.ReportStackTrace1()) {
+                writer.StringElement(
+                    MakeStringSpan(isc.ToIdString(aB.ReportStackTrace1())));
+              }
+              if (aB.ReportStackTrace2()) {
+                writer.StringElement(
+                    MakeStringSpan(isc.ToIdString(aB.ReportStackTrace2())));
+              }
+            }
+            writer.EndArray();
+          }
+
+          if (aNum > 1) {
+            writer.IntProperty("num", aNum);
+          }
+        }
+        writer.EndObject();
+      };
+
+      // Live blocks.
+      if (!gOptions->IsScanMode()) {
+        // At this point we typically have many LiveBlocks that differ only in
+        // their address. Aggregate them to reduce the size of the output file.
+        AggregatedLiveBlockTable agg(8192);
+        for (auto iter = gLiveBlockTable->iter(); !iter.done(); iter.next()) {
+          const LiveBlock& b = iter.get();
+          b.AddStackTracesToTable(usedStackTraces);
+
+          if (AggregatedLiveBlockTable::AddPtr p = agg.lookupForAdd(&b)) {
+            p->value() += 1;
+          } else {
+            MOZ_ALWAYS_TRUE(agg.add(p, &b, 1));
+          }
+        }
+
+        // Now iterate over the aggregated table.
+        for (auto iter = agg.iter(); !iter.done(); iter.next()) {
+          const LiveBlock& b = *iter.get().key();
+          size_t num = iter.get().value();
+          writeLiveBlock(b, num);
+        }
+
+      } else {
+        // In scan mode we cannot aggregate because we print each live block's
+        // address and contents.
+        for (auto iter = gLiveBlockTable->iter(); !iter.done(); iter.next()) {
+          const LiveBlock& b = iter.get();
+          b.AddStackTracesToTable(usedStackTraces);
+
+          writeLiveBlock(b, 1);
+        }
+      }
+
+      // Dead blocks.
+      for (auto iter = gDeadBlockTable->iter(); !iter.done(); iter.next()) {
+        const DeadBlock& b = iter.get().key();
+        b.AddStackTracesToTable(usedStackTraces);
+
+        size_t num = iter.get().value();
+        MOZ_ASSERT(num > 0);
+
+        writer.StartObjectElement(writer.SingleLineStyle);
+        {
+          writer.IntProperty("req", b.ReqSize());
+          if (b.SlopSize() > 0) {
+            writer.IntProperty("slop", b.SlopSize());
+          }
+          if (b.AllocStackTrace()) {
+            writer.StringProperty(
+                "alloc", MakeStringSpan(isc.ToIdString(b.AllocStackTrace())));
+          }
+
+          if (num > 1) {
+            writer.IntProperty("num", num);
+          }
+        }
+        writer.EndObject();
+      }
+    }
+    writer.EndArray();
+
+    StatusMsg("  Constructing the stack trace table...\n");
+
+    writer.StartObjectProperty("traceTable");
+    {
+      for (auto iter = usedStackTraces.iter(); !iter.done(); iter.next()) {
+        const StackTrace* const st = iter.get();
+        writer.StartArrayProperty(MakeStringSpan(isc.ToIdString(st)),
+                                  writer.SingleLineStyle);
+        {
+          for (uint32_t i = 0; i < st->Length(); i++) {
+            const void* pc = st->Pc(i);
+            writer.StringElement(MakeStringSpan(isc.ToIdString(pc)));
+            MOZ_ALWAYS_TRUE(usedPcs.put(pc));
+          }
+        }
+        writer.EndArray();
+      }
+    }
+    writer.EndObject();
+
+    StatusMsg("  Constructing the stack frame table...\n");
+
+    writer.StartObjectProperty("frameTable");
+    {
+      static const size_t locBufLen = 1024;
+      char locBuf[locBufLen];
+
+      for (auto iter = usedPcs.iter(); !iter.done(); iter.next()) {
+        const void* const pc = iter.get();
+
+        // Use 0 for the frame number. See the JSON format description comment
+        // in DMD.h to understand why.
+        locService->GetLocation(0, pc, locBuf, locBufLen);
+        writer.StringProperty(MakeStringSpan(isc.ToIdString(pc)),
+                              MakeStringSpan(locBuf));
+      }
+    }
+    writer.EndObject();
+
+    iscSize = isc.sizeOfExcludingThis(MallocSizeOf);
+  }
+  writer.End();
+
+  if (gOptions->ShowDumpStats()) {
+    Sizes sizes;
+    SizeOfInternal(&sizes);
+
+    static const size_t kBufLen = 64;
+    char buf1[kBufLen];
+    char buf2[kBufLen];
+    char buf3[kBufLen];
+
+    StatusMsg("  Execution measurements {\n");
+
+    StatusMsg("    Data structures that persist after Dump() ends {\n");
+
+    StatusMsg("      Used stack traces:    %10s bytes\n",
+              Show(sizes.mStackTracesUsed, buf1, kBufLen));
+
+    StatusMsg("      Unused stack traces:  %10s bytes\n",
+              Show(sizes.mStackTracesUnused, buf1, kBufLen));
+
+    StatusMsg("      Stack trace table:    %10s bytes (%s entries, %s used)\n",
+              Show(sizes.mStackTraceTable, buf1, kBufLen),
+              Show(gStackTraceTable->capacity(), buf2, kBufLen),
+              Show(gStackTraceTable->count(), buf3, kBufLen));
+
+    StatusMsg("      Live block table:     %10s bytes (%s entries, %s used)\n",
+              Show(sizes.mLiveBlockTable, buf1, kBufLen),
+              Show(gLiveBlockTable->capacity(), buf2, kBufLen),
+              Show(gLiveBlockTable->count(), buf3, kBufLen));
+
+    StatusMsg("      Dead block table:     %10s bytes (%s entries, %s used)\n",
+              Show(sizes.mDeadBlockTable, buf1, kBufLen),
+              Show(gDeadBlockTable->capacity(), buf2, kBufLen),
+              Show(gDeadBlockTable->count(), buf3, kBufLen));
+
+    StatusMsg("    }\n");
+    StatusMsg("    Data structures that are destroyed after Dump() ends {\n");
+
+    StatusMsg(
+        "      Location service:      %10s bytes\n",
+        Show(locService->SizeOfIncludingThis(MallocSizeOf), buf1, kBufLen));
+    StatusMsg("      Used stack traces set: %10s bytes\n",
+              Show(usedStackTraces.shallowSizeOfExcludingThis(MallocSizeOf),
+                   buf1, kBufLen));
+    StatusMsg(
+        "      Used PCs set:          %10s bytes\n",
+        Show(usedPcs.shallowSizeOfExcludingThis(MallocSizeOf), buf1, kBufLen));
+    StatusMsg("      Pointer ID map:        %10s bytes\n",
+              Show(iscSize, buf1, kBufLen));
+
+    StatusMsg("    }\n");
+    StatusMsg("    Counts {\n");
+
+    size_t hits = locService->NumCacheHits();
+    size_t misses = locService->NumCacheMisses();
+    size_t requests = hits + misses;
+    StatusMsg("      Location service:    %10s requests\n",
+              Show(requests, buf1, kBufLen));
+
+    size_t count = locService->CacheCount();
+    size_t capacity = locService->CacheCapacity();
+    StatusMsg(
+        "      Location service cache:  "
+        "%4.1f%% hit rate, %.1f%% occupancy at end\n",
+        Percent(hits, requests), Percent(count, capacity));
+
+    StatusMsg("    }\n");
+    StatusMsg("  }\n");
+  }
+
+  InfallibleAllocPolicy::delete_(locService);
+
+  StatusMsg("}\n");
+}
+
+void DMDFuncs::Analyze(UniquePtr<JSONWriteFunc> aWriter) {
+  AnalyzeImpl(std::move(aWriter));
+  ClearReports();
+}
+
+//---------------------------------------------------------------------------
+// Testing
+//---------------------------------------------------------------------------
+
+void DMDFuncs::ResetEverything(const char* aOptions) {
+  AutoLockState lock;
+
+  // Reset options.
+  InfallibleAllocPolicy::delete_(gOptions);
+  gOptions = InfallibleAllocPolicy::new_<Options>(aOptions);
+
+  // Clear all existing blocks.
+  gLiveBlockTable->clear();
+  gDeadBlockTable->clear();
+
+  // Reset gBernoulli to a deterministic state. (Its current state depends on
+  // all previous trials.)
+  ResetBernoulli();
+}
+
+}  // namespace dmd
+}  // namespace mozilla
diff --git a/memory/replace/dmd/DMD.h b/memory/replace/dmd/DMD.h
new file mode 100644
index 0000000000..c057047800
--- /dev/null
+++ b/memory/replace/dmd/DMD.h
@@ -0,0 +1,291 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef DMD_h___
+#define DMD_h___
+
+#include <stdarg.h>
+#include <string.h>
+
+#include <utility>
+
+#include "mozilla/DebugOnly.h"
+#include "mozilla/Types.h"
+#include "mozilla/UniquePtr.h"
+#include "replace_malloc_bridge.h"
+
+namespace mozilla {
+
+class JSONWriteFunc;
+
+namespace dmd {
+
+struct Sizes {
+  size_t mStackTracesUsed;
+  size_t mStackTracesUnused;
+  size_t mStackTraceTable;
+  size_t mLiveBlockTable;
+  size_t mDeadBlockTable;
+
+  Sizes() { Clear(); }
+  void Clear() { memset(this, 0, sizeof(Sizes)); }
+};
+
+// See further below for a description of each method. The DMDFuncs class
+// should contain a virtual method for each of them (except IsRunning,
+// which can be inferred from the DMDFuncs singleton existing).
+struct DMDFuncs {
+  virtual void Report(const void*);
+
+  virtual void ReportOnAlloc(const void*);
+
+  virtual void ClearReports();
+
+  virtual void Analyze(UniquePtr<JSONWriteFunc>);
+
+  virtual void SizeOf(Sizes*);
+
+  virtual void StatusMsg(const char*, va_list) MOZ_FORMAT_PRINTF(2, 0);
+
+  virtual void ResetEverything(const char*);
+
+#ifndef REPLACE_MALLOC_IMPL
+  // We deliberately don't use ReplaceMalloc::GetDMDFuncs here, because if we
+  // did, the following would happen.
+  // - The code footprint of each call to Get() larger as GetDMDFuncs ends
+  //   up inlined.
+  // - When no replace-malloc library is loaded, the number of instructions
+  //   executed is equivalent, but don't necessarily fit in the same cache
+  //   line.
+  // - When a non-DMD replace-malloc library is loaded, the overhead is
+  //   higher because there is first a check for the replace malloc bridge
+  //   and then for the DMDFuncs singleton.
+  // Initializing the DMDFuncs singleton on the first access makes the
+  // overhead even worse. Either Get() is inlined and massive, or it isn't
+  // and a simple value check becomes a function call.
+  static DMDFuncs* Get() { return sSingleton.Get(); }
+
+ private:
+  // Wrapper class keeping a pointer to the DMD functions. It is statically
+  // initialized because it needs to be set early enough.
+  // Debug builds also check that it's never accessed before the static
+  // initialization actually occured, which could be the case if some other
+  // static initializer ended up calling into DMD.
+  class Singleton {
+   public:
+    Singleton()
+        : mValue(ReplaceMalloc::GetDMDFuncs())
+#  ifdef DEBUG
+          ,
+          mInitialized(true)
+#  endif
+    {
+    }
+
+    DMDFuncs* Get() {
+      MOZ_ASSERT(mInitialized);
+      return mValue;
+    }
+
+   private:
+    DMDFuncs* mValue;
+#  ifdef DEBUG
+    bool mInitialized;
+#  endif
+  };
+
+  // This singleton pointer must be defined on the program side. In Gecko,
+  // this is done in xpcom/base/nsMemoryInfoDumper.cpp.
+  static /* DMDFuncs:: */ Singleton sSingleton;
+#endif
+};
+
+#ifndef REPLACE_MALLOC_IMPL
+// Mark a heap block as reported by a memory reporter.
+inline void Report(const void* aPtr) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->Report(aPtr);
+  }
+}
+
+// Mark a heap block as reported immediately on allocation.
+inline void ReportOnAlloc(const void* aPtr) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->ReportOnAlloc(aPtr);
+  }
+}
+
+// Clears existing reportedness data from any prior runs of the memory
+// reporters.  The following sequence should be used.
+// - ClearReports()
+// - run the memory reporters
+// - Analyze()
+// This sequence avoids spurious twice-reported warnings.
+inline void ClearReports() {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->ClearReports();
+  }
+}
+
+// Determines which heap blocks have been reported, and dumps JSON output
+// (via |aWriter|) describing the heap.
+//
+// The following sample output contains comments that explain the format and
+// design choices. The output files can be quite large, so a number of
+// decisions were made to minimize size, such as using short property names and
+// omitting properties whenever possible.
+//
+// {
+//   // The version number of the format, which will be incremented each time
+//   // backwards-incompatible changes are made. A mandatory integer.
+//   //
+//   // Version history:
+//   // - 1: Bug 1044709
+//   // - 2: Bug 1094552
+//   // - 3: Bug 1100851
+//   // - 4: Bug 1121830
+//   // - 5: Bug 1253512
+//   "version": 5,
+//
+//   // Information about how DMD was invoked. A mandatory object.
+//   "invocation": {
+//     // The contents of the $DMD environment variable. A string, or |null| if
+//     // $DMD is undefined.
+//     "dmdEnvVar": "--mode=dark-matter",
+//
+//     // The profiling mode. A mandatory string taking one of the following
+//     // values: "live", "dark-matter", "cumulative", "scan".
+//     "mode": "dark-matter",
+//   },
+//
+//   // Details of all analyzed heap blocks. A mandatory array.
+//   "blockList": [
+//     // An example of a heap block.
+//     {
+//       // Requested size, in bytes. This is a mandatory integer.
+//       "req": 3584,
+//
+//       // Requested slop size, in bytes. This is mandatory if it is non-zero,
+//       // but omitted otherwise.
+//       "slop": 512,
+//
+//       // The stack trace at which the block was allocated. An optional
+//       // string that indexes into the "traceTable" object. If omitted, no
+//       // allocation stack trace was recorded for the block.
+//       "alloc": "A",
+//
+//       // One or more stack traces at which this heap block was reported by a
+//       // memory reporter. An optional array that will only be present in
+//       // "dark-matter" mode. The elements are strings that index into
+//       // the "traceTable" object.
+//       "reps": ["B"]
+//
+//       // The number of heap blocks with exactly the above properties. This
+//       // is mandatory if it is greater than one, but omitted otherwise.
+//       // (Blocks with identical properties don't have to be aggregated via
+//       // this property, but it can greatly reduce output file size.)
+//       "num": 5,
+//
+//       // The address of the block. This is mandatory in "scan" mode, but
+//       // omitted otherwise.
+//       "addr": "4e4e4e4e",
+//
+//       // The contents of the block, read one word at a time. This is
+//       // mandatory in "scan" mode for blocks at least one word long, but
+//       // omitted otherwise.
+//       "contents": ["0", "6", "7f7f7f7f", "0"]
+//     }
+//   ],
+//
+//   // The stack traces referenced by elements of the "blockList" array. This
+//   // could be an array, but making it an object makes it easier to see
+//   // which stacks correspond to which references in the "blockList" array.
+//   "traceTable": {
+//     // Each property corresponds to a stack trace mentioned in the "blocks"
+//     // object. Each element is an index into the "frameTable" object.
+//     "A": ["D", "E"],
+//     "B": ["F", "G"]
+//   },
+//
+//   // The stack frames referenced by the "traceTable" object. The
+//   // descriptions can be quite long, so they are stored separately from the
+//   // "traceTable" object so that each one only has to be written once.
+//   // This could also be an array, but again, making it an object makes it
+//   // easier to see which frames correspond to which references in the
+//   // "traceTable" object.
+//   "frameTable": {
+//     // Each property key is a frame key mentioned in the "traceTable" object.
+//     // Each property value is a string containing a frame description. Each
+//     // frame description must be in a format recognized by `fix_stacks.py`,
+//     // which requires a frame number at the start. Because each stack frame
+//     // description in this table can be shared between multiple stack
+//     // traces, we use a dummy value of #00. The proper frame number can be
+//     // reconstructed later by scripts that output stack traces in a
+//     // conventional non-shared format.
+//     "D": "#00: foo (Foo.cpp:123)",
+//     "E": "#00: bar (Bar.cpp:234)",
+//     "F": "#00: baz (Baz.cpp:345)",
+//     "G": "#00: quux (Quux.cpp:456)"
+//   }
+// }
+//
+// Implementation note: normally, this function wouldn't be templated, but in
+// that case, the function is compiled, which makes the destructor for the
+// UniquePtr fire up, and that needs JSONWriteFunc to be fully defined. That,
+// in turn, requires to include JSONWriter.h, which includes
+// double-conversion.h, which ends up breaking various things built with
+// -Werror for various reasons.
+//
+template <typename JSONWriteFunc>
+inline void Analyze(UniquePtr<JSONWriteFunc> aWriteFunc) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->Analyze(std::move(aWriteFunc));
+  }
+}
+
+// Gets the size of various data structures.  Used to implement a memory
+// reporter for DMD.
+inline void SizeOf(Sizes* aSizes) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->SizeOf(aSizes);
+  }
+}
+
+// Prints a status message prefixed with "DMD[<pid>]". Use sparingly.
+MOZ_FORMAT_PRINTF(1, 2)
+inline void StatusMsg(const char* aFmt, ...) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    va_list ap;
+    va_start(ap, aFmt);
+    funcs->StatusMsg(aFmt, ap);
+    va_end(ap);
+  }
+}
+
+// Indicates whether or not DMD is running.
+inline bool IsRunning() { return !!DMDFuncs::Get(); }
+
+// Resets all DMD options and then sets new ones according to those specified
+// in |aOptions|. Also clears all recorded data about allocations. Only used
+// for testing purposes.
+inline void ResetEverything(const char* aOptions) {
+  DMDFuncs* funcs = DMDFuncs::Get();
+  if (funcs) {
+    funcs->ResetEverything(aOptions);
+  }
+}
+#endif
+
+}  // namespace dmd
+}  // namespace mozilla
+
+#endif /* DMD_h___ */
diff --git a/memory/replace/dmd/README b/memory/replace/dmd/README
new file mode 100644
index 0000000000..537893358a
--- /dev/null
+++ b/memory/replace/dmd/README
@@ -0,0 +1,3 @@
+This is DMD.  See
+https://firefox-source-docs.mozilla.org/performance/memory/dmd.html for
+details on how to use it.
diff --git a/memory/replace/dmd/block_analyzer.py b/memory/replace/dmd/block_analyzer.py
new file mode 100644
index 0000000000..eb26c9f5f5
--- /dev/null
+++ b/memory/replace/dmd/block_analyzer.py
@@ -0,0 +1,293 @@
+#!/usr/bin/env python3
+
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+# From a scan mode DMD log, extract some information about a
+# particular block, such as its allocation stack or which other blocks
+# contain pointers to it. This can be useful when investigating leaks
+# caused by unknown references to refcounted objects.
+
+import argparse
+import gzip
+import json
+import re
+import sys
+
+# The DMD output version this script handles.
+outputVersion = 5
+
+# If --ignore-alloc-fns is specified, stack frames containing functions that
+# match these strings will be removed from the *start* of stack traces. (Once
+# we hit a non-matching frame, any subsequent frames won't be removed even if
+# they do match.)
+allocatorFns = [
+    "malloc (",
+    "replace_malloc",
+    "replace_calloc",
+    "replace_realloc",
+    "replace_memalign",
+    "replace_posix_memalign",
+    "malloc_zone_malloc",
+    "moz_xmalloc",
+    "moz_xcalloc",
+    "moz_xrealloc",
+    "operator new(",
+    "operator new[](",
+    "g_malloc",
+    "g_slice_alloc",
+    "callocCanGC",
+    "reallocCanGC",
+    "vpx_malloc",
+    "vpx_calloc",
+    "vpx_realloc",
+    "vpx_memalign",
+    "js_malloc",
+    "js_calloc",
+    "js_realloc",
+    "pod_malloc",
+    "pod_calloc",
+    "pod_realloc",
+    "nsTArrayInfallibleAllocator::Malloc",
+    "Allocator<ReplaceMallocBase>::malloc(",
+    "mozilla::dmd::StackTrace::Get(",
+    "mozilla::dmd::AllocCallback(",
+    "mozilla::dom::DOMArena::Allocate(",
+    # This one necessary to fully filter some sequences of allocation functions
+    # that happen in practice. Note that ??? entries that follow non-allocation
+    # functions won't be stripped, as explained above.
+    "???",
+]
+
+####
+
+# Command line arguments
+
+
+def range_1_24(string):
+    value = int(string)
+    if value < 1 or value > 24:
+        msg = "{:s} is not in the range 1..24".format(string)
+        raise argparse.ArgumentTypeError(msg)
+    return value
+
+
+parser = argparse.ArgumentParser(
+    description="Analyze the heap graph to find out things about an object. \
+By default this prints out information about blocks that point to the given block."
+)
+
+parser.add_argument("dmd_log_file_name", help="clamped DMD log file name")
+
+parser.add_argument("block", help="address of the block of interest")
+
+parser.add_argument(
+    "--info",
+    dest="info",
+    action="store_true",
+    default=False,
+    help="Print out information about the block.",
+)
+
+parser.add_argument(
+    "-sfl",
+    "--max-stack-frame-length",
+    type=int,
+    default=300,
+    help="Maximum number of characters to print from each stack frame",
+)
+
+parser.add_argument(
+    "-a",
+    "--ignore-alloc-fns",
+    action="store_true",
+    help="ignore allocation functions at the start of traces",
+)
+
+parser.add_argument(
+    "-f",
+    "--max-frames",
+    type=range_1_24,
+    default=8,
+    help="maximum number of frames to consider in each trace",
+)
+
+parser.add_argument(
+    "-c",
+    "--chain-reports",
+    action="store_true",
+    help="if only one block is found to hold onto the object, report "
+    "the next one, too",
+)
+
+
+####
+
+
+class BlockData:
+    def __init__(self, json_block):
+        self.addr = json_block["addr"]
+
+        if "contents" in json_block:
+            contents = json_block["contents"]
+        else:
+            contents = []
+        self.contents = []
+        for c in contents:
+            self.contents.append(int(c, 16))
+
+        self.req_size = json_block["req"]
+
+        self.alloc_stack = json_block["alloc"]
+
+
+def print_trace_segment(args, stacks, block):
+    (traceTable, frameTable) = stacks
+
+    for l in traceTable[block.alloc_stack]:
+        # The 5: is to remove the bogus leading "#00: " from the stack frame.
+        print("  " + frameTable[l][5 : args.max_stack_frame_length])
+
+
+def show_referrers(args, blocks, stacks, block):
+    visited = set([])
+
+    anyFound = False
+
+    while True:
+        referrers = {}
+
+        for b, data in blocks.items():
+            which_edge = 0
+            for e in data.contents:
+                if e == block:
+                    # 8 is the number of bytes per word on a 64-bit system.
+                    # XXX This means that this output will be wrong for logs from 32-bit systems!
+                    referrers.setdefault(b, []).append(8 * which_edge)
+                    anyFound = True
+                which_edge += 1
+
+        for r in referrers:
+            sys.stdout.write(
+                "0x{} size = {} bytes".format(blocks[r].addr, blocks[r].req_size)
+            )
+            plural = "s" if len(referrers[r]) > 1 else ""
+            sys.stdout.write(
+                " at byte offset"
+                + plural
+                + " "
+                + (", ".join(str(x) for x in referrers[r]))
+            )
+            print()
+            print_trace_segment(args, stacks, blocks[r])
+            print()
+
+        if args.chain_reports:
+            if len(referrers) == 0:
+                sys.stdout.write("Found no more referrers.\n")
+                break
+            if len(referrers) > 1:
+                sys.stdout.write("Found too many referrers.\n")
+                break
+
+            sys.stdout.write("Chaining to next referrer.\n\n")
+            for r in referrers:
+                block = r
+            if block in visited:
+                sys.stdout.write("Found a loop.\n")
+                break
+            visited.add(block)
+        else:
+            break
+
+    if not anyFound:
+        print("No referrers found.")
+
+
+def show_block_info(args, blocks, stacks, block):
+    b = blocks[block]
+    sys.stdout.write("block: 0x{}\n".format(b.addr))
+    sys.stdout.write("requested size: {} bytes\n".format(b.req_size))
+    sys.stdout.write("\n")
+    sys.stdout.write("block contents: ")
+    for c in b.contents:
+        v = "0" if c == 0 else blocks[c].addr
+        sys.stdout.write("0x{} ".format(v))
+    sys.stdout.write("\n\n")
+    sys.stdout.write("allocation stack:\n")
+    print_trace_segment(args, stacks, b)
+    return
+
+
+def cleanupTraceTable(args, frameTable, traceTable):
+    # Remove allocation functions at the start of traces.
+    if args.ignore_alloc_fns:
+        # Build a regexp that matches every function in allocatorFns.
+        escapedAllocatorFns = map(re.escape, allocatorFns)
+        fn_re = re.compile("|".join(escapedAllocatorFns))
+
+        # Remove allocator fns from each stack trace.
+        for traceKey, frameKeys in traceTable.items():
+            numSkippedFrames = 0
+            for frameKey in frameKeys:
+                frameDesc = frameTable[frameKey]
+                if re.search(fn_re, frameDesc):
+                    numSkippedFrames += 1
+                else:
+                    break
+            if numSkippedFrames > 0:
+                traceTable[traceKey] = frameKeys[numSkippedFrames:]
+
+    # Trim the number of frames.
+    for traceKey, frameKeys in traceTable.items():
+        if len(frameKeys) > args.max_frames:
+            traceTable[traceKey] = frameKeys[: args.max_frames]
+
+
+def loadGraph(options):
+    # Handle gzipped input if necessary.
+    isZipped = options.dmd_log_file_name.endswith(".gz")
+    opener = gzip.open if isZipped else open
+
+    with opener(options.dmd_log_file_name, "rb") as f:
+        j = json.load(f)
+
+    if j["version"] != outputVersion:
+        raise Exception("'version' property isn't '{:d}'".format(outputVersion))
+
+    block_list = j["blockList"]
+    blocks = {}
+
+    for json_block in block_list:
+        blocks[int(json_block["addr"], 16)] = BlockData(json_block)
+
+    traceTable = j["traceTable"]
+    frameTable = j["frameTable"]
+
+    cleanupTraceTable(options, frameTable, traceTable)
+
+    return (blocks, (traceTable, frameTable))
+
+
+def analyzeLogs():
+    options = parser.parse_args()
+
+    (blocks, stacks) = loadGraph(options)
+
+    block = int(options.block, 16)
+
+    if block not in blocks:
+        print("Object " + options.block + " not found in traces.")
+        print("It could still be the target of some nodes.")
+        return
+
+    if options.info:
+        show_block_info(options, blocks, stacks, block)
+        return
+
+    show_referrers(options, blocks, stacks, block)
+
+
+if __name__ == "__main__":
+    analyzeLogs()
diff --git a/memory/replace/dmd/dmd.py b/memory/replace/dmd/dmd.py
new file mode 100755
index 0000000000..ae772d15ff
--- /dev/null
+++ b/memory/replace/dmd/dmd.py
@@ -0,0 +1,1027 @@
+#! /usr/bin/env python3
+#
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+"""This script analyzes a JSON file emitted by DMD."""
+
+import argparse
+import collections
+import gzip
+import io
+import json
+import os
+import platform
+import re
+import shutil
+import sys
+import tempfile
+from bisect import bisect_right
+from functools import cmp_to_key
+from typing import Callable
+
+# The DMD output version this script handles.
+outputVersion = 5
+
+# If --ignore-alloc-fns is specified, stack frames containing functions that
+# match these strings will be removed from the *start* of stack traces. (Once
+# we hit a non-matching frame, any subsequent frames won't be removed even if
+# they do match.)
+allocatorFns = [
+    # Matches malloc, replace_malloc, moz_xmalloc, vpx_malloc, js_malloc,
+    # pod_malloc, malloc_zone_*, g_malloc.
+    "malloc",
+    # Matches calloc, replace_calloc, moz_xcalloc, vpx_calloc, js_calloc,
+    # pod_calloc, malloc_zone_calloc, pod_callocCanGC.
+    "calloc",
+    # Matches realloc, replace_realloc, moz_xrealloc, vpx_realloc, js_realloc,
+    # pod_realloc, pod_reallocCanGC.
+    "realloc",
+    # Matches memalign, posix_memalign, replace_memalign, replace_posix_memalign,
+    # moz_xmemalign, vpx_memalign, malloc_zone_memalign.
+    "memalign",
+    "operator new(",
+    "operator new[](",
+    "g_slice_alloc",
+    # This one is necessary to fully filter some sequences of allocation
+    # functions that happen in practice. Note that ??? entries that follow
+    # non-allocation functions won't be stripped, as explained above.
+    "???",
+    # Match DMD internals.
+    "mozilla::dmd::AllocCallback",
+    "mozilla::dmd::StackTrace::Get",
+]
+
+
+def cmp(a, b):
+    return (a > b) - (a < b)
+
+
+class Record(object):
+    """A record is an aggregation of heap blocks that have identical stack
+    traces. It can also be used to represent the difference between two
+    records."""
+
+    def __init__(self):
+        self.numBlocks = 0
+        self.reqSize = 0
+        self.slopSize = 0
+        self.usableSize = 0
+        self.allocatedAtDesc = None
+        self.reportedAtDescs = []
+        self.usableSizes = collections.defaultdict(int)
+
+    def isZero(self, args):
+        return (
+            self.numBlocks == 0
+            and self.reqSize == 0
+            and self.slopSize == 0
+            and self.usableSize == 0
+            and len(self.usableSizes) == 0
+        )
+
+    def negate(self):
+        self.numBlocks = -self.numBlocks
+        self.reqSize = -self.reqSize
+        self.slopSize = -self.slopSize
+        self.usableSize = -self.usableSize
+
+        negatedUsableSizes = collections.defaultdict(int)
+        for usableSize, count in self.usableSizes.items():
+            negatedUsableSizes[-usableSize] = count
+        self.usableSizes = negatedUsableSizes
+
+    def subtract(self, r):
+        # We should only be calling this on records with matching stack traces.
+        # Check this.
+        assert self.allocatedAtDesc == r.allocatedAtDesc
+        assert self.reportedAtDescs == r.reportedAtDescs
+
+        self.numBlocks -= r.numBlocks
+        self.reqSize -= r.reqSize
+        self.slopSize -= r.slopSize
+        self.usableSize -= r.usableSize
+
+        usableSizes1 = self.usableSizes
+        usableSizes2 = r.usableSizes
+        usableSizes3 = collections.defaultdict(int)
+        for usableSize in usableSizes1:
+            counts1 = usableSizes1[usableSize]
+            if usableSize in usableSizes2:
+                counts2 = usableSizes2[usableSize]
+                del usableSizes2[usableSize]
+                counts3 = counts1 - counts2
+                if counts3 != 0:
+                    if counts3 < 0:
+                        usableSize = -usableSize
+                        counts3 = -counts3
+                    usableSizes3[usableSize] = counts3
+            else:
+                usableSizes3[usableSize] = counts1
+
+        for usableSize in usableSizes2:
+            usableSizes3[-usableSize] = usableSizes2[usableSize]
+
+        self.usableSizes = usableSizes3
+
+    @staticmethod
+    def cmpByUsableSize(r1, r2):
+        # Sort by usable size, then by req size.
+        return cmp(abs(r1.usableSize), abs(r2.usableSize)) or Record.cmpByReqSize(
+            r1, r2
+        )
+
+    @staticmethod
+    def cmpByReqSize(r1, r2):
+        # Sort by req size.
+        return cmp(abs(r1.reqSize), abs(r2.reqSize))
+
+    @staticmethod
+    def cmpBySlopSize(r1, r2):
+        # Sort by slop size.
+        return cmp(abs(r1.slopSize), abs(r2.slopSize))
+
+    @staticmethod
+    def cmpByNumBlocks(r1, r2):
+        # Sort by block counts, then by usable size.
+        return cmp(abs(r1.numBlocks), abs(r2.numBlocks)) or Record.cmpByUsableSize(
+            r1, r2
+        )
+
+
+sortByChoices = {
+    "usable": Record.cmpByUsableSize,  # the default
+    "req": Record.cmpByReqSize,
+    "slop": Record.cmpBySlopSize,
+    "num-blocks": Record.cmpByNumBlocks,
+}
+
+
+def parseCommandLine():
+    # 24 is the maximum number of frames that DMD will produce.
+    def range_1_24(string):
+        value = int(string)
+        if value < 1 or value > 24:
+            msg = "{:s} is not in the range 1..24".format(string)
+            raise argparse.ArgumentTypeError(msg)
+        return value
+
+    description = """
+Analyze heap data produced by DMD.
+If one file is specified, analyze it; if two files are specified, analyze the
+difference.
+Input files can be gzipped.
+Write to stdout unless -o/--output is specified.
+Stack traces are fixed to show function names, filenames and line numbers
+unless --no-fix-stacks is specified; stack fixing modifies the original file
+and may take some time. If specified, the BREAKPAD_SYMBOLS_PATH environment
+variable is used to find breakpad symbols for stack fixing.
+"""
+    p = argparse.ArgumentParser(description=description)
+
+    p.add_argument(
+        "-o",
+        "--output",
+        type=argparse.FileType("w"),
+        help="output file; stdout if unspecified",
+    )
+
+    p.add_argument(
+        "-f",
+        "--max-frames",
+        type=range_1_24,
+        default=8,
+        help="maximum number of frames to consider in each trace",
+    )
+
+    p.add_argument(
+        "-s",
+        "--sort-by",
+        choices=sortByChoices.keys(),
+        default="usable",
+        help="sort the records by a particular metric",
+    )
+
+    p.add_argument(
+        "-a",
+        "--ignore-alloc-fns",
+        action="store_true",
+        help="ignore allocation functions at the start of traces",
+    )
+
+    p.add_argument("--no-fix-stacks", action="store_true", help="do not fix stacks")
+
+    p.add_argument(
+        "--clamp-contents",
+        action="store_true",
+        help="for a scan mode log, clamp addresses to the start of live blocks, "
+        "or zero if not in one",
+    )
+
+    p.add_argument(
+        "--print-clamp-stats",
+        action="store_true",
+        help="print information about the results of pointer clamping; mostly "
+        "useful for debugging clamping",
+    )
+
+    p.add_argument(
+        "--filter-stacks-for-testing",
+        action="store_true",
+        help="filter stack traces; only useful for testing purposes",
+    )
+
+    p.add_argument(
+        "--filter",
+        default=[],
+        action="append",
+        help="Only print entries that have a stack that matches the filter. "
+        "A filter may be negated by prefixing it with `!`. "
+        "If multiple filters are specified, all of them must match.",
+    )
+
+    p.add_argument("input_file", help="a file produced by DMD")
+
+    p.add_argument(
+        "input_file2",
+        nargs="?",
+        help="a file produced by DMD; if present, it is diff'd with input_file",
+    )
+
+    return p.parse_args(sys.argv[1:])
+
+
+# Fix stacks if necessary: first write the output to a tempfile, then replace
+# the original file with it.
+def fixStackTraces(inputFilename, isZipped, opener):
+    # This append() call is needed to make the import statements work when this
+    # script is installed as a symlink.
+    sys.path.append(os.path.dirname(__file__))
+
+    bpsyms = os.environ.get("BREAKPAD_SYMBOLS_PATH", None)
+    sysname = platform.system()
+    if bpsyms and os.path.exists(bpsyms):
+        import fix_stacks as fixModule
+
+        def fix(line):
+            return fixModule.fixSymbols(line, jsonMode=True, breakpadSymsDir=bpsyms)
+
+    elif sysname in ("Linux", "Darwin", "Windows"):
+        import fix_stacks as fixModule
+
+        def fix(line):
+            return fixModule.fixSymbols(line, jsonMode=True)
+
+    else:
+        return
+
+    # Fix stacks, writing output to a temporary file, and then overwrite the
+    # original file.
+    tmpFile = tempfile.NamedTemporaryFile(delete=False)
+
+    # If the input is gzipped, then the output (written initially to |tmpFile|)
+    # should be gzipped as well.
+    #
+    # And we want to set its pre-gzipped filename to '' rather than the name of
+    # the temporary file, so that programs like the Unix 'file' utility don't
+    # say that it was called 'tmp6ozTxE' (or something like that) before it was
+    # zipped. So that explains the |filename=''| parameter.
+    #
+    # But setting the filename like that clobbers |tmpFile.name|, so we must
+    # get that now in order to move |tmpFile| at the end.
+    tmpFilename = tmpFile.name
+    if isZipped:
+        tmpFile = gzip.GzipFile(filename="", fileobj=tmpFile, mode="wb")
+
+    with opener(inputFilename, "rb") as inputFile:
+        for line in inputFile:
+            tmpFile.write(fix(line))
+
+    tmpFile.close()
+
+    shutil.move(tmpFilename, inputFilename)
+
+
+def getDigestFromFile(args, inputFile):
+    # Handle gzipped input if necessary.
+    isZipped = inputFile.endswith(".gz")
+    opener = gzip.open if isZipped else open
+
+    # Fix stack traces unless otherwise instructed.
+    if not args.no_fix_stacks:
+        fixStackTraces(inputFile, isZipped, opener)
+
+    if args.clamp_contents:
+        clampBlockList(args, inputFile, isZipped, opener)
+
+    with opener(inputFile, "rb") as f:
+        j = json.load(f)
+
+    if j["version"] != outputVersion:
+        raise Exception("'version' property isn't '{:d}'".format(outputVersion))
+
+    # Extract the main parts of the JSON object.
+    invocation = j["invocation"]
+    dmdEnvVar = invocation["dmdEnvVar"]
+    mode = invocation["mode"]
+    blockList = j["blockList"]
+    traceTable = j["traceTable"]
+    frameTable = j["frameTable"]
+
+    # Insert the necessary entries for unrecorded stack traces. Note that 'ut'
+    # and 'uf' will not overlap with any keys produced by DMD's
+    # ToIdStringConverter::Base32() function.
+    unrecordedTraceID = "ut"
+    unrecordedFrameID = "uf"
+    traceTable[unrecordedTraceID] = [unrecordedFrameID]
+    frameTable[
+        unrecordedFrameID
+    ] = "#00: (no stack trace recorded due to --stacks=partial)"
+
+    # For the purposes of this script, 'scan' behaves like 'live'.
+    if mode == "scan":
+        mode = "live"
+
+    if mode not in ["live", "dark-matter", "cumulative"]:
+        raise Exception("bad 'mode' property: '{:s}'".format(mode))
+
+    # Remove allocation functions at the start of traces.
+    if args.ignore_alloc_fns:
+        # Build a regexp that matches every function in allocatorFns.
+        escapedAllocatorFns = map(re.escape, allocatorFns)
+        fn_re = re.compile("|".join(escapedAllocatorFns))
+
+        # Remove allocator fns from each stack trace.
+        for traceKey, frameKeys in traceTable.items():
+            numSkippedFrames = 0
+            for frameKey in frameKeys:
+                frameDesc = frameTable[frameKey]
+                if re.search(fn_re, frameDesc):
+                    numSkippedFrames += 1
+                else:
+                    break
+            if numSkippedFrames > 0:
+                traceTable[traceKey] = frameKeys[numSkippedFrames:]
+
+    # Trim the number of frames.
+    for traceKey, frameKeys in traceTable.items():
+        if len(frameKeys) > args.max_frames:
+            del frameKeys[args.max_frames :]
+
+    def buildTraceDescription(traceTable, frameTable, traceKey):
+        frameKeys = traceTable[traceKey]
+        fmt = "    #{:02d}{:}"
+
+        if args.filter_stacks_for_testing:
+            # This option is used by `test_dmd.js`, which runs the code in
+            # `SmokeDMD.cpp`. When running that test, there is too much
+            # variation in the stack traces across different machines and
+            # platforms to do exact output matching. However, every stack trace
+            # should have at least three frames that contain `DMD` (in one of
+            # `DMD.cpp`, `SmokeDMD.cpp`, `SmokeDMD`, or `SmokeDMD.exe`). Some
+            # example frames from automation (where `..` indicates excised path
+            # segments):
+            #
+            # Linux debug, with stack fixing using breakpad syms:
+            # `#01: replace_realloc(void*, unsigned long) [../dmd/DMD.cpp:1110]`
+            #
+            # Linux opt, with native stack fixing:
+            # `#02: TestFull(char const*, int, char const*, int) (../dmd/test/SmokeDMD.cpp:165)`
+            #
+            # Mac opt, with native stack fixing:
+            # `#03: RunTests() (../build/tests/bin/SmokeDMD + 0x21f9)`
+            #
+            # Windows opt, with native stack fixing failing due to a missing PDB:
+            # `#04: ??? (..\\build\\tests\\bin\\SmokeDMD.exe + 0x1c58)`
+            #
+            # If we see three such frames, we replace the entire stack trace
+            # with a single, predictable frame. This imprecise matching will at
+            # least detect if stack fixing fails completely.
+            dmd_frame_matches = 0
+            for frameKey in frameKeys:
+                frameDesc = frameTable[frameKey]
+                if "DMD" in frameDesc:
+                    dmd_frame_matches += 1
+                    if dmd_frame_matches >= 3:
+                        return [fmt.format(1, ": ... DMD.cpp ...")]
+
+        # The frame number is always '#00' (see DMD.h for why), so we have to
+        # replace that with the correct frame number.
+        desc = []
+        for n, frameKey in enumerate(traceTable[traceKey], start=1):
+            desc.append(fmt.format(n, frameTable[frameKey][3:]))
+        return desc
+
+    # Aggregate blocks into records. All sufficiently similar blocks go into a
+    # single record.
+
+    if mode in ["live", "cumulative"]:
+        liveOrCumulativeRecords = collections.defaultdict(Record)
+    elif mode == "dark-matter":
+        unreportedRecords = collections.defaultdict(Record)
+        onceReportedRecords = collections.defaultdict(Record)
+        twiceReportedRecords = collections.defaultdict(Record)
+
+    heapUsableSize = 0
+    heapBlocks = 0
+
+    recordKeyPartCache = {}
+
+    for block in blockList:
+        # For each block we compute a |recordKey|, and all blocks with the same
+        # |recordKey| are aggregated into a single record. The |recordKey| is
+        # derived from the block's 'alloc' and 'reps' (if present) stack
+        # traces.
+        #
+        # We use frame descriptions (e.g. "#00: foo (X.cpp:99)") when comparing
+        # traces for equality. We can't use trace keys or frame keys because
+        # they're not comparable across different DMD runs (which is relevant
+        # when doing diffs).
+        #
+        # Using frame descriptions also fits in with the stack trimming done
+        # for --max-frames, which requires that stack traces with common
+        # beginnings but different endings to be considered equivalent. E.g. if
+        # we have distinct traces T1:[A:D1,B:D2,C:D3] and T2:[X:D1,Y:D2,Z:D4]
+        # and we trim the final frame of each they should be considered
+        # equivalent because the untrimmed frame descriptions (D1 and D2)
+        # match.
+        #
+        # Having said all that, during a single invocation of dmd.py on a
+        # single DMD file, for a single frameKey value the record key will
+        # always be the same, and we might encounter it 1000s of times. So we
+        # cache prior results for speed.
+        def makeRecordKeyPart(traceKey):
+            if traceKey in recordKeyPartCache:
+                return recordKeyPartCache[traceKey]
+
+            recordKeyPart = str(
+                list(map(lambda frameKey: frameTable[frameKey], traceTable[traceKey]))
+            )
+            recordKeyPartCache[traceKey] = recordKeyPart
+            return recordKeyPart
+
+        allocatedAtTraceKey = block.get("alloc", unrecordedTraceID)
+        if mode in ["live", "cumulative"]:
+            recordKey = makeRecordKeyPart(allocatedAtTraceKey)
+            records = liveOrCumulativeRecords
+        elif mode == "dark-matter":
+            recordKey = makeRecordKeyPart(allocatedAtTraceKey)
+            if "reps" in block:
+                reportedAtTraceKeys = block["reps"]
+                for reportedAtTraceKey in reportedAtTraceKeys:
+                    recordKey += makeRecordKeyPart(reportedAtTraceKey)
+                if len(reportedAtTraceKeys) == 1:
+                    records = onceReportedRecords
+                else:
+                    records = twiceReportedRecords
+            else:
+                records = unreportedRecords
+
+        record = records[recordKey]
+
+        if "req" not in block:
+            raise Exception("'req' property missing in block'")
+
+        reqSize = block["req"]
+        slopSize = block.get("slop", 0)
+
+        if "num" in block:
+            num = block["num"]
+        else:
+            num = 1
+
+        usableSize = reqSize + slopSize
+        heapUsableSize += num * usableSize
+        heapBlocks += num
+
+        record.numBlocks += num
+        record.reqSize += num * reqSize
+        record.slopSize += num * slopSize
+        record.usableSize += num * usableSize
+        if record.allocatedAtDesc is None:
+            record.allocatedAtDesc = buildTraceDescription(
+                traceTable, frameTable, allocatedAtTraceKey
+            )
+
+        if mode in ["live", "cumulative"]:
+            pass
+        elif mode == "dark-matter":
+            if "reps" in block and record.reportedAtDescs == []:
+
+                def f(k):
+                    return buildTraceDescription(traceTable, frameTable, k)
+
+                record.reportedAtDescs = list(map(f, reportedAtTraceKeys))
+        record.usableSizes[usableSize] += num
+
+    # All the processed data for a single DMD file is called a "digest".
+    digest = {}
+    digest["dmdEnvVar"] = dmdEnvVar
+    digest["mode"] = mode
+    digest["heapUsableSize"] = heapUsableSize
+    digest["heapBlocks"] = heapBlocks
+    if mode in ["live", "cumulative"]:
+        digest["liveOrCumulativeRecords"] = liveOrCumulativeRecords
+    elif mode == "dark-matter":
+        digest["unreportedRecords"] = unreportedRecords
+        digest["onceReportedRecords"] = onceReportedRecords
+        digest["twiceReportedRecords"] = twiceReportedRecords
+    return digest
+
+
+def diffRecords(args, records1, records2):
+    records3 = {}
+
+    # Process records1.
+    for k in records1:
+        r1 = records1[k]
+        if k in records2:
+            # This record is present in both records1 and records2.
+            r2 = records2[k]
+            del records2[k]
+            r2.subtract(r1)
+            if not r2.isZero(args):
+                records3[k] = r2
+        else:
+            # This record is present only in records1.
+            r1.negate()
+            records3[k] = r1
+
+    for k in records2:
+        # This record is present only in records2.
+        records3[k] = records2[k]
+
+    return records3
+
+
+def diffDigests(args, d1, d2):
+    if d1["mode"] != d2["mode"]:
+        raise Exception("the input files have different 'mode' properties")
+
+    d3 = {}
+    d3["dmdEnvVar"] = (d1["dmdEnvVar"], d2["dmdEnvVar"])
+    d3["mode"] = d1["mode"]
+    d3["heapUsableSize"] = d2["heapUsableSize"] - d1["heapUsableSize"]
+    d3["heapBlocks"] = d2["heapBlocks"] - d1["heapBlocks"]
+    if d1["mode"] in ["live", "cumulative"]:
+        d3["liveOrCumulativeRecords"] = diffRecords(
+            args, d1["liveOrCumulativeRecords"], d2["liveOrCumulativeRecords"]
+        )
+    elif d1["mode"] == "dark-matter":
+        d3["unreportedRecords"] = diffRecords(
+            args, d1["unreportedRecords"], d2["unreportedRecords"]
+        )
+        d3["onceReportedRecords"] = diffRecords(
+            args, d1["onceReportedRecords"], d2["onceReportedRecords"]
+        )
+        d3["twiceReportedRecords"] = diffRecords(
+            args, d1["twiceReportedRecords"], d2["twiceReportedRecords"]
+        )
+    return d3
+
+
+def printDigest(args, digest):
+    dmdEnvVar = digest["dmdEnvVar"]
+    mode = digest["mode"]
+    heapUsableSize = digest["heapUsableSize"]
+    heapBlocks = digest["heapBlocks"]
+    if mode in ["live", "cumulative"]:
+        liveOrCumulativeRecords = digest["liveOrCumulativeRecords"]
+    elif mode == "dark-matter":
+        unreportedRecords = digest["unreportedRecords"]
+        onceReportedRecords = digest["onceReportedRecords"]
+        twiceReportedRecords = digest["twiceReportedRecords"]
+
+    separator = "#" + "-" * 65 + "\n"
+
+    def number(n):
+        """Format a number with comma as a separator."""
+        return "{:,d}".format(n)
+
+    def perc(m, n):
+        return 0 if n == 0 else (100 * m / n)
+
+    def plural(n):
+        return "" if n == 1 else "s"
+
+    # Prints to stdout, or to file if -o/--output was specified.
+    def out(*arguments, **kwargs):
+        print(*arguments, file=args.output, **kwargs)
+
+    def printStack(traceDesc):
+        for frameDesc in traceDesc:
+            out(frameDesc)
+
+    def printRecords(recordKind, records, heapUsableSize):
+        RecordKind = recordKind.capitalize()
+        out(separator)
+        numRecords = len(records)
+        cmpRecords = sortByChoices[args.sort_by]
+        sortedRecords = sorted(
+            records.values(), key=cmp_to_key(cmpRecords), reverse=True
+        )
+        kindBlocks = 0
+        kindUsableSize = 0
+        maxRecord = 1000
+
+        def is_match(rec: Record, key: str):
+            return any(key in desc for desc in rec.allocatedAtDesc)
+
+        for arg in args.filter:
+            key: str
+            cond: Callable[[Record], bool]
+            if arg.startswith("\\"):
+                # just in case you really need to start a filter with '!' (or '\')
+                key = arg[1:]
+                cond = is_match
+            elif arg.startswith("!"):
+                key = arg[1:]
+
+                def cond(rec, key):
+                    return not is_match(rec, key)  # noqa: E731
+
+            else:
+                key = arg
+                cond = is_match
+            sortedRecords = [rec for rec in sortedRecords if cond(rec, key)]
+
+        # First iteration: get totals, etc.
+        for record in sortedRecords:
+            kindBlocks += record.numBlocks
+            kindUsableSize += record.usableSize
+
+        # Second iteration: print.
+        if numRecords == 0:
+            out("# no {:} heap blocks\n".format(recordKind))
+
+        kindCumulativeUsableSize = 0
+        for i, record in enumerate(sortedRecords, start=1):
+            # Stop printing at the |maxRecord|th record.
+            if i == maxRecord:
+                out(
+                    "# {:}: stopping after {:,d} heap block records\n".format(
+                        RecordKind, i
+                    )
+                )
+                break
+
+            kindCumulativeUsableSize += record.usableSize
+
+            out(RecordKind + " {")
+            out(
+                "  {:} block{:} in heap block record {:,d} of {:,d}".format(
+                    number(record.numBlocks), plural(record.numBlocks), i, numRecords
+                )
+            )
+            out(
+                "  {:} bytes ({:} requested / {:} slop)".format(
+                    number(record.usableSize),
+                    number(record.reqSize),
+                    number(record.slopSize),
+                )
+            )
+
+            usableSizes = sorted(
+                record.usableSizes.items(), key=lambda x: abs(x[0]), reverse=True
+            )
+            hasSingleBlock = len(usableSizes) == 1 and usableSizes[0][1] == 1
+
+            if not hasSingleBlock:
+                out("  Individual block sizes: ", end="")
+                if len(usableSizes) == 0:
+                    out("(no change)", end="")
+                else:
+                    isFirst = True
+                    for usableSize, count in usableSizes:
+                        if not isFirst:
+                            out("; ", end="")
+                        out("{:}".format(number(usableSize)), end="")
+                        if count > 1:
+                            out(" x {:,d}".format(count), end="")
+                        isFirst = False
+                out()
+
+            out(
+                "  {:4.2f}% of the heap ({:4.2f}% cumulative)".format(
+                    perc(record.usableSize, heapUsableSize),
+                    perc(kindCumulativeUsableSize, heapUsableSize),
+                )
+            )
+            if mode in ["live", "cumulative"]:
+                pass
+            elif mode == "dark-matter":
+                out(
+                    "  {:4.2f}% of {:} ({:4.2f}% cumulative)".format(
+                        perc(record.usableSize, kindUsableSize),
+                        recordKind,
+                        perc(kindCumulativeUsableSize, kindUsableSize),
+                    )
+                )
+            out("  Allocated at {")
+            printStack(record.allocatedAtDesc)
+            out("  }")
+            if mode in ["live", "cumulative"]:
+                pass
+            elif mode == "dark-matter":
+                for n, reportedAtDesc in enumerate(record.reportedAtDescs):
+                    again = "again " if n > 0 else ""
+                    out("  Reported {:}at {{".format(again))
+                    printStack(reportedAtDesc)
+                    out("  }")
+            out("}\n")
+
+        return (kindUsableSize, kindBlocks)
+
+    def printInvocation(n, dmdEnvVar, mode):
+        out("Invocation{:} {{".format(n))
+        if dmdEnvVar is None:
+            out("  $DMD is undefined")
+        else:
+            out("  $DMD = '" + dmdEnvVar + "'")
+        out("  Mode = '" + mode + "'")
+        out("}\n")
+
+    # Print command line. Strip dirs so the output is deterministic, which is
+    # needed for testing.
+    out(separator, end="")
+    out("# " + " ".join(map(os.path.basename, sys.argv)) + "\n")
+
+    # Print invocation(s).
+    if type(dmdEnvVar) is not tuple:
+        printInvocation("", dmdEnvVar, mode)
+    else:
+        printInvocation(" 1", dmdEnvVar[0], mode)
+        printInvocation(" 2", dmdEnvVar[1], mode)
+
+    # Print records.
+    if mode in ["live", "cumulative"]:
+        liveOrCumulativeUsableSize, liveOrCumulativeBlocks = printRecords(
+            mode, liveOrCumulativeRecords, heapUsableSize
+        )
+    elif mode == "dark-matter":
+        twiceReportedUsableSize, twiceReportedBlocks = printRecords(
+            "twice-reported", twiceReportedRecords, heapUsableSize
+        )
+
+        unreportedUsableSize, unreportedBlocks = printRecords(
+            "unreported", unreportedRecords, heapUsableSize
+        )
+
+        onceReportedUsableSize, onceReportedBlocks = printRecords(
+            "once-reported", onceReportedRecords, heapUsableSize
+        )
+
+    # Print summary.
+    out(separator)
+    out("Summary {")
+    if mode in ["live", "cumulative"]:
+        out(
+            "  Total: {:} bytes in {:} blocks".format(
+                number(liveOrCumulativeUsableSize), number(liveOrCumulativeBlocks)
+            )
+        )
+    elif mode == "dark-matter":
+        fmt = "  {:15} {:>12} bytes ({:6.2f}%) in {:>7} blocks ({:6.2f}%)"
+        out(fmt.format("Total:", number(heapUsableSize), 100, number(heapBlocks), 100))
+        out(
+            fmt.format(
+                "Unreported:",
+                number(unreportedUsableSize),
+                perc(unreportedUsableSize, heapUsableSize),
+                number(unreportedBlocks),
+                perc(unreportedBlocks, heapBlocks),
+            )
+        )
+        out(
+            fmt.format(
+                "Once-reported:",
+                number(onceReportedUsableSize),
+                perc(onceReportedUsableSize, heapUsableSize),
+                number(onceReportedBlocks),
+                perc(onceReportedBlocks, heapBlocks),
+            )
+        )
+        out(
+            fmt.format(
+                "Twice-reported:",
+                number(twiceReportedUsableSize),
+                perc(twiceReportedUsableSize, heapUsableSize),
+                number(twiceReportedBlocks),
+                perc(twiceReportedBlocks, heapBlocks),
+            )
+        )
+    out("}\n")
+
+
+#############################
+# Pretty printer for DMD JSON
+#############################
+
+
+def prettyPrintDmdJson(out, j):
+    out.write("{\n")
+
+    out.write(' "version": {0},\n'.format(j["version"]))
+    out.write(' "invocation": ')
+    json.dump(j["invocation"], out, sort_keys=True)
+    out.write(",\n")
+
+    out.write(' "blockList": [')
+    first = True
+    for b in j["blockList"]:
+        out.write("" if first else ",")
+        out.write("\n  ")
+        json.dump(b, out, sort_keys=True)
+        first = False
+    out.write("\n ],\n")
+
+    out.write(' "traceTable": {')
+    first = True
+    for k, l in j["traceTable"].items():
+        out.write("" if first else ",")
+        out.write('\n  "{0}": {1}'.format(k, json.dumps(l)))
+        first = False
+    out.write("\n },\n")
+
+    out.write(' "frameTable": {')
+    first = True
+    for k, v in j["frameTable"].items():
+        out.write("" if first else ",")
+        out.write('\n  "{0}": {1}'.format(k, json.dumps(v)))
+        first = False
+    out.write("\n }\n")
+
+    out.write("}\n")
+
+
+##################################################################
+# Code for clamping addresses using conservative pointer analysis.
+##################################################################
+
+# Start is the address of the first byte of the block, while end is
+# the address of the first byte after the final byte in the block.
+class AddrRange:
+    def __init__(self, block, length):
+        self.block = block
+        self.start = int(block, 16)
+        self.length = length
+        self.end = self.start + self.length
+
+        assert self.start > 0
+        assert length >= 0
+
+
+class ClampStats:
+    def __init__(self):
+        # Number of pointers already pointing to the start of a block.
+        self.startBlockPtr = 0
+
+        # Number of pointers pointing to the middle of a block. These
+        # are clamped to the start of the block they point into.
+        self.midBlockPtr = 0
+
+        # Number of null pointers.
+        self.nullPtr = 0
+
+        # Number of non-null pointers that didn't point into the middle
+        # of any blocks. These are clamped to null.
+        self.nonNullNonBlockPtr = 0
+
+    def clampedBlockAddr(self, sameAddress):
+        if sameAddress:
+            self.startBlockPtr += 1
+        else:
+            self.midBlockPtr += 1
+
+    def nullAddr(self):
+        self.nullPtr += 1
+
+    def clampedNonBlockAddr(self):
+        self.nonNullNonBlockPtr += 1
+
+    def log(self):
+        sys.stderr.write("Results:\n")
+        sys.stderr.write(
+            "  Number of pointers already pointing to start of blocks: "
+            + str(self.startBlockPtr)
+            + "\n"
+        )
+        sys.stderr.write(
+            "  Number of pointers clamped to start of blocks: "
+            + str(self.midBlockPtr)
+            + "\n"
+        )
+        sys.stderr.write(
+            "  Number of non-null pointers not pointing into blocks "
+            "clamped to null: " + str(self.nonNullNonBlockPtr) + "\n"
+        )
+        sys.stderr.write("  Number of null pointers: " + str(self.nullPtr) + "\n")
+
+
+# Search the block ranges array for a block that address points into.
+# The search is carried out in an array of starting addresses for each blocks
+# because it is faster.
+def clampAddress(blockRanges, blockStarts, clampStats, address):
+    i = bisect_right(blockStarts, address)
+
+    # Any addresses completely out of the range should have been eliminated already.
+    assert i > 0
+    r = blockRanges[i - 1]
+    assert r.start <= address
+
+    if address >= r.end:
+        assert address < blockRanges[i].start
+        clampStats.clampedNonBlockAddr()
+        return "0"
+
+    clampStats.clampedBlockAddr(r.start == address)
+    return r.block
+
+
+def clampBlockList(args, inputFileName, isZipped, opener):
+    # XXX This isn't very efficient because we end up reading and writing
+    # the file multiple times.
+    with opener(inputFileName, "rb") as f:
+        j = json.load(f)
+
+    if j["version"] != outputVersion:
+        raise Exception("'version' property isn't '{:d}'".format(outputVersion))
+
+    # Check that the invocation is reasonable for contents clamping.
+    invocation = j["invocation"]
+    if invocation["mode"] != "scan":
+        raise Exception("Log was taken in mode " + invocation["mode"] + " not scan")
+
+    sys.stderr.write("Creating block range list.\n")
+    blockList = j["blockList"]
+    blockRanges = []
+    for block in blockList:
+        blockRanges.append(AddrRange(block["addr"], block["req"]))
+    blockRanges.sort(key=lambda r: r.start)
+
+    # Make sure there are no overlapping blocks.
+    prevRange = blockRanges[0]
+    for currRange in blockRanges[1:]:
+        assert prevRange.end <= currRange.start
+        prevRange = currRange
+
+    sys.stderr.write("Clamping block contents.\n")
+    clampStats = ClampStats()
+    firstAddr = blockRanges[0].start
+    lastAddr = blockRanges[-1].end
+
+    blockStarts = []
+    for r in blockRanges:
+        blockStarts.append(r.start)
+
+    for block in blockList:
+        # Small blocks don't have any contents.
+        if "contents" not in block:
+            continue
+
+        cont = block["contents"]
+        for i in range(len(cont)):
+            address = int(cont[i], 16)
+
+            if address == 0:
+                clampStats.nullAddr()
+                continue
+
+            # If the address is before the first block or after the last
+            # block then it can't be within a block.
+            if address < firstAddr or address >= lastAddr:
+                clampStats.clampedNonBlockAddr()
+                cont[i] = "0"
+                continue
+
+            cont[i] = clampAddress(blockRanges, blockStarts, clampStats, address)
+
+        # Remove any trailing nulls.
+        while len(cont) and cont[-1] == "0":
+            cont.pop()
+
+    if args.print_clamp_stats:
+        clampStats.log()
+
+    sys.stderr.write("Saving file.\n")
+    tmpFile = tempfile.NamedTemporaryFile(delete=False)
+    tmpFilename = tmpFile.name
+    if isZipped:
+        tmpFile = gzip.GzipFile(filename="", fileobj=tmpFile, mode="wb")
+    prettyPrintDmdJson(io.TextIOWrapper(tmpFile, encoding="utf-8"), j)
+    tmpFile.close()
+    shutil.move(tmpFilename, inputFileName)
+
+
+def main():
+    args = parseCommandLine()
+    digest = getDigestFromFile(args, args.input_file)
+    if args.input_file2:
+        digest2 = getDigestFromFile(args, args.input_file2)
+        digest = diffDigests(args, digest, digest2)
+    printDigest(args, digest)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/memory/replace/dmd/moz.build b/memory/replace/dmd/moz.build
new file mode 100644
index 0000000000..6f3121df48
--- /dev/null
+++ b/memory/replace/dmd/moz.build
@@ -0,0 +1,37 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+EXPORTS += [
+    "DMD.h",
+]
+
+UNIFIED_SOURCES += [
+    "DMD.cpp",
+]
+
+if not CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+    UNIFIED_SOURCES += [
+        "/mfbt/HashFunctions.cpp",
+        "/mfbt/JSONWriter.cpp",
+        "/mfbt/Poison.cpp",
+        "/mozglue/misc/StackWalk.cpp",
+    ]
+    if CONFIG["OS_ARCH"] == "WINNT":
+        OS_LIBS += [
+            "dbghelp",
+        ]
+
+ReplaceMalloc("dmd")
+
+DEFINES["MOZ_NO_MOZALLOC"] = True
+DEFINES["IMPL_MFBT"] = True
+
+if CONFIG["MOZ_OPTIMIZE"]:
+    DEFINES["MOZ_OPTIMIZE"] = True
+
+DisableStlWrapping()
+
+TEST_DIRS += ["test"]
diff --git a/memory/replace/dmd/test/SmokeDMD.cpp b/memory/replace/dmd/test/SmokeDMD.cpp
new file mode 100644
index 0000000000..c72e92a543
--- /dev/null
+++ b/memory/replace/dmd/test/SmokeDMD.cpp
@@ -0,0 +1,378 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this file,
+ * You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// This program is used by the DMD xpcshell test. It is run under DMD and
+// produces some output. The xpcshell test then post-processes and checks this
+// output.
+//
+// Note that this file does not have "Test" or "test" in its name, because that
+// will cause the build system to not record breakpad symbols for it, which
+// will stop the post-processing (which includes stack fixing) from working
+// correctly.
+
+// This is required on some systems such as Fedora to allow
+// building with -O0 together with --warnings-as-errors due to
+// a check in /usr/include/features.h
+#undef _FORTIFY_SOURCE
+
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/JSONWriter.h"
+#include "mozilla/Sprintf.h"
+#include "mozilla/UniquePtr.h"
+#include "DMD.h"
+
+using mozilla::MakeUnique;
+using namespace mozilla::dmd;
+
+DMDFuncs::Singleton DMDFuncs::sSingleton;
+
+class FpWriteFunc final : public mozilla::JSONWriteFunc {
+ public:
+  explicit FpWriteFunc(const char* aFilename) {
+    mFp = fopen(aFilename, "w");
+    if (!mFp) {
+      fprintf(stderr, "SmokeDMD: can't create %s file: %s\n", aFilename,
+              strerror(errno));
+      exit(1);
+    }
+  }
+
+  ~FpWriteFunc() { fclose(mFp); }
+
+  void Write(const mozilla::Span<const char>& aStr) final {
+    for (const char c : aStr) {
+      fputc(c, mFp);
+    }
+  }
+
+ private:
+  FILE* mFp;
+};
+
+// This stops otherwise-unused variables from being optimized away.
+static void UseItOrLoseIt(void* aPtr, int aSeven) {
+  char buf[64];
+  int n = SprintfLiteral(buf, "%p\n", aPtr);
+  if (n == 20 + aSeven) {
+    fprintf(stderr, "well, that is surprising");
+  }
+}
+
+// This function checks that heap blocks that have the same stack trace but
+// different (or no) reporters get aggregated separately.
+void Foo(int aSeven) {
+  char* a[6];
+  for (int i = 0; i < aSeven - 1; i++) {
+    a[i] = (char*)malloc(128 - 16 * i);
+    UseItOrLoseIt(a[i], aSeven);
+  }
+
+  // Oddly, some versions of clang will cause identical stack traces to be
+  // generated for adjacent calls to Report(), which breaks the test. Inserting
+  // the UseItOrLoseIt() calls in between is enough to prevent this.
+
+  Report(a[2]);  // reported
+
+  UseItOrLoseIt(a[2], aSeven);
+
+  for (int i = 0; i < aSeven - 5; i++) {
+    Report(a[i]);  // reported
+    UseItOrLoseIt(a[i], aSeven);
+  }
+
+  UseItOrLoseIt(a[2], aSeven);
+
+  Report(a[3]);  // reported
+
+  // a[4], a[5] unreported
+}
+
+void TestEmpty(const char* aTestName, const char* aMode) {
+  char filename[128];
+  SprintfLiteral(filename, "complete-%s-%s.json", aTestName, aMode);
+  auto f = MakeUnique<FpWriteFunc>(filename);
+
+  char options[128];
+  SprintfLiteral(options, "--mode=%s --stacks=full", aMode);
+  ResetEverything(options);
+
+  // Zero for everything.
+  Analyze(std::move(f));
+}
+
+void TestFull(const char* aTestName, int aNum, const char* aMode, int aSeven) {
+  char filename[128];
+  SprintfLiteral(filename, "complete-%s%d-%s.json", aTestName, aNum, aMode);
+  auto f = MakeUnique<FpWriteFunc>(filename);
+
+  // The --show-dump-stats=yes is there just to give that option some basic
+  // testing, e.g. ensure it doesn't crash. It's hard to test much beyond that.
+  char options[128];
+  SprintfLiteral(options, "--mode=%s --stacks=full --show-dump-stats=yes",
+                 aMode);
+  ResetEverything(options);
+
+  // Analyze 1: 1 freed, 9 out of 10 unreported.
+  // Analyze 2: still present and unreported.
+  int i;
+  char* a = nullptr;
+  for (i = 0; i < aSeven + 3; i++) {
+    a = (char*)malloc(100);
+    UseItOrLoseIt(a, aSeven);
+  }
+  free(a);
+
+  // A no-op.
+  free(nullptr);
+
+  // Note: 16 bytes is the smallest requested size that gives consistent
+  // behaviour across all platforms with jemalloc.
+  // Analyze 1: reported.
+  // Analyze 2: thrice-reported.
+  char* a2 = (char*)malloc(16);
+  Report(a2);
+
+  // Analyze 1: reported.
+  // Analyze 2: reportedness carries over, due to ReportOnAlloc.
+  char* b = (char*)malloc(10);
+  ReportOnAlloc(b);
+
+  // ReportOnAlloc, then freed.
+  // Analyze 1: freed, irrelevant.
+  // Analyze 2: freed, irrelevant.
+  char* b2 = (char*)malloc(16);
+  ReportOnAlloc(b2);
+  free(b2);
+
+  // Analyze 1: reported 4 times.
+  // Analyze 2: freed, irrelevant.
+  char* c = (char*)calloc(10, 3);
+  Report(c);
+  for (int i = 0; i < aSeven - 4; i++) {
+    Report(c);
+  }
+
+  // Analyze 1: ignored.
+  // Analyze 2: irrelevant.
+  Report((void*)(intptr_t)i);
+
+  // jemalloc rounds this up to 8192.
+  // Analyze 1: reported.
+  // Analyze 2: freed.
+  char* e = (char*)malloc(4096);
+  e = (char*)realloc(e, 7169);
+  Report(e);
+
+  // First realloc is like malloc;  second realloc is shrinking.
+  // Analyze 1: reported.
+  // Analyze 2: re-reported.
+  char* e2 = (char*)realloc(nullptr, 1024);
+  e2 = (char*)realloc(e2, 512);
+  Report(e2);
+
+  // First realloc is like malloc;  second realloc creates a min-sized block.
+  // XXX: on Windows, second realloc frees the block.
+  // Analyze 1: reported.
+  // Analyze 2: freed, irrelevant.
+  char* e3 = (char*)realloc(nullptr, 1023);
+  // e3 = (char*) realloc(e3, 0);
+  MOZ_ASSERT(e3);
+  Report(e3);
+
+  // Analyze 1: freed, irrelevant.
+  // Analyze 2: freed, irrelevant.
+  char* f1 = (char*)malloc(64);
+  UseItOrLoseIt(f1, aSeven);
+  free(f1);
+
+  // Analyze 1: ignored.
+  // Analyze 2: irrelevant.
+  Report((void*)(intptr_t)0x0);
+
+  // Analyze 1: mixture of reported and unreported.
+  // Analyze 2: all unreported.
+  Foo(aSeven);
+
+  // Analyze 1: twice-reported.
+  // Analyze 2: twice-reported.
+  char* g1 = (char*)malloc(77);
+  ReportOnAlloc(g1);
+  ReportOnAlloc(g1);
+
+  // Analyze 1: mixture of reported and unreported.
+  // Analyze 2: all unreported.
+  // Nb: this Foo() call is deliberately not adjacent to the previous one. See
+  // the comment about adjacent calls in Foo() for more details.
+  Foo(aSeven);
+
+  // Analyze 1: twice-reported.
+  // Analyze 2: once-reported.
+  char* g2 = (char*)malloc(78);
+  Report(g2);
+  ReportOnAlloc(g2);
+
+  // Analyze 1: twice-reported.
+  // Analyze 2: once-reported.
+  char* g3 = (char*)malloc(79);
+  ReportOnAlloc(g3);
+  Report(g3);
+
+  // All the odd-ball ones.
+  // Analyze 1: all unreported.
+  // Analyze 2: all freed, irrelevant.
+  // XXX: no memalign on Mac
+  // void* w = memalign(64, 65);           // rounds up to 128
+  // UseItOrLoseIt(w, aSeven);
+
+  // XXX: posix_memalign doesn't work on B2G
+  // void* x;
+  // posix_memalign(&y, 128, 129);         // rounds up to 256
+  // UseItOrLoseIt(x, aSeven);
+
+  // XXX: valloc doesn't work on Windows.
+  // void* y = valloc(1);                  // rounds up to 4096
+  // UseItOrLoseIt(y, aSeven);
+
+  // XXX: C11 only
+  // void* z = aligned_alloc(64, 256);
+  // UseItOrLoseIt(z, aSeven);
+
+  if (aNum == 1) {
+    // Analyze 1.
+    Analyze(std::move(f));
+  }
+
+  ClearReports();
+
+  //---------
+
+  Report(a2);
+  Report(a2);
+  free(c);
+  free(e);
+  Report(e2);
+  free(e3);
+  // free(w);
+  // free(x);
+  // free(y);
+  // free(z);
+
+  // Do some allocations that will only show up in cumulative mode.
+  for (int i = 0; i < 100; i++) {
+    void* v = malloc(128);
+    UseItOrLoseIt(v, aSeven);
+    free(v);
+  }
+
+  if (aNum == 2) {
+    // Analyze 2.
+    Analyze(std::move(f));
+  }
+}
+
+void TestPartial(const char* aTestName, const char* aMode, int aSeven) {
+  char filename[128];
+  SprintfLiteral(filename, "complete-%s-%s.json", aTestName, aMode);
+  auto f = MakeUnique<FpWriteFunc>(filename);
+
+  char options[128];
+  SprintfLiteral(options, "--mode=%s", aMode);
+  ResetEverything(options);
+
+  int kTenThousand = aSeven + 9993;
+  char* s;
+
+  // The output of this function is deterministic but it relies on the
+  // probability and seeds given to the FastBernoulliTrial instance in
+  // ResetBernoulli(). If they change, the output will change too.
+
+  // Expected fraction with stacks: (1 - (1 - 0.003) ** 16) = 0.0469.
+  // So we expect about 0.0469 * 10000 == 469.
+  // We actually get 511.
+  for (int i = 0; i < kTenThousand; i++) {
+    s = (char*)malloc(16);
+    UseItOrLoseIt(s, aSeven);
+  }
+
+  // Expected fraction with stacks: (1 - (1 - 0.003) ** 128) = 0.3193.
+  // So we expect about 0.3193 * 10000 == 3193.
+  // We actually get 3136.
+  for (int i = 0; i < kTenThousand; i++) {
+    s = (char*)malloc(128);
+    UseItOrLoseIt(s, aSeven);
+  }
+
+  // Expected fraction with stacks: (1 - (1 - 0.003) ** 1024) = 0.9539.
+  // So we expect about 0.9539 * 10000 == 9539.
+  // We actually get 9531.
+  for (int i = 0; i < kTenThousand; i++) {
+    s = (char*)malloc(1024);
+    UseItOrLoseIt(s, aSeven);
+  }
+
+  Analyze(std::move(f));
+}
+
+void TestScan(int aSeven) {
+  auto f = MakeUnique<FpWriteFunc>("basic-scan.json");
+
+  ResetEverything("--mode=scan");
+
+  uintptr_t* p = (uintptr_t*)malloc(6 * sizeof(uintptr_t));
+  UseItOrLoseIt(p, aSeven);
+
+  // Hard-coded values checked by scan-test.py
+  p[0] = 0x123;                         // outside a block, small value
+  p[1] = 0x0;                           // null
+  p[2] = (uintptr_t)((uint8_t*)p - 1);  // pointer outside a block, but nearby
+  p[3] = (uintptr_t)p;                  // pointer to start of a block
+  p[4] = (uintptr_t)((uint8_t*)p + 1);  // pointer into a block
+  p[5] = 0x0;                           // trailing null
+
+  Analyze(std::move(f));
+}
+
+void RunTests() {
+  // This test relies on the compiler not doing various optimizations, such as
+  // eliding unused malloc() calls or unrolling loops with fixed iteration
+  // counts. So we compile it with -O0 (or equivalent), which probably prevents
+  // that. We also use the following variable for various loop iteration
+  // counts, just in case compilers might unroll very small loops even with
+  // -O0.
+  int seven = 7;
+
+  // Make sure that DMD is actually running; it is initialized on the first
+  // allocation.
+  int* x = (int*)malloc(100);
+  UseItOrLoseIt(x, seven);
+  MOZ_RELEASE_ASSERT(IsRunning());
+
+  // Please keep this in sync with run_test in test_dmd.js.
+
+  TestEmpty("empty", "live");
+  TestEmpty("empty", "dark-matter");
+  TestEmpty("empty", "cumulative");
+
+  TestFull("full", 1, "live", seven);
+  TestFull("full", 1, "dark-matter", seven);
+
+  TestFull("full", 2, "dark-matter", seven);
+  TestFull("full", 2, "cumulative", seven);
+
+  TestPartial("partial", "live", seven);
+
+  TestScan(seven);
+}
+
+int main() {
+  RunTests();
+
+  return 0;
+}
diff --git a/memory/replace/dmd/test/basic-scan-32-expected.txt b/memory/replace/dmd/test/basic-scan-32-expected.txt
new file mode 100644
index 0000000000..9f6f4db325
--- /dev/null
+++ b/memory/replace/dmd/test/basic-scan-32-expected.txt
@@ -0,0 +1,25 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o basic-scan-32-actual.txt --clamp-contents basic-scan.json
+
+Invocation {
+  $DMD = '--mode=scan'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  1 block in heap block record 1 of 1
+  32 bytes (24 requested / 8 slop)
+  100.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 32 bytes in 1 blocks
+}
+
diff --git a/memory/replace/dmd/test/basic-scan-64-expected.txt b/memory/replace/dmd/test/basic-scan-64-expected.txt
new file mode 100644
index 0000000000..59effc07b7
--- /dev/null
+++ b/memory/replace/dmd/test/basic-scan-64-expected.txt
@@ -0,0 +1,25 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o basic-scan-64-actual.txt --clamp-contents basic-scan.json
+
+Invocation {
+  $DMD = '--mode=scan'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  1 block in heap block record 1 of 1
+  48 bytes (48 requested / 0 slop)
+  100.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 48 bytes in 1 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-empty-cumulative-expected.txt b/memory/replace/dmd/test/complete-empty-cumulative-expected.txt
new file mode 100644
index 0000000000..2486015d0b
--- /dev/null
+++ b/memory/replace/dmd/test/complete-empty-cumulative-expected.txt
@@ -0,0 +1,18 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-empty-cumulative-actual.txt complete-empty-cumulative.json
+
+Invocation {
+  $DMD = '--mode=cumulative --stacks=full'
+  Mode = 'cumulative'
+}
+
+#-----------------------------------------------------------------
+
+# no cumulative heap blocks
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 0 bytes in 0 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-empty-dark-matter-expected.txt b/memory/replace/dmd/test/complete-empty-dark-matter-expected.txt
new file mode 100644
index 0000000000..0020cddde3
--- /dev/null
+++ b/memory/replace/dmd/test/complete-empty-dark-matter-expected.txt
@@ -0,0 +1,29 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-empty-dark-matter-actual.txt complete-empty-dark-matter.json
+
+Invocation {
+  $DMD = '--mode=dark-matter --stacks=full'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+# no twice-reported heap blocks
+
+#-----------------------------------------------------------------
+
+# no unreported heap blocks
+
+#-----------------------------------------------------------------
+
+# no once-reported heap blocks
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:                     0 bytes (100.00%) in       0 blocks (100.00%)
+  Unreported:                0 bytes (  0.00%) in       0 blocks (  0.00%)
+  Once-reported:             0 bytes (  0.00%) in       0 blocks (  0.00%)
+  Twice-reported:            0 bytes (  0.00%) in       0 blocks (  0.00%)
+}
+
diff --git a/memory/replace/dmd/test/complete-empty-live-expected.txt b/memory/replace/dmd/test/complete-empty-live-expected.txt
new file mode 100644
index 0000000000..d0d1721965
--- /dev/null
+++ b/memory/replace/dmd/test/complete-empty-live-expected.txt
@@ -0,0 +1,18 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-empty-live-actual.txt complete-empty-live.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+# no live heap blocks
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 0 bytes in 0 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-full1-dark-matter-expected.txt b/memory/replace/dmd/test/complete-full1-dark-matter-expected.txt
new file mode 100644
index 0000000000..2c7d6b6343
--- /dev/null
+++ b/memory/replace/dmd/test/complete-full1-dark-matter-expected.txt
@@ -0,0 +1,265 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-full1-dark-matter-actual.txt complete-full1-dark-matter.json
+
+Invocation {
+  $DMD = '--mode=dark-matter --stacks=full --show-dump-stats=yes'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+Twice-reported {
+  1 block in heap block record 1 of 4
+  80 bytes (79 requested / 1 slop)
+  0.66% of the heap (0.66% cumulative)
+  29.41% of twice-reported (29.41% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Twice-reported {
+  1 block in heap block record 2 of 4
+  80 bytes (78 requested / 2 slop)
+  0.66% of the heap (1.32% cumulative)
+  29.41% of twice-reported (58.82% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Twice-reported {
+  1 block in heap block record 3 of 4
+  80 bytes (77 requested / 3 slop)
+  0.66% of the heap (1.98% cumulative)
+  29.41% of twice-reported (88.24% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Twice-reported {
+  1 block in heap block record 4 of 4
+  32 bytes (30 requested / 2 slop)
+  0.26% of the heap (2.25% cumulative)
+  11.76% of twice-reported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Unreported {
+  9 blocks in heap block record 1 of 3
+  1,008 bytes (900 requested / 108 slop)
+  Individual block sizes: 112 x 9
+  8.33% of the heap (8.33% cumulative)
+  81.82% of unreported (81.82% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Unreported {
+  2 blocks in heap block record 2 of 3
+  112 bytes (112 requested / 0 slop)
+  Individual block sizes: 64; 48
+  0.93% of the heap (9.26% cumulative)
+  9.09% of unreported (90.91% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Unreported {
+  2 blocks in heap block record 3 of 3
+  112 bytes (112 requested / 0 slop)
+  Individual block sizes: 64; 48
+  0.93% of the heap (10.19% cumulative)
+  9.09% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Once-reported {
+  1 block in heap block record 1 of 11
+  8,192 bytes (7,169 requested / 1,023 slop)
+  67.72% of the heap (67.72% cumulative)
+  77.34% of once-reported (77.34% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 2 of 11
+  1,024 bytes (1,023 requested / 1 slop)
+  8.47% of the heap (76.19% cumulative)
+  9.67% of once-reported (87.01% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 3 of 11
+  512 bytes (512 requested / 0 slop)
+  4.23% of the heap (80.42% cumulative)
+  4.83% of once-reported (91.84% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  2 blocks in heap block record 4 of 11
+  240 bytes (240 requested / 0 slop)
+  Individual block sizes: 128; 112
+  1.98% of the heap (82.41% cumulative)
+  2.27% of once-reported (94.11% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  2 blocks in heap block record 5 of 11
+  240 bytes (240 requested / 0 slop)
+  Individual block sizes: 128; 112
+  1.98% of the heap (84.39% cumulative)
+  2.27% of once-reported (96.37% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 6 of 11
+  96 bytes (96 requested / 0 slop)
+  0.79% of the heap (85.19% cumulative)
+  0.91% of once-reported (97.28% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 7 of 11
+  96 bytes (96 requested / 0 slop)
+  0.79% of the heap (85.98% cumulative)
+  0.91% of once-reported (98.19% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 8 of 11
+  80 bytes (80 requested / 0 slop)
+  0.66% of the heap (86.64% cumulative)
+  0.76% of once-reported (98.94% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 9 of 11
+  80 bytes (80 requested / 0 slop)
+  0.66% of the heap (87.30% cumulative)
+  0.76% of once-reported (99.70% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 10 of 11
+  16 bytes (16 requested / 0 slop)
+  0.13% of the heap (87.43% cumulative)
+  0.15% of once-reported (99.85% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 11 of 11
+  16 bytes (10 requested / 6 slop)
+  0.13% of the heap (87.57% cumulative)
+  0.15% of once-reported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:                12,096 bytes (100.00%) in      30 blocks (100.00%)
+  Unreported:            1,232 bytes ( 10.19%) in      13 blocks ( 43.33%)
+  Once-reported:        10,592 bytes ( 87.57%) in      13 blocks ( 43.33%)
+  Twice-reported:          272 bytes (  2.25%) in       4 blocks ( 13.33%)
+}
+
diff --git a/memory/replace/dmd/test/complete-full1-live-expected.txt b/memory/replace/dmd/test/complete-full1-live-expected.txt
new file mode 100644
index 0000000000..eaa1883e1f
--- /dev/null
+++ b/memory/replace/dmd/test/complete-full1-live-expected.txt
@@ -0,0 +1,127 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-full1-live-actual.txt complete-full1-live.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full --show-dump-stats=yes'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  1 block in heap block record 1 of 12
+  8,192 bytes (7,169 requested / 1,023 slop)
+  67.72% of the heap (67.72% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 2 of 12
+  1,024 bytes (1,023 requested / 1 slop)
+  8.47% of the heap (76.19% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  9 blocks in heap block record 3 of 12
+  1,008 bytes (900 requested / 108 slop)
+  Individual block sizes: 112 x 9
+  8.33% of the heap (84.52% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  6 blocks in heap block record 4 of 12
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  4.37% of the heap (88.89% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  6 blocks in heap block record 5 of 12
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  4.37% of the heap (93.25% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 6 of 12
+  512 bytes (512 requested / 0 slop)
+  4.23% of the heap (97.49% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 7 of 12
+  80 bytes (79 requested / 1 slop)
+  0.66% of the heap (98.15% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 8 of 12
+  80 bytes (78 requested / 2 slop)
+  0.66% of the heap (98.81% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 9 of 12
+  80 bytes (77 requested / 3 slop)
+  0.66% of the heap (99.47% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 10 of 12
+  32 bytes (30 requested / 2 slop)
+  0.26% of the heap (99.74% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 11 of 12
+  16 bytes (16 requested / 0 slop)
+  0.13% of the heap (99.87% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  1 block in heap block record 12 of 12
+  16 bytes (10 requested / 6 slop)
+  0.13% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 12,096 bytes in 30 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-full2-cumulative-expected.txt b/memory/replace/dmd/test/complete-full2-cumulative-expected.txt
new file mode 100644
index 0000000000..5a225b9b8e
--- /dev/null
+++ b/memory/replace/dmd/test/complete-full2-cumulative-expected.txt
@@ -0,0 +1,173 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-full2-cumulative-actual.txt complete-full2-cumulative.json
+
+Invocation {
+  $DMD = '--mode=cumulative --stacks=full --show-dump-stats=yes'
+  Mode = 'cumulative'
+}
+
+#-----------------------------------------------------------------
+
+Cumulative {
+  100 blocks in heap block record 1 of 17
+  12,800 bytes (12,800 requested / 0 slop)
+  Individual block sizes: 128 x 100
+  42.37% of the heap (42.37% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 2 of 17
+  8,192 bytes (7,169 requested / 1,023 slop)
+  27.12% of the heap (69.49% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 3 of 17
+  4,096 bytes (4,096 requested / 0 slop)
+  13.56% of the heap (83.05% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  10 blocks in heap block record 4 of 17
+  1,120 bytes (1,000 requested / 120 slop)
+  Individual block sizes: 112 x 10
+  3.71% of the heap (86.76% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 5 of 17
+  1,024 bytes (1,024 requested / 0 slop)
+  3.39% of the heap (90.15% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 6 of 17
+  1,024 bytes (1,023 requested / 1 slop)
+  3.39% of the heap (93.54% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  6 blocks in heap block record 7 of 17
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  1.75% of the heap (95.29% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  6 blocks in heap block record 8 of 17
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  1.75% of the heap (97.03% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 9 of 17
+  512 bytes (512 requested / 0 slop)
+  1.69% of the heap (98.73% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 10 of 17
+  80 bytes (79 requested / 1 slop)
+  0.26% of the heap (98.99% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 11 of 17
+  80 bytes (78 requested / 2 slop)
+  0.26% of the heap (99.26% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 12 of 17
+  80 bytes (77 requested / 3 slop)
+  0.26% of the heap (99.52% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 13 of 17
+  64 bytes (64 requested / 0 slop)
+  0.21% of the heap (99.74% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 14 of 17
+  32 bytes (30 requested / 2 slop)
+  0.11% of the heap (99.84% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 15 of 17
+  16 bytes (16 requested / 0 slop)
+  0.05% of the heap (99.89% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 16 of 17
+  16 bytes (16 requested / 0 slop)
+  0.05% of the heap (99.95% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Cumulative {
+  1 block in heap block record 17 of 17
+  16 bytes (10 requested / 6 slop)
+  0.05% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 30,208 bytes in 135 blocks
+}
+
diff --git a/memory/replace/dmd/test/complete-full2-dark-matter-expected.txt b/memory/replace/dmd/test/complete-full2-dark-matter-expected.txt
new file mode 100644
index 0000000000..5f9585a8c6
--- /dev/null
+++ b/memory/replace/dmd/test/complete-full2-dark-matter-expected.txt
@@ -0,0 +1,140 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-full2-dark-matter-actual.txt complete-full2-dark-matter.json
+
+Invocation {
+  $DMD = '--mode=dark-matter --stacks=full --show-dump-stats=yes'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+Twice-reported {
+  1 block in heap block record 1 of 2
+  80 bytes (77 requested / 3 slop)
+  2.81% of the heap (2.81% cumulative)
+  83.33% of twice-reported (83.33% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Twice-reported {
+  1 block in heap block record 2 of 2
+  16 bytes (16 requested / 0 slop)
+  0.56% of the heap (3.37% cumulative)
+  16.67% of twice-reported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+  Reported again at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Unreported {
+  9 blocks in heap block record 1 of 3
+  1,008 bytes (900 requested / 108 slop)
+  Individual block sizes: 112 x 9
+  35.39% of the heap (35.39% cumulative)
+  48.84% of unreported (48.84% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Unreported {
+  6 blocks in heap block record 2 of 3
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  18.54% of the heap (53.93% cumulative)
+  25.58% of unreported (74.42% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Unreported {
+  6 blocks in heap block record 3 of 3
+  528 bytes (528 requested / 0 slop)
+  Individual block sizes: 128; 112; 96; 80; 64; 48
+  18.54% of the heap (72.47% cumulative)
+  25.58% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Once-reported {
+  1 block in heap block record 1 of 4
+  512 bytes (512 requested / 0 slop)
+  17.98% of the heap (17.98% cumulative)
+  74.42% of once-reported (74.42% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 2 of 4
+  80 bytes (79 requested / 1 slop)
+  2.81% of the heap (20.79% cumulative)
+  11.63% of once-reported (86.05% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 3 of 4
+  80 bytes (78 requested / 2 slop)
+  2.81% of the heap (23.60% cumulative)
+  11.63% of once-reported (97.67% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Once-reported {
+  1 block in heap block record 4 of 4
+  16 bytes (10 requested / 6 slop)
+  0.56% of the heap (24.16% cumulative)
+  2.33% of once-reported (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+  Reported at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:                 2,848 bytes (100.00%) in      27 blocks (100.00%)
+  Unreported:            2,064 bytes ( 72.47%) in      21 blocks ( 77.78%)
+  Once-reported:           688 bytes ( 24.16%) in       4 blocks ( 14.81%)
+  Twice-reported:           96 bytes (  3.37%) in       2 blocks (  7.41%)
+}
+
diff --git a/memory/replace/dmd/test/complete-partial-live-expected.txt b/memory/replace/dmd/test/complete-partial-live-expected.txt
new file mode 100644
index 0000000000..e7f27b0ee6
--- /dev/null
+++ b/memory/replace/dmd/test/complete-partial-live-expected.txt
@@ -0,0 +1,56 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o complete-partial-live-actual.txt complete-partial-live.json
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  9,531 blocks in heap block record 1 of 4
+  9,759,744 bytes (9,759,744 requested / 0 slop)
+  Individual block sizes: 1,024 x 9,531
+  83.56% of the heap (83.56% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  16,822 blocks in heap block record 2 of 4
+  1,510,672 bytes (1,510,672 requested / 0 slop)
+  Individual block sizes: 1,024 x 469; 128 x 6,864; 16 x 9,489
+  12.93% of the heap (96.49% cumulative)
+  Allocated at {
+    #01: (no stack trace recorded due to --stacks=partial)
+  }
+}
+
+Live {
+  3,136 blocks in heap block record 3 of 4
+  401,408 bytes (401,408 requested / 0 slop)
+  Individual block sizes: 128 x 3,136
+  3.44% of the heap (99.93% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+Live {
+  511 blocks in heap block record 4 of 4
+  8,176 bytes (8,176 requested / 0 slop)
+  Individual block sizes: 16 x 511
+  0.07% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: ... DMD.cpp ...
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 11,680,000 bytes in 30,000 blocks
+}
+
diff --git a/memory/replace/dmd/test/moz.build b/memory/replace/dmd/test/moz.build
new file mode 100644
index 0000000000..d6ec431c70
--- /dev/null
+++ b/memory/replace/dmd/test/moz.build
@@ -0,0 +1,26 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+GeckoSimplePrograms(
+    [
+        "SmokeDMD",
+    ],
+    linkage=None,
+)
+
+# See the comment at the top of SmokeDMD.cpp:RunTests().
+if CONFIG["CC_TYPE"] == "clang-cl":
+    CXXFLAGS += ["-Od", "-clang:-fno-lto"]
+else:
+    CXXFLAGS += ["-O0", "-fno-lto"]
+
+DEFINES["MOZ_NO_MOZALLOC"] = True
+
+DisableStlWrapping()
+
+XPCSHELL_TESTS_MANIFESTS += [
+    "xpcshell.ini",
+]
diff --git a/memory/replace/dmd/test/scan-test.py b/memory/replace/dmd/test/scan-test.py
new file mode 100644
index 0000000000..c282b02693
--- /dev/null
+++ b/memory/replace/dmd/test/scan-test.py
@@ -0,0 +1,102 @@
+#! /usr/bin/env python
+#
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+"""Testing for the JSON file emitted by DMD heap scan mode when running SmokeDMD."""
+
+import argparse
+import gzip
+import json
+import sys
+
+# The DMD output version this script handles.
+outputVersion = 5
+
+
+def parseCommandLine():
+    description = """
+Ensure that DMD heap scan mode creates the correct output when run with SmokeDMD.
+This is only for testing. Input files can be gzipped.
+"""
+    p = argparse.ArgumentParser(description=description)
+
+    p.add_argument(
+        "--clamp-contents",
+        action="store_true",
+        help="expect that the contents of the JSON input file have had "
+        "their addresses clamped",
+    )
+
+    p.add_argument("input_file", help="a file produced by DMD")
+
+    return p.parse_args(sys.argv[1:])
+
+
+def checkScanContents(contents, expected):
+    if len(contents) != len(expected):
+        raise Exception(
+            "Expected "
+            + str(len(expected))
+            + " things in contents but found "
+            + str(len(contents))
+        )
+
+    for i in range(len(expected)):
+        if contents[i] != expected[i]:
+            raise Exception(
+                "Expected to find "
+                + expected[i]
+                + " at offset "
+                + str(i)
+                + " but found "
+                + contents[i]
+            )
+
+
+def main():
+    args = parseCommandLine()
+
+    # Handle gzipped input if necessary.
+    isZipped = args.input_file.endswith(".gz")
+    opener = gzip.open if isZipped else open
+
+    with opener(args.input_file, "rb") as f:
+        j = json.load(f)
+
+    if j["version"] != outputVersion:
+        raise Exception("'version' property isn't '{:d}'".format(outputVersion))
+
+    invocation = j["invocation"]
+
+    mode = invocation["mode"]
+    if mode != "scan":
+        raise Exception("bad 'mode' property: '{:s}'".format(mode))
+
+    blockList = j["blockList"]
+
+    if len(blockList) != 1:
+        raise Exception("Expected only one block")
+
+    b = blockList[0]
+
+    # The expected values are based on hard-coded values in SmokeDMD.cpp.
+    if args.clamp_contents:
+        expected = ["0", "0", "0", b["addr"], b["addr"]]
+    else:
+        addr = int(b["addr"], 16)
+        expected = [
+            "123",
+            "0",
+            str(format(addr - 1, "x")),
+            b["addr"],
+            str(format(addr + 1, "x")),
+            "0",
+        ]
+
+    checkScanContents(b["contents"], expected)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/memory/replace/dmd/test/script-diff-dark-matter-expected.txt b/memory/replace/dmd/test/script-diff-dark-matter-expected.txt
new file mode 100644
index 0000000000..b1fc28bac5
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-dark-matter-expected.txt
@@ -0,0 +1,127 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-diff-dark-matter-actual.txt script-diff-dark-matter1.json script-diff-dark-matter2.json
+
+Invocation 1 {
+  $DMD = '--mode=dark-matter'
+  Mode = 'dark-matter'
+}
+
+Invocation 2 {
+  $DMD = '1'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+Twice-reported {
+  -1 blocks in heap block record 1 of 1
+  -1,088 bytes (-1,064 requested / -24 slop)
+  Individual block sizes: -1,024; -127; 63
+  15.46% of the heap (15.46% cumulative)
+  100.00% of twice-reported (100.00% cumulative)
+  Allocated at {
+    #01: F (F.cpp:99)
+  }
+  Reported at {
+    #01: R1 (R1.cpp:99)
+  }
+  Reported again at {
+    #01: R2 (R2.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Unreported {
+  4 blocks in heap block record 1 of 5
+  16,384 bytes (16,384 requested / 0 slop)
+  Individual block sizes: 4,096 x 4
+  -232.76% of the heap (-232.76% cumulative)
+  371.01% of unreported (371.01% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+Unreported {
+  7 blocks in heap block record 2 of 5
+  -11,968 bytes (-12,016 requested / 48 slop)
+  Individual block sizes: -15,360; 2,048; 512 x 2; 128; -127; 64 x 4; 63
+  170.02% of the heap (-62.74% cumulative)
+  -271.01% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: F (F.cpp:99)
+  }
+}
+
+Unreported {
+  0 blocks in heap block record 3 of 5
+  0 bytes (-384 requested / 384 slop)
+  Individual block sizes: (no change)
+  -0.00% of the heap (-62.74% cumulative)
+  0.00% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+Unreported {
+  -2 blocks in heap block record 4 of 5
+  0 bytes (0 requested / 0 slop)
+  Individual block sizes: 8,192 x 2; -4,096 x 4
+  -0.00% of the heap (-62.74% cumulative)
+  0.00% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Unreported {
+  0 blocks in heap block record 5 of 5
+  0 bytes (0 requested / 0 slop)
+  Individual block sizes: 20,480; -16,384; -8,192; 4,096
+  -0.00% of the heap (-62.74% cumulative)
+  0.00% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: (no stack trace recorded due to --stacks=partial)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Once-reported {
+  -3 blocks in heap block record 1 of 2
+  -10,240 bytes (-10,192 requested / -48 slop)
+  Individual block sizes: -4,096 x 2; -2,048
+  145.48% of the heap (145.48% cumulative)
+  98.77% of once-reported (98.77% cumulative)
+  Allocated at {
+    #01: D (D.cpp:99)
+  }
+  Reported at {
+    #01: R1 (R1.cpp:99)
+  }
+}
+
+Once-reported {
+  -1 blocks in heap block record 2 of 2
+  -127 bytes (-151 requested / 24 slop)
+  1.80% of the heap (147.28% cumulative)
+  1.23% of once-reported (100.00% cumulative)
+  Allocated at {
+    #01: F (F.cpp:99)
+  }
+  Reported at {
+    #01: R1 (R1.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:                -7,039 bytes (100.00%) in       4 blocks (100.00%)
+  Unreported:            4,416 bytes (-62.74%) in       9 blocks (225.00%)
+  Once-reported:       -10,367 bytes (147.28%) in      -4 blocks (-100.00%)
+  Twice-reported:       -1,088 bytes ( 15.46%) in      -1 blocks (-25.00%)
+}
+
diff --git a/memory/replace/dmd/test/script-diff-dark-matter1.json b/memory/replace/dmd/test/script-diff-dark-matter1.json
new file mode 100644
index 0000000000..c8edafebe9
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-dark-matter1.json
@@ -0,0 +1,51 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "--mode=dark-matter",
+  "mode": "dark-matter"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A", "num": 4},
+
+  {"req": 4096, "alloc": "B", "num": 3},
+  {"req": 4096, "alloc": "B"},
+
+  {"req": 4096, "alloc": "C", "num": 2},
+  {"req": 4096, "alloc": "C", "num": 2},
+
+  {"req": 4096,             "alloc": "D", "reps": ["R1"], "num": 2},
+  {"req": 2000, "slop": 48, "alloc": "D", "reps": ["R1"]},
+
+  {"req": 15360,            "alloc": "F"},
+  {"req": 512,              "alloc": "F", "num": 2},
+  {"req": 127,              "alloc": "F"},
+  {"req": 1024,             "alloc": "F", "reps": ["R1"]},
+  {"req": 127,              "alloc": "F", "reps": ["R1"]},
+  {"req": 1000, "slop": 24, "alloc": "F", "reps": ["R1", "R2"]},
+  {"req": 127,              "alloc": "F", "reps": ["R1", "R2"]},
+
+  {"req": 4096 },
+  {"req": 8192 },
+  {"req": 16384 }
+ ],
+ "traceTable": {
+  "A": ["AA"],
+  "B": ["BB"],
+  "C": ["CC"],
+  "D": ["DD"],
+  "E": ["EE"],
+  "F": ["FF"],
+  "R1": ["RR1"],
+  "R2": ["RR2"]
+ },
+ "frameTable": {
+  "AA": "#00: A (A.cpp:99)",
+  "BB": "#00: B (B.cpp:99)",
+  "CC": "#00: C (C.cpp:99)",
+  "DD": "#00: D (D.cpp:99)",
+  "EE": "#00: E (E.cpp:99)",
+  "FF": "#00: F (F.cpp:99)",
+  "RR1": "#00: R1 (R1.cpp:99)",
+  "RR2": "#00: R2 (R2.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-diff-dark-matter2.json b/memory/replace/dmd/test/script-diff-dark-matter2.json
new file mode 100644
index 0000000000..94b8888b89
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-dark-matter2.json
@@ -0,0 +1,51 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "1",
+  "mode": "dark-matter"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A", "num": 4},
+
+  {"req": 8192, "alloc": "B"},
+  {"req": 8192, "alloc": "B"},
+
+  {"req": 4000, "slop": 96, "alloc": "C", "num": 4},
+
+  {"req": 4096, "alloc": "E", "num": 4},
+
+  {"req": 2000, "slop": 48, "alloc": "F"},
+  {"req": 1000, "slop": 24, "alloc": "F", "reps": ["R1"]},
+  {"req": 512,              "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 128,              "alloc": "F"},
+  {"req": 63,               "alloc": "F", "reps": ["R1", "R2"]},
+  {"req": 64,               "alloc": "F", "num": 4},
+  {"req": 63,               "alloc": "F"},
+
+  {"req": 4096, "num": 2 },
+  {"req": 20480 }
+ ],
+ "traceTable": {
+  "A": ["AA"],
+  "B": ["BB"],
+  "C": ["CC"],
+  "D": ["DD"],
+  "E": ["EE"],
+  "F": ["FF"],
+  "R1": ["RR1"],
+  "R2": ["RR2"]
+ },
+ "frameTable": {
+  "AA": "#00: A (A.cpp:99)",
+  "BB": "#00: B (B.cpp:99)",
+  "CC": "#00: C (C.cpp:99)",
+  "DD": "#00: D (D.cpp:99)",
+  "EE": "#00: E (E.cpp:99)",
+  "FF": "#00: F (F.cpp:99)",
+  "RR1": "#00: R1 (R1.cpp:99)",
+  "RR2": "#00: R2 (R2.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-diff-live-expected.txt b/memory/replace/dmd/test/script-diff-live-expected.txt
new file mode 100644
index 0000000000..20208c0768
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-live-expected.txt
@@ -0,0 +1,81 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-diff-live-actual.txt script-diff-live1.json script-diff-live2.json
+
+Invocation 1 {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+Invocation 2 {
+  $DMD = '--mode=live --stacks=partial'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  4 blocks in heap block record 1 of 6
+  16,384 bytes (16,384 requested / 0 slop)
+  Individual block sizes: 4,096 x 4
+  -232.76% of the heap (-232.76% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 6
+  -13,183 bytes (-13,231 requested / 48 slop)
+  Individual block sizes: -15,360; 2,048; -1,024; 512 x 2; 128; -127 x 3; 64 x 4; 63 x 2
+  187.29% of the heap (-45.48% cumulative)
+  Allocated at {
+    #01: F (F.cpp:99)
+  }
+}
+
+Live {
+  -3 blocks in heap block record 3 of 6
+  -10,240 bytes (-10,192 requested / -48 slop)
+  Individual block sizes: -4,096 x 2; -2,048
+  145.48% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: D (D.cpp:99)
+  }
+}
+
+Live {
+  0 blocks in heap block record 4 of 6
+  0 bytes (-384 requested / 384 slop)
+  Individual block sizes: (no change)
+  -0.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+Live {
+  -2 blocks in heap block record 5 of 6
+  0 bytes (0 requested / 0 slop)
+  Individual block sizes: 8,192 x 2; -4,096 x 4
+  -0.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  0 blocks in heap block record 6 of 6
+  0 bytes (0 requested / 0 slop)
+  Individual block sizes: 20,480; -16,384; -8,192; 4,096
+  -0.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: (no stack trace recorded due to --stacks=partial)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: -7,039 bytes in 4 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-diff-live1.json b/memory/replace/dmd/test/script-diff-live1.json
new file mode 100644
index 0000000000..87e07aed5e
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-live1.json
@@ -0,0 +1,51 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "--mode=live",
+  "mode": "live"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A", "num": 4},
+
+  {"req": 4096, "alloc": "B", "num": 4},
+
+  {"req": 4096, "alloc": "C", "num": 4},
+
+  {"req": 4096,             "alloc": "D"},
+  {"req": 4096,             "alloc": "D"},
+  {"req": 2000, "slop": 48, "alloc": "D"},
+
+  {"req": 15360,            "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 512,              "alloc": "F"},
+  {"req": 127,              "alloc": "F"},
+  {"req": 1024,             "alloc": "F"},
+  {"req": 127,              "alloc": "F"},
+  {"req": 1000, "slop": 24, "alloc": "F"},
+  {"req": 127,              "alloc": "F"},
+
+  {"req": 4096 },
+  {"req": 8192 },
+  {"req": 16384 }
+ ],
+ "traceTable": {
+  "A": ["AA"],
+  "B": ["BB"],
+  "C": ["CC"],
+  "D": ["DD"],
+  "E": ["EE"],
+  "F": ["FF"],
+  "R1": ["RR1"],
+  "R2": ["RR2"]
+ },
+ "frameTable": {
+  "AA": "#00: A (A.cpp:99)",
+  "BB": "#00: B (B.cpp:99)",
+  "CC": "#00: C (C.cpp:99)",
+  "DD": "#00: D (D.cpp:99)",
+  "EE": "#00: E (E.cpp:99)",
+  "FF": "#00: F (F.cpp:99)",
+  "RR1": "#00: R1 (R1.cpp:99)",
+  "RR2": "#00: R2 (R2.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-diff-live2.json b/memory/replace/dmd/test/script-diff-live2.json
new file mode 100644
index 0000000000..4c7476f4c3
--- /dev/null
+++ b/memory/replace/dmd/test/script-diff-live2.json
@@ -0,0 +1,53 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "--mode=live --stacks=partial",
+  "mode": "live"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A", "num": 3},
+  {"req": 4096, "alloc": "A"},
+
+  {"req": 8192, "alloc": "B"},
+  {"req": 8192, "alloc": "B"},
+
+  {"req": 4000, "slop": 96, "alloc": "C", "num": 4},
+
+  {"req": 4096, "alloc": "E"},
+  {"req": 4096, "alloc": "E"},
+  {"req": 4096, "alloc": "E"},
+  {"req": 4096, "alloc": "E"},
+
+  {"req": 2000, "slop": 48, "alloc": "F"},
+  {"req": 1000, "slop": 24, "alloc": "F"},
+  {"req": 512,              "alloc": "F", "num": 4},
+  {"req": 128,              "alloc": "F"},
+  {"req": 63,               "alloc": "F"},
+  {"req": 64,               "alloc": "F", "num": 4},
+  {"req": 63,               "alloc": "F"},
+
+  {"req": 4096 },
+  {"req": 4096 },
+  {"req": 20480 }
+ ],
+ "traceTable": {
+  "A": ["AA"],
+  "B": ["BB"],
+  "C": ["CC"],
+  "D": ["DD"],
+  "E": ["EE"],
+  "F": ["FF"],
+  "R1": ["RR1"],
+  "R2": ["RR2"]
+ },
+ "frameTable": {
+  "AA": "#00: A (A.cpp:99)",
+  "BB": "#00: B (B.cpp:99)",
+  "CC": "#00: C (C.cpp:99)",
+  "DD": "#00: D (D.cpp:99)",
+  "EE": "#00: E (E.cpp:99)",
+  "FF": "#00: F (F.cpp:99)",
+  "RR1": "#00: R1 (R1.cpp:99)",
+  "RR2": "#00: R2 (R2.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-ignore-alloc-fns-expected.txt b/memory/replace/dmd/test/script-ignore-alloc-fns-expected.txt
new file mode 100644
index 0000000000..9428ef45fb
--- /dev/null
+++ b/memory/replace/dmd/test/script-ignore-alloc-fns-expected.txt
@@ -0,0 +1,72 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-ignore-alloc-fns-actual.txt --ignore-alloc-fns script-ignore-alloc-fns.json
+
+Invocation {
+  $DMD = '1'
+  Mode = 'dark-matter'
+}
+
+#-----------------------------------------------------------------
+
+# no twice-reported heap blocks
+
+#-----------------------------------------------------------------
+
+Unreported {
+  1 block in heap block record 1 of 4
+  1,048,576 bytes (1,048,576 requested / 0 slop)
+  93.22% of the heap (93.22% cumulative)
+  93.22% of unreported (93.22% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+Unreported {
+  1 block in heap block record 2 of 4
+  65,536 bytes (65,536 requested / 0 slop)
+  5.83% of the heap (99.05% cumulative)
+  5.83% of unreported (99.05% cumulative)
+  Allocated at {
+    #01: js::jit::JitRuntime::initialize(JSContext*) (Ion.cpp:301)
+  }
+}
+
+Unreported {
+  1 block in heap block record 3 of 4
+  8,192 bytes (8,000 requested / 192 slop)
+  0.73% of the heap (99.78% cumulative)
+  0.73% of unreported (99.78% cumulative)
+  Allocated at {
+    #01: mozilla::Vector::growStorageBy(unsigned long) (Vector.h:802)
+    #02: D (D.cpp:99)
+  }
+}
+
+Unreported {
+  1 block in heap block record 4 of 4
+  2,500 bytes (2,500 requested / 0 slop)
+  0.22% of the heap (100.00% cumulative)
+  0.22% of unreported (100.00% cumulative)
+  Allocated at {
+    #01: g_type_create_instance (/usr/lib/x86_64-linux-gnu/libgobject-2.0.so.0)
+    #02: not_an_alloc_function_so_alloc_functions_below_here_will_not_be_stripped (blah)
+    #03: replace_posix_memalign (replace_malloc.h:120)
+    #04: ??? (/lib/x86_64-linux-gnu/libglib-2.0.so.0)
+    #05: another_non_alloc_function (blah)
+  }
+}
+
+#-----------------------------------------------------------------
+
+# no once-reported heap blocks
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total:             1,124,804 bytes (100.00%) in       4 blocks (100.00%)
+  Unreported:        1,124,804 bytes (100.00%) in       4 blocks (100.00%)
+  Once-reported:             0 bytes (  0.00%) in       0 blocks (  0.00%)
+  Twice-reported:            0 bytes (  0.00%) in       0 blocks (  0.00%)
+}
+
diff --git a/memory/replace/dmd/test/script-ignore-alloc-fns.json b/memory/replace/dmd/test/script-ignore-alloc-fns.json
new file mode 100644
index 0000000000..900d33e759
--- /dev/null
+++ b/memory/replace/dmd/test/script-ignore-alloc-fns.json
@@ -0,0 +1,46 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "1",
+  "mode": "dark-matter"
+ },
+ "blockList": [
+  {"req": 1048576,           "alloc": "A"},
+  {"req": 65536,             "alloc": "B"},
+  {"req": 8000, "slop": 192, "alloc": "C"},
+  {"req": 2500,              "alloc": "D"}
+ ],
+ "traceTable": {
+  "A": ["AA", "AB", "AC", "AD"],
+  "B": ["BA", "BB", "BC"],
+  "C": ["CA", "CB", "CC", "CD"],
+  "D": ["DA", "DB", "DD", "DD", "DE", "DF", "DG", "DH", "DI", "DJ"]
+ },
+ "frameTable": {
+  "AA": "#00: replace_malloc (DMD.cpp:1106)",
+  "AB": "#00: moz_xmalloc (mozalloc.cpp:68)",
+  "AC": "#00: operator new(unsigned long) (mozalloc.h:208)",
+  "AD": "#00: A (A.cpp:99)",
+
+  "BA": "#00: replace_calloc (DMD.cpp:1125)",
+  "BB": "#00: js_calloc(unsigned long) (Utility.h:107)",
+  "BC": "#06: js::jit::JitRuntime::initialize(JSContext*) (Ion.cpp:301)",
+
+  "CA": "#00: replace_realloc (DMD.cpp:1153)",
+  "CB": "#00: bool* mozilla::MallocAllocPolicy::pod_realloc<bool>(bool*, unsigned long, unsigned long) (AllocPolicy.h:74)",
+  "CC": "#00: mozilla::Vector::growStorageBy(unsigned long) (Vector.h:802)",
+  "CD": "#00: D (D.cpp:99)",
+
+  "DA": "#00: replace_memalign (DMD.cpp:1181)",
+  "DB": "#00: replace_posix_memalign (replace_malloc.h:120)",
+  "DC": "#00: ??? (/lib/x86_64-linux-gnu/libglib-2.0.so.0)",
+  "DD": "#00: g_slice_alloc (/lib/x86_64-linux-gnu/libglib-2.0.so.0)",
+  "DE": "#00: g_slice_alloc0 (/lib/x86_64-linux-gnu/libglib-2.0.so.0)",
+  "DF": "#00: g_type_create_instance (/usr/lib/x86_64-linux-gnu/libgobject-2.0.so.0)",
+  "DG": "#00: not_an_alloc_function_so_alloc_functions_below_here_will_not_be_stripped (blah)",
+  "DH": "#00: replace_posix_memalign (replace_malloc.h:120)",
+  "DI": "#00: ??? (/lib/x86_64-linux-gnu/libglib-2.0.so.0)",
+  "DJ": "#00: another_non_alloc_function (blah)"
+ }
+}
+
diff --git a/memory/replace/dmd/test/script-max-frames-1-expected.txt b/memory/replace/dmd/test/script-max-frames-1-expected.txt
new file mode 100644
index 0000000000..65a00762bb
--- /dev/null
+++ b/memory/replace/dmd/test/script-max-frames-1-expected.txt
@@ -0,0 +1,26 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-max-frames-1-actual.txt --max-frames=1 script-max-frames.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  4 blocks in heap block record 1 of 1
+  4,416 bytes (4,404 requested / 12 slop)
+  Individual block sizes: 4,096; 128; 112; 80
+  100.00% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 4,416 bytes in 4 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-max-frames-3-expected.txt b/memory/replace/dmd/test/script-max-frames-3-expected.txt
new file mode 100644
index 0000000000..5df4914738
--- /dev/null
+++ b/memory/replace/dmd/test/script-max-frames-3-expected.txt
@@ -0,0 +1,48 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-max-frames-3-actual.txt --max-frames=3 --no-fix-stacks script-max-frames.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  2 blocks in heap block record 1 of 3
+  4,224 bytes (4,224 requested / 0 slop)
+  Individual block sizes: 4,096; 128
+  95.65% of the heap (95.65% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: F (F.cpp:99)
+    #03: G (G.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 2 of 3
+  112 bytes (100 requested / 12 slop)
+  2.54% of the heap (98.19% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: X (X.cpp:99)
+    #03: Y (Y.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 3 of 3
+  80 bytes (80 requested / 0 slop)
+  1.81% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 4,416 bytes in 4 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-max-frames-8-expected.txt b/memory/replace/dmd/test/script-max-frames-8-expected.txt
new file mode 100644
index 0000000000..174992d5b8
--- /dev/null
+++ b/memory/replace/dmd/test/script-max-frames-8-expected.txt
@@ -0,0 +1,69 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-max-frames-8-actual.txt script-max-frames.json
+
+Invocation {
+  $DMD = '--mode=live --stacks=full'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  1 block in heap block record 1 of 4
+  4,096 bytes (4,096 requested / 0 slop)
+  92.75% of the heap (92.75% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: F (F.cpp:99)
+    #03: G (G.cpp:99)
+    #04: H (H.cpp:99)
+    #05: I (I.cpp:99)
+    #06: J (J.cpp:99)
+    #07: K (K.cpp:99)
+    #08: L (L.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 2 of 4
+  128 bytes (128 requested / 0 slop)
+  2.90% of the heap (95.65% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: F (F.cpp:99)
+    #03: G (G.cpp:99)
+    #04: R (R.cpp:99)
+    #05: S (S.cpp:99)
+    #06: T (T.cpp:99)
+    #07: U (U.cpp:99)
+    #08: V (V.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 3 of 4
+  112 bytes (100 requested / 12 slop)
+  2.54% of the heap (98.19% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+    #02: X (X.cpp:99)
+    #03: Y (Y.cpp:99)
+    #04: Z (Z.cpp:99)
+  }
+}
+
+Live {
+  1 block in heap block record 4 of 4
+  80 bytes (80 requested / 0 slop)
+  1.81% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: E (E.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 4,416 bytes in 4 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-max-frames.json b/memory/replace/dmd/test/script-max-frames.json
new file mode 100644
index 0000000000..690d50fa7e
--- /dev/null
+++ b/memory/replace/dmd/test/script-max-frames.json
@@ -0,0 +1,43 @@
+{
+ "version": 5,
+ "invocation": {
+  "dmdEnvVar": "--mode=live --stacks=full",
+  "mode": "live"
+ },
+ "blockList": [
+  {"req": 4096, "alloc": "A"},
+  {"req": 128, "alloc": "B"},
+  {"req": 100, "slop":12, "alloc": "C"},
+  {"req": 80, "alloc": "D"}
+ ],
+ "traceTable": {
+  "A": ["E", "F", "G", "H", "I", "J", "K", "L", "M", "N", "O", "P"],
+  "B": ["E", "F", "G", "R", "S", "T", "U", "V"],
+  "C": ["E", "X", "Y", "Z"],
+  "D": ["E"]
+ },
+ "frameTable": {
+  "E": "#00: E (E.cpp:99)",
+  "F": "#00: F (F.cpp:99)",
+  "G": "#00: G (G.cpp:99)",
+  "H": "#00: H (H.cpp:99)",
+  "I": "#00: I (I.cpp:99)",
+  "J": "#00: J (J.cpp:99)",
+  "K": "#00: K (K.cpp:99)",
+  "L": "#00: L (L.cpp:99)",
+  "M": "#00: M (M.cpp:99)",
+  "N": "#00: N (N.cpp:99)",
+  "O": "#00: O (O.cpp:99)",
+  "P": "#00: P (P.cpp:99)",
+  "Q": "#00: Q (Q.cpp:99)",
+  "R": "#00: R (R.cpp:99)",
+  "S": "#00: S (S.cpp:99)",
+  "T": "#00: T (T.cpp:99)",
+  "U": "#00: U (U.cpp:99)",
+  "V": "#00: V (V.cpp:99)",
+  "W": "#00: W (W.cpp:99)",
+  "X": "#00: X (X.cpp:99)",
+  "Y": "#00: Y (Y.cpp:99)",
+  "Z": "#00: Z (Z.cpp:99)"
+ }
+}
diff --git a/memory/replace/dmd/test/script-sort-by-num-blocks-expected.txt b/memory/replace/dmd/test/script-sort-by-num-blocks-expected.txt
new file mode 100644
index 0000000000..8de03d953b
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by-num-blocks-expected.txt
@@ -0,0 +1,46 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-sort-by-num-blocks-actual.txt --sort-by=num-blocks script-sort-by.json.gz
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  8 blocks in heap block record 1 of 3
+  16,384 bytes (8,200 requested / 8,184 slop)
+  Individual block sizes: 2,048 x 8
+  33.32% of the heap (33.32% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 3
+  16,400 bytes (12,016 requested / 4,384 slop)
+  Individual block sizes: 4,096 x 4; 16
+  33.35% of the heap (66.67% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 3 of 3
+  16,392 bytes (16,392 requested / 0 slop)
+  Individual block sizes: 4,096 x 4; 8
+  33.33% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 49,176 bytes in 18 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-sort-by-req-expected.txt b/memory/replace/dmd/test/script-sort-by-req-expected.txt
new file mode 100644
index 0000000000..3ab21ba8f7
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by-req-expected.txt
@@ -0,0 +1,46 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-sort-by-req-actual.txt --sort-by=req --no-fix-stacks script-sort-by.json.gz
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  5 blocks in heap block record 1 of 3
+  16,392 bytes (16,392 requested / 0 slop)
+  Individual block sizes: 4,096 x 4; 8
+  33.33% of the heap (33.33% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 3
+  16,400 bytes (12,016 requested / 4,384 slop)
+  Individual block sizes: 4,096 x 4; 16
+  33.35% of the heap (66.68% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  8 blocks in heap block record 3 of 3
+  16,384 bytes (8,200 requested / 8,184 slop)
+  Individual block sizes: 2,048 x 8
+  33.32% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 49,176 bytes in 18 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-sort-by-slop-expected.txt b/memory/replace/dmd/test/script-sort-by-slop-expected.txt
new file mode 100644
index 0000000000..c325c7ed40
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by-slop-expected.txt
@@ -0,0 +1,46 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-sort-by-slop-actual.txt --sort-by=slop script-sort-by.json.gz
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  8 blocks in heap block record 1 of 3
+  16,384 bytes (8,200 requested / 8,184 slop)
+  Individual block sizes: 2,048 x 8
+  33.32% of the heap (33.32% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 3
+  16,400 bytes (12,016 requested / 4,384 slop)
+  Individual block sizes: 4,096 x 4; 16
+  33.35% of the heap (66.67% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 3 of 3
+  16,392 bytes (16,392 requested / 0 slop)
+  Individual block sizes: 4,096 x 4; 8
+  33.33% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 49,176 bytes in 18 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-sort-by-usable-expected.txt b/memory/replace/dmd/test/script-sort-by-usable-expected.txt
new file mode 100644
index 0000000000..8239a4759e
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by-usable-expected.txt
@@ -0,0 +1,46 @@
+#-----------------------------------------------------------------
+# dmd.py --filter-stacks-for-testing -o script-sort-by-usable-actual.txt --sort-by=usable script-sort-by.json.gz
+
+Invocation {
+  $DMD = '--mode=live'
+  Mode = 'live'
+}
+
+#-----------------------------------------------------------------
+
+Live {
+  5 blocks in heap block record 1 of 3
+  16,400 bytes (12,016 requested / 4,384 slop)
+  Individual block sizes: 4,096 x 4; 16
+  33.35% of the heap (33.35% cumulative)
+  Allocated at {
+    #01: B (B.cpp:99)
+  }
+}
+
+Live {
+  5 blocks in heap block record 2 of 3
+  16,392 bytes (16,392 requested / 0 slop)
+  Individual block sizes: 4,096 x 4; 8
+  33.33% of the heap (66.68% cumulative)
+  Allocated at {
+    #01: A (A.cpp:99)
+  }
+}
+
+Live {
+  8 blocks in heap block record 3 of 3
+  16,384 bytes (8,200 requested / 8,184 slop)
+  Individual block sizes: 2,048 x 8
+  33.32% of the heap (100.00% cumulative)
+  Allocated at {
+    #01: C (C.cpp:99)
+  }
+}
+
+#-----------------------------------------------------------------
+
+Summary {
+  Total: 49,176 bytes in 18 blocks
+}
+
diff --git a/memory/replace/dmd/test/script-sort-by.json.gz b/memory/replace/dmd/test/script-sort-by.json.gz
new file mode 100644
index 0000000000..b2308bab40
--- /dev/null
+++ b/memory/replace/dmd/test/script-sort-by.json.gz
diff --git a/memory/replace/dmd/test/test_dmd.js b/memory/replace/dmd/test/test_dmd.js
new file mode 100644
index 0000000000..d96444ea05
--- /dev/null
+++ b/memory/replace/dmd/test/test_dmd.js
@@ -0,0 +1,224 @@
+/* -*- indent-tabs-mode: nil; js-indent-level: 2 -*-*/
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this file,
+ * You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+"use strict";
+
+const { FileUtils } = ChromeUtils.importESModule(
+  "resource://gre/modules/FileUtils.sys.mjs"
+);
+
+// The xpcshell test harness sets PYTHON so we can read it here.
+var gPythonName = Services.env.get("PYTHON");
+
+// If we're testing locally, the executable file is in "CurProcD". Otherwise,
+// it is in another location that we have to find.
+function getExecutable(aFilename) {
+  let file = FileUtils.getFile("CurProcD", [aFilename]);
+  if (!file.exists()) {
+    file = FileUtils.getFile("CurWorkD", []);
+    while (file.path.includes("xpcshell")) {
+      file = file.parent;
+    }
+    file.append("bin");
+    file.append(aFilename);
+  }
+  return file;
+}
+
+var gIsWindows = Services.appinfo.OS === "WINNT";
+var gDmdTestFile = getExecutable("SmokeDMD" + (gIsWindows ? ".exe" : ""));
+
+var gDmdScriptFile = getExecutable("dmd.py");
+
+var gScanTestFile = FileUtils.getFile("CurWorkD", ["scan-test.py"]);
+
+function readFile(aFile) {
+  let fstream = Cc["@mozilla.org/network/file-input-stream;1"].createInstance(
+    Ci.nsIFileInputStream
+  );
+  let cstream = Cc["@mozilla.org/intl/converter-input-stream;1"].createInstance(
+    Ci.nsIConverterInputStream
+  );
+  fstream.init(aFile, -1, 0, 0);
+  cstream.init(fstream, "UTF-8", 0, 0);
+
+  let data = "";
+  let str = {};
+  let read = 0;
+  do {
+    // Read as much as we can and put it in str.value.
+    read = cstream.readString(0xffffffff, str);
+    data += str.value;
+  } while (read != 0);
+
+  cstream.close(); // this closes fstream
+  return data.replace(/\r/g, ""); // normalize line endings
+}
+
+function runProcess(aExeFile, aArgs) {
+  let process = Cc["@mozilla.org/process/util;1"].createInstance(Ci.nsIProcess);
+  process.init(aExeFile);
+  process.run(/* blocking = */ true, aArgs, aArgs.length);
+  return process.exitValue;
+}
+
+function test(aPrefix, aArgs) {
+  // DMD writes the JSON files to CurWorkD, so we do likewise here with
+  // |actualFile| for consistency. It is removed once we've finished.
+  let expectedFile = FileUtils.getFile("CurWorkD", [aPrefix + "-expected.txt"]);
+  let actualFile = FileUtils.getFile("CurWorkD", [aPrefix + "-actual.txt"]);
+
+  // Run dmd.py on the JSON file, producing |actualFile|.
+
+  let args = [
+    gDmdScriptFile.path,
+    "--filter-stacks-for-testing",
+    "-o",
+    actualFile.path,
+  ].concat(aArgs);
+
+  runProcess(new FileUtils.File(gPythonName), args);
+
+  // Compare |expectedFile| with |actualFile|. We produce nice diffs with
+  // /usr/bin/diff on systems that have it (Mac and Linux). Otherwise (Windows)
+  // we do a string compare of the file contents and then print them both if
+  // they don't match.
+
+  let success;
+  try {
+    let rv = runProcess(new FileUtils.File("/usr/bin/diff"), [
+      "-u",
+      expectedFile.path,
+      actualFile.path,
+    ]);
+    success = rv == 0;
+  } catch (e) {
+    let expectedData = readFile(expectedFile);
+    let actualData = readFile(actualFile);
+    success = expectedData === actualData;
+    if (!success) {
+      expectedData = expectedData.split("\n");
+      actualData = actualData.split("\n");
+      for (let i = 0; i < expectedData.length; i++) {
+        print("EXPECTED:" + expectedData[i]);
+      }
+      for (let i = 0; i < actualData.length; i++) {
+        print("  ACTUAL:" + actualData[i]);
+      }
+    }
+  }
+
+  ok(success, aPrefix);
+
+  actualFile.remove(true);
+}
+
+// Run scan-test.py on the JSON file and see if it succeeds.
+function scanTest(aJsonFilePath, aExtraArgs) {
+  let args = [gScanTestFile.path, aJsonFilePath].concat(aExtraArgs);
+
+  return runProcess(new FileUtils.File(gPythonName), args) == 0;
+}
+
+function run_test() {
+  let jsonFile, jsonFile2;
+
+  // These tests do complete end-to-end testing of DMD, i.e. both the C++ code
+  // that generates the JSON output, and the script that post-processes that
+  // output.
+  //
+  // Run these synchronously, because test() updates the complete*.json files
+  // in-place (to fix stacks) when it runs dmd.py, and that's not safe to do
+  // asynchronously.
+
+  Services.env.set("DMD", "1");
+
+  runProcess(gDmdTestFile, []);
+
+  function test2(aTestName, aMode) {
+    let name = "complete-" + aTestName + "-" + aMode;
+    jsonFile = FileUtils.getFile("CurWorkD", [name + ".json"]);
+    test(name, [jsonFile.path]);
+    jsonFile.remove(true);
+  }
+
+  // Please keep this in sync with RunTests() in SmokeDMD.cpp.
+
+  test2("empty", "live");
+  test2("empty", "dark-matter");
+  test2("empty", "cumulative");
+
+  test2("full1", "live");
+  test2("full1", "dark-matter");
+
+  test2("full2", "dark-matter");
+  test2("full2", "cumulative");
+
+  test2("partial", "live");
+
+  // Heap scan testing.
+  jsonFile = FileUtils.getFile("CurWorkD", ["basic-scan.json"]);
+  ok(scanTest(jsonFile.path), "Basic scan test");
+
+  let is64Bit = Services.appinfo.is64Bit;
+  let basicScanFileName = "basic-scan-" + (is64Bit ? "64" : "32");
+  test(basicScanFileName, ["--clamp-contents", jsonFile.path]);
+  ok(
+    scanTest(jsonFile.path, ["--clamp-contents"]),
+    "Scan with address clamping"
+  );
+
+  // Run the generic test a second time to ensure that the first time produced
+  // valid JSON output. "--clamp-contents" is passed in so we don't have to have
+  // more variants of the files.
+  test(basicScanFileName, ["--clamp-contents", jsonFile.path]);
+  jsonFile.remove(true);
+
+  // These tests only test the post-processing script. They use hand-written
+  // JSON files as input. Ideally the JSON files would contain comments
+  // explaining how they work, but JSON doesn't allow comments, so I've put
+  // explanations here.
+
+  // This just tests that stack traces of various lengths are truncated
+  // appropriately. The number of records in the output is different for each
+  // of the tested values.
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-max-frames.json"]);
+  test("script-max-frames-8", [jsonFile.path]); // --max-frames=8 is the default
+  test("script-max-frames-3", [
+    "--max-frames=3",
+    "--no-fix-stacks",
+    jsonFile.path,
+  ]);
+  test("script-max-frames-1", ["--max-frames=1", jsonFile.path]);
+
+  // This file has three records that are shown in a different order for each
+  // of the different sort values. It also tests the handling of gzipped JSON
+  // files.
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-sort-by.json.gz"]);
+  test("script-sort-by-usable", ["--sort-by=usable", jsonFile.path]);
+  test("script-sort-by-req", [
+    "--sort-by=req",
+    "--no-fix-stacks",
+    jsonFile.path,
+  ]);
+  test("script-sort-by-slop", ["--sort-by=slop", jsonFile.path]);
+  test("script-sort-by-num-blocks", ["--sort-by=num-blocks", jsonFile.path]);
+
+  // This file has several real stack traces taken from Firefox execution, each
+  // of which tests a different allocator function (or functions).
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-ignore-alloc-fns.json"]);
+  test("script-ignore-alloc-fns", ["--ignore-alloc-fns", jsonFile.path]);
+
+  // This tests "live"-mode diffs.
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-diff-live1.json"]);
+  jsonFile2 = FileUtils.getFile("CurWorkD", ["script-diff-live2.json"]);
+  test("script-diff-live", [jsonFile.path, jsonFile2.path]);
+
+  // This tests "dark-matter"-mode diffs.
+  jsonFile = FileUtils.getFile("CurWorkD", ["script-diff-dark-matter1.json"]);
+  jsonFile2 = FileUtils.getFile("CurWorkD", ["script-diff-dark-matter2.json"]);
+  test("script-diff-dark-matter", [jsonFile.path, jsonFile2.path]);
+}
diff --git a/memory/replace/dmd/test/xpcshell.ini b/memory/replace/dmd/test/xpcshell.ini
new file mode 100644
index 0000000000..2baeeed270
--- /dev/null
+++ b/memory/replace/dmd/test/xpcshell.ini
@@ -0,0 +1,35 @@
+[DEFAULT]
+support-files =
+  basic-scan-32-expected.txt
+  basic-scan-64-expected.txt
+  complete-empty-live-expected.txt
+  complete-empty-dark-matter-expected.txt
+  complete-empty-cumulative-expected.txt
+  complete-full1-live-expected.txt
+  complete-full1-dark-matter-expected.txt
+  complete-full2-dark-matter-expected.txt
+  complete-full2-cumulative-expected.txt
+  complete-partial-live-expected.txt
+  scan-test.py
+  script-max-frames.json
+  script-max-frames-8-expected.txt
+  script-max-frames-3-expected.txt
+  script-max-frames-1-expected.txt
+  script-sort-by.json.gz
+  script-sort-by-usable-expected.txt
+  script-sort-by-req-expected.txt
+  script-sort-by-slop-expected.txt
+  script-sort-by-num-blocks-expected.txt
+  script-ignore-alloc-fns.json
+  script-ignore-alloc-fns-expected.txt
+  script-diff-live1.json
+  script-diff-live2.json
+  script-diff-live-expected.txt
+  script-diff-dark-matter1.json
+  script-diff-dark-matter2.json
+  script-diff-dark-matter-expected.txt
+
+[test_dmd.js]
+dmd = true
+skip-if =
+  !(os=='linux' || os=='mac' || (os=='win' && !pgo))
diff --git a/memory/replace/logalloc/FdPrintf.cpp b/memory/replace/logalloc/FdPrintf.cpp
new file mode 100644
index 0000000000..4a8e48af78
--- /dev/null
+++ b/memory/replace/logalloc/FdPrintf.cpp
@@ -0,0 +1,200 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <cstdarg>
+
+#ifdef _WIN32
+#  include <windows.h>
+#else
+#  include <unistd.h>
+#endif
+#include <cmath>
+#include <cstring>
+#include "mozilla/Assertions.h"
+#include "mozilla/Unused.h"
+#include "FdPrintf.h"
+
+/* Template class allowing a limited number of increments on a value */
+template <typename T>
+class CheckedIncrement {
+ public:
+  CheckedIncrement(T aValue, size_t aMaxIncrement)
+      : mValue(aValue), mMaxIncrement(aMaxIncrement) {}
+
+  T operator++(int) {
+    if (!mMaxIncrement) {
+      MOZ_CRASH("overflow detected");
+    }
+    mMaxIncrement--;
+    return mValue++;
+  }
+
+  T& operator++() {
+    (*this)++;
+    return mValue;
+  }
+
+  void advance(T end) {
+    // Only makes sense if T is a pointer type.
+    size_t diff = end - mValue;
+    if (diff > mMaxIncrement) {
+      MOZ_CRASH("overflow detected");
+    }
+    mMaxIncrement -= diff;
+    mValue = end;
+  };
+
+  void rewind(T pos) {
+    size_t diff = mValue - pos;
+    mMaxIncrement += diff;
+    mValue = pos;
+  }
+
+  operator T() { return mValue; }
+  T value() { return mValue; }
+
+ private:
+  T mValue;
+  size_t mMaxIncrement;
+};
+
+template <typename T>
+static unsigned NumDigits(T n) {
+  if (n < 1) {
+    // We want one digit, it will be 0.
+    return 1;
+  }
+
+  double l = log10(static_cast<double>(n));
+  double cl = ceil(l);
+  return l == cl ? unsigned(cl) + 1 : unsigned(cl);
+}
+
+static void LeftPad(CheckedIncrement<char*>& b, size_t pad) {
+  while (pad-- > 0) {
+    *(b++) = ' ';
+  }
+}
+
+// Write the digits into the buffer.
+static void WriteDigits(CheckedIncrement<char*>& b, size_t i,
+                        size_t num_digits) {
+  size_t x = pow(10, double(num_digits - 1));
+  do {
+    *(b++) = "0123456789"[(i / x) % 10];
+    x /= 10;
+  } while (x > 0);
+}
+
+void FdPrintf(intptr_t aFd, const char* aFormat, ...) {
+  if (aFd == 0) {
+    return;
+  }
+  char buf[256];
+  CheckedIncrement<char*> b(buf, sizeof(buf));
+  CheckedIncrement<const char*> f(aFormat, strlen(aFormat) + 1);
+  va_list ap;
+  va_start(ap, aFormat);
+  while (true) {
+    switch (*f) {
+      case '\0':
+        goto out;
+
+      case '%': {
+        // The start of the format specifier is used if this specifier is
+        // invalid.
+        const char* start = f;
+
+        // Read the field width
+        f++;
+        char* end = nullptr;
+        size_t width = strtoul(f, &end, 10);
+        // If strtol can't find a number that's okay, that means 0 in our
+        // case, but we must advance f).
+        f.advance(end);
+
+        switch (*f) {
+          case 'z': {
+            if (*(++f) == 'u') {
+              size_t i = va_arg(ap, size_t);
+
+              size_t num_digits = NumDigits(i);
+              LeftPad(b, width > num_digits ? width - num_digits : 0);
+              WriteDigits(b, i, num_digits);
+            } else {
+              // If the format specifier is unknown then write out '%' and
+              // rewind to the beginning of the specifier causing it to be
+              // printed normally.
+              *(b++) = '%';
+              f.rewind(start);
+            }
+            break;
+          }
+
+          case 'p': {
+            intptr_t ptr = va_arg(ap, intptr_t);
+            *(b++) = '0';
+            *(b++) = 'x';
+            int x = sizeof(intptr_t) * 8;
+            bool wrote_msb = false;
+            do {
+              x -= 4;
+              size_t hex_digit = ptr >> x & 0xf;
+              if (hex_digit || wrote_msb) {
+                *(b++) = "0123456789abcdef"[hex_digit];
+                wrote_msb = true;
+              }
+            } while (x > 0);
+            if (!wrote_msb) {
+              *(b++) = '0';
+            }
+            break;
+          }
+
+          case 's': {
+            const char* str = va_arg(ap, const char*);
+            size_t len = strlen(str);
+
+            LeftPad(b, width > len ? width - len : 0);
+
+            while (*str) {
+              *(b++) = *(str++);
+            }
+
+            break;
+          }
+
+          case '%':
+            // Print a single raw '%'.
+            *(b++) = '%';
+            break;
+
+          default:
+            // If the format specifier is unknown then write out '%' and
+            // rewind to the beginning of the specifier causing it to be
+            // printed normally.
+            *(b++) = '%';
+            f.rewind(start);
+            break;
+        }
+        break;
+      }
+      default:
+        *(b++) = *f;
+        break;
+    }
+    f++;
+  }
+out:
+#ifdef _WIN32
+  // See comment in FdPrintf.h as to why WriteFile is used.
+  DWORD written;
+  WriteFile(reinterpret_cast<HANDLE>(aFd), buf, b - buf, &written, nullptr);
+#else
+  MOZ_UNUSED(write(aFd, buf, b - buf));
+#endif
+  va_end(ap);
+}
diff --git a/memory/replace/logalloc/FdPrintf.h b/memory/replace/logalloc/FdPrintf.h
new file mode 100644
index 0000000000..f390d57ed5
--- /dev/null
+++ b/memory/replace/logalloc/FdPrintf.h
@@ -0,0 +1,27 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef __FdPrintf_h__
+#define __FdPrintf_h__
+
+/* We can't use libc's (f)printf because it would reenter in replace_malloc,
+ * So use a custom and simplified version.  Only %p, %zu, %s and %% are
+ * supported, %zu, %s, support width specifiers.
+ *
+ * /!\ This function used a fixed-size internal buffer. The caller is
+ * expected to not use a format string that may overflow.
+ * The aFd argument is a file descriptor on UNIX and a native win32 file
+ * handle on Windows (from CreateFile). We can't use the windows POSIX
+ * APIs is that they don't support O_APPEND in a multi-process-safe way,
+ * while CreateFile does.
+ */
+extern void FdPrintf(intptr_t aFd, const char* aFormat, ...)
+#ifdef __GNUC__
+    __attribute__((format(printf, 2, 3)))
+#endif
+    ;
+
+#endif /* __FdPrintf_h__ */
diff --git a/memory/replace/logalloc/LogAlloc.cpp b/memory/replace/logalloc/LogAlloc.cpp
new file mode 100644
index 0000000000..a976b0c674
--- /dev/null
+++ b/memory/replace/logalloc/LogAlloc.cpp
@@ -0,0 +1,238 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <cstdlib>
+#include <fcntl.h>
+
+#ifdef _WIN32
+#  include <windows.h>
+#  include <io.h>
+#  include <process.h>
+#else
+#  include <unistd.h>
+#  include <pthread.h>
+#endif
+
+#include "replace_malloc.h"
+#include "FdPrintf.h"
+#include "Mutex.h"
+
+static malloc_table_t sFuncs;
+static intptr_t sFd = 0;
+static bool sStdoutOrStderr = false;
+
+static Mutex sMutex MOZ_UNANNOTATED;
+
+#ifndef _WIN32
+static void prefork() MOZ_NO_THREAD_SAFETY_ANALYSIS { sMutex.Lock(); }
+static void postfork_parent() MOZ_NO_THREAD_SAFETY_ANALYSIS { sMutex.Unlock(); }
+static void postfork_child() { sMutex.Init(); }
+#endif
+
+static size_t GetPid() { return size_t(getpid()); }
+
+static size_t GetTid() {
+#if defined(_WIN32)
+  return size_t(GetCurrentThreadId());
+#else
+  return size_t(pthread_self());
+#endif
+}
+
+#ifdef ANDROID
+/* Android doesn't have pthread_atfork defined in pthread.h */
+extern "C" MOZ_EXPORT int pthread_atfork(void (*)(void), void (*)(void),
+                                         void (*)(void));
+#endif
+
+class LogAllocBridge : public ReplaceMallocBridge {
+  virtual void InitDebugFd(mozilla::DebugFdRegistry& aRegistry) override {
+    if (!sStdoutOrStderr) {
+      aRegistry.RegisterHandle(sFd);
+    }
+  }
+};
+
+/* Do a simple, text-form, log of all calls to replace-malloc functions.
+ * Use locking to guarantee that an allocation that did happen is logged
+ * before any other allocation/free happens.
+ */
+
+static void* replace_malloc(size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.malloc(aSize);
+  FdPrintf(sFd, "%zu %zu malloc(%zu)=%p\n", GetPid(), GetTid(), aSize, ptr);
+  return ptr;
+}
+
+static int replace_posix_memalign(void** aPtr, size_t aAlignment,
+                                  size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  int ret = sFuncs.posix_memalign(aPtr, aAlignment, aSize);
+  FdPrintf(sFd, "%zu %zu posix_memalign(%zu,%zu)=%p\n", GetPid(), GetTid(),
+           aAlignment, aSize, (ret == 0) ? *aPtr : nullptr);
+  return ret;
+}
+
+static void* replace_aligned_alloc(size_t aAlignment, size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.aligned_alloc(aAlignment, aSize);
+  FdPrintf(sFd, "%zu %zu aligned_alloc(%zu,%zu)=%p\n", GetPid(), GetTid(),
+           aAlignment, aSize, ptr);
+  return ptr;
+}
+
+static void* replace_calloc(size_t aNum, size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.calloc(aNum, aSize);
+  FdPrintf(sFd, "%zu %zu calloc(%zu,%zu)=%p\n", GetPid(), GetTid(), aNum, aSize,
+           ptr);
+  return ptr;
+}
+
+static void* replace_realloc(void* aPtr, size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* new_ptr = sFuncs.realloc(aPtr, aSize);
+  FdPrintf(sFd, "%zu %zu realloc(%p,%zu)=%p\n", GetPid(), GetTid(), aPtr, aSize,
+           new_ptr);
+  return new_ptr;
+}
+
+static void replace_free(void* aPtr) {
+  MutexAutoLock lock(sMutex);
+  FdPrintf(sFd, "%zu %zu free(%p)\n", GetPid(), GetTid(), aPtr);
+  sFuncs.free(aPtr);
+}
+
+static void* replace_memalign(size_t aAlignment, size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.memalign(aAlignment, aSize);
+  FdPrintf(sFd, "%zu %zu memalign(%zu,%zu)=%p\n", GetPid(), GetTid(),
+           aAlignment, aSize, ptr);
+  return ptr;
+}
+
+static void* replace_valloc(size_t aSize) {
+  MutexAutoLock lock(sMutex);
+  void* ptr = sFuncs.valloc(aSize);
+  FdPrintf(sFd, "%zu %zu valloc(%zu)=%p\n", GetPid(), GetTid(), aSize, ptr);
+  return ptr;
+}
+
+static void replace_jemalloc_stats(jemalloc_stats_t* aStats,
+                                   jemalloc_bin_stats_t* aBinStats) {
+  MutexAutoLock lock(sMutex);
+  sFuncs.jemalloc_stats_internal(aStats, aBinStats);
+  FdPrintf(sFd, "%zu %zu jemalloc_stats()\n", GetPid(), GetTid());
+}
+
+void replace_init(malloc_table_t* aTable, ReplaceMallocBridge** aBridge) {
+  /* Initialize output file descriptor from the MALLOC_LOG environment
+   * variable. Numbers up to 9999 are considered as a preopened file
+   * descriptor number. Other values are considered as a file name. */
+#ifdef _WIN32
+  wchar_t* log = _wgetenv(L"MALLOC_LOG");
+#else
+  char* log = getenv("MALLOC_LOG");
+#endif
+  if (log && *log) {
+    int fd = 0;
+    const auto* fd_num = log;
+    while (*fd_num) {
+      /* Reject non digits. */
+      if (*fd_num < '0' || *fd_num > '9') {
+        fd = -1;
+        break;
+      }
+      fd = fd * 10 + (*fd_num - '0');
+      /* Reject values >= 10000. */
+      if (fd >= 10000) {
+        fd = -1;
+        break;
+      }
+      fd_num++;
+    }
+    if (fd == 1 || fd == 2) {
+      sStdoutOrStderr = true;
+    }
+#ifdef _WIN32
+    // See comment in FdPrintf.h as to why CreateFile is used.
+    HANDLE handle;
+    if (fd > 0) {
+      handle = reinterpret_cast<HANDLE>(_get_osfhandle(fd));
+    } else {
+      handle =
+          CreateFileW(log, FILE_APPEND_DATA, FILE_SHARE_READ | FILE_SHARE_WRITE,
+                      nullptr, OPEN_ALWAYS, FILE_ATTRIBUTE_NORMAL, nullptr);
+    }
+    if (handle != INVALID_HANDLE_VALUE) {
+      sFd = reinterpret_cast<intptr_t>(handle);
+    }
+#else
+    if (fd == -1) {
+      fd = open(log, O_WRONLY | O_CREAT | O_APPEND, 0644);
+    }
+    if (fd > 0) {
+      sFd = fd;
+    }
+#endif
+  }
+
+  // Don't initialize if we weren't passed a valid MALLOC_LOG.
+  if (sFd == 0) {
+    return;
+  }
+
+  sMutex.Init();
+  static LogAllocBridge bridge;
+  sFuncs = *aTable;
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC_BASE
+#define MALLOC_DECL(name, ...) aTable->name = replace_##name;
+#include "malloc_decls.h"
+  aTable->jemalloc_stats_internal = replace_jemalloc_stats;
+  if (!getenv("MALLOC_LOG_MINIMAL")) {
+    aTable->posix_memalign = replace_posix_memalign;
+    aTable->aligned_alloc = replace_aligned_alloc;
+    aTable->valloc = replace_valloc;
+  }
+  *aBridge = &bridge;
+
+#ifndef _WIN32
+  /* When another thread has acquired a lock before forking, the child
+   * process will inherit the lock state but the thread, being nonexistent
+   * in the child process, will never release it, leading to a dead-lock
+   * whenever the child process gets the lock. We thus need to ensure no
+   * other thread is holding the lock before forking, by acquiring it
+   * ourselves, and releasing it after forking in the parent process and
+   * resetting it to its initial state in the child process. The latter is
+   * important because some implementations (notably macOS) prevent a lock from
+   * being unlocked by a different thread than the one which locked it in the
+   * first place.
+   * Windows doesn't have this problem since there is no fork().
+   * The real allocator, however, might be doing the same thing (jemalloc
+   * does). But pthread_atfork `prepare` handlers (first argument) are
+   * processed in reverse order they were established. But replace_init
+   * runs before the real allocator has had any chance to initialize and
+   * call pthread_atfork itself. This leads to its prefork running before
+   * ours. This leads to a race condition that can lead to a deadlock like
+   * the following:
+   *   - thread A forks.
+   *   - libc calls real allocator's prefork, so thread A holds the real
+   *     allocator lock.
+   *   - thread B calls malloc, which calls our replace_malloc.
+   *   - consequently, thread B holds our lock.
+   *   - thread B then proceeds to call the real allocator's malloc, and
+   *     waits for the real allocator's lock, which thread A holds.
+   *   - libc calls our prefork, so thread A waits for our lock, which
+   *     thread B holds.
+   * To avoid this race condition, the real allocator's prefork must be
+   * called after ours, which means it needs to be registered before ours.
+   * So trick the real allocator into initializing itself without more side
+   * effects by calling malloc with a size it can't possibly allocate. */
+  sFuncs.malloc(-1);
+  pthread_atfork(prefork, postfork_parent, postfork_child);
+#endif
+}
diff --git a/memory/replace/logalloc/README b/memory/replace/logalloc/README
new file mode 100644
index 0000000000..c2e8cf66ce
--- /dev/null
+++ b/memory/replace/logalloc/README
@@ -0,0 +1,95 @@
+Logalloc is a replace-malloc library for Firefox (see
+memory/build/replace_malloc.h) that dumps a log of memory allocations to a
+given file descriptor or file name. That log can then be replayed against
+Firefox's default memory allocator independently or through another
+replace-malloc library, allowing the testing of other allocators under the
+exact same workload.
+
+To get an allocation log the following environment variable when starting
+Firefox:
+  MALLOC_LOG=/path/to/log-file
+  or
+  MALLOC_LOG=number
+
+When MALLOC_LOG is a number below 10000, it is considered as a file
+descriptor number that is fed to Firefox when it is started. Otherwise,
+it is considered as a file name.
+
+As those allocation logs can grow large quite quickly, it can be useful
+to pipe the output to a compression tool.
+
+MALLOC_LOG=1 would send to Firefox's stdout, MALLOC_LOG=2 would send to
+its stderr. Since in both cases that could be mixed with other output
+from Firefox, it is usually better to use another file descriptor
+by shell redirections, such as:
+
+  MALLOC_LOG=3 firefox 3>&1 1>&2 | gzip -c > log.gz
+
+(3>&1 copies the `| gzip` pipe file descriptor to file descriptor #3, 1>&2
+then copies stderr to stdout. This leads to: fd1 and fd2 sending to stderr
+of the parent process (the shell), and fd3 sending to gzip.)
+
+Each line of the allocations log is formatted as follows:
+  <pid> <tid> <function>([<args>])[=<result>]
+where <args> is a comma separated list of values. The number of <args> and
+the presence of <result> depend on the <function>.
+
+Example log:
+  18545 18545 malloc(32)=0x7f90495120e0
+  18545 18545 calloc(1,148)=0x7f9049537480
+  18545 18545 realloc(0x7f90495120e0,64)=0x7f9049536680
+  18545 18545 posix_memalign(256,240)=0x7f9049583300
+  18545 18545 jemalloc_stats()
+  18545 18545 free(0x7f9049536680)
+
+This log can be replayed with the logalloc-replay tool in
+memory/replace/logalloc/replay. However, as the goal of that tool is to
+reproduce the recorded memory allocations, it needs to avoid as much as
+possible doing its own allocations for bookkeeping. Reading the logs as
+they are would require data structures and memory allocations. As a
+consequence, the logs need to be preprocessed beforehand.
+
+The logalloc_munge.py script is responsible for that preprocessing. It simply
+takes a raw log on its stdin, and outputs the preprocessed log on its stdout.
+It replaces pointer addresses with indexes the logalloc-replay tool can use
+in a large (almost) linear array of allocation tracking slots (prefixed with
+'#'). It also replaces the pids with numbers starting from 1 (such as the
+first seen pid number is 1, the second is 2, etc.).
+
+The above example log would become the following, once preprocessed:
+  1 1 malloc(32)=#1
+  1 1 calloc(1,148)=#2
+  1 1 realloc(#1,64)=#1
+  1 1 posix_memalign(256,240)=#3
+  1 1 jemalloc_stats()
+  1 1 free(#1)
+
+The logalloc-replay tool then takes the preprocessed log on its stdin and
+replays the allocations printed there, but will only replay those with the
+same process id as the first line (which normally is 1).
+
+As the log files are simple text files, though, it is easy to separate out
+the different processes log with e.g. grep, and feed the separate processes
+logs to logalloc-replay.
+
+The logalloc-replay program won't output anything unless jemalloc_stats
+records appears in the log. You can expect those to be recorded when going
+to about:memory in Firefox, but they can also be added after preprocessing.
+
+Here is an example of what one can do:
+
+  gunzip -c log.gz | python logalloc_munge.py | \
+  awk '$1 == "2" { print $0 } !(NR % 10000) { print "2 1 jemalloc_stats()" }' | \
+    ./logalloc-replay
+
+The above command replays the allocations of process #2, with some stats
+output every 10000 records.
+
+The logalloc-replay tool itself being hooked with replace-malloc, it is possible
+to set LD_PRELOAD/DYLD_INSERT_LIBRARIES/MOZ_REPLACE_MALLOC_LIB and replay a log
+through a different allocator. For example:
+
+  LD_PRELOAD=libreplace_jemalloc.so logalloc-replay < log
+
+Will replay the log against jemalloc4 (which is, as of writing, what
+libreplace_jemalloc.so contains).
diff --git a/memory/replace/logalloc/moz.build b/memory/replace/logalloc/moz.build
new file mode 100644
index 0000000000..c52d9e69e0
--- /dev/null
+++ b/memory/replace/logalloc/moz.build
@@ -0,0 +1,30 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+ReplaceMalloc("logalloc")
+
+SOURCES += [
+    "FdPrintf.cpp",
+    "LogAlloc.cpp",
+]
+
+DisableStlWrapping()
+NO_PGO = True
+DEFINES["MOZ_NO_MOZALLOC"] = True
+
+LOCAL_INCLUDES += [
+    "/memory/build",
+]
+
+# Android doesn't have pthread_atfork, but we have our own in mozglue.
+if CONFIG["OS_TARGET"] == "Android" and FORCE_SHARED_LIB:
+    USE_LIBS += [
+        "mozglue",
+    ]
+
+DIRS += [
+    "replay",
+]
diff --git a/memory/replace/logalloc/replay/Makefile.in b/memory/replace/logalloc/replay/Makefile.in
new file mode 100644
index 0000000000..73659add98
--- /dev/null
+++ b/memory/replace/logalloc/replay/Makefile.in
@@ -0,0 +1,48 @@
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+ifdef MOZ_CODE_COVERAGE
+SKIP = 1
+endif
+
+ifdef CROSS_COMPILE
+SKIP = 1
+endif
+
+ifneq ($(SKIP),1)
+
+ifeq ($(OS_TARGET),WINNT)
+LOGALLOC_VAR = MOZ_REPLACE_MALLOC_LIB
+else
+ifeq ($(OS_TARGET),Darwin)
+LOGALLOC_VAR = DYLD_INSERT_LIBRARIES
+else
+LOGALLOC_VAR = LD_PRELOAD
+endif
+endif
+
+ifndef MOZ_REPLACE_MALLOC_STATIC
+LOGALLOC = $(LOGALLOC_VAR)=$(CURDIR)/../$(DLL_PREFIX)logalloc$(DLL_SUFFIX)
+endif
+
+expected_output.log: $(srcdir)/replay.log
+# The logalloc-replay program will only replay entries from the first pid,
+# so the expected output only contains entries beginning with "1 "
+	grep "^1 " $< > $@
+
+check:: $(srcdir)/replay.log expected_output.log $(srcdir)/expected_output_minimal.log
+# Test with MALLOC_LOG as a file descriptor number
+# We filter out anything happening before the first jemalloc_stats (first
+# command in replay.log) because starting with libstdc++ 5, a static
+# initializer in the STL allocates memory, which we obviously don't have
+# in expected_output.log.
+	MALLOC_LOG=1 $(LOGALLOC) ./$(PROGRAM) < $< | sed -n '/jemalloc_stats/,$$p' | $(PYTHON3) $(srcdir)/logalloc_munge.py | diff -w - expected_output.log
+# Test with MALLOC_LOG as a file name
+	$(RM) test_output.log
+	MALLOC_LOG=test_output.log $(LOGALLOC) ./$(PROGRAM) < $<
+	sed -n '/jemalloc_stats/,$$p' test_output.log | $(PYTHON3) $(srcdir)/logalloc_munge.py | diff -w - expected_output.log
+
+	MALLOC_LOG=1 MALLOC_LOG_MINIMAL=1 $(LOGALLOC) ./$(PROGRAM) < $< | sed -n '/jemalloc_stats/,$$p' | $(PYTHON3) $(srcdir)/logalloc_munge.py | diff -w - $(srcdir)/expected_output_minimal.log
+
+endif
diff --git a/memory/replace/logalloc/replay/Replay.cpp b/memory/replace/logalloc/replay/Replay.cpp
new file mode 100644
index 0000000000..b5ad0c540e
--- /dev/null
+++ b/memory/replace/logalloc/replay/Replay.cpp
@@ -0,0 +1,1159 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#define MOZ_MEMORY_IMPL
+#include "mozmemory_wrap.h"
+
+#ifdef _WIN32
+#  include <windows.h>
+#  include <io.h>
+typedef intptr_t ssize_t;
+#else
+#  include <sys/mman.h>
+#  include <unistd.h>
+#endif
+#ifdef XP_LINUX
+#  include <fcntl.h>
+#  include <stdlib.h>
+#endif
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/MathAlgorithms.h"
+#include "mozilla/Maybe.h"
+#include "FdPrintf.h"
+
+using namespace mozilla;
+
+static void die(const char* message) {
+  /* Here, it doesn't matter that fprintf may allocate memory. */
+  fprintf(stderr, "%s\n", message);
+  exit(1);
+}
+
+#ifdef XP_LINUX
+static size_t sPageSize = []() { return sysconf(_SC_PAGESIZE); }();
+#endif
+
+/* We don't want to be using malloc() to allocate our internal tracking
+ * data, because that would change the parameters of what is being measured,
+ * so we want to use data types that directly use mmap/VirtualAlloc. */
+template <typename T, size_t Len>
+class MappedArray {
+ public:
+  MappedArray() : mPtr(nullptr) {
+#ifdef XP_LINUX
+    MOZ_RELEASE_ASSERT(!((sizeof(T) * Len) & (sPageSize - 1)),
+                       "MappedArray size must be a multiple of the page size");
+#endif
+  }
+
+  ~MappedArray() {
+    if (mPtr) {
+#ifdef _WIN32
+      VirtualFree(mPtr, sizeof(T) * Len, MEM_RELEASE);
+#elif defined(XP_LINUX)
+      munmap(reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(mPtr) -
+                                     sPageSize),
+             sizeof(T) * Len + sPageSize * 2);
+#else
+      munmap(mPtr, sizeof(T) * Len);
+#endif
+    }
+  }
+
+  T& operator[](size_t aIndex) const {
+    if (mPtr) {
+      return mPtr[aIndex];
+    }
+
+#ifdef _WIN32
+    mPtr = reinterpret_cast<T*>(VirtualAlloc(
+        nullptr, sizeof(T) * Len, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE));
+    if (mPtr == nullptr) {
+      die("VirtualAlloc error");
+    }
+#else
+    size_t data_size = sizeof(T) * Len;
+    size_t size = data_size;
+#  ifdef XP_LINUX
+    // See below
+    size += sPageSize * 2;
+#  endif
+    mPtr = reinterpret_cast<T*>(mmap(nullptr, size, PROT_READ | PROT_WRITE,
+                                     MAP_ANON | MAP_PRIVATE, -1, 0));
+    if (mPtr == MAP_FAILED) {
+      die("Mmap error");
+    }
+#  ifdef XP_LINUX
+    // On Linux we request a page on either side of the allocation and
+    // mprotect them.  This prevents mappings in /proc/self/smaps from being
+    // merged and allows us to parse this file to calculate the allocator's RSS.
+    MOZ_ASSERT(0 == mprotect(mPtr, sPageSize, 0));
+    MOZ_ASSERT(0 == mprotect(reinterpret_cast<void*>(
+                                 reinterpret_cast<uintptr_t>(mPtr) + data_size +
+                                 sPageSize),
+                             sPageSize, 0));
+    mPtr = reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(mPtr) + sPageSize);
+#  endif
+#endif
+    return mPtr[aIndex];
+  }
+
+  bool ownsMapping(uintptr_t addr) const { return addr == (uintptr_t)mPtr; }
+
+  bool allocated() const { return !!mPtr; }
+
+ private:
+  mutable T* mPtr;
+};
+
+/* Type for records of allocations. */
+struct MemSlot {
+  void* mPtr;
+
+  // mRequest is only valid if mPtr is non-null.  It doesn't need to be cleared
+  // when memory is freed or realloc()ed.
+  size_t mRequest;
+};
+
+/* An almost infinite list of slots.
+ * In essence, this is a linked list of arrays of groups of slots.
+ * Each group is 1MB. On 64-bits, one group allows to store 64k allocations.
+ * Each MemSlotList instance can store 1023 such groups, which means more
+ * than 67M allocations. In case more would be needed, we chain to another
+ * MemSlotList, and so on.
+ * Using 1023 groups makes the MemSlotList itself page sized on 32-bits
+ * and 2 pages-sized on 64-bits.
+ */
+class MemSlotList {
+  static constexpr size_t kGroups = 1024 - 1;
+  static constexpr size_t kGroupSize = (1024 * 1024) / sizeof(MemSlot);
+
+  MappedArray<MemSlot, kGroupSize> mSlots[kGroups];
+  MappedArray<MemSlotList, 1> mNext;
+
+ public:
+  MemSlot& operator[](size_t aIndex) const {
+    if (aIndex < kGroupSize * kGroups) {
+      return mSlots[aIndex / kGroupSize][aIndex % kGroupSize];
+    }
+    aIndex -= kGroupSize * kGroups;
+    return mNext[0][aIndex];
+  }
+
+  // Ask if any of the memory-mapped buffers use this range.
+  bool ownsMapping(uintptr_t aStart) const {
+    for (const auto& slot : mSlots) {
+      if (slot.allocated() && slot.ownsMapping(aStart)) {
+        return true;
+      }
+    }
+    return mNext.ownsMapping(aStart) ||
+           (mNext.allocated() && mNext[0].ownsMapping(aStart));
+  }
+};
+
+/* Helper class for memory buffers */
+class Buffer {
+ public:
+  Buffer() : mBuf(nullptr), mLength(0) {}
+
+  Buffer(const void* aBuf, size_t aLength)
+      : mBuf(reinterpret_cast<const char*>(aBuf)), mLength(aLength) {}
+
+  /* Constructor for string literals. */
+  template <size_t Size>
+  explicit Buffer(const char (&aStr)[Size]) : mBuf(aStr), mLength(Size - 1) {}
+
+  /* Returns a sub-buffer up-to but not including the given aNeedle character.
+   * The "parent" buffer itself is altered to begin after the aNeedle
+   * character.
+   * If the aNeedle character is not found, return the entire buffer, and empty
+   * the "parent" buffer. */
+  Buffer SplitChar(char aNeedle) {
+    char* buf = const_cast<char*>(mBuf);
+    char* c = reinterpret_cast<char*>(memchr(buf, aNeedle, mLength));
+    if (!c) {
+      return Split(mLength);
+    }
+
+    Buffer result = Split(c - buf);
+    // Remove the aNeedle character itself.
+    Split(1);
+    return result;
+  }
+
+  // Advance to the position after aNeedle.  This is like SplitChar but does not
+  // return the skipped portion.
+  void Skip(char aNeedle, unsigned nTimes = 1) {
+    for (unsigned i = 0; i < nTimes; i++) {
+      SplitChar(aNeedle);
+    }
+  }
+
+  void SkipWhitespace() {
+    while (mLength > 0) {
+      if (!IsSpace(mBuf[0])) {
+        break;
+      }
+      mBuf++;
+      mLength--;
+    }
+  }
+
+  static bool IsSpace(char c) {
+    switch (c) {
+      case ' ':
+      case '\t':
+      case '\n':
+      case '\v':
+      case '\f':
+      case '\r':
+        return true;
+    }
+    return false;
+  }
+
+  /* Returns a sub-buffer of at most aLength characters. The "parent" buffer is
+   * amputated of those aLength characters. If the "parent" buffer is smaller
+   * than aLength, then its length is used instead. */
+  Buffer Split(size_t aLength) {
+    Buffer result(mBuf, std::min(aLength, mLength));
+    mLength -= result.mLength;
+    mBuf += result.mLength;
+    return result;
+  }
+
+  /* Move the buffer (including its content) to the memory address of the aOther
+   * buffer. */
+  void Slide(Buffer aOther) {
+    memmove(const_cast<char*>(aOther.mBuf), mBuf, mLength);
+    mBuf = aOther.mBuf;
+  }
+
+  /* Returns whether the two involved buffers have the same content. */
+  bool operator==(Buffer aOther) {
+    return mLength == aOther.mLength &&
+           (mBuf == aOther.mBuf || !strncmp(mBuf, aOther.mBuf, mLength));
+  }
+
+  bool operator!=(Buffer aOther) { return !(*this == aOther); }
+
+  /* Returns true if the buffer is not empty. */
+  explicit operator bool() { return mLength; }
+
+  char operator[](size_t n) const { return mBuf[n]; }
+
+  /* Returns the memory location of the buffer. */
+  const char* get() { return mBuf; }
+
+  /* Returns the memory location of the end of the buffer (technically, the
+   * first byte after the buffer). */
+  const char* GetEnd() { return mBuf + mLength; }
+
+  /* Extend the buffer over the content of the other buffer, assuming it is
+   * adjacent. */
+  void Extend(Buffer aOther) {
+    MOZ_ASSERT(aOther.mBuf == GetEnd());
+    mLength += aOther.mLength;
+  }
+
+  size_t Length() const { return mLength; }
+
+ private:
+  const char* mBuf;
+  size_t mLength;
+};
+
+/* Helper class to read from a file descriptor line by line. */
+class FdReader {
+ public:
+  explicit FdReader(int aFd, bool aNeedClose = false)
+      : mFd(aFd),
+        mNeedClose(aNeedClose),
+        mData(&mRawBuf, 0),
+        mBuf(&mRawBuf, sizeof(mRawBuf)) {}
+
+  FdReader(FdReader&& aOther) noexcept
+      : mFd(aOther.mFd),
+        mNeedClose(aOther.mNeedClose),
+        mData(&mRawBuf, 0),
+        mBuf(&mRawBuf, sizeof(mRawBuf)) {
+    memcpy(mRawBuf, aOther.mRawBuf, sizeof(mRawBuf));
+    aOther.mFd = -1;
+    aOther.mNeedClose = false;
+    aOther.mData = Buffer();
+    aOther.mBuf = Buffer();
+  }
+
+  FdReader& operator=(const FdReader&) = delete;
+  FdReader(const FdReader&) = delete;
+
+  ~FdReader() {
+    if (mNeedClose) {
+      close(mFd);
+    }
+  }
+
+  /* Read a line from the file descriptor and returns it as a Buffer instance */
+  Buffer ReadLine() {
+    while (true) {
+      Buffer result = mData.SplitChar('\n');
+
+      /* There are essentially three different cases here:
+       * - '\n' was found "early". In this case, the end of the result buffer
+       *   is before the beginning of the mData buffer (since SplitChar
+       *   amputated it).
+       * - '\n' was found as the last character of mData. In this case, mData
+       *   is empty, but still points at the end of mBuf. result points to what
+       *   used to be in mData, without the last character.
+       * - '\n' was not found. In this case too, mData is empty and points at
+       *   the end of mBuf. But result points to the entire buffer that used to
+       *   be pointed by mData.
+       * Only in the latter case do both result and mData's end match, and it's
+       * the only case where we need to refill the buffer.
+       */
+      if (result.GetEnd() != mData.GetEnd()) {
+        return result;
+      }
+
+      /* Since SplitChar emptied mData, make it point to what it had before. */
+      mData = result;
+
+      /* And move it to the beginning of the read buffer. */
+      mData.Slide(mBuf);
+
+      FillBuffer();
+
+      if (!mData) {
+        return Buffer();
+      }
+    }
+  }
+
+ private:
+  /* Fill the read buffer. */
+  void FillBuffer() {
+    size_t size = mBuf.GetEnd() - mData.GetEnd();
+    Buffer remainder(mData.GetEnd(), size);
+
+    ssize_t len = 1;
+    while (remainder && len > 0) {
+      len = ::read(mFd, const_cast<char*>(remainder.get()), size);
+      if (len < 0) {
+        die("Read error");
+      }
+      size -= len;
+      mData.Extend(remainder.Split(len));
+    }
+  }
+
+  /* File descriptor to read from. */
+  int mFd;
+  bool mNeedClose;
+
+  /* Part of data that was read from the file descriptor but not returned with
+   * ReadLine yet. */
+  Buffer mData;
+  /* Buffer representation of mRawBuf */
+  Buffer mBuf;
+  /* read() buffer */
+  char mRawBuf[4096];
+};
+
+MOZ_BEGIN_EXTERN_C
+
+/* Function declarations for all the replace_malloc _impl functions.
+ * See memory/build/replace_malloc.c */
+#define MALLOC_DECL(name, return_type, ...) \
+  return_type name##_impl(__VA_ARGS__);
+#define MALLOC_FUNCS MALLOC_FUNCS_MALLOC
+#include "malloc_decls.h"
+
+#define MALLOC_DECL(name, return_type, ...) return_type name(__VA_ARGS__);
+#define MALLOC_FUNCS MALLOC_FUNCS_JEMALLOC
+#include "malloc_decls.h"
+
+#ifdef ANDROID
+
+/* mozjemalloc and jemalloc use pthread_atfork, which Android doesn't have.
+ * While gecko has one in libmozglue, the replay program can't use that.
+ * Since we're not going to fork anyways, make it a dummy function. */
+int pthread_atfork(void (*aPrepare)(void), void (*aParent)(void),
+                   void (*aChild)(void)) {
+  return 0;
+}
+#endif
+
+MOZ_END_EXTERN_C
+
+template <unsigned Base = 10>
+size_t parseNumber(Buffer aBuf) {
+  if (!aBuf) {
+    die("Malformed input");
+  }
+
+  size_t result = 0;
+  for (const char *c = aBuf.get(), *end = aBuf.GetEnd(); c < end; c++) {
+    result *= Base;
+    if ((*c >= '0' && *c <= '9')) {
+      result += *c - '0';
+    } else if (Base == 16 && *c >= 'a' && *c <= 'f') {
+      result += *c - 'a' + 10;
+    } else if (Base == 16 && *c >= 'A' && *c <= 'F') {
+      result += *c - 'A' + 10;
+    } else {
+      die("Malformed input");
+    }
+  }
+  return result;
+}
+
+static size_t percent(size_t a, size_t b) {
+  if (!b) {
+    return 0;
+  }
+  return size_t(round(double(a) / double(b) * 100.0));
+}
+
+class Distribution {
+ public:
+  // Default constructor used for array initialisation.
+  Distribution()
+      : mMaxSize(0),
+        mNextSmallest(0),
+        mShift(0),
+        mArrayOffset(0),
+        mArraySlots(0),
+        mTotalRequests(0),
+        mRequests{0} {}
+
+  Distribution(size_t max_size, size_t next_smallest, size_t bucket_size)
+      : mMaxSize(max_size),
+        mNextSmallest(next_smallest),
+        mShift(CeilingLog2(bucket_size)),
+        mArrayOffset(1 + next_smallest),
+        mArraySlots((max_size - next_smallest) >> mShift),
+        mTotalRequests(0),
+        mRequests{
+            0,
+        } {
+    MOZ_ASSERT(mMaxSize);
+    MOZ_RELEASE_ASSERT(mArraySlots <= MAX_NUM_BUCKETS);
+  }
+
+  Distribution& operator=(const Distribution& aOther) = default;
+
+  void addRequest(size_t request) {
+    MOZ_ASSERT(mMaxSize);
+
+    mRequests[(request - mArrayOffset) >> mShift]++;
+    mTotalRequests++;
+  }
+
+  void printDist(intptr_t std_err) {
+    MOZ_ASSERT(mMaxSize);
+
+    // The translation to turn a slot index into a memory request size.
+    const size_t array_offset_add = (1 << mShift) + mNextSmallest;
+
+    FdPrintf(std_err, "\n%zu-bin Distribution:\n", mMaxSize);
+    FdPrintf(std_err, "   request   :  count percent\n");
+    size_t range_start = mNextSmallest + 1;
+    for (size_t j = 0; j < mArraySlots; j++) {
+      size_t range_end = (j << mShift) + array_offset_add;
+      FdPrintf(std_err, "%5zu - %5zu: %6zu %6zu%%\n", range_start, range_end,
+               mRequests[j], percent(mRequests[j], mTotalRequests));
+      range_start = range_end + 1;
+    }
+  }
+
+  size_t maxSize() const { return mMaxSize; }
+
+ private:
+  static constexpr size_t MAX_NUM_BUCKETS = 16;
+
+  // If size is zero this distribution is uninitialised.
+  size_t mMaxSize;
+  size_t mNextSmallest;
+
+  // Parameters to convert a size into a slot number.
+  unsigned mShift;
+  unsigned mArrayOffset;
+
+  // The number of slots.
+  unsigned mArraySlots;
+
+  size_t mTotalRequests;
+  size_t mRequests[MAX_NUM_BUCKETS];
+};
+
+#ifdef XP_LINUX
+struct MemoryMap {
+  uintptr_t mStart;
+  uintptr_t mEnd;
+  bool mReadable;
+  bool mPrivate;
+  bool mAnon;
+  bool mIsStack;
+  bool mIsSpecial;
+  size_t mRSS;
+
+  bool IsCandidate() const {
+    // Candidates mappings are:
+    //  * anonymous
+    //  * they are private (not shared),
+    //  * anonymous or "[heap]" (not another area such as stack),
+    //
+    // The only mappings we're falsely including are the .bss segments for
+    // shared libraries.
+    return mReadable && mPrivate && mAnon && !mIsStack && !mIsSpecial;
+  }
+};
+
+class SMapsReader : private FdReader {
+ private:
+  explicit SMapsReader(FdReader&& reader) : FdReader(std::move(reader)) {}
+
+ public:
+  static Maybe<SMapsReader> open() {
+    int fd = ::open(FILENAME, O_RDONLY);
+    if (fd < 0) {
+      perror(FILENAME);
+      return mozilla::Nothing();
+    }
+
+    return Some(SMapsReader(FdReader(fd, true)));
+  }
+
+  Maybe<MemoryMap> readMap(intptr_t aStdErr) {
+    // This is not very tolerant of format changes because things like
+    // parseNumber will crash if they get a bad value.  TODO: make this
+    // soft-fail.
+
+    Buffer line = ReadLine();
+    if (!line) {
+      return Nothing();
+    }
+
+    // We're going to be at the start of an entry, start tokenising the first
+    // line.
+
+    // Range
+    Buffer range = line.SplitChar(' ');
+    uintptr_t range_start = parseNumber<16>(range.SplitChar('-'));
+    uintptr_t range_end = parseNumber<16>(range);
+
+    // Mode.
+    Buffer mode = line.SplitChar(' ');
+    if (mode.Length() != 4) {
+      FdPrintf(aStdErr, "Couldn't parse SMAPS file\n");
+      return Nothing();
+    }
+    bool readable = mode[0] == 'r';
+    bool private_ = mode[3] == 'p';
+
+    // Offset, device and inode.
+    line.SkipWhitespace();
+    bool zero_offset = !parseNumber<16>(line.SplitChar(' '));
+    line.SkipWhitespace();
+    bool no_device = line.SplitChar(' ') == Buffer("00:00");
+    line.SkipWhitespace();
+    bool zero_inode = !parseNumber(line.SplitChar(' '));
+    bool is_anon = zero_offset && no_device && zero_inode;
+
+    // Filename, or empty for anon mappings.
+    line.SkipWhitespace();
+    Buffer filename = line.SplitChar(' ');
+
+    bool is_stack;
+    bool is_special;
+    if (filename && filename[0] == '[') {
+      is_stack = filename == Buffer("[stack]");
+      is_special = filename == Buffer("[vdso]") ||
+                   filename == Buffer("[vvar]") ||
+                   filename == Buffer("[vsyscall]");
+    } else {
+      is_stack = false;
+      is_special = false;
+    }
+
+    size_t rss = 0;
+    while ((line = ReadLine())) {
+      Buffer field = line.SplitChar(':');
+      if (field == Buffer("VmFlags")) {
+        // This is the last field, at least in the current format. Break this
+        // loop to read the next mapping.
+        break;
+      }
+
+      if (field == Buffer("Rss")) {
+        line.SkipWhitespace();
+        Buffer value = line.SplitChar(' ');
+        rss = parseNumber(value) * 1024;
+      }
+    }
+
+    return Some(MemoryMap({range_start, range_end, readable, private_, is_anon,
+                           is_stack, is_special, rss}));
+  }
+
+  static constexpr char FILENAME[] = "/proc/self/smaps";
+};
+#endif  // XP_LINUX
+
+/* Class to handle dispatching the replay function calls to replace-malloc. */
+class Replay {
+ public:
+  Replay() {
+#ifdef _WIN32
+    // See comment in FdPrintf.h as to why native win32 handles are used.
+    mStdErr = reinterpret_cast<intptr_t>(GetStdHandle(STD_ERROR_HANDLE));
+#else
+    mStdErr = fileno(stderr);
+#endif
+#ifdef XP_LINUX
+    BuildInitialMapInfo();
+#endif
+  }
+
+  void enableSlopCalculation() { mCalculateSlop = true; }
+  void enableMemset() { mDoMemset = true; }
+
+  MemSlot& operator[](size_t index) const { return mSlots[index]; }
+
+  void malloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::malloc_impl(size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      MaybeCommit(aSlot);
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void posix_memalign(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t alignment = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    void* ptr;
+    if (::posix_memalign_impl(&ptr, alignment, size) == 0) {
+      aSlot.mPtr = ptr;
+      aSlot.mRequest = size;
+      MaybeCommit(aSlot);
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    } else {
+      aSlot.mPtr = nullptr;
+    }
+  }
+
+  void aligned_alloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t alignment = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::aligned_alloc_impl(alignment, size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      MaybeCommit(aSlot);
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void calloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t num = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::calloc_impl(num, size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = num * size;
+      MaybeCommit(aSlot);
+      if (mCalculateSlop) {
+        mTotalRequestedSize += num * size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void realloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    Buffer dummy = aArgs.SplitChar('#');
+    if (dummy) {
+      die("Malformed input");
+    }
+    size_t slot_id = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    MemSlot& old_slot = (*this)[slot_id];
+    void* old_ptr = old_slot.mPtr;
+    old_slot.mPtr = nullptr;
+    aSlot.mPtr = ::realloc_impl(old_ptr, size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      MaybeCommit(aSlot);
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void free(Buffer& aArgs, Buffer& aResult) {
+    if (aResult) {
+      die("Malformed input");
+    }
+    mOps++;
+    Buffer dummy = aArgs.SplitChar('#');
+    if (dummy) {
+      die("Malformed input");
+    }
+    size_t slot_id = parseNumber(aArgs);
+    MemSlot& slot = (*this)[slot_id];
+    ::free_impl(slot.mPtr);
+    slot.mPtr = nullptr;
+  }
+
+  void memalign(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t alignment = parseNumber(aArgs.SplitChar(','));
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::memalign_impl(alignment, size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      MaybeCommit(aSlot);
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void valloc(Buffer& aArgs, Buffer& aResult) {
+    MemSlot& aSlot = SlotForResult(aResult);
+    mOps++;
+    size_t size = parseNumber(aArgs);
+    aSlot.mPtr = ::valloc_impl(size);
+    if (aSlot.mPtr) {
+      aSlot.mRequest = size;
+      MaybeCommit(aSlot);
+      if (mCalculateSlop) {
+        mTotalRequestedSize += size;
+        mTotalAllocatedSize += ::malloc_usable_size_impl(aSlot.mPtr);
+      }
+    }
+  }
+
+  void jemalloc_stats(Buffer& aArgs, Buffer& aResult) {
+    if (aArgs || aResult) {
+      die("Malformed input");
+    }
+    mOps++;
+    jemalloc_stats_t stats;
+    // Using a variable length array here is a GCC & Clang extension. But it
+    // allows us to place this on the stack and not alter jemalloc's profiling.
+    const size_t num_bins = ::jemalloc_stats_num_bins();
+    const size_t MAX_NUM_BINS = 100;
+    if (num_bins > MAX_NUM_BINS) {
+      die("Exceeded maximum number of jemalloc stats bins");
+    }
+    jemalloc_bin_stats_t bin_stats[MAX_NUM_BINS] = {{0}};
+    ::jemalloc_stats_internal(&stats, bin_stats);
+
+#ifdef XP_LINUX
+    size_t rss = get_rss();
+#endif
+
+    size_t num_objects = 0;
+    size_t num_sloppy_objects = 0;
+    size_t total_allocated = 0;
+    size_t total_slop = 0;
+    size_t large_slop = 0;
+    size_t large_used = 0;
+    size_t huge_slop = 0;
+    size_t huge_used = 0;
+    size_t bin_slop[MAX_NUM_BINS] = {0};
+
+    for (size_t slot_id = 0; slot_id < mNumUsedSlots; slot_id++) {
+      MemSlot& slot = mSlots[slot_id];
+      if (slot.mPtr) {
+        size_t used = ::malloc_usable_size_impl(slot.mPtr);
+        size_t slop = used - slot.mRequest;
+        total_allocated += used;
+        total_slop += slop;
+        num_objects++;
+        if (slop) {
+          num_sloppy_objects++;
+        }
+
+        if (used <=
+            (stats.subpage_max ? stats.subpage_max : stats.quantum_wide_max)) {
+          // We know that this is an inefficient linear search, but there's a
+          // small number of bins and this is simple.
+          for (unsigned i = 0; i < num_bins; i++) {
+            auto& bin = bin_stats[i];
+            if (used == bin.size) {
+              bin_slop[i] += slop;
+              break;
+            }
+          }
+        } else if (used <= stats.large_max) {
+          large_slop += slop;
+          large_used += used;
+        } else {
+          huge_slop += slop;
+          huge_used += used;
+        }
+      }
+    }
+
+    // This formula corresponds to the calculation of wasted (from committed and
+    // the other parameters) within jemalloc_stats()
+    size_t committed = stats.allocated + stats.waste + stats.page_cache +
+                       stats.bookkeeping + stats.bin_unused;
+
+    FdPrintf(mStdErr, "\n");
+    FdPrintf(mStdErr, "Objects:          %9zu\n", num_objects);
+    FdPrintf(mStdErr, "Slots:            %9zu\n", mNumUsedSlots);
+    FdPrintf(mStdErr, "Ops:              %9zu\n", mOps);
+    FdPrintf(mStdErr, "mapped:           %9zu\n", stats.mapped);
+    FdPrintf(mStdErr, "committed:        %9zu\n", committed);
+#ifdef XP_LINUX
+    if (rss) {
+      FdPrintf(mStdErr, "rss:              %9zu\n", rss);
+    }
+#endif
+    FdPrintf(mStdErr, "allocated:        %9zu\n", stats.allocated);
+    FdPrintf(mStdErr, "waste:            %9zu\n", stats.waste);
+    FdPrintf(mStdErr, "dirty:            %9zu\n", stats.page_cache);
+    FdPrintf(mStdErr, "bookkeep:         %9zu\n", stats.bookkeeping);
+    FdPrintf(mStdErr, "bin-unused:       %9zu\n", stats.bin_unused);
+    FdPrintf(mStdErr, "quantum-max:      %9zu\n", stats.quantum_max);
+    FdPrintf(mStdErr, "quantum-wide-max: %9zu\n", stats.quantum_wide_max);
+    FdPrintf(mStdErr, "subpage-max:      %9zu\n", stats.subpage_max);
+    FdPrintf(mStdErr, "large-max:        %9zu\n", stats.large_max);
+    if (mCalculateSlop) {
+      size_t slop = mTotalAllocatedSize - mTotalRequestedSize;
+      FdPrintf(mStdErr,
+               "Total slop for all allocations: %zuKiB/%zuKiB (%zu%%)\n",
+               slop / 1024, mTotalAllocatedSize / 1024,
+               percent(slop, mTotalAllocatedSize));
+    }
+    FdPrintf(mStdErr, "Live sloppy objects: %zu/%zu (%zu%%)\n",
+             num_sloppy_objects, num_objects,
+             percent(num_sloppy_objects, num_objects));
+    FdPrintf(mStdErr, "Live sloppy bytes: %zuKiB/%zuKiB (%zu%%)\n",
+             total_slop / 1024, total_allocated / 1024,
+             percent(total_slop, total_allocated));
+
+    FdPrintf(mStdErr, "\n%8s %11s %10s %8s %9s %9s %8s\n", "bin-size",
+             "unused (c)", "total (c)", "used (c)", "non-full (r)", "total (r)",
+             "used (r)");
+    for (unsigned i = 0; i < num_bins; i++) {
+      auto& bin = bin_stats[i];
+      MOZ_ASSERT(bin.size);
+      FdPrintf(mStdErr, "%8zu %8zuKiB %7zuKiB %7zu%% %12zu %9zu %7zu%%\n",
+               bin.size, bin.bytes_unused / 1024, bin.bytes_total / 1024,
+               percent(bin.bytes_total - bin.bytes_unused, bin.bytes_total),
+               bin.num_non_full_runs, bin.num_runs,
+               percent(bin.num_runs - bin.num_non_full_runs, bin.num_runs));
+    }
+
+    FdPrintf(mStdErr, "\n%5s %8s %9s %7s\n", "bin", "slop", "used", "percent");
+    for (unsigned i = 0; i < num_bins; i++) {
+      auto& bin = bin_stats[i];
+      size_t used = bin.bytes_total - bin.bytes_unused;
+      FdPrintf(mStdErr, "%5zu %8zu %9zu %6zu%%\n", bin.size, bin_slop[i], used,
+               percent(bin_slop[i], used));
+    }
+    FdPrintf(mStdErr, "%5s %8zu %9zu %6zu%%\n", "large", large_slop, large_used,
+             percent(large_slop, large_used));
+    FdPrintf(mStdErr, "%5s %8zu %9zu %6zu%%\n", "huge", huge_slop, huge_used,
+             percent(huge_slop, huge_used));
+
+    print_distributions(stats, bin_stats);
+  }
+
+ private:
+  /*
+   * Create and print frequency distributions of memory requests.
+   */
+  void print_distributions(jemalloc_stats_t& stats,
+                           jemalloc_bin_stats_t* bin_stats) {
+    const size_t num_bins = ::jemalloc_stats_num_bins();
+
+    // We compute distributions for all of the bins for small allocations
+    // (num_bins) plus two more distributions for larger allocations.
+    Distribution dists[num_bins + 2];
+
+    unsigned last_size = 0;
+    unsigned num_dists = 0;
+    for (unsigned i = 0; i < num_bins; i++) {
+      auto& bin = bin_stats[i];
+      auto& dist = dists[num_dists++];
+
+      MOZ_ASSERT(bin.size);
+      if (bin.size <= 16) {
+        // 1 byte buckets.
+        dist = Distribution(bin.size, last_size, 1);
+      } else if (bin.size <= stats.quantum_max) {
+        // 4 buckets, (4 bytes per bucket with a 16 byte quantum).
+        dist = Distribution(bin.size, last_size, stats.quantum / 4);
+      } else if (bin.size <= stats.quantum_wide_max) {
+        // 8 buckets, (32 bytes per bucket with a 256 byte quantum-wide).
+        dist = Distribution(bin.size, last_size, stats.quantum_wide / 8);
+      } else {
+        // 16 buckets.
+        dist = Distribution(bin.size, last_size, (bin.size - last_size) / 16);
+      }
+      last_size = bin.size;
+    }
+
+    // 16 buckets.
+    dists[num_dists] = Distribution(stats.page_size, last_size,
+                                    (stats.page_size - last_size) / 16);
+    num_dists++;
+
+    // Buckets are 1/4 of the page size (12 buckets).
+    dists[num_dists] =
+        Distribution(stats.page_size * 4, stats.page_size, stats.page_size / 4);
+    num_dists++;
+
+    MOZ_RELEASE_ASSERT(num_dists <= num_bins + 2);
+
+    for (size_t slot_id = 0; slot_id < mNumUsedSlots; slot_id++) {
+      MemSlot& slot = mSlots[slot_id];
+      if (slot.mPtr) {
+        for (size_t i = 0; i < num_dists; i++) {
+          if (slot.mRequest <= dists[i].maxSize()) {
+            dists[i].addRequest(slot.mRequest);
+            break;
+          }
+        }
+      }
+    }
+
+    for (unsigned i = 0; i < num_dists; i++) {
+      dists[i].printDist(mStdErr);
+    }
+  }
+
+#ifdef XP_LINUX
+  size_t get_rss() {
+    if (mGetRSSFailed) {
+      return 0;
+    }
+
+    // On Linux we can determine the RSS of the heap area by examining the
+    // smaps file.
+    mozilla::Maybe<SMapsReader> reader = SMapsReader::open();
+    if (!reader) {
+      mGetRSSFailed = true;
+      return 0;
+    }
+
+    size_t rss = 0;
+    while (Maybe<MemoryMap> map = reader->readMap(mStdErr)) {
+      if (map->IsCandidate() && !mSlots.ownsMapping(map->mStart) &&
+          !InitialMapsContains(map->mStart)) {
+        rss += map->mRSS;
+      }
+    }
+
+    return rss;
+  }
+
+  bool InitialMapsContains(uintptr_t aRangeStart) {
+    for (unsigned i = 0; i < mNumInitialMaps; i++) {
+      MOZ_ASSERT(i < MAX_INITIAL_MAPS);
+
+      if (mInitialMaps[i] == aRangeStart) {
+        return true;
+      }
+    }
+    return false;
+  }
+
+ public:
+  void BuildInitialMapInfo() {
+    if (mGetRSSFailed) {
+      return;
+    }
+
+    Maybe<SMapsReader> reader = SMapsReader::open();
+    if (!reader) {
+      mGetRSSFailed = true;
+      return;
+    }
+
+    while (Maybe<MemoryMap> map = reader->readMap(mStdErr)) {
+      if (map->IsCandidate()) {
+        if (mNumInitialMaps >= MAX_INITIAL_MAPS) {
+          FdPrintf(mStdErr, "Too many initial mappings, can't compute RSS\n");
+          mGetRSSFailed = false;
+          return;
+        }
+
+        mInitialMaps[mNumInitialMaps++] = map->mStart;
+      }
+    }
+  }
+#endif
+
+ private:
+  MemSlot& SlotForResult(Buffer& aResult) {
+    /* Parse result value and get the corresponding slot. */
+    Buffer dummy = aResult.SplitChar('=');
+    Buffer dummy2 = aResult.SplitChar('#');
+    if (dummy || dummy2) {
+      die("Malformed input");
+    }
+
+    size_t slot_id = parseNumber(aResult);
+    mNumUsedSlots = std::max(mNumUsedSlots, slot_id + 1);
+
+    return mSlots[slot_id];
+  }
+
+  void MaybeCommit(MemSlot& aSlot) {
+    if (mDoMemset) {
+      // Write any byte, 0x55 isn't significant.
+      memset(aSlot.mPtr, 0x55, aSlot.mRequest);
+    }
+  }
+
+  intptr_t mStdErr;
+  size_t mOps = 0;
+
+  // The number of slots that have been used. It is used to iterate over slots
+  // without accessing those we haven't initialised.
+  size_t mNumUsedSlots = 0;
+
+  MemSlotList mSlots;
+  size_t mTotalRequestedSize = 0;
+  size_t mTotalAllocatedSize = 0;
+  // Whether to calculate slop for all allocations over the runtime of a
+  // process.
+  bool mCalculateSlop = false;
+  bool mDoMemset = false;
+
+#ifdef XP_LINUX
+  // If we have a failure reading smaps info then this is used to disable that
+  // feature.
+  bool mGetRSSFailed = false;
+
+  // The initial memory mappings are recorded here at start up.  We exclude
+  // memory in these mappings when computing RSS.  We assume they do not grow
+  // and that no regions are allocated near them, this is true because they'll
+  // only record the .bss and .data segments from our binary and shared objects
+  // or regions that logalloc-replay has created for MappedArrays.
+  //
+  // 64 should be enough for anybody.
+  static constexpr unsigned MAX_INITIAL_MAPS = 64;
+  uintptr_t mInitialMaps[MAX_INITIAL_MAPS];
+  unsigned mNumInitialMaps = 0;
+#endif  // XP_LINUX
+};
+
+static Replay replay;
+
+int main(int argc, const char* argv[]) {
+  size_t first_pid = 0;
+  FdReader reader(0);
+
+  for (int i = 1; i < argc; i++) {
+    const char* option = argv[i];
+    if (strcmp(option, "-s") == 0) {
+      // Do accounting to calculate allocation slop.
+      replay.enableSlopCalculation();
+    } else if (strcmp(option, "-c") == 0) {
+      // Touch memory as we allocate it.
+      replay.enableMemset();
+    } else {
+      fprintf(stderr, "Unknown command line option: %s\n", option);
+      return EXIT_FAILURE;
+    }
+  }
+
+  /* Read log from stdin and dispatch function calls to the Replay instance.
+   * The log format is essentially:
+   *   <pid> <tid> <function>([<args>])[=<result>]
+   * <args> is a comma separated list of arguments.
+   *
+   * The logs are expected to be preprocessed so that allocations are
+   * attributed a tracking slot. The input is trusted not to have crazy
+   * values for these slot numbers.
+   *
+   * <result>, as well as some of the args to some of the function calls are
+   * such slot numbers.
+   */
+  while (true) {
+    Buffer line = reader.ReadLine();
+
+    if (!line) {
+      break;
+    }
+
+    size_t pid = parseNumber(line.SplitChar(' '));
+    if (!first_pid) {
+      first_pid = pid;
+    }
+
+    /* The log may contain data for several processes, only entries for the
+     * very first that appears are treated. */
+    if (first_pid != pid) {
+      continue;
+    }
+
+    /* The log contains thread ids for manual analysis, but we just ignore them
+     * for now. */
+    parseNumber(line.SplitChar(' '));
+
+    Buffer func = line.SplitChar('(');
+    Buffer args = line.SplitChar(')');
+
+    if (func == Buffer("jemalloc_stats")) {
+      replay.jemalloc_stats(args, line);
+    } else if (func == Buffer("free")) {
+      replay.free(args, line);
+    } else if (func == Buffer("malloc")) {
+      replay.malloc(args, line);
+    } else if (func == Buffer("posix_memalign")) {
+      replay.posix_memalign(args, line);
+    } else if (func == Buffer("aligned_alloc")) {
+      replay.aligned_alloc(args, line);
+    } else if (func == Buffer("calloc")) {
+      replay.calloc(args, line);
+    } else if (func == Buffer("realloc")) {
+      replay.realloc(args, line);
+    } else if (func == Buffer("memalign")) {
+      replay.memalign(args, line);
+    } else if (func == Buffer("valloc")) {
+      replay.valloc(args, line);
+    } else {
+      die("Malformed input");
+    }
+  }
+
+  return 0;
+}
diff --git a/memory/replace/logalloc/replay/expected_output_minimal.log b/memory/replace/logalloc/replay/expected_output_minimal.log
new file mode 100644
index 0000000000..332fe20957
--- /dev/null
+++ b/memory/replace/logalloc/replay/expected_output_minimal.log
@@ -0,0 +1,17 @@
+1 1 jemalloc_stats()
+1 1 malloc(42)=#1
+1 1 malloc(24)=#2
+1 1 free(#1)
+1 1 memalign(4096,1024)=#1
+1 1 calloc(4,42)=#3
+1 1 free(#2)
+1 1 realloc(#3,84)=#2
+1 1 memalign(256,1024)=#3
+1 1 memalign(512,1024)=#4
+1 1 memalign(4096,1024)=#5
+1 1 jemalloc_stats()
+1 1 free(#5)
+1 1 free(#4)
+1 1 free(#3)
+1 1 free(#2)
+1 1 free(#1)
diff --git a/memory/replace/logalloc/replay/logalloc_munge.py b/memory/replace/logalloc/replay/logalloc_munge.py
new file mode 100644
index 0000000000..52d0032463
--- /dev/null
+++ b/memory/replace/logalloc/replay/logalloc_munge.py
@@ -0,0 +1,147 @@
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+"""
+This script takes a log from the replace-malloc logalloc library on stdin
+and munges it so that it can be used with the logalloc-replay tool.
+
+Given the following output:
+  13663 malloc(42)=0x7f0c33502040
+  13663 malloc(24)=0x7f0c33503040
+  13663 free(0x7f0c33502040)
+The resulting output is:
+  1 malloc(42)=#1
+  1 malloc(24)=#2
+  1 free(#1)
+
+See README for more details.
+"""
+
+import sys
+from collections import defaultdict, deque
+
+
+class IdMapping(object):
+    """Class to map values to ids.
+
+    Each value is associated to an increasing id, starting from 1.
+    When a value is removed, its id is recycled and will be reused for
+    subsequent values.
+    """
+
+    def __init__(self):
+        self.id = 1
+        self._values = {}
+        self._recycle = deque()
+
+    def __getitem__(self, value):
+        if value not in self._values:
+            if self._recycle:
+                self._values[value] = self._recycle.popleft()
+            else:
+                self._values[value] = self.id
+                self.id += 1
+        return self._values[value]
+
+    def __delitem__(self, value):
+        if value == 0:
+            return
+        self._recycle.append(self._values[value])
+        del self._values[value]
+
+    def __contains__(self, value):
+        return value == 0 or value in self._values
+
+
+class Ignored(Exception):
+    pass
+
+
+def split_log_line(line):
+    try:
+        # The format for each line is:
+        # <pid> [<tid>] <function>([<args>])[=<result>]
+        #
+        # The original format didn't include the tid, so we try to parse
+        # lines whether they have one or not.
+        pid, func_call = line.split(" ", 1)
+        call, result = func_call.split(")")
+        func, args = call.split("(")
+        args = args.split(",") if args else []
+        if result:
+            if result[0] != "=":
+                raise Ignored("Malformed input")
+            result = result[1:]
+        if " " in func:
+            tid, func = func.split(" ", 1)
+        else:
+            tid = pid
+        return pid, tid, func, args, result
+    except Exception:
+        raise Ignored("Malformed input")
+
+
+NUM_ARGUMENTS = {
+    "jemalloc_stats": 0,
+    "free": 1,
+    "malloc": 1,
+    "posix_memalign": 2,
+    "aligned_alloc": 2,
+    "calloc": 2,
+    "realloc": 2,
+    "memalign": 2,
+    "valloc": 1,
+}
+
+
+def main():
+    pids = IdMapping()
+    processes = defaultdict(lambda: {"pointers": IdMapping(), "tids": IdMapping()})
+    for line in sys.stdin:
+        line = line.strip()
+
+        try:
+            pid, tid, func, args, result = split_log_line(line)
+
+            # Replace pid with an id.
+            pid = pids[int(pid)]
+
+            process = processes[pid]
+            tid = process["tids"][int(tid)]
+
+            pointers = process["pointers"]
+
+            if func not in NUM_ARGUMENTS:
+                raise Ignored("Unknown function")
+
+            if len(args) != NUM_ARGUMENTS[func]:
+                raise Ignored("Malformed input")
+
+            if func in ("jemalloc_stats", "free") and result:
+                raise Ignored("Malformed input")
+
+            if func in ("free", "realloc"):
+                ptr = int(args[0], 16)
+                if ptr and ptr not in pointers:
+                    raise Ignored("Did not see an alloc for pointer")
+                args[0] = "#%d" % pointers[ptr]
+                del pointers[ptr]
+
+            if result:
+                result = int(result, 16)
+                if not result:
+                    raise Ignored("Result is NULL")
+                result = "#%d" % pointers[result]
+
+            print(
+                "%d %d %s(%s)%s"
+                % (pid, tid, func, ",".join(args), "=%s" % result if result else "")
+            )
+
+        except Exception as e:
+            print('Ignored "%s": %s' % (line, e), file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/memory/replace/logalloc/replay/moz.build b/memory/replace/logalloc/replay/moz.build
new file mode 100644
index 0000000000..1d39864699
--- /dev/null
+++ b/memory/replace/logalloc/replay/moz.build
@@ -0,0 +1,92 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+Program("logalloc-replay")
+
+SOURCES += [
+    "/mfbt/Assertions.cpp",
+    "/mfbt/Poison.cpp",
+    "/mfbt/RandomNum.cpp",
+    "/mfbt/TaggedAnonymousMemory.cpp",
+    "/mfbt/Unused.cpp",
+    "/mozglue/misc/StackWalk.cpp",
+    "Replay.cpp",
+]
+
+if CONFIG["OS_TARGET"] == "WINNT":
+    SOURCES += [
+        "/mozglue/misc/ProcessType.cpp",
+    ]
+
+if CONFIG["OS_TARGET"] == "Linux":
+    LDFLAGS += ["-static-libstdc++"]
+
+if CONFIG["OS_TARGET"] == "Darwin":
+    # Work around "warning: 'aligned_alloc' is only available on macOS 10.15 or newer"
+    # when building with MACOSX_DEPLOYMENT_TARGET < 10.15 with >= 10.15 SDK.
+    # We have our own definition of the function, so it doesn't matter what the SDK says.
+    SOURCES["Replay.cpp"].flags += ["-Wno-unguarded-availability-new"]
+
+if CONFIG["MOZ_REPLACE_MALLOC_STATIC"] and (CONFIG["MOZ_DMD"] or CONFIG["MOZ_PHC"]):
+    UNIFIED_SOURCES += [
+        "/mfbt/HashFunctions.cpp",
+        "/mfbt/JSONWriter.cpp",
+    ]
+
+if CONFIG["OS_ARCH"] == "WINNT":
+    OS_LIBS += [
+        "advapi32",
+        "dbghelp",
+    ]
+
+if CONFIG["MOZ_LINKER"] and CONFIG["MOZ_WIDGET_TOOLKIT"] == "android":
+    LOCAL_INCLUDES += [
+        "/mozglue/linker",
+    ]
+    DEFINES["__wrap_dladdr"] = "dladdr"
+
+
+if CONFIG["MOZ_BUILD_APP"] == "memory":
+    EXPORTS.mozilla += [
+        "/mozglue/misc/StackWalk.h",
+    ]
+
+if CONFIG["MOZ_BUILD_APP"] == "memory" or CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+    UNIFIED_SOURCES += [
+        "/mfbt/double-conversion/double-conversion/bignum-dtoa.cc",
+        "/mfbt/double-conversion/double-conversion/bignum.cc",
+        "/mfbt/double-conversion/double-conversion/cached-powers.cc",
+        "/mfbt/double-conversion/double-conversion/double-to-string.cc",
+        "/mfbt/double-conversion/double-conversion/fast-dtoa.cc",
+        "/mfbt/double-conversion/double-conversion/fixed-dtoa.cc",
+        "/mfbt/double-conversion/double-conversion/string-to-double.cc",
+        "/mfbt/double-conversion/double-conversion/strtod.cc",
+        "/mozglue/misc/Printf.cpp",
+    ]
+
+if not CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+    SOURCES += [
+        "../FdPrintf.cpp",
+    ]
+
+LOCAL_INCLUDES += [
+    "..",
+]
+
+# Link replace-malloc and the default allocator.
+USE_LIBS += [
+    "memory",
+]
+
+# The memory library defines this, so it's needed here too.
+DEFINES["IMPL_MFBT"] = True
+
+if CONFIG["MOZ_NEEDS_LIBATOMIC"]:
+    OS_LIBS += ["atomic"]
+
+DisableStlWrapping()
+
+include("/mozglue/build/replace_malloc.mozbuild")
diff --git a/memory/replace/logalloc/replay/replay.log b/memory/replace/logalloc/replay/replay.log
new file mode 100644
index 0000000000..f1e6de788b
--- /dev/null
+++ b/memory/replace/logalloc/replay/replay.log
@@ -0,0 +1,18 @@
+1 1 jemalloc_stats()
+1 1 malloc(42)=#1
+1 1 malloc(24)=#2
+2 2 malloc(42)=#1
+1 1 free(#1)
+1 1 posix_memalign(4096,1024)=#1
+1 1 calloc(4,42)=#3
+1 1 free(#2)
+1 1 realloc(#3,84)=#2
+1 1 aligned_alloc(256,1024)=#3
+1 1 memalign(512,1024)=#4
+1 1 valloc(1024)=#5
+1 1 jemalloc_stats()
+1 1 free(#5)
+1 1 free(#4)
+1 1 free(#3)
+1 1 free(#2)
+1 1 free(#1)
diff --git a/memory/replace/moz.build b/memory/replace/moz.build
new file mode 100644
index 0000000000..a76c5aac58
--- /dev/null
+++ b/memory/replace/moz.build
@@ -0,0 +1,25 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+@template
+def ReplaceMalloc(name):
+    if CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+        DEFINES["MOZ_REPLACE_MALLOC_PREFIX"] = name.replace("-", "_")
+        FINAL_LIBRARY = "memory"
+    else:
+        SharedLibrary(name)
+
+
+DIRS += [
+    "logalloc",
+]
+
+if CONFIG["MOZ_DMD"]:
+    DIRS += ["dmd"]
+
+if CONFIG["MOZ_PHC"]:
+    DIRS += ["phc"]
diff --git a/memory/replace/phc/PHC.cpp b/memory/replace/phc/PHC.cpp
new file mode 100644
index 0000000000..b28c36232b
--- /dev/null
+++ b/memory/replace/phc/PHC.cpp
@@ -0,0 +1,1609 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// PHC is a probabilistic heap checker. A tiny fraction of randomly chosen heap
+// allocations are subject to some expensive checking via the use of OS page
+// access protection. A failed check triggers a crash, whereupon useful
+// information about the failure is put into the crash report. The cost and
+// coverage for each user is minimal, but spread over the entire user base the
+// coverage becomes significant.
+//
+// The idea comes from Chromium, where it is called GWP-ASAN. (Firefox uses PHC
+// as the name because GWP-ASAN is long, awkward, and doesn't have any
+// particular meaning.)
+//
+// In the current implementation up to 64 allocations per process can become
+// PHC allocations. These allocations must be page-sized or smaller. Each PHC
+// allocation gets its own page, and when the allocation is freed its page is
+// marked inaccessible until the page is reused for another allocation. This
+// means that a use-after-free defect (which includes double-frees) will be
+// caught if the use occurs before the page is reused for another allocation.
+// The crash report will contain stack traces for the allocation site, the free
+// site, and the use-after-free site, which is often enough to diagnose the
+// defect.
+//
+// Also, each PHC allocation is followed by a guard page. The PHC allocation is
+// positioned so that its end abuts the guard page (or as close as possible,
+// given alignment constraints). This means that a bounds violation at the end
+// of the allocation (overflow) will be caught. The crash report will contain
+// stack traces for the allocation site and the bounds violation use site,
+// which is often enough to diagnose the defect.
+//
+// (A bounds violation at the start of the allocation (underflow) will not be
+// caught, unless it is sufficiently large to hit the preceding allocation's
+// guard page, which is not that likely. It would be possible to look more
+// assiduously for underflow by randomly placing some allocations at the end of
+// the page and some at the start of the page, and GWP-ASAN does this. PHC does
+// not, however, because overflow is likely to be much more common than
+// underflow in practice.)
+//
+// We use a simple heuristic to categorize a guard page access as overflow or
+// underflow: if the address falls in the lower half of the guard page, we
+// assume it is overflow, otherwise we assume it is underflow. More
+// sophisticated heuristics are possible, but this one is very simple, and it is
+// likely that most overflows/underflows in practice are very close to the page
+// boundary.
+//
+// The design space for the randomization strategy is large. The current
+// implementation has a large random delay before it starts operating, and a
+// small random delay between each PHC allocation attempt. Each freed PHC
+// allocation is quarantined for a medium random delay before being reused, in
+// order to increase the chance of catching UAFs.
+//
+// The basic cost of PHC's operation is as follows.
+//
+// - The physical memory cost is 64 pages plus some metadata (including stack
+//   traces) for each page. This amounts to 256 KiB per process on
+//   architectures with 4 KiB pages and 1024 KiB on macOS/AArch64 which uses
+//   16 KiB pages.
+//
+// - The virtual memory cost is the physical memory cost plus the guard pages:
+//   another 64 pages. This amounts to another 256 KiB per process on
+//   architectures with 4 KiB pages and 1024 KiB on macOS/AArch64 which uses
+//   16 KiB pages. PHC is currently only enabled on 64-bit platforms so the
+//   impact of the virtual memory usage is negligible.
+//
+// - Every allocation requires a size check and a decrement-and-check of an
+//   atomic counter. When the counter reaches zero a PHC allocation can occur,
+//   which involves marking a page as accessible and getting a stack trace for
+//   the allocation site. Otherwise, mozjemalloc performs the allocation.
+//
+// - Every deallocation requires a range check on the pointer to see if it
+//   involves a PHC allocation. (The choice to only do PHC allocations that are
+//   a page or smaller enables this range check, because the 64 pages are
+//   contiguous. Allowing larger allocations would make this more complicated,
+//   and we definitely don't want something as slow as a hash table lookup on
+//   every deallocation.) PHC deallocations involve marking a page as
+//   inaccessible and getting a stack trace for the deallocation site.
+//
+// Note that calls to realloc(), free(), and malloc_usable_size() will
+// immediately crash if the given pointer falls within a page allocation's
+// page, but does not point to the start of the allocation itself.
+//
+//   void* p = malloc(64);
+//   free(p + 1);     // p+1 doesn't point to the allocation start; crash
+//
+// Such crashes will not have the PHC fields in the crash report.
+//
+// PHC-specific tests can be run with the following commands:
+// - gtests: `./mach gtest '*PHC*'`
+// - xpcshell-tests: `./mach test toolkit/crashreporter/test/unit`
+//   - This runs some non-PHC tests as well.
+
+#include "PHC.h"
+
+#include <stdlib.h>
+#include <time.h>
+
+#include <algorithm>
+
+#ifdef XP_WIN
+#  include <process.h>
+#else
+#  include <sys/mman.h>
+#  include <sys/types.h>
+#  include <pthread.h>
+#  include <unistd.h>
+#endif
+
+#include "replace_malloc.h"
+#include "FdPrintf.h"
+#include "Mutex.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/Atomics.h"
+#include "mozilla/Attributes.h"
+#include "mozilla/CheckedInt.h"
+#include "mozilla/Maybe.h"
+#include "mozilla/StackWalk.h"
+#include "mozilla/ThreadLocal.h"
+#include "mozilla/XorShift128PlusRNG.h"
+
+using namespace mozilla;
+
+//---------------------------------------------------------------------------
+// Utilities
+//---------------------------------------------------------------------------
+
+#ifdef ANDROID
+// Android doesn't have pthread_atfork defined in pthread.h.
+extern "C" MOZ_EXPORT int pthread_atfork(void (*)(void), void (*)(void),
+                                         void (*)(void));
+#endif
+
+#ifndef DISALLOW_COPY_AND_ASSIGN
+#  define DISALLOW_COPY_AND_ASSIGN(T) \
+    T(const T&);                      \
+    void operator=(const T&)
+#endif
+
+static malloc_table_t sMallocTable;
+
+// This class provides infallible operations for the small number of heap
+// allocations that PHC does for itself. It would be nice if we could use the
+// InfallibleAllocPolicy from mozalloc, but PHC cannot use mozalloc.
+class InfallibleAllocPolicy {
+ public:
+  static void AbortOnFailure(const void* aP) {
+    if (!aP) {
+      MOZ_CRASH("PHC failed to allocate");
+    }
+  }
+
+  template <class T>
+  static T* new_() {
+    void* p = sMallocTable.malloc(sizeof(T));
+    AbortOnFailure(p);
+    return new (p) T;
+  }
+};
+
+//---------------------------------------------------------------------------
+// Stack traces
+//---------------------------------------------------------------------------
+
+// This code is similar to the equivalent code within DMD.
+
+class StackTrace : public phc::StackTrace {
+ public:
+  StackTrace() : phc::StackTrace() {}
+
+  void Clear() { mLength = 0; }
+
+  void Fill();
+
+ private:
+  static void StackWalkCallback(uint32_t aFrameNumber, void* aPc, void* aSp,
+                                void* aClosure) {
+    StackTrace* st = (StackTrace*)aClosure;
+    MOZ_ASSERT(st->mLength < kMaxFrames);
+    st->mPcs[st->mLength] = aPc;
+    st->mLength++;
+    MOZ_ASSERT(st->mLength == aFrameNumber);
+  }
+};
+
+// WARNING WARNING WARNING: this function must only be called when GMut::sMutex
+// is *not* locked, otherwise we might get deadlocks.
+//
+// How? On Windows, MozStackWalk() can lock a mutex, M, from the shared library
+// loader. Another thread might call malloc() while holding M locked (when
+// loading a shared library) and try to lock GMut::sMutex, causing a deadlock.
+// So GMut::sMutex can't be locked during the call to MozStackWalk(). (For
+// details, see https://bugzilla.mozilla.org/show_bug.cgi?id=374829#c8. On
+// Linux, something similar can happen; see bug 824340. So we just disallow it
+// on all platforms.)
+//
+// In DMD, to avoid this problem we temporarily unlock the equivalent mutex for
+// the MozStackWalk() call. But that's grotty, and things are a bit different
+// here, so we just require that stack traces be obtained before locking
+// GMut::sMutex.
+//
+// Unfortunately, there is no reliable way at compile-time or run-time to ensure
+// this pre-condition. Hence this large comment.
+//
+void StackTrace::Fill() {
+  mLength = 0;
+
+#if defined(XP_WIN) && defined(_M_IX86)
+  // This avoids MozStackWalk(), which causes unusably slow startup on Win32
+  // when it is called during static initialization (see bug 1241684).
+  //
+  // This code is cribbed from the Gecko Profiler, which also uses
+  // FramePointerStackWalk() on Win32: Registers::SyncPopulate() for the
+  // frame pointer, and GetStackTop() for the stack end.
+  CONTEXT context;
+  RtlCaptureContext(&context);
+  void** fp = reinterpret_cast<void**>(context.Ebp);
+
+  PNT_TIB pTib = reinterpret_cast<PNT_TIB>(NtCurrentTeb());
+  void* stackEnd = static_cast<void*>(pTib->StackBase);
+  FramePointerStackWalk(StackWalkCallback, kMaxFrames, this, fp, stackEnd);
+#elif defined(XP_MACOSX)
+  // This avoids MozStackWalk(), which has become unusably slow on Mac due to
+  // changes in libunwind.
+  //
+  // This code is cribbed from the Gecko Profiler, which also uses
+  // FramePointerStackWalk() on Mac: Registers::SyncPopulate() for the frame
+  // pointer, and GetStackTop() for the stack end.
+#  pragma GCC diagnostic push
+#  pragma GCC diagnostic ignored "-Wframe-address"
+  void** fp = reinterpret_cast<void**>(__builtin_frame_address(1));
+#  pragma GCC diagnostic pop
+  void* stackEnd = pthread_get_stackaddr_np(pthread_self());
+  FramePointerStackWalk(StackWalkCallback, kMaxFrames, this, fp, stackEnd);
+#else
+  MozStackWalk(StackWalkCallback, nullptr, kMaxFrames, this);
+#endif
+}
+
+//---------------------------------------------------------------------------
+// Logging
+//---------------------------------------------------------------------------
+
+// Change this to 1 to enable some PHC logging. Useful for debugging.
+#define PHC_LOGGING 0
+
+#if PHC_LOGGING
+
+static size_t GetPid() { return size_t(getpid()); }
+
+static size_t GetTid() {
+#  if defined(XP_WIN)
+  return size_t(GetCurrentThreadId());
+#  else
+  return size_t(pthread_self());
+#  endif
+}
+
+#  if defined(XP_WIN)
+#    define LOG_STDERR \
+      reinterpret_cast<intptr_t>(GetStdHandle(STD_ERROR_HANDLE))
+#  else
+#    define LOG_STDERR 2
+#  endif
+#  define LOG(fmt, ...)                                                \
+    FdPrintf(LOG_STDERR, "PHC[%zu,%zu,~%zu] " fmt, GetPid(), GetTid(), \
+             size_t(GAtomic::Now()), __VA_ARGS__)
+
+#else
+
+#  define LOG(fmt, ...)
+
+#endif  // PHC_LOGGING
+
+//---------------------------------------------------------------------------
+// Global state
+//---------------------------------------------------------------------------
+
+// Throughout this entire file time is measured as the number of sub-page
+// allocations performed (by PHC and mozjemalloc combined). `Time` is 64-bit
+// because we could have more than 2**32 allocations in a long-running session.
+// `Delay` is 32-bit because the delays used within PHC are always much smaller
+// than 2**32.
+using Time = uint64_t;   // A moment in time.
+using Delay = uint32_t;  // A time duration.
+
+// PHC only runs if the page size is 4 KiB; anything more is uncommon and would
+// use too much memory. So we hardwire this size for all platforms but macOS
+// on ARM processors. For the latter we make an exception because the minimum
+// page size supported is 16KiB so there's no way to go below that.
+static const size_t kPageSize =
+#if defined(XP_MACOSX) && defined(__aarch64__)
+    16384
+#else
+    4096
+#endif
+    ;
+
+// There are two kinds of page.
+// - Allocation pages, from which allocations are made.
+// - Guard pages, which are never touched by PHC.
+//
+// These page kinds are interleaved; each allocation page has a guard page on
+// either side.
+static const size_t kNumAllocPages = 64;
+static const size_t kNumAllPages = kNumAllocPages * 2 + 1;
+
+// The total size of the allocation pages and guard pages.
+static const size_t kAllPagesSize = kNumAllPages * kPageSize;
+
+// The junk value used to fill new allocation in debug builds. It's same value
+// as the one used by mozjemalloc. PHC applies it unconditionally in debug
+// builds. Unlike mozjemalloc, PHC doesn't consult the MALLOC_OPTIONS
+// environment variable to possibly change that behaviour.
+//
+// Also note that, unlike mozjemalloc, PHC doesn't have a poison value for freed
+// allocations because freed allocations are protected by OS page protection.
+#ifdef DEBUG
+const uint8_t kAllocJunk = 0xe4;
+#endif
+
+// The maximum time.
+static const Time kMaxTime = ~(Time(0));
+
+// The average delay before doing any page allocations at the start of a
+// process. Note that roughly 1 million allocations occur in the main process
+// while starting the browser. The delay range is 1..kAvgFirstAllocDelay*2.
+static const Delay kAvgFirstAllocDelay = 512 * 1024;
+
+// The average delay until the next attempted page allocation, once we get past
+// the first delay. The delay range is 1..kAvgAllocDelay*2.
+static const Delay kAvgAllocDelay = 16 * 1024;
+
+// The average delay before reusing a freed page. Should be significantly larger
+// than kAvgAllocDelay, otherwise there's not much point in having it. The delay
+// range is (kAvgAllocDelay / 2)..(kAvgAllocDelay / 2 * 3). This is different to
+// the other delay ranges in not having a minimum of 1, because that's such a
+// short delay that there is a high likelihood of bad stacks in any crash
+// report.
+static const Delay kAvgPageReuseDelay = 256 * 1024;
+
+// Truncate aRnd to the range (1 .. AvgDelay*2). If aRnd is random, this
+// results in an average value of aAvgDelay + 0.5, which is close enough to
+// aAvgDelay. aAvgDelay must be a power-of-two (otherwise it will crash) for
+// speed.
+template <Delay AvgDelay>
+constexpr Delay Rnd64ToDelay(uint64_t aRnd) {
+  static_assert(IsPowerOfTwo(AvgDelay), "must be a power of two");
+
+  return aRnd % (AvgDelay * 2) + 1;
+}
+
+// Maps a pointer to a PHC-specific structure:
+// - Nothing
+// - A guard page (it is unspecified which one)
+// - An allocation page (with an index < kNumAllocPages)
+//
+// The standard way of handling a PtrKind is to check IsNothing(), and if that
+// fails, to check IsGuardPage(), and if that fails, to call AllocPage().
+class PtrKind {
+ private:
+  enum class Tag : uint8_t {
+    Nothing,
+    GuardPage,
+    AllocPage,
+  };
+
+  Tag mTag;
+  uintptr_t mIndex;  // Only used if mTag == Tag::AllocPage.
+
+ public:
+  // Detect what a pointer points to. This constructor must be fast because it
+  // is called for every call to free(), realloc(), malloc_usable_size(), and
+  // jemalloc_ptr_info().
+  PtrKind(const void* aPtr, const uint8_t* aPagesStart,
+          const uint8_t* aPagesLimit) {
+    if (!(aPagesStart <= aPtr && aPtr < aPagesLimit)) {
+      mTag = Tag::Nothing;
+    } else {
+      uintptr_t offset = static_cast<const uint8_t*>(aPtr) - aPagesStart;
+      uintptr_t allPageIndex = offset / kPageSize;
+      MOZ_ASSERT(allPageIndex < kNumAllPages);
+      if (allPageIndex & 1) {
+        // Odd-indexed pages are allocation pages.
+        uintptr_t allocPageIndex = allPageIndex / 2;
+        MOZ_ASSERT(allocPageIndex < kNumAllocPages);
+        mTag = Tag::AllocPage;
+        mIndex = allocPageIndex;
+      } else {
+        // Even-numbered pages are guard pages.
+        mTag = Tag::GuardPage;
+      }
+    }
+  }
+
+  bool IsNothing() const { return mTag == Tag::Nothing; }
+  bool IsGuardPage() const { return mTag == Tag::GuardPage; }
+
+  // This should only be called after IsNothing() and IsGuardPage() have been
+  // checked and failed.
+  uintptr_t AllocPageIndex() const {
+    MOZ_RELEASE_ASSERT(mTag == Tag::AllocPage);
+    return mIndex;
+  }
+};
+
+// Shared, atomic, mutable global state.
+class GAtomic {
+ public:
+  static void Init(Delay aFirstDelay) {
+    sAllocDelay = aFirstDelay;
+
+    LOG("Initial sAllocDelay <- %zu\n", size_t(aFirstDelay));
+  }
+
+  static Time Now() { return sNow; }
+
+  static void IncrementNow() { sNow++; }
+
+  // Decrements the delay and returns the decremented value.
+  static int32_t DecrementDelay() { return --sAllocDelay; }
+
+  static void SetAllocDelay(Delay aAllocDelay) { sAllocDelay = aAllocDelay; }
+
+ private:
+  // The current time. Relaxed semantics because it's primarily used for
+  // determining if an allocation can be recycled yet and therefore it doesn't
+  // need to be exact.
+  static Atomic<Time, Relaxed> sNow;
+
+  // Delay until the next attempt at a page allocation. See the comment in
+  // MaybePageAlloc() for an explanation of why it is a signed integer, and why
+  // it uses ReleaseAcquire semantics.
+  static Atomic<Delay, ReleaseAcquire> sAllocDelay;
+};
+
+Atomic<Time, Relaxed> GAtomic::sNow;
+Atomic<Delay, ReleaseAcquire> GAtomic::sAllocDelay;
+
+// Shared, immutable global state. Initialized by replace_init() and never
+// changed after that. replace_init() runs early enough that no synchronization
+// is needed.
+class GConst {
+ private:
+  // The bounds of the allocated pages.
+  uint8_t* const mPagesStart;
+  uint8_t* const mPagesLimit;
+
+  // Allocates the allocation pages and the guard pages, contiguously.
+  uint8_t* AllocAllPages() {
+    // Allocate the pages so that they are inaccessible. They are never freed,
+    // because it would happen at process termination when it would be of little
+    // use.
+    void* pages =
+#ifdef XP_WIN
+        VirtualAlloc(nullptr, kAllPagesSize, MEM_RESERVE, PAGE_NOACCESS);
+#else
+        mmap(nullptr, kAllPagesSize, PROT_NONE, MAP_ANONYMOUS | MAP_PRIVATE, -1,
+             0);
+#endif
+    if (!pages) {
+      MOZ_CRASH();
+    }
+
+    return static_cast<uint8_t*>(pages);
+  }
+
+ public:
+  GConst()
+      : mPagesStart(AllocAllPages()), mPagesLimit(mPagesStart + kAllPagesSize) {
+    LOG("AllocAllPages at %p..%p\n", mPagesStart, mPagesLimit);
+  }
+
+  class PtrKind PtrKind(const void* aPtr) {
+    class PtrKind pk(aPtr, mPagesStart, mPagesLimit);
+    return pk;
+  }
+
+  bool IsInFirstGuardPage(const void* aPtr) {
+    return mPagesStart <= aPtr && aPtr < mPagesStart + kPageSize;
+  }
+
+  // Get the address of the allocation page referred to via an index. Used when
+  // marking the page as accessible/inaccessible.
+  uint8_t* AllocPagePtr(uintptr_t aIndex) {
+    MOZ_ASSERT(aIndex < kNumAllocPages);
+    // Multiply by two and add one to account for allocation pages *and* guard
+    // pages.
+    return mPagesStart + (2 * aIndex + 1) * kPageSize;
+  }
+};
+
+static GConst* gConst;
+
+// On MacOS, the first __thread/thread_local access calls malloc, which leads
+// to an infinite loop. So we use pthread-based TLS instead, which somehow
+// doesn't have this problem.
+#if !defined(XP_DARWIN)
+#  define PHC_THREAD_LOCAL(T) MOZ_THREAD_LOCAL(T)
+#else
+#  define PHC_THREAD_LOCAL(T) \
+    detail::ThreadLocal<T, detail::ThreadLocalKeyStorage>
+#endif
+
+// Thread-local state.
+class GTls {
+  GTls(const GTls&) = delete;
+
+  const GTls& operator=(const GTls&) = delete;
+
+  // When true, PHC does as little as possible.
+  //
+  // (a) It does not allocate any new page allocations.
+  //
+  // (b) It avoids doing any operations that might call malloc/free/etc., which
+  //     would cause re-entry into PHC. (In practice, MozStackWalk() is the
+  //     only such operation.) Note that calls to the functions in sMallocTable
+  //     are ok.
+  //
+  // For example, replace_malloc() will just fall back to mozjemalloc. However,
+  // operations involving existing allocations are more complex, because those
+  // existing allocations may be page allocations. For example, if
+  // replace_free() is passed a page allocation on a PHC-disabled thread, it
+  // will free the page allocation in the usual way, but it will get a dummy
+  // freeStack in order to avoid calling MozStackWalk(), as per (b) above.
+  //
+  // This single disabling mechanism has two distinct uses.
+  //
+  // - It's used to prevent re-entry into PHC, which can cause correctness
+  //   problems. For example, consider this sequence.
+  //
+  //   1. enter replace_free()
+  //   2. which calls PageFree()
+  //   3. which calls MozStackWalk()
+  //   4. which locks a mutex M, and then calls malloc
+  //   5. enter replace_malloc()
+  //   6. which calls MaybePageAlloc()
+  //   7. which calls MozStackWalk()
+  //   8. which (re)locks a mutex M --> deadlock
+  //
+  //   We avoid this sequence by "disabling" the thread in PageFree() (at step
+  //   2), which causes MaybePageAlloc() to fail, avoiding the call to
+  //   MozStackWalk() (at step 7).
+  //
+  //   In practice, realloc or free of a PHC allocation is unlikely on a thread
+  //   that is disabled because of this use: MozStackWalk() will probably only
+  //   realloc/free allocations that it allocated itself, but those won't be
+  //   page allocations because PHC is disabled before calling MozStackWalk().
+  //
+  //   (Note that MaybePageAlloc() could safely do a page allocation so long as
+  //   it avoided calling MozStackWalk() by getting a dummy allocStack. But it
+  //   wouldn't be useful, and it would prevent the second use below.)
+  //
+  // - It's used to prevent PHC allocations in some tests that rely on
+  //   mozjemalloc's exact allocation behaviour, which PHC does not replicate
+  //   exactly. (Note that (b) isn't necessary for this use -- MozStackWalk()
+  //   could be safely called -- but it is necessary for the first use above.)
+  //
+  static PHC_THREAD_LOCAL(bool) tlsIsDisabled;
+
+ public:
+  static void Init() {
+    if (!tlsIsDisabled.init()) {
+      MOZ_CRASH();
+    }
+  }
+
+  static void DisableOnCurrentThread() {
+    MOZ_ASSERT(!GTls::tlsIsDisabled.get());
+    tlsIsDisabled.set(true);
+  }
+
+  static void EnableOnCurrentThread() {
+    MOZ_ASSERT(GTls::tlsIsDisabled.get());
+    tlsIsDisabled.set(false);
+  }
+
+  static bool IsDisabledOnCurrentThread() { return tlsIsDisabled.get(); }
+};
+
+PHC_THREAD_LOCAL(bool) GTls::tlsIsDisabled;
+
+class AutoDisableOnCurrentThread {
+  AutoDisableOnCurrentThread(const AutoDisableOnCurrentThread&) = delete;
+
+  const AutoDisableOnCurrentThread& operator=(
+      const AutoDisableOnCurrentThread&) = delete;
+
+ public:
+  explicit AutoDisableOnCurrentThread() { GTls::DisableOnCurrentThread(); }
+  ~AutoDisableOnCurrentThread() { GTls::EnableOnCurrentThread(); }
+};
+
+// This type is used as a proof-of-lock token, to make it clear which functions
+// require sMutex to be locked.
+using GMutLock = const MutexAutoLock&;
+
+// Shared, mutable global state. Protected by sMutex; all accessing functions
+// take a GMutLock as proof that sMutex is held.
+class GMut {
+  enum class AllocPageState {
+    NeverAllocated = 0,
+    InUse = 1,
+    Freed = 2,
+  };
+
+  // Metadata for each allocation page.
+  class AllocPageInfo {
+   public:
+    AllocPageInfo()
+        : mState(AllocPageState::NeverAllocated),
+          mArenaId(),
+          mBaseAddr(nullptr),
+          mAllocStack(),
+          mFreeStack(),
+          mReuseTime(0) {}
+
+    // The current allocation page state.
+    AllocPageState mState;
+
+    // The arena that the allocation is nominally from. This isn't meaningful
+    // within PHC, which has no arenas. But it is necessary for reallocation of
+    // page allocations as normal allocations, such as in this code:
+    //
+    //   p = moz_arena_malloc(arenaId, 4096);
+    //   realloc(p, 8192);
+    //
+    // The realloc is more than one page, and thus too large for PHC to handle.
+    // Therefore, if PHC handles the first allocation, it must ask mozjemalloc
+    // to allocate the 8192 bytes in the correct arena, and to do that, it must
+    // call sMallocTable.moz_arena_malloc with the correct arenaId under the
+    // covers. Therefore it must record that arenaId.
+    //
+    // This field is also needed for jemalloc_ptr_info() to work, because it
+    // also returns the arena ID (but only in debug builds).
+    //
+    // - NeverAllocated: must be 0.
+    // - InUse | Freed: can be any valid arena ID value.
+    Maybe<arena_id_t> mArenaId;
+
+    // The starting address of the allocation. Will not be the same as the page
+    // address unless the allocation is a full page.
+    // - NeverAllocated: must be 0.
+    // - InUse | Freed: must be within the allocation page.
+    uint8_t* mBaseAddr;
+
+    // Usable size is computed as the number of bytes between the pointer and
+    // the end of the allocation page. This might be bigger than the requested
+    // size, especially if an outsized alignment is requested.
+    size_t UsableSize() const {
+      return mState == AllocPageState::NeverAllocated
+                 ? 0
+                 : kPageSize - (reinterpret_cast<uintptr_t>(mBaseAddr) &
+                                (kPageSize - 1));
+    }
+
+    // The allocation stack.
+    // - NeverAllocated: Nothing.
+    // - InUse | Freed: Some.
+    Maybe<StackTrace> mAllocStack;
+
+    // The free stack.
+    // - NeverAllocated | InUse: Nothing.
+    // - Freed: Some.
+    Maybe<StackTrace> mFreeStack;
+
+    // The time at which the page is available for reuse, as measured against
+    // GAtomic::sNow. When the page is in use this value will be kMaxTime.
+    // - NeverAllocated: must be 0.
+    // - InUse: must be kMaxTime.
+    // - Freed: must be > 0 and < kMaxTime.
+    Time mReuseTime;
+  };
+
+ public:
+  // The mutex that protects the other members.
+  static Mutex sMutex MOZ_UNANNOTATED;
+
+  GMut()
+      : mRNG(RandomSeed<0>(), RandomSeed<1>()),
+        mAllocPages(),
+        mNumPageAllocs(0),
+        mPageAllocHits(0),
+        mPageAllocMisses(0) {
+    sMutex.Init();
+  }
+
+  uint64_t Random64(GMutLock) { return mRNG.next(); }
+
+  bool IsPageInUse(GMutLock, uintptr_t aIndex) {
+    return mAllocPages[aIndex].mState == AllocPageState::InUse;
+  }
+
+  // Is the page free? And if so, has enough time passed that we can use it?
+  bool IsPageAllocatable(GMutLock, uintptr_t aIndex, Time aNow) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    return page.mState != AllocPageState::InUse && aNow >= page.mReuseTime;
+  }
+
+  Maybe<arena_id_t> PageArena(GMutLock aLock, uintptr_t aIndex) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageInUse(aLock, page);
+
+    return page.mArenaId;
+  }
+
+  size_t PageUsableSize(GMutLock aLock, uintptr_t aIndex) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageInUse(aLock, page);
+
+    return page.UsableSize();
+  }
+
+  void SetPageInUse(GMutLock aLock, uintptr_t aIndex,
+                    const Maybe<arena_id_t>& aArenaId, uint8_t* aBaseAddr,
+                    const StackTrace& aAllocStack) {
+    AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageNotInUse(aLock, page);
+
+    page.mState = AllocPageState::InUse;
+    page.mArenaId = aArenaId;
+    page.mBaseAddr = aBaseAddr;
+    page.mAllocStack = Some(aAllocStack);
+    page.mFreeStack = Nothing();
+    page.mReuseTime = kMaxTime;
+
+    mNumPageAllocs++;
+    MOZ_RELEASE_ASSERT(mNumPageAllocs <= kNumAllocPages);
+  }
+
+  void ResizePageInUse(GMutLock aLock, uintptr_t aIndex,
+                       const Maybe<arena_id_t>& aArenaId, uint8_t* aNewBaseAddr,
+                       const StackTrace& aAllocStack) {
+    AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageInUse(aLock, page);
+
+    // page.mState is not changed.
+    if (aArenaId.isSome()) {
+      // Crash if the arenas don't match.
+      MOZ_RELEASE_ASSERT(page.mArenaId == aArenaId);
+    }
+    page.mBaseAddr = aNewBaseAddr;
+    // We could just keep the original alloc stack, but the realloc stack is
+    // more recent and therefore seems more useful.
+    page.mAllocStack = Some(aAllocStack);
+    // page.mFreeStack is not changed.
+    // page.mReuseTime is not changed.
+  };
+
+  void SetPageFreed(GMutLock aLock, uintptr_t aIndex,
+                    const Maybe<arena_id_t>& aArenaId,
+                    const StackTrace& aFreeStack, Delay aReuseDelay) {
+    AllocPageInfo& page = mAllocPages[aIndex];
+    AssertAllocPageInUse(aLock, page);
+
+    page.mState = AllocPageState::Freed;
+
+    // page.mArenaId is left unchanged, for jemalloc_ptr_info() calls that
+    // occur after freeing (e.g. in the PtrInfo test in TestJemalloc.cpp).
+    if (aArenaId.isSome()) {
+      // Crash if the arenas don't match.
+      MOZ_RELEASE_ASSERT(page.mArenaId == aArenaId);
+    }
+
+    // page.musableSize is left unchanged, for reporting on UAF, and for
+    // jemalloc_ptr_info() calls that occur after freeing (e.g. in the PtrInfo
+    // test in TestJemalloc.cpp).
+
+    // page.mAllocStack is left unchanged, for reporting on UAF.
+
+    page.mFreeStack = Some(aFreeStack);
+    page.mReuseTime = GAtomic::Now() + aReuseDelay;
+
+    MOZ_RELEASE_ASSERT(mNumPageAllocs > 0);
+    mNumPageAllocs--;
+  }
+
+  static void CrashOnGuardPage(void* aPtr) {
+    // An operation on a guard page? This is a bounds violation. Deliberately
+    // touch the page in question, to cause a crash that triggers the usual PHC
+    // machinery.
+    LOG("CrashOnGuardPage(%p), bounds violation\n", aPtr);
+    *static_cast<uint8_t*>(aPtr) = 0;
+    MOZ_CRASH("unreachable");
+  }
+
+  void EnsureValidAndInUse(GMutLock, void* aPtr, uintptr_t aIndex)
+      MOZ_REQUIRES(sMutex) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+
+    // The pointer must point to the start of the allocation.
+    MOZ_RELEASE_ASSERT(page.mBaseAddr == aPtr);
+
+    if (page.mState == AllocPageState::Freed) {
+      LOG("EnsureValidAndInUse(%p), use-after-free\n", aPtr);
+      // An operation on a freed page? This is a particular kind of
+      // use-after-free. Deliberately touch the page in question, in order to
+      // cause a crash that triggers the usual PHC machinery. But unlock sMutex
+      // first, because that self-same PHC machinery needs to re-lock it, and
+      // the crash causes non-local control flow so sMutex won't be unlocked
+      // the normal way in the caller.
+      sMutex.Unlock();
+      *static_cast<uint8_t*>(aPtr) = 0;
+      MOZ_CRASH("unreachable");
+    }
+  }
+
+  void FillAddrInfo(GMutLock, uintptr_t aIndex, const void* aBaseAddr,
+                    bool isGuardPage, phc::AddrInfo& aOut) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    if (isGuardPage) {
+      aOut.mKind = phc::AddrInfo::Kind::GuardPage;
+    } else {
+      switch (page.mState) {
+        case AllocPageState::NeverAllocated:
+          aOut.mKind = phc::AddrInfo::Kind::NeverAllocatedPage;
+          break;
+
+        case AllocPageState::InUse:
+          aOut.mKind = phc::AddrInfo::Kind::InUsePage;
+          break;
+
+        case AllocPageState::Freed:
+          aOut.mKind = phc::AddrInfo::Kind::FreedPage;
+          break;
+
+        default:
+          MOZ_CRASH();
+      }
+    }
+    aOut.mBaseAddr = page.mBaseAddr;
+    aOut.mUsableSize = page.UsableSize();
+    aOut.mAllocStack = page.mAllocStack;
+    aOut.mFreeStack = page.mFreeStack;
+  }
+
+  void FillJemallocPtrInfo(GMutLock, const void* aPtr, uintptr_t aIndex,
+                           jemalloc_ptr_info_t* aInfo) {
+    const AllocPageInfo& page = mAllocPages[aIndex];
+    switch (page.mState) {
+      case AllocPageState::NeverAllocated:
+        break;
+
+      case AllocPageState::InUse: {
+        // Only return TagLiveAlloc if the pointer is within the bounds of the
+        // allocation's usable size.
+        uint8_t* base = page.mBaseAddr;
+        uint8_t* limit = base + page.UsableSize();
+        if (base <= aPtr && aPtr < limit) {
+          *aInfo = {TagLiveAlloc, page.mBaseAddr, page.UsableSize(),
+                    page.mArenaId.valueOr(0)};
+          return;
+        }
+        break;
+      }
+
+      case AllocPageState::Freed: {
+        // Only return TagFreedAlloc if the pointer is within the bounds of the
+        // former allocation's usable size.
+        uint8_t* base = page.mBaseAddr;
+        uint8_t* limit = base + page.UsableSize();
+        if (base <= aPtr && aPtr < limit) {
+          *aInfo = {TagFreedAlloc, page.mBaseAddr, page.UsableSize(),
+                    page.mArenaId.valueOr(0)};
+          return;
+        }
+        break;
+      }
+
+      default:
+        MOZ_CRASH();
+    }
+
+    // Pointers into guard pages will end up here, as will pointers into
+    // allocation pages that aren't within the allocation's bounds.
+    *aInfo = {TagUnknown, nullptr, 0, 0};
+  }
+
+#ifndef XP_WIN
+  static void prefork() MOZ_NO_THREAD_SAFETY_ANALYSIS { sMutex.Lock(); }
+  static void postfork_parent() MOZ_NO_THREAD_SAFETY_ANALYSIS {
+    sMutex.Unlock();
+  }
+  static void postfork_child() { sMutex.Init(); }
+#endif
+
+  void IncPageAllocHits(GMutLock) { mPageAllocHits++; }
+  void IncPageAllocMisses(GMutLock) { mPageAllocMisses++; }
+
+  size_t NumPageAllocs(GMutLock) { return mNumPageAllocs; }
+
+  size_t PageAllocHits(GMutLock) { return mPageAllocHits; }
+  size_t PageAllocAttempts(GMutLock) {
+    return mPageAllocHits + mPageAllocMisses;
+  }
+
+  // This is an integer because FdPrintf only supports integer printing.
+  size_t PageAllocHitRate(GMutLock) {
+    return mPageAllocHits * 100 / (mPageAllocHits + mPageAllocMisses);
+  }
+
+ private:
+  template <int N>
+  uint64_t RandomSeed() {
+    // An older version of this code used RandomUint64() here, but on Mac that
+    // function uses arc4random(), which can allocate, which would cause
+    // re-entry, which would be bad. So we just use time() and a local variable
+    // address. These are mediocre sources of entropy, but good enough for PHC.
+    static_assert(N == 0 || N == 1, "must be 0 or 1");
+    uint64_t seed;
+    if (N == 0) {
+      time_t t = time(nullptr);
+      seed = t ^ (t << 32);
+    } else {
+      seed = uintptr_t(&seed) ^ (uintptr_t(&seed) << 32);
+    }
+    return seed;
+  }
+
+  void AssertAllocPageInUse(GMutLock, const AllocPageInfo& aPage) {
+    MOZ_ASSERT(aPage.mState == AllocPageState::InUse);
+    // There is nothing to assert about aPage.mArenaId.
+    MOZ_ASSERT(aPage.mBaseAddr);
+    MOZ_ASSERT(aPage.UsableSize() > 0);
+    MOZ_ASSERT(aPage.mAllocStack.isSome());
+    MOZ_ASSERT(aPage.mFreeStack.isNothing());
+    MOZ_ASSERT(aPage.mReuseTime == kMaxTime);
+  }
+
+  void AssertAllocPageNotInUse(GMutLock, const AllocPageInfo& aPage) {
+    // We can assert a lot about `NeverAllocated` pages, but not much about
+    // `Freed` pages.
+#ifdef DEBUG
+    bool isFresh = aPage.mState == AllocPageState::NeverAllocated;
+    MOZ_ASSERT(isFresh || aPage.mState == AllocPageState::Freed);
+    MOZ_ASSERT_IF(isFresh, aPage.mArenaId == Nothing());
+    MOZ_ASSERT(isFresh == (aPage.mBaseAddr == nullptr));
+    MOZ_ASSERT(isFresh == (aPage.mAllocStack.isNothing()));
+    MOZ_ASSERT(isFresh == (aPage.mFreeStack.isNothing()));
+    MOZ_ASSERT(aPage.mReuseTime != kMaxTime);
+#endif
+  }
+
+  // RNG for deciding which allocations to treat specially. It doesn't need to
+  // be high quality.
+  //
+  // This is a raw pointer for the reason explained in the comment above
+  // GMut's constructor. Don't change it to UniquePtr or anything like that.
+  non_crypto::XorShift128PlusRNG mRNG;
+
+  AllocPageInfo mAllocPages[kNumAllocPages];
+
+  // How many page allocs are currently in use (the max is kNumAllocPages).
+  size_t mNumPageAllocs;
+
+  // How many allocations that could have been page allocs actually were? As
+  // constrained kNumAllocPages. If the hit ratio isn't close to 100% it's
+  // likely that the global constants are poorly chosen.
+  size_t mPageAllocHits;
+  size_t mPageAllocMisses;
+};
+
+Mutex GMut::sMutex;
+
+static GMut* gMut;
+
+//---------------------------------------------------------------------------
+// Page allocation operations
+//---------------------------------------------------------------------------
+
+// Attempt a page allocation if the time and the size are right. Allocated
+// memory is zeroed if aZero is true. On failure, the caller should attempt a
+// normal allocation via sMallocTable. Can be called in a context where
+// GMut::sMutex is locked.
+static void* MaybePageAlloc(const Maybe<arena_id_t>& aArenaId, size_t aReqSize,
+                            size_t aAlignment, bool aZero) {
+  MOZ_ASSERT(IsPowerOfTwo(aAlignment));
+
+  if (aReqSize > kPageSize) {
+    return nullptr;
+  }
+
+  GAtomic::IncrementNow();
+
+  // Decrement the delay. If it's zero, we do a page allocation and reset the
+  // delay to a random number. Because the assignment to the random number isn't
+  // atomic w.r.t. the decrement, we might have a sequence like this:
+  //
+  //     Thread 1                      Thread 2           Thread 3
+  //     --------                      --------           --------
+  // (a) newDelay = --sAllocDelay (-> 0)
+  // (b)                               --sAllocDelay (-> -1)
+  // (c) (newDelay != 0) fails
+  // (d)                                                  --sAllocDelay (-> -2)
+  // (e) sAllocDelay = new_random_number()
+  //
+  // It's critical that sAllocDelay has ReleaseAcquire semantics, because that
+  // guarantees that exactly one thread will see sAllocDelay have the value 0.
+  // (Relaxed semantics wouldn't guarantee that.)
+  //
+  // It's also nice that sAllocDelay is signed, given that we can decrement to
+  // below zero. (Strictly speaking, an unsigned integer would also work due
+  // to wrapping, but a signed integer is conceptually cleaner.)
+  //
+  // Finally, note that the decrements that occur between (a) and (e) above are
+  // effectively ignored, because (e) clobbers them. This shouldn't be a
+  // problem; it effectively just adds a little more randomness to
+  // new_random_number(). An early version of this code tried to account for
+  // these decrements by doing `sAllocDelay += new_random_number()`. However, if
+  // new_random_value() is small, the number of decrements between (a) and (e)
+  // can easily exceed it, whereupon sAllocDelay ends up negative after
+  // `sAllocDelay += new_random_number()`, and the zero-check never succeeds
+  // again. (At least, not until sAllocDelay wraps around on overflow, which
+  // would take a very long time indeed.)
+  //
+  int32_t newDelay = GAtomic::DecrementDelay();
+  if (newDelay != 0) {
+    return nullptr;
+  }
+
+  if (GTls::IsDisabledOnCurrentThread()) {
+    return nullptr;
+  }
+
+  // Disable on this thread *before* getting the stack trace.
+  AutoDisableOnCurrentThread disable;
+
+  // Get the stack trace *before* locking the mutex. If we return nullptr then
+  // it was a waste, but it's not so frequent, and doing a stack walk while
+  // the mutex is locked is problematic (see the big comment on
+  // StackTrace::Fill() for details).
+  StackTrace allocStack;
+  allocStack.Fill();
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  Time now = GAtomic::Now();
+  Delay newAllocDelay = Rnd64ToDelay<kAvgAllocDelay>(gMut->Random64(lock));
+
+  // We start at a random page alloc and wrap around, to ensure pages get even
+  // amounts of use.
+  uint8_t* ptr = nullptr;
+  uint8_t* pagePtr = nullptr;
+  for (uintptr_t n = 0, i = size_t(gMut->Random64(lock)) % kNumAllocPages;
+       n < kNumAllocPages; n++, i = (i + 1) % kNumAllocPages) {
+    if (!gMut->IsPageAllocatable(lock, i, now)) {
+      continue;
+    }
+
+    pagePtr = gConst->AllocPagePtr(i);
+    MOZ_ASSERT(pagePtr);
+    bool ok =
+#ifdef XP_WIN
+        !!VirtualAlloc(pagePtr, kPageSize, MEM_COMMIT, PAGE_READWRITE);
+#else
+        mprotect(pagePtr, kPageSize, PROT_READ | PROT_WRITE) == 0;
+#endif
+    size_t usableSize = sMallocTable.malloc_good_size(aReqSize);
+    if (ok) {
+      MOZ_ASSERT(usableSize > 0);
+
+      // Put the allocation as close to the end of the page as possible,
+      // allowing for alignment requirements.
+      ptr = pagePtr + kPageSize - usableSize;
+      if (aAlignment != 1) {
+        ptr = reinterpret_cast<uint8_t*>(
+            (reinterpret_cast<uintptr_t>(ptr) & ~(aAlignment - 1)));
+      }
+
+      gMut->SetPageInUse(lock, i, aArenaId, ptr, allocStack);
+
+      if (aZero) {
+        memset(ptr, 0, usableSize);
+      } else {
+#ifdef DEBUG
+        memset(ptr, kAllocJunk, usableSize);
+#endif
+      }
+    }
+
+    gMut->IncPageAllocHits(lock);
+    LOG("PageAlloc(%zu, %zu) -> %p[%zu]/%p (%zu) (z%zu), sAllocDelay <- %zu, "
+        "fullness %zu/%zu, hits %zu/%zu (%zu%%)\n",
+        aReqSize, aAlignment, pagePtr, i, ptr, usableSize, size_t(aZero),
+        size_t(newAllocDelay), gMut->NumPageAllocs(lock), kNumAllocPages,
+        gMut->PageAllocHits(lock), gMut->PageAllocAttempts(lock),
+        gMut->PageAllocHitRate(lock));
+    break;
+  }
+
+  if (!pagePtr) {
+    // No pages are available, or VirtualAlloc/mprotect failed.
+    gMut->IncPageAllocMisses(lock);
+    LOG("No PageAlloc(%zu, %zu), sAllocDelay <- %zu, fullness %zu/%zu, hits "
+        "%zu/%zu "
+        "(%zu%%)\n",
+        aReqSize, aAlignment, size_t(newAllocDelay), gMut->NumPageAllocs(lock),
+        kNumAllocPages, gMut->PageAllocHits(lock),
+        gMut->PageAllocAttempts(lock), gMut->PageAllocHitRate(lock));
+  }
+
+  // Set the new alloc delay.
+  GAtomic::SetAllocDelay(newAllocDelay);
+
+  return ptr;
+}
+
+static void FreePage(GMutLock aLock, uintptr_t aIndex,
+                     const Maybe<arena_id_t>& aArenaId,
+                     const StackTrace& aFreeStack, Delay aReuseDelay) {
+  void* pagePtr = gConst->AllocPagePtr(aIndex);
+#ifdef XP_WIN
+  if (!VirtualFree(pagePtr, kPageSize, MEM_DECOMMIT)) {
+    return;
+  }
+#else
+  if (mmap(pagePtr, kPageSize, PROT_NONE, MAP_FIXED | MAP_PRIVATE | MAP_ANON,
+           -1, 0) == MAP_FAILED) {
+    return;
+  }
+#endif
+
+  gMut->SetPageFreed(aLock, aIndex, aArenaId, aFreeStack, aReuseDelay);
+}
+
+//---------------------------------------------------------------------------
+// replace-malloc machinery
+//---------------------------------------------------------------------------
+
+// This handles malloc, moz_arena_malloc, and realloc-with-a-nullptr.
+MOZ_ALWAYS_INLINE static void* PageMalloc(const Maybe<arena_id_t>& aArenaId,
+                                          size_t aReqSize) {
+  void* ptr = MaybePageAlloc(aArenaId, aReqSize, /* aAlignment */ 1,
+                             /* aZero */ false);
+  return ptr ? ptr
+             : (aArenaId.isSome()
+                    ? sMallocTable.moz_arena_malloc(*aArenaId, aReqSize)
+                    : sMallocTable.malloc(aReqSize));
+}
+
+static void* replace_malloc(size_t aReqSize) {
+  return PageMalloc(Nothing(), aReqSize);
+}
+
+static Delay ReuseDelay(GMutLock aLock) {
+  return (kAvgPageReuseDelay / 2) +
+         Rnd64ToDelay<kAvgPageReuseDelay / 2>(gMut->Random64(aLock));
+}
+
+// This handles both calloc and moz_arena_calloc.
+MOZ_ALWAYS_INLINE static void* PageCalloc(const Maybe<arena_id_t>& aArenaId,
+                                          size_t aNum, size_t aReqSize) {
+  CheckedInt<size_t> checkedSize = CheckedInt<size_t>(aNum) * aReqSize;
+  if (!checkedSize.isValid()) {
+    return nullptr;
+  }
+
+  void* ptr = MaybePageAlloc(aArenaId, checkedSize.value(), /* aAlignment */ 1,
+                             /* aZero */ true);
+  return ptr ? ptr
+             : (aArenaId.isSome()
+                    ? sMallocTable.moz_arena_calloc(*aArenaId, aNum, aReqSize)
+                    : sMallocTable.calloc(aNum, aReqSize));
+}
+
+static void* replace_calloc(size_t aNum, size_t aReqSize) {
+  return PageCalloc(Nothing(), aNum, aReqSize);
+}
+
+// This function handles both realloc and moz_arena_realloc.
+//
+// As always, realloc is complicated, and doubly so when there are two
+// different kinds of allocations in play. Here are the possible transitions,
+// and what we do in practice.
+//
+// - normal-to-normal: This is straightforward and obviously necessary.
+//
+// - normal-to-page: This is disallowed because it would require getting the
+//   arenaId of the normal allocation, which isn't possible in non-DEBUG builds
+//   for security reasons.
+//
+// - page-to-page: This is done whenever possible, i.e. whenever the new size
+//   is less than or equal to 4 KiB. This choice counterbalances the
+//   disallowing of normal-to-page allocations, in order to avoid biasing
+//   towards or away from page allocations. It always occurs in-place.
+//
+// - page-to-normal: this is done only when necessary, i.e. only when the new
+//   size is greater than 4 KiB. This choice naturally flows from the
+//   prior choice on page-to-page transitions.
+//
+// In summary: realloc doesn't change the allocation kind unless it must.
+//
+MOZ_ALWAYS_INLINE static void* PageRealloc(const Maybe<arena_id_t>& aArenaId,
+                                           void* aOldPtr, size_t aNewSize) {
+  if (!aOldPtr) {
+    // Null pointer. Treat like malloc(aNewSize).
+    return PageMalloc(aArenaId, aNewSize);
+  }
+
+  PtrKind pk = gConst->PtrKind(aOldPtr);
+  if (pk.IsNothing()) {
+    // A normal-to-normal transition.
+    return aArenaId.isSome()
+               ? sMallocTable.moz_arena_realloc(*aArenaId, aOldPtr, aNewSize)
+               : sMallocTable.realloc(aOldPtr, aNewSize);
+  }
+
+  if (pk.IsGuardPage()) {
+    GMut::CrashOnGuardPage(aOldPtr);
+  }
+
+  // At this point we know we have an allocation page.
+  uintptr_t index = pk.AllocPageIndex();
+
+  // A page-to-something transition.
+
+  // Note that `disable` has no effect unless it is emplaced below.
+  Maybe<AutoDisableOnCurrentThread> disable;
+  // Get the stack trace *before* locking the mutex.
+  StackTrace stack;
+  if (GTls::IsDisabledOnCurrentThread()) {
+    // PHC is disabled on this thread. Leave the stack empty.
+  } else {
+    // Disable on this thread *before* getting the stack trace.
+    disable.emplace();
+    stack.Fill();
+  }
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  // Check for realloc() of a freed block.
+  gMut->EnsureValidAndInUse(lock, aOldPtr, index);
+
+  if (aNewSize <= kPageSize) {
+    // A page-to-page transition. Just keep using the page allocation. We do
+    // this even if the thread is disabled, because it doesn't create a new
+    // page allocation. Note that ResizePageInUse() checks aArenaId.
+    //
+    // Move the bytes with memmove(), because the old allocation and the new
+    // allocation overlap. Move the usable size rather than the requested size,
+    // because the user might have used malloc_usable_size() and filled up the
+    // usable size.
+    size_t oldUsableSize = gMut->PageUsableSize(lock, index);
+    size_t newUsableSize = sMallocTable.malloc_good_size(aNewSize);
+    uint8_t* pagePtr = gConst->AllocPagePtr(index);
+    uint8_t* newPtr = pagePtr + kPageSize - newUsableSize;
+    memmove(newPtr, aOldPtr, std::min(oldUsableSize, aNewSize));
+    gMut->ResizePageInUse(lock, index, aArenaId, newPtr, stack);
+    LOG("PageRealloc-Reuse(%p, %zu) -> %p\n", aOldPtr, aNewSize, newPtr);
+    return newPtr;
+  }
+
+  // A page-to-normal transition (with the new size greater than page-sized).
+  // (Note that aArenaId is checked below.)
+  void* newPtr;
+  if (aArenaId.isSome()) {
+    newPtr = sMallocTable.moz_arena_malloc(*aArenaId, aNewSize);
+  } else {
+    Maybe<arena_id_t> oldArenaId = gMut->PageArena(lock, index);
+    newPtr = (oldArenaId.isSome()
+                  ? sMallocTable.moz_arena_malloc(*oldArenaId, aNewSize)
+                  : sMallocTable.malloc(aNewSize));
+  }
+  if (!newPtr) {
+    return nullptr;
+  }
+
+  MOZ_ASSERT(aNewSize > kPageSize);
+
+  Delay reuseDelay = ReuseDelay(lock);
+
+  // Copy the usable size rather than the requested size, because the user
+  // might have used malloc_usable_size() and filled up the usable size. Note
+  // that FreePage() checks aArenaId (via SetPageFreed()).
+  size_t oldUsableSize = gMut->PageUsableSize(lock, index);
+  memcpy(newPtr, aOldPtr, std::min(oldUsableSize, aNewSize));
+  FreePage(lock, index, aArenaId, stack, reuseDelay);
+  LOG("PageRealloc-Free(%p[%zu], %zu) -> %p, %zu delay, reuse at ~%zu\n",
+      aOldPtr, index, aNewSize, newPtr, size_t(reuseDelay),
+      size_t(GAtomic::Now()) + reuseDelay);
+
+  return newPtr;
+}
+
+static void* replace_realloc(void* aOldPtr, size_t aNewSize) {
+  return PageRealloc(Nothing(), aOldPtr, aNewSize);
+}
+
+// This handles both free and moz_arena_free.
+MOZ_ALWAYS_INLINE static void PageFree(const Maybe<arena_id_t>& aArenaId,
+                                       void* aPtr) {
+  PtrKind pk = gConst->PtrKind(aPtr);
+  if (pk.IsNothing()) {
+    // Not a page allocation.
+    return aArenaId.isSome() ? sMallocTable.moz_arena_free(*aArenaId, aPtr)
+                             : sMallocTable.free(aPtr);
+  }
+
+  if (pk.IsGuardPage()) {
+    GMut::CrashOnGuardPage(aPtr);
+  }
+
+  // At this point we know we have an allocation page.
+  uintptr_t index = pk.AllocPageIndex();
+
+  // Note that `disable` has no effect unless it is emplaced below.
+  Maybe<AutoDisableOnCurrentThread> disable;
+  // Get the stack trace *before* locking the mutex.
+  StackTrace freeStack;
+  if (GTls::IsDisabledOnCurrentThread()) {
+    // PHC is disabled on this thread. Leave the stack empty.
+  } else {
+    // Disable on this thread *before* getting the stack trace.
+    disable.emplace();
+    freeStack.Fill();
+  }
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  // Check for a double-free.
+  gMut->EnsureValidAndInUse(lock, aPtr, index);
+
+  // Note that FreePage() checks aArenaId (via SetPageFreed()).
+  Delay reuseDelay = ReuseDelay(lock);
+  FreePage(lock, index, aArenaId, freeStack, reuseDelay);
+
+  LOG("PageFree(%p[%zu]), %zu delay, reuse at ~%zu, fullness %zu/%zu\n", aPtr,
+      index, size_t(reuseDelay), size_t(GAtomic::Now()) + reuseDelay,
+      gMut->NumPageAllocs(lock), kNumAllocPages);
+}
+
+static void replace_free(void* aPtr) { return PageFree(Nothing(), aPtr); }
+
+// This handles memalign and moz_arena_memalign.
+MOZ_ALWAYS_INLINE static void* PageMemalign(const Maybe<arena_id_t>& aArenaId,
+                                            size_t aAlignment,
+                                            size_t aReqSize) {
+  MOZ_RELEASE_ASSERT(IsPowerOfTwo(aAlignment));
+
+  // PHC can't satisfy an alignment greater than a page size, so fall back to
+  // mozjemalloc in that case.
+  void* ptr = nullptr;
+  if (aAlignment <= kPageSize) {
+    ptr = MaybePageAlloc(aArenaId, aReqSize, aAlignment, /* aZero */ false);
+  }
+  return ptr ? ptr
+             : (aArenaId.isSome()
+                    ? sMallocTable.moz_arena_memalign(*aArenaId, aAlignment,
+                                                      aReqSize)
+                    : sMallocTable.memalign(aAlignment, aReqSize));
+}
+
+static void* replace_memalign(size_t aAlignment, size_t aReqSize) {
+  return PageMemalign(Nothing(), aAlignment, aReqSize);
+}
+
+static size_t replace_malloc_usable_size(usable_ptr_t aPtr) {
+  PtrKind pk = gConst->PtrKind(aPtr);
+  if (pk.IsNothing()) {
+    // Not a page allocation. Measure it normally.
+    return sMallocTable.malloc_usable_size(aPtr);
+  }
+
+  if (pk.IsGuardPage()) {
+    GMut::CrashOnGuardPage(const_cast<void*>(aPtr));
+  }
+
+  // At this point we know we have an allocation page.
+  uintptr_t index = pk.AllocPageIndex();
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  // Check for malloc_usable_size() of a freed block.
+  gMut->EnsureValidAndInUse(lock, const_cast<void*>(aPtr), index);
+
+  return gMut->PageUsableSize(lock, index);
+}
+
+void replace_jemalloc_stats(jemalloc_stats_t* aStats,
+                            jemalloc_bin_stats_t* aBinStats) {
+  sMallocTable.jemalloc_stats_internal(aStats, aBinStats);
+
+  // Add all the pages to `mapped`.
+  size_t mapped = kAllPagesSize;
+  aStats->mapped += mapped;
+
+  size_t allocated = 0;
+  {
+    MutexAutoLock lock(GMut::sMutex);
+
+    // Add usable space of in-use allocations to `allocated`.
+    for (size_t i = 0; i < kNumAllocPages; i++) {
+      if (gMut->IsPageInUse(lock, i)) {
+        allocated += gMut->PageUsableSize(lock, i);
+      }
+    }
+  }
+  aStats->allocated += allocated;
+
+  // Waste is the gap between `allocated` and `mapped`.
+  size_t waste = mapped - allocated;
+  aStats->waste += waste;
+
+  // aStats.page_cache and aStats.bin_unused are left unchanged because PHC
+  // doesn't have anything corresponding to those.
+
+  // gConst and gMut are normal heap allocations, so they're measured by
+  // mozjemalloc as `allocated`. Move them into `bookkeeping`.
+  size_t bookkeeping = sMallocTable.malloc_usable_size(gConst) +
+                       sMallocTable.malloc_usable_size(gMut);
+  aStats->allocated -= bookkeeping;
+  aStats->bookkeeping += bookkeeping;
+}
+
+void replace_jemalloc_ptr_info(const void* aPtr, jemalloc_ptr_info_t* aInfo) {
+  // We need to implement this properly, because various code locations do
+  // things like checking that allocations are in the expected arena.
+  PtrKind pk = gConst->PtrKind(aPtr);
+  if (pk.IsNothing()) {
+    // Not a page allocation.
+    return sMallocTable.jemalloc_ptr_info(aPtr, aInfo);
+  }
+
+  if (pk.IsGuardPage()) {
+    // Treat a guard page as unknown because there's no better alternative.
+    *aInfo = {TagUnknown, nullptr, 0, 0};
+    return;
+  }
+
+  // At this point we know we have an allocation page.
+  uintptr_t index = pk.AllocPageIndex();
+
+  MutexAutoLock lock(GMut::sMutex);
+
+  gMut->FillJemallocPtrInfo(lock, aPtr, index, aInfo);
+#if DEBUG
+  LOG("JemallocPtrInfo(%p[%zu]) -> {%zu, %p, %zu, %zu}\n", aPtr, index,
+      size_t(aInfo->tag), aInfo->addr, aInfo->size, aInfo->arenaId);
+#else
+  LOG("JemallocPtrInfo(%p[%zu]) -> {%zu, %p, %zu}\n", aPtr, index,
+      size_t(aInfo->tag), aInfo->addr, aInfo->size);
+#endif
+}
+
+arena_id_t replace_moz_create_arena_with_params(arena_params_t* aParams) {
+  // No need to do anything special here.
+  return sMallocTable.moz_create_arena_with_params(aParams);
+}
+
+void replace_moz_dispose_arena(arena_id_t aArenaId) {
+  // No need to do anything special here.
+  return sMallocTable.moz_dispose_arena(aArenaId);
+}
+
+void* replace_moz_arena_malloc(arena_id_t aArenaId, size_t aReqSize) {
+  return PageMalloc(Some(aArenaId), aReqSize);
+}
+
+void* replace_moz_arena_calloc(arena_id_t aArenaId, size_t aNum,
+                               size_t aReqSize) {
+  return PageCalloc(Some(aArenaId), aNum, aReqSize);
+}
+
+void* replace_moz_arena_realloc(arena_id_t aArenaId, void* aOldPtr,
+                                size_t aNewSize) {
+  return PageRealloc(Some(aArenaId), aOldPtr, aNewSize);
+}
+
+void replace_moz_arena_free(arena_id_t aArenaId, void* aPtr) {
+  return PageFree(Some(aArenaId), aPtr);
+}
+
+void* replace_moz_arena_memalign(arena_id_t aArenaId, size_t aAlignment,
+                                 size_t aReqSize) {
+  return PageMemalign(Some(aArenaId), aAlignment, aReqSize);
+}
+
+class PHCBridge : public ReplaceMallocBridge {
+  virtual bool IsPHCAllocation(const void* aPtr, phc::AddrInfo* aOut) override {
+    PtrKind pk = gConst->PtrKind(aPtr);
+    if (pk.IsNothing()) {
+      return false;
+    }
+
+    bool isGuardPage = false;
+    if (pk.IsGuardPage()) {
+      if ((uintptr_t(aPtr) % kPageSize) < (kPageSize / 2)) {
+        // The address is in the lower half of a guard page, so it's probably an
+        // overflow. But first check that it is not on the very first guard
+        // page, in which case it cannot be an overflow, and we ignore it.
+        if (gConst->IsInFirstGuardPage(aPtr)) {
+          return false;
+        }
+
+        // Get the allocation page preceding this guard page.
+        pk = gConst->PtrKind(static_cast<const uint8_t*>(aPtr) - kPageSize);
+
+      } else {
+        // The address is in the upper half of a guard page, so it's probably an
+        // underflow. Get the allocation page following this guard page.
+        pk = gConst->PtrKind(static_cast<const uint8_t*>(aPtr) + kPageSize);
+      }
+
+      // Make a note of the fact that we hit a guard page.
+      isGuardPage = true;
+    }
+
+    // At this point we know we have an allocation page.
+    uintptr_t index = pk.AllocPageIndex();
+
+    if (aOut) {
+      MutexAutoLock lock(GMut::sMutex);
+      gMut->FillAddrInfo(lock, index, aPtr, isGuardPage, *aOut);
+      LOG("IsPHCAllocation: %zu, %p, %zu, %zu, %zu\n", size_t(aOut->mKind),
+          aOut->mBaseAddr, aOut->mUsableSize,
+          aOut->mAllocStack.isSome() ? aOut->mAllocStack->mLength : 0,
+          aOut->mFreeStack.isSome() ? aOut->mFreeStack->mLength : 0);
+    }
+    return true;
+  }
+
+  virtual void DisablePHCOnCurrentThread() override {
+    GTls::DisableOnCurrentThread();
+    LOG("DisablePHCOnCurrentThread: %zu\n", 0ul);
+  }
+
+  virtual void ReenablePHCOnCurrentThread() override {
+    GTls::EnableOnCurrentThread();
+    LOG("ReenablePHCOnCurrentThread: %zu\n", 0ul);
+  }
+
+  virtual bool IsPHCEnabledOnCurrentThread() override {
+    bool enabled = !GTls::IsDisabledOnCurrentThread();
+    LOG("IsPHCEnabledOnCurrentThread: %zu\n", size_t(enabled));
+    return enabled;
+  }
+};
+
+// WARNING: this function runs *very* early -- before all static initializers
+// have run. For this reason, non-scalar globals (gConst, gMut) are allocated
+// dynamically (so we can guarantee their construction in this function) rather
+// than statically. GAtomic and GTls contain simple static data that doesn't
+// involve static initializers so they don't need to be allocated dynamically.
+void replace_init(malloc_table_t* aMallocTable, ReplaceMallocBridge** aBridge) {
+  // Don't run PHC if the page size isn't 4 KiB.
+  jemalloc_stats_t stats;
+  aMallocTable->jemalloc_stats_internal(&stats, nullptr);
+  if (stats.page_size != kPageSize) {
+    return;
+  }
+
+  sMallocTable = *aMallocTable;
+
+  // The choices of which functions to replace are complex enough that we set
+  // them individually instead of using MALLOC_FUNCS/malloc_decls.h.
+
+  aMallocTable->malloc = replace_malloc;
+  aMallocTable->calloc = replace_calloc;
+  aMallocTable->realloc = replace_realloc;
+  aMallocTable->free = replace_free;
+  aMallocTable->memalign = replace_memalign;
+
+  // posix_memalign, aligned_alloc & valloc: unset, which means they fall back
+  // to replace_memalign.
+  aMallocTable->malloc_usable_size = replace_malloc_usable_size;
+  // default malloc_good_size: the default suffices.
+
+  aMallocTable->jemalloc_stats_internal = replace_jemalloc_stats;
+  // jemalloc_purge_freed_pages: the default suffices.
+  // jemalloc_free_dirty_pages: the default suffices.
+  // jemalloc_thread_local_arena: the default suffices.
+  aMallocTable->jemalloc_ptr_info = replace_jemalloc_ptr_info;
+
+  aMallocTable->moz_create_arena_with_params =
+      replace_moz_create_arena_with_params;
+  aMallocTable->moz_dispose_arena = replace_moz_dispose_arena;
+  aMallocTable->moz_arena_malloc = replace_moz_arena_malloc;
+  aMallocTable->moz_arena_calloc = replace_moz_arena_calloc;
+  aMallocTable->moz_arena_realloc = replace_moz_arena_realloc;
+  aMallocTable->moz_arena_free = replace_moz_arena_free;
+  aMallocTable->moz_arena_memalign = replace_moz_arena_memalign;
+
+  static PHCBridge bridge;
+  *aBridge = &bridge;
+
+#ifndef XP_WIN
+  // Avoid deadlocks when forking by acquiring our state lock prior to forking
+  // and releasing it after forking. See |LogAlloc|'s |replace_init| for
+  // in-depth details.
+  //
+  // Note: This must run after attempting an allocation so as to give the
+  // system malloc a chance to insert its own atfork handler.
+  sMallocTable.malloc(-1);
+  pthread_atfork(GMut::prefork, GMut::postfork_parent, GMut::postfork_child);
+#endif
+
+  // gConst and gMut are never freed. They live for the life of the process.
+  gConst = InfallibleAllocPolicy::new_<GConst>();
+  GTls::Init();
+  gMut = InfallibleAllocPolicy::new_<GMut>();
+  {
+    MutexAutoLock lock(GMut::sMutex);
+    Delay firstAllocDelay =
+        Rnd64ToDelay<kAvgFirstAllocDelay>(gMut->Random64(lock));
+    GAtomic::Init(firstAllocDelay);
+  }
+}
diff --git a/memory/replace/phc/PHC.h b/memory/replace/phc/PHC.h
new file mode 100644
index 0000000000..139a0e81c9
--- /dev/null
+++ b/memory/replace/phc/PHC.h
@@ -0,0 +1,102 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef PHC_h
+#define PHC_h
+
+#include "mozilla/Assertions.h"
+#include "mozilla/Maybe.h"
+#include <stdint.h>
+#include <stdlib.h>
+
+namespace mozilla {
+namespace phc {
+
+// Note: a stack trace may have no frames due to a collection problem.
+//
+// Also note: a more compact stack trace representation could be achieved with
+// some effort.
+struct StackTrace {
+ public:
+  static const size_t kMaxFrames = 16;
+
+  // The number of PCs in the stack trace.
+  size_t mLength;
+
+  // The PCs in the stack trace. Only the first mLength are initialized.
+  const void* mPcs[kMaxFrames];
+
+ public:
+  StackTrace() : mLength(0) {}
+};
+
+// Info from PHC about an address in memory.
+class AddrInfo {
+ public:
+  enum class Kind {
+    // The address is not in PHC-managed memory.
+    Unknown = 0,
+
+    // The address is within a PHC page that has never been allocated. A crash
+    // involving such an address is unlikely in practice, because it would
+    // require the crash to happen quite early.
+    NeverAllocatedPage = 1,
+
+    // The address is within a PHC page that is in use.
+    InUsePage = 2,
+
+    // The address is within a PHC page that has been allocated and then freed.
+    // A crash involving such an address most likely indicates a
+    // use-after-free. (A sufficiently wild write -- e.g. a large buffer
+    // overflow -- could also trigger it, but this is less likely.)
+    FreedPage = 3,
+
+    // The address is within a PHC guard page. A crash involving such an
+    // address most likely indicates a buffer overflow. (Again, a sufficiently
+    // wild write could unluckily trigger it, but this is less likely.)
+    GuardPage = 4,
+  };
+
+  // The page kind.
+  Kind mKind;
+
+  // The starting address of the allocation.
+  // - Unknown | NeverAllocatedPage: nullptr.
+  // - InUsePage | FreedPage: the address of the allocation within the page.
+  // - GuardPage: the mBaseAddr value from the preceding allocation page.
+  const void* mBaseAddr;
+
+  // The usable size, which could be bigger than the requested size.
+  // - Unknown | NeverAllocatePage: 0.
+  // - InUsePage | FreedPage: the usable size of the allocation within the page.
+  // - GuardPage: the mUsableSize value from the preceding allocation page.
+  size_t mUsableSize;
+
+  // The allocation stack.
+  // - Unknown | NeverAllocatedPage: Nothing.
+  // - InUsePage | FreedPage: Some.
+  // - GuardPage: the mAllocStack value from the preceding allocation page.
+  mozilla::Maybe<StackTrace> mAllocStack;
+
+  // The free stack.
+  // - Unknown | NeverAllocatedPage | InUsePage: Nothing.
+  // - FreedPage: Some.
+  // - GuardPage: the mFreeStack value from the preceding allocation page.
+  mozilla::Maybe<StackTrace> mFreeStack;
+
+  // Default to no PHC info.
+  AddrInfo()
+      : mKind(Kind::Unknown),
+        mBaseAddr(nullptr),
+        mUsableSize(0),
+        mAllocStack(),
+        mFreeStack() {}
+};
+
+}  // namespace phc
+}  // namespace mozilla
+
+#endif /* PHC_h */
diff --git a/memory/replace/phc/moz.build b/memory/replace/phc/moz.build
new file mode 100644
index 0000000000..e2d233642c
--- /dev/null
+++ b/memory/replace/phc/moz.build
@@ -0,0 +1,50 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+ReplaceMalloc("phc")
+
+DEFINES["MOZ_NO_MOZALLOC"] = True
+DEFINES["IMPL_MFBT"] = True
+
+LOCAL_INCLUDES += [
+    "../logalloc",
+    "/memory/build",
+]
+
+EXPORTS += [
+    "PHC.h",
+]
+
+UNIFIED_SOURCES += [
+    "PHC.cpp",
+]
+
+if CONFIG["MOZ_BUILD_APP"] == "memory":
+    UNIFIED_SOURCES += [
+        "/mfbt/double-conversion/double-conversion/bignum-dtoa.cc",
+        "/mfbt/double-conversion/double-conversion/bignum.cc",
+        "/mfbt/double-conversion/double-conversion/cached-powers.cc",
+        "/mfbt/double-conversion/double-conversion/double-to-string.cc",
+        "/mfbt/double-conversion/double-conversion/fast-dtoa.cc",
+        "/mfbt/double-conversion/double-conversion/fixed-dtoa.cc",
+        "/mfbt/double-conversion/double-conversion/string-to-double.cc",
+        "/mfbt/double-conversion/double-conversion/strtod.cc",
+        "/mozglue/misc/Printf.cpp",
+    ]
+
+if not CONFIG["MOZ_REPLACE_MALLOC_STATIC"]:
+    UNIFIED_SOURCES += [
+        "../logalloc/FdPrintf.cpp",
+        "/mozglue/misc/StackWalk.cpp",
+    ]
+    if CONFIG["OS_ARCH"] == "WINNT":
+        OS_LIBS += [
+            "dbghelp",
+        ]
+
+TEST_DIRS += ["test"]
+
+DisableStlWrapping()
diff --git a/memory/replace/phc/test/gtest/TestPHC.cpp b/memory/replace/phc/test/gtest/TestPHC.cpp
new file mode 100644
index 0000000000..b83db800f2
--- /dev/null
+++ b/memory/replace/phc/test/gtest/TestPHC.cpp
@@ -0,0 +1,303 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "gtest/gtest.h"
+
+#include "mozmemory.h"
+#include "replace_malloc_bridge.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/mozalloc.h"
+#include "../../PHC.h"
+
+using namespace mozilla;
+
+bool PHCInfoEq(phc::AddrInfo& aInfo, phc::AddrInfo::Kind aKind, void* aBaseAddr,
+               size_t aUsableSize, bool aHasAllocStack, bool aHasFreeStack) {
+  return aInfo.mKind == aKind && aInfo.mBaseAddr == aBaseAddr &&
+         aInfo.mUsableSize == aUsableSize &&
+         // Proper stack traces will have at least 3 elements.
+         (aHasAllocStack ? (aInfo.mAllocStack->mLength > 2)
+                         : (aInfo.mAllocStack.isNothing())) &&
+         (aHasFreeStack ? (aInfo.mFreeStack->mLength > 2)
+                        : (aInfo.mFreeStack.isNothing()));
+}
+
+bool JeInfoEq(jemalloc_ptr_info_t& aInfo, PtrInfoTag aTag, void* aAddr,
+              size_t aSize, arena_id_t arenaId) {
+  return aInfo.tag == aTag && aInfo.addr == aAddr && aInfo.size == aSize
+#ifdef MOZ_DEBUG
+         && aInfo.arenaId == arenaId
+#endif
+      ;
+}
+
+uint8_t* GetPHCAllocation(size_t aSize, size_t aAlignment = 1) {
+  // A crude but effective way to get a PHC allocation.
+  for (int i = 0; i < 2000000; i++) {
+    void* p = (aAlignment == 1) ? moz_xmalloc(aSize)
+                                : moz_xmemalign(aAlignment, aSize);
+    if (ReplaceMalloc::IsPHCAllocation(p, nullptr)) {
+      return (uint8_t*)p;
+    }
+    free(p);
+  }
+  return nullptr;
+}
+
+static const size_t kPageSize = 4096;
+
+TEST(PHC, TestPHCAllocations)
+{
+  // First, check that allocations of various sizes all get put at the end of
+  // their page as expected. Also, check their sizes are as expected.
+
+#define ASSERT_POS(n1, n2)                                      \
+  p = (uint8_t*)moz_xrealloc(p, (n1));                          \
+  ASSERT_EQ((reinterpret_cast<uintptr_t>(p) & (kPageSize - 1)), \
+            kPageSize - (n2));                                  \
+  ASSERT_EQ(moz_malloc_usable_size(p), (n2));
+
+  uint8_t* p = GetPHCAllocation(1);
+  if (!p) {
+    MOZ_CRASH("failed to get a PHC allocation");
+  }
+
+  // On Win64 the smallest possible allocation is 16 bytes. On other platforms
+  // it is 8 bytes.
+#if defined(XP_WIN) && defined(HAVE_64BIT_BUILD)
+  ASSERT_POS(8U, 16U);
+#else
+  ASSERT_POS(8U, 8U);
+#endif
+  ASSERT_POS(16U, 16U);
+  ASSERT_POS(32U, 32U);
+  ASSERT_POS(64U, 64U);
+  ASSERT_POS(128U, 128U);
+  ASSERT_POS(256U, 256U);
+  ASSERT_POS(512U, 512U);
+  ASSERT_POS(1024U, 1024U);
+  ASSERT_POS(2048U, 2048U);
+  ASSERT_POS(4096U, 4096U);
+
+  free(p);
+
+#undef ASSERT_POS
+
+  // Second, do similar checking with allocations of various alignments. Also
+  // check that their sizes (which are different to allocations with normal
+  // alignment) are the same as the sizes of equivalent non-PHC allocations.
+
+#define ASSERT_ALIGN(a1, a2)                                    \
+  p = (uint8_t*)GetPHCAllocation(8, (a1));                      \
+  ASSERT_EQ((reinterpret_cast<uintptr_t>(p) & (kPageSize - 1)), \
+            kPageSize - (a2));                                  \
+  ASSERT_EQ(moz_malloc_usable_size(p), (a2));                   \
+  free(p);                                                      \
+  p = (uint8_t*)moz_xmemalign((a1), 8);                         \
+  ASSERT_EQ(moz_malloc_usable_size(p), (a2));                   \
+  free(p);
+
+  // On Win64 the smallest possible allocation is 16 bytes. On other platforms
+  // it is 8 bytes.
+#if defined(XP_WIN) && defined(HAVE_64BIT_BUILD)
+  ASSERT_ALIGN(8U, 16U);
+#else
+  ASSERT_ALIGN(8U, 8U);
+#endif
+  ASSERT_ALIGN(16U, 16U);
+  ASSERT_ALIGN(32U, 32U);
+  ASSERT_ALIGN(64U, 64U);
+  ASSERT_ALIGN(128U, 128U);
+  ASSERT_ALIGN(256U, 256U);
+  ASSERT_ALIGN(512U, 512U);
+  ASSERT_ALIGN(1024U, 1024U);
+  ASSERT_ALIGN(2048U, 2048U);
+  ASSERT_ALIGN(4096U, 4096U);
+
+#undef ASSERT_ALIGN
+}
+
+TEST(PHC, TestPHCInfo)
+{
+  int stackVar;
+  phc::AddrInfo phcInfo;
+  jemalloc_ptr_info_t jeInfo;
+
+  // Test a default AddrInfo.
+  ASSERT_TRUE(PHCInfoEq(phcInfo, phc::AddrInfo::Kind::Unknown, nullptr, 0ul,
+                        false, false));
+
+  // Test some non-PHC allocation addresses.
+  ASSERT_FALSE(ReplaceMalloc::IsPHCAllocation(nullptr, &phcInfo));
+  ASSERT_TRUE(PHCInfoEq(phcInfo, phc::AddrInfo::Kind::Unknown, nullptr, 0,
+                        false, false));
+  ASSERT_FALSE(ReplaceMalloc::IsPHCAllocation(&stackVar, &phcInfo));
+  ASSERT_TRUE(PHCInfoEq(phcInfo, phc::AddrInfo::Kind::Unknown, nullptr, 0,
+                        false, false));
+
+  uint8_t* p = GetPHCAllocation(32);
+  if (!p) {
+    MOZ_CRASH("failed to get a PHC allocation");
+  }
+
+  // Test an in-use PHC allocation: first byte within it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::InUsePage, p, 32ul, true, false));
+  ASSERT_EQ(moz_malloc_usable_size(p), 32ul);
+  jemalloc_ptr_info(p, &jeInfo);
+
+  // Test an in-use PHC allocation: last byte within it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 31, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::InUsePage, p, 32ul, true, false));
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagLiveAlloc, p, 32, 0));
+  jemalloc_ptr_info(p + 31, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagLiveAlloc, p, 32, 0));
+
+  // Test an in-use PHC allocation: last byte before it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p - 1, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::InUsePage, p, 32ul, true, false));
+  jemalloc_ptr_info(p - 1, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: first byte on its allocation page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 32 - kPageSize, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::InUsePage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 32 - kPageSize, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: first byte in the following guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 32, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 32, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: last byte in the lower half of the
+  // following guard page.
+  ASSERT_TRUE(
+      ReplaceMalloc::IsPHCAllocation(p + 32 + (kPageSize / 2 - 1), &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 32 + (kPageSize / 2 - 1), &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: last byte in the preceding guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 31 - kPageSize, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 31 - kPageSize, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test an in-use PHC allocation: first byte in the upper half of the
+  // preceding guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(
+      p + 31 - kPageSize - (kPageSize / 2 - 1), &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, false));
+  jemalloc_ptr_info(p + 31 - kPageSize - (kPageSize / 2 - 1), &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  free(p);
+
+  // Test a freed PHC allocation: first byte within it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::FreedPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagFreedAlloc, p, 32, 0));
+
+  // Test a freed PHC allocation: last byte within it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 31, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::FreedPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 31, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagFreedAlloc, p, 32, 0));
+
+  // Test a freed PHC allocation: last byte before it.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p - 1, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::FreedPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p - 1, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: first byte on its allocation page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 32 - kPageSize, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::FreedPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 32 - kPageSize, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: first byte in the following guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 32, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 32, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: last byte in the lower half of the following
+  // guard page.
+  ASSERT_TRUE(
+      ReplaceMalloc::IsPHCAllocation(p + 32 + (kPageSize / 2 - 1), &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 32 + (kPageSize / 2 - 1), &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: last byte in the preceding guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p + 31 - kPageSize, &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 31 - kPageSize, &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // Test a freed PHC allocation: first byte in the upper half of the preceding
+  // guard page.
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(
+      p + 31 - kPageSize - (kPageSize / 2 - 1), &phcInfo));
+  ASSERT_TRUE(
+      PHCInfoEq(phcInfo, phc::AddrInfo::Kind::GuardPage, p, 32ul, true, true));
+  jemalloc_ptr_info(p + 31 - kPageSize - (kPageSize / 2 - 1), &jeInfo);
+  ASSERT_TRUE(JeInfoEq(jeInfo, TagUnknown, nullptr, 0, 0));
+
+  // There are no tests for `mKind == NeverAllocatedPage` because it's not
+  // possible to reliably get ahold of such a page.
+}
+
+TEST(PHC, TestPHCDisabling)
+{
+  uint8_t* p = GetPHCAllocation(32);
+  uint8_t* q = GetPHCAllocation(32);
+  if (!p || !q) {
+    MOZ_CRASH("failed to get a PHC allocation");
+  }
+
+  ASSERT_TRUE(ReplaceMalloc::IsPHCEnabledOnCurrentThread());
+  ReplaceMalloc::DisablePHCOnCurrentThread();
+  ASSERT_FALSE(ReplaceMalloc::IsPHCEnabledOnCurrentThread());
+
+  // Test realloc() on a PHC allocation while PHC is disabled on the thread.
+  uint8_t* p2 = (uint8_t*)realloc(p, 128);
+  // The small realloc is fulfilled within the same page, but it does move.
+  ASSERT_TRUE(p2 == p - 96);
+  ASSERT_TRUE(ReplaceMalloc::IsPHCAllocation(p2, nullptr));
+  uint8_t* p3 = (uint8_t*)realloc(p2, 8192);
+  // The big realloc is not in-place, and the result is not a PHC allocation.
+  ASSERT_TRUE(p3 != p2);
+  ASSERT_FALSE(ReplaceMalloc::IsPHCAllocation(p3, nullptr));
+  free(p3);
+
+  // Test free() on a PHC allocation while PHC is disabled on the thread.
+  free(q);
+
+  // These must not be PHC allocations.
+  uint8_t* r = GetPHCAllocation(32);  // This will fail.
+  ASSERT_FALSE(!!r);
+
+  ReplaceMalloc::ReenablePHCOnCurrentThread();
+  ASSERT_TRUE(ReplaceMalloc::IsPHCEnabledOnCurrentThread());
+}
diff --git a/memory/replace/phc/test/gtest/moz.build b/memory/replace/phc/test/gtest/moz.build
new file mode 100644
index 0000000000..82ccaaf9c6
--- /dev/null
+++ b/memory/replace/phc/test/gtest/moz.build
@@ -0,0 +1,15 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+UNIFIED_SOURCES += [
+    "TestPHC.cpp",
+]
+
+LOCAL_INCLUDES += [
+    "../../",
+]
+
+FINAL_LIBRARY = "xul-gtest"
diff --git a/memory/replace/phc/test/moz.build b/memory/replace/phc/test/moz.build
new file mode 100644
index 0000000000..8208ae849d
--- /dev/null
+++ b/memory/replace/phc/test/moz.build
@@ -0,0 +1,9 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+# The gtests won't work in a SpiderMonkey-only build.
+if CONFIG["MOZ_WIDGET_TOOLKIT"]:
+    TEST_DIRS += ["gtest"]
diff --git a/memory/volatile/VolatileBuffer.h b/memory/volatile/VolatileBuffer.h
new file mode 100644
index 0000000000..badd7f3cec
--- /dev/null
+++ b/memory/volatile/VolatileBuffer.h
@@ -0,0 +1,161 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifndef mozalloc_VolatileBuffer_h
+#define mozalloc_VolatileBuffer_h
+
+#include "mozilla/mozalloc.h"
+#include "mozilla/Mutex.h"
+#include "mozilla/RefPtr.h"
+#include "mozilla/MemoryReporting.h"
+#include "mozilla/RefCounted.h"
+
+/* VolatileBuffer
+ *
+ * This class represents a piece of memory that can potentially be reclaimed
+ * by the OS when not in use. As long as there are one or more
+ * VolatileBufferPtrs holding on to a VolatileBuffer, the memory will remain
+ * available. However, when there are no VolatileBufferPtrs holding a
+ * VolatileBuffer, the OS can purge the pages if it wants to. The OS can make
+ * better decisions about what pages to purge than we can.
+ *
+ * VolatileBuffers may not always be volatile - if the allocation is too small,
+ * or if the OS doesn't support the feature, or if the OS doesn't want to,
+ * the buffer will be allocated on heap.
+ *
+ * VolatileBuffer allocations are fallible. They are intended for uses where
+ * one may allocate large buffers for caching data. Init() must be called
+ * exactly once.
+ *
+ * After getting a reference to VolatileBuffer using VolatileBufferPtr,
+ * WasPurged() can be used to check if the OS purged any pages in the buffer.
+ * The OS cannot purge a buffer immediately after a VolatileBuffer is
+ * initialized. At least one VolatileBufferPtr must be created before the
+ * buffer can be purged, so the first use of VolatileBufferPtr does not need
+ * to check WasPurged().
+ *
+ * When a buffer is purged, some or all of the buffer is zeroed out. This
+ * API cannot tell which parts of the buffer were lost.
+ *
+ * VolatileBuffer and VolatileBufferPtr are threadsafe.
+ */
+
+namespace mozilla {
+
+class VolatileBuffer {
+  friend class VolatileBufferPtr_base;
+
+ public:
+  MOZ_DECLARE_REFCOUNTED_TYPENAME(VolatileBuffer)
+  NS_INLINE_DECL_THREADSAFE_REFCOUNTING(VolatileBuffer)
+
+  VolatileBuffer();
+
+  /* aAlignment must be a multiple of the pointer size */
+  bool Init(size_t aSize, size_t aAlignment = sizeof(void*));
+
+  size_t HeapSizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const;
+  size_t NonHeapSizeOfExcludingThis() const;
+  bool OnHeap() const;
+
+ protected:
+  bool Lock(void** aBuf);
+  void Unlock();
+
+ private:
+  ~VolatileBuffer();
+
+  /**
+   * Protects mLockCount, mFirstLock, and changes to the volatility of our
+   * buffer.  Other member variables are read-only except in Init() and the
+   * destructor.
+   */
+  Mutex mMutex MOZ_UNANNOTATED;
+
+  void* mBuf;
+  size_t mSize;
+  int mLockCount;
+#if defined(ANDROID)
+  int mFd;
+#elif defined(XP_DARWIN)
+  bool mHeap;
+#elif defined(XP_WIN)
+  bool mHeap;
+  bool mFirstLock;
+#endif
+};
+
+class VolatileBufferPtr_base {
+ public:
+  explicit VolatileBufferPtr_base(VolatileBuffer* vbuf)
+      : mVBuf(vbuf), mMapping(nullptr), mPurged(false) {
+    Lock();
+  }
+
+  ~VolatileBufferPtr_base() { Unlock(); }
+
+  bool WasBufferPurged() const { return mPurged; }
+
+ protected:
+  RefPtr<VolatileBuffer> mVBuf;
+  void* mMapping;
+
+  void Set(VolatileBuffer* vbuf) {
+    Unlock();
+    mVBuf = vbuf;
+    Lock();
+  }
+
+ private:
+  bool mPurged;
+
+  void Lock() {
+    if (mVBuf) {
+      mPurged = !mVBuf->Lock(&mMapping);
+    } else {
+      mMapping = nullptr;
+      mPurged = false;
+    }
+  }
+
+  void Unlock() {
+    if (mVBuf) {
+      mVBuf->Unlock();
+    }
+  }
+};
+
+template <class T>
+class VolatileBufferPtr : public VolatileBufferPtr_base {
+ public:
+  explicit VolatileBufferPtr(VolatileBuffer* vbuf)
+      : VolatileBufferPtr_base(vbuf) {}
+  VolatileBufferPtr() : VolatileBufferPtr_base(nullptr) {}
+
+  VolatileBufferPtr(VolatileBufferPtr&& aOther)
+      : VolatileBufferPtr_base(aOther.mVBuf) {
+    aOther.Set(nullptr);
+  }
+
+  operator T*() const { return (T*)mMapping; }
+
+  VolatileBufferPtr& operator=(VolatileBuffer* aVBuf) {
+    Set(aVBuf);
+    return *this;
+  }
+
+  VolatileBufferPtr& operator=(VolatileBufferPtr&& aOther) {
+    MOZ_ASSERT(this != &aOther, "Self-moves are prohibited");
+    Set(aOther.mVBuf);
+    aOther.Set(nullptr);
+    return *this;
+  }
+
+ private:
+  VolatileBufferPtr(VolatileBufferPtr const& vbufptr) = delete;
+};
+
+}  // namespace mozilla
+
+#endif /* mozalloc_VolatileBuffer_h */
diff --git a/memory/volatile/VolatileBufferAshmem.cpp b/memory/volatile/VolatileBufferAshmem.cpp
new file mode 100644
index 0000000000..99a0d1307f
--- /dev/null
+++ b/memory/volatile/VolatileBufferAshmem.cpp
@@ -0,0 +1,120 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "VolatileBuffer.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/mozalloc.h"
+
+#include <fcntl.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include "mozilla/Ashmem.h"
+
+#ifdef MOZ_MEMORY
+extern "C" int posix_memalign(void** memptr, size_t alignment, size_t size);
+#endif
+
+#define MIN_VOLATILE_ALLOC_SIZE 8192
+
+namespace mozilla {
+
+VolatileBuffer::VolatileBuffer()
+    : mMutex("VolatileBuffer"),
+      mBuf(nullptr),
+      mSize(0),
+      mLockCount(0),
+      mFd(-1) {}
+
+bool VolatileBuffer::Init(size_t aSize, size_t aAlignment) {
+  MOZ_ASSERT(!mSize && !mBuf, "Init called twice");
+  MOZ_ASSERT(!(aAlignment % sizeof(void*)),
+             "Alignment must be multiple of pointer size");
+
+  mSize = aSize;
+  if (aSize < MIN_VOLATILE_ALLOC_SIZE) {
+    goto heap_alloc;
+  }
+
+  mFd = mozilla::android::ashmem_create(nullptr, mSize);
+  if (mFd < 0) {
+    goto heap_alloc;
+  }
+
+  mBuf = mmap(nullptr, mSize, PROT_READ | PROT_WRITE, MAP_SHARED, mFd, 0);
+  if (mBuf != MAP_FAILED) {
+    return true;
+  }
+
+heap_alloc:
+  mBuf = nullptr;
+  if (mFd >= 0) {
+    close(mFd);
+    mFd = -1;
+  }
+
+#ifdef MOZ_MEMORY
+  posix_memalign(&mBuf, aAlignment, aSize);
+#else
+  mBuf = memalign(aAlignment, aSize);
+#endif
+  return !!mBuf;
+}
+
+VolatileBuffer::~VolatileBuffer() {
+  MOZ_ASSERT(mLockCount == 0, "Being destroyed with non-zero lock count?");
+
+  if (OnHeap()) {
+    free(mBuf);
+  } else {
+    munmap(mBuf, mSize);
+    close(mFd);
+  }
+}
+
+bool VolatileBuffer::Lock(void** aBuf) {
+  MutexAutoLock lock(mMutex);
+
+  MOZ_ASSERT(mBuf, "Attempting to lock an uninitialized VolatileBuffer");
+
+  *aBuf = mBuf;
+  if (++mLockCount > 1 || OnHeap()) {
+    return true;
+  }
+
+  // Zero offset and zero length means we want to pin/unpin the entire thing.
+  struct ashmem_pin pin = {0, 0};
+  return ioctl(mFd, ASHMEM_PIN, &pin) == ASHMEM_NOT_PURGED;
+}
+
+void VolatileBuffer::Unlock() {
+  MutexAutoLock lock(mMutex);
+
+  MOZ_ASSERT(mLockCount > 0, "VolatileBuffer unlocked too many times!");
+  if (--mLockCount || OnHeap()) {
+    return;
+  }
+
+  struct ashmem_pin pin = {0, 0};
+  ioctl(mFd, ASHMEM_UNPIN, &pin);
+}
+
+bool VolatileBuffer::OnHeap() const { return mFd < 0; }
+
+size_t VolatileBuffer::HeapSizeOfExcludingThis(
+    MallocSizeOf aMallocSizeOf) const {
+  return OnHeap() ? aMallocSizeOf(mBuf) : 0;
+}
+
+size_t VolatileBuffer::NonHeapSizeOfExcludingThis() const {
+  if (OnHeap()) {
+    return 0;
+  }
+
+  return (mSize + (PAGE_SIZE - 1)) & PAGE_MASK;
+}
+
+}  // namespace mozilla
diff --git a/memory/volatile/VolatileBufferFallback.cpp b/memory/volatile/VolatileBufferFallback.cpp
new file mode 100644
index 0000000000..d4311788ff
--- /dev/null
+++ b/memory/volatile/VolatileBufferFallback.cpp
@@ -0,0 +1,67 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "VolatileBuffer.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/mozalloc.h"
+
+#ifdef MOZ_MEMORY
+int posix_memalign(void** memptr, size_t alignment, size_t size);
+#endif
+
+namespace mozilla {
+
+VolatileBuffer::VolatileBuffer()
+    : mMutex("VolatileBuffer"), mBuf(nullptr), mSize(0), mLockCount(0) {}
+
+bool VolatileBuffer::Init(size_t aSize, size_t aAlignment) {
+  MOZ_ASSERT(!mSize && !mBuf, "Init called twice");
+  MOZ_ASSERT(!(aAlignment % sizeof(void*)),
+             "Alignment must be multiple of pointer size");
+
+  mSize = aSize;
+#if defined(MOZ_MEMORY) || defined(HAVE_POSIX_MEMALIGN)
+  if (posix_memalign(&mBuf, aAlignment, aSize) != 0) {
+    return false;
+  }
+#else
+#  error "No memalign implementation found"
+#endif
+  return !!mBuf;
+}
+
+VolatileBuffer::~VolatileBuffer() {
+  MOZ_ASSERT(mLockCount == 0, "Being destroyed with non-zero lock count?");
+
+  free(mBuf);
+}
+
+bool VolatileBuffer::Lock(void** aBuf) {
+  MutexAutoLock lock(mMutex);
+
+  MOZ_ASSERT(mBuf, "Attempting to lock an uninitialized VolatileBuffer");
+
+  *aBuf = mBuf;
+  mLockCount++;
+
+  return true;
+}
+
+void VolatileBuffer::Unlock() {
+  MutexAutoLock lock(mMutex);
+
+  mLockCount--;
+  MOZ_ASSERT(mLockCount >= 0, "VolatileBuffer unlocked too many times!");
+}
+
+bool VolatileBuffer::OnHeap() const { return true; }
+
+size_t VolatileBuffer::HeapSizeOfExcludingThis(
+    MallocSizeOf aMallocSizeOf) const {
+  return aMallocSizeOf(mBuf);
+}
+
+size_t VolatileBuffer::NonHeapSizeOfExcludingThis() const { return 0; }
+
+}  // namespace mozilla
diff --git a/memory/volatile/VolatileBufferOSX.cpp b/memory/volatile/VolatileBufferOSX.cpp
new file mode 100644
index 0000000000..c6299d1db0
--- /dev/null
+++ b/memory/volatile/VolatileBufferOSX.cpp
@@ -0,0 +1,105 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "VolatileBuffer.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/DebugOnly.h"
+#include "mozilla/mozalloc.h"
+
+#include <mach/mach.h>
+#include <sys/mman.h>
+#include <unistd.h>
+
+#define MIN_VOLATILE_ALLOC_SIZE 8192
+
+namespace mozilla {
+
+VolatileBuffer::VolatileBuffer()
+    : mMutex("VolatileBuffer"),
+      mBuf(nullptr),
+      mSize(0),
+      mLockCount(0),
+      mHeap(false) {}
+
+bool VolatileBuffer::Init(size_t aSize, size_t aAlignment) {
+  MOZ_ASSERT(!mSize && !mBuf, "Init called twice");
+  MOZ_ASSERT(!(aAlignment % sizeof(void*)),
+             "Alignment must be multiple of pointer size");
+
+  mSize = aSize;
+
+  kern_return_t ret = 0;
+  if (aSize < MIN_VOLATILE_ALLOC_SIZE) {
+    goto heap_alloc;
+  }
+
+  ret = vm_allocate(mach_task_self(), (vm_address_t*)&mBuf, mSize,
+                    VM_FLAGS_PURGABLE | VM_FLAGS_ANYWHERE);
+  if (ret == KERN_SUCCESS) {
+    return true;
+  }
+
+heap_alloc:
+  (void)posix_memalign(&mBuf, aAlignment, aSize);
+  mHeap = true;
+  return !!mBuf;
+}
+
+VolatileBuffer::~VolatileBuffer() {
+  MOZ_ASSERT(mLockCount == 0, "Being destroyed with non-zero lock count?");
+
+  if (OnHeap()) {
+    free(mBuf);
+  } else {
+    vm_deallocate(mach_task_self(), (vm_address_t)mBuf, mSize);
+  }
+}
+
+bool VolatileBuffer::Lock(void** aBuf) {
+  MutexAutoLock lock(mMutex);
+
+  MOZ_ASSERT(mBuf, "Attempting to lock an uninitialized VolatileBuffer");
+
+  *aBuf = mBuf;
+  if (++mLockCount > 1 || OnHeap()) {
+    return true;
+  }
+
+  int state = VM_PURGABLE_NONVOLATILE;
+  kern_return_t ret = vm_purgable_control(mach_task_self(), (vm_address_t)mBuf,
+                                          VM_PURGABLE_SET_STATE, &state);
+  return ret == KERN_SUCCESS && !(state & VM_PURGABLE_EMPTY);
+}
+
+void VolatileBuffer::Unlock() {
+  MutexAutoLock lock(mMutex);
+
+  MOZ_ASSERT(mLockCount > 0, "VolatileBuffer unlocked too many times!");
+  if (--mLockCount || OnHeap()) {
+    return;
+  }
+
+  int state = VM_PURGABLE_VOLATILE | VM_VOLATILE_GROUP_DEFAULT;
+  DebugOnly<kern_return_t> ret = vm_purgable_control(
+      mach_task_self(), (vm_address_t)mBuf, VM_PURGABLE_SET_STATE, &state);
+  MOZ_ASSERT(ret == KERN_SUCCESS, "Failed to set buffer as purgable");
+}
+
+bool VolatileBuffer::OnHeap() const { return mHeap; }
+
+size_t VolatileBuffer::HeapSizeOfExcludingThis(
+    MallocSizeOf aMallocSizeOf) const {
+  return OnHeap() ? aMallocSizeOf(mBuf) : 0;
+}
+
+size_t VolatileBuffer::NonHeapSizeOfExcludingThis() const {
+  if (OnHeap()) {
+    return 0;
+  }
+
+  unsigned long pagemask = getpagesize() - 1;
+  return (mSize + pagemask) & ~pagemask;
+}
+
+}  // namespace mozilla
diff --git a/memory/volatile/VolatileBufferWindows.cpp b/memory/volatile/VolatileBufferWindows.cpp
new file mode 100644
index 0000000000..6615f89299
--- /dev/null
+++ b/memory/volatile/VolatileBufferWindows.cpp
@@ -0,0 +1,136 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "VolatileBuffer.h"
+#include "mozilla/Assertions.h"
+#include "mozilla/DebugOnly.h"
+#include "mozilla/mozalloc.h"
+#include "mozilla/WindowsVersion.h"
+
+#include <windows.h>
+
+#ifdef MOZ_MEMORY
+extern "C" int posix_memalign(void** memptr, size_t alignment, size_t size);
+#endif
+
+#ifndef MEM_RESET_UNDO
+#  define MEM_RESET_UNDO 0x1000000
+#endif
+
+#define MIN_VOLATILE_ALLOC_SIZE 8192
+
+namespace mozilla {
+
+VolatileBuffer::VolatileBuffer()
+    : mMutex("VolatileBuffer"),
+      mBuf(nullptr),
+      mSize(0),
+      mLockCount(0),
+      mHeap(false),
+      mFirstLock(true) {}
+
+bool VolatileBuffer::Init(size_t aSize, size_t aAlignment) {
+  MOZ_ASSERT(!mSize && !mBuf, "Init called twice");
+  MOZ_ASSERT(!(aAlignment % sizeof(void*)),
+             "Alignment must be multiple of pointer size");
+
+  mSize = aSize;
+  if (aSize < MIN_VOLATILE_ALLOC_SIZE) {
+    goto heap_alloc;
+  }
+
+  static bool sUndoSupported = IsWin8OrLater();
+  if (!sUndoSupported) {
+    goto heap_alloc;
+  }
+
+  mBuf = VirtualAllocEx(GetCurrentProcess(), nullptr, mSize,
+                        MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE);
+  if (mBuf) {
+    return true;
+  }
+
+heap_alloc:
+#ifdef MOZ_MEMORY
+  posix_memalign(&mBuf, aAlignment, aSize);
+#else
+  mBuf = _aligned_malloc(aSize, aAlignment);
+#endif
+  mHeap = true;
+  return !!mBuf;
+}
+
+VolatileBuffer::~VolatileBuffer() {
+  MOZ_ASSERT(mLockCount == 0, "Being destroyed with non-zero lock count?");
+
+  if (OnHeap()) {
+#ifdef MOZ_MEMORY
+    free(mBuf);
+#else
+    _aligned_free(mBuf);
+#endif
+  } else {
+    VirtualFreeEx(GetCurrentProcess(), mBuf, 0, MEM_RELEASE);
+  }
+}
+
+bool VolatileBuffer::Lock(void** aBuf) {
+  MutexAutoLock lock(mMutex);
+
+  MOZ_ASSERT(mBuf, "Attempting to lock an uninitialized VolatileBuffer");
+
+  *aBuf = mBuf;
+  if (++mLockCount > 1 || OnHeap()) {
+    return true;
+  }
+
+  // MEM_RESET_UNDO's behavior is undefined when called on memory that
+  // hasn't been MEM_RESET.
+  if (mFirstLock) {
+    mFirstLock = false;
+    return true;
+  }
+
+  void* addr = VirtualAllocEx(GetCurrentProcess(), mBuf, mSize, MEM_RESET_UNDO,
+                              PAGE_READWRITE);
+  return !!addr;
+}
+
+void VolatileBuffer::Unlock() {
+  MutexAutoLock lock(mMutex);
+
+  MOZ_ASSERT(mLockCount > 0, "VolatileBuffer unlocked too many times!");
+  if (--mLockCount || OnHeap()) {
+    return;
+  }
+
+  DebugOnly<void*> addr = VirtualAllocEx(GetCurrentProcess(), mBuf, mSize,
+                                         MEM_RESET, PAGE_READWRITE);
+  MOZ_ASSERT(addr, "Failed to MEM_RESET");
+}
+
+bool VolatileBuffer::OnHeap() const { return mHeap; }
+
+size_t VolatileBuffer::HeapSizeOfExcludingThis(
+    MallocSizeOf aMallocSizeOf) const {
+  if (OnHeap()) {
+#ifdef MOZ_MEMORY
+    return aMallocSizeOf(mBuf);
+#else
+    return mSize;
+#endif
+  }
+
+  return 0;
+}
+
+size_t VolatileBuffer::NonHeapSizeOfExcludingThis() const {
+  if (OnHeap()) {
+    return 0;
+  }
+
+  return (mSize + 4095) & ~4095;
+}
+
+}  // namespace mozilla
diff --git a/memory/volatile/moz.build b/memory/volatile/moz.build
new file mode 100644
index 0000000000..9cfe43e5d0
--- /dev/null
+++ b/memory/volatile/moz.build
@@ -0,0 +1,31 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+NoVisibilityFlags()
+
+EXPORTS.mozilla += [
+    "VolatileBuffer.h",
+]
+
+if CONFIG["OS_TARGET"] == "Android":
+    UNIFIED_SOURCES += [
+        "VolatileBufferAshmem.cpp",
+    ]
+elif CONFIG["OS_TARGET"] == "Darwin":
+    UNIFIED_SOURCES += [
+        "VolatileBufferOSX.cpp",
+    ]
+elif CONFIG["OS_TARGET"] == "WINNT":
+    UNIFIED_SOURCES += [
+        "VolatileBufferWindows.cpp",
+    ]
+else:
+    UNIFIED_SOURCES += [
+        "VolatileBufferFallback.cpp",
+    ]
+
+FINAL_LIBRARY = "xul"
+
+TEST_DIRS += ["tests"]
diff --git a/memory/volatile/tests/TestVolatileBuffer.cpp b/memory/volatile/tests/TestVolatileBuffer.cpp
new file mode 100644
index 0000000000..e9e9699ec6
--- /dev/null
+++ b/memory/volatile/tests/TestVolatileBuffer.cpp
@@ -0,0 +1,100 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "gtest/gtest.h"
+#include "mozilla/VolatileBuffer.h"
+#include <string.h>
+
+#if defined(ANDROID)
+#  include <fcntl.h>
+#  include <linux/ashmem.h>
+#  include <sys/ioctl.h>
+#  include <sys/stat.h>
+#  include <sys/types.h>
+#elif defined(XP_DARWIN)
+#  include <mach/mach.h>
+#endif
+
+using namespace mozilla;
+
+TEST(VolatileBufferTest, HeapVolatileBuffersWork)
+{
+  RefPtr<VolatileBuffer> heapbuf = new VolatileBuffer();
+
+  ASSERT_TRUE(heapbuf)
+  << "Failed to create VolatileBuffer";
+  ASSERT_TRUE(heapbuf->Init(512))
+  << "Failed to initialize VolatileBuffer";
+
+  VolatileBufferPtr<char> ptr(heapbuf);
+
+  EXPECT_FALSE(ptr.WasBufferPurged())
+      << "Buffer should not be purged immediately after initialization";
+  EXPECT_TRUE(ptr) << "Couldn't get pointer from VolatileBufferPtr";
+}
+
+TEST(VolatileBufferTest, RealVolatileBuffersWork)
+{
+  RefPtr<VolatileBuffer> buf = new VolatileBuffer();
+
+  ASSERT_TRUE(buf)
+  << "Failed to create VolatileBuffer";
+  ASSERT_TRUE(buf->Init(16384))
+  << "Failed to initialize VolatileBuffer";
+
+  const char teststr[] = "foobar";
+
+  {
+    VolatileBufferPtr<char> ptr(buf);
+
+    EXPECT_FALSE(ptr.WasBufferPurged())
+        << "Buffer should not be purged immediately after initialization";
+    EXPECT_TRUE(ptr) << "Couldn't get pointer from VolatileBufferPtr";
+
+    {
+      VolatileBufferPtr<char> ptr2(buf);
+
+      EXPECT_FALSE(ptr.WasBufferPurged())
+          << "Failed to lock buffer again while currently locked";
+      ASSERT_TRUE(ptr2)
+      << "Didn't get a pointer on the second lock";
+
+      strcpy(ptr2, teststr);
+    }
+  }
+
+  {
+    VolatileBufferPtr<char> ptr(buf);
+
+    EXPECT_FALSE(ptr.WasBufferPurged())
+        << "Buffer was immediately purged after unlock";
+    EXPECT_STREQ(ptr, teststr) << "Buffer failed to retain data after unlock";
+  }
+
+  // Test purging if we know how to
+#if defined(XP_DARWIN)
+  int state;
+  vm_purgable_control(mach_task_self(), (vm_address_t)NULL,
+                      VM_PURGABLE_PURGE_ALL, &state);
+#else
+  return;
+#endif
+
+  EXPECT_GT(buf->NonHeapSizeOfExcludingThis(), 0ul)
+      << "Buffer should not be allocated on heap";
+
+  {
+    VolatileBufferPtr<char> ptr(buf);
+
+    EXPECT_TRUE(ptr.WasBufferPurged())
+        << "Buffer should not be unpurged after forced purge";
+    EXPECT_STRNE(ptr, teststr) << "Purge did not actually purge data";
+  }
+
+  {
+    VolatileBufferPtr<char> ptr(buf);
+
+    EXPECT_FALSE(ptr.WasBufferPurged()) << "Buffer still purged after lock";
+  }
+}
diff --git a/memory/volatile/tests/moz.build b/memory/volatile/tests/moz.build
new file mode 100644
index 0000000000..31201db9d4
--- /dev/null
+++ b/memory/volatile/tests/moz.build
@@ -0,0 +1,11 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+UNIFIED_SOURCES = [
+    "TestVolatileBuffer.cpp",
+]
+
+FINAL_LIBRARY = "xul-gtest"