Adding upstream version 86.0.1.upstream/86.0.1upstream

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

74 files changed, 35423 insertions, 0 deletions

diff --git a/mozglue/baseprofiler/core/EHABIStackWalk.cpp b/mozglue/baseprofiler/core/EHABIStackWalk.cpp
new file mode 100644
index 0000000000..0c2c855c9b
--- /dev/null
+++ b/mozglue/baseprofiler/core/EHABIStackWalk.cpp
@@ -0,0 +1,592 @@
+/* -*- 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 is an implementation of stack unwinding according to a subset
+ * of the ARM Exception Handling ABI, as described in:
+ *   http://infocenter.arm.com/help/topic/com.arm.doc.ihi0038a/IHI0038A_ehabi.pdf
+ *
+ * This handles only the ARM-defined "personality routines" (chapter
+ * 9), and don't track the value of FP registers, because profiling
+ * needs only chain of PC/SP values.
+ *
+ * Because the exception handling info may not be accurate for all
+ * possible places where an async signal could occur (e.g., in a
+ * prologue or epilogue), this bounds-checks all stack accesses.
+ *
+ * This file uses "struct" for structures in the exception tables and
+ * "class" otherwise.  We should avoid violating the C++11
+ * standard-layout rules in the former.
+ */
+
+#include "BaseProfiler.h"
+
+#include "EHABIStackWalk.h"
+
+#include "BaseProfilerSharedLibraries.h"
+#include "platform.h"
+
+#include "mozilla/Atomics.h"
+#include "mozilla/DebugOnly.h"
+#include "mozilla/EndianUtils.h"
+
+#include <algorithm>
+#include <elf.h>
+#include <stdint.h>
+#include <vector>
+#include <string>
+
+#ifndef PT_ARM_EXIDX
+#  define PT_ARM_EXIDX 0x70000001
+#endif
+
+namespace mozilla {
+namespace baseprofiler {
+
+struct PRel31 {
+  uint32_t mBits;
+  bool topBit() const { return mBits & 0x80000000; }
+  uint32_t value() const { return mBits & 0x7fffffff; }
+  int32_t offset() const { return (static_cast<int32_t>(mBits) << 1) >> 1; }
+  const void* compute() const {
+    return reinterpret_cast<const char*>(this) + offset();
+  }
+
+ private:
+  PRel31(const PRel31& copied) = delete;
+  PRel31() = delete;
+};
+
+struct EHEntry {
+  PRel31 startPC;
+  PRel31 exidx;
+
+ private:
+  EHEntry(const EHEntry& copied) = delete;
+  EHEntry() = delete;
+};
+
+class EHState {
+  // Note that any core register can be used as a "frame pointer" to
+  // influence the unwinding process, so this must track all of them.
+  uint32_t mRegs[16];
+
+ public:
+  bool unwind(const EHEntry* aEntry, const void* stackBase);
+  uint32_t& operator[](int i) { return mRegs[i]; }
+  const uint32_t& operator[](int i) const { return mRegs[i]; }
+  explicit EHState(const mcontext_t&);
+};
+
+enum { R_SP = 13, R_LR = 14, R_PC = 15 };
+
+class EHTable {
+  uint32_t mStartPC;
+  uint32_t mEndPC;
+  uint32_t mBaseAddress;
+  const EHEntry* mEntriesBegin;
+  const EHEntry* mEntriesEnd;
+  std::string mName;
+
+ public:
+  EHTable(const void* aELF, size_t aSize, const std::string& aName);
+  const EHEntry* lookup(uint32_t aPC) const;
+  bool isValid() const { return mEntriesEnd != mEntriesBegin; }
+  const std::string& name() const { return mName; }
+  uint32_t startPC() const { return mStartPC; }
+  uint32_t endPC() const { return mEndPC; }
+  uint32_t baseAddress() const { return mBaseAddress; }
+};
+
+class EHAddrSpace {
+  std::vector<uint32_t> mStarts;
+  std::vector<EHTable> mTables;
+  static Atomic<const EHAddrSpace*> sCurrent;
+
+ public:
+  explicit EHAddrSpace(const std::vector<EHTable>& aTables);
+  const EHTable* lookup(uint32_t aPC) const;
+  static void Update();
+  static const EHAddrSpace* Get();
+};
+
+void EHABIStackWalkInit() { EHAddrSpace::Update(); }
+
+size_t EHABIStackWalk(const mcontext_t& aContext, void* stackBase, void** aSPs,
+                      void** aPCs, const size_t aNumFrames) {
+  const EHAddrSpace* space = EHAddrSpace::Get();
+  EHState state(aContext);
+  size_t count = 0;
+
+  while (count < aNumFrames) {
+    uint32_t pc = state[R_PC], sp = state[R_SP];
+    aPCs[count] = reinterpret_cast<void*>(pc);
+    aSPs[count] = reinterpret_cast<void*>(sp);
+    count++;
+
+    if (!space) break;
+    // TODO: cache these lookups.  Binary-searching libxul is
+    // expensive (possibly more expensive than doing the actual
+    // unwind), and even a small cache should help.
+    const EHTable* table = space->lookup(pc);
+    if (!table) break;
+    const EHEntry* entry = table->lookup(pc);
+    if (!entry) break;
+    if (!state.unwind(entry, stackBase)) break;
+  }
+
+  return count;
+}
+
+class EHInterp {
+ public:
+  // Note that stackLimit is exclusive and stackBase is inclusive
+  // (i.e, stackLimit < SP <= stackBase), following the convention
+  // set by the AAPCS spec.
+  EHInterp(EHState& aState, const EHEntry* aEntry, uint32_t aStackLimit,
+           uint32_t aStackBase)
+      : mState(aState),
+        mStackLimit(aStackLimit),
+        mStackBase(aStackBase),
+        mNextWord(0),
+        mWordsLeft(0),
+        mFailed(false) {
+    const PRel31& exidx = aEntry->exidx;
+    uint32_t firstWord;
+
+    if (exidx.mBits == 1) {  // EXIDX_CANTUNWIND
+      mFailed = true;
+      return;
+    }
+    if (exidx.topBit()) {
+      firstWord = exidx.mBits;
+    } else {
+      mNextWord = reinterpret_cast<const uint32_t*>(exidx.compute());
+      firstWord = *mNextWord++;
+    }
+
+    switch (firstWord >> 24) {
+      case 0x80:  // short
+        mWord = firstWord << 8;
+        mBytesLeft = 3;
+        break;
+      case 0x81:
+      case 0x82:  // long; catch descriptor size ignored
+        mWord = firstWord << 16;
+        mBytesLeft = 2;
+        mWordsLeft = (firstWord >> 16) & 0xff;
+        break;
+      default:
+        // unknown personality
+        mFailed = true;
+    }
+  }
+
+  bool unwind();
+
+ private:
+  // TODO: GCC has been observed not CSEing repeated reads of
+  // mState[R_SP] with writes to mFailed between them, suggesting that
+  // it hasn't determined that they can't alias and is thus missing
+  // optimization opportunities.  So, we may want to flatten EHState
+  // into this class; this may also make the code simpler.
+  EHState& mState;
+  uint32_t mStackLimit;
+  uint32_t mStackBase;
+  const uint32_t* mNextWord;
+  uint32_t mWord;
+  uint8_t mWordsLeft;
+  uint8_t mBytesLeft;
+  bool mFailed;
+
+  enum {
+    I_ADDSP = 0x00,  // 0sxxxxxx (subtract if s)
+    M_ADDSP = 0x80,
+    I_POPMASK = 0x80,  // 1000iiii iiiiiiii (if any i set)
+    M_POPMASK = 0xf0,
+    I_MOVSP = 0x90,  // 1001nnnn
+    M_MOVSP = 0xf0,
+    I_POPN = 0xa0,  // 1010lnnn
+    M_POPN = 0xf0,
+    I_FINISH = 0xb0,    // 10110000
+    I_POPLO = 0xb1,     // 10110001 0000iiii (if any i set)
+    I_ADDSPBIG = 0xb2,  // 10110010 uleb128
+    I_POPFDX = 0xb3,    // 10110011 sssscccc
+    I_POPFDX8 = 0xb8,   // 10111nnn
+    M_POPFDX8 = 0xf8,
+    // "Intel Wireless MMX" extensions omitted.
+    I_POPFDD = 0xc8,  // 1100100h sssscccc
+    M_POPFDD = 0xfe,
+    I_POPFDD8 = 0xd0,  // 11010nnn
+    M_POPFDD8 = 0xf8
+  };
+
+  uint8_t next() {
+    if (mBytesLeft == 0) {
+      if (mWordsLeft == 0) {
+        return I_FINISH;
+      }
+      mWordsLeft--;
+      mWord = *mNextWord++;
+      mBytesLeft = 4;
+    }
+    mBytesLeft--;
+    mWord = (mWord << 8) | (mWord >> 24);  // rotate
+    return mWord;
+  }
+
+  uint32_t& vSP() { return mState[R_SP]; }
+  uint32_t* ptrSP() { return reinterpret_cast<uint32_t*>(vSP()); }
+
+  void checkStackBase() {
+    if (vSP() > mStackBase) mFailed = true;
+  }
+  void checkStackLimit() {
+    if (vSP() <= mStackLimit) mFailed = true;
+  }
+  void checkStackAlign() {
+    if ((vSP() & 3) != 0) mFailed = true;
+  }
+  void checkStack() {
+    checkStackBase();
+    checkStackLimit();
+    checkStackAlign();
+  }
+
+  void popRange(uint8_t first, uint8_t last, uint16_t mask) {
+    bool hasSP = false;
+    uint32_t tmpSP;
+    if (mask == 0) mFailed = true;
+    for (uint8_t r = first; r <= last; ++r) {
+      if (mask & 1) {
+        if (r == R_SP) {
+          hasSP = true;
+          tmpSP = *ptrSP();
+        } else
+          mState[r] = *ptrSP();
+        vSP() += 4;
+        checkStackBase();
+        if (mFailed) return;
+      }
+      mask >>= 1;
+    }
+    if (hasSP) {
+      vSP() = tmpSP;
+      checkStack();
+    }
+  }
+};
+
+bool EHState::unwind(const EHEntry* aEntry, const void* stackBasePtr) {
+  // The unwinding program cannot set SP to less than the initial value.
+  uint32_t stackLimit = mRegs[R_SP] - 4;
+  uint32_t stackBase = reinterpret_cast<uint32_t>(stackBasePtr);
+  EHInterp interp(*this, aEntry, stackLimit, stackBase);
+  return interp.unwind();
+}
+
+bool EHInterp::unwind() {
+  mState[R_PC] = 0;
+  checkStack();
+  while (!mFailed) {
+    uint8_t insn = next();
+#if DEBUG_EHABI_UNWIND
+    LOG("unwind insn = %02x", (unsigned)insn);
+#endif
+    // Try to put the common cases first.
+
+    // 00xxxxxx: vsp = vsp + (xxxxxx << 2) + 4
+    // 01xxxxxx: vsp = vsp - (xxxxxx << 2) - 4
+    if ((insn & M_ADDSP) == I_ADDSP) {
+      uint32_t offset = ((insn & 0x3f) << 2) + 4;
+      if (insn & 0x40) {
+        vSP() -= offset;
+        checkStackLimit();
+      } else {
+        vSP() += offset;
+        checkStackBase();
+      }
+      continue;
+    }
+
+    // 10100nnn: Pop r4-r[4+nnn]
+    // 10101nnn: Pop r4-r[4+nnn], r14
+    if ((insn & M_POPN) == I_POPN) {
+      uint8_t n = (insn & 0x07) + 1;
+      bool lr = insn & 0x08;
+      uint32_t* ptr = ptrSP();
+      vSP() += (n + (lr ? 1 : 0)) * 4;
+      checkStackBase();
+      for (uint8_t r = 4; r < 4 + n; ++r) mState[r] = *ptr++;
+      if (lr) mState[R_LR] = *ptr++;
+      continue;
+    }
+
+    // 1011000: Finish
+    if (insn == I_FINISH) {
+      if (mState[R_PC] == 0) {
+        mState[R_PC] = mState[R_LR];
+        // Non-standard change (bug 916106): Prevent the caller from
+        // re-using LR.  Since the caller is by definition not a leaf
+        // routine, it will have to restore LR from somewhere to
+        // return to its own caller, so we can safely zero it here.
+        // This makes a difference only if an error in unwinding
+        // (e.g., caused by starting from within a prologue/epilogue)
+        // causes us to load a pointer to a leaf routine as LR; if we
+        // don't do something, we'll go into an infinite loop of
+        // "returning" to that same function.
+        mState[R_LR] = 0;
+      }
+      return true;
+    }
+
+    // 1001nnnn: Set vsp = r[nnnn]
+    if ((insn & M_MOVSP) == I_MOVSP) {
+      vSP() = mState[insn & 0x0f];
+      checkStack();
+      continue;
+    }
+
+    // 11001000 sssscccc: Pop VFP regs D[16+ssss]-D[16+ssss+cccc] (as FLDMFDD)
+    // 11001001 sssscccc: Pop VFP regs D[ssss]-D[ssss+cccc] (as FLDMFDD)
+    if ((insn & M_POPFDD) == I_POPFDD) {
+      uint8_t n = (next() & 0x0f) + 1;
+      // Note: if the 16+ssss+cccc > 31, the encoding is reserved.
+      // As the space is currently unused, we don't try to check.
+      vSP() += 8 * n;
+      checkStackBase();
+      continue;
+    }
+
+    // 11010nnn: Pop VFP regs D[8]-D[8+nnn] (as FLDMFDD)
+    if ((insn & M_POPFDD8) == I_POPFDD8) {
+      uint8_t n = (insn & 0x07) + 1;
+      vSP() += 8 * n;
+      checkStackBase();
+      continue;
+    }
+
+    // 10110010 uleb128: vsp = vsp + 0x204 + (uleb128 << 2)
+    if (insn == I_ADDSPBIG) {
+      uint32_t acc = 0;
+      uint8_t shift = 0;
+      uint8_t byte;
+      do {
+        if (shift >= 32) return false;
+        byte = next();
+        acc |= (byte & 0x7f) << shift;
+        shift += 7;
+      } while (byte & 0x80);
+      uint32_t offset = 0x204 + (acc << 2);
+      // The calculations above could have overflowed.
+      // But the one we care about is this:
+      if (vSP() + offset < vSP()) mFailed = true;
+      vSP() += offset;
+      // ...so that this is the only other check needed:
+      checkStackBase();
+      continue;
+    }
+
+    // 1000iiii iiiiiiii (i not all 0): Pop under masks {r15-r12}, {r11-r4}
+    if ((insn & M_POPMASK) == I_POPMASK) {
+      popRange(4, 15, ((insn & 0x0f) << 8) | next());
+      continue;
+    }
+
+    // 1011001 0000iiii (i not all 0): Pop under mask {r3-r0}
+    if (insn == I_POPLO) {
+      popRange(0, 3, next() & 0x0f);
+      continue;
+    }
+
+    // 10110011 sssscccc: Pop VFP regs D[ssss]-D[ssss+cccc] (as FLDMFDX)
+    if (insn == I_POPFDX) {
+      uint8_t n = (next() & 0x0f) + 1;
+      vSP() += 8 * n + 4;
+      checkStackBase();
+      continue;
+    }
+
+    // 10111nnn: Pop VFP regs D[8]-D[8+nnn] (as FLDMFDX)
+    if ((insn & M_POPFDX8) == I_POPFDX8) {
+      uint8_t n = (insn & 0x07) + 1;
+      vSP() += 8 * n + 4;
+      checkStackBase();
+      continue;
+    }
+
+    // unhandled instruction
+#ifdef DEBUG_EHABI_UNWIND
+    LOG("Unhandled EHABI instruction 0x%02x", insn);
+#endif
+    mFailed = true;
+  }
+  return false;
+}
+
+bool operator<(const EHTable& lhs, const EHTable& rhs) {
+  return lhs.startPC() < rhs.startPC();
+}
+
+// Async signal unsafe.
+EHAddrSpace::EHAddrSpace(const std::vector<EHTable>& aTables)
+    : mTables(aTables) {
+  std::sort(mTables.begin(), mTables.end());
+  DebugOnly<uint32_t> lastEnd = 0;
+  for (std::vector<EHTable>::iterator i = mTables.begin(); i != mTables.end();
+       ++i) {
+    MOZ_ASSERT(i->startPC() >= lastEnd);
+    mStarts.push_back(i->startPC());
+    lastEnd = i->endPC();
+  }
+}
+
+const EHTable* EHAddrSpace::lookup(uint32_t aPC) const {
+  ptrdiff_t i = (std::upper_bound(mStarts.begin(), mStarts.end(), aPC) -
+                 mStarts.begin()) -
+                1;
+
+  if (i < 0 || aPC >= mTables[i].endPC()) return 0;
+  return &mTables[i];
+}
+
+const EHEntry* EHTable::lookup(uint32_t aPC) const {
+  MOZ_ASSERT(aPC >= mStartPC);
+  if (aPC >= mEndPC) return nullptr;
+
+  const EHEntry* begin = mEntriesBegin;
+  const EHEntry* end = mEntriesEnd;
+  MOZ_ASSERT(begin < end);
+  if (aPC < reinterpret_cast<uint32_t>(begin->startPC.compute()))
+    return nullptr;
+
+  while (end - begin > 1) {
+#ifdef EHABI_UNWIND_MORE_ASSERTS
+    if ((end - 1)->startPC.compute() < begin->startPC.compute()) {
+      MOZ_CRASH("unsorted exidx");
+    }
+#endif
+    const EHEntry* mid = begin + (end - begin) / 2;
+    if (aPC < reinterpret_cast<uint32_t>(mid->startPC.compute()))
+      end = mid;
+    else
+      begin = mid;
+  }
+  return begin;
+}
+
+#if MOZ_LITTLE_ENDIAN()
+static const unsigned char hostEndian = ELFDATA2LSB;
+#elif MOZ_BIG_ENDIAN()
+static const unsigned char hostEndian = ELFDATA2MSB;
+#else
+#  error "No endian?"
+#endif
+
+// Async signal unsafe: std::vector::reserve, std::string copy ctor.
+EHTable::EHTable(const void* aELF, size_t aSize, const std::string& aName)
+    : mStartPC(~0),  // largest uint32_t
+      mEndPC(0),
+      mEntriesBegin(nullptr),
+      mEntriesEnd(nullptr),
+      mName(aName) {
+  const uint32_t fileHeaderAddr = reinterpret_cast<uint32_t>(aELF);
+
+  if (aSize < sizeof(Elf32_Ehdr)) return;
+
+  const Elf32_Ehdr& file = *(reinterpret_cast<Elf32_Ehdr*>(fileHeaderAddr));
+  if (memcmp(&file.e_ident[EI_MAG0], ELFMAG, SELFMAG) != 0 ||
+      file.e_ident[EI_CLASS] != ELFCLASS32 ||
+      file.e_ident[EI_DATA] != hostEndian ||
+      file.e_ident[EI_VERSION] != EV_CURRENT || file.e_machine != EM_ARM ||
+      file.e_version != EV_CURRENT)
+    // e_flags?
+    return;
+
+  MOZ_ASSERT(file.e_phoff + file.e_phnum * file.e_phentsize <= aSize);
+  const Elf32_Phdr *exidxHdr = 0, *zeroHdr = 0;
+  for (unsigned i = 0; i < file.e_phnum; ++i) {
+    const Elf32_Phdr& phdr = *(reinterpret_cast<Elf32_Phdr*>(
+        fileHeaderAddr + file.e_phoff + i * file.e_phentsize));
+    if (phdr.p_type == PT_ARM_EXIDX) {
+      exidxHdr = &phdr;
+    } else if (phdr.p_type == PT_LOAD) {
+      if (phdr.p_offset == 0) {
+        zeroHdr = &phdr;
+      }
+      if (phdr.p_flags & PF_X) {
+        mStartPC = std::min(mStartPC, phdr.p_vaddr);
+        mEndPC = std::max(mEndPC, phdr.p_vaddr + phdr.p_memsz);
+      }
+    }
+  }
+  if (!exidxHdr) return;
+  if (!zeroHdr) return;
+  mBaseAddress = fileHeaderAddr - zeroHdr->p_vaddr;
+  mStartPC += mBaseAddress;
+  mEndPC += mBaseAddress;
+  mEntriesBegin =
+      reinterpret_cast<const EHEntry*>(mBaseAddress + exidxHdr->p_vaddr);
+  mEntriesEnd = reinterpret_cast<const EHEntry*>(
+      mBaseAddress + exidxHdr->p_vaddr + exidxHdr->p_memsz);
+}
+
+Atomic<const EHAddrSpace*> EHAddrSpace::sCurrent(nullptr);
+
+// Async signal safe; can fail if Update() hasn't returned yet.
+const EHAddrSpace* EHAddrSpace::Get() { return sCurrent; }
+
+// Collect unwinding information from loaded objects.  Calls after the
+// first have no effect.  Async signal unsafe.
+void EHAddrSpace::Update() {
+  const EHAddrSpace* space = sCurrent;
+  if (space) return;
+
+  SharedLibraryInfo info = SharedLibraryInfo::GetInfoForSelf();
+  std::vector<EHTable> tables;
+
+  for (size_t i = 0; i < info.GetSize(); ++i) {
+    const SharedLibrary& lib = info.GetEntry(i);
+    // FIXME: This isn't correct if the start address isn't p_offset 0, because
+    // the start address will not point at the file header. But this is worked
+    // around by magic number checks in the EHTable constructor.
+    EHTable tab(reinterpret_cast<const void*>(lib.GetStart()),
+                lib.GetEnd() - lib.GetStart(), lib.GetDebugPath());
+    if (tab.isValid()) tables.push_back(tab);
+  }
+  space = new EHAddrSpace(tables);
+
+  if (!sCurrent.compareExchange(nullptr, space)) {
+    delete space;
+    space = sCurrent;
+  }
+}
+
+EHState::EHState(const mcontext_t& context) {
+#ifdef linux
+  mRegs[0] = context.arm_r0;
+  mRegs[1] = context.arm_r1;
+  mRegs[2] = context.arm_r2;
+  mRegs[3] = context.arm_r3;
+  mRegs[4] = context.arm_r4;
+  mRegs[5] = context.arm_r5;
+  mRegs[6] = context.arm_r6;
+  mRegs[7] = context.arm_r7;
+  mRegs[8] = context.arm_r8;
+  mRegs[9] = context.arm_r9;
+  mRegs[10] = context.arm_r10;
+  mRegs[11] = context.arm_fp;
+  mRegs[12] = context.arm_ip;
+  mRegs[13] = context.arm_sp;
+  mRegs[14] = context.arm_lr;
+  mRegs[15] = context.arm_pc;
+#else
+#  error "Unhandled OS for ARM EHABI unwinding"
+#endif
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/EHABIStackWalk.h b/mozglue/baseprofiler/core/EHABIStackWalk.h
new file mode 100644
index 0000000000..d5f4edc0d7
--- /dev/null
+++ b/mozglue/baseprofiler/core/EHABIStackWalk.h
@@ -0,0 +1,30 @@
+/* -*- 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 is an implementation of stack unwinding according to a subset
+ * of the ARM Exception Handling ABI; see the comment at the top of
+ * the .cpp file for details.
+ */
+
+#ifndef mozilla_EHABIStackWalk_h__
+#define mozilla_EHABIStackWalk_h__
+
+#include <stddef.h>
+#include <ucontext.h>
+
+namespace mozilla {
+namespace baseprofiler {
+
+void EHABIStackWalkInit();
+
+size_t EHABIStackWalk(const mcontext_t& aContext, void* stackBase, void** aSPs,
+                      void** aPCs, size_t aNumFrames);
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif
diff --git a/mozglue/baseprofiler/core/PageInformation.cpp b/mozglue/baseprofiler/core/PageInformation.cpp
new file mode 100644
index 0000000000..7ce47e86d7
--- /dev/null
+++ b/mozglue/baseprofiler/core/PageInformation.cpp
@@ -0,0 +1,50 @@
+/* -*- Mode: C++; tab-width: 2; 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 "PageInformation.h"
+
+#include "BaseProfiler.h"
+
+#include "mozilla/BaseProfileJSONWriter.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+PageInformation::PageInformation(uint64_t aBrowsingContextID,
+                                 uint64_t aInnerWindowID,
+                                 const std::string& aUrl,
+                                 uint64_t aEmbedderInnerWindowID)
+    : mBrowsingContextID(aBrowsingContextID),
+      mInnerWindowID(aInnerWindowID),
+      mUrl(aUrl),
+      mEmbedderInnerWindowID(aEmbedderInnerWindowID),
+      mRefCnt(0) {}
+
+bool PageInformation::Equals(PageInformation* aOtherPageInfo) const {
+  // It's enough to check inner window IDs because they are unique for each
+  // page. Therefore, we don't have to check browsing context ID or url.
+  return InnerWindowID() == aOtherPageInfo->InnerWindowID();
+}
+
+void PageInformation::StreamJSON(SpliceableJSONWriter& aWriter) const {
+  aWriter.StartObjectElement();
+  // Here, we are converting uint64_t to double. Both Browsing Context and Inner
+  // Window IDs are creating using `nsContentUtils::GenerateProcessSpecificId`,
+  // which is specifically designed to only use 53 of the 64 bits to be lossless
+  // when passed into and out of JS as a double.
+  aWriter.DoubleProperty("browsingContextID", BrowsingContextID());
+  aWriter.DoubleProperty("innerWindowID", InnerWindowID());
+  aWriter.StringProperty("url", Url());
+  aWriter.DoubleProperty("embedderInnerWindowID", EmbedderInnerWindowID());
+  aWriter.EndObject();
+}
+
+size_t PageInformation::SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+  return aMallocSizeOf(this);
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/PageInformation.h b/mozglue/baseprofiler/core/PageInformation.h
new file mode 100644
index 0000000000..158b172bdf
--- /dev/null
+++ b/mozglue/baseprofiler/core/PageInformation.h
@@ -0,0 +1,77 @@
+/* -*- Mode: C++; tab-width: 2; 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 PageInformation_h
+#define PageInformation_h
+
+#include "mozilla/Atomics.h"
+#include "mozilla/Maybe.h"
+#include "mozilla/MemoryReporting.h"
+
+#include <string>
+
+namespace mozilla {
+namespace baseprofiler {
+
+class SpliceableJSONWriter;
+
+// This class contains information that's relevant to a single page only
+// while the page information is important and registered with the profiler,
+// but regardless of whether the profiler is running. All accesses to it are
+// protected by the profiler state lock.
+// When the page gets unregistered, we keep the profiler buffer position
+// to determine if we are still using this page. If not, we unregister
+// it in the next page registration.
+class PageInformation final {
+ public:
+  PageInformation(uint64_t aBrowsingContextID, uint64_t aInnerWindowID,
+                  const std::string& aUrl, uint64_t aEmbedderInnerWindowID);
+
+  // Using hand-rolled ref-counting, because RefCounted.h macros don't produce
+  // the same code between mozglue and libxul, see bug 1536656.
+  MFBT_API void AddRef() const { ++mRefCnt; }
+  MFBT_API void Release() const {
+    MOZ_ASSERT(int32_t(mRefCnt) > 0);
+    if (--mRefCnt) {
+      delete this;
+    }
+  }
+
+  size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const;
+  bool Equals(PageInformation* aOtherPageInfo) const;
+  void StreamJSON(SpliceableJSONWriter& aWriter) const;
+
+  uint64_t InnerWindowID() const { return mInnerWindowID; }
+  uint64_t BrowsingContextID() const { return mBrowsingContextID; }
+  const std::string& Url() const { return mUrl; }
+  uint64_t EmbedderInnerWindowID() const { return mEmbedderInnerWindowID; }
+
+  Maybe<uint64_t> BufferPositionWhenUnregistered() const {
+    return mBufferPositionWhenUnregistered;
+  }
+
+  void NotifyUnregistered(uint64_t aBufferPosition) {
+    mBufferPositionWhenUnregistered = Some(aBufferPosition);
+  }
+
+ private:
+  const uint64_t mBrowsingContextID;
+  const uint64_t mInnerWindowID;
+  const std::string mUrl;
+  const uint64_t mEmbedderInnerWindowID;
+
+  // Holds the buffer position when page is unregistered.
+  // It's used to determine if we still use this page in the profiler or
+  // not.
+  Maybe<uint64_t> mBufferPositionWhenUnregistered;
+
+  mutable Atomic<int32_t, MemoryOrdering::ReleaseAcquire> mRefCnt;
+};
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif  // PageInformation_h
diff --git a/mozglue/baseprofiler/core/PlatformMacros.h b/mozglue/baseprofiler/core/PlatformMacros.h
new file mode 100644
index 0000000000..c72e94c128
--- /dev/null
+++ b/mozglue/baseprofiler/core/PlatformMacros.h
@@ -0,0 +1,130 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* 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 PLATFORM_MACROS_H
+#define PLATFORM_MACROS_H
+
+// Define platform selection macros in a consistent way. Don't add anything
+// else to this file, so it can remain freestanding. The primary factorisation
+// is on (ARCH,OS) pairs ("PLATforms") but ARCH_ and OS_ macros are defined
+// too, since they are sometimes convenient.
+//
+// Note: "GP" is short for "Gecko Profiler".
+
+#undef GP_PLAT_x86_android
+#undef GP_PLAT_amd64_android
+#undef GP_PLAT_arm_android
+#undef GP_PLAT_arm64_android
+#undef GP_PLAT_x86_linux
+#undef GP_PLAT_amd64_linux
+#undef GP_PLAT_arm_linux
+#undef GP_PLAT_mips64_linux
+#undef GP_PLAT_amd64_darwin
+#undef GP_PLAT_arm64_darwin
+#undef GP_PLAT_x86_windows
+#undef GP_PLAT_amd64_windows
+#undef GP_PLAT_arm64_windows
+
+#undef GP_ARCH_x86
+#undef GP_ARCH_amd64
+#undef GP_ARCH_arm
+#undef GP_ARCH_arm64
+#undef GP_ARCH_mips64
+
+#undef GP_OS_android
+#undef GP_OS_linux
+#undef GP_OS_darwin
+#undef GP_OS_windows
+
+// We test __ANDROID__ before __linux__ because __linux__ is defined on both
+// Android and Linux, whereas GP_OS_android is not defined on vanilla Linux.
+
+#if defined(__ANDROID__) && defined(__i386__)
+#  define GP_PLAT_x86_android 1
+#  define GP_ARCH_x86 1
+#  define GP_OS_android 1
+
+#elif defined(__ANDROID__) && defined(__x86_64__)
+#  define GP_PLAT_amd64_android 1
+#  define GP_ARCH_amd64 1
+#  define GP_OS_android 1
+
+#elif defined(__ANDROID__) && defined(__arm__)
+#  define GP_PLAT_arm_android 1
+#  define GP_ARCH_arm 1
+#  define GP_OS_android 1
+
+#elif defined(__ANDROID__) && defined(__aarch64__)
+#  define GP_PLAT_arm64_android 1
+#  define GP_ARCH_arm64 1
+#  define GP_OS_android 1
+
+#elif defined(__linux__) && defined(__i386__)
+#  define GP_PLAT_x86_linux 1
+#  define GP_ARCH_x86 1
+#  define GP_OS_linux 1
+
+#elif defined(__linux__) && defined(__x86_64__)
+#  define GP_PLAT_amd64_linux 1
+#  define GP_ARCH_amd64 1
+#  define GP_OS_linux 1
+
+#elif defined(__linux__) && defined(__arm__)
+#  define GP_PLAT_arm_linux 1
+#  define GP_ARCH_arm 1
+#  define GP_OS_linux 1
+
+#elif defined(__linux__) && defined(__aarch64__)
+#  define GP_PLAT_arm64_linux 1
+#  define GP_ARCH_arm64 1
+#  define GP_OS_linux 1
+
+#elif defined(__linux__) && defined(__mips64)
+#  define GP_PLAT_mips64_linux 1
+#  define GP_ARCH_mips64 1
+#  define GP_OS_linux 1
+
+#elif defined(__APPLE__) && defined(__aarch64__)
+#  define GP_PLAT_arm64_darwin 1
+#  define GP_ARCH_arm64 1
+#  define GP_OS_darwin 1
+
+#elif defined(__APPLE__) && defined(__x86_64__)
+#  define GP_PLAT_amd64_darwin 1
+#  define GP_ARCH_amd64 1
+#  define GP_OS_darwin 1
+
+#elif defined(__FreeBSD__) && defined(__x86_64__)
+#  define GP_PLAT_amd64_freebsd 1
+#  define GP_ARCH_amd64 1
+#  define GP_OS_freebsd 1
+
+#elif defined(__FreeBSD__) && defined(__aarch64__)
+#  define GP_PLAT_arm64_freebsd 1
+#  define GP_ARCH_arm64 1
+#  define GP_OS_freebsd 1
+
+#elif (defined(_MSC_VER) || defined(__MINGW32__)) && \
+    (defined(_M_IX86) || defined(__i386__))
+#  define GP_PLAT_x86_windows 1
+#  define GP_ARCH_x86 1
+#  define GP_OS_windows 1
+
+#elif (defined(_MSC_VER) || defined(__MINGW32__)) && \
+    (defined(_M_X64) || defined(__x86_64__))
+#  define GP_PLAT_amd64_windows 1
+#  define GP_ARCH_amd64 1
+#  define GP_OS_windows 1
+
+#elif defined(_MSC_VER) && defined(_M_ARM64)
+#  define GP_PLAT_arm64_windows 1
+#  define GP_ARCH_arm64 1
+#  define GP_OS_windows 1
+
+#else
+#  error "Unsupported platform"
+#endif
+
+#endif /* ndef PLATFORM_MACROS_H */
diff --git a/mozglue/baseprofiler/core/ProfileBuffer.cpp b/mozglue/baseprofiler/core/ProfileBuffer.cpp
new file mode 100644
index 0000000000..f39244ee91
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfileBuffer.cpp
@@ -0,0 +1,210 @@
+/* -*- Mode: C++; tab-width: 2; 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 "ProfileBuffer.h"
+
+#include "mozilla/MathAlgorithms.h"
+
+#include "BaseProfiler.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+ProfileBuffer::ProfileBuffer(ProfileChunkedBuffer& aBuffer)
+    : mEntries(aBuffer) {
+  // Assume the given buffer is in-session.
+  MOZ_ASSERT(mEntries.IsInSession());
+}
+
+/* static */
+ProfileBufferBlockIndex ProfileBuffer::AddEntry(
+    ProfileChunkedBuffer& aProfileChunkedBuffer,
+    const ProfileBufferEntry& aEntry) {
+  switch (aEntry.GetKind()) {
+#define SWITCH_KIND(KIND, TYPE, SIZE)                          \
+  case ProfileBufferEntry::Kind::KIND: {                       \
+    return aProfileChunkedBuffer.PutFrom(&aEntry, 1 + (SIZE)); \
+    break;                                                     \
+  }
+
+    FOR_EACH_PROFILE_BUFFER_ENTRY_KIND(SWITCH_KIND)
+
+#undef SWITCH_KIND
+    default:
+      MOZ_ASSERT(false, "Unhandled baseprofiler::ProfilerBuffer entry KIND");
+      return ProfileBufferBlockIndex{};
+  }
+}
+
+// Called from signal, call only reentrant functions
+uint64_t ProfileBuffer::AddEntry(const ProfileBufferEntry& aEntry) {
+  return AddEntry(mEntries, aEntry).ConvertToProfileBufferIndex();
+}
+
+/* static */
+ProfileBufferBlockIndex ProfileBuffer::AddThreadIdEntry(
+    ProfileChunkedBuffer& aProfileChunkedBuffer, int aThreadId) {
+  return AddEntry(aProfileChunkedBuffer,
+                  ProfileBufferEntry::ThreadId(aThreadId));
+}
+
+uint64_t ProfileBuffer::AddThreadIdEntry(int aThreadId) {
+  return AddThreadIdEntry(mEntries, aThreadId).ConvertToProfileBufferIndex();
+}
+
+void ProfileBuffer::CollectCodeLocation(
+    const char* aLabel, const char* aStr, uint32_t aFrameFlags,
+    uint64_t aInnerWindowID, const Maybe<uint32_t>& aLineNumber,
+    const Maybe<uint32_t>& aColumnNumber,
+    const Maybe<ProfilingCategoryPair>& aCategoryPair) {
+  AddEntry(ProfileBufferEntry::Label(aLabel));
+  AddEntry(ProfileBufferEntry::FrameFlags(uint64_t(aFrameFlags)));
+
+  if (aStr) {
+    // Store the string using one or more DynamicStringFragment entries.
+    size_t strLen = strlen(aStr) + 1;  // +1 for the null terminator
+    // If larger than the prescribed limit, we will cut the string and end it
+    // with an ellipsis.
+    const bool tooBig = strLen > kMaxFrameKeyLength;
+    if (tooBig) {
+      strLen = kMaxFrameKeyLength;
+    }
+    char chars[ProfileBufferEntry::kNumChars];
+    for (size_t j = 0;; j += ProfileBufferEntry::kNumChars) {
+      // Store up to kNumChars characters in the entry.
+      size_t len = ProfileBufferEntry::kNumChars;
+      const bool last = j + len >= strLen;
+      if (last) {
+        // Only the last entry may be smaller than kNumChars.
+        len = strLen - j;
+        if (tooBig) {
+          // That last entry is part of a too-big string, replace the end
+          // characters with an ellipsis "...".
+          len = std::max(len, size_t(4));
+          chars[len - 4] = '.';
+          chars[len - 3] = '.';
+          chars[len - 2] = '.';
+          chars[len - 1] = '\0';
+          // Make sure the memcpy will not overwrite our ellipsis!
+          len -= 4;
+        }
+      }
+      memcpy(chars, &aStr[j], len);
+      AddEntry(ProfileBufferEntry::DynamicStringFragment(chars));
+      if (last) {
+        break;
+      }
+    }
+  }
+
+  if (aInnerWindowID) {
+    AddEntry(ProfileBufferEntry::InnerWindowID(aInnerWindowID));
+  }
+
+  if (aLineNumber) {
+    AddEntry(ProfileBufferEntry::LineNumber(*aLineNumber));
+  }
+
+  if (aColumnNumber) {
+    AddEntry(ProfileBufferEntry::ColumnNumber(*aColumnNumber));
+  }
+
+  if (aCategoryPair.isSome()) {
+    AddEntry(ProfileBufferEntry::CategoryPair(int(*aCategoryPair)));
+  }
+}
+
+size_t ProfileBuffer::SizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const {
+  // Measurement of the following members may be added later if DMD finds it
+  // is worthwhile:
+  // - memory pointed to by the elements within mEntries
+  return mEntries.SizeOfExcludingThis(aMallocSizeOf);
+}
+
+size_t ProfileBuffer::SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+  return aMallocSizeOf(this) + SizeOfExcludingThis(aMallocSizeOf);
+}
+
+void ProfileBuffer::CollectOverheadStats(TimeDuration aSamplingTime,
+                                         TimeDuration aLocking,
+                                         TimeDuration aCleaning,
+                                         TimeDuration aCounters,
+                                         TimeDuration aThreads) {
+  double timeUs = aSamplingTime.ToMilliseconds() * 1000.0;
+  if (mFirstSamplingTimeUs == 0.0) {
+    mFirstSamplingTimeUs = timeUs;
+  } else {
+    // Note that we'll have 1 fewer interval than other numbers (because
+    // we need both ends of an interval to know its duration). The final
+    // difference should be insignificant over the expected many thousands
+    // of iterations.
+    mIntervalsUs.Count(timeUs - mLastSamplingTimeUs);
+  }
+  mLastSamplingTimeUs = timeUs;
+  // Time to take the lock before sampling.
+  double lockingUs = aLocking.ToMilliseconds() * 1000.0;
+  // Time to discard expired data.
+  double cleaningUs = aCleaning.ToMilliseconds() * 1000.0;
+  // Time to gather all counters.
+  double countersUs = aCounters.ToMilliseconds() * 1000.0;
+  // Time to sample all threads.
+  double threadsUs = aThreads.ToMilliseconds() * 1000.0;
+
+  // Add to our gathered stats.
+  mOverheadsUs.Count(lockingUs + cleaningUs + countersUs + threadsUs);
+  mLockingsUs.Count(lockingUs);
+  mCleaningsUs.Count(cleaningUs);
+  mCountersUs.Count(countersUs);
+  mThreadsUs.Count(threadsUs);
+
+  // Record details in buffer.
+  AddEntry(ProfileBufferEntry::ProfilerOverheadTime(timeUs));
+  AddEntry(ProfileBufferEntry::ProfilerOverheadDuration(lockingUs));
+  AddEntry(ProfileBufferEntry::ProfilerOverheadDuration(cleaningUs));
+  AddEntry(ProfileBufferEntry::ProfilerOverheadDuration(countersUs));
+  AddEntry(ProfileBufferEntry::ProfilerOverheadDuration(threadsUs));
+}
+
+ProfilerBufferInfo ProfileBuffer::GetProfilerBufferInfo() const {
+  return {BufferRangeStart(),
+          BufferRangeEnd(),
+          static_cast<uint32_t>(*mEntries.BufferLength() /
+                                8),  // 8 bytes per entry.
+          mIntervalsUs,
+          mOverheadsUs,
+          mLockingsUs,
+          mCleaningsUs,
+          mCountersUs,
+          mThreadsUs};
+}
+
+/* ProfileBufferCollector */
+
+void ProfileBufferCollector::CollectNativeLeafAddr(void* aAddr) {
+  mBuf.AddEntry(ProfileBufferEntry::NativeLeafAddr(aAddr));
+}
+
+void ProfileBufferCollector::CollectProfilingStackFrame(
+    const ProfilingStackFrame& aFrame) {
+  // WARNING: this function runs within the profiler's "critical section".
+
+  MOZ_ASSERT(aFrame.isLabelFrame() ||
+             (aFrame.isJsFrame() && !aFrame.isOSRFrame()));
+
+  const char* label = aFrame.label();
+  const char* dynamicString = aFrame.dynamicString();
+  Maybe<uint32_t> line;
+  Maybe<uint32_t> column;
+
+  MOZ_ASSERT(aFrame.isLabelFrame());
+
+  mBuf.CollectCodeLocation(label, dynamicString, aFrame.flags(),
+                           aFrame.realmID(), line, column,
+                           Some(aFrame.categoryPair()));
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/ProfileBuffer.h b/mozglue/baseprofiler/core/ProfileBuffer.h
new file mode 100644
index 0000000000..b7a0af5e93
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfileBuffer.h
@@ -0,0 +1,186 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* 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 MOZ_PROFILE_BUFFER_H
+#define MOZ_PROFILE_BUFFER_H
+
+#include "ProfileBufferEntry.h"
+
+#include "mozilla/Maybe.h"
+#include "mozilla/PowerOfTwo.h"
+#include "mozilla/ProfileBufferChunkManagerSingle.h"
+#include "mozilla/ProfileChunkedBuffer.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+// Class storing most profiling data in a ProfileChunkedBuffer.
+//
+// This class is used as a queue of entries which, after construction, never
+// allocates. This makes it safe to use in the profiler's "critical section".
+class ProfileBuffer final {
+ public:
+  // ProfileBuffer constructor
+  // @param aBuffer The in-session ProfileChunkedBuffer to use as buffer
+  // manager.
+  explicit ProfileBuffer(ProfileChunkedBuffer& aBuffer);
+
+  ProfileChunkedBuffer& UnderlyingChunkedBuffer() const { return mEntries; }
+
+  bool IsThreadSafe() const { return mEntries.IsThreadSafe(); }
+
+  // Add |aEntry| to the buffer, ignoring what kind of entry it is.
+  // Returns the position of the entry.
+  uint64_t AddEntry(const ProfileBufferEntry& aEntry);
+
+  // Add to the buffer a sample start (ThreadId) entry for aThreadId.
+  // Returns the position of the entry.
+  uint64_t AddThreadIdEntry(int aThreadId);
+
+  void CollectCodeLocation(const char* aLabel, const char* aStr,
+                           uint32_t aFrameFlags, uint64_t aInnerWindowID,
+                           const Maybe<uint32_t>& aLineNumber,
+                           const Maybe<uint32_t>& aColumnNumber,
+                           const Maybe<ProfilingCategoryPair>& aCategoryPair);
+
+  // Maximum size of a frameKey string that we'll handle.
+  static const size_t kMaxFrameKeyLength = 512;
+
+  // Stream JSON for samples in the buffer to aWriter, using the supplied
+  // UniqueStacks object.
+  // Only streams samples for the given thread ID and which were taken at or
+  // after aSinceTime. If ID is 0, ignore the stored thread ID; this should only
+  // be used when the buffer contains only one sample.
+  // Return the thread ID of the streamed sample(s), or 0.
+  int StreamSamplesToJSON(SpliceableJSONWriter& aWriter, int aThreadId,
+                          double aSinceTime, UniqueStacks& aUniqueStacks) const;
+
+  void StreamMarkersToJSON(SpliceableJSONWriter& aWriter, int aThreadId,
+                           const TimeStamp& aProcessStartTime,
+                           double aSinceTime,
+                           UniqueStacks& aUniqueStacks) const;
+  void StreamPausedRangesToJSON(SpliceableJSONWriter& aWriter,
+                                double aSinceTime) const;
+  void StreamProfilerOverheadToJSON(SpliceableJSONWriter& aWriter,
+                                    const TimeStamp& aProcessStartTime,
+                                    double aSinceTime) const;
+  void StreamCountersToJSON(SpliceableJSONWriter& aWriter,
+                            const TimeStamp& aProcessStartTime,
+                            double aSinceTime) const;
+
+  // Find (via |aLastSample|) the most recent sample for the thread denoted by
+  // |aThreadId| and clone it, patching in the current time as appropriate.
+  // Mutate |aLastSample| to point to the newly inserted sample.
+  // Returns whether duplication was successful.
+  bool DuplicateLastSample(int aThreadId, const TimeStamp& aProcessStartTime,
+                           Maybe<uint64_t>& aLastSample);
+
+  void DiscardSamplesBeforeTime(double aTime);
+
+  size_t SizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const;
+  size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const;
+
+  void CollectOverheadStats(TimeDuration aSamplingTime, TimeDuration aLocking,
+                            TimeDuration aCleaning, TimeDuration aCounters,
+                            TimeDuration aThreads);
+
+  ProfilerBufferInfo GetProfilerBufferInfo() const;
+
+ private:
+  // Add |aEntry| to the provider ProfileChunkedBuffer.
+  // `static` because it may be used to add an entry to a `ProfileChunkedBuffer`
+  // that is not attached to a `ProfileBuffer`.
+  static ProfileBufferBlockIndex AddEntry(
+      ProfileChunkedBuffer& aProfileChunkedBuffer,
+      const ProfileBufferEntry& aEntry);
+
+  // Add a sample start (ThreadId) entry for aThreadId to the provided
+  // ProfileChunkedBuffer. Returns the position of the entry.
+  // `static` because it may be used to add an entry to a `ProfileChunkedBuffer`
+  // that is not attached to a `ProfileBuffer`.
+  static ProfileBufferBlockIndex AddThreadIdEntry(
+      ProfileChunkedBuffer& aProfileChunkedBuffer, int aThreadId);
+
+  // The storage in which this ProfileBuffer stores its entries.
+  ProfileChunkedBuffer& mEntries;
+
+ public:
+  // `BufferRangeStart()` and `BufferRangeEnd()` return `uint64_t` values
+  // corresponding to the first entry and past the last entry stored in
+  // `mEntries`.
+  //
+  // The returned values are not guaranteed to be stable, because other threads
+  // may also be accessing the buffer concurrently. But they will always
+  // increase, and can therefore give an indication of how far these values have
+  // *at least* reached. In particular:
+  // - Entries whose index is strictly less that `BufferRangeStart()` have been
+  //   discarded by now, so any related data may also be safely discarded.
+  // - It is safe to try and read entries at any index strictly less than
+  //   `BufferRangeEnd()` -- but note that these reads may fail by the time you
+  //   request them, as old entries get overwritten by new ones.
+  uint64_t BufferRangeStart() const { return mEntries.GetState().mRangeStart; }
+  uint64_t BufferRangeEnd() const { return mEntries.GetState().mRangeEnd; }
+
+ private:
+  // Single pre-allocated chunk (to avoid spurious mallocs), used when:
+  // - Duplicating sleeping stacks (hence scExpectedMaximumStackSize).
+  // - Adding JIT info.
+  // - Streaming stacks to JSON.
+  // Mutable because it's accessed from non-multithreaded const methods.
+  mutable ProfileBufferChunkManagerSingle mWorkerChunkManager{
+      ProfileBufferChunk::Create(
+          ProfileBufferChunk::SizeofChunkMetadata() +
+          ProfileBufferChunkManager::scExpectedMaximumStackSize)};
+
+  // Time from launch (us) when first sampling was recorded.
+  double mFirstSamplingTimeUs = 0.0;
+  // Time from launch (us) when last sampling was recorded.
+  double mLastSamplingTimeUs = 0.0;
+  // Sampling stats: Interval (us) between successive samplings.
+  ProfilerStats mIntervalsUs;
+  // Sampling stats: Total duration (us) of each sampling. (Split detail below.)
+  ProfilerStats mOverheadsUs;
+  // Sampling stats: Time (us) to acquire the lock before sampling.
+  ProfilerStats mLockingsUs;
+  // Sampling stats: Time (us) to discard expired data.
+  ProfilerStats mCleaningsUs;
+  // Sampling stats: Time (us) to collect counter data.
+  ProfilerStats mCountersUs;
+  // Sampling stats: Time (us) to sample thread stacks.
+  ProfilerStats mThreadsUs;
+};
+
+/**
+ * Helper type used to implement ProfilerStackCollector. This type is used as
+ * the collector for MergeStacks by ProfileBuffer. It holds a reference to the
+ * buffer, as well as additional feature flags which are needed to control the
+ * data collection strategy
+ */
+class ProfileBufferCollector final : public ProfilerStackCollector {
+ public:
+  ProfileBufferCollector(ProfileBuffer& aBuf, uint64_t aSamplePos)
+      : mBuf(aBuf), mSamplePositionInBuffer(aSamplePos) {}
+
+  Maybe<uint64_t> SamplePositionInBuffer() override {
+    return Some(mSamplePositionInBuffer);
+  }
+
+  Maybe<uint64_t> BufferRangeStart() override {
+    return Some(mBuf.BufferRangeStart());
+  }
+
+  virtual void CollectNativeLeafAddr(void* aAddr) override;
+  virtual void CollectProfilingStackFrame(
+      const ProfilingStackFrame& aFrame) override;
+
+ private:
+  ProfileBuffer& mBuf;
+  uint64_t mSamplePositionInBuffer;
+};
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif
diff --git a/mozglue/baseprofiler/core/ProfileBufferEntry.cpp b/mozglue/baseprofiler/core/ProfileBufferEntry.cpp
new file mode 100644
index 0000000000..3d3f68b655
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfileBufferEntry.cpp
@@ -0,0 +1,1337 @@
+/* -*- 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 "ProfileBufferEntry.h"
+
+#include <ostream>
+#include <type_traits>
+
+#include "mozilla/Logging.h"
+#include "mozilla/ScopeExit.h"
+#include "mozilla/Sprintf.h"
+#include "mozilla/StackWalk.h"
+
+#include "BaseProfiler.h"
+#include "mozilla/BaseProfilerMarkers.h"
+#include "platform.h"
+#include "ProfileBuffer.h"
+#include "ProfilerBacktrace.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN ProfileBufferEntry
+
+ProfileBufferEntry::ProfileBufferEntry()
+    : mKind(Kind::INVALID), mStorage{0, 0, 0, 0, 0, 0, 0, 0} {}
+
+// aString must be a static string.
+ProfileBufferEntry::ProfileBufferEntry(Kind aKind, const char* aString)
+    : mKind(aKind) {
+  memcpy(mStorage, &aString, sizeof(aString));
+}
+
+ProfileBufferEntry::ProfileBufferEntry(Kind aKind, char aChars[kNumChars])
+    : mKind(aKind) {
+  memcpy(mStorage, aChars, kNumChars);
+}
+
+ProfileBufferEntry::ProfileBufferEntry(Kind aKind, void* aPtr) : mKind(aKind) {
+  memcpy(mStorage, &aPtr, sizeof(aPtr));
+}
+
+ProfileBufferEntry::ProfileBufferEntry(Kind aKind, double aDouble)
+    : mKind(aKind) {
+  memcpy(mStorage, &aDouble, sizeof(aDouble));
+}
+
+ProfileBufferEntry::ProfileBufferEntry(Kind aKind, int aInt) : mKind(aKind) {
+  memcpy(mStorage, &aInt, sizeof(aInt));
+}
+
+ProfileBufferEntry::ProfileBufferEntry(Kind aKind, int64_t aInt64)
+    : mKind(aKind) {
+  memcpy(mStorage, &aInt64, sizeof(aInt64));
+}
+
+ProfileBufferEntry::ProfileBufferEntry(Kind aKind, uint64_t aUint64)
+    : mKind(aKind) {
+  memcpy(mStorage, &aUint64, sizeof(aUint64));
+}
+
+const char* ProfileBufferEntry::GetString() const {
+  const char* result;
+  memcpy(&result, mStorage, sizeof(result));
+  return result;
+}
+
+void* ProfileBufferEntry::GetPtr() const {
+  void* result;
+  memcpy(&result, mStorage, sizeof(result));
+  return result;
+}
+
+double ProfileBufferEntry::GetDouble() const {
+  double result;
+  memcpy(&result, mStorage, sizeof(result));
+  return result;
+}
+
+int ProfileBufferEntry::GetInt() const {
+  int result;
+  memcpy(&result, mStorage, sizeof(result));
+  return result;
+}
+
+int64_t ProfileBufferEntry::GetInt64() const {
+  int64_t result;
+  memcpy(&result, mStorage, sizeof(result));
+  return result;
+}
+
+uint64_t ProfileBufferEntry::GetUint64() const {
+  uint64_t result;
+  memcpy(&result, mStorage, sizeof(result));
+  return result;
+}
+
+void ProfileBufferEntry::CopyCharsInto(char (&aOutArray)[kNumChars]) const {
+  memcpy(aOutArray, mStorage, kNumChars);
+}
+
+// END ProfileBufferEntry
+////////////////////////////////////////////////////////////////////////
+
+// As mentioned in ProfileBufferEntry.h, the JSON format contains many
+// arrays whose elements are laid out according to various schemas to help
+// de-duplication. This RAII class helps write these arrays by keeping track of
+// the last non-null element written and adding the appropriate number of null
+// elements when writing new non-null elements. It also automatically opens and
+// closes an array element on the given JSON writer.
+//
+// You grant the AutoArraySchemaWriter exclusive access to the JSONWriter and
+// the UniqueJSONStrings objects for the lifetime of AutoArraySchemaWriter. Do
+// not access them independently while the AutoArraySchemaWriter is alive.
+// If you need to add complex objects, call FreeFormElement(), which will give
+// you temporary access to the writer.
+//
+// Example usage:
+//
+//     // Define the schema of elements in this type of array: [FOO, BAR, BAZ]
+//     enum Schema : uint32_t {
+//       FOO = 0,
+//       BAR = 1,
+//       BAZ = 2
+//     };
+//
+//     AutoArraySchemaWriter writer(someJsonWriter, someUniqueStrings);
+//     if (shouldWriteFoo) {
+//       writer.IntElement(FOO, getFoo());
+//     }
+//     ... etc ...
+//
+//     The elements need to be added in-order.
+class MOZ_RAII AutoArraySchemaWriter {
+ public:
+  explicit AutoArraySchemaWriter(SpliceableJSONWriter& aWriter)
+      : mJSONWriter(aWriter), mNextFreeIndex(0) {
+    mJSONWriter.StartArrayElement(SpliceableJSONWriter::SingleLineStyle);
+  }
+
+  ~AutoArraySchemaWriter() { mJSONWriter.EndArray(); }
+
+  template <typename T>
+  void IntElement(uint32_t aIndex, T aValue) {
+    static_assert(!std::is_same_v<T, uint64_t>,
+                  "Narrowing uint64 -> int64 conversion not allowed");
+    FillUpTo(aIndex);
+    mJSONWriter.IntElement(static_cast<int64_t>(aValue));
+  }
+
+  void DoubleElement(uint32_t aIndex, double aValue) {
+    FillUpTo(aIndex);
+    mJSONWriter.DoubleElement(aValue);
+  }
+
+  void BoolElement(uint32_t aIndex, bool aValue) {
+    FillUpTo(aIndex);
+    mJSONWriter.BoolElement(aValue);
+  }
+
+ protected:
+  SpliceableJSONWriter& Writer() { return mJSONWriter; }
+
+  void FillUpTo(uint32_t aIndex) {
+    MOZ_ASSERT(aIndex >= mNextFreeIndex);
+    mJSONWriter.NullElements(aIndex - mNextFreeIndex);
+    mNextFreeIndex = aIndex + 1;
+  }
+
+ private:
+  SpliceableJSONWriter& mJSONWriter;
+  uint32_t mNextFreeIndex;
+};
+
+// Same as AutoArraySchemaWriter, but this can also write strings (output as
+// indexes into the table of unique strings).
+class MOZ_RAII AutoArraySchemaWithStringsWriter : public AutoArraySchemaWriter {
+ public:
+  AutoArraySchemaWithStringsWriter(SpliceableJSONWriter& aWriter,
+                                   UniqueJSONStrings& aStrings)
+      : AutoArraySchemaWriter(aWriter), mStrings(aStrings) {}
+
+  void StringElement(uint32_t aIndex, const Span<const char>& aValue) {
+    FillUpTo(aIndex);
+    mStrings.WriteElement(Writer(), aValue);
+  }
+
+ private:
+  UniqueJSONStrings& mStrings;
+};
+
+UniqueStacks::StackKey UniqueStacks::BeginStack(const FrameKey& aFrame) {
+  return StackKey(GetOrAddFrameIndex(aFrame));
+}
+
+UniqueStacks::StackKey UniqueStacks::AppendFrame(const StackKey& aStack,
+                                                 const FrameKey& aFrame) {
+  return StackKey(aStack, GetOrAddStackIndex(aStack),
+                  GetOrAddFrameIndex(aFrame));
+}
+
+bool UniqueStacks::FrameKey::NormalFrameData::operator==(
+    const NormalFrameData& aOther) const {
+  return mLocation == aOther.mLocation &&
+         mRelevantForJS == aOther.mRelevantForJS &&
+         mInnerWindowID == aOther.mInnerWindowID && mLine == aOther.mLine &&
+         mColumn == aOther.mColumn && mCategoryPair == aOther.mCategoryPair;
+}
+
+UniqueStacks::UniqueStacks() : mUniqueStrings(MakeUnique<UniqueJSONStrings>()) {
+  mFrameTableWriter.StartBareList();
+  mStackTableWriter.StartBareList();
+}
+
+uint32_t UniqueStacks::GetOrAddStackIndex(const StackKey& aStack) {
+  uint32_t count = mStackToIndexMap.count();
+  auto entry = mStackToIndexMap.lookupForAdd(aStack);
+  if (entry) {
+    MOZ_ASSERT(entry->value() < count);
+    return entry->value();
+  }
+
+  MOZ_RELEASE_ASSERT(mStackToIndexMap.add(entry, aStack, count));
+  StreamStack(aStack);
+  return count;
+}
+
+uint32_t UniqueStacks::GetOrAddFrameIndex(const FrameKey& aFrame) {
+  uint32_t count = mFrameToIndexMap.count();
+  auto entry = mFrameToIndexMap.lookupForAdd(aFrame);
+  if (entry) {
+    MOZ_ASSERT(entry->value() < count);
+    return entry->value();
+  }
+
+  MOZ_RELEASE_ASSERT(mFrameToIndexMap.add(entry, aFrame, count));
+  StreamNonJITFrame(aFrame);
+  return count;
+}
+
+void UniqueStacks::SpliceFrameTableElements(SpliceableJSONWriter& aWriter) {
+  mFrameTableWriter.EndBareList();
+  aWriter.TakeAndSplice(mFrameTableWriter.TakeChunkedWriteFunc());
+}
+
+void UniqueStacks::SpliceStackTableElements(SpliceableJSONWriter& aWriter) {
+  mStackTableWriter.EndBareList();
+  aWriter.TakeAndSplice(mStackTableWriter.TakeChunkedWriteFunc());
+}
+
+void UniqueStacks::StreamStack(const StackKey& aStack) {
+  enum Schema : uint32_t { PREFIX = 0, FRAME = 1 };
+
+  AutoArraySchemaWriter writer(mStackTableWriter);
+  if (aStack.mPrefixStackIndex.isSome()) {
+    writer.IntElement(PREFIX, *aStack.mPrefixStackIndex);
+  }
+  writer.IntElement(FRAME, aStack.mFrameIndex);
+}
+
+void UniqueStacks::StreamNonJITFrame(const FrameKey& aFrame) {
+  using NormalFrameData = FrameKey::NormalFrameData;
+
+  enum Schema : uint32_t {
+    LOCATION = 0,
+    RELEVANT_FOR_JS = 1,
+    INNER_WINDOW_ID = 2,
+    IMPLEMENTATION = 3,
+    OPTIMIZATIONS = 4,
+    LINE = 5,
+    COLUMN = 6,
+    CATEGORY = 7,
+    SUBCATEGORY = 8
+  };
+
+  AutoArraySchemaWithStringsWriter writer(mFrameTableWriter, *mUniqueStrings);
+
+  const NormalFrameData& data = aFrame.mData.as<NormalFrameData>();
+  writer.StringElement(LOCATION, data.mLocation);
+  writer.BoolElement(RELEVANT_FOR_JS, data.mRelevantForJS);
+
+  // It's okay to convert uint64_t to double here because DOM always creates IDs
+  // that are convertible to double.
+  writer.DoubleElement(INNER_WINDOW_ID, data.mInnerWindowID);
+
+  if (data.mLine.isSome()) {
+    writer.IntElement(LINE, *data.mLine);
+  }
+  if (data.mColumn.isSome()) {
+    writer.IntElement(COLUMN, *data.mColumn);
+  }
+  if (data.mCategoryPair.isSome()) {
+    const ProfilingCategoryPairInfo& info =
+        GetProfilingCategoryPairInfo(*data.mCategoryPair);
+    writer.IntElement(CATEGORY, uint32_t(info.mCategory));
+    writer.IntElement(SUBCATEGORY, info.mSubcategoryIndex);
+  }
+}
+
+struct CStringWriteFunc : public JSONWriteFunc {
+  std::string& mBuffer;  // The struct must not outlive this buffer
+  explicit CStringWriteFunc(std::string& aBuffer) : mBuffer(aBuffer) {}
+
+  void Write(const Span<const char>& aStr) override {
+    mBuffer.append(aStr.data(), aStr.size());
+  }
+};
+
+struct ProfileSample {
+  uint32_t mStack;
+  double mTime;
+  Maybe<double> mResponsiveness;
+};
+
+static void WriteSample(SpliceableJSONWriter& aWriter,
+                        const ProfileSample& aSample) {
+  enum Schema : uint32_t {
+    STACK = 0,
+    TIME = 1,
+    EVENT_DELAY = 2,
+  };
+
+  AutoArraySchemaWriter writer(aWriter);
+
+  writer.IntElement(STACK, aSample.mStack);
+
+  writer.DoubleElement(TIME, aSample.mTime);
+
+  if (aSample.mResponsiveness.isSome()) {
+    writer.DoubleElement(EVENT_DELAY, *aSample.mResponsiveness);
+  }
+}
+
+class EntryGetter {
+ public:
+  explicit EntryGetter(ProfileChunkedBuffer::Reader& aReader,
+                       uint64_t aInitialReadPos = 0)
+      : mBlockIt(
+            aReader.At(ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+                aInitialReadPos))),
+        mBlockItEnd(aReader.end()) {
+    if (!ReadLegacyOrEnd()) {
+      // Find and read the next non-legacy entry.
+      Next();
+    }
+  }
+
+  bool Has() const { return mBlockIt != mBlockItEnd; }
+
+  const ProfileBufferEntry& Get() const {
+    MOZ_ASSERT(Has(), "Caller should have checked `Has()` before `Get()`");
+    return mEntry;
+  }
+
+  void Next() {
+    MOZ_ASSERT(Has(), "Caller should have checked `Has()` before `Next()`");
+    for (;;) {
+      ++mBlockIt;
+      if (ReadLegacyOrEnd()) {
+        // Either we're at the end, or we could read a legacy entry -> Done.
+        break;
+      }
+      // Otherwise loop around until we hit the end or a legacy entry.
+    }
+  }
+
+  ProfileBufferBlockIndex CurBlockIndex() const {
+    return mBlockIt.CurrentBlockIndex();
+  }
+
+  uint64_t CurPos() const {
+    return CurBlockIndex().ConvertToProfileBufferIndex();
+  }
+
+ private:
+  // Try to read the entry at the current `mBlockIt` position.
+  // * If we're at the end of the buffer, just return `true`.
+  // * If there is a "legacy" entry (containing a real `ProfileBufferEntry`),
+  //   read it into `mEntry`, and return `true` as well.
+  // * Otherwise the entry contains a "modern" type that cannot be read into
+  // `mEntry`, return `false` (so `EntryGetter` can skip to another entry).
+  bool ReadLegacyOrEnd() {
+    if (!Has()) {
+      return true;
+    }
+    // Read the entry "kind", which is always at the start of all entries.
+    ProfileBufferEntryReader aER = *mBlockIt;
+    auto type = static_cast<ProfileBufferEntry::Kind>(
+        aER.ReadObject<ProfileBufferEntry::KindUnderlyingType>());
+    MOZ_ASSERT(static_cast<ProfileBufferEntry::KindUnderlyingType>(type) <
+               static_cast<ProfileBufferEntry::KindUnderlyingType>(
+                   ProfileBufferEntry::Kind::MODERN_LIMIT));
+    if (type >= ProfileBufferEntry::Kind::LEGACY_LIMIT) {
+      aER.SetRemainingBytes(0);
+      return false;
+    }
+    // Here, we have a legacy item, we need to read it from the start.
+    // Because the above `ReadObject` moved the reader, we ned to reset it to
+    // the start of the entry before reading the whole entry.
+    aER = *mBlockIt;
+    aER.ReadBytes(&mEntry, aER.RemainingBytes());
+    return true;
+  }
+
+  ProfileBufferEntry mEntry;
+  ProfileChunkedBuffer::BlockIterator mBlockIt;
+  const ProfileChunkedBuffer::BlockIterator mBlockItEnd;
+};
+
+// The following grammar shows legal sequences of profile buffer entries.
+// The sequences beginning with a ThreadId entry are known as "samples".
+//
+// (
+//   ( /* Samples */
+//     ThreadId
+//     Time
+//     ( NativeLeafAddr
+//     | Label FrameFlags? DynamicStringFragment* LineNumber? CategoryPair?
+//     | JitReturnAddr
+//     )+
+//     Responsiveness?
+//   )
+//   | MarkerData
+//   | ( /* Counters */
+//       CounterId
+//       Time
+//       (
+//         CounterKey
+//         Count
+//         Number?
+//       )*
+//     )
+//   | CollectionStart
+//   | CollectionEnd
+//   | Pause
+//   | Resume
+//   | ( ProfilerOverheadTime /* Sampling start timestamp */
+//       ProfilerOverheadDuration /* Lock acquisition */
+//       ProfilerOverheadDuration /* Expired data cleaning */
+//       ProfilerOverheadDuration /* Counters */
+//       ProfilerOverheadDuration /* Threads */
+//     )
+// )*
+//
+// The most complicated part is the stack entry sequence that begins with
+// Label. Here are some examples.
+//
+// - ProfilingStack frames without a dynamic string:
+//
+//     Label("js::RunScript")
+//     CategoryPair(ProfilingCategoryPair::JS)
+//
+//     Label("XREMain::XRE_main")
+//     LineNumber(4660)
+//     CategoryPair(ProfilingCategoryPair::OTHER)
+//
+//     Label("ElementRestyler::ComputeStyleChangeFor")
+//     LineNumber(3003)
+//     CategoryPair(ProfilingCategoryPair::CSS)
+//
+// - ProfilingStack frames with a dynamic string:
+//
+//     Label("nsObserverService::NotifyObservers")
+//     FrameFlags(uint64_t(ProfilingStackFrame::Flags::IS_LABEL_FRAME))
+//     DynamicStringFragment("domwindo")
+//     DynamicStringFragment("wopened")
+//     LineNumber(291)
+//     CategoryPair(ProfilingCategoryPair::OTHER)
+//
+//     Label("")
+//     FrameFlags(uint64_t(ProfilingStackFrame::Flags::IS_JS_FRAME))
+//     DynamicStringFragment("closeWin")
+//     DynamicStringFragment("dow (chr")
+//     DynamicStringFragment("ome://gl")
+//     DynamicStringFragment("obal/con")
+//     DynamicStringFragment("tent/glo")
+//     DynamicStringFragment("balOverl")
+//     DynamicStringFragment("ay.js:5)")
+//     DynamicStringFragment("")          # this string holds the closing '\0'
+//     LineNumber(25)
+//     CategoryPair(ProfilingCategoryPair::JS)
+//
+//     Label("")
+//     FrameFlags(uint64_t(ProfilingStackFrame::Flags::IS_JS_FRAME))
+//     DynamicStringFragment("bound (s")
+//     DynamicStringFragment("elf-host")
+//     DynamicStringFragment("ed:914)")
+//     LineNumber(945)
+//     CategoryPair(ProfilingCategoryPair::JS)
+//
+// - A profiling stack frame with an overly long dynamic string:
+//
+//     Label("")
+//     FrameFlags(uint64_t(ProfilingStackFrame::Flags::IS_LABEL_FRAME))
+//     DynamicStringFragment("(too lon")
+//     DynamicStringFragment("g)")
+//     LineNumber(100)
+//     CategoryPair(ProfilingCategoryPair::NETWORK)
+//
+// - A wasm JIT frame:
+//
+//     Label("")
+//     FrameFlags(uint64_t(0))
+//     DynamicStringFragment("wasm-fun")
+//     DynamicStringFragment("ction[87")
+//     DynamicStringFragment("36] (blo")
+//     DynamicStringFragment("b:http:/")
+//     DynamicStringFragment("/webasse")
+//     DynamicStringFragment("mbly.org")
+//     DynamicStringFragment("/3dc5759")
+//     DynamicStringFragment("4-ce58-4")
+//     DynamicStringFragment("626-975b")
+//     DynamicStringFragment("-08ad116")
+//     DynamicStringFragment("30bc1:38")
+//     DynamicStringFragment("29856)")
+//
+// - A JS frame in a synchronous sample:
+//
+//     Label("")
+//     FrameFlags(uint64_t(ProfilingStackFrame::Flags::IS_LABEL_FRAME))
+//     DynamicStringFragment("u (https")
+//     DynamicStringFragment("://perf-")
+//     DynamicStringFragment("html.io/")
+//     DynamicStringFragment("ac0da204")
+//     DynamicStringFragment("aaa44d75")
+//     DynamicStringFragment("a800.bun")
+//     DynamicStringFragment("dle.js:2")
+//     DynamicStringFragment("5)")
+
+// Because this is a format entirely internal to the Profiler, any parsing
+// error indicates a bug in the ProfileBuffer writing or the parser itself,
+// or possibly flaky hardware.
+#define ERROR_AND_CONTINUE(msg)                            \
+  {                                                        \
+    fprintf(stderr, "ProfileBuffer parse error: %s", msg); \
+    MOZ_ASSERT(false, msg);                                \
+    continue;                                              \
+  }
+
+int ProfileBuffer::StreamSamplesToJSON(SpliceableJSONWriter& aWriter,
+                                       int aThreadId, double aSinceTime,
+                                       UniqueStacks& aUniqueStacks) const {
+  UniquePtr<char[]> dynStrBuf = MakeUnique<char[]>(kMaxFrameKeyLength);
+
+  return mEntries.Read([&](ProfileChunkedBuffer::Reader* aReader) {
+    MOZ_ASSERT(aReader,
+               "ProfileChunkedBuffer cannot be out-of-session when sampler is "
+               "running");
+
+    int processedThreadId = 0;
+
+    EntryGetter e(*aReader);
+
+    for (;;) {
+      // This block skips entries until we find the start of the next sample.
+      // This is useful in three situations.
+      //
+      // - The circular buffer overwrites old entries, so when we start parsing
+      //   we might be in the middle of a sample, and we must skip forward to
+      //   the start of the next sample.
+      //
+      // - We skip samples that don't have an appropriate ThreadId or Time.
+      //
+      // - We skip range Pause, Resume, CollectionStart, Counter and
+      //   CollectionEnd entries between samples.
+      while (e.Has()) {
+        if (e.Get().IsThreadId()) {
+          break;
+        }
+        e.Next();
+      }
+
+      if (!e.Has()) {
+        break;
+      }
+
+      // Due to the skip_to_next_sample block above, if we have an entry here it
+      // must be a ThreadId entry.
+      MOZ_ASSERT(e.Get().IsThreadId());
+
+      int threadId = e.Get().GetInt();
+      e.Next();
+
+      // Ignore samples that are for the wrong thread.
+      if (threadId != aThreadId && aThreadId != 0) {
+        continue;
+      }
+
+      MOZ_ASSERT(aThreadId != 0 || processedThreadId == 0,
+                 "aThreadId==0 should only be used with 1-sample buffer");
+
+      ProfileSample sample;
+
+      if (e.Has() && e.Get().IsTime()) {
+        sample.mTime = e.Get().GetDouble();
+        e.Next();
+
+        // Ignore samples that are too old.
+        if (sample.mTime < aSinceTime) {
+          continue;
+        }
+      } else {
+        ERROR_AND_CONTINUE("expected a Time entry");
+      }
+
+      UniqueStacks::StackKey stack =
+          aUniqueStacks.BeginStack(UniqueStacks::FrameKey("(root)"));
+
+      int numFrames = 0;
+      while (e.Has()) {
+        if (e.Get().IsNativeLeafAddr()) {
+          numFrames++;
+
+          void* pc = e.Get().GetPtr();
+          e.Next();
+
+          static const uint32_t BUF_SIZE = 256;
+          char buf[BUF_SIZE];
+
+          // Bug 753041: We need a double cast here to tell GCC that we don't
+          // want to sign extend 32-bit addresses starting with 0xFXXXXXX.
+          unsigned long long pcULL = (unsigned long long)(uintptr_t)pc;
+          SprintfLiteral(buf, "%#llx", pcULL);
+
+          // If the "MOZ_PROFILER_SYMBOLICATE" env-var is set, we add a local
+          // symbolication description to the PC address. This is off by
+          // default, and mainly intended for local development.
+          static const bool preSymbolicate = []() {
+            const char* symbolicate = getenv("MOZ_PROFILER_SYMBOLICATE");
+            return symbolicate && symbolicate[0] != '\0';
+          }();
+          if (preSymbolicate) {
+            MozCodeAddressDetails details;
+            if (MozDescribeCodeAddress(pc, &details)) {
+              // Replace \0 terminator with space.
+              const uint32_t pcLen = strlen(buf);
+              buf[pcLen] = ' ';
+              // Add description after space. Note: Using a frame number of 0,
+              // as using `numFrames` wouldn't help here, and would prevent
+              // combining same function calls that happen at different depths.
+              // TODO: Remove unsightly "#00: " if too annoying. :-)
+              MozFormatCodeAddressDetails(
+                  buf + pcLen + 1, BUF_SIZE - (pcLen + 1), 0, pc, &details);
+            }
+          }
+
+          stack = aUniqueStacks.AppendFrame(stack, UniqueStacks::FrameKey(buf));
+
+        } else if (e.Get().IsLabel()) {
+          numFrames++;
+
+          const char* label = e.Get().GetString();
+          e.Next();
+
+          using FrameFlags = ProfilingStackFrame::Flags;
+          uint32_t frameFlags = 0;
+          if (e.Has() && e.Get().IsFrameFlags()) {
+            frameFlags = uint32_t(e.Get().GetUint64());
+            e.Next();
+          }
+
+          bool relevantForJS =
+              frameFlags & uint32_t(FrameFlags::RELEVANT_FOR_JS);
+
+          // Copy potential dynamic string fragments into dynStrBuf, so that
+          // dynStrBuf will then contain the entire dynamic string.
+          size_t i = 0;
+          dynStrBuf[0] = '\0';
+          while (e.Has()) {
+            if (e.Get().IsDynamicStringFragment()) {
+              char chars[ProfileBufferEntry::kNumChars];
+              e.Get().CopyCharsInto(chars);
+              for (char c : chars) {
+                if (i < kMaxFrameKeyLength) {
+                  dynStrBuf[i] = c;
+                  i++;
+                }
+              }
+              e.Next();
+            } else {
+              break;
+            }
+          }
+          dynStrBuf[kMaxFrameKeyLength - 1] = '\0';
+          bool hasDynamicString = (i != 0);
+
+          std::string frameLabel;
+          if (label[0] != '\0' && hasDynamicString) {
+            if (frameFlags & uint32_t(FrameFlags::STRING_TEMPLATE_METHOD)) {
+              frameLabel += label;
+              frameLabel += '.';
+              frameLabel += dynStrBuf.get();
+            } else if (frameFlags &
+                       uint32_t(FrameFlags::STRING_TEMPLATE_GETTER)) {
+              frameLabel += "get ";
+              frameLabel += label;
+              frameLabel += '.';
+              frameLabel += dynStrBuf.get();
+            } else if (frameFlags &
+                       uint32_t(FrameFlags::STRING_TEMPLATE_SETTER)) {
+              frameLabel += "set ";
+              frameLabel += label;
+              frameLabel += '.';
+              frameLabel += dynStrBuf.get();
+            } else {
+              frameLabel += label;
+              frameLabel += ' ';
+              frameLabel += dynStrBuf.get();
+            }
+          } else if (hasDynamicString) {
+            frameLabel += dynStrBuf.get();
+          } else {
+            frameLabel += label;
+          }
+
+          uint64_t innerWindowID = 0;
+          if (e.Has() && e.Get().IsInnerWindowID()) {
+            innerWindowID = uint64_t(e.Get().GetUint64());
+            e.Next();
+          }
+
+          Maybe<unsigned> line;
+          if (e.Has() && e.Get().IsLineNumber()) {
+            line = Some(unsigned(e.Get().GetInt()));
+            e.Next();
+          }
+
+          Maybe<unsigned> column;
+          if (e.Has() && e.Get().IsColumnNumber()) {
+            column = Some(unsigned(e.Get().GetInt()));
+            e.Next();
+          }
+
+          Maybe<ProfilingCategoryPair> categoryPair;
+          if (e.Has() && e.Get().IsCategoryPair()) {
+            categoryPair =
+                Some(ProfilingCategoryPair(uint32_t(e.Get().GetInt())));
+            e.Next();
+          }
+
+          stack = aUniqueStacks.AppendFrame(
+              stack, UniqueStacks::FrameKey(std::move(frameLabel),
+                                            relevantForJS, innerWindowID, line,
+                                            column, categoryPair));
+
+        } else {
+          break;
+        }
+      }
+
+      if (numFrames == 0) {
+        // It is possible to have empty stacks if native stackwalking is
+        // disabled. Skip samples with empty stacks. (See Bug 1497985).
+        // Thus, don't use ERROR_AND_CONTINUE, but just continue.
+        continue;
+      }
+
+      sample.mStack = aUniqueStacks.GetOrAddStackIndex(stack);
+
+      if (e.Has() && e.Get().IsResponsiveness()) {
+        sample.mResponsiveness = Some(e.Get().GetDouble());
+        e.Next();
+      }
+
+      WriteSample(aWriter, sample);
+
+      processedThreadId = threadId;
+    }
+
+    return processedThreadId;
+  });
+}
+
+void ProfileBuffer::StreamMarkersToJSON(SpliceableJSONWriter& aWriter,
+                                        int aThreadId,
+                                        const TimeStamp& aProcessStartTime,
+                                        double aSinceTime,
+                                        UniqueStacks& aUniqueStacks) const {
+  mEntries.ReadEach([&](ProfileBufferEntryReader& aER) {
+    auto type = static_cast<ProfileBufferEntry::Kind>(
+        aER.ReadObject<ProfileBufferEntry::KindUnderlyingType>());
+    MOZ_ASSERT(static_cast<ProfileBufferEntry::KindUnderlyingType>(type) <
+               static_cast<ProfileBufferEntry::KindUnderlyingType>(
+                   ProfileBufferEntry::Kind::MODERN_LIMIT));
+    bool entryWasFullyRead = false;
+
+    if (type == ProfileBufferEntry::Kind::Marker) {
+      entryWasFullyRead = ::mozilla::base_profiler_markers_detail::
+          DeserializeAfterKindAndStream(
+              aER, aWriter, aThreadId,
+              [&](ProfileChunkedBuffer& aChunkedBuffer) {
+                ProfilerBacktrace backtrace("", &aChunkedBuffer);
+                backtrace.StreamJSON(aWriter, TimeStamp::ProcessCreation(),
+                                     aUniqueStacks);
+              });
+    }
+
+    if (!entryWasFullyRead) {
+      // Not a marker, or marker for another thread.
+      // We probably didn't read the whole entry, so we need to skip to the end.
+      aER.SetRemainingBytes(0);
+    }
+  });
+}
+
+void ProfileBuffer::StreamProfilerOverheadToJSON(
+    SpliceableJSONWriter& aWriter, const TimeStamp& aProcessStartTime,
+    double aSinceTime) const {
+  mEntries.Read([&](ProfileChunkedBuffer::Reader* aReader) {
+    MOZ_ASSERT(aReader,
+               "ProfileChunkedBuffer cannot be out-of-session when sampler is "
+               "running");
+
+    EntryGetter e(*aReader);
+
+    enum Schema : uint32_t {
+      TIME = 0,
+      LOCKING = 1,
+      MARKER_CLEANING = 2,
+      COUNTERS = 3,
+      THREADS = 4
+    };
+
+    aWriter.StartObjectProperty("profilerOverhead");
+    aWriter.StartObjectProperty("samples");
+    // Stream all sampling overhead data. We skip other entries, because we
+    // process them in StreamSamplesToJSON()/etc.
+    {
+      JSONSchemaWriter schema(aWriter);
+      schema.WriteField("time");
+      schema.WriteField("locking");
+      schema.WriteField("expiredMarkerCleaning");
+      schema.WriteField("counters");
+      schema.WriteField("threads");
+    }
+
+    aWriter.StartArrayProperty("data");
+    double firstTime = 0.0;
+    double lastTime = 0.0;
+    ProfilerStats intervals, overheads, lockings, cleanings, counters, threads;
+    while (e.Has()) {
+      // valid sequence: ProfilerOverheadTime, ProfilerOverheadDuration * 4
+      if (e.Get().IsProfilerOverheadTime()) {
+        double time = e.Get().GetDouble();
+        if (time >= aSinceTime) {
+          e.Next();
+          if (!e.Has() || !e.Get().IsProfilerOverheadDuration()) {
+            ERROR_AND_CONTINUE(
+                "expected a ProfilerOverheadDuration entry after "
+                "ProfilerOverheadTime");
+          }
+          double locking = e.Get().GetDouble();
+          e.Next();
+          if (!e.Has() || !e.Get().IsProfilerOverheadDuration()) {
+            ERROR_AND_CONTINUE(
+                "expected a ProfilerOverheadDuration entry after "
+                "ProfilerOverheadTime,ProfilerOverheadDuration");
+          }
+          double cleaning = e.Get().GetDouble();
+          e.Next();
+          if (!e.Has() || !e.Get().IsProfilerOverheadDuration()) {
+            ERROR_AND_CONTINUE(
+                "expected a ProfilerOverheadDuration entry after "
+                "ProfilerOverheadTime,ProfilerOverheadDuration*2");
+          }
+          double counter = e.Get().GetDouble();
+          e.Next();
+          if (!e.Has() || !e.Get().IsProfilerOverheadDuration()) {
+            ERROR_AND_CONTINUE(
+                "expected a ProfilerOverheadDuration entry after "
+                "ProfilerOverheadTime,ProfilerOverheadDuration*3");
+          }
+          double thread = e.Get().GetDouble();
+
+          if (firstTime == 0.0) {
+            firstTime = time;
+          } else {
+            // Note that we'll have 1 fewer interval than other numbers (because
+            // we need both ends of an interval to know its duration). The final
+            // difference should be insignificant over the expected many
+            // thousands of iterations.
+            intervals.Count(time - lastTime);
+          }
+          lastTime = time;
+          overheads.Count(locking + cleaning + counter + thread);
+          lockings.Count(locking);
+          cleanings.Count(cleaning);
+          counters.Count(counter);
+          threads.Count(thread);
+
+          AutoArraySchemaWriter writer(aWriter);
+          writer.DoubleElement(TIME, time);
+          writer.DoubleElement(LOCKING, locking);
+          writer.DoubleElement(MARKER_CLEANING, cleaning);
+          writer.DoubleElement(COUNTERS, counter);
+          writer.DoubleElement(THREADS, thread);
+        }
+      }
+      e.Next();
+    }
+    aWriter.EndArray();   // data
+    aWriter.EndObject();  // samples
+
+    // Only output statistics if there is at least one full interval (and
+    // therefore at least two samplings.)
+    if (intervals.n > 0) {
+      aWriter.StartObjectProperty("statistics");
+      aWriter.DoubleProperty("profiledDuration", lastTime - firstTime);
+      aWriter.IntProperty("samplingCount", overheads.n);
+      aWriter.DoubleProperty("overheadDurations", overheads.sum);
+      aWriter.DoubleProperty("overheadPercentage",
+                             overheads.sum / (lastTime - firstTime));
+#define PROFILER_STATS(name, var)                           \
+  aWriter.DoubleProperty("mean" name, (var).sum / (var).n); \
+  aWriter.DoubleProperty("min" name, (var).min);            \
+  aWriter.DoubleProperty("max" name, (var).max);
+      PROFILER_STATS("Interval", intervals);
+      PROFILER_STATS("Overhead", overheads);
+      PROFILER_STATS("Lockings", lockings);
+      PROFILER_STATS("Cleaning", cleanings);
+      PROFILER_STATS("Counter", counters);
+      PROFILER_STATS("Thread", threads);
+#undef PROFILER_STATS
+      aWriter.EndObject();  // statistics
+    }
+    aWriter.EndObject();  // profilerOverhead
+  });
+}
+
+struct CounterKeyedSample {
+  double mTime;
+  uint64_t mNumber;
+  int64_t mCount;
+};
+
+using CounterKeyedSamples = Vector<CounterKeyedSample>;
+
+using CounterMap = HashMap<uint64_t, CounterKeyedSamples>;
+
+// HashMap lookup, if not found, a default value is inserted.
+// Returns reference to (existing or new) value inside the HashMap.
+template <typename HashM, typename Key>
+static auto& LookupOrAdd(HashM& aMap, Key&& aKey) {
+  auto addPtr = aMap.lookupForAdd(aKey);
+  if (!addPtr) {
+    MOZ_RELEASE_ASSERT(aMap.add(addPtr, std::forward<Key>(aKey),
+                                typename HashM::Entry::ValueType{}));
+    MOZ_ASSERT(!!addPtr);
+  }
+  return addPtr->value();
+}
+
+void ProfileBuffer::StreamCountersToJSON(SpliceableJSONWriter& aWriter,
+                                         const TimeStamp& aProcessStartTime,
+                                         double aSinceTime) const {
+  // Because this is a format entirely internal to the Profiler, any parsing
+  // error indicates a bug in the ProfileBuffer writing or the parser itself,
+  // or possibly flaky hardware.
+
+  mEntries.Read([&](ProfileChunkedBuffer::Reader* aReader) {
+    MOZ_ASSERT(aReader,
+               "ProfileChunkedBuffer cannot be out-of-session when sampler is "
+               "running");
+
+    EntryGetter e(*aReader);
+
+    enum Schema : uint32_t { TIME = 0, NUMBER = 1, COUNT = 2 };
+
+    // Stream all counters. We skip other entries, because we process them in
+    // StreamSamplesToJSON()/etc.
+    //
+    // Valid sequence in the buffer:
+    // CounterID
+    // Time
+    // ( CounterKey Count Number? )*
+    //
+    // And the JSON (example):
+    // "counters": {
+    //  "name": "malloc",
+    //  "category": "Memory",
+    //  "description": "Amount of allocated memory",
+    //  "sample_groups": {
+    //   "id": 0,
+    //   "samples": {
+    //    "schema": {"time": 0, "number": 1, "count": 2},
+    //    "data": [
+    //     [
+    //      16117.033968000002,
+    //      2446216,
+    //      6801320
+    //     ],
+    //     [
+    //      16118.037638,
+    //      2446216,
+    //      6801320
+    //     ],
+    //    ],
+    //   }
+    //  }
+    // },
+
+    // Build the map of counters and populate it
+    HashMap<void*, CounterMap> counters;
+
+    while (e.Has()) {
+      // skip all non-Counters, including if we start in the middle of a counter
+      if (e.Get().IsCounterId()) {
+        void* id = e.Get().GetPtr();
+        CounterMap& counter = LookupOrAdd(counters, id);
+        e.Next();
+        if (!e.Has() || !e.Get().IsTime()) {
+          ERROR_AND_CONTINUE("expected a Time entry");
+        }
+        double time = e.Get().GetDouble();
+        if (time >= aSinceTime) {
+          e.Next();
+          while (e.Has() && e.Get().IsCounterKey()) {
+            uint64_t key = e.Get().GetUint64();
+            CounterKeyedSamples& data = LookupOrAdd(counter, key);
+            e.Next();
+            if (!e.Has() || !e.Get().IsCount()) {
+              ERROR_AND_CONTINUE("expected a Count entry");
+            }
+            int64_t count = e.Get().GetUint64();
+            e.Next();
+            uint64_t number;
+            if (!e.Has() || !e.Get().IsNumber()) {
+              number = 0;
+            } else {
+              number = e.Get().GetInt64();
+            }
+            CounterKeyedSample sample = {time, number, count};
+            MOZ_RELEASE_ASSERT(data.append(sample));
+          }
+        } else {
+          // skip counter sample - only need to skip the initial counter
+          // id, then let the loop at the top skip the rest
+        }
+      }
+      e.Next();
+    }
+    // we have a map of a map of counter entries; dump them to JSON
+    if (counters.count() == 0) {
+      return;
+    }
+
+    aWriter.StartArrayProperty("counters");
+    for (auto iter = counters.iter(); !iter.done(); iter.next()) {
+      CounterMap& counter = iter.get().value();
+      const BaseProfilerCount* base_counter =
+          static_cast<const BaseProfilerCount*>(iter.get().key());
+
+      aWriter.Start();
+      aWriter.StringProperty("name", MakeStringSpan(base_counter->mLabel));
+      aWriter.StringProperty("category",
+                             MakeStringSpan(base_counter->mCategory));
+      aWriter.StringProperty("description",
+                             MakeStringSpan(base_counter->mDescription));
+
+      aWriter.StartArrayProperty("sample_groups");
+      for (auto counter_iter = counter.iter(); !counter_iter.done();
+           counter_iter.next()) {
+        CounterKeyedSamples& samples = counter_iter.get().value();
+        uint64_t key = counter_iter.get().key();
+
+        size_t size = samples.length();
+        if (size == 0) {
+          continue;
+        }
+
+        aWriter.StartObjectElement();
+        {
+          aWriter.IntProperty("id", static_cast<int64_t>(key));
+          aWriter.StartObjectProperty("samples");
+          {
+            // XXX Can we assume a missing count means 0?
+            JSONSchemaWriter schema(aWriter);
+            schema.WriteField("time");
+            schema.WriteField("number");
+            schema.WriteField("count");
+          }
+
+          aWriter.StartArrayProperty("data");
+          uint64_t previousNumber = 0;
+          int64_t previousCount = 0;
+          for (size_t i = 0; i < size; i++) {
+            // Encode as deltas, and only encode if different than the last
+            // sample
+            if (i == 0 || samples[i].mNumber != previousNumber ||
+                samples[i].mCount != previousCount) {
+              MOZ_ASSERT(i == 0 || samples[i].mTime >= samples[i - 1].mTime);
+              MOZ_ASSERT(samples[i].mNumber >= previousNumber);
+              MOZ_ASSERT(samples[i].mNumber - previousNumber <=
+                         uint64_t(std::numeric_limits<int64_t>::max()));
+
+              AutoArraySchemaWriter writer(aWriter);
+              writer.DoubleElement(TIME, samples[i].mTime);
+              writer.IntElement(
+                  NUMBER,
+                  static_cast<int64_t>(samples[i].mNumber - previousNumber));
+              writer.IntElement(COUNT, samples[i].mCount - previousCount);
+              previousNumber = samples[i].mNumber;
+              previousCount = samples[i].mCount;
+            }
+          }
+          aWriter.EndArray();   // data
+          aWriter.EndObject();  // samples
+        }
+        aWriter.EndObject();  // sample_groups item
+      }
+      aWriter.EndArray();  // sample groups
+      aWriter.End();       // for each counter
+    }
+    aWriter.EndArray();  // counters
+  });
+}
+
+#undef ERROR_AND_CONTINUE
+
+static void AddPausedRange(SpliceableJSONWriter& aWriter, const char* aReason,
+                           const Maybe<double>& aStartTime,
+                           const Maybe<double>& aEndTime) {
+  aWriter.Start();
+  if (aStartTime) {
+    aWriter.DoubleProperty("startTime", *aStartTime);
+  } else {
+    aWriter.NullProperty("startTime");
+  }
+  if (aEndTime) {
+    aWriter.DoubleProperty("endTime", *aEndTime);
+  } else {
+    aWriter.NullProperty("endTime");
+  }
+  aWriter.StringProperty("reason", MakeStringSpan(aReason));
+  aWriter.End();
+}
+
+void ProfileBuffer::StreamPausedRangesToJSON(SpliceableJSONWriter& aWriter,
+                                             double aSinceTime) const {
+  mEntries.Read([&](ProfileChunkedBuffer::Reader* aReader) {
+    MOZ_ASSERT(aReader,
+               "ProfileChunkedBuffer cannot be out-of-session when sampler is "
+               "running");
+
+    EntryGetter e(*aReader);
+
+    Maybe<double> currentPauseStartTime;
+    Maybe<double> currentCollectionStartTime;
+
+    while (e.Has()) {
+      if (e.Get().IsPause()) {
+        currentPauseStartTime = Some(e.Get().GetDouble());
+      } else if (e.Get().IsResume()) {
+        AddPausedRange(aWriter, "profiler-paused", currentPauseStartTime,
+                       Some(e.Get().GetDouble()));
+        currentPauseStartTime = Nothing();
+      } else if (e.Get().IsCollectionStart()) {
+        currentCollectionStartTime = Some(e.Get().GetDouble());
+      } else if (e.Get().IsCollectionEnd()) {
+        AddPausedRange(aWriter, "collecting", currentCollectionStartTime,
+                       Some(e.Get().GetDouble()));
+        currentCollectionStartTime = Nothing();
+      }
+      e.Next();
+    }
+
+    if (currentPauseStartTime) {
+      AddPausedRange(aWriter, "profiler-paused", currentPauseStartTime,
+                     Nothing());
+    }
+    if (currentCollectionStartTime) {
+      AddPausedRange(aWriter, "collecting", currentCollectionStartTime,
+                     Nothing());
+    }
+  });
+}
+
+bool ProfileBuffer::DuplicateLastSample(int aThreadId,
+                                        const TimeStamp& aProcessStartTime,
+                                        Maybe<uint64_t>& aLastSample) {
+  if (!aLastSample) {
+    return false;
+  }
+
+  ProfileChunkedBuffer tempBuffer(
+      ProfileChunkedBuffer::ThreadSafety::WithoutMutex, mWorkerChunkManager);
+
+  auto retrieveWorkerChunk = MakeScopeExit(
+      [&]() { mWorkerChunkManager.Reset(tempBuffer.GetAllChunks()); });
+
+  const bool ok = mEntries.Read([&](ProfileChunkedBuffer::Reader* aReader) {
+    MOZ_ASSERT(aReader,
+               "ProfileChunkedBuffer cannot be out-of-session when sampler is "
+               "running");
+
+    EntryGetter e(*aReader, *aLastSample);
+
+    if (e.CurPos() != *aLastSample) {
+      // The last sample is no longer within the buffer range, so we cannot
+      // use it. Reset the stored buffer position to Nothing().
+      aLastSample.reset();
+      return false;
+    }
+
+    MOZ_RELEASE_ASSERT(e.Has() && e.Get().IsThreadId() &&
+                       e.Get().GetInt() == aThreadId);
+
+    e.Next();
+
+    // Go through the whole entry and duplicate it, until we find the next
+    // one.
+    while (e.Has()) {
+      switch (e.Get().GetKind()) {
+        case ProfileBufferEntry::Kind::Pause:
+        case ProfileBufferEntry::Kind::Resume:
+        case ProfileBufferEntry::Kind::PauseSampling:
+        case ProfileBufferEntry::Kind::ResumeSampling:
+        case ProfileBufferEntry::Kind::CollectionStart:
+        case ProfileBufferEntry::Kind::CollectionEnd:
+        case ProfileBufferEntry::Kind::ThreadId:
+          // We're done.
+          return true;
+        case ProfileBufferEntry::Kind::Time:
+          // Copy with new time
+          AddEntry(tempBuffer,
+                   ProfileBufferEntry::Time(
+                       (TimeStamp::NowUnfuzzed() - aProcessStartTime)
+                           .ToMilliseconds()));
+          break;
+        case ProfileBufferEntry::Kind::CounterKey:
+        case ProfileBufferEntry::Kind::Number:
+        case ProfileBufferEntry::Kind::Count:
+        case ProfileBufferEntry::Kind::Responsiveness:
+          // Don't copy anything not part of a thread's stack sample
+          break;
+        case ProfileBufferEntry::Kind::CounterId:
+          // CounterId is normally followed by Time - if so, we'd like
+          // to skip it.  If we duplicate Time, it won't hurt anything, just
+          // waste buffer space (and this can happen if the CounterId has
+          // fallen off the end of the buffer, but Time (and Number/Count)
+          // are still in the buffer).
+          e.Next();
+          if (e.Has() && e.Get().GetKind() != ProfileBufferEntry::Kind::Time) {
+            // this would only happen if there was an invalid sequence
+            // in the buffer.  Don't skip it.
+            continue;
+          }
+          // we've skipped Time
+          break;
+        case ProfileBufferEntry::Kind::ProfilerOverheadTime:
+          // ProfilerOverheadTime is normally followed by
+          // ProfilerOverheadDuration*4 - if so, we'd like to skip it. Don't
+          // duplicate, as we are in the middle of a sampling and will soon
+          // capture its own overhead.
+          e.Next();
+          // A missing Time would only happen if there was an invalid
+          // sequence in the buffer. Don't skip unexpected entry.
+          if (e.Has() &&
+              e.Get().GetKind() !=
+                  ProfileBufferEntry::Kind::ProfilerOverheadDuration) {
+            continue;
+          }
+          e.Next();
+          if (e.Has() &&
+              e.Get().GetKind() !=
+                  ProfileBufferEntry::Kind::ProfilerOverheadDuration) {
+            continue;
+          }
+          e.Next();
+          if (e.Has() &&
+              e.Get().GetKind() !=
+                  ProfileBufferEntry::Kind::ProfilerOverheadDuration) {
+            continue;
+          }
+          e.Next();
+          if (e.Has() &&
+              e.Get().GetKind() !=
+                  ProfileBufferEntry::Kind::ProfilerOverheadDuration) {
+            continue;
+          }
+          // we've skipped ProfilerOverheadTime and
+          // ProfilerOverheadDuration*4.
+          break;
+        default: {
+          // Copy anything else we don't know about.
+          AddEntry(tempBuffer, e.Get());
+          break;
+        }
+      }
+      e.Next();
+    }
+    return true;
+  });
+
+  if (!ok) {
+    return false;
+  }
+
+  // If the buffer was big enough, there won't be any cleared blocks.
+  if (tempBuffer.GetState().mClearedBlockCount != 0) {
+    // No need to try to read stack again as it won't fit. Reset the stored
+    // buffer position to Nothing().
+    aLastSample.reset();
+    return false;
+  }
+
+  aLastSample = Some(AddThreadIdEntry(aThreadId));
+
+  tempBuffer.Read([&](ProfileChunkedBuffer::Reader* aReader) {
+    MOZ_ASSERT(aReader, "tempBuffer cannot be out-of-session");
+
+    EntryGetter e(*aReader);
+
+    while (e.Has()) {
+      AddEntry(e.Get());
+      e.Next();
+    }
+  });
+
+  return true;
+}
+
+void ProfileBuffer::DiscardSamplesBeforeTime(double aTime) {
+  // This function does nothing!
+  // The duration limit will be removed from Firefox, see bug 1632365.
+  Unused << aTime;
+}
+
+// END ProfileBuffer
+////////////////////////////////////////////////////////////////////////
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/ProfileBufferEntry.h b/mozglue/baseprofiler/core/ProfileBufferEntry.h
new file mode 100644
index 0000000000..6422a34a85
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfileBufferEntry.h
@@ -0,0 +1,358 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferEntry_h
+#define ProfileBufferEntry_h
+
+#include "BaseProfilingCategory.h"
+#include "gtest/MozGtestFriend.h"
+#include "mozilla/BaseProfileJSONWriter.h"
+#include "mozilla/HashFunctions.h"
+#include "mozilla/HashTable.h"
+#include "mozilla/Maybe.h"
+#include "mozilla/ProfileBufferEntryKinds.h"
+#include "mozilla/UniquePtr.h"
+#include "mozilla/Variant.h"
+#include "mozilla/Vector.h"
+
+#include <string>
+
+namespace mozilla {
+namespace baseprofiler {
+
+class ProfileBufferEntry {
+ public:
+  using KindUnderlyingType = ::mozilla::ProfileBufferEntryKindUnderlyingType;
+  using Kind = ::mozilla::ProfileBufferEntryKind;
+
+  ProfileBufferEntry();
+
+  static constexpr size_t kNumChars = ::mozilla::ProfileBufferEntryNumChars;
+
+ private:
+  // aString must be a static string.
+  ProfileBufferEntry(Kind aKind, const char* aString);
+  ProfileBufferEntry(Kind aKind, char aChars[kNumChars]);
+  ProfileBufferEntry(Kind aKind, void* aPtr);
+  ProfileBufferEntry(Kind aKind, double aDouble);
+  ProfileBufferEntry(Kind aKind, int64_t aInt64);
+  ProfileBufferEntry(Kind aKind, uint64_t aUint64);
+  ProfileBufferEntry(Kind aKind, int aInt);
+
+ public:
+#define CTOR(KIND, TYPE, SIZE)                   \
+  static ProfileBufferEntry KIND(TYPE aVal) {    \
+    return ProfileBufferEntry(Kind::KIND, aVal); \
+  }
+  FOR_EACH_PROFILE_BUFFER_ENTRY_KIND(CTOR)
+#undef CTOR
+
+  Kind GetKind() const { return mKind; }
+
+#define IS_KIND(KIND, TYPE, SIZE) \
+  bool Is##KIND() const { return mKind == Kind::KIND; }
+  FOR_EACH_PROFILE_BUFFER_ENTRY_KIND(IS_KIND)
+#undef IS_KIND
+
+ private:
+  FRIEND_TEST(ThreadProfile, InsertOneEntry);
+  FRIEND_TEST(ThreadProfile, InsertOneEntryWithTinyBuffer);
+  FRIEND_TEST(ThreadProfile, InsertEntriesNoWrap);
+  FRIEND_TEST(ThreadProfile, InsertEntriesWrap);
+  FRIEND_TEST(ThreadProfile, MemoryMeasure);
+  friend class ProfileBuffer;
+
+  Kind mKind;
+  uint8_t mStorage[kNumChars];
+
+  const char* GetString() const;
+  void* GetPtr() const;
+  double GetDouble() const;
+  int GetInt() const;
+  int64_t GetInt64() const;
+  uint64_t GetUint64() const;
+  void CopyCharsInto(char (&aOutArray)[kNumChars]) const;
+};
+
+// Packed layout: 1 byte for the tag + 8 bytes for the value.
+static_assert(sizeof(ProfileBufferEntry) == 9, "bad ProfileBufferEntry size");
+
+class UniqueStacks {
+ public:
+  struct FrameKey {
+    explicit FrameKey(const char* aLocation)
+        : mData(NormalFrameData{std::string(aLocation), false, 0, Nothing(),
+                                Nothing()}) {}
+
+    FrameKey(std::string&& aLocation, bool aRelevantForJS,
+             uint64_t aInnerWindowID, const Maybe<unsigned>& aLine,
+             const Maybe<unsigned>& aColumn,
+             const Maybe<ProfilingCategoryPair>& aCategoryPair)
+        : mData(NormalFrameData{aLocation, aRelevantForJS, aInnerWindowID,
+                                aLine, aColumn, aCategoryPair}) {}
+
+    FrameKey(const FrameKey& aToCopy) = default;
+
+    uint32_t Hash() const;
+    bool operator==(const FrameKey& aOther) const {
+      return mData == aOther.mData;
+    }
+
+    struct NormalFrameData {
+      bool operator==(const NormalFrameData& aOther) const;
+
+      std::string mLocation;
+      bool mRelevantForJS;
+      uint64_t mInnerWindowID;
+      Maybe<unsigned> mLine;
+      Maybe<unsigned> mColumn;
+      Maybe<ProfilingCategoryPair> mCategoryPair;
+    };
+    Variant<NormalFrameData> mData;
+  };
+
+  struct FrameKeyHasher {
+    using Lookup = FrameKey;
+
+    static HashNumber hash(const FrameKey& aLookup) {
+      HashNumber hash = 0;
+      if (aLookup.mData.is<FrameKey::NormalFrameData>()) {
+        const FrameKey::NormalFrameData& data =
+            aLookup.mData.as<FrameKey::NormalFrameData>();
+        if (!data.mLocation.empty()) {
+          hash = AddToHash(hash, HashString(data.mLocation.c_str()));
+        }
+        hash = AddToHash(hash, data.mRelevantForJS);
+        hash = mozilla::AddToHash(hash, data.mInnerWindowID);
+        if (data.mLine.isSome()) {
+          hash = AddToHash(hash, *data.mLine);
+        }
+        if (data.mColumn.isSome()) {
+          hash = AddToHash(hash, *data.mColumn);
+        }
+        if (data.mCategoryPair.isSome()) {
+          hash = AddToHash(hash, static_cast<uint32_t>(*data.mCategoryPair));
+        }
+      }
+      return hash;
+    }
+
+    static bool match(const FrameKey& aKey, const FrameKey& aLookup) {
+      return aKey == aLookup;
+    }
+
+    static void rekey(FrameKey& aKey, const FrameKey& aNewKey) {
+      aKey = aNewKey;
+    }
+  };
+
+  struct StackKey {
+    Maybe<uint32_t> mPrefixStackIndex;
+    uint32_t mFrameIndex;
+
+    explicit StackKey(uint32_t aFrame)
+        : mFrameIndex(aFrame), mHash(HashGeneric(aFrame)) {}
+
+    StackKey(const StackKey& aPrefix, uint32_t aPrefixStackIndex,
+             uint32_t aFrame)
+        : mPrefixStackIndex(Some(aPrefixStackIndex)),
+          mFrameIndex(aFrame),
+          mHash(AddToHash(aPrefix.mHash, aFrame)) {}
+
+    HashNumber Hash() const { return mHash; }
+
+    bool operator==(const StackKey& aOther) const {
+      return mPrefixStackIndex == aOther.mPrefixStackIndex &&
+             mFrameIndex == aOther.mFrameIndex;
+    }
+
+   private:
+    HashNumber mHash;
+  };
+
+  struct StackKeyHasher {
+    using Lookup = StackKey;
+
+    static HashNumber hash(const StackKey& aLookup) { return aLookup.Hash(); }
+
+    static bool match(const StackKey& aKey, const StackKey& aLookup) {
+      return aKey == aLookup;
+    }
+
+    static void rekey(StackKey& aKey, const StackKey& aNewKey) {
+      aKey = aNewKey;
+    }
+  };
+
+  UniqueStacks();
+
+  // Return a StackKey for aFrame as the stack's root frame (no prefix).
+  [[nodiscard]] StackKey BeginStack(const FrameKey& aFrame);
+
+  // Return a new StackKey that is obtained by appending aFrame to aStack.
+  [[nodiscard]] StackKey AppendFrame(const StackKey& aStack,
+                                     const FrameKey& aFrame);
+
+  [[nodiscard]] uint32_t GetOrAddFrameIndex(const FrameKey& aFrame);
+  [[nodiscard]] uint32_t GetOrAddStackIndex(const StackKey& aStack);
+
+  void SpliceFrameTableElements(SpliceableJSONWriter& aWriter);
+  void SpliceStackTableElements(SpliceableJSONWriter& aWriter);
+
+ private:
+  void StreamNonJITFrame(const FrameKey& aFrame);
+  void StreamStack(const StackKey& aStack);
+
+ public:
+  UniquePtr<UniqueJSONStrings> mUniqueStrings;
+
+ private:
+  SpliceableChunkedJSONWriter mFrameTableWriter;
+  HashMap<FrameKey, uint32_t, FrameKeyHasher> mFrameToIndexMap;
+
+  SpliceableChunkedJSONWriter mStackTableWriter;
+  HashMap<StackKey, uint32_t, StackKeyHasher> mStackToIndexMap;
+};
+
+//
+// Thread profile JSON Format
+// --------------------------
+//
+// The profile contains much duplicate information. The output JSON of the
+// profile attempts to deduplicate strings, frames, and stack prefixes, to cut
+// down on size and to increase JSON streaming speed. Deduplicated values are
+// streamed as indices into their respective tables.
+//
+// Further, arrays of objects with the same set of properties (e.g., samples,
+// frames) are output as arrays according to a schema instead of an object
+// with property names. A property that is not present is represented in the
+// array as null or undefined.
+//
+// The format of the thread profile JSON is shown by the following example
+// with 1 sample and 1 marker:
+//
+// {
+//   "name": "Foo",
+//   "tid": 42,
+//   "samples":
+//   {
+//     "schema":
+//     {
+//       "stack": 0,          /* index into stackTable */
+//       "time": 1,           /* number */
+//       "eventDelay": 2, /* number */
+//     },
+//     "data":
+//     [
+//       [ 1, 0.0, 0.0 ]      /* { stack: 1, time: 0.0, eventDelay: 0.0 } */
+//     ]
+//   },
+//
+//   "markers":
+//   {
+//     "schema":
+//     {
+//       "name": 0,           /* index into stringTable */
+//       "time": 1,           /* number */
+//       "data": 2            /* arbitrary JSON */
+//     },
+//     "data":
+//     [
+//       [ 3, 0.1 ]           /* { name: 'example marker', time: 0.1 } */
+//     ]
+//   },
+//
+//   "stackTable":
+//   {
+//     "schema":
+//     {
+//       "prefix": 0,         /* index into stackTable */
+//       "frame": 1           /* index into frameTable */
+//     },
+//     "data":
+//     [
+//       [ null, 0 ],         /* (root) */
+//       [ 0,    1 ]          /* (root) > foo.js */
+//     ]
+//   },
+//
+//   "frameTable":
+//   {
+//     "schema":
+//     {
+//       "location": 0,       /* index into stringTable */
+//       "relevantForJS": 1,  /* bool */
+//       "innerWindowID": 2,  /* inner window ID of global JS `window` object */
+//       "implementation": 3, /* index into stringTable */
+//       "optimizations": 4,  /* arbitrary JSON */
+//       "line": 5,           /* number */
+//       "column": 6,         /* number */
+//       "category": 7,       /* index into profile.meta.categories */
+//       "subcategory": 8     /* index into
+//       profile.meta.categories[category].subcategories */
+//     },
+//     "data":
+//     [
+//       [ 0 ],               /* { location: '(root)' } */
+//       [ 1, 2 ]             /* { location: 'foo.js',
+//                                 implementation: 'baseline' } */
+//     ]
+//   },
+//
+//   "stringTable":
+//   [
+//     "(root)",
+//     "foo.js",
+//     "baseline",
+//     "example marker"
+//   ]
+// }
+//
+// Process:
+// {
+//   "name": "Bar",
+//   "pid": 24,
+//   "threads":
+//   [
+//     <0-N threads from above>
+//   ],
+//   "counters": /* includes the memory counter */
+//   [
+//     {
+//       "name": "qwerty",
+//       "category": "uiop",
+//       "description": "this is qwerty uiop",
+//       "sample_groups:
+//       [
+//         {
+//           "id": 42, /* number (thread id, or object identifier (tab), etc) */
+//           "samples:
+//           {
+//             "schema":
+//             {
+//               "time": 1,   /* number */
+//               "number": 2, /* number (of times the counter was touched) */
+//               "count": 3   /* number (total for the counter) */
+//             },
+//             "data":
+//             [
+//               [ 0.1, 1824,
+//                 454622 ]   /* { time: 0.1, number: 1824, count: 454622 } */
+//             ]
+//           },
+//         },
+//         /* more sample-group objects with different id's */
+//       ]
+//     },
+//     /* more counters */
+//   ],
+// }
+//
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif /* ndef ProfileBufferEntry_h */
diff --git a/mozglue/baseprofiler/core/ProfileJSONWriter.cpp b/mozglue/baseprofiler/core/ProfileJSONWriter.cpp
new file mode 100644
index 0000000000..966ff2f515
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfileJSONWriter.cpp
@@ -0,0 +1,52 @@
+/* -*- Mode: C++; tab-width: 2; 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/BaseProfileJSONWriter.h"
+
+namespace mozilla::baseprofiler {
+
+UniqueJSONStrings::UniqueJSONStrings(JSONWriter::CollectionStyle aStyle) {
+  mStringTableWriter.StartBareList(aStyle);
+}
+
+UniqueJSONStrings::UniqueJSONStrings(const UniqueJSONStrings& aOther,
+                                     JSONWriter::CollectionStyle aStyle) {
+  mStringTableWriter.StartBareList(aStyle);
+  uint32_t count = aOther.mStringHashToIndexMap.count();
+  if (count != 0) {
+    MOZ_RELEASE_ASSERT(mStringHashToIndexMap.reserve(count));
+    for (auto iter = aOther.mStringHashToIndexMap.iter(); !iter.done();
+         iter.next()) {
+      mStringHashToIndexMap.putNewInfallible(iter.get().key(),
+                                             iter.get().value());
+    }
+    mStringTableWriter.CopyAndSplice(
+        aOther.mStringTableWriter.ChunkedWriteFunc());
+  }
+}
+
+UniqueJSONStrings::~UniqueJSONStrings() = default;
+
+void UniqueJSONStrings::SpliceStringTableElements(
+    SpliceableJSONWriter& aWriter) {
+  aWriter.TakeAndSplice(mStringTableWriter.TakeChunkedWriteFunc());
+}
+
+uint32_t UniqueJSONStrings::GetOrAddIndex(const Span<const char>& aStr) {
+  uint32_t count = mStringHashToIndexMap.count();
+  HashNumber hash = HashString(aStr.data(), aStr.size());
+  auto entry = mStringHashToIndexMap.lookupForAdd(hash);
+  if (entry) {
+    MOZ_ASSERT(entry->value() < count);
+    return entry->value();
+  }
+
+  MOZ_RELEASE_ASSERT(mStringHashToIndexMap.add(entry, hash, count));
+  mStringTableWriter.StringElement(aStr);
+  return count;
+}
+
+}  // namespace mozilla::baseprofiler
diff --git a/mozglue/baseprofiler/core/ProfiledThreadData.cpp b/mozglue/baseprofiler/core/ProfiledThreadData.cpp
new file mode 100644
index 0000000000..4dc600d97c
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfiledThreadData.cpp
@@ -0,0 +1,187 @@
+/* -*- Mode: C++; tab-width: 2; 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 "ProfiledThreadData.h"
+
+#include "BaseProfiler.h"
+#include "ProfileBuffer.h"
+
+#include "mozilla/BaseProfileJSONWriter.h"
+
+#if defined(GP_OS_darwin)
+#  include <pthread.h>
+#endif
+
+namespace mozilla {
+namespace baseprofiler {
+
+ProfiledThreadData::ProfiledThreadData(ThreadInfo* aThreadInfo)
+    : mThreadInfo(aThreadInfo) {}
+
+ProfiledThreadData::~ProfiledThreadData() {}
+
+void ProfiledThreadData::StreamJSON(const ProfileBuffer& aBuffer,
+                                    SpliceableJSONWriter& aWriter,
+                                    const std::string& aProcessName,
+                                    const std::string& aETLDplus1,
+                                    const TimeStamp& aProcessStartTime,
+                                    double aSinceTime) {
+  UniqueStacks uniqueStacks;
+
+  MOZ_ASSERT(uniqueStacks.mUniqueStrings);
+  aWriter.SetUniqueStrings(*uniqueStacks.mUniqueStrings);
+
+  aWriter.Start();
+  {
+    StreamSamplesAndMarkers(mThreadInfo->Name(), mThreadInfo->ThreadId(),
+                            aBuffer, aWriter, aProcessName, aETLDplus1,
+                            aProcessStartTime, mThreadInfo->RegisterTime(),
+                            mUnregisterTime, aSinceTime, uniqueStacks);
+
+    aWriter.StartObjectProperty("stackTable");
+    {
+      {
+        JSONSchemaWriter schema(aWriter);
+        schema.WriteField("prefix");
+        schema.WriteField("frame");
+      }
+
+      aWriter.StartArrayProperty("data");
+      { uniqueStacks.SpliceStackTableElements(aWriter); }
+      aWriter.EndArray();
+    }
+    aWriter.EndObject();
+
+    aWriter.StartObjectProperty("frameTable");
+    {
+      {
+        JSONSchemaWriter schema(aWriter);
+        schema.WriteField("location");
+        schema.WriteField("relevantForJS");
+        schema.WriteField("innerWindowID");
+        schema.WriteField("implementation");
+        schema.WriteField("optimizations");
+        schema.WriteField("line");
+        schema.WriteField("column");
+        schema.WriteField("category");
+        schema.WriteField("subcategory");
+      }
+
+      aWriter.StartArrayProperty("data");
+      { uniqueStacks.SpliceFrameTableElements(aWriter); }
+      aWriter.EndArray();
+    }
+    aWriter.EndObject();
+
+    aWriter.StartArrayProperty("stringTable");
+    {
+      std::move(*uniqueStacks.mUniqueStrings)
+          .SpliceStringTableElements(aWriter);
+    }
+    aWriter.EndArray();
+  }
+  aWriter.End();
+
+  aWriter.ResetUniqueStrings();
+}
+
+int StreamSamplesAndMarkers(
+    const char* aName, int aThreadId, const ProfileBuffer& aBuffer,
+    SpliceableJSONWriter& aWriter, const std::string& aProcessName,
+    const std::string& aETLDplus1, const TimeStamp& aProcessStartTime,
+    const TimeStamp& aRegisterTime, const TimeStamp& aUnregisterTime,
+    double aSinceTime, UniqueStacks& aUniqueStacks) {
+  int processedThreadId = 0;
+
+  aWriter.StringProperty(
+      "processType",
+      "(unknown)" /* XRE_GeckoProcessTypeToString(XRE_GetProcessType()) */);
+
+  {
+    std::string name = aName;
+    // We currently need to distinguish threads output by Base Profiler from
+    // those in Gecko Profiler, as the frontend could get confused and lose
+    // tracks with the same name.
+    // TODO: As part of the profilers de-duplication, thread data from both
+    // profilers should end up in the same track, at which point this won't be
+    // necessary anymore. See meta bug 1557566.
+    name += " (pre-xul)";
+    aWriter.StringProperty("name", name);
+  }
+
+  // Use given process name (if any).
+  if (!aProcessName.empty()) {
+    aWriter.StringProperty("processName", aProcessName);
+  }
+  if (!aETLDplus1.empty()) {
+    aWriter.StringProperty("eTLD+1", aETLDplus1);
+  }
+
+  if (aRegisterTime) {
+    aWriter.DoubleProperty(
+        "registerTime", (aRegisterTime - aProcessStartTime).ToMilliseconds());
+  } else {
+    aWriter.NullProperty("registerTime");
+  }
+
+  if (aUnregisterTime) {
+    aWriter.DoubleProperty(
+        "unregisterTime",
+        (aUnregisterTime - aProcessStartTime).ToMilliseconds());
+  } else {
+    aWriter.NullProperty("unregisterTime");
+  }
+
+  aWriter.StartObjectProperty("samples");
+  {
+    {
+      JSONSchemaWriter schema(aWriter);
+      schema.WriteField("stack");
+      schema.WriteField("time");
+      schema.WriteField("eventDelay");
+    }
+
+    aWriter.StartArrayProperty("data");
+    {
+      processedThreadId = aBuffer.StreamSamplesToJSON(
+          aWriter, aThreadId, aSinceTime, aUniqueStacks);
+    }
+    aWriter.EndArray();
+  }
+  aWriter.EndObject();
+
+  aWriter.StartObjectProperty("markers");
+  {
+    {
+      JSONSchemaWriter schema(aWriter);
+      schema.WriteField("name");
+      schema.WriteField("startTime");
+      schema.WriteField("endTime");
+      schema.WriteField("phase");
+      schema.WriteField("category");
+      schema.WriteField("data");
+    }
+
+    aWriter.StartArrayProperty("data");
+    {
+      aBuffer.StreamMarkersToJSON(aWriter, aThreadId, aProcessStartTime,
+                                  aSinceTime, aUniqueStacks);
+    }
+    aWriter.EndArray();
+  }
+  aWriter.EndObject();
+
+  aWriter.IntProperty("pid",
+                      static_cast<int64_t>(profiler_current_process_id()));
+  aWriter.IntProperty(
+      "tid",
+      static_cast<int64_t>(aThreadId != 0 ? aThreadId : processedThreadId));
+
+  return processedThreadId;
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/ProfiledThreadData.h b/mozglue/baseprofiler/core/ProfiledThreadData.h
new file mode 100644
index 0000000000..c45c02a7bb
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfiledThreadData.h
@@ -0,0 +1,119 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfiledThreadData_h
+#define ProfiledThreadData_h
+
+#include "BaseProfilingStack.h"
+#include "platform.h"
+#include "ProfileBufferEntry.h"
+#include "ThreadInfo.h"
+
+#include "mozilla/RefPtr.h"
+#include "mozilla/TimeStamp.h"
+#include "mozilla/UniquePtr.h"
+
+#include <string>
+
+namespace mozilla {
+namespace baseprofiler {
+
+class ProfileBuffer;
+
+// This class contains information about a thread that is only relevant while
+// the profiler is running, for any threads (both alive and dead) whose thread
+// name matches the "thread filter" in the current profiler run.
+// ProfiledThreadData objects may be kept alive even after the thread is
+// unregistered, as long as there is still data for that thread in the profiler
+// buffer.
+//
+// Accesses to this class are protected by the profiler state lock.
+//
+// Created as soon as the following are true for the thread:
+//  - The profiler is running, and
+//  - the thread matches the profiler's thread filter, and
+//  - the thread is registered with the profiler.
+// So it gets created in response to either (1) the profiler being started (for
+// an existing registered thread) or (2) the thread being registered (if the
+// profiler is already running).
+//
+// The thread may be unregistered during the lifetime of ProfiledThreadData.
+// If that happens, NotifyUnregistered() is called.
+//
+// This class is the right place to store buffer positions. Profiler buffer
+// positions become invalid if the profiler buffer is destroyed, which happens
+// when the profiler is stopped.
+class ProfiledThreadData final {
+ public:
+  explicit ProfiledThreadData(ThreadInfo* aThreadInfo);
+  ~ProfiledThreadData();
+
+  void NotifyUnregistered(uint64_t aBufferPosition) {
+    mLastSample = Nothing();
+    MOZ_ASSERT(!mBufferPositionWhenReceivedJSContext,
+               "JSContext should have been cleared before the thread was "
+               "unregistered");
+    mUnregisterTime = TimeStamp::NowUnfuzzed();
+    mBufferPositionWhenUnregistered = Some(aBufferPosition);
+  }
+  Maybe<uint64_t> BufferPositionWhenUnregistered() {
+    return mBufferPositionWhenUnregistered;
+  }
+
+  Maybe<uint64_t>& LastSample() { return mLastSample; }
+
+  void StreamJSON(const ProfileBuffer& aBuffer, SpliceableJSONWriter& aWriter,
+                  const std::string& aProcessName,
+                  const std::string& aETLDplus1,
+                  const TimeStamp& aProcessStartTime, double aSinceTime);
+
+  const RefPtr<ThreadInfo> Info() const { return mThreadInfo; }
+
+  void NotifyReceivedJSContext(uint64_t aCurrentBufferPosition) {
+    mBufferPositionWhenReceivedJSContext = Some(aCurrentBufferPosition);
+  }
+
+ private:
+  // Group A:
+  // The following fields are interesting for the entire lifetime of a
+  // ProfiledThreadData object.
+
+  // This thread's thread info.
+  const RefPtr<ThreadInfo> mThreadInfo;
+
+  // Group B:
+  // The following fields are only used while this thread is alive and
+  // registered. They become Nothing() once the thread is unregistered.
+
+  // When sampling, this holds the position in ActivePS::mBuffer of the most
+  // recent sample for this thread, or Nothing() if there is no sample for this
+  // thread in the buffer.
+  Maybe<uint64_t> mLastSample;
+
+  // Only non-Nothing() if the thread currently has a JSContext.
+  Maybe<uint64_t> mBufferPositionWhenReceivedJSContext;
+
+  // Group C:
+  // The following fields are only used once this thread has been unregistered.
+
+  Maybe<uint64_t> mBufferPositionWhenUnregistered;
+  TimeStamp mUnregisterTime;
+};
+
+// Stream all samples and markers from aBuffer with the given aThreadId (or 0
+// for everything, which is assumed to be a single backtrace sample.)
+// Returns the thread id of the output sample(s), or 0 if none was present.
+int StreamSamplesAndMarkers(
+    const char* aName, int aThreadId, const ProfileBuffer& aBuffer,
+    SpliceableJSONWriter& aWriter, const std::string& aProcessName,
+    const std::string& aETLDplus1, const TimeStamp& aProcessStartTime,
+    const TimeStamp& aRegisterTime, const TimeStamp& aUnregisterTime,
+    double aSinceTime, UniqueStacks& aUniqueStacks);
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif  // ProfiledThreadData_h
diff --git a/mozglue/baseprofiler/core/ProfilerBacktrace.cpp b/mozglue/baseprofiler/core/ProfilerBacktrace.cpp
new file mode 100644
index 0000000000..166e72fd9c
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfilerBacktrace.cpp
@@ -0,0 +1,123 @@
+/* -*- 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 "ProfilerBacktrace.h"
+
+#include "BaseProfiler.h"
+#include "ProfileBuffer.h"
+#include "ProfiledThreadData.h"
+#include "ThreadInfo.h"
+
+#include "mozilla/BaseProfileJSONWriter.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+ProfilerBacktrace::ProfilerBacktrace(
+    const char* aName,
+    UniquePtr<ProfileChunkedBuffer> aProfileChunkedBufferStorage,
+    UniquePtr<ProfileBuffer> aProfileBufferStorageOrNull /* = nullptr */)
+    : mName(aName),
+      mOptionalProfileChunkedBufferStorage(
+          std::move(aProfileChunkedBufferStorage)),
+      mProfileChunkedBuffer(mOptionalProfileChunkedBufferStorage.get()),
+      mOptionalProfileBufferStorage(std::move(aProfileBufferStorageOrNull)),
+      mProfileBuffer(mOptionalProfileBufferStorage.get()) {
+  if (mProfileBuffer) {
+    MOZ_RELEASE_ASSERT(mProfileChunkedBuffer,
+                       "If we take ownership of a ProfileBuffer, we must also "
+                       "receive ownership of a ProfileChunkedBuffer");
+    MOZ_RELEASE_ASSERT(
+        mProfileChunkedBuffer == &mProfileBuffer->UnderlyingChunkedBuffer(),
+        "If we take ownership of a ProfileBuffer, we must also receive "
+        "ownership of its ProfileChunkedBuffer");
+  }
+  MOZ_ASSERT(
+      !mProfileChunkedBuffer || !mProfileChunkedBuffer->IsThreadSafe(),
+      "ProfilerBacktrace only takes a non-thread-safe ProfileChunkedBuffer");
+}
+
+ProfilerBacktrace::ProfilerBacktrace(
+    const char* aName,
+    ProfileChunkedBuffer* aExternalProfileChunkedBufferOrNull /* = nullptr */,
+    ProfileBuffer* aExternalProfileBufferOrNull /* = nullptr */)
+    : mName(aName),
+      mProfileChunkedBuffer(aExternalProfileChunkedBufferOrNull),
+      mProfileBuffer(aExternalProfileBufferOrNull) {
+  if (!mProfileChunkedBuffer) {
+    if (mProfileBuffer) {
+      // We don't have a ProfileChunkedBuffer but we have a ProfileBuffer, use
+      // the latter's ProfileChunkedBuffer.
+      mProfileChunkedBuffer = &mProfileBuffer->UnderlyingChunkedBuffer();
+      MOZ_ASSERT(!mProfileChunkedBuffer->IsThreadSafe(),
+                 "ProfilerBacktrace only takes a non-thread-safe "
+                 "ProfileChunkedBuffer");
+    }
+  } else {
+    if (mProfileBuffer) {
+      MOZ_RELEASE_ASSERT(
+          mProfileChunkedBuffer == &mProfileBuffer->UnderlyingChunkedBuffer(),
+          "If we reference both ProfileChunkedBuffer and ProfileBuffer, they "
+          "must already be connected");
+    }
+    MOZ_ASSERT(!mProfileChunkedBuffer->IsThreadSafe(),
+               "ProfilerBacktrace only takes a non-thread-safe "
+               "ProfileChunkedBuffer");
+  }
+}
+
+ProfilerBacktrace::~ProfilerBacktrace() {}
+
+int ProfilerBacktrace::StreamJSON(SpliceableJSONWriter& aWriter,
+                                  const TimeStamp& aProcessStartTime,
+                                  UniqueStacks& aUniqueStacks) {
+  int processedThreadId = 0;
+
+  // Unlike ProfiledThreadData::StreamJSON, we don't need to call
+  // ProfileBuffer::AddJITInfoForRange because ProfileBuffer does not contain
+  // any JitReturnAddr entries. For synchronous samples, JIT frames get expanded
+  // at sample time.
+  if (mProfileBuffer) {
+    processedThreadId = StreamSamplesAndMarkers(
+        mName.c_str(), 0, *mProfileBuffer, aWriter, "", "", aProcessStartTime,
+        /* aRegisterTime */ TimeStamp(),
+        /* aUnregisterTime */ TimeStamp(),
+        /* aSinceTime */ 0, aUniqueStacks);
+  } else if (mProfileChunkedBuffer) {
+    ProfileBuffer profileBuffer(*mProfileChunkedBuffer);
+    processedThreadId = StreamSamplesAndMarkers(
+        mName.c_str(), 0, profileBuffer, aWriter, "", "", aProcessStartTime,
+        /* aRegisterTime */ TimeStamp(),
+        /* aUnregisterTime */ TimeStamp(),
+        /* aSinceTime */ 0, aUniqueStacks);
+  }
+  // If there are no buffers, the backtrace is empty and nothing is streamed.
+
+  return processedThreadId;
+}
+
+}  // namespace baseprofiler
+
+// static
+template <typename Destructor>
+UniquePtr<baseprofiler::ProfilerBacktrace, Destructor>
+ProfileBufferEntryReader::
+    Deserializer<UniquePtr<baseprofiler::ProfilerBacktrace, Destructor>>::Read(
+        ProfileBufferEntryReader& aER) {
+  auto profileChunkedBuffer = aER.ReadObject<UniquePtr<ProfileChunkedBuffer>>();
+  if (!profileChunkedBuffer) {
+    return nullptr;
+  }
+  MOZ_ASSERT(
+      !profileChunkedBuffer->IsThreadSafe(),
+      "ProfilerBacktrace only stores non-thread-safe ProfileChunkedBuffers");
+  std::string name = aER.ReadObject<std::string>();
+  return UniquePtr<baseprofiler::ProfilerBacktrace, Destructor>{
+      new baseprofiler::ProfilerBacktrace(name.c_str(),
+                                          std::move(profileChunkedBuffer))};
+};
+
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/ProfilerBacktrace.h b/mozglue/baseprofiler/core/ProfilerBacktrace.h
new file mode 100644
index 0000000000..b5365cd4ac
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfilerBacktrace.h
@@ -0,0 +1,162 @@
+/* -*- Mode: C++; tab-width: 2; 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 __PROFILER_BACKTRACE_H
+#define __PROFILER_BACKTRACE_H
+
+#include "mozilla/UniquePtrExtensions.h"
+
+#include <string>
+
+namespace mozilla {
+
+class ProfileChunkedBuffer;
+class TimeStamp;
+
+namespace baseprofiler {
+
+class ProfileBuffer;
+class SpliceableJSONWriter;
+class ThreadInfo;
+class UniqueStacks;
+
+// ProfilerBacktrace encapsulates a synchronous sample.
+// It can work with a ProfileBuffer and/or a ProfileChunkedBuffer (if both, they
+// must already be linked together). The ProfileChunkedBuffer contains all the
+// data; the ProfileBuffer is not strictly needed, only provide it if it is
+// already available at the call site.
+// And these buffers can either be:
+// - owned here, so that the ProfilerBacktrace object can be kept for later
+//   use), OR
+// - referenced through pointers (in cases where the backtrace is immediately
+//   streamed out, so we only need temporary references to external buffers);
+//   these pointers may be null for empty backtraces.
+class ProfilerBacktrace {
+ public:
+  // Take ownership of external buffers and use them to keep, and to stream a
+  // backtrace. If a ProfileBuffer is given, its underlying chunked buffer must
+  // be provided as well.
+  explicit ProfilerBacktrace(
+      const char* aName,
+      UniquePtr<ProfileChunkedBuffer> aProfileChunkedBufferStorage,
+      UniquePtr<ProfileBuffer> aProfileBufferStorageOrNull = nullptr);
+
+  // Take pointers to external buffers and use them to stream a backtrace.
+  // If null, the backtrace is effectively empty.
+  // If both are provided, they must already be connected.
+  explicit ProfilerBacktrace(
+      const char* aName,
+      ProfileChunkedBuffer* aExternalProfileChunkedBufferOrNull = nullptr,
+      ProfileBuffer* aExternalProfileBufferOrNull = nullptr);
+
+  ~ProfilerBacktrace();
+
+  [[nodiscard]] bool IsEmpty() const {
+    return !mProfileChunkedBuffer ||
+           ProfileBufferEntryWriter::Serializer<ProfileChunkedBuffer>::Bytes(
+               *mProfileChunkedBuffer) <= ULEB128Size(0u);
+  }
+
+  // ProfilerBacktraces' stacks are deduplicated in the context of the
+  // profile that contains the backtrace as a marker payload.
+  //
+  // That is, markers that contain backtraces should not need their own stack,
+  // frame, and string tables. They should instead reuse their parent
+  // profile's tables.
+  int StreamJSON(SpliceableJSONWriter& aWriter,
+                 const TimeStamp& aProcessStartTime,
+                 UniqueStacks& aUniqueStacks);
+
+ private:
+  // Used to de/serialize a ProfilerBacktrace.
+  friend ProfileBufferEntryWriter::Serializer<ProfilerBacktrace>;
+  friend ProfileBufferEntryReader::Deserializer<ProfilerBacktrace>;
+
+  std::string mName;
+
+  // `ProfileChunkedBuffer` in which `mProfileBuffer` stores its data; must be
+  // located before `mProfileBuffer` so that it's destroyed after.
+  UniquePtr<ProfileChunkedBuffer> mOptionalProfileChunkedBufferStorage;
+  // If null, there is no need to check mProfileBuffer's (if present) underlying
+  // buffer because this is done when constructed.
+  ProfileChunkedBuffer* mProfileChunkedBuffer;
+
+  UniquePtr<ProfileBuffer> mOptionalProfileBufferStorage;
+  ProfileBuffer* mProfileBuffer;
+};
+
+}  // namespace baseprofiler
+
+// Format: [ UniquePtr<BlockRingsBuffer> | name ]
+// Initial len==0 marks a nullptr or empty backtrace.
+template <>
+struct ProfileBufferEntryWriter::Serializer<baseprofiler::ProfilerBacktrace> {
+  static Length Bytes(const baseprofiler::ProfilerBacktrace& aBacktrace) {
+    if (!aBacktrace.mProfileChunkedBuffer) {
+      // No buffer.
+      return ULEB128Size(0u);
+    }
+    auto bufferBytes = SumBytes(*aBacktrace.mProfileChunkedBuffer);
+    if (bufferBytes <= ULEB128Size(0u)) {
+      // Empty buffer.
+      return ULEB128Size(0u);
+    }
+    return bufferBytes + SumBytes(aBacktrace.mName);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const baseprofiler::ProfilerBacktrace& aBacktrace) {
+    if (!aBacktrace.mProfileChunkedBuffer ||
+        SumBytes(*aBacktrace.mProfileChunkedBuffer) <= ULEB128Size(0u)) {
+      // No buffer, or empty buffer.
+      aEW.WriteULEB128(0u);
+      return;
+    }
+    aEW.WriteObject(*aBacktrace.mProfileChunkedBuffer);
+    aEW.WriteObject(aBacktrace.mName);
+  }
+};
+
+template <typename Destructor>
+struct ProfileBufferEntryWriter::Serializer<
+    UniquePtr<baseprofiler::ProfilerBacktrace, Destructor>> {
+  static Length Bytes(const UniquePtr<baseprofiler::ProfilerBacktrace,
+                                      Destructor>& aBacktrace) {
+    if (!aBacktrace) {
+      // Null backtrace pointer (treated like an empty backtrace).
+      return ULEB128Size(0u);
+    }
+    return SumBytes(*aBacktrace);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const UniquePtr<baseprofiler::ProfilerBacktrace,
+                                    Destructor>& aBacktrace) {
+    if (!aBacktrace) {
+      // Null backtrace pointer (treated like an empty backtrace).
+      aEW.WriteULEB128(0u);
+      return;
+    }
+    aEW.WriteObject(*aBacktrace);
+  }
+};
+
+template <typename Destructor>
+struct ProfileBufferEntryReader::Deserializer<
+    UniquePtr<baseprofiler::ProfilerBacktrace, Destructor>> {
+  static void ReadInto(
+      ProfileBufferEntryReader& aER,
+      UniquePtr<baseprofiler::ProfilerBacktrace, Destructor>& aBacktrace) {
+    aBacktrace = Read(aER);
+  }
+
+  static UniquePtr<baseprofiler::ProfilerBacktrace, Destructor> Read(
+      ProfileBufferEntryReader& aER);
+};
+
+}  // namespace mozilla
+
+#endif  // __PROFILER_BACKTRACE_H
diff --git a/mozglue/baseprofiler/core/ProfilerMarkers.cpp b/mozglue/baseprofiler/core/ProfilerMarkers.cpp
new file mode 100644
index 0000000000..bff2a9ebdd
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfilerMarkers.cpp
@@ -0,0 +1,221 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* 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/BaseProfilerMarkers.h"
+
+#include "mozilla/Likely.h"
+
+#include <limits>
+
+namespace mozilla {
+namespace base_profiler_markers_detail {
+
+// We need an atomic type that can hold a `DeserializerTag`. (Atomic doesn't
+// work with too-small types.)
+using DeserializerTagAtomic = unsigned;
+
+// Number of currently-registered deserializers and other marker type functions.
+static Atomic<DeserializerTagAtomic, MemoryOrdering::Relaxed>
+    sDeserializerCount{0};
+
+// This needs to be big enough to handle all possible marker types. If one day
+// this needs to be higher, the underlying DeserializerTag type will have to be
+// changed.
+static constexpr DeserializerTagAtomic DeserializerMax = 250;
+
+static_assert(
+    DeserializerMax <= std::numeric_limits<Streaming::DeserializerTag>::max(),
+    "The maximum number of deserializers must fit in the DeserializerTag type");
+
+// Array of marker type functions.
+// 1-based, i.e.: [0] -> tag 1, [DeserializerMax - 1] -> tag DeserializerMax.
+// Elements are added at the next available atomically-incremented
+// `sDeserializerCount` (minus 1) whenever a new marker type is used in a
+// Firefox session; the content is kept between profiler runs in that session.
+// There is theoretically a race between the increment and the time the entry is
+// fully written, but in practice all new elements are written (during
+// profiling, using a marker type for the first time) long before they are read
+// (after profiling is paused).
+static Streaming::MarkerTypeFunctions
+    sMarkerTypeFunctions1Based[DeserializerMax];
+
+/* static */ Streaming::DeserializerTag Streaming::TagForMarkerTypeFunctions(
+    Streaming::MarkerDataDeserializer aDeserializer,
+    Streaming::MarkerTypeNameFunction aMarkerTypeNameFunction,
+    Streaming::MarkerSchemaFunction aMarkerSchemaFunction) {
+  MOZ_RELEASE_ASSERT(!!aDeserializer);
+  MOZ_RELEASE_ASSERT(!!aMarkerTypeNameFunction);
+  MOZ_RELEASE_ASSERT(!!aMarkerSchemaFunction);
+
+  DeserializerTagAtomic tag = ++sDeserializerCount;
+  MOZ_RELEASE_ASSERT(
+      tag <= DeserializerMax,
+      "Too many deserializers, consider increasing DeserializerMax. "
+      "Or is a deserializer stored again and again?");
+  sMarkerTypeFunctions1Based[tag - 1] = {aDeserializer, aMarkerTypeNameFunction,
+                                         aMarkerSchemaFunction};
+
+  return static_cast<DeserializerTag>(tag);
+}
+
+/* static */ Streaming::MarkerDataDeserializer Streaming::DeserializerForTag(
+    Streaming::DeserializerTag aTag) {
+  MOZ_RELEASE_ASSERT(
+      aTag > 0 && static_cast<DeserializerTagAtomic>(aTag) <=
+                      static_cast<DeserializerTagAtomic>(sDeserializerCount),
+      "Out-of-range tag value");
+  return sMarkerTypeFunctions1Based[aTag - 1].mMarkerDataDeserializer;
+}
+
+/* static */ Span<const Streaming::MarkerTypeFunctions>
+Streaming::MarkerTypeFunctionsArray() {
+  return {sMarkerTypeFunctions1Based, sDeserializerCount};
+}
+
+}  // namespace base_profiler_markers_detail
+
+void MarkerSchema::Stream(JSONWriter& aWriter,
+                          const Span<const char>& aName) && {
+  // The caller should have started a JSON array, in which we can add an object
+  // that defines a marker schema.
+
+  if (mLocations.empty()) {
+    // SpecialFrontendLocation case, don't output anything for this type.
+    return;
+  }
+
+  aWriter.StartObjectElement();
+  {
+    aWriter.StringProperty("name", aName);
+
+    if (!mChartLabel.empty()) {
+      aWriter.StringProperty("chartLabel", mChartLabel);
+    }
+
+    if (!mTooltipLabel.empty()) {
+      aWriter.StringProperty("tooltipLabel", mTooltipLabel);
+    }
+
+    if (!mTableLabel.empty()) {
+      aWriter.StringProperty("tableLabel", mTableLabel);
+    }
+
+    aWriter.StartArrayProperty("display");
+    {
+      for (Location location : mLocations) {
+        aWriter.StringElement(LocationToStringSpan(location));
+      }
+    }
+    aWriter.EndArray();
+
+    aWriter.StartArrayProperty("data");
+    {
+      for (const DataRow& row : mData) {
+        aWriter.StartObjectElement();
+        {
+          row.match(
+              [&aWriter](const DynamicData& aData) {
+                aWriter.StringProperty("key", aData.mKey);
+                if (aData.mLabel) {
+                  aWriter.StringProperty("label", *aData.mLabel);
+                }
+                aWriter.StringProperty("format",
+                                       FormatToStringSpan(aData.mFormat));
+                if (aData.mSearchable) {
+                  aWriter.BoolProperty(
+                      "searchable",
+                      *aData.mSearchable == Searchable::searchable);
+                }
+              },
+              [&aWriter](const StaticData& aStaticData) {
+                aWriter.StringProperty("label", aStaticData.mLabel);
+                aWriter.StringProperty("value", aStaticData.mValue);
+              });
+        }
+        aWriter.EndObject();
+      }
+    }
+    aWriter.EndArray();
+  }
+  aWriter.EndObject();
+}
+
+/* static */
+Span<const char> MarkerSchema::LocationToStringSpan(
+    MarkerSchema::Location aLocation) {
+  switch (aLocation) {
+    case Location::markerChart:
+      return mozilla::MakeStringSpan("marker-chart");
+    case Location::markerTable:
+      return mozilla::MakeStringSpan("marker-table");
+    case Location::timelineOverview:
+      return mozilla::MakeStringSpan("timeline-overview");
+    case Location::timelineMemory:
+      return mozilla::MakeStringSpan("timeline-memory");
+    case Location::timelineIPC:
+      return mozilla::MakeStringSpan("timeline-ipc");
+    case Location::timelineFileIO:
+      return mozilla::MakeStringSpan("timeline-fileio");
+    case Location::stackChart:
+      return mozilla::MakeStringSpan("stack-chart");
+    default:
+      MOZ_CRASH("Unexpected Location enum");
+      return {};
+  }
+}
+
+/* static */
+Span<const char> MarkerSchema::FormatToStringSpan(
+    MarkerSchema::Format aFormat) {
+  switch (aFormat) {
+    case Format::url:
+      return mozilla::MakeStringSpan("url");
+    case Format::filePath:
+      return mozilla::MakeStringSpan("file-path");
+    case Format::string:
+      return mozilla::MakeStringSpan("string");
+    case Format::duration:
+      return mozilla::MakeStringSpan("duration");
+    case Format::time:
+      return mozilla::MakeStringSpan("time");
+    case Format::seconds:
+      return mozilla::MakeStringSpan("seconds");
+    case Format::milliseconds:
+      return mozilla::MakeStringSpan("milliseconds");
+    case Format::microseconds:
+      return mozilla::MakeStringSpan("microseconds");
+    case Format::nanoseconds:
+      return mozilla::MakeStringSpan("nanoseconds");
+    case Format::bytes:
+      return mozilla::MakeStringSpan("bytes");
+    case Format::percentage:
+      return mozilla::MakeStringSpan("percentage");
+    case Format::integer:
+      return mozilla::MakeStringSpan("integer");
+    case Format::decimal:
+      return mozilla::MakeStringSpan("decimal");
+    default:
+      MOZ_CRASH("Unexpected Format enum");
+      return {};
+  }
+}
+
+}  // namespace mozilla
+
+namespace mozilla::baseprofiler {
+template MFBT_API ProfileBufferBlockIndex AddMarker(const ProfilerString8View&,
+                                                    const MarkerCategory&,
+                                                    MarkerOptions&&,
+                                                    markers::TextMarker,
+                                                    const std::string&);
+
+template MFBT_API ProfileBufferBlockIndex
+AddMarkerToBuffer(ProfileChunkedBuffer&, const ProfilerString8View&,
+                  const MarkerCategory&, MarkerOptions&&, markers::NoPayload);
+
+template MFBT_API ProfileBufferBlockIndex AddMarkerToBuffer(
+    ProfileChunkedBuffer&, const ProfilerString8View&, const MarkerCategory&,
+    MarkerOptions&&, markers::TextMarker, const std::string&);
+}  // namespace mozilla::baseprofiler
diff --git a/mozglue/baseprofiler/core/ProfilingCategory.cpp b/mozglue/baseprofiler/core/ProfilingCategory.cpp
new file mode 100644
index 0000000000..8ff2b15555
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfilingCategory.cpp
@@ -0,0 +1,71 @@
+/* -*- 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 "BaseProfilingCategory.h"
+
+#include "mozilla/ArrayUtils.h"
+#include "mozilla/Assertions.h"
+
+#include "BaseProfiler.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+// clang-format off
+
+// ProfilingSubcategory_X:
+// One enum for each category X, listing that category's subcategories. This
+// allows the sProfilingCategoryInfo macro construction below to look up a
+// per-category index for a subcategory.
+#define SUBCATEGORY_ENUMS_BEGIN_CATEGORY(name, labelAsString, color) \
+  enum class ProfilingSubcategory_##name : uint32_t {
+#define SUBCATEGORY_ENUMS_SUBCATEGORY(category, name, labelAsString) \
+    name,
+#define SUBCATEGORY_ENUMS_END_CATEGORY \
+  };
+MOZ_PROFILING_CATEGORY_LIST(SUBCATEGORY_ENUMS_BEGIN_CATEGORY,
+                            SUBCATEGORY_ENUMS_SUBCATEGORY,
+                            SUBCATEGORY_ENUMS_END_CATEGORY)
+#undef SUBCATEGORY_ENUMS_BEGIN_CATEGORY
+#undef SUBCATEGORY_ENUMS_SUBCATEGORY
+#undef SUBCATEGORY_ENUMS_END_CATEGORY
+
+// sProfilingCategoryPairInfo:
+// A list of ProfilingCategoryPairInfos with the same order as
+// ProfilingCategoryPair, which can be used to map a ProfilingCategoryPair to
+// its information.
+#define CATEGORY_INFO_BEGIN_CATEGORY(name, labelAsString, color)
+#define CATEGORY_INFO_SUBCATEGORY(category, name, labelAsString) \
+  {ProfilingCategory::category,                                  \
+   uint32_t(ProfilingSubcategory_##category::name), labelAsString},
+#define CATEGORY_INFO_END_CATEGORY
+const ProfilingCategoryPairInfo sProfilingCategoryPairInfo[] = {
+  MOZ_PROFILING_CATEGORY_LIST(CATEGORY_INFO_BEGIN_CATEGORY,
+                              CATEGORY_INFO_SUBCATEGORY,
+                              CATEGORY_INFO_END_CATEGORY)
+};
+#undef CATEGORY_INFO_BEGIN_CATEGORY
+#undef CATEGORY_INFO_SUBCATEGORY
+#undef CATEGORY_INFO_END_CATEGORY
+
+// clang-format on
+
+const ProfilingCategoryPairInfo& GetProfilingCategoryPairInfo(
+    ProfilingCategoryPair aCategoryPair) {
+  static_assert(
+      MOZ_ARRAY_LENGTH(sProfilingCategoryPairInfo) ==
+          uint32_t(ProfilingCategoryPair::COUNT),
+      "sProfilingCategoryPairInfo and ProfilingCategory need to have the "
+      "same order and the same length");
+
+  uint32_t categoryPairIndex = uint32_t(aCategoryPair);
+  MOZ_RELEASE_ASSERT(categoryPairIndex <=
+                     uint32_t(ProfilingCategoryPair::LAST));
+  return sProfilingCategoryPairInfo[categoryPairIndex];
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/ProfilingStack.cpp b/mozglue/baseprofiler/core/ProfilingStack.cpp
new file mode 100644
index 0000000000..f5cd2ddd04
--- /dev/null
+++ b/mozglue/baseprofiler/core/ProfilingStack.cpp
@@ -0,0 +1,52 @@
+/* -*- 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 "BaseProfilingStack.h"
+
+#include <algorithm>
+
+#include "mozilla/IntegerRange.h"
+#include "mozilla/UniquePtr.h"
+#include "mozilla/UniquePtrExtensions.h"
+
+#include "BaseProfiler.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+ProfilingStack::~ProfilingStack() {
+  // The label macros keep a reference to the ProfilingStack to avoid a TLS
+  // access. If these are somehow not all cleared we will get a
+  // use-after-free so better to crash now.
+  MOZ_RELEASE_ASSERT(stackPointer == 0);
+
+  delete[] frames;
+}
+
+void ProfilingStack::ensureCapacitySlow() {
+  MOZ_ASSERT(stackPointer >= capacity);
+  const uint32_t kInitialCapacity = 128;
+
+  uint32_t sp = stackPointer;
+  auto newCapacity =
+      std::max(sp + 1, capacity ? capacity * 2 : kInitialCapacity);
+
+  auto* newFrames = new ProfilingStackFrame[newCapacity];
+
+  // It's important that `frames` / `capacity` / `stackPointer` remain
+  // consistent here at all times.
+  for (auto i : IntegerRange(capacity)) {
+    newFrames[i] = frames[i];
+  }
+
+  ProfilingStackFrame* oldFrames = frames;
+  frames = newFrames;
+  capacity = newCapacity;
+  delete[] oldFrames;
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/RegisteredThread.cpp b/mozglue/baseprofiler/core/RegisteredThread.cpp
new file mode 100644
index 0000000000..85a7fc2c6d
--- /dev/null
+++ b/mozglue/baseprofiler/core/RegisteredThread.cpp
@@ -0,0 +1,42 @@
+/* -*- Mode: C++; tab-width: 2; 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 "RegisteredThread.h"
+
+#include "BaseProfiler.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+RegisteredThread::RegisteredThread(ThreadInfo* aInfo, void* aStackTop)
+    : mRacyRegisteredThread(aInfo->ThreadId()),
+      mPlatformData(AllocPlatformData(aInfo->ThreadId())),
+      mStackTop(aStackTop),
+      mThreadInfo(aInfo) {
+  // We don't have to guess on mac
+#if defined(GP_OS_darwin)
+  pthread_t self = pthread_self();
+  mStackTop = pthread_get_stackaddr_np(self);
+#endif
+}
+
+RegisteredThread::~RegisteredThread() {}
+
+size_t RegisteredThread::SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+  size_t n = aMallocSizeOf(this);
+
+  // Measurement of the following members may be added later if DMD finds it
+  // is worthwhile:
+  // - mPlatformData
+  //
+  // The following members are not measured:
+  // - mThreadInfo: because it is non-owning
+
+  return n;
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/RegisteredThread.h b/mozglue/baseprofiler/core/RegisteredThread.h
new file mode 100644
index 0000000000..6ae12b823f
--- /dev/null
+++ b/mozglue/baseprofiler/core/RegisteredThread.h
@@ -0,0 +1,166 @@
+/* -*- Mode: C++; tab-width: 2; 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 RegisteredThread_h
+#define RegisteredThread_h
+
+#include "platform.h"
+#include "ThreadInfo.h"
+
+#include "mozilla/UniquePtr.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+// This class contains the state for a single thread that is accessible without
+// protection from gPSMutex in platform.cpp. Because there is no external
+// protection against data races, it must provide internal protection. Hence
+// the "Racy" prefix.
+//
+class RacyRegisteredThread final {
+ public:
+  explicit RacyRegisteredThread(int aThreadId)
+      : mThreadId(aThreadId), mSleep(AWAKE), mIsBeingProfiled(false) {}
+
+  ~RacyRegisteredThread() {}
+
+  void SetIsBeingProfiled(bool aIsBeingProfiled) {
+    mIsBeingProfiled = aIsBeingProfiled;
+  }
+
+  bool IsBeingProfiled() const { return mIsBeingProfiled; }
+
+  // This is called on every profiler restart. Put things that should happen at
+  // that time here.
+  void ReinitializeOnResume() {
+    // This is needed to cause an initial sample to be taken from sleeping
+    // threads that had been observed prior to the profiler stopping and
+    // restarting. Otherwise sleeping threads would not have any samples to
+    // copy forward while sleeping.
+    (void)mSleep.compareExchange(SLEEPING_OBSERVED, SLEEPING_NOT_OBSERVED);
+  }
+
+  // This returns true for the second and subsequent calls in each sleep cycle.
+  bool CanDuplicateLastSampleDueToSleep() {
+    if (mSleep == AWAKE) {
+      return false;
+    }
+
+    if (mSleep.compareExchange(SLEEPING_NOT_OBSERVED, SLEEPING_OBSERVED)) {
+      return false;
+    }
+
+    return true;
+  }
+
+  // Call this whenever the current thread sleeps. Calling it twice in a row
+  // without an intervening setAwake() call is an error.
+  void SetSleeping() {
+    MOZ_ASSERT(mSleep == AWAKE);
+    mSleep = SLEEPING_NOT_OBSERVED;
+  }
+
+  // Call this whenever the current thread wakes. Calling it twice in a row
+  // without an intervening setSleeping() call is an error.
+  void SetAwake() {
+    MOZ_ASSERT(mSleep != AWAKE);
+    mSleep = AWAKE;
+  }
+
+  bool IsSleeping() { return mSleep != AWAKE; }
+
+  int ThreadId() const { return mThreadId; }
+
+  class ProfilingStack& ProfilingStack() {
+    return mProfilingStack;
+  }
+  const class ProfilingStack& ProfilingStack() const { return mProfilingStack; }
+
+ private:
+  class ProfilingStack mProfilingStack;
+
+  // mThreadId contains the thread ID of the current thread. It is safe to read
+  // this from multiple threads concurrently, as it will never be mutated.
+  const int mThreadId;
+
+  // mSleep tracks whether the thread is sleeping, and if so, whether it has
+  // been previously observed. This is used for an optimization: in some cases,
+  // when a thread is asleep, we duplicate the previous sample, which is
+  // cheaper than taking a new sample.
+  //
+  // mSleep is atomic because it is accessed from multiple threads.
+  //
+  // - It is written only by this thread, via setSleeping() and setAwake().
+  //
+  // - It is read by SamplerThread::Run().
+  //
+  // There are two cases where racing between threads can cause an issue.
+  //
+  // - If CanDuplicateLastSampleDueToSleep() returns false but that result is
+  //   invalidated before being acted upon, we will take a full sample
+  //   unnecessarily. This is additional work but won't cause any correctness
+  //   issues. (In actual fact, this case is impossible. In order to go from
+  //   CanDuplicateLastSampleDueToSleep() returning false to it returning true
+  //   requires an intermediate call to it in order for mSleep to go from
+  //   SLEEPING_NOT_OBSERVED to SLEEPING_OBSERVED.)
+  //
+  // - If CanDuplicateLastSampleDueToSleep() returns true but that result is
+  //   invalidated before being acted upon -- i.e. the thread wakes up before
+  //   DuplicateLastSample() is called -- we will duplicate the previous
+  //   sample. This is inaccurate, but only slightly... we will effectively
+  //   treat the thread as having slept a tiny bit longer than it really did.
+  //
+  // This latter inaccuracy could be avoided by moving the
+  // CanDuplicateLastSampleDueToSleep() check within the thread-freezing code,
+  // e.g. the section where Tick() is called. But that would reduce the
+  // effectiveness of the optimization because more code would have to be run
+  // before we can tell that duplication is allowed.
+  //
+  static const int AWAKE = 0;
+  static const int SLEEPING_NOT_OBSERVED = 1;
+  static const int SLEEPING_OBSERVED = 2;
+  Atomic<int> mSleep;
+
+  // Is this thread being profiled? (e.g., should markers be recorded?)
+  Atomic<bool, MemoryOrdering::Relaxed> mIsBeingProfiled;
+};
+
+// This class contains information that's relevant to a single thread only
+// while that thread is running and registered with the profiler, but
+// regardless of whether the profiler is running. All accesses to it are
+// protected by the profiler state lock.
+class RegisteredThread final {
+ public:
+  RegisteredThread(ThreadInfo* aInfo, void* aStackTop);
+  ~RegisteredThread();
+
+  class RacyRegisteredThread& RacyRegisteredThread() {
+    return mRacyRegisteredThread;
+  }
+  const class RacyRegisteredThread& RacyRegisteredThread() const {
+    return mRacyRegisteredThread;
+  }
+
+  PlatformData* GetPlatformData() const { return mPlatformData.get(); }
+  const void* StackTop() const { return mStackTop; }
+
+  size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const;
+
+  const RefPtr<ThreadInfo> Info() const { return mThreadInfo; }
+
+ private:
+  class RacyRegisteredThread mRacyRegisteredThread;
+
+  const UniquePlatformData mPlatformData;
+  const void* mStackTop;
+
+  const RefPtr<ThreadInfo> mThreadInfo;
+};
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif  // RegisteredThread_h
diff --git a/mozglue/baseprofiler/core/ThreadInfo.h b/mozglue/baseprofiler/core/ThreadInfo.h
new file mode 100644
index 0000000000..4be84a45a9
--- /dev/null
+++ b/mozglue/baseprofiler/core/ThreadInfo.h
@@ -0,0 +1,61 @@
+/* -*- Mode: C++; tab-width: 2; 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 ThreadInfo_h
+#define ThreadInfo_h
+
+#include "mozilla/Atomics.h"
+#include "mozilla/TimeStamp.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+// This class contains information about a thread which needs to be stored
+// across restarts of the profiler and which can be useful even after the
+// thread has stopped running.
+// It uses threadsafe refcounting and only contains immutable data.
+class ThreadInfo final {
+ public:
+  ThreadInfo(const char* aName, int aThreadId, bool aIsMainThread,
+             const TimeStamp& aRegisterTime = TimeStamp::NowUnfuzzed())
+      : mName(aName),
+        mRegisterTime(aRegisterTime),
+        mThreadId(aThreadId),
+        mIsMainThread(aIsMainThread),
+        mRefCnt(0) {
+    // I don't know if we can assert this. But we should warn.
+    MOZ_ASSERT(aThreadId >= 0, "native thread ID is < 0");
+    MOZ_ASSERT(aThreadId <= INT32_MAX, "native thread ID is > INT32_MAX");
+  }
+
+  // Using hand-rolled ref-counting, because RefCounted.h macros don't produce
+  // the same code between mozglue and libxul, see bug 1536656.
+  MFBT_API void AddRef() const { ++mRefCnt; }
+  MFBT_API void Release() const {
+    MOZ_ASSERT(int32_t(mRefCnt) > 0);
+    if (--mRefCnt == 0) {
+      delete this;
+    }
+  }
+
+  const char* Name() const { return mName.c_str(); }
+  TimeStamp RegisterTime() const { return mRegisterTime; }
+  int ThreadId() const { return mThreadId; }
+  bool IsMainThread() const { return mIsMainThread; }
+
+ private:
+  const std::string mName;
+  const TimeStamp mRegisterTime;
+  const int mThreadId;
+  const bool mIsMainThread;
+
+  mutable Atomic<int32_t, MemoryOrdering::ReleaseAcquire> mRefCnt;
+};
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif  // ThreadInfo_h
diff --git a/mozglue/baseprofiler/core/VTuneProfiler.cpp b/mozglue/baseprofiler/core/VTuneProfiler.cpp
new file mode 100644
index 0000000000..2911c39f08
--- /dev/null
+++ b/mozglue/baseprofiler/core/VTuneProfiler.cpp
@@ -0,0 +1,92 @@
+/* -*- Mode: C++; tab-width: 2; 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/. */
+
+#ifdef XP_WIN
+#  undef UNICODE
+#  undef _UNICODE
+#endif
+
+#include "VTuneProfiler.h"
+
+#include <memory>
+
+#include "BaseProfiler.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+VTuneProfiler* VTuneProfiler::mInstance = nullptr;
+
+void VTuneProfiler::Initialize() {
+  // This is just a 'dirty trick' to find out if the ittnotify DLL was found.
+  // If it wasn't this function always returns 0, otherwise it returns
+  // incrementing numbers, if the library was found this wastes 2 events but
+  // that should be okay.
+  // TODO re-implement here if vtune is needed
+  // __itt_event testEvent =
+  //     __itt_event_create("Test event", strlen("Test event"));
+  // testEvent = __itt_event_create("Test event 2", strlen("Test event 2"));
+
+  // if (testEvent) {
+  //   mInstance = new VTuneProfiler();
+  // }
+}
+
+void VTuneProfiler::Shutdown() {}
+
+void VTuneProfiler::TraceInternal(const char* aName, TracingKind aKind) {
+  // TODO re-implement here if vtune is needed
+  // std::string str(aName);
+
+  // auto iter = mStrings.find(str);
+
+  // __itt_event event;
+  // if (iter != mStrings.end()) {
+  //   event = iter->second;
+  // } else {
+  //   event = __itt_event_create(aName, str.length());
+  //   mStrings.insert({str, event});
+  // }
+
+  // if (aKind == TRACING_INTERVAL_START || aKind == TRACING_EVENT) {
+  //   // VTune will consider starts not matched with an end to be single point
+  //   in
+  //   // time events.
+  //   __itt_event_start(event);
+  // } else {
+  //   __itt_event_end(event);
+  // }
+}
+
+void VTuneProfiler::RegisterThreadInternal(const char* aName) {
+  // TODO re-implement here if vtune is needed
+  // std::string str(aName);
+
+  // if (!str.compare("Main Thread (Base Profiler)")) {
+  //   // Process main thread.
+  //   switch (XRE_GetProcessType()) {
+  //     case GeckoProcessType::GeckoProcessType_Default:
+  //       __itt_thread_set_name("Main Process");
+  //       break;
+  //     case GeckoProcessType::GeckoProcessType_Content:
+  //       __itt_thread_set_name("Content Process");
+  //       break;
+  //     case GeckoProcessType::GeckoProcessType_GMPlugin:
+  //       __itt_thread_set_name("Plugin Process");
+  //       break;
+  //     case GeckoProcessType::GeckoProcessType_GPU:
+  //       __itt_thread_set_name("GPU Process");
+  //       break;
+  //     default:
+  //       __itt_thread_set_name("Unknown Process");
+  //   }
+  //   return;
+  // }
+  // __itt_thread_set_name(aName);
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/VTuneProfiler.h b/mozglue/baseprofiler/core/VTuneProfiler.h
new file mode 100644
index 0000000000..cf94ab7242
--- /dev/null
+++ b/mozglue/baseprofiler/core/VTuneProfiler.h
@@ -0,0 +1,84 @@
+/* -*- Mode: C++; tab-width: 2; 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 VTuneProfiler_h
+#define VTuneProfiler_h
+
+// The intent here is to add 0 overhead for regular users. In order to build
+// the VTune profiler code at all --enable-vtune-instrumentation needs to be
+// set as a build option. Even then, when none of the environment variables
+// is specified that allow us to find the ittnotify DLL, these functions
+// should be minimal overhead. When starting Firefox under VTune, these
+// env vars will be automatically defined, otherwise INTEL_LIBITTNOTIFY32/64
+// should be set to point at the ittnotify DLL.
+#ifndef MOZ_VTUNE_INSTRUMENTATION
+
+#  define VTUNE_INIT()
+#  define VTUNE_SHUTDOWN()
+
+#  define VTUNE_TRACING(name, kind)
+#  define VTUNE_REGISTER_THREAD(name)
+
+#else
+
+#  include "BaseProfiler.h"
+
+// This is the regular Intel header, these functions are actually defined for
+// us inside js/src/vtune by an intel C file which actually dynamically resolves
+// them to the correct DLL. Through libxul these will 'magically' resolve.
+#  include "vtune/ittnotify.h"
+
+#  include <stddef.h>
+#  include <unordered_map>
+#  include <string>
+
+namespace mozilla {
+namespace baseprofiler {
+
+class VTuneProfiler {
+ public:
+  static void Initialize();
+  static void Shutdown();
+
+  enum TracingKind {
+    TRACING_EVENT,
+    TRACING_INTERVAL_START,
+    TRACING_INTERVAL_END,
+  };
+
+  static void Trace(const char* aName, TracingKind aKind) {
+    if (mInstance) {
+      mInstance->TraceInternal(aName, aKind);
+    }
+  }
+  static void RegisterThread(const char* aName) {
+    if (mInstance) {
+      mInstance->RegisterThreadInternal(aName);
+    }
+  }
+
+ private:
+  void TraceInternal(const char* aName, TracingKind aKind);
+  void RegisterThreadInternal(const char* aName);
+
+  // This is null when the ittnotify DLL could not be found.
+  static VTuneProfiler* mInstance;
+
+  std::unordered_map<std::string, __itt_event> mStrings;
+};
+
+#  define VTUNE_INIT() VTuneProfiler::Initialize()
+#  define VTUNE_SHUTDOWN() VTuneProfiler::Shutdown()
+
+#  define VTUNE_TRACING(name, kind) VTuneProfiler::Trace(name, kind)
+#  define VTUNE_REGISTER_THREAD(name) VTuneProfiler::RegisterThread(name)
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif
+
+#endif /* VTuneProfiler_h */
diff --git a/mozglue/baseprofiler/core/platform-linux-android.cpp b/mozglue/baseprofiler/core/platform-linux-android.cpp
new file mode 100644
index 0000000000..210bc4dd31
--- /dev/null
+++ b/mozglue/baseprofiler/core/platform-linux-android.cpp
@@ -0,0 +1,550 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+// Copyright (c) 2006-2011 The Chromium Authors. All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//  * Redistributions of source code must retain the above copyright
+//    notice, this list of conditions and the following disclaimer.
+//  * Redistributions in binary form must reproduce the above copyright
+//    notice, this list of conditions and the following disclaimer in
+//    the documentation and/or other materials provided with the
+//    distribution.
+//  * Neither the name of Google, Inc. nor the names of its contributors
+//    may be used to endorse or promote products derived from this
+//    software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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 OWNER OR CONTRIBUTORS 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 file is used for both Linux and Android.
+
+#include <stdio.h>
+#include <math.h>
+
+#include <pthread.h>
+#if defined(GP_OS_freebsd)
+#  include <sys/thr.h>
+#endif
+#include <semaphore.h>
+#include <signal.h>
+#include <sys/time.h>
+#include <sys/resource.h>
+#include <sys/syscall.h>
+#include <sys/types.h>
+#include <stdlib.h>
+#include <sched.h>
+#include <ucontext.h>
+// Ubuntu Dapper requires memory pages to be marked as
+// executable. Otherwise, OS raises an exception when executing code
+// in that page.
+#include <sys/types.h>  // mmap & munmap
+#include <sys/mman.h>   // mmap & munmap
+#include <sys/stat.h>   // open
+#include <fcntl.h>      // open
+#include <unistd.h>     // sysconf
+#include <semaphore.h>
+#ifdef __GLIBC__
+#  include <execinfo.h>  // backtrace, backtrace_symbols
+#endif                   // def __GLIBC__
+#include <strings.h>     // index
+#include <errno.h>
+#include <stdarg.h>
+
+#include "prenv.h"
+#include "mozilla/PodOperations.h"
+#include "mozilla/DebugOnly.h"
+
+#include <string.h>
+#include <list>
+
+using namespace mozilla;
+
+namespace mozilla {
+namespace baseprofiler {
+
+int profiler_current_process_id() { return getpid(); }
+
+int profiler_current_thread_id() {
+#if defined(GP_OS_linux) || defined(GP_OS_android)
+  // glibc doesn't provide a wrapper for gettid().
+  return static_cast<int>(static_cast<pid_t>(syscall(SYS_gettid)));
+#elif defined(GP_OS_freebsd)
+  long id;
+  (void)thr_self(&id);
+  return static_cast<int>(id);
+#else
+#  error "bad platform"
+#endif
+}
+
+static int64_t MicrosecondsSince1970() {
+  struct timeval tv;
+  gettimeofday(&tv, NULL);
+  return int64_t(tv.tv_sec) * 1000000 + int64_t(tv.tv_usec);
+}
+
+void* GetStackTop(void* aGuess) { return aGuess; }
+
+static void PopulateRegsFromContext(Registers& aRegs, ucontext_t* aContext) {
+  aRegs.mContext = aContext;
+  mcontext_t& mcontext = aContext->uc_mcontext;
+
+  // Extracting the sample from the context is extremely machine dependent.
+#if defined(GP_PLAT_x86_linux) || defined(GP_PLAT_x86_android)
+  aRegs.mPC = reinterpret_cast<Address>(mcontext.gregs[REG_EIP]);
+  aRegs.mSP = reinterpret_cast<Address>(mcontext.gregs[REG_ESP]);
+  aRegs.mFP = reinterpret_cast<Address>(mcontext.gregs[REG_EBP]);
+  aRegs.mLR = 0;
+#elif defined(GP_PLAT_amd64_linux) || defined(GP_PLAT_amd64_android)
+  aRegs.mPC = reinterpret_cast<Address>(mcontext.gregs[REG_RIP]);
+  aRegs.mSP = reinterpret_cast<Address>(mcontext.gregs[REG_RSP]);
+  aRegs.mFP = reinterpret_cast<Address>(mcontext.gregs[REG_RBP]);
+  aRegs.mLR = 0;
+#elif defined(GP_PLAT_amd64_freebsd)
+  aRegs.mPC = reinterpret_cast<Address>(mcontext.mc_rip);
+  aRegs.mSP = reinterpret_cast<Address>(mcontext.mc_rsp);
+  aRegs.mFP = reinterpret_cast<Address>(mcontext.mc_rbp);
+  aRegs.mLR = 0;
+#elif defined(GP_PLAT_arm_linux) || defined(GP_PLAT_arm_android)
+  aRegs.mPC = reinterpret_cast<Address>(mcontext.arm_pc);
+  aRegs.mSP = reinterpret_cast<Address>(mcontext.arm_sp);
+  aRegs.mFP = reinterpret_cast<Address>(mcontext.arm_fp);
+  aRegs.mLR = reinterpret_cast<Address>(mcontext.arm_lr);
+#elif defined(GP_PLAT_arm64_linux) || defined(GP_PLAT_arm64_android)
+  aRegs.mPC = reinterpret_cast<Address>(mcontext.pc);
+  aRegs.mSP = reinterpret_cast<Address>(mcontext.sp);
+  aRegs.mFP = reinterpret_cast<Address>(mcontext.regs[29]);
+  aRegs.mLR = reinterpret_cast<Address>(mcontext.regs[30]);
+#elif defined(GP_PLAT_arm64_freebsd)
+  aRegs.mPC = reinterpret_cast<Address>(mcontext.mc_gpregs.gp_elr);
+  aRegs.mSP = reinterpret_cast<Address>(mcontext.mc_gpregs.gp_sp);
+  aRegs.mFP = reinterpret_cast<Address>(mcontext.mc_gpregs.gp_x[29]);
+  aRegs.mLR = reinterpret_cast<Address>(mcontext.mc_gpregs.gp_lr);
+#elif defined(GP_PLAT_mips64_linux) || defined(GP_PLAT_mips64_android)
+  aRegs.mPC = reinterpret_cast<Address>(mcontext.pc);
+  aRegs.mSP = reinterpret_cast<Address>(mcontext.gregs[29]);
+  aRegs.mFP = reinterpret_cast<Address>(mcontext.gregs[30]);
+
+#else
+#  error "bad platform"
+#endif
+}
+
+#if defined(GP_OS_android)
+#  define SYS_tgkill __NR_tgkill
+#endif
+
+#if defined(GP_OS_linux) || defined(GP_OS_android)
+int tgkill(pid_t tgid, pid_t tid, int signalno) {
+  return syscall(SYS_tgkill, tgid, tid, signalno);
+}
+#endif
+
+#if defined(GP_OS_freebsd)
+#  define tgkill thr_kill2
+#endif
+
+class PlatformData {
+ public:
+  explicit PlatformData(int aThreadId) {}
+
+  ~PlatformData() {}
+};
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN Sampler target specifics
+
+// The only way to reliably interrupt a Linux thread and inspect its register
+// and stack state is by sending a signal to it, and doing the work inside the
+// signal handler.  But we don't want to run much code inside the signal
+// handler, since POSIX severely restricts what we can do in signal handlers.
+// So we use a system of semaphores to suspend the thread and allow the
+// sampler thread to do all the work of unwinding and copying out whatever
+// data it wants.
+//
+// A four-message protocol is used to reliably suspend and later resume the
+// thread to be sampled (the samplee):
+//
+// Sampler (signal sender) thread              Samplee (thread to be sampled)
+//
+// Prepare the SigHandlerCoordinator
+// and point sSigHandlerCoordinator at it
+//
+// send SIGPROF to samplee ------- MSG 1 ----> (enter signal handler)
+// wait(mMessage2)                             Copy register state
+//                                               into sSigHandlerCoordinator
+//                         <------ MSG 2 ----- post(mMessage2)
+// Samplee is now suspended.                   wait(mMessage3)
+//   Examine its stack/register
+//   state at leisure
+//
+// Release samplee:
+//   post(mMessage3)       ------- MSG 3 ----->
+// wait(mMessage4)                              Samplee now resumes.  Tell
+//                                                the sampler that we are done.
+//                         <------ MSG 4 ------ post(mMessage4)
+// Now we know the samplee's signal             (leave signal handler)
+//   handler has finished using
+//   sSigHandlerCoordinator.  We can
+//   safely reuse it for some other thread.
+//
+
+// A type used to coordinate between the sampler (signal sending) thread and
+// the thread currently being sampled (the samplee, which receives the
+// signals).
+//
+// The first message is sent using a SIGPROF signal delivery.  The subsequent
+// three are sent using sem_wait/sem_post pairs.  They are named accordingly
+// in the following struct.
+struct SigHandlerCoordinator {
+  SigHandlerCoordinator() {
+    PodZero(&mUContext);
+    int r = sem_init(&mMessage2, /* pshared */ 0, 0);
+    r |= sem_init(&mMessage3, /* pshared */ 0, 0);
+    r |= sem_init(&mMessage4, /* pshared */ 0, 0);
+    MOZ_ASSERT(r == 0);
+  }
+
+  ~SigHandlerCoordinator() {
+    int r = sem_destroy(&mMessage2);
+    r |= sem_destroy(&mMessage3);
+    r |= sem_destroy(&mMessage4);
+    MOZ_ASSERT(r == 0);
+  }
+
+  sem_t mMessage2;       // To sampler: "context is in sSigHandlerCoordinator"
+  sem_t mMessage3;       // To samplee: "resume"
+  sem_t mMessage4;       // To sampler: "finished with sSigHandlerCoordinator"
+  ucontext_t mUContext;  // Context at signal
+};
+
+struct SigHandlerCoordinator* Sampler::sSigHandlerCoordinator = nullptr;
+
+static void SigprofHandler(int aSignal, siginfo_t* aInfo, void* aContext) {
+  // Avoid TSan warning about clobbering errno.
+  int savedErrno = errno;
+
+  MOZ_ASSERT(aSignal == SIGPROF);
+  MOZ_ASSERT(Sampler::sSigHandlerCoordinator);
+
+  // By sending us this signal, the sampler thread has sent us message 1 in
+  // the comment above, with the meaning "|sSigHandlerCoordinator| is ready
+  // for use, please copy your register context into it."
+  Sampler::sSigHandlerCoordinator->mUContext =
+      *static_cast<ucontext_t*>(aContext);
+
+  // Send message 2: tell the sampler thread that the context has been copied
+  // into |sSigHandlerCoordinator->mUContext|.  sem_post can never fail by
+  // being interrupted by a signal, so there's no loop around this call.
+  int r = sem_post(&Sampler::sSigHandlerCoordinator->mMessage2);
+  MOZ_ASSERT(r == 0);
+
+  // At this point, the sampler thread assumes we are suspended, so we must
+  // not touch any global state here.
+
+  // Wait for message 3: the sampler thread tells us to resume.
+  while (true) {
+    r = sem_wait(&Sampler::sSigHandlerCoordinator->mMessage3);
+    if (r == -1 && errno == EINTR) {
+      // Interrupted by a signal.  Try again.
+      continue;
+    }
+    // We don't expect any other kind of failure
+    MOZ_ASSERT(r == 0);
+    break;
+  }
+
+  // Send message 4: tell the sampler thread that we are finished accessing
+  // |sSigHandlerCoordinator|.  After this point it is not safe to touch
+  // |sSigHandlerCoordinator|.
+  r = sem_post(&Sampler::sSigHandlerCoordinator->mMessage4);
+  MOZ_ASSERT(r == 0);
+
+  errno = savedErrno;
+}
+
+Sampler::Sampler(PSLockRef aLock)
+    : mMyPid(profiler_current_process_id())
+      // We don't know what the sampler thread's ID will be until it runs, so
+      // set mSamplerTid to a dummy value and fill it in for real in
+      // SuspendAndSampleAndResumeThread().
+      ,
+      mSamplerTid(-1) {
+#if defined(USE_EHABI_STACKWALK)
+  EHABIStackWalkInit();
+#endif
+
+  // NOTE: We don't initialize LUL here, instead initializing it in
+  // SamplerThread's constructor. This is because with the
+  // profiler_suspend_and_sample_thread entry point, we want to be able to
+  // sample without waiting for LUL to be initialized.
+
+  // Request profiling signals.
+  struct sigaction sa;
+  sa.sa_sigaction = SigprofHandler;
+  sigemptyset(&sa.sa_mask);
+  sa.sa_flags = SA_RESTART | SA_SIGINFO;
+  if (sigaction(SIGPROF, &sa, &mOldSigprofHandler) != 0) {
+    MOZ_CRASH("Error installing SIGPROF handler in the profiler");
+  }
+}
+
+void Sampler::Disable(PSLockRef aLock) {
+  // Restore old signal handler. This is global state so it's important that
+  // we do it now, while gPSMutex is locked.
+  sigaction(SIGPROF, &mOldSigprofHandler, 0);
+}
+
+template <typename Func>
+void Sampler::SuspendAndSampleAndResumeThread(
+    PSLockRef aLock, const RegisteredThread& aRegisteredThread,
+    const TimeStamp& aNow, const Func& aProcessRegs) {
+  // Only one sampler thread can be sampling at once.  So we expect to have
+  // complete control over |sSigHandlerCoordinator|.
+  MOZ_ASSERT(!sSigHandlerCoordinator);
+
+  if (mSamplerTid == -1) {
+    mSamplerTid = profiler_current_thread_id();
+  }
+  int sampleeTid = aRegisteredThread.Info()->ThreadId();
+  MOZ_RELEASE_ASSERT(sampleeTid != mSamplerTid);
+
+  //----------------------------------------------------------------//
+  // Suspend the samplee thread and get its context.
+
+  SigHandlerCoordinator coord;  // on sampler thread's stack
+  sSigHandlerCoordinator = &coord;
+
+  // Send message 1 to the samplee (the thread to be sampled), by
+  // signalling at it.
+  // This could fail if the thread doesn't exist anymore.
+  int r = tgkill(mMyPid, sampleeTid, SIGPROF);
+  if (r == 0) {
+    // Wait for message 2 from the samplee, indicating that the context
+    // is available and that the thread is suspended.
+    while (true) {
+      r = sem_wait(&sSigHandlerCoordinator->mMessage2);
+      if (r == -1 && errno == EINTR) {
+        // Interrupted by a signal.  Try again.
+        continue;
+      }
+      // We don't expect any other kind of failure.
+      MOZ_ASSERT(r == 0);
+      break;
+    }
+
+    //----------------------------------------------------------------//
+    // Sample the target thread.
+
+    // WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+    //
+    // The profiler's "critical section" begins here.  In the critical section,
+    // we must not do any dynamic memory allocation, nor try to acquire any lock
+    // or any other unshareable resource.  This is because the thread to be
+    // sampled has been suspended at some entirely arbitrary point, and we have
+    // no idea which unsharable resources (locks, essentially) it holds.  So any
+    // attempt to acquire any lock, including the implied locks used by the
+    // malloc implementation, risks deadlock.  This includes TimeStamp::Now(),
+    // which gets a lock on Windows.
+
+    // The samplee thread is now frozen and sSigHandlerCoordinator->mUContext is
+    // valid.  We can poke around in it and unwind its stack as we like.
+
+    // Extract the current register values.
+    Registers regs;
+    PopulateRegsFromContext(regs, &sSigHandlerCoordinator->mUContext);
+    aProcessRegs(regs, aNow);
+
+    //----------------------------------------------------------------//
+    // Resume the target thread.
+
+    // Send message 3 to the samplee, which tells it to resume.
+    r = sem_post(&sSigHandlerCoordinator->mMessage3);
+    MOZ_ASSERT(r == 0);
+
+    // Wait for message 4 from the samplee, which tells us that it has
+    // finished with |sSigHandlerCoordinator|.
+    while (true) {
+      r = sem_wait(&sSigHandlerCoordinator->mMessage4);
+      if (r == -1 && errno == EINTR) {
+        continue;
+      }
+      MOZ_ASSERT(r == 0);
+      break;
+    }
+
+    // The profiler's critical section ends here.  After this point, none of the
+    // critical section limitations documented above apply.
+    //
+    // WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+  }
+
+  // This isn't strictly necessary, but doing so does help pick up anomalies
+  // in which the signal handler is running when it shouldn't be.
+  sSigHandlerCoordinator = nullptr;
+}
+
+// END Sampler target specifics
+////////////////////////////////////////////////////////////////////////
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN SamplerThread target specifics
+
+static void* ThreadEntry(void* aArg) {
+  auto thread = static_cast<SamplerThread*>(aArg);
+  thread->Run();
+  return nullptr;
+}
+
+SamplerThread::SamplerThread(PSLockRef aLock, uint32_t aActivityGeneration,
+                             double aIntervalMilliseconds)
+    : mSampler(aLock),
+      mActivityGeneration(aActivityGeneration),
+      mIntervalMicroseconds(
+          std::max(1, int(floor(aIntervalMilliseconds * 1000 + 0.5)))) {
+#if defined(USE_LUL_STACKWALK)
+  lul::LUL* lul = CorePS::Lul(aLock);
+  if (!lul) {
+    CorePS::SetLul(aLock, MakeUnique<lul::LUL>(logging_sink_for_LUL));
+    // Read all the unwind info currently available.
+    lul = CorePS::Lul(aLock);
+    read_procmaps(lul);
+
+    // Switch into unwind mode. After this point, we can't add or remove any
+    // unwind info to/from this LUL instance. The only thing we can do with
+    // it is Unwind() calls.
+    lul->EnableUnwinding();
+
+    // Has a test been requested?
+    if (getenv("MOZ_PROFILER_LUL_TEST")) {
+      int nTests = 0, nTestsPassed = 0;
+      RunLulUnitTests(&nTests, &nTestsPassed, lul);
+    }
+  }
+#endif
+
+  // Start the sampling thread. It repeatedly sends a SIGPROF signal. Sending
+  // the signal ourselves instead of relying on itimer provides much better
+  // accuracy.
+  if (pthread_create(&mThread, nullptr, ThreadEntry, this) != 0) {
+    MOZ_CRASH("pthread_create failed");
+  }
+}
+
+SamplerThread::~SamplerThread() { pthread_join(mThread, nullptr); }
+
+void SamplerThread::SleepMicro(uint32_t aMicroseconds) {
+  if (aMicroseconds >= 1000000) {
+    // Use usleep for larger intervals, because the nanosleep
+    // code below only supports intervals < 1 second.
+    MOZ_ALWAYS_TRUE(!::usleep(aMicroseconds));
+    return;
+  }
+
+  struct timespec ts;
+  ts.tv_sec = 0;
+  ts.tv_nsec = aMicroseconds * 1000UL;
+
+  int rv = ::nanosleep(&ts, &ts);
+
+  while (rv != 0 && errno == EINTR) {
+    // Keep waiting in case of interrupt.
+    // nanosleep puts the remaining time back into ts.
+    rv = ::nanosleep(&ts, &ts);
+  }
+
+  MOZ_ASSERT(!rv, "nanosleep call failed");
+}
+
+void SamplerThread::Stop(PSLockRef aLock) {
+  // Restore old signal handler. This is global state so it's important that
+  // we do it now, while gPSMutex is locked. It's safe to do this now even
+  // though this SamplerThread is still alive, because the next time the main
+  // loop of Run() iterates it won't get past the mActivityGeneration check,
+  // and so won't send any signals.
+  mSampler.Disable(aLock);
+}
+
+// END SamplerThread target specifics
+////////////////////////////////////////////////////////////////////////
+
+#if defined(GP_OS_linux) || defined(GP_OS_freebsd)
+
+// We use pthread_atfork() to temporarily disable signal delivery during any
+// fork() call. Without that, fork() can be repeatedly interrupted by signal
+// delivery, requiring it to be repeatedly restarted, which can lead to *long*
+// delays. See bug 837390.
+//
+// We provide no paf_child() function to run in the child after forking. This
+// is fine because we always immediately exec() after fork(), and exec()
+// clobbers all process state. (At one point we did have a paf_child()
+// function, but it caused problems related to locking gPSMutex. See bug
+// 1348374.)
+//
+// Unfortunately all this is only doable on non-Android because Bionic doesn't
+// have pthread_atfork.
+
+// In the parent, before the fork, record IsSamplingPaused, and then pause.
+static void paf_prepare() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (ActivePS::Exists(lock)) {
+    ActivePS::SetWasSamplingPaused(lock, ActivePS::IsSamplingPaused(lock));
+    ActivePS::SetIsSamplingPaused(lock, true);
+  }
+}
+
+// In the parent, after the fork, return IsSamplingPaused to the pre-fork state.
+static void paf_parent() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (ActivePS::Exists(lock)) {
+    ActivePS::SetIsSamplingPaused(lock, ActivePS::WasSamplingPaused(lock));
+    ActivePS::SetWasSamplingPaused(lock, false);
+  }
+}
+
+static void PlatformInit(PSLockRef aLock) {
+  // Set up the fork handlers.
+  pthread_atfork(paf_prepare, paf_parent, nullptr);
+}
+
+#else
+
+static void PlatformInit(PSLockRef aLock) {}
+
+#endif
+
+#if defined(HAVE_NATIVE_UNWIND)
+// Context used by synchronous samples. It's safe to have a single one because
+// only one synchronous sample can be taken at a time (due to
+// profiler_get_backtrace()'s PSAutoLock).
+// ucontext_t sSyncUContext;
+
+void Registers::SyncPopulate() {
+  // TODO port getcontext from breakpad, if profiler_get_backtrace is needed.
+  MOZ_CRASH("profiler_get_backtrace() unsupported");
+  // if (!getcontext(&sSyncUContext)) {
+  //   PopulateRegsFromContext(*this, &sSyncUContext);
+  // }
+}
+#endif
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/platform-macos.cpp b/mozglue/baseprofiler/core/platform-macos.cpp
new file mode 100644
index 0000000000..fc847886ee
--- /dev/null
+++ b/mozglue/baseprofiler/core/platform-macos.cpp
@@ -0,0 +1,233 @@
+/* -*- 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 <unistd.h>
+#include <sys/mman.h>
+#include <mach/mach_init.h>
+#include <mach-o/getsect.h>
+
+#include <AvailabilityMacros.h>
+
+#include <pthread.h>
+#include <semaphore.h>
+#include <signal.h>
+#include <libkern/OSAtomic.h>
+#include <mach/mach.h>
+#include <mach/semaphore.h>
+#include <mach/task.h>
+#include <mach/thread_act.h>
+#include <mach/vm_statistics.h>
+#include <sys/time.h>
+#include <sys/resource.h>
+#include <sys/syscall.h>
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#include <stdarg.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <math.h>
+
+// this port is based off of v8 svn revision 9837
+
+namespace mozilla {
+namespace baseprofiler {
+
+int profiler_current_process_id() { return getpid(); }
+
+int profiler_current_thread_id() {
+  return static_cast<int>(static_cast<pid_t>(syscall(SYS_thread_selfid)));
+}
+
+static int64_t MicrosecondsSince1970() {
+  struct timeval tv;
+  gettimeofday(&tv, NULL);
+  return int64_t(tv.tv_sec) * 1000000 + int64_t(tv.tv_usec);
+}
+
+void* GetStackTop(void* aGuess) {
+  pthread_t thread = pthread_self();
+  return pthread_get_stackaddr_np(thread);
+}
+
+class PlatformData {
+ public:
+  explicit PlatformData(int aThreadId) : mProfiledThread(mach_thread_self()) {}
+
+  ~PlatformData() {
+    // Deallocate Mach port for thread.
+    mach_port_deallocate(mach_task_self(), mProfiledThread);
+  }
+
+  thread_act_t ProfiledThread() { return mProfiledThread; }
+
+ private:
+  // Note: for mProfiledThread Mach primitives are used instead of pthread's
+  // because the latter doesn't provide thread manipulation primitives required.
+  // For details, consult "Mac OS X Internals" book, Section 7.3.
+  thread_act_t mProfiledThread;
+};
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN Sampler target specifics
+
+Sampler::Sampler(PSLockRef aLock) {}
+
+void Sampler::Disable(PSLockRef aLock) {}
+
+template <typename Func>
+void Sampler::SuspendAndSampleAndResumeThread(
+    PSLockRef aLock, const RegisteredThread& aRegisteredThread,
+    const TimeStamp& aNow, const Func& aProcessRegs) {
+  thread_act_t samplee_thread =
+      aRegisteredThread.GetPlatformData()->ProfiledThread();
+
+  //----------------------------------------------------------------//
+  // Suspend the samplee thread and get its context.
+
+  // We're using thread_suspend on OS X because pthread_kill (which is what we
+  // at one time used on Linux) has less consistent performance and causes
+  // strange crashes, see bug 1166778 and bug 1166808.  thread_suspend
+  // is also just a lot simpler to use.
+
+  if (KERN_SUCCESS != thread_suspend(samplee_thread)) {
+    return;
+  }
+
+  //----------------------------------------------------------------//
+  // Sample the target thread.
+
+  // WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+  //
+  // The profiler's "critical section" begins here.  We must be very careful
+  // what we do here, or risk deadlock.  See the corresponding comment in
+  // platform-linux-android.cpp for details.
+
+#if defined(__x86_64__)
+  thread_state_flavor_t flavor = x86_THREAD_STATE64;
+  x86_thread_state64_t state;
+  mach_msg_type_number_t count = x86_THREAD_STATE64_COUNT;
+#  if __DARWIN_UNIX03
+#    define REGISTER_FIELD(name) __r##name
+#  else
+#    define REGISTER_FIELD(name) r##name
+#  endif  // __DARWIN_UNIX03
+#elif defined(__aarch64__)
+  thread_state_flavor_t flavor = ARM_THREAD_STATE64;
+  arm_thread_state64_t state;
+  mach_msg_type_number_t count = ARM_THREAD_STATE64_COUNT;
+#  if __DARWIN_UNIX03
+#    define REGISTER_FIELD(name) __##name
+#  else
+#    define REGISTER_FIELD(name) name
+#  endif  // __DARWIN_UNIX03
+#else
+#  error "unknown architecture"
+#endif
+
+  if (thread_get_state(samplee_thread, flavor,
+                       reinterpret_cast<natural_t*>(&state),
+                       &count) == KERN_SUCCESS) {
+    Registers regs;
+#if defined(__x86_64__)
+    regs.mPC = reinterpret_cast<Address>(state.REGISTER_FIELD(ip));
+    regs.mSP = reinterpret_cast<Address>(state.REGISTER_FIELD(sp));
+    regs.mFP = reinterpret_cast<Address>(state.REGISTER_FIELD(bp));
+#elif defined(__aarch64__)
+    regs.mPC = reinterpret_cast<Address>(state.REGISTER_FIELD(pc));
+    regs.mSP = reinterpret_cast<Address>(state.REGISTER_FIELD(sp));
+    regs.mFP = reinterpret_cast<Address>(state.REGISTER_FIELD(fp));
+#else
+#  error "unknown architecture"
+#endif
+    regs.mLR = 0;
+
+    aProcessRegs(regs, aNow);
+  }
+
+#undef REGISTER_FIELD
+
+  //----------------------------------------------------------------//
+  // Resume the target thread.
+
+  thread_resume(samplee_thread);
+
+  // The profiler's critical section ends here.
+  //
+  // WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+}
+
+// END Sampler target specifics
+////////////////////////////////////////////////////////////////////////
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN SamplerThread target specifics
+
+static void* ThreadEntry(void* aArg) {
+  auto thread = static_cast<SamplerThread*>(aArg);
+  thread->Run();
+  return nullptr;
+}
+
+SamplerThread::SamplerThread(PSLockRef aLock, uint32_t aActivityGeneration,
+                             double aIntervalMilliseconds)
+    : mSampler(aLock),
+      mActivityGeneration(aActivityGeneration),
+      mIntervalMicroseconds(
+          std::max(1, int(floor(aIntervalMilliseconds * 1000 + 0.5)))),
+      mThread{nullptr} {
+  pthread_attr_t* attr_ptr = nullptr;
+  if (pthread_create(&mThread, attr_ptr, ThreadEntry, this) != 0) {
+    MOZ_CRASH("pthread_create failed");
+  }
+}
+
+SamplerThread::~SamplerThread() { pthread_join(mThread, nullptr); }
+
+void SamplerThread::SleepMicro(uint32_t aMicroseconds) {
+  usleep(aMicroseconds);
+  // FIXME: the OSX 10.12 page for usleep says "The usleep() function is
+  // obsolescent.  Use nanosleep(2) instead."  This implementation could be
+  // merged with the linux-android version.  Also, this doesn't handle the
+  // case where the usleep call is interrupted by a signal.
+}
+
+void SamplerThread::Stop(PSLockRef aLock) { mSampler.Disable(aLock); }
+
+// END SamplerThread target specifics
+////////////////////////////////////////////////////////////////////////
+
+static void PlatformInit(PSLockRef aLock) {}
+
+#if defined(HAVE_NATIVE_UNWIND)
+void Registers::SyncPopulate() {
+#  if defined(__x86_64__)
+  asm(
+      // Compute caller's %rsp by adding to %rbp:
+      // 8 bytes for previous %rbp, 8 bytes for return address
+      "leaq 0x10(%%rbp), %0\n\t"
+      // Dereference %rbp to get previous %rbp
+      "movq (%%rbp), %1\n\t"
+      : "=r"(mSP), "=r"(mFP));
+#  elif defined(__aarch64__)
+  asm(
+      // Compute caller's sp by adding to fp:
+      // 8 bytes for previous fp, 8 bytes for return address
+      "add %0, x29, #0x10\n\t"
+      // Dereference fp to get previous fp
+      "ldr %1, [x29]\n\t"
+      : "=r"(mSP), "=r"(mFP));
+#  else
+#    error "unknown architecture"
+#  endif
+  mPC = reinterpret_cast<Address>(
+      __builtin_extract_return_addr(__builtin_return_address(0)));
+  mLR = 0;
+}
+#endif
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/platform-win32.cpp b/mozglue/baseprofiler/core/platform-win32.cpp
new file mode 100644
index 0000000000..22b8a8462b
--- /dev/null
+++ b/mozglue/baseprofiler/core/platform-win32.cpp
@@ -0,0 +1,351 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+// Copyright (c) 2006-2011 The Chromium Authors. All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//  * Redistributions of source code must retain the above copyright
+//    notice, this list of conditions and the following disclaimer.
+//  * Redistributions in binary form must reproduce the above copyright
+//    notice, this list of conditions and the following disclaimer in
+//    the documentation and/or other materials provided with the
+//    distribution.
+//  * Neither the name of Google, Inc. nor the names of its contributors
+//    may be used to endorse or promote products derived from this
+//    software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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 OWNER OR CONTRIBUTORS 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.
+
+#include <windows.h>
+#include <mmsystem.h>
+#include <process.h>
+
+#include "nsWindowsDllInterceptor.h"
+#include "mozilla/StackWalk_windows.h"
+#include "mozilla/WindowsVersion.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+int profiler_current_process_id() { return _getpid(); }
+
+int profiler_current_thread_id() {
+  DWORD threadId = GetCurrentThreadId();
+  MOZ_ASSERT(threadId <= INT32_MAX, "native thread ID is > INT32_MAX");
+  return int(threadId);
+}
+
+static int64_t MicrosecondsSince1970() {
+  int64_t prt;
+  FILETIME ft;
+  SYSTEMTIME st;
+
+  GetSystemTime(&st);
+  SystemTimeToFileTime(&st, &ft);
+  static_assert(sizeof(ft) == sizeof(prt), "Expect FILETIME to be 64 bits");
+  memcpy(&prt, &ft, sizeof(prt));
+  const int64_t epochBias = 116444736000000000LL;
+  prt = (prt - epochBias) / 10;
+
+  return prt;
+}
+
+void* GetStackTop(void* aGuess) {
+  PNT_TIB pTib = reinterpret_cast<PNT_TIB>(NtCurrentTeb());
+  return reinterpret_cast<void*>(pTib->StackBase);
+}
+
+static void PopulateRegsFromContext(Registers& aRegs, CONTEXT* aContext) {
+#if defined(GP_ARCH_amd64)
+  aRegs.mPC = reinterpret_cast<Address>(aContext->Rip);
+  aRegs.mSP = reinterpret_cast<Address>(aContext->Rsp);
+  aRegs.mFP = reinterpret_cast<Address>(aContext->Rbp);
+#elif defined(GP_ARCH_x86)
+  aRegs.mPC = reinterpret_cast<Address>(aContext->Eip);
+  aRegs.mSP = reinterpret_cast<Address>(aContext->Esp);
+  aRegs.mFP = reinterpret_cast<Address>(aContext->Ebp);
+#elif defined(GP_ARCH_arm64)
+  aRegs.mPC = reinterpret_cast<Address>(aContext->Pc);
+  aRegs.mSP = reinterpret_cast<Address>(aContext->Sp);
+  aRegs.mFP = reinterpret_cast<Address>(aContext->Fp);
+#else
+#  error "bad arch"
+#endif
+  aRegs.mLR = 0;
+}
+
+// Gets a real (i.e. not pseudo) handle for the current thread, with the
+// permissions needed for profiling.
+// @return a real HANDLE for the current thread.
+static HANDLE GetRealCurrentThreadHandleForProfiling() {
+  HANDLE realCurrentThreadHandle;
+  if (!::DuplicateHandle(
+          ::GetCurrentProcess(), ::GetCurrentThread(), ::GetCurrentProcess(),
+          &realCurrentThreadHandle,
+          THREAD_GET_CONTEXT | THREAD_SUSPEND_RESUME | THREAD_QUERY_INFORMATION,
+          FALSE, 0)) {
+    return nullptr;
+  }
+
+  return realCurrentThreadHandle;
+}
+
+class PlatformData {
+ public:
+  // Get a handle to the calling thread. This is the thread that we are
+  // going to profile. We need a real handle because we are going to use it in
+  // the sampler thread.
+  explicit PlatformData(int aThreadId)
+      : mProfiledThread(GetRealCurrentThreadHandleForProfiling()) {
+    MOZ_ASSERT(aThreadId == ::GetCurrentThreadId());
+  }
+
+  ~PlatformData() {
+    if (mProfiledThread != nullptr) {
+      CloseHandle(mProfiledThread);
+      mProfiledThread = nullptr;
+    }
+  }
+
+  HANDLE ProfiledThread() { return mProfiledThread; }
+
+ private:
+  HANDLE mProfiledThread;
+};
+
+#if defined(USE_MOZ_STACK_WALK)
+HANDLE
+GetThreadHandle(PlatformData* aData) { return aData->ProfiledThread(); }
+#endif
+
+static const HANDLE kNoThread = INVALID_HANDLE_VALUE;
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN Sampler target specifics
+
+Sampler::Sampler(PSLockRef aLock) {}
+
+void Sampler::Disable(PSLockRef aLock) {}
+
+template <typename Func>
+void Sampler::SuspendAndSampleAndResumeThread(
+    PSLockRef aLock, const RegisteredThread& aRegisteredThread,
+    const TimeStamp& aNow, const Func& aProcessRegs) {
+  HANDLE profiled_thread =
+      aRegisteredThread.GetPlatformData()->ProfiledThread();
+  if (profiled_thread == nullptr) {
+    return;
+  }
+
+  // Context used for sampling the register state of the profiled thread.
+  CONTEXT context;
+  memset(&context, 0, sizeof(context));
+
+  //----------------------------------------------------------------//
+  // Suspend the samplee thread and get its context.
+
+  static const DWORD kSuspendFailed = static_cast<DWORD>(-1);
+  if (SuspendThread(profiled_thread) == kSuspendFailed) {
+    return;
+  }
+
+  // SuspendThread is asynchronous, so the thread may still be running.
+  // Call GetThreadContext first to ensure the thread is really suspended.
+  // See https://blogs.msdn.microsoft.com/oldnewthing/20150205-00/?p=44743.
+
+  // Using only CONTEXT_CONTROL is faster but on 64-bit it causes crashes in
+  // RtlVirtualUnwind (see bug 1120126) so we set all the flags.
+#if defined(GP_ARCH_amd64)
+  context.ContextFlags = CONTEXT_FULL;
+#else
+  context.ContextFlags = CONTEXT_CONTROL;
+#endif
+  if (!GetThreadContext(profiled_thread, &context)) {
+    ResumeThread(profiled_thread);
+    return;
+  }
+
+  //----------------------------------------------------------------//
+  // Sample the target thread.
+
+  // WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+  //
+  // The profiler's "critical section" begins here.  We must be very careful
+  // what we do here, or risk deadlock.  See the corresponding comment in
+  // platform-linux-android.cpp for details.
+
+  Registers regs;
+  PopulateRegsFromContext(regs, &context);
+  aProcessRegs(regs, aNow);
+
+  //----------------------------------------------------------------//
+  // Resume the target thread.
+
+  ResumeThread(profiled_thread);
+
+  // The profiler's critical section ends here.
+  //
+  // WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+}
+
+// END Sampler target specifics
+////////////////////////////////////////////////////////////////////////
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN SamplerThread target specifics
+
+static unsigned int __stdcall ThreadEntry(void* aArg) {
+  auto thread = static_cast<SamplerThread*>(aArg);
+  thread->Run();
+  return 0;
+}
+
+SamplerThread::SamplerThread(PSLockRef aLock, uint32_t aActivityGeneration,
+                             double aIntervalMilliseconds)
+    : mSampler(aLock),
+      mActivityGeneration(aActivityGeneration),
+      mIntervalMicroseconds(
+          std::max(1, int(floor(aIntervalMilliseconds * 1000 + 0.5)))) {
+  // By default we'll not adjust the timer resolution which tends to be
+  // around 16ms. However, if the requested interval is sufficiently low
+  // we'll try to adjust the resolution to match.
+  if (mIntervalMicroseconds < 10 * 1000) {
+    ::timeBeginPeriod(mIntervalMicroseconds / 1000);
+  }
+
+  // Create a new thread. It is important to use _beginthreadex() instead of
+  // the Win32 function CreateThread(), because the CreateThread() does not
+  // initialize thread-specific structures in the C runtime library.
+  mThread = reinterpret_cast<HANDLE>(_beginthreadex(nullptr,
+                                                    /* stack_size */ 0,
+                                                    ThreadEntry, this,
+                                                    /* initflag */ 0, nullptr));
+  if (mThread == 0) {
+    MOZ_CRASH("_beginthreadex failed");
+  }
+}
+
+SamplerThread::~SamplerThread() {
+  WaitForSingleObject(mThread, INFINITE);
+
+  // Close our own handle for the thread.
+  if (mThread != kNoThread) {
+    CloseHandle(mThread);
+  }
+}
+
+void SamplerThread::SleepMicro(uint32_t aMicroseconds) {
+  // For now, keep the old behaviour of minimum Sleep(1), even for
+  // smaller-than-usual sleeps after an overshoot, unless the user has
+  // explicitly opted into a sub-millisecond profiler interval.
+  if (mIntervalMicroseconds >= 1000) {
+    ::Sleep(std::max(1u, aMicroseconds / 1000));
+  } else {
+    TimeStamp start = TimeStamp::NowUnfuzzed();
+    TimeStamp end = start + TimeDuration::FromMicroseconds(aMicroseconds);
+
+    // First, sleep for as many whole milliseconds as possible.
+    if (aMicroseconds >= 1000) {
+      ::Sleep(aMicroseconds / 1000);
+    }
+
+    // Then, spin until enough time has passed.
+    while (TimeStamp::NowUnfuzzed() < end) {
+      YieldProcessor();
+    }
+  }
+}
+
+void SamplerThread::Stop(PSLockRef aLock) {
+  // Disable any timer resolution changes we've made. Do it now while
+  // gPSMutex is locked, i.e. before any other SamplerThread can be created
+  // and call ::timeBeginPeriod().
+  //
+  // It's safe to do this now even though this SamplerThread is still alive,
+  // because the next time the main loop of Run() iterates it won't get past
+  // the mActivityGeneration check, and so it won't make any more ::Sleep()
+  // calls.
+  if (mIntervalMicroseconds < 10 * 1000) {
+    ::timeEndPeriod(mIntervalMicroseconds / 1000);
+  }
+
+  mSampler.Disable(aLock);
+}
+
+// END SamplerThread target specifics
+////////////////////////////////////////////////////////////////////////
+
+static void PlatformInit(PSLockRef aLock) {}
+
+#if defined(HAVE_NATIVE_UNWIND)
+void Registers::SyncPopulate() {
+  CONTEXT context;
+  RtlCaptureContext(&context);
+  PopulateRegsFromContext(*this, &context);
+}
+#endif
+
+#if defined(GP_PLAT_amd64_windows)
+static WindowsDllInterceptor NtDllIntercept;
+
+typedef NTSTATUS(NTAPI* LdrUnloadDll_func)(HMODULE module);
+static WindowsDllInterceptor::FuncHookType<LdrUnloadDll_func> stub_LdrUnloadDll;
+
+static NTSTATUS NTAPI patched_LdrUnloadDll(HMODULE module) {
+  // Prevent the stack walker from suspending this thread when LdrUnloadDll
+  // holds the RtlLookupFunctionEntry lock.
+  AutoSuppressStackWalking suppress;
+  return stub_LdrUnloadDll(module);
+}
+
+// These pointers are disguised as PVOID to avoid pulling in obscure headers
+typedef PVOID(WINAPI* LdrResolveDelayLoadedAPI_func)(
+    PVOID ParentModuleBase, PVOID DelayloadDescriptor, PVOID FailureDllHook,
+    PVOID FailureSystemHook, PVOID ThunkAddress, ULONG Flags);
+static WindowsDllInterceptor::FuncHookType<LdrResolveDelayLoadedAPI_func>
+    stub_LdrResolveDelayLoadedAPI;
+
+static PVOID WINAPI patched_LdrResolveDelayLoadedAPI(
+    PVOID ParentModuleBase, PVOID DelayloadDescriptor, PVOID FailureDllHook,
+    PVOID FailureSystemHook, PVOID ThunkAddress, ULONG Flags) {
+  // Prevent the stack walker from suspending this thread when
+  // LdrResolveDelayLoadAPI holds the RtlLookupFunctionEntry lock.
+  AutoSuppressStackWalking suppress;
+  return stub_LdrResolveDelayLoadedAPI(ParentModuleBase, DelayloadDescriptor,
+                                       FailureDllHook, FailureSystemHook,
+                                       ThunkAddress, Flags);
+}
+
+MFBT_API void InitializeWin64ProfilerHooks() {
+  // This function could be called by both profilers, but we only want to run
+  // it once.
+  static bool ran = false;
+  if (ran) {
+    return;
+  }
+  ran = true;
+
+  NtDllIntercept.Init("ntdll.dll");
+  stub_LdrUnloadDll.Set(NtDllIntercept, "LdrUnloadDll", &patched_LdrUnloadDll);
+  if (IsWin8OrLater()) {  // LdrResolveDelayLoadedAPI was introduced in Win8
+    stub_LdrResolveDelayLoadedAPI.Set(NtDllIntercept,
+                                      "LdrResolveDelayLoadedAPI",
+                                      &patched_LdrResolveDelayLoadedAPI);
+  }
+}
+#endif  // defined(GP_PLAT_amd64_windows)
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/platform.cpp b/mozglue/baseprofiler/core/platform.cpp
new file mode 100644
index 0000000000..14c48ce649
--- /dev/null
+++ b/mozglue/baseprofiler/core/platform.cpp
@@ -0,0 +1,3712 @@
+/* -*- 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/. */
+
+// There are three kinds of samples done by the profiler.
+//
+// - A "periodic" sample is the most complex kind. It is done in response to a
+//   timer while the profiler is active. It involves writing a stack trace plus
+//   a variety of other values (memory measurements, responsiveness
+//   measurements, etc.) into the main ProfileBuffer. The sampling is done from
+//   off-thread, and so SuspendAndSampleAndResumeThread() is used to get the
+//   register values.
+//
+// - A "synchronous" sample is a simpler kind. It is done in response to an API
+//   call (profiler_get_backtrace()). It involves writing a stack trace and
+//   little else into a temporary ProfileBuffer, and wrapping that up in a
+//   ProfilerBacktrace that can be subsequently used in a marker. The sampling
+//   is done on-thread, and so Registers::SyncPopulate() is used to get the
+//   register values.
+//
+// - A "backtrace" sample is the simplest kind. It is done in response to an
+//   API call (profiler_suspend_and_sample_thread()). It involves getting a
+//   stack trace via a ProfilerStackCollector; it does not write to a
+//   ProfileBuffer. The sampling is done from off-thread, and so uses
+//   SuspendAndSampleAndResumeThread() to get the register values.
+
+#include "platform.h"
+
+#include <algorithm>
+#include <errno.h>
+#include <fstream>
+#include <ostream>
+#include <set>
+#include <sstream>
+
+// #include "memory_hooks.h"
+#include "mozilla/ArrayUtils.h"
+#include "mozilla/Atomics.h"
+#include "mozilla/AutoProfilerLabel.h"
+#include "mozilla/BaseProfilerDetail.h"
+#include "mozilla/DoubleConversion.h"
+#include "mozilla/Printf.h"
+#include "mozilla/ProfileBufferChunkManagerSingle.h"
+#include "mozilla/ProfileBufferChunkManagerWithLocalLimit.h"
+#include "mozilla/ProfileChunkedBuffer.h"
+#include "mozilla/Services.h"
+#include "mozilla/Span.h"
+#include "mozilla/StackWalk.h"
+#include "mozilla/StaticPtr.h"
+#include "mozilla/ThreadLocal.h"
+#include "mozilla/TimeStamp.h"
+#include "mozilla/Tuple.h"
+#include "mozilla/UniquePtr.h"
+#include "mozilla/Vector.h"
+#include "prdtoa.h"
+#include "prtime.h"
+
+#include "BaseProfiler.h"
+#include "BaseProfilingCategory.h"
+#include "PageInformation.h"
+#include "ProfiledThreadData.h"
+#include "ProfilerBacktrace.h"
+#include "ProfileBuffer.h"
+#include "RegisteredThread.h"
+#include "BaseProfilerSharedLibraries.h"
+#include "ThreadInfo.h"
+#include "VTuneProfiler.h"
+
+// Win32 builds always have frame pointers, so FramePointerStackWalk() always
+// works.
+#if defined(GP_PLAT_x86_windows)
+#  define HAVE_NATIVE_UNWIND
+#  define USE_FRAME_POINTER_STACK_WALK
+#endif
+
+// Win64 builds always omit frame pointers, so we use the slower
+// MozStackWalk(), which works in that case.
+#if defined(GP_PLAT_amd64_windows)
+#  define HAVE_NATIVE_UNWIND
+#  define USE_MOZ_STACK_WALK
+#endif
+
+// AArch64 Win64 doesn't seem to use frame pointers, so we use the slower
+// MozStackWalk().
+#if defined(GP_PLAT_arm64_windows)
+#  define HAVE_NATIVE_UNWIND
+#  define USE_MOZ_STACK_WALK
+#endif
+
+// Mac builds only have frame pointers when MOZ_PROFILING is specified, so
+// FramePointerStackWalk() only works in that case. We don't use MozStackWalk()
+// on Mac.
+#if defined(GP_OS_darwin) && defined(MOZ_PROFILING)
+#  define HAVE_NATIVE_UNWIND
+#  define USE_FRAME_POINTER_STACK_WALK
+#endif
+
+// No stack-walking in baseprofiler on linux, android, bsd.
+// APIs now make it easier to capture backtraces from the Base Profiler, which
+// is currently not supported on these platform, and would lead to a MOZ_CRASH
+// in Registers::SyncPopulate(). `#if 0` added in bug 1658232, follow-up bugs
+// should be referenced in meta bug 1557568.
+#if 0
+// Android builds use the ARM Exception Handling ABI to unwind.
+#  if defined(GP_PLAT_arm_linux) || defined(GP_PLAT_arm_android)
+#    define HAVE_NATIVE_UNWIND
+#    define USE_EHABI_STACKWALK
+#    include "EHABIStackWalk.h"
+#  endif
+
+// Linux/BSD builds use LUL, which uses DWARF info to unwind stacks.
+#  if defined(GP_PLAT_amd64_linux) || defined(GP_PLAT_x86_linux) ||       \
+      defined(GP_PLAT_amd64_android) || defined(GP_PLAT_x86_android) ||   \
+      defined(GP_PLAT_mips64_linux) || defined(GP_PLAT_arm64_linux) ||    \
+      defined(GP_PLAT_arm64_android) || defined(GP_PLAT_amd64_freebsd) || \
+      defined(GP_PLAT_arm64_freebsd)
+#    define HAVE_NATIVE_UNWIND
+#    define USE_LUL_STACKWALK
+#    include "lul/LulMain.h"
+#    include "lul/platform-linux-lul.h"
+
+// On linux we use LUL for periodic samples and synchronous samples, but we use
+// FramePointerStackWalk for backtrace samples when MOZ_PROFILING is enabled.
+// (See the comment at the top of the file for a definition of
+// periodic/synchronous/backtrace.).
+//
+// FramePointerStackWalk can produce incomplete stacks when the current entry is
+// in a shared library without framepointers, however LUL can take a long time
+// to initialize, which is undesirable for consumers of
+// profiler_suspend_and_sample_thread like the Background Hang Reporter.
+#    if defined(MOZ_PROFILING)
+#      define USE_FRAME_POINTER_STACK_WALK
+#    endif
+#  endif
+#endif
+
+// We can only stackwalk without expensive initialization on platforms which
+// support FramePointerStackWalk or MozStackWalk. LUL Stackwalking requires
+// initializing LUL, and EHABIStackWalk requires initializing EHABI, both of
+// which can be expensive.
+#if defined(USE_FRAME_POINTER_STACK_WALK) || defined(USE_MOZ_STACK_WALK)
+#  define HAVE_FASTINIT_NATIVE_UNWIND
+#endif
+
+#ifdef MOZ_VALGRIND
+#  include <valgrind/memcheck.h>
+#else
+#  define VALGRIND_MAKE_MEM_DEFINED(_addr, _len) ((void)0)
+#endif
+
+#if defined(GP_OS_linux) || defined(GP_OS_android) || defined(GP_OS_freebsd)
+#  include <ucontext.h>
+#endif
+
+namespace mozilla {
+namespace baseprofiler {
+
+using detail::RacyFeatures;
+
+bool LogTest(int aLevelToTest) {
+  static const int maxLevel = getenv("MOZ_BASE_PROFILER_VERBOSE_LOGGING") ? 5
+                              : getenv("MOZ_BASE_PROFILER_DEBUG_LOGGING") ? 4
+                              : getenv("MOZ_BASE_PROFILER_LOGGING")       ? 3
+                                                                          : 0;
+  return aLevelToTest <= maxLevel;
+}
+
+void PrintToConsole(const char* aFmt, ...) {
+  va_list args;
+  va_start(args, aFmt);
+#if defined(ANDROID)
+  __android_log_vprint(ANDROID_LOG_INFO, "Gecko", aFmt, args);
+#else
+  vfprintf(stderr, aFmt, args);
+#endif
+  va_end(args);
+}
+
+// Statically initialized to 0, then set once from profiler_init(), which should
+// be called from the main thread before any other use of the profiler.
+int scProfilerMainThreadId;
+
+constexpr static bool ValidateFeatures() {
+  int expectedFeatureNumber = 0;
+
+  // Feature numbers should start at 0 and increase by 1 each.
+#define CHECK_FEATURE(n_, str_, Name_, desc_) \
+  if ((n_) != expectedFeatureNumber) {        \
+    return false;                             \
+  }                                           \
+  ++expectedFeatureNumber;
+
+  BASE_PROFILER_FOR_EACH_FEATURE(CHECK_FEATURE)
+
+#undef CHECK_FEATURE
+
+  return true;
+}
+
+static_assert(ValidateFeatures(), "Feature list is invalid");
+
+// Return all features that are available on this platform.
+static uint32_t AvailableFeatures() {
+  uint32_t features = 0;
+
+#define ADD_FEATURE(n_, str_, Name_, desc_) \
+  ProfilerFeature::Set##Name_(features);
+
+  // Add all the possible features.
+  BASE_PROFILER_FOR_EACH_FEATURE(ADD_FEATURE)
+
+#undef ADD_FEATURE
+
+  // Now remove features not supported on this platform/configuration.
+  ProfilerFeature::ClearJava(features);
+  ProfilerFeature::ClearJS(features);
+  ProfilerFeature::ClearScreenshots(features);
+#if !defined(HAVE_NATIVE_UNWIND)
+  ProfilerFeature::ClearStackWalk(features);
+#endif
+  ProfilerFeature::ClearTaskTracer(features);
+  ProfilerFeature::ClearJSTracer(features);
+
+  return features;
+}
+
+// Default features common to all contexts (even if not available).
+static uint32_t DefaultFeatures() {
+  return ProfilerFeature::Java | ProfilerFeature::JS | ProfilerFeature::Leaf |
+         ProfilerFeature::StackWalk | ProfilerFeature::Threads;
+}
+
+// Extra default features when MOZ_PROFILER_STARTUP is set (even if not
+// available).
+static uint32_t StartupExtraDefaultFeatures() {
+  // Enable CPUUtilization by default for startup profiles as it is useful to
+  // see when startup alternates between CPU intensive tasks and being blocked.
+  // Enable mainthreadio by default for startup profiles as startup is heavy on
+  // I/O operations, and main thread I/O is really important to see there.
+  return ProfilerFeature::CPUUtilization | ProfilerFeature::MainThreadIO;
+}
+
+class MOZ_RAII PSAutoTryLock;
+
+// The auto-lock/unlock mutex that guards accesses to CorePS and ActivePS.
+// Use `PSAutoLock lock;` to take the lock until the end of the enclosing block.
+// External profilers may use this same lock for their own data, but as the lock
+// is non-recursive, *only* `f(PSLockRef, ...)` functions below should be
+// called, to avoid double-locking.
+class MOZ_RAII PSAutoLock {
+ public:
+  PSAutoLock() { gPSMutex.Lock(); }
+
+  ~PSAutoLock() { gPSMutex.Unlock(); }
+
+  PSAutoLock(const PSAutoLock&) = delete;
+  void operator=(const PSAutoLock&) = delete;
+
+  [[nodiscard]] static bool IsLockedOnCurrentThread() {
+    return gPSMutex.IsLockedOnCurrentThread();
+  }
+
+ private:
+  // Allow PSAutoTryLock to access gPSMutex, and to call the following
+  // `PSAutoLock(int)` constructor through `Maybe<const PSAutoLock>::emplace()`.
+  friend class PSAutoTryLock;
+  friend class Maybe<const PSAutoLock>;
+
+  // Special constructor for an already-locked gPSMutex. The `int` parameter is
+  // necessary to distinguish it from the main constructor.
+  explicit PSAutoLock(int) { gPSMutex.AssertCurrentThreadOwns(); }
+
+  static detail::BaseProfilerMutex gPSMutex;
+};
+
+// RAII class that attempts to lock the profiler mutex. Example usage:
+//   PSAutoTryLock tryLock;
+//   if (tryLock.IsLocked()) { locked_foo(tryLock.LockRef()); }
+class MOZ_RAII PSAutoTryLock {
+ public:
+  PSAutoTryLock() {
+    if (PSAutoLock::gPSMutex.TryLock()) {
+      mMaybePSAutoLock.emplace(0);
+    }
+  }
+
+  // Return true if the mutex was aquired and locked.
+  [[nodiscard]] bool IsLocked() const { return mMaybePSAutoLock.isSome(); }
+
+  // Assuming the mutex is locked, return a reference to a `PSAutoLock` for that
+  // mutex, which can be passed as proof-of-lock.
+  [[nodiscard]] const PSAutoLock& LockRef() const {
+    MOZ_ASSERT(IsLocked());
+    return mMaybePSAutoLock.ref();
+  }
+
+ private:
+  // `mMaybePSAutoLock` is `Nothing` if locking failed, otherwise it contains a
+  // `const PSAutoLock` holding the locked mutex, and whose reference may be
+  // passed to functions expecting a proof-of-lock.
+  Maybe<const PSAutoLock> mMaybePSAutoLock;
+};
+
+detail::BaseProfilerMutex PSAutoLock::gPSMutex;
+
+// Only functions that take a PSLockRef arg can access CorePS's and ActivePS's
+// fields.
+typedef const PSAutoLock& PSLockRef;
+
+#define PS_GET(type_, name_)      \
+  static type_ name_(PSLockRef) { \
+    MOZ_ASSERT(sInstance);        \
+    return sInstance->m##name_;   \
+  }
+
+#define PS_GET_LOCKLESS(type_, name_) \
+  static type_ name_() {              \
+    MOZ_ASSERT(sInstance);            \
+    return sInstance->m##name_;       \
+  }
+
+#define PS_GET_AND_SET(type_, name_)                  \
+  PS_GET(type_, name_)                                \
+  static void Set##name_(PSLockRef, type_ a##name_) { \
+    MOZ_ASSERT(sInstance);                            \
+    sInstance->m##name_ = a##name_;                   \
+  }
+
+// All functions in this file can run on multiple threads unless they have an
+// NS_IsMainThread() assertion.
+
+// This class contains the profiler's core global state, i.e. that which is
+// valid even when the profiler is not active. Most profile operations can't do
+// anything useful when this class is not instantiated, so we release-assert
+// its non-nullness in all such operations.
+//
+// Accesses to CorePS are guarded by gPSMutex. Getters and setters take a
+// PSAutoLock reference as an argument as proof that the gPSMutex is currently
+// locked. This makes it clear when gPSMutex is locked and helps avoid
+// accidental unlocked accesses to global state. There are ways to circumvent
+// this mechanism, but please don't do so without *very* good reason and a
+// detailed explanation.
+//
+// The exceptions to this rule:
+//
+// - mProcessStartTime, because it's immutable;
+//
+// - each thread's RacyRegisteredThread object is accessible without locking via
+//   TLSRegisteredThread::RacyRegisteredThread().
+class CorePS {
+ private:
+  CorePS()
+      : mProcessStartTime(TimeStamp::ProcessCreation()),
+        // This needs its own mutex, because it is used concurrently from
+        // functions guarded by gPSMutex as well as others without safety (e.g.,
+        // profiler_add_marker). It is *not* used inside the critical section of
+        // the sampler, because mutexes cannot be used there.
+        mCoreBuffer(ProfileChunkedBuffer::ThreadSafety::WithMutex)
+#ifdef USE_LUL_STACKWALK
+        ,
+        mLul(nullptr)
+#endif
+  {
+  }
+
+  ~CorePS() {}
+
+ public:
+  static void Create(PSLockRef aLock) {
+    MOZ_ASSERT(!sInstance);
+    sInstance = new CorePS();
+  }
+
+  static void Destroy(PSLockRef aLock) {
+    MOZ_ASSERT(sInstance);
+    delete sInstance;
+    sInstance = nullptr;
+  }
+
+  // Unlike ActivePS::Exists(), CorePS::Exists() can be called without gPSMutex
+  // being locked. This is because CorePS is instantiated so early on the main
+  // thread that we don't have to worry about it being racy.
+  static bool Exists() { return !!sInstance; }
+
+  static void AddSizeOf(PSLockRef, MallocSizeOf aMallocSizeOf,
+                        size_t& aProfSize, size_t& aLulSize) {
+    MOZ_ASSERT(sInstance);
+
+    aProfSize += aMallocSizeOf(sInstance);
+
+    for (auto& registeredThread : sInstance->mRegisteredThreads) {
+      aProfSize += registeredThread->SizeOfIncludingThis(aMallocSizeOf);
+    }
+
+    for (auto& registeredPage : sInstance->mRegisteredPages) {
+      aProfSize += registeredPage->SizeOfIncludingThis(aMallocSizeOf);
+    }
+
+    // Measurement of the following things may be added later if DMD finds it
+    // is worthwhile:
+    // - CorePS::mRegisteredThreads itself (its elements' children are
+    // measured above)
+    // - CorePS::mRegisteredPages itself (its elements' children are
+    // measured above)
+    // - CorePS::mInterposeObserver
+
+#if defined(USE_LUL_STACKWALK)
+    if (sInstance->mLul) {
+      aLulSize += sInstance->mLul->SizeOfIncludingThis(aMallocSizeOf);
+    }
+#endif
+  }
+
+  // No PSLockRef is needed for this field because it's immutable.
+  PS_GET_LOCKLESS(const TimeStamp&, ProcessStartTime)
+
+  // No PSLockRef is needed for this field because it's thread-safe.
+  PS_GET_LOCKLESS(ProfileChunkedBuffer&, CoreBuffer)
+
+  PS_GET(const Vector<UniquePtr<RegisteredThread>>&, RegisteredThreads)
+
+  static void AppendRegisteredThread(
+      PSLockRef, UniquePtr<RegisteredThread>&& aRegisteredThread) {
+    MOZ_ASSERT(sInstance);
+    MOZ_RELEASE_ASSERT(
+        sInstance->mRegisteredThreads.append(std::move(aRegisteredThread)));
+  }
+
+  static void RemoveRegisteredThread(PSLockRef,
+                                     RegisteredThread* aRegisteredThread) {
+    MOZ_ASSERT(sInstance);
+    // Remove aRegisteredThread from mRegisteredThreads.
+    for (UniquePtr<RegisteredThread>& rt : sInstance->mRegisteredThreads) {
+      if (rt.get() == aRegisteredThread) {
+        sInstance->mRegisteredThreads.erase(&rt);
+        return;
+      }
+    }
+  }
+
+  PS_GET(Vector<RefPtr<PageInformation>>&, RegisteredPages)
+
+  static void AppendRegisteredPage(PSLockRef,
+                                   RefPtr<PageInformation>&& aRegisteredPage) {
+    MOZ_ASSERT(sInstance);
+    struct RegisteredPageComparator {
+      PageInformation* aA;
+      bool operator()(PageInformation* aB) const { return aA->Equals(aB); }
+    };
+
+    auto foundPageIter = std::find_if(
+        sInstance->mRegisteredPages.begin(), sInstance->mRegisteredPages.end(),
+        RegisteredPageComparator{aRegisteredPage.get()});
+
+    if (foundPageIter != sInstance->mRegisteredPages.end()) {
+      if ((*foundPageIter)->Url() == "about:blank") {
+        // When a BrowsingContext is loaded, the first url loaded in it will be
+        // about:blank, and if the principal matches, the first document loaded
+        // in it will share an inner window. That's why we should delete the
+        // intermittent about:blank if they share the inner window.
+        sInstance->mRegisteredPages.erase(foundPageIter);
+      } else {
+        // Do not register the same page again.
+        return;
+      }
+    }
+    MOZ_RELEASE_ASSERT(
+        sInstance->mRegisteredPages.append(std::move(aRegisteredPage)));
+  }
+
+  static void RemoveRegisteredPage(PSLockRef,
+                                   uint64_t aRegisteredInnerWindowID) {
+    MOZ_ASSERT(sInstance);
+    // Remove RegisteredPage from mRegisteredPages by given inner window ID.
+    sInstance->mRegisteredPages.eraseIf([&](const RefPtr<PageInformation>& rd) {
+      return rd->InnerWindowID() == aRegisteredInnerWindowID;
+    });
+  }
+
+  static void ClearRegisteredPages(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    sInstance->mRegisteredPages.clear();
+  }
+
+  PS_GET(const Vector<BaseProfilerCount*>&, Counters)
+
+  static void AppendCounter(PSLockRef, BaseProfilerCount* aCounter) {
+    MOZ_ASSERT(sInstance);
+    // we don't own the counter; they may be stored in static objects
+    MOZ_RELEASE_ASSERT(sInstance->mCounters.append(aCounter));
+  }
+
+  static void RemoveCounter(PSLockRef, BaseProfilerCount* aCounter) {
+    // we may be called to remove a counter after the profiler is stopped or
+    // late in shutdown.
+    if (sInstance) {
+      auto* counter = std::find(sInstance->mCounters.begin(),
+                                sInstance->mCounters.end(), aCounter);
+      MOZ_RELEASE_ASSERT(counter != sInstance->mCounters.end());
+      sInstance->mCounters.erase(counter);
+    }
+  }
+
+#ifdef USE_LUL_STACKWALK
+  static lul::LUL* Lul(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    return sInstance->mLul.get();
+  }
+  static void SetLul(PSLockRef, UniquePtr<lul::LUL> aLul) {
+    MOZ_ASSERT(sInstance);
+    sInstance->mLul = std::move(aLul);
+  }
+#endif
+
+  PS_GET_AND_SET(const std::string&, ProcessName)
+  PS_GET_AND_SET(const std::string&, ETLDplus1)
+
+ private:
+  // The singleton instance
+  static CorePS* sInstance;
+
+  // The time that the process started.
+  const TimeStamp mProcessStartTime;
+
+  // The thread-safe blocks-oriented buffer into which all profiling data is
+  // recorded.
+  // ActivePS controls the lifetime of the underlying contents buffer: When
+  // ActivePS does not exist, mCoreBuffer is empty and rejects all reads&writes;
+  // see ActivePS for further details.
+  // Note: This needs to live here outside of ActivePS, because some producers
+  // are indirectly controlled (e.g., by atomic flags) and therefore may still
+  // attempt to write some data shortly after ActivePS has shutdown and deleted
+  // the underlying buffer in memory.
+  ProfileChunkedBuffer mCoreBuffer;
+
+  // Info on all the registered threads.
+  // ThreadIds in mRegisteredThreads are unique.
+  Vector<UniquePtr<RegisteredThread>> mRegisteredThreads;
+
+  // Info on all the registered pages.
+  // InnerWindowIDs in mRegisteredPages are unique.
+  Vector<RefPtr<PageInformation>> mRegisteredPages;
+
+  // Non-owning pointers to all active counters
+  Vector<BaseProfilerCount*> mCounters;
+
+#ifdef USE_LUL_STACKWALK
+  // LUL's state. Null prior to the first activation, non-null thereafter.
+  UniquePtr<lul::LUL> mLul;
+#endif
+
+  // Process name, provided by child process initialization code.
+  std::string mProcessName;
+  // Private name, provided by child process initialization code (eTLD+1 in
+  // fission)
+  std::string mETLDplus1;
+};
+
+CorePS* CorePS::sInstance = nullptr;
+
+ProfileChunkedBuffer& profiler_get_core_buffer() {
+  MOZ_ASSERT(CorePS::Exists());
+  return CorePS::CoreBuffer();
+}
+
+class SamplerThread;
+
+static SamplerThread* NewSamplerThread(PSLockRef aLock, uint32_t aGeneration,
+                                       double aInterval);
+
+struct LiveProfiledThreadData {
+  RegisteredThread* mRegisteredThread;
+  UniquePtr<ProfiledThreadData> mProfiledThreadData;
+};
+
+// The buffer size is provided as a number of "entries", this is their size in
+// bytes.
+constexpr static uint32_t scBytesPerEntry = 8;
+
+// This class contains the profiler's global state that is valid only when the
+// profiler is active. When not instantiated, the profiler is inactive.
+//
+// Accesses to ActivePS are guarded by gPSMutex, in much the same fashion as
+// CorePS.
+//
+class ActivePS {
+ private:
+  // We need to decide how many chunks of what size we want to fit in the given
+  // total maximum capacity for this process, in the (likely) context of
+  // multiple processes doing the same choice and having an inter-process
+  // mechanism to control the overal memory limit.
+
+  // Minimum chunk size allowed, enough for at least one stack.
+  constexpr static uint32_t scMinimumChunkSize =
+      2 * ProfileBufferChunkManager::scExpectedMaximumStackSize;
+
+  // Ideally we want at least 2 unreleased chunks to work with (1 current and 1
+  // next), and 2 released chunks (so that one can be recycled when old, leaving
+  // one with some data).
+  constexpr static uint32_t scMinimumNumberOfChunks = 4;
+
+  // And we want to limit chunks to a maximum size, which is a compromise
+  // between:
+  // - A big size, which helps with reducing the rate of allocations and IPCs.
+  // - A small size, which helps with equalizing the duration of recorded data
+  //   (as the inter-process controller will discard the oldest chunks in all
+  //   Firefox processes).
+  constexpr static uint32_t scMaximumChunkSize = 1024 * 1024;
+
+ public:
+  // We should be able to store at least the minimum number of the smallest-
+  // possible chunks.
+  constexpr static uint32_t scMinimumBufferSize =
+      scMinimumNumberOfChunks * scMinimumChunkSize;
+  constexpr static uint32_t scMinimumBufferEntries =
+      scMinimumBufferSize / scBytesPerEntry;
+
+  // Limit to 2GiB.
+  constexpr static uint32_t scMaximumBufferSize = 2u * 1024u * 1024u * 1024u;
+  constexpr static uint32_t scMaximumBufferEntries =
+      scMaximumBufferSize / scBytesPerEntry;
+
+  constexpr static uint32_t ClampToAllowedEntries(uint32_t aEntries) {
+    if (aEntries <= scMinimumBufferEntries) {
+      return scMinimumBufferEntries;
+    }
+    if (aEntries >= scMaximumBufferEntries) {
+      return scMaximumBufferEntries;
+    }
+    return aEntries;
+  }
+
+ private:
+  constexpr static uint32_t ChunkSizeForEntries(uint32_t aEntries) {
+    return uint32_t(std::min(size_t(ClampToAllowedEntries(aEntries)) *
+                                 scBytesPerEntry / scMinimumNumberOfChunks,
+                             size_t(scMaximumChunkSize)));
+  }
+
+  static uint32_t AdjustFeatures(uint32_t aFeatures, uint32_t aFilterCount) {
+    // Filter out any features unavailable in this platform/configuration.
+    aFeatures &= AvailableFeatures();
+
+    // Always enable ProfilerFeature::Threads if we have a filter, because
+    // users sometimes ask to filter by a list of threads but forget to
+    // explicitly specify ProfilerFeature::Threads.
+    if (aFilterCount > 0) {
+      aFeatures |= ProfilerFeature::Threads;
+    }
+
+    // Some features imply others.
+    if (aFeatures & ProfilerFeature::FileIOAll) {
+      aFeatures |= ProfilerFeature::MainThreadIO | ProfilerFeature::FileIO;
+    } else if (aFeatures & ProfilerFeature::FileIO) {
+      aFeatures |= ProfilerFeature::MainThreadIO;
+    }
+
+    return aFeatures;
+  }
+
+  ActivePS(PSLockRef aLock, PowerOfTwo32 aCapacity, double aInterval,
+           uint32_t aFeatures, const char** aFilters, uint32_t aFilterCount,
+           const Maybe<double>& aDuration)
+      : mGeneration(sNextGeneration++),
+        mCapacity(aCapacity),
+        mDuration(aDuration),
+        mInterval(aInterval),
+        mFeatures(AdjustFeatures(aFeatures, aFilterCount)),
+        mProfileBufferChunkManager(
+            size_t(ClampToAllowedEntries(aCapacity.Value())) * scBytesPerEntry,
+            ChunkSizeForEntries(aCapacity.Value())),
+        mProfileBuffer([this]() -> ProfileChunkedBuffer& {
+          CorePS::CoreBuffer().SetChunkManager(mProfileBufferChunkManager);
+          return CorePS::CoreBuffer();
+        }()),
+        // The new sampler thread doesn't start sampling immediately because the
+        // main loop within Run() is blocked until this function's caller
+        // unlocks gPSMutex.
+        mSamplerThread(NewSamplerThread(aLock, mGeneration, aInterval)),
+        mIsPaused(false),
+        mIsSamplingPaused(false)
+#if defined(GP_OS_linux) || defined(GP_OS_freebsd)
+        ,
+        mWasSamplingPaused(false)
+#endif
+  {
+    // Deep copy aFilters.
+    MOZ_ALWAYS_TRUE(mFilters.resize(aFilterCount));
+    for (uint32_t i = 0; i < aFilterCount; ++i) {
+      mFilters[i] = aFilters[i];
+    }
+  }
+
+  ~ActivePS() { CorePS::CoreBuffer().ResetChunkManager(); }
+
+  bool ThreadSelected(const char* aThreadName) {
+    if (mFilters.empty()) {
+      return true;
+    }
+
+    std::string name = aThreadName;
+    std::transform(name.begin(), name.end(), name.begin(), ::tolower);
+
+    for (uint32_t i = 0; i < mFilters.length(); ++i) {
+      std::string filter = mFilters[i];
+
+      if (filter == "*") {
+        return true;
+      }
+
+      std::transform(filter.begin(), filter.end(), filter.begin(), ::tolower);
+
+      // Crude, non UTF-8 compatible, case insensitive substring search
+      if (name.find(filter) != std::string::npos) {
+        return true;
+      }
+
+      // If the filter starts with pid:, check for a pid match
+      if (filter.find("pid:") == 0) {
+        std::string mypid = std::to_string(profiler_current_process_id());
+        if (filter.compare(4, std::string::npos, mypid) == 0) {
+          return true;
+        }
+      }
+    }
+
+    return false;
+  }
+
+ public:
+  static void Create(PSLockRef aLock, PowerOfTwo32 aCapacity, double aInterval,
+                     uint32_t aFeatures, const char** aFilters,
+                     uint32_t aFilterCount, const Maybe<double>& aDuration) {
+    MOZ_ASSERT(!sInstance);
+    sInstance = new ActivePS(aLock, aCapacity, aInterval, aFeatures, aFilters,
+                             aFilterCount, aDuration);
+  }
+
+  [[nodiscard]] static SamplerThread* Destroy(PSLockRef aLock) {
+    MOZ_ASSERT(sInstance);
+    auto samplerThread = sInstance->mSamplerThread;
+    delete sInstance;
+    sInstance = nullptr;
+
+    return samplerThread;
+  }
+
+  static bool Exists(PSLockRef) { return !!sInstance; }
+
+  static bool Equals(PSLockRef, PowerOfTwo32 aCapacity,
+                     const Maybe<double>& aDuration, double aInterval,
+                     uint32_t aFeatures, const char** aFilters,
+                     uint32_t aFilterCount) {
+    MOZ_ASSERT(sInstance);
+    if (sInstance->mCapacity != aCapacity ||
+        sInstance->mDuration != aDuration ||
+        sInstance->mInterval != aInterval ||
+        sInstance->mFeatures != aFeatures ||
+        sInstance->mFilters.length() != aFilterCount) {
+      return false;
+    }
+
+    for (uint32_t i = 0; i < sInstance->mFilters.length(); ++i) {
+      if (strcmp(sInstance->mFilters[i].c_str(), aFilters[i]) != 0) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  static size_t SizeOf(PSLockRef, MallocSizeOf aMallocSizeOf) {
+    MOZ_ASSERT(sInstance);
+
+    size_t n = aMallocSizeOf(sInstance);
+
+    n += sInstance->mProfileBuffer.SizeOfExcludingThis(aMallocSizeOf);
+
+    // Measurement of the following members may be added later if DMD finds it
+    // is worthwhile:
+    // - mLiveProfiledThreads (both the array itself, and the contents)
+    // - mDeadProfiledThreads (both the array itself, and the contents)
+    //
+
+    return n;
+  }
+
+  static bool ShouldProfileThread(PSLockRef aLock, ThreadInfo* aInfo) {
+    MOZ_ASSERT(sInstance);
+    return ((aInfo->IsMainThread() || FeatureThreads(aLock)) &&
+            sInstance->ThreadSelected(aInfo->Name()));
+  }
+
+  PS_GET(uint32_t, Generation)
+
+  PS_GET(PowerOfTwo32, Capacity)
+
+  PS_GET(Maybe<double>, Duration)
+
+  PS_GET(double, Interval)
+
+  PS_GET(uint32_t, Features)
+
+#define PS_GET_FEATURE(n_, str_, Name_, desc_)                \
+  static bool Feature##Name_(PSLockRef) {                     \
+    MOZ_ASSERT(sInstance);                                    \
+    return ProfilerFeature::Has##Name_(sInstance->mFeatures); \
+  }
+
+  BASE_PROFILER_FOR_EACH_FEATURE(PS_GET_FEATURE)
+
+#undef PS_GET_FEATURE
+
+  PS_GET(const Vector<std::string>&, Filters)
+
+  static void FulfillChunkRequests(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    sInstance->mProfileBufferChunkManager.FulfillChunkRequests();
+  }
+
+  static ProfileBuffer& Buffer(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    return sInstance->mProfileBuffer;
+  }
+
+  static const Vector<LiveProfiledThreadData>& LiveProfiledThreads(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    return sInstance->mLiveProfiledThreads;
+  }
+
+  // Returns an array containing (RegisteredThread*, ProfiledThreadData*) pairs
+  // for all threads that should be included in a profile, both for threads
+  // that are still registered, and for threads that have been unregistered but
+  // still have data in the buffer.
+  // For threads that have already been unregistered, the RegisteredThread
+  // pointer will be null.
+  // The returned array is sorted by thread register time.
+  // Do not hold on to the return value across thread registration or profiler
+  // restarts.
+  static Vector<std::pair<RegisteredThread*, ProfiledThreadData*>>
+  ProfiledThreads(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    Vector<std::pair<RegisteredThread*, ProfiledThreadData*>> array;
+    MOZ_RELEASE_ASSERT(
+        array.initCapacity(sInstance->mLiveProfiledThreads.length() +
+                           sInstance->mDeadProfiledThreads.length()));
+    for (auto& t : sInstance->mLiveProfiledThreads) {
+      MOZ_RELEASE_ASSERT(array.append(
+          std::make_pair(t.mRegisteredThread, t.mProfiledThreadData.get())));
+    }
+    for (auto& t : sInstance->mDeadProfiledThreads) {
+      MOZ_RELEASE_ASSERT(
+          array.append(std::make_pair((RegisteredThread*)nullptr, t.get())));
+    }
+
+    std::sort(array.begin(), array.end(),
+              [](const std::pair<RegisteredThread*, ProfiledThreadData*>& a,
+                 const std::pair<RegisteredThread*, ProfiledThreadData*>& b) {
+                return a.second->Info()->RegisterTime() <
+                       b.second->Info()->RegisterTime();
+              });
+    return array;
+  }
+
+  static Vector<RefPtr<PageInformation>> ProfiledPages(PSLockRef aLock) {
+    MOZ_ASSERT(sInstance);
+    Vector<RefPtr<PageInformation>> array;
+    for (auto& d : CorePS::RegisteredPages(aLock)) {
+      MOZ_RELEASE_ASSERT(array.append(d));
+    }
+    for (auto& d : sInstance->mDeadProfiledPages) {
+      MOZ_RELEASE_ASSERT(array.append(d));
+    }
+    // We don't need to sort the pages like threads since we won't show them
+    // as a list.
+    return array;
+  }
+
+  // Do a linear search through mLiveProfiledThreads to find the
+  // ProfiledThreadData object for a RegisteredThread.
+  static ProfiledThreadData* GetProfiledThreadData(
+      PSLockRef, RegisteredThread* aRegisteredThread) {
+    MOZ_ASSERT(sInstance);
+    for (const LiveProfiledThreadData& thread :
+         sInstance->mLiveProfiledThreads) {
+      if (thread.mRegisteredThread == aRegisteredThread) {
+        return thread.mProfiledThreadData.get();
+      }
+    }
+    return nullptr;
+  }
+
+  static ProfiledThreadData* AddLiveProfiledThread(
+      PSLockRef, RegisteredThread* aRegisteredThread,
+      UniquePtr<ProfiledThreadData>&& aProfiledThreadData) {
+    MOZ_ASSERT(sInstance);
+    MOZ_RELEASE_ASSERT(
+        sInstance->mLiveProfiledThreads.append(LiveProfiledThreadData{
+            aRegisteredThread, std::move(aProfiledThreadData)}));
+
+    // Return a weak pointer to the ProfiledThreadData object.
+    return sInstance->mLiveProfiledThreads.back().mProfiledThreadData.get();
+  }
+
+  static void UnregisterThread(PSLockRef aLockRef,
+                               RegisteredThread* aRegisteredThread) {
+    MOZ_ASSERT(sInstance);
+
+    DiscardExpiredDeadProfiledThreads(aLockRef);
+
+    // Find the right entry in the mLiveProfiledThreads array and remove the
+    // element, moving the ProfiledThreadData object for the thread into the
+    // mDeadProfiledThreads array.
+    // The thread's RegisteredThread object gets destroyed here.
+    for (size_t i = 0; i < sInstance->mLiveProfiledThreads.length(); i++) {
+      LiveProfiledThreadData& thread = sInstance->mLiveProfiledThreads[i];
+      if (thread.mRegisteredThread == aRegisteredThread) {
+        thread.mProfiledThreadData->NotifyUnregistered(
+            sInstance->mProfileBuffer.BufferRangeEnd());
+        MOZ_RELEASE_ASSERT(sInstance->mDeadProfiledThreads.append(
+            std::move(thread.mProfiledThreadData)));
+        sInstance->mLiveProfiledThreads.erase(
+            &sInstance->mLiveProfiledThreads[i]);
+        return;
+      }
+    }
+  }
+
+  PS_GET_AND_SET(bool, IsPaused)
+
+  // True if sampling is paused (though generic `SetIsPaused()` or specific
+  // `SetIsSamplingPaused()`).
+  static bool IsSamplingPaused(PSLockRef lock) {
+    MOZ_ASSERT(sInstance);
+    return IsPaused(lock) || sInstance->mIsSamplingPaused;
+  }
+
+  static void SetIsSamplingPaused(PSLockRef, bool aIsSamplingPaused) {
+    MOZ_ASSERT(sInstance);
+    sInstance->mIsSamplingPaused = aIsSamplingPaused;
+  }
+
+#if defined(GP_OS_linux) || defined(GP_OS_freebsd)
+  PS_GET_AND_SET(bool, WasSamplingPaused)
+#endif
+
+  static void DiscardExpiredDeadProfiledThreads(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    uint64_t bufferRangeStart = sInstance->mProfileBuffer.BufferRangeStart();
+    // Discard any dead threads that were unregistered before bufferRangeStart.
+    sInstance->mDeadProfiledThreads.eraseIf(
+        [bufferRangeStart](
+            const UniquePtr<ProfiledThreadData>& aProfiledThreadData) {
+          Maybe<uint64_t> bufferPosition =
+              aProfiledThreadData->BufferPositionWhenUnregistered();
+          MOZ_RELEASE_ASSERT(bufferPosition,
+                             "should have unregistered this thread");
+          return *bufferPosition < bufferRangeStart;
+        });
+  }
+
+  static void UnregisterPage(PSLockRef aLock,
+                             uint64_t aRegisteredInnerWindowID) {
+    MOZ_ASSERT(sInstance);
+    auto& registeredPages = CorePS::RegisteredPages(aLock);
+    for (size_t i = 0; i < registeredPages.length(); i++) {
+      RefPtr<PageInformation>& page = registeredPages[i];
+      if (page->InnerWindowID() == aRegisteredInnerWindowID) {
+        page->NotifyUnregistered(sInstance->mProfileBuffer.BufferRangeEnd());
+        MOZ_RELEASE_ASSERT(
+            sInstance->mDeadProfiledPages.append(std::move(page)));
+        registeredPages.erase(&registeredPages[i--]);
+      }
+    }
+  }
+
+  static void DiscardExpiredPages(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    uint64_t bufferRangeStart = sInstance->mProfileBuffer.BufferRangeStart();
+    // Discard any dead pages that were unregistered before
+    // bufferRangeStart.
+    sInstance->mDeadProfiledPages.eraseIf(
+        [bufferRangeStart](const RefPtr<PageInformation>& aProfiledPage) {
+          Maybe<uint64_t> bufferPosition =
+              aProfiledPage->BufferPositionWhenUnregistered();
+          MOZ_RELEASE_ASSERT(bufferPosition,
+                             "should have unregistered this page");
+          return *bufferPosition < bufferRangeStart;
+        });
+  }
+
+  static void ClearUnregisteredPages(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    sInstance->mDeadProfiledPages.clear();
+  }
+
+  static void ClearExpiredExitProfiles(PSLockRef) {
+    MOZ_ASSERT(sInstance);
+    uint64_t bufferRangeStart = sInstance->mProfileBuffer.BufferRangeStart();
+    // Discard exit profiles that were gathered before our buffer RangeStart.
+    sInstance->mExitProfiles.eraseIf(
+        [bufferRangeStart](const ExitProfile& aExitProfile) {
+          return aExitProfile.mBufferPositionAtGatherTime < bufferRangeStart;
+        });
+  }
+
+  static void AddExitProfile(PSLockRef aLock, const std::string& aExitProfile) {
+    MOZ_ASSERT(sInstance);
+
+    ClearExpiredExitProfiles(aLock);
+
+    MOZ_RELEASE_ASSERT(sInstance->mExitProfiles.append(
+        ExitProfile{aExitProfile, sInstance->mProfileBuffer.BufferRangeEnd()}));
+  }
+
+  static Vector<std::string> MoveExitProfiles(PSLockRef aLock) {
+    MOZ_ASSERT(sInstance);
+
+    ClearExpiredExitProfiles(aLock);
+
+    Vector<std::string> profiles;
+    MOZ_RELEASE_ASSERT(
+        profiles.initCapacity(sInstance->mExitProfiles.length()));
+    for (auto& profile : sInstance->mExitProfiles) {
+      MOZ_RELEASE_ASSERT(profiles.append(std::move(profile.mJSON)));
+    }
+    sInstance->mExitProfiles.clear();
+    return profiles;
+  }
+
+ private:
+  // The singleton instance.
+  static ActivePS* sInstance;
+
+  // We need to track activity generations. If we didn't we could have the
+  // following scenario.
+  //
+  // - profiler_stop() locks gPSMutex, de-instantiates ActivePS, unlocks
+  //   gPSMutex, deletes the SamplerThread (which does a join).
+  //
+  // - profiler_start() runs on a different thread, locks gPSMutex,
+  //   re-instantiates ActivePS, unlocks gPSMutex -- all before the join
+  //   completes.
+  //
+  // - SamplerThread::Run() locks gPSMutex, sees that ActivePS is instantiated,
+  //   and continues as if the start/stop pair didn't occur. Also
+  //   profiler_stop() is stuck, unable to finish.
+  //
+  // By checking ActivePS *and* the generation, we can avoid this scenario.
+  // sNextGeneration is used to track the next generation number; it is static
+  // because it must persist across different ActivePS instantiations.
+  const uint32_t mGeneration;
+  static uint32_t sNextGeneration;
+
+  // The maximum number of 8-byte entries in mProfileBuffer.
+  const PowerOfTwo32 mCapacity;
+
+  // The maximum duration of entries in mProfileBuffer, in seconds.
+  const Maybe<double> mDuration;
+
+  // The interval between samples, measured in milliseconds.
+  const double mInterval;
+
+  // The profile features that are enabled.
+  const uint32_t mFeatures;
+
+  // Substrings of names of threads we want to profile.
+  Vector<std::string> mFilters;
+
+  // The chunk manager used by `mProfileBuffer` below.
+  ProfileBufferChunkManagerWithLocalLimit mProfileBufferChunkManager;
+
+  // The buffer into which all samples are recorded.
+  ProfileBuffer mProfileBuffer;
+
+  // ProfiledThreadData objects for any threads that were profiled at any point
+  // during this run of the profiler:
+  //  - mLiveProfiledThreads contains all threads that are still registered, and
+  //  - mDeadProfiledThreads contains all threads that have already been
+  //    unregistered but for which there is still data in the profile buffer.
+  Vector<LiveProfiledThreadData> mLiveProfiledThreads;
+  Vector<UniquePtr<ProfiledThreadData>> mDeadProfiledThreads;
+
+  // Info on all the dead pages.
+  // Registered pages are being moved to this array after unregistration.
+  // We are keeping them in case we need them in the profile data.
+  // We are removing them when we ensure that we won't need them anymore.
+  Vector<RefPtr<PageInformation>> mDeadProfiledPages;
+
+  // The current sampler thread. This class is not responsible for destroying
+  // the SamplerThread object; the Destroy() method returns it so the caller
+  // can destroy it.
+  SamplerThread* const mSamplerThread;
+
+  // Is the profiler fully paused?
+  bool mIsPaused;
+
+  // Is the profiler periodic sampling paused?
+  bool mIsSamplingPaused;
+
+#if defined(GP_OS_linux) || defined(GP_OS_freebsd)
+  // Used to record whether the sampler was paused just before forking. False
+  // at all times except just before/after forking.
+  bool mWasSamplingPaused;
+#endif
+
+  struct ExitProfile {
+    std::string mJSON;
+    uint64_t mBufferPositionAtGatherTime;
+  };
+  Vector<ExitProfile> mExitProfiles;
+};
+
+ActivePS* ActivePS::sInstance = nullptr;
+uint32_t ActivePS::sNextGeneration = 0;
+
+#undef PS_GET
+#undef PS_GET_LOCKLESS
+#undef PS_GET_AND_SET
+
+Atomic<uint32_t, MemoryOrdering::Relaxed> RacyFeatures::sActiveAndFeatures(0);
+
+/* static */
+void RacyFeatures::SetActive(uint32_t aFeatures) {
+  sActiveAndFeatures = Active | aFeatures;
+}
+
+/* static */
+void RacyFeatures::SetInactive() { sActiveAndFeatures = 0; }
+
+/* static */
+bool RacyFeatures::IsActive() { return uint32_t(sActiveAndFeatures) & Active; }
+
+/* static */
+void RacyFeatures::SetPaused() { sActiveAndFeatures |= Paused; }
+
+/* static */
+void RacyFeatures::SetUnpaused() { sActiveAndFeatures &= ~Paused; }
+
+/* static */
+void RacyFeatures::SetSamplingPaused() { sActiveAndFeatures |= SamplingPaused; }
+
+/* static */
+void RacyFeatures::SetSamplingUnpaused() {
+  sActiveAndFeatures &= ~SamplingPaused;
+}
+
+/* static */
+bool RacyFeatures::IsActiveWithFeature(uint32_t aFeature) {
+  uint32_t af = sActiveAndFeatures;  // copy it first
+  return (af & Active) && (af & aFeature);
+}
+
+/* static */
+bool RacyFeatures::IsActiveAndUnpaused() {
+  uint32_t af = sActiveAndFeatures;  // copy it first
+  return (af & Active) && !(af & Paused);
+}
+
+/* static */
+bool RacyFeatures::IsActiveAndSamplingUnpaused() {
+  uint32_t af = sActiveAndFeatures;  // copy it first
+  return (af & Active) && !(af & (Paused | SamplingPaused));
+}
+
+// Each live thread has a RegisteredThread, and we store a reference to it in
+// TLS. This class encapsulates that TLS.
+class TLSRegisteredThread {
+ public:
+  static bool Init(PSLockRef) {
+    bool ok1 = sRegisteredThread.init();
+    bool ok2 = AutoProfilerLabel::sProfilingStack.init();
+    return ok1 && ok2;
+  }
+
+  // Get the entire RegisteredThread. Accesses are guarded by gPSMutex.
+  static class RegisteredThread* RegisteredThread(PSLockRef) {
+    return sRegisteredThread.get();
+  }
+
+  // Get only the RacyRegisteredThread. Accesses are not guarded by gPSMutex.
+  static class RacyRegisteredThread* RacyRegisteredThread() {
+    class RegisteredThread* registeredThread = sRegisteredThread.get();
+    return registeredThread ? &registeredThread->RacyRegisteredThread()
+                            : nullptr;
+  }
+
+  // Get only the ProfilingStack. Accesses are not guarded by gPSMutex.
+  // RacyRegisteredThread() can also be used to get the ProfilingStack, but that
+  // is marginally slower because it requires an extra pointer indirection.
+  static ProfilingStack* Stack() {
+    return AutoProfilerLabel::sProfilingStack.get();
+  }
+
+  static void SetRegisteredThread(PSLockRef,
+                                  class RegisteredThread* aRegisteredThread) {
+    sRegisteredThread.set(aRegisteredThread);
+    AutoProfilerLabel::sProfilingStack.set(
+        aRegisteredThread
+            ? &aRegisteredThread->RacyRegisteredThread().ProfilingStack()
+            : nullptr);
+  }
+
+ private:
+  // This is a non-owning reference to the RegisteredThread;
+  // CorePS::mRegisteredThreads is the owning reference. On thread
+  // deregistration, this reference is cleared and the RegisteredThread is
+  // destroyed.
+  static MOZ_THREAD_LOCAL(class RegisteredThread*) sRegisteredThread;
+};
+
+MOZ_THREAD_LOCAL(RegisteredThread*) TLSRegisteredThread::sRegisteredThread;
+
+/* static */
+ProfilingStack* AutoProfilerLabel::GetProfilingStack() {
+  return sProfilingStack.get();
+}
+
+// Although you can access a thread's ProfilingStack via
+// TLSRegisteredThread::sRegisteredThread, we also have a second TLS pointer
+// directly to the ProfilingStack. Here's why.
+//
+// - We need to be able to push to and pop from the ProfilingStack in
+//   AutoProfilerLabel.
+//
+// - The class functions are hot and must be defined in BaseProfiler.h so they
+//   can be inlined.
+//
+// - We don't want to expose TLSRegisteredThread (and RegisteredThread) in
+//   BaseProfiler.h.
+//
+// This second pointer isn't ideal, but does provide a way to satisfy those
+// constraints. TLSRegisteredThread is responsible for updating it.
+MOZ_THREAD_LOCAL(ProfilingStack*) AutoProfilerLabel::sProfilingStack;
+
+// The name of the main thread.
+static const char* const kMainThreadName = "GeckoMain";
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN sampling/unwinding code
+
+// The registers used for stack unwinding and a few other sampling purposes.
+// The ctor does nothing; users are responsible for filling in the fields.
+class Registers {
+ public:
+  Registers() : mPC{nullptr}, mSP{nullptr}, mFP{nullptr}, mLR{nullptr} {}
+
+#if defined(HAVE_NATIVE_UNWIND)
+  // Fills in mPC, mSP, mFP, mLR, and mContext for a synchronous sample.
+  void SyncPopulate();
+#endif
+
+  void Clear() { memset(this, 0, sizeof(*this)); }
+
+  // These fields are filled in by
+  // Sampler::SuspendAndSampleAndResumeThread() for periodic and backtrace
+  // samples, and by SyncPopulate() for synchronous samples.
+  Address mPC;  // Instruction pointer.
+  Address mSP;  // Stack pointer.
+  Address mFP;  // Frame pointer.
+  Address mLR;  // ARM link register.
+#if defined(GP_OS_linux) || defined(GP_OS_android) || defined(GP_OS_freebsd)
+  // This contains all the registers, which means it duplicates the four fields
+  // above. This is ok.
+  ucontext_t* mContext;  // The context from the signal handler.
+#endif
+};
+
+// Setting MAX_NATIVE_FRAMES too high risks the unwinder wasting a lot of time
+// looping on corrupted stacks.
+static const size_t MAX_NATIVE_FRAMES = 1024;
+
+struct NativeStack {
+  void* mPCs[MAX_NATIVE_FRAMES];
+  void* mSPs[MAX_NATIVE_FRAMES];
+  size_t mCount;  // Number of frames filled.
+
+  NativeStack() : mPCs(), mSPs(), mCount(0) {}
+};
+
+// Merges the profiling stack and native stack, outputting the details to
+// aCollector.
+static void MergeStacks(uint32_t aFeatures, bool aIsSynchronous,
+                        const RegisteredThread& aRegisteredThread,
+                        const Registers& aRegs, const NativeStack& aNativeStack,
+                        ProfilerStackCollector& aCollector) {
+  // WARNING: this function runs within the profiler's "critical section".
+  // WARNING: this function might be called while the profiler is inactive, and
+  //          cannot rely on ActivePS.
+
+  const ProfilingStack& profilingStack =
+      aRegisteredThread.RacyRegisteredThread().ProfilingStack();
+  const ProfilingStackFrame* profilingStackFrames = profilingStack.frames;
+  uint32_t profilingStackFrameCount = profilingStack.stackSize();
+
+  Maybe<uint64_t> samplePosInBuffer;
+  if (!aIsSynchronous) {
+    // aCollector.SamplePositionInBuffer() will return Nothing() when
+    // profiler_suspend_and_sample_thread is called from the background hang
+    // reporter.
+    samplePosInBuffer = aCollector.SamplePositionInBuffer();
+  }
+  // While the profiling stack array is ordered oldest-to-youngest, the JS and
+  // native arrays are ordered youngest-to-oldest. We must add frames to aInfo
+  // oldest-to-youngest. Thus, iterate over the profiling stack forwards and JS
+  // and native arrays backwards. Note: this means the terminating condition
+  // jsIndex and nativeIndex is being < 0.
+  uint32_t profilingStackIndex = 0;
+  int32_t nativeIndex = aNativeStack.mCount - 1;
+
+  uint8_t* lastLabelFrameStackAddr = nullptr;
+
+  // Iterate as long as there is at least one frame remaining.
+  while (profilingStackIndex != profilingStackFrameCount || nativeIndex >= 0) {
+    // There are 1 to 3 frames available. Find and add the oldest.
+    uint8_t* profilingStackAddr = nullptr;
+    uint8_t* nativeStackAddr = nullptr;
+
+    if (profilingStackIndex != profilingStackFrameCount) {
+      const ProfilingStackFrame& profilingStackFrame =
+          profilingStackFrames[profilingStackIndex];
+
+      if (profilingStackFrame.isLabelFrame() ||
+          profilingStackFrame.isSpMarkerFrame()) {
+        lastLabelFrameStackAddr = (uint8_t*)profilingStackFrame.stackAddress();
+      }
+
+      // Skip any JS_OSR frames. Such frames are used when the JS interpreter
+      // enters a jit frame on a loop edge (via on-stack-replacement, or OSR).
+      // To avoid both the profiling stack frame and jit frame being recorded
+      // (and showing up twice), the interpreter marks the interpreter
+      // profiling stack frame as JS_OSR to ensure that it doesn't get counted.
+      if (profilingStackFrame.isOSRFrame()) {
+        profilingStackIndex++;
+        continue;
+      }
+
+      MOZ_ASSERT(lastLabelFrameStackAddr);
+      profilingStackAddr = lastLabelFrameStackAddr;
+    }
+
+    if (nativeIndex >= 0) {
+      nativeStackAddr = (uint8_t*)aNativeStack.mSPs[nativeIndex];
+    }
+
+    // If there's a native stack frame which has the same SP as a profiling
+    // stack frame, pretend we didn't see the native stack frame.  Ditto for a
+    // native stack frame which has the same SP as a JS stack frame.  In effect
+    // this means profiling stack frames or JS frames trump conflicting native
+    // frames.
+    if (nativeStackAddr && (profilingStackAddr == nativeStackAddr)) {
+      nativeStackAddr = nullptr;
+      nativeIndex--;
+      MOZ_ASSERT(profilingStackAddr);
+    }
+
+    // Sanity checks.
+    MOZ_ASSERT_IF(profilingStackAddr, profilingStackAddr != nativeStackAddr);
+    MOZ_ASSERT_IF(nativeStackAddr, nativeStackAddr != profilingStackAddr);
+
+    // Check to see if profiling stack frame is top-most.
+    if (profilingStackAddr > nativeStackAddr) {
+      MOZ_ASSERT(profilingStackIndex < profilingStackFrameCount);
+      const ProfilingStackFrame& profilingStackFrame =
+          profilingStackFrames[profilingStackIndex];
+
+      // Sp marker frames are just annotations and should not be recorded in
+      // the profile.
+      if (!profilingStackFrame.isSpMarkerFrame()) {
+        aCollector.CollectProfilingStackFrame(profilingStackFrame);
+      }
+      profilingStackIndex++;
+      continue;
+    }
+
+    // If we reach here, there must be a native stack frame and it must be the
+    // greatest frame.
+    if (nativeStackAddr) {
+      MOZ_ASSERT(nativeIndex >= 0);
+      void* addr = (void*)aNativeStack.mPCs[nativeIndex];
+      aCollector.CollectNativeLeafAddr(addr);
+    }
+    if (nativeIndex >= 0) {
+      nativeIndex--;
+    }
+  }
+}
+
+#if defined(GP_OS_windows) && defined(USE_MOZ_STACK_WALK)
+static HANDLE GetThreadHandle(PlatformData* aData);
+#endif
+
+#if defined(USE_FRAME_POINTER_STACK_WALK) || defined(USE_MOZ_STACK_WALK)
+static void StackWalkCallback(uint32_t aFrameNumber, void* aPC, void* aSP,
+                              void* aClosure) {
+  NativeStack* nativeStack = static_cast<NativeStack*>(aClosure);
+  MOZ_ASSERT(nativeStack->mCount < MAX_NATIVE_FRAMES);
+  nativeStack->mSPs[nativeStack->mCount] = aSP;
+  nativeStack->mPCs[nativeStack->mCount] = aPC;
+  nativeStack->mCount++;
+}
+#endif
+
+#if defined(USE_FRAME_POINTER_STACK_WALK)
+static void DoFramePointerBacktrace(PSLockRef aLock,
+                                    const RegisteredThread& aRegisteredThread,
+                                    const Registers& aRegs,
+                                    NativeStack& aNativeStack) {
+  // WARNING: this function runs within the profiler's "critical section".
+  // WARNING: this function might be called while the profiler is inactive, and
+  //          cannot rely on ActivePS.
+
+  // Start with the current function. We use 0 as the frame number here because
+  // the FramePointerStackWalk() call below will use 1..N. This is a bit weird
+  // but it doesn't matter because StackWalkCallback() doesn't use the frame
+  // number argument.
+  StackWalkCallback(/* frameNum */ 0, aRegs.mPC, aRegs.mSP, &aNativeStack);
+
+  uint32_t maxFrames = uint32_t(MAX_NATIVE_FRAMES - aNativeStack.mCount);
+
+  const void* stackEnd = aRegisteredThread.StackTop();
+  if (aRegs.mFP >= aRegs.mSP && aRegs.mFP <= stackEnd) {
+    FramePointerStackWalk(StackWalkCallback, /* skipFrames */ 0, maxFrames,
+                          &aNativeStack, reinterpret_cast<void**>(aRegs.mFP),
+                          const_cast<void*>(stackEnd));
+  }
+}
+#endif
+
+#if defined(USE_MOZ_STACK_WALK)
+static void DoMozStackWalkBacktrace(PSLockRef aLock,
+                                    const RegisteredThread& aRegisteredThread,
+                                    const Registers& aRegs,
+                                    NativeStack& aNativeStack) {
+  // WARNING: this function runs within the profiler's "critical section".
+  // WARNING: this function might be called while the profiler is inactive, and
+  //          cannot rely on ActivePS.
+
+  // Start with the current function. We use 0 as the frame number here because
+  // the MozStackWalkThread() call below will use 1..N. This is a bit weird but
+  // it doesn't matter because StackWalkCallback() doesn't use the frame number
+  // argument.
+  StackWalkCallback(/* frameNum */ 0, aRegs.mPC, aRegs.mSP, &aNativeStack);
+
+  uint32_t maxFrames = uint32_t(MAX_NATIVE_FRAMES - aNativeStack.mCount);
+
+  HANDLE thread = GetThreadHandle(aRegisteredThread.GetPlatformData());
+  MOZ_ASSERT(thread);
+  MozStackWalkThread(StackWalkCallback, /* skipFrames */ 0, maxFrames,
+                     &aNativeStack, thread, /* context */ nullptr);
+}
+#endif
+
+#ifdef USE_EHABI_STACKWALK
+static void DoEHABIBacktrace(PSLockRef aLock,
+                             const RegisteredThread& aRegisteredThread,
+                             const Registers& aRegs,
+                             NativeStack& aNativeStack) {
+  // WARNING: this function runs within the profiler's "critical section".
+  // WARNING: this function might be called while the profiler is inactive, and
+  //          cannot rely on ActivePS.
+
+  aNativeStack.mCount =
+      EHABIStackWalk(aRegs.mContext->uc_mcontext,
+                     const_cast<void*>(aRegisteredThread.StackTop()),
+                     aNativeStack.mSPs, aNativeStack.mPCs, MAX_NATIVE_FRAMES);
+}
+#endif
+
+#ifdef USE_LUL_STACKWALK
+
+// See the comment at the callsite for why this function is necessary.
+#  if defined(MOZ_HAVE_ASAN_BLACKLIST)
+MOZ_ASAN_BLACKLIST static void ASAN_memcpy(void* aDst, const void* aSrc,
+                                           size_t aLen) {
+  // The obvious thing to do here is call memcpy(). However, although
+  // ASAN_memcpy() is not instrumented by ASAN, memcpy() still is, and the
+  // false positive still manifests! So we must implement memcpy() ourselves
+  // within this function.
+  char* dst = static_cast<char*>(aDst);
+  const char* src = static_cast<const char*>(aSrc);
+
+  for (size_t i = 0; i < aLen; i++) {
+    dst[i] = src[i];
+  }
+}
+#  endif
+
+static void DoLULBacktrace(PSLockRef aLock,
+                           const RegisteredThread& aRegisteredThread,
+                           const Registers& aRegs, NativeStack& aNativeStack) {
+  // WARNING: this function runs within the profiler's "critical section".
+  // WARNING: this function might be called while the profiler is inactive, and
+  //          cannot rely on ActivePS.
+
+  const mcontext_t* mc = &aRegs.mContext->uc_mcontext;
+
+  lul::UnwindRegs startRegs;
+  memset(&startRegs, 0, sizeof(startRegs));
+
+#  if defined(GP_PLAT_amd64_linux) || defined(GP_PLAT_amd64_android)
+  startRegs.xip = lul::TaggedUWord(mc->gregs[REG_RIP]);
+  startRegs.xsp = lul::TaggedUWord(mc->gregs[REG_RSP]);
+  startRegs.xbp = lul::TaggedUWord(mc->gregs[REG_RBP]);
+#  elif defined(GP_PLAT_amd64_freebsd)
+  startRegs.xip = lul::TaggedUWord(mc->mc_rip);
+  startRegs.xsp = lul::TaggedUWord(mc->mc_rsp);
+  startRegs.xbp = lul::TaggedUWord(mc->mc_rbp);
+#  elif defined(GP_PLAT_arm_linux) || defined(GP_PLAT_arm_android)
+  startRegs.r15 = lul::TaggedUWord(mc->arm_pc);
+  startRegs.r14 = lul::TaggedUWord(mc->arm_lr);
+  startRegs.r13 = lul::TaggedUWord(mc->arm_sp);
+  startRegs.r12 = lul::TaggedUWord(mc->arm_ip);
+  startRegs.r11 = lul::TaggedUWord(mc->arm_fp);
+  startRegs.r7 = lul::TaggedUWord(mc->arm_r7);
+#  elif defined(GP_PLAT_arm64_linux) || defined(GP_PLAT_arm64_android)
+  startRegs.pc = lul::TaggedUWord(mc->pc);
+  startRegs.x29 = lul::TaggedUWord(mc->regs[29]);
+  startRegs.x30 = lul::TaggedUWord(mc->regs[30]);
+  startRegs.sp = lul::TaggedUWord(mc->sp);
+#  elif defined(GP_PLAT_arm64_freebsd)
+  startRegs.pc = lul::TaggedUWord(mc->mc_gpregs.gp_elr);
+  startRegs.x29 = lul::TaggedUWord(mc->mc_gpregs.gp_x[29]);
+  startRegs.x30 = lul::TaggedUWord(mc->mc_gpregs.gp_lr);
+  startRegs.sp = lul::TaggedUWord(mc->mc_gpregs.gp_sp);
+#  elif defined(GP_PLAT_x86_linux) || defined(GP_PLAT_x86_android)
+  startRegs.xip = lul::TaggedUWord(mc->gregs[REG_EIP]);
+  startRegs.xsp = lul::TaggedUWord(mc->gregs[REG_ESP]);
+  startRegs.xbp = lul::TaggedUWord(mc->gregs[REG_EBP]);
+#  elif defined(GP_PLAT_mips64_linux)
+  startRegs.pc = lul::TaggedUWord(mc->pc);
+  startRegs.sp = lul::TaggedUWord(mc->gregs[29]);
+  startRegs.fp = lul::TaggedUWord(mc->gregs[30]);
+#  else
+#    error "Unknown plat"
+#  endif
+
+  // Copy up to N_STACK_BYTES from rsp-REDZONE upwards, but not going past the
+  // stack's registered top point.  Do some basic sanity checks too.  This
+  // assumes that the TaggedUWord holding the stack pointer value is valid, but
+  // it should be, since it was constructed that way in the code just above.
+
+  // We could construct |stackImg| so that LUL reads directly from the stack in
+  // question, rather than from a copy of it.  That would reduce overhead and
+  // space use a bit.  However, it gives a problem with dynamic analysis tools
+  // (ASan, TSan, Valgrind) which is that such tools will report invalid or
+  // racing memory accesses, and such accesses will be reported deep inside LUL.
+  // By taking a copy here, we can either sanitise the copy (for Valgrind) or
+  // copy it using an unchecked memcpy (for ASan, TSan).  That way we don't have
+  // to try and suppress errors inside LUL.
+  //
+  // N_STACK_BYTES is set to 160KB.  This is big enough to hold all stacks
+  // observed in some minutes of testing, whilst keeping the size of this
+  // function (DoNativeBacktrace)'s frame reasonable.  Most stacks observed in
+  // practice are small, 4KB or less, and so the copy costs are insignificant
+  // compared to other profiler overhead.
+  //
+  // |stackImg| is allocated on this (the sampling thread's) stack.  That
+  // implies that the frame for this function is at least N_STACK_BYTES large.
+  // In general it would be considered unacceptable to have such a large frame
+  // on a stack, but it only exists for the unwinder thread, and so is not
+  // expected to be a problem.  Allocating it on the heap is troublesome because
+  // this function runs whilst the sampled thread is suspended, so any heap
+  // allocation risks deadlock.  Allocating it as a global variable is not
+  // thread safe, which would be a problem if we ever allow multiple sampler
+  // threads.  Hence allocating it on the stack seems to be the least-worst
+  // option.
+
+  lul::StackImage stackImg;
+
+  {
+#  if defined(GP_PLAT_amd64_linux) || defined(GP_PLAT_amd64_android) || \
+      defined(GP_PLAT_amd64_freebsd)
+    uintptr_t rEDZONE_SIZE = 128;
+    uintptr_t start = startRegs.xsp.Value() - rEDZONE_SIZE;
+#  elif defined(GP_PLAT_arm_linux) || defined(GP_PLAT_arm_android)
+    uintptr_t rEDZONE_SIZE = 0;
+    uintptr_t start = startRegs.r13.Value() - rEDZONE_SIZE;
+#  elif defined(GP_PLAT_arm64_linux) || defined(GP_PLAT_arm64_android) || \
+      defined(GP_PLAT_arm64_freebsd)
+    uintptr_t rEDZONE_SIZE = 0;
+    uintptr_t start = startRegs.sp.Value() - rEDZONE_SIZE;
+#  elif defined(GP_PLAT_x86_linux) || defined(GP_PLAT_x86_android)
+    uintptr_t rEDZONE_SIZE = 0;
+    uintptr_t start = startRegs.xsp.Value() - rEDZONE_SIZE;
+#  elif defined(GP_PLAT_mips64_linux)
+    uintptr_t rEDZONE_SIZE = 0;
+    uintptr_t start = startRegs.sp.Value() - rEDZONE_SIZE;
+#  else
+#    error "Unknown plat"
+#  endif
+    uintptr_t end = reinterpret_cast<uintptr_t>(aRegisteredThread.StackTop());
+    uintptr_t ws = sizeof(void*);
+    start &= ~(ws - 1);
+    end &= ~(ws - 1);
+    uintptr_t nToCopy = 0;
+    if (start < end) {
+      nToCopy = end - start;
+      if (nToCopy > lul::N_STACK_BYTES) nToCopy = lul::N_STACK_BYTES;
+    }
+    MOZ_ASSERT(nToCopy <= lul::N_STACK_BYTES);
+    stackImg.mLen = nToCopy;
+    stackImg.mStartAvma = start;
+    if (nToCopy > 0) {
+      // If this is a vanilla memcpy(), ASAN makes the following complaint:
+      //
+      //   ERROR: AddressSanitizer: stack-buffer-underflow ...
+      //   ...
+      //   HINT: this may be a false positive if your program uses some custom
+      //   stack unwind mechanism or swapcontext
+      //
+      // This code is very much a custom stack unwind mechanism! So we use an
+      // alternative memcpy() implementation that is ignored by ASAN.
+#  if defined(MOZ_HAVE_ASAN_BLACKLIST)
+      ASAN_memcpy(&stackImg.mContents[0], (void*)start, nToCopy);
+#  else
+      memcpy(&stackImg.mContents[0], (void*)start, nToCopy);
+#  endif
+      (void)VALGRIND_MAKE_MEM_DEFINED(&stackImg.mContents[0], nToCopy);
+    }
+  }
+
+  size_t framePointerFramesAcquired = 0;
+  lul::LUL* lul = CorePS::Lul(aLock);
+  lul->Unwind(reinterpret_cast<uintptr_t*>(aNativeStack.mPCs),
+              reinterpret_cast<uintptr_t*>(aNativeStack.mSPs),
+              &aNativeStack.mCount, &framePointerFramesAcquired,
+              MAX_NATIVE_FRAMES, &startRegs, &stackImg);
+
+  // Update stats in the LUL stats object.  Unfortunately this requires
+  // three global memory operations.
+  lul->mStats.mContext += 1;
+  lul->mStats.mCFI += aNativeStack.mCount - 1 - framePointerFramesAcquired;
+  lul->mStats.mFP += framePointerFramesAcquired;
+}
+
+#endif
+
+#ifdef HAVE_NATIVE_UNWIND
+static void DoNativeBacktrace(PSLockRef aLock,
+                              const RegisteredThread& aRegisteredThread,
+                              const Registers& aRegs,
+                              NativeStack& aNativeStack) {
+  // This method determines which stackwalker is used for periodic and
+  // synchronous samples. (Backtrace samples are treated differently, see
+  // profiler_suspend_and_sample_thread() for details). The only part of the
+  // ordering that matters is that LUL must precede FRAME_POINTER, because on
+  // Linux they can both be present.
+#  if defined(USE_LUL_STACKWALK)
+  DoLULBacktrace(aLock, aRegisteredThread, aRegs, aNativeStack);
+#  elif defined(USE_EHABI_STACKWALK)
+  DoEHABIBacktrace(aLock, aRegisteredThread, aRegs, aNativeStack);
+#  elif defined(USE_FRAME_POINTER_STACK_WALK)
+  DoFramePointerBacktrace(aLock, aRegisteredThread, aRegs, aNativeStack);
+#  elif defined(USE_MOZ_STACK_WALK)
+  DoMozStackWalkBacktrace(aLock, aRegisteredThread, aRegs, aNativeStack);
+#  else
+#    error "Invalid configuration"
+#  endif
+}
+#endif
+
+// Writes some components shared by periodic and synchronous profiles to
+// ActivePS's ProfileBuffer. (This should only be called from DoSyncSample()
+// and DoPeriodicSample().)
+//
+// The grammar for entry sequences is in a comment above
+// ProfileBuffer::StreamSamplesToJSON.
+static inline void DoSharedSample(PSLockRef aLock, bool aIsSynchronous,
+                                  RegisteredThread& aRegisteredThread,
+                                  const Registers& aRegs, uint64_t aSamplePos,
+                                  ProfileBuffer& aBuffer) {
+  // WARNING: this function runs within the profiler's "critical section".
+
+  MOZ_ASSERT(!aBuffer.IsThreadSafe(),
+             "Mutexes cannot be used inside this critical section");
+
+  MOZ_RELEASE_ASSERT(ActivePS::Exists(aLock));
+
+  ProfileBufferCollector collector(aBuffer, aSamplePos);
+  NativeStack nativeStack;
+#if defined(HAVE_NATIVE_UNWIND)
+  if (ActivePS::FeatureStackWalk(aLock)) {
+    DoNativeBacktrace(aLock, aRegisteredThread, aRegs, nativeStack);
+
+    MergeStacks(ActivePS::Features(aLock), aIsSynchronous, aRegisteredThread,
+                aRegs, nativeStack, collector);
+  } else
+#endif
+  {
+    MergeStacks(ActivePS::Features(aLock), aIsSynchronous, aRegisteredThread,
+                aRegs, nativeStack, collector);
+
+    // We can't walk the whole native stack, but we can record the top frame.
+    if (ActivePS::FeatureLeaf(aLock)) {
+      aBuffer.AddEntry(ProfileBufferEntry::NativeLeafAddr((void*)aRegs.mPC));
+    }
+  }
+}
+
+// Writes the components of a synchronous sample to the given ProfileBuffer.
+static void DoSyncSample(PSLockRef aLock, RegisteredThread& aRegisteredThread,
+                         const TimeStamp& aNow, const Registers& aRegs,
+                         ProfileBuffer& aBuffer) {
+  // WARNING: this function runs within the profiler's "critical section".
+
+  uint64_t samplePos =
+      aBuffer.AddThreadIdEntry(aRegisteredThread.Info()->ThreadId());
+
+  TimeDuration delta = aNow - CorePS::ProcessStartTime();
+  aBuffer.AddEntry(ProfileBufferEntry::Time(delta.ToMilliseconds()));
+
+  DoSharedSample(aLock, /* aIsSynchronous = */ true, aRegisteredThread, aRegs,
+                 samplePos, aBuffer);
+}
+
+// Writes the components of a periodic sample to ActivePS's ProfileBuffer.
+// The ThreadId entry is already written in the main ProfileBuffer, its location
+// is `aSamplePos`, we can write the rest to `aBuffer` (which may be different).
+static void DoPeriodicSample(PSLockRef aLock,
+                             RegisteredThread& aRegisteredThread,
+                             ProfiledThreadData& aProfiledThreadData,
+                             const Registers& aRegs, uint64_t aSamplePos,
+                             ProfileBuffer& aBuffer) {
+  // WARNING: this function runs within the profiler's "critical section".
+
+  DoSharedSample(aLock, /* aIsSynchronous = */ false, aRegisteredThread, aRegs,
+                 aSamplePos, aBuffer);
+}
+
+// END sampling/unwinding code
+////////////////////////////////////////////////////////////////////////
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN saving/streaming code
+
+const static uint64_t kJS_MAX_SAFE_UINTEGER = +9007199254740991ULL;
+
+static int64_t SafeJSInteger(uint64_t aValue) {
+  return aValue <= kJS_MAX_SAFE_UINTEGER ? int64_t(aValue) : -1;
+}
+
+static void AddSharedLibraryInfoToStream(JSONWriter& aWriter,
+                                         const SharedLibrary& aLib) {
+  aWriter.StartObjectElement();
+  aWriter.IntProperty("start", SafeJSInteger(aLib.GetStart()));
+  aWriter.IntProperty("end", SafeJSInteger(aLib.GetEnd()));
+  aWriter.IntProperty("offset", SafeJSInteger(aLib.GetOffset()));
+  aWriter.StringProperty("name", aLib.GetModuleName());
+  aWriter.StringProperty("path", aLib.GetModulePath());
+  aWriter.StringProperty("debugName", aLib.GetDebugName());
+  aWriter.StringProperty("debugPath", aLib.GetDebugPath());
+  aWriter.StringProperty("breakpadId", aLib.GetBreakpadId());
+  aWriter.StringProperty("arch", aLib.GetArch());
+  aWriter.EndObject();
+}
+
+void AppendSharedLibraries(JSONWriter& aWriter) {
+  SharedLibraryInfo info = SharedLibraryInfo::GetInfoForSelf();
+  info.SortByAddress();
+  for (size_t i = 0; i < info.GetSize(); i++) {
+    AddSharedLibraryInfoToStream(aWriter, info.GetEntry(i));
+  }
+}
+
+static void StreamCategories(SpliceableJSONWriter& aWriter) {
+  // Same order as ProfilingCategory. Format:
+  // [
+  //   {
+  //     name: "Idle",
+  //     color: "transparent",
+  //     subcategories: ["Other"],
+  //   },
+  //   {
+  //     name: "Other",
+  //     color: "grey",
+  //     subcategories: [
+  //       "JSM loading",
+  //       "Subprocess launching",
+  //       "DLL loading"
+  //     ]
+  //   },
+  //   ...
+  // ]
+
+#define CATEGORY_JSON_BEGIN_CATEGORY(name, labelAsString, color) \
+  aWriter.Start();                                               \
+  aWriter.StringProperty("name", labelAsString);                 \
+  aWriter.StringProperty("color", color);                        \
+  aWriter.StartArrayProperty("subcategories");
+#define CATEGORY_JSON_SUBCATEGORY(supercategory, name, labelAsString) \
+  aWriter.StringElement(labelAsString);
+#define CATEGORY_JSON_END_CATEGORY \
+  aWriter.EndArray();              \
+  aWriter.EndObject();
+
+  MOZ_PROFILING_CATEGORY_LIST(CATEGORY_JSON_BEGIN_CATEGORY,
+                              CATEGORY_JSON_SUBCATEGORY,
+                              CATEGORY_JSON_END_CATEGORY)
+
+#undef CATEGORY_JSON_BEGIN_CATEGORY
+#undef CATEGORY_JSON_SUBCATEGORY
+#undef CATEGORY_JSON_END_CATEGORY
+}
+
+static void StreamMarkerSchema(SpliceableJSONWriter& aWriter) {
+  // Get an array view with all registered marker-type-specific functions.
+  Span<const base_profiler_markers_detail::Streaming::MarkerTypeFunctions>
+      markerTypeFunctionsArray =
+          base_profiler_markers_detail::Streaming::MarkerTypeFunctionsArray();
+  // List of streamed marker names, this is used to spot duplicates.
+  std::set<std::string> names;
+  // Stream the display schema for each different one. (Duplications may come
+  // from the same code potentially living in different libraries.)
+  for (const auto& markerTypeFunctions : markerTypeFunctionsArray) {
+    auto name = markerTypeFunctions.mMarkerTypeNameFunction();
+    // std::set.insert(T&&) returns a pair, its `second` is true if the element
+    // was actually inserted (i.e., it was not there yet.)
+    const bool didInsert =
+        names.insert(std::string(name.data(), name.size())).second;
+    if (didInsert) {
+      markerTypeFunctions.mMarkerSchemaFunction().Stream(aWriter, name);
+    }
+  }
+}
+
+static int64_t MicrosecondsSince1970();
+
+static void StreamMetaJSCustomObject(PSLockRef aLock,
+                                     SpliceableJSONWriter& aWriter,
+                                     bool aIsShuttingDown) {
+  MOZ_RELEASE_ASSERT(CorePS::Exists() && ActivePS::Exists(aLock));
+
+  aWriter.IntProperty("version", 19);
+
+  // The "startTime" field holds the number of milliseconds since midnight
+  // January 1, 1970 GMT. This grotty code computes (Now - (Now -
+  // ProcessStartTime)) to convert CorePS::ProcessStartTime() into that form.
+  TimeDuration delta = TimeStamp::NowUnfuzzed() - CorePS::ProcessStartTime();
+  aWriter.DoubleProperty(
+      "startTime", MicrosecondsSince1970() / 1000.0 - delta.ToMilliseconds());
+
+  // Write the shutdownTime field. Unlike startTime, shutdownTime is not an
+  // absolute time stamp: It's relative to startTime. This is consistent with
+  // all other (non-"startTime") times anywhere in the profile JSON.
+  if (aIsShuttingDown) {
+    aWriter.DoubleProperty("shutdownTime", profiler_time());
+  } else {
+    aWriter.NullProperty("shutdownTime");
+  }
+
+  aWriter.StartArrayProperty("categories");
+  StreamCategories(aWriter);
+  aWriter.EndArray();
+
+  aWriter.StartArrayProperty("markerSchema");
+  StreamMarkerSchema(aWriter);
+  aWriter.EndArray();
+
+  if (!profiler_is_main_thread()) {
+    // Leave the rest of the properties out if we're not on the main thread.
+    // At the moment, the only case in which this function is called on a
+    // background thread is if we're in a content process and are going to
+    // send this profile to the parent process. In that case, the parent
+    // process profile's "meta" object already has the rest of the properties,
+    // and the parent process profile is dumped on that process's main thread.
+    return;
+  }
+
+  aWriter.DoubleProperty("interval", ActivePS::Interval(aLock));
+  aWriter.IntProperty("stackwalk", ActivePS::FeatureStackWalk(aLock));
+
+#ifdef DEBUG
+  aWriter.IntProperty("debug", 1);
+#else
+  aWriter.IntProperty("debug", 0);
+#endif
+
+  aWriter.IntProperty("gcpoison", 0);
+
+  aWriter.IntProperty("asyncstack", 0);
+
+  aWriter.IntProperty("processType", 0);
+}
+
+static void StreamPages(PSLockRef aLock, SpliceableJSONWriter& aWriter) {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+  ActivePS::DiscardExpiredPages(aLock);
+  for (const auto& page : ActivePS::ProfiledPages(aLock)) {
+    page->StreamJSON(aWriter);
+  }
+}
+
+static void locked_profiler_stream_json_for_this_process(
+    PSLockRef aLock, SpliceableJSONWriter& aWriter, double aSinceTime,
+    bool aIsShuttingDown, bool aOnlyThreads = false) {
+  LOG("locked_profiler_stream_json_for_this_process");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists() && ActivePS::Exists(aLock));
+
+  AUTO_PROFILER_STATS(base_locked_profiler_stream_json_for_this_process);
+
+  const double collectionStartMs = profiler_time();
+
+  ProfileBuffer& buffer = ActivePS::Buffer(aLock);
+
+  // If there is a set "Window length", discard older data.
+  Maybe<double> durationS = ActivePS::Duration(aLock);
+  if (durationS.isSome()) {
+    const double durationStartMs = collectionStartMs - *durationS * 1000;
+    buffer.DiscardSamplesBeforeTime(durationStartMs);
+  }
+
+  if (!aOnlyThreads) {
+    // Put shared library info
+    aWriter.StartArrayProperty("libs");
+    AppendSharedLibraries(aWriter);
+    aWriter.EndArray();
+
+    // Put meta data
+    aWriter.StartObjectProperty("meta");
+    { StreamMetaJSCustomObject(aLock, aWriter, aIsShuttingDown); }
+    aWriter.EndObject();
+
+    // Put page data
+    aWriter.StartArrayProperty("pages");
+    { StreamPages(aLock, aWriter); }
+    aWriter.EndArray();
+
+    buffer.StreamProfilerOverheadToJSON(aWriter, CorePS::ProcessStartTime(),
+                                        aSinceTime);
+    buffer.StreamCountersToJSON(aWriter, CorePS::ProcessStartTime(),
+                                aSinceTime);
+
+    // Lists the samples for each thread profile
+    aWriter.StartArrayProperty("threads");
+  }
+
+  // if aOnlyThreads is true, the only output will be the threads array items.
+  {
+    ActivePS::DiscardExpiredDeadProfiledThreads(aLock);
+    Vector<std::pair<RegisteredThread*, ProfiledThreadData*>> threads =
+        ActivePS::ProfiledThreads(aLock);
+    for (auto& thread : threads) {
+      ProfiledThreadData* profiledThreadData = thread.second;
+      profiledThreadData->StreamJSON(
+          buffer, aWriter, CorePS::ProcessName(aLock), CorePS::ETLDplus1(aLock),
+          CorePS::ProcessStartTime(), aSinceTime);
+    }
+  }
+
+  if (!aOnlyThreads) {
+    aWriter.EndArray();
+
+    aWriter.StartArrayProperty("pausedRanges");
+    { buffer.StreamPausedRangesToJSON(aWriter, aSinceTime); }
+    aWriter.EndArray();
+  }
+
+  const double collectionEndMs = profiler_time();
+
+  // Record timestamps for the collection into the buffer, so that consumers
+  // know why we didn't collect any samples for its duration.
+  // We put these entries into the buffer after we've collected the profile,
+  // so they'll be visible for the *next* profile collection (if they haven't
+  // been overwritten due to buffer wraparound by then).
+  buffer.AddEntry(ProfileBufferEntry::CollectionStart(collectionStartMs));
+  buffer.AddEntry(ProfileBufferEntry::CollectionEnd(collectionEndMs));
+}
+
+bool profiler_stream_json_for_this_process(SpliceableJSONWriter& aWriter,
+                                           double aSinceTime,
+                                           bool aIsShuttingDown,
+                                           bool aOnlyThreads) {
+  LOG("profiler_stream_json_for_this_process");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (!ActivePS::Exists(lock)) {
+    return false;
+  }
+
+  locked_profiler_stream_json_for_this_process(lock, aWriter, aSinceTime,
+                                               aIsShuttingDown, aOnlyThreads);
+  return true;
+}
+
+// END saving/streaming code
+////////////////////////////////////////////////////////////////////////
+
+static char FeatureCategory(uint32_t aFeature) {
+  if (aFeature & DefaultFeatures()) {
+    if (aFeature & AvailableFeatures()) {
+      return 'D';
+    }
+    return 'd';
+  }
+
+  if (aFeature & StartupExtraDefaultFeatures()) {
+    if (aFeature & AvailableFeatures()) {
+      return 'S';
+    }
+    return 's';
+  }
+
+  if (aFeature & AvailableFeatures()) {
+    return '-';
+  }
+  return 'x';
+}
+
+static void PrintUsageThenExit(int aExitCode) {
+  PrintToConsole(
+      "\n"
+      "Profiler environment variable usage:\n"
+      "\n"
+      "  MOZ_BASE_PROFILER_HELP\n"
+      "  If set to any value, prints this message.\n"
+      "  (Only BaseProfiler features are known here; Use MOZ_PROFILER_HELP\n"
+      "  for Gecko Profiler help, with more features).\n"
+      "\n"
+      "  MOZ_BASE_PROFILER_{,DEBUG_,VERBOSE}LOGGING\n"
+      "  Enables BaseProfiler logging to stdout. The levels of logging\n"
+      "  available are MOZ_BASE_PROFILER_LOGGING' (least verbose),\n"
+      "  '..._DEBUG_LOGGING', '..._VERBOSE_LOGGING' (most verbose)\n"
+      "\n"
+      "  MOZ_PROFILER_STARTUP\n"
+      "  If set to any value other than '' or '0'/'N'/'n', starts the\n"
+      "  profiler immediately on start-up.\n"
+      "  Useful if you want profile code that runs very early.\n"
+      "\n"
+      "  MOZ_PROFILER_STARTUP_ENTRIES=<%u..%u>\n"
+      "  If MOZ_PROFILER_STARTUP is set, specifies the number of entries\n"
+      "  per process in the profiler's circular buffer when the profiler is\n"
+      "  first started.\n"
+      "  If unset, the platform default is used:\n"
+      "  %u entries per process, or %u when MOZ_PROFILER_STARTUP is set.\n"
+      "  (%u bytes per entry -> %u or %u total bytes per process)\n"
+      "\n"
+      "  MOZ_PROFILER_STARTUP_DURATION=<1..>\n"
+      "  If MOZ_PROFILER_STARTUP is set, specifies the maximum life time\n"
+      "  of entries in the the profiler's circular buffer when the profiler\n"
+      "  is first started, in seconds.\n"
+      "  If unset, the life time of the entries will only be restricted by\n"
+      "  MOZ_PROFILER_STARTUP_ENTRIES (or its default value), and no\n"
+      "  additional time duration restriction will be applied.\n"
+      "\n"
+      "  MOZ_PROFILER_STARTUP_INTERVAL=<1..1000>\n"
+      "  If MOZ_PROFILER_STARTUP is set, specifies the sample interval,\n"
+      "  measured in milliseconds, when the profiler is first started.\n"
+      "  If unset, the platform default is used.\n"
+      "\n"
+      "  MOZ_PROFILER_STARTUP_FEATURES_BITFIELD=<Number>\n"
+      "  If MOZ_PROFILER_STARTUP is set, specifies the profiling\n"
+      "  features, as the integer value of the features bitfield.\n"
+      "  If unset, the value from MOZ_PROFILER_STARTUP_FEATURES is used.\n"
+      "\n"
+      "  MOZ_PROFILER_STARTUP_FEATURES=<Features>\n"
+      "  If MOZ_PROFILER_STARTUP is set, specifies the profiling\n"
+      "  features, as a comma-separated list of strings.\n"
+      "  Ignored if MOZ_PROFILER_STARTUP_FEATURES_BITFIELD is set.\n"
+      "  If unset, the platform default is used.\n"
+      "\n"
+      "    Features: (x=unavailable, D/d=default/unavailable,\n"
+      "               S/s=MOZ_PROFILER_STARTUP extra "
+      "default/unavailable)\n",
+      unsigned(ActivePS::scMinimumBufferEntries),
+      unsigned(ActivePS::scMaximumBufferEntries),
+      unsigned(BASE_PROFILER_DEFAULT_ENTRIES.Value()),
+      unsigned(BASE_PROFILER_DEFAULT_STARTUP_ENTRIES.Value()),
+      unsigned(scBytesPerEntry),
+      unsigned(BASE_PROFILER_DEFAULT_ENTRIES.Value() * scBytesPerEntry),
+      unsigned(BASE_PROFILER_DEFAULT_STARTUP_ENTRIES.Value() *
+               scBytesPerEntry));
+
+#define PRINT_FEATURE(n_, str_, Name_, desc_)             \
+  PrintToConsole("    %c %7u: \"%s\" (%s)\n",             \
+                 FeatureCategory(ProfilerFeature::Name_), \
+                 ProfilerFeature::Name_, str_, desc_);
+
+  BASE_PROFILER_FOR_EACH_FEATURE(PRINT_FEATURE)
+
+#undef PRINT_FEATURE
+
+  PrintToConsole(
+      "    -        \"default\" (All above D+S defaults)\n"
+      "\n"
+      "  MOZ_PROFILER_STARTUP_FILTERS=<Filters>\n"
+      "  If MOZ_PROFILER_STARTUP is set, specifies the thread filters, as "
+      "a\n"
+      "  comma-separated list of strings. A given thread will be sampled if\n"
+      "  any of the filters is a case-insensitive substring of the thread\n"
+      "  name. If unset, a default is used.\n"
+      "\n"
+      "  MOZ_PROFILER_SHUTDOWN\n"
+      "  If set, the profiler saves a profile to the named file on shutdown.\n"
+      "\n"
+      "  MOZ_PROFILER_SYMBOLICATE\n"
+      "  If set, the profiler will pre-symbolicate profiles.\n"
+      "  *Note* This will add a significant pause when gathering data, and\n"
+      "  is intended mainly for local development.\n"
+      "\n"
+      "  MOZ_PROFILER_LUL_TEST\n"
+      "  If set to any value, runs LUL unit tests at startup.\n"
+      "\n"
+      "  This platform %s native unwinding.\n"
+      "\n",
+#if defined(HAVE_NATIVE_UNWIND)
+      "supports"
+#else
+      "does not support"
+#endif
+  );
+
+  exit(aExitCode);
+}
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN Sampler
+
+#if defined(GP_OS_linux) || defined(GP_OS_android)
+struct SigHandlerCoordinator;
+#endif
+
+// Sampler performs setup and teardown of the state required to sample with the
+// profiler. Sampler may exist when ActivePS is not present.
+//
+// SuspendAndSampleAndResumeThread must only be called from a single thread,
+// and must not sample the thread it is being called from. A separate Sampler
+// instance must be used for each thread which wants to capture samples.
+
+// WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+//
+// With the exception of SamplerThread, all Sampler objects must be Disable-d
+// before releasing the lock which was used to create them. This avoids races
+// on linux with the SIGPROF signal handler.
+
+class Sampler {
+ public:
+  // Sets up the profiler such that it can begin sampling.
+  explicit Sampler(PSLockRef aLock);
+
+  // Disable the sampler, restoring it to its previous state. This must be
+  // called once, and only once, before the Sampler is destroyed.
+  void Disable(PSLockRef aLock);
+
+  // This method suspends and resumes the samplee thread. It calls the passed-in
+  // function-like object aProcessRegs (passing it a populated |const
+  // Registers&| arg) while the samplee thread is suspended.
+  //
+  // Func must be a function-like object of type `void()`.
+  template <typename Func>
+  void SuspendAndSampleAndResumeThread(
+      PSLockRef aLock, const RegisteredThread& aRegisteredThread,
+      const TimeStamp& aNow, const Func& aProcessRegs);
+
+ private:
+#if defined(GP_OS_linux) || defined(GP_OS_android) || defined(GP_OS_freebsd)
+  // Used to restore the SIGPROF handler when ours is removed.
+  struct sigaction mOldSigprofHandler;
+
+  // This process' ID. Needed as an argument for tgkill in
+  // SuspendAndSampleAndResumeThread.
+  int mMyPid;
+
+  // The sampler thread's ID.  Used to assert that it is not sampling itself,
+  // which would lead to deadlock.
+  int mSamplerTid;
+
+ public:
+  // This is the one-and-only variable used to communicate between the sampler
+  // thread and the samplee thread's signal handler. It's static because the
+  // samplee thread's signal handler is static.
+  static struct SigHandlerCoordinator* sSigHandlerCoordinator;
+#endif
+};
+
+// END Sampler
+////////////////////////////////////////////////////////////////////////
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN SamplerThread
+
+// The sampler thread controls sampling and runs whenever the profiler is
+// active. It periodically runs through all registered threads, finds those
+// that should be sampled, then pauses and samples them.
+
+class SamplerThread {
+ public:
+  // Creates a sampler thread, but doesn't start it.
+  SamplerThread(PSLockRef aLock, uint32_t aActivityGeneration,
+                double aIntervalMilliseconds);
+  ~SamplerThread();
+
+  // This runs on (is!) the sampler thread.
+  void Run();
+
+  // This runs on the main thread.
+  void Stop(PSLockRef aLock);
+
+ private:
+  // This suspends the calling thread for the given number of microseconds.
+  // Best effort timing.
+  void SleepMicro(uint32_t aMicroseconds);
+
+  // The sampler used to suspend and sample threads.
+  Sampler mSampler;
+
+  // The activity generation, for detecting when the sampler thread must stop.
+  const uint32_t mActivityGeneration;
+
+  // The interval between samples, measured in microseconds.
+  const int mIntervalMicroseconds;
+
+  // The OS-specific handle for the sampler thread.
+#if defined(GP_OS_windows)
+  HANDLE mThread;
+#elif defined(GP_OS_darwin) || defined(GP_OS_linux) || \
+    defined(GP_OS_android) || defined(GP_OS_freebsd)
+  pthread_t mThread;
+#endif
+
+  SamplerThread(const SamplerThread&) = delete;
+  void operator=(const SamplerThread&) = delete;
+};
+
+// This function is required because we need to create a SamplerThread within
+// ActivePS's constructor, but SamplerThread is defined after ActivePS. It
+// could probably be removed by moving some code around.
+static SamplerThread* NewSamplerThread(PSLockRef aLock, uint32_t aGeneration,
+                                       double aInterval) {
+  return new SamplerThread(aLock, aGeneration, aInterval);
+}
+
+// This function is the sampler thread.  This implementation is used for all
+// targets.
+void SamplerThread::Run() {
+  // TODO: If possible, name this thread later on, after NSPR becomes available.
+  // PR_SetCurrentThreadName("SamplerThread");
+
+  // Features won't change during this SamplerThread's lifetime, so we can read
+  // them once and store them locally.
+  const uint32_t features = []() -> uint32_t {
+    PSAutoLock lock;
+    if (!ActivePS::Exists(lock)) {
+      // If there is no active profiler, it doesn't matter what we return,
+      // because this thread will exit before any feature is used.
+      return 0;
+    }
+    return ActivePS::Features(lock);
+  }();
+
+  // Not *no*-stack-sampling means we do want stack sampling.
+  const bool stackSampling = !ProfilerFeature::HasNoStackSampling(features);
+
+  // Use local BlocksRingBuffer&ProfileBuffer to capture the stack.
+  // (This is to avoid touching the CorePS::CoreBuffer lock while
+  // a thread is suspended, because that thread could be working with
+  // the CorePS::CoreBuffer as well.)
+  ProfileBufferChunkManagerSingle localChunkManager(
+      ProfileBufferChunkManager::scExpectedMaximumStackSize);
+  ProfileChunkedBuffer localBuffer(
+      ProfileChunkedBuffer::ThreadSafety::WithoutMutex, localChunkManager);
+  ProfileBuffer localProfileBuffer(localBuffer);
+
+  // Will be kept between collections, to know what each collection does.
+  auto previousState = localBuffer.GetState();
+
+  // This will be positive if we are running behind schedule (sampling less
+  // frequently than desired) and negative if we are ahead of schedule.
+  TimeDuration lastSleepOvershoot = 0;
+  TimeStamp sampleStart = TimeStamp::NowUnfuzzed();
+
+  while (true) {
+    // This scope is for |lock|. It ends before we sleep below.
+    {
+      PSAutoLock lock;
+      TimeStamp lockAcquired = TimeStamp::NowUnfuzzed();
+
+      if (!ActivePS::Exists(lock)) {
+        return;
+      }
+
+      // At this point profiler_stop() might have been called, and
+      // profiler_start() might have been called on another thread. If this
+      // happens the generation won't match.
+      if (ActivePS::Generation(lock) != mActivityGeneration) {
+        return;
+      }
+
+      ActivePS::ClearExpiredExitProfiles(lock);
+
+      TimeStamp expiredMarkersCleaned = TimeStamp::NowUnfuzzed();
+
+      if (!ActivePS::IsSamplingPaused(lock)) {
+        TimeDuration delta = sampleStart - CorePS::ProcessStartTime();
+        ProfileBuffer& buffer = ActivePS::Buffer(lock);
+
+        // handle per-process generic counters
+        const Vector<BaseProfilerCount*>& counters = CorePS::Counters(lock);
+        for (auto& counter : counters) {
+          // create Buffer entries for each counter
+          buffer.AddEntry(ProfileBufferEntry::CounterId(counter));
+          buffer.AddEntry(ProfileBufferEntry::Time(delta.ToMilliseconds()));
+          // XXX support keyed maps of counts
+          // In the future, we'll support keyed counters - for example, counters
+          // with a key which is a thread ID. For "simple" counters we'll just
+          // use a key of 0.
+          int64_t count;
+          uint64_t number;
+          counter->Sample(count, number);
+          buffer.AddEntry(ProfileBufferEntry::CounterKey(0));
+          buffer.AddEntry(ProfileBufferEntry::Count(count));
+          if (number) {
+            buffer.AddEntry(ProfileBufferEntry::Number(number));
+          }
+        }
+        TimeStamp countersSampled = TimeStamp::NowUnfuzzed();
+
+        if (stackSampling) {
+          const Vector<LiveProfiledThreadData>& liveThreads =
+              ActivePS::LiveProfiledThreads(lock);
+
+          for (auto& thread : liveThreads) {
+            RegisteredThread* registeredThread = thread.mRegisteredThread;
+            ProfiledThreadData* profiledThreadData =
+                thread.mProfiledThreadData.get();
+            RefPtr<ThreadInfo> info = registeredThread->Info();
+
+            // If the thread is asleep and has been sampled before in the same
+            // sleep episode, find and copy the previous sample, as that's
+            // cheaper than taking a new sample.
+            if (registeredThread->RacyRegisteredThread()
+                    .CanDuplicateLastSampleDueToSleep()) {
+              bool dup_ok = ActivePS::Buffer(lock).DuplicateLastSample(
+                  info->ThreadId(), CorePS::ProcessStartTime(),
+                  profiledThreadData->LastSample());
+              if (dup_ok) {
+                continue;
+              }
+            }
+
+            AUTO_PROFILER_STATS(base_SamplerThread_Run_DoPeriodicSample);
+
+            TimeStamp now = TimeStamp::NowUnfuzzed();
+
+            // Add the thread ID now, so we know its position in the main
+            // buffer, which is used by some JS data. (DoPeriodicSample only
+            // knows about the temporary local buffer.)
+            uint64_t samplePos =
+                buffer.AddThreadIdEntry(registeredThread->Info()->ThreadId());
+            profiledThreadData->LastSample() = Some(samplePos);
+
+            // Also add the time, so it's always there after the thread ID, as
+            // expected by the parser. (Other stack data is optional.)
+            TimeDuration delta = now - CorePS::ProcessStartTime();
+            buffer.AddEntry(ProfileBufferEntry::Time(delta.ToMilliseconds()));
+
+            mSampler.SuspendAndSampleAndResumeThread(
+                lock, *registeredThread, now,
+                [&](const Registers& aRegs, const TimeStamp& aNow) {
+                  DoPeriodicSample(lock, *registeredThread, *profiledThreadData,
+                                   aRegs, samplePos, localProfileBuffer);
+                });
+
+            // If data is complete, copy it into the global buffer.
+            auto state = localBuffer.GetState();
+            if (state.mClearedBlockCount != previousState.mClearedBlockCount) {
+              LOG("Stack sample too big for local storage, needed %u bytes",
+                  unsigned(state.mRangeEnd - previousState.mRangeEnd));
+            } else if (state.mRangeEnd - previousState.mRangeEnd >=
+                       *CorePS::CoreBuffer().BufferLength()) {
+              LOG("Stack sample too big for profiler storage, needed %u bytes",
+                  unsigned(state.mRangeEnd - previousState.mRangeEnd));
+            } else {
+              CorePS::CoreBuffer().AppendContents(localBuffer);
+            }
+
+            // Clean up for the next run.
+            localBuffer.Clear();
+            previousState = localBuffer.GetState();
+          }
+        }
+
+#if defined(USE_LUL_STACKWALK)
+        // The LUL unwind object accumulates frame statistics. Periodically we
+        // should poke it to give it a chance to print those statistics.  This
+        // involves doing I/O (fprintf, __android_log_print, etc.) and so
+        // can't safely be done from the critical section inside
+        // SuspendAndSampleAndResumeThread, which is why it is done here.
+        CorePS::Lul(lock)->MaybeShowStats();
+#endif
+        TimeStamp threadsSampled = TimeStamp::NowUnfuzzed();
+
+        {
+          AUTO_PROFILER_STATS(Sampler_FulfillChunkRequests);
+          ActivePS::FulfillChunkRequests(lock);
+        }
+
+        buffer.CollectOverheadStats(delta, lockAcquired - sampleStart,
+                                    expiredMarkersCleaned - lockAcquired,
+                                    countersSampled - expiredMarkersCleaned,
+                                    threadsSampled - countersSampled);
+      }
+    }
+    // gPSMutex is not held after this point.
+
+    // Calculate how long a sleep to request.  After the sleep, measure how
+    // long we actually slept and take the difference into account when
+    // calculating the sleep interval for the next iteration.  This is an
+    // attempt to keep "to schedule" in the presence of inaccuracy of the
+    // actual sleep intervals.
+    TimeStamp targetSleepEndTime =
+        sampleStart + TimeDuration::FromMicroseconds(mIntervalMicroseconds);
+    TimeStamp beforeSleep = TimeStamp::NowUnfuzzed();
+    TimeDuration targetSleepDuration = targetSleepEndTime - beforeSleep;
+    double sleepTime = std::max(
+        0.0, (targetSleepDuration - lastSleepOvershoot).ToMicroseconds());
+    SleepMicro(static_cast<uint32_t>(sleepTime));
+    sampleStart = TimeStamp::NowUnfuzzed();
+    lastSleepOvershoot =
+        sampleStart - (beforeSleep + TimeDuration::FromMicroseconds(sleepTime));
+  }
+}
+
+// Temporary closing namespaces from enclosing platform.cpp.
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+// We #include these files directly because it means those files can use
+// declarations from this file trivially.  These provide target-specific
+// implementations of all SamplerThread methods except Run().
+#if defined(GP_OS_windows)
+#  include "platform-win32.cpp"
+#elif defined(GP_OS_darwin)
+#  include "platform-macos.cpp"
+#elif defined(GP_OS_linux) || defined(GP_OS_android) || defined(GP_OS_freebsd)
+#  include "platform-linux-android.cpp"
+#else
+#  error "bad platform"
+#endif
+
+namespace mozilla {
+namespace baseprofiler {
+
+UniquePlatformData AllocPlatformData(int aThreadId) {
+  return UniquePlatformData(new PlatformData(aThreadId));
+}
+
+void PlatformDataDestructor::operator()(PlatformData* aData) { delete aData; }
+
+// END SamplerThread
+////////////////////////////////////////////////////////////////////////
+
+////////////////////////////////////////////////////////////////////////
+// BEGIN externally visible functions
+
+static uint32_t ParseFeature(const char* aFeature, bool aIsStartup) {
+  if (strcmp(aFeature, "default") == 0) {
+    return (aIsStartup ? (DefaultFeatures() | StartupExtraDefaultFeatures())
+                       : DefaultFeatures()) &
+           AvailableFeatures();
+  }
+
+#define PARSE_FEATURE_BIT(n_, str_, Name_, desc_) \
+  if (strcmp(aFeature, str_) == 0) {              \
+    return ProfilerFeature::Name_;                \
+  }
+
+  BASE_PROFILER_FOR_EACH_FEATURE(PARSE_FEATURE_BIT)
+
+#undef PARSE_FEATURE_BIT
+
+  PrintToConsole("\nUnrecognized feature \"%s\".\n\n", aFeature);
+  // Since we may have an old feature we don't implement anymore, don't exit
+  PrintUsageThenExit(0);
+  return 0;
+}
+
+uint32_t ParseFeaturesFromStringArray(const char** aFeatures,
+                                      uint32_t aFeatureCount,
+                                      bool aIsStartup /* = false */) {
+  uint32_t features = 0;
+  for (size_t i = 0; i < aFeatureCount; i++) {
+    features |= ParseFeature(aFeatures[i], aIsStartup);
+  }
+  return features;
+}
+
+// Find the RegisteredThread for the current thread. This should only be called
+// in places where TLSRegisteredThread can't be used.
+static RegisteredThread* FindCurrentThreadRegisteredThread(PSLockRef aLock) {
+  int id = profiler_current_thread_id();
+  const Vector<UniquePtr<RegisteredThread>>& registeredThreads =
+      CorePS::RegisteredThreads(aLock);
+  for (auto& registeredThread : registeredThreads) {
+    if (registeredThread->Info()->ThreadId() == id) {
+      return registeredThread.get();
+    }
+  }
+
+  return nullptr;
+}
+
+static ProfilingStack* locked_register_thread(PSLockRef aLock,
+                                              const char* aName,
+                                              void* aStackTop) {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  MOZ_ASSERT(!FindCurrentThreadRegisteredThread(aLock));
+
+  VTUNE_REGISTER_THREAD(aName);
+
+  if (!TLSRegisteredThread::Init(aLock)) {
+    return nullptr;
+  }
+
+  RefPtr<ThreadInfo> info = new ThreadInfo(aName, profiler_current_thread_id(),
+                                           profiler_is_main_thread());
+  UniquePtr<RegisteredThread> registeredThread =
+      MakeUnique<RegisteredThread>(info, aStackTop);
+
+  TLSRegisteredThread::SetRegisteredThread(aLock, registeredThread.get());
+
+  if (ActivePS::Exists(aLock) && ActivePS::ShouldProfileThread(aLock, info)) {
+    registeredThread->RacyRegisteredThread().SetIsBeingProfiled(true);
+    ActivePS::AddLiveProfiledThread(aLock, registeredThread.get(),
+                                    MakeUnique<ProfiledThreadData>(info));
+  }
+
+  ProfilingStack* profilingStack =
+      &registeredThread->RacyRegisteredThread().ProfilingStack();
+
+  CorePS::AppendRegisteredThread(aLock, std::move(registeredThread));
+
+  return profilingStack;
+}
+
+static void locked_profiler_start(PSLockRef aLock, PowerOfTwo32 aCapacity,
+                                  double aInterval, uint32_t aFeatures,
+                                  const char** aFilters, uint32_t aFilterCount,
+                                  const Maybe<double>& aDuration);
+
+static Vector<const char*> SplitAtCommas(const char* aString,
+                                         UniquePtr<char[]>& aStorage) {
+  size_t len = strlen(aString);
+  aStorage = MakeUnique<char[]>(len + 1);
+  PodCopy(aStorage.get(), aString, len + 1);
+
+  // Iterate over all characters in aStorage and split at commas, by
+  // overwriting commas with the null char.
+  Vector<const char*> array;
+  size_t currentElementStart = 0;
+  for (size_t i = 0; i <= len; i++) {
+    if (aStorage[i] == ',') {
+      aStorage[i] = '\0';
+    }
+    if (aStorage[i] == '\0') {
+      MOZ_RELEASE_ASSERT(array.append(&aStorage[currentElementStart]));
+      currentElementStart = i + 1;
+    }
+  }
+  return array;
+}
+
+void profiler_init(void* aStackTop) {
+  LOG("profiler_init");
+
+  scProfilerMainThreadId = profiler_current_thread_id();
+
+  VTUNE_INIT();
+
+  MOZ_RELEASE_ASSERT(!CorePS::Exists());
+
+  if (getenv("MOZ_BASE_PROFILER_HELP")) {
+    PrintUsageThenExit(0);  // terminates execution
+  }
+
+  SharedLibraryInfo::Initialize();
+
+  uint32_t features = DefaultFeatures() & AvailableFeatures();
+
+  UniquePtr<char[]> filterStorage;
+
+  Vector<const char*> filters;
+  MOZ_RELEASE_ASSERT(filters.append(kMainThreadName));
+
+  PowerOfTwo32 capacity = BASE_PROFILER_DEFAULT_ENTRIES;
+  Maybe<double> duration = Nothing();
+  double interval = BASE_PROFILER_DEFAULT_INTERVAL;
+
+  {
+    PSAutoLock lock;
+
+    // We've passed the possible failure point. Instantiate CorePS, which
+    // indicates that the profiler has initialized successfully.
+    CorePS::Create(lock);
+
+    Unused << locked_register_thread(lock, kMainThreadName, aStackTop);
+
+    // Platform-specific initialization.
+    PlatformInit(lock);
+
+    // (Linux-only) We could create CorePS::mLul and read unwind info into it
+    // at this point. That would match the lifetime implied by destruction of
+    // it in profiler_shutdown() just below. However, that gives a big delay on
+    // startup, even if no profiling is actually to be done. So, instead, it is
+    // created on demand at the first call to PlatformStart().
+
+    const char* startupEnv = getenv("MOZ_PROFILER_STARTUP");
+    if (!startupEnv || startupEnv[0] == '\0' ||
+        ((startupEnv[0] == '0' || startupEnv[0] == 'N' ||
+          startupEnv[0] == 'n') &&
+         startupEnv[1] == '\0')) {
+      return;
+    }
+
+    // Hidden option to stop Base Profiler, mostly due to Talos intermittents,
+    // see https://bugzilla.mozilla.org/show_bug.cgi?id=1638851#c3
+    // TODO: Investigate root cause and remove this in bugs 1648324 and 1648325.
+    if (getenv("MOZ_PROFILER_STARTUP_NO_BASE")) {
+      return;
+    }
+
+    LOG("- MOZ_PROFILER_STARTUP is set");
+
+    // Startup default capacity may be different.
+    capacity = BASE_PROFILER_DEFAULT_STARTUP_ENTRIES;
+
+    const char* startupCapacity = getenv("MOZ_PROFILER_STARTUP_ENTRIES");
+    if (startupCapacity && startupCapacity[0] != '\0') {
+      errno = 0;
+      long capacityLong = strtol(startupCapacity, nullptr, 10);
+      // `long` could be 32 or 64 bits, so we force a 64-bit comparison with
+      // the maximum 32-bit signed number (as more than that is clamped down to
+      // 2^31 anyway).
+      if (errno == 0 && capacityLong > 0 &&
+          static_cast<uint64_t>(capacityLong) <=
+              static_cast<uint64_t>(INT32_MAX)) {
+        capacity = PowerOfTwo32(ActivePS::ClampToAllowedEntries(
+            static_cast<uint32_t>(capacityLong)));
+        LOG("- MOZ_PROFILER_STARTUP_ENTRIES = %u", unsigned(capacity.Value()));
+      } else {
+        PrintToConsole("- MOZ_PROFILER_STARTUP_ENTRIES not a valid integer: %s",
+                       startupCapacity);
+        PrintUsageThenExit(1);
+      }
+    }
+
+    const char* startupDuration = getenv("MOZ_PROFILER_STARTUP_DURATION");
+    if (startupDuration && startupDuration[0] != '\0') {
+      // The duration is a floating point number. Use StringToDouble rather than
+      // strtod, so that "." is used as the decimal separator regardless of OS
+      // locale.
+      auto durationVal = StringToDouble(std::string(startupDuration));
+      if (durationVal && *durationVal >= 0.0) {
+        if (*durationVal > 0.0) {
+          duration = Some(*durationVal);
+        }
+        LOG("- MOZ_PROFILER_STARTUP_DURATION = %f", *durationVal);
+      } else {
+        PrintToConsole("- MOZ_PROFILER_STARTUP_DURATION not a valid float: %s",
+                       startupDuration);
+        PrintUsageThenExit(1);
+      }
+    }
+
+    const char* startupInterval = getenv("MOZ_PROFILER_STARTUP_INTERVAL");
+    if (startupInterval && startupInterval[0] != '\0') {
+      // The interval is a floating point number. Use StringToDouble rather than
+      // strtod, so that "." is used as the decimal separator regardless of OS
+      // locale.
+      auto intervalValue = StringToDouble(MakeStringSpan(startupInterval));
+      if (intervalValue && *intervalValue > 0.0 && *intervalValue <= 1000.0) {
+        interval = *intervalValue;
+        LOG("- MOZ_PROFILER_STARTUP_INTERVAL = %f", interval);
+      } else {
+        PrintToConsole("- MOZ_PROFILER_STARTUP_INTERVAL not a valid float: %s",
+                       startupInterval);
+        PrintUsageThenExit(1);
+      }
+    }
+
+    features |= StartupExtraDefaultFeatures() & AvailableFeatures();
+
+    const char* startupFeaturesBitfield =
+        getenv("MOZ_PROFILER_STARTUP_FEATURES_BITFIELD");
+    if (startupFeaturesBitfield && startupFeaturesBitfield[0] != '\0') {
+      errno = 0;
+      features = strtol(startupFeaturesBitfield, nullptr, 10);
+      if (errno == 0 && features != 0) {
+        LOG("- MOZ_PROFILER_STARTUP_FEATURES_BITFIELD = %d", features);
+      } else {
+        PrintToConsole(
+            "- MOZ_PROFILER_STARTUP_FEATURES_BITFIELD not a valid integer: %s",
+            startupFeaturesBitfield);
+        PrintUsageThenExit(1);
+      }
+    } else {
+      const char* startupFeatures = getenv("MOZ_PROFILER_STARTUP_FEATURES");
+      if (startupFeatures && startupFeatures[0] != '\0') {
+        // Interpret startupFeatures as a list of feature strings, separated by
+        // commas.
+        UniquePtr<char[]> featureStringStorage;
+        Vector<const char*> featureStringArray =
+            SplitAtCommas(startupFeatures, featureStringStorage);
+        features = ParseFeaturesFromStringArray(featureStringArray.begin(),
+                                                featureStringArray.length(),
+                                                /* aIsStartup */ true);
+        LOG("- MOZ_PROFILER_STARTUP_FEATURES = %d", features);
+      }
+    }
+
+    const char* startupFilters = getenv("MOZ_PROFILER_STARTUP_FILTERS");
+    if (startupFilters && startupFilters[0] != '\0') {
+      filters = SplitAtCommas(startupFilters, filterStorage);
+      LOG("- MOZ_PROFILER_STARTUP_FILTERS = %s", startupFilters);
+    }
+
+    locked_profiler_start(lock, capacity, interval, features, filters.begin(),
+                          filters.length(), duration);
+  }
+
+  // TODO: Install memory counter if it is possible from mozglue.
+  // #if defined(MOZ_REPLACE_MALLOC) && defined(MOZ_PROFILER_MEMORY)
+  //   // start counting memory allocations (outside of lock because this may
+  //   call
+  //   // profiler_add_sampled_counter which would attempt to take the lock.)
+  //   mozilla::profiler::install_memory_counter(true);
+  // #endif
+}
+
+static void locked_profiler_save_profile_to_file(PSLockRef aLock,
+                                                 const char* aFilename,
+                                                 bool aIsShuttingDown);
+
+static SamplerThread* locked_profiler_stop(PSLockRef aLock);
+
+void profiler_shutdown() {
+  LOG("profiler_shutdown");
+
+  VTUNE_SHUTDOWN();
+
+  MOZ_RELEASE_ASSERT(profiler_is_main_thread());
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  // If the profiler is active we must get a handle to the SamplerThread before
+  // ActivePS is destroyed, in order to delete it.
+  SamplerThread* samplerThread = nullptr;
+  {
+    PSAutoLock lock;
+
+    // Save the profile on shutdown if requested.
+    if (ActivePS::Exists(lock)) {
+      const char* filename = getenv("MOZ_PROFILER_SHUTDOWN");
+      if (filename) {
+        locked_profiler_save_profile_to_file(lock, filename,
+                                             /* aIsShuttingDown */ true);
+      }
+
+      samplerThread = locked_profiler_stop(lock);
+    }
+
+    CorePS::Destroy(lock);
+
+    // We just destroyed CorePS and the ThreadInfos it contains, so we can
+    // clear this thread's TLSRegisteredThread.
+    TLSRegisteredThread::SetRegisteredThread(lock, nullptr);
+  }
+
+  // We do these operations with gPSMutex unlocked. The comments in
+  // profiler_stop() explain why.
+  if (samplerThread) {
+    delete samplerThread;
+  }
+}
+
+static bool WriteProfileToJSONWriter(SpliceableChunkedJSONWriter& aWriter,
+                                     double aSinceTime, bool aIsShuttingDown,
+                                     bool aOnlyThreads = false) {
+  LOG("WriteProfileToJSONWriter");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  if (!aOnlyThreads) {
+    aWriter.Start();
+    {
+      if (!profiler_stream_json_for_this_process(
+              aWriter, aSinceTime, aIsShuttingDown, aOnlyThreads)) {
+        return false;
+      }
+
+      // Don't include profiles from other processes because this is a
+      // synchronous function.
+      aWriter.StartArrayProperty("processes");
+      aWriter.EndArray();
+    }
+    aWriter.End();
+  } else {
+    aWriter.StartBareList();
+    if (!profiler_stream_json_for_this_process(aWriter, aSinceTime,
+                                               aIsShuttingDown, aOnlyThreads)) {
+      return false;
+    }
+    aWriter.EndBareList();
+  }
+  return true;
+}
+
+void profiler_set_process_name(const std::string& aProcessName,
+                               const std::string* aETLDplus1) {
+  LOG("profiler_set_process_name(\"%s\", \"%s\")", aProcessName.c_str(),
+      aETLDplus1 ? aETLDplus1->c_str() : "<none>");
+  PSAutoLock lock;
+  CorePS::SetProcessName(lock, aProcessName);
+  if (aETLDplus1) {
+    CorePS::SetETLDplus1(lock, *aETLDplus1);
+  }
+}
+
+UniquePtr<char[]> profiler_get_profile(double aSinceTime, bool aIsShuttingDown,
+                                       bool aOnlyThreads) {
+  LOG("profiler_get_profile");
+
+  SpliceableChunkedJSONWriter b;
+  if (!WriteProfileToJSONWriter(b, aSinceTime, aIsShuttingDown, aOnlyThreads)) {
+    return nullptr;
+  }
+  return b.ChunkedWriteFunc().CopyData();
+}
+
+void profiler_get_profile_json_into_lazily_allocated_buffer(
+    const std::function<char*(size_t)>& aAllocator, double aSinceTime,
+    bool aIsShuttingDown) {
+  LOG("profiler_get_profile_json_into_lazily_allocated_buffer");
+
+  SpliceableChunkedJSONWriter b;
+  if (!WriteProfileToJSONWriter(b, aSinceTime, aIsShuttingDown)) {
+    return;
+  }
+
+  b.ChunkedWriteFunc().CopyDataIntoLazilyAllocatedBuffer(aAllocator);
+}
+
+void profiler_get_start_params(int* aCapacity, Maybe<double>* aDuration,
+                               double* aInterval, uint32_t* aFeatures,
+                               Vector<const char*>* aFilters) {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  if (!aCapacity || !aDuration || !aInterval || !aFeatures || !aFilters) {
+    return;
+  }
+
+  PSAutoLock lock;
+
+  if (!ActivePS::Exists(lock)) {
+    *aCapacity = 0;
+    *aDuration = Nothing();
+    *aInterval = 0;
+    *aFeatures = 0;
+    aFilters->clear();
+    return;
+  }
+
+  *aCapacity = ActivePS::Capacity(lock).Value();
+  *aDuration = ActivePS::Duration(lock);
+  *aInterval = ActivePS::Interval(lock);
+  *aFeatures = ActivePS::Features(lock);
+
+  const Vector<std::string>& filters = ActivePS::Filters(lock);
+  MOZ_ALWAYS_TRUE(aFilters->resize(filters.length()));
+  for (uint32_t i = 0; i < filters.length(); ++i) {
+    (*aFilters)[i] = filters[i].c_str();
+  }
+}
+
+void GetProfilerEnvVarsForChildProcess(
+    std::function<void(const char* key, const char* value)>&& aSetEnv) {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (!ActivePS::Exists(lock)) {
+    aSetEnv("MOZ_PROFILER_STARTUP", "");
+    return;
+  }
+
+  aSetEnv("MOZ_PROFILER_STARTUP", "1");
+  auto capacityString =
+      Smprintf("%u", unsigned(ActivePS::Capacity(lock).Value()));
+  aSetEnv("MOZ_PROFILER_STARTUP_ENTRIES", capacityString.get());
+
+  // Use AppendFloat instead of Smprintf with %f because the decimal
+  // separator used by %f is locale-dependent. But the string we produce needs
+  // to be parseable by strtod, which only accepts the period character as a
+  // decimal separator. AppendFloat always uses the period character.
+  std::string intervalString = std::to_string(ActivePS::Interval(lock));
+  aSetEnv("MOZ_PROFILER_STARTUP_INTERVAL", intervalString.c_str());
+
+  auto featuresString = Smprintf("%d", ActivePS::Features(lock));
+  aSetEnv("MOZ_PROFILER_STARTUP_FEATURES_BITFIELD", featuresString.get());
+
+  std::string filtersString;
+  const Vector<std::string>& filters = ActivePS::Filters(lock);
+  for (uint32_t i = 0; i < filters.length(); ++i) {
+    filtersString += filters[i];
+    if (i != filters.length() - 1) {
+      filtersString += ",";
+    }
+  }
+  aSetEnv("MOZ_PROFILER_STARTUP_FILTERS", filtersString.c_str());
+}
+
+void profiler_received_exit_profile(const std::string& aExitProfile) {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+  PSAutoLock lock;
+  if (!ActivePS::Exists(lock)) {
+    return;
+  }
+  ActivePS::AddExitProfile(lock, aExitProfile);
+}
+
+Vector<std::string> profiler_move_exit_profiles() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+  PSAutoLock lock;
+  Vector<std::string> profiles;
+  if (ActivePS::Exists(lock)) {
+    profiles = ActivePS::MoveExitProfiles(lock);
+  }
+  return profiles;
+}
+
+static void locked_profiler_save_profile_to_file(PSLockRef aLock,
+                                                 const char* aFilename,
+                                                 bool aIsShuttingDown = false) {
+  LOG("locked_profiler_save_profile_to_file(%s)", aFilename);
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists() && ActivePS::Exists(aLock));
+
+  std::ofstream stream;
+  stream.open(aFilename);
+  if (stream.is_open()) {
+    SpliceableJSONWriter w(MakeUnique<OStreamJSONWriteFunc>(stream));
+    w.Start();
+    {
+      locked_profiler_stream_json_for_this_process(aLock, w, /* sinceTime */ 0,
+                                                   aIsShuttingDown);
+
+      w.StartArrayProperty("processes");
+      Vector<std::string> exitProfiles = ActivePS::MoveExitProfiles(aLock);
+      for (auto& exitProfile : exitProfiles) {
+        if (!exitProfile.empty()) {
+          w.Splice(exitProfile);
+        }
+      }
+      w.EndArray();
+    }
+    w.End();
+
+    stream.close();
+  }
+}
+
+void profiler_save_profile_to_file(const char* aFilename) {
+  LOG("profiler_save_profile_to_file(%s)", aFilename);
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (!ActivePS::Exists(lock)) {
+    return;
+  }
+
+  locked_profiler_save_profile_to_file(lock, aFilename);
+}
+
+uint32_t profiler_get_available_features() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+  return AvailableFeatures();
+}
+
+Maybe<ProfilerBufferInfo> profiler_get_buffer_info() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (!ActivePS::Exists(lock)) {
+    return Nothing();
+  }
+
+  return Some(ActivePS::Buffer(lock).GetProfilerBufferInfo());
+}
+
+// This basically duplicates AutoProfilerLabel's constructor.
+static void* MozGlueBaseLabelEnter(const char* aLabel,
+                                   const char* aDynamicString, void* aSp) {
+  ProfilingStack* profilingStack = AutoProfilerLabel::sProfilingStack.get();
+  if (profilingStack) {
+    profilingStack->pushLabelFrame(aLabel, aDynamicString, aSp,
+                                   ProfilingCategoryPair::OTHER);
+  }
+  return profilingStack;
+}
+
+// This basically duplicates AutoProfilerLabel's destructor.
+static void MozGlueBaseLabelExit(void* sProfilingStack) {
+  if (sProfilingStack) {
+    reinterpret_cast<ProfilingStack*>(sProfilingStack)->pop();
+  }
+}
+
+static void locked_profiler_start(PSLockRef aLock, PowerOfTwo32 aCapacity,
+                                  double aInterval, uint32_t aFeatures,
+                                  const char** aFilters, uint32_t aFilterCount,
+                                  const Maybe<double>& aDuration) {
+  if (LOG_TEST) {
+    LOG("locked_profiler_start");
+    LOG("- capacity  = %d", int(aCapacity.Value()));
+    LOG("- duration  = %.2f", aDuration ? *aDuration : -1);
+    LOG("- interval = %.2f", aInterval);
+
+#define LOG_FEATURE(n_, str_, Name_, desc_)     \
+  if (ProfilerFeature::Has##Name_(aFeatures)) { \
+    LOG("- feature  = %s", str_);               \
+  }
+
+    BASE_PROFILER_FOR_EACH_FEATURE(LOG_FEATURE)
+
+#undef LOG_FEATURE
+
+    for (uint32_t i = 0; i < aFilterCount; i++) {
+      LOG("- threads  = %s", aFilters[i]);
+    }
+  }
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists() && !ActivePS::Exists(aLock));
+
+#if defined(GP_PLAT_amd64_windows)
+  InitializeWin64ProfilerHooks();
+#endif
+
+  // Fall back to the default values if the passed-in values are unreasonable.
+  // We want to be able to store at least one full stack.
+  // TODO: Review magic numbers.
+  PowerOfTwo32 capacity =
+      (aCapacity.Value() >=
+       ProfileBufferChunkManager::scExpectedMaximumStackSize / scBytesPerEntry)
+          ? aCapacity
+          : BASE_PROFILER_DEFAULT_ENTRIES;
+  Maybe<double> duration = aDuration;
+
+  if (aDuration && *aDuration <= 0) {
+    duration = Nothing();
+  }
+  double interval = aInterval > 0 ? aInterval : BASE_PROFILER_DEFAULT_INTERVAL;
+
+  ActivePS::Create(aLock, capacity, interval, aFeatures, aFilters, aFilterCount,
+                   duration);
+
+  // Set up profiling for each registered thread, if appropriate.
+  const Vector<UniquePtr<RegisteredThread>>& registeredThreads =
+      CorePS::RegisteredThreads(aLock);
+  for (auto& registeredThread : registeredThreads) {
+    RefPtr<ThreadInfo> info = registeredThread->Info();
+
+    if (ActivePS::ShouldProfileThread(aLock, info)) {
+      registeredThread->RacyRegisteredThread().SetIsBeingProfiled(true);
+      ActivePS::AddLiveProfiledThread(aLock, registeredThread.get(),
+                                      MakeUnique<ProfiledThreadData>(info));
+      registeredThread->RacyRegisteredThread().ReinitializeOnResume();
+    }
+  }
+
+  // Setup support for pushing/popping labels in mozglue.
+  RegisterProfilerLabelEnterExit(MozGlueBaseLabelEnter, MozGlueBaseLabelExit);
+
+  // At the very end, set up RacyFeatures.
+  RacyFeatures::SetActive(ActivePS::Features(aLock));
+}
+
+void profiler_start(PowerOfTwo32 aCapacity, double aInterval,
+                    uint32_t aFeatures, const char** aFilters,
+                    uint32_t aFilterCount, const Maybe<double>& aDuration) {
+  LOG("profiler_start");
+
+  SamplerThread* samplerThread = nullptr;
+  {
+    PSAutoLock lock;
+
+    // Initialize if necessary.
+    if (!CorePS::Exists()) {
+      profiler_init(nullptr);
+    }
+
+    // Reset the current state if the profiler is running.
+    if (ActivePS::Exists(lock)) {
+      samplerThread = locked_profiler_stop(lock);
+    }
+
+    locked_profiler_start(lock, aCapacity, aInterval, aFeatures, aFilters,
+                          aFilterCount, aDuration);
+  }
+
+  // TODO: Install memory counter if it is possible from mozglue.
+  // #if defined(MOZ_REPLACE_MALLOC) && defined(MOZ_PROFILER_MEMORY)
+  //   // start counting memory allocations (outside of lock because this may
+  //   call
+  //   // profiler_add_sampled_counter which would attempt to take the lock.)
+  //   mozilla::profiler::install_memory_counter(true);
+  // #endif
+
+  // We do these operations with gPSMutex unlocked. The comments in
+  // profiler_stop() explain why.
+  if (samplerThread) {
+    delete samplerThread;
+  }
+}
+
+void profiler_ensure_started(PowerOfTwo32 aCapacity, double aInterval,
+                             uint32_t aFeatures, const char** aFilters,
+                             uint32_t aFilterCount,
+                             const Maybe<double>& aDuration) {
+  LOG("profiler_ensure_started");
+
+  // bool startedProfiler = false; (See TODO below)
+  SamplerThread* samplerThread = nullptr;
+  {
+    PSAutoLock lock;
+
+    // Initialize if necessary.
+    if (!CorePS::Exists()) {
+      profiler_init(nullptr);
+    }
+
+    if (ActivePS::Exists(lock)) {
+      // The profiler is active.
+      if (!ActivePS::Equals(lock, aCapacity, aDuration, aInterval, aFeatures,
+                            aFilters, aFilterCount)) {
+        // Stop and restart with different settings.
+        samplerThread = locked_profiler_stop(lock);
+        locked_profiler_start(lock, aCapacity, aInterval, aFeatures, aFilters,
+                              aFilterCount, aDuration);
+        // startedProfiler = true; (See TODO below)
+      }
+    } else {
+      // The profiler is stopped.
+      locked_profiler_start(lock, aCapacity, aInterval, aFeatures, aFilters,
+                            aFilterCount, aDuration);
+      // startedProfiler = true; (See TODO below)
+    }
+  }
+
+  // TODO: Install memory counter if it is possible from mozglue.
+  // #if defined(MOZ_REPLACE_MALLOC) && defined(MOZ_PROFILER_MEMORY)
+  //   // start counting memory allocations (outside of lock because this may
+  //   // call profiler_add_sampled_counter which would attempt to take the
+  //   // lock.)
+  //   mozilla::profiler::install_memory_counter(true);
+  // #endif
+
+  // We do these operations with gPSMutex unlocked. The comments in
+  // profiler_stop() explain why.
+  if (samplerThread) {
+    delete samplerThread;
+  }
+}
+
+[[nodiscard]] static SamplerThread* locked_profiler_stop(PSLockRef aLock) {
+  LOG("locked_profiler_stop");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists() && ActivePS::Exists(aLock));
+
+  // At the very start, clear RacyFeatures.
+  RacyFeatures::SetInactive();
+
+  // TODO: Uninstall memory counter if it is possible from mozglue.
+  // #if defined(MOZ_REPLACE_MALLOC) && defined(MOZ_PROFILER_MEMORY)
+  //   mozilla::profiler::install_memory_counter(false);
+  // #endif
+
+  // Remove support for pushing/popping labels in mozglue.
+  RegisterProfilerLabelEnterExit(nullptr, nullptr);
+
+  // Stop sampling live threads.
+  const Vector<LiveProfiledThreadData>& liveProfiledThreads =
+      ActivePS::LiveProfiledThreads(aLock);
+  for (auto& thread : liveProfiledThreads) {
+    RegisteredThread* registeredThread = thread.mRegisteredThread;
+    registeredThread->RacyRegisteredThread().SetIsBeingProfiled(false);
+  }
+
+  // The Stop() call doesn't actually stop Run(); that happens in this
+  // function's caller when the sampler thread is destroyed. Stop() just gives
+  // the SamplerThread a chance to do some cleanup with gPSMutex locked.
+  SamplerThread* samplerThread = ActivePS::Destroy(aLock);
+  samplerThread->Stop(aLock);
+
+  return samplerThread;
+}
+
+void profiler_stop() {
+  LOG("profiler_stop");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  SamplerThread* samplerThread;
+  {
+    PSAutoLock lock;
+
+    if (!ActivePS::Exists(lock)) {
+      return;
+    }
+
+    samplerThread = locked_profiler_stop(lock);
+  }
+
+  // We delete with gPSMutex unlocked. Otherwise we would get a deadlock: we
+  // would be waiting here with gPSMutex locked for SamplerThread::Run() to
+  // return so the join operation within the destructor can complete, but Run()
+  // needs to lock gPSMutex to return.
+  //
+  // Because this call occurs with gPSMutex unlocked, it -- including the final
+  // iteration of Run()'s loop -- must be able detect deactivation and return
+  // in a way that's safe with respect to other gPSMutex-locking operations
+  // that may have occurred in the meantime.
+  delete samplerThread;
+}
+
+bool profiler_is_paused() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (!ActivePS::Exists(lock)) {
+    return false;
+  }
+
+  return ActivePS::IsPaused(lock);
+}
+
+void profiler_pause() {
+  LOG("profiler_pause");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  {
+    PSAutoLock lock;
+
+    if (!ActivePS::Exists(lock)) {
+      return;
+    }
+
+    RacyFeatures::SetPaused();
+    ActivePS::SetIsPaused(lock, true);
+    ActivePS::Buffer(lock).AddEntry(ProfileBufferEntry::Pause(profiler_time()));
+  }
+}
+
+void profiler_resume() {
+  LOG("profiler_resume");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  {
+    PSAutoLock lock;
+
+    if (!ActivePS::Exists(lock)) {
+      return;
+    }
+
+    ActivePS::Buffer(lock).AddEntry(
+        ProfileBufferEntry::Resume(profiler_time()));
+    ActivePS::SetIsPaused(lock, false);
+    RacyFeatures::SetUnpaused();
+  }
+}
+
+bool profiler_is_sampling_paused() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (!ActivePS::Exists(lock)) {
+    return false;
+  }
+
+  return ActivePS::IsSamplingPaused(lock);
+}
+
+void profiler_pause_sampling() {
+  LOG("profiler_pause_sampling");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  {
+    PSAutoLock lock;
+
+    if (!ActivePS::Exists(lock)) {
+      return;
+    }
+
+    RacyFeatures::SetSamplingPaused();
+    ActivePS::SetIsSamplingPaused(lock, true);
+    ActivePS::Buffer(lock).AddEntry(
+        ProfileBufferEntry::PauseSampling(profiler_time()));
+  }
+}
+
+void profiler_resume_sampling() {
+  LOG("profiler_resume_sampling");
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  {
+    PSAutoLock lock;
+
+    if (!ActivePS::Exists(lock)) {
+      return;
+    }
+
+    ActivePS::Buffer(lock).AddEntry(
+        ProfileBufferEntry::ResumeSampling(profiler_time()));
+    ActivePS::SetIsSamplingPaused(lock, false);
+    RacyFeatures::SetSamplingUnpaused();
+  }
+}
+
+bool profiler_feature_active(uint32_t aFeature) {
+  // This function runs both on and off the main thread.
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  // This function is hot enough that we use RacyFeatures, not ActivePS.
+  return RacyFeatures::IsActiveWithFeature(aFeature);
+}
+
+void profiler_add_sampled_counter(BaseProfilerCount* aCounter) {
+  DEBUG_LOG("profiler_add_sampled_counter(%s)", aCounter->mLabel);
+  PSAutoLock lock;
+  CorePS::AppendCounter(lock, aCounter);
+}
+
+void profiler_remove_sampled_counter(BaseProfilerCount* aCounter) {
+  DEBUG_LOG("profiler_remove_sampled_counter(%s)", aCounter->mLabel);
+  PSAutoLock lock;
+  // Note: we don't enforce a final sample, though we could do so if the
+  // profiler was active
+  CorePS::RemoveCounter(lock, aCounter);
+}
+
+ProfilingStack* profiler_register_thread(const char* aName,
+                                         void* aGuessStackTop) {
+  DEBUG_LOG("profiler_register_thread(%s)", aName);
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (RegisteredThread* thread = FindCurrentThreadRegisteredThread(lock);
+      thread) {
+    LOG("profiler_register_thread(%s) - thread %d already registered as %s",
+        aName, profiler_current_thread_id(), thread->Info()->Name());
+    // TODO: Use new name. This is currently not possible because the
+    // RegisteredThread's ThreadInfo cannot be changed.
+    // In the meantime, we record a marker that could be used in the frontend.
+    std::string text("Thread ");
+    text += std::to_string(profiler_current_thread_id());
+    text += " \"";
+    text += thread->Info()->Name();
+    text += "\" attempted to re-register as \"";
+    text += aName;
+    text += "\"";
+    BASE_PROFILER_MARKER_TEXT("profiler_register_thread again", OTHER_Profiling,
+                              MarkerThreadId::MainThread(), text);
+
+    return &thread->RacyRegisteredThread().ProfilingStack();
+  }
+
+  void* stackTop = GetStackTop(aGuessStackTop);
+  return locked_register_thread(lock, aName, stackTop);
+}
+
+void profiler_unregister_thread() {
+  if (!CorePS::Exists()) {
+    // This function can be called after the main thread has already shut down.
+    return;
+  }
+
+  PSAutoLock lock;
+
+  RegisteredThread* registeredThread = FindCurrentThreadRegisteredThread(lock);
+  MOZ_RELEASE_ASSERT(registeredThread ==
+                     TLSRegisteredThread::RegisteredThread(lock));
+  if (registeredThread) {
+    RefPtr<ThreadInfo> info = registeredThread->Info();
+
+    DEBUG_LOG("profiler_unregister_thread: %s", info->Name());
+
+    if (ActivePS::Exists(lock)) {
+      ActivePS::UnregisterThread(lock, registeredThread);
+    }
+
+    // Clear the pointer to the RegisteredThread object that we're about to
+    // destroy.
+    TLSRegisteredThread::SetRegisteredThread(lock, nullptr);
+
+    // Remove the thread from the list of registered threads. This deletes the
+    // registeredThread object.
+    CorePS::RemoveRegisteredThread(lock, registeredThread);
+  } else {
+    LOG("profiler_unregister_thread() - thread %d already unregistered",
+        profiler_current_thread_id());
+    // We cannot record a marker on this thread because it was already
+    // unregistered. Send it to the main thread (unless this *is* already the
+    // main thread, which has been unregistered); this may be useful to catch
+    // mismatched register/unregister pairs in Firefox.
+    if (int tid = profiler_current_thread_id();
+        tid != profiler_main_thread_id()) {
+      BASE_PROFILER_MARKER_TEXT("profiler_unregister_thread again",
+                                OTHER_Profiling, MarkerThreadId::MainThread(),
+                                std::to_string(profiler_current_thread_id()));
+    }
+    // There are two ways FindCurrentThreadRegisteredThread() might have failed.
+    //
+    // - TLSRegisteredThread::Init() failed in locked_register_thread().
+    //
+    // - We've already called profiler_unregister_thread() for this thread.
+    //   (Whether or not it should, this does happen in practice.)
+    //
+    // Either way, TLSRegisteredThread should be empty.
+    MOZ_RELEASE_ASSERT(!TLSRegisteredThread::RegisteredThread(lock));
+  }
+}
+
+void profiler_register_page(uint64_t aBrowsingContextID,
+                            uint64_t aInnerWindowID, const std::string& aUrl,
+                            uint64_t aEmbedderInnerWindowID) {
+  DEBUG_LOG("profiler_register_page(%" PRIu64 ", %" PRIu64 ", %s, %" PRIu64 ")",
+            aBrowsingContextID, aInnerWindowID, aUrl.c_str(),
+            aEmbedderInnerWindowID);
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  // When a Browsing context is first loaded, the first url loaded in it will be
+  // about:blank. Because of that, this call keeps the first non-about:blank
+  // registration of window and discards the previous one.
+  RefPtr<PageInformation> pageInfo = new PageInformation(
+      aBrowsingContextID, aInnerWindowID, aUrl, aEmbedderInnerWindowID);
+  CorePS::AppendRegisteredPage(lock, std::move(pageInfo));
+
+  // After appending the given page to CorePS, look for the expired
+  // pages and remove them if there are any.
+  if (ActivePS::Exists(lock)) {
+    ActivePS::DiscardExpiredPages(lock);
+  }
+}
+
+void profiler_unregister_page(uint64_t aRegisteredInnerWindowID) {
+  if (!CorePS::Exists()) {
+    // This function can be called after the main thread has already shut down.
+    return;
+  }
+
+  PSAutoLock lock;
+
+  // During unregistration, if the profiler is active, we have to keep the
+  // page information since there may be some markers associated with the given
+  // page. But if profiler is not active. we have no reason to keep the
+  // page information here because there can't be any marker associated with it.
+  if (ActivePS::Exists(lock)) {
+    ActivePS::UnregisterPage(lock, aRegisteredInnerWindowID);
+  } else {
+    CorePS::RemoveRegisteredPage(lock, aRegisteredInnerWindowID);
+  }
+}
+
+void profiler_clear_all_pages() {
+  if (!CorePS::Exists()) {
+    // This function can be called after the main thread has already shut down.
+    return;
+  }
+
+  {
+    PSAutoLock lock;
+    CorePS::ClearRegisteredPages(lock);
+    if (ActivePS::Exists(lock)) {
+      ActivePS::ClearUnregisteredPages(lock);
+    }
+  }
+}
+
+void profiler_thread_sleep() {
+  // This function runs both on and off the main thread.
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  RacyRegisteredThread* racyRegisteredThread =
+      TLSRegisteredThread::RacyRegisteredThread();
+  if (!racyRegisteredThread) {
+    return;
+  }
+
+  racyRegisteredThread->SetSleeping();
+}
+
+void profiler_thread_wake() {
+  // This function runs both on and off the main thread.
+
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  RacyRegisteredThread* racyRegisteredThread =
+      TLSRegisteredThread::RacyRegisteredThread();
+  if (!racyRegisteredThread) {
+    return;
+  }
+
+  racyRegisteredThread->SetAwake();
+}
+
+bool detail::IsThreadBeingProfiled() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  const RacyRegisteredThread* racyRegisteredThread =
+      TLSRegisteredThread::RacyRegisteredThread();
+  return racyRegisteredThread && racyRegisteredThread->IsBeingProfiled();
+}
+
+bool profiler_thread_is_sleeping() {
+  MOZ_RELEASE_ASSERT(profiler_is_main_thread());
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  RacyRegisteredThread* racyRegisteredThread =
+      TLSRegisteredThread::RacyRegisteredThread();
+  if (!racyRegisteredThread) {
+    return false;
+  }
+  return racyRegisteredThread->IsSleeping();
+}
+
+double profiler_time() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  TimeDuration delta = TimeStamp::NowUnfuzzed() - CorePS::ProcessStartTime();
+  return delta.ToMilliseconds();
+}
+
+bool profiler_capture_backtrace_into(ProfileChunkedBuffer& aChunkedBuffer) {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  PSAutoLock lock;
+
+  if (!ActivePS::Exists(lock)) {
+    return false;
+  }
+
+  RegisteredThread* registeredThread =
+      TLSRegisteredThread::RegisteredThread(lock);
+  if (!registeredThread) {
+    MOZ_ASSERT(registeredThread);
+    return false;
+  }
+
+  ProfileBuffer profileBuffer(aChunkedBuffer);
+
+  Registers regs;
+#if defined(HAVE_NATIVE_UNWIND)
+  regs.SyncPopulate();
+#else
+  regs.Clear();
+#endif
+
+  DoSyncSample(lock, *registeredThread, TimeStamp::NowUnfuzzed(), regs,
+               profileBuffer);
+
+  return true;
+}
+
+UniquePtr<ProfileChunkedBuffer> profiler_capture_backtrace() {
+  MOZ_RELEASE_ASSERT(CorePS::Exists());
+
+  // Quick is-active check before allocating a buffer.
+  if (!profiler_is_active()) {
+    return nullptr;
+  }
+
+  auto buffer = MakeUnique<ProfileChunkedBuffer>(
+      ProfileChunkedBuffer::ThreadSafety::WithoutMutex,
+      MakeUnique<ProfileBufferChunkManagerSingle>(
+          ProfileBufferChunkManager::scExpectedMaximumStackSize));
+
+  if (!profiler_capture_backtrace_into(*buffer)) {
+    return nullptr;
+  }
+
+  return buffer;
+}
+
+UniqueProfilerBacktrace profiler_get_backtrace() {
+  UniquePtr<ProfileChunkedBuffer> buffer = profiler_capture_backtrace();
+
+  if (!buffer) {
+    return nullptr;
+  }
+
+  return UniqueProfilerBacktrace(
+      new ProfilerBacktrace("SyncProfile", std::move(buffer)));
+}
+
+void ProfilerBacktraceDestructor::operator()(ProfilerBacktrace* aBacktrace) {
+  delete aBacktrace;
+}
+
+bool profiler_is_locked_on_current_thread() {
+  // This function is used to help users avoid calling `profiler_...` functions
+  // when the profiler may already have a lock in place, which would prevent a
+  // 2nd recursive lock (resulting in a crash or a never-ending wait).
+  // So we must return `true` for any of:
+  // - The main profiler mutex, used by most functions, and/or
+  // - The buffer mutex, used directly in some functions without locking the
+  //   main mutex, e.g., marker-related functions.
+  return PSAutoLock::IsLockedOnCurrentThread() ||
+         CorePS::CoreBuffer().IsThreadSafeAndLockedOnCurrentThread();
+}
+
+// This is a simplified version of profiler_add_marker that can be easily passed
+// into the JS engine.
+void profiler_add_js_marker(const char* aMarkerName, const char* aMarkerText) {
+  BASE_PROFILER_MARKER_TEXT(
+      ProfilerString8View::WrapNullTerminatedString(aMarkerName), JS, {},
+      ProfilerString8View::WrapNullTerminatedString(aMarkerText));
+}
+
+// NOTE: aCollector's methods will be called while the target thread is paused.
+// Doing things in those methods like allocating -- which may try to claim
+// locks -- is a surefire way to deadlock.
+void profiler_suspend_and_sample_thread(int aThreadId, uint32_t aFeatures,
+                                        ProfilerStackCollector& aCollector,
+                                        bool aSampleNative /* = true */) {
+  // Lock the profiler mutex
+  PSAutoLock lock;
+
+  const Vector<UniquePtr<RegisteredThread>>& registeredThreads =
+      CorePS::RegisteredThreads(lock);
+  for (auto& thread : registeredThreads) {
+    RefPtr<ThreadInfo> info = thread->Info();
+    RegisteredThread& registeredThread = *thread.get();
+
+    if (info->ThreadId() == aThreadId) {
+      if (info->IsMainThread()) {
+        aCollector.SetIsMainThread();
+      }
+
+      // Allocate the space for the native stack
+      NativeStack nativeStack;
+
+      // Suspend, sample, and then resume the target thread.
+      Sampler sampler(lock);
+      TimeStamp now = TimeStamp::NowUnfuzzed();
+      sampler.SuspendAndSampleAndResumeThread(
+          lock, registeredThread, now,
+          [&](const Registers& aRegs, const TimeStamp& aNow) {
+            // The target thread is now suspended. Collect a native
+            // backtrace, and call the callback.
+            bool isSynchronous = false;
+#if defined(HAVE_FASTINIT_NATIVE_UNWIND)
+            if (aSampleNative) {
+          // We can only use FramePointerStackWalk or MozStackWalk from
+          // suspend_and_sample_thread as other stackwalking methods may not be
+          // initialized.
+#  if defined(USE_FRAME_POINTER_STACK_WALK)
+              DoFramePointerBacktrace(lock, registeredThread, aRegs,
+                                      nativeStack);
+#  elif defined(USE_MOZ_STACK_WALK)
+              DoMozStackWalkBacktrace(lock, registeredThread, aRegs,
+                                      nativeStack);
+#  else
+#    error "Invalid configuration"
+#  endif
+
+              MergeStacks(aFeatures, isSynchronous, registeredThread, aRegs,
+                          nativeStack, aCollector);
+            } else
+#endif
+            {
+              MergeStacks(aFeatures, isSynchronous, registeredThread, aRegs,
+                          nativeStack, aCollector);
+
+              if (ProfilerFeature::HasLeaf(aFeatures)) {
+                aCollector.CollectNativeLeafAddr((void*)aRegs.mPC);
+              }
+            }
+          });
+
+      // NOTE: Make sure to disable the sampler before it is destroyed, in case
+      // the profiler is running at the same time.
+      sampler.Disable(lock);
+      break;
+    }
+  }
+}
+
+// END externally visible functions
+////////////////////////////////////////////////////////////////////////
+
+}  // namespace baseprofiler
+}  // namespace mozilla
diff --git a/mozglue/baseprofiler/core/platform.h b/mozglue/baseprofiler/core/platform.h
new file mode 100644
index 0000000000..1913a0def6
--- /dev/null
+++ b/mozglue/baseprofiler/core/platform.h
@@ -0,0 +1,132 @@
+// Copyright (c) 2006-2011 The Chromium Authors. All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//  * Redistributions of source code must retain the above copyright
+//    notice, this list of conditions and the following disclaimer.
+//  * Redistributions in binary form must reproduce the above copyright
+//    notice, this list of conditions and the following disclaimer in
+//    the documentation and/or other materials provided with the
+//    distribution.
+//  * Neither the name of Google, Inc. nor the names of its contributors
+//    may be used to endorse or promote products derived from this
+//    software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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 OWNER OR CONTRIBUTORS 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 TOOLS_PLATFORM_H_
+#define TOOLS_PLATFORM_H_
+
+#include "PlatformMacros.h"
+
+#include "BaseProfiler.h"
+
+#include "mozilla/Logging.h"
+#include "mozilla/UniquePtr.h"
+#include "mozilla/Vector.h"
+
+#include <functional>
+#include <stdint.h>
+#include <string>
+
+namespace mozilla {
+namespace baseprofiler {
+bool LogTest(int aLevelToTest);
+void PrintToConsole(const char* aFmt, ...) MOZ_FORMAT_PRINTF(1, 2);
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+// These are for MOZ_BASE_PROFILER_LOGGING and above. It's the default logging
+// level for the profiler, and should be used sparingly.
+#define LOG_TEST ::mozilla::baseprofiler::LogTest(3)
+#define LOG(arg, ...)                                           \
+  do {                                                          \
+    if (LOG_TEST) {                                             \
+      ::mozilla::baseprofiler::PrintToConsole(                  \
+          "[I %d/%d] " arg "\n", profiler_current_process_id(), \
+          profiler_current_thread_id(), ##__VA_ARGS__);         \
+    }                                                           \
+  } while (0)
+
+// These are for MOZ_BASE_PROFILER_DEBUG_LOGGING. It should be used for logging
+// that is somewhat more verbose than LOG.
+#define DEBUG_LOG_TEST ::mozilla::baseprofiler::LogTest(4)
+#define DEBUG_LOG(arg, ...)                                     \
+  do {                                                          \
+    if (DEBUG_LOG_TEST) {                                       \
+      ::mozilla::baseprofiler::PrintToConsole(                  \
+          "[D %d/%d] " arg "\n", profiler_current_process_id(), \
+          profiler_current_thread_id(), ##__VA_ARGS__);         \
+    }                                                           \
+  } while (0)
+
+// These are for MOZ_BASE_PROFILER_VERBOSE_LOGGING. It should be used for
+// logging that is somewhat more verbose than DEBUG_LOG.
+#define VERBOSE_LOG_TEST ::mozilla::baseprofiler::LogTest(5)
+#define VERBOSE_LOG(arg, ...)                                   \
+  do {                                                          \
+    if (VERBOSE_LOG_TEST) {                                     \
+      ::mozilla::baseprofiler::PrintToConsole(                  \
+          "[V %d/%d] " arg "\n", profiler_current_process_id(), \
+          profiler_current_thread_id(), ##__VA_ARGS__);         \
+    }                                                           \
+  } while (0)
+
+namespace mozilla {
+
+class JSONWriter;
+
+namespace baseprofiler {
+
+typedef uint8_t* Address;
+
+class PlatformData;
+
+// We can't new/delete the type safely without defining it
+// (-Wdelete-incomplete).  Use these to hide the details from clients.
+struct PlatformDataDestructor {
+  void operator()(PlatformData*);
+};
+
+typedef UniquePtr<PlatformData, PlatformDataDestructor> UniquePlatformData;
+UniquePlatformData AllocPlatformData(int aThreadId);
+
+// Convert the array of strings to a bitfield.
+uint32_t ParseFeaturesFromStringArray(const char** aFeatures,
+                                      uint32_t aFeatureCount,
+                                      bool aIsStartup = false);
+
+void profiler_get_profile_json_into_lazily_allocated_buffer(
+    const std::function<char*(size_t)>& aAllocator, double aSinceTime,
+    bool aIsShuttingDown);
+
+// Flags to conveniently track various JS instrumentations.
+enum class JSInstrumentationFlags {
+  StackSampling = 0x1,
+  TraceLogging = 0x2,
+  Allocations = 0x4,
+};
+
+// Record an exit profile from a child process.
+void profiler_received_exit_profile(const std::string& aExitProfile);
+
+// Extract all received exit profiles that have not yet expired (i.e., they
+// still intersect with this process' buffer range).
+Vector<std::string> profiler_move_exit_profiles();
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif /* ndef TOOLS_PLATFORM_H_ */
diff --git a/mozglue/baseprofiler/core/shared-libraries-linux.cc b/mozglue/baseprofiler/core/shared-libraries-linux.cc
new file mode 100644
index 0000000000..c38e72378a
--- /dev/null
+++ b/mozglue/baseprofiler/core/shared-libraries-linux.cc
@@ -0,0 +1,835 @@
+/* -*- 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 "BaseProfilerSharedLibraries.h"
+
+#define PATH_MAX_TOSTRING(x) #x
+#define PATH_MAX_STRING(x) PATH_MAX_TOSTRING(x)
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <limits.h>
+#include <unistd.h>
+#include <fstream>
+#include "platform.h"
+#include "mozilla/Sprintf.h"
+#include "mozilla/Unused.h"
+
+#include <algorithm>
+#include <arpa/inet.h>
+#include <dlfcn.h>
+#include <elf.h>
+#include <fcntl.h>
+#if defined(GP_OS_linux) || defined(GP_OS_android)
+#  include <features.h>
+#endif
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <vector>
+
+#if defined(GP_OS_linux) || defined(GP_OS_android) || defined(GP_OS_freebsd)
+#  include <link.h>  // dl_phdr_info, ElfW()
+#else
+#  error "Unexpected configuration"
+#endif
+
+#if defined(GP_OS_android)
+extern "C" MOZ_EXPORT __attribute__((weak)) int dl_iterate_phdr(
+    int (*callback)(struct dl_phdr_info* info, size_t size, void* data),
+    void* data);
+#endif
+
+#if defined(GP_OS_freebsd) && !defined(ElfW)
+#  define ElfW(type) Elf_##type
+#endif
+
+// ----------------------------------------------------------------------------
+// Starting imports from toolkit/crashreporter/google-breakpad/, as needed by
+// this file when moved to mozglue.
+
+// Imported from
+// toolkit/crashreporter/google-breakpad/src/common/memory_range.h.
+// A lightweight wrapper with a pointer and a length to encapsulate a contiguous
+// range of memory. It provides helper methods for checked access of a subrange
+// of the memory. Its implemementation does not allocate memory or call into
+// libc functions, and is thus safer to use in a crashed environment.
+class MemoryRange {
+ public:
+  MemoryRange() : data_(NULL), length_(0) {}
+
+  MemoryRange(const void* data, size_t length) { Set(data, length); }
+
+  // Returns true if this memory range contains no data.
+  bool IsEmpty() const {
+    // Set() guarantees that |length_| is zero if |data_| is NULL.
+    return length_ == 0;
+  }
+
+  // Resets to an empty range.
+  void Reset() {
+    data_ = NULL;
+    length_ = 0;
+  }
+
+  // Sets this memory range to point to |data| and its length to |length|.
+  void Set(const void* data, size_t length) {
+    data_ = reinterpret_cast<const uint8_t*>(data);
+    // Always set |length_| to zero if |data_| is NULL.
+    length_ = data ? length : 0;
+  }
+
+  // Returns true if this range covers a subrange of |sub_length| bytes
+  // at |sub_offset| bytes of this memory range, or false otherwise.
+  bool Covers(size_t sub_offset, size_t sub_length) const {
+    // The following checks verify that:
+    // 1. sub_offset is within [ 0 .. length_ - 1 ]
+    // 2. sub_offset + sub_length is within
+    //    [ sub_offset .. length_ ]
+    return sub_offset < length_ && sub_offset + sub_length >= sub_offset &&
+           sub_offset + sub_length <= length_;
+  }
+
+  // Returns a raw data pointer to a subrange of |sub_length| bytes at
+  // |sub_offset| bytes of this memory range, or NULL if the subrange
+  // is out of bounds.
+  const void* GetData(size_t sub_offset, size_t sub_length) const {
+    return Covers(sub_offset, sub_length) ? (data_ + sub_offset) : NULL;
+  }
+
+  // Same as the two-argument version of GetData() but uses sizeof(DataType)
+  // as the subrange length and returns an |DataType| pointer for convenience.
+  template <typename DataType>
+  const DataType* GetData(size_t sub_offset) const {
+    return reinterpret_cast<const DataType*>(
+        GetData(sub_offset, sizeof(DataType)));
+  }
+
+  // Returns a raw pointer to the |element_index|-th element of an array
+  // of elements of length |element_size| starting at |sub_offset| bytes
+  // of this memory range, or NULL if the element is out of bounds.
+  const void* GetArrayElement(size_t element_offset, size_t element_size,
+                              unsigned element_index) const {
+    size_t sub_offset = element_offset + element_index * element_size;
+    return GetData(sub_offset, element_size);
+  }
+
+  // Same as the three-argument version of GetArrayElement() but deduces
+  // the element size using sizeof(ElementType) and returns an |ElementType|
+  // pointer for convenience.
+  template <typename ElementType>
+  const ElementType* GetArrayElement(size_t element_offset,
+                                     unsigned element_index) const {
+    return reinterpret_cast<const ElementType*>(
+        GetArrayElement(element_offset, sizeof(ElementType), element_index));
+  }
+
+  // Returns a subrange of |sub_length| bytes at |sub_offset| bytes of
+  // this memory range, or an empty range if the subrange is out of bounds.
+  MemoryRange Subrange(size_t sub_offset, size_t sub_length) const {
+    return Covers(sub_offset, sub_length)
+               ? MemoryRange(data_ + sub_offset, sub_length)
+               : MemoryRange();
+  }
+
+  // Returns a pointer to the beginning of this memory range.
+  const uint8_t* data() const { return data_; }
+
+  // Returns the length, in bytes, of this memory range.
+  size_t length() const { return length_; }
+
+ private:
+  // Pointer to the beginning of this memory range.
+  const uint8_t* data_;
+
+  // Length, in bytes, of this memory range.
+  size_t length_;
+};
+
+// Imported from
+// toolkit/crashreporter/google-breakpad/src/common/linux/memory_mapped_file.h
+// and inlined .cc.
+// A utility class for mapping a file into memory for read-only access of the
+// file content. Its implementation avoids calling into libc functions by
+// directly making system calls for open, close, mmap, and munmap.
+class MemoryMappedFile {
+ public:
+  MemoryMappedFile() {}
+
+  // Constructor that calls Map() to map a file at |path| into memory.
+  // If Map() fails, the object behaves as if it is default constructed.
+  MemoryMappedFile(const char* path, size_t offset) { Map(path, offset); }
+
+  MemoryMappedFile(const MemoryMappedFile&) = delete;
+  MemoryMappedFile& operator=(const MemoryMappedFile&) = delete;
+
+  ~MemoryMappedFile() {}
+
+  // Maps a file at |path| into memory, which can then be accessed via
+  // content() as a MemoryRange object or via data(), and returns true on
+  // success. Mapping an empty file will succeed but with data() and size()
+  // returning NULL and 0, respectively. An existing mapping is unmapped
+  // before a new mapping is created.
+  bool Map(const char* path, size_t offset) {
+    Unmap();
+
+    int fd = open(path, O_RDONLY, 0);
+    if (fd == -1) {
+      return false;
+    }
+
+#if defined(__x86_64__) || defined(__aarch64__) || \
+    (defined(__mips__) && _MIPS_SIM == _ABI64) ||  \
+    !(defined(GP_OS_linux) || defined(GP_OS_android))
+
+    struct stat st;
+    if (fstat(fd, &st) == -1 || st.st_size < 0) {
+#else
+    struct stat64 st;
+    if (fstat64(fd, &st) == -1 || st.st_size < 0) {
+#endif
+      close(fd);
+      return false;
+    }
+
+    // Strangely file size can be negative, but we check above that it is not.
+    size_t file_len = static_cast<size_t>(st.st_size);
+    // If the file does not extend beyond the offset, simply use an empty
+    // MemoryRange and return true. Don't bother to call mmap()
+    // even though mmap() can handle an empty file on some platforms.
+    if (offset >= file_len) {
+      close(fd);
+      return true;
+    }
+
+    void* data = mmap(NULL, file_len, PROT_READ, MAP_PRIVATE, fd, offset);
+    close(fd);
+    if (data == MAP_FAILED) {
+      return false;
+    }
+
+    content_.Set(data, file_len - offset);
+    return true;
+  }
+
+  // Unmaps the memory for the mapped file. It's a no-op if no file is
+  // mapped.
+  void Unmap() {
+    if (content_.data()) {
+      munmap(const_cast<uint8_t*>(content_.data()), content_.length());
+      content_.Set(NULL, 0);
+    }
+  }
+
+  // Returns a MemoryRange object that covers the memory for the mapped
+  // file. The MemoryRange object is empty if no file is mapped.
+  const MemoryRange& content() const { return content_; }
+
+  // Returns a pointer to the beginning of the memory for the mapped file.
+  // or NULL if no file is mapped or the mapped file is empty.
+  const void* data() const { return content_.data(); }
+
+  // Returns the size in bytes of the mapped file, or zero if no file
+  // is mapped.
+  size_t size() const { return content_.length(); }
+
+ private:
+  // Mapped file content as a MemoryRange object.
+  MemoryRange content_;
+};
+
+// Imported from
+// toolkit/crashreporter/google-breakpad/src/common/linux/file_id.h and inlined
+// .cc.
+// GNU binutils' ld defaults to 'sha1', which is 160 bits == 20 bytes,
+// so this is enough to fit that, which most binaries will use.
+// This is just a sensible default for vectors so most callers can get away with
+// stack allocation.
+static const size_t kDefaultBuildIdSize = 20;
+
+// Used in a few places for backwards-compatibility.
+typedef struct {
+  uint32_t data1;
+  uint16_t data2;
+  uint16_t data3;
+  uint8_t data4[8];
+} MDGUID; /* GUID */
+
+const size_t kMDGUIDSize = sizeof(MDGUID);
+
+class FileID {
+ public:
+  explicit FileID(const char* path) : path_(path) {}
+  ~FileID() {}
+
+  // Load the identifier for the elf file path specified in the constructor into
+  // |identifier|.
+  //
+  // The current implementation will look for a .note.gnu.build-id
+  // section and use that as the file id, otherwise it falls back to
+  // XORing the first 4096 bytes of the .text section to generate an identifier.
+  bool ElfFileIdentifier(std::vector<uint8_t>& identifier) {
+    MemoryMappedFile mapped_file(path_.c_str(), 0);
+    if (!mapped_file.data())  // Should probably check if size >= ElfW(Ehdr)?
+      return false;
+
+    return ElfFileIdentifierFromMappedFile(mapped_file.data(), identifier);
+  }
+
+  // Traits classes so consumers can write templatized code to deal
+  // with specific ELF bits.
+  struct ElfClass32 {
+    typedef Elf32_Addr Addr;
+    typedef Elf32_Ehdr Ehdr;
+    typedef Elf32_Nhdr Nhdr;
+    typedef Elf32_Phdr Phdr;
+    typedef Elf32_Shdr Shdr;
+    typedef Elf32_Half Half;
+    typedef Elf32_Off Off;
+    typedef Elf32_Sym Sym;
+    typedef Elf32_Word Word;
+
+    static const int kClass = ELFCLASS32;
+    static const uint16_t kMachine = EM_386;
+    static const size_t kAddrSize = sizeof(Elf32_Addr);
+    static constexpr const char* kMachineName = "x86";
+  };
+
+  struct ElfClass64 {
+    typedef Elf64_Addr Addr;
+    typedef Elf64_Ehdr Ehdr;
+    typedef Elf64_Nhdr Nhdr;
+    typedef Elf64_Phdr Phdr;
+    typedef Elf64_Shdr Shdr;
+    typedef Elf64_Half Half;
+    typedef Elf64_Off Off;
+    typedef Elf64_Sym Sym;
+    typedef Elf64_Word Word;
+
+    static const int kClass = ELFCLASS64;
+    static const uint16_t kMachine = EM_X86_64;
+    static const size_t kAddrSize = sizeof(Elf64_Addr);
+    static constexpr const char* kMachineName = "x86_64";
+  };
+
+  // Internal helper method, exposed for convenience for callers
+  // that already have more info.
+  template <typename ElfClass>
+  static const typename ElfClass::Shdr* FindElfSectionByName(
+      const char* name, typename ElfClass::Word section_type,
+      const typename ElfClass::Shdr* sections, const char* section_names,
+      const char* names_end, int nsection) {
+    if (!name || !sections || nsection == 0) {
+      return NULL;
+    }
+
+    int name_len = strlen(name);
+    if (name_len == 0) return NULL;
+
+    for (int i = 0; i < nsection; ++i) {
+      const char* section_name = section_names + sections[i].sh_name;
+      if (sections[i].sh_type == section_type &&
+          names_end - section_name >= name_len + 1 &&
+          strcmp(name, section_name) == 0) {
+        return sections + i;
+      }
+    }
+    return NULL;
+  }
+
+  struct ElfSegment {
+    const void* start;
+    size_t size;
+  };
+
+  // Convert an offset from an Elf header into a pointer to the mapped
+  // address in the current process. Takes an extra template parameter
+  // to specify the return type to avoid having to dynamic_cast the
+  // result.
+  template <typename ElfClass, typename T>
+  static const T* GetOffset(const typename ElfClass::Ehdr* elf_header,
+                            typename ElfClass::Off offset) {
+    return reinterpret_cast<const T*>(reinterpret_cast<uintptr_t>(elf_header) +
+                                      offset);
+  }
+
+// ELF note name and desc are 32-bits word padded.
+#define NOTE_PADDING(a) ((a + 3) & ~3)
+
+  static bool ElfClassBuildIDNoteIdentifier(const void* section, size_t length,
+                                            std::vector<uint8_t>& identifier) {
+    static_assert(sizeof(ElfClass32::Nhdr) == sizeof(ElfClass64::Nhdr),
+                  "Elf32_Nhdr and Elf64_Nhdr should be the same");
+    typedef typename ElfClass32::Nhdr Nhdr;
+
+    const void* section_end = reinterpret_cast<const char*>(section) + length;
+    const Nhdr* note_header = reinterpret_cast<const Nhdr*>(section);
+    while (reinterpret_cast<const void*>(note_header) < section_end) {
+      if (note_header->n_type == NT_GNU_BUILD_ID) break;
+      note_header = reinterpret_cast<const Nhdr*>(
+          reinterpret_cast<const char*>(note_header) + sizeof(Nhdr) +
+          NOTE_PADDING(note_header->n_namesz) +
+          NOTE_PADDING(note_header->n_descsz));
+    }
+    if (reinterpret_cast<const void*>(note_header) >= section_end ||
+        note_header->n_descsz == 0) {
+      return false;
+    }
+
+    const uint8_t* build_id = reinterpret_cast<const uint8_t*>(note_header) +
+                              sizeof(Nhdr) +
+                              NOTE_PADDING(note_header->n_namesz);
+    identifier.insert(identifier.end(), build_id,
+                      build_id + note_header->n_descsz);
+
+    return true;
+  }
+
+  template <typename ElfClass>
+  static bool FindElfClassSection(const char* elf_base,
+                                  const char* section_name,
+                                  typename ElfClass::Word section_type,
+                                  const void** section_start,
+                                  size_t* section_size) {
+    typedef typename ElfClass::Ehdr Ehdr;
+    typedef typename ElfClass::Shdr Shdr;
+
+    if (!elf_base || !section_start || !section_size) {
+      return false;
+    }
+
+    if (strncmp(elf_base, ELFMAG, SELFMAG) != 0) {
+      return false;
+    }
+
+    const Ehdr* elf_header = reinterpret_cast<const Ehdr*>(elf_base);
+    if (elf_header->e_ident[EI_CLASS] != ElfClass::kClass) {
+      return false;
+    }
+
+    const Shdr* sections =
+        GetOffset<ElfClass, Shdr>(elf_header, elf_header->e_shoff);
+    const Shdr* section_names = sections + elf_header->e_shstrndx;
+    const char* names =
+        GetOffset<ElfClass, char>(elf_header, section_names->sh_offset);
+    const char* names_end = names + section_names->sh_size;
+
+    const Shdr* section =
+        FindElfSectionByName<ElfClass>(section_name, section_type, sections,
+                                       names, names_end, elf_header->e_shnum);
+
+    if (section != NULL && section->sh_size > 0) {
+      *section_start = elf_base + section->sh_offset;
+      *section_size = section->sh_size;
+    }
+
+    return true;
+  }
+
+  template <typename ElfClass>
+  static bool FindElfClassSegment(const char* elf_base,
+                                  typename ElfClass::Word segment_type,
+                                  std::vector<ElfSegment>* segments) {
+    typedef typename ElfClass::Ehdr Ehdr;
+    typedef typename ElfClass::Phdr Phdr;
+
+    if (!elf_base || !segments) {
+      return false;
+    }
+
+    if (strncmp(elf_base, ELFMAG, SELFMAG) != 0) {
+      return false;
+    }
+
+    const Ehdr* elf_header = reinterpret_cast<const Ehdr*>(elf_base);
+    if (elf_header->e_ident[EI_CLASS] != ElfClass::kClass) {
+      return false;
+    }
+
+    const Phdr* phdrs =
+        GetOffset<ElfClass, Phdr>(elf_header, elf_header->e_phoff);
+
+    for (int i = 0; i < elf_header->e_phnum; ++i) {
+      if (phdrs[i].p_type == segment_type) {
+        ElfSegment seg = {};
+        seg.start = elf_base + phdrs[i].p_offset;
+        seg.size = phdrs[i].p_filesz;
+        segments->push_back(seg);
+      }
+    }
+
+    return true;
+  }
+
+  static bool IsValidElf(const void* elf_base) {
+    return strncmp(reinterpret_cast<const char*>(elf_base), ELFMAG, SELFMAG) ==
+           0;
+  }
+
+  static int ElfClass(const void* elf_base) {
+    const ElfW(Ehdr)* elf_header =
+        reinterpret_cast<const ElfW(Ehdr)*>(elf_base);
+
+    return elf_header->e_ident[EI_CLASS];
+  }
+
+  static bool FindElfSection(const void* elf_mapped_base,
+                             const char* section_name, uint32_t section_type,
+                             const void** section_start, size_t* section_size) {
+    if (!elf_mapped_base || !section_start || !section_size) {
+      return false;
+    }
+
+    *section_start = NULL;
+    *section_size = 0;
+
+    if (!IsValidElf(elf_mapped_base)) return false;
+
+    int cls = ElfClass(elf_mapped_base);
+    const char* elf_base = static_cast<const char*>(elf_mapped_base);
+
+    if (cls == ELFCLASS32) {
+      return FindElfClassSection<ElfClass32>(elf_base, section_name,
+                                             section_type, section_start,
+                                             section_size) &&
+             *section_start != NULL;
+    } else if (cls == ELFCLASS64) {
+      return FindElfClassSection<ElfClass64>(elf_base, section_name,
+                                             section_type, section_start,
+                                             section_size) &&
+             *section_start != NULL;
+    }
+
+    return false;
+  }
+
+  static bool FindElfSegments(const void* elf_mapped_base,
+                              uint32_t segment_type,
+                              std::vector<ElfSegment>* segments) {
+    if (!elf_mapped_base || !segments) {
+      return false;
+    }
+
+    if (!IsValidElf(elf_mapped_base)) return false;
+
+    int cls = ElfClass(elf_mapped_base);
+    const char* elf_base = static_cast<const char*>(elf_mapped_base);
+
+    if (cls == ELFCLASS32) {
+      return FindElfClassSegment<ElfClass32>(elf_base, segment_type, segments);
+    } else if (cls == ELFCLASS64) {
+      return FindElfClassSegment<ElfClass64>(elf_base, segment_type, segments);
+    }
+
+    return false;
+  }
+
+  // Attempt to locate a .note.gnu.build-id section in an ELF binary
+  // and copy it into |identifier|.
+  static bool FindElfBuildIDNote(const void* elf_mapped_base,
+                                 std::vector<uint8_t>& identifier) {
+    // lld normally creates 2 PT_NOTEs, gold normally creates 1.
+    std::vector<ElfSegment> segs;
+    if (FindElfSegments(elf_mapped_base, PT_NOTE, &segs)) {
+      for (ElfSegment& seg : segs) {
+        if (ElfClassBuildIDNoteIdentifier(seg.start, seg.size, identifier)) {
+          return true;
+        }
+      }
+    }
+
+    void* note_section;
+    size_t note_size;
+    if (FindElfSection(elf_mapped_base, ".note.gnu.build-id", SHT_NOTE,
+                       (const void**)&note_section, &note_size)) {
+      return ElfClassBuildIDNoteIdentifier(note_section, note_size, identifier);
+    }
+
+    return false;
+  }
+
+  // Attempt to locate the .text section of an ELF binary and generate
+  // a simple hash by XORing the first page worth of bytes into |identifier|.
+  static bool HashElfTextSection(const void* elf_mapped_base,
+                                 std::vector<uint8_t>& identifier) {
+    identifier.resize(kMDGUIDSize);
+
+    void* text_section;
+    size_t text_size;
+    if (!FindElfSection(elf_mapped_base, ".text", SHT_PROGBITS,
+                        (const void**)&text_section, &text_size) ||
+        text_size == 0) {
+      return false;
+    }
+
+    // Only provide |kMDGUIDSize| bytes to keep identifiers produced by this
+    // function backwards-compatible.
+    memset(&identifier[0], 0, kMDGUIDSize);
+    const uint8_t* ptr = reinterpret_cast<const uint8_t*>(text_section);
+    const uint8_t* ptr_end =
+        ptr + std::min(text_size, static_cast<size_t>(4096));
+    while (ptr < ptr_end) {
+      for (unsigned i = 0; i < kMDGUIDSize; i++) identifier[i] ^= ptr[i];
+      ptr += kMDGUIDSize;
+    }
+    return true;
+  }
+
+  // Load the identifier for the elf file mapped into memory at |base| into
+  // |identifier|. Return false if the identifier could not be created for this
+  // file.
+  static bool ElfFileIdentifierFromMappedFile(
+      const void* base, std::vector<uint8_t>& identifier) {
+    // Look for a build id note first.
+    if (FindElfBuildIDNote(base, identifier)) return true;
+
+    // Fall back on hashing the first page of the text section.
+    return HashElfTextSection(base, identifier);
+  }
+
+  // These three functions are not ever called in an unsafe context, so it's OK
+  // to allocate memory and use libc.
+  static std::string bytes_to_hex_string(const uint8_t* bytes, size_t count) {
+    std::string result;
+    for (unsigned int idx = 0; idx < count; ++idx) {
+      char buf[3];
+      SprintfLiteral(buf, "%02X", bytes[idx]);
+      result.append(buf);
+    }
+    return result;
+  }
+
+  // Convert the |identifier| data to a string.  The string will
+  // be formatted as a UUID in all uppercase without dashes.
+  // (e.g., 22F065BBFC9C49F780FE26A7CEBD7BCE).
+  static std::string ConvertIdentifierToUUIDString(
+      const std::vector<uint8_t>& identifier) {
+    uint8_t identifier_swapped[kMDGUIDSize] = {0};
+
+    // Endian-ness swap to match dump processor expectation.
+    memcpy(identifier_swapped, &identifier[0],
+           std::min(kMDGUIDSize, identifier.size()));
+    uint32_t* data1 = reinterpret_cast<uint32_t*>(identifier_swapped);
+    *data1 = htonl(*data1);
+    uint16_t* data2 = reinterpret_cast<uint16_t*>(identifier_swapped + 4);
+    *data2 = htons(*data2);
+    uint16_t* data3 = reinterpret_cast<uint16_t*>(identifier_swapped + 6);
+    *data3 = htons(*data3);
+
+    return bytes_to_hex_string(identifier_swapped, kMDGUIDSize);
+  }
+
+  // Convert the entire |identifier| data to a hex string.
+  static std::string ConvertIdentifierToString(
+      const std::vector<uint8_t>& identifier) {
+    return bytes_to_hex_string(&identifier[0], identifier.size());
+  }
+
+ private:
+  // Storage for the path specified
+  std::string path_;
+};
+
+// End of imports from toolkit/crashreporter/google-breakpad/.
+// ----------------------------------------------------------------------------
+
+struct LoadedLibraryInfo {
+  LoadedLibraryInfo(const char* aName, unsigned long aBaseAddress,
+                    unsigned long aFirstMappingStart,
+                    unsigned long aLastMappingEnd)
+      : mName(aName),
+        mBaseAddress(aBaseAddress),
+        mFirstMappingStart(aFirstMappingStart),
+        mLastMappingEnd(aLastMappingEnd) {}
+
+  std::string mName;
+  unsigned long mBaseAddress;
+  unsigned long mFirstMappingStart;
+  unsigned long mLastMappingEnd;
+};
+
+static std::string IDtoUUIDString(const std::vector<uint8_t>& aIdentifier) {
+  std::string uuid = FileID::ConvertIdentifierToUUIDString(aIdentifier);
+  // This is '0', not '\0', since it represents the breakpad id age.
+  uuid += '0';
+  return uuid;
+}
+
+// Get the breakpad Id for the binary file pointed by bin_name
+static std::string getId(const char* bin_name) {
+  std::vector<uint8_t> identifier;
+  identifier.reserve(kDefaultBuildIdSize);
+
+  FileID file_id(bin_name);
+  if (file_id.ElfFileIdentifier(identifier)) {
+    return IDtoUUIDString(identifier);
+  }
+
+  return {};
+}
+
+static SharedLibrary SharedLibraryAtPath(const char* path,
+                                         unsigned long libStart,
+                                         unsigned long libEnd,
+                                         unsigned long offset = 0) {
+  std::string pathStr = path;
+
+  size_t pos = pathStr.rfind('\\');
+  std::string nameStr =
+      (pos != std::string::npos) ? pathStr.substr(pos + 1) : pathStr;
+
+  return SharedLibrary(libStart, libEnd, offset, getId(path), nameStr, pathStr,
+                       nameStr, pathStr, std::string{}, "");
+}
+
+static int dl_iterate_callback(struct dl_phdr_info* dl_info, size_t size,
+                               void* data) {
+  auto libInfoList = reinterpret_cast<std::vector<LoadedLibraryInfo>*>(data);
+
+  if (dl_info->dlpi_phnum <= 0) return 0;
+
+  unsigned long baseAddress = dl_info->dlpi_addr;
+  unsigned long firstMappingStart = -1;
+  unsigned long lastMappingEnd = 0;
+
+  for (size_t i = 0; i < dl_info->dlpi_phnum; i++) {
+    if (dl_info->dlpi_phdr[i].p_type != PT_LOAD) {
+      continue;
+    }
+    unsigned long start = dl_info->dlpi_addr + dl_info->dlpi_phdr[i].p_vaddr;
+    unsigned long end = start + dl_info->dlpi_phdr[i].p_memsz;
+    if (start < firstMappingStart) {
+      firstMappingStart = start;
+    }
+    if (end > lastMappingEnd) {
+      lastMappingEnd = end;
+    }
+  }
+
+  libInfoList->push_back(LoadedLibraryInfo(dl_info->dlpi_name, baseAddress,
+                                           firstMappingStart, lastMappingEnd));
+
+  return 0;
+}
+
+SharedLibraryInfo SharedLibraryInfo::GetInfoForSelf() {
+  SharedLibraryInfo info;
+
+#if defined(GP_OS_linux)
+  // We need to find the name of the executable (exeName, exeNameLen) and the
+  // address of its executable section (exeExeAddr) in the running image.
+  char exeName[PATH_MAX];
+  memset(exeName, 0, sizeof(exeName));
+
+  ssize_t exeNameLen = readlink("/proc/self/exe", exeName, sizeof(exeName) - 1);
+  if (exeNameLen == -1) {
+    // readlink failed for whatever reason.  Note this, but keep going.
+    exeName[0] = '\0';
+    exeNameLen = 0;
+    // LOG("SharedLibraryInfo::GetInfoForSelf(): readlink failed");
+  } else {
+    // Assert no buffer overflow.
+    MOZ_RELEASE_ASSERT(exeNameLen >= 0 &&
+                       exeNameLen < static_cast<ssize_t>(sizeof(exeName)));
+  }
+
+  unsigned long exeExeAddr = 0;
+#endif
+
+#if defined(GP_OS_android)
+  // If dl_iterate_phdr doesn't exist, we give up immediately.
+  if (!dl_iterate_phdr) {
+    // On ARM Android, dl_iterate_phdr is provided by the custom linker.
+    // So if libxul was loaded by the system linker (e.g. as part of
+    // xpcshell when running tests), it won't be available and we should
+    // not call it.
+    return info;
+  }
+#endif
+
+#if defined(GP_OS_linux) || defined(GP_OS_android)
+  // Read info from /proc/self/maps. We ignore most of it.
+  pid_t pid = mozilla::baseprofiler::profiler_current_process_id();
+  char path[PATH_MAX];
+  SprintfLiteral(path, "/proc/%d/maps", pid);
+  std::ifstream maps(path);
+  std::string line;
+  while (std::getline(maps, line)) {
+    int ret;
+    unsigned long start;
+    unsigned long end;
+    char perm[6 + 1] = "";
+    unsigned long offset;
+    char modulePath[PATH_MAX + 1] = "";
+    ret = sscanf(line.c_str(),
+                 "%lx-%lx %6s %lx %*s %*x %" PATH_MAX_STRING(PATH_MAX) "s\n",
+                 &start, &end, perm, &offset, modulePath);
+    if (!strchr(perm, 'x')) {
+      // Ignore non executable entries
+      continue;
+    }
+    if (ret != 5 && ret != 4) {
+      // LOG("SharedLibraryInfo::GetInfoForSelf(): "
+      //     "reading /proc/self/maps failed");
+      continue;
+    }
+
+#  if defined(GP_OS_linux)
+    // Try to establish the main executable's load address.
+    if (exeNameLen > 0 && strcmp(modulePath, exeName) == 0) {
+      exeExeAddr = start;
+    }
+#  elif defined(GP_OS_android)
+    // Use /proc/pid/maps to get the dalvik-jit section since it has no
+    // associated phdrs.
+    if (0 == strcmp(modulePath, "/dev/ashmem/dalvik-jit-code-cache")) {
+      info.AddSharedLibrary(
+          SharedLibraryAtPath(modulePath, start, end, offset));
+      if (info.GetSize() > 10000) {
+        // LOG("SharedLibraryInfo::GetInfoForSelf(): "
+        //     "implausibly large number of mappings acquired");
+        break;
+      }
+    }
+#  endif
+  }
+#endif
+
+  std::vector<LoadedLibraryInfo> libInfoList;
+
+  // We collect the bulk of the library info using dl_iterate_phdr.
+  dl_iterate_phdr(dl_iterate_callback, &libInfoList);
+
+  for (const auto& libInfo : libInfoList) {
+    info.AddSharedLibrary(
+        SharedLibraryAtPath(libInfo.mName.c_str(), libInfo.mFirstMappingStart,
+                            libInfo.mLastMappingEnd,
+                            libInfo.mFirstMappingStart - libInfo.mBaseAddress));
+  }
+
+#if defined(GP_OS_linux)
+  // Make another pass over the information we just harvested from
+  // dl_iterate_phdr.  If we see a nameless object mapped at what we earlier
+  // established to be the main executable's load address, attach the
+  // executable's name to that entry.
+  for (size_t i = 0; i < info.GetSize(); i++) {
+    SharedLibrary& lib = info.GetMutableEntry(i);
+    if (lib.GetStart() <= exeExeAddr && exeExeAddr <= lib.GetEnd() &&
+        lib.GetDebugPath().empty()) {
+      lib = SharedLibraryAtPath(exeName, lib.GetStart(), lib.GetEnd(),
+                                lib.GetOffset());
+
+      // We only expect to see one such entry.
+      break;
+    }
+  }
+#endif
+
+  return info;
+}
+
+void SharedLibraryInfo::Initialize() { /* do nothing */
+}
diff --git a/mozglue/baseprofiler/core/shared-libraries-macos.cc b/mozglue/baseprofiler/core/shared-libraries-macos.cc
new file mode 100644
index 0000000000..13e66f9f26
--- /dev/null
+++ b/mozglue/baseprofiler/core/shared-libraries-macos.cc
@@ -0,0 +1,182 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* 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 "BaseProfilerSharedLibraries.h"
+
+#include "platform.h"
+
+#include "mozilla/Unused.h"
+#include <AvailabilityMacros.h>
+
+#include <dlfcn.h>
+#include <mach-o/arch.h>
+#include <mach-o/dyld_images.h>
+#include <mach-o/dyld.h>
+#include <mach-o/loader.h>
+#include <mach/mach_init.h>
+#include <mach/mach_traps.h>
+#include <mach/task_info.h>
+#include <mach/task.h>
+#include <sstream>
+#include <stdlib.h>
+#include <string.h>
+#include <vector>
+
+// Architecture specific abstraction.
+#if defined(GP_ARCH_x86)
+typedef mach_header platform_mach_header;
+typedef segment_command mach_segment_command_type;
+#  define MACHO_MAGIC_NUMBER MH_MAGIC
+#  define CMD_SEGMENT LC_SEGMENT
+#  define seg_size uint32_t
+#else
+typedef mach_header_64 platform_mach_header;
+typedef segment_command_64 mach_segment_command_type;
+#  define MACHO_MAGIC_NUMBER MH_MAGIC_64
+#  define CMD_SEGMENT LC_SEGMENT_64
+#  define seg_size uint64_t
+#endif
+
+struct NativeSharedLibrary {
+  const platform_mach_header* header;
+  std::string path;
+};
+static std::vector<NativeSharedLibrary>* sSharedLibrariesList = nullptr;
+
+class MOZ_RAII SharedLibrariesLock {
+ public:
+  SharedLibrariesLock() { sSharedLibrariesMutex.Lock(); }
+
+  ~SharedLibrariesLock() { sSharedLibrariesMutex.Unlock(); }
+
+  SharedLibrariesLock(const SharedLibrariesLock&) = delete;
+  void operator=(const SharedLibrariesLock&) = delete;
+
+ private:
+  static mozilla::baseprofiler::detail::BaseProfilerMutex sSharedLibrariesMutex;
+};
+
+mozilla::baseprofiler::detail::BaseProfilerMutex
+    SharedLibrariesLock::sSharedLibrariesMutex;
+
+static void SharedLibraryAddImage(const struct mach_header* mh,
+                                  intptr_t vmaddr_slide) {
+  // NOTE: Presumably for backwards-compatibility reasons, this function accepts
+  // a mach_header even on 64-bit where it ought to be a mach_header_64. We cast
+  // it to the right type here.
+  auto header = reinterpret_cast<const platform_mach_header*>(mh);
+
+  Dl_info info;
+  if (!dladdr(header, &info)) {
+    return;
+  }
+
+  SharedLibrariesLock lock;
+  if (!sSharedLibrariesList) {
+    return;
+  }
+
+  NativeSharedLibrary lib = {header, info.dli_fname};
+  sSharedLibrariesList->push_back(lib);
+}
+
+static void SharedLibraryRemoveImage(const struct mach_header* mh,
+                                     intptr_t vmaddr_slide) {
+  // NOTE: Presumably for backwards-compatibility reasons, this function accepts
+  // a mach_header even on 64-bit where it ought to be a mach_header_64. We cast
+  // it to the right type here.
+  auto header = reinterpret_cast<const platform_mach_header*>(mh);
+
+  SharedLibrariesLock lock;
+  if (!sSharedLibrariesList) {
+    return;
+  }
+
+  uint32_t count = sSharedLibrariesList->size();
+  for (uint32_t i = 0; i < count; ++i) {
+    if ((*sSharedLibrariesList)[i].header == header) {
+      sSharedLibrariesList->erase(sSharedLibrariesList->begin() + i);
+      return;
+    }
+  }
+}
+
+void SharedLibraryInfo::Initialize() {
+  // NOTE: We intentionally leak this memory here. We're allocating dynamically
+  // in order to avoid static initializers.
+  sSharedLibrariesList = new std::vector<NativeSharedLibrary>();
+
+  _dyld_register_func_for_add_image(SharedLibraryAddImage);
+  _dyld_register_func_for_remove_image(SharedLibraryRemoveImage);
+}
+
+static void addSharedLibrary(const platform_mach_header* header,
+                             const char* path, SharedLibraryInfo& info) {
+  const struct load_command* cmd =
+      reinterpret_cast<const struct load_command*>(header + 1);
+
+  seg_size size = 0;
+  unsigned long long start = reinterpret_cast<unsigned long long>(header);
+  // Find the cmd segment in the macho image. It will contain the offset we care
+  // about.
+  const uint8_t* uuid_bytes = nullptr;
+  for (unsigned int i = 0;
+       cmd && (i < header->ncmds) && (uuid_bytes == nullptr || size == 0);
+       ++i) {
+    if (cmd->cmd == CMD_SEGMENT) {
+      const mach_segment_command_type* seg =
+          reinterpret_cast<const mach_segment_command_type*>(cmd);
+
+      if (!strcmp(seg->segname, "__TEXT")) {
+        size = seg->vmsize;
+      }
+    } else if (cmd->cmd == LC_UUID) {
+      const uuid_command* ucmd = reinterpret_cast<const uuid_command*>(cmd);
+      uuid_bytes = ucmd->uuid;
+    }
+
+    cmd = reinterpret_cast<const struct load_command*>(
+        reinterpret_cast<const char*>(cmd) + cmd->cmdsize);
+  }
+
+  std::string uuid;
+  if (uuid_bytes != nullptr) {
+    static constexpr char digits[16] = {'0', '1', '2', '3', '4', '5', '6', '7',
+                                        '8', '9', 'A', 'B', 'C', 'D', 'E', 'F'};
+    for (int i = 0; i < 15; ++i) {
+      uint8_t byte = uuid_bytes[i];
+      uuid += digits[byte >> 4];
+      uuid += digits[byte & 0xFu];
+    }
+    // breakpad id age.
+    uuid += '0';
+  }
+
+  std::string pathStr = path;
+
+  size_t pos = pathStr.rfind('\\');
+  std::string nameStr =
+      (pos != std::string::npos) ? pathStr.substr(pos + 1) : pathStr;
+
+  const NXArchInfo* archInfo =
+      NXGetArchInfoFromCpuType(header->cputype, header->cpusubtype);
+
+  info.AddSharedLibrary(SharedLibrary(start, start + size, 0, uuid, nameStr,
+                                      pathStr, nameStr, pathStr, std::string{},
+                                      archInfo ? archInfo->name : ""));
+}
+
+// Translate the statically stored sSharedLibrariesList information into a
+// SharedLibraryInfo object.
+SharedLibraryInfo SharedLibraryInfo::GetInfoForSelf() {
+  SharedLibrariesLock lock;
+  SharedLibraryInfo sharedLibraryInfo;
+
+  for (auto& info : *sSharedLibrariesList) {
+    addSharedLibrary(info.header, info.path.c_str(), sharedLibraryInfo);
+  }
+
+  return sharedLibraryInfo;
+}
diff --git a/mozglue/baseprofiler/core/shared-libraries-win32.cc b/mozglue/baseprofiler/core/shared-libraries-win32.cc
new file mode 100644
index 0000000000..5bf7408193
--- /dev/null
+++ b/mozglue/baseprofiler/core/shared-libraries-win32.cc
@@ -0,0 +1,277 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* 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 <windows.h>
+#include <dbghelp.h>
+#include <sstream>
+#include <psapi.h>
+
+#include "BaseProfilerSharedLibraries.h"
+
+#include "mozilla/glue/WindowsUnicode.h"
+#include "mozilla/Unused.h"
+#include "mozilla/WindowsVersion.h"
+
+#include <cctype>
+#include <string>
+
+#define CV_SIGNATURE 0x53445352  // 'SDSR'
+
+struct CodeViewRecord70 {
+  uint32_t signature;
+  GUID pdbSignature;
+  uint32_t pdbAge;
+  // A UTF-8 string, according to
+  // https://github.com/Microsoft/microsoft-pdb/blob/082c5290e5aff028ae84e43affa8be717aa7af73/PDB/dbi/locator.cpp#L785
+  char pdbFileName[1];
+};
+
+static constexpr char digits[16] = {'0', '1', '2', '3', '4', '5', '6', '7',
+                                    '8', '9', 'A', 'B', 'C', 'D', 'E', 'F'};
+
+static void AppendHex(const unsigned char* aBegin, const unsigned char* aEnd,
+                      std::string& aOut) {
+  for (const unsigned char* p = aBegin; p < aEnd; ++p) {
+    unsigned char c = *p;
+    aOut += digits[c >> 4];
+    aOut += digits[c & 0xFu];
+  }
+}
+
+static constexpr bool WITH_PADDING = true;
+static constexpr bool WITHOUT_PADDING = false;
+template <typename T>
+static void AppendHex(T aValue, std::string& aOut, bool aWithPadding) {
+  for (int i = sizeof(T) * 2 - 1; i >= 0; --i) {
+    unsigned nibble = (aValue >> (i * 4)) & 0xFu;
+    // If no-padding requested, skip starting zeroes -- unless we're on the very
+    // last nibble (so we don't output a blank).
+    if (!aWithPadding && i != 0) {
+      if (nibble == 0) {
+        // Requested no padding, skip zeroes.
+        continue;
+      }
+      // Requested no padding, got first non-zero, pretend we now want padding
+      // so we don't skip zeroes anymore.
+      aWithPadding = true;
+    }
+    aOut += digits[nibble];
+  }
+}
+
+static bool GetPdbInfo(uintptr_t aStart, std::string& aSignature,
+                       uint32_t& aAge, char** aPdbName) {
+  if (!aStart) {
+    return false;
+  }
+
+  PIMAGE_DOS_HEADER dosHeader = reinterpret_cast<PIMAGE_DOS_HEADER>(aStart);
+  if (dosHeader->e_magic != IMAGE_DOS_SIGNATURE) {
+    return false;
+  }
+
+  PIMAGE_NT_HEADERS ntHeaders =
+      reinterpret_cast<PIMAGE_NT_HEADERS>(aStart + dosHeader->e_lfanew);
+  if (ntHeaders->Signature != IMAGE_NT_SIGNATURE) {
+    return false;
+  }
+
+  uint32_t relativeVirtualAddress =
+      ntHeaders->OptionalHeader.DataDirectory[IMAGE_DIRECTORY_ENTRY_DEBUG]
+          .VirtualAddress;
+  if (!relativeVirtualAddress) {
+    return false;
+  }
+
+  PIMAGE_DEBUG_DIRECTORY debugDirectory =
+      reinterpret_cast<PIMAGE_DEBUG_DIRECTORY>(aStart + relativeVirtualAddress);
+  if (!debugDirectory || debugDirectory->Type != IMAGE_DEBUG_TYPE_CODEVIEW) {
+    return false;
+  }
+
+  CodeViewRecord70* debugInfo = reinterpret_cast<CodeViewRecord70*>(
+      aStart + debugDirectory->AddressOfRawData);
+  if (!debugInfo || debugInfo->signature != CV_SIGNATURE) {
+    return false;
+  }
+
+  aAge = debugInfo->pdbAge;
+  GUID& pdbSignature = debugInfo->pdbSignature;
+  AppendHex(pdbSignature.Data1, aSignature, WITH_PADDING);
+  AppendHex(pdbSignature.Data2, aSignature, WITH_PADDING);
+  AppendHex(pdbSignature.Data3, aSignature, WITH_PADDING);
+  AppendHex(reinterpret_cast<const unsigned char*>(&pdbSignature.Data4),
+            reinterpret_cast<const unsigned char*>(&pdbSignature.Data4) +
+                sizeof(pdbSignature.Data4),
+            aSignature);
+
+  // The PDB file name could be different from module filename, so report both
+  // e.g. The PDB for C:\Windows\SysWOW64\ntdll.dll is wntdll.pdb
+  *aPdbName = debugInfo->pdbFileName;
+
+  return true;
+}
+
+static std::string GetVersion(wchar_t* dllPath) {
+  DWORD infoSize = GetFileVersionInfoSizeW(dllPath, nullptr);
+  if (infoSize == 0) {
+    return {};
+  }
+
+  mozilla::UniquePtr<unsigned char[]> infoData =
+      mozilla::MakeUnique<unsigned char[]>(infoSize);
+  if (!GetFileVersionInfoW(dllPath, 0, infoSize, infoData.get())) {
+    return {};
+  }
+
+  VS_FIXEDFILEINFO* vInfo;
+  UINT vInfoLen;
+  if (!VerQueryValueW(infoData.get(), L"\\", (LPVOID*)&vInfo, &vInfoLen)) {
+    return {};
+  }
+  if (!vInfo) {
+    return {};
+  }
+
+  return std::to_string(vInfo->dwFileVersionMS >> 16) + '.' +
+         std::to_string(vInfo->dwFileVersionMS & 0xFFFF) + '.' +
+         std::to_string(vInfo->dwFileVersionLS >> 16) + '.' +
+         std::to_string(vInfo->dwFileVersionLS & 0xFFFF);
+}
+
+SharedLibraryInfo SharedLibraryInfo::GetInfoForSelf() {
+  SharedLibraryInfo sharedLibraryInfo;
+
+  HANDLE hProcess = GetCurrentProcess();
+  mozilla::UniquePtr<HMODULE[]> hMods;
+  size_t modulesNum = 0;
+  if (hProcess != NULL) {
+    DWORD modulesSize;
+    if (!EnumProcessModules(hProcess, nullptr, 0, &modulesSize)) {
+      return sharedLibraryInfo;
+    }
+    modulesNum = modulesSize / sizeof(HMODULE);
+    hMods = mozilla::MakeUnique<HMODULE[]>(modulesNum);
+    if (!EnumProcessModules(hProcess, hMods.get(), modulesNum * sizeof(HMODULE),
+                            &modulesSize)) {
+      return sharedLibraryInfo;
+    }
+    // The list may have shrunk between calls
+    if (modulesSize / sizeof(HMODULE) < modulesNum) {
+      modulesNum = modulesSize / sizeof(HMODULE);
+    }
+  }
+
+  for (unsigned int i = 0; i < modulesNum; i++) {
+    wchar_t modulePath[MAX_PATH + 1];
+    if (!GetModuleFileNameExW(hProcess, hMods[i], modulePath,
+                              std::size(modulePath))) {
+      continue;
+    }
+    mozilla::UniquePtr<char[]> utf8ModulePath(
+        mozilla::glue::WideToUTF8(modulePath));
+    if (!utf8ModulePath) {
+      continue;
+    }
+
+    MODULEINFO module = {0};
+    if (!GetModuleInformation(hProcess, hMods[i], &module,
+                              sizeof(MODULEINFO))) {
+      continue;
+    }
+
+    std::string modulePathStr(utf8ModulePath.get());
+    size_t pos = modulePathStr.find_last_of("\\/");
+    std::string moduleNameStr = (pos != std::string::npos)
+                                    ? modulePathStr.substr(pos + 1)
+                                    : modulePathStr;
+
+    // Hackaround for Bug 1607574.  Nvidia's shim driver nvd3d9wrap[x].dll
+    // detours LoadLibraryExW when it's loaded and the detour function causes
+    // AV when the code tries to access data pointing to an address within
+    // unloaded nvinit[x].dll.
+    // The crashing code is executed when a given parameter is "detoured.dll"
+    // and OS version is older than 6.2.  We hit that crash at the following
+    // call to LoadLibraryEx even if we specify LOAD_LIBRARY_AS_DATAFILE.
+    // We work around it by skipping LoadLibraryEx, and add a library info with
+    // a dummy breakpad id instead.
+#if !defined(_M_ARM64)
+#  if defined(_M_AMD64)
+    LPCWSTR kNvidiaShimDriver = L"nvd3d9wrapx.dll";
+    LPCWSTR kNvidiaInitDriver = L"nvinitx.dll";
+#  elif defined(_M_IX86)
+    LPCWSTR kNvidiaShimDriver = L"nvd3d9wrap.dll";
+    LPCWSTR kNvidiaInitDriver = L"nvinit.dll";
+#  endif
+    constexpr std::string_view detoured_dll = "detoured.dll";
+    if (std::equal(moduleNameStr.cbegin(), moduleNameStr.cend(),
+                   detoured_dll.cbegin(), detoured_dll.cend(),
+                   [](char aModuleChar, char aDetouredChar) {
+                     return std::tolower(aModuleChar) == aDetouredChar;
+                   }) &&
+        !mozilla::IsWin8OrLater() && ::GetModuleHandleW(kNvidiaShimDriver) &&
+        !::GetModuleHandleW(kNvidiaInitDriver)) {
+      const std::string pdbNameStr = "detoured.pdb";
+      SharedLibrary shlib((uintptr_t)module.lpBaseOfDll,
+                          (uintptr_t)module.lpBaseOfDll + module.SizeOfImage,
+                          0,  // DLLs are always mapped at offset 0 on Windows
+                          "000000000000000000000000000000000", moduleNameStr,
+                          modulePathStr, pdbNameStr, pdbNameStr, "", "");
+      sharedLibraryInfo.AddSharedLibrary(shlib);
+      continue;
+    }
+#endif  // !defined(_M_ARM64)
+
+    std::string breakpadId;
+    // Load the module again to make sure that its handle will remain
+    // valid as we attempt to read the PDB information from it.  We load the
+    // DLL as a datafile so that if the module actually gets unloaded between
+    // the call to EnumProcessModules and the following LoadLibraryEx, we
+    // don't end up running the now newly loaded module's DllMain function. If
+    // the module is already loaded, LoadLibraryEx just increments its
+    // refcount.
+    //
+    // Note that because of the race condition above, merely loading the DLL
+    // again is not safe enough, therefore we also need to make sure that we
+    // can read the memory mapped at the base address before we can safely
+    // proceed to actually access those pages.
+    HMODULE handleLock =
+        LoadLibraryExW(modulePath, NULL, LOAD_LIBRARY_AS_DATAFILE);
+    MEMORY_BASIC_INFORMATION vmemInfo = {0};
+    std::string pdbSig;
+    uint32_t pdbAge;
+    std::string pdbPathStr;
+    std::string pdbNameStr;
+    char* pdbName = nullptr;
+    if (handleLock &&
+        sizeof(vmemInfo) ==
+            VirtualQuery(module.lpBaseOfDll, &vmemInfo, sizeof(vmemInfo)) &&
+        vmemInfo.State == MEM_COMMIT &&
+        GetPdbInfo((uintptr_t)module.lpBaseOfDll, pdbSig, pdbAge, &pdbName)) {
+      MOZ_ASSERT(breakpadId.empty());
+      breakpadId += pdbSig;
+      AppendHex(pdbAge, breakpadId, WITHOUT_PADDING);
+
+      pdbPathStr = pdbName;
+      size_t pos = pdbPathStr.find_last_of("\\/");
+      pdbNameStr =
+          (pos != std::string::npos) ? pdbPathStr.substr(pos + 1) : pdbPathStr;
+    }
+
+    SharedLibrary shlib((uintptr_t)module.lpBaseOfDll,
+                        (uintptr_t)module.lpBaseOfDll + module.SizeOfImage,
+                        0,  // DLLs are always mapped at offset 0 on Windows
+                        breakpadId, moduleNameStr, modulePathStr, pdbNameStr,
+                        pdbPathStr, GetVersion(modulePath), "");
+    sharedLibraryInfo.AddSharedLibrary(shlib);
+
+    FreeLibrary(handleLock);  // ok to free null handles
+  }
+
+  return sharedLibraryInfo;
+}
+
+void SharedLibraryInfo::Initialize() { /* do nothing */
+}
diff --git a/mozglue/baseprofiler/core/vtune/ittnotify.h b/mozglue/baseprofiler/core/vtune/ittnotify.h
new file mode 100644
index 0000000000..04adf9eb5e
--- /dev/null
+++ b/mozglue/baseprofiler/core/vtune/ittnotify.h
@@ -0,0 +1,4127 @@
+// clang-format off
+
+/* <copyright>
+  This file is provided under a dual BSD/GPLv2 license.  When using or
+  redistributing this file, you may do so under either license.
+
+  GPL LICENSE SUMMARY
+
+  Copyright (c) 2005-2014 Intel Corporation. All rights reserved.
+
+  This program is free software; you can redistribute it and/or modify
+  it under the terms of version 2 of the GNU General Public License as
+  published by the Free Software Foundation.
+
+  This program is distributed in the hope that it will be useful, but
+  WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+  General Public License for more details.
+
+  You should have received a copy of the GNU General Public License
+  along with this program; if not, write to the Free Software
+  Foundation, Inc., 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
+  The full GNU General Public License is included in this distribution
+  in the file called LICENSE.GPL.
+
+  Contact Information:
+  http://software.intel.com/en-us/articles/intel-vtune-amplifier-xe/
+
+  BSD LICENSE
+
+  Copyright (c) 2005-2014 Intel Corporation. All rights reserved.
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions
+  are met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in
+      the documentation and/or other materials provided with the
+      distribution.
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived
+      from this software without specific prior written permission.
+
+  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+  "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
+  OWNER OR CONTRIBUTORS 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.
+</copyright> */
+#ifndef _ITTNOTIFY_H_
+#  define _ITTNOTIFY_H_
+
+/**
+@file
+@brief Public User API functions and types
+@mainpage
+
+The ITT API is used to annotate a user's program with additional information
+that can be used by correctness and performance tools. The user inserts
+calls in their program. Those calls generate information that is collected
+at runtime, and used by Intel(R) Threading Tools.
+
+@section API Concepts
+The following general concepts are used throughout the API.
+
+@subsection Unicode Support
+Many API functions take character string arguments. On Windows, there
+are two versions of each such function. The function name is suffixed
+by W if Unicode support is enabled, and by A otherwise. Any API function
+that takes a character string argument adheres to this convention.
+
+@subsection Conditional Compilation
+Many users prefer having an option to modify ITT API code when linking it
+inside their runtimes. ITT API header file provides a mechanism to replace
+ITT API function names inside your code with empty strings. To do this,
+define the macros INTEL_NO_ITTNOTIFY_API during compilation and remove the
+static library from the linker script.
+
+@subsection Domains
+[see domains]
+Domains provide a way to separate notification for different modules or
+libraries in a program. Domains are specified by dotted character strings,
+e.g. TBB.Internal.Control.
+
+A mechanism (to be specified) is provided to enable and disable
+domains. By default, all domains are enabled.
+@subsection Named Entities and Instances
+Named entities (frames, regions, tasks, and markers) communicate
+information about the program to the analysis tools. A named entity often
+refers to a section of program code, or to some set of logical concepts
+that the programmer wants to group together.
+
+Named entities relate to the programmer's static view of the program. When
+the program actually executes, many instances of a given named entity
+may be created.
+
+The API annotations denote instances of named entities. The actual
+named entities are displayed using the analysis tools. In other words,
+the named entities come into existence when instances are created.
+
+Instances of named entities may have instance identifiers (IDs). Some
+API calls use instance identifiers to create relationships between
+different instances of named entities. Other API calls associate data
+with instances of named entities.
+
+Some named entities must always have instance IDs. In particular, regions
+and frames always have IDs. Task and markers need IDs only if the ID is
+needed in another API call (such as adding a relation or metadata).
+
+The lifetime of instance IDs is distinct from the lifetime of
+instances. This allows various relationships to be specified separate
+from the actual execution of instances. This flexibility comes at the
+expense of extra API calls.
+
+The same ID may not be reused for different instances, unless a previous
+[ref] __itt_id_destroy call for that ID has been issued.
+*/
+
+/** @cond exclude_from_documentation */
+#ifndef ITT_OS_WIN
+#  define ITT_OS_WIN   1
+#endif /* ITT_OS_WIN */
+
+#ifndef ITT_OS_LINUX
+#  define ITT_OS_LINUX 2
+#endif /* ITT_OS_LINUX */
+
+#ifndef ITT_OS_MAC
+#  define ITT_OS_MAC   3
+#endif /* ITT_OS_MAC */
+
+#ifndef ITT_OS_FREEBSD
+#  define ITT_OS_FREEBSD   4
+#endif /* ITT_OS_FREEBSD */
+
+#ifndef ITT_OS
+#  if defined WIN32 || defined _WIN32
+#    define ITT_OS ITT_OS_WIN
+#  elif defined( __APPLE__ ) && defined( __MACH__ )
+#    define ITT_OS ITT_OS_MAC
+#  elif defined( __FreeBSD__ )
+#    define ITT_OS ITT_OS_FREEBSD
+#  else
+#    define ITT_OS ITT_OS_LINUX
+#  endif
+#endif /* ITT_OS */
+
+#ifndef ITT_PLATFORM_WIN
+#  define ITT_PLATFORM_WIN 1
+#endif /* ITT_PLATFORM_WIN */
+
+#ifndef ITT_PLATFORM_POSIX
+#  define ITT_PLATFORM_POSIX 2
+#endif /* ITT_PLATFORM_POSIX */
+
+#ifndef ITT_PLATFORM_MAC
+#  define ITT_PLATFORM_MAC 3
+#endif /* ITT_PLATFORM_MAC */
+
+#ifndef ITT_PLATFORM_FREEBSD
+#  define ITT_PLATFORM_FREEBSD 4
+#endif /* ITT_PLATFORM_FREEBSD */
+
+#ifndef ITT_PLATFORM
+#  if ITT_OS==ITT_OS_WIN
+#    define ITT_PLATFORM ITT_PLATFORM_WIN
+#  elif ITT_OS==ITT_OS_MAC
+#    define ITT_PLATFORM ITT_PLATFORM_MAC
+#  elif ITT_OS==ITT_OS_FREEBSD
+#    define ITT_PLATFORM ITT_PLATFORM_FREEBSD
+#  else
+#    define ITT_PLATFORM ITT_PLATFORM_POSIX
+#  endif
+#endif /* ITT_PLATFORM */
+
+#if defined(_UNICODE) && !defined(UNICODE)
+#define UNICODE
+#endif
+
+#include <stddef.h>
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#include <tchar.h>
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#include <stdint.h>
+#if defined(UNICODE) || defined(_UNICODE)
+#include <wchar.h>
+#endif /* UNICODE || _UNICODE */
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+#ifndef ITTAPI_CDECL
+#  if ITT_PLATFORM==ITT_PLATFORM_WIN
+#    define ITTAPI_CDECL __cdecl
+#  else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#    if defined _M_IX86 || defined __i386__
+#      define ITTAPI_CDECL __attribute__ ((cdecl))
+#    else  /* _M_IX86 || __i386__ */
+#      define ITTAPI_CDECL /* actual only on x86 platform */
+#    endif /* _M_IX86 || __i386__ */
+#  endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* ITTAPI_CDECL */
+
+#ifndef STDCALL
+#  if ITT_PLATFORM==ITT_PLATFORM_WIN
+#    define STDCALL __stdcall
+#  else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#    if defined _M_IX86 || defined __i386__
+#      define STDCALL __attribute__ ((stdcall))
+#    else  /* _M_IX86 || __i386__ */
+#      define STDCALL /* supported only on x86 platform */
+#    endif /* _M_IX86 || __i386__ */
+#  endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* STDCALL */
+
+#define ITTAPI    ITTAPI_CDECL
+#define LIBITTAPI ITTAPI_CDECL
+
+/* TODO: Temporary for compatibility! */
+#define ITTAPI_CALL    ITTAPI_CDECL
+#define LIBITTAPI_CALL ITTAPI_CDECL
+
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+/* use __forceinline (VC++ specific) */
+#define ITT_INLINE           __forceinline
+#define ITT_INLINE_ATTRIBUTE /* nothing */
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+/*
+ * Generally, functions are not inlined unless optimization is specified.
+ * For functions declared inline, this attribute inlines the function even
+ * if no optimization level was specified.
+ */
+#ifdef __STRICT_ANSI__
+#define ITT_INLINE           static
+#define ITT_INLINE_ATTRIBUTE __attribute__((unused))
+#else  /* __STRICT_ANSI__ */
+#define ITT_INLINE           static inline
+#define ITT_INLINE_ATTRIBUTE __attribute__((always_inline, unused))
+#endif /* __STRICT_ANSI__ */
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+/** @endcond */
+
+#ifdef INTEL_ITTNOTIFY_ENABLE_LEGACY
+#  if ITT_PLATFORM==ITT_PLATFORM_WIN
+#    pragma message("WARNING!!! Deprecated API is used. Please undefine INTEL_ITTNOTIFY_ENABLE_LEGACY macro")
+#  else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#    warning "Deprecated API is used. Please undefine INTEL_ITTNOTIFY_ENABLE_LEGACY macro"
+#  endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#  include "vtune/legacy/ittnotify.h"
+#endif /* INTEL_ITTNOTIFY_ENABLE_LEGACY */
+
+/** @cond exclude_from_documentation */
+/* Helper macro for joining tokens */
+#define ITT_JOIN_AUX(p,n) p##n
+#define ITT_JOIN(p,n)     ITT_JOIN_AUX(p,n)
+
+#ifdef ITT_MAJOR
+#undef ITT_MAJOR
+#endif
+#ifdef ITT_MINOR
+#undef ITT_MINOR
+#endif
+#define ITT_MAJOR     3
+#define ITT_MINOR     0
+
+/* Standard versioning of a token with major and minor version numbers */
+#define ITT_VERSIONIZE(x)    \
+    ITT_JOIN(x,              \
+    ITT_JOIN(_,              \
+    ITT_JOIN(ITT_MAJOR,      \
+    ITT_JOIN(_, ITT_MINOR))))
+
+#ifndef INTEL_ITTNOTIFY_PREFIX
+#  define INTEL_ITTNOTIFY_PREFIX __itt_
+#endif /* INTEL_ITTNOTIFY_PREFIX */
+#ifndef INTEL_ITTNOTIFY_POSTFIX
+#  define INTEL_ITTNOTIFY_POSTFIX _ptr_
+#endif /* INTEL_ITTNOTIFY_POSTFIX */
+
+#define ITTNOTIFY_NAME_AUX(n) ITT_JOIN(INTEL_ITTNOTIFY_PREFIX,n)
+#define ITTNOTIFY_NAME(n)     ITT_VERSIONIZE(ITTNOTIFY_NAME_AUX(ITT_JOIN(n,INTEL_ITTNOTIFY_POSTFIX)))
+
+#define ITTNOTIFY_VOID(n) (!ITTNOTIFY_NAME(n)) ? (void)0 : ITTNOTIFY_NAME(n)
+#define ITTNOTIFY_DATA(n) (!ITTNOTIFY_NAME(n)) ?       0 : ITTNOTIFY_NAME(n)
+
+#define ITTNOTIFY_VOID_D0(n,d)       (!(d)->flags) ? (void)0 : (!ITTNOTIFY_NAME(n)) ? (void)0 : ITTNOTIFY_NAME(n)(d)
+#define ITTNOTIFY_VOID_D1(n,d,x)     (!(d)->flags) ? (void)0 : (!ITTNOTIFY_NAME(n)) ? (void)0 : ITTNOTIFY_NAME(n)(d,x)
+#define ITTNOTIFY_VOID_D2(n,d,x,y)   (!(d)->flags) ? (void)0 : (!ITTNOTIFY_NAME(n)) ? (void)0 : ITTNOTIFY_NAME(n)(d,x,y)
+#define ITTNOTIFY_VOID_D3(n,d,x,y,z) (!(d)->flags) ? (void)0 : (!ITTNOTIFY_NAME(n)) ? (void)0 : ITTNOTIFY_NAME(n)(d,x,y,z)
+#define ITTNOTIFY_VOID_D4(n,d,x,y,z,a)     (!(d)->flags) ? (void)0 : (!ITTNOTIFY_NAME(n)) ? (void)0 : ITTNOTIFY_NAME(n)(d,x,y,z,a)
+#define ITTNOTIFY_VOID_D5(n,d,x,y,z,a,b)   (!(d)->flags) ? (void)0 : (!ITTNOTIFY_NAME(n)) ? (void)0 : ITTNOTIFY_NAME(n)(d,x,y,z,a,b)
+#define ITTNOTIFY_VOID_D6(n,d,x,y,z,a,b,c) (!(d)->flags) ? (void)0 : (!ITTNOTIFY_NAME(n)) ? (void)0 : ITTNOTIFY_NAME(n)(d,x,y,z,a,b,c)
+#define ITTNOTIFY_DATA_D0(n,d)       (!(d)->flags) ?       0 : (!ITTNOTIFY_NAME(n)) ?       0 : ITTNOTIFY_NAME(n)(d)
+#define ITTNOTIFY_DATA_D1(n,d,x)     (!(d)->flags) ?       0 : (!ITTNOTIFY_NAME(n)) ?       0 : ITTNOTIFY_NAME(n)(d,x)
+#define ITTNOTIFY_DATA_D2(n,d,x,y)   (!(d)->flags) ?       0 : (!ITTNOTIFY_NAME(n)) ?       0 : ITTNOTIFY_NAME(n)(d,x,y)
+#define ITTNOTIFY_DATA_D3(n,d,x,y,z) (!(d)->flags) ?       0 : (!ITTNOTIFY_NAME(n)) ?       0 : ITTNOTIFY_NAME(n)(d,x,y,z)
+#define ITTNOTIFY_DATA_D4(n,d,x,y,z,a)     (!(d)->flags) ? 0 : (!ITTNOTIFY_NAME(n)) ?       0 : ITTNOTIFY_NAME(n)(d,x,y,z,a)
+#define ITTNOTIFY_DATA_D5(n,d,x,y,z,a,b)   (!(d)->flags) ? 0 : (!ITTNOTIFY_NAME(n)) ?       0 : ITTNOTIFY_NAME(n)(d,x,y,z,a,b)
+#define ITTNOTIFY_DATA_D6(n,d,x,y,z,a,b,c) (!(d)->flags) ? 0 : (!ITTNOTIFY_NAME(n)) ?       0 : ITTNOTIFY_NAME(n)(d,x,y,z,a,b,c)
+
+#ifdef ITT_STUB
+#undef ITT_STUB
+#endif
+#ifdef ITT_STUBV
+#undef ITT_STUBV
+#endif
+#define ITT_STUBV(api,type,name,args)                             \
+    typedef type (api* ITT_JOIN(ITTNOTIFY_NAME(name),_t)) args;   \
+    extern ITT_JOIN(ITTNOTIFY_NAME(name),_t) ITTNOTIFY_NAME(name);
+#define ITT_STUB ITT_STUBV
+/** @endcond */
+
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+
+/** @cond exclude_from_gpa_documentation */
+/**
+ * @defgroup public Public API
+ * @{
+ * @}
+ */
+
+/**
+ * @defgroup control Collection Control
+ * @ingroup public
+ * General behavior: application continues to run, but no profiling information is being collected
+ *
+ * Pausing occurs not only for the current thread but for all process as well as spawned processes
+ * - Intel(R) Parallel Inspector and Intel(R) Inspector XE:
+ *   - Does not analyze or report errors that involve memory access.
+ *   - Other errors are reported as usual. Pausing data collection in
+ *     Intel(R) Parallel Inspector and Intel(R) Inspector XE
+ *     only pauses tracing and analyzing memory access.
+ *     It does not pause tracing or analyzing threading APIs.
+ *   .
+ * - Intel(R) Parallel Amplifier and Intel(R) VTune(TM) Amplifier XE:
+ *   - Does continue to record when new threads are started.
+ *   .
+ * - Other effects:
+ *   - Possible reduction of runtime overhead.
+ *   .
+ * @{
+ */
+/** @brief Pause collection */
+void ITTAPI __itt_pause(void);
+/** @brief Resume collection */
+void ITTAPI __itt_resume(void);
+/** @brief Detach collection */
+void ITTAPI __itt_detach(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, pause,  (void))
+ITT_STUBV(ITTAPI, void, resume, (void))
+ITT_STUBV(ITTAPI, void, detach, (void))
+#define __itt_pause      ITTNOTIFY_VOID(pause)
+#define __itt_pause_ptr  ITTNOTIFY_NAME(pause)
+#define __itt_resume     ITTNOTIFY_VOID(resume)
+#define __itt_resume_ptr ITTNOTIFY_NAME(resume)
+#define __itt_detach     ITTNOTIFY_VOID(detach)
+#define __itt_detach_ptr ITTNOTIFY_NAME(detach)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_pause()
+#define __itt_pause_ptr  0
+#define __itt_resume()
+#define __itt_resume_ptr 0
+#define __itt_detach()
+#define __itt_detach_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_pause_ptr  0
+#define __itt_resume_ptr 0
+#define __itt_detach_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} control group */
+/** @endcond */
+
+/**
+ * @defgroup threads Threads
+ * @ingroup public
+ * Give names to threads
+ * @{
+ */
+/**
+ * @brief Sets thread name of calling thread
+ * @param[in] name - name of thread
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+void ITTAPI __itt_thread_set_nameA(const char    *name);
+void ITTAPI __itt_thread_set_nameW(const wchar_t *name);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_thread_set_name     __itt_thread_set_nameW
+#  define __itt_thread_set_name_ptr __itt_thread_set_nameW_ptr
+#else /* UNICODE */
+#  define __itt_thread_set_name     __itt_thread_set_nameA
+#  define __itt_thread_set_name_ptr __itt_thread_set_nameA_ptr
+#endif /* UNICODE */
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+void ITTAPI __itt_thread_set_name(const char *name);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUBV(ITTAPI, void, thread_set_nameA, (const char    *name))
+ITT_STUBV(ITTAPI, void, thread_set_nameW, (const wchar_t *name))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUBV(ITTAPI, void, thread_set_name,  (const char    *name))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_thread_set_nameA     ITTNOTIFY_VOID(thread_set_nameA)
+#define __itt_thread_set_nameA_ptr ITTNOTIFY_NAME(thread_set_nameA)
+#define __itt_thread_set_nameW     ITTNOTIFY_VOID(thread_set_nameW)
+#define __itt_thread_set_nameW_ptr ITTNOTIFY_NAME(thread_set_nameW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_thread_set_name     ITTNOTIFY_VOID(thread_set_name)
+#define __itt_thread_set_name_ptr ITTNOTIFY_NAME(thread_set_name)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_thread_set_nameA(name)
+#define __itt_thread_set_nameA_ptr 0
+#define __itt_thread_set_nameW(name)
+#define __itt_thread_set_nameW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_thread_set_name(name)
+#define __itt_thread_set_name_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_thread_set_nameA_ptr 0
+#define __itt_thread_set_nameW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_thread_set_name_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @cond exclude_from_gpa_documentation */
+
+/**
+ * @brief Mark current thread as ignored from this point on, for the duration of its existence.
+ */
+void ITTAPI __itt_thread_ignore(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, thread_ignore, (void))
+#define __itt_thread_ignore     ITTNOTIFY_VOID(thread_ignore)
+#define __itt_thread_ignore_ptr ITTNOTIFY_NAME(thread_ignore)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_thread_ignore()
+#define __itt_thread_ignore_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_thread_ignore_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} threads group */
+
+/**
+ * @defgroup suppress Error suppression
+ * @ingroup public
+ * General behavior: application continues to run, but errors are suppressed
+ *
+ * @{
+ */
+
+/*****************************************************************//**
+ * @name group of functions used for error suppression in correctness tools
+ *********************************************************************/
+/** @{ */
+/**
+ * @hideinitializer
+ * @brief possible value for suppression mask
+ */
+#define __itt_suppress_all_errors 0x7fffffff
+
+/**
+ * @hideinitializer
+ * @brief possible value for suppression mask (suppresses errors from threading analysis)
+ */
+#define __itt_suppress_threading_errors 0x000000ff
+
+/**
+ * @hideinitializer
+ * @brief possible value for suppression mask (suppresses errors from memory analysis)
+ */
+#define __itt_suppress_memory_errors 0x0000ff00
+
+/**
+ * @brief Start suppressing errors identified in mask on this thread
+ */
+void ITTAPI __itt_suppress_push(unsigned int mask);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, suppress_push, (unsigned int mask))
+#define __itt_suppress_push     ITTNOTIFY_VOID(suppress_push)
+#define __itt_suppress_push_ptr ITTNOTIFY_NAME(suppress_push)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_suppress_push(mask)
+#define __itt_suppress_push_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_suppress_push_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Undo the effects of the matching call to __itt_suppress_push
+ */
+void ITTAPI __itt_suppress_pop(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, suppress_pop, (void))
+#define __itt_suppress_pop     ITTNOTIFY_VOID(suppress_pop)
+#define __itt_suppress_pop_ptr ITTNOTIFY_NAME(suppress_pop)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_suppress_pop()
+#define __itt_suppress_pop_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_suppress_pop_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @enum __itt_model_disable
+ * @brief Enumerator for the disable methods
+ */
+typedef enum __itt_suppress_mode {
+    __itt_unsuppress_range,
+    __itt_suppress_range
+} __itt_suppress_mode_t;
+
+/**
+ * @brief Mark a range of memory for error suppression or unsuppression for error types included in mask
+ */
+void ITTAPI __itt_suppress_mark_range(__itt_suppress_mode_t mode, unsigned int mask, void * address, size_t size);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, suppress_mark_range, (__itt_suppress_mode_t mode, unsigned int mask, void * address, size_t size))
+#define __itt_suppress_mark_range     ITTNOTIFY_VOID(suppress_mark_range)
+#define __itt_suppress_mark_range_ptr ITTNOTIFY_NAME(suppress_mark_range)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_suppress_mark_range(mask)
+#define __itt_suppress_mark_range_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_suppress_mark_range_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Undo the effect of a matching call to __itt_suppress_mark_range.   If not matching
+ *        call is found, nothing is changed.
+ */
+void ITTAPI __itt_suppress_clear_range(__itt_suppress_mode_t mode, unsigned int mask, void * address, size_t size);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, suppress_clear_range, (__itt_suppress_mode_t mode, unsigned int mask, void * address, size_t size))
+#define __itt_suppress_clear_range     ITTNOTIFY_VOID(suppress_clear_range)
+#define __itt_suppress_clear_range_ptr ITTNOTIFY_NAME(suppress_clear_range)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_suppress_clear_range(mask)
+#define __itt_suppress_clear_range_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_suppress_clear_range_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} */
+/** @} suppress group */
+
+/**
+ * @defgroup sync Synchronization
+ * @ingroup public
+ * Indicate user-written synchronization code
+ * @{
+ */
+/**
+ * @hideinitializer
+ * @brief possible value of attribute argument for sync object type
+ */
+#define __itt_attr_barrier 1
+
+/**
+ * @hideinitializer
+ * @brief possible value of attribute argument for sync object type
+ */
+#define __itt_attr_mutex   2
+
+/**
+@brief Name a synchronization object
+@param[in] addr       Handle for the synchronization object. You should
+use a real address to uniquely identify the synchronization object.
+@param[in] objtype    null-terminated object type string. If NULL is
+passed, the name will be "User Synchronization".
+@param[in] objname    null-terminated object name string. If NULL,
+no name will be assigned to the object.
+@param[in] attribute  one of [#__itt_attr_barrier, #__itt_attr_mutex]
+ */
+
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+void ITTAPI __itt_sync_createA(void *addr, const char    *objtype, const char    *objname, int attribute);
+void ITTAPI __itt_sync_createW(void *addr, const wchar_t *objtype, const wchar_t *objname, int attribute);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_sync_create     __itt_sync_createW
+#  define __itt_sync_create_ptr __itt_sync_createW_ptr
+#else /* UNICODE */
+#  define __itt_sync_create     __itt_sync_createA
+#  define __itt_sync_create_ptr __itt_sync_createA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+void ITTAPI __itt_sync_create (void *addr, const char *objtype, const char *objname, int attribute);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUBV(ITTAPI, void, sync_createA, (void *addr, const char    *objtype, const char    *objname, int attribute))
+ITT_STUBV(ITTAPI, void, sync_createW, (void *addr, const wchar_t *objtype, const wchar_t *objname, int attribute))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUBV(ITTAPI, void, sync_create,  (void *addr, const char*    objtype, const char*    objname, int attribute))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_sync_createA     ITTNOTIFY_VOID(sync_createA)
+#define __itt_sync_createA_ptr ITTNOTIFY_NAME(sync_createA)
+#define __itt_sync_createW     ITTNOTIFY_VOID(sync_createW)
+#define __itt_sync_createW_ptr ITTNOTIFY_NAME(sync_createW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_sync_create     ITTNOTIFY_VOID(sync_create)
+#define __itt_sync_create_ptr ITTNOTIFY_NAME(sync_create)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_sync_createA(addr, objtype, objname, attribute)
+#define __itt_sync_createA_ptr 0
+#define __itt_sync_createW(addr, objtype, objname, attribute)
+#define __itt_sync_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_sync_create(addr, objtype, objname, attribute)
+#define __itt_sync_create_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_sync_createA_ptr 0
+#define __itt_sync_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_sync_create_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+@brief Rename a synchronization object
+
+You can use the rename call to assign or reassign a name to a given
+synchronization object.
+@param[in] addr  handle for the synchronization object.
+@param[in] name  null-terminated object name string.
+*/
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+void ITTAPI __itt_sync_renameA(void *addr, const char    *name);
+void ITTAPI __itt_sync_renameW(void *addr, const wchar_t *name);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_sync_rename     __itt_sync_renameW
+#  define __itt_sync_rename_ptr __itt_sync_renameW_ptr
+#else /* UNICODE */
+#  define __itt_sync_rename     __itt_sync_renameA
+#  define __itt_sync_rename_ptr __itt_sync_renameA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+void ITTAPI __itt_sync_rename(void *addr, const char *name);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUBV(ITTAPI, void, sync_renameA, (void *addr, const char    *name))
+ITT_STUBV(ITTAPI, void, sync_renameW, (void *addr, const wchar_t *name))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUBV(ITTAPI, void, sync_rename,  (void *addr, const char    *name))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_sync_renameA     ITTNOTIFY_VOID(sync_renameA)
+#define __itt_sync_renameA_ptr ITTNOTIFY_NAME(sync_renameA)
+#define __itt_sync_renameW     ITTNOTIFY_VOID(sync_renameW)
+#define __itt_sync_renameW_ptr ITTNOTIFY_NAME(sync_renameW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_sync_rename     ITTNOTIFY_VOID(sync_rename)
+#define __itt_sync_rename_ptr ITTNOTIFY_NAME(sync_rename)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_sync_renameA(addr, name)
+#define __itt_sync_renameA_ptr 0
+#define __itt_sync_renameW(addr, name)
+#define __itt_sync_renameW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_sync_rename(addr, name)
+#define __itt_sync_rename_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_sync_renameA_ptr 0
+#define __itt_sync_renameW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_sync_rename_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ @brief Destroy a synchronization object.
+ @param addr Handle for the synchronization object.
+ */
+void ITTAPI __itt_sync_destroy(void *addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, sync_destroy, (void *addr))
+#define __itt_sync_destroy     ITTNOTIFY_VOID(sync_destroy)
+#define __itt_sync_destroy_ptr ITTNOTIFY_NAME(sync_destroy)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_sync_destroy(addr)
+#define __itt_sync_destroy_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_sync_destroy_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/*****************************************************************//**
+ * @name group of functions is used for performance measurement tools
+ *********************************************************************/
+/** @{ */
+/**
+ * @brief Enter spin loop on user-defined sync object
+ */
+void ITTAPI __itt_sync_prepare(void* addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, sync_prepare, (void *addr))
+#define __itt_sync_prepare     ITTNOTIFY_VOID(sync_prepare)
+#define __itt_sync_prepare_ptr ITTNOTIFY_NAME(sync_prepare)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_sync_prepare(addr)
+#define __itt_sync_prepare_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_sync_prepare_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Quit spin loop without acquiring spin object
+ */
+void ITTAPI __itt_sync_cancel(void *addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, sync_cancel, (void *addr))
+#define __itt_sync_cancel     ITTNOTIFY_VOID(sync_cancel)
+#define __itt_sync_cancel_ptr ITTNOTIFY_NAME(sync_cancel)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_sync_cancel(addr)
+#define __itt_sync_cancel_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_sync_cancel_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Successful spin loop completion (sync object acquired)
+ */
+void ITTAPI __itt_sync_acquired(void *addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, sync_acquired, (void *addr))
+#define __itt_sync_acquired     ITTNOTIFY_VOID(sync_acquired)
+#define __itt_sync_acquired_ptr ITTNOTIFY_NAME(sync_acquired)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_sync_acquired(addr)
+#define __itt_sync_acquired_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_sync_acquired_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Start sync object releasing code. Is called before the lock release call.
+ */
+void ITTAPI __itt_sync_releasing(void* addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, sync_releasing, (void *addr))
+#define __itt_sync_releasing     ITTNOTIFY_VOID(sync_releasing)
+#define __itt_sync_releasing_ptr ITTNOTIFY_NAME(sync_releasing)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_sync_releasing(addr)
+#define __itt_sync_releasing_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_sync_releasing_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} */
+
+/** @} sync group */
+
+/**************************************************************//**
+ * @name group of functions is used for correctness checking tools
+ ******************************************************************/
+/** @{ */
+/**
+ * @ingroup legacy
+ * @deprecated Legacy API
+ * @brief Fast synchronization which does no require spinning.
+ * - This special function is to be used by TBB and OpenMP libraries only when they know
+ *   there is no spin but they need to suppress TC warnings about shared variable modifications.
+ * - It only has corresponding pointers in static library and does not have corresponding function
+ *   in dynamic library.
+ * @see void __itt_sync_prepare(void* addr);
+ */
+void ITTAPI __itt_fsync_prepare(void* addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, fsync_prepare, (void *addr))
+#define __itt_fsync_prepare     ITTNOTIFY_VOID(fsync_prepare)
+#define __itt_fsync_prepare_ptr ITTNOTIFY_NAME(fsync_prepare)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_fsync_prepare(addr)
+#define __itt_fsync_prepare_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_fsync_prepare_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup legacy
+ * @deprecated Legacy API
+ * @brief Fast synchronization which does no require spinning.
+ * - This special function is to be used by TBB and OpenMP libraries only when they know
+ *   there is no spin but they need to suppress TC warnings about shared variable modifications.
+ * - It only has corresponding pointers in static library and does not have corresponding function
+ *   in dynamic library.
+ * @see void __itt_sync_cancel(void *addr);
+ */
+void ITTAPI __itt_fsync_cancel(void *addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, fsync_cancel, (void *addr))
+#define __itt_fsync_cancel     ITTNOTIFY_VOID(fsync_cancel)
+#define __itt_fsync_cancel_ptr ITTNOTIFY_NAME(fsync_cancel)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_fsync_cancel(addr)
+#define __itt_fsync_cancel_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_fsync_cancel_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup legacy
+ * @deprecated Legacy API
+ * @brief Fast synchronization which does no require spinning.
+ * - This special function is to be used by TBB and OpenMP libraries only when they know
+ *   there is no spin but they need to suppress TC warnings about shared variable modifications.
+ * - It only has corresponding pointers in static library and does not have corresponding function
+ *   in dynamic library.
+ * @see void __itt_sync_acquired(void *addr);
+ */
+void ITTAPI __itt_fsync_acquired(void *addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, fsync_acquired, (void *addr))
+#define __itt_fsync_acquired     ITTNOTIFY_VOID(fsync_acquired)
+#define __itt_fsync_acquired_ptr ITTNOTIFY_NAME(fsync_acquired)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_fsync_acquired(addr)
+#define __itt_fsync_acquired_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_fsync_acquired_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup legacy
+ * @deprecated Legacy API
+ * @brief Fast synchronization which does no require spinning.
+ * - This special function is to be used by TBB and OpenMP libraries only when they know
+ *   there is no spin but they need to suppress TC warnings about shared variable modifications.
+ * - It only has corresponding pointers in static library and does not have corresponding function
+ *   in dynamic library.
+ * @see void __itt_sync_releasing(void* addr);
+ */
+void ITTAPI __itt_fsync_releasing(void* addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, fsync_releasing, (void *addr))
+#define __itt_fsync_releasing     ITTNOTIFY_VOID(fsync_releasing)
+#define __itt_fsync_releasing_ptr ITTNOTIFY_NAME(fsync_releasing)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_fsync_releasing(addr)
+#define __itt_fsync_releasing_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_fsync_releasing_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} */
+
+/**
+ * @defgroup model Modeling by Intel(R) Parallel Advisor
+ * @ingroup public
+ * This is the subset of itt used for modeling by Intel(R) Parallel Advisor.
+ * This API is called ONLY using annotate.h, by "Annotation" macros
+ * the user places in their sources during the parallelism modeling steps.
+ *
+ * site_begin/end and task_begin/end take the address of handle variables,
+ * which are writeable by the API.  Handles must be 0 initialized prior
+ * to the first call to begin, or may cause a run-time failure.
+ * The handles are initialized in a multi-thread safe way by the API if
+ * the handle is 0.  The commonly expected idiom is one static handle to
+ * identify a site or task.  If a site or task of the same name has already
+ * been started during this collection, the same handle MAY be returned,
+ * but is not required to be - it is unspecified if data merging is done
+ * based on name.  These routines also take an instance variable.  Like
+ * the lexical instance, these must be 0 initialized.  Unlike the lexical
+ * instance, this is used to track a single dynamic instance.
+ *
+ * API used by the Intel(R) Parallel Advisor to describe potential concurrency
+ * and related activities. User-added source annotations expand to calls
+ * to these procedures to enable modeling of a hypothetical concurrent
+ * execution serially.
+ * @{
+ */
+#if !defined(_ADVISOR_ANNOTATE_H_) || defined(ANNOTATE_EXPAND_NULL)
+
+typedef void* __itt_model_site;             /*!< @brief handle for lexical site     */
+typedef void* __itt_model_site_instance;    /*!< @brief handle for dynamic instance */
+typedef void* __itt_model_task;             /*!< @brief handle for lexical site     */
+typedef void* __itt_model_task_instance;    /*!< @brief handle for dynamic instance */
+
+/**
+ * @enum __itt_model_disable
+ * @brief Enumerator for the disable methods
+ */
+typedef enum {
+    __itt_model_disable_observation,
+    __itt_model_disable_collection
+} __itt_model_disable;
+
+#endif /* !_ADVISOR_ANNOTATE_H_ || ANNOTATE_EXPAND_NULL */
+
+/**
+ * @brief ANNOTATE_SITE_BEGIN/ANNOTATE_SITE_END support.
+ *
+ * site_begin/end model a potential concurrency site.
+ * site instances may be recursively nested with themselves.
+ * site_end exits the most recently started but unended site for the current
+ * thread.  The handle passed to end may be used to validate structure.
+ * Instances of a site encountered on different threads concurrently
+ * are considered completely distinct. If the site name for two different
+ * lexical sites match, it is unspecified whether they are treated as the
+ * same or different for data presentation.
+ */
+void ITTAPI __itt_model_site_begin(__itt_model_site *site, __itt_model_site_instance *instance, const char *name);
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+void ITTAPI __itt_model_site_beginW(const wchar_t *name);
+#endif
+void ITTAPI __itt_model_site_beginA(const char *name);
+void ITTAPI __itt_model_site_beginAL(const char *name, size_t siteNameLen);
+void ITTAPI __itt_model_site_end  (__itt_model_site *site, __itt_model_site_instance *instance);
+void ITTAPI __itt_model_site_end_2(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_site_begin,  (__itt_model_site *site, __itt_model_site_instance *instance, const char *name))
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUBV(ITTAPI, void, model_site_beginW,  (const wchar_t *name))
+#endif
+ITT_STUBV(ITTAPI, void, model_site_beginA,  (const char *name))
+ITT_STUBV(ITTAPI, void, model_site_beginAL,  (const char *name, size_t siteNameLen))
+ITT_STUBV(ITTAPI, void, model_site_end,    (__itt_model_site *site, __itt_model_site_instance *instance))
+ITT_STUBV(ITTAPI, void, model_site_end_2,  (void))
+#define __itt_model_site_begin      ITTNOTIFY_VOID(model_site_begin)
+#define __itt_model_site_begin_ptr  ITTNOTIFY_NAME(model_site_begin)
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_model_site_beginW      ITTNOTIFY_VOID(model_site_beginW)
+#define __itt_model_site_beginW_ptr  ITTNOTIFY_NAME(model_site_beginW)
+#endif
+#define __itt_model_site_beginA      ITTNOTIFY_VOID(model_site_beginA)
+#define __itt_model_site_beginA_ptr  ITTNOTIFY_NAME(model_site_beginA)
+#define __itt_model_site_beginAL      ITTNOTIFY_VOID(model_site_beginAL)
+#define __itt_model_site_beginAL_ptr  ITTNOTIFY_NAME(model_site_beginAL)
+#define __itt_model_site_end        ITTNOTIFY_VOID(model_site_end)
+#define __itt_model_site_end_ptr    ITTNOTIFY_NAME(model_site_end)
+#define __itt_model_site_end_2        ITTNOTIFY_VOID(model_site_end_2)
+#define __itt_model_site_end_2_ptr    ITTNOTIFY_NAME(model_site_end_2)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_site_begin(site, instance, name)
+#define __itt_model_site_begin_ptr  0
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_model_site_beginW(name)
+#define __itt_model_site_beginW_ptr  0
+#endif
+#define __itt_model_site_beginA(name)
+#define __itt_model_site_beginA_ptr  0
+#define __itt_model_site_beginAL(name, siteNameLen)
+#define __itt_model_site_beginAL_ptr  0
+#define __itt_model_site_end(site, instance)
+#define __itt_model_site_end_ptr    0
+#define __itt_model_site_end_2()
+#define __itt_model_site_end_2_ptr    0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_site_begin_ptr  0
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_model_site_beginW_ptr  0
+#endif
+#define __itt_model_site_beginA_ptr  0
+#define __itt_model_site_beginAL_ptr  0
+#define __itt_model_site_end_ptr    0
+#define __itt_model_site_end_2_ptr    0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief ANNOTATE_TASK_BEGIN/ANNOTATE_TASK_END support
+ *
+ * task_begin/end model a potential task, which is contained within the most
+ * closely enclosing dynamic site.  task_end exits the most recently started
+ * but unended task.  The handle passed to end may be used to validate
+ * structure.  It is unspecified if bad dynamic nesting is detected.  If it
+ * is, it should be encoded in the resulting data collection.  The collector
+ * should not fail due to construct nesting issues, nor attempt to directly
+ * indicate the problem.
+ */
+void ITTAPI __itt_model_task_begin(__itt_model_task *task, __itt_model_task_instance *instance, const char *name);
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+void ITTAPI __itt_model_task_beginW(const wchar_t *name);
+void ITTAPI __itt_model_iteration_taskW(const wchar_t *name);
+#endif
+void ITTAPI __itt_model_task_beginA(const char *name);
+void ITTAPI __itt_model_task_beginAL(const char *name, size_t taskNameLen);
+void ITTAPI __itt_model_iteration_taskA(const char *name);
+void ITTAPI __itt_model_iteration_taskAL(const char *name, size_t taskNameLen);
+void ITTAPI __itt_model_task_end  (__itt_model_task *task, __itt_model_task_instance *instance);
+void ITTAPI __itt_model_task_end_2(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_task_begin,  (__itt_model_task *task, __itt_model_task_instance *instance, const char *name))
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUBV(ITTAPI, void, model_task_beginW,  (const wchar_t *name))
+ITT_STUBV(ITTAPI, void, model_iteration_taskW, (const wchar_t *name))
+#endif
+ITT_STUBV(ITTAPI, void, model_task_beginA,  (const char *name))
+ITT_STUBV(ITTAPI, void, model_task_beginAL,  (const char *name, size_t taskNameLen))
+ITT_STUBV(ITTAPI, void, model_iteration_taskA,  (const char *name))
+ITT_STUBV(ITTAPI, void, model_iteration_taskAL,  (const char *name, size_t taskNameLen))
+ITT_STUBV(ITTAPI, void, model_task_end,    (__itt_model_task *task, __itt_model_task_instance *instance))
+ITT_STUBV(ITTAPI, void, model_task_end_2,  (void))
+#define __itt_model_task_begin      ITTNOTIFY_VOID(model_task_begin)
+#define __itt_model_task_begin_ptr  ITTNOTIFY_NAME(model_task_begin)
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_model_task_beginW     ITTNOTIFY_VOID(model_task_beginW)
+#define __itt_model_task_beginW_ptr ITTNOTIFY_NAME(model_task_beginW)
+#define __itt_model_iteration_taskW     ITTNOTIFY_VOID(model_iteration_taskW)
+#define __itt_model_iteration_taskW_ptr ITTNOTIFY_NAME(model_iteration_taskW)
+#endif
+#define __itt_model_task_beginA    ITTNOTIFY_VOID(model_task_beginA)
+#define __itt_model_task_beginA_ptr ITTNOTIFY_NAME(model_task_beginA)
+#define __itt_model_task_beginAL    ITTNOTIFY_VOID(model_task_beginAL)
+#define __itt_model_task_beginAL_ptr ITTNOTIFY_NAME(model_task_beginAL)
+#define __itt_model_iteration_taskA    ITTNOTIFY_VOID(model_iteration_taskA)
+#define __itt_model_iteration_taskA_ptr ITTNOTIFY_NAME(model_iteration_taskA)
+#define __itt_model_iteration_taskAL    ITTNOTIFY_VOID(model_iteration_taskAL)
+#define __itt_model_iteration_taskAL_ptr ITTNOTIFY_NAME(model_iteration_taskAL)
+#define __itt_model_task_end        ITTNOTIFY_VOID(model_task_end)
+#define __itt_model_task_end_ptr    ITTNOTIFY_NAME(model_task_end)
+#define __itt_model_task_end_2        ITTNOTIFY_VOID(model_task_end_2)
+#define __itt_model_task_end_2_ptr    ITTNOTIFY_NAME(model_task_end_2)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_task_begin(task, instance, name)
+#define __itt_model_task_begin_ptr  0
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_model_task_beginW(name)
+#define __itt_model_task_beginW_ptr  0
+#endif
+#define __itt_model_task_beginA(name)
+#define __itt_model_task_beginA_ptr  0
+#define __itt_model_task_beginAL(name, siteNameLen)
+#define __itt_model_task_beginAL_ptr  0
+#define __itt_model_iteration_taskA(name)
+#define __itt_model_iteration_taskA_ptr  0
+#define __itt_model_iteration_taskAL(name, siteNameLen)
+#define __itt_model_iteration_taskAL_ptr  0
+#define __itt_model_task_end(task, instance)
+#define __itt_model_task_end_ptr    0
+#define __itt_model_task_end_2()
+#define __itt_model_task_end_2_ptr    0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_task_begin_ptr  0
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_model_task_beginW_ptr 0
+#endif
+#define __itt_model_task_beginA_ptr  0
+#define __itt_model_task_beginAL_ptr  0
+#define __itt_model_iteration_taskA_ptr    0
+#define __itt_model_iteration_taskAL_ptr    0
+#define __itt_model_task_end_ptr    0
+#define __itt_model_task_end_2_ptr    0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief ANNOTATE_LOCK_ACQUIRE/ANNOTATE_LOCK_RELEASE support
+ *
+ * lock_acquire/release model a potential lock for both lockset and
+ * performance modeling.  Each unique address is modeled as a separate
+ * lock, with invalid addresses being valid lock IDs.  Specifically:
+ * no storage is accessed by the API at the specified address - it is only
+ * used for lock identification.  Lock acquires may be self-nested and are
+ * unlocked by a corresponding number of releases.
+ * (These closely correspond to __itt_sync_acquired/__itt_sync_releasing,
+ * but may not have identical semantics.)
+ */
+void ITTAPI __itt_model_lock_acquire(void *lock);
+void ITTAPI __itt_model_lock_acquire_2(void *lock);
+void ITTAPI __itt_model_lock_release(void *lock);
+void ITTAPI __itt_model_lock_release_2(void *lock);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_lock_acquire, (void *lock))
+ITT_STUBV(ITTAPI, void, model_lock_acquire_2, (void *lock))
+ITT_STUBV(ITTAPI, void, model_lock_release, (void *lock))
+ITT_STUBV(ITTAPI, void, model_lock_release_2, (void *lock))
+#define __itt_model_lock_acquire     ITTNOTIFY_VOID(model_lock_acquire)
+#define __itt_model_lock_acquire_ptr ITTNOTIFY_NAME(model_lock_acquire)
+#define __itt_model_lock_acquire_2     ITTNOTIFY_VOID(model_lock_acquire_2)
+#define __itt_model_lock_acquire_2_ptr ITTNOTIFY_NAME(model_lock_acquire_2)
+#define __itt_model_lock_release     ITTNOTIFY_VOID(model_lock_release)
+#define __itt_model_lock_release_ptr ITTNOTIFY_NAME(model_lock_release)
+#define __itt_model_lock_release_2     ITTNOTIFY_VOID(model_lock_release_2)
+#define __itt_model_lock_release_2_ptr ITTNOTIFY_NAME(model_lock_release_2)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_lock_acquire(lock)
+#define __itt_model_lock_acquire_ptr 0
+#define __itt_model_lock_acquire_2(lock)
+#define __itt_model_lock_acquire_2_ptr 0
+#define __itt_model_lock_release(lock)
+#define __itt_model_lock_release_ptr 0
+#define __itt_model_lock_release_2(lock)
+#define __itt_model_lock_release_2_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_lock_acquire_ptr 0
+#define __itt_model_lock_acquire_2_ptr 0
+#define __itt_model_lock_release_ptr 0
+#define __itt_model_lock_release_2_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief ANNOTATE_RECORD_ALLOCATION/ANNOTATE_RECORD_DEALLOCATION support
+ *
+ * record_allocation/deallocation describe user-defined memory allocator
+ * behavior, which may be required for correctness modeling to understand
+ * when storage is not expected to be actually reused across threads.
+ */
+void ITTAPI __itt_model_record_allocation  (void *addr, size_t size);
+void ITTAPI __itt_model_record_deallocation(void *addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_record_allocation,   (void *addr, size_t size))
+ITT_STUBV(ITTAPI, void, model_record_deallocation, (void *addr))
+#define __itt_model_record_allocation       ITTNOTIFY_VOID(model_record_allocation)
+#define __itt_model_record_allocation_ptr   ITTNOTIFY_NAME(model_record_allocation)
+#define __itt_model_record_deallocation     ITTNOTIFY_VOID(model_record_deallocation)
+#define __itt_model_record_deallocation_ptr ITTNOTIFY_NAME(model_record_deallocation)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_record_allocation(addr, size)
+#define __itt_model_record_allocation_ptr   0
+#define __itt_model_record_deallocation(addr)
+#define __itt_model_record_deallocation_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_record_allocation_ptr   0
+#define __itt_model_record_deallocation_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief ANNOTATE_INDUCTION_USES support
+ *
+ * Note particular storage is inductive through the end of the current site
+ */
+void ITTAPI __itt_model_induction_uses(void* addr, size_t size);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_induction_uses, (void *addr, size_t size))
+#define __itt_model_induction_uses     ITTNOTIFY_VOID(model_induction_uses)
+#define __itt_model_induction_uses_ptr ITTNOTIFY_NAME(model_induction_uses)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_induction_uses(addr, size)
+#define __itt_model_induction_uses_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_induction_uses_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief ANNOTATE_REDUCTION_USES support
+ *
+ * Note particular storage is used for reduction through the end
+ * of the current site
+ */
+void ITTAPI __itt_model_reduction_uses(void* addr, size_t size);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_reduction_uses, (void *addr, size_t size))
+#define __itt_model_reduction_uses     ITTNOTIFY_VOID(model_reduction_uses)
+#define __itt_model_reduction_uses_ptr ITTNOTIFY_NAME(model_reduction_uses)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_reduction_uses(addr, size)
+#define __itt_model_reduction_uses_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_reduction_uses_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief ANNOTATE_OBSERVE_USES support
+ *
+ * Have correctness modeling record observations about uses of storage
+ * through the end of the current site
+ */
+void ITTAPI __itt_model_observe_uses(void* addr, size_t size);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_observe_uses, (void *addr, size_t size))
+#define __itt_model_observe_uses     ITTNOTIFY_VOID(model_observe_uses)
+#define __itt_model_observe_uses_ptr ITTNOTIFY_NAME(model_observe_uses)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_observe_uses(addr, size)
+#define __itt_model_observe_uses_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_observe_uses_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief ANNOTATE_CLEAR_USES support
+ *
+ * Clear the special handling of a piece of storage related to induction,
+ * reduction or observe_uses
+ */
+void ITTAPI __itt_model_clear_uses(void* addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_clear_uses, (void *addr))
+#define __itt_model_clear_uses     ITTNOTIFY_VOID(model_clear_uses)
+#define __itt_model_clear_uses_ptr ITTNOTIFY_NAME(model_clear_uses)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_clear_uses(addr)
+#define __itt_model_clear_uses_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_clear_uses_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief ANNOTATE_DISABLE_*_PUSH/ANNOTATE_DISABLE_*_POP support
+ *
+ * disable_push/disable_pop push and pop disabling based on a parameter.
+ * Disabling observations stops processing of memory references during
+ * correctness modeling, and all annotations that occur in the disabled
+ * region.  This allows description of code that is expected to be handled
+ * specially during conversion to parallelism or that is not recognized
+ * by tools (e.g. some kinds of synchronization operations.)
+ * This mechanism causes all annotations in the disabled region, other
+ * than disable_push and disable_pop, to be ignored.  (For example, this
+ * might validly be used to disable an entire parallel site and the contained
+ * tasks and locking in it for data collection purposes.)
+ * The disable for collection is a more expensive operation, but reduces
+ * collector overhead significantly.  This applies to BOTH correctness data
+ * collection and performance data collection.  For example, a site
+ * containing a task might only enable data collection for the first 10
+ * iterations.  Both performance and correctness data should reflect this,
+ * and the program should run as close to full speed as possible when
+ * collection is disabled.
+ */
+void ITTAPI __itt_model_disable_push(__itt_model_disable x);
+void ITTAPI __itt_model_disable_pop(void);
+void ITTAPI __itt_model_aggregate_task(size_t x);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, model_disable_push, (__itt_model_disable x))
+ITT_STUBV(ITTAPI, void, model_disable_pop,  (void))
+ITT_STUBV(ITTAPI, void, model_aggregate_task, (size_t x))
+#define __itt_model_disable_push     ITTNOTIFY_VOID(model_disable_push)
+#define __itt_model_disable_push_ptr ITTNOTIFY_NAME(model_disable_push)
+#define __itt_model_disable_pop      ITTNOTIFY_VOID(model_disable_pop)
+#define __itt_model_disable_pop_ptr  ITTNOTIFY_NAME(model_disable_pop)
+#define __itt_model_aggregate_task      ITTNOTIFY_VOID(model_aggregate_task)
+#define __itt_model_aggregate_task_ptr  ITTNOTIFY_NAME(model_aggregate_task)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_model_disable_push(x)
+#define __itt_model_disable_push_ptr 0
+#define __itt_model_disable_pop()
+#define __itt_model_disable_pop_ptr 0
+#define __itt_model_aggregate_task(x)
+#define __itt_model_aggregate_task_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_model_disable_push_ptr 0
+#define __itt_model_disable_pop_ptr 0
+#define __itt_model_aggregate_task_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} model group */
+
+/**
+ * @defgroup heap Heap
+ * @ingroup public
+ * Heap group
+ * @{
+ */
+
+typedef void* __itt_heap_function;
+
+/**
+ * @brief Create an identification for heap function
+ * @return non-zero identifier or NULL
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+__itt_heap_function ITTAPI __itt_heap_function_createA(const char*    name, const char*    domain);
+__itt_heap_function ITTAPI __itt_heap_function_createW(const wchar_t* name, const wchar_t* domain);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_heap_function_create     __itt_heap_function_createW
+#  define __itt_heap_function_create_ptr __itt_heap_function_createW_ptr
+#else
+#  define __itt_heap_function_create     __itt_heap_function_createA
+#  define __itt_heap_function_create_ptr __itt_heap_function_createA_ptr
+#endif /* UNICODE */
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+__itt_heap_function ITTAPI __itt_heap_function_create(const char* name, const char* domain);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, __itt_heap_function, heap_function_createA, (const char*    name, const char*    domain))
+ITT_STUB(ITTAPI, __itt_heap_function, heap_function_createW, (const wchar_t* name, const wchar_t* domain))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, __itt_heap_function, heap_function_create,  (const char*    name, const char*    domain))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_heap_function_createA     ITTNOTIFY_DATA(heap_function_createA)
+#define __itt_heap_function_createA_ptr ITTNOTIFY_NAME(heap_function_createA)
+#define __itt_heap_function_createW     ITTNOTIFY_DATA(heap_function_createW)
+#define __itt_heap_function_createW_ptr ITTNOTIFY_NAME(heap_function_createW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_heap_function_create      ITTNOTIFY_DATA(heap_function_create)
+#define __itt_heap_function_create_ptr  ITTNOTIFY_NAME(heap_function_create)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_heap_function_createA(name, domain) (__itt_heap_function)0
+#define __itt_heap_function_createA_ptr 0
+#define __itt_heap_function_createW(name, domain) (__itt_heap_function)0
+#define __itt_heap_function_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_heap_function_create(name, domain)  (__itt_heap_function)0
+#define __itt_heap_function_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_heap_function_createA_ptr 0
+#define __itt_heap_function_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_heap_function_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Record an allocation begin occurrence.
+ */
+void ITTAPI __itt_heap_allocate_begin(__itt_heap_function h, size_t size, int initialized);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_allocate_begin, (__itt_heap_function h, size_t size, int initialized))
+#define __itt_heap_allocate_begin     ITTNOTIFY_VOID(heap_allocate_begin)
+#define __itt_heap_allocate_begin_ptr ITTNOTIFY_NAME(heap_allocate_begin)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_allocate_begin(h, size, initialized)
+#define __itt_heap_allocate_begin_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_allocate_begin_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Record an allocation end occurrence.
+ */
+void ITTAPI __itt_heap_allocate_end(__itt_heap_function h, void** addr, size_t size, int initialized);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_allocate_end, (__itt_heap_function h, void** addr, size_t size, int initialized))
+#define __itt_heap_allocate_end     ITTNOTIFY_VOID(heap_allocate_end)
+#define __itt_heap_allocate_end_ptr ITTNOTIFY_NAME(heap_allocate_end)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_allocate_end(h, addr, size, initialized)
+#define __itt_heap_allocate_end_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_allocate_end_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Record an free begin occurrence.
+ */
+void ITTAPI __itt_heap_free_begin(__itt_heap_function h, void* addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_free_begin, (__itt_heap_function h, void* addr))
+#define __itt_heap_free_begin     ITTNOTIFY_VOID(heap_free_begin)
+#define __itt_heap_free_begin_ptr ITTNOTIFY_NAME(heap_free_begin)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_free_begin(h, addr)
+#define __itt_heap_free_begin_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_free_begin_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Record an free end occurrence.
+ */
+void ITTAPI __itt_heap_free_end(__itt_heap_function h, void* addr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_free_end, (__itt_heap_function h, void* addr))
+#define __itt_heap_free_end     ITTNOTIFY_VOID(heap_free_end)
+#define __itt_heap_free_end_ptr ITTNOTIFY_NAME(heap_free_end)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_free_end(h, addr)
+#define __itt_heap_free_end_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_free_end_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Record an reallocation begin occurrence.
+ */
+void ITTAPI __itt_heap_reallocate_begin(__itt_heap_function h, void* addr, size_t new_size, int initialized);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_reallocate_begin, (__itt_heap_function h, void* addr, size_t new_size, int initialized))
+#define __itt_heap_reallocate_begin     ITTNOTIFY_VOID(heap_reallocate_begin)
+#define __itt_heap_reallocate_begin_ptr ITTNOTIFY_NAME(heap_reallocate_begin)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_reallocate_begin(h, addr, new_size, initialized)
+#define __itt_heap_reallocate_begin_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_reallocate_begin_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Record an reallocation end occurrence.
+ */
+void ITTAPI __itt_heap_reallocate_end(__itt_heap_function h, void* addr, void** new_addr, size_t new_size, int initialized);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_reallocate_end, (__itt_heap_function h, void* addr, void** new_addr, size_t new_size, int initialized))
+#define __itt_heap_reallocate_end     ITTNOTIFY_VOID(heap_reallocate_end)
+#define __itt_heap_reallocate_end_ptr ITTNOTIFY_NAME(heap_reallocate_end)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_reallocate_end(h, addr, new_addr, new_size, initialized)
+#define __itt_heap_reallocate_end_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_reallocate_end_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @brief internal access begin */
+void ITTAPI __itt_heap_internal_access_begin(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_internal_access_begin,  (void))
+#define __itt_heap_internal_access_begin      ITTNOTIFY_VOID(heap_internal_access_begin)
+#define __itt_heap_internal_access_begin_ptr  ITTNOTIFY_NAME(heap_internal_access_begin)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_internal_access_begin()
+#define __itt_heap_internal_access_begin_ptr  0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_internal_access_begin_ptr  0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @brief internal access end */
+void ITTAPI __itt_heap_internal_access_end(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_internal_access_end, (void))
+#define __itt_heap_internal_access_end     ITTNOTIFY_VOID(heap_internal_access_end)
+#define __itt_heap_internal_access_end_ptr ITTNOTIFY_NAME(heap_internal_access_end)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_internal_access_end()
+#define __itt_heap_internal_access_end_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_internal_access_end_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @brief record memory growth begin */
+void ITTAPI __itt_heap_record_memory_growth_begin(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_record_memory_growth_begin,  (void))
+#define __itt_heap_record_memory_growth_begin      ITTNOTIFY_VOID(heap_record_memory_growth_begin)
+#define __itt_heap_record_memory_growth_begin_ptr  ITTNOTIFY_NAME(heap_record_memory_growth_begin)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_record_memory_growth_begin()
+#define __itt_heap_record_memory_growth_begin_ptr  0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_record_memory_growth_begin_ptr  0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @brief record memory growth end */
+void ITTAPI __itt_heap_record_memory_growth_end(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_record_memory_growth_end, (void))
+#define __itt_heap_record_memory_growth_end     ITTNOTIFY_VOID(heap_record_memory_growth_end)
+#define __itt_heap_record_memory_growth_end_ptr ITTNOTIFY_NAME(heap_record_memory_growth_end)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_record_memory_growth_end()
+#define __itt_heap_record_memory_growth_end_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_record_memory_growth_end_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Specify the type of heap detection/reporting to modify.
+ */
+/**
+ * @hideinitializer
+ * @brief Report on memory leaks.
+ */
+#define __itt_heap_leaks 0x00000001
+
+/**
+ * @hideinitializer
+ * @brief Report on memory growth.
+ */
+#define __itt_heap_growth 0x00000002
+
+
+/** @brief heap reset detection */
+void ITTAPI __itt_heap_reset_detection(unsigned int reset_mask);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_reset_detection,  (unsigned int reset_mask))
+#define __itt_heap_reset_detection      ITTNOTIFY_VOID(heap_reset_detection)
+#define __itt_heap_reset_detection_ptr  ITTNOTIFY_NAME(heap_reset_detection)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_reset_detection()
+#define __itt_heap_reset_detection_ptr  0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_reset_detection_ptr  0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @brief report */
+void ITTAPI __itt_heap_record(unsigned int record_mask);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, heap_record, (unsigned int record_mask))
+#define __itt_heap_record     ITTNOTIFY_VOID(heap_record)
+#define __itt_heap_record_ptr ITTNOTIFY_NAME(heap_record)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_heap_record()
+#define __itt_heap_record_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_heap_record_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @} heap group */
+/** @endcond */
+/* ========================================================================== */
+
+/**
+ * @defgroup domains Domains
+ * @ingroup public
+ * Domains group
+ * @{
+ */
+
+/** @cond exclude_from_documentation */
+#pragma pack(push, 8)
+
+typedef struct ___itt_domain
+{
+    volatile int flags; /*!< Zero if disabled, non-zero if enabled. The meaning of different non-zero values is reserved to the runtime */
+    const char* nameA;  /*!< Copy of original name in ASCII. */
+#if defined(UNICODE) || defined(_UNICODE)
+    const wchar_t* nameW; /*!< Copy of original name in UNICODE. */
+#else  /* UNICODE || _UNICODE */
+    void* nameW;
+#endif /* UNICODE || _UNICODE */
+    int   extra1; /*!< Reserved to the runtime */
+    void* extra2; /*!< Reserved to the runtime */
+    struct ___itt_domain* next;
+} __itt_domain;
+
+#pragma pack(pop)
+/** @endcond */
+
+/**
+ * @ingroup domains
+ * @brief Create a domain.
+ * Create domain using some domain name: the URI naming style is recommended.
+ * Because the set of domains is expected to be static over the application's
+ * execution time, there is no mechanism to destroy a domain.
+ * Any domain can be accessed by any thread in the process, regardless of
+ * which thread created the domain. This call is thread-safe.
+ * @param[in] name name of domain
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+__itt_domain* ITTAPI __itt_domain_createA(const char    *name);
+__itt_domain* ITTAPI __itt_domain_createW(const wchar_t *name);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_domain_create     __itt_domain_createW
+#  define __itt_domain_create_ptr __itt_domain_createW_ptr
+#else /* UNICODE */
+#  define __itt_domain_create     __itt_domain_createA
+#  define __itt_domain_create_ptr __itt_domain_createA_ptr
+#endif /* UNICODE */
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+__itt_domain* ITTAPI __itt_domain_create(const char *name);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, __itt_domain*, domain_createA, (const char    *name))
+ITT_STUB(ITTAPI, __itt_domain*, domain_createW, (const wchar_t *name))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, __itt_domain*, domain_create,  (const char    *name))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_domain_createA     ITTNOTIFY_DATA(domain_createA)
+#define __itt_domain_createA_ptr ITTNOTIFY_NAME(domain_createA)
+#define __itt_domain_createW     ITTNOTIFY_DATA(domain_createW)
+#define __itt_domain_createW_ptr ITTNOTIFY_NAME(domain_createW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_domain_create     ITTNOTIFY_DATA(domain_create)
+#define __itt_domain_create_ptr ITTNOTIFY_NAME(domain_create)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_domain_createA(name) (__itt_domain*)0
+#define __itt_domain_createA_ptr 0
+#define __itt_domain_createW(name) (__itt_domain*)0
+#define __itt_domain_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_domain_create(name)  (__itt_domain*)0
+#define __itt_domain_create_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_domain_createA_ptr 0
+#define __itt_domain_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_domain_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} domains group */
+
+/**
+ * @defgroup ids IDs
+ * @ingroup public
+ * IDs group
+ * @{
+ */
+
+/** @cond exclude_from_documentation */
+#pragma pack(push, 8)
+
+typedef struct ___itt_id
+{
+    unsigned long long d1, d2, d3;
+} __itt_id;
+
+#pragma pack(pop)
+/** @endcond */
+
+const __itt_id __itt_null = { 0, 0, 0 };
+
+/**
+ * @ingroup ids
+ * @brief A convenience function is provided to create an ID without domain control.
+ * @brief This is a convenience function to initialize an __itt_id structure. This function
+ * does not affect the collector runtime in any way. After you make the ID with this
+ * function, you still must create it with the __itt_id_create function before using the ID
+ * to identify a named entity.
+ * @param[in] addr The address of object; high QWORD of the ID value.
+ * @param[in] extra The extra data to unique identify object; low QWORD of the ID value.
+ */
+
+ITT_INLINE __itt_id ITTAPI __itt_id_make(void* addr, unsigned long long extra) ITT_INLINE_ATTRIBUTE;
+ITT_INLINE __itt_id ITTAPI __itt_id_make(void* addr, unsigned long long extra)
+{
+    __itt_id id = __itt_null;
+    id.d1 = (unsigned long long)((uintptr_t)addr);
+    id.d2 = (unsigned long long)extra;
+    id.d3 = (unsigned long long)0; /* Reserved. Must be zero */
+    return id;
+}
+
+/**
+ * @ingroup ids
+ * @brief Create an instance of identifier.
+ * This establishes the beginning of the lifetime of an instance of
+ * the given ID in the trace. Once this lifetime starts, the ID
+ * can be used to tag named entity instances in calls such as
+ * __itt_task_begin, and to specify relationships among
+ * identified named entity instances, using the \ref relations APIs.
+ * Instance IDs are not domain specific!
+ * @param[in] domain The domain controlling the execution of this call.
+ * @param[in] id The ID to create.
+ */
+void ITTAPI __itt_id_create(const __itt_domain *domain, __itt_id id);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, id_create, (const __itt_domain *domain, __itt_id id))
+#define __itt_id_create(d,x) ITTNOTIFY_VOID_D1(id_create,d,x)
+#define __itt_id_create_ptr  ITTNOTIFY_NAME(id_create)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_id_create(domain,id)
+#define __itt_id_create_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_id_create_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup ids
+ * @brief Destroy an instance of identifier.
+ * This ends the lifetime of the current instance of the given ID value in the trace.
+ * Any relationships that are established after this lifetime ends are invalid.
+ * This call must be performed before the given ID value can be reused for a different
+ * named entity instance.
+ * @param[in] domain The domain controlling the execution of this call.
+ * @param[in] id The ID to destroy.
+ */
+void ITTAPI __itt_id_destroy(const __itt_domain *domain, __itt_id id);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, id_destroy, (const __itt_domain *domain, __itt_id id))
+#define __itt_id_destroy(d,x) ITTNOTIFY_VOID_D1(id_destroy,d,x)
+#define __itt_id_destroy_ptr  ITTNOTIFY_NAME(id_destroy)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_id_destroy(domain,id)
+#define __itt_id_destroy_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_id_destroy_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} ids group */
+
+/**
+ * @defgroup handless String Handles
+ * @ingroup public
+ * String Handles group
+ * @{
+ */
+
+/** @cond exclude_from_documentation */
+#pragma pack(push, 8)
+
+typedef struct ___itt_string_handle
+{
+    const char* strA; /*!< Copy of original string in ASCII. */
+#if defined(UNICODE) || defined(_UNICODE)
+    const wchar_t* strW; /*!< Copy of original string in UNICODE. */
+#else  /* UNICODE || _UNICODE */
+    void* strW;
+#endif /* UNICODE || _UNICODE */
+    int   extra1; /*!< Reserved. Must be zero   */
+    void* extra2; /*!< Reserved. Must be zero   */
+    struct ___itt_string_handle* next;
+} __itt_string_handle;
+
+#pragma pack(pop)
+/** @endcond */
+
+/**
+ * @ingroup handles
+ * @brief Create a string handle.
+ * Create and return handle value that can be associated with a string.
+ * Consecutive calls to __itt_string_handle_create with the same name
+ * return the same value. Because the set of string handles is expected to remain
+ * static during the application's execution time, there is no mechanism to destroy a string handle.
+ * Any string handle can be accessed by any thread in the process, regardless of which thread created
+ * the string handle. This call is thread-safe.
+ * @param[in] name The input string
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+__itt_string_handle* ITTAPI __itt_string_handle_createA(const char    *name);
+__itt_string_handle* ITTAPI __itt_string_handle_createW(const wchar_t *name);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_string_handle_create     __itt_string_handle_createW
+#  define __itt_string_handle_create_ptr __itt_string_handle_createW_ptr
+#else /* UNICODE */
+#  define __itt_string_handle_create     __itt_string_handle_createA
+#  define __itt_string_handle_create_ptr __itt_string_handle_createA_ptr
+#endif /* UNICODE */
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+__itt_string_handle* ITTAPI __itt_string_handle_create(const char *name);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, __itt_string_handle*, string_handle_createA, (const char    *name))
+ITT_STUB(ITTAPI, __itt_string_handle*, string_handle_createW, (const wchar_t *name))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, __itt_string_handle*, string_handle_create,  (const char    *name))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_string_handle_createA     ITTNOTIFY_DATA(string_handle_createA)
+#define __itt_string_handle_createA_ptr ITTNOTIFY_NAME(string_handle_createA)
+#define __itt_string_handle_createW     ITTNOTIFY_DATA(string_handle_createW)
+#define __itt_string_handle_createW_ptr ITTNOTIFY_NAME(string_handle_createW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_string_handle_create     ITTNOTIFY_DATA(string_handle_create)
+#define __itt_string_handle_create_ptr ITTNOTIFY_NAME(string_handle_create)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_string_handle_createA(name) (__itt_string_handle*)0
+#define __itt_string_handle_createA_ptr 0
+#define __itt_string_handle_createW(name) (__itt_string_handle*)0
+#define __itt_string_handle_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_string_handle_create(name)  (__itt_string_handle*)0
+#define __itt_string_handle_create_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_string_handle_createA_ptr 0
+#define __itt_string_handle_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_string_handle_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} handles group */
+
+/** @cond exclude_from_documentation */
+typedef unsigned long long __itt_timestamp;
+/** @endcond */
+
+#define __itt_timestamp_none ((__itt_timestamp)-1LL)
+
+/** @cond exclude_from_gpa_documentation */
+
+/**
+ * @ingroup timestamps
+ * @brief Return timestamp corresponding to the current moment.
+ * This returns the timestamp in the format that is the most relevant for the current
+ * host or platform (RDTSC, QPC, and others). You can use the "<" operator to
+ * compare __itt_timestamp values.
+ */
+__itt_timestamp ITTAPI __itt_get_timestamp(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(ITTAPI, __itt_timestamp, get_timestamp, (void))
+#define __itt_get_timestamp      ITTNOTIFY_DATA(get_timestamp)
+#define __itt_get_timestamp_ptr  ITTNOTIFY_NAME(get_timestamp)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_get_timestamp()
+#define __itt_get_timestamp_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_get_timestamp_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} timestamps */
+/** @endcond */
+
+/** @cond exclude_from_gpa_documentation */
+
+/**
+ * @defgroup regions Regions
+ * @ingroup public
+ * Regions group
+ * @{
+ */
+/**
+ * @ingroup regions
+ * @brief Begin of region instance.
+ * Successive calls to __itt_region_begin with the same ID are ignored
+ * until a call to __itt_region_end with the same ID
+ * @param[in] domain The domain for this region instance
+ * @param[in] id The instance ID for this region instance. Must not be __itt_null
+ * @param[in] parentid The instance ID for the parent of this region instance, or __itt_null
+ * @param[in] name The name of this region
+ */
+void ITTAPI __itt_region_begin(const __itt_domain *domain, __itt_id id, __itt_id parentid, __itt_string_handle *name);
+
+/**
+ * @ingroup regions
+ * @brief End of region instance.
+ * The first call to __itt_region_end with a given ID ends the
+ * region. Successive calls with the same ID are ignored, as are
+ * calls that do not have a matching __itt_region_begin call.
+ * @param[in] domain The domain for this region instance
+ * @param[in] id The instance ID for this region instance
+ */
+void ITTAPI __itt_region_end(const __itt_domain *domain, __itt_id id);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, region_begin, (const __itt_domain *domain, __itt_id id, __itt_id parentid, __itt_string_handle *name))
+ITT_STUBV(ITTAPI, void, region_end,   (const __itt_domain *domain, __itt_id id))
+#define __itt_region_begin(d,x,y,z) ITTNOTIFY_VOID_D3(region_begin,d,x,y,z)
+#define __itt_region_begin_ptr      ITTNOTIFY_NAME(region_begin)
+#define __itt_region_end(d,x)       ITTNOTIFY_VOID_D1(region_end,d,x)
+#define __itt_region_end_ptr        ITTNOTIFY_NAME(region_end)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_region_begin(d,x,y,z)
+#define __itt_region_begin_ptr 0
+#define __itt_region_end(d,x)
+#define __itt_region_end_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_region_begin_ptr 0
+#define __itt_region_end_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} regions group */
+
+/**
+ * @defgroup frames Frames
+ * @ingroup public
+ * Frames are similar to regions, but are intended to be easier to use and to implement.
+ * In particular:
+ * - Frames always represent periods of elapsed time
+ * - By default, frames have no nesting relationships
+ * @{
+ */
+
+/**
+ * @ingroup frames
+ * @brief Begin a frame instance.
+ * Successive calls to __itt_frame_begin with the
+ * same ID are ignored until a call to __itt_frame_end with the same ID.
+ * @param[in] domain The domain for this frame instance
+ * @param[in] id The instance ID for this frame instance or NULL
+ */
+void ITTAPI __itt_frame_begin_v3(const __itt_domain *domain, __itt_id *id);
+
+/**
+ * @ingroup frames
+ * @brief End a frame instance.
+ * The first call to __itt_frame_end with a given ID
+ * ends the frame. Successive calls with the same ID are ignored, as are
+ * calls that do not have a matching __itt_frame_begin call.
+ * @param[in] domain The domain for this frame instance
+ * @param[in] id The instance ID for this frame instance or NULL for current
+ */
+void ITTAPI __itt_frame_end_v3(const __itt_domain *domain, __itt_id *id);
+
+/**
+ * @ingroup frames
+ * @brief Submits a frame instance.
+ * Successive calls to __itt_frame_begin or __itt_frame_submit with the
+ * same ID are ignored until a call to __itt_frame_end or __itt_frame_submit
+ * with the same ID.
+ * Passing special __itt_timestamp_none value as "end" argument means
+ * take the current timestamp as the end timestamp.
+ * @param[in] domain The domain for this frame instance
+ * @param[in] id The instance ID for this frame instance or NULL
+ * @param[in] begin Timestamp of the beginning of the frame
+ * @param[in] end Timestamp of the end of the frame
+ */
+void ITTAPI __itt_frame_submit_v3(const __itt_domain *domain, __itt_id *id,
+    __itt_timestamp begin, __itt_timestamp end);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, frame_begin_v3,  (const __itt_domain *domain, __itt_id *id))
+ITT_STUBV(ITTAPI, void, frame_end_v3,    (const __itt_domain *domain, __itt_id *id))
+ITT_STUBV(ITTAPI, void, frame_submit_v3, (const __itt_domain *domain, __itt_id *id, __itt_timestamp begin, __itt_timestamp end))
+#define __itt_frame_begin_v3(d,x)      ITTNOTIFY_VOID_D1(frame_begin_v3,d,x)
+#define __itt_frame_begin_v3_ptr       ITTNOTIFY_NAME(frame_begin_v3)
+#define __itt_frame_end_v3(d,x)        ITTNOTIFY_VOID_D1(frame_end_v3,d,x)
+#define __itt_frame_end_v3_ptr         ITTNOTIFY_NAME(frame_end_v3)
+#define __itt_frame_submit_v3(d,x,b,e) ITTNOTIFY_VOID_D3(frame_submit_v3,d,x,b,e)
+#define __itt_frame_submit_v3_ptr      ITTNOTIFY_NAME(frame_submit_v3)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_frame_begin_v3(domain,id)
+#define __itt_frame_begin_v3_ptr 0
+#define __itt_frame_end_v3(domain,id)
+#define __itt_frame_end_v3_ptr   0
+#define __itt_frame_submit_v3(domain,id,begin,end)
+#define __itt_frame_submit_v3_ptr   0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_frame_begin_v3_ptr 0
+#define __itt_frame_end_v3_ptr   0
+#define __itt_frame_submit_v3_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} frames group */
+/** @endcond */
+
+/**
+ * @defgroup taskgroup Task Group
+ * @ingroup public
+ * Task Group
+ * @{
+ */
+/**
+ * @ingroup task_groups
+ * @brief Denotes a task_group instance.
+ * Successive calls to __itt_task_group with the same ID are ignored.
+ * @param[in] domain The domain for this task_group instance
+ * @param[in] id The instance ID for this task_group instance. Must not be __itt_null.
+ * @param[in] parentid The instance ID for the parent of this task_group instance, or __itt_null.
+ * @param[in] name The name of this task_group
+ */
+void ITTAPI __itt_task_group(const __itt_domain *domain, __itt_id id, __itt_id parentid, __itt_string_handle *name);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, task_group, (const __itt_domain *domain, __itt_id id, __itt_id parentid, __itt_string_handle *name))
+#define __itt_task_group(d,x,y,z) ITTNOTIFY_VOID_D3(task_group,d,x,y,z)
+#define __itt_task_group_ptr      ITTNOTIFY_NAME(task_group)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_task_group(d,x,y,z)
+#define __itt_task_group_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_task_group_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} taskgroup group */
+
+/**
+ * @defgroup tasks Tasks
+ * @ingroup public
+ * A task instance represents a piece of work performed by a particular
+ * thread for a period of time. A call to __itt_task_begin creates a
+ * task instance. This becomes the current instance for that task on that
+ * thread. A following call to __itt_task_end on the same thread ends the
+ * instance. There may be multiple simultaneous instances of tasks with the
+ * same name on different threads. If an ID is specified, the task instance
+ * receives that ID. Nested tasks are allowed.
+ *
+ * Note: The task is defined by the bracketing of __itt_task_begin and
+ * __itt_task_end on the same thread. If some scheduling mechanism causes
+ * task switching (the thread executes a different user task) or task
+ * switching (the user task switches to a different thread) then this breaks
+ * the notion of  current instance. Additional API calls are required to
+ * deal with that possibility.
+ * @{
+ */
+
+/**
+ * @ingroup tasks
+ * @brief Begin a task instance.
+ * @param[in] domain The domain for this task
+ * @param[in] taskid The instance ID for this task instance, or __itt_null
+ * @param[in] parentid The parent instance to which this task instance belongs, or __itt_null
+ * @param[in] name The name of this task
+ */
+void ITTAPI __itt_task_begin(const __itt_domain *domain, __itt_id taskid, __itt_id parentid, __itt_string_handle *name);
+
+/**
+ * @ingroup tasks
+ * @brief Begin a task instance.
+ * @param[in] domain The domain for this task
+ * @param[in] taskid The identifier for this task instance (may be 0)
+ * @param[in] parentid The parent of this task (may be 0)
+ * @param[in] fn The pointer to the function you are tracing
+ */
+void ITTAPI __itt_task_begin_fn(const __itt_domain *domain, __itt_id taskid, __itt_id parentid, void* fn);
+
+/**
+ * @ingroup tasks
+ * @brief End the current task instance.
+ * @param[in] domain The domain for this task
+ */
+void ITTAPI __itt_task_end(const __itt_domain *domain);
+
+/**
+ * @ingroup tasks
+ * @brief Begin an overlapped task instance.
+ * @param[in] domain The domain for this task.
+ * @param[in] taskid The identifier for this task instance, *cannot* be __itt_null.
+ * @param[in] parentid The parent of this task, or __itt_null.
+ * @param[in] name The name of this task.
+ */
+void ITTAPI __itt_task_begin_overlapped(const __itt_domain* domain, __itt_id taskid, __itt_id parentid, __itt_string_handle* name);
+
+/**
+ * @ingroup tasks
+ * @brief End an overlapped task instance.
+ * @param[in] domain The domain for this task
+ * @param[in] taskid Explicit ID of finished task
+ */
+void ITTAPI __itt_task_end_overlapped(const __itt_domain *domain, __itt_id taskid);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, task_begin,    (const __itt_domain *domain, __itt_id id, __itt_id parentid, __itt_string_handle *name))
+ITT_STUBV(ITTAPI, void, task_begin_fn, (const __itt_domain *domain, __itt_id id, __itt_id parentid, void* fn))
+ITT_STUBV(ITTAPI, void, task_end,      (const __itt_domain *domain))
+ITT_STUBV(ITTAPI, void, task_begin_overlapped, (const __itt_domain *domain, __itt_id taskid, __itt_id parentid, __itt_string_handle *name))
+ITT_STUBV(ITTAPI, void, task_end_overlapped,   (const __itt_domain *domain, __itt_id taskid))
+#define __itt_task_begin(d,x,y,z)    ITTNOTIFY_VOID_D3(task_begin,d,x,y,z)
+#define __itt_task_begin_ptr         ITTNOTIFY_NAME(task_begin)
+#define __itt_task_begin_fn(d,x,y,z) ITTNOTIFY_VOID_D3(task_begin_fn,d,x,y,z)
+#define __itt_task_begin_fn_ptr      ITTNOTIFY_NAME(task_begin_fn)
+#define __itt_task_end(d)            ITTNOTIFY_VOID_D0(task_end,d)
+#define __itt_task_end_ptr           ITTNOTIFY_NAME(task_end)
+#define __itt_task_begin_overlapped(d,x,y,z) ITTNOTIFY_VOID_D3(task_begin_overlapped,d,x,y,z)
+#define __itt_task_begin_overlapped_ptr      ITTNOTIFY_NAME(task_begin_overlapped)
+#define __itt_task_end_overlapped(d,x)       ITTNOTIFY_VOID_D1(task_end_overlapped,d,x)
+#define __itt_task_end_overlapped_ptr        ITTNOTIFY_NAME(task_end_overlapped)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_task_begin(domain,id,parentid,name)
+#define __itt_task_begin_ptr    0
+#define __itt_task_begin_fn(domain,id,parentid,fn)
+#define __itt_task_begin_fn_ptr 0
+#define __itt_task_end(domain)
+#define __itt_task_end_ptr      0
+#define __itt_task_begin_overlapped(domain,taskid,parentid,name)
+#define __itt_task_begin_overlapped_ptr         0
+#define __itt_task_end_overlapped(domain,taskid)
+#define __itt_task_end_overlapped_ptr           0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_task_begin_ptr    0
+#define __itt_task_begin_fn_ptr 0
+#define __itt_task_end_ptr      0
+#define __itt_task_begin_overlapped_ptr 0
+#define __itt_task_end_overlapped_ptr   0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} tasks group */
+
+
+/**
+ * @defgroup markers Markers
+ * Markers represent a single discreet event in time. Markers have a scope,
+ * described by an enumerated type __itt_scope. Markers are created by
+ * the API call __itt_marker. A marker instance can be given an ID for use in
+ * adding metadata.
+ * @{
+ */
+
+/**
+ * @brief Describes the scope of an event object in the trace.
+ */
+typedef enum
+{
+    __itt_scope_unknown = 0,
+    __itt_scope_global,
+    __itt_scope_track_group,
+    __itt_scope_track,
+    __itt_scope_task,
+    __itt_scope_marker
+} __itt_scope;
+
+/** @cond exclude_from_documentation */
+#define __itt_marker_scope_unknown  __itt_scope_unknown
+#define __itt_marker_scope_global   __itt_scope_global
+#define __itt_marker_scope_process  __itt_scope_track_group
+#define __itt_marker_scope_thread   __itt_scope_track
+#define __itt_marker_scope_task     __itt_scope_task
+/** @endcond */
+
+/**
+ * @ingroup markers
+ * @brief Create a marker instance
+ * @param[in] domain The domain for this marker
+ * @param[in] id The instance ID for this marker or __itt_null
+ * @param[in] name The name for this marker
+ * @param[in] scope The scope for this marker
+ */
+void ITTAPI __itt_marker(const __itt_domain *domain, __itt_id id, __itt_string_handle *name, __itt_scope scope);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, marker, (const __itt_domain *domain, __itt_id id, __itt_string_handle *name, __itt_scope scope))
+#define __itt_marker(d,x,y,z) ITTNOTIFY_VOID_D3(marker,d,x,y,z)
+#define __itt_marker_ptr      ITTNOTIFY_NAME(marker)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_marker(domain,id,name,scope)
+#define __itt_marker_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_marker_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} markers group */
+
+/**
+ * @defgroup metadata Metadata
+ * The metadata API is used to attach extra information to named
+ * entities. Metadata can be attached to an identified named entity by ID,
+ * or to the current entity (which is always a task).
+ *
+ * Conceptually metadata has a type (what kind of metadata), a key (the
+ * name of the metadata), and a value (the actual data). The encoding of
+ * the value depends on the type of the metadata.
+ *
+ * The type of metadata is specified by an enumerated type __itt_metdata_type.
+ * @{
+ */
+
+/**
+ * @ingroup parameters
+ * @brief describes the type of metadata
+ */
+typedef enum {
+    __itt_metadata_unknown = 0,
+    __itt_metadata_u64,     /**< Unsigned 64-bit integer */
+    __itt_metadata_s64,     /**< Signed 64-bit integer */
+    __itt_metadata_u32,     /**< Unsigned 32-bit integer */
+    __itt_metadata_s32,     /**< Signed 32-bit integer */
+    __itt_metadata_u16,     /**< Unsigned 16-bit integer */
+    __itt_metadata_s16,     /**< Signed 16-bit integer */
+    __itt_metadata_float,   /**< Signed 32-bit floating-point */
+    __itt_metadata_double   /**< SIgned 64-bit floating-point */
+} __itt_metadata_type;
+
+/**
+ * @ingroup parameters
+ * @brief Add metadata to an instance of a named entity.
+ * @param[in] domain The domain controlling the call
+ * @param[in] id The identifier of the instance to which the metadata is to be added, or __itt_null to add to the current task
+ * @param[in] key The name of the metadata
+ * @param[in] type The type of the metadata
+ * @param[in] count The number of elements of the given type. If count == 0, no metadata will be added.
+ * @param[in] data The metadata itself
+*/
+void ITTAPI __itt_metadata_add(const __itt_domain *domain, __itt_id id, __itt_string_handle *key, __itt_metadata_type type, size_t count, void *data);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, metadata_add, (const __itt_domain *domain, __itt_id id, __itt_string_handle *key, __itt_metadata_type type, size_t count, void *data))
+#define __itt_metadata_add(d,x,y,z,a,b) ITTNOTIFY_VOID_D5(metadata_add,d,x,y,z,a,b)
+#define __itt_metadata_add_ptr          ITTNOTIFY_NAME(metadata_add)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_metadata_add(d,x,y,z,a,b)
+#define __itt_metadata_add_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_metadata_add_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup parameters
+ * @brief Add string metadata to an instance of a named entity.
+ * @param[in] domain The domain controlling the call
+ * @param[in] id The identifier of the instance to which the metadata is to be added, or __itt_null to add to the current task
+ * @param[in] key The name of the metadata
+ * @param[in] data The metadata itself
+ * @param[in] length The number of characters in the string, or -1 if the length is unknown but the string is null-terminated
+*/
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+void ITTAPI __itt_metadata_str_addA(const __itt_domain *domain, __itt_id id, __itt_string_handle *key, const char *data, size_t length);
+void ITTAPI __itt_metadata_str_addW(const __itt_domain *domain, __itt_id id, __itt_string_handle *key, const wchar_t *data, size_t length);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_metadata_str_add     __itt_metadata_str_addW
+#  define __itt_metadata_str_add_ptr __itt_metadata_str_addW_ptr
+#else /* UNICODE */
+#  define __itt_metadata_str_add     __itt_metadata_str_addA
+#  define __itt_metadata_str_add_ptr __itt_metadata_str_addA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+void ITTAPI __itt_metadata_str_add(const __itt_domain *domain, __itt_id id, __itt_string_handle *key, const char *data, size_t length);
+#endif
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUBV(ITTAPI, void, metadata_str_addA, (const __itt_domain *domain, __itt_id id, __itt_string_handle *key, const char *data, size_t length))
+ITT_STUBV(ITTAPI, void, metadata_str_addW, (const __itt_domain *domain, __itt_id id, __itt_string_handle *key, const wchar_t *data, size_t length))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUBV(ITTAPI, void, metadata_str_add, (const __itt_domain *domain, __itt_id id, __itt_string_handle *key, const char *data, size_t length))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_metadata_str_addA(d,x,y,z,a) ITTNOTIFY_VOID_D4(metadata_str_addA,d,x,y,z,a)
+#define __itt_metadata_str_addA_ptr        ITTNOTIFY_NAME(metadata_str_addA)
+#define __itt_metadata_str_addW(d,x,y,z,a) ITTNOTIFY_VOID_D4(metadata_str_addW,d,x,y,z,a)
+#define __itt_metadata_str_addW_ptr        ITTNOTIFY_NAME(metadata_str_addW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_metadata_str_add(d,x,y,z,a)  ITTNOTIFY_VOID_D4(metadata_str_add,d,x,y,z,a)
+#define __itt_metadata_str_add_ptr         ITTNOTIFY_NAME(metadata_str_add)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_metadata_str_addA(d,x,y,z,a)
+#define __itt_metadata_str_addA_ptr 0
+#define __itt_metadata_str_addW(d,x,y,z,a)
+#define __itt_metadata_str_addW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_metadata_str_add(d,x,y,z,a)
+#define __itt_metadata_str_add_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_metadata_str_addA_ptr 0
+#define __itt_metadata_str_addW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_metadata_str_add_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup parameters
+ * @brief Add metadata to an instance of a named entity.
+ * @param[in] domain The domain controlling the call
+ * @param[in] scope The scope of the instance to which the metadata is to be added
+
+ * @param[in] id The identifier of the instance to which the metadata is to be added, or __itt_null to add to the current task
+
+ * @param[in] key The name of the metadata
+ * @param[in] type The type of the metadata
+ * @param[in] count The number of elements of the given type. If count == 0, no metadata will be added.
+ * @param[in] data The metadata itself
+*/
+void ITTAPI __itt_metadata_add_with_scope(const __itt_domain *domain, __itt_scope scope, __itt_string_handle *key, __itt_metadata_type type, size_t count, void *data);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, metadata_add_with_scope, (const __itt_domain *domain, __itt_scope scope, __itt_string_handle *key, __itt_metadata_type type, size_t count, void *data))
+#define __itt_metadata_add_with_scope(d,x,y,z,a,b) ITTNOTIFY_VOID_D5(metadata_add_with_scope,d,x,y,z,a,b)
+#define __itt_metadata_add_with_scope_ptr          ITTNOTIFY_NAME(metadata_add_with_scope)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_metadata_add_with_scope(d,x,y,z,a,b)
+#define __itt_metadata_add_with_scope_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_metadata_add_with_scope_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup parameters
+ * @brief Add string metadata to an instance of a named entity.
+ * @param[in] domain The domain controlling the call
+ * @param[in] scope The scope of the instance to which the metadata is to be added
+
+ * @param[in] id The identifier of the instance to which the metadata is to be added, or __itt_null to add to the current task
+
+ * @param[in] key The name of the metadata
+ * @param[in] data The metadata itself
+ * @param[in] length The number of characters in the string, or -1 if the length is unknown but the string is null-terminated
+*/
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+void ITTAPI __itt_metadata_str_add_with_scopeA(const __itt_domain *domain, __itt_scope scope, __itt_string_handle *key, const char *data, size_t length);
+void ITTAPI __itt_metadata_str_add_with_scopeW(const __itt_domain *domain, __itt_scope scope, __itt_string_handle *key, const wchar_t *data, size_t length);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_metadata_str_add_with_scope     __itt_metadata_str_add_with_scopeW
+#  define __itt_metadata_str_add_with_scope_ptr __itt_metadata_str_add_with_scopeW_ptr
+#else /* UNICODE */
+#  define __itt_metadata_str_add_with_scope     __itt_metadata_str_add_with_scopeA
+#  define __itt_metadata_str_add_with_scope_ptr __itt_metadata_str_add_with_scopeA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+void ITTAPI __itt_metadata_str_add_with_scope(const __itt_domain *domain, __itt_scope scope, __itt_string_handle *key, const char *data, size_t length);
+#endif
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUBV(ITTAPI, void, metadata_str_add_with_scopeA, (const __itt_domain *domain, __itt_scope scope, __itt_string_handle *key, const char *data, size_t length))
+ITT_STUBV(ITTAPI, void, metadata_str_add_with_scopeW, (const __itt_domain *domain, __itt_scope scope, __itt_string_handle *key, const wchar_t *data, size_t length))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUBV(ITTAPI, void, metadata_str_add_with_scope, (const __itt_domain *domain, __itt_scope scope, __itt_string_handle *key, const char *data, size_t length))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_metadata_str_add_with_scopeA(d,x,y,z,a) ITTNOTIFY_VOID_D4(metadata_str_add_with_scopeA,d,x,y,z,a)
+#define __itt_metadata_str_add_with_scopeA_ptr        ITTNOTIFY_NAME(metadata_str_add_with_scopeA)
+#define __itt_metadata_str_add_with_scopeW(d,x,y,z,a) ITTNOTIFY_VOID_D4(metadata_str_add_with_scopeW,d,x,y,z,a)
+#define __itt_metadata_str_add_with_scopeW_ptr        ITTNOTIFY_NAME(metadata_str_add_with_scopeW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_metadata_str_add_with_scope(d,x,y,z,a)  ITTNOTIFY_VOID_D4(metadata_str_add_with_scope,d,x,y,z,a)
+#define __itt_metadata_str_add_with_scope_ptr         ITTNOTIFY_NAME(metadata_str_add_with_scope)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_metadata_str_add_with_scopeA(d,x,y,z,a)
+#define __itt_metadata_str_add_with_scopeA_ptr  0
+#define __itt_metadata_str_add_with_scopeW(d,x,y,z,a)
+#define __itt_metadata_str_add_with_scopeW_ptr  0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_metadata_str_add_with_scope(d,x,y,z,a)
+#define __itt_metadata_str_add_with_scope_ptr   0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_metadata_str_add_with_scopeA_ptr  0
+#define __itt_metadata_str_add_with_scopeW_ptr  0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_metadata_str_add_with_scope_ptr   0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @} metadata group */
+
+/**
+ * @defgroup relations Relations
+ * Instances of named entities can be explicitly associated with other
+ * instances using instance IDs and the relationship API calls.
+ *
+ * @{
+ */
+
+/**
+ * @ingroup relations
+ * @brief The kind of relation between two instances is specified by the enumerated type __itt_relation.
+ * Relations between instances can be added with an API call. The relation
+ * API uses instance IDs. Relations can be added before or after the actual
+ * instances are created and persist independently of the instances. This
+ * is the motivation for having different lifetimes for instance IDs and
+ * the actual instances.
+ */
+typedef enum
+{
+    __itt_relation_is_unknown = 0,
+    __itt_relation_is_dependent_on,         /**< "A is dependent on B" means that A cannot start until B completes */
+    __itt_relation_is_sibling_of,           /**< "A is sibling of B" means that A and B were created as a group */
+    __itt_relation_is_parent_of,            /**< "A is parent of B" means that A created B */
+    __itt_relation_is_continuation_of,      /**< "A is continuation of B" means that A assumes the dependencies of B */
+    __itt_relation_is_child_of,             /**< "A is child of B" means that A was created by B (inverse of is_parent_of) */
+    __itt_relation_is_continued_by,         /**< "A is continued by B" means that B assumes the dependencies of A (inverse of is_continuation_of) */
+    __itt_relation_is_predecessor_to        /**< "A is predecessor to B" means that B cannot start until A completes (inverse of is_dependent_on) */
+} __itt_relation;
+
+/**
+ * @ingroup relations
+ * @brief Add a relation to the current task instance.
+ * The current task instance is the head of the relation.
+ * @param[in] domain The domain controlling this call
+ * @param[in] relation The kind of relation
+ * @param[in] tail The ID for the tail of the relation
+ */
+void ITTAPI __itt_relation_add_to_current(const __itt_domain *domain, __itt_relation relation, __itt_id tail);
+
+/**
+ * @ingroup relations
+ * @brief Add a relation between two instance identifiers.
+ * @param[in] domain The domain controlling this call
+ * @param[in] head The ID for the head of the relation
+ * @param[in] relation The kind of relation
+ * @param[in] tail The ID for the tail of the relation
+ */
+void ITTAPI __itt_relation_add(const __itt_domain *domain, __itt_id head, __itt_relation relation, __itt_id tail);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, relation_add_to_current, (const __itt_domain *domain, __itt_relation relation, __itt_id tail))
+ITT_STUBV(ITTAPI, void, relation_add,            (const __itt_domain *domain, __itt_id head, __itt_relation relation, __itt_id tail))
+#define __itt_relation_add_to_current(d,x,y) ITTNOTIFY_VOID_D2(relation_add_to_current,d,x,y)
+#define __itt_relation_add_to_current_ptr    ITTNOTIFY_NAME(relation_add_to_current)
+#define __itt_relation_add(d,x,y,z)          ITTNOTIFY_VOID_D3(relation_add,d,x,y,z)
+#define __itt_relation_add_ptr               ITTNOTIFY_NAME(relation_add)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_relation_add_to_current(d,x,y)
+#define __itt_relation_add_to_current_ptr 0
+#define __itt_relation_add(d,x,y,z)
+#define __itt_relation_add_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_relation_add_to_current_ptr 0
+#define __itt_relation_add_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} relations group */
+
+/** @cond exclude_from_documentation */
+#pragma pack(push, 8)
+
+typedef struct ___itt_clock_info
+{
+    unsigned long long clock_freq; /*!< Clock domain frequency */
+    unsigned long long clock_base; /*!< Clock domain base timestamp */
+} __itt_clock_info;
+
+#pragma pack(pop)
+/** @endcond */
+
+/** @cond exclude_from_documentation */
+typedef void (ITTAPI *__itt_get_clock_info_fn)(__itt_clock_info* clock_info, void* data);
+/** @endcond */
+
+/** @cond exclude_from_documentation */
+#pragma pack(push, 8)
+
+typedef struct ___itt_clock_domain
+{
+    __itt_clock_info info;      /*!< Most recent clock domain info */
+    __itt_get_clock_info_fn fn; /*!< Callback function pointer */
+    void* fn_data;              /*!< Input argument for the callback function */
+    int   extra1;               /*!< Reserved. Must be zero */
+    void* extra2;               /*!< Reserved. Must be zero */
+    struct ___itt_clock_domain* next;
+} __itt_clock_domain;
+
+#pragma pack(pop)
+/** @endcond */
+
+/**
+ * @ingroup clockdomains
+ * @brief Create a clock domain.
+ * Certain applications require the capability to trace their application using
+ * a clock domain different than the CPU, for instance the instrumentation of events
+ * that occur on a GPU.
+ * Because the set of domains is expected to be static over the application's execution time,
+ * there is no mechanism to destroy a domain.
+ * Any domain can be accessed by any thread in the process, regardless of which thread created
+ * the domain. This call is thread-safe.
+ * @param[in] fn A pointer to a callback function which retrieves alternative CPU timestamps
+ * @param[in] fn_data Argument for a callback function; may be NULL
+ */
+__itt_clock_domain* ITTAPI __itt_clock_domain_create(__itt_get_clock_info_fn fn, void* fn_data);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(ITTAPI, __itt_clock_domain*, clock_domain_create, (__itt_get_clock_info_fn fn, void* fn_data))
+#define __itt_clock_domain_create     ITTNOTIFY_DATA(clock_domain_create)
+#define __itt_clock_domain_create_ptr ITTNOTIFY_NAME(clock_domain_create)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_clock_domain_create(fn,fn_data) (__itt_clock_domain*)0
+#define __itt_clock_domain_create_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_clock_domain_create_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup clockdomains
+ * @brief Recalculate clock domains frequences and clock base timestamps.
+ */
+void ITTAPI __itt_clock_domain_reset(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, clock_domain_reset, (void))
+#define __itt_clock_domain_reset     ITTNOTIFY_VOID(clock_domain_reset)
+#define __itt_clock_domain_reset_ptr ITTNOTIFY_NAME(clock_domain_reset)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_clock_domain_reset()
+#define __itt_clock_domain_reset_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_clock_domain_reset_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup clockdomain
+ * @brief Create an instance of identifier. This establishes the beginning of the lifetime of
+ * an instance of the given ID in the trace. Once this lifetime starts, the ID can be used to
+ * tag named entity instances in calls such as __itt_task_begin, and to specify relationships among
+ * identified named entity instances, using the \ref relations APIs.
+ * @param[in] domain The domain controlling the execution of this call.
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] id The ID to create.
+ */
+void ITTAPI __itt_id_create_ex(const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id id);
+
+/**
+ * @ingroup clockdomain
+ * @brief Destroy an instance of identifier. This ends the lifetime of the current instance of the
+ * given ID value in the trace. Any relationships that are established after this lifetime ends are
+ * invalid. This call must be performed before the given ID value can be reused for a different
+ * named entity instance.
+ * @param[in] domain The domain controlling the execution of this call.
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] id The ID to destroy.
+ */
+void ITTAPI __itt_id_destroy_ex(const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id id);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, id_create_ex,  (const __itt_domain *domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id id))
+ITT_STUBV(ITTAPI, void, id_destroy_ex, (const __itt_domain *domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id id))
+#define __itt_id_create_ex(d,x,y,z)  ITTNOTIFY_VOID_D3(id_create_ex,d,x,y,z)
+#define __itt_id_create_ex_ptr       ITTNOTIFY_NAME(id_create_ex)
+#define __itt_id_destroy_ex(d,x,y,z) ITTNOTIFY_VOID_D3(id_destroy_ex,d,x,y,z)
+#define __itt_id_destroy_ex_ptr      ITTNOTIFY_NAME(id_destroy_ex)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_id_create_ex(domain,clock_domain,timestamp,id)
+#define __itt_id_create_ex_ptr    0
+#define __itt_id_destroy_ex(domain,clock_domain,timestamp,id)
+#define __itt_id_destroy_ex_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_id_create_ex_ptr    0
+#define __itt_id_destroy_ex_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup clockdomain
+ * @brief Begin a task instance.
+ * @param[in] domain The domain for this task
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] taskid The instance ID for this task instance, or __itt_null
+ * @param[in] parentid The parent instance to which this task instance belongs, or __itt_null
+ * @param[in] name The name of this task
+ */
+void ITTAPI __itt_task_begin_ex(const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id taskid, __itt_id parentid, __itt_string_handle* name);
+
+/**
+ * @ingroup clockdomain
+ * @brief Begin a task instance.
+ * @param[in] domain The domain for this task
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] taskid The identifier for this task instance, or __itt_null
+ * @param[in] parentid The parent of this task, or __itt_null
+ * @param[in] fn The pointer to the function you are tracing
+ */
+void ITTAPI __itt_task_begin_fn_ex(const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id taskid, __itt_id parentid, void* fn);
+
+/**
+ * @ingroup clockdomain
+ * @brief End the current task instance.
+ * @param[in] domain The domain for this task
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ */
+void ITTAPI __itt_task_end_ex(const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, task_begin_ex,        (const __itt_domain *domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id id, __itt_id parentid, __itt_string_handle *name))
+ITT_STUBV(ITTAPI, void, task_begin_fn_ex,     (const __itt_domain *domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id id, __itt_id parentid, void* fn))
+ITT_STUBV(ITTAPI, void, task_end_ex,          (const __itt_domain *domain, __itt_clock_domain* clock_domain, unsigned long long timestamp))
+#define __itt_task_begin_ex(d,x,y,z,a,b)      ITTNOTIFY_VOID_D5(task_begin_ex,d,x,y,z,a,b)
+#define __itt_task_begin_ex_ptr               ITTNOTIFY_NAME(task_begin_ex)
+#define __itt_task_begin_fn_ex(d,x,y,z,a,b)   ITTNOTIFY_VOID_D5(task_begin_fn_ex,d,x,y,z,a,b)
+#define __itt_task_begin_fn_ex_ptr            ITTNOTIFY_NAME(task_begin_fn_ex)
+#define __itt_task_end_ex(d,x,y)              ITTNOTIFY_VOID_D2(task_end_ex,d,x,y)
+#define __itt_task_end_ex_ptr                 ITTNOTIFY_NAME(task_end_ex)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_task_begin_ex(domain,clock_domain,timestamp,id,parentid,name)
+#define __itt_task_begin_ex_ptr          0
+#define __itt_task_begin_fn_ex(domain,clock_domain,timestamp,id,parentid,fn)
+#define __itt_task_begin_fn_ex_ptr       0
+#define __itt_task_end_ex(domain,clock_domain,timestamp)
+#define __itt_task_end_ex_ptr            0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_task_begin_ex_ptr          0
+#define __itt_task_begin_fn_ex_ptr       0
+#define __itt_task_end_ex_ptr            0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @defgroup counters Counters
+ * @ingroup public
+ * Counters are user-defined objects with a monotonically increasing
+ * value. Counter values are 64-bit unsigned integers.
+ * Counters have names that can be displayed in
+ * the tools.
+ * @{
+ */
+
+/**
+ * @brief opaque structure for counter identification
+ */
+/** @cond exclude_from_documentation */
+
+typedef struct ___itt_counter* __itt_counter;
+
+/**
+ * @brief Create an unsigned 64 bits integer counter with given name/domain
+ *
+ * After __itt_counter_create() is called, __itt_counter_inc(id), __itt_counter_inc_delta(id, delta),
+ * __itt_counter_set_value(id, value_ptr) or __itt_counter_set_value_ex(id, clock_domain, timestamp, value_ptr)
+ * can be used to change the value of the counter, where value_ptr is a pointer to an unsigned 64 bits integer
+ *
+ * The call is equal to __itt_counter_create_typed(name, domain, __itt_metadata_u64)
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+__itt_counter ITTAPI __itt_counter_createA(const char    *name, const char    *domain);
+__itt_counter ITTAPI __itt_counter_createW(const wchar_t *name, const wchar_t *domain);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_counter_create     __itt_counter_createW
+#  define __itt_counter_create_ptr __itt_counter_createW_ptr
+#else /* UNICODE */
+#  define __itt_counter_create     __itt_counter_createA
+#  define __itt_counter_create_ptr __itt_counter_createA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+__itt_counter ITTAPI __itt_counter_create(const char *name, const char *domain);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, __itt_counter, counter_createA, (const char    *name, const char    *domain))
+ITT_STUB(ITTAPI, __itt_counter, counter_createW, (const wchar_t *name, const wchar_t *domain))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, __itt_counter, counter_create,  (const char *name, const char *domain))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_counter_createA     ITTNOTIFY_DATA(counter_createA)
+#define __itt_counter_createA_ptr ITTNOTIFY_NAME(counter_createA)
+#define __itt_counter_createW     ITTNOTIFY_DATA(counter_createW)
+#define __itt_counter_createW_ptr ITTNOTIFY_NAME(counter_createW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_counter_create     ITTNOTIFY_DATA(counter_create)
+#define __itt_counter_create_ptr ITTNOTIFY_NAME(counter_create)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_counter_createA(name, domain)
+#define __itt_counter_createA_ptr 0
+#define __itt_counter_createW(name, domain)
+#define __itt_counter_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_counter_create(name, domain)
+#define __itt_counter_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_counter_createA_ptr 0
+#define __itt_counter_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_counter_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Increment the unsigned 64 bits integer counter value
+ *
+ * Calling this function to non-unsigned 64 bits integer counters has no effect
+ */
+void ITTAPI __itt_counter_inc(__itt_counter id);
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_inc, (__itt_counter id))
+#define __itt_counter_inc     ITTNOTIFY_VOID(counter_inc)
+#define __itt_counter_inc_ptr ITTNOTIFY_NAME(counter_inc)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_inc(id)
+#define __itt_counter_inc_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_inc_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/**
+ * @brief Increment the unsigned 64 bits integer counter value with x
+ *
+ * Calling this function to non-unsigned 64 bits integer counters has no effect
+ */
+void ITTAPI __itt_counter_inc_delta(__itt_counter id, unsigned long long value);
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_inc_delta, (__itt_counter id, unsigned long long value))
+#define __itt_counter_inc_delta     ITTNOTIFY_VOID(counter_inc_delta)
+#define __itt_counter_inc_delta_ptr ITTNOTIFY_NAME(counter_inc_delta)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_inc_delta(id, value)
+#define __itt_counter_inc_delta_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_inc_delta_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Decrement the unsigned 64 bits integer counter value
+ *
+ * Calling this function to non-unsigned 64 bits integer counters has no effect
+ */
+void ITTAPI __itt_counter_dec(__itt_counter id);
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_dec, (__itt_counter id))
+#define __itt_counter_dec     ITTNOTIFY_VOID(counter_dec)
+#define __itt_counter_dec_ptr ITTNOTIFY_NAME(counter_dec)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_dec(id)
+#define __itt_counter_dec_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_dec_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/**
+ * @brief Decrement the unsigned 64 bits integer counter value with x
+ *
+ * Calling this function to non-unsigned 64 bits integer counters has no effect
+ */
+void ITTAPI __itt_counter_dec_delta(__itt_counter id, unsigned long long value);
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_dec_delta, (__itt_counter id, unsigned long long value))
+#define __itt_counter_dec_delta     ITTNOTIFY_VOID(counter_dec_delta)
+#define __itt_counter_dec_delta_ptr ITTNOTIFY_NAME(counter_dec_delta)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_dec_delta(id, value)
+#define __itt_counter_dec_delta_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_dec_delta_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup counters
+ * @brief Increment a counter by one.
+ * The first call with a given name creates a counter by that name and sets its
+ * value to zero. Successive calls increment the counter value.
+ * @param[in] domain The domain controlling the call. Counter names are not domain specific.
+ *            The domain argument is used only to enable or disable the API calls.
+ * @param[in] name The name of the counter
+ */
+void ITTAPI __itt_counter_inc_v3(const __itt_domain *domain, __itt_string_handle *name);
+
+/**
+ * @ingroup counters
+ * @brief Increment a counter by the value specified in delta.
+ * @param[in] domain The domain controlling the call. Counter names are not domain specific.
+ *            The domain argument is used only to enable or disable the API calls.
+ * @param[in] name The name of the counter
+ * @param[in] delta The amount by which to increment the counter
+ */
+void ITTAPI __itt_counter_inc_delta_v3(const __itt_domain *domain, __itt_string_handle *name, unsigned long long delta);
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_inc_v3,       (const __itt_domain *domain, __itt_string_handle *name))
+ITT_STUBV(ITTAPI, void, counter_inc_delta_v3, (const __itt_domain *domain, __itt_string_handle *name, unsigned long long delta))
+#define __itt_counter_inc_v3(d,x)         ITTNOTIFY_VOID_D1(counter_inc_v3,d,x)
+#define __itt_counter_inc_v3_ptr          ITTNOTIFY_NAME(counter_inc_v3)
+#define __itt_counter_inc_delta_v3(d,x,y) ITTNOTIFY_VOID_D2(counter_inc_delta_v3,d,x,y)
+#define __itt_counter_inc_delta_v3_ptr    ITTNOTIFY_NAME(counter_inc_delta_v3)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_inc_v3(domain,name)
+#define __itt_counter_inc_v3_ptr       0
+#define __itt_counter_inc_delta_v3(domain,name,delta)
+#define __itt_counter_inc_delta_v3_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_inc_v3_ptr       0
+#define __itt_counter_inc_delta_v3_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+
+/**
+ * @ingroup counters
+ * @brief Decrement a counter by one.
+ * The first call with a given name creates a counter by that name and sets its
+ * value to zero. Successive calls decrement the counter value.
+ * @param[in] domain The domain controlling the call. Counter names are not domain specific.
+ *            The domain argument is used only to enable or disable the API calls.
+ * @param[in] name The name of the counter
+ */
+void ITTAPI __itt_counter_dec_v3(const __itt_domain *domain, __itt_string_handle *name);
+
+/**
+ * @ingroup counters
+ * @brief Decrement a counter by the value specified in delta.
+ * @param[in] domain The domain controlling the call. Counter names are not domain specific.
+ *            The domain argument is used only to enable or disable the API calls.
+ * @param[in] name The name of the counter
+ * @param[in] delta The amount by which to decrement the counter
+ */
+void ITTAPI __itt_counter_dec_delta_v3(const __itt_domain *domain, __itt_string_handle *name, unsigned long long delta);
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_dec_v3,       (const __itt_domain *domain, __itt_string_handle *name))
+ITT_STUBV(ITTAPI, void, counter_dec_delta_v3, (const __itt_domain *domain, __itt_string_handle *name, unsigned long long delta))
+#define __itt_counter_dec_v3(d,x)         ITTNOTIFY_VOID_D1(counter_dec_v3,d,x)
+#define __itt_counter_dec_v3_ptr          ITTNOTIFY_NAME(counter_dec_v3)
+#define __itt_counter_dec_delta_v3(d,x,y) ITTNOTIFY_VOID_D2(counter_dec_delta_v3,d,x,y)
+#define __itt_counter_dec_delta_v3_ptr    ITTNOTIFY_NAME(counter_dec_delta_v3)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_dec_v3(domain,name)
+#define __itt_counter_dec_v3_ptr       0
+#define __itt_counter_dec_delta_v3(domain,name,delta)
+#define __itt_counter_dec_delta_v3_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_dec_v3_ptr       0
+#define __itt_counter_dec_delta_v3_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @} counters group */
+
+
+/**
+ * @brief Set the counter value
+ */
+void ITTAPI __itt_counter_set_value(__itt_counter id, void *value_ptr);
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_set_value, (__itt_counter id, void *value_ptr))
+#define __itt_counter_set_value     ITTNOTIFY_VOID(counter_set_value)
+#define __itt_counter_set_value_ptr ITTNOTIFY_NAME(counter_set_value)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_set_value(id, value_ptr)
+#define __itt_counter_set_value_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_set_value_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Set the counter value
+ */
+void ITTAPI __itt_counter_set_value_ex(__itt_counter id, __itt_clock_domain *clock_domain, unsigned long long timestamp, void *value_ptr);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_set_value_ex, (__itt_counter id, __itt_clock_domain *clock_domain, unsigned long long timestamp, void *value_ptr))
+#define __itt_counter_set_value_ex     ITTNOTIFY_VOID(counter_set_value_ex)
+#define __itt_counter_set_value_ex_ptr ITTNOTIFY_NAME(counter_set_value_ex)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_set_value_ex(id, clock_domain, timestamp, value_ptr)
+#define __itt_counter_set_value_ex_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_set_value_ex_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Create a typed counter with given name/domain
+ *
+ * After __itt_counter_create_typed() is called, __itt_counter_inc(id), __itt_counter_inc_delta(id, delta),
+ * __itt_counter_set_value(id, value_ptr) or __itt_counter_set_value_ex(id, clock_domain, timestamp, value_ptr)
+ * can be used to change the value of the counter
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+__itt_counter ITTAPI __itt_counter_create_typedA(const char    *name, const char    *domain, __itt_metadata_type type);
+__itt_counter ITTAPI __itt_counter_create_typedW(const wchar_t *name, const wchar_t *domain, __itt_metadata_type type);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_counter_create_typed     __itt_counter_create_typedW
+#  define __itt_counter_create_typed_ptr __itt_counter_create_typedW_ptr
+#else /* UNICODE */
+#  define __itt_counter_create_typed     __itt_counter_create_typedA
+#  define __itt_counter_create_typed_ptr __itt_counter_create_typedA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+__itt_counter ITTAPI __itt_counter_create_typed(const char *name, const char *domain, __itt_metadata_type type);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, __itt_counter, counter_create_typedA, (const char    *name, const char    *domain, __itt_metadata_type type))
+ITT_STUB(ITTAPI, __itt_counter, counter_create_typedW, (const wchar_t *name, const wchar_t *domain, __itt_metadata_type type))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, __itt_counter, counter_create_typed,  (const char *name, const char *domain, __itt_metadata_type type))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_counter_create_typedA     ITTNOTIFY_DATA(counter_create_typedA)
+#define __itt_counter_create_typedA_ptr ITTNOTIFY_NAME(counter_create_typedA)
+#define __itt_counter_create_typedW     ITTNOTIFY_DATA(counter_create_typedW)
+#define __itt_counter_create_typedW_ptr ITTNOTIFY_NAME(counter_create_typedW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_counter_create_typed     ITTNOTIFY_DATA(counter_create_typed)
+#define __itt_counter_create_typed_ptr ITTNOTIFY_NAME(counter_create_typed)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_counter_create_typedA(name, domain, type)
+#define __itt_counter_create_typedA_ptr 0
+#define __itt_counter_create_typedW(name, domain, type)
+#define __itt_counter_create_typedW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_counter_create_typed(name, domain, type)
+#define __itt_counter_create_typed_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_counter_create_typedA_ptr 0
+#define __itt_counter_create_typedW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_counter_create_typed_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Destroy the counter identified by the pointer previously returned by __itt_counter_create() or
+ * __itt_counter_create_typed()
+ */
+void ITTAPI __itt_counter_destroy(__itt_counter id);
+
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, counter_destroy, (__itt_counter id))
+#define __itt_counter_destroy     ITTNOTIFY_VOID(counter_destroy)
+#define __itt_counter_destroy_ptr ITTNOTIFY_NAME(counter_destroy)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_counter_destroy(id)
+#define __itt_counter_destroy_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_counter_destroy_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} counters group */
+
+/**
+ * @ingroup markers
+ * @brief Create a marker instance.
+ * @param[in] domain The domain for this marker
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] id The instance ID for this marker, or __itt_null
+ * @param[in] name The name for this marker
+ * @param[in] scope The scope for this marker
+ */
+void ITTAPI __itt_marker_ex(const __itt_domain *domain,  __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id id, __itt_string_handle *name, __itt_scope scope);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, marker_ex,    (const __itt_domain *domain,  __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id id, __itt_string_handle *name, __itt_scope scope))
+#define __itt_marker_ex(d,x,y,z,a,b)    ITTNOTIFY_VOID_D5(marker_ex,d,x,y,z,a,b)
+#define __itt_marker_ex_ptr             ITTNOTIFY_NAME(marker_ex)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_marker_ex(domain,clock_domain,timestamp,id,name,scope)
+#define __itt_marker_ex_ptr    0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_marker_ex_ptr    0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @ingroup clockdomain
+ * @brief Add a relation to the current task instance.
+ * The current task instance is the head of the relation.
+ * @param[in] domain The domain controlling this call
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] relation The kind of relation
+ * @param[in] tail The ID for the tail of the relation
+ */
+void ITTAPI __itt_relation_add_to_current_ex(const __itt_domain *domain,  __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_relation relation, __itt_id tail);
+
+/**
+ * @ingroup clockdomain
+ * @brief Add a relation between two instance identifiers.
+ * @param[in] domain The domain controlling this call
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] head The ID for the head of the relation
+ * @param[in] relation The kind of relation
+ * @param[in] tail The ID for the tail of the relation
+ */
+void ITTAPI __itt_relation_add_ex(const __itt_domain *domain,  __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id head, __itt_relation relation, __itt_id tail);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, relation_add_to_current_ex, (const __itt_domain *domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_relation relation, __itt_id tail))
+ITT_STUBV(ITTAPI, void, relation_add_ex,            (const __itt_domain *domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id head, __itt_relation relation, __itt_id tail))
+#define __itt_relation_add_to_current_ex(d,x,y,z,a) ITTNOTIFY_VOID_D4(relation_add_to_current_ex,d,x,y,z,a)
+#define __itt_relation_add_to_current_ex_ptr        ITTNOTIFY_NAME(relation_add_to_current_ex)
+#define __itt_relation_add_ex(d,x,y,z,a,b)          ITTNOTIFY_VOID_D5(relation_add_ex,d,x,y,z,a,b)
+#define __itt_relation_add_ex_ptr                   ITTNOTIFY_NAME(relation_add_ex)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_relation_add_to_current_ex(domain,clock_domain,timestame,relation,tail)
+#define __itt_relation_add_to_current_ex_ptr 0
+#define __itt_relation_add_ex(domain,clock_domain,timestamp,head,relation,tail)
+#define __itt_relation_add_ex_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_relation_add_to_current_ex_ptr 0
+#define __itt_relation_add_ex_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @cond exclude_from_documentation */
+typedef enum ___itt_track_group_type
+{
+    __itt_track_group_type_normal = 0
+} __itt_track_group_type;
+/** @endcond */
+
+/** @cond exclude_from_documentation */
+#pragma pack(push, 8)
+
+typedef struct ___itt_track_group
+{
+    __itt_string_handle* name;     /*!< Name of the track group */
+    struct ___itt_track* track;    /*!< List of child tracks    */
+    __itt_track_group_type tgtype; /*!< Type of the track group */
+    int   extra1;                  /*!< Reserved. Must be zero  */
+    void* extra2;                  /*!< Reserved. Must be zero  */
+    struct ___itt_track_group* next;
+} __itt_track_group;
+
+#pragma pack(pop)
+/** @endcond */
+
+/**
+ * @brief Placeholder for custom track types. Currently, "normal" custom track
+ * is the only available track type.
+ */
+typedef enum ___itt_track_type
+{
+    __itt_track_type_normal = 0
+#ifdef INTEL_ITTNOTIFY_API_PRIVATE
+    , __itt_track_type_queue
+#endif /* INTEL_ITTNOTIFY_API_PRIVATE */
+} __itt_track_type;
+
+/** @cond exclude_from_documentation */
+#pragma pack(push, 8)
+
+typedef struct ___itt_track
+{
+    __itt_string_handle* name; /*!< Name of the track group */
+    __itt_track_group* group;  /*!< Parent group to a track */
+    __itt_track_type ttype;    /*!< Type of the track       */
+    int   extra1;              /*!< Reserved. Must be zero  */
+    void* extra2;              /*!< Reserved. Must be zero  */
+    struct ___itt_track* next;
+} __itt_track;
+
+#pragma pack(pop)
+/** @endcond */
+
+/**
+ * @brief Create logical track group.
+ */
+__itt_track_group* ITTAPI __itt_track_group_create(__itt_string_handle* name, __itt_track_group_type track_group_type);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(ITTAPI, __itt_track_group*, track_group_create, (__itt_string_handle* name, __itt_track_group_type track_group_type))
+#define __itt_track_group_create     ITTNOTIFY_DATA(track_group_create)
+#define __itt_track_group_create_ptr ITTNOTIFY_NAME(track_group_create)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_track_group_create(name)  (__itt_track_group*)0
+#define __itt_track_group_create_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_track_group_create_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Create logical track.
+ */
+__itt_track* ITTAPI __itt_track_create(__itt_track_group* track_group, __itt_string_handle* name, __itt_track_type track_type);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(ITTAPI, __itt_track*, track_create, (__itt_track_group* track_group,__itt_string_handle* name, __itt_track_type track_type))
+#define __itt_track_create     ITTNOTIFY_DATA(track_create)
+#define __itt_track_create_ptr ITTNOTIFY_NAME(track_create)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_track_create(track_group,name,track_type)  (__itt_track*)0
+#define __itt_track_create_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_track_create_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Set the logical track.
+ */
+void ITTAPI __itt_set_track(__itt_track* track);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, set_track, (__itt_track *track))
+#define __itt_set_track     ITTNOTIFY_VOID(set_track)
+#define __itt_set_track_ptr ITTNOTIFY_NAME(set_track)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_set_track(track)
+#define __itt_set_track_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_set_track_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/* ========================================================================== */
+/** @cond exclude_from_gpa_documentation */
+/**
+ * @defgroup events Events
+ * @ingroup public
+ * Events group
+ * @{
+ */
+/** @brief user event type */
+typedef int __itt_event;
+
+/**
+ * @brief Create an event notification
+ * @note name or namelen being null/name and namelen not matching, user event feature not enabled
+ * @return non-zero event identifier upon success and __itt_err otherwise
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+__itt_event LIBITTAPI __itt_event_createA(const char    *name, int namelen);
+__itt_event LIBITTAPI __itt_event_createW(const wchar_t *name, int namelen);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_event_create     __itt_event_createW
+#  define __itt_event_create_ptr __itt_event_createW_ptr
+#else
+#  define __itt_event_create     __itt_event_createA
+#  define __itt_event_create_ptr __itt_event_createA_ptr
+#endif /* UNICODE */
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+__itt_event LIBITTAPI __itt_event_create(const char *name, int namelen);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(LIBITTAPI, __itt_event, event_createA, (const char    *name, int namelen))
+ITT_STUB(LIBITTAPI, __itt_event, event_createW, (const wchar_t *name, int namelen))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(LIBITTAPI, __itt_event, event_create,  (const char    *name, int namelen))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_event_createA     ITTNOTIFY_DATA(event_createA)
+#define __itt_event_createA_ptr ITTNOTIFY_NAME(event_createA)
+#define __itt_event_createW     ITTNOTIFY_DATA(event_createW)
+#define __itt_event_createW_ptr ITTNOTIFY_NAME(event_createW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_event_create      ITTNOTIFY_DATA(event_create)
+#define __itt_event_create_ptr  ITTNOTIFY_NAME(event_create)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_event_createA(name, namelen) (__itt_event)0
+#define __itt_event_createA_ptr 0
+#define __itt_event_createW(name, namelen) (__itt_event)0
+#define __itt_event_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_event_create(name, namelen)  (__itt_event)0
+#define __itt_event_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_event_createA_ptr 0
+#define __itt_event_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_event_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Record an event occurrence.
+ * @return __itt_err upon failure (invalid event id/user event feature not enabled)
+ */
+int LIBITTAPI __itt_event_start(__itt_event event);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(LIBITTAPI, int, event_start, (__itt_event event))
+#define __itt_event_start     ITTNOTIFY_DATA(event_start)
+#define __itt_event_start_ptr ITTNOTIFY_NAME(event_start)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_event_start(event) (int)0
+#define __itt_event_start_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_event_start_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Record an event end occurrence.
+ * @note It is optional if events do not have durations.
+ * @return __itt_err upon failure (invalid event id/user event feature not enabled)
+ */
+int LIBITTAPI __itt_event_end(__itt_event event);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(LIBITTAPI, int, event_end, (__itt_event event))
+#define __itt_event_end     ITTNOTIFY_DATA(event_end)
+#define __itt_event_end_ptr ITTNOTIFY_NAME(event_end)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_event_end(event) (int)0
+#define __itt_event_end_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_event_end_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} events group */
+
+
+/**
+ * @defgroup arrays Arrays Visualizer
+ * @ingroup public
+ * Visualize arrays
+ * @{
+ */
+
+/**
+ * @enum __itt_av_data_type
+ * @brief Defines types of arrays data (for C/C++ intrinsic types)
+ */
+typedef enum
+{
+    __itt_e_first = 0,
+    __itt_e_char = 0,  /* 1-byte integer */
+    __itt_e_uchar,     /* 1-byte unsigned integer */
+    __itt_e_int16,     /* 2-byte integer */
+    __itt_e_uint16,    /* 2-byte unsigned integer  */
+    __itt_e_int32,     /* 4-byte integer */
+    __itt_e_uint32,    /* 4-byte unsigned integer */
+    __itt_e_int64,     /* 8-byte integer */
+    __itt_e_uint64,    /* 8-byte unsigned integer */
+    __itt_e_float,     /* 4-byte floating */
+    __itt_e_double,    /* 8-byte floating */
+    __itt_e_last = __itt_e_double
+} __itt_av_data_type;
+
+/**
+ * @brief Save an array data to a file.
+ * Output format is defined by the file extension. The csv and bmp formats are supported (bmp - for 2-dimensional array only).
+ * @param[in] data - pointer to the array data
+ * @param[in] rank - the rank of the array
+ * @param[in] dimensions - pointer to an array of integers, which specifies the array dimensions.
+ * The size of dimensions must be equal to the rank
+ * @param[in] type - the type of the array, specified as one of the __itt_av_data_type values (for intrinsic types)
+ * @param[in] filePath - the file path; the output format is defined by the file extension
+ * @param[in] columnOrder - defines how the array is stored in the linear memory.
+ * It should be 1 for column-major order (e.g. in FORTRAN) or 0 - for row-major order (e.g. in C).
+ */
+
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+int ITTAPI __itt_av_saveA(void *data, int rank, const int *dimensions, int type, const char *filePath, int columnOrder);
+int ITTAPI __itt_av_saveW(void *data, int rank, const int *dimensions, int type, const wchar_t *filePath, int columnOrder);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_av_save     __itt_av_saveW
+#  define __itt_av_save_ptr __itt_av_saveW_ptr
+#else /* UNICODE */
+#  define __itt_av_save     __itt_av_saveA
+#  define __itt_av_save_ptr __itt_av_saveA_ptr
+#endif /* UNICODE */
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+int ITTAPI __itt_av_save(void *data, int rank, const int *dimensions, int type, const char *filePath, int columnOrder);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, int, av_saveA, (void *data, int rank, const int *dimensions, int type, const char *filePath, int columnOrder))
+ITT_STUB(ITTAPI, int, av_saveW, (void *data, int rank, const int *dimensions, int type, const wchar_t *filePath, int columnOrder))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, int, av_save,  (void *data, int rank, const int *dimensions, int type, const char *filePath, int columnOrder))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_av_saveA     ITTNOTIFY_DATA(av_saveA)
+#define __itt_av_saveA_ptr ITTNOTIFY_NAME(av_saveA)
+#define __itt_av_saveW     ITTNOTIFY_DATA(av_saveW)
+#define __itt_av_saveW_ptr ITTNOTIFY_NAME(av_saveW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_av_save     ITTNOTIFY_DATA(av_save)
+#define __itt_av_save_ptr ITTNOTIFY_NAME(av_save)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_av_saveA(name)
+#define __itt_av_saveA_ptr 0
+#define __itt_av_saveW(name)
+#define __itt_av_saveW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_av_save(name)
+#define __itt_av_save_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_av_saveA_ptr 0
+#define __itt_av_saveW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_av_save_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+void ITTAPI __itt_enable_attach(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, enable_attach, (void))
+#define __itt_enable_attach     ITTNOTIFY_VOID(enable_attach)
+#define __itt_enable_attach_ptr ITTNOTIFY_NAME(enable_attach)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_enable_attach()
+#define __itt_enable_attach_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_enable_attach_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @cond exclude_from_gpa_documentation */
+
+/** @} arrays group */
+
+/** @endcond */
+
+/**
+ * @brief Module load info
+ * This API is used to report necessary information in case of module relocation
+ * @param[in] start_addr - relocated module start address
+ * @param[in] end_addr - relocated module end address
+ * @param[in] path - file system path to the module
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+void ITTAPI __itt_module_loadA(void *start_addr, void *end_addr, const char *path);
+void ITTAPI __itt_module_loadW(void *start_addr, void *end_addr, const wchar_t *path);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_module_load     __itt_module_loadW
+#  define __itt_module_load_ptr __itt_module_loadW_ptr
+#else /* UNICODE */
+#  define __itt_module_load     __itt_module_loadA
+#  define __itt_module_load_ptr __itt_module_loadA_ptr
+#endif /* UNICODE */
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+void ITTAPI __itt_module_load(void *start_addr, void *end_addr, const char *path);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, void, module_loadA, (void *start_addr, void *end_addr, const char *path))
+ITT_STUB(ITTAPI, void, module_loadW, (void *start_addr, void *end_addr, const wchar_t *path))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, void, module_load,  (void *start_addr, void *end_addr, const char *path))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_module_loadA     ITTNOTIFY_VOID(module_loadA)
+#define __itt_module_loadA_ptr ITTNOTIFY_NAME(module_loadA)
+#define __itt_module_loadW     ITTNOTIFY_VOID(module_loadW)
+#define __itt_module_loadW_ptr ITTNOTIFY_NAME(module_loadW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_module_load     ITTNOTIFY_VOID(module_load)
+#define __itt_module_load_ptr ITTNOTIFY_NAME(module_load)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_module_loadA(start_addr, end_addr, path)
+#define __itt_module_loadA_ptr 0
+#define __itt_module_loadW(start_addr, end_addr, path)
+#define __itt_module_loadW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_module_load(start_addr, end_addr, path)
+#define __itt_module_load_ptr 0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_module_loadA_ptr 0
+#define __itt_module_loadW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_module_load_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+
+
+#ifdef __cplusplus
+}
+#endif /* __cplusplus */
+
+#endif /* _ITTNOTIFY_H_ */
+
+#ifdef INTEL_ITTNOTIFY_API_PRIVATE
+
+#ifndef _ITTNOTIFY_PRIVATE_
+#define _ITTNOTIFY_PRIVATE_
+
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+
+/**
+ * @ingroup clockdomain
+ * @brief Begin an overlapped task instance.
+ * @param[in] domain The domain for this task
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] taskid The identifier for this task instance, *cannot* be __itt_null.
+ * @param[in] parentid The parent of this task, or __itt_null.
+ * @param[in] name The name of this task.
+ */
+void ITTAPI __itt_task_begin_overlapped_ex(const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id taskid, __itt_id parentid, __itt_string_handle* name);
+
+/**
+ * @ingroup clockdomain
+ * @brief End an overlapped task instance.
+ * @param[in] domain The domain for this task
+ * @param[in] clock_domain The clock domain controlling the execution of this call.
+ * @param[in] timestamp The user defined timestamp.
+ * @param[in] taskid Explicit ID of finished task
+ */
+void ITTAPI __itt_task_end_overlapped_ex(const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id taskid);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, task_begin_overlapped_ex,       (const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id taskid, __itt_id parentid, __itt_string_handle* name))
+ITT_STUBV(ITTAPI, void, task_end_overlapped_ex,         (const __itt_domain* domain, __itt_clock_domain* clock_domain, unsigned long long timestamp, __itt_id taskid))
+#define __itt_task_begin_overlapped_ex(d,x,y,z,a,b)     ITTNOTIFY_VOID_D5(task_begin_overlapped_ex,d,x,y,z,a,b)
+#define __itt_task_begin_overlapped_ex_ptr              ITTNOTIFY_NAME(task_begin_overlapped_ex)
+#define __itt_task_end_overlapped_ex(d,x,y,z)           ITTNOTIFY_VOID_D3(task_end_overlapped_ex,d,x,y,z)
+#define __itt_task_end_overlapped_ex_ptr                ITTNOTIFY_NAME(task_end_overlapped_ex)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_task_begin_overlapped_ex(domain,clock_domain,timestamp,taskid,parentid,name)
+#define __itt_task_begin_overlapped_ex_ptr      0
+#define __itt_task_end_overlapped_ex(domain,clock_domain,timestamp,taskid)
+#define __itt_task_end_overlapped_ex_ptr        0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_task_begin_overlapped_ex_ptr      0
+#define __itt_task_end_overlapped_ptr           0
+#define __itt_task_end_overlapped_ex_ptr        0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @defgroup makrs_internal Marks
+ * @ingroup internal
+ * Marks group
+ * @warning Internal API:
+ *   - It is not shipped to outside of Intel
+ *   - It is delivered to internal Intel teams using e-mail or SVN access only
+ * @{
+ */
+/** @brief user mark type */
+typedef int __itt_mark_type;
+
+/**
+ * @brief Creates a user mark type with the specified name using char or Unicode string.
+ * @param[in] name - name of mark to create
+ * @return Returns a handle to the mark type
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+__itt_mark_type ITTAPI __itt_mark_createA(const char    *name);
+__itt_mark_type ITTAPI __itt_mark_createW(const wchar_t *name);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_mark_create     __itt_mark_createW
+#  define __itt_mark_create_ptr __itt_mark_createW_ptr
+#else /* UNICODE */
+#  define __itt_mark_create     __itt_mark_createA
+#  define __itt_mark_create_ptr __itt_mark_createA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+__itt_mark_type ITTAPI __itt_mark_create(const char *name);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, __itt_mark_type, mark_createA, (const char    *name))
+ITT_STUB(ITTAPI, __itt_mark_type, mark_createW, (const wchar_t *name))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, __itt_mark_type, mark_create,  (const char *name))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_mark_createA     ITTNOTIFY_DATA(mark_createA)
+#define __itt_mark_createA_ptr ITTNOTIFY_NAME(mark_createA)
+#define __itt_mark_createW     ITTNOTIFY_DATA(mark_createW)
+#define __itt_mark_createW_ptr ITTNOTIFY_NAME(mark_createW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark_create      ITTNOTIFY_DATA(mark_create)
+#define __itt_mark_create_ptr  ITTNOTIFY_NAME(mark_create)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_mark_createA(name) (__itt_mark_type)0
+#define __itt_mark_createA_ptr 0
+#define __itt_mark_createW(name) (__itt_mark_type)0
+#define __itt_mark_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark_create(name)  (__itt_mark_type)0
+#define __itt_mark_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_mark_createA_ptr 0
+#define __itt_mark_createW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark_create_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Creates a "discrete" user mark type of the specified type and an optional parameter using char or Unicode string.
+ *
+ * - The mark of "discrete" type is placed to collection results in case of success. It appears in overtime view(s) as a special tick sign.
+ * - The call is "synchronous" - function returns after mark is actually added to results.
+ * - This function is useful, for example, to mark different phases of application
+ *   (beginning of the next mark automatically meand end of current region).
+ * - Can be used together with "continuous" marks (see below) at the same collection session
+ * @param[in] mt - mark, created by __itt_mark_create(const char* name) function
+ * @param[in] parameter - string parameter of mark
+ * @return Returns zero value in case of success, non-zero value otherwise.
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+int ITTAPI __itt_markA(__itt_mark_type mt, const char    *parameter);
+int ITTAPI __itt_markW(__itt_mark_type mt, const wchar_t *parameter);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_mark     __itt_markW
+#  define __itt_mark_ptr __itt_markW_ptr
+#else /* UNICODE  */
+#  define __itt_mark     __itt_markA
+#  define __itt_mark_ptr __itt_markA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+int ITTAPI __itt_mark(__itt_mark_type mt, const char *parameter);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, int, markA, (__itt_mark_type mt, const char    *parameter))
+ITT_STUB(ITTAPI, int, markW, (__itt_mark_type mt, const wchar_t *parameter))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, int, mark,  (__itt_mark_type mt, const char *parameter))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_markA     ITTNOTIFY_DATA(markA)
+#define __itt_markA_ptr ITTNOTIFY_NAME(markA)
+#define __itt_markW     ITTNOTIFY_DATA(markW)
+#define __itt_markW_ptr ITTNOTIFY_NAME(markW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark      ITTNOTIFY_DATA(mark)
+#define __itt_mark_ptr  ITTNOTIFY_NAME(mark)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_markA(mt, parameter) (int)0
+#define __itt_markA_ptr 0
+#define __itt_markW(mt, parameter) (int)0
+#define __itt_markW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark(mt, parameter)  (int)0
+#define __itt_mark_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_markA_ptr 0
+#define __itt_markW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Use this if necessary to create a "discrete" user event type (mark) for process
+ * rather then for one thread
+ * @see int __itt_mark(__itt_mark_type mt, const char* parameter);
+ */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+int ITTAPI __itt_mark_globalA(__itt_mark_type mt, const char    *parameter);
+int ITTAPI __itt_mark_globalW(__itt_mark_type mt, const wchar_t *parameter);
+#if defined(UNICODE) || defined(_UNICODE)
+#  define __itt_mark_global     __itt_mark_globalW
+#  define __itt_mark_global_ptr __itt_mark_globalW_ptr
+#else /* UNICODE  */
+#  define __itt_mark_global     __itt_mark_globalA
+#  define __itt_mark_global_ptr __itt_mark_globalA_ptr
+#endif /* UNICODE */
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+int ITTAPI __itt_mark_global(__itt_mark_type mt, const char *parameter);
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+ITT_STUB(ITTAPI, int, mark_globalA, (__itt_mark_type mt, const char    *parameter))
+ITT_STUB(ITTAPI, int, mark_globalW, (__itt_mark_type mt, const wchar_t *parameter))
+#else  /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+ITT_STUB(ITTAPI, int, mark_global,  (__itt_mark_type mt, const char *parameter))
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_mark_globalA     ITTNOTIFY_DATA(mark_globalA)
+#define __itt_mark_globalA_ptr ITTNOTIFY_NAME(mark_globalA)
+#define __itt_mark_globalW     ITTNOTIFY_DATA(mark_globalW)
+#define __itt_mark_globalW_ptr ITTNOTIFY_NAME(mark_globalW)
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark_global      ITTNOTIFY_DATA(mark_global)
+#define __itt_mark_global_ptr  ITTNOTIFY_NAME(mark_global)
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_mark_globalA(mt, parameter) (int)0
+#define __itt_mark_globalA_ptr 0
+#define __itt_mark_globalW(mt, parameter) (int)0
+#define __itt_mark_globalW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark_global(mt, parameter)  (int)0
+#define __itt_mark_global_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#if ITT_PLATFORM==ITT_PLATFORM_WIN
+#define __itt_mark_globalA_ptr 0
+#define __itt_mark_globalW_ptr 0
+#else /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#define __itt_mark_global_ptr  0
+#endif /* ITT_PLATFORM==ITT_PLATFORM_WIN */
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Creates an "end" point for "continuous" mark with specified name.
+ *
+ * - Returns zero value in case of success, non-zero value otherwise.
+ *   Also returns non-zero value when preceding "begin" point for the
+ *   mark with the same name failed to be created or not created.
+ * - The mark of "continuous" type is placed to collection results in
+ *   case of success. It appears in overtime view(s) as a special tick
+ *   sign (different from "discrete" mark) together with line from
+ *   corresponding "begin" mark to "end" mark.
+ * @note Continuous marks can overlap and be nested inside each other.
+ * Discrete mark can be nested inside marked region
+ * @param[in] mt - mark, created by __itt_mark_create(const char* name) function
+ * @return Returns zero value in case of success, non-zero value otherwise.
+ */
+int ITTAPI __itt_mark_off(__itt_mark_type mt);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(ITTAPI, int, mark_off, (__itt_mark_type mt))
+#define __itt_mark_off     ITTNOTIFY_DATA(mark_off)
+#define __itt_mark_off_ptr ITTNOTIFY_NAME(mark_off)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_mark_off(mt) (int)0
+#define __itt_mark_off_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_mark_off_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Use this if necessary to create an "end" point for mark of process
+ * @see int __itt_mark_off(__itt_mark_type mt);
+ */
+int ITTAPI __itt_mark_global_off(__itt_mark_type mt);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(ITTAPI, int, mark_global_off, (__itt_mark_type mt))
+#define __itt_mark_global_off     ITTNOTIFY_DATA(mark_global_off)
+#define __itt_mark_global_off_ptr ITTNOTIFY_NAME(mark_global_off)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_mark_global_off(mt) (int)0
+#define __itt_mark_global_off_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_mark_global_off_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+/** @} marks group */
+
+/**
+ * @defgroup counters_internal Counters
+ * @ingroup internal
+ * Counters group
+ * @{
+ */
+
+
+/**
+ * @defgroup stitch Stack Stitching
+ * @ingroup internal
+ * Stack Stitching group
+ * @{
+ */
+/**
+ * @brief opaque structure for counter identification
+ */
+typedef struct ___itt_caller *__itt_caller;
+
+/**
+ * @brief Create the stitch point e.g. a point in call stack where other stacks should be stitched to.
+ * The function returns a unique identifier which is used to match the cut points with corresponding stitch points.
+ */
+__itt_caller ITTAPI __itt_stack_caller_create(void);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUB(ITTAPI, __itt_caller, stack_caller_create, (void))
+#define __itt_stack_caller_create     ITTNOTIFY_DATA(stack_caller_create)
+#define __itt_stack_caller_create_ptr ITTNOTIFY_NAME(stack_caller_create)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_stack_caller_create() (__itt_caller)0
+#define __itt_stack_caller_create_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_stack_caller_create_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Destroy the inforamtion about stitch point identified by the pointer previously returned by __itt_stack_caller_create()
+ */
+void ITTAPI __itt_stack_caller_destroy(__itt_caller id);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, stack_caller_destroy, (__itt_caller id))
+#define __itt_stack_caller_destroy     ITTNOTIFY_VOID(stack_caller_destroy)
+#define __itt_stack_caller_destroy_ptr ITTNOTIFY_NAME(stack_caller_destroy)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_stack_caller_destroy(id)
+#define __itt_stack_caller_destroy_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_stack_caller_destroy_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief Sets the cut point. Stack from each event which occurs after this call will be cut
+ * at the same stack level the function was called and stitched to the corresponding stitch point.
+ */
+void ITTAPI __itt_stack_callee_enter(__itt_caller id);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, stack_callee_enter, (__itt_caller id))
+#define __itt_stack_callee_enter     ITTNOTIFY_VOID(stack_callee_enter)
+#define __itt_stack_callee_enter_ptr ITTNOTIFY_NAME(stack_callee_enter)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_stack_callee_enter(id)
+#define __itt_stack_callee_enter_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_stack_callee_enter_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/**
+ * @brief This function eliminates the cut point which was set by latest __itt_stack_callee_enter().
+ */
+void ITTAPI __itt_stack_callee_leave(__itt_caller id);
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+ITT_STUBV(ITTAPI, void, stack_callee_leave, (__itt_caller id))
+#define __itt_stack_callee_leave     ITTNOTIFY_VOID(stack_callee_leave)
+#define __itt_stack_callee_leave_ptr ITTNOTIFY_NAME(stack_callee_leave)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_stack_callee_leave(id)
+#define __itt_stack_callee_leave_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_stack_callee_leave_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+/** @} stitch group */
+
+/* ***************************************************************************************************************************** */
+
+#include <stdarg.h>
+
+/** @cond exclude_from_documentation */
+typedef enum __itt_error_code
+{
+    __itt_error_success       = 0, /*!< no error */
+    __itt_error_no_module     = 1, /*!< module can't be loaded */
+    /* %1$s -- library name; win: %2$d -- system error code; unx: %2$s -- system error message. */
+    __itt_error_no_symbol     = 2, /*!< symbol not found */
+    /* %1$s -- library name, %2$s -- symbol name. */
+    __itt_error_unknown_group = 3, /*!< unknown group specified */
+    /* %1$s -- env var name, %2$s -- group name. */
+    __itt_error_cant_read_env = 4, /*!< GetEnvironmentVariable() failed */
+    /* %1$s -- env var name, %2$d -- system error. */
+    __itt_error_env_too_long  = 5, /*!< variable value too long */
+    /* %1$s -- env var name, %2$d -- actual length of the var, %3$d -- max allowed length. */
+    __itt_error_system        = 6  /*!< pthread_mutexattr_init or pthread_mutex_init failed */
+    /* %1$s -- function name, %2$d -- errno. */
+} __itt_error_code;
+
+typedef void (__itt_error_handler_t)(__itt_error_code code, va_list);
+__itt_error_handler_t* __itt_set_error_handler(__itt_error_handler_t*);
+
+const char* ITTAPI __itt_api_version(void);
+/** @endcond */
+
+/** @cond exclude_from_documentation */
+#ifndef INTEL_NO_MACRO_BODY
+#ifndef INTEL_NO_ITTNOTIFY_API
+#define __itt_error_handler ITT_JOIN(INTEL_ITTNOTIFY_PREFIX, error_handler)
+void __itt_error_handler(__itt_error_code code, va_list args);
+extern const int ITTNOTIFY_NAME(err);
+#define __itt_err ITTNOTIFY_NAME(err)
+ITT_STUB(ITTAPI, const char*, api_version, (void))
+#define __itt_api_version     ITTNOTIFY_DATA(api_version)
+#define __itt_api_version_ptr ITTNOTIFY_NAME(api_version)
+#else  /* INTEL_NO_ITTNOTIFY_API */
+#define __itt_api_version()   (const char*)0
+#define __itt_api_version_ptr 0
+#endif /* INTEL_NO_ITTNOTIFY_API */
+#else  /* INTEL_NO_MACRO_BODY */
+#define __itt_api_version_ptr 0
+#endif /* INTEL_NO_MACRO_BODY */
+/** @endcond */
+
+#ifdef __cplusplus
+}
+#endif /* __cplusplus */
+
+#endif /* _ITTNOTIFY_PRIVATE_ */
+
+#endif /* INTEL_ITTNOTIFY_API_PRIVATE */
+
+// clang-format on
diff --git a/mozglue/baseprofiler/lul/AutoObjectMapper.cpp b/mozglue/baseprofiler/lul/AutoObjectMapper.cpp
new file mode 100644
index 0000000000..0037c943aa
--- /dev/null
+++ b/mozglue/baseprofiler/lul/AutoObjectMapper.cpp
@@ -0,0 +1,80 @@
+/* -*- 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 <sys/mman.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/Sprintf.h"
+
+#include "BaseProfiler.h"
+#include "PlatformMacros.h"
+#include "AutoObjectMapper.h"
+
+// A helper function for creating failure error messages in
+// AutoObjectMapper*::Map.
+static void failedToMessage(void (*aLog)(const char*), const char* aHowFailed,
+                            std::string aFileName) {
+  char buf[300];
+  SprintfLiteral(buf, "AutoObjectMapper::Map: Failed to %s \'%s\'", aHowFailed,
+                 aFileName.c_str());
+  buf[sizeof(buf) - 1] = 0;
+  aLog(buf);
+}
+
+AutoObjectMapperPOSIX::AutoObjectMapperPOSIX(void (*aLog)(const char*))
+    : mImage(nullptr), mSize(0), mLog(aLog), mIsMapped(false) {}
+
+AutoObjectMapperPOSIX::~AutoObjectMapperPOSIX() {
+  if (!mIsMapped) {
+    // There's nothing to do.
+    MOZ_ASSERT(!mImage);
+    MOZ_ASSERT(mSize == 0);
+    return;
+  }
+  MOZ_ASSERT(mSize > 0);
+  // The following assertion doesn't necessarily have to be true,
+  // but we assume (reasonably enough) that no mmap facility would
+  // be crazy enough to map anything at page zero.
+  MOZ_ASSERT(mImage);
+  munmap(mImage, mSize);
+}
+
+bool AutoObjectMapperPOSIX::Map(/*OUT*/ void** start, /*OUT*/ size_t* length,
+                                std::string fileName) {
+  MOZ_ASSERT(!mIsMapped);
+
+  int fd = open(fileName.c_str(), O_RDONLY);
+  if (fd == -1) {
+    failedToMessage(mLog, "open", fileName);
+    return false;
+  }
+
+  struct stat st;
+  int err = fstat(fd, &st);
+  size_t sz = (err == 0) ? st.st_size : 0;
+  if (err != 0 || sz == 0) {
+    failedToMessage(mLog, "fstat", fileName);
+    close(fd);
+    return false;
+  }
+
+  void* image = mmap(nullptr, sz, PROT_READ, MAP_SHARED, fd, 0);
+  if (image == MAP_FAILED) {
+    failedToMessage(mLog, "mmap", fileName);
+    close(fd);
+    return false;
+  }
+
+  close(fd);
+  mIsMapped = true;
+  mImage = *start = image;
+  mSize = *length = sz;
+  return true;
+}
diff --git a/mozglue/baseprofiler/lul/AutoObjectMapper.h b/mozglue/baseprofiler/lul/AutoObjectMapper.h
new file mode 100644
index 0000000000..f63aa43e0e
--- /dev/null
+++ b/mozglue/baseprofiler/lul/AutoObjectMapper.h
@@ -0,0 +1,64 @@
+/* -*- 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 AutoObjectMapper_h
+#define AutoObjectMapper_h
+
+#include <string>
+
+#include "mozilla/Attributes.h"
+#include "PlatformMacros.h"
+
+// A (nearly-) RAII class that maps an object in and then unmaps it on
+// destruction.  This base class version uses the "normal" POSIX
+// functions: open, fstat, close, mmap, munmap.
+
+class MOZ_STACK_CLASS AutoObjectMapperPOSIX {
+ public:
+  // The constructor does not attempt to map the file, because that
+  // might fail.  Instead, once the object has been constructed,
+  // call Map() to attempt the mapping.  There is no corresponding
+  // Unmap() since the unmapping is done in the destructor.  Failure
+  // messages are sent to |aLog|.
+  explicit AutoObjectMapperPOSIX(void (*aLog)(const char*));
+
+  // Unmap the file on destruction of this object.
+  ~AutoObjectMapperPOSIX();
+
+  // Map |fileName| into the address space and return the mapping
+  // extents.  If the file is zero sized this will fail.  The file is
+  // mapped read-only and private.  Returns true iff the mapping
+  // succeeded, in which case *start and *length hold its extent.
+  // Once a call to Map succeeds, all subsequent calls to it will
+  // fail.
+  bool Map(/*OUT*/ void** start, /*OUT*/ size_t* length, std::string fileName);
+
+ protected:
+  // If we are currently holding a mapped object, these record the
+  // mapped address range.
+  void* mImage;
+  size_t mSize;
+
+  // A logging sink, for complaining about mapping failures.
+  void (*mLog)(const char*);
+
+ private:
+  // Are we currently holding a mapped object?  This is private to
+  // the base class.  Derived classes need to have their own way to
+  // track whether they are holding a mapped object.
+  bool mIsMapped;
+
+  // Disable copying and assignment.
+  AutoObjectMapperPOSIX(const AutoObjectMapperPOSIX&);
+  AutoObjectMapperPOSIX& operator=(const AutoObjectMapperPOSIX&);
+  // Disable heap allocation of this class.
+  void* operator new(size_t);
+  void* operator new[](size_t);
+  void operator delete(void*);
+  void operator delete[](void*);
+};
+
+#endif  // AutoObjectMapper_h
diff --git a/mozglue/baseprofiler/lul/LulCommon.cpp b/mozglue/baseprofiler/lul/LulCommon.cpp
new file mode 100644
index 0000000000..f014892a57
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulCommon.cpp
@@ -0,0 +1,102 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+// Copyright (c) 2011, 2013 Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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
+// OWNER OR CONTRIBUTORS 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.
+
+// Original author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
+
+// This file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/module.cc
+//   src/common/unique_string.cc
+
+// There's no internal-only interface for LulCommon.  Hence include
+// the external interface directly.
+#include "LulCommonExt.h"
+
+#include <stdlib.h>
+#include <string.h>
+
+#include <string>
+#include <map>
+
+#include "BaseProfiler.h"
+
+namespace lul {
+
+using std::string;
+
+////////////////////////////////////////////////////////////////
+// Module
+//
+Module::Module(const string& name, const string& os, const string& architecture,
+               const string& id)
+    : name_(name), os_(os), architecture_(architecture), id_(id) {}
+
+Module::~Module() {}
+
+////////////////////////////////////////////////////////////////
+// UniqueString
+//
+class UniqueString {
+ public:
+  explicit UniqueString(string str) { str_ = strdup(str.c_str()); }
+  ~UniqueString() { free(reinterpret_cast<void*>(const_cast<char*>(str_))); }
+  const char* str_;
+};
+
+const char* FromUniqueString(const UniqueString* ustr) { return ustr->str_; }
+
+bool IsEmptyUniqueString(const UniqueString* ustr) {
+  return (ustr->str_)[0] == '\0';
+}
+
+////////////////////////////////////////////////////////////////
+// UniqueStringUniverse
+//
+UniqueStringUniverse::~UniqueStringUniverse() {
+  for (std::map<string, UniqueString*>::iterator it = map_.begin();
+       it != map_.end(); it++) {
+    delete it->second;
+  }
+}
+
+const UniqueString* UniqueStringUniverse::ToUniqueString(string str) {
+  std::map<string, UniqueString*>::iterator it = map_.find(str);
+  if (it == map_.end()) {
+    UniqueString* ustr = new UniqueString(str);
+    map_[str] = ustr;
+    return ustr;
+  } else {
+    return it->second;
+  }
+}
+
+}  // namespace lul
diff --git a/mozglue/baseprofiler/lul/LulCommonExt.h b/mozglue/baseprofiler/lul/LulCommonExt.h
new file mode 100644
index 0000000000..b20a7321ff
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulCommonExt.h
@@ -0,0 +1,509 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+// Copyright (c) 2006, 2010, 2012, 2013 Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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
+// OWNER OR CONTRIBUTORS 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.
+
+// Original author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
+
+// module.h: Define google_breakpad::Module. A Module holds debugging
+// information, and can write that information out as a Breakpad
+// symbol file.
+
+//  (C) Copyright Greg Colvin and Beman Dawes 1998, 1999.
+//  Copyright (c) 2001, 2002 Peter Dimov
+//
+//  Permission to copy, use, modify, sell and distribute this software
+//  is granted provided this copyright notice appears in all copies.
+//  This software is provided "as is" without express or implied
+//  warranty, and with no claim as to its suitability for any purpose.
+//
+//  See http://www.boost.org/libs/smart_ptr/scoped_ptr.htm for documentation.
+//
+
+// This file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/unique_string.h
+//   src/common/scoped_ptr.h
+//   src/common/module.h
+
+// External interface for the "Common" component of LUL.
+
+#ifndef LulCommonExt_h
+#define LulCommonExt_h
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdint.h>
+
+#include <string>
+#include <map>
+#include <vector>
+#include <cstddef>  // for std::ptrdiff_t
+
+#include "mozilla/Assertions.h"
+
+namespace lul {
+
+using std::map;
+using std::string;
+
+////////////////////////////////////////////////////////////////
+// UniqueString
+//
+
+// Abstract type
+class UniqueString;
+
+// Get the contained C string (debugging only)
+const char* FromUniqueString(const UniqueString*);
+
+// Is the given string empty (that is, "") ?
+bool IsEmptyUniqueString(const UniqueString*);
+
+////////////////////////////////////////////////////////////////
+// UniqueStringUniverse
+//
+
+// All UniqueStrings live in some specific UniqueStringUniverse.
+class UniqueStringUniverse {
+ public:
+  UniqueStringUniverse() {}
+  ~UniqueStringUniverse();
+  // Convert a |string| to a UniqueString, that lives in this universe.
+  const UniqueString* ToUniqueString(string str);
+
+ private:
+  map<string, UniqueString*> map_;
+};
+
+////////////////////////////////////////////////////////////////
+// GUID
+//
+
+typedef struct {
+  uint32_t data1;
+  uint16_t data2;
+  uint16_t data3;
+  uint8_t data4[8];
+} MDGUID;  // GUID
+
+typedef MDGUID GUID;
+
+////////////////////////////////////////////////////////////////
+// scoped_ptr
+//
+
+//  scoped_ptr mimics a built-in pointer except that it guarantees deletion
+//  of the object pointed to, either on destruction of the scoped_ptr or via
+//  an explicit reset(). scoped_ptr is a simple solution for simple needs;
+//  use shared_ptr or std::auto_ptr if your needs are more complex.
+
+//  *** NOTE ***
+//  If your scoped_ptr is a class member of class FOO pointing to a
+//  forward declared type BAR (as shown below), then you MUST use a non-inlined
+//  version of the destructor.  The destructor of a scoped_ptr (called from
+//  FOO's destructor) must have a complete definition of BAR in order to
+//  destroy it.  Example:
+//
+//  -- foo.h --
+//  class BAR;
+//
+//  class FOO {
+//   public:
+//    FOO();
+//    ~FOO();  // Required for sources that instantiate class FOO to compile!
+//
+//   private:
+//    scoped_ptr<BAR> bar_;
+//  };
+//
+//  -- foo.cc --
+//  #include "foo.h"
+//  FOO::~FOO() {} // Empty, but must be non-inlined to FOO's class definition.
+
+//  scoped_ptr_malloc added by Google
+//  When one of these goes out of scope, instead of doing a delete or
+//  delete[], it calls free().  scoped_ptr_malloc<char> is likely to see
+//  much more use than any other specializations.
+
+//  release() added by Google
+//  Use this to conditionally transfer ownership of a heap-allocated object
+//  to the caller, usually on method success.
+
+template <typename T>
+class scoped_ptr {
+ private:
+  T* ptr;
+
+  scoped_ptr(scoped_ptr const&);
+  scoped_ptr& operator=(scoped_ptr const&);
+
+ public:
+  typedef T element_type;
+
+  explicit scoped_ptr(T* p = 0) : ptr(p) {}
+
+  ~scoped_ptr() { delete ptr; }
+
+  void reset(T* p = 0) {
+    if (ptr != p) {
+      delete ptr;
+      ptr = p;
+    }
+  }
+
+  T& operator*() const {
+    MOZ_ASSERT(ptr != 0);
+    return *ptr;
+  }
+
+  T* operator->() const {
+    MOZ_ASSERT(ptr != 0);
+    return ptr;
+  }
+
+  bool operator==(T* p) const { return ptr == p; }
+
+  bool operator!=(T* p) const { return ptr != p; }
+
+  T* get() const { return ptr; }
+
+  void swap(scoped_ptr& b) {
+    T* tmp = b.ptr;
+    b.ptr = ptr;
+    ptr = tmp;
+  }
+
+  T* release() {
+    T* tmp = ptr;
+    ptr = 0;
+    return tmp;
+  }
+
+ private:
+  // no reason to use these: each scoped_ptr should have its own object
+  template <typename U>
+  bool operator==(scoped_ptr<U> const& p) const;
+  template <typename U>
+  bool operator!=(scoped_ptr<U> const& p) const;
+};
+
+template <typename T>
+inline void swap(scoped_ptr<T>& a, scoped_ptr<T>& b) {
+  a.swap(b);
+}
+
+template <typename T>
+inline bool operator==(T* p, const scoped_ptr<T>& b) {
+  return p == b.get();
+}
+
+template <typename T>
+inline bool operator!=(T* p, const scoped_ptr<T>& b) {
+  return p != b.get();
+}
+
+//  scoped_array extends scoped_ptr to arrays. Deletion of the array pointed to
+//  is guaranteed, either on destruction of the scoped_array or via an explicit
+//  reset(). Use shared_array or std::vector if your needs are more complex.
+
+template <typename T>
+class scoped_array {
+ private:
+  T* ptr;
+
+  scoped_array(scoped_array const&);
+  scoped_array& operator=(scoped_array const&);
+
+ public:
+  typedef T element_type;
+
+  explicit scoped_array(T* p = 0) : ptr(p) {}
+
+  ~scoped_array() { delete[] ptr; }
+
+  void reset(T* p = 0) {
+    if (ptr != p) {
+      delete[] ptr;
+      ptr = p;
+    }
+  }
+
+  T& operator[](std::ptrdiff_t i) const {
+    MOZ_ASSERT(ptr != 0);
+    MOZ_ASSERT(i >= 0);
+    return ptr[i];
+  }
+
+  bool operator==(T* p) const { return ptr == p; }
+
+  bool operator!=(T* p) const { return ptr != p; }
+
+  T* get() const { return ptr; }
+
+  void swap(scoped_array& b) {
+    T* tmp = b.ptr;
+    b.ptr = ptr;
+    ptr = tmp;
+  }
+
+  T* release() {
+    T* tmp = ptr;
+    ptr = 0;
+    return tmp;
+  }
+
+ private:
+  // no reason to use these: each scoped_array should have its own object
+  template <typename U>
+  bool operator==(scoped_array<U> const& p) const;
+  template <typename U>
+  bool operator!=(scoped_array<U> const& p) const;
+};
+
+template <class T>
+inline void swap(scoped_array<T>& a, scoped_array<T>& b) {
+  a.swap(b);
+}
+
+template <typename T>
+inline bool operator==(T* p, const scoped_array<T>& b) {
+  return p == b.get();
+}
+
+template <typename T>
+inline bool operator!=(T* p, const scoped_array<T>& b) {
+  return p != b.get();
+}
+
+// This class wraps the c library function free() in a class that can be
+// passed as a template argument to scoped_ptr_malloc below.
+class ScopedPtrMallocFree {
+ public:
+  inline void operator()(void* x) const { free(x); }
+};
+
+// scoped_ptr_malloc<> is similar to scoped_ptr<>, but it accepts a
+// second template argument, the functor used to free the object.
+
+template <typename T, typename FreeProc = ScopedPtrMallocFree>
+class scoped_ptr_malloc {
+ private:
+  T* ptr;
+
+  scoped_ptr_malloc(scoped_ptr_malloc const&);
+  scoped_ptr_malloc& operator=(scoped_ptr_malloc const&);
+
+ public:
+  typedef T element_type;
+
+  explicit scoped_ptr_malloc(T* p = 0) : ptr(p) {}
+
+  ~scoped_ptr_malloc() { free_((void*)ptr); }
+
+  void reset(T* p = 0) {
+    if (ptr != p) {
+      free_((void*)ptr);
+      ptr = p;
+    }
+  }
+
+  T& operator*() const {
+    MOZ_ASSERT(ptr != 0);
+    return *ptr;
+  }
+
+  T* operator->() const {
+    MOZ_ASSERT(ptr != 0);
+    return ptr;
+  }
+
+  bool operator==(T* p) const { return ptr == p; }
+
+  bool operator!=(T* p) const { return ptr != p; }
+
+  T* get() const { return ptr; }
+
+  void swap(scoped_ptr_malloc& b) {
+    T* tmp = b.ptr;
+    b.ptr = ptr;
+    ptr = tmp;
+  }
+
+  T* release() {
+    T* tmp = ptr;
+    ptr = 0;
+    return tmp;
+  }
+
+ private:
+  // no reason to use these: each scoped_ptr_malloc should have its own object
+  template <typename U, typename GP>
+  bool operator==(scoped_ptr_malloc<U, GP> const& p) const;
+  template <typename U, typename GP>
+  bool operator!=(scoped_ptr_malloc<U, GP> const& p) const;
+
+  static FreeProc const free_;
+};
+
+template <typename T, typename FP>
+FP const scoped_ptr_malloc<T, FP>::free_ = FP();
+
+template <typename T, typename FP>
+inline void swap(scoped_ptr_malloc<T, FP>& a, scoped_ptr_malloc<T, FP>& b) {
+  a.swap(b);
+}
+
+template <typename T, typename FP>
+inline bool operator==(T* p, const scoped_ptr_malloc<T, FP>& b) {
+  return p == b.get();
+}
+
+template <typename T, typename FP>
+inline bool operator!=(T* p, const scoped_ptr_malloc<T, FP>& b) {
+  return p != b.get();
+}
+
+////////////////////////////////////////////////////////////////
+// Module
+//
+
+// A Module represents the contents of a module, and supports methods
+// for adding information produced by parsing STABS or DWARF data
+// --- possibly both from the same file --- and then writing out the
+// unified contents as a Breakpad-format symbol file.
+class Module {
+ public:
+  // The type of addresses and sizes in a symbol table.
+  typedef uint64_t Address;
+
+  // Representation of an expression.  This can either be a postfix
+  // expression, in which case it is stored as a string, or a simple
+  // expression of the form (identifier + imm) or *(identifier + imm).
+  // It can also be invalid (denoting "no value").
+  enum ExprHow { kExprInvalid = 1, kExprPostfix, kExprSimple, kExprSimpleMem };
+
+  struct Expr {
+    // Construct a simple-form expression
+    Expr(const UniqueString* ident, long offset, bool deref) {
+      if (IsEmptyUniqueString(ident)) {
+        Expr();
+      } else {
+        postfix_ = "";
+        ident_ = ident;
+        offset_ = offset;
+        how_ = deref ? kExprSimpleMem : kExprSimple;
+      }
+    }
+
+    // Construct an invalid expression
+    Expr() {
+      postfix_ = "";
+      ident_ = nullptr;
+      offset_ = 0;
+      how_ = kExprInvalid;
+    }
+
+    // Return the postfix expression string, either directly,
+    // if this is a postfix expression, or by synthesising it
+    // for a simple expression.
+    std::string getExprPostfix() const {
+      switch (how_) {
+        case kExprPostfix:
+          return postfix_;
+        case kExprSimple:
+        case kExprSimpleMem: {
+          char buf[40];
+          sprintf(buf, " %ld %c%s", labs(offset_), offset_ < 0 ? '-' : '+',
+                  how_ == kExprSimple ? "" : " ^");
+          return std::string(FromUniqueString(ident_)) + std::string(buf);
+        }
+        case kExprInvalid:
+        default:
+          MOZ_ASSERT(0 && "getExprPostfix: invalid Module::Expr type");
+          return "Expr::genExprPostfix: kExprInvalid";
+      }
+    }
+
+    // The identifier that gives the starting value for simple expressions.
+    const UniqueString* ident_;
+    // The offset to add for simple expressions.
+    long offset_;
+    // The Postfix expression string to evaluate for non-simple expressions.
+    std::string postfix_;
+    // The operation expressed by this expression.
+    ExprHow how_;
+  };
+
+  // A map from register names to expressions that recover
+  // their values. This can represent a complete set of rules to
+  // follow at some address, or a set of changes to be applied to an
+  // extant set of rules.
+  // NOTE! there are two completely different types called RuleMap.  This
+  // is one of them.
+  typedef std::map<const UniqueString*, Expr> RuleMap;
+
+  // A map from addresses to RuleMaps, representing changes that take
+  // effect at given addresses.
+  typedef std::map<Address, RuleMap> RuleChangeMap;
+
+  // A range of 'STACK CFI' stack walking information. An instance of
+  // this structure corresponds to a 'STACK CFI INIT' record and the
+  // subsequent 'STACK CFI' records that fall within its range.
+  struct StackFrameEntry {
+    // The starting address and number of bytes of machine code this
+    // entry covers.
+    Address address, size;
+
+    // The initial register recovery rules, in force at the starting
+    // address.
+    RuleMap initial_rules;
+
+    // A map from addresses to rule changes. To find the rules in
+    // force at a given address, start with initial_rules, and then
+    // apply the changes given in this map for all addresses up to and
+    // including the address you're interested in.
+    RuleChangeMap rule_changes;
+  };
+
+  // Create a new module with the given name, operating system,
+  // architecture, and ID string.
+  Module(const std::string& name, const std::string& os,
+         const std::string& architecture, const std::string& id);
+  ~Module();
+
+ private:
+  // Module header entries.
+  std::string name_, os_, architecture_, id_;
+};
+
+}  // namespace lul
+
+#endif  // LulCommonExt_h
diff --git a/mozglue/baseprofiler/lul/LulDwarf.cpp b/mozglue/baseprofiler/lul/LulDwarf.cpp
new file mode 100644
index 0000000000..c83296fc62
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulDwarf.cpp
@@ -0,0 +1,2252 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+// Copyright (c) 2010 Google Inc. All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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
+// OWNER OR CONTRIBUTORS 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.
+
+// CFI reader author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
+// Original author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
+
+// Implementation of dwarf2reader::LineInfo, dwarf2reader::CompilationUnit,
+// and dwarf2reader::CallFrameInfo. See dwarf2reader.h for details.
+
+// This file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/dwarf/bytereader.cc
+//   src/common/dwarf/dwarf2reader.cc
+//   src/common/dwarf_cfi_to_module.cc
+
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+
+#include <map>
+#include <stack>
+#include <string>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/Sprintf.h"
+
+#include "LulCommonExt.h"
+#include "LulDwarfInt.h"
+
+// Set this to 1 for verbose logging
+#define DEBUG_DWARF 0
+
+namespace lul {
+
+using std::string;
+
+ByteReader::ByteReader(enum Endianness endian)
+    : offset_reader_(NULL),
+      address_reader_(NULL),
+      endian_(endian),
+      address_size_(0),
+      offset_size_(0),
+      have_section_base_(),
+      have_text_base_(),
+      have_data_base_(),
+      have_function_base_() {}
+
+ByteReader::~ByteReader() {}
+
+void ByteReader::SetOffsetSize(uint8 size) {
+  offset_size_ = size;
+  MOZ_ASSERT(size == 4 || size == 8);
+  if (size == 4) {
+    this->offset_reader_ = &ByteReader::ReadFourBytes;
+  } else {
+    this->offset_reader_ = &ByteReader::ReadEightBytes;
+  }
+}
+
+void ByteReader::SetAddressSize(uint8 size) {
+  address_size_ = size;
+  MOZ_ASSERT(size == 4 || size == 8);
+  if (size == 4) {
+    this->address_reader_ = &ByteReader::ReadFourBytes;
+  } else {
+    this->address_reader_ = &ByteReader::ReadEightBytes;
+  }
+}
+
+uint64 ByteReader::ReadInitialLength(const char* start, size_t* len) {
+  const uint64 initial_length = ReadFourBytes(start);
+  start += 4;
+
+  // In DWARF2/3, if the initial length is all 1 bits, then the offset
+  // size is 8 and we need to read the next 8 bytes for the real length.
+  if (initial_length == 0xffffffff) {
+    SetOffsetSize(8);
+    *len = 12;
+    return ReadOffset(start);
+  } else {
+    SetOffsetSize(4);
+    *len = 4;
+  }
+  return initial_length;
+}
+
+bool ByteReader::ValidEncoding(DwarfPointerEncoding encoding) const {
+  if (encoding == DW_EH_PE_omit) return true;
+  if (encoding == DW_EH_PE_aligned) return true;
+  if ((encoding & 0x7) > DW_EH_PE_udata8) return false;
+  if ((encoding & 0x70) > DW_EH_PE_funcrel) return false;
+  return true;
+}
+
+bool ByteReader::UsableEncoding(DwarfPointerEncoding encoding) const {
+  switch (encoding & 0x70) {
+    case DW_EH_PE_absptr:
+      return true;
+    case DW_EH_PE_pcrel:
+      return have_section_base_;
+    case DW_EH_PE_textrel:
+      return have_text_base_;
+    case DW_EH_PE_datarel:
+      return have_data_base_;
+    case DW_EH_PE_funcrel:
+      return have_function_base_;
+    default:
+      return false;
+  }
+}
+
+uint64 ByteReader::ReadEncodedPointer(const char* buffer,
+                                      DwarfPointerEncoding encoding,
+                                      size_t* len) const {
+  // UsableEncoding doesn't approve of DW_EH_PE_omit, so we shouldn't
+  // see it here.
+  MOZ_ASSERT(encoding != DW_EH_PE_omit);
+
+  // The Linux Standards Base 4.0 does not make this clear, but the
+  // GNU tools (gcc/unwind-pe.h; readelf/dwarf.c; gdb/dwarf2-frame.c)
+  // agree that aligned pointers are always absolute, machine-sized,
+  // machine-signed pointers.
+  if (encoding == DW_EH_PE_aligned) {
+    MOZ_ASSERT(have_section_base_);
+
+    // We don't need to align BUFFER in *our* address space. Rather, we
+    // need to find the next position in our buffer that would be aligned
+    // when the .eh_frame section the buffer contains is loaded into the
+    // program's memory. So align assuming that buffer_base_ gets loaded at
+    // address section_base_, where section_base_ itself may or may not be
+    // aligned.
+
+    // First, find the offset to START from the closest prior aligned
+    // address.
+    uint64 skew = section_base_ & (AddressSize() - 1);
+    // Now find the offset from that aligned address to buffer.
+    uint64 offset = skew + (buffer - buffer_base_);
+    // Round up to the next boundary.
+    uint64 aligned = (offset + AddressSize() - 1) & -AddressSize();
+    // Convert back to a pointer.
+    const char* aligned_buffer = buffer_base_ + (aligned - skew);
+    // Finally, store the length and actually fetch the pointer.
+    *len = aligned_buffer - buffer + AddressSize();
+    return ReadAddress(aligned_buffer);
+  }
+
+  // Extract the value first, ignoring whether it's a pointer or an
+  // offset relative to some base.
+  uint64 offset;
+  switch (encoding & 0x0f) {
+    case DW_EH_PE_absptr:
+      // DW_EH_PE_absptr is weird, as it is used as a meaningful value for
+      // both the high and low nybble of encoding bytes. When it appears in
+      // the high nybble, it means that the pointer is absolute, not an
+      // offset from some base address. When it appears in the low nybble,
+      // as here, it means that the pointer is stored as a normal
+      // machine-sized and machine-signed address. A low nybble of
+      // DW_EH_PE_absptr does not imply that the pointer is absolute; it is
+      // correct for us to treat the value as an offset from a base address
+      // if the upper nybble is not DW_EH_PE_absptr.
+      offset = ReadAddress(buffer);
+      *len = AddressSize();
+      break;
+
+    case DW_EH_PE_uleb128:
+      offset = ReadUnsignedLEB128(buffer, len);
+      break;
+
+    case DW_EH_PE_udata2:
+      offset = ReadTwoBytes(buffer);
+      *len = 2;
+      break;
+
+    case DW_EH_PE_udata4:
+      offset = ReadFourBytes(buffer);
+      *len = 4;
+      break;
+
+    case DW_EH_PE_udata8:
+      offset = ReadEightBytes(buffer);
+      *len = 8;
+      break;
+
+    case DW_EH_PE_sleb128:
+      offset = ReadSignedLEB128(buffer, len);
+      break;
+
+    case DW_EH_PE_sdata2:
+      offset = ReadTwoBytes(buffer);
+      // Sign-extend from 16 bits.
+      offset = (offset ^ 0x8000) - 0x8000;
+      *len = 2;
+      break;
+
+    case DW_EH_PE_sdata4:
+      offset = ReadFourBytes(buffer);
+      // Sign-extend from 32 bits.
+      offset = (offset ^ 0x80000000ULL) - 0x80000000ULL;
+      *len = 4;
+      break;
+
+    case DW_EH_PE_sdata8:
+      // No need to sign-extend; this is the full width of our type.
+      offset = ReadEightBytes(buffer);
+      *len = 8;
+      break;
+
+    default:
+      abort();
+  }
+
+  // Find the appropriate base address.
+  uint64 base;
+  switch (encoding & 0x70) {
+    case DW_EH_PE_absptr:
+      base = 0;
+      break;
+
+    case DW_EH_PE_pcrel:
+      MOZ_ASSERT(have_section_base_);
+      base = section_base_ + (buffer - buffer_base_);
+      break;
+
+    case DW_EH_PE_textrel:
+      MOZ_ASSERT(have_text_base_);
+      base = text_base_;
+      break;
+
+    case DW_EH_PE_datarel:
+      MOZ_ASSERT(have_data_base_);
+      base = data_base_;
+      break;
+
+    case DW_EH_PE_funcrel:
+      MOZ_ASSERT(have_function_base_);
+      base = function_base_;
+      break;
+
+    default:
+      abort();
+  }
+
+  uint64 pointer = base + offset;
+
+  // Remove inappropriate upper bits.
+  if (AddressSize() == 4)
+    pointer = pointer & 0xffffffff;
+  else
+    MOZ_ASSERT(AddressSize() == sizeof(uint64));
+
+  return pointer;
+}
+
+// A DWARF rule for recovering the address or value of a register, or
+// computing the canonical frame address. There is one subclass of this for
+// each '*Rule' member function in CallFrameInfo::Handler.
+//
+// It's annoying that we have to handle Rules using pointers (because
+// the concrete instances can have an arbitrary size). They're small,
+// so it would be much nicer if we could just handle them by value
+// instead of fretting about ownership and destruction.
+//
+// It seems like all these could simply be instances of std::tr1::bind,
+// except that we need instances to be EqualityComparable, too.
+//
+// This could logically be nested within State, but then the qualified names
+// get horrendous.
+class CallFrameInfo::Rule {
+ public:
+  virtual ~Rule() {}
+
+  // Tell HANDLER that, at ADDRESS in the program, REG can be
+  // recovered using this rule. If REG is kCFARegister, then this rule
+  // describes how to compute the canonical frame address. Return what the
+  // HANDLER member function returned.
+  virtual bool Handle(Handler* handler, uint64 address, int reg) const = 0;
+
+  // Equality on rules. We use these to decide which rules we need
+  // to report after a DW_CFA_restore_state instruction.
+  virtual bool operator==(const Rule& rhs) const = 0;
+
+  bool operator!=(const Rule& rhs) const { return !(*this == rhs); }
+
+  // Return a pointer to a copy of this rule.
+  virtual Rule* Copy() const = 0;
+
+  // If this is a base+offset rule, change its base register to REG.
+  // Otherwise, do nothing. (Ugly, but required for DW_CFA_def_cfa_register.)
+  virtual void SetBaseRegister(unsigned reg) {}
+
+  // If this is a base+offset rule, change its offset to OFFSET. Otherwise,
+  // do nothing. (Ugly, but required for DW_CFA_def_cfa_offset.)
+  virtual void SetOffset(long long offset) {}
+
+  // A RTTI workaround, to make it possible to implement equality
+  // comparisons on classes derived from this one.
+  enum CFIRTag {
+    CFIR_UNDEFINED_RULE,
+    CFIR_SAME_VALUE_RULE,
+    CFIR_OFFSET_RULE,
+    CFIR_VAL_OFFSET_RULE,
+    CFIR_REGISTER_RULE,
+    CFIR_EXPRESSION_RULE,
+    CFIR_VAL_EXPRESSION_RULE
+  };
+
+  // Produce the tag that identifies the child class of this object.
+  virtual CFIRTag getTag() const = 0;
+};
+
+// Rule: the value the register had in the caller cannot be recovered.
+class CallFrameInfo::UndefinedRule : public CallFrameInfo::Rule {
+ public:
+  UndefinedRule() {}
+  ~UndefinedRule() {}
+  CFIRTag getTag() const override { return CFIR_UNDEFINED_RULE; }
+  bool Handle(Handler* handler, uint64 address, int reg) const override {
+    return handler->UndefinedRule(address, reg);
+  }
+  bool operator==(const Rule& rhs) const override {
+    if (rhs.getTag() != CFIR_UNDEFINED_RULE) return false;
+    return true;
+  }
+  Rule* Copy() const override { return new UndefinedRule(*this); }
+};
+
+// Rule: the register's value is the same as that it had in the caller.
+class CallFrameInfo::SameValueRule : public CallFrameInfo::Rule {
+ public:
+  SameValueRule() {}
+  ~SameValueRule() {}
+  CFIRTag getTag() const override { return CFIR_SAME_VALUE_RULE; }
+  bool Handle(Handler* handler, uint64 address, int reg) const override {
+    return handler->SameValueRule(address, reg);
+  }
+  bool operator==(const Rule& rhs) const override {
+    if (rhs.getTag() != CFIR_SAME_VALUE_RULE) return false;
+    return true;
+  }
+  Rule* Copy() const override { return new SameValueRule(*this); }
+};
+
+// Rule: the register is saved at OFFSET from BASE_REGISTER.  BASE_REGISTER
+// may be CallFrameInfo::Handler::kCFARegister.
+class CallFrameInfo::OffsetRule : public CallFrameInfo::Rule {
+ public:
+  OffsetRule(int base_register, long offset)
+      : base_register_(base_register), offset_(offset) {}
+  ~OffsetRule() {}
+  CFIRTag getTag() const override { return CFIR_OFFSET_RULE; }
+  bool Handle(Handler* handler, uint64 address, int reg) const override {
+    return handler->OffsetRule(address, reg, base_register_, offset_);
+  }
+  bool operator==(const Rule& rhs) const override {
+    if (rhs.getTag() != CFIR_OFFSET_RULE) return false;
+    const OffsetRule* our_rhs = static_cast<const OffsetRule*>(&rhs);
+    return (base_register_ == our_rhs->base_register_ &&
+            offset_ == our_rhs->offset_);
+  }
+  Rule* Copy() const override { return new OffsetRule(*this); }
+  // We don't actually need SetBaseRegister or SetOffset here, since they
+  // are only ever applied to CFA rules, for DW_CFA_def_cfa_offset, and it
+  // doesn't make sense to use OffsetRule for computing the CFA: it
+  // computes the address at which a register is saved, not a value.
+ private:
+  int base_register_;
+  long offset_;
+};
+
+// Rule: the value the register had in the caller is the value of
+// BASE_REGISTER plus offset. BASE_REGISTER may be
+// CallFrameInfo::Handler::kCFARegister.
+class CallFrameInfo::ValOffsetRule : public CallFrameInfo::Rule {
+ public:
+  ValOffsetRule(int base_register, long offset)
+      : base_register_(base_register), offset_(offset) {}
+  ~ValOffsetRule() {}
+  CFIRTag getTag() const override { return CFIR_VAL_OFFSET_RULE; }
+  bool Handle(Handler* handler, uint64 address, int reg) const override {
+    return handler->ValOffsetRule(address, reg, base_register_, offset_);
+  }
+  bool operator==(const Rule& rhs) const override {
+    if (rhs.getTag() != CFIR_VAL_OFFSET_RULE) return false;
+    const ValOffsetRule* our_rhs = static_cast<const ValOffsetRule*>(&rhs);
+    return (base_register_ == our_rhs->base_register_ &&
+            offset_ == our_rhs->offset_);
+  }
+  Rule* Copy() const override { return new ValOffsetRule(*this); }
+  void SetBaseRegister(unsigned reg) override { base_register_ = reg; }
+  void SetOffset(long long offset) override { offset_ = offset; }
+
+ private:
+  int base_register_;
+  long offset_;
+};
+
+// Rule: the register has been saved in another register REGISTER_NUMBER_.
+class CallFrameInfo::RegisterRule : public CallFrameInfo::Rule {
+ public:
+  explicit RegisterRule(int register_number)
+      : register_number_(register_number) {}
+  ~RegisterRule() {}
+  CFIRTag getTag() const override { return CFIR_REGISTER_RULE; }
+  bool Handle(Handler* handler, uint64 address, int reg) const override {
+    return handler->RegisterRule(address, reg, register_number_);
+  }
+  bool operator==(const Rule& rhs) const override {
+    if (rhs.getTag() != CFIR_REGISTER_RULE) return false;
+    const RegisterRule* our_rhs = static_cast<const RegisterRule*>(&rhs);
+    return (register_number_ == our_rhs->register_number_);
+  }
+  Rule* Copy() const override { return new RegisterRule(*this); }
+
+ private:
+  int register_number_;
+};
+
+// Rule: EXPRESSION evaluates to the address at which the register is saved.
+class CallFrameInfo::ExpressionRule : public CallFrameInfo::Rule {
+ public:
+  explicit ExpressionRule(const string& expression) : expression_(expression) {}
+  ~ExpressionRule() {}
+  CFIRTag getTag() const override { return CFIR_EXPRESSION_RULE; }
+  bool Handle(Handler* handler, uint64 address, int reg) const override {
+    return handler->ExpressionRule(address, reg, expression_);
+  }
+  bool operator==(const Rule& rhs) const override {
+    if (rhs.getTag() != CFIR_EXPRESSION_RULE) return false;
+    const ExpressionRule* our_rhs = static_cast<const ExpressionRule*>(&rhs);
+    return (expression_ == our_rhs->expression_);
+  }
+  Rule* Copy() const override { return new ExpressionRule(*this); }
+
+ private:
+  string expression_;
+};
+
+// Rule: EXPRESSION evaluates to the previous value of the register.
+class CallFrameInfo::ValExpressionRule : public CallFrameInfo::Rule {
+ public:
+  explicit ValExpressionRule(const string& expression)
+      : expression_(expression) {}
+  ~ValExpressionRule() {}
+  CFIRTag getTag() const override { return CFIR_VAL_EXPRESSION_RULE; }
+  bool Handle(Handler* handler, uint64 address, int reg) const override {
+    return handler->ValExpressionRule(address, reg, expression_);
+  }
+  bool operator==(const Rule& rhs) const override {
+    if (rhs.getTag() != CFIR_VAL_EXPRESSION_RULE) return false;
+    const ValExpressionRule* our_rhs =
+        static_cast<const ValExpressionRule*>(&rhs);
+    return (expression_ == our_rhs->expression_);
+  }
+  Rule* Copy() const override { return new ValExpressionRule(*this); }
+
+ private:
+  string expression_;
+};
+
+// A map from register numbers to rules.
+class CallFrameInfo::RuleMap {
+ public:
+  RuleMap() : cfa_rule_(NULL) {}
+  RuleMap(const RuleMap& rhs) : cfa_rule_(NULL) { *this = rhs; }
+  ~RuleMap() { Clear(); }
+
+  RuleMap& operator=(const RuleMap& rhs);
+
+  // Set the rule for computing the CFA to RULE. Take ownership of RULE.
+  void SetCFARule(Rule* rule) {
+    delete cfa_rule_;
+    cfa_rule_ = rule;
+  }
+
+  // Return the current CFA rule. Unlike RegisterRule, this RuleMap retains
+  // ownership of the rule. We use this for DW_CFA_def_cfa_offset and
+  // DW_CFA_def_cfa_register, and for detecting references to the CFA before
+  // a rule for it has been established.
+  Rule* CFARule() const { return cfa_rule_; }
+
+  // Return the rule for REG, or NULL if there is none. The caller takes
+  // ownership of the result.
+  Rule* RegisterRule(int reg) const;
+
+  // Set the rule for computing REG to RULE. Take ownership of RULE.
+  void SetRegisterRule(int reg, Rule* rule);
+
+  // Make all the appropriate calls to HANDLER as if we were changing from
+  // this RuleMap to NEW_RULES at ADDRESS. We use this to implement
+  // DW_CFA_restore_state, where lots of rules can change simultaneously.
+  // Return true if all handlers returned true; otherwise, return false.
+  bool HandleTransitionTo(Handler* handler, uint64 address,
+                          const RuleMap& new_rules) const;
+
+ private:
+  // A map from register numbers to Rules.
+  typedef std::map<int, Rule*> RuleByNumber;
+
+  // Remove all register rules and clear cfa_rule_.
+  void Clear();
+
+  // The rule for computing the canonical frame address. This RuleMap owns
+  // this rule.
+  Rule* cfa_rule_;
+
+  // A map from register numbers to postfix expressions to recover
+  // their values. This RuleMap owns the Rules the map refers to.
+  RuleByNumber registers_;
+};
+
+CallFrameInfo::RuleMap& CallFrameInfo::RuleMap::operator=(const RuleMap& rhs) {
+  Clear();
+  // Since each map owns the rules it refers to, assignment must copy them.
+  if (rhs.cfa_rule_) cfa_rule_ = rhs.cfa_rule_->Copy();
+  for (RuleByNumber::const_iterator it = rhs.registers_.begin();
+       it != rhs.registers_.end(); it++)
+    registers_[it->first] = it->second->Copy();
+  return *this;
+}
+
+CallFrameInfo::Rule* CallFrameInfo::RuleMap::RegisterRule(int reg) const {
+  MOZ_ASSERT(reg != Handler::kCFARegister);
+  RuleByNumber::const_iterator it = registers_.find(reg);
+  if (it != registers_.end())
+    return it->second->Copy();
+  else
+    return NULL;
+}
+
+void CallFrameInfo::RuleMap::SetRegisterRule(int reg, Rule* rule) {
+  MOZ_ASSERT(reg != Handler::kCFARegister);
+  MOZ_ASSERT(rule);
+  Rule** slot = &registers_[reg];
+  delete *slot;
+  *slot = rule;
+}
+
+bool CallFrameInfo::RuleMap::HandleTransitionTo(
+    Handler* handler, uint64 address, const RuleMap& new_rules) const {
+  // Transition from cfa_rule_ to new_rules.cfa_rule_.
+  if (cfa_rule_ && new_rules.cfa_rule_) {
+    if (*cfa_rule_ != *new_rules.cfa_rule_ &&
+        !new_rules.cfa_rule_->Handle(handler, address, Handler::kCFARegister))
+      return false;
+  } else if (cfa_rule_) {
+    // this RuleMap has a CFA rule but new_rules doesn't.
+    // CallFrameInfo::Handler has no way to handle this --- and shouldn't;
+    // it's garbage input. The instruction interpreter should have
+    // detected this and warned, so take no action here.
+  } else if (new_rules.cfa_rule_) {
+    // This shouldn't be possible: NEW_RULES is some prior state, and
+    // there's no way to remove entries.
+    MOZ_ASSERT(0);
+  } else {
+    // Both CFA rules are empty.  No action needed.
+  }
+
+  // Traverse the two maps in order by register number, and report
+  // whatever differences we find.
+  RuleByNumber::const_iterator old_it = registers_.begin();
+  RuleByNumber::const_iterator new_it = new_rules.registers_.begin();
+  while (old_it != registers_.end() && new_it != new_rules.registers_.end()) {
+    if (old_it->first < new_it->first) {
+      // This RuleMap has an entry for old_it->first, but NEW_RULES
+      // doesn't.
+      //
+      // This isn't really the right thing to do, but since CFI generally
+      // only mentions callee-saves registers, and GCC's convention for
+      // callee-saves registers is that they are unchanged, it's a good
+      // approximation.
+      if (!handler->SameValueRule(address, old_it->first)) return false;
+      old_it++;
+    } else if (old_it->first > new_it->first) {
+      // NEW_RULES has entry for new_it->first, but this RuleMap
+      // doesn't. This shouldn't be possible: NEW_RULES is some prior
+      // state, and there's no way to remove entries.
+      MOZ_ASSERT(0);
+    } else {
+      // Both maps have an entry for this register. Report the new
+      // rule if it is different.
+      if (*old_it->second != *new_it->second &&
+          !new_it->second->Handle(handler, address, new_it->first))
+        return false;
+      new_it++;
+      old_it++;
+    }
+  }
+  // Finish off entries from this RuleMap with no counterparts in new_rules.
+  while (old_it != registers_.end()) {
+    if (!handler->SameValueRule(address, old_it->first)) return false;
+    old_it++;
+  }
+  // Since we only make transitions from a rule set to some previously
+  // saved rule set, and we can only add rules to the map, NEW_RULES
+  // must have fewer rules than *this.
+  MOZ_ASSERT(new_it == new_rules.registers_.end());
+
+  return true;
+}
+
+// Remove all register rules and clear cfa_rule_.
+void CallFrameInfo::RuleMap::Clear() {
+  delete cfa_rule_;
+  cfa_rule_ = NULL;
+  for (RuleByNumber::iterator it = registers_.begin(); it != registers_.end();
+       it++)
+    delete it->second;
+  registers_.clear();
+}
+
+// The state of the call frame information interpreter as it processes
+// instructions from a CIE and FDE.
+class CallFrameInfo::State {
+ public:
+  // Create a call frame information interpreter state with the given
+  // reporter, reader, handler, and initial call frame info address.
+  State(ByteReader* reader, Handler* handler, Reporter* reporter,
+        uint64 address)
+      : reader_(reader),
+        handler_(handler),
+        reporter_(reporter),
+        address_(address),
+        entry_(NULL),
+        cursor_(NULL),
+        saved_rules_(NULL) {}
+
+  ~State() {
+    if (saved_rules_) delete saved_rules_;
+  }
+
+  // Interpret instructions from CIE, save the resulting rule set for
+  // DW_CFA_restore instructions, and return true. On error, report
+  // the problem to reporter_ and return false.
+  bool InterpretCIE(const CIE& cie);
+
+  // Interpret instructions from FDE, and return true. On error,
+  // report the problem to reporter_ and return false.
+  bool InterpretFDE(const FDE& fde);
+
+ private:
+  // The operands of a CFI instruction, for ParseOperands.
+  struct Operands {
+    unsigned register_number;  // A register number.
+    uint64 offset;             // An offset or address.
+    long signed_offset;        // A signed offset.
+    string expression;         // A DWARF expression.
+  };
+
+  // Parse CFI instruction operands from STATE's instruction stream as
+  // described by FORMAT. On success, populate OPERANDS with the
+  // results, and return true. On failure, report the problem and
+  // return false.
+  //
+  // Each character of FORMAT should be one of the following:
+  //
+  //   'r'  unsigned LEB128 register number (OPERANDS->register_number)
+  //   'o'  unsigned LEB128 offset          (OPERANDS->offset)
+  //   's'  signed LEB128 offset            (OPERANDS->signed_offset)
+  //   'a'  machine-size address            (OPERANDS->offset)
+  //        (If the CIE has a 'z' augmentation string, 'a' uses the
+  //        encoding specified by the 'R' argument.)
+  //   '1'  a one-byte offset               (OPERANDS->offset)
+  //   '2'  a two-byte offset               (OPERANDS->offset)
+  //   '4'  a four-byte offset              (OPERANDS->offset)
+  //   '8'  an eight-byte offset            (OPERANDS->offset)
+  //   'e'  a DW_FORM_block holding a       (OPERANDS->expression)
+  //        DWARF expression
+  bool ParseOperands(const char* format, Operands* operands);
+
+  // Interpret one CFI instruction from STATE's instruction stream, update
+  // STATE, report any rule changes to handler_, and return true. On
+  // failure, report the problem and return false.
+  bool DoInstruction();
+
+  // The following Do* member functions are subroutines of DoInstruction,
+  // factoring out the actual work of operations that have several
+  // different encodings.
+
+  // Set the CFA rule to be the value of BASE_REGISTER plus OFFSET, and
+  // return true. On failure, report and return false. (Used for
+  // DW_CFA_def_cfa and DW_CFA_def_cfa_sf.)
+  bool DoDefCFA(unsigned base_register, long offset);
+
+  // Change the offset of the CFA rule to OFFSET, and return true. On
+  // failure, report and return false. (Subroutine for
+  // DW_CFA_def_cfa_offset and DW_CFA_def_cfa_offset_sf.)
+  bool DoDefCFAOffset(long offset);
+
+  // Specify that REG can be recovered using RULE, and return true. On
+  // failure, report and return false.
+  bool DoRule(unsigned reg, Rule* rule);
+
+  // Specify that REG can be found at OFFSET from the CFA, and return true.
+  // On failure, report and return false. (Subroutine for DW_CFA_offset,
+  // DW_CFA_offset_extended, and DW_CFA_offset_extended_sf.)
+  bool DoOffset(unsigned reg, long offset);
+
+  // Specify that the caller's value for REG is the CFA plus OFFSET,
+  // and return true. On failure, report and return false. (Subroutine
+  // for DW_CFA_val_offset and DW_CFA_val_offset_sf.)
+  bool DoValOffset(unsigned reg, long offset);
+
+  // Restore REG to the rule established in the CIE, and return true. On
+  // failure, report and return false. (Subroutine for DW_CFA_restore and
+  // DW_CFA_restore_extended.)
+  bool DoRestore(unsigned reg);
+
+  // Return the section offset of the instruction at cursor. For use
+  // in error messages.
+  uint64 CursorOffset() { return entry_->offset + (cursor_ - entry_->start); }
+
+  // Report that entry_ is incomplete, and return false. For brevity.
+  bool ReportIncomplete() {
+    reporter_->Incomplete(entry_->offset, entry_->kind);
+    return false;
+  }
+
+  // For reading multi-byte values with the appropriate endianness.
+  ByteReader* reader_;
+
+  // The handler to which we should report the data we find.
+  Handler* handler_;
+
+  // For reporting problems in the info we're parsing.
+  Reporter* reporter_;
+
+  // The code address to which the next instruction in the stream applies.
+  uint64 address_;
+
+  // The entry whose instructions we are currently processing. This is
+  // first a CIE, and then an FDE.
+  const Entry* entry_;
+
+  // The next instruction to process.
+  const char* cursor_;
+
+  // The current set of rules.
+  RuleMap rules_;
+
+  // The set of rules established by the CIE, used by DW_CFA_restore
+  // and DW_CFA_restore_extended. We set this after interpreting the
+  // CIE's instructions.
+  RuleMap cie_rules_;
+
+  // A stack of saved states, for DW_CFA_remember_state and
+  // DW_CFA_restore_state.
+  std::stack<RuleMap>* saved_rules_;
+};
+
+bool CallFrameInfo::State::InterpretCIE(const CIE& cie) {
+  entry_ = &cie;
+  cursor_ = entry_->instructions;
+  while (cursor_ < entry_->end)
+    if (!DoInstruction()) return false;
+  // Note the rules established by the CIE, for use by DW_CFA_restore
+  // and DW_CFA_restore_extended.
+  cie_rules_ = rules_;
+  return true;
+}
+
+bool CallFrameInfo::State::InterpretFDE(const FDE& fde) {
+  entry_ = &fde;
+  cursor_ = entry_->instructions;
+  while (cursor_ < entry_->end)
+    if (!DoInstruction()) return false;
+  return true;
+}
+
+bool CallFrameInfo::State::ParseOperands(const char* format,
+                                         Operands* operands) {
+  size_t len;
+  const char* operand;
+
+  for (operand = format; *operand; operand++) {
+    size_t bytes_left = entry_->end - cursor_;
+    switch (*operand) {
+      case 'r':
+        operands->register_number = reader_->ReadUnsignedLEB128(cursor_, &len);
+        if (len > bytes_left) return ReportIncomplete();
+        cursor_ += len;
+        break;
+
+      case 'o':
+        operands->offset = reader_->ReadUnsignedLEB128(cursor_, &len);
+        if (len > bytes_left) return ReportIncomplete();
+        cursor_ += len;
+        break;
+
+      case 's':
+        operands->signed_offset = reader_->ReadSignedLEB128(cursor_, &len);
+        if (len > bytes_left) return ReportIncomplete();
+        cursor_ += len;
+        break;
+
+      case 'a':
+        operands->offset = reader_->ReadEncodedPointer(
+            cursor_, entry_->cie->pointer_encoding, &len);
+        if (len > bytes_left) return ReportIncomplete();
+        cursor_ += len;
+        break;
+
+      case '1':
+        if (1 > bytes_left) return ReportIncomplete();
+        operands->offset = static_cast<unsigned char>(*cursor_++);
+        break;
+
+      case '2':
+        if (2 > bytes_left) return ReportIncomplete();
+        operands->offset = reader_->ReadTwoBytes(cursor_);
+        cursor_ += 2;
+        break;
+
+      case '4':
+        if (4 > bytes_left) return ReportIncomplete();
+        operands->offset = reader_->ReadFourBytes(cursor_);
+        cursor_ += 4;
+        break;
+
+      case '8':
+        if (8 > bytes_left) return ReportIncomplete();
+        operands->offset = reader_->ReadEightBytes(cursor_);
+        cursor_ += 8;
+        break;
+
+      case 'e': {
+        size_t expression_length = reader_->ReadUnsignedLEB128(cursor_, &len);
+        if (len > bytes_left || expression_length > bytes_left - len)
+          return ReportIncomplete();
+        cursor_ += len;
+        operands->expression = string(cursor_, expression_length);
+        cursor_ += expression_length;
+        break;
+      }
+
+      default:
+        MOZ_ASSERT(0);
+    }
+  }
+
+  return true;
+}
+
+bool CallFrameInfo::State::DoInstruction() {
+  CIE* cie = entry_->cie;
+  Operands ops;
+
+  // Our entry's kind should have been set by now.
+  MOZ_ASSERT(entry_->kind != kUnknown);
+
+  // We shouldn't have been invoked unless there were more
+  // instructions to parse.
+  MOZ_ASSERT(cursor_ < entry_->end);
+
+  unsigned opcode = *cursor_++;
+  if ((opcode & 0xc0) != 0) {
+    switch (opcode & 0xc0) {
+      // Advance the address.
+      case DW_CFA_advance_loc: {
+        size_t code_offset = opcode & 0x3f;
+        address_ += code_offset * cie->code_alignment_factor;
+        break;
+      }
+
+      // Find a register at an offset from the CFA.
+      case DW_CFA_offset:
+        if (!ParseOperands("o", &ops) ||
+            !DoOffset(opcode & 0x3f, ops.offset * cie->data_alignment_factor))
+          return false;
+        break;
+
+      // Restore the rule established for a register by the CIE.
+      case DW_CFA_restore:
+        if (!DoRestore(opcode & 0x3f)) return false;
+        break;
+
+      // The 'if' above should have excluded this possibility.
+      default:
+        MOZ_ASSERT(0);
+    }
+
+    // Return here, so the big switch below won't be indented.
+    return true;
+  }
+
+  switch (opcode) {
+    // Set the address.
+    case DW_CFA_set_loc:
+      if (!ParseOperands("a", &ops)) return false;
+      address_ = ops.offset;
+      break;
+
+    // Advance the address.
+    case DW_CFA_advance_loc1:
+      if (!ParseOperands("1", &ops)) return false;
+      address_ += ops.offset * cie->code_alignment_factor;
+      break;
+
+    // Advance the address.
+    case DW_CFA_advance_loc2:
+      if (!ParseOperands("2", &ops)) return false;
+      address_ += ops.offset * cie->code_alignment_factor;
+      break;
+
+    // Advance the address.
+    case DW_CFA_advance_loc4:
+      if (!ParseOperands("4", &ops)) return false;
+      address_ += ops.offset * cie->code_alignment_factor;
+      break;
+
+    // Advance the address.
+    case DW_CFA_MIPS_advance_loc8:
+      if (!ParseOperands("8", &ops)) return false;
+      address_ += ops.offset * cie->code_alignment_factor;
+      break;
+
+    // Compute the CFA by adding an offset to a register.
+    case DW_CFA_def_cfa:
+      if (!ParseOperands("ro", &ops) ||
+          !DoDefCFA(ops.register_number, ops.offset))
+        return false;
+      break;
+
+    // Compute the CFA by adding an offset to a register.
+    case DW_CFA_def_cfa_sf:
+      if (!ParseOperands("rs", &ops) ||
+          !DoDefCFA(ops.register_number,
+                    ops.signed_offset * cie->data_alignment_factor))
+        return false;
+      break;
+
+    // Change the base register used to compute the CFA.
+    case DW_CFA_def_cfa_register: {
+      Rule* cfa_rule = rules_.CFARule();
+      if (!cfa_rule) {
+        reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset());
+        return false;
+      }
+      if (!ParseOperands("r", &ops)) return false;
+      cfa_rule->SetBaseRegister(ops.register_number);
+      if (!cfa_rule->Handle(handler_, address_, Handler::kCFARegister))
+        return false;
+      break;
+    }
+
+    // Change the offset used to compute the CFA.
+    case DW_CFA_def_cfa_offset:
+      if (!ParseOperands("o", &ops) || !DoDefCFAOffset(ops.offset))
+        return false;
+      break;
+
+    // Change the offset used to compute the CFA.
+    case DW_CFA_def_cfa_offset_sf:
+      if (!ParseOperands("s", &ops) ||
+          !DoDefCFAOffset(ops.signed_offset * cie->data_alignment_factor))
+        return false;
+      break;
+
+    // Specify an expression whose value is the CFA.
+    case DW_CFA_def_cfa_expression: {
+      if (!ParseOperands("e", &ops)) return false;
+      Rule* rule = new ValExpressionRule(ops.expression);
+      rules_.SetCFARule(rule);
+      if (!rule->Handle(handler_, address_, Handler::kCFARegister))
+        return false;
+      break;
+    }
+
+    // The register's value cannot be recovered.
+    case DW_CFA_undefined: {
+      if (!ParseOperands("r", &ops) ||
+          !DoRule(ops.register_number, new UndefinedRule()))
+        return false;
+      break;
+    }
+
+    // The register's value is unchanged from its value in the caller.
+    case DW_CFA_same_value: {
+      if (!ParseOperands("r", &ops) ||
+          !DoRule(ops.register_number, new SameValueRule()))
+        return false;
+      break;
+    }
+
+    // Find a register at an offset from the CFA.
+    case DW_CFA_offset_extended:
+      if (!ParseOperands("ro", &ops) ||
+          !DoOffset(ops.register_number,
+                    ops.offset * cie->data_alignment_factor))
+        return false;
+      break;
+
+    // The register is saved at an offset from the CFA.
+    case DW_CFA_offset_extended_sf:
+      if (!ParseOperands("rs", &ops) ||
+          !DoOffset(ops.register_number,
+                    ops.signed_offset * cie->data_alignment_factor))
+        return false;
+      break;
+
+    // The register is saved at an offset from the CFA.
+    case DW_CFA_GNU_negative_offset_extended:
+      if (!ParseOperands("ro", &ops) ||
+          !DoOffset(ops.register_number,
+                    -ops.offset * cie->data_alignment_factor))
+        return false;
+      break;
+
+    // The register's value is the sum of the CFA plus an offset.
+    case DW_CFA_val_offset:
+      if (!ParseOperands("ro", &ops) ||
+          !DoValOffset(ops.register_number,
+                       ops.offset * cie->data_alignment_factor))
+        return false;
+      break;
+
+    // The register's value is the sum of the CFA plus an offset.
+    case DW_CFA_val_offset_sf:
+      if (!ParseOperands("rs", &ops) ||
+          !DoValOffset(ops.register_number,
+                       ops.signed_offset * cie->data_alignment_factor))
+        return false;
+      break;
+
+    // The register has been saved in another register.
+    case DW_CFA_register: {
+      if (!ParseOperands("ro", &ops) ||
+          !DoRule(ops.register_number, new RegisterRule(ops.offset)))
+        return false;
+      break;
+    }
+
+    // An expression yields the address at which the register is saved.
+    case DW_CFA_expression: {
+      if (!ParseOperands("re", &ops) ||
+          !DoRule(ops.register_number, new ExpressionRule(ops.expression)))
+        return false;
+      break;
+    }
+
+    // An expression yields the caller's value for the register.
+    case DW_CFA_val_expression: {
+      if (!ParseOperands("re", &ops) ||
+          !DoRule(ops.register_number, new ValExpressionRule(ops.expression)))
+        return false;
+      break;
+    }
+
+    // Restore the rule established for a register by the CIE.
+    case DW_CFA_restore_extended:
+      if (!ParseOperands("r", &ops) || !DoRestore(ops.register_number))
+        return false;
+      break;
+
+    // Save the current set of rules on a stack.
+    case DW_CFA_remember_state:
+      if (!saved_rules_) {
+        saved_rules_ = new std::stack<RuleMap>();
+      }
+      saved_rules_->push(rules_);
+      break;
+
+    // Pop the current set of rules off the stack.
+    case DW_CFA_restore_state: {
+      if (!saved_rules_ || saved_rules_->empty()) {
+        reporter_->EmptyStateStack(entry_->offset, entry_->kind,
+                                   CursorOffset());
+        return false;
+      }
+      const RuleMap& new_rules = saved_rules_->top();
+      if (rules_.CFARule() && !new_rules.CFARule()) {
+        reporter_->ClearingCFARule(entry_->offset, entry_->kind,
+                                   CursorOffset());
+        return false;
+      }
+      rules_.HandleTransitionTo(handler_, address_, new_rules);
+      rules_ = new_rules;
+      saved_rules_->pop();
+      break;
+    }
+
+    // No operation.  (Padding instruction.)
+    case DW_CFA_nop:
+      break;
+
+    // A SPARC register window save: Registers 8 through 15 (%o0-%o7)
+    // are saved in registers 24 through 31 (%i0-%i7), and registers
+    // 16 through 31 (%l0-%l7 and %i0-%i7) are saved at CFA offsets
+    // (0-15 * the register size). The register numbers must be
+    // hard-coded. A GNU extension, and not a pretty one.
+    case DW_CFA_GNU_window_save: {
+      // Save %o0-%o7 in %i0-%i7.
+      for (int i = 8; i < 16; i++)
+        if (!DoRule(i, new RegisterRule(i + 16))) return false;
+      // Save %l0-%l7 and %i0-%i7 at the CFA.
+      for (int i = 16; i < 32; i++)
+        // Assume that the byte reader's address size is the same as
+        // the architecture's register size. !@#%*^ hilarious.
+        if (!DoRule(i, new OffsetRule(Handler::kCFARegister,
+                                      (i - 16) * reader_->AddressSize())))
+          return false;
+      break;
+    }
+
+    // I'm not sure what this is. GDB doesn't use it for unwinding.
+    case DW_CFA_GNU_args_size:
+      if (!ParseOperands("o", &ops)) return false;
+      break;
+
+    // An opcode we don't recognize.
+    default: {
+      reporter_->BadInstruction(entry_->offset, entry_->kind, CursorOffset());
+      return false;
+    }
+  }
+
+  return true;
+}
+
+bool CallFrameInfo::State::DoDefCFA(unsigned base_register, long offset) {
+  Rule* rule = new ValOffsetRule(base_register, offset);
+  rules_.SetCFARule(rule);
+  return rule->Handle(handler_, address_, Handler::kCFARegister);
+}
+
+bool CallFrameInfo::State::DoDefCFAOffset(long offset) {
+  Rule* cfa_rule = rules_.CFARule();
+  if (!cfa_rule) {
+    reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset());
+    return false;
+  }
+  cfa_rule->SetOffset(offset);
+  return cfa_rule->Handle(handler_, address_, Handler::kCFARegister);
+}
+
+bool CallFrameInfo::State::DoRule(unsigned reg, Rule* rule) {
+  rules_.SetRegisterRule(reg, rule);
+  return rule->Handle(handler_, address_, reg);
+}
+
+bool CallFrameInfo::State::DoOffset(unsigned reg, long offset) {
+  if (!rules_.CFARule()) {
+    reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset());
+    return false;
+  }
+  return DoRule(reg, new OffsetRule(Handler::kCFARegister, offset));
+}
+
+bool CallFrameInfo::State::DoValOffset(unsigned reg, long offset) {
+  if (!rules_.CFARule()) {
+    reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset());
+    return false;
+  }
+  return DoRule(reg, new ValOffsetRule(Handler::kCFARegister, offset));
+}
+
+bool CallFrameInfo::State::DoRestore(unsigned reg) {
+  // DW_CFA_restore and DW_CFA_restore_extended don't make sense in a CIE.
+  if (entry_->kind == kCIE) {
+    reporter_->RestoreInCIE(entry_->offset, CursorOffset());
+    return false;
+  }
+  Rule* rule = cie_rules_.RegisterRule(reg);
+  if (!rule) {
+    // This isn't really the right thing to do, but since CFI generally
+    // only mentions callee-saves registers, and GCC's convention for
+    // callee-saves registers is that they are unchanged, it's a good
+    // approximation.
+    rule = new SameValueRule();
+  }
+  return DoRule(reg, rule);
+}
+
+bool CallFrameInfo::ReadEntryPrologue(const char* cursor, Entry* entry) {
+  const char* buffer_end = buffer_ + buffer_length_;
+
+  // Initialize enough of ENTRY for use in error reporting.
+  entry->offset = cursor - buffer_;
+  entry->start = cursor;
+  entry->kind = kUnknown;
+  entry->end = NULL;
+
+  // Read the initial length. This sets reader_'s offset size.
+  size_t length_size;
+  uint64 length = reader_->ReadInitialLength(cursor, &length_size);
+  if (length_size > size_t(buffer_end - cursor)) return ReportIncomplete(entry);
+  cursor += length_size;
+
+  // In a .eh_frame section, a length of zero marks the end of the series
+  // of entries.
+  if (length == 0 && eh_frame_) {
+    entry->kind = kTerminator;
+    entry->end = cursor;
+    return true;
+  }
+
+  // Validate the length.
+  if (length > size_t(buffer_end - cursor)) return ReportIncomplete(entry);
+
+  // The length is the number of bytes after the initial length field;
+  // we have that position handy at this point, so compute the end
+  // now. (If we're parsing 64-bit-offset DWARF on a 32-bit machine,
+  // and the length didn't fit in a size_t, we would have rejected it
+  // above.)
+  entry->end = cursor + length;
+
+  // Parse the next field: either the offset of a CIE or a CIE id.
+  size_t offset_size = reader_->OffsetSize();
+  if (offset_size > size_t(entry->end - cursor)) return ReportIncomplete(entry);
+  entry->id = reader_->ReadOffset(cursor);
+
+  // Don't advance cursor past id field yet; in .eh_frame data we need
+  // the id's position to compute the section offset of an FDE's CIE.
+
+  // Now we can decide what kind of entry this is.
+  if (eh_frame_) {
+    // In .eh_frame data, an ID of zero marks the entry as a CIE, and
+    // anything else is an offset from the id field of the FDE to the start
+    // of the CIE.
+    if (entry->id == 0) {
+      entry->kind = kCIE;
+    } else {
+      entry->kind = kFDE;
+      // Turn the offset from the id into an offset from the buffer's start.
+      entry->id = (cursor - buffer_) - entry->id;
+    }
+  } else {
+    // In DWARF CFI data, an ID of ~0 (of the appropriate width, given the
+    // offset size for the entry) marks the entry as a CIE, and anything
+    // else is the offset of the CIE from the beginning of the section.
+    if (offset_size == 4)
+      entry->kind = (entry->id == 0xffffffff) ? kCIE : kFDE;
+    else {
+      MOZ_ASSERT(offset_size == 8);
+      entry->kind = (entry->id == 0xffffffffffffffffULL) ? kCIE : kFDE;
+    }
+  }
+
+  // Now advance cursor past the id.
+  cursor += offset_size;
+
+  // The fields specific to this kind of entry start here.
+  entry->fields = cursor;
+
+  entry->cie = NULL;
+
+  return true;
+}
+
+bool CallFrameInfo::ReadCIEFields(CIE* cie) {
+  const char* cursor = cie->fields;
+  size_t len;
+
+  MOZ_ASSERT(cie->kind == kCIE);
+
+  // Prepare for early exit.
+  cie->version = 0;
+  cie->augmentation.clear();
+  cie->code_alignment_factor = 0;
+  cie->data_alignment_factor = 0;
+  cie->return_address_register = 0;
+  cie->has_z_augmentation = false;
+  cie->pointer_encoding = DW_EH_PE_absptr;
+  cie->instructions = 0;
+
+  // Parse the version number.
+  if (cie->end - cursor < 1) return ReportIncomplete(cie);
+  cie->version = reader_->ReadOneByte(cursor);
+  cursor++;
+
+  // If we don't recognize the version, we can't parse any more fields of the
+  // CIE. For DWARF CFI, we handle versions 1 through 4 (there was never a
+  // version 2 of CFI data). For .eh_frame, we handle versions 1 and 4 as well;
+  // the difference between those versions seems to be the same as for
+  // .debug_frame.
+  if (cie->version < 1 || cie->version > 4) {
+    reporter_->UnrecognizedVersion(cie->offset, cie->version);
+    return false;
+  }
+
+  const char* augmentation_start = cursor;
+  const void* augmentation_end =
+      memchr(augmentation_start, '\0', cie->end - augmentation_start);
+  if (!augmentation_end) return ReportIncomplete(cie);
+  cursor = static_cast<const char*>(augmentation_end);
+  cie->augmentation = string(augmentation_start, cursor - augmentation_start);
+  // Skip the terminating '\0'.
+  cursor++;
+
+  // Is this CFI augmented?
+  if (!cie->augmentation.empty()) {
+    // Is it an augmentation we recognize?
+    if (cie->augmentation[0] == DW_Z_augmentation_start) {
+      // Linux C++ ABI 'z' augmentation, used for exception handling data.
+      cie->has_z_augmentation = true;
+    } else {
+      // Not an augmentation we recognize. Augmentations can have arbitrary
+      // effects on the form of rest of the content, so we have to give up.
+      reporter_->UnrecognizedAugmentation(cie->offset, cie->augmentation);
+      return false;
+    }
+  }
+
+  if (cie->version >= 4) {
+    // Check that the address_size and segment_size fields are plausible.
+    if (cie->end - cursor < 2) {
+      return ReportIncomplete(cie);
+    }
+    uint8_t address_size = reader_->ReadOneByte(cursor);
+    cursor++;
+    if (address_size != sizeof(void*)) {
+      // This is not per-se invalid CFI.  But we can reasonably expect to
+      // be running on a target of the same word size as the CFI is for,
+      // so we reject this case.
+      reporter_->InvalidDwarf4Artefact(cie->offset, "Invalid address_size");
+      return false;
+    }
+    uint8_t segment_size = reader_->ReadOneByte(cursor);
+    cursor++;
+    if (segment_size != 0) {
+      // This is also not per-se invalid CFI, but we don't currently handle
+      // the case of non-zero |segment_size|.
+      reporter_->InvalidDwarf4Artefact(cie->offset, "Invalid segment_size");
+      return false;
+    }
+    // We only continue parsing if |segment_size| is zero.  If this routine
+    // is ever changed to allow non-zero |segment_size|, then
+    // ReadFDEFields() below will have to be changed to match, per comments
+    // there.
+  }
+
+  // Parse the code alignment factor.
+  cie->code_alignment_factor = reader_->ReadUnsignedLEB128(cursor, &len);
+  if (size_t(cie->end - cursor) < len) return ReportIncomplete(cie);
+  cursor += len;
+
+  // Parse the data alignment factor.
+  cie->data_alignment_factor = reader_->ReadSignedLEB128(cursor, &len);
+  if (size_t(cie->end - cursor) < len) return ReportIncomplete(cie);
+  cursor += len;
+
+  // Parse the return address register. This is a ubyte in version 1, and
+  // a ULEB128 in version 3.
+  if (cie->version == 1) {
+    if (cursor >= cie->end) return ReportIncomplete(cie);
+    cie->return_address_register = uint8(*cursor++);
+  } else {
+    cie->return_address_register = reader_->ReadUnsignedLEB128(cursor, &len);
+    if (size_t(cie->end - cursor) < len) return ReportIncomplete(cie);
+    cursor += len;
+  }
+
+  // If we have a 'z' augmentation string, find the augmentation data and
+  // use the augmentation string to parse it.
+  if (cie->has_z_augmentation) {
+    uint64_t data_size = reader_->ReadUnsignedLEB128(cursor, &len);
+    if (size_t(cie->end - cursor) < len + data_size)
+      return ReportIncomplete(cie);
+    cursor += len;
+    const char* data = cursor;
+    cursor += data_size;
+    const char* data_end = cursor;
+
+    cie->has_z_lsda = false;
+    cie->has_z_personality = false;
+    cie->has_z_signal_frame = false;
+
+    // Walk the augmentation string, and extract values from the
+    // augmentation data as the string directs.
+    for (size_t i = 1; i < cie->augmentation.size(); i++) {
+      switch (cie->augmentation[i]) {
+        case DW_Z_has_LSDA:
+          // The CIE's augmentation data holds the language-specific data
+          // area pointer's encoding, and the FDE's augmentation data holds
+          // the pointer itself.
+          cie->has_z_lsda = true;
+          // Fetch the LSDA encoding from the augmentation data.
+          if (data >= data_end) return ReportIncomplete(cie);
+          cie->lsda_encoding = DwarfPointerEncoding(*data++);
+          if (!reader_->ValidEncoding(cie->lsda_encoding)) {
+            reporter_->InvalidPointerEncoding(cie->offset, cie->lsda_encoding);
+            return false;
+          }
+          // Don't check if the encoding is usable here --- we haven't
+          // read the FDE's fields yet, so we're not prepared for
+          // DW_EH_PE_funcrel, although that's a fine encoding for the
+          // LSDA to use, since it appears in the FDE.
+          break;
+
+        case DW_Z_has_personality_routine:
+          // The CIE's augmentation data holds the personality routine
+          // pointer's encoding, followed by the pointer itself.
+          cie->has_z_personality = true;
+          // Fetch the personality routine pointer's encoding from the
+          // augmentation data.
+          if (data >= data_end) return ReportIncomplete(cie);
+          cie->personality_encoding = DwarfPointerEncoding(*data++);
+          if (!reader_->ValidEncoding(cie->personality_encoding)) {
+            reporter_->InvalidPointerEncoding(cie->offset,
+                                              cie->personality_encoding);
+            return false;
+          }
+          if (!reader_->UsableEncoding(cie->personality_encoding)) {
+            reporter_->UnusablePointerEncoding(cie->offset,
+                                               cie->personality_encoding);
+            return false;
+          }
+          // Fetch the personality routine's pointer itself from the data.
+          cie->personality_address = reader_->ReadEncodedPointer(
+              data, cie->personality_encoding, &len);
+          if (len > size_t(data_end - data)) return ReportIncomplete(cie);
+          data += len;
+          break;
+
+        case DW_Z_has_FDE_address_encoding:
+          // The CIE's augmentation data holds the pointer encoding to use
+          // for addresses in the FDE.
+          if (data >= data_end) return ReportIncomplete(cie);
+          cie->pointer_encoding = DwarfPointerEncoding(*data++);
+          if (!reader_->ValidEncoding(cie->pointer_encoding)) {
+            reporter_->InvalidPointerEncoding(cie->offset,
+                                              cie->pointer_encoding);
+            return false;
+          }
+          if (!reader_->UsableEncoding(cie->pointer_encoding)) {
+            reporter_->UnusablePointerEncoding(cie->offset,
+                                               cie->pointer_encoding);
+            return false;
+          }
+          break;
+
+        case DW_Z_is_signal_trampoline:
+          // Frames using this CIE are signal delivery frames.
+          cie->has_z_signal_frame = true;
+          break;
+
+        default:
+          // An augmentation we don't recognize.
+          reporter_->UnrecognizedAugmentation(cie->offset, cie->augmentation);
+          return false;
+      }
+    }
+  }
+
+  // The CIE's instructions start here.
+  cie->instructions = cursor;
+
+  return true;
+}
+
+bool CallFrameInfo::ReadFDEFields(FDE* fde) {
+  const char* cursor = fde->fields;
+  size_t size;
+
+  // At this point, for Dwarf 4 and above, we are assuming that the
+  // associated CIE has its |segment_size| field equal to zero.  This is
+  // checked for in ReadCIEFields() above.  If ReadCIEFields() is ever
+  // changed to allow non-zero |segment_size| CIEs then we will have to read
+  // the segment_selector value at this point.
+
+  fde->address =
+      reader_->ReadEncodedPointer(cursor, fde->cie->pointer_encoding, &size);
+  if (size > size_t(fde->end - cursor)) return ReportIncomplete(fde);
+  cursor += size;
+  reader_->SetFunctionBase(fde->address);
+
+  // For the length, we strip off the upper nybble of the encoding used for
+  // the starting address.
+  DwarfPointerEncoding length_encoding =
+      DwarfPointerEncoding(fde->cie->pointer_encoding & 0x0f);
+  fde->size = reader_->ReadEncodedPointer(cursor, length_encoding, &size);
+  if (size > size_t(fde->end - cursor)) return ReportIncomplete(fde);
+  cursor += size;
+
+  // If the CIE has a 'z' augmentation string, then augmentation data
+  // appears here.
+  if (fde->cie->has_z_augmentation) {
+    uint64_t data_size = reader_->ReadUnsignedLEB128(cursor, &size);
+    if (size_t(fde->end - cursor) < size + data_size)
+      return ReportIncomplete(fde);
+    cursor += size;
+
+    // In the abstract, we should walk the augmentation string, and extract
+    // items from the FDE's augmentation data as we encounter augmentation
+    // string characters that specify their presence: the ordering of items
+    // in the augmentation string determines the arrangement of values in
+    // the augmentation data.
+    //
+    // In practice, there's only ever one value in FDE augmentation data
+    // that we support --- the LSDA pointer --- and we have to bail if we
+    // see any unrecognized augmentation string characters. So if there is
+    // anything here at all, we know what it is, and where it starts.
+    if (fde->cie->has_z_lsda) {
+      // Check whether the LSDA's pointer encoding is usable now: only once
+      // we've parsed the FDE's starting address do we call reader_->
+      // SetFunctionBase, so that the DW_EH_PE_funcrel encoding becomes
+      // usable.
+      if (!reader_->UsableEncoding(fde->cie->lsda_encoding)) {
+        reporter_->UnusablePointerEncoding(fde->cie->offset,
+                                           fde->cie->lsda_encoding);
+        return false;
+      }
+
+      fde->lsda_address =
+          reader_->ReadEncodedPointer(cursor, fde->cie->lsda_encoding, &size);
+      if (size > data_size) return ReportIncomplete(fde);
+      // Ideally, we would also complain here if there were unconsumed
+      // augmentation data.
+    }
+
+    cursor += data_size;
+  }
+
+  // The FDE's instructions start after those.
+  fde->instructions = cursor;
+
+  return true;
+}
+
+bool CallFrameInfo::Start() {
+  const char* buffer_end = buffer_ + buffer_length_;
+  const char* cursor;
+  bool all_ok = true;
+  const char* entry_end;
+  bool ok;
+
+  // Traverse all the entries in buffer_, skipping CIEs and offering
+  // FDEs to the handler.
+  for (cursor = buffer_; cursor < buffer_end;
+       cursor = entry_end, all_ok = all_ok && ok) {
+    FDE fde;
+
+    // Make it easy to skip this entry with 'continue': assume that
+    // things are not okay until we've checked all the data, and
+    // prepare the address of the next entry.
+    ok = false;
+
+    // Read the entry's prologue.
+    if (!ReadEntryPrologue(cursor, &fde)) {
+      if (!fde.end) {
+        // If we couldn't even figure out this entry's extent, then we
+        // must stop processing entries altogether.
+        all_ok = false;
+        break;
+      }
+      entry_end = fde.end;
+      continue;
+    }
+
+    // The next iteration picks up after this entry.
+    entry_end = fde.end;
+
+    // Did we see an .eh_frame terminating mark?
+    if (fde.kind == kTerminator) {
+      // If there appears to be more data left in the section after the
+      // terminating mark, warn the user. But this is just a warning;
+      // we leave all_ok true.
+      if (fde.end < buffer_end) reporter_->EarlyEHTerminator(fde.offset);
+      break;
+    }
+
+    // In this loop, we skip CIEs. We only parse them fully when we
+    // parse an FDE that refers to them. This limits our memory
+    // consumption (beyond the buffer itself) to that needed to
+    // process the largest single entry.
+    if (fde.kind != kFDE) {
+      ok = true;
+      continue;
+    }
+
+    // Validate the CIE pointer.
+    if (fde.id > buffer_length_) {
+      reporter_->CIEPointerOutOfRange(fde.offset, fde.id);
+      continue;
+    }
+
+    CIE cie;
+
+    // Parse this FDE's CIE header.
+    if (!ReadEntryPrologue(buffer_ + fde.id, &cie)) continue;
+    // This had better be an actual CIE.
+    if (cie.kind != kCIE) {
+      reporter_->BadCIEId(fde.offset, fde.id);
+      continue;
+    }
+    if (!ReadCIEFields(&cie)) continue;
+
+    // We now have the values that govern both the CIE and the FDE.
+    cie.cie = &cie;
+    fde.cie = &cie;
+
+    // Parse the FDE's header.
+    if (!ReadFDEFields(&fde)) continue;
+
+    // Call Entry to ask the consumer if they're interested.
+    if (!handler_->Entry(fde.offset, fde.address, fde.size, cie.version,
+                         cie.augmentation, cie.return_address_register)) {
+      // The handler isn't interested in this entry. That's not an error.
+      ok = true;
+      continue;
+    }
+
+    if (cie.has_z_augmentation) {
+      // Report the personality routine address, if we have one.
+      if (cie.has_z_personality) {
+        if (!handler_->PersonalityRoutine(
+                cie.personality_address,
+                IsIndirectEncoding(cie.personality_encoding)))
+          continue;
+      }
+
+      // Report the language-specific data area address, if we have one.
+      if (cie.has_z_lsda) {
+        if (!handler_->LanguageSpecificDataArea(
+                fde.lsda_address, IsIndirectEncoding(cie.lsda_encoding)))
+          continue;
+      }
+
+      // If this is a signal-handling frame, report that.
+      if (cie.has_z_signal_frame) {
+        if (!handler_->SignalHandler()) continue;
+      }
+    }
+
+    // Interpret the CIE's instructions, and then the FDE's instructions.
+    State state(reader_, handler_, reporter_, fde.address);
+    ok = state.InterpretCIE(cie) && state.InterpretFDE(fde);
+
+    // Tell the ByteReader that the function start address from the
+    // FDE header is no longer valid.
+    reader_->ClearFunctionBase();
+
+    // Report the end of the entry.
+    handler_->End();
+  }
+
+  return all_ok;
+}
+
+const char* CallFrameInfo::KindName(EntryKind kind) {
+  if (kind == CallFrameInfo::kUnknown)
+    return "entry";
+  else if (kind == CallFrameInfo::kCIE)
+    return "common information entry";
+  else if (kind == CallFrameInfo::kFDE)
+    return "frame description entry";
+  else {
+    MOZ_ASSERT(kind == CallFrameInfo::kTerminator);
+    return ".eh_frame sequence terminator";
+  }
+}
+
+bool CallFrameInfo::ReportIncomplete(Entry* entry) {
+  reporter_->Incomplete(entry->offset, entry->kind);
+  return false;
+}
+
+void CallFrameInfo::Reporter::Incomplete(uint64 offset,
+                                         CallFrameInfo::EntryKind kind) {
+  char buf[300];
+  SprintfLiteral(buf, "%s: CFI %s at offset 0x%llx in '%s': entry ends early\n",
+                 filename_.c_str(), CallFrameInfo::KindName(kind), offset,
+                 section_.c_str());
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::EarlyEHTerminator(uint64 offset) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI at offset 0x%llx in '%s': saw end-of-data marker"
+                 " before end of section contents\n",
+                 filename_.c_str(), offset, section_.c_str());
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::CIEPointerOutOfRange(uint64 offset,
+                                                   uint64 cie_offset) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI frame description entry at offset 0x%llx in '%s':"
+                 " CIE pointer is out of range: 0x%llx\n",
+                 filename_.c_str(), offset, section_.c_str(), cie_offset);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::BadCIEId(uint64 offset, uint64 cie_offset) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI frame description entry at offset 0x%llx in '%s':"
+                 " CIE pointer does not point to a CIE: 0x%llx\n",
+                 filename_.c_str(), offset, section_.c_str(), cie_offset);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::UnrecognizedVersion(uint64 offset, int version) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI frame description entry at offset 0x%llx in '%s':"
+                 " CIE specifies unrecognized version: %d\n",
+                 filename_.c_str(), offset, section_.c_str(), version);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::UnrecognizedAugmentation(uint64 offset,
+                                                       const string& aug) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI frame description entry at offset 0x%llx in '%s':"
+                 " CIE specifies unrecognized augmentation: '%s'\n",
+                 filename_.c_str(), offset, section_.c_str(), aug.c_str());
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::InvalidDwarf4Artefact(uint64 offset,
+                                                    const char* what) {
+  char* what_safe = strndup(what, 100);
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI frame description entry at offset 0x%llx in '%s':"
+                 " CIE specifies invalid Dwarf4 artefact: %s\n",
+                 filename_.c_str(), offset, section_.c_str(), what_safe);
+  log_(buf);
+  free(what_safe);
+}
+
+void CallFrameInfo::Reporter::InvalidPointerEncoding(uint64 offset,
+                                                     uint8 encoding) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI common information entry at offset 0x%llx in '%s':"
+                 " 'z' augmentation specifies invalid pointer encoding: "
+                 "0x%02x\n",
+                 filename_.c_str(), offset, section_.c_str(), encoding);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::UnusablePointerEncoding(uint64 offset,
+                                                      uint8 encoding) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI common information entry at offset 0x%llx in '%s':"
+                 " 'z' augmentation specifies a pointer encoding for which"
+                 " we have no base address: 0x%02x\n",
+                 filename_.c_str(), offset, section_.c_str(), encoding);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::RestoreInCIE(uint64 offset, uint64 insn_offset) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI common information entry at offset 0x%llx in '%s':"
+                 " the DW_CFA_restore instruction at offset 0x%llx"
+                 " cannot be used in a common information entry\n",
+                 filename_.c_str(), offset, section_.c_str(), insn_offset);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::BadInstruction(uint64 offset,
+                                             CallFrameInfo::EntryKind kind,
+                                             uint64 insn_offset) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI %s at offset 0x%llx in section '%s':"
+                 " the instruction at offset 0x%llx is unrecognized\n",
+                 filename_.c_str(), CallFrameInfo::KindName(kind), offset,
+                 section_.c_str(), insn_offset);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::NoCFARule(uint64 offset,
+                                        CallFrameInfo::EntryKind kind,
+                                        uint64 insn_offset) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI %s at offset 0x%llx in section '%s':"
+                 " the instruction at offset 0x%llx assumes that a CFA rule "
+                 "has been set, but none has been set\n",
+                 filename_.c_str(), CallFrameInfo::KindName(kind), offset,
+                 section_.c_str(), insn_offset);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::EmptyStateStack(uint64 offset,
+                                              CallFrameInfo::EntryKind kind,
+                                              uint64 insn_offset) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI %s at offset 0x%llx in section '%s':"
+                 " the DW_CFA_restore_state instruction at offset 0x%llx"
+                 " should pop a saved state from the stack, but the stack "
+                 "is empty\n",
+                 filename_.c_str(), CallFrameInfo::KindName(kind), offset,
+                 section_.c_str(), insn_offset);
+  log_(buf);
+}
+
+void CallFrameInfo::Reporter::ClearingCFARule(uint64 offset,
+                                              CallFrameInfo::EntryKind kind,
+                                              uint64 insn_offset) {
+  char buf[300];
+  SprintfLiteral(buf,
+                 "%s: CFI %s at offset 0x%llx in section '%s':"
+                 " the DW_CFA_restore_state instruction at offset 0x%llx"
+                 " would clear the CFA rule in effect\n",
+                 filename_.c_str(), CallFrameInfo::KindName(kind), offset,
+                 section_.c_str(), insn_offset);
+  log_(buf);
+}
+
+unsigned int DwarfCFIToModule::RegisterNames::I386() {
+  /*
+   8 "$eax", "$ecx", "$edx", "$ebx", "$esp", "$ebp", "$esi", "$edi",
+   3 "$eip", "$eflags", "$unused1",
+   8 "$st0", "$st1", "$st2", "$st3", "$st4", "$st5", "$st6", "$st7",
+   2 "$unused2", "$unused3",
+   8 "$xmm0", "$xmm1", "$xmm2", "$xmm3", "$xmm4", "$xmm5", "$xmm6", "$xmm7",
+   8 "$mm0", "$mm1", "$mm2", "$mm3", "$mm4", "$mm5", "$mm6", "$mm7",
+   3 "$fcw", "$fsw", "$mxcsr",
+   8 "$es", "$cs", "$ss", "$ds", "$fs", "$gs", "$unused4", "$unused5",
+   2 "$tr", "$ldtr"
+  */
+  return 8 + 3 + 8 + 2 + 8 + 8 + 3 + 8 + 2;
+}
+
+unsigned int DwarfCFIToModule::RegisterNames::X86_64() {
+  /*
+   8 "$rax", "$rdx", "$rcx", "$rbx", "$rsi", "$rdi", "$rbp", "$rsp",
+   8 "$r8",  "$r9",  "$r10", "$r11", "$r12", "$r13", "$r14", "$r15",
+   1 "$rip",
+   8 "$xmm0","$xmm1","$xmm2", "$xmm3", "$xmm4", "$xmm5", "$xmm6", "$xmm7",
+   8 "$xmm8","$xmm9","$xmm10","$xmm11","$xmm12","$xmm13","$xmm14","$xmm15",
+   8 "$st0", "$st1", "$st2", "$st3", "$st4", "$st5", "$st6", "$st7",
+   8 "$mm0", "$mm1", "$mm2", "$mm3", "$mm4", "$mm5", "$mm6", "$mm7",
+   1 "$rflags",
+   8 "$es", "$cs", "$ss", "$ds", "$fs", "$gs", "$unused1", "$unused2",
+   4 "$fs.base", "$gs.base", "$unused3", "$unused4",
+   2 "$tr", "$ldtr",
+   3 "$mxcsr", "$fcw", "$fsw"
+  */
+  return 8 + 8 + 1 + 8 + 8 + 8 + 8 + 1 + 8 + 4 + 2 + 3;
+}
+
+// Per ARM IHI 0040A, section 3.1
+unsigned int DwarfCFIToModule::RegisterNames::ARM() {
+  /*
+   8 "r0",  "r1",  "r2",  "r3",  "r4",  "r5",  "r6",  "r7",
+   8 "r8",  "r9",  "r10", "r11", "r12", "sp",  "lr",  "pc",
+   8 "f0",  "f1",  "f2",  "f3",  "f4",  "f5",  "f6",  "f7",
+   8 "fps", "cpsr", "",   "",    "",    "",    "",    "",
+   8 "",    "",    "",    "",    "",    "",    "",    "",
+   8 "",    "",    "",    "",    "",    "",    "",    "",
+   8 "",    "",    "",    "",    "",    "",    "",    "",
+   8 "",    "",    "",    "",    "",    "",    "",    "",
+   8 "s0",  "s1",  "s2",  "s3",  "s4",  "s5",  "s6",  "s7",
+   8 "s8",  "s9",  "s10", "s11", "s12", "s13", "s14", "s15",
+   8 "s16", "s17", "s18", "s19", "s20", "s21", "s22", "s23",
+   8 "s24", "s25", "s26", "s27", "s28", "s29", "s30", "s31",
+   8 "f0",  "f1",  "f2",  "f3",  "f4",  "f5",  "f6",  "f7"
+  */
+  return 13 * 8;
+}
+
+// Per ARM IHI 0057A, section 3.1
+unsigned int DwarfCFIToModule::RegisterNames::ARM64() {
+  /*
+   8 "x0",  "x1",  "x2",  "x3",  "x4",  "x5",  "x6",  "x7",
+   8 "x8",  "x9",  "x10", "x11", "x12", "x13", "x14", "x15",
+   8 "x16"  "x17", "x18", "x19", "x20", "x21", "x22", "x23",
+   8 "x24", "x25", "x26", "x27", "x28", "x29",  "x30","sp",
+   8 "",    "",    "",    "",    "",    "",    "",    "",
+   8 "",    "",    "",    "",    "",    "",    "",    "",
+   8 "",    "",    "",    "",    "",    "",    "",    "",
+   8 "",    "",    "",    "",    "",    "",    "",    "",
+   8 "v0",  "v1",  "v2",  "v3",  "v4",  "v5",  "v6",  "v7",
+   8 "v8",  "v9",  "v10", "v11", "v12", "v13", "v14", "v15",
+   8 "v16", "v17", "v18", "v19", "v20", "v21", "v22,  "v23",
+   8 "v24", "x25", "x26,  "x27", "v28", "v29", "v30", "v31",
+  */
+  return 12 * 8;
+}
+
+unsigned int DwarfCFIToModule::RegisterNames::MIPS() {
+  /*
+   8 "$zero", "$at",  "$v0",  "$v1",  "$a0",   "$a1",  "$a2",  "$a3",
+   8 "$t0",   "$t1",  "$t2",  "$t3",  "$t4",   "$t5",  "$t6",  "$t7",
+   8 "$s0",   "$s1",  "$s2",  "$s3",  "$s4",   "$s5",  "$s6",  "$s7",
+   8 "$t8",   "$t9",  "$k0",  "$k1",  "$gp",   "$sp",  "$fp",  "$ra",
+   9 "$lo",   "$hi",  "$pc",  "$f0",  "$f1",   "$f2",  "$f3",  "$f4",  "$f5",
+   8 "$f6",   "$f7",  "$f8",  "$f9",  "$f10",  "$f11", "$f12", "$f13",
+   7 "$f14",  "$f15", "$f16", "$f17", "$f18",  "$f19", "$f20",
+   7 "$f21",  "$f22", "$f23", "$f24", "$f25",  "$f26", "$f27",
+   6 "$f28",  "$f29", "$f30", "$f31", "$fcsr", "$fir"
+  */
+  return 8 + 8 + 8 + 8 + 9 + 8 + 7 + 7 + 6;
+}
+
+// See prototype for comments.
+int32_t parseDwarfExpr(Summariser* summ, const ByteReader* reader, string expr,
+                       bool debug, bool pushCfaAtStart, bool derefAtEnd) {
+  const char* cursor = expr.c_str();
+  const char* end1 = cursor + expr.length();
+
+  char buf[100];
+  if (debug) {
+    SprintfLiteral(buf, "LUL.DW  << DwarfExpr, len is %d\n",
+                   (int)(end1 - cursor));
+    summ->Log(buf);
+  }
+
+  // Add a marker for the start of this expression.  In it, indicate
+  // whether or not the CFA should be pushed onto the stack prior to
+  // evaluation.
+  int32_t start_ix =
+      summ->AddPfxInstr(PfxInstr(PX_Start, pushCfaAtStart ? 1 : 0));
+  MOZ_ASSERT(start_ix >= 0);
+
+  while (cursor < end1) {
+    uint8 opc = reader->ReadOneByte(cursor);
+    cursor++;
+
+    const char* nm = nullptr;
+    PfxExprOp pxop = PX_End;
+
+    switch (opc) {
+      case DW_OP_lit0 ... DW_OP_lit31: {
+        int32_t simm32 = (int32_t)(opc - DW_OP_lit0);
+        if (debug) {
+          SprintfLiteral(buf, "LUL.DW   DW_OP_lit%d\n", (int)simm32);
+          summ->Log(buf);
+        }
+        (void)summ->AddPfxInstr(PfxInstr(PX_SImm32, simm32));
+        break;
+      }
+
+      case DW_OP_breg0 ... DW_OP_breg31: {
+        size_t len;
+        int64_t n = reader->ReadSignedLEB128(cursor, &len);
+        cursor += len;
+        DW_REG_NUMBER reg = (DW_REG_NUMBER)(opc - DW_OP_breg0);
+        if (debug) {
+          SprintfLiteral(buf, "LUL.DW   DW_OP_breg%d %lld\n", (int)reg,
+                         (long long int)n);
+          summ->Log(buf);
+        }
+        // PfxInstr only allows a 32 bit signed offset.  So we
+        // must fail if the immediate is out of range.
+        if (n < INT32_MIN || INT32_MAX < n) goto fail;
+        (void)summ->AddPfxInstr(PfxInstr(PX_DwReg, reg));
+        (void)summ->AddPfxInstr(PfxInstr(PX_SImm32, (int32_t)n));
+        (void)summ->AddPfxInstr(PfxInstr(PX_Add));
+        break;
+      }
+
+      case DW_OP_const4s: {
+        uint64_t u64 = reader->ReadFourBytes(cursor);
+        cursor += 4;
+        // u64 is guaranteed by |ReadFourBytes| to be in the
+        // range 0 .. FFFFFFFF inclusive.  But to be safe:
+        uint32_t u32 = (uint32_t)(u64 & 0xFFFFFFFF);
+        int32_t s32 = (int32_t)u32;
+        if (debug) {
+          SprintfLiteral(buf, "LUL.DW   DW_OP_const4s %d\n", (int)s32);
+          summ->Log(buf);
+        }
+        (void)summ->AddPfxInstr(PfxInstr(PX_SImm32, s32));
+        break;
+      }
+
+      case DW_OP_deref:
+        nm = "deref";
+        pxop = PX_Deref;
+        goto no_operands;
+      case DW_OP_and:
+        nm = "and";
+        pxop = PX_And;
+        goto no_operands;
+      case DW_OP_plus:
+        nm = "plus";
+        pxop = PX_Add;
+        goto no_operands;
+      case DW_OP_minus:
+        nm = "minus";
+        pxop = PX_Sub;
+        goto no_operands;
+      case DW_OP_shl:
+        nm = "shl";
+        pxop = PX_Shl;
+        goto no_operands;
+      case DW_OP_ge:
+        nm = "ge";
+        pxop = PX_CmpGES;
+        goto no_operands;
+      no_operands:
+        MOZ_ASSERT(nm && pxop != PX_End);
+        if (debug) {
+          SprintfLiteral(buf, "LUL.DW   DW_OP_%s\n", nm);
+          summ->Log(buf);
+        }
+        (void)summ->AddPfxInstr(PfxInstr(pxop));
+        break;
+
+      default:
+        if (debug) {
+          SprintfLiteral(buf, "LUL.DW   unknown opc %d\n", (int)opc);
+          summ->Log(buf);
+        }
+        goto fail;
+
+    }  // switch (opc)
+
+  }  // while (cursor < end1)
+
+  MOZ_ASSERT(cursor >= end1);
+
+  if (cursor > end1) {
+    // We overran the Dwarf expression.  Give up.
+    goto fail;
+  }
+
+  // For DW_CFA_expression, what the expression denotes is the address
+  // of where the previous value is located.  The caller of this routine
+  // may therefore request one last dereference before the end marker is
+  // inserted.
+  if (derefAtEnd) {
+    (void)summ->AddPfxInstr(PfxInstr(PX_Deref));
+  }
+
+  // Insert an end marker, and declare success.
+  (void)summ->AddPfxInstr(PfxInstr(PX_End));
+  if (debug) {
+    SprintfLiteral(buf,
+                   "LUL.DW   conversion of dwarf expression succeeded, "
+                   "ix = %d\n",
+                   (int)start_ix);
+    summ->Log(buf);
+    summ->Log("LUL.DW  >>\n");
+  }
+  return start_ix;
+
+fail:
+  if (debug) {
+    summ->Log("LUL.DW   conversion of dwarf expression failed\n");
+    summ->Log("LUL.DW  >>\n");
+  }
+  return -1;
+}
+
+bool DwarfCFIToModule::Entry(size_t offset, uint64 address, uint64 length,
+                             uint8 version, const string& augmentation,
+                             unsigned return_address) {
+  if (DEBUG_DWARF) {
+    char buf[100];
+    SprintfLiteral(buf, "LUL.DW DwarfCFIToModule::Entry 0x%llx,+%lld\n",
+                   address, length);
+    summ_->Log(buf);
+  }
+
+  summ_->Entry(address, length);
+
+  // If dwarf2reader::CallFrameInfo can handle this version and
+  // augmentation, then we should be okay with that, so there's no
+  // need to check them here.
+
+  // Get ready to collect entries.
+  return_address_ = return_address;
+
+  // Breakpad STACK CFI records must provide a .ra rule, but DWARF CFI
+  // may not establish any rule for .ra if the return address column
+  // is an ordinary register, and that register holds the return
+  // address on entry to the function. So establish an initial .ra
+  // rule citing the return address register.
+  if (return_address_ < num_dw_regs_) {
+    summ_->Rule(address, return_address_, NODEREF, return_address, 0);
+  }
+
+  return true;
+}
+
+const UniqueString* DwarfCFIToModule::RegisterName(int i) {
+  if (i < 0) {
+    MOZ_ASSERT(i == kCFARegister);
+    return usu_->ToUniqueString(".cfa");
+  }
+  unsigned reg = i;
+  if (reg == return_address_) return usu_->ToUniqueString(".ra");
+
+  char buf[30];
+  SprintfLiteral(buf, "dwarf_reg_%u", reg);
+  return usu_->ToUniqueString(buf);
+}
+
+bool DwarfCFIToModule::UndefinedRule(uint64 address, int reg) {
+  reporter_->UndefinedNotSupported(entry_offset_, RegisterName(reg));
+  // Treat this as a non-fatal error.
+  return true;
+}
+
+bool DwarfCFIToModule::SameValueRule(uint64 address, int reg) {
+  if (DEBUG_DWARF) {
+    char buf[100];
+    SprintfLiteral(buf, "LUL.DW  0x%llx: old r%d = Same\n", address, reg);
+    summ_->Log(buf);
+  }
+  // reg + 0
+  summ_->Rule(address, reg, NODEREF, reg, 0);
+  return true;
+}
+
+bool DwarfCFIToModule::OffsetRule(uint64 address, int reg, int base_register,
+                                  long offset) {
+  if (DEBUG_DWARF) {
+    char buf[100];
+    SprintfLiteral(buf, "LUL.DW  0x%llx: old r%d = *(r%d + %ld)\n", address,
+                   reg, base_register, offset);
+    summ_->Log(buf);
+  }
+  // *(base_register + offset)
+  summ_->Rule(address, reg, DEREF, base_register, offset);
+  return true;
+}
+
+bool DwarfCFIToModule::ValOffsetRule(uint64 address, int reg, int base_register,
+                                     long offset) {
+  if (DEBUG_DWARF) {
+    char buf[100];
+    SprintfLiteral(buf, "LUL.DW  0x%llx: old r%d = r%d + %ld\n", address, reg,
+                   base_register, offset);
+    summ_->Log(buf);
+  }
+  // base_register + offset
+  summ_->Rule(address, reg, NODEREF, base_register, offset);
+  return true;
+}
+
+bool DwarfCFIToModule::RegisterRule(uint64 address, int reg,
+                                    int base_register) {
+  if (DEBUG_DWARF) {
+    char buf[100];
+    SprintfLiteral(buf, "LUL.DW  0x%llx: old r%d = r%d\n", address, reg,
+                   base_register);
+    summ_->Log(buf);
+  }
+  // base_register + 0
+  summ_->Rule(address, reg, NODEREF, base_register, 0);
+  return true;
+}
+
+bool DwarfCFIToModule::ExpressionRule(uint64 address, int reg,
+                                      const string& expression) {
+  bool debug = !!DEBUG_DWARF;
+  int32_t start_ix =
+      parseDwarfExpr(summ_, reader_, expression, debug, true /*pushCfaAtStart*/,
+                     true /*derefAtEnd*/);
+  if (start_ix >= 0) {
+    summ_->Rule(address, reg, PFXEXPR, 0, start_ix);
+  } else {
+    // Parsing of the Dwarf expression failed.  Treat this as a
+    // non-fatal error, hence return |true| even on this path.
+    reporter_->ExpressionCouldNotBeSummarised(entry_offset_, RegisterName(reg));
+  }
+  return true;
+}
+
+bool DwarfCFIToModule::ValExpressionRule(uint64 address, int reg,
+                                         const string& expression) {
+  bool debug = !!DEBUG_DWARF;
+  int32_t start_ix =
+      parseDwarfExpr(summ_, reader_, expression, debug, true /*pushCfaAtStart*/,
+                     false /*!derefAtEnd*/);
+  if (start_ix >= 0) {
+    summ_->Rule(address, reg, PFXEXPR, 0, start_ix);
+  } else {
+    // Parsing of the Dwarf expression failed.  Treat this as a
+    // non-fatal error, hence return |true| even on this path.
+    reporter_->ExpressionCouldNotBeSummarised(entry_offset_, RegisterName(reg));
+  }
+  return true;
+}
+
+bool DwarfCFIToModule::End() {
+  // module_->AddStackFrameEntry(entry_);
+  if (DEBUG_DWARF) {
+    summ_->Log("LUL.DW DwarfCFIToModule::End()\n");
+  }
+  summ_->End();
+  return true;
+}
+
+void DwarfCFIToModule::Reporter::UndefinedNotSupported(
+    size_t offset, const UniqueString* reg) {
+  char buf[300];
+  SprintfLiteral(buf, "DwarfCFIToModule::Reporter::UndefinedNotSupported()\n");
+  log_(buf);
+  // BPLOG(INFO) << file_ << ", section '" << section_
+  //  << "': the call frame entry at offset 0x"
+  //  << std::setbase(16) << offset << std::setbase(10)
+  //  << " sets the rule for register '" << FromUniqueString(reg)
+  //  << "' to 'undefined', but the Breakpad symbol file format cannot "
+  //  << " express this";
+}
+
+// FIXME: move this somewhere sensible
+static bool is_power_of_2(uint64_t n) {
+  int i, nSetBits = 0;
+  for (i = 0; i < 8 * (int)sizeof(n); i++) {
+    if ((n & ((uint64_t)1) << i) != 0) nSetBits++;
+  }
+  return nSetBits <= 1;
+}
+
+void DwarfCFIToModule::Reporter::ExpressionCouldNotBeSummarised(
+    size_t offset, const UniqueString* reg) {
+  static uint64_t n_complaints = 0;  // This isn't threadsafe
+  n_complaints++;
+  if (!is_power_of_2(n_complaints)) return;
+  char buf[300];
+  SprintfLiteral(buf,
+                 "DwarfCFIToModule::Reporter::"
+                 "ExpressionCouldNotBeSummarised(shown %llu times)\n",
+                 (unsigned long long int)n_complaints);
+  log_(buf);
+}
+
+}  // namespace lul
diff --git a/mozglue/baseprofiler/lul/LulDwarfExt.h b/mozglue/baseprofiler/lul/LulDwarfExt.h
new file mode 100644
index 0000000000..dcd2500e5a
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulDwarfExt.h
@@ -0,0 +1,1289 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+// Copyright 2006, 2010 Google Inc. All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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
+// OWNER OR CONTRIBUTORS 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.
+
+// Original author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
+
+// This file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/dwarf/types.h
+//   src/common/dwarf/dwarf2enums.h
+//   src/common/dwarf/bytereader.h
+//   src/common/dwarf_cfi_to_module.h
+//   src/common/dwarf/dwarf2reader.h
+
+#ifndef LulDwarfExt_h
+#define LulDwarfExt_h
+
+#include "LulDwarfSummariser.h"
+
+#include "mozilla/Assertions.h"
+
+#include <stdint.h>
+#include <string>
+
+typedef signed char int8;
+typedef short int16;
+typedef int int32;
+typedef long long int64;
+
+typedef unsigned char uint8;
+typedef unsigned short uint16;
+typedef unsigned int uint32;
+typedef unsigned long long uint64;
+
+#ifdef __PTRDIFF_TYPE__
+typedef __PTRDIFF_TYPE__ intptr;
+typedef unsigned __PTRDIFF_TYPE__ uintptr;
+#else
+#  error "Can't find pointer-sized integral types."
+#endif
+
+namespace lul {
+
+class UniqueString;
+
+// Exception handling frame description pointer formats, as described
+// by the Linux Standard Base Core Specification 4.0, section 11.5,
+// DWARF Extensions.
+enum DwarfPointerEncoding {
+  DW_EH_PE_absptr = 0x00,
+  DW_EH_PE_omit = 0xff,
+  DW_EH_PE_uleb128 = 0x01,
+  DW_EH_PE_udata2 = 0x02,
+  DW_EH_PE_udata4 = 0x03,
+  DW_EH_PE_udata8 = 0x04,
+  DW_EH_PE_sleb128 = 0x09,
+  DW_EH_PE_sdata2 = 0x0A,
+  DW_EH_PE_sdata4 = 0x0B,
+  DW_EH_PE_sdata8 = 0x0C,
+  DW_EH_PE_pcrel = 0x10,
+  DW_EH_PE_textrel = 0x20,
+  DW_EH_PE_datarel = 0x30,
+  DW_EH_PE_funcrel = 0x40,
+  DW_EH_PE_aligned = 0x50,
+
+  // The GNU toolchain sources define this enum value as well,
+  // simply to help classify the lower nybble values into signed and
+  // unsigned groups.
+  DW_EH_PE_signed = 0x08,
+
+  // This is not documented in LSB 4.0, but it is used in both the
+  // Linux and OS X toolchains. It can be added to any other
+  // encoding (except DW_EH_PE_aligned), and indicates that the
+  // encoded value represents the address at which the true address
+  // is stored, not the true address itself.
+  DW_EH_PE_indirect = 0x80
+};
+
+// We can't use the obvious name of LITTLE_ENDIAN and BIG_ENDIAN
+// because it conflicts with a macro
+enum Endianness { ENDIANNESS_BIG, ENDIANNESS_LITTLE };
+
+// A ByteReader knows how to read single- and multi-byte values of
+// various endiannesses, sizes, and encodings, as used in DWARF
+// debugging information and Linux C++ exception handling data.
+class ByteReader {
+ public:
+  // Construct a ByteReader capable of reading one-, two-, four-, and
+  // eight-byte values according to ENDIANNESS, absolute machine-sized
+  // addresses, DWARF-style "initial length" values, signed and
+  // unsigned LEB128 numbers, and Linux C++ exception handling data's
+  // encoded pointers.
+  explicit ByteReader(enum Endianness endianness);
+  virtual ~ByteReader();
+
+  // Read a single byte from BUFFER and return it as an unsigned 8 bit
+  // number.
+  uint8 ReadOneByte(const char* buffer) const;
+
+  // Read two bytes from BUFFER and return them as an unsigned 16 bit
+  // number, using this ByteReader's endianness.
+  uint16 ReadTwoBytes(const char* buffer) const;
+
+  // Read four bytes from BUFFER and return them as an unsigned 32 bit
+  // number, using this ByteReader's endianness. This function returns
+  // a uint64 so that it is compatible with ReadAddress and
+  // ReadOffset. The number it returns will never be outside the range
+  // of an unsigned 32 bit integer.
+  uint64 ReadFourBytes(const char* buffer) const;
+
+  // Read eight bytes from BUFFER and return them as an unsigned 64
+  // bit number, using this ByteReader's endianness.
+  uint64 ReadEightBytes(const char* buffer) const;
+
+  // Read an unsigned LEB128 (Little Endian Base 128) number from
+  // BUFFER and return it as an unsigned 64 bit integer. Set LEN to
+  // the number of bytes read.
+  //
+  // The unsigned LEB128 representation of an integer N is a variable
+  // number of bytes:
+  //
+  // - If N is between 0 and 0x7f, then its unsigned LEB128
+  //   representation is a single byte whose value is N.
+  //
+  // - Otherwise, its unsigned LEB128 representation is (N & 0x7f) |
+  //   0x80, followed by the unsigned LEB128 representation of N /
+  //   128, rounded towards negative infinity.
+  //
+  // In other words, we break VALUE into groups of seven bits, put
+  // them in little-endian order, and then write them as eight-bit
+  // bytes with the high bit on all but the last.
+  uint64 ReadUnsignedLEB128(const char* buffer, size_t* len) const;
+
+  // Read a signed LEB128 number from BUFFER and return it as an
+  // signed 64 bit integer. Set LEN to the number of bytes read.
+  //
+  // The signed LEB128 representation of an integer N is a variable
+  // number of bytes:
+  //
+  // - If N is between -0x40 and 0x3f, then its signed LEB128
+  //   representation is a single byte whose value is N in two's
+  //   complement.
+  //
+  // - Otherwise, its signed LEB128 representation is (N & 0x7f) |
+  //   0x80, followed by the signed LEB128 representation of N / 128,
+  //   rounded towards negative infinity.
+  //
+  // In other words, we break VALUE into groups of seven bits, put
+  // them in little-endian order, and then write them as eight-bit
+  // bytes with the high bit on all but the last.
+  int64 ReadSignedLEB128(const char* buffer, size_t* len) const;
+
+  // Indicate that addresses on this architecture are SIZE bytes long. SIZE
+  // must be either 4 or 8. (DWARF allows addresses to be any number of
+  // bytes in length from 1 to 255, but we only support 32- and 64-bit
+  // addresses at the moment.) You must call this before using the
+  // ReadAddress member function.
+  //
+  // For data in a .debug_info section, or something that .debug_info
+  // refers to like line number or macro data, the compilation unit
+  // header's address_size field indicates the address size to use. Call
+  // frame information doesn't indicate its address size (a shortcoming of
+  // the spec); you must supply the appropriate size based on the
+  // architecture of the target machine.
+  void SetAddressSize(uint8 size);
+
+  // Return the current address size, in bytes. This is either 4,
+  // indicating 32-bit addresses, or 8, indicating 64-bit addresses.
+  uint8 AddressSize() const { return address_size_; }
+
+  // Read an address from BUFFER and return it as an unsigned 64 bit
+  // integer, respecting this ByteReader's endianness and address size. You
+  // must call SetAddressSize before calling this function.
+  uint64 ReadAddress(const char* buffer) const;
+
+  // DWARF actually defines two slightly different formats: 32-bit DWARF
+  // and 64-bit DWARF. This is *not* related to the size of registers or
+  // addresses on the target machine; it refers only to the size of section
+  // offsets and data lengths appearing in the DWARF data. One only needs
+  // 64-bit DWARF when the debugging data itself is larger than 4GiB.
+  // 32-bit DWARF can handle x86_64 or PPC64 code just fine, unless the
+  // debugging data itself is very large.
+  //
+  // DWARF information identifies itself as 32-bit or 64-bit DWARF: each
+  // compilation unit and call frame information entry begins with an
+  // "initial length" field, which, in addition to giving the length of the
+  // data, also indicates the size of section offsets and lengths appearing
+  // in that data. The ReadInitialLength member function, below, reads an
+  // initial length and sets the ByteReader's offset size as a side effect.
+  // Thus, in the normal process of reading DWARF data, the appropriate
+  // offset size is set automatically. So, you should only need to call
+  // SetOffsetSize if you are using the same ByteReader to jump from the
+  // midst of one block of DWARF data into another.
+
+  // Read a DWARF "initial length" field from START, and return it as
+  // an unsigned 64 bit integer, respecting this ByteReader's
+  // endianness. Set *LEN to the length of the initial length in
+  // bytes, either four or twelve. As a side effect, set this
+  // ByteReader's offset size to either 4 (if we see a 32-bit DWARF
+  // initial length) or 8 (if we see a 64-bit DWARF initial length).
+  //
+  // A DWARF initial length is either:
+  //
+  // - a byte count stored as an unsigned 32-bit value less than
+  //   0xffffff00, indicating that the data whose length is being
+  //   measured uses the 32-bit DWARF format, or
+  //
+  // - The 32-bit value 0xffffffff, followed by a 64-bit byte count,
+  //   indicating that the data whose length is being measured uses
+  //   the 64-bit DWARF format.
+  uint64 ReadInitialLength(const char* start, size_t* len);
+
+  // Read an offset from BUFFER and return it as an unsigned 64 bit
+  // integer, respecting the ByteReader's endianness. In 32-bit DWARF, the
+  // offset is 4 bytes long; in 64-bit DWARF, the offset is eight bytes
+  // long. You must call ReadInitialLength or SetOffsetSize before calling
+  // this function; see the comments above for details.
+  uint64 ReadOffset(const char* buffer) const;
+
+  // Return the current offset size, in bytes.
+  // A return value of 4 indicates that we are reading 32-bit DWARF.
+  // A return value of 8 indicates that we are reading 64-bit DWARF.
+  uint8 OffsetSize() const { return offset_size_; }
+
+  // Indicate that section offsets and lengths are SIZE bytes long. SIZE
+  // must be either 4 (meaning 32-bit DWARF) or 8 (meaning 64-bit DWARF).
+  // Usually, you should not call this function yourself; instead, let a
+  // call to ReadInitialLength establish the data's offset size
+  // automatically.
+  void SetOffsetSize(uint8 size);
+
+  // The Linux C++ ABI uses a variant of DWARF call frame information
+  // for exception handling. This data is included in the program's
+  // address space as the ".eh_frame" section, and intepreted at
+  // runtime to walk the stack, find exception handlers, and run
+  // cleanup code. The format is mostly the same as DWARF CFI, with
+  // some adjustments made to provide the additional
+  // exception-handling data, and to make the data easier to work with
+  // in memory --- for example, to allow it to be placed in read-only
+  // memory even when describing position-independent code.
+  //
+  // In particular, exception handling data can select a number of
+  // different encodings for pointers that appear in the data, as
+  // described by the DwarfPointerEncoding enum. There are actually
+  // four axes(!) to the encoding:
+  //
+  // - The pointer size: pointers can be 2, 4, or 8 bytes long, or use
+  //   the DWARF LEB128 encoding.
+  //
+  // - The pointer's signedness: pointers can be signed or unsigned.
+  //
+  // - The pointer's base address: the data stored in the exception
+  //   handling data can be the actual address (that is, an absolute
+  //   pointer), or relative to one of a number of different base
+  //   addreses --- including that of the encoded pointer itself, for
+  //   a form of "pc-relative" addressing.
+  //
+  // - The pointer may be indirect: it may be the address where the
+  //   true pointer is stored. (This is used to refer to things via
+  //   global offset table entries, program linkage table entries, or
+  //   other tricks used in position-independent code.)
+  //
+  // There are also two options that fall outside that matrix
+  // altogether: the pointer may be omitted, or it may have padding to
+  // align it on an appropriate address boundary. (That last option
+  // may seem like it should be just another axis, but it is not.)
+
+  // Indicate that the exception handling data is loaded starting at
+  // SECTION_BASE, and that the start of its buffer in our own memory
+  // is BUFFER_BASE. This allows us to find the address that a given
+  // byte in our buffer would have when loaded into the program the
+  // data describes. We need this to resolve DW_EH_PE_pcrel pointers.
+  void SetCFIDataBase(uint64 section_base, const char* buffer_base);
+
+  // Indicate that the base address of the program's ".text" section
+  // is TEXT_BASE. We need this to resolve DW_EH_PE_textrel pointers.
+  void SetTextBase(uint64 text_base);
+
+  // Indicate that the base address for DW_EH_PE_datarel pointers is
+  // DATA_BASE. The proper value depends on the ABI; it is usually the
+  // address of the global offset table, held in a designated register in
+  // position-independent code. You will need to look at the startup code
+  // for the target system to be sure. I tried; my eyes bled.
+  void SetDataBase(uint64 data_base);
+
+  // Indicate that the base address for the FDE we are processing is
+  // FUNCTION_BASE. This is the start address of DW_EH_PE_funcrel
+  // pointers. (This encoding does not seem to be used by the GNU
+  // toolchain.)
+  void SetFunctionBase(uint64 function_base);
+
+  // Indicate that we are no longer processing any FDE, so any use of
+  // a DW_EH_PE_funcrel encoding is an error.
+  void ClearFunctionBase();
+
+  // Return true if ENCODING is a valid pointer encoding.
+  bool ValidEncoding(DwarfPointerEncoding encoding) const;
+
+  // Return true if we have all the information we need to read a
+  // pointer that uses ENCODING. This checks that the appropriate
+  // SetFooBase function for ENCODING has been called.
+  bool UsableEncoding(DwarfPointerEncoding encoding) const;
+
+  // Read an encoded pointer from BUFFER using ENCODING; return the
+  // absolute address it represents, and set *LEN to the pointer's
+  // length in bytes, including any padding for aligned pointers.
+  //
+  // This function calls 'abort' if ENCODING is invalid or refers to a
+  // base address this reader hasn't been given, so you should check
+  // with ValidEncoding and UsableEncoding first if you would rather
+  // die in a more helpful way.
+  uint64 ReadEncodedPointer(const char* buffer, DwarfPointerEncoding encoding,
+                            size_t* len) const;
+
+ private:
+  // Function pointer type for our address and offset readers.
+  typedef uint64 (ByteReader::*AddressReader)(const char*) const;
+
+  // Read an offset from BUFFER and return it as an unsigned 64 bit
+  // integer.  DWARF2/3 define offsets as either 4 or 8 bytes,
+  // generally depending on the amount of DWARF2/3 info present.
+  // This function pointer gets set by SetOffsetSize.
+  AddressReader offset_reader_;
+
+  // Read an address from BUFFER and return it as an unsigned 64 bit
+  // integer.  DWARF2/3 allow addresses to be any size from 0-255
+  // bytes currently.  Internally we support 4 and 8 byte addresses,
+  // and will CHECK on anything else.
+  // This function pointer gets set by SetAddressSize.
+  AddressReader address_reader_;
+
+  Endianness endian_;
+  uint8 address_size_;
+  uint8 offset_size_;
+
+  // Base addresses for Linux C++ exception handling data's encoded pointers.
+  bool have_section_base_, have_text_base_, have_data_base_;
+  bool have_function_base_;
+  uint64 section_base_;
+  uint64 text_base_, data_base_, function_base_;
+  const char* buffer_base_;
+};
+
+inline uint8 ByteReader::ReadOneByte(const char* buffer) const {
+  return buffer[0];
+}
+
+inline uint16 ByteReader::ReadTwoBytes(const char* signed_buffer) const {
+  const unsigned char* buffer =
+      reinterpret_cast<const unsigned char*>(signed_buffer);
+  const uint16 buffer0 = buffer[0];
+  const uint16 buffer1 = buffer[1];
+  if (endian_ == ENDIANNESS_LITTLE) {
+    return buffer0 | buffer1 << 8;
+  } else {
+    return buffer1 | buffer0 << 8;
+  }
+}
+
+inline uint64 ByteReader::ReadFourBytes(const char* signed_buffer) const {
+  const unsigned char* buffer =
+      reinterpret_cast<const unsigned char*>(signed_buffer);
+  const uint32 buffer0 = buffer[0];
+  const uint32 buffer1 = buffer[1];
+  const uint32 buffer2 = buffer[2];
+  const uint32 buffer3 = buffer[3];
+  if (endian_ == ENDIANNESS_LITTLE) {
+    return buffer0 | buffer1 << 8 | buffer2 << 16 | buffer3 << 24;
+  } else {
+    return buffer3 | buffer2 << 8 | buffer1 << 16 | buffer0 << 24;
+  }
+}
+
+inline uint64 ByteReader::ReadEightBytes(const char* signed_buffer) const {
+  const unsigned char* buffer =
+      reinterpret_cast<const unsigned char*>(signed_buffer);
+  const uint64 buffer0 = buffer[0];
+  const uint64 buffer1 = buffer[1];
+  const uint64 buffer2 = buffer[2];
+  const uint64 buffer3 = buffer[3];
+  const uint64 buffer4 = buffer[4];
+  const uint64 buffer5 = buffer[5];
+  const uint64 buffer6 = buffer[6];
+  const uint64 buffer7 = buffer[7];
+  if (endian_ == ENDIANNESS_LITTLE) {
+    return buffer0 | buffer1 << 8 | buffer2 << 16 | buffer3 << 24 |
+           buffer4 << 32 | buffer5 << 40 | buffer6 << 48 | buffer7 << 56;
+  } else {
+    return buffer7 | buffer6 << 8 | buffer5 << 16 | buffer4 << 24 |
+           buffer3 << 32 | buffer2 << 40 | buffer1 << 48 | buffer0 << 56;
+  }
+}
+
+// Read an unsigned LEB128 number.  Each byte contains 7 bits of
+// information, plus one bit saying whether the number continues or
+// not.
+
+inline uint64 ByteReader::ReadUnsignedLEB128(const char* buffer,
+                                             size_t* len) const {
+  uint64 result = 0;
+  size_t num_read = 0;
+  unsigned int shift = 0;
+  unsigned char byte;
+
+  do {
+    byte = *buffer++;
+    num_read++;
+
+    result |= (static_cast<uint64>(byte & 0x7f)) << shift;
+
+    shift += 7;
+
+  } while (byte & 0x80);
+
+  *len = num_read;
+
+  return result;
+}
+
+// Read a signed LEB128 number.  These are like regular LEB128
+// numbers, except the last byte may have a sign bit set.
+
+inline int64 ByteReader::ReadSignedLEB128(const char* buffer,
+                                          size_t* len) const {
+  int64 result = 0;
+  unsigned int shift = 0;
+  size_t num_read = 0;
+  unsigned char byte;
+
+  do {
+    byte = *buffer++;
+    num_read++;
+    result |= (static_cast<uint64>(byte & 0x7f) << shift);
+    shift += 7;
+  } while (byte & 0x80);
+
+  if ((shift < 8 * sizeof(result)) && (byte & 0x40))
+    result |= -((static_cast<int64>(1)) << shift);
+  *len = num_read;
+  return result;
+}
+
+inline uint64 ByteReader::ReadOffset(const char* buffer) const {
+  MOZ_ASSERT(this->offset_reader_);
+  return (this->*offset_reader_)(buffer);
+}
+
+inline uint64 ByteReader::ReadAddress(const char* buffer) const {
+  MOZ_ASSERT(this->address_reader_);
+  return (this->*address_reader_)(buffer);
+}
+
+inline void ByteReader::SetCFIDataBase(uint64 section_base,
+                                       const char* buffer_base) {
+  section_base_ = section_base;
+  buffer_base_ = buffer_base;
+  have_section_base_ = true;
+}
+
+inline void ByteReader::SetTextBase(uint64 text_base) {
+  text_base_ = text_base;
+  have_text_base_ = true;
+}
+
+inline void ByteReader::SetDataBase(uint64 data_base) {
+  data_base_ = data_base;
+  have_data_base_ = true;
+}
+
+inline void ByteReader::SetFunctionBase(uint64 function_base) {
+  function_base_ = function_base;
+  have_function_base_ = true;
+}
+
+inline void ByteReader::ClearFunctionBase() { have_function_base_ = false; }
+
+// (derived from)
+// dwarf_cfi_to_module.h: Define the DwarfCFIToModule class, which
+// accepts parsed DWARF call frame info and adds it to a Summariser object.
+
+// This class is a reader for DWARF's Call Frame Information.  CFI
+// describes how to unwind stack frames --- even for functions that do
+// not follow fixed conventions for saving registers, whose frame size
+// varies as they execute, etc.
+//
+// CFI describes, at each machine instruction, how to compute the
+// stack frame's base address, how to find the return address, and
+// where to find the saved values of the caller's registers (if the
+// callee has stashed them somewhere to free up the registers for its
+// own use).
+//
+// For example, suppose we have a function whose machine code looks
+// like this (imagine an assembly language that looks like C, for a
+// machine with 32-bit registers, and a stack that grows towards lower
+// addresses):
+//
+// func:                                ; entry point; return address at sp
+// func+0:      sp = sp - 16            ; allocate space for stack frame
+// func+1:      sp[12] = r0             ; save r0 at sp+12
+// ...                                  ; other code, not frame-related
+// func+10:     sp -= 4; *sp = x        ; push some x on the stack
+// ...                                  ; other code, not frame-related
+// func+20:     r0 = sp[16]             ; restore saved r0
+// func+21:     sp += 20                ; pop whole stack frame
+// func+22:     pc = *sp; sp += 4       ; pop return address and jump to it
+//
+// DWARF CFI is (a very compressed representation of) a table with a
+// row for each machine instruction address and a column for each
+// register showing how to restore it, if possible.
+//
+// A special column named "CFA", for "Canonical Frame Address", tells how
+// to compute the base address of the frame; registers' entries may
+// refer to the CFA in describing where the registers are saved.
+//
+// Another special column, named "RA", represents the return address.
+//
+// For example, here is a complete (uncompressed) table describing the
+// function above:
+//
+//     insn      cfa    r0      r1 ...  ra
+//     =======================================
+//     func+0:   sp                     cfa[0]
+//     func+1:   sp+16                  cfa[0]
+//     func+2:   sp+16  cfa[-4]         cfa[0]
+//     func+11:  sp+20  cfa[-4]         cfa[0]
+//     func+21:  sp+20                  cfa[0]
+//     func+22:  sp                     cfa[0]
+//
+// Some things to note here:
+//
+// - Each row describes the state of affairs *before* executing the
+//   instruction at the given address.  Thus, the row for func+0
+//   describes the state before we allocate the stack frame.  In the
+//   next row, the formula for computing the CFA has changed,
+//   reflecting that allocation.
+//
+// - The other entries are written in terms of the CFA; this allows
+//   them to remain unchanged as the stack pointer gets bumped around.
+//   For example, the rule for recovering the return address (the "ra"
+//   column) remains unchanged throughout the function, even as the
+//   stack pointer takes on three different offsets from the return
+//   address.
+//
+// - Although we haven't shown it, most calling conventions designate
+//   "callee-saves" and "caller-saves" registers. The callee must
+//   preserve the values of callee-saves registers; if it uses them,
+//   it must save their original values somewhere, and restore them
+//   before it returns. In contrast, the callee is free to trash
+//   caller-saves registers; if the callee uses these, it will
+//   probably not bother to save them anywhere, and the CFI will
+//   probably mark their values as "unrecoverable".
+//
+//   (However, since the caller cannot assume the callee was going to
+//   save them, caller-saves registers are probably dead in the caller
+//   anyway, so compilers usually don't generate CFA for caller-saves
+//   registers.)
+//
+// - Exactly where the CFA points is a matter of convention that
+//   depends on the architecture and ABI in use. In the example, the
+//   CFA is the value the stack pointer had upon entry to the
+//   function, pointing at the saved return address. But on the x86,
+//   the call frame information generated by GCC follows the
+//   convention that the CFA is the address *after* the saved return
+//   address.
+//
+//   But by definition, the CFA remains constant throughout the
+//   lifetime of the frame. This makes it a useful value for other
+//   columns to refer to. It is also gives debuggers a useful handle
+//   for identifying a frame.
+//
+// If you look at the table above, you'll notice that a given entry is
+// often the same as the one immediately above it: most instructions
+// change only one or two aspects of the stack frame, if they affect
+// it at all. The DWARF format takes advantage of this fact, and
+// reduces the size of the data by mentioning only the addresses and
+// columns at which changes take place. So for the above, DWARF CFI
+// data would only actually mention the following:
+//
+//     insn      cfa    r0      r1 ...  ra
+//     =======================================
+//     func+0:   sp                     cfa[0]
+//     func+1:   sp+16
+//     func+2:          cfa[-4]
+//     func+11:  sp+20
+//     func+21:         r0
+//     func+22:  sp
+//
+// In fact, this is the way the parser reports CFI to the consumer: as
+// a series of statements of the form, "At address X, column Y changed
+// to Z," and related conventions for describing the initial state.
+//
+// Naturally, it would be impractical to have to scan the entire
+// program's CFI, noting changes as we go, just to recover the
+// unwinding rules in effect at one particular instruction. To avoid
+// this, CFI data is grouped into "entries", each of which covers a
+// specified range of addresses and begins with a complete statement
+// of the rules for all recoverable registers at that starting
+// address. Each entry typically covers a single function.
+//
+// Thus, to compute the contents of a given row of the table --- that
+// is, rules for recovering the CFA, RA, and registers at a given
+// instruction --- the consumer should find the entry that covers that
+// instruction's address, start with the initial state supplied at the
+// beginning of the entry, and work forward until it has processed all
+// the changes up to and including those for the present instruction.
+//
+// There are seven kinds of rules that can appear in an entry of the
+// table:
+//
+// - "undefined": The given register is not preserved by the callee;
+//   its value cannot be recovered.
+//
+// - "same value": This register has the same value it did in the callee.
+//
+// - offset(N): The register is saved at offset N from the CFA.
+//
+// - val_offset(N): The value the register had in the caller is the
+//   CFA plus offset N. (This is usually only useful for describing
+//   the stack pointer.)
+//
+// - register(R): The register's value was saved in another register R.
+//
+// - expression(E): Evaluating the DWARF expression E using the
+//   current frame's registers' values yields the address at which the
+//   register was saved.
+//
+// - val_expression(E): Evaluating the DWARF expression E using the
+//   current frame's registers' values yields the value the register
+//   had in the caller.
+
+class CallFrameInfo {
+ public:
+  // The different kinds of entries one finds in CFI. Used internally,
+  // and for error reporting.
+  enum EntryKind { kUnknown, kCIE, kFDE, kTerminator };
+
+  // The handler class to which the parser hands the parsed call frame
+  // information.  Defined below.
+  class Handler;
+
+  // A reporter class, which CallFrameInfo uses to report errors
+  // encountered while parsing call frame information.  Defined below.
+  class Reporter;
+
+  // Create a DWARF CFI parser. BUFFER points to the contents of the
+  // .debug_frame section to parse; BUFFER_LENGTH is its length in bytes.
+  // REPORTER is an error reporter the parser should use to report
+  // problems. READER is a ByteReader instance that has the endianness and
+  // address size set properly. Report the data we find to HANDLER.
+  //
+  // This class can also parse Linux C++ exception handling data, as found
+  // in '.eh_frame' sections. This data is a variant of DWARF CFI that is
+  // placed in loadable segments so that it is present in the program's
+  // address space, and is interpreted by the C++ runtime to search the
+  // call stack for a handler interested in the exception being thrown,
+  // actually pop the frames, and find cleanup code to run.
+  //
+  // There are two differences between the call frame information described
+  // in the DWARF standard and the exception handling data Linux places in
+  // the .eh_frame section:
+  //
+  // - Exception handling data uses uses a different format for call frame
+  //   information entry headers. The distinguished CIE id, the way FDEs
+  //   refer to their CIEs, and the way the end of the series of entries is
+  //   determined are all slightly different.
+  //
+  //   If the constructor's EH_FRAME argument is true, then the
+  //   CallFrameInfo parses the entry headers as Linux C++ exception
+  //   handling data. If EH_FRAME is false or omitted, the CallFrameInfo
+  //   parses standard DWARF call frame information.
+  //
+  // - Linux C++ exception handling data uses CIE augmentation strings
+  //   beginning with 'z' to specify the presence of additional data after
+  //   the CIE and FDE headers and special encodings used for addresses in
+  //   frame description entries.
+  //
+  //   CallFrameInfo can handle 'z' augmentations in either DWARF CFI or
+  //   exception handling data if you have supplied READER with the base
+  //   addresses needed to interpret the pointer encodings that 'z'
+  //   augmentations can specify. See the ByteReader interface for details
+  //   about the base addresses. See the CallFrameInfo::Handler interface
+  //   for details about the additional information one might find in
+  //   'z'-augmented data.
+  //
+  // Thus:
+  //
+  // - If you are parsing standard DWARF CFI, as found in a .debug_frame
+  //   section, you should pass false for the EH_FRAME argument, or omit
+  //   it, and you need not worry about providing READER with the
+  //   additional base addresses.
+  //
+  // - If you want to parse Linux C++ exception handling data from a
+  //   .eh_frame section, you should pass EH_FRAME as true, and call
+  //   READER's Set*Base member functions before calling our Start method.
+  //
+  // - If you want to parse DWARF CFI that uses the 'z' augmentations
+  //   (although I don't think any toolchain ever emits such data), you
+  //   could pass false for EH_FRAME, but call READER's Set*Base members.
+  //
+  // The extensions the Linux C++ ABI makes to DWARF for exception
+  // handling are described here, rather poorly:
+  // http://refspecs.linux-foundation.org/LSB_4.0.0/LSB-Core-generic/LSB-Core-generic/dwarfext.html
+  // http://refspecs.linux-foundation.org/LSB_4.0.0/LSB-Core-generic/LSB-Core-generic/ehframechpt.html
+  //
+  // The mechanics of C++ exception handling, personality routines,
+  // and language-specific data areas are described here, rather nicely:
+  // http://www.codesourcery.com/public/cxx-abi/abi-eh.html
+
+  CallFrameInfo(const char* buffer, size_t buffer_length, ByteReader* reader,
+                Handler* handler, Reporter* reporter, bool eh_frame = false)
+      : buffer_(buffer),
+        buffer_length_(buffer_length),
+        reader_(reader),
+        handler_(handler),
+        reporter_(reporter),
+        eh_frame_(eh_frame) {}
+
+  ~CallFrameInfo() {}
+
+  // Parse the entries in BUFFER, reporting what we find to HANDLER.
+  // Return true if we reach the end of the section successfully, or
+  // false if we encounter an error.
+  bool Start();
+
+  // Return the textual name of KIND. For error reporting.
+  static const char* KindName(EntryKind kind);
+
+ private:
+  struct CIE;
+
+  // A CFI entry, either an FDE or a CIE.
+  struct Entry {
+    // The starting offset of the entry in the section, for error
+    // reporting.
+    size_t offset;
+
+    // The start of this entry in the buffer.
+    const char* start;
+
+    // Which kind of entry this is.
+    //
+    // We want to be able to use this for error reporting even while we're
+    // in the midst of parsing. Error reporting code may assume that kind,
+    // offset, and start fields are valid, although kind may be kUnknown.
+    EntryKind kind;
+
+    // The end of this entry's common prologue (initial length and id), and
+    // the start of this entry's kind-specific fields.
+    const char* fields;
+
+    // The start of this entry's instructions.
+    const char* instructions;
+
+    // The address past the entry's last byte in the buffer. (Note that
+    // since offset points to the entry's initial length field, and the
+    // length field is the number of bytes after that field, this is not
+    // simply buffer_ + offset + length.)
+    const char* end;
+
+    // For both DWARF CFI and .eh_frame sections, this is the CIE id in a
+    // CIE, and the offset of the associated CIE in an FDE.
+    uint64 id;
+
+    // The CIE that applies to this entry, if we've parsed it. If this is a
+    // CIE, then this field points to this structure.
+    CIE* cie;
+  };
+
+  // A common information entry (CIE).
+  struct CIE : public Entry {
+    uint8 version;                     // CFI data version number
+    std::string augmentation;          // vendor format extension markers
+    uint64 code_alignment_factor;      // scale for code address adjustments
+    int data_alignment_factor;         // scale for stack pointer adjustments
+    unsigned return_address_register;  // which register holds the return addr
+
+    // True if this CIE includes Linux C++ ABI 'z' augmentation data.
+    bool has_z_augmentation;
+
+    // Parsed 'z' augmentation data. These are meaningful only if
+    // has_z_augmentation is true.
+    bool has_z_lsda;          // The 'z' augmentation included 'L'.
+    bool has_z_personality;   // The 'z' augmentation included 'P'.
+    bool has_z_signal_frame;  // The 'z' augmentation included 'S'.
+
+    // If has_z_lsda is true, this is the encoding to be used for language-
+    // specific data area pointers in FDEs.
+    DwarfPointerEncoding lsda_encoding;
+
+    // If has_z_personality is true, this is the encoding used for the
+    // personality routine pointer in the augmentation data.
+    DwarfPointerEncoding personality_encoding;
+
+    // If has_z_personality is true, this is the address of the personality
+    // routine --- or, if personality_encoding & DW_EH_PE_indirect, the
+    // address where the personality routine's address is stored.
+    uint64 personality_address;
+
+    // This is the encoding used for addresses in the FDE header and
+    // in DW_CFA_set_loc instructions. This is always valid, whether
+    // or not we saw a 'z' augmentation string; its default value is
+    // DW_EH_PE_absptr, which is what normal DWARF CFI uses.
+    DwarfPointerEncoding pointer_encoding;
+  };
+
+  // A frame description entry (FDE).
+  struct FDE : public Entry {
+    uint64 address;  // start address of described code
+    uint64 size;     // size of described code, in bytes
+
+    // If cie->has_z_lsda is true, then this is the language-specific data
+    // area's address --- or its address's address, if cie->lsda_encoding
+    // has the DW_EH_PE_indirect bit set.
+    uint64 lsda_address;
+  };
+
+  // Internal use.
+  class Rule;
+  class UndefinedRule;
+  class SameValueRule;
+  class OffsetRule;
+  class ValOffsetRule;
+  class RegisterRule;
+  class ExpressionRule;
+  class ValExpressionRule;
+  class RuleMap;
+  class State;
+
+  // Parse the initial length and id of a CFI entry, either a CIE, an FDE,
+  // or a .eh_frame end-of-data mark. CURSOR points to the beginning of the
+  // data to parse. On success, populate ENTRY as appropriate, and return
+  // true. On failure, report the problem, and return false. Even if we
+  // return false, set ENTRY->end to the first byte after the entry if we
+  // were able to figure that out, or NULL if we weren't.
+  bool ReadEntryPrologue(const char* cursor, Entry* entry);
+
+  // Parse the fields of a CIE after the entry prologue, including any 'z'
+  // augmentation data. Assume that the 'Entry' fields of CIE are
+  // populated; use CIE->fields and CIE->end as the start and limit for
+  // parsing. On success, populate the rest of *CIE, and return true; on
+  // failure, report the problem and return false.
+  bool ReadCIEFields(CIE* cie);
+
+  // Parse the fields of an FDE after the entry prologue, including any 'z'
+  // augmentation data. Assume that the 'Entry' fields of *FDE are
+  // initialized; use FDE->fields and FDE->end as the start and limit for
+  // parsing. Assume that FDE->cie is fully initialized. On success,
+  // populate the rest of *FDE, and return true; on failure, report the
+  // problem and return false.
+  bool ReadFDEFields(FDE* fde);
+
+  // Report that ENTRY is incomplete, and return false. This is just a
+  // trivial wrapper for invoking reporter_->Incomplete; it provides a
+  // little brevity.
+  bool ReportIncomplete(Entry* entry);
+
+  // Return true if ENCODING has the DW_EH_PE_indirect bit set.
+  static bool IsIndirectEncoding(DwarfPointerEncoding encoding) {
+    return encoding & DW_EH_PE_indirect;
+  }
+
+  // The contents of the DWARF .debug_info section we're parsing.
+  const char* buffer_;
+  size_t buffer_length_;
+
+  // For reading multi-byte values with the appropriate endianness.
+  ByteReader* reader_;
+
+  // The handler to which we should report the data we find.
+  Handler* handler_;
+
+  // For reporting problems in the info we're parsing.
+  Reporter* reporter_;
+
+  // True if we are processing .eh_frame-format data.
+  bool eh_frame_;
+};
+
+// The handler class for CallFrameInfo.  The a CFI parser calls the
+// member functions of a handler object to report the data it finds.
+class CallFrameInfo::Handler {
+ public:
+  // The pseudo-register number for the canonical frame address.
+  enum { kCFARegister = DW_REG_CFA };
+
+  Handler() {}
+  virtual ~Handler() {}
+
+  // The parser has found CFI for the machine code at ADDRESS,
+  // extending for LENGTH bytes. OFFSET is the offset of the frame
+  // description entry in the section, for use in error messages.
+  // VERSION is the version number of the CFI format. AUGMENTATION is
+  // a string describing any producer-specific extensions present in
+  // the data. RETURN_ADDRESS is the number of the register that holds
+  // the address to which the function should return.
+  //
+  // Entry should return true to process this CFI, or false to skip to
+  // the next entry.
+  //
+  // The parser invokes Entry for each Frame Description Entry (FDE)
+  // it finds.  The parser doesn't report Common Information Entries
+  // to the handler explicitly; instead, if the handler elects to
+  // process a given FDE, the parser reiterates the appropriate CIE's
+  // contents at the beginning of the FDE's rules.
+  virtual bool Entry(size_t offset, uint64 address, uint64 length,
+                     uint8 version, const std::string& augmentation,
+                     unsigned return_address) = 0;
+
+  // When the Entry function returns true, the parser calls these
+  // handler functions repeatedly to describe the rules for recovering
+  // registers at each instruction in the given range of machine code.
+  // Immediately after a call to Entry, the handler should assume that
+  // the rule for each callee-saves register is "unchanged" --- that
+  // is, that the register still has the value it had in the caller.
+  //
+  // If a *Rule function returns true, we continue processing this entry's
+  // instructions. If a *Rule function returns false, we stop evaluating
+  // instructions, and skip to the next entry. Either way, we call End
+  // before going on to the next entry.
+  //
+  // In all of these functions, if the REG parameter is kCFARegister, then
+  // the rule describes how to find the canonical frame address.
+  // kCFARegister may be passed as a BASE_REGISTER argument, meaning that
+  // the canonical frame address should be used as the base address for the
+  // computation. All other REG values will be positive.
+
+  // At ADDRESS, register REG's value is not recoverable.
+  virtual bool UndefinedRule(uint64 address, int reg) = 0;
+
+  // At ADDRESS, register REG's value is the same as that it had in
+  // the caller.
+  virtual bool SameValueRule(uint64 address, int reg) = 0;
+
+  // At ADDRESS, register REG has been saved at offset OFFSET from
+  // BASE_REGISTER.
+  virtual bool OffsetRule(uint64 address, int reg, int base_register,
+                          long offset) = 0;
+
+  // At ADDRESS, the caller's value of register REG is the current
+  // value of BASE_REGISTER plus OFFSET. (This rule doesn't provide an
+  // address at which the register's value is saved.)
+  virtual bool ValOffsetRule(uint64 address, int reg, int base_register,
+                             long offset) = 0;
+
+  // At ADDRESS, register REG has been saved in BASE_REGISTER. This differs
+  // from ValOffsetRule(ADDRESS, REG, BASE_REGISTER, 0), in that
+  // BASE_REGISTER is the "home" for REG's saved value: if you want to
+  // assign to a variable whose home is REG in the calling frame, you
+  // should put the value in BASE_REGISTER.
+  virtual bool RegisterRule(uint64 address, int reg, int base_register) = 0;
+
+  // At ADDRESS, the DWARF expression EXPRESSION yields the address at
+  // which REG was saved.
+  virtual bool ExpressionRule(uint64 address, int reg,
+                              const std::string& expression) = 0;
+
+  // At ADDRESS, the DWARF expression EXPRESSION yields the caller's
+  // value for REG. (This rule doesn't provide an address at which the
+  // register's value is saved.)
+  virtual bool ValExpressionRule(uint64 address, int reg,
+                                 const std::string& expression) = 0;
+
+  // Indicate that the rules for the address range reported by the
+  // last call to Entry are complete.  End should return true if
+  // everything is okay, or false if an error has occurred and parsing
+  // should stop.
+  virtual bool End() = 0;
+
+  // Handler functions for Linux C++ exception handling data. These are
+  // only called if the data includes 'z' augmentation strings.
+
+  // The Linux C++ ABI uses an extension of the DWARF CFI format to
+  // walk the stack to propagate exceptions from the throw to the
+  // appropriate catch, and do the appropriate cleanups along the way.
+  // CFI entries used for exception handling have two additional data
+  // associated with them:
+  //
+  // - The "language-specific data area" describes which exception
+  //   types the function has 'catch' clauses for, and indicates how
+  //   to go about re-entering the function at the appropriate catch
+  //   clause. If the exception is not caught, it describes the
+  //   destructors that must run before the frame is popped.
+  //
+  // - The "personality routine" is responsible for interpreting the
+  //   language-specific data area's contents, and deciding whether
+  //   the exception should continue to propagate down the stack,
+  //   perhaps after doing some cleanup for this frame, or whether the
+  //   exception will be caught here.
+  //
+  // In principle, the language-specific data area is opaque to
+  // everybody but the personality routine. In practice, these values
+  // may be useful or interesting to readers with extra context, and
+  // we have to at least skip them anyway, so we might as well report
+  // them to the handler.
+
+  // This entry's exception handling personality routine's address is
+  // ADDRESS. If INDIRECT is true, then ADDRESS is the address at
+  // which the routine's address is stored. The default definition for
+  // this handler function simply returns true, allowing parsing of
+  // the entry to continue.
+  virtual bool PersonalityRoutine(uint64 address, bool indirect) {
+    return true;
+  }
+
+  // This entry's language-specific data area (LSDA) is located at
+  // ADDRESS. If INDIRECT is true, then ADDRESS is the address at
+  // which the area's address is stored. The default definition for
+  // this handler function simply returns true, allowing parsing of
+  // the entry to continue.
+  virtual bool LanguageSpecificDataArea(uint64 address, bool indirect) {
+    return true;
+  }
+
+  // This entry describes a signal trampoline --- this frame is the
+  // caller of a signal handler. The default definition for this
+  // handler function simply returns true, allowing parsing of the
+  // entry to continue.
+  //
+  // The best description of the rationale for and meaning of signal
+  // trampoline CFI entries seems to be in the GCC bug database:
+  // http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26208
+  virtual bool SignalHandler() { return true; }
+};
+
+// The CallFrameInfo class makes calls on an instance of this class to
+// report errors or warn about problems in the data it is parsing.
+// These messages are sent to the message sink |aLog| provided to the
+// constructor.
+class CallFrameInfo::Reporter {
+ public:
+  // Create an error reporter which attributes troubles to the section
+  // named SECTION in FILENAME.
+  //
+  // Normally SECTION would be .debug_frame, but the Mac puts CFI data
+  // in a Mach-O section named __debug_frame. If we support
+  // Linux-style exception handling data, we could be reading an
+  // .eh_frame section.
+  Reporter(void (*aLog)(const char*), const std::string& filename,
+           const std::string& section = ".debug_frame")
+      : log_(aLog), filename_(filename), section_(section) {}
+  virtual ~Reporter() {}
+
+  // The CFI entry at OFFSET ends too early to be well-formed. KIND
+  // indicates what kind of entry it is; KIND can be kUnknown if we
+  // haven't parsed enough of the entry to tell yet.
+  virtual void Incomplete(uint64 offset, CallFrameInfo::EntryKind kind);
+
+  // The .eh_frame data has a four-byte zero at OFFSET where the next
+  // entry's length would be; this is a terminator. However, the buffer
+  // length as given to the CallFrameInfo constructor says there should be
+  // more data.
+  virtual void EarlyEHTerminator(uint64 offset);
+
+  // The FDE at OFFSET refers to the CIE at CIE_OFFSET, but the
+  // section is not that large.
+  virtual void CIEPointerOutOfRange(uint64 offset, uint64 cie_offset);
+
+  // The FDE at OFFSET refers to the CIE at CIE_OFFSET, but the entry
+  // there is not a CIE.
+  virtual void BadCIEId(uint64 offset, uint64 cie_offset);
+
+  // The FDE at OFFSET refers to a CIE with version number VERSION,
+  // which we don't recognize. We cannot parse DWARF CFI if it uses
+  // a version number we don't recognize.
+  virtual void UnrecognizedVersion(uint64 offset, int version);
+
+  // The FDE at OFFSET refers to a CIE with augmentation AUGMENTATION,
+  // which we don't recognize. We cannot parse DWARF CFI if it uses
+  // augmentations we don't recognize.
+  virtual void UnrecognizedAugmentation(uint64 offset,
+                                        const std::string& augmentation);
+
+  // The FDE at OFFSET contains an invalid or otherwise unusable Dwarf4
+  // specific field (currently, only "address_size" or "segment_size").
+  // Parsing DWARF CFI with unexpected values here seems dubious at best,
+  // so we stop.  WHAT gives a little more information about what is wrong.
+  virtual void InvalidDwarf4Artefact(uint64 offset, const char* what);
+
+  // The pointer encoding ENCODING, specified by the CIE at OFFSET, is not
+  // a valid encoding.
+  virtual void InvalidPointerEncoding(uint64 offset, uint8 encoding);
+
+  // The pointer encoding ENCODING, specified by the CIE at OFFSET, depends
+  // on a base address which has not been supplied.
+  virtual void UnusablePointerEncoding(uint64 offset, uint8 encoding);
+
+  // The CIE at OFFSET contains a DW_CFA_restore instruction at
+  // INSN_OFFSET, which may not appear in a CIE.
+  virtual void RestoreInCIE(uint64 offset, uint64 insn_offset);
+
+  // The entry at OFFSET, of kind KIND, has an unrecognized
+  // instruction at INSN_OFFSET.
+  virtual void BadInstruction(uint64 offset, CallFrameInfo::EntryKind kind,
+                              uint64 insn_offset);
+
+  // The instruction at INSN_OFFSET in the entry at OFFSET, of kind
+  // KIND, establishes a rule that cites the CFA, but we have not
+  // established a CFA rule yet.
+  virtual void NoCFARule(uint64 offset, CallFrameInfo::EntryKind kind,
+                         uint64 insn_offset);
+
+  // The instruction at INSN_OFFSET in the entry at OFFSET, of kind
+  // KIND, is a DW_CFA_restore_state instruction, but the stack of
+  // saved states is empty.
+  virtual void EmptyStateStack(uint64 offset, CallFrameInfo::EntryKind kind,
+                               uint64 insn_offset);
+
+  // The DW_CFA_remember_state instruction at INSN_OFFSET in the entry
+  // at OFFSET, of kind KIND, would restore a state that has no CFA
+  // rule, whereas the current state does have a CFA rule. This is
+  // bogus input, which the CallFrameInfo::Handler interface doesn't
+  // (and shouldn't) have any way to report.
+  virtual void ClearingCFARule(uint64 offset, CallFrameInfo::EntryKind kind,
+                               uint64 insn_offset);
+
+ private:
+  // A logging sink function, as supplied by LUL's user.
+  void (*log_)(const char*);
+
+ protected:
+  // The name of the file whose CFI we're reading.
+  std::string filename_;
+
+  // The name of the CFI section in that file.
+  std::string section_;
+};
+
+using lul::CallFrameInfo;
+using lul::Summariser;
+
+// A class that accepts parsed call frame information from the DWARF
+// CFI parser and populates a google_breakpad::Module object with the
+// contents.
+class DwarfCFIToModule : public CallFrameInfo::Handler {
+ public:
+  // DwarfCFIToModule uses an instance of this class to report errors
+  // detected while converting DWARF CFI to Breakpad STACK CFI records.
+  class Reporter {
+   public:
+    // Create a reporter that writes messages to the message sink
+    // |aLog|. FILE is the name of the file we're processing, and
+    // SECTION is the name of the section within that file that we're
+    // looking at (.debug_frame, .eh_frame, etc.).
+    Reporter(void (*aLog)(const char*), const std::string& file,
+             const std::string& section)
+        : log_(aLog), file_(file), section_(section) {}
+    virtual ~Reporter() {}
+
+    // The DWARF CFI entry at OFFSET says that REG is undefined, but the
+    // Breakpad symbol file format cannot express this.
+    virtual void UndefinedNotSupported(size_t offset, const UniqueString* reg);
+
+    // The DWARF CFI entry at OFFSET says that REG uses a DWARF
+    // expression to find its value, but parseDwarfExpr could not
+    // convert it to a sequence of PfxInstrs.
+    virtual void ExpressionCouldNotBeSummarised(size_t offset,
+                                                const UniqueString* reg);
+
+   private:
+    // A logging sink function, as supplied by LUL's user.
+    void (*log_)(const char*);
+
+   protected:
+    std::string file_, section_;
+  };
+
+  // Register name tables. If TABLE is a vector returned by one of these
+  // functions, then TABLE[R] is the name of the register numbered R in
+  // DWARF call frame information.
+  class RegisterNames {
+   public:
+    // Intel's "x86" or IA-32.
+    static unsigned int I386();
+
+    // AMD x86_64, AMD64, Intel EM64T, or Intel 64
+    static unsigned int X86_64();
+
+    // ARM.
+    static unsigned int ARM();
+
+    // AARCH64.
+    static unsigned int ARM64();
+
+    // MIPS.
+    static unsigned int MIPS();
+  };
+
+  // Create a handler for the dwarf2reader::CallFrameInfo parser that
+  // records the stack unwinding information it receives in SUMM.
+  //
+  // Use REGISTER_NAMES[I] as the name of register number I; *this
+  // keeps a reference to the vector, so the vector should remain
+  // alive for as long as the DwarfCFIToModule does.
+  //
+  // Use REPORTER for reporting problems encountered in the conversion
+  // process.
+  DwarfCFIToModule(const unsigned int num_dw_regs, Reporter* reporter,
+                   ByteReader* reader,
+                   /*MOD*/ UniqueStringUniverse* usu,
+                   /*OUT*/ Summariser* summ)
+      : summ_(summ),
+        usu_(usu),
+        num_dw_regs_(num_dw_regs),
+        reporter_(reporter),
+        reader_(reader),
+        return_address_(-1) {}
+  virtual ~DwarfCFIToModule() {}
+
+  virtual bool Entry(size_t offset, uint64 address, uint64 length,
+                     uint8 version, const std::string& augmentation,
+                     unsigned return_address) override;
+  virtual bool UndefinedRule(uint64 address, int reg) override;
+  virtual bool SameValueRule(uint64 address, int reg) override;
+  virtual bool OffsetRule(uint64 address, int reg, int base_register,
+                          long offset) override;
+  virtual bool ValOffsetRule(uint64 address, int reg, int base_register,
+                             long offset) override;
+  virtual bool RegisterRule(uint64 address, int reg,
+                            int base_register) override;
+  virtual bool ExpressionRule(uint64 address, int reg,
+                              const std::string& expression) override;
+  virtual bool ValExpressionRule(uint64 address, int reg,
+                                 const std::string& expression) override;
+  virtual bool End() override;
+
+ private:
+  // Return the name to use for register I.
+  const UniqueString* RegisterName(int i);
+
+  // The Summariser to which we should give entries
+  Summariser* summ_;
+
+  // Universe for creating UniqueStrings in, should that be necessary.
+  UniqueStringUniverse* usu_;
+
+  // The number of Dwarf-defined register names for this architecture.
+  const unsigned int num_dw_regs_;
+
+  // The reporter to use to report problems.
+  Reporter* reporter_;
+
+  // The ByteReader to use for parsing Dwarf expressions.
+  ByteReader* reader_;
+
+  // The section offset of the current frame description entry, for
+  // use in error messages.
+  size_t entry_offset_;
+
+  // The return address column for that entry.
+  unsigned return_address_;
+};
+
+// Convert the Dwarf expression in |expr| into PfxInstrs stored in the
+// SecMap referred to by |summ|, and return the index of the starting
+// PfxInstr added, which must be >= 0.  In case of failure return -1.
+int32_t parseDwarfExpr(Summariser* summ, const ByteReader* reader,
+                       std::string expr, bool debug, bool pushCfaAtStart,
+                       bool derefAtEnd);
+
+}  // namespace lul
+
+#endif  // LulDwarfExt_h
diff --git a/mozglue/baseprofiler/lul/LulDwarfInt.h b/mozglue/baseprofiler/lul/LulDwarfInt.h
new file mode 100644
index 0000000000..b72c6e08e3
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulDwarfInt.h
@@ -0,0 +1,193 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+// Copyright (c) 2008, 2010 Google Inc. All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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
+// OWNER OR CONTRIBUTORS 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.
+
+// CFI reader author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
+
+// This file is derived from the following file in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/dwarf/dwarf2enums.h
+
+#ifndef LulDwarfInt_h
+#define LulDwarfInt_h
+
+#include "LulCommonExt.h"
+#include "LulDwarfExt.h"
+
+namespace lul {
+
+// These enums do not follow the google3 style only because they are
+// known universally (specs, other implementations) by the names in
+// exactly this capitalization.
+// Tag names and codes.
+
+// Call Frame Info instructions.
+enum DwarfCFI {
+  DW_CFA_advance_loc = 0x40,
+  DW_CFA_offset = 0x80,
+  DW_CFA_restore = 0xc0,
+  DW_CFA_nop = 0x00,
+  DW_CFA_set_loc = 0x01,
+  DW_CFA_advance_loc1 = 0x02,
+  DW_CFA_advance_loc2 = 0x03,
+  DW_CFA_advance_loc4 = 0x04,
+  DW_CFA_offset_extended = 0x05,
+  DW_CFA_restore_extended = 0x06,
+  DW_CFA_undefined = 0x07,
+  DW_CFA_same_value = 0x08,
+  DW_CFA_register = 0x09,
+  DW_CFA_remember_state = 0x0a,
+  DW_CFA_restore_state = 0x0b,
+  DW_CFA_def_cfa = 0x0c,
+  DW_CFA_def_cfa_register = 0x0d,
+  DW_CFA_def_cfa_offset = 0x0e,
+  DW_CFA_def_cfa_expression = 0x0f,
+  DW_CFA_expression = 0x10,
+  DW_CFA_offset_extended_sf = 0x11,
+  DW_CFA_def_cfa_sf = 0x12,
+  DW_CFA_def_cfa_offset_sf = 0x13,
+  DW_CFA_val_offset = 0x14,
+  DW_CFA_val_offset_sf = 0x15,
+  DW_CFA_val_expression = 0x16,
+
+  // Opcodes in this range are reserved for user extensions.
+  DW_CFA_lo_user = 0x1c,
+  DW_CFA_hi_user = 0x3f,
+
+  // SGI/MIPS specific.
+  DW_CFA_MIPS_advance_loc8 = 0x1d,
+
+  // GNU extensions.
+  DW_CFA_GNU_window_save = 0x2d,
+  DW_CFA_GNU_args_size = 0x2e,
+  DW_CFA_GNU_negative_offset_extended = 0x2f
+};
+
+// Exception handling 'z' augmentation letters.
+enum DwarfZAugmentationCodes {
+  // If the CFI augmentation string begins with 'z', then the CIE and FDE
+  // have an augmentation data area just before the instructions, whose
+  // contents are determined by the subsequent augmentation letters.
+  DW_Z_augmentation_start = 'z',
+
+  // If this letter is present in a 'z' augmentation string, the CIE
+  // augmentation data includes a pointer encoding, and the FDE
+  // augmentation data includes a language-specific data area pointer,
+  // represented using that encoding.
+  DW_Z_has_LSDA = 'L',
+
+  // If this letter is present in a 'z' augmentation string, the CIE
+  // augmentation data includes a pointer encoding, followed by a pointer
+  // to a personality routine, represented using that encoding.
+  DW_Z_has_personality_routine = 'P',
+
+  // If this letter is present in a 'z' augmentation string, the CIE
+  // augmentation data includes a pointer encoding describing how the FDE's
+  // initial location, address range, and DW_CFA_set_loc operands are
+  // encoded.
+  DW_Z_has_FDE_address_encoding = 'R',
+
+  // If this letter is present in a 'z' augmentation string, then code
+  // addresses covered by FDEs that cite this CIE are signal delivery
+  // trampolines. Return addresses of frames in trampolines should not be
+  // adjusted as described in section 6.4.4 of the DWARF 3 spec.
+  DW_Z_is_signal_trampoline = 'S'
+};
+
+// Expression opcodes
+enum DwarfExpressionOpcodes {
+  DW_OP_addr = 0x03,
+  DW_OP_deref = 0x06,
+  DW_OP_const1s = 0x09,
+  DW_OP_const2u = 0x0a,
+  DW_OP_const2s = 0x0b,
+  DW_OP_const4u = 0x0c,
+  DW_OP_const4s = 0x0d,
+  DW_OP_const8u = 0x0e,
+  DW_OP_const8s = 0x0f,
+  DW_OP_constu = 0x10,
+  DW_OP_consts = 0x11,
+  DW_OP_dup = 0x12,
+  DW_OP_drop = 0x13,
+  DW_OP_over = 0x14,
+  DW_OP_pick = 0x15,
+  DW_OP_swap = 0x16,
+  DW_OP_rot = 0x17,
+  DW_OP_xderef = 0x18,
+  DW_OP_abs = 0x19,
+  DW_OP_and = 0x1a,
+  DW_OP_div = 0x1b,
+  DW_OP_minus = 0x1c,
+  DW_OP_mod = 0x1d,
+  DW_OP_mul = 0x1e,
+  DW_OP_neg = 0x1f,
+  DW_OP_not = 0x20,
+  DW_OP_or = 0x21,
+  DW_OP_plus = 0x22,
+  DW_OP_plus_uconst = 0x23,
+  DW_OP_shl = 0x24,
+  DW_OP_shr = 0x25,
+  DW_OP_shra = 0x26,
+  DW_OP_xor = 0x27,
+  DW_OP_skip = 0x2f,
+  DW_OP_bra = 0x28,
+  DW_OP_eq = 0x29,
+  DW_OP_ge = 0x2a,
+  DW_OP_gt = 0x2b,
+  DW_OP_le = 0x2c,
+  DW_OP_lt = 0x2d,
+  DW_OP_ne = 0x2e,
+  DW_OP_lit0 = 0x30,
+  DW_OP_lit31 = 0x4f,
+  DW_OP_reg0 = 0x50,
+  DW_OP_reg31 = 0x6f,
+  DW_OP_breg0 = 0x70,
+  DW_OP_breg31 = 0x8f,
+  DW_OP_regx = 0x90,
+  DW_OP_fbreg = 0x91,
+  DW_OP_bregx = 0x92,
+  DW_OP_piece = 0x93,
+  DW_OP_deref_size = 0x94,
+  DW_OP_xderef_size = 0x95,
+  DW_OP_nop = 0x96,
+  DW_OP_push_object_address = 0x97,
+  DW_OP_call2 = 0x98,
+  DW_OP_call4 = 0x99,
+  DW_OP_call_ref = 0x9a,
+  DW_OP_form_tls_address = 0x9b,
+  DW_OP_call_frame_cfa = 0x9c,
+  DW_OP_bit_piece = 0x9d,
+  DW_OP_lo_user = 0xe0,
+  DW_OP_hi_user = 0xff
+};
+
+}  // namespace lul
+
+#endif  // LulDwarfInt_h
diff --git a/mozglue/baseprofiler/lul/LulDwarfSummariser.cpp b/mozglue/baseprofiler/lul/LulDwarfSummariser.cpp
new file mode 100644
index 0000000000..ff0f212f6c
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulDwarfSummariser.cpp
@@ -0,0 +1,553 @@
+/* -*- 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 "LulDwarfSummariser.h"
+
+#include "mozilla/Assertions.h"
+#include "mozilla/Sprintf.h"
+
+#include "LulDwarfExt.h"
+
+// Set this to 1 for verbose logging
+#define DEBUG_SUMMARISER 0
+
+namespace lul {
+
+// Do |s64|'s lowest 32 bits sign extend back to |s64| itself?
+static inline bool fitsIn32Bits(int64 s64) {
+  return s64 == ((s64 & 0xffffffff) ^ 0x80000000) - 0x80000000;
+}
+
+// Check a LExpr prefix expression, starting at pfxInstrs[start] up to
+// the next PX_End instruction, to ensure that:
+// * It only mentions registers that are tracked on this target
+// * The start point is sane
+// If the expression is ok, return NULL.  Else return a pointer
+// a const char* holding a bit of text describing the problem.
+static const char* checkPfxExpr(const vector<PfxInstr>* pfxInstrs,
+                                int64_t start) {
+  size_t nInstrs = pfxInstrs->size();
+  if (start < 0 || start >= (ssize_t)nInstrs) {
+    return "bogus start point";
+  }
+  size_t i;
+  for (i = start; i < nInstrs; i++) {
+    PfxInstr pxi = (*pfxInstrs)[i];
+    if (pxi.mOpcode == PX_End) break;
+    if (pxi.mOpcode == PX_DwReg &&
+        !registerIsTracked((DW_REG_NUMBER)pxi.mOperand)) {
+      return "uses untracked reg";
+    }
+  }
+  return nullptr;  // success
+}
+
+Summariser::Summariser(SecMap* aSecMap, uintptr_t aTextBias,
+                       void (*aLog)(const char*))
+    : mSecMap(aSecMap), mTextBias(aTextBias), mLog(aLog) {
+  mCurrAddr = 0;
+  mMax1Addr = 0;  // Gives an empty range.
+
+  // Initialise the running RuleSet to "haven't got a clue" status.
+  new (&mCurrRules) RuleSet();
+}
+
+void Summariser::Entry(uintptr_t aAddress, uintptr_t aLength) {
+  aAddress += mTextBias;
+  if (DEBUG_SUMMARISER) {
+    char buf[100];
+    SprintfLiteral(buf, "LUL Entry(%llx, %llu)\n",
+                   (unsigned long long int)aAddress,
+                   (unsigned long long int)aLength);
+    mLog(buf);
+  }
+  // This throws away any previous summary, that is, assumes
+  // that the previous summary, if any, has been properly finished
+  // by a call to End().
+  mCurrAddr = aAddress;
+  mMax1Addr = aAddress + aLength;
+  new (&mCurrRules) RuleSet();
+}
+
+void Summariser::Rule(uintptr_t aAddress, int aNewReg, LExprHow how,
+                      int16_t oldReg, int64_t offset) {
+  aAddress += mTextBias;
+  if (DEBUG_SUMMARISER) {
+    char buf[100];
+    if (how == NODEREF || how == DEREF) {
+      bool deref = how == DEREF;
+      SprintfLiteral(buf, "LUL  0x%llx  old-r%d = %sr%d + %lld%s\n",
+                     (unsigned long long int)aAddress, aNewReg,
+                     deref ? "*(" : "", (int)oldReg, (long long int)offset,
+                     deref ? ")" : "");
+    } else if (how == PFXEXPR) {
+      SprintfLiteral(buf, "LUL  0x%llx  old-r%d = pfx-expr-at %lld\n",
+                     (unsigned long long int)aAddress, aNewReg,
+                     (long long int)offset);
+    } else {
+      SprintfLiteral(buf, "LUL  0x%llx  old-r%d = (invalid LExpr!)\n",
+                     (unsigned long long int)aAddress, aNewReg);
+    }
+    mLog(buf);
+  }
+
+  if (mCurrAddr < aAddress) {
+    // Flush the existing summary first.
+    mCurrRules.mAddr = mCurrAddr;
+    mCurrRules.mLen = aAddress - mCurrAddr;
+    mSecMap->AddRuleSet(&mCurrRules);
+    if (DEBUG_SUMMARISER) {
+      mLog("LUL  ");
+      mCurrRules.Print(mLog);
+      mLog("\n");
+    }
+    mCurrAddr = aAddress;
+  }
+
+  // If for some reason summarisation fails, either or both of these
+  // become non-null and point at constant text describing the
+  // problem.  Using two rather than just one avoids complications of
+  // having to concatenate two strings to produce a complete error message.
+  const char* reason1 = nullptr;
+  const char* reason2 = nullptr;
+
+  // |offset| needs to be a 32 bit value that sign extends to 64 bits
+  // on a 64 bit target.  We will need to incorporate |offset| into
+  // any LExpr made here.  So we may as well check it right now.
+  if (!fitsIn32Bits(offset)) {
+    reason1 = "offset not in signed 32-bit range";
+    goto cant_summarise;
+  }
+
+  // FIXME: factor out common parts of the arch-dependent summarisers.
+
+#if defined(GP_ARCH_arm)
+
+  // ----------------- arm ----------------- //
+
+  // Now, can we add the rule to our summary?  This depends on whether
+  // the registers and the overall expression are representable.  This
+  // is the heart of the summarisation process.
+  switch (aNewReg) {
+    case DW_REG_CFA:
+      // This is a rule that defines the CFA.  The only forms we
+      // choose to represent are: r7/11/12/13 + offset.  The offset
+      // must fit into 32 bits since 'uintptr_t' is 32 bit on ARM,
+      // hence there is no need to check it for overflow.
+      if (how != NODEREF) {
+        reason1 = "rule for DW_REG_CFA: invalid |how|";
+        goto cant_summarise;
+      }
+      switch (oldReg) {
+        case DW_REG_ARM_R7:
+        case DW_REG_ARM_R11:
+        case DW_REG_ARM_R12:
+        case DW_REG_ARM_R13:
+          break;
+        default:
+          reason1 = "rule for DW_REG_CFA: invalid |oldReg|";
+          goto cant_summarise;
+      }
+      mCurrRules.mCfaExpr = LExpr(how, oldReg, offset);
+      break;
+
+    case DW_REG_ARM_R7:
+    case DW_REG_ARM_R11:
+    case DW_REG_ARM_R12:
+    case DW_REG_ARM_R13:
+    case DW_REG_ARM_R14:
+    case DW_REG_ARM_R15: {
+      // This is a new rule for R7, R11, R12, R13 (SP), R14 (LR) or
+      // R15 (the return address).
+      switch (how) {
+        case NODEREF:
+        case DEREF:
+          // Check the old register is one we're tracking.
+          if (!registerIsTracked((DW_REG_NUMBER)oldReg) &&
+              oldReg != DW_REG_CFA) {
+            reason1 = "rule for R7/11/12/13/14/15: uses untracked reg";
+            goto cant_summarise;
+          }
+          break;
+        case PFXEXPR: {
+          // Check that the prefix expression only mentions tracked registers.
+          const vector<PfxInstr>* pfxInstrs = mSecMap->GetPfxInstrs();
+          reason2 = checkPfxExpr(pfxInstrs, offset);
+          if (reason2) {
+            reason1 = "rule for R7/11/12/13/14/15: ";
+            goto cant_summarise;
+          }
+          break;
+        }
+        default:
+          goto cant_summarise;
+      }
+      LExpr expr = LExpr(how, oldReg, offset);
+      switch (aNewReg) {
+        case DW_REG_ARM_R7:
+          mCurrRules.mR7expr = expr;
+          break;
+        case DW_REG_ARM_R11:
+          mCurrRules.mR11expr = expr;
+          break;
+        case DW_REG_ARM_R12:
+          mCurrRules.mR12expr = expr;
+          break;
+        case DW_REG_ARM_R13:
+          mCurrRules.mR13expr = expr;
+          break;
+        case DW_REG_ARM_R14:
+          mCurrRules.mR14expr = expr;
+          break;
+        case DW_REG_ARM_R15:
+          mCurrRules.mR15expr = expr;
+          break;
+        default:
+          MOZ_ASSERT(0);
+      }
+      break;
+    }
+
+    default:
+      // Leave |reason1| and |reason2| unset here.  This program point
+      // is reached so often that it causes a flood of "Can't
+      // summarise" messages.  In any case, we don't really care about
+      // the fact that this summary would produce a new value for a
+      // register that we're not tracking.  We do on the other hand
+      // care if the summary's expression *uses* a register that we're
+      // not tracking.  But in that case one of the above failures
+      // should tell us which.
+      goto cant_summarise;
+  }
+
+  // Mark callee-saved registers (r4 .. r11) as unchanged, if there is
+  // no other information about them.  FIXME: do this just once, at
+  // the point where the ruleset is committed.
+  if (mCurrRules.mR7expr.mHow == UNKNOWN) {
+    mCurrRules.mR7expr = LExpr(NODEREF, DW_REG_ARM_R7, 0);
+  }
+  if (mCurrRules.mR11expr.mHow == UNKNOWN) {
+    mCurrRules.mR11expr = LExpr(NODEREF, DW_REG_ARM_R11, 0);
+  }
+  if (mCurrRules.mR12expr.mHow == UNKNOWN) {
+    mCurrRules.mR12expr = LExpr(NODEREF, DW_REG_ARM_R12, 0);
+  }
+
+  // The old r13 (SP) value before the call is always the same as the
+  // CFA.
+  mCurrRules.mR13expr = LExpr(NODEREF, DW_REG_CFA, 0);
+
+  // If there's no information about R15 (the return address), say
+  // it's a copy of R14 (the link register).
+  if (mCurrRules.mR15expr.mHow == UNKNOWN) {
+    mCurrRules.mR15expr = LExpr(NODEREF, DW_REG_ARM_R14, 0);
+  }
+
+#elif defined(GP_ARCH_arm64)
+
+  // ----------------- arm64 ----------------- //
+
+  switch (aNewReg) {
+    case DW_REG_CFA:
+      if (how != NODEREF) {
+        reason1 = "rule for DW_REG_CFA: invalid |how|";
+        goto cant_summarise;
+      }
+      switch (oldReg) {
+        case DW_REG_AARCH64_X29:
+        case DW_REG_AARCH64_SP:
+          break;
+        default:
+          reason1 = "rule for DW_REG_CFA: invalid |oldReg|";
+          goto cant_summarise;
+      }
+      mCurrRules.mCfaExpr = LExpr(how, oldReg, offset);
+      break;
+
+    case DW_REG_AARCH64_X29:
+    case DW_REG_AARCH64_X30:
+    case DW_REG_AARCH64_SP: {
+      switch (how) {
+        case NODEREF:
+        case DEREF:
+          // Check the old register is one we're tracking.
+          if (!registerIsTracked((DW_REG_NUMBER)oldReg) &&
+              oldReg != DW_REG_CFA) {
+            reason1 = "rule for X29/X30/SP: uses untracked reg";
+            goto cant_summarise;
+          }
+          break;
+        case PFXEXPR: {
+          // Check that the prefix expression only mentions tracked registers.
+          const vector<PfxInstr>* pfxInstrs = mSecMap->GetPfxInstrs();
+          reason2 = checkPfxExpr(pfxInstrs, offset);
+          if (reason2) {
+            reason1 = "rule for X29/X30/SP: ";
+            goto cant_summarise;
+          }
+          break;
+        }
+        default:
+          goto cant_summarise;
+      }
+      LExpr expr = LExpr(how, oldReg, offset);
+      switch (aNewReg) {
+        case DW_REG_AARCH64_X29:
+          mCurrRules.mX29expr = expr;
+          break;
+        case DW_REG_AARCH64_X30:
+          mCurrRules.mX30expr = expr;
+          break;
+        case DW_REG_AARCH64_SP:
+          mCurrRules.mSPexpr = expr;
+          break;
+        default:
+          MOZ_ASSERT(0);
+      }
+      break;
+    }
+    default:
+      // Leave |reason1| and |reason2| unset here, for the reasons explained
+      // in the analogous point
+      goto cant_summarise;
+  }
+
+  if (mCurrRules.mX29expr.mHow == UNKNOWN) {
+    mCurrRules.mX29expr = LExpr(NODEREF, DW_REG_AARCH64_X29, 0);
+  }
+  if (mCurrRules.mX30expr.mHow == UNKNOWN) {
+    mCurrRules.mX30expr = LExpr(NODEREF, DW_REG_AARCH64_X30, 0);
+  }
+  // On aarch64, it seems the old SP value before the call is always the
+  // same as the CFA.  Therefore, in the absence of any other way to
+  // recover the SP, specify that the CFA should be copied.
+  if (mCurrRules.mSPexpr.mHow == UNKNOWN) {
+    mCurrRules.mSPexpr = LExpr(NODEREF, DW_REG_CFA, 0);
+  }
+#elif defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+
+  // ---------------- x64/x86 ---------------- //
+
+  // Now, can we add the rule to our summary?  This depends on whether
+  // the registers and the overall expression are representable.  This
+  // is the heart of the summarisation process.
+  switch (aNewReg) {
+    case DW_REG_CFA: {
+      // This is a rule that defines the CFA.  The only forms we choose to
+      // represent are: = SP+offset, = FP+offset, or =prefix-expr.
+      switch (how) {
+        case NODEREF:
+          if (oldReg != DW_REG_INTEL_XSP && oldReg != DW_REG_INTEL_XBP) {
+            reason1 = "rule for DW_REG_CFA: invalid |oldReg|";
+            goto cant_summarise;
+          }
+          break;
+        case DEREF:
+          reason1 = "rule for DW_REG_CFA: invalid |how|";
+          goto cant_summarise;
+        case PFXEXPR: {
+          // Check that the prefix expression only mentions tracked registers.
+          const vector<PfxInstr>* pfxInstrs = mSecMap->GetPfxInstrs();
+          reason2 = checkPfxExpr(pfxInstrs, offset);
+          if (reason2) {
+            reason1 = "rule for CFA: ";
+            goto cant_summarise;
+          }
+          break;
+        }
+        default:
+          goto cant_summarise;
+      }
+      mCurrRules.mCfaExpr = LExpr(how, oldReg, offset);
+      break;
+    }
+
+    case DW_REG_INTEL_XSP:
+    case DW_REG_INTEL_XBP:
+    case DW_REG_INTEL_XIP: {
+      // This is a new rule for XSP, XBP or XIP (the return address).
+      switch (how) {
+        case NODEREF:
+        case DEREF:
+          // Check the old register is one we're tracking.
+          if (!registerIsTracked((DW_REG_NUMBER)oldReg) &&
+              oldReg != DW_REG_CFA) {
+            reason1 = "rule for XSP/XBP/XIP: uses untracked reg";
+            goto cant_summarise;
+          }
+          break;
+        case PFXEXPR: {
+          // Check that the prefix expression only mentions tracked registers.
+          const vector<PfxInstr>* pfxInstrs = mSecMap->GetPfxInstrs();
+          reason2 = checkPfxExpr(pfxInstrs, offset);
+          if (reason2) {
+            reason1 = "rule for XSP/XBP/XIP: ";
+            goto cant_summarise;
+          }
+          break;
+        }
+        default:
+          goto cant_summarise;
+      }
+      LExpr expr = LExpr(how, oldReg, offset);
+      switch (aNewReg) {
+        case DW_REG_INTEL_XBP:
+          mCurrRules.mXbpExpr = expr;
+          break;
+        case DW_REG_INTEL_XSP:
+          mCurrRules.mXspExpr = expr;
+          break;
+        case DW_REG_INTEL_XIP:
+          mCurrRules.mXipExpr = expr;
+          break;
+        default:
+          MOZ_CRASH("impossible value for aNewReg");
+      }
+      break;
+    }
+
+    default:
+      // Leave |reason1| and |reason2| unset here, for the reasons
+      // explained in the analogous point in the ARM case just above.
+      goto cant_summarise;
+  }
+
+  // On Intel, it seems the old SP value before the call is always the
+  // same as the CFA.  Therefore, in the absence of any other way to
+  // recover the SP, specify that the CFA should be copied.
+  if (mCurrRules.mXspExpr.mHow == UNKNOWN) {
+    mCurrRules.mXspExpr = LExpr(NODEREF, DW_REG_CFA, 0);
+  }
+
+  // Also, gcc says "Undef" for BP when it is unchanged.
+  if (mCurrRules.mXbpExpr.mHow == UNKNOWN) {
+    mCurrRules.mXbpExpr = LExpr(NODEREF, DW_REG_INTEL_XBP, 0);
+  }
+
+#elif defined(GP_ARCH_mips64)
+  // ---------------- mips ---------------- //
+  //
+  // Now, can we add the rule to our summary?  This depends on whether
+  // the registers and the overall expression are representable.  This
+  // is the heart of the summarisation process.
+  switch (aNewReg) {
+    case DW_REG_CFA:
+      // This is a rule that defines the CFA.  The only forms we can
+      // represent are: = SP+offset or = FP+offset.
+      if (how != NODEREF) {
+        reason1 = "rule for DW_REG_CFA: invalid |how|";
+        goto cant_summarise;
+      }
+      if (oldReg != DW_REG_MIPS_SP && oldReg != DW_REG_MIPS_FP) {
+        reason1 = "rule for DW_REG_CFA: invalid |oldReg|";
+        goto cant_summarise;
+      }
+      mCurrRules.mCfaExpr = LExpr(how, oldReg, offset);
+      break;
+
+    case DW_REG_MIPS_SP:
+    case DW_REG_MIPS_FP:
+    case DW_REG_MIPS_PC: {
+      // This is a new rule for SP, FP or PC (the return address).
+      switch (how) {
+        case NODEREF:
+        case DEREF:
+          // Check the old register is one we're tracking.
+          if (!registerIsTracked((DW_REG_NUMBER)oldReg) &&
+              oldReg != DW_REG_CFA) {
+            reason1 = "rule for SP/FP/PC: uses untracked reg";
+            goto cant_summarise;
+          }
+          break;
+        case PFXEXPR: {
+          // Check that the prefix expression only mentions tracked registers.
+          const vector<PfxInstr>* pfxInstrs = mSecMap->GetPfxInstrs();
+          reason2 = checkPfxExpr(pfxInstrs, offset);
+          if (reason2) {
+            reason1 = "rule for SP/FP/PC: ";
+            goto cant_summarise;
+          }
+          break;
+        }
+        default:
+          goto cant_summarise;
+      }
+      LExpr expr = LExpr(how, oldReg, offset);
+      switch (aNewReg) {
+        case DW_REG_MIPS_FP:
+          mCurrRules.mFPexpr = expr;
+          break;
+        case DW_REG_MIPS_SP:
+          mCurrRules.mSPexpr = expr;
+          break;
+        case DW_REG_MIPS_PC:
+          mCurrRules.mPCexpr = expr;
+          break;
+        default:
+          MOZ_CRASH("impossible value for aNewReg");
+      }
+      break;
+    }
+    default:
+      // Leave |reason1| and |reason2| unset here, for the reasons
+      // explained in the analogous point in the ARM case just above.
+      goto cant_summarise;
+  }
+
+  // On MIPS, it seems the old SP value before the call is always the
+  // same as the CFA.  Therefore, in the absence of any other way to
+  // recover the SP, specify that the CFA should be copied.
+  if (mCurrRules.mSPexpr.mHow == UNKNOWN) {
+    mCurrRules.mSPexpr = LExpr(NODEREF, DW_REG_CFA, 0);
+  }
+
+  // Also, gcc says "Undef" for FP when it is unchanged.
+  if (mCurrRules.mFPexpr.mHow == UNKNOWN) {
+    mCurrRules.mFPexpr = LExpr(NODEREF, DW_REG_MIPS_FP, 0);
+  }
+
+#else
+
+#  error "Unsupported arch"
+#endif
+
+  return;
+
+cant_summarise:
+  if (reason1 || reason2) {
+    char buf[200];
+    SprintfLiteral(buf,
+                   "LUL  can't summarise: "
+                   "SVMA=0x%llx: %s%s, expr=LExpr(%s,%u,%lld)\n",
+                   (unsigned long long int)(aAddress - mTextBias),
+                   reason1 ? reason1 : "", reason2 ? reason2 : "",
+                   NameOf_LExprHow(how), (unsigned int)oldReg,
+                   (long long int)offset);
+    mLog(buf);
+  }
+}
+
+uint32_t Summariser::AddPfxInstr(PfxInstr pfxi) {
+  return mSecMap->AddPfxInstr(pfxi);
+}
+
+void Summariser::End() {
+  if (DEBUG_SUMMARISER) {
+    mLog("LUL End\n");
+  }
+  if (mCurrAddr < mMax1Addr) {
+    mCurrRules.mAddr = mCurrAddr;
+    mCurrRules.mLen = mMax1Addr - mCurrAddr;
+    mSecMap->AddRuleSet(&mCurrRules);
+    if (DEBUG_SUMMARISER) {
+      mLog("LUL  ");
+      mCurrRules.Print(mLog);
+      mLog("\n");
+    }
+  }
+}
+
+}  // namespace lul
diff --git a/mozglue/baseprofiler/lul/LulDwarfSummariser.h b/mozglue/baseprofiler/lul/LulDwarfSummariser.h
new file mode 100644
index 0000000000..30f1ba23c1
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulDwarfSummariser.h
@@ -0,0 +1,64 @@
+/* -*- 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 LulDwarfSummariser_h
+#define LulDwarfSummariser_h
+
+#include "LulMainInt.h"
+
+namespace lul {
+
+class Summariser {
+ public:
+  Summariser(SecMap* aSecMap, uintptr_t aTextBias, void (*aLog)(const char*));
+
+  virtual void Entry(uintptr_t aAddress, uintptr_t aLength);
+  virtual void End();
+
+  // Tell the summariser that the value for |aNewReg| at |aAddress| is
+  // recovered using the LExpr that can be constructed using the
+  // components |how|, |oldReg| and |offset|.  The summariser will
+  // inspect the components and may reject them for various reasons,
+  // but the hope is that it will find them acceptable and record this
+  // rule permanently.
+  virtual void Rule(uintptr_t aAddress, int aNewReg, LExprHow how,
+                    int16_t oldReg, int64_t offset);
+
+  virtual uint32_t AddPfxInstr(PfxInstr pfxi);
+
+  // Send output to the logging sink, for debugging.
+  virtual void Log(const char* str) { mLog(str); }
+
+ private:
+  // The SecMap in which we park the finished summaries (RuleSets) and
+  // also any PfxInstrs derived from Dwarf expressions.
+  SecMap* mSecMap;
+
+  // Running state for the current summary (RuleSet) under construction.
+  RuleSet mCurrRules;
+
+  // The start of the address range to which the RuleSet under
+  // construction applies.
+  uintptr_t mCurrAddr;
+
+  // The highest address, plus one, for which the RuleSet under
+  // construction could possibly apply.  If there are no further
+  // incoming events then mCurrRules will eventually be emitted
+  // as-is, for the range mCurrAddr.. mMax1Addr - 1, if that is
+  // nonempty.
+  uintptr_t mMax1Addr;
+
+  // The bias value (to add to the SVMAs, to get AVMAs) to be used
+  // when adding entries into mSecMap.
+  uintptr_t mTextBias;
+
+  // A logging sink, for debugging.
+  void (*mLog)(const char* aFmt);
+};
+
+}  // namespace lul
+
+#endif  // LulDwarfSummariser_h
diff --git a/mozglue/baseprofiler/lul/LulElf.cpp b/mozglue/baseprofiler/lul/LulElf.cpp
new file mode 100644
index 0000000000..b5d5e772be
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulElf.cpp
@@ -0,0 +1,873 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+// Copyright (c) 2006, 2011, 2012 Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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
+// OWNER OR CONTRIBUTORS 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.
+
+// Restructured in 2009 by: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
+
+// (derived from)
+// dump_symbols.cc: implement google_breakpad::WriteSymbolFile:
+// Find all the debugging info in a file and dump it as a Breakpad symbol file.
+//
+// dump_symbols.h: Read debugging information from an ELF file, and write
+// it out as a Breakpad symbol file.
+
+// This file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/linux/dump_symbols.cc
+//   src/common/linux/elfutils.cc
+//   src/common/linux/file_id.cc
+
+#include <errno.h>
+#include <fcntl.h>
+#include <libgen.h>
+#include <stdio.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#include <arpa/inet.h>
+
+#include <cstdlib>
+#include <set>
+#include <string>
+#include <vector>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/Sprintf.h"
+
+#include "PlatformMacros.h"
+#include "LulCommonExt.h"
+#include "LulDwarfExt.h"
+#include "LulElfInt.h"
+#include "LulMainInt.h"
+
+#if defined(GP_PLAT_arm_android) && !defined(SHT_ARM_EXIDX)
+// bionic and older glibsc don't define it
+#  define SHT_ARM_EXIDX (SHT_LOPROC + 1)
+#endif
+
+// Old Linux header doesn't define EM_AARCH64
+#ifndef EM_AARCH64
+#  define EM_AARCH64 183
+#endif
+
+// This namespace contains helper functions.
+namespace {
+
+using lul::DwarfCFIToModule;
+using lul::FindElfSectionByName;
+using lul::GetOffset;
+using lul::IsValidElf;
+using lul::Module;
+using lul::scoped_ptr;
+using lul::Summariser;
+using lul::UniqueStringUniverse;
+using std::set;
+using std::string;
+using std::vector;
+
+//
+// FDWrapper
+//
+// Wrapper class to make sure opened file is closed.
+//
+class FDWrapper {
+ public:
+  explicit FDWrapper(int fd) : fd_(fd) {}
+  ~FDWrapper() {
+    if (fd_ != -1) close(fd_);
+  }
+  int get() { return fd_; }
+  int release() {
+    int fd = fd_;
+    fd_ = -1;
+    return fd;
+  }
+
+ private:
+  int fd_;
+};
+
+//
+// MmapWrapper
+//
+// Wrapper class to make sure mapped regions are unmapped.
+//
+class MmapWrapper {
+ public:
+  MmapWrapper() : is_set_(false), base_(NULL), size_(0) {}
+  ~MmapWrapper() {
+    if (is_set_ && base_ != NULL) {
+      MOZ_ASSERT(size_ > 0);
+      munmap(base_, size_);
+    }
+  }
+  void set(void* mapped_address, size_t mapped_size) {
+    is_set_ = true;
+    base_ = mapped_address;
+    size_ = mapped_size;
+  }
+  void release() {
+    MOZ_ASSERT(is_set_);
+    is_set_ = false;
+    base_ = NULL;
+    size_ = 0;
+  }
+
+ private:
+  bool is_set_;
+  void* base_;
+  size_t size_;
+};
+
+// Set NUM_DW_REGNAMES to be the number of Dwarf register names
+// appropriate to the machine architecture given in HEADER.  Return
+// true on success, or false if HEADER's machine architecture is not
+// supported.
+template <typename ElfClass>
+bool DwarfCFIRegisterNames(const typename ElfClass::Ehdr* elf_header,
+                           unsigned int* num_dw_regnames) {
+  switch (elf_header->e_machine) {
+    case EM_386:
+      *num_dw_regnames = DwarfCFIToModule::RegisterNames::I386();
+      return true;
+    case EM_ARM:
+      *num_dw_regnames = DwarfCFIToModule::RegisterNames::ARM();
+      return true;
+    case EM_X86_64:
+      *num_dw_regnames = DwarfCFIToModule::RegisterNames::X86_64();
+      return true;
+    case EM_MIPS:
+      *num_dw_regnames = DwarfCFIToModule::RegisterNames::MIPS();
+      return true;
+    case EM_AARCH64:
+      *num_dw_regnames = DwarfCFIToModule::RegisterNames::ARM64();
+      return true;
+    default:
+      MOZ_ASSERT(0);
+      return false;
+  }
+}
+
+template <typename ElfClass>
+bool LoadDwarfCFI(const string& dwarf_filename,
+                  const typename ElfClass::Ehdr* elf_header,
+                  const char* section_name,
+                  const typename ElfClass::Shdr* section, const bool eh_frame,
+                  const typename ElfClass::Shdr* got_section,
+                  const typename ElfClass::Shdr* text_section,
+                  const bool big_endian, SecMap* smap, uintptr_t text_bias,
+                  UniqueStringUniverse* usu, void (*log)(const char*)) {
+  // Find the appropriate set of register names for this file's
+  // architecture.
+  unsigned int num_dw_regs = 0;
+  if (!DwarfCFIRegisterNames<ElfClass>(elf_header, &num_dw_regs)) {
+    fprintf(stderr,
+            "%s: unrecognized ELF machine architecture '%d';"
+            " cannot convert DWARF call frame information\n",
+            dwarf_filename.c_str(), elf_header->e_machine);
+    return false;
+  }
+
+  const lul::Endianness endianness =
+      big_endian ? lul::ENDIANNESS_BIG : lul::ENDIANNESS_LITTLE;
+
+  // Find the call frame information and its size.
+  const char* cfi = GetOffset<ElfClass, char>(elf_header, section->sh_offset);
+  size_t cfi_size = section->sh_size;
+
+  // Plug together the parser, handler, and their entourages.
+
+  // Here's a summariser, which will receive the output of the
+  // parser, create summaries, and add them to |smap|.
+  Summariser summ(smap, text_bias, log);
+
+  lul::ByteReader reader(endianness);
+  reader.SetAddressSize(ElfClass::kAddrSize);
+
+  DwarfCFIToModule::Reporter module_reporter(log, dwarf_filename, section_name);
+  DwarfCFIToModule handler(num_dw_regs, &module_reporter, &reader, usu, &summ);
+
+  // Provide the base addresses for .eh_frame encoded pointers, if
+  // possible.
+  reader.SetCFIDataBase(section->sh_addr, cfi);
+  if (got_section) reader.SetDataBase(got_section->sh_addr);
+  if (text_section) reader.SetTextBase(text_section->sh_addr);
+
+  lul::CallFrameInfo::Reporter dwarf_reporter(log, dwarf_filename,
+                                              section_name);
+  lul::CallFrameInfo parser(cfi, cfi_size, &reader, &handler, &dwarf_reporter,
+                            eh_frame);
+  parser.Start();
+
+  return true;
+}
+
+bool LoadELF(const string& obj_file, MmapWrapper* map_wrapper,
+             void** elf_header) {
+  int obj_fd = open(obj_file.c_str(), O_RDONLY);
+  if (obj_fd < 0) {
+    fprintf(stderr, "Failed to open ELF file '%s': %s\n", obj_file.c_str(),
+            strerror(errno));
+    return false;
+  }
+  FDWrapper obj_fd_wrapper(obj_fd);
+  struct stat st;
+  if (fstat(obj_fd, &st) != 0 && st.st_size <= 0) {
+    fprintf(stderr, "Unable to fstat ELF file '%s': %s\n", obj_file.c_str(),
+            strerror(errno));
+    return false;
+  }
+  // Mapping it read-only is good enough.  In any case, mapping it
+  // read-write confuses Valgrind's debuginfo acquire/discard
+  // heuristics, making it hard to profile the profiler.
+  void* obj_base = mmap(nullptr, st.st_size, PROT_READ, MAP_PRIVATE, obj_fd, 0);
+  if (obj_base == MAP_FAILED) {
+    fprintf(stderr, "Failed to mmap ELF file '%s': %s\n", obj_file.c_str(),
+            strerror(errno));
+    return false;
+  }
+  map_wrapper->set(obj_base, st.st_size);
+  *elf_header = obj_base;
+  if (!IsValidElf(*elf_header)) {
+    fprintf(stderr, "Not a valid ELF file: %s\n", obj_file.c_str());
+    return false;
+  }
+  return true;
+}
+
+// Get the endianness of ELF_HEADER. If it's invalid, return false.
+template <typename ElfClass>
+bool ElfEndianness(const typename ElfClass::Ehdr* elf_header,
+                   bool* big_endian) {
+  if (elf_header->e_ident[EI_DATA] == ELFDATA2LSB) {
+    *big_endian = false;
+    return true;
+  }
+  if (elf_header->e_ident[EI_DATA] == ELFDATA2MSB) {
+    *big_endian = true;
+    return true;
+  }
+
+  fprintf(stderr, "bad data encoding in ELF header: %d\n",
+          elf_header->e_ident[EI_DATA]);
+  return false;
+}
+
+//
+// LoadSymbolsInfo
+//
+// Holds the state between the two calls to LoadSymbols() in case it's necessary
+// to follow the .gnu_debuglink section and load debug information from a
+// different file.
+//
+template <typename ElfClass>
+class LoadSymbolsInfo {
+ public:
+  typedef typename ElfClass::Addr Addr;
+
+  explicit LoadSymbolsInfo(const vector<string>& dbg_dirs)
+      : debug_dirs_(dbg_dirs), has_loading_addr_(false) {}
+
+  // Keeps track of which sections have been loaded so sections don't
+  // accidentally get loaded twice from two different files.
+  void LoadedSection(const string& section) {
+    if (loaded_sections_.count(section) == 0) {
+      loaded_sections_.insert(section);
+    } else {
+      fprintf(stderr, "Section %s has already been loaded.\n", section.c_str());
+    }
+  }
+
+  string debuglink_file() const { return debuglink_file_; }
+
+ private:
+  const vector<string>& debug_dirs_;  // Directories in which to
+                                      // search for the debug ELF file.
+
+  string debuglink_file_;  // Full path to the debug ELF file.
+
+  bool has_loading_addr_;  // Indicate if LOADING_ADDR_ is valid.
+
+  set<string> loaded_sections_;  // Tracks the Loaded ELF sections
+                                 // between calls to LoadSymbols().
+};
+
+// Find the preferred loading address of the binary.
+template <typename ElfClass>
+typename ElfClass::Addr GetLoadingAddress(
+    const typename ElfClass::Phdr* program_headers, int nheader) {
+  typedef typename ElfClass::Phdr Phdr;
+
+  // For non-PIC executables (e_type == ET_EXEC), the load address is
+  // the start address of the first PT_LOAD segment.  (ELF requires
+  // the segments to be sorted by load address.)  For PIC executables
+  // and dynamic libraries (e_type == ET_DYN), this address will
+  // normally be zero.
+  for (int i = 0; i < nheader; ++i) {
+    const Phdr& header = program_headers[i];
+    if (header.p_type == PT_LOAD) return header.p_vaddr;
+  }
+  return 0;
+}
+
+template <typename ElfClass>
+bool LoadSymbols(const string& obj_file, const bool big_endian,
+                 const typename ElfClass::Ehdr* elf_header,
+                 const bool read_gnu_debug_link,
+                 LoadSymbolsInfo<ElfClass>* info, SecMap* smap, void* rx_avma,
+                 size_t rx_size, UniqueStringUniverse* usu,
+                 void (*log)(const char*)) {
+  typedef typename ElfClass::Phdr Phdr;
+  typedef typename ElfClass::Shdr Shdr;
+
+  char buf[500];
+  SprintfLiteral(buf, "LoadSymbols: BEGIN   %s\n", obj_file.c_str());
+  buf[sizeof(buf) - 1] = 0;
+  log(buf);
+
+  // This is how the text bias is calculated.
+  // BEGIN CALCULATE BIAS
+  uintptr_t loading_addr = GetLoadingAddress<ElfClass>(
+      GetOffset<ElfClass, Phdr>(elf_header, elf_header->e_phoff),
+      elf_header->e_phnum);
+  uintptr_t text_bias = ((uintptr_t)rx_avma) - loading_addr;
+  SprintfLiteral(buf, "LoadSymbols:   rx_avma=%llx, text_bias=%llx",
+                 (unsigned long long int)(uintptr_t)rx_avma,
+                 (unsigned long long int)text_bias);
+  buf[sizeof(buf) - 1] = 0;
+  log(buf);
+  // END CALCULATE BIAS
+
+  const Shdr* sections =
+      GetOffset<ElfClass, Shdr>(elf_header, elf_header->e_shoff);
+  const Shdr* section_names = sections + elf_header->e_shstrndx;
+  const char* names =
+      GetOffset<ElfClass, char>(elf_header, section_names->sh_offset);
+  const char* names_end = names + section_names->sh_size;
+  bool found_usable_info = false;
+
+  // Dwarf Call Frame Information (CFI) is actually independent from
+  // the other DWARF debugging information, and can be used alone.
+  const Shdr* dwarf_cfi_section =
+      FindElfSectionByName<ElfClass>(".debug_frame", SHT_PROGBITS, sections,
+                                     names, names_end, elf_header->e_shnum);
+  if (dwarf_cfi_section) {
+    // Ignore the return value of this function; even without call frame
+    // information, the other debugging information could be perfectly
+    // useful.
+    info->LoadedSection(".debug_frame");
+    bool result = LoadDwarfCFI<ElfClass>(obj_file, elf_header, ".debug_frame",
+                                         dwarf_cfi_section, false, 0, 0,
+                                         big_endian, smap, text_bias, usu, log);
+    found_usable_info = found_usable_info || result;
+    if (result) log("LoadSymbols:   read CFI from .debug_frame");
+  }
+
+  // Linux C++ exception handling information can also provide
+  // unwinding data.
+  const Shdr* eh_frame_section =
+      FindElfSectionByName<ElfClass>(".eh_frame", SHT_PROGBITS, sections, names,
+                                     names_end, elf_header->e_shnum);
+  if (eh_frame_section) {
+    // Pointers in .eh_frame data may be relative to the base addresses of
+    // certain sections. Provide those sections if present.
+    const Shdr* got_section = FindElfSectionByName<ElfClass>(
+        ".got", SHT_PROGBITS, sections, names, names_end, elf_header->e_shnum);
+    const Shdr* text_section = FindElfSectionByName<ElfClass>(
+        ".text", SHT_PROGBITS, sections, names, names_end, elf_header->e_shnum);
+    info->LoadedSection(".eh_frame");
+    // As above, ignore the return value of this function.
+    bool result = LoadDwarfCFI<ElfClass>(
+        obj_file, elf_header, ".eh_frame", eh_frame_section, true, got_section,
+        text_section, big_endian, smap, text_bias, usu, log);
+    found_usable_info = found_usable_info || result;
+    if (result) log("LoadSymbols:   read CFI from .eh_frame");
+  }
+
+  SprintfLiteral(buf, "LoadSymbols: END     %s\n", obj_file.c_str());
+  buf[sizeof(buf) - 1] = 0;
+  log(buf);
+
+  return found_usable_info;
+}
+
+// Return the breakpad symbol file identifier for the architecture of
+// ELF_HEADER.
+template <typename ElfClass>
+const char* ElfArchitecture(const typename ElfClass::Ehdr* elf_header) {
+  typedef typename ElfClass::Half Half;
+  Half arch = elf_header->e_machine;
+  switch (arch) {
+    case EM_386:
+      return "x86";
+    case EM_ARM:
+      return "arm";
+    case EM_AARCH64:
+      return "arm64";
+    case EM_MIPS:
+      return "mips";
+    case EM_PPC64:
+      return "ppc64";
+    case EM_PPC:
+      return "ppc";
+    case EM_S390:
+      return "s390";
+    case EM_SPARC:
+      return "sparc";
+    case EM_SPARCV9:
+      return "sparcv9";
+    case EM_X86_64:
+      return "x86_64";
+    default:
+      return NULL;
+  }
+}
+
+// Format the Elf file identifier in IDENTIFIER as a UUID with the
+// dashes removed.
+string FormatIdentifier(unsigned char identifier[16]) {
+  char identifier_str[40];
+  lul::FileID::ConvertIdentifierToString(identifier, identifier_str,
+                                         sizeof(identifier_str));
+  string id_no_dash;
+  for (int i = 0; identifier_str[i] != '\0'; ++i)
+    if (identifier_str[i] != '-') id_no_dash += identifier_str[i];
+  // Add an extra "0" by the end.  PDB files on Windows have an 'age'
+  // number appended to the end of the file identifier; this isn't
+  // really used or necessary on other platforms, but be consistent.
+  id_no_dash += '0';
+  return id_no_dash;
+}
+
+// Return the non-directory portion of FILENAME: the portion after the
+// last slash, or the whole filename if there are no slashes.
+string BaseFileName(const string& filename) {
+  // Lots of copies!  basename's behavior is less than ideal.
+  char* c_filename = strdup(filename.c_str());
+  string base = basename(c_filename);
+  free(c_filename);
+  return base;
+}
+
+template <typename ElfClass>
+bool ReadSymbolDataElfClass(const typename ElfClass::Ehdr* elf_header,
+                            const string& obj_filename,
+                            const vector<string>& debug_dirs, SecMap* smap,
+                            void* rx_avma, size_t rx_size,
+                            UniqueStringUniverse* usu,
+                            void (*log)(const char*)) {
+  typedef typename ElfClass::Ehdr Ehdr;
+
+  unsigned char identifier[16];
+  if (!lul ::FileID::ElfFileIdentifierFromMappedFile(elf_header, identifier)) {
+    fprintf(stderr, "%s: unable to generate file identifier\n",
+            obj_filename.c_str());
+    return false;
+  }
+
+  const char* architecture = ElfArchitecture<ElfClass>(elf_header);
+  if (!architecture) {
+    fprintf(stderr, "%s: unrecognized ELF machine architecture: %d\n",
+            obj_filename.c_str(), elf_header->e_machine);
+    return false;
+  }
+
+  // Figure out what endianness this file is.
+  bool big_endian;
+  if (!ElfEndianness<ElfClass>(elf_header, &big_endian)) return false;
+
+  string name = BaseFileName(obj_filename);
+  string os = "Linux";
+  string id = FormatIdentifier(identifier);
+
+  LoadSymbolsInfo<ElfClass> info(debug_dirs);
+  if (!LoadSymbols<ElfClass>(obj_filename, big_endian, elf_header,
+                             !debug_dirs.empty(), &info, smap, rx_avma, rx_size,
+                             usu, log)) {
+    const string debuglink_file = info.debuglink_file();
+    if (debuglink_file.empty()) return false;
+
+    // Load debuglink ELF file.
+    fprintf(stderr, "Found debugging info in %s\n", debuglink_file.c_str());
+    MmapWrapper debug_map_wrapper;
+    Ehdr* debug_elf_header = NULL;
+    if (!LoadELF(debuglink_file, &debug_map_wrapper,
+                 reinterpret_cast<void**>(&debug_elf_header)))
+      return false;
+    // Sanity checks to make sure everything matches up.
+    const char* debug_architecture =
+        ElfArchitecture<ElfClass>(debug_elf_header);
+    if (!debug_architecture) {
+      fprintf(stderr, "%s: unrecognized ELF machine architecture: %d\n",
+              debuglink_file.c_str(), debug_elf_header->e_machine);
+      return false;
+    }
+    if (strcmp(architecture, debug_architecture)) {
+      fprintf(stderr,
+              "%s with ELF machine architecture %s does not match "
+              "%s with ELF architecture %s\n",
+              debuglink_file.c_str(), debug_architecture, obj_filename.c_str(),
+              architecture);
+      return false;
+    }
+
+    bool debug_big_endian;
+    if (!ElfEndianness<ElfClass>(debug_elf_header, &debug_big_endian))
+      return false;
+    if (debug_big_endian != big_endian) {
+      fprintf(stderr, "%s and %s does not match in endianness\n",
+              obj_filename.c_str(), debuglink_file.c_str());
+      return false;
+    }
+
+    if (!LoadSymbols<ElfClass>(debuglink_file, debug_big_endian,
+                               debug_elf_header, false, &info, smap, rx_avma,
+                               rx_size, usu, log)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+}  // namespace
+
+namespace lul {
+
+bool ReadSymbolDataInternal(const uint8_t* obj_file, const string& obj_filename,
+                            const vector<string>& debug_dirs, SecMap* smap,
+                            void* rx_avma, size_t rx_size,
+                            UniqueStringUniverse* usu,
+                            void (*log)(const char*)) {
+  if (!IsValidElf(obj_file)) {
+    fprintf(stderr, "Not a valid ELF file: %s\n", obj_filename.c_str());
+    return false;
+  }
+
+  int elfclass = ElfClass(obj_file);
+  if (elfclass == ELFCLASS32) {
+    return ReadSymbolDataElfClass<ElfClass32>(
+        reinterpret_cast<const Elf32_Ehdr*>(obj_file), obj_filename, debug_dirs,
+        smap, rx_avma, rx_size, usu, log);
+  }
+  if (elfclass == ELFCLASS64) {
+    return ReadSymbolDataElfClass<ElfClass64>(
+        reinterpret_cast<const Elf64_Ehdr*>(obj_file), obj_filename, debug_dirs,
+        smap, rx_avma, rx_size, usu, log);
+  }
+
+  return false;
+}
+
+bool ReadSymbolData(const string& obj_file, const vector<string>& debug_dirs,
+                    SecMap* smap, void* rx_avma, size_t rx_size,
+                    UniqueStringUniverse* usu, void (*log)(const char*)) {
+  MmapWrapper map_wrapper;
+  void* elf_header = NULL;
+  if (!LoadELF(obj_file, &map_wrapper, &elf_header)) return false;
+
+  return ReadSymbolDataInternal(reinterpret_cast<uint8_t*>(elf_header),
+                                obj_file, debug_dirs, smap, rx_avma, rx_size,
+                                usu, log);
+}
+
+namespace {
+
+template <typename ElfClass>
+void FindElfClassSection(const char* elf_base, const char* section_name,
+                         typename ElfClass::Word section_type,
+                         const void** section_start, int* section_size) {
+  typedef typename ElfClass::Ehdr Ehdr;
+  typedef typename ElfClass::Shdr Shdr;
+
+  MOZ_ASSERT(elf_base);
+  MOZ_ASSERT(section_start);
+  MOZ_ASSERT(section_size);
+
+  MOZ_ASSERT(strncmp(elf_base, ELFMAG, SELFMAG) == 0);
+
+  const Ehdr* elf_header = reinterpret_cast<const Ehdr*>(elf_base);
+  MOZ_ASSERT(elf_header->e_ident[EI_CLASS] == ElfClass::kClass);
+
+  const Shdr* sections =
+      GetOffset<ElfClass, Shdr>(elf_header, elf_header->e_shoff);
+  const Shdr* section_names = sections + elf_header->e_shstrndx;
+  const char* names =
+      GetOffset<ElfClass, char>(elf_header, section_names->sh_offset);
+  const char* names_end = names + section_names->sh_size;
+
+  const Shdr* section =
+      FindElfSectionByName<ElfClass>(section_name, section_type, sections,
+                                     names, names_end, elf_header->e_shnum);
+
+  if (section != NULL && section->sh_size > 0) {
+    *section_start = elf_base + section->sh_offset;
+    *section_size = section->sh_size;
+  }
+}
+
+template <typename ElfClass>
+void FindElfClassSegment(const char* elf_base,
+                         typename ElfClass::Word segment_type,
+                         const void** segment_start, int* segment_size) {
+  typedef typename ElfClass::Ehdr Ehdr;
+  typedef typename ElfClass::Phdr Phdr;
+
+  MOZ_ASSERT(elf_base);
+  MOZ_ASSERT(segment_start);
+  MOZ_ASSERT(segment_size);
+
+  MOZ_ASSERT(strncmp(elf_base, ELFMAG, SELFMAG) == 0);
+
+  const Ehdr* elf_header = reinterpret_cast<const Ehdr*>(elf_base);
+  MOZ_ASSERT(elf_header->e_ident[EI_CLASS] == ElfClass::kClass);
+
+  const Phdr* phdrs =
+      GetOffset<ElfClass, Phdr>(elf_header, elf_header->e_phoff);
+
+  for (int i = 0; i < elf_header->e_phnum; ++i) {
+    if (phdrs[i].p_type == segment_type) {
+      *segment_start = elf_base + phdrs[i].p_offset;
+      *segment_size = phdrs[i].p_filesz;
+      return;
+    }
+  }
+}
+
+}  // namespace
+
+bool IsValidElf(const void* elf_base) {
+  return strncmp(reinterpret_cast<const char*>(elf_base), ELFMAG, SELFMAG) == 0;
+}
+
+int ElfClass(const void* elf_base) {
+  const ElfW(Ehdr)* elf_header = reinterpret_cast<const ElfW(Ehdr)*>(elf_base);
+
+  return elf_header->e_ident[EI_CLASS];
+}
+
+bool FindElfSection(const void* elf_mapped_base, const char* section_name,
+                    uint32_t section_type, const void** section_start,
+                    int* section_size, int* elfclass) {
+  MOZ_ASSERT(elf_mapped_base);
+  MOZ_ASSERT(section_start);
+  MOZ_ASSERT(section_size);
+
+  *section_start = NULL;
+  *section_size = 0;
+
+  if (!IsValidElf(elf_mapped_base)) return false;
+
+  int cls = ElfClass(elf_mapped_base);
+  if (elfclass) {
+    *elfclass = cls;
+  }
+
+  const char* elf_base = static_cast<const char*>(elf_mapped_base);
+
+  if (cls == ELFCLASS32) {
+    FindElfClassSection<ElfClass32>(elf_base, section_name, section_type,
+                                    section_start, section_size);
+    return *section_start != NULL;
+  } else if (cls == ELFCLASS64) {
+    FindElfClassSection<ElfClass64>(elf_base, section_name, section_type,
+                                    section_start, section_size);
+    return *section_start != NULL;
+  }
+
+  return false;
+}
+
+bool FindElfSegment(const void* elf_mapped_base, uint32_t segment_type,
+                    const void** segment_start, int* segment_size,
+                    int* elfclass) {
+  MOZ_ASSERT(elf_mapped_base);
+  MOZ_ASSERT(segment_start);
+  MOZ_ASSERT(segment_size);
+
+  *segment_start = NULL;
+  *segment_size = 0;
+
+  if (!IsValidElf(elf_mapped_base)) return false;
+
+  int cls = ElfClass(elf_mapped_base);
+  if (elfclass) {
+    *elfclass = cls;
+  }
+
+  const char* elf_base = static_cast<const char*>(elf_mapped_base);
+
+  if (cls == ELFCLASS32) {
+    FindElfClassSegment<ElfClass32>(elf_base, segment_type, segment_start,
+                                    segment_size);
+    return *segment_start != NULL;
+  } else if (cls == ELFCLASS64) {
+    FindElfClassSegment<ElfClass64>(elf_base, segment_type, segment_start,
+                                    segment_size);
+    return *segment_start != NULL;
+  }
+
+  return false;
+}
+
+// (derived from)
+// file_id.cc: Return a unique identifier for a file
+//
+// See file_id.h for documentation
+//
+
+// ELF note name and desc are 32-bits word padded.
+#define NOTE_PADDING(a) ((a + 3) & ~3)
+
+// These functions are also used inside the crashed process, so be safe
+// and use the syscall/libc wrappers instead of direct syscalls or libc.
+
+template <typename ElfClass>
+static bool ElfClassBuildIDNoteIdentifier(const void* section, int length,
+                                          uint8_t identifier[kMDGUIDSize]) {
+  typedef typename ElfClass::Nhdr Nhdr;
+
+  const void* section_end = reinterpret_cast<const char*>(section) + length;
+  const Nhdr* note_header = reinterpret_cast<const Nhdr*>(section);
+  while (reinterpret_cast<const void*>(note_header) < section_end) {
+    if (note_header->n_type == NT_GNU_BUILD_ID) break;
+    note_header = reinterpret_cast<const Nhdr*>(
+        reinterpret_cast<const char*>(note_header) + sizeof(Nhdr) +
+        NOTE_PADDING(note_header->n_namesz) +
+        NOTE_PADDING(note_header->n_descsz));
+  }
+  if (reinterpret_cast<const void*>(note_header) >= section_end ||
+      note_header->n_descsz == 0) {
+    return false;
+  }
+
+  const char* build_id = reinterpret_cast<const char*>(note_header) +
+                         sizeof(Nhdr) + NOTE_PADDING(note_header->n_namesz);
+  // Copy as many bits of the build ID as will fit
+  // into the GUID space.
+  memset(identifier, 0, kMDGUIDSize);
+  memcpy(identifier, build_id,
+         std::min(kMDGUIDSize, (size_t)note_header->n_descsz));
+
+  return true;
+}
+
+// Attempt to locate a .note.gnu.build-id section in an ELF binary
+// and copy as many bytes of it as will fit into |identifier|.
+static bool FindElfBuildIDNote(const void* elf_mapped_base,
+                               uint8_t identifier[kMDGUIDSize]) {
+  void* note_section;
+  int note_size, elfclass;
+  if ((!FindElfSegment(elf_mapped_base, PT_NOTE, (const void**)&note_section,
+                       &note_size, &elfclass) ||
+       note_size == 0) &&
+      (!FindElfSection(elf_mapped_base, ".note.gnu.build-id", SHT_NOTE,
+                       (const void**)&note_section, &note_size, &elfclass) ||
+       note_size == 0)) {
+    return false;
+  }
+
+  if (elfclass == ELFCLASS32) {
+    return ElfClassBuildIDNoteIdentifier<ElfClass32>(note_section, note_size,
+                                                     identifier);
+  } else if (elfclass == ELFCLASS64) {
+    return ElfClassBuildIDNoteIdentifier<ElfClass64>(note_section, note_size,
+                                                     identifier);
+  }
+
+  return false;
+}
+
+// Attempt to locate the .text section of an ELF binary and generate
+// a simple hash by XORing the first page worth of bytes into |identifier|.
+static bool HashElfTextSection(const void* elf_mapped_base,
+                               uint8_t identifier[kMDGUIDSize]) {
+  void* text_section;
+  int text_size;
+  if (!FindElfSection(elf_mapped_base, ".text", SHT_PROGBITS,
+                      (const void**)&text_section, &text_size, NULL) ||
+      text_size == 0) {
+    return false;
+  }
+
+  memset(identifier, 0, kMDGUIDSize);
+  const uint8_t* ptr = reinterpret_cast<const uint8_t*>(text_section);
+  const uint8_t* ptr_end = ptr + std::min(text_size, 4096);
+  while (ptr < ptr_end) {
+    for (unsigned i = 0; i < kMDGUIDSize; i++) identifier[i] ^= ptr[i];
+    ptr += kMDGUIDSize;
+  }
+  return true;
+}
+
+// static
+bool FileID::ElfFileIdentifierFromMappedFile(const void* base,
+                                             uint8_t identifier[kMDGUIDSize]) {
+  // Look for a build id note first.
+  if (FindElfBuildIDNote(base, identifier)) return true;
+
+  // Fall back on hashing the first page of the text section.
+  return HashElfTextSection(base, identifier);
+}
+
+// static
+void FileID::ConvertIdentifierToString(const uint8_t identifier[kMDGUIDSize],
+                                       char* buffer, int buffer_length) {
+  uint8_t identifier_swapped[kMDGUIDSize];
+
+  // Endian-ness swap to match dump processor expectation.
+  memcpy(identifier_swapped, identifier, kMDGUIDSize);
+  uint32_t* data1 = reinterpret_cast<uint32_t*>(identifier_swapped);
+  *data1 = htonl(*data1);
+  uint16_t* data2 = reinterpret_cast<uint16_t*>(identifier_swapped + 4);
+  *data2 = htons(*data2);
+  uint16_t* data3 = reinterpret_cast<uint16_t*>(identifier_swapped + 6);
+  *data3 = htons(*data3);
+
+  int buffer_idx = 0;
+  for (unsigned int idx = 0;
+       (buffer_idx < buffer_length) && (idx < kMDGUIDSize); ++idx) {
+    int hi = (identifier_swapped[idx] >> 4) & 0x0F;
+    int lo = (identifier_swapped[idx]) & 0x0F;
+
+    if (idx == 4 || idx == 6 || idx == 8 || idx == 10)
+      buffer[buffer_idx++] = '-';
+
+    buffer[buffer_idx++] = (hi >= 10) ? 'A' + hi - 10 : '0' + hi;
+    buffer[buffer_idx++] = (lo >= 10) ? 'A' + lo - 10 : '0' + lo;
+  }
+
+  // NULL terminate
+  buffer[(buffer_idx < buffer_length) ? buffer_idx : buffer_idx - 1] = 0;
+}
+
+}  // namespace lul
diff --git a/mozglue/baseprofiler/lul/LulElfExt.h b/mozglue/baseprofiler/lul/LulElfExt.h
new file mode 100644
index 0000000000..73d9ff7f15
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulElfExt.h
@@ -0,0 +1,69 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+// Copyright (c) 2006, 2011, 2012 Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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
+// OWNER OR CONTRIBUTORS 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 file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/linux/dump_symbols.h
+
+#ifndef LulElfExt_h
+#define LulElfExt_h
+
+// These two functions are the external interface to the
+// ELF/Dwarf/EXIDX reader.
+
+#include "LulMainInt.h"
+
+using lul::SecMap;
+
+namespace lul {
+
+class UniqueStringUniverse;
+
+// Find all the unwind information in OBJ_FILE, an ELF executable
+// or shared library, and add it to SMAP.
+bool ReadSymbolData(const std::string& obj_file,
+                    const std::vector<std::string>& debug_dirs, SecMap* smap,
+                    void* rx_avma, size_t rx_size, UniqueStringUniverse* usu,
+                    void (*log)(const char*));
+
+// The same as ReadSymbolData, except that OBJ_FILE is assumed to
+// point to a mapped-in image of OBJ_FILENAME.
+bool ReadSymbolDataInternal(const uint8_t* obj_file,
+                            const std::string& obj_filename,
+                            const std::vector<std::string>& debug_dirs,
+                            SecMap* smap, void* rx_avma, size_t rx_size,
+                            UniqueStringUniverse* usu,
+                            void (*log)(const char*));
+
+}  // namespace lul
+
+#endif  // LulElfExt_h
diff --git a/mozglue/baseprofiler/lul/LulElfInt.h b/mozglue/baseprofiler/lul/LulElfInt.h
new file mode 100644
index 0000000000..31ffba8ff0
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulElfInt.h
@@ -0,0 +1,218 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+// Copyright (c) 2006, 2012, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "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
+// OWNER OR CONTRIBUTORS 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 file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/android/include/elf.h
+//   src/common/linux/elfutils.h
+//   src/common/linux/file_id.h
+//   src/common/linux/elfutils-inl.h
+
+#ifndef LulElfInt_h
+#define LulElfInt_h
+
+// This header defines functions etc internal to the ELF reader.  It
+// should not be included outside of LulElf.cpp.
+
+#include <elf.h>
+#include <stdlib.h>
+
+#include "mozilla/Assertions.h"
+
+#include "PlatformMacros.h"
+
+// (derived from)
+// elfutils.h: Utilities for dealing with ELF files.
+//
+#include <link.h>
+
+#if defined(GP_OS_android)
+
+// From toolkit/crashreporter/google-breakpad/src/common/android/include/elf.h
+// The Android headers don't always define this constant.
+#  ifndef EM_X86_64
+#    define EM_X86_64 62
+#  endif
+
+#  ifndef EM_PPC64
+#    define EM_PPC64 21
+#  endif
+
+#  ifndef EM_S390
+#    define EM_S390 22
+#  endif
+
+#  ifndef NT_GNU_BUILD_ID
+#    define NT_GNU_BUILD_ID 3
+#  endif
+
+#  ifndef ElfW
+#    define ElfW(type) _ElfW(Elf, ELFSIZE, type)
+#    define _ElfW(e, w, t) _ElfW_1(e, w, _##t)
+#    define _ElfW_1(e, w, t) e##w##t
+#  endif
+
+#endif
+
+#if defined(GP_OS_freebsd)
+
+#  ifndef ElfW
+#    define ElfW(type) Elf_##type
+#  endif
+
+#endif
+
+namespace lul {
+
+// Traits classes so consumers can write templatized code to deal
+// with specific ELF bits.
+struct ElfClass32 {
+  typedef Elf32_Addr Addr;
+  typedef Elf32_Ehdr Ehdr;
+  typedef Elf32_Nhdr Nhdr;
+  typedef Elf32_Phdr Phdr;
+  typedef Elf32_Shdr Shdr;
+  typedef Elf32_Half Half;
+  typedef Elf32_Off Off;
+  typedef Elf32_Word Word;
+  static const int kClass = ELFCLASS32;
+  static const size_t kAddrSize = sizeof(Elf32_Addr);
+};
+
+struct ElfClass64 {
+  typedef Elf64_Addr Addr;
+  typedef Elf64_Ehdr Ehdr;
+  typedef Elf64_Nhdr Nhdr;
+  typedef Elf64_Phdr Phdr;
+  typedef Elf64_Shdr Shdr;
+  typedef Elf64_Half Half;
+  typedef Elf64_Off Off;
+  typedef Elf64_Word Word;
+  static const int kClass = ELFCLASS64;
+  static const size_t kAddrSize = sizeof(Elf64_Addr);
+};
+
+bool IsValidElf(const void* elf_header);
+int ElfClass(const void* elf_base);
+
+// Attempt to find a section named |section_name| of type |section_type|
+// in the ELF binary data at |elf_mapped_base|. On success, returns true
+// and sets |*section_start| to point to the start of the section data,
+// and |*section_size| to the size of the section's data. If |elfclass|
+// is not NULL, set |*elfclass| to the ELF file class.
+bool FindElfSection(const void* elf_mapped_base, const char* section_name,
+                    uint32_t section_type, const void** section_start,
+                    int* section_size, int* elfclass);
+
+// Internal helper method, exposed for convenience for callers
+// that already have more info.
+template <typename ElfClass>
+const typename ElfClass::Shdr* FindElfSectionByName(
+    const char* name, typename ElfClass::Word section_type,
+    const typename ElfClass::Shdr* sections, const char* section_names,
+    const char* names_end, int nsection);
+
+// Attempt to find the first segment of type |segment_type| in the ELF
+// binary data at |elf_mapped_base|. On success, returns true and sets
+// |*segment_start| to point to the start of the segment data, and
+// and |*segment_size| to the size of the segment's data. If |elfclass|
+// is not NULL, set |*elfclass| to the ELF file class.
+bool FindElfSegment(const void* elf_mapped_base, uint32_t segment_type,
+                    const void** segment_start, int* segment_size,
+                    int* elfclass);
+
+// Convert an offset from an Elf header into a pointer to the mapped
+// address in the current process. Takes an extra template parameter
+// to specify the return type to avoid having to dynamic_cast the
+// result.
+template <typename ElfClass, typename T>
+const T* GetOffset(const typename ElfClass::Ehdr* elf_header,
+                   typename ElfClass::Off offset);
+
+// (derived from)
+// file_id.h: Return a unique identifier for a file
+//
+
+static const size_t kMDGUIDSize = sizeof(MDGUID);
+
+class FileID {
+ public:
+  // Load the identifier for the elf file mapped into memory at |base| into
+  // |identifier|.  Return false if the identifier could not be created for the
+  // file.
+  static bool ElfFileIdentifierFromMappedFile(const void* base,
+                                              uint8_t identifier[kMDGUIDSize]);
+
+  // Convert the |identifier| data to a NULL terminated string.  The string will
+  // be formatted as a UUID (e.g., 22F065BB-FC9C-49F7-80FE-26A7CEBD7BCE).
+  // The |buffer| should be at least 37 bytes long to receive all of the data
+  // and termination.  Shorter buffers will contain truncated data.
+  static void ConvertIdentifierToString(const uint8_t identifier[kMDGUIDSize],
+                                        char* buffer, int buffer_length);
+};
+
+template <typename ElfClass, typename T>
+const T* GetOffset(const typename ElfClass::Ehdr* elf_header,
+                   typename ElfClass::Off offset) {
+  return reinterpret_cast<const T*>(reinterpret_cast<uintptr_t>(elf_header) +
+                                    offset);
+}
+
+template <typename ElfClass>
+const typename ElfClass::Shdr* FindElfSectionByName(
+    const char* name, typename ElfClass::Word section_type,
+    const typename ElfClass::Shdr* sections, const char* section_names,
+    const char* names_end, int nsection) {
+  MOZ_ASSERT(name != NULL);
+  MOZ_ASSERT(sections != NULL);
+  MOZ_ASSERT(nsection > 0);
+
+  int name_len = strlen(name);
+  if (name_len == 0) return NULL;
+
+  for (int i = 0; i < nsection; ++i) {
+    const char* section_name = section_names + sections[i].sh_name;
+    if (sections[i].sh_type == section_type &&
+        names_end - section_name >= name_len + 1 &&
+        strcmp(name, section_name) == 0) {
+      return sections + i;
+    }
+  }
+  return NULL;
+}
+
+}  // namespace lul
+
+// And finally, the external interface, offered to LulMain.cpp
+#include "LulElfExt.h"
+
+#endif  // LulElfInt_h
diff --git a/mozglue/baseprofiler/lul/LulMain.cpp b/mozglue/baseprofiler/lul/LulMain.cpp
new file mode 100644
index 0000000000..f513920a0b
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulMain.cpp
@@ -0,0 +1,1956 @@
+/* -*- 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 "LulMain.h"
+
+#include <string.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <unistd.h>  // write(), only for testing LUL
+
+#include <algorithm>  // std::sort
+#include <string>
+#include <utility>
+
+#include "mozilla/Assertions.h"
+#include "mozilla/ArrayUtils.h"
+#include "mozilla/CheckedInt.h"
+#include "mozilla/DebugOnly.h"
+#include "mozilla/MemoryChecking.h"
+#include "mozilla/Sprintf.h"
+#include "mozilla/UniquePtr.h"
+#include "mozilla/Unused.h"
+
+#include "BaseProfiler.h"
+#include "LulCommonExt.h"
+#include "LulElfExt.h"
+#include "LulMainInt.h"
+
+using mozilla::baseprofiler::profiler_current_process_id;
+using mozilla::baseprofiler::profiler_current_thread_id;
+
+// Set this to 1 for verbose logging
+#define DEBUG_MAIN 0
+
+namespace lul {
+
+using mozilla::CheckedInt;
+using mozilla::DebugOnly;
+using mozilla::MallocSizeOf;
+using mozilla::Unused;
+using std::pair;
+using std::string;
+using std::vector;
+
+// WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+//
+// Some functions in this file are marked RUNS IN NO-MALLOC CONTEXT.
+// Any such function -- and, hence, the transitive closure of those
+// reachable from it -- must not do any dynamic memory allocation.
+// Doing so risks deadlock.  There is exactly one root function for
+// the transitive closure: Lul::Unwind.
+//
+// WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING
+
+////////////////////////////////////////////////////////////////
+// RuleSet                                                    //
+////////////////////////////////////////////////////////////////
+
+static const char* NameOf_DW_REG(int16_t aReg) {
+  switch (aReg) {
+    case DW_REG_CFA:
+      return "cfa";
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+    case DW_REG_INTEL_XBP:
+      return "xbp";
+    case DW_REG_INTEL_XSP:
+      return "xsp";
+    case DW_REG_INTEL_XIP:
+      return "xip";
+#elif defined(GP_ARCH_arm)
+    case DW_REG_ARM_R7:
+      return "r7";
+    case DW_REG_ARM_R11:
+      return "r11";
+    case DW_REG_ARM_R12:
+      return "r12";
+    case DW_REG_ARM_R13:
+      return "r13";
+    case DW_REG_ARM_R14:
+      return "r14";
+    case DW_REG_ARM_R15:
+      return "r15";
+#elif defined(GP_ARCH_arm64)
+    case DW_REG_AARCH64_X29:
+      return "x29";
+    case DW_REG_AARCH64_X30:
+      return "x30";
+    case DW_REG_AARCH64_SP:
+      return "sp";
+#elif defined(GP_ARCH_mips64)
+    case DW_REG_MIPS_SP:
+      return "sp";
+    case DW_REG_MIPS_FP:
+      return "fp";
+    case DW_REG_MIPS_PC:
+      return "pc";
+#else
+#  error "Unsupported arch"
+#endif
+    default:
+      return "???";
+  }
+}
+
+string LExpr::ShowRule(const char* aNewReg) const {
+  char buf[64];
+  string res = string(aNewReg) + "=";
+  switch (mHow) {
+    case UNKNOWN:
+      res += "Unknown";
+      break;
+    case NODEREF:
+      SprintfLiteral(buf, "%s+%d", NameOf_DW_REG(mReg), (int)mOffset);
+      res += buf;
+      break;
+    case DEREF:
+      SprintfLiteral(buf, "*(%s+%d)", NameOf_DW_REG(mReg), (int)mOffset);
+      res += buf;
+      break;
+    case PFXEXPR:
+      SprintfLiteral(buf, "PfxExpr-at-%d", (int)mOffset);
+      res += buf;
+      break;
+    default:
+      res += "???";
+      break;
+  }
+  return res;
+}
+
+void RuleSet::Print(void (*aLog)(const char*)) const {
+  char buf[96];
+  SprintfLiteral(buf, "[%llx .. %llx]: let ", (unsigned long long int)mAddr,
+                 (unsigned long long int)(mAddr + mLen - 1));
+  string res = string(buf);
+  res += mCfaExpr.ShowRule("cfa");
+  res += " in";
+  // For each reg we care about, print the recovery expression.
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+  res += mXipExpr.ShowRule(" RA");
+  res += mXspExpr.ShowRule(" SP");
+  res += mXbpExpr.ShowRule(" BP");
+#elif defined(GP_ARCH_arm)
+  res += mR15expr.ShowRule(" R15");
+  res += mR7expr.ShowRule(" R7");
+  res += mR11expr.ShowRule(" R11");
+  res += mR12expr.ShowRule(" R12");
+  res += mR13expr.ShowRule(" R13");
+  res += mR14expr.ShowRule(" R14");
+#elif defined(GP_ARCH_arm64)
+  res += mX29expr.ShowRule(" X29");
+  res += mX30expr.ShowRule(" X30");
+  res += mSPexpr.ShowRule(" SP");
+#elif defined(GP_ARCH_mips64)
+  res += mPCexpr.ShowRule(" PC");
+  res += mSPexpr.ShowRule(" SP");
+  res += mFPexpr.ShowRule(" FP");
+#else
+#  error "Unsupported arch"
+#endif
+  aLog(res.c_str());
+}
+
+LExpr* RuleSet::ExprForRegno(DW_REG_NUMBER aRegno) {
+  switch (aRegno) {
+    case DW_REG_CFA:
+      return &mCfaExpr;
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+    case DW_REG_INTEL_XIP:
+      return &mXipExpr;
+    case DW_REG_INTEL_XSP:
+      return &mXspExpr;
+    case DW_REG_INTEL_XBP:
+      return &mXbpExpr;
+#elif defined(GP_ARCH_arm)
+    case DW_REG_ARM_R15:
+      return &mR15expr;
+    case DW_REG_ARM_R14:
+      return &mR14expr;
+    case DW_REG_ARM_R13:
+      return &mR13expr;
+    case DW_REG_ARM_R12:
+      return &mR12expr;
+    case DW_REG_ARM_R11:
+      return &mR11expr;
+    case DW_REG_ARM_R7:
+      return &mR7expr;
+#elif defined(GP_ARCH_arm64)
+    case DW_REG_AARCH64_X29:
+      return &mX29expr;
+    case DW_REG_AARCH64_X30:
+      return &mX30expr;
+    case DW_REG_AARCH64_SP:
+      return &mSPexpr;
+#elif defined(GP_ARCH_mips64)
+    case DW_REG_MIPS_SP:
+      return &mSPexpr;
+    case DW_REG_MIPS_FP:
+      return &mFPexpr;
+    case DW_REG_MIPS_PC:
+      return &mPCexpr;
+#else
+#  error "Unknown arch"
+#endif
+    default:
+      return nullptr;
+  }
+}
+
+RuleSet::RuleSet() {
+  mAddr = 0;
+  mLen = 0;
+  // The only other fields are of type LExpr and those are initialised
+  // by LExpr::LExpr().
+}
+
+////////////////////////////////////////////////////////////////
+// SecMap                                                     //
+////////////////////////////////////////////////////////////////
+
+// See header file LulMainInt.h for comments about invariants.
+
+SecMap::SecMap(void (*aLog)(const char*))
+    : mSummaryMinAddr(1), mSummaryMaxAddr(0), mUsable(true), mLog(aLog) {}
+
+SecMap::~SecMap() { mRuleSets.clear(); }
+
+// RUNS IN NO-MALLOC CONTEXT
+RuleSet* SecMap::FindRuleSet(uintptr_t ia) {
+  // Binary search mRuleSets to find one that brackets |ia|.
+  // lo and hi need to be signed, else the loop termination tests
+  // don't work properly.  Note that this works correctly even when
+  // mRuleSets.size() == 0.
+
+  // Can't do this until the array has been sorted and preened.
+  MOZ_ASSERT(mUsable);
+
+  long int lo = 0;
+  long int hi = (long int)mRuleSets.size() - 1;
+  while (true) {
+    // current unsearched space is from lo to hi, inclusive.
+    if (lo > hi) {
+      // not found
+      return nullptr;
+    }
+    long int mid = lo + ((hi - lo) / 2);
+    RuleSet* mid_ruleSet = &mRuleSets[mid];
+    uintptr_t mid_minAddr = mid_ruleSet->mAddr;
+    uintptr_t mid_maxAddr = mid_minAddr + mid_ruleSet->mLen - 1;
+    if (ia < mid_minAddr) {
+      hi = mid - 1;
+      continue;
+    }
+    if (ia > mid_maxAddr) {
+      lo = mid + 1;
+      continue;
+    }
+    MOZ_ASSERT(mid_minAddr <= ia && ia <= mid_maxAddr);
+    return mid_ruleSet;
+  }
+  // NOTREACHED
+}
+
+// Add a RuleSet to the collection.  The rule is copied in.  Calling
+// this makes the map non-searchable.
+void SecMap::AddRuleSet(const RuleSet* rs) {
+  mUsable = false;
+  mRuleSets.push_back(*rs);
+}
+
+// Add a PfxInstr to the vector of such instrs, and return the index
+// in the vector.  Calling this makes the map non-searchable.
+uint32_t SecMap::AddPfxInstr(PfxInstr pfxi) {
+  mUsable = false;
+  mPfxInstrs.push_back(pfxi);
+  return mPfxInstrs.size() - 1;
+}
+
+static bool CmpRuleSetsByAddrLE(const RuleSet& rs1, const RuleSet& rs2) {
+  return rs1.mAddr < rs2.mAddr;
+}
+
+// Prepare the map for searching.  Completely remove any which don't
+// fall inside the specified range [start, +len).
+void SecMap::PrepareRuleSets(uintptr_t aStart, size_t aLen) {
+  if (mRuleSets.empty()) {
+    return;
+  }
+
+  MOZ_ASSERT(aLen > 0);
+  if (aLen == 0) {
+    // This should never happen.
+    mRuleSets.clear();
+    return;
+  }
+
+  // Sort by start addresses.
+  std::sort(mRuleSets.begin(), mRuleSets.end(), CmpRuleSetsByAddrLE);
+
+  // Detect any entry not completely contained within [start, +len).
+  // Set its length to zero, so that the next pass will remove it.
+  for (size_t i = 0; i < mRuleSets.size(); ++i) {
+    RuleSet* rs = &mRuleSets[i];
+    if (rs->mLen > 0 &&
+        (rs->mAddr < aStart || rs->mAddr + rs->mLen > aStart + aLen)) {
+      rs->mLen = 0;
+    }
+  }
+
+  // Iteratively truncate any overlaps and remove any zero length
+  // entries that might result, or that may have been present
+  // initially.  Unless the input is seriously screwy, this is
+  // expected to iterate only once.
+  while (true) {
+    size_t i;
+    size_t n = mRuleSets.size();
+    size_t nZeroLen = 0;
+
+    if (n == 0) {
+      break;
+    }
+
+    for (i = 1; i < n; ++i) {
+      RuleSet* prev = &mRuleSets[i - 1];
+      RuleSet* here = &mRuleSets[i];
+      MOZ_ASSERT(prev->mAddr <= here->mAddr);
+      if (prev->mAddr + prev->mLen > here->mAddr) {
+        prev->mLen = here->mAddr - prev->mAddr;
+      }
+      if (prev->mLen == 0) nZeroLen++;
+    }
+
+    if (mRuleSets[n - 1].mLen == 0) {
+      nZeroLen++;
+    }
+
+    // At this point, the entries are in-order and non-overlapping.
+    // If none of them are zero-length, we are done.
+    if (nZeroLen == 0) {
+      break;
+    }
+
+    // Slide back the entries to remove the zero length ones.
+    size_t j = 0;  // The write-point.
+    for (i = 0; i < n; ++i) {
+      if (mRuleSets[i].mLen == 0) {
+        continue;
+      }
+      if (j != i) mRuleSets[j] = mRuleSets[i];
+      ++j;
+    }
+    MOZ_ASSERT(i == n);
+    MOZ_ASSERT(nZeroLen <= n);
+    MOZ_ASSERT(j == n - nZeroLen);
+    while (nZeroLen > 0) {
+      mRuleSets.pop_back();
+      nZeroLen--;
+    }
+
+    MOZ_ASSERT(mRuleSets.size() == j);
+  }
+
+  size_t n = mRuleSets.size();
+
+#ifdef DEBUG
+  // Do a final check on the rules: their address ranges must be
+  // ascending, non overlapping, non zero sized.
+  if (n > 0) {
+    MOZ_ASSERT(mRuleSets[0].mLen > 0);
+    for (size_t i = 1; i < n; ++i) {
+      RuleSet* prev = &mRuleSets[i - 1];
+      RuleSet* here = &mRuleSets[i];
+      MOZ_ASSERT(prev->mAddr < here->mAddr);
+      MOZ_ASSERT(here->mLen > 0);
+      MOZ_ASSERT(prev->mAddr + prev->mLen <= here->mAddr);
+    }
+  }
+#endif
+
+  // Set the summary min and max address values.
+  if (n == 0) {
+    // Use the values defined in comments in the class declaration.
+    mSummaryMinAddr = 1;
+    mSummaryMaxAddr = 0;
+  } else {
+    mSummaryMinAddr = mRuleSets[0].mAddr;
+    mSummaryMaxAddr = mRuleSets[n - 1].mAddr + mRuleSets[n - 1].mLen - 1;
+  }
+  char buf[150];
+  SprintfLiteral(buf, "PrepareRuleSets: %d entries, smin/smax 0x%llx, 0x%llx\n",
+                 (int)n, (unsigned long long int)mSummaryMinAddr,
+                 (unsigned long long int)mSummaryMaxAddr);
+  buf[sizeof(buf) - 1] = 0;
+  mLog(buf);
+
+  // Is now usable for binary search.
+  mUsable = true;
+
+#if 0
+  mLog("\nRulesets after preening\n");
+  for (size_t i = 0; i < mRuleSets.size(); ++i) {
+    mRuleSets[i].Print(mLog);
+    mLog("\n");
+  }
+  mLog("\n");
+#endif
+}
+
+bool SecMap::IsEmpty() { return mRuleSets.empty(); }
+
+size_t SecMap::SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+  size_t n = aMallocSizeOf(this);
+
+  // It's conceivable that these calls would be unsafe with some
+  // implementations of std::vector, but it seems to be working for now...
+  n += aMallocSizeOf(mRuleSets.data());
+  n += aMallocSizeOf(mPfxInstrs.data());
+
+  return n;
+}
+
+////////////////////////////////////////////////////////////////
+// SegArray                                                   //
+////////////////////////////////////////////////////////////////
+
+// A SegArray holds a set of address ranges that together exactly
+// cover an address range, with no overlaps or holes.  Each range has
+// an associated value, which in this case has been specialised to be
+// a simple boolean.  The representation is kept to minimal canonical
+// form in which adjacent ranges with the same associated value are
+// merged together.  Each range is represented by a |struct Seg|.
+//
+// SegArrays are used to keep track of which parts of the address
+// space are known to contain instructions.
+class SegArray {
+ public:
+  void add(uintptr_t lo, uintptr_t hi, bool val) {
+    if (lo > hi) {
+      return;
+    }
+    split_at(lo);
+    if (hi < UINTPTR_MAX) {
+      split_at(hi + 1);
+    }
+    std::vector<Seg>::size_type iLo, iHi, i;
+    iLo = find(lo);
+    iHi = find(hi);
+    for (i = iLo; i <= iHi; ++i) {
+      mSegs[i].val = val;
+    }
+    preen();
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  bool getBoundingCodeSegment(/*OUT*/ uintptr_t* rx_min,
+                              /*OUT*/ uintptr_t* rx_max, uintptr_t addr) {
+    std::vector<Seg>::size_type i = find(addr);
+    if (!mSegs[i].val) {
+      return false;
+    }
+    *rx_min = mSegs[i].lo;
+    *rx_max = mSegs[i].hi;
+    return true;
+  }
+
+  SegArray() {
+    Seg s(0, UINTPTR_MAX, false);
+    mSegs.push_back(s);
+  }
+
+ private:
+  struct Seg {
+    Seg(uintptr_t lo, uintptr_t hi, bool val) : lo(lo), hi(hi), val(val) {}
+    uintptr_t lo;
+    uintptr_t hi;
+    bool val;
+  };
+
+  void preen() {
+    for (std::vector<Seg>::iterator iter = mSegs.begin();
+         iter < mSegs.end() - 1; ++iter) {
+      if (iter[0].val != iter[1].val) {
+        continue;
+      }
+      iter[0].hi = iter[1].hi;
+      mSegs.erase(iter + 1);
+      // Back up one, so as not to miss an opportunity to merge
+      // with the entry after this one.
+      --iter;
+    }
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  std::vector<Seg>::size_type find(uintptr_t a) {
+    long int lo = 0;
+    long int hi = (long int)mSegs.size();
+    while (true) {
+      // The unsearched space is lo .. hi inclusive.
+      if (lo > hi) {
+        // Not found.  This can't happen.
+        return (std::vector<Seg>::size_type)(-1);
+      }
+      long int mid = lo + ((hi - lo) / 2);
+      uintptr_t mid_lo = mSegs[mid].lo;
+      uintptr_t mid_hi = mSegs[mid].hi;
+      if (a < mid_lo) {
+        hi = mid - 1;
+        continue;
+      }
+      if (a > mid_hi) {
+        lo = mid + 1;
+        continue;
+      }
+      return (std::vector<Seg>::size_type)mid;
+    }
+  }
+
+  void split_at(uintptr_t a) {
+    std::vector<Seg>::size_type i = find(a);
+    if (mSegs[i].lo == a) {
+      return;
+    }
+    mSegs.insert(mSegs.begin() + i + 1, mSegs[i]);
+    mSegs[i].hi = a - 1;
+    mSegs[i + 1].lo = a;
+  }
+
+  void show() {
+    printf("<< %d entries:\n", (int)mSegs.size());
+    for (std::vector<Seg>::iterator iter = mSegs.begin(); iter < mSegs.end();
+         ++iter) {
+      printf("  %016llx  %016llx  %s\n", (unsigned long long int)(*iter).lo,
+             (unsigned long long int)(*iter).hi,
+             (*iter).val ? "true" : "false");
+    }
+    printf(">>\n");
+  }
+
+  std::vector<Seg> mSegs;
+};
+
+////////////////////////////////////////////////////////////////
+// PriMap                                                     //
+////////////////////////////////////////////////////////////////
+
+class PriMap {
+ public:
+  explicit PriMap(void (*aLog)(const char*)) : mLog(aLog) {}
+
+  // RUNS IN NO-MALLOC CONTEXT
+  pair<const RuleSet*, const vector<PfxInstr>*> Lookup(uintptr_t ia) {
+    SecMap* sm = FindSecMap(ia);
+    return pair<const RuleSet*, const vector<PfxInstr>*>(
+        sm ? sm->FindRuleSet(ia) : nullptr, sm ? sm->GetPfxInstrs() : nullptr);
+  }
+
+  // Add a secondary map.  No overlaps allowed w.r.t. existing
+  // secondary maps.
+  void AddSecMap(mozilla::UniquePtr<SecMap>&& aSecMap) {
+    // We can't add an empty SecMap to the PriMap.  But that's OK
+    // since we'd never be able to find anything in it anyway.
+    if (aSecMap->IsEmpty()) {
+      return;
+    }
+
+    // Iterate through the SecMaps and find the right place for this
+    // one.  At the same time, ensure that the in-order
+    // non-overlapping invariant is preserved (and, generally, holds).
+    // FIXME: this gives a cost that is O(N^2) in the total number of
+    // shared objects in the system.  ToDo: better.
+    MOZ_ASSERT(aSecMap->mSummaryMinAddr <= aSecMap->mSummaryMaxAddr);
+
+    size_t num_secMaps = mSecMaps.size();
+    uintptr_t i;
+    for (i = 0; i < num_secMaps; ++i) {
+      mozilla::UniquePtr<SecMap>& sm_i = mSecMaps[i];
+      MOZ_ASSERT(sm_i->mSummaryMinAddr <= sm_i->mSummaryMaxAddr);
+      if (aSecMap->mSummaryMinAddr < sm_i->mSummaryMaxAddr) {
+        // |aSecMap| needs to be inserted immediately before mSecMaps[i].
+        break;
+      }
+    }
+    MOZ_ASSERT(i <= num_secMaps);
+    if (i == num_secMaps) {
+      // It goes at the end.
+      mSecMaps.push_back(std::move(aSecMap));
+    } else {
+      std::vector<mozilla::UniquePtr<SecMap>>::iterator iter =
+          mSecMaps.begin() + i;
+      mSecMaps.insert(iter, std::move(aSecMap));
+    }
+    char buf[100];
+    SprintfLiteral(buf, "AddSecMap: now have %d SecMaps\n",
+                   (int)mSecMaps.size());
+    buf[sizeof(buf) - 1] = 0;
+    mLog(buf);
+  }
+
+  // Remove and delete any SecMaps in the mapping, that intersect
+  // with the specified address range.
+  void RemoveSecMapsInRange(uintptr_t avma_min, uintptr_t avma_max) {
+    MOZ_ASSERT(avma_min <= avma_max);
+    size_t num_secMaps = mSecMaps.size();
+    if (num_secMaps > 0) {
+      intptr_t i;
+      // Iterate from end to start over the vector, so as to ensure
+      // that the special case where |avma_min| and |avma_max| denote
+      // the entire address space, can be completed in time proportional
+      // to the number of elements in the map.
+      for (i = (intptr_t)num_secMaps - 1; i >= 0; i--) {
+        mozilla::UniquePtr<SecMap>& sm_i = mSecMaps[i];
+        if (sm_i->mSummaryMaxAddr < avma_min ||
+            avma_max < sm_i->mSummaryMinAddr) {
+          // There's no overlap.  Move on.
+          continue;
+        }
+        // We need to remove mSecMaps[i] and slide all those above it
+        // downwards to cover the hole.
+        mSecMaps.erase(mSecMaps.begin() + i);
+      }
+    }
+  }
+
+  // Return the number of currently contained SecMaps.
+  size_t CountSecMaps() { return mSecMaps.size(); }
+
+  size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+    size_t n = aMallocSizeOf(this);
+
+    // It's conceivable that this call would be unsafe with some
+    // implementations of std::vector, but it seems to be working for now...
+    n += aMallocSizeOf(mSecMaps.data());
+
+    for (size_t i = 0; i < mSecMaps.size(); i++) {
+      n += mSecMaps[i]->SizeOfIncludingThis(aMallocSizeOf);
+    }
+
+    return n;
+  }
+
+ private:
+  // RUNS IN NO-MALLOC CONTEXT
+  SecMap* FindSecMap(uintptr_t ia) {
+    // Binary search mSecMaps to find one that brackets |ia|.
+    // lo and hi need to be signed, else the loop termination tests
+    // don't work properly.
+    long int lo = 0;
+    long int hi = (long int)mSecMaps.size() - 1;
+    while (true) {
+      // current unsearched space is from lo to hi, inclusive.
+      if (lo > hi) {
+        // not found
+        return nullptr;
+      }
+      long int mid = lo + ((hi - lo) / 2);
+      mozilla::UniquePtr<SecMap>& mid_secMap = mSecMaps[mid];
+      uintptr_t mid_minAddr = mid_secMap->mSummaryMinAddr;
+      uintptr_t mid_maxAddr = mid_secMap->mSummaryMaxAddr;
+      if (ia < mid_minAddr) {
+        hi = mid - 1;
+        continue;
+      }
+      if (ia > mid_maxAddr) {
+        lo = mid + 1;
+        continue;
+      }
+      MOZ_ASSERT(mid_minAddr <= ia && ia <= mid_maxAddr);
+      return mid_secMap.get();
+    }
+    // NOTREACHED
+  }
+
+ private:
+  // sorted array of per-object ranges, non overlapping, non empty
+  std::vector<mozilla::UniquePtr<SecMap>> mSecMaps;
+
+  // a logging sink, for debugging.
+  void (*mLog)(const char*);
+};
+
+////////////////////////////////////////////////////////////////
+// LUL                                                        //
+////////////////////////////////////////////////////////////////
+
+#define LUL_LOG(_str)                                           \
+  do {                                                          \
+    char buf[200];                                              \
+    SprintfLiteral(buf, "LUL: pid %d tid %d lul-obj %p: %s",    \
+                   profiler_current_process_id(),               \
+                   profiler_current_thread_id(), this, (_str)); \
+    buf[sizeof(buf) - 1] = 0;                                   \
+    mLog(buf);                                                  \
+  } while (0)
+
+LUL::LUL(void (*aLog)(const char*))
+    : mLog(aLog),
+      mAdminMode(true),
+      mAdminThreadId(profiler_current_thread_id()),
+      mPriMap(new PriMap(aLog)),
+      mSegArray(new SegArray()),
+      mUSU(new UniqueStringUniverse()) {
+  LUL_LOG("LUL::LUL: Created object");
+}
+
+LUL::~LUL() {
+  LUL_LOG("LUL::~LUL: Destroyed object");
+  delete mPriMap;
+  delete mSegArray;
+  mLog = nullptr;
+  delete mUSU;
+}
+
+void LUL::MaybeShowStats() {
+  // This is racey in the sense that it can't guarantee that
+  //   n_new == n_new_Context + n_new_CFI + n_new_Scanned
+  // if it should happen that mStats is updated by some other thread
+  // in between computation of n_new and n_new_{Context,CFI,FP}.
+  // But it's just stats printing, so we don't really care.
+  uint32_t n_new = mStats - mStatsPrevious;
+  if (n_new >= 5000) {
+    uint32_t n_new_Context = mStats.mContext - mStatsPrevious.mContext;
+    uint32_t n_new_CFI = mStats.mCFI - mStatsPrevious.mCFI;
+    uint32_t n_new_FP = mStats.mFP - mStatsPrevious.mFP;
+    mStatsPrevious = mStats;
+    char buf[200];
+    SprintfLiteral(buf,
+                   "LUL frame stats: TOTAL %5u"
+                   "    CTX %4u    CFI %4u    FP %4u",
+                   n_new, n_new_Context, n_new_CFI, n_new_FP);
+    buf[sizeof(buf) - 1] = 0;
+    mLog(buf);
+  }
+}
+
+size_t LUL::SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+  size_t n = aMallocSizeOf(this);
+  n += mPriMap->SizeOfIncludingThis(aMallocSizeOf);
+
+  // Measurement of the following members may be added later if DMD finds it
+  // is worthwhile:
+  // - mSegArray
+  // - mUSU
+
+  return n;
+}
+
+void LUL::EnableUnwinding() {
+  LUL_LOG("LUL::EnableUnwinding");
+  // Don't assert for Admin mode here.  That is, tolerate a call here
+  // if we are already in Unwinding mode.
+  MOZ_RELEASE_ASSERT(profiler_current_thread_id() == mAdminThreadId);
+
+  mAdminMode = false;
+}
+
+void LUL::NotifyAfterMap(uintptr_t aRXavma, size_t aSize, const char* aFileName,
+                         const void* aMappedImage) {
+  MOZ_RELEASE_ASSERT(mAdminMode);
+  MOZ_RELEASE_ASSERT(profiler_current_thread_id() == mAdminThreadId);
+
+  mLog(":\n");
+  char buf[200];
+  SprintfLiteral(buf, "NotifyMap %llx %llu %s\n",
+                 (unsigned long long int)aRXavma, (unsigned long long int)aSize,
+                 aFileName);
+  buf[sizeof(buf) - 1] = 0;
+  mLog(buf);
+
+  // Ignore obviously-stupid notifications.
+  if (aSize > 0) {
+    // Here's a new mapping, for this object.
+    mozilla::UniquePtr<SecMap> smap = mozilla::MakeUnique<SecMap>(mLog);
+
+    // Read CFI or EXIDX unwind data into |smap|.
+    if (!aMappedImage) {
+      (void)lul::ReadSymbolData(string(aFileName), std::vector<string>(),
+                                smap.get(), (void*)aRXavma, aSize, mUSU, mLog);
+    } else {
+      (void)lul::ReadSymbolDataInternal(
+          (const uint8_t*)aMappedImage, string(aFileName),
+          std::vector<string>(), smap.get(), (void*)aRXavma, aSize, mUSU, mLog);
+    }
+
+    mLog("NotifyMap .. preparing entries\n");
+
+    smap->PrepareRuleSets(aRXavma, aSize);
+
+    SprintfLiteral(buf, "NotifyMap got %lld entries\n",
+                   (long long int)smap->Size());
+    buf[sizeof(buf) - 1] = 0;
+    mLog(buf);
+
+    // Add it to the primary map (the top level set of mapped objects).
+    mPriMap->AddSecMap(std::move(smap));
+
+    // Tell the segment array about the mapping, so that the stack
+    // scan and __kernel_syscall mechanisms know where valid code is.
+    mSegArray->add(aRXavma, aRXavma + aSize - 1, true);
+  }
+}
+
+void LUL::NotifyExecutableArea(uintptr_t aRXavma, size_t aSize) {
+  MOZ_RELEASE_ASSERT(mAdminMode);
+  MOZ_RELEASE_ASSERT(profiler_current_thread_id() == mAdminThreadId);
+
+  mLog(":\n");
+  char buf[200];
+  SprintfLiteral(buf, "NotifyExecutableArea %llx %llu\n",
+                 (unsigned long long int)aRXavma,
+                 (unsigned long long int)aSize);
+  buf[sizeof(buf) - 1] = 0;
+  mLog(buf);
+
+  // Ignore obviously-stupid notifications.
+  if (aSize > 0) {
+    // Tell the segment array about the mapping, so that the stack
+    // scan and __kernel_syscall mechanisms know where valid code is.
+    mSegArray->add(aRXavma, aRXavma + aSize - 1, true);
+  }
+}
+
+void LUL::NotifyBeforeUnmap(uintptr_t aRXavmaMin, uintptr_t aRXavmaMax) {
+  MOZ_RELEASE_ASSERT(mAdminMode);
+  MOZ_RELEASE_ASSERT(profiler_current_thread_id() == mAdminThreadId);
+
+  mLog(":\n");
+  char buf[100];
+  SprintfLiteral(buf, "NotifyUnmap %016llx-%016llx\n",
+                 (unsigned long long int)aRXavmaMin,
+                 (unsigned long long int)aRXavmaMax);
+  buf[sizeof(buf) - 1] = 0;
+  mLog(buf);
+
+  MOZ_ASSERT(aRXavmaMin <= aRXavmaMax);
+
+  // Remove from the primary map, any secondary maps that intersect
+  // with the address range.  Also delete the secondary maps.
+  mPriMap->RemoveSecMapsInRange(aRXavmaMin, aRXavmaMax);
+
+  // Tell the segment array that the address range no longer
+  // contains valid code.
+  mSegArray->add(aRXavmaMin, aRXavmaMax, false);
+
+  SprintfLiteral(buf, "NotifyUnmap: now have %d SecMaps\n",
+                 (int)mPriMap->CountSecMaps());
+  buf[sizeof(buf) - 1] = 0;
+  mLog(buf);
+}
+
+size_t LUL::CountMappings() {
+  MOZ_RELEASE_ASSERT(mAdminMode);
+  MOZ_RELEASE_ASSERT(profiler_current_thread_id() == mAdminThreadId);
+
+  return mPriMap->CountSecMaps();
+}
+
+// RUNS IN NO-MALLOC CONTEXT
+static TaggedUWord DerefTUW(TaggedUWord aAddr, const StackImage* aStackImg) {
+  if (!aAddr.Valid()) {
+    return TaggedUWord();
+  }
+
+  // Lower limit check.  |aAddr.Value()| is the lowest requested address
+  // and |aStackImg->mStartAvma| is the lowest address we actually have,
+  // so the comparison is straightforward.
+  if (aAddr.Value() < aStackImg->mStartAvma) {
+    return TaggedUWord();
+  }
+
+  // Upper limit check.  We must compute the highest requested address
+  // and the highest address we actually have, but being careful to
+  // avoid overflow.  In particular if |aAddr| is 0xFFF...FFF or the
+  // 3/7 values below that, then we will get overflow.  See bug #1245477.
+  typedef CheckedInt<uintptr_t> CheckedUWord;
+  CheckedUWord highest_requested_plus_one =
+      CheckedUWord(aAddr.Value()) + CheckedUWord(sizeof(uintptr_t));
+  CheckedUWord highest_available_plus_one =
+      CheckedUWord(aStackImg->mStartAvma) + CheckedUWord(aStackImg->mLen);
+  if (!highest_requested_plus_one.isValid()     // overflow?
+      || !highest_available_plus_one.isValid()  // overflow?
+      || (highest_requested_plus_one.value() >
+          highest_available_plus_one.value())) {  // in range?
+    return TaggedUWord();
+  }
+
+  return TaggedUWord(
+      *(uintptr_t*)(&aStackImg
+                         ->mContents[aAddr.Value() - aStackImg->mStartAvma]));
+}
+
+// RUNS IN NO-MALLOC CONTEXT
+static TaggedUWord EvaluateReg(int16_t aReg, const UnwindRegs* aOldRegs,
+                               TaggedUWord aCFA) {
+  switch (aReg) {
+    case DW_REG_CFA:
+      return aCFA;
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+    case DW_REG_INTEL_XBP:
+      return aOldRegs->xbp;
+    case DW_REG_INTEL_XSP:
+      return aOldRegs->xsp;
+    case DW_REG_INTEL_XIP:
+      return aOldRegs->xip;
+#elif defined(GP_ARCH_arm)
+    case DW_REG_ARM_R7:
+      return aOldRegs->r7;
+    case DW_REG_ARM_R11:
+      return aOldRegs->r11;
+    case DW_REG_ARM_R12:
+      return aOldRegs->r12;
+    case DW_REG_ARM_R13:
+      return aOldRegs->r13;
+    case DW_REG_ARM_R14:
+      return aOldRegs->r14;
+    case DW_REG_ARM_R15:
+      return aOldRegs->r15;
+#elif defined(GP_ARCH_arm64)
+    case DW_REG_AARCH64_X29:
+      return aOldRegs->x29;
+    case DW_REG_AARCH64_X30:
+      return aOldRegs->x30;
+    case DW_REG_AARCH64_SP:
+      return aOldRegs->sp;
+#elif defined(GP_ARCH_mips64)
+    case DW_REG_MIPS_SP:
+      return aOldRegs->sp;
+    case DW_REG_MIPS_FP:
+      return aOldRegs->fp;
+    case DW_REG_MIPS_PC:
+      return aOldRegs->pc;
+#else
+#  error "Unsupported arch"
+#endif
+    default:
+      MOZ_ASSERT(0);
+      return TaggedUWord();
+  }
+}
+
+// RUNS IN NO-MALLOC CONTEXT
+// See prototype for comment.
+TaggedUWord EvaluatePfxExpr(int32_t start, const UnwindRegs* aOldRegs,
+                            TaggedUWord aCFA, const StackImage* aStackImg,
+                            const vector<PfxInstr>& aPfxInstrs) {
+  // A small evaluation stack, and a stack pointer, which points to
+  // the highest numbered in-use element.
+  const int N_STACK = 10;
+  TaggedUWord stack[N_STACK];
+  int stackPointer = -1;
+  for (int i = 0; i < N_STACK; i++) stack[i] = TaggedUWord();
+
+#define PUSH(_tuw)                                             \
+  do {                                                         \
+    if (stackPointer >= N_STACK - 1) goto fail; /* overflow */ \
+    stack[++stackPointer] = (_tuw);                            \
+  } while (0)
+
+#define POP(_lval)                                   \
+  do {                                               \
+    if (stackPointer < 0) goto fail; /* underflow */ \
+    _lval = stack[stackPointer--];                   \
+  } while (0)
+
+  // Cursor in the instruction sequence.
+  size_t curr = start + 1;
+
+  // Check the start point is sane.
+  size_t nInstrs = aPfxInstrs.size();
+  if (start < 0 || (size_t)start >= nInstrs) goto fail;
+
+  {
+    // The instruction sequence must start with PX_Start.  If not,
+    // something is seriously wrong.
+    PfxInstr first = aPfxInstrs[start];
+    if (first.mOpcode != PX_Start) goto fail;
+
+    // Push the CFA on the stack to start with (or not), as required by
+    // the original DW_OP_*expression* CFI.
+    if (first.mOperand != 0) PUSH(aCFA);
+  }
+
+  while (true) {
+    if (curr >= nInstrs) goto fail;  // ran off the end of the sequence
+
+    PfxInstr pfxi = aPfxInstrs[curr++];
+    if (pfxi.mOpcode == PX_End) break;  // we're done
+
+    switch (pfxi.mOpcode) {
+      case PX_Start:
+        // This should appear only at the start of the sequence.
+        goto fail;
+      case PX_End:
+        // We just took care of that, so we shouldn't see it again.
+        MOZ_ASSERT(0);
+        goto fail;
+      case PX_SImm32:
+        PUSH(TaggedUWord((intptr_t)pfxi.mOperand));
+        break;
+      case PX_DwReg: {
+        DW_REG_NUMBER reg = (DW_REG_NUMBER)pfxi.mOperand;
+        MOZ_ASSERT(reg != DW_REG_CFA);
+        PUSH(EvaluateReg(reg, aOldRegs, aCFA));
+        break;
+      }
+      case PX_Deref: {
+        TaggedUWord addr;
+        POP(addr);
+        PUSH(DerefTUW(addr, aStackImg));
+        break;
+      }
+      case PX_Add: {
+        TaggedUWord x, y;
+        POP(x);
+        POP(y);
+        PUSH(y + x);
+        break;
+      }
+      case PX_Sub: {
+        TaggedUWord x, y;
+        POP(x);
+        POP(y);
+        PUSH(y - x);
+        break;
+      }
+      case PX_And: {
+        TaggedUWord x, y;
+        POP(x);
+        POP(y);
+        PUSH(y & x);
+        break;
+      }
+      case PX_Or: {
+        TaggedUWord x, y;
+        POP(x);
+        POP(y);
+        PUSH(y | x);
+        break;
+      }
+      case PX_CmpGES: {
+        TaggedUWord x, y;
+        POP(x);
+        POP(y);
+        PUSH(y.CmpGEs(x));
+        break;
+      }
+      case PX_Shl: {
+        TaggedUWord x, y;
+        POP(x);
+        POP(y);
+        PUSH(y << x);
+        break;
+      }
+      default:
+        MOZ_ASSERT(0);
+        goto fail;
+    }
+  }  // while (true)
+
+  // Evaluation finished.  The top value on the stack is the result.
+  if (stackPointer >= 0) {
+    return stack[stackPointer];
+  }
+  // Else fall through
+
+fail:
+  return TaggedUWord();
+
+#undef PUSH
+#undef POP
+}
+
+// RUNS IN NO-MALLOC CONTEXT
+TaggedUWord LExpr::EvaluateExpr(const UnwindRegs* aOldRegs, TaggedUWord aCFA,
+                                const StackImage* aStackImg,
+                                const vector<PfxInstr>* aPfxInstrs) const {
+  switch (mHow) {
+    case UNKNOWN:
+      return TaggedUWord();
+    case NODEREF: {
+      TaggedUWord tuw = EvaluateReg(mReg, aOldRegs, aCFA);
+      tuw = tuw + TaggedUWord((intptr_t)mOffset);
+      return tuw;
+    }
+    case DEREF: {
+      TaggedUWord tuw = EvaluateReg(mReg, aOldRegs, aCFA);
+      tuw = tuw + TaggedUWord((intptr_t)mOffset);
+      return DerefTUW(tuw, aStackImg);
+    }
+    case PFXEXPR: {
+      MOZ_ASSERT(aPfxInstrs);
+      if (!aPfxInstrs) {
+        return TaggedUWord();
+      }
+      return EvaluatePfxExpr(mOffset, aOldRegs, aCFA, aStackImg, *aPfxInstrs);
+    }
+    default:
+      MOZ_ASSERT(0);
+      return TaggedUWord();
+  }
+}
+
+// RUNS IN NO-MALLOC CONTEXT
+static void UseRuleSet(/*MOD*/ UnwindRegs* aRegs, const StackImage* aStackImg,
+                       const RuleSet* aRS, const vector<PfxInstr>* aPfxInstrs) {
+  // Take a copy of regs, since we'll need to refer to the old values
+  // whilst computing the new ones.
+  UnwindRegs old_regs = *aRegs;
+
+  // Mark all the current register values as invalid, so that the
+  // caller can see, on our return, which ones have been computed
+  // anew.  If we don't even manage to compute a new PC value, then
+  // the caller will have to abandon the unwind.
+  // FIXME: Create and use instead: aRegs->SetAllInvalid();
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+  aRegs->xbp = TaggedUWord();
+  aRegs->xsp = TaggedUWord();
+  aRegs->xip = TaggedUWord();
+#elif defined(GP_ARCH_arm)
+  aRegs->r7 = TaggedUWord();
+  aRegs->r11 = TaggedUWord();
+  aRegs->r12 = TaggedUWord();
+  aRegs->r13 = TaggedUWord();
+  aRegs->r14 = TaggedUWord();
+  aRegs->r15 = TaggedUWord();
+#elif defined(GP_ARCH_arm64)
+  aRegs->x29 = TaggedUWord();
+  aRegs->x30 = TaggedUWord();
+  aRegs->sp = TaggedUWord();
+  aRegs->pc = TaggedUWord();
+#elif defined(GP_ARCH_mips64)
+  aRegs->sp = TaggedUWord();
+  aRegs->fp = TaggedUWord();
+  aRegs->pc = TaggedUWord();
+#else
+#  error "Unsupported arch"
+#endif
+
+  // This is generally useful.
+  const TaggedUWord inval = TaggedUWord();
+
+  // First, compute the CFA.
+  TaggedUWord cfa = aRS->mCfaExpr.EvaluateExpr(&old_regs, inval /*old cfa*/,
+                                               aStackImg, aPfxInstrs);
+
+  // If we didn't manage to compute the CFA, well .. that's ungood,
+  // but keep going anyway.  It'll be OK provided none of the register
+  // value rules mention the CFA.  In any case, compute the new values
+  // for each register that we're tracking.
+
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+  aRegs->xbp =
+      aRS->mXbpExpr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->xsp =
+      aRS->mXspExpr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->xip =
+      aRS->mXipExpr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+#elif defined(GP_ARCH_arm)
+  aRegs->r7 = aRS->mR7expr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->r11 =
+      aRS->mR11expr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->r12 =
+      aRS->mR12expr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->r13 =
+      aRS->mR13expr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->r14 =
+      aRS->mR14expr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->r15 =
+      aRS->mR15expr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+#elif defined(GP_ARCH_arm64)
+  aRegs->x29 =
+      aRS->mX29expr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->x30 =
+      aRS->mX30expr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->sp = aRS->mSPexpr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+#elif defined(GP_ARCH_mips64)
+  aRegs->sp = aRS->mSPexpr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->fp = aRS->mFPexpr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+  aRegs->pc = aRS->mPCexpr.EvaluateExpr(&old_regs, cfa, aStackImg, aPfxInstrs);
+#else
+#  error "Unsupported arch"
+#endif
+
+  // We're done.  Any regs for which we didn't manage to compute a
+  // new value will now be marked as invalid.
+}
+
+// RUNS IN NO-MALLOC CONTEXT
+void LUL::Unwind(/*OUT*/ uintptr_t* aFramePCs,
+                 /*OUT*/ uintptr_t* aFrameSPs,
+                 /*OUT*/ size_t* aFramesUsed,
+                 /*OUT*/ size_t* aFramePointerFramesAcquired,
+                 size_t aFramesAvail, UnwindRegs* aStartRegs,
+                 StackImage* aStackImg) {
+  MOZ_RELEASE_ASSERT(!mAdminMode);
+
+  /////////////////////////////////////////////////////////
+  // BEGIN UNWIND
+
+  *aFramesUsed = 0;
+
+  UnwindRegs regs = *aStartRegs;
+  TaggedUWord last_valid_sp = TaggedUWord();
+
+  while (true) {
+    if (DEBUG_MAIN) {
+      char buf[300];
+      mLog("\n");
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+      SprintfLiteral(
+          buf, "LoopTop: rip %d/%llx  rsp %d/%llx  rbp %d/%llx\n",
+          (int)regs.xip.Valid(), (unsigned long long int)regs.xip.Value(),
+          (int)regs.xsp.Valid(), (unsigned long long int)regs.xsp.Value(),
+          (int)regs.xbp.Valid(), (unsigned long long int)regs.xbp.Value());
+      buf[sizeof(buf) - 1] = 0;
+      mLog(buf);
+#elif defined(GP_ARCH_arm)
+      SprintfLiteral(
+          buf,
+          "LoopTop: r15 %d/%llx  r7 %d/%llx  r11 %d/%llx"
+          "  r12 %d/%llx  r13 %d/%llx  r14 %d/%llx\n",
+          (int)regs.r15.Valid(), (unsigned long long int)regs.r15.Value(),
+          (int)regs.r7.Valid(), (unsigned long long int)regs.r7.Value(),
+          (int)regs.r11.Valid(), (unsigned long long int)regs.r11.Value(),
+          (int)regs.r12.Valid(), (unsigned long long int)regs.r12.Value(),
+          (int)regs.r13.Valid(), (unsigned long long int)regs.r13.Value(),
+          (int)regs.r14.Valid(), (unsigned long long int)regs.r14.Value());
+      buf[sizeof(buf) - 1] = 0;
+      mLog(buf);
+#elif defined(GP_ARCH_arm64)
+      SprintfLiteral(
+          buf,
+          "LoopTop: pc %d/%llx  x29 %d/%llx  x30 %d/%llx"
+          "  sp %d/%llx\n",
+          (int)regs.pc.Valid(), (unsigned long long int)regs.pc.Value(),
+          (int)regs.x29.Valid(), (unsigned long long int)regs.x29.Value(),
+          (int)regs.x30.Valid(), (unsigned long long int)regs.x30.Value(),
+          (int)regs.sp.Valid(), (unsigned long long int)regs.sp.Value());
+      buf[sizeof(buf) - 1] = 0;
+      mLog(buf);
+#elif defined(GP_ARCH_mips64)
+      SprintfLiteral(
+          buf, "LoopTop: pc %d/%llx  sp %d/%llx  fp %d/%llx\n",
+          (int)regs.pc.Valid(), (unsigned long long int)regs.pc.Value(),
+          (int)regs.sp.Valid(), (unsigned long long int)regs.sp.Value(),
+          (int)regs.fp.Valid(), (unsigned long long int)regs.fp.Value());
+      buf[sizeof(buf) - 1] = 0;
+      mLog(buf);
+#else
+#  error "Unsupported arch"
+#endif
+    }
+
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+    TaggedUWord ia = regs.xip;
+    TaggedUWord sp = regs.xsp;
+#elif defined(GP_ARCH_arm)
+    TaggedUWord ia = (*aFramesUsed == 0 ? regs.r15 : regs.r14);
+    TaggedUWord sp = regs.r13;
+#elif defined(GP_ARCH_arm64)
+    TaggedUWord ia = (*aFramesUsed == 0 ? regs.pc : regs.x30);
+    TaggedUWord sp = regs.sp;
+#elif defined(GP_ARCH_mips64)
+    TaggedUWord ia = regs.pc;
+    TaggedUWord sp = regs.sp;
+#else
+#  error "Unsupported arch"
+#endif
+
+    if (*aFramesUsed >= aFramesAvail) {
+      break;
+    }
+
+    // If we don't have a valid value for the PC, give up.
+    if (!ia.Valid()) {
+      break;
+    }
+
+    // If this is the innermost frame, record the SP value, which
+    // presumably is valid.  If this isn't the innermost frame, and we
+    // have a valid SP value, check that its SP value isn't less that
+    // the one we've seen so far, so as to catch potential SP value
+    // cycles.
+    if (*aFramesUsed == 0) {
+      last_valid_sp = sp;
+    } else {
+      MOZ_ASSERT(last_valid_sp.Valid());
+      if (sp.Valid()) {
+        if (sp.Value() < last_valid_sp.Value()) {
+          // Hmm, SP going in the wrong direction.  Let's stop.
+          break;
+        }
+        // Remember where we got to.
+        last_valid_sp = sp;
+      }
+    }
+
+    // For the innermost frame, the IA value is what we need.  For all
+    // other frames, it's actually the return address, so back up one
+    // byte so as to get it into the calling instruction.
+    aFramePCs[*aFramesUsed] = ia.Value() - (*aFramesUsed == 0 ? 0 : 1);
+    aFrameSPs[*aFramesUsed] = sp.Valid() ? sp.Value() : 0;
+    (*aFramesUsed)++;
+
+    // Find the RuleSet for the current IA, if any.  This will also
+    // query the backing (secondary) maps if it isn't found in the
+    // thread-local cache.
+
+    // If this isn't the innermost frame, back up into the calling insn.
+    if (*aFramesUsed > 1) {
+      ia = ia + TaggedUWord((uintptr_t)(-1));
+    }
+
+    pair<const RuleSet*, const vector<PfxInstr>*> ruleset_and_pfxinstrs =
+        mPriMap->Lookup(ia.Value());
+    const RuleSet* ruleset = ruleset_and_pfxinstrs.first;
+    const vector<PfxInstr>* pfxinstrs = ruleset_and_pfxinstrs.second;
+
+    if (DEBUG_MAIN) {
+      char buf[100];
+      SprintfLiteral(buf, "ruleset for 0x%llx = %p\n",
+                     (unsigned long long int)ia.Value(), ruleset);
+      buf[sizeof(buf) - 1] = 0;
+      mLog(buf);
+    }
+
+#if defined(GP_PLAT_x86_android) || defined(GP_PLAT_x86_linux)
+    /////////////////////////////////////////////
+    ////
+    // On 32 bit x86-linux, syscalls are often done via the VDSO
+    // function __kernel_vsyscall, which doesn't have a corresponding
+    // object that we can read debuginfo from.  That effectively kills
+    // off all stack traces for threads blocked in syscalls.  Hence
+    // special-case by looking at the code surrounding the program
+    // counter.
+    //
+    // 0xf7757420 <__kernel_vsyscall+0>:	push   %ecx
+    // 0xf7757421 <__kernel_vsyscall+1>:	push   %edx
+    // 0xf7757422 <__kernel_vsyscall+2>:	push   %ebp
+    // 0xf7757423 <__kernel_vsyscall+3>:	mov    %esp,%ebp
+    // 0xf7757425 <__kernel_vsyscall+5>:	sysenter
+    // 0xf7757427 <__kernel_vsyscall+7>:	nop
+    // 0xf7757428 <__kernel_vsyscall+8>:	nop
+    // 0xf7757429 <__kernel_vsyscall+9>:	nop
+    // 0xf775742a <__kernel_vsyscall+10>:	nop
+    // 0xf775742b <__kernel_vsyscall+11>:	nop
+    // 0xf775742c <__kernel_vsyscall+12>:	nop
+    // 0xf775742d <__kernel_vsyscall+13>:	nop
+    // 0xf775742e <__kernel_vsyscall+14>:	int    $0x80
+    // 0xf7757430 <__kernel_vsyscall+16>:	pop    %ebp
+    // 0xf7757431 <__kernel_vsyscall+17>:	pop    %edx
+    // 0xf7757432 <__kernel_vsyscall+18>:	pop    %ecx
+    // 0xf7757433 <__kernel_vsyscall+19>:	ret
+    //
+    // In cases where the sampled thread is blocked in a syscall, its
+    // program counter will point at "pop %ebp".  Hence we look for
+    // the sequence "int $0x80; pop %ebp; pop %edx; pop %ecx; ret", and
+    // the corresponding register-recovery actions are:
+    //    new_ebp = *(old_esp + 0)
+    //    new eip = *(old_esp + 12)
+    //    new_esp = old_esp + 16
+    //
+    // It may also be the case that the program counter points two
+    // nops before the "int $0x80", viz, is __kernel_vsyscall+12, in
+    // the case where the syscall has been restarted but the thread
+    // hasn't been rescheduled.  The code below doesn't handle that;
+    // it could easily be made to.
+    //
+    if (!ruleset && *aFramesUsed == 1 && ia.Valid() && sp.Valid()) {
+      uintptr_t insns_min, insns_max;
+      uintptr_t eip = ia.Value();
+      bool b = mSegArray->getBoundingCodeSegment(&insns_min, &insns_max, eip);
+      if (b && eip - 2 >= insns_min && eip + 3 <= insns_max) {
+        uint8_t* eipC = (uint8_t*)eip;
+        if (eipC[-2] == 0xCD && eipC[-1] == 0x80 && eipC[0] == 0x5D &&
+            eipC[1] == 0x5A && eipC[2] == 0x59 && eipC[3] == 0xC3) {
+          TaggedUWord sp_plus_0 = sp;
+          TaggedUWord sp_plus_12 = sp;
+          TaggedUWord sp_plus_16 = sp;
+          sp_plus_12 = sp_plus_12 + TaggedUWord(12);
+          sp_plus_16 = sp_plus_16 + TaggedUWord(16);
+          TaggedUWord new_ebp = DerefTUW(sp_plus_0, aStackImg);
+          TaggedUWord new_eip = DerefTUW(sp_plus_12, aStackImg);
+          TaggedUWord new_esp = sp_plus_16;
+          if (new_ebp.Valid() && new_eip.Valid() && new_esp.Valid()) {
+            regs.xbp = new_ebp;
+            regs.xip = new_eip;
+            regs.xsp = new_esp;
+            continue;
+          }
+        }
+      }
+    }
+    ////
+    /////////////////////////////////////////////
+#endif  // defined(GP_PLAT_x86_android) || defined(GP_PLAT_x86_linux)
+
+    // So, do we have a ruleset for this address?  If so, use it now.
+    if (ruleset) {
+      if (DEBUG_MAIN) {
+        ruleset->Print(mLog);
+        mLog("\n");
+      }
+      // Use the RuleSet to compute the registers for the previous
+      // frame.  |regs| is modified in-place.
+      UseRuleSet(&regs, aStackImg, ruleset, pfxinstrs);
+      continue;
+    }
+
+#if defined(GP_PLAT_amd64_linux) || defined(GP_PLAT_x86_linux) ||     \
+    defined(GP_PLAT_amd64_android) || defined(GP_PLAT_x86_android) || \
+    defined(GP_PLAT_amd64_freebsd)
+    // There's no RuleSet for the specified address.  On amd64/x86_linux, see if
+    // it's possible to recover the caller's frame by using the frame pointer.
+
+    // We seek to compute (new_IP, new_SP, new_BP) from (old_BP, stack image),
+    // and assume the following layout:
+    //
+    //                 <--- new_SP
+    //   +----------+
+    //   |  new_IP  |  (return address)
+    //   +----------+
+    //   |  new_BP  |  <--- old_BP
+    //   +----------+
+    //   |   ....   |
+    //   |   ....   |
+    //   |   ....   |
+    //   +----------+  <---- old_SP (arbitrary, but must be <= old_BP)
+
+    const size_t wordSzB = sizeof(uintptr_t);
+    TaggedUWord old_xsp = regs.xsp;
+
+    // points at new_BP ?
+    TaggedUWord old_xbp = regs.xbp;
+    // points at new_IP ?
+    TaggedUWord old_xbp_plus1 = regs.xbp + TaggedUWord(1 * wordSzB);
+    // is the new_SP ?
+    TaggedUWord old_xbp_plus2 = regs.xbp + TaggedUWord(2 * wordSzB);
+
+    if (old_xbp.Valid() && old_xbp.IsAligned() && old_xsp.Valid() &&
+        old_xsp.IsAligned() && old_xsp.Value() <= old_xbp.Value()) {
+      // We don't need to do any range, alignment or validity checks for
+      // addresses passed to DerefTUW, since that performs them itself, and
+      // returns an invalid value on failure.  Any such value will poison
+      // subsequent uses, and we do a final check for validity before putting
+      // the computed values into |regs|.
+      TaggedUWord new_xbp = DerefTUW(old_xbp, aStackImg);
+      if (new_xbp.Valid() && new_xbp.IsAligned() &&
+          old_xbp.Value() < new_xbp.Value()) {
+        TaggedUWord new_xip = DerefTUW(old_xbp_plus1, aStackImg);
+        TaggedUWord new_xsp = old_xbp_plus2;
+        if (new_xbp.Valid() && new_xip.Valid() && new_xsp.Valid()) {
+          regs.xbp = new_xbp;
+          regs.xip = new_xip;
+          regs.xsp = new_xsp;
+          (*aFramePointerFramesAcquired)++;
+          continue;
+        }
+      }
+    }
+#elif defined(GP_ARCH_arm64)
+    // Here is an example of generated code for prologue and epilogue..
+    //
+    // stp     x29, x30, [sp, #-16]!
+    // mov     x29, sp
+    // ...
+    // ldp     x29, x30, [sp], #16
+    // ret
+    //
+    // Next is another example of generated code.
+    //
+    // stp     x20, x19, [sp, #-32]!
+    // stp     x29, x30, [sp, #16]
+    // add     x29, sp, #0x10
+    // ...
+    // ldp     x29, x30, [sp, #16]
+    // ldp     x20, x19, [sp], #32
+    // ret
+    //
+    // Previous x29 and x30 register are stored in the address of x29 register.
+    // But since sp register value depends on local variables, we cannot compute
+    // previous sp register from current sp/fp/lr register and there is no
+    // regular rule for sp register in prologue. But since return address is lr
+    // register, if x29 is valid, we will get return address without sp
+    // register.
+    //
+    // So we assume the following layout that if no rule set. x29 is frame
+    // pointer, so we will be able to compute x29 and x30 .
+    //
+    //   +----------+  <--- new_sp (cannot compute)
+    //   |   ....   |
+    //   +----------+
+    //   |  new_lr  |  (return address)
+    //   +----------+
+    //   |  new_fp  |  <--- old_fp
+    //   +----------+
+    //   |   ....   |
+    //   |   ....   |
+    //   +----------+  <---- old_sp (arbitrary, but unused)
+
+    TaggedUWord old_fp = regs.x29;
+    if (old_fp.Valid() && old_fp.IsAligned() && last_valid_sp.Valid() &&
+        last_valid_sp.Value() <= old_fp.Value()) {
+      TaggedUWord new_fp = DerefTUW(old_fp, aStackImg);
+      if (new_fp.Valid() && new_fp.IsAligned() &&
+          old_fp.Value() < new_fp.Value()) {
+        TaggedUWord old_fp_plus1 = old_fp + TaggedUWord(8);
+        TaggedUWord new_lr = DerefTUW(old_fp_plus1, aStackImg);
+        if (new_lr.Valid()) {
+          regs.x29 = new_fp;
+          regs.x30 = new_lr;
+          // When using frame pointer to walk stack, we cannot compute sp
+          // register since we cannot compute sp register from fp/lr/sp
+          // register, and there is no regular rule to compute previous sp
+          // register. So mark as invalid.
+          regs.sp = TaggedUWord();
+          (*aFramePointerFramesAcquired)++;
+          continue;
+        }
+      }
+    }
+#endif  // defined(GP_PLAT_amd64_linux) || defined(GP_PLAT_x86_linux) ||
+        // defined(GP_PLAT_amd64_android) || defined(GP_PLAT_x86_android)
+
+    // We failed to recover a frame either using CFI or FP chasing, and we
+    // have no other ways to recover the frame.  So we have to give up.
+    break;
+
+  }  // top level unwind loop
+
+  // END UNWIND
+  /////////////////////////////////////////////////////////
+}
+
+////////////////////////////////////////////////////////////////
+// LUL Unit Testing                                           //
+////////////////////////////////////////////////////////////////
+
+static const int LUL_UNIT_TEST_STACK_SIZE = 32768;
+
+#if defined(GP_ARCH_mips64)
+static __attribute__((noinline)) unsigned long __getpc(void) {
+  unsigned long rtaddr;
+  __asm__ volatile("move %0, $31" : "=r"(rtaddr));
+  return rtaddr;
+}
+#endif
+
+// This function is innermost in the test call sequence.  It uses LUL
+// to unwind, and compares the result with the sequence specified in
+// the director string.  These need to agree in order for the test to
+// pass.  In order not to screw up the results, this function needs
+// to have a not-very big stack frame, since we're only presenting
+// the innermost LUL_UNIT_TEST_STACK_SIZE bytes of stack to LUL, and
+// that chunk unavoidably includes the frame for this function.
+//
+// This function must not be inlined into its callers.  Doing so will
+// cause the expected-vs-actual backtrace consistency checking to
+// fail.  Prints summary results to |aLUL|'s logging sink and also
+// returns a boolean indicating whether or not the test passed.
+static __attribute__((noinline)) bool GetAndCheckStackTrace(
+    LUL* aLUL, const char* dstring) {
+  // Get hold of the current unwind-start registers.
+  UnwindRegs startRegs;
+  memset(&startRegs, 0, sizeof(startRegs));
+#if defined(GP_ARCH_amd64)
+  volatile uintptr_t block[3];
+  MOZ_ASSERT(sizeof(block) == 24);
+  __asm__ __volatile__(
+      "leaq 0(%%rip), %%r15"
+      "\n\t"
+      "movq %%r15, 0(%0)"
+      "\n\t"
+      "movq %%rsp, 8(%0)"
+      "\n\t"
+      "movq %%rbp, 16(%0)"
+      "\n"
+      :
+      : "r"(&block[0])
+      : "memory", "r15");
+  startRegs.xip = TaggedUWord(block[0]);
+  startRegs.xsp = TaggedUWord(block[1]);
+  startRegs.xbp = TaggedUWord(block[2]);
+  const uintptr_t REDZONE_SIZE = 128;
+  uintptr_t start = block[1] - REDZONE_SIZE;
+#elif defined(GP_PLAT_x86_linux) || defined(GP_PLAT_x86_android)
+  volatile uintptr_t block[3];
+  MOZ_ASSERT(sizeof(block) == 12);
+  __asm__ __volatile__(
+      ".byte 0xE8,0x00,0x00,0x00,0x00" /*call next insn*/
+      "\n\t"
+      "popl %%edi"
+      "\n\t"
+      "movl %%edi, 0(%0)"
+      "\n\t"
+      "movl %%esp, 4(%0)"
+      "\n\t"
+      "movl %%ebp, 8(%0)"
+      "\n"
+      :
+      : "r"(&block[0])
+      : "memory", "edi");
+  startRegs.xip = TaggedUWord(block[0]);
+  startRegs.xsp = TaggedUWord(block[1]);
+  startRegs.xbp = TaggedUWord(block[2]);
+  const uintptr_t REDZONE_SIZE = 0;
+  uintptr_t start = block[1] - REDZONE_SIZE;
+#elif defined(GP_PLAT_arm_linux) || defined(GP_PLAT_arm_android)
+  volatile uintptr_t block[6];
+  MOZ_ASSERT(sizeof(block) == 24);
+  __asm__ __volatile__(
+      "mov r0, r15"
+      "\n\t"
+      "str r0,  [%0, #0]"
+      "\n\t"
+      "str r14, [%0, #4]"
+      "\n\t"
+      "str r13, [%0, #8]"
+      "\n\t"
+      "str r12, [%0, #12]"
+      "\n\t"
+      "str r11, [%0, #16]"
+      "\n\t"
+      "str r7,  [%0, #20]"
+      "\n"
+      :
+      : "r"(&block[0])
+      : "memory", "r0");
+  startRegs.r15 = TaggedUWord(block[0]);
+  startRegs.r14 = TaggedUWord(block[1]);
+  startRegs.r13 = TaggedUWord(block[2]);
+  startRegs.r12 = TaggedUWord(block[3]);
+  startRegs.r11 = TaggedUWord(block[4]);
+  startRegs.r7 = TaggedUWord(block[5]);
+  const uintptr_t REDZONE_SIZE = 0;
+  uintptr_t start = block[1] - REDZONE_SIZE;
+#elif defined(GP_ARCH_arm64)
+  volatile uintptr_t block[4];
+  MOZ_ASSERT(sizeof(block) == 32);
+  __asm__ __volatile__(
+      "adr x0, . \n\t"
+      "str x0, [%0, #0] \n\t"
+      "str x29, [%0, #8] \n\t"
+      "str x30, [%0, #16] \n\t"
+      "mov x0, sp \n\t"
+      "str x0, [%0, #24] \n\t"
+      :
+      : "r"(&block[0])
+      : "memory", "x0");
+  startRegs.pc = TaggedUWord(block[0]);
+  startRegs.x29 = TaggedUWord(block[1]);
+  startRegs.x30 = TaggedUWord(block[2]);
+  startRegs.sp = TaggedUWord(block[3]);
+  const uintptr_t REDZONE_SIZE = 0;
+  uintptr_t start = block[1] - REDZONE_SIZE;
+#elif defined(GP_ARCH_mips64)
+  volatile uintptr_t block[3];
+  MOZ_ASSERT(sizeof(block) == 24);
+  __asm__ __volatile__(
+      "sd $29, 8(%0)     \n"
+      "sd $30, 16(%0)    \n"
+      :
+      : "r"(block)
+      : "memory");
+  block[0] = __getpc();
+  startRegs.pc = TaggedUWord(block[0]);
+  startRegs.sp = TaggedUWord(block[1]);
+  startRegs.fp = TaggedUWord(block[2]);
+  const uintptr_t REDZONE_SIZE = 0;
+  uintptr_t start = block[1] - REDZONE_SIZE;
+#else
+#  error "Unsupported platform"
+#endif
+
+  // Get hold of the innermost LUL_UNIT_TEST_STACK_SIZE bytes of the
+  // stack.
+  uintptr_t end = start + LUL_UNIT_TEST_STACK_SIZE;
+  uintptr_t ws = sizeof(void*);
+  start &= ~(ws - 1);
+  end &= ~(ws - 1);
+  uintptr_t nToCopy = end - start;
+  if (nToCopy > lul::N_STACK_BYTES) {
+    nToCopy = lul::N_STACK_BYTES;
+  }
+  MOZ_ASSERT(nToCopy <= lul::N_STACK_BYTES);
+  StackImage* stackImg = new StackImage();
+  stackImg->mLen = nToCopy;
+  stackImg->mStartAvma = start;
+  if (nToCopy > 0) {
+    MOZ_MAKE_MEM_DEFINED((void*)start, nToCopy);
+    memcpy(&stackImg->mContents[0], (void*)start, nToCopy);
+  }
+
+  // Unwind it.
+  const int MAX_TEST_FRAMES = 64;
+  uintptr_t framePCs[MAX_TEST_FRAMES];
+  uintptr_t frameSPs[MAX_TEST_FRAMES];
+  size_t framesAvail = mozilla::ArrayLength(framePCs);
+  size_t framesUsed = 0;
+  size_t framePointerFramesAcquired = 0;
+  aLUL->Unwind(&framePCs[0], &frameSPs[0], &framesUsed,
+               &framePointerFramesAcquired, framesAvail, &startRegs, stackImg);
+
+  delete stackImg;
+
+  // if (0) {
+  //  // Show what we have.
+  //  fprintf(stderr, "Got %d frames:\n", (int)framesUsed);
+  //  for (size_t i = 0; i < framesUsed; i++) {
+  //    fprintf(stderr, "  [%2d]   SP %p   PC %p\n",
+  //            (int)i, (void*)frameSPs[i], (void*)framePCs[i]);
+  //  }
+  //  fprintf(stderr, "\n");
+  //}
+
+  // Check to see if there's a consistent binding between digits in
+  // the director string ('1' .. '8') and the PC values acquired by
+  // the unwind.  If there isn't, the unwinding has failed somehow.
+  uintptr_t binding[8];  // binding for '1' .. binding for '8'
+  memset((void*)binding, 0, sizeof(binding));
+
+  // The general plan is to work backwards along the director string
+  // and forwards along the framePCs array.  Doing so corresponds to
+  // working outwards from the innermost frame of the recursive test set.
+  const char* cursor = dstring;
+
+  // Find the end.  This leaves |cursor| two bytes past the first
+  // character we want to look at -- see comment below.
+  while (*cursor) cursor++;
+
+  // Counts the number of consistent frames.
+  size_t nConsistent = 0;
+
+  // Iterate back to the start of the director string.  The starting
+  // points are a bit complex.  We can't use framePCs[0] because that
+  // contains the PC in this frame (above).  We can't use framePCs[1]
+  // because that will contain the PC at return point in the recursive
+  // test group (TestFn[1-8]) for their call "out" to this function,
+  // GetAndCheckStackTrace.  Although LUL will compute a correct
+  // return address, that will not be the same return address as for a
+  // recursive call out of the the function to another function in the
+  // group.  Hence we can only start consistency checking at
+  // framePCs[2].
+  //
+  // To be consistent, then, we must ignore the last element in the
+  // director string as that corresponds to framePCs[1].  Hence the
+  // start points are: framePCs[2] and the director string 2 bytes
+  // before the terminating zero.
+  //
+  // Also as a result of this, the number of consistent frames counted
+  // will always be one less than the length of the director string
+  // (not including its terminating zero).
+  size_t frameIx;
+  for (cursor = cursor - 2, frameIx = 2;
+       cursor >= dstring && frameIx < framesUsed; cursor--, frameIx++) {
+    char c = *cursor;
+    uintptr_t pc = framePCs[frameIx];
+    // If this doesn't hold, the director string is ill-formed.
+    MOZ_ASSERT(c >= '1' && c <= '8');
+    int n = ((int)c) - ((int)'1');
+    if (binding[n] == 0) {
+      // There's no binding for |c| yet, so install |pc| and carry on.
+      binding[n] = pc;
+      nConsistent++;
+      continue;
+    }
+    // There's a pre-existing binding for |c|.  Check it's consistent.
+    if (binding[n] != pc) {
+      // Not consistent.  Give up now.
+      break;
+    }
+    // Consistent.  Keep going.
+    nConsistent++;
+  }
+
+  // So, did we succeed?
+  bool passed = nConsistent + 1 == strlen(dstring);
+
+  // Show the results.
+  char buf[200];
+  SprintfLiteral(buf, "LULUnitTest:   dstring = %s\n", dstring);
+  buf[sizeof(buf) - 1] = 0;
+  aLUL->mLog(buf);
+  SprintfLiteral(buf, "LULUnitTest:     %d consistent, %d in dstring: %s\n",
+                 (int)nConsistent, (int)strlen(dstring),
+                 passed ? "PASS" : "FAIL");
+  buf[sizeof(buf) - 1] = 0;
+  aLUL->mLog(buf);
+
+  return passed;
+}
+
+// Macro magic to create a set of 8 mutually recursive functions with
+// varying frame sizes.  These will recurse amongst themselves as
+// specified by |strP|, the directory string, and call
+// GetAndCheckStackTrace when the string becomes empty, passing it the
+// original value of the string.  This checks the result, printing
+// results on |aLUL|'s logging sink, and also returns a boolean
+// indicating whether or not the results are acceptable (correct).
+
+#define DECL_TEST_FN(NAME) \
+  bool NAME(LUL* aLUL, const char* strPorig, const char* strP);
+
+#define GEN_TEST_FN(NAME, FRAMESIZE)                                          \
+  bool NAME(LUL* aLUL, const char* strPorig, const char* strP) {              \
+    /* Create a frame of size (at least) FRAMESIZE, so that the */            \
+    /* 8 functions created by this macro offer some variation in frame */     \
+    /* sizes.  This isn't as simple as it might seem, since a clever */       \
+    /* optimizing compiler (eg, clang-5) detects that the array is unused */  \
+    /* and removes it.  We try to defeat this by passing it to a function */  \
+    /* in a different compilation unit, and hoping that clang does not */     \
+    /* notice that the call is a no-op. */                                    \
+    char space[FRAMESIZE];                                                    \
+    Unused << write(1, space, 0); /* write zero bytes of |space| to stdout */ \
+                                                                              \
+    if (*strP == '\0') {                                                      \
+      /* We've come to the end of the director string. */                     \
+      /* Take a stack snapshot. */                                            \
+      return GetAndCheckStackTrace(aLUL, strPorig);                           \
+    } else {                                                                  \
+      /* Recurse onwards.  This is a bit subtle.  The obvious */              \
+      /* thing to do here is call onwards directly, from within the */        \
+      /* arms of the case statement.  That gives a problem in that */         \
+      /* there will be multiple return points inside each function when */    \
+      /* unwinding, so it will be difficult to check for consistency */       \
+      /* against the director string.  Instead, we make an indirect */        \
+      /* call, so as to guarantee that there is only one call site */         \
+      /* within each function.  This does assume that the compiler */         \
+      /* won't transform it back to the simple direct-call form. */           \
+      /* To discourage it from doing so, the call is bracketed with */        \
+      /* __asm__ __volatile__ sections so as to make it not-movable. */       \
+      bool (*nextFn)(LUL*, const char*, const char*) = NULL;                  \
+      switch (*strP) {                                                        \
+        case '1':                                                             \
+          nextFn = TestFn1;                                                   \
+          break;                                                              \
+        case '2':                                                             \
+          nextFn = TestFn2;                                                   \
+          break;                                                              \
+        case '3':                                                             \
+          nextFn = TestFn3;                                                   \
+          break;                                                              \
+        case '4':                                                             \
+          nextFn = TestFn4;                                                   \
+          break;                                                              \
+        case '5':                                                             \
+          nextFn = TestFn5;                                                   \
+          break;                                                              \
+        case '6':                                                             \
+          nextFn = TestFn6;                                                   \
+          break;                                                              \
+        case '7':                                                             \
+          nextFn = TestFn7;                                                   \
+          break;                                                              \
+        case '8':                                                             \
+          nextFn = TestFn8;                                                   \
+          break;                                                              \
+        default:                                                              \
+          nextFn = TestFn8;                                                   \
+          break;                                                              \
+      }                                                                       \
+      /* "use" |space| immediately after the recursive call, */               \
+      /* so as to dissuade clang from deallocating the space while */         \
+      /* the call is active, or otherwise messing with the stack frame. */    \
+      __asm__ __volatile__("" ::: "cc", "memory");                            \
+      bool passed = nextFn(aLUL, strPorig, strP + 1);                         \
+      Unused << write(1, space, 0);                                           \
+      __asm__ __volatile__("" ::: "cc", "memory");                            \
+      return passed;                                                          \
+    }                                                                         \
+  }
+
+// The test functions are mutually recursive, so it is necessary to
+// declare them before defining them.
+DECL_TEST_FN(TestFn1)
+DECL_TEST_FN(TestFn2)
+DECL_TEST_FN(TestFn3)
+DECL_TEST_FN(TestFn4)
+DECL_TEST_FN(TestFn5)
+DECL_TEST_FN(TestFn6)
+DECL_TEST_FN(TestFn7)
+DECL_TEST_FN(TestFn8)
+
+GEN_TEST_FN(TestFn1, 123)
+GEN_TEST_FN(TestFn2, 456)
+GEN_TEST_FN(TestFn3, 789)
+GEN_TEST_FN(TestFn4, 23)
+GEN_TEST_FN(TestFn5, 47)
+GEN_TEST_FN(TestFn6, 117)
+GEN_TEST_FN(TestFn7, 1)
+GEN_TEST_FN(TestFn8, 99)
+
+// This starts the test sequence going.  Call here to generate a
+// sequence of calls as directed by the string |dstring|.  The call
+// sequence will, from its innermost frame, finish by calling
+// GetAndCheckStackTrace() and passing it |dstring|.
+// GetAndCheckStackTrace() will unwind the stack, check consistency
+// of those results against |dstring|, and print a pass/fail message
+// to aLUL's logging sink.  It also updates the counters in *aNTests
+// and aNTestsPassed.
+__attribute__((noinline)) void TestUnw(/*OUT*/ int* aNTests,
+                                       /*OUT*/ int* aNTestsPassed, LUL* aLUL,
+                                       const char* dstring) {
+  // Ensure that the stack has at least this much space on it.  This
+  // makes it safe to saw off the top LUL_UNIT_TEST_STACK_SIZE bytes
+  // and hand it to LUL.  Safe in the sense that no segfault can
+  // happen because the stack is at least this big.  This is all
+  // somewhat dubious in the sense that a sufficiently clever compiler
+  // (clang, for one) can figure out that space[] is unused and delete
+  // it from the frame.  Hence the somewhat elaborate hoop jumping to
+  // fill it up before the call and to at least appear to use the
+  // value afterwards.
+  int i;
+  volatile char space[LUL_UNIT_TEST_STACK_SIZE];
+  for (i = 0; i < LUL_UNIT_TEST_STACK_SIZE; i++) {
+    space[i] = (char)(i & 0x7F);
+  }
+
+  // Really run the test.
+  bool passed = TestFn1(aLUL, dstring, dstring);
+
+  // Appear to use space[], by visiting the value to compute some kind
+  // of checksum, and then (apparently) using the checksum.
+  int sum = 0;
+  for (i = 0; i < LUL_UNIT_TEST_STACK_SIZE; i++) {
+    // If this doesn't fool LLVM, I don't know what will.
+    sum += space[i] - 3 * i;
+  }
+  __asm__ __volatile__("" : : "r"(sum));
+
+  // Update the counters.
+  (*aNTests)++;
+  if (passed) {
+    (*aNTestsPassed)++;
+  }
+}
+
+void RunLulUnitTests(/*OUT*/ int* aNTests, /*OUT*/ int* aNTestsPassed,
+                     LUL* aLUL) {
+  aLUL->mLog(":\n");
+  aLUL->mLog("LULUnitTest: BEGIN\n");
+  *aNTests = *aNTestsPassed = 0;
+  TestUnw(aNTests, aNTestsPassed, aLUL, "11111111");
+  TestUnw(aNTests, aNTestsPassed, aLUL, "11222211");
+  TestUnw(aNTests, aNTestsPassed, aLUL, "111222333");
+  TestUnw(aNTests, aNTestsPassed, aLUL, "1212121231212331212121212121212");
+  TestUnw(aNTests, aNTestsPassed, aLUL, "31415827271828325332173258");
+  TestUnw(aNTests, aNTestsPassed, aLUL,
+          "123456781122334455667788777777777777777777777");
+  aLUL->mLog("LULUnitTest: END\n");
+  aLUL->mLog(":\n");
+}
+
+}  // namespace lul
diff --git a/mozglue/baseprofiler/lul/LulMain.h b/mozglue/baseprofiler/lul/LulMain.h
new file mode 100644
index 0000000000..b0cb7f4f1e
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulMain.h
@@ -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/. */
+
+#ifndef LulMain_h
+#define LulMain_h
+
+#include "PlatformMacros.h"
+#include "mozilla/Atomics.h"
+#include "mozilla/MemoryReporting.h"
+
+// LUL: A Lightweight Unwind Library.
+// This file provides the end-user (external) interface for LUL.
+
+// Some comments about naming in the implementation.  These are safe
+// to ignore if you are merely using LUL, but are important if you
+// hack on its internals.
+//
+// Debuginfo readers in general have tended to use the word "address"
+// to mean several different things.  This sometimes makes them
+// difficult to understand and maintain.  LUL tries hard to avoid
+// using the word "address" and instead uses the following more
+// precise terms:
+//
+// * SVMA ("Stated Virtual Memory Address"): this is an address of a
+//   symbol (etc) as it is stated in the symbol table, or other
+//   metadata, of an object.  Such values are typically small and
+//   start from zero or thereabouts, unless the object has been
+//   prelinked.
+//
+// * AVMA ("Actual Virtual Memory Address"): this is the address of a
+//   symbol (etc) in a running process, that is, once the associated
+//   object has been mapped into a process.  Such values are typically
+//   much larger than SVMAs, since objects can get mapped arbitrarily
+//   far along the address space.
+//
+// * "Bias": the difference between AVMA and SVMA for a given symbol
+//   (specifically, AVMA - SVMA).  The bias is always an integral
+//   number of pages.  Once we know the bias for a given object's
+//   text section (for example), we can compute the AVMAs of all of
+//   its text symbols by adding the bias to their SVMAs.
+//
+// * "Image address": typically, to read debuginfo from an object we
+//   will temporarily mmap in the file so as to read symbol tables
+//   etc.  Addresses in this temporary mapping are called "Image
+//   addresses".  Note that the temporary mapping is entirely
+//   unrelated to the mappings of the file that the dynamic linker
+//   must perform merely in order to get the program to run.  Hence
+//   image addresses are unrelated to either SVMAs or AVMAs.
+
+namespace lul {
+
+// A machine word plus validity tag.
+class TaggedUWord {
+ public:
+  // RUNS IN NO-MALLOC CONTEXT
+  // Construct a valid one.
+  explicit TaggedUWord(uintptr_t w) : mValue(w), mValid(true) {}
+
+  // RUNS IN NO-MALLOC CONTEXT
+  // Construct an invalid one.
+  TaggedUWord() : mValue(0), mValid(false) {}
+
+  // RUNS IN NO-MALLOC CONTEXT
+  TaggedUWord operator+(TaggedUWord rhs) const {
+    return (Valid() && rhs.Valid()) ? TaggedUWord(Value() + rhs.Value())
+                                    : TaggedUWord();
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  TaggedUWord operator-(TaggedUWord rhs) const {
+    return (Valid() && rhs.Valid()) ? TaggedUWord(Value() - rhs.Value())
+                                    : TaggedUWord();
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  TaggedUWord operator&(TaggedUWord rhs) const {
+    return (Valid() && rhs.Valid()) ? TaggedUWord(Value() & rhs.Value())
+                                    : TaggedUWord();
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  TaggedUWord operator|(TaggedUWord rhs) const {
+    return (Valid() && rhs.Valid()) ? TaggedUWord(Value() | rhs.Value())
+                                    : TaggedUWord();
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  TaggedUWord CmpGEs(TaggedUWord rhs) const {
+    if (Valid() && rhs.Valid()) {
+      intptr_t s1 = (intptr_t)Value();
+      intptr_t s2 = (intptr_t)rhs.Value();
+      return TaggedUWord(s1 >= s2 ? 1 : 0);
+    }
+    return TaggedUWord();
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  TaggedUWord operator<<(TaggedUWord rhs) const {
+    if (Valid() && rhs.Valid()) {
+      uintptr_t shift = rhs.Value();
+      if (shift < 8 * sizeof(uintptr_t)) return TaggedUWord(Value() << shift);
+    }
+    return TaggedUWord();
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  // Is equal?  Note: non-validity on either side gives non-equality.
+  bool operator==(TaggedUWord other) const {
+    return (mValid && other.Valid()) ? (mValue == other.Value()) : false;
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  // Is it word-aligned?
+  bool IsAligned() const {
+    return mValid && (mValue & (sizeof(uintptr_t) - 1)) == 0;
+  }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  uintptr_t Value() const { return mValue; }
+
+  // RUNS IN NO-MALLOC CONTEXT
+  bool Valid() const { return mValid; }
+
+ private:
+  uintptr_t mValue;
+  bool mValid;
+};
+
+// The registers, with validity tags, that will be unwound.
+
+struct UnwindRegs {
+#if defined(GP_ARCH_arm)
+  TaggedUWord r7;
+  TaggedUWord r11;
+  TaggedUWord r12;
+  TaggedUWord r13;
+  TaggedUWord r14;
+  TaggedUWord r15;
+#elif defined(GP_ARCH_arm64)
+  TaggedUWord x29;
+  TaggedUWord x30;
+  TaggedUWord sp;
+  TaggedUWord pc;
+#elif defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+  TaggedUWord xbp;
+  TaggedUWord xsp;
+  TaggedUWord xip;
+#elif defined(GP_ARCH_mips64)
+  TaggedUWord sp;
+  TaggedUWord fp;
+  TaggedUWord pc;
+#else
+#  error "Unknown plat"
+#endif
+};
+
+// The maximum number of bytes in a stack snapshot.  This value can be increased
+// if necessary, but testing showed that 160k is enough to obtain good
+// backtraces on x86_64 Linux.  Most backtraces fit comfortably into 4-8k of
+// stack space, but we do have some very deep stacks occasionally.  Please see
+// the comments in DoNativeBacktrace as to why it's OK to have this value be so
+// large.
+static const size_t N_STACK_BYTES = 160 * 1024;
+
+// The stack chunk image that will be unwound.
+struct StackImage {
+  // [start_avma, +len) specify the address range in the buffer.
+  // Obviously we require 0 <= len <= N_STACK_BYTES.
+  uintptr_t mStartAvma;
+  size_t mLen;
+  uint8_t mContents[N_STACK_BYTES];
+};
+
+// Statistics collection for the unwinder.
+template <typename T>
+class LULStats {
+ public:
+  LULStats() : mContext(0), mCFI(0), mFP(0) {}
+
+  template <typename S>
+  explicit LULStats(const LULStats<S>& aOther)
+      : mContext(aOther.mContext), mCFI(aOther.mCFI), mFP(aOther.mFP) {}
+
+  template <typename S>
+  LULStats<T>& operator=(const LULStats<S>& aOther) {
+    mContext = aOther.mContext;
+    mCFI = aOther.mCFI;
+    mFP = aOther.mFP;
+    return *this;
+  }
+
+  template <typename S>
+  uint32_t operator-(const LULStats<S>& aOther) {
+    return (mContext - aOther.mContext) + (mCFI - aOther.mCFI) +
+           (mFP - aOther.mFP);
+  }
+
+  T mContext;  // Number of context frames
+  T mCFI;      // Number of CFI/EXIDX frames
+  T mFP;       // Number of frame-pointer recovered frames
+};
+
+// The core unwinder library class.  Just one of these is needed, and
+// it can be shared by multiple unwinder threads.
+//
+// The library operates in one of two modes.
+//
+// * Admin mode.  The library is this state after creation.  In Admin
+//   mode, no unwinding may be performed.  It is however allowable to
+//   perform administrative tasks -- primarily, loading of unwind info
+//   -- in this mode.  In particular, it is safe for the library to
+//   perform dynamic memory allocation in this mode.  Safe in the
+//   sense that there is no risk of deadlock against unwinding threads
+//   that might -- because of where they have been sampled -- hold the
+//   system's malloc lock.
+//
+// * Unwind mode.  In this mode, calls to ::Unwind may be made, but
+//   nothing else.  ::Unwind guarantees not to make any dynamic memory
+//   requests, so as to guarantee that the calling thread won't
+//   deadlock in the case where it already holds the system's malloc lock.
+//
+// The library is created in Admin mode.  After debuginfo is loaded,
+// the caller must switch it into Unwind mode by calling
+// ::EnableUnwinding.  There is no way to switch it back to Admin mode
+// after that.  To safely switch back to Admin mode would require the
+// caller (or other external agent) to guarantee that there are no
+// pending ::Unwind calls.
+
+class PriMap;
+class SegArray;
+class UniqueStringUniverse;
+
+class LUL {
+ public:
+  // Create; supply a logging sink.  Sets the object in Admin mode.
+  explicit LUL(void (*aLog)(const char*));
+
+  // Destroy.  Caller is responsible for ensuring that no other
+  // threads are in Unwind calls.  All resources are freed and all
+  // registered unwinder threads are deregistered.  Can be called
+  // either in Admin or Unwind mode.
+  ~LUL();
+
+  // Notify the library that unwinding is now allowed and so
+  // admin-mode calls are no longer allowed.  The object is initially
+  // created in admin mode.  The only possible transition is
+  // admin->unwinding, therefore.
+  void EnableUnwinding();
+
+  // Notify of a new r-x mapping, and load the associated unwind info.
+  // The filename is strdup'd and used for debug printing.  If
+  // aMappedImage is NULL, this function will mmap/munmap the file
+  // itself, so as to be able to read the unwind info.  If
+  // aMappedImage is non-NULL then it is assumed to point to a
+  // called-supplied and caller-managed mapped image of the file.
+  // May only be called in Admin mode.
+  void NotifyAfterMap(uintptr_t aRXavma, size_t aSize, const char* aFileName,
+                      const void* aMappedImage);
+
+  // In rare cases we know an executable area exists but don't know
+  // what the associated file is.  This call notifies LUL of such
+  // areas.  This is important for correct functioning of stack
+  // scanning and of the x86-{linux,android} special-case
+  // __kernel_syscall function handling.
+  // This must be called only after the code area in
+  // question really has been mapped.
+  // May only be called in Admin mode.
+  void NotifyExecutableArea(uintptr_t aRXavma, size_t aSize);
+
+  // Notify that a mapped area has been unmapped; discard any
+  // associated unwind info.  Acquires mRWlock for writing.  Note that
+  // to avoid segfaulting the stack-scan unwinder, which inspects code
+  // areas, this must be called before the code area in question is
+  // really unmapped.  Note that, unlike NotifyAfterMap(), this
+  // function takes the start and end addresses of the range to be
+  // unmapped, rather than a start and a length parameter.  This is so
+  // as to make it possible to notify an unmap for the entire address
+  // space using a single call.
+  // May only be called in Admin mode.
+  void NotifyBeforeUnmap(uintptr_t aAvmaMin, uintptr_t aAvmaMax);
+
+  // Apply NotifyBeforeUnmap to the entire address space.  This causes
+  // LUL to discard all unwind and executable-area information for the
+  // entire address space.
+  // May only be called in Admin mode.
+  void NotifyBeforeUnmapAll() { NotifyBeforeUnmap(0, UINTPTR_MAX); }
+
+  // Returns the number of mappings currently registered.
+  // May only be called in Admin mode.
+  size_t CountMappings();
+
+  // Unwind |aStackImg| starting with the context in |aStartRegs|.
+  // Write the number of frames recovered in *aFramesUsed.  Put
+  // the PC values in aFramePCs[0 .. *aFramesUsed-1] and
+  // the SP values in aFrameSPs[0 .. *aFramesUsed-1].
+  // |aFramesAvail| is the size of the two output arrays and hence the
+  // largest possible value of *aFramesUsed.  PC values are always
+  // valid, and the unwind will stop when the PC becomes invalid, but
+  // the SP values might be invalid, in which case the value zero will
+  // be written in the relevant frameSPs[] slot.
+  //
+  // This function assumes that the SP values increase as it unwinds
+  // away from the innermost frame -- that is, that the stack grows
+  // down.  It monitors SP values as it unwinds to check they
+  // decrease, so as to avoid looping on corrupted stacks.
+  //
+  // May only be called in Unwind mode.  Multiple threads may unwind
+  // at once.  LUL user is responsible for ensuring that no thread makes
+  // any Admin calls whilst in Unwind mode.
+  // MOZ_CRASHes if the calling thread is not registered for unwinding.
+  //
+  // The calling thread must previously have been registered via a call to
+  // RegisterSampledThread.
+  void Unwind(/*OUT*/ uintptr_t* aFramePCs,
+              /*OUT*/ uintptr_t* aFrameSPs,
+              /*OUT*/ size_t* aFramesUsed,
+              /*OUT*/ size_t* aFramePointerFramesAcquired, size_t aFramesAvail,
+              UnwindRegs* aStartRegs, StackImage* aStackImg);
+
+  // The logging sink.  Call to send debug strings to the caller-
+  // specified destination.  Can only be called by the Admin thread.
+  void (*mLog)(const char*);
+
+  // Statistics relating to unwinding.  These have to be atomic since
+  // unwinding can occur on different threads simultaneously.
+  LULStats<mozilla::Atomic<uint32_t>> mStats;
+
+  // Possibly show the statistics.  This may not be called from any
+  // registered sampling thread, since it involves I/O.
+  void MaybeShowStats();
+
+  size_t SizeOfIncludingThis(mozilla::MallocSizeOf) const;
+
+ private:
+  // The statistics counters at the point where they were last printed.
+  LULStats<uint32_t> mStatsPrevious;
+
+  // Are we in admin mode?  Initially |true| but changes to |false|
+  // once unwinding begins.
+  bool mAdminMode;
+
+  // The thread ID associated with admin mode.  This is the only thread
+  // that is allowed do perform non-Unwind calls on this object.  Conversely,
+  // no registered Unwinding thread may be the admin thread.  This is so
+  // as to clearly partition the one thread that may do dynamic memory
+  // allocation from the threads that are being sampled, since the latter
+  // absolutely may not do dynamic memory allocation.
+  int mAdminThreadId;
+
+  // The top level mapping from code address ranges to postprocessed
+  // unwind info.  Basically a sorted array of (addr, len, info)
+  // records.  This field is updated by NotifyAfterMap and NotifyBeforeUnmap.
+  PriMap* mPriMap;
+
+  // An auxiliary structure that records which address ranges are
+  // mapped r-x, for the benefit of the stack scanner.
+  SegArray* mSegArray;
+
+  // A UniqueStringUniverse that holds all the strdup'd strings created
+  // whilst reading unwind information.  This is included so as to make
+  // it possible to free them in ~LUL.
+  UniqueStringUniverse* mUSU;
+};
+
+// Run unit tests on an initialised, loaded-up LUL instance, and print
+// summary results on |aLUL|'s logging sink.  Also return the number
+// of tests run in *aNTests and the number that passed in
+// *aNTestsPassed.
+void RunLulUnitTests(/*OUT*/ int* aNTests, /*OUT*/ int* aNTestsPassed,
+                     LUL* aLUL);
+
+}  // namespace lul
+
+#endif  // LulMain_h
diff --git a/mozglue/baseprofiler/lul/LulMainInt.h b/mozglue/baseprofiler/lul/LulMainInt.h
new file mode 100644
index 0000000000..c2ee45d73d
--- /dev/null
+++ b/mozglue/baseprofiler/lul/LulMainInt.h
@@ -0,0 +1,420 @@
+/* -*- 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 LulMainInt_h
+#define LulMainInt_h
+
+#include "PlatformMacros.h"
+#include "LulMain.h"  // for TaggedUWord
+
+#include "mozilla/Assertions.h"
+
+#include <string>
+#include <vector>
+
+// This file is provides internal interface inside LUL.  If you are an
+// end-user of LUL, do not include it in your code.  The end-user
+// interface is in LulMain.h.
+
+namespace lul {
+
+using std::vector;
+
+////////////////////////////////////////////////////////////////
+// DW_REG_ constants                                          //
+////////////////////////////////////////////////////////////////
+
+// These are the Dwarf CFI register numbers, as (presumably) defined
+// in the ELF ABI supplements for each architecture.
+
+enum DW_REG_NUMBER {
+  // No real register has this number.  It's convenient to be able to
+  // treat the CFA (Canonical Frame Address) as "just another
+  // register", though.
+  DW_REG_CFA = -1,
+#if defined(GP_ARCH_arm)
+  // ARM registers
+  DW_REG_ARM_R7 = 7,
+  DW_REG_ARM_R11 = 11,
+  DW_REG_ARM_R12 = 12,
+  DW_REG_ARM_R13 = 13,
+  DW_REG_ARM_R14 = 14,
+  DW_REG_ARM_R15 = 15,
+#elif defined(GP_ARCH_arm64)
+  // aarch64 registers
+  DW_REG_AARCH64_X29 = 29,
+  DW_REG_AARCH64_X30 = 30,
+  DW_REG_AARCH64_SP = 31,
+#elif defined(GP_ARCH_amd64)
+  // Because the X86 (32 bit) and AMD64 (64 bit) summarisers are
+  // combined, a merged set of register constants is needed.
+  DW_REG_INTEL_XBP = 6,
+  DW_REG_INTEL_XSP = 7,
+  DW_REG_INTEL_XIP = 16,
+#elif defined(GP_ARCH_x86)
+  DW_REG_INTEL_XBP = 5,
+  DW_REG_INTEL_XSP = 4,
+  DW_REG_INTEL_XIP = 8,
+#elif defined(GP_ARCH_mips64)
+  DW_REG_MIPS_SP = 29,
+  DW_REG_MIPS_FP = 30,
+  DW_REG_MIPS_PC = 34,
+#else
+#  error "Unknown arch"
+#endif
+};
+
+////////////////////////////////////////////////////////////////
+// PfxExpr                                                    //
+////////////////////////////////////////////////////////////////
+
+enum PfxExprOp {
+  //             meaning of mOperand     effect on stack
+  PX_Start,   // bool start-with-CFA?    start, with CFA on stack, or not
+  PX_End,     // none                    stop; result is at top of stack
+  PX_SImm32,  // int32                   push signed int32
+  PX_DwReg,   // DW_REG_NUMBER           push value of the specified reg
+  PX_Deref,   // none                    pop X ; push *X
+  PX_Add,     // none                    pop X ; pop Y ; push Y + X
+  PX_Sub,     // none                    pop X ; pop Y ; push Y - X
+  PX_And,     // none                    pop X ; pop Y ; push Y & X
+  PX_Or,      // none                    pop X ; pop Y ; push Y | X
+  PX_CmpGES,  // none                    pop X ; pop Y ; push (Y >=s X) ? 1 : 0
+  PX_Shl      // none                    pop X ; pop Y ; push Y << X
+};
+
+struct PfxInstr {
+  PfxInstr(PfxExprOp opcode, int32_t operand)
+      : mOpcode(opcode), mOperand(operand) {}
+  explicit PfxInstr(PfxExprOp opcode) : mOpcode(opcode), mOperand(0) {}
+  bool operator==(const PfxInstr& other) const {
+    return mOpcode == other.mOpcode && mOperand == other.mOperand;
+  }
+  PfxExprOp mOpcode;
+  int32_t mOperand;
+};
+
+static_assert(sizeof(PfxInstr) <= 8, "PfxInstr size changed unexpectedly");
+
+// Evaluate the prefix expression whose PfxInstrs start at aPfxInstrs[start].
+// In the case of any mishap (stack over/underflow, running off the end of
+// the instruction vector, obviously malformed sequences),
+// return an invalid TaggedUWord.
+// RUNS IN NO-MALLOC CONTEXT
+TaggedUWord EvaluatePfxExpr(int32_t start, const UnwindRegs* aOldRegs,
+                            TaggedUWord aCFA, const StackImage* aStackImg,
+                            const vector<PfxInstr>& aPfxInstrs);
+
+////////////////////////////////////////////////////////////////
+// LExpr                                                      //
+////////////////////////////////////////////////////////////////
+
+// An expression -- very primitive.  Denotes either "register +
+// offset", a dereferenced version of the same, or a reference to a
+// prefix expression stored elsewhere.  So as to allow convenient
+// handling of Dwarf-derived unwind info, the register may also denote
+// the CFA.  A large number of these need to be stored, so we ensure
+// it fits into 8 bytes.  See comment below on RuleSet to see how
+// expressions fit into the bigger picture.
+
+enum LExprHow {
+  UNKNOWN = 0,  // This LExpr denotes no value.
+  NODEREF,      // Value is  (mReg + mOffset).
+  DEREF,        // Value is *(mReg + mOffset).
+  PFXEXPR       // Value is EvaluatePfxExpr(secMap->mPfxInstrs[mOffset])
+};
+
+inline static const char* NameOf_LExprHow(LExprHow how) {
+  switch (how) {
+    case UNKNOWN:
+      return "UNKNOWN";
+    case NODEREF:
+      return "NODEREF";
+    case DEREF:
+      return "DEREF";
+    case PFXEXPR:
+      return "PFXEXPR";
+    default:
+      return "LExpr-??";
+  }
+}
+
+struct LExpr {
+  // Denotes an expression with no value.
+  LExpr() : mHow(UNKNOWN), mReg(0), mOffset(0) {}
+
+  // Denotes any expressible expression.
+  LExpr(LExprHow how, int16_t reg, int32_t offset)
+      : mHow(how), mReg(reg), mOffset(offset) {
+    switch (how) {
+      case UNKNOWN:
+        MOZ_ASSERT(reg == 0 && offset == 0);
+        break;
+      case NODEREF:
+        break;
+      case DEREF:
+        break;
+      case PFXEXPR:
+        MOZ_ASSERT(reg == 0 && offset >= 0);
+        break;
+      default:
+        MOZ_ASSERT(0, "LExpr::LExpr: invalid how");
+    }
+  }
+
+  // Change the offset for an expression that references memory.
+  LExpr add_delta(long delta) {
+    MOZ_ASSERT(mHow == NODEREF);
+    // If this is a non-debug build and the above assertion would have
+    // failed, at least return LExpr() so that the machinery that uses
+    // the resulting expression fails in a repeatable way.
+    return (mHow == NODEREF) ? LExpr(mHow, mReg, mOffset + delta)
+                             : LExpr();  // Gone bad
+  }
+
+  // Dereference an expression that denotes a memory address.
+  LExpr deref() {
+    MOZ_ASSERT(mHow == NODEREF);
+    // Same rationale as for add_delta().
+    return (mHow == NODEREF) ? LExpr(DEREF, mReg, mOffset)
+                             : LExpr();  // Gone bad
+  }
+
+  // Print a rule for recovery of |aNewReg| whose recovered value
+  // is this LExpr.
+  std::string ShowRule(const char* aNewReg) const;
+
+  // Evaluate this expression, producing a TaggedUWord.  |aOldRegs|
+  // holds register values that may be referred to by the expression.
+  // |aCFA| holds the CFA value, if any, that applies.  |aStackImg|
+  // contains a chuck of stack that will be consulted if the expression
+  // references memory.  |aPfxInstrs| holds the vector of PfxInstrs
+  // that will be consulted if this is a PFXEXPR.
+  // RUNS IN NO-MALLOC CONTEXT
+  TaggedUWord EvaluateExpr(const UnwindRegs* aOldRegs, TaggedUWord aCFA,
+                           const StackImage* aStackImg,
+                           const vector<PfxInstr>* aPfxInstrs) const;
+
+  // Representation of expressions.  If |mReg| is DW_REG_CFA (-1) then
+  // it denotes the CFA.  All other allowed values for |mReg| are
+  // nonnegative and are DW_REG_ values.
+  LExprHow mHow : 8;
+  int16_t mReg;     // A DW_REG_ value
+  int32_t mOffset;  // 32-bit signed offset should be more than enough.
+};
+
+static_assert(sizeof(LExpr) <= 8, "LExpr size changed unexpectedly");
+
+////////////////////////////////////////////////////////////////
+// RuleSet                                                    //
+////////////////////////////////////////////////////////////////
+
+// This is platform-dependent.  For some address range, describes how
+// to recover the CFA and then how to recover the registers for the
+// previous frame.
+//
+// The set of LExprs contained in a given RuleSet describe a DAG which
+// says how to compute the caller's registers ("new registers") from
+// the callee's registers ("old registers").  The DAG can contain a
+// single internal node, which is the value of the CFA for the callee.
+// It would be possible to construct a DAG that omits the CFA, but
+// including it makes the summarisers simpler, and the Dwarf CFI spec
+// has the CFA as a central concept.
+//
+// For this to make sense, |mCfaExpr| can't have
+// |mReg| == DW_REG_CFA since we have no previous value for the CFA.
+// All of the other |Expr| fields can -- and usually do -- specify
+// |mReg| == DW_REG_CFA.
+//
+// With that in place, the unwind algorithm proceeds as follows.
+//
+// (0) Initially: we have values for the old registers, and a memory
+//     image.
+//
+// (1) Compute the CFA by evaluating |mCfaExpr|.  Add the computed
+//     value to the set of "old registers".
+//
+// (2) Compute values for the registers by evaluating all of the other
+//     |Expr| fields in the RuleSet.  These can depend on both the old
+//     register values and the just-computed CFA.
+//
+// If we are unwinding without computing a CFA, perhaps because the
+// RuleSets are derived from EXIDX instead of Dwarf, then
+// |mCfaExpr.mHow| will be LExpr::UNKNOWN, so the computed value will
+// be invalid -- that is, TaggedUWord() -- and so any attempt to use
+// that will result in the same value.  But that's OK because the
+// RuleSet would make no sense if depended on the CFA but specified no
+// way to compute it.
+//
+// A RuleSet is not allowed to cover zero address range.  Having zero
+// length would break binary searching in SecMaps and PriMaps.
+
+class RuleSet {
+ public:
+  RuleSet();
+  void Print(void (*aLog)(const char*)) const;
+
+  // Find the LExpr* for a given DW_REG_ value in this class.
+  LExpr* ExprForRegno(DW_REG_NUMBER aRegno);
+
+  uintptr_t mAddr;
+  uintptr_t mLen;
+  // How to compute the CFA.
+  LExpr mCfaExpr;
+  // How to compute caller register values.  These may reference the
+  // value defined by |mCfaExpr|.
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+  LExpr mXipExpr;  // return address
+  LExpr mXspExpr;
+  LExpr mXbpExpr;
+#elif defined(GP_ARCH_arm)
+  LExpr mR15expr;  // return address
+  LExpr mR14expr;
+  LExpr mR13expr;
+  LExpr mR12expr;
+  LExpr mR11expr;
+  LExpr mR7expr;
+#elif defined(GP_ARCH_arm64)
+  LExpr mX29expr;  // frame pointer register
+  LExpr mX30expr;  // link register
+  LExpr mSPexpr;
+#elif defined(GP_ARCH_mips64)
+  LExpr mPCexpr;
+  LExpr mFPexpr;
+  LExpr mSPexpr;
+#else
+#  error "Unknown arch"
+#endif
+};
+
+// Returns |true| for Dwarf register numbers which are members
+// of the set of registers that LUL unwinds on this target.
+static inline bool registerIsTracked(DW_REG_NUMBER reg) {
+  switch (reg) {
+#if defined(GP_ARCH_amd64) || defined(GP_ARCH_x86)
+    case DW_REG_INTEL_XBP:
+    case DW_REG_INTEL_XSP:
+    case DW_REG_INTEL_XIP:
+      return true;
+#elif defined(GP_ARCH_arm)
+    case DW_REG_ARM_R7:
+    case DW_REG_ARM_R11:
+    case DW_REG_ARM_R12:
+    case DW_REG_ARM_R13:
+    case DW_REG_ARM_R14:
+    case DW_REG_ARM_R15:
+      return true;
+#elif defined(GP_ARCH_arm64)
+    case DW_REG_AARCH64_X29:
+    case DW_REG_AARCH64_X30:
+    case DW_REG_AARCH64_SP:
+      return true;
+#elif defined(GP_ARCH_mips64)
+    case DW_REG_MIPS_FP:
+    case DW_REG_MIPS_SP:
+    case DW_REG_MIPS_PC:
+      return true;
+#else
+#  error "Unknown arch"
+#endif
+    default:
+      return false;
+  }
+}
+
+////////////////////////////////////////////////////////////////
+// SecMap                                                     //
+////////////////////////////////////////////////////////////////
+
+// A SecMap may have zero address range, temporarily, whilst RuleSets
+// are being added to it.  But adding a zero-range SecMap to a PriMap
+// will make it impossible to maintain the total order of the PriMap
+// entries, and so that can't be allowed to happen.
+
+class SecMap {
+ public:
+  // These summarise the contained mRuleSets, in that they give
+  // exactly the lowest and highest addresses that any of the entries
+  // in this SecMap cover.  Hence invariants:
+  //
+  // mRuleSets is nonempty
+  //    <=> mSummaryMinAddr <= mSummaryMaxAddr
+  //        && mSummaryMinAddr == mRuleSets[0].mAddr
+  //        && mSummaryMaxAddr == mRuleSets[#rulesets-1].mAddr
+  //                              + mRuleSets[#rulesets-1].mLen - 1;
+  //
+  // This requires that no RuleSet has zero length.
+  //
+  // mRuleSets is empty
+  //    <=> mSummaryMinAddr > mSummaryMaxAddr
+  //
+  // This doesn't constrain mSummaryMinAddr and mSummaryMaxAddr uniquely,
+  // so let's use mSummaryMinAddr == 1 and mSummaryMaxAddr == 0 to denote
+  // this case.
+
+  explicit SecMap(void (*aLog)(const char*));
+  ~SecMap();
+
+  // Binary search mRuleSets to find one that brackets |ia|, or nullptr
+  // if none is found.  It's not allowable to do this until PrepareRuleSets
+  // has been called first.
+  RuleSet* FindRuleSet(uintptr_t ia);
+
+  // Add a RuleSet to the collection.  The rule is copied in.  Calling
+  // this makes the map non-searchable.
+  void AddRuleSet(const RuleSet* rs);
+
+  // Add a PfxInstr to the vector of such instrs, and return the index
+  // in the vector.  Calling this makes the map non-searchable.
+  uint32_t AddPfxInstr(PfxInstr pfxi);
+
+  // Returns the entire vector of PfxInstrs.
+  const vector<PfxInstr>* GetPfxInstrs() { return &mPfxInstrs; }
+
+  // Prepare the map for searching.  Also, remove any rules for code
+  // address ranges which don't fall inside [start, +len).  |len| may
+  // not be zero.
+  void PrepareRuleSets(uintptr_t start, size_t len);
+
+  bool IsEmpty();
+
+  size_t Size() { return mRuleSets.size(); }
+
+  size_t SizeOfIncludingThis(mozilla::MallocSizeOf aMallocSizeOf) const;
+
+  // The min and max addresses of the addresses in the contained
+  // RuleSets.  See comment above for invariants.
+  uintptr_t mSummaryMinAddr;
+  uintptr_t mSummaryMaxAddr;
+
+ private:
+  // False whilst adding entries; true once it is safe to call FindRuleSet.
+  // Transition (false->true) is caused by calling PrepareRuleSets().
+  bool mUsable;
+
+  // A vector of RuleSets, sorted, nonoverlapping (post Prepare()).
+  vector<RuleSet> mRuleSets;
+
+  // A vector of PfxInstrs, which are referred to by the RuleSets.
+  // These are provided as a representation of Dwarf expressions
+  // (DW_CFA_val_expression, DW_CFA_expression, DW_CFA_def_cfa_expression),
+  // are relatively expensive to evaluate, and and are therefore
+  // expected to be used only occasionally.
+  //
+  // The vector holds a bunch of separate PfxInstr programs, each one
+  // starting with a PX_Start and terminated by a PX_End, all
+  // concatenated together.  When a RuleSet can't recover a value
+  // using a self-contained LExpr, it uses a PFXEXPR whose mOffset is
+  // the index in this vector of start of the necessary PfxInstr program.
+  vector<PfxInstr> mPfxInstrs;
+
+  // A logging sink, for debugging.
+  void (*mLog)(const char*);
+};
+
+}  // namespace lul
+
+#endif  // ndef LulMainInt_h
diff --git a/mozglue/baseprofiler/lul/platform-linux-lul.cpp b/mozglue/baseprofiler/lul/platform-linux-lul.cpp
new file mode 100644
index 0000000000..a9ee65858d
--- /dev/null
+++ b/mozglue/baseprofiler/lul/platform-linux-lul.cpp
@@ -0,0 +1,76 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* 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 <signal.h>
+#include <string.h>
+#include <stdlib.h>
+#include <time.h>
+
+#include "AutoObjectMapper.h"
+#include "BaseProfiler.h"
+#include "BaseProfilerSharedLibraries.h"
+#include "platform.h"
+#include "PlatformMacros.h"
+#include "LulMain.h"
+
+// Contains miscellaneous helpers that are used to connect the Gecko Profiler
+// and LUL.
+
+// Find out, in a platform-dependent way, where the code modules got
+// mapped in the process' virtual address space, and get |aLUL| to
+// load unwind info for them.
+void read_procmaps(lul::LUL* aLUL) {
+  MOZ_ASSERT(aLUL->CountMappings() == 0);
+
+#if defined(GP_OS_linux) || defined(GP_OS_android) || defined(GP_OS_freebsd)
+  SharedLibraryInfo info = SharedLibraryInfo::GetInfoForSelf();
+
+  for (size_t i = 0; i < info.GetSize(); i++) {
+    const SharedLibrary& lib = info.GetEntry(i);
+
+    std::string nativePath = lib.GetDebugPath();
+
+    // We can use the standard POSIX-based mapper.
+    AutoObjectMapperPOSIX mapper(aLUL->mLog);
+
+    // Ask |mapper| to map the object.  Then hand its mapped address
+    // to NotifyAfterMap().
+    void* image = nullptr;
+    size_t size = 0;
+    bool ok = mapper.Map(&image, &size, nativePath);
+    if (ok && image && size > 0) {
+      aLUL->NotifyAfterMap(lib.GetStart(), lib.GetEnd() - lib.GetStart(),
+                           nativePath.c_str(), image);
+    } else if (!ok && lib.GetDebugName().empty()) {
+      // The object has no name and (as a consequence) the mapper failed to map
+      // it.  This happens on Linux, where GetInfoForSelf() produces such a
+      // mapping for the VDSO.  This is a problem on x86-{linux,android} because
+      // lack of knowledge about the mapped area inhibits LUL's special
+      // __kernel_syscall handling.  Hence notify |aLUL| at least of the
+      // mapping, even though it can't read any unwind information for the area.
+      aLUL->NotifyExecutableArea(lib.GetStart(), lib.GetEnd() - lib.GetStart());
+    }
+
+    // |mapper| goes out of scope at this point and so its destructor
+    // unmaps the object.
+  }
+
+#else
+#  error "Unknown platform"
+#endif
+}
+
+// LUL needs a callback for its logging sink.
+void logging_sink_for_LUL(const char* str) {
+  // These are only printed when Verbose logging is enabled (e.g. with
+  // MOZ_BASE_PROFILER_VERBOSE_LOGGING=1). This is because LUL's logging is much
+  // more verbose than the rest of the profiler's logging, which occurs at the
+  // Info (3) and Debug (4) levels.
+  // FIXME: This causes a build failure in memory/replace/dmd/test/SmokeDMD (!)
+  // and other places, because it doesn't link the implementation in
+  // platform.cpp.
+  // VERBOSE_LOG("[%d] %s", profiler_current_process_id(), str);
+}
diff --git a/mozglue/baseprofiler/lul/platform-linux-lul.h b/mozglue/baseprofiler/lul/platform-linux-lul.h
new file mode 100644
index 0000000000..b54e80edcf
--- /dev/null
+++ b/mozglue/baseprofiler/lul/platform-linux-lul.h
@@ -0,0 +1,21 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* 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 MOZ_PLATFORM_LINUX_LUL_H
+#define MOZ_PLATFORM_LINUX_LUL_H
+
+#include "platform.h"
+
+#include "BaseProfiler.h"
+
+// Find out, in a platform-dependent way, where the code modules got
+// mapped in the process' virtual address space, and get |aLUL| to
+// load unwind info for them.
+void read_procmaps(lul::LUL* aLUL);
+
+// LUL needs a callback for its logging sink.
+void logging_sink_for_LUL(const char* str);
+
+#endif /* ndef MOZ_PLATFORM_LINUX_LUL_H */
diff --git a/mozglue/baseprofiler/moz.build b/mozglue/baseprofiler/moz.build
new file mode 100644
index 0000000000..db2949863f
--- /dev/null
+++ b/mozglue/baseprofiler/moz.build
@@ -0,0 +1,118 @@
+# -*- 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/.
+
+# This is pretty much a copy from tools/profiler, cut down to exclude anything
+# that cannot work in mozglue (because they are totally dependent on libxul-
+# specific code).
+# All exported headers now prefixed with "Base" to avoid #include name clashes.
+
+if CONFIG["MOZ_GECKO_PROFILER"]:
+    DEFINES["IMPL_MFBT"] = True
+    EXPORTS += [
+        "public/BaseProfilerSharedLibraries.h",
+        "public/BaseProfilingCategory.h",
+        "public/BaseProfilingStack.h",
+        "public/ProfilingCategoryList.h",
+    ]
+    UNIFIED_SOURCES += [
+        "core/PageInformation.cpp",
+        "core/platform.cpp",
+        "core/ProfileBuffer.cpp",
+        "core/ProfileBufferEntry.cpp",
+        "core/ProfiledThreadData.cpp",
+        "core/ProfileJSONWriter.cpp",
+        "core/ProfilerBacktrace.cpp",
+        "core/ProfilerMarkers.cpp",
+        "core/ProfilingCategory.cpp",
+        "core/ProfilingStack.cpp",
+        "core/RegisteredThread.cpp",
+    ]
+
+    if CONFIG["OS_TARGET"] in ("Android", "Linux", "FreeBSD"):
+        if CONFIG["CPU_ARCH"] in ("arm", "aarch64", "x86", "x86_64", "mips64"):
+            UNIFIED_SOURCES += [
+                "lul/AutoObjectMapper.cpp",
+                "lul/LulCommon.cpp",
+                "lul/LulDwarf.cpp",
+                "lul/LulDwarfSummariser.cpp",
+                "lul/LulElf.cpp",
+                "lul/LulMain.cpp",
+                "lul/platform-linux-lul.cpp",
+            ]
+        # These files cannot be built in unified mode because of name clashes with mozglue headers on Android.
+        SOURCES += [
+            "core/shared-libraries-linux.cc",
+        ]
+        if CONFIG["CPU_ARCH"] == "arm" and CONFIG["OS_TARGET"] != "FreeBSD":
+            SOURCES += [
+                "core/EHABIStackWalk.cpp",
+            ]
+    elif CONFIG["OS_TARGET"] == "Darwin":
+        UNIFIED_SOURCES += [
+            "core/shared-libraries-macos.cc",
+        ]
+    elif CONFIG["OS_TARGET"] == "WINNT":
+        SOURCES += [
+            "core/shared-libraries-win32.cc",
+        ]
+
+    LOCAL_INCLUDES += [
+        "/mozglue/baseprofiler/core/",
+        "/mozglue/linker",
+    ]
+
+    if CONFIG["OS_TARGET"] == "Android":
+        DEFINES["ANDROID_NDK_MAJOR_VERSION"] = CONFIG["ANDROID_NDK_MAJOR_VERSION"]
+        DEFINES["ANDROID_NDK_MINOR_VERSION"] = CONFIG["ANDROID_NDK_MINOR_VERSION"]
+
+    FINAL_LIBRARY = "mozglue"
+
+# Many of the headers in this directory are usable in non-MOZ_GECKO_PROFILER
+# builds.
+# BaseProfiler.h and BaseProfilerCounts.h only contain no-op macros in that
+# case.
+EXPORTS += [
+    "public/BaseProfiler.h",
+]
+
+EXPORTS.mozilla += [
+    "public/BaseProfileJSONWriter.h",
+    "public/BaseProfilerCounts.h",
+    "public/BaseProfilerDetail.h",
+    "public/BaseProfilerMarkers.h",
+    "public/BaseProfilerMarkersDetail.h",
+    "public/BaseProfilerMarkersPrerequisites.h",
+    "public/BaseProfilerMarkerTypes.h",
+    "public/BlocksRingBuffer.h",
+    "public/leb128iterator.h",
+    "public/ModuloBuffer.h",
+    "public/PowerOfTwo.h",
+    "public/ProfileBufferChunk.h",
+    "public/ProfileBufferChunkManager.h",
+    "public/ProfileBufferChunkManagerSingle.h",
+    "public/ProfileBufferChunkManagerWithLocalLimit.h",
+    "public/ProfileBufferControlledChunkManager.h",
+    "public/ProfileBufferEntryKinds.h",
+    "public/ProfileBufferEntrySerialization.h",
+    "public/ProfileBufferIndex.h",
+    "public/ProfileChunkedBuffer.h",
+]
+
+if CONFIG["MOZ_VTUNE"]:
+    DEFINES["MOZ_VTUNE_INSTRUMENTATION"] = True
+    UNIFIED_SOURCES += [
+        "core/VTuneProfiler.cpp",
+    ]
+
+
+if CONFIG["CC_TYPE"] in ("clang", "gcc"):
+    CXXFLAGS += [
+        "-Wno-error=shadow",
+        "-Wno-ignored-qualifiers",  # due to use of breakpad headers
+    ]
+
+with Files("**"):
+    BUG_COMPONENT = ("Core", "Gecko Profiler")
diff --git a/mozglue/baseprofiler/public/BaseProfileJSONWriter.h b/mozglue/baseprofiler/public/BaseProfileJSONWriter.h
new file mode 100644
index 0000000000..5dcf06f3f3
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfileJSONWriter.h
@@ -0,0 +1,388 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* 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 BASEPROFILEJSONWRITER_H
+#define BASEPROFILEJSONWRITER_H
+
+#include "mozilla/HashFunctions.h"
+#include "mozilla/HashTable.h"
+#include "mozilla/JSONWriter.h"
+#include "mozilla/TimeStamp.h"
+#include "mozilla/UniquePtr.h"
+
+#include <functional>
+#include <ostream>
+#include <string_view>
+
+namespace mozilla {
+namespace baseprofiler {
+
+class SpliceableJSONWriter;
+
+// On average, profile JSONs are large enough such that we want to avoid
+// reallocating its buffer when expanding. Additionally, the contents of the
+// profile are not accessed until the profile is entirely written. For these
+// reasons we use a chunked writer that keeps an array of chunks, which is
+// concatenated together after writing is finished.
+class ChunkedJSONWriteFunc final : public JSONWriteFunc {
+ public:
+  friend class SpliceableJSONWriter;
+
+  ChunkedJSONWriteFunc() : mChunkPtr{nullptr}, mChunkEnd{nullptr} {
+    AllocChunk(kChunkSize);
+  }
+
+  bool IsEmpty() const {
+    MOZ_ASSERT_IF(!mChunkPtr, !mChunkEnd && mChunkList.length() == 0 &&
+                                  mChunkLengths.length() == 0);
+    return !mChunkPtr;
+  }
+
+  void Write(const Span<const char>& aStr) override {
+    MOZ_ASSERT(mChunkPtr >= mChunkList.back().get() && mChunkPtr <= mChunkEnd);
+    MOZ_ASSERT(mChunkEnd >= mChunkList.back().get() + mChunkLengths.back());
+    MOZ_ASSERT(*mChunkPtr == '\0');
+
+    // Most strings to be written are small, but subprocess profiles (e.g.,
+    // from the content process in e10s) may be huge. If the string is larger
+    // than a chunk, allocate its own chunk.
+    char* newPtr;
+    if (aStr.size() >= kChunkSize) {
+      AllocChunk(aStr.size() + 1);
+      newPtr = mChunkPtr + aStr.size();
+    } else {
+      newPtr = mChunkPtr + aStr.size();
+      if (newPtr >= mChunkEnd) {
+        AllocChunk(kChunkSize);
+        newPtr = mChunkPtr + aStr.size();
+      }
+    }
+
+    memcpy(mChunkPtr, aStr.data(), aStr.size());
+    *newPtr = '\0';
+    mChunkPtr = newPtr;
+    mChunkLengths.back() += aStr.size();
+  }
+  void CopyDataIntoLazilyAllocatedBuffer(
+      const std::function<char*(size_t)>& aAllocator) const {
+    // Request a buffer for the full content plus a null terminator.
+    MOZ_ASSERT(mChunkLengths.length() == mChunkList.length());
+    size_t totalLen = 1;
+    for (size_t i = 0; i < mChunkLengths.length(); i++) {
+      MOZ_ASSERT(strlen(mChunkList[i].get()) == mChunkLengths[i]);
+      totalLen += mChunkLengths[i];
+    }
+    char* ptr = aAllocator(totalLen);
+
+    if (!ptr) {
+      // Failed to allocate memory.
+      return;
+    }
+
+    for (size_t i = 0; i < mChunkList.length(); i++) {
+      size_t len = mChunkLengths[i];
+      memcpy(ptr, mChunkList[i].get(), len);
+      ptr += len;
+    }
+    *ptr = '\0';
+  }
+  UniquePtr<char[]> CopyData() const {
+    UniquePtr<char[]> c;
+    CopyDataIntoLazilyAllocatedBuffer([&](size_t allocationSize) {
+      c = MakeUnique<char[]>(allocationSize);
+      return c.get();
+    });
+    return c;
+  }
+  void Take(ChunkedJSONWriteFunc&& aOther) {
+    for (size_t i = 0; i < aOther.mChunkList.length(); i++) {
+      MOZ_ALWAYS_TRUE(mChunkLengths.append(aOther.mChunkLengths[i]));
+      MOZ_ALWAYS_TRUE(mChunkList.append(std::move(aOther.mChunkList[i])));
+    }
+    mChunkPtr = mChunkList.back().get() + mChunkLengths.back();
+    mChunkEnd = mChunkPtr;
+    aOther.mChunkPtr = nullptr;
+    aOther.mChunkEnd = nullptr;
+    aOther.mChunkList.clear();
+    aOther.mChunkLengths.clear();
+  }
+
+ private:
+  void AllocChunk(size_t aChunkSize) {
+    MOZ_ASSERT(mChunkLengths.length() == mChunkList.length());
+    UniquePtr<char[]> newChunk = MakeUnique<char[]>(aChunkSize);
+    mChunkPtr = newChunk.get();
+    mChunkEnd = mChunkPtr + aChunkSize;
+    *mChunkPtr = '\0';
+    MOZ_ALWAYS_TRUE(mChunkLengths.append(0));
+    MOZ_ALWAYS_TRUE(mChunkList.append(std::move(newChunk)));
+  }
+
+  static const size_t kChunkSize = 4096 * 512;
+
+  // Pointer for writing inside the current chunk.
+  //
+  // The current chunk is always at the back of mChunkList, i.e.,
+  // mChunkList.back() <= mChunkPtr <= mChunkEnd.
+  char* mChunkPtr;
+
+  // Pointer to the end of the current chunk.
+  //
+  // The current chunk is always at the back of mChunkList, i.e.,
+  // mChunkEnd >= mChunkList.back() + mChunkLengths.back().
+  char* mChunkEnd;
+
+  // List of chunks and their lengths.
+  //
+  // For all i, the length of the string in mChunkList[i] is
+  // mChunkLengths[i].
+  Vector<UniquePtr<char[]>> mChunkList;
+  Vector<size_t> mChunkLengths;
+};
+
+struct OStreamJSONWriteFunc final : public JSONWriteFunc {
+  explicit OStreamJSONWriteFunc(std::ostream& aStream) : mStream(aStream) {}
+
+  void Write(const Span<const char>& aStr) override {
+    std::string_view sv(aStr.data(), aStr.size());
+    mStream << sv;
+  }
+
+  std::ostream& mStream;
+};
+
+class UniqueJSONStrings;
+
+class SpliceableJSONWriter : public JSONWriter {
+ public:
+  explicit SpliceableJSONWriter(UniquePtr<JSONWriteFunc> aWriter)
+      : JSONWriter(std::move(aWriter)) {}
+
+  void StartBareList(CollectionStyle aStyle = MultiLineStyle) {
+    StartCollection(scEmptyString, scEmptyString, aStyle);
+  }
+
+  void EndBareList() { EndCollection(scEmptyString); }
+
+  // This function must be used to correctly stream timestamps in profiles.
+  // Null timestamps don't output anything.
+  void TimeProperty(const Span<const char>& aName, const TimeStamp& aTime) {
+    if (!aTime.IsNull()) {
+      DoubleProperty(aName,
+                     (aTime - TimeStamp::ProcessCreation()).ToMilliseconds());
+    }
+  }
+
+  void NullElements(uint32_t aCount) {
+    for (uint32_t i = 0; i < aCount; i++) {
+      NullElement();
+    }
+  }
+
+  void Splice(const Span<const char>& aStr) {
+    Separator();
+    WriteFunc()->Write(aStr);
+    mNeedComma[mDepth] = true;
+  }
+
+  void Splice(const char* aStr, size_t aLen) {
+    Separator();
+    WriteFunc()->Write(Span<const char>(aStr, aLen));
+    mNeedComma[mDepth] = true;
+  }
+
+  // Splice the given JSON directly in, without quoting.
+  void SplicedJSONProperty(const Span<const char>& aMaybePropertyName,
+                           const Span<const char>& aJsonValue) {
+    Scalar(aMaybePropertyName, aJsonValue);
+  }
+
+  void CopyAndSplice(const ChunkedJSONWriteFunc& aFunc) {
+    Separator();
+    for (size_t i = 0; i < aFunc.mChunkList.length(); i++) {
+      WriteFunc()->Write(
+          Span<const char>(aFunc.mChunkList[i].get(), aFunc.mChunkLengths[i]));
+    }
+    mNeedComma[mDepth] = true;
+  }
+
+  // Takes the chunks from aFunc and write them. If move is not possible
+  // (e.g., using OStreamJSONWriteFunc), aFunc's chunks are copied and its
+  // storage cleared.
+  virtual void TakeAndSplice(ChunkedJSONWriteFunc&& aFunc) {
+    Separator();
+    for (size_t i = 0; i < aFunc.mChunkList.length(); i++) {
+      WriteFunc()->Write(
+          Span<const char>(aFunc.mChunkList[i].get(), aFunc.mChunkLengths[i]));
+    }
+    aFunc.mChunkPtr = nullptr;
+    aFunc.mChunkEnd = nullptr;
+    aFunc.mChunkList.clear();
+    aFunc.mChunkLengths.clear();
+    mNeedComma[mDepth] = true;
+  }
+
+  // Set (or reset) the pointer to a UniqueJSONStrings.
+  void SetUniqueStrings(UniqueJSONStrings& aUniqueStrings) {
+    MOZ_RELEASE_ASSERT(!mUniqueStrings);
+    mUniqueStrings = &aUniqueStrings;
+  }
+
+  // Set (or reset) the pointer to a UniqueJSONStrings.
+  void ResetUniqueStrings() {
+    MOZ_RELEASE_ASSERT(mUniqueStrings);
+    mUniqueStrings = nullptr;
+  }
+
+  // Add `aStr` to the unique-strings list (if not already there), and write its
+  // index as a named object property.
+  inline void UniqueStringProperty(const Span<const char>& aName,
+                                   const Span<const char>& aStr);
+
+  // Add `aStr` to the unique-strings list (if not already there), and write its
+  // index as an array element.
+  inline void UniqueStringElement(const Span<const char>& aStr);
+
+ private:
+  UniqueJSONStrings* mUniqueStrings = nullptr;
+};
+
+class SpliceableChunkedJSONWriter final : public SpliceableJSONWriter {
+ public:
+  explicit SpliceableChunkedJSONWriter()
+      : SpliceableJSONWriter(MakeUnique<ChunkedJSONWriteFunc>()) {}
+
+  // Access the ChunkedJSONWriteFunc as reference-to-const, usually to copy data
+  // out.
+  const ChunkedJSONWriteFunc& ChunkedWriteFunc() const {
+    MOZ_ASSERT(!mTaken);
+    // The WriteFunc was non-fallibly allocated as a ChunkedJSONWriteFunc in the
+    // only constructor above, so it's safe to cast to ChunkedJSONWriteFunc*.
+    return *static_cast<const ChunkedJSONWriteFunc*>(WriteFunc());
+  }
+
+  // Access the ChunkedJSONWriteFunc as rvalue-reference, usually to take its
+  // data out. This writer shouldn't be used anymore after this.
+  ChunkedJSONWriteFunc&& TakeChunkedWriteFunc() {
+#ifdef DEBUG
+    MOZ_ASSERT(!mTaken);
+    mTaken = true;
+#endif  //
+    // The WriteFunc was non-fallibly allocated as a ChunkedJSONWriteFunc in the
+    // only constructor above, so it's safe to cast to ChunkedJSONWriteFunc*.
+    return std::move(*static_cast<ChunkedJSONWriteFunc*>(WriteFunc()));
+  }
+
+  // Adopts the chunks from aFunc without copying.
+  void TakeAndSplice(ChunkedJSONWriteFunc&& aFunc) override {
+    MOZ_ASSERT(!mTaken);
+    Separator();
+    // The WriteFunc was non-fallibly allocated as a ChunkedJSONWriteFunc in the
+    // only constructor above, so it's safe to cast to ChunkedJSONWriteFunc*.
+    static_cast<ChunkedJSONWriteFunc*>(WriteFunc())->Take(std::move(aFunc));
+    mNeedComma[mDepth] = true;
+  }
+
+#ifdef DEBUG
+ private:
+  bool mTaken = false;
+#endif  //
+};
+
+class JSONSchemaWriter {
+  JSONWriter& mWriter;
+  uint32_t mIndex;
+
+ public:
+  explicit JSONSchemaWriter(JSONWriter& aWriter) : mWriter(aWriter), mIndex(0) {
+    aWriter.StartObjectProperty("schema",
+                                SpliceableJSONWriter::SingleLineStyle);
+  }
+
+  void WriteField(const Span<const char>& aName) {
+    mWriter.IntProperty(aName, mIndex++);
+  }
+
+  template <size_t Np1>
+  void WriteField(const char (&aName)[Np1]) {
+    WriteField(Span<const char>(aName, Np1 - 1));
+  }
+
+  ~JSONSchemaWriter() { mWriter.EndObject(); }
+};
+
+// This class helps create an indexed list of unique strings, and inserts the
+// index as a JSON value. The collected list of unique strings can later be
+// inserted as a JSON array.
+// This can be useful for elements/properties with many repeated strings.
+//
+// With only JSONWriter w,
+// `w.WriteElement("a"); w.WriteElement("b"); w.WriteElement("a");`
+// when done inside a JSON array, will generate:
+// `["a", "b", "c"]`
+//
+// With UniqueStrings u,
+// `u.WriteElement(w, "a"); u.WriteElement(w, "b"); u.WriteElement(w, "a");`
+// when done inside a JSON array, will generate:
+// `[0, 1, 0]`
+// and later, `u.SpliceStringTableElements(w)` (inside a JSON array), will
+// output the corresponding indexed list of unique strings:
+// `["a", "b"]`
+class UniqueJSONStrings {
+ public:
+  // Start an empty list of unique strings.
+  MFBT_API explicit UniqueJSONStrings(
+      JSONWriter::CollectionStyle aStyle = JSONWriter::MultiLineStyle);
+
+  // Start with a copy of the strings from another list.
+  MFBT_API explicit UniqueJSONStrings(
+      const UniqueJSONStrings& aOther,
+      JSONWriter::CollectionStyle aStyle = JSONWriter::MultiLineStyle);
+
+  MFBT_API ~UniqueJSONStrings();
+
+  // Add `aStr` to the list (if not already there), and write its index as a
+  // named object property.
+  void WriteProperty(JSONWriter& aWriter, const Span<const char>& aName,
+                     const Span<const char>& aStr) {
+    aWriter.IntProperty(aName, GetOrAddIndex(aStr));
+  }
+
+  // Add `aStr` to the list (if not already there), and write its index as an
+  // array element.
+  void WriteElement(JSONWriter& aWriter, const Span<const char>& aStr) {
+    aWriter.IntElement(GetOrAddIndex(aStr));
+  }
+
+  // Splice all collected unique strings into an array. This should only be done
+  // once, and then this UniqueStrings shouldn't be used anymore.
+  MFBT_API void SpliceStringTableElements(SpliceableJSONWriter& aWriter);
+
+ private:
+  // If `aStr` is already listed, return its index.
+  // Otherwise add it to the list and return the new index.
+  MFBT_API uint32_t GetOrAddIndex(const Span<const char>& aStr);
+
+  SpliceableChunkedJSONWriter mStringTableWriter;
+  HashMap<HashNumber, uint32_t> mStringHashToIndexMap;
+};
+
+void SpliceableJSONWriter::UniqueStringProperty(const Span<const char>& aName,
+                                                const Span<const char>& aStr) {
+  MOZ_RELEASE_ASSERT(mUniqueStrings);
+  mUniqueStrings->WriteProperty(*this, aName, aStr);
+}
+
+// Add `aStr` to the list (if not already there), and write its index as an
+// array element.
+void SpliceableJSONWriter::UniqueStringElement(const Span<const char>& aStr) {
+  MOZ_RELEASE_ASSERT(mUniqueStrings);
+  mUniqueStrings->WriteElement(*this, aStr);
+}
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif  // BASEPROFILEJSONWRITER_H
diff --git a/mozglue/baseprofiler/public/BaseProfiler.h b/mozglue/baseprofiler/public/BaseProfiler.h
new file mode 100644
index 0000000000..4bf1705041
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfiler.h
@@ -0,0 +1,964 @@
+/* -*- Mode: C++; tab-width: 2; 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/. */
+
+// The Gecko Profiler is an always-on profiler that takes fast and low overhead
+// samples of the program execution using only userspace functionality for
+// portability. The goal of this module is to provide performance data in a
+// generic cross-platform way without requiring custom tools or kernel support.
+//
+// Samples are collected to form a timeline with optional timeline event
+// (markers) used for filtering. The samples include both native stacks and
+// platform-independent "label stack" frames.
+
+#ifndef BaseProfiler_h
+#define BaseProfiler_h
+
+// This file is safe to include unconditionally, and only defines
+// empty macros if MOZ_GECKO_PROFILER is not set.
+
+// These headers are also safe to include unconditionally, with empty macros if
+// MOZ_GECKO_PROFILER is not set.
+#include "mozilla/BaseProfilerCounts.h"
+
+// BaseProfilerMarkers.h is #included in the middle of this header!
+// #include "mozilla/BaseProfilerMarkers.h"
+
+#ifndef MOZ_GECKO_PROFILER
+
+#  include "mozilla/BaseProfilerMarkers.h"
+#  include "mozilla/UniquePtr.h"
+
+// This file can be #included unconditionally. However, everything within this
+// file must be guarded by a #ifdef MOZ_GECKO_PROFILER, *except* for the
+// following macros and functions, which encapsulate the most common operations
+// and thus avoid the need for many #ifdefs.
+
+#  define AUTO_BASE_PROFILER_INIT
+
+#  define BASE_PROFILER_REGISTER_THREAD(name)
+#  define BASE_PROFILER_UNREGISTER_THREAD()
+#  define AUTO_BASE_PROFILER_REGISTER_THREAD(name)
+
+#  define AUTO_BASE_PROFILER_THREAD_SLEEP
+#  define AUTO_BASE_PROFILER_THREAD_WAKE
+
+#  define AUTO_BASE_PROFILER_LABEL(label, categoryPair)
+#  define AUTO_BASE_PROFILER_LABEL_CATEGORY_PAIR(categoryPair)
+#  define AUTO_BASE_PROFILER_LABEL_DYNAMIC_CSTR(label, categoryPair, cStr)
+#  define AUTO_BASE_PROFILER_LABEL_DYNAMIC_STRING(label, categoryPair, str)
+#  define AUTO_BASE_PROFILER_LABEL_FAST(label, categoryPair, ctx)
+#  define AUTO_BASE_PROFILER_LABEL_DYNAMIC_FAST(label, dynamicString, \
+                                                categoryPair, ctx, flags)
+
+#  define AUTO_PROFILER_STATS(name)
+
+// Function stubs for when MOZ_GECKO_PROFILER is not defined.
+
+namespace mozilla {
+// This won't be used, it's just there to allow the empty definition of
+// `profiler_capture_backtrace`.
+class ProfileChunkedBuffer {};
+
+namespace baseprofiler {
+// This won't be used, it's just there to allow the empty definition of
+// `profiler_get_backtrace`.
+struct ProfilerBacktrace {};
+using UniqueProfilerBacktrace = UniquePtr<ProfilerBacktrace>;
+
+// Get/Capture-backtrace functions can return nullptr or false, the result
+// should be fed to another empty macro or stub anyway.
+
+static inline UniqueProfilerBacktrace profiler_get_backtrace() {
+  return nullptr;
+}
+
+static inline bool profiler_capture_backtrace_into(
+    ProfileChunkedBuffer& aChunkedBuffer) {
+  return false;
+}
+
+static inline UniquePtr<ProfileChunkedBuffer> profiler_capture_backtrace() {
+  return nullptr;
+}
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#else  // !MOZ_GECKO_PROFILER
+
+#  include "BaseProfilingStack.h"
+
+#  include "mozilla/Assertions.h"
+#  include "mozilla/Atomics.h"
+#  include "mozilla/Attributes.h"
+#  include "mozilla/Maybe.h"
+#  include "mozilla/PowerOfTwo.h"
+#  include "mozilla/Sprintf.h"
+#  include "mozilla/ThreadLocal.h"
+#  include "mozilla/TimeStamp.h"
+#  include "mozilla/UniquePtr.h"
+
+#  include <functional>
+#  include <stdint.h>
+#  include <string>
+
+namespace mozilla {
+
+class MallocAllocPolicy;
+class ProfileChunkedBuffer;
+template <class T, size_t MinInlineCapacity, class AllocPolicy>
+class Vector;
+
+namespace baseprofiler {
+
+class ProfilerBacktrace;
+class SpliceableJSONWriter;
+
+// Macros used by the AUTO_PROFILER_* macros below.
+#  define BASE_PROFILER_RAII_PASTE(id, line) id##line
+#  define BASE_PROFILER_RAII_EXPAND(id, line) BASE_PROFILER_RAII_PASTE(id, line)
+#  define BASE_PROFILER_RAII BASE_PROFILER_RAII_EXPAND(raiiObject, __LINE__)
+
+//---------------------------------------------------------------------------
+// Profiler features
+//---------------------------------------------------------------------------
+
+// Higher-order macro containing all the feature info in one place. Define
+// |MACRO| appropriately to extract the relevant parts. Note that the number
+// values are used internally only and so can be changed without consequence.
+// Any changes to this list should also be applied to the feature list in
+// toolkit/components/extensions/schemas/geckoProfiler.json.
+#  define BASE_PROFILER_FOR_EACH_FEATURE(MACRO)                                \
+    MACRO(0, "java", Java, "Profile Java code, Android only")                  \
+                                                                               \
+    MACRO(1, "js", JS,                                                         \
+          "Get the JS engine to expose the JS stack to the profiler")          \
+                                                                               \
+    /* The DevTools profiler doesn't want the native addresses. */             \
+    MACRO(2, "leaf", Leaf, "Include the C++ leaf node if not stackwalking")    \
+                                                                               \
+    MACRO(3, "mainthreadio", MainThreadIO, "Add main thread file I/O")         \
+                                                                               \
+    MACRO(4, "fileio", FileIO,                                                 \
+          "Add file I/O from all profiled threads, implies mainthreadio")      \
+                                                                               \
+    MACRO(5, "fileioall", FileIOAll,                                           \
+          "Add file I/O from all threads, implies fileio")                     \
+                                                                               \
+    MACRO(6, "noiostacks", NoIOStacks,                                         \
+          "File I/O markers do not capture stacks, to reduce overhead")        \
+                                                                               \
+    MACRO(7, "screenshots", Screenshots,                                       \
+          "Take a snapshot of the window on every composition")                \
+                                                                               \
+    MACRO(8, "seqstyle", SequentialStyle,                                      \
+          "Disable parallel traversal in styling")                             \
+                                                                               \
+    MACRO(9, "stackwalk", StackWalk,                                           \
+          "Walk the C++ stack, not available on all platforms")                \
+                                                                               \
+    MACRO(10, "tasktracer", TaskTracer,                                        \
+          "Start profiling with feature TaskTracer")                           \
+                                                                               \
+    MACRO(11, "threads", Threads, "Profile the registered secondary threads")  \
+                                                                               \
+    MACRO(12, "jstracer", JSTracer, "Enable tracing of the JavaScript engine") \
+                                                                               \
+    MACRO(13, "jsallocations", JSAllocations,                                  \
+          "Have the JavaScript engine track allocations")                      \
+                                                                               \
+    MACRO(14, "nostacksampling", NoStackSampling,                              \
+          "Disable all stack sampling: Cancels \"js\", \"leaf\", "             \
+          "\"stackwalk\" and labels")                                          \
+                                                                               \
+    MACRO(15, "preferencereads", PreferenceReads,                              \
+          "Track when preferences are read")                                   \
+                                                                               \
+    MACRO(16, "nativeallocations", NativeAllocations,                          \
+          "Collect the stacks from a smaller subset of all native "            \
+          "allocations, biasing towards collecting larger allocations")        \
+                                                                               \
+    MACRO(17, "ipcmessages", IPCMessages,                                      \
+          "Have the IPC layer track cross-process messages")                   \
+                                                                               \
+    MACRO(18, "audiocallbacktracing", AudioCallbackTracing,                    \
+          "Audio callback tracing")                                            \
+                                                                               \
+    MACRO(19, "cpu", CPUUtilization, "CPU utilization")
+
+struct ProfilerFeature {
+#  define DECLARE(n_, str_, Name_, desc_)                     \
+    static constexpr uint32_t Name_ = (1u << n_);             \
+    static constexpr bool Has##Name_(uint32_t aFeatures) {    \
+      return aFeatures & Name_;                               \
+    }                                                         \
+    static constexpr void Set##Name_(uint32_t& aFeatures) {   \
+      aFeatures |= Name_;                                     \
+    }                                                         \
+    static constexpr void Clear##Name_(uint32_t& aFeatures) { \
+      aFeatures &= ~Name_;                                    \
+    }
+
+  // Define a bitfield constant, a getter, and two setters for each feature.
+  BASE_PROFILER_FOR_EACH_FEATURE(DECLARE)
+
+#  undef DECLARE
+};
+
+namespace detail {
+
+// RacyFeatures is only defined in this header file so that its methods can
+// be inlined into profiler_is_active(). Please do not use anything from the
+// detail namespace outside the profiler.
+
+// Within the profiler's code, the preferred way to check profiler activeness
+// and features is via ActivePS(). However, that requires locking gPSMutex.
+// There are some hot operations where absolute precision isn't required, so we
+// duplicate the activeness/feature state in a lock-free manner in this class.
+class RacyFeatures {
+ public:
+  MFBT_API static void SetActive(uint32_t aFeatures);
+
+  MFBT_API static void SetInactive();
+
+  MFBT_API static void SetPaused();
+
+  MFBT_API static void SetUnpaused();
+
+  MFBT_API static void SetSamplingPaused();
+
+  MFBT_API static void SetSamplingUnpaused();
+
+  MFBT_API static bool IsActive();
+
+  MFBT_API static bool IsActiveWithFeature(uint32_t aFeature);
+
+  // True if profiler is active, and not fully paused.
+  // Note that periodic sampling *could* be paused!
+  MFBT_API static bool IsActiveAndUnpaused();
+
+  // True if profiler is active, and sampling is not paused (though generic
+  // `SetPaused()` or specific `SetSamplingPaused()`).
+  MFBT_API static bool IsActiveAndSamplingUnpaused();
+
+ private:
+  static constexpr uint32_t Active = 1u << 31;
+  static constexpr uint32_t Paused = 1u << 30;
+  static constexpr uint32_t SamplingPaused = 1u << 29;
+
+// Ensure Active/Paused don't overlap with any of the feature bits.
+#  define NO_OVERLAP(n_, str_, Name_, desc_)                \
+    static_assert(ProfilerFeature::Name_ != SamplingPaused, \
+                  "bad feature value");
+
+  BASE_PROFILER_FOR_EACH_FEATURE(NO_OVERLAP);
+
+#  undef NO_OVERLAP
+
+  // We combine the active bit with the feature bits so they can be read or
+  // written in a single atomic operation.
+  // TODO: Could this be MFBT_DATA for better inlining optimization?
+  static Atomic<uint32_t, MemoryOrdering::Relaxed> sActiveAndFeatures;
+};
+
+MFBT_API bool IsThreadBeingProfiled();
+
+}  // namespace detail
+
+//---------------------------------------------------------------------------
+// Start and stop the profiler
+//---------------------------------------------------------------------------
+
+static constexpr PowerOfTwo32 BASE_PROFILER_DEFAULT_ENTRIES =
+#  if !defined(GP_PLAT_arm_android)
+    MakePowerOfTwo32<1024 * 1024>();  // 1M entries = 8MB
+#  else
+    MakePowerOfTwo32<128 * 1024>();  // 128k entries = 1MB
+#  endif
+
+// Startup profiling usually need to capture more data, especially on slow
+// systems.
+static constexpr PowerOfTwo32 BASE_PROFILER_DEFAULT_STARTUP_ENTRIES =
+#  if !defined(GP_PLAT_arm_android)
+    MakePowerOfTwo32<4 * 1024 * 1024>();  // 4M entries = 32MB
+#  else
+    MakePowerOfTwo32<256 * 1024>();  // 256k entries = 2MB
+#  endif
+
+#  define BASE_PROFILER_DEFAULT_DURATION 20
+#  define BASE_PROFILER_DEFAULT_INTERVAL 1
+
+// Initialize the profiler. If MOZ_PROFILER_STARTUP is set the profiler will
+// also be started. This call must happen before any other profiler calls
+// (except profiler_start(), which will call profiler_init() if it hasn't
+// already run).
+MFBT_API void profiler_init(void* stackTop);
+
+#  define AUTO_BASE_PROFILER_INIT \
+    ::mozilla::baseprofiler::AutoProfilerInit BASE_PROFILER_RAII
+
+// Clean up the profiler module, stopping it if required. This function may
+// also save a shutdown profile if requested. No profiler calls should happen
+// after this point and all profiling stack labels should have been popped.
+MFBT_API void profiler_shutdown();
+
+// Start the profiler -- initializing it first if necessary -- with the
+// selected options. Stops and restarts the profiler if it is already active.
+// After starting the profiler is "active". The samples will be recorded in a
+// circular buffer.
+//   "aCapacity" is the maximum number of 8-byte entries in the profiler's
+//               circular buffer.
+//   "aInterval" the sampling interval, measured in millseconds.
+//   "aFeatures" is the feature set. Features unsupported by this
+//               platform/configuration are ignored.
+//   "aFilters" is the list of thread filters. Threads that do not match any
+//              of the filters are not profiled. A filter matches a thread if
+//              (a) the thread name contains the filter as a case-insensitive
+//                  substring, or
+//              (b) the filter is of the form "pid:<n>" where n is the process
+//                  id of the process that the thread is running in.
+//   "aDuration" is the duration of entries in the profiler's circular buffer.
+MFBT_API void profiler_start(PowerOfTwo32 aCapacity, double aInterval,
+                             uint32_t aFeatures, const char** aFilters,
+                             uint32_t aFilterCount,
+                             const Maybe<double>& aDuration = Nothing());
+
+// Stop the profiler and discard the profile without saving it. A no-op if the
+// profiler is inactive. After stopping the profiler is "inactive".
+MFBT_API void profiler_stop();
+
+// If the profiler is inactive, start it. If it's already active, restart it if
+// the requested settings differ from the current settings. Both the check and
+// the state change are performed while the profiler state is locked.
+// The only difference to profiler_start is that the current buffer contents are
+// not discarded if the profiler is already running with the requested settings.
+MFBT_API void profiler_ensure_started(
+    PowerOfTwo32 aCapacity, double aInterval, uint32_t aFeatures,
+    const char** aFilters, uint32_t aFilterCount,
+    const Maybe<double>& aDuration = Nothing());
+
+//---------------------------------------------------------------------------
+// Control the profiler
+//---------------------------------------------------------------------------
+
+// Register/unregister threads with the profiler. Both functions operate the
+// same whether the profiler is active or inactive.
+#  define BASE_PROFILER_REGISTER_THREAD(name)                             \
+    do {                                                                  \
+      char stackTop;                                                      \
+      ::mozilla::baseprofiler::profiler_register_thread(name, &stackTop); \
+    } while (0)
+#  define BASE_PROFILER_UNREGISTER_THREAD() \
+    ::mozilla::baseprofiler::profiler_unregister_thread()
+MFBT_API ProfilingStack* profiler_register_thread(const char* name,
+                                                  void* guessStackTop);
+MFBT_API void profiler_unregister_thread();
+
+// Registers a DOM Window (the JS global `window`) with the profiler. Each
+// Window _roughly_ corresponds to a single document loaded within a
+// BrowsingContext. The unique IDs for both the Window and BrowsingContext are
+// recorded to allow correlating different Windows loaded within the same tab or
+// frame element.
+//
+// We register pages for each navigations but we do not register
+// history.pushState or history.replaceState since they correspond to the same
+// Inner Window ID. When a Browsing context is first loaded, the first url
+// loaded in it will be about:blank. Because of that, this call keeps the first
+// non-about:blank registration of window and discards the previous one.
+//
+//   "aBrowsingContextID"     is the ID of the browsing context that document
+//                            belongs to. That's used to determine the tab of
+//                            that page.
+//   "aInnerWindowID"         is the ID of the `window` global object of that
+//                            document.
+//   "aUrl"                   is the URL of the page.
+//   "aEmbedderInnerWindowID" is the inner window id of embedder. It's used to
+//                            determine sub documents of a page.
+MFBT_API void profiler_register_page(uint64_t aBrowsingContextID,
+                                     uint64_t aInnerWindowID,
+                                     const std::string& aUrl,
+                                     uint64_t aEmbedderInnerWindowID);
+// Unregister page with the profiler.
+//
+// Take a Inner Window ID and unregister the page entry that has the same ID.
+MFBT_API void profiler_unregister_page(uint64_t aRegisteredInnerWindowID);
+
+// Remove all registered and unregistered pages in the profiler.
+void profiler_clear_all_pages();
+
+class BaseProfilerCount;
+MFBT_API void profiler_add_sampled_counter(BaseProfilerCount* aCounter);
+MFBT_API void profiler_remove_sampled_counter(BaseProfilerCount* aCounter);
+
+// Register and unregister a thread within a scope.
+#  define AUTO_BASE_PROFILER_REGISTER_THREAD(name) \
+    ::mozilla::baseprofiler::AutoProfilerRegisterThread BASE_PROFILER_RAII(name)
+
+// Pause and resume the profiler. No-ops if the profiler is inactive. While
+// paused the profile will not take any samples and will not record any data
+// into its buffers. The profiler remains fully initialized in this state.
+// This feature will keep JavaScript profiling enabled, thus allowing toggling
+// the profiler without invalidating the JIT.
+MFBT_API void profiler_pause();
+MFBT_API void profiler_resume();
+
+// Only pause and resume the periodic sampling loop, including stack sampling,
+// counters, and profiling overheads.
+MFBT_API void profiler_pause_sampling();
+MFBT_API void profiler_resume_sampling();
+
+// These functions tell the profiler that a thread went to sleep so that we can
+// avoid sampling it while it's sleeping. Calling profiler_thread_sleep()
+// twice without an intervening profiler_thread_wake() is an error. All three
+// functions operate the same whether the profiler is active or inactive.
+MFBT_API void profiler_thread_sleep();
+MFBT_API void profiler_thread_wake();
+
+// Mark a thread as asleep/awake within a scope.
+#  define AUTO_BASE_PROFILER_THREAD_SLEEP \
+    ::mozilla::baseprofiler::AutoProfilerThreadSleep BASE_PROFILER_RAII
+#  define AUTO_BASE_PROFILER_THREAD_WAKE \
+    ::mozilla::baseprofiler::AutoProfilerThreadWake BASE_PROFILER_RAII
+
+//---------------------------------------------------------------------------
+// Get information from the profiler
+//---------------------------------------------------------------------------
+
+// Is the profiler active? Note: the return value of this function can become
+// immediately out-of-date. E.g. the profile might be active but then
+// profiler_stop() is called immediately afterward. One common and reasonable
+// pattern of usage is the following:
+//
+//   if (profiler_is_active()) {
+//     ExpensiveData expensiveData = CreateExpensiveData();
+//     PROFILER_OPERATION(expensiveData);
+//   }
+//
+// where PROFILER_OPERATION is a no-op if the profiler is inactive. In this
+// case the profiler_is_active() check is just an optimization -- it prevents
+// us calling CreateExpensiveData() unnecessarily in most cases, but the
+// expensive data will end up being created but not used if another thread
+// stops the profiler between the CreateExpensiveData() and PROFILER_OPERATION
+// calls.
+inline bool profiler_is_active() {
+  return baseprofiler::detail::RacyFeatures::IsActive();
+}
+
+// Same as profiler_is_active(), but with the same extra checks that determine
+// if the profiler would currently store markers. So this should be used before
+// doing some potentially-expensive work that's used in a marker. E.g.:
+//
+//   if (profiler_can_accept_markers()) {
+//     BASE_PROFILER_MARKER(name, OTHER, SomeMarkerType, expensivePayload);
+//   }
+inline bool profiler_can_accept_markers() {
+  return baseprofiler::detail::RacyFeatures::IsActiveAndUnpaused();
+}
+
+// Is the profiler active, and is the current thread being profiled?
+// (Same caveats and recommented usage as profiler_is_active().)
+inline bool profiler_thread_is_being_profiled() {
+  return profiler_is_active() && baseprofiler::detail::IsThreadBeingProfiled();
+}
+
+// Is the profiler active and paused? Returns false if the profiler is inactive.
+MFBT_API bool profiler_is_paused();
+
+// Is the profiler active and sampling is paused? Returns false if the profiler
+// is inactive.
+MFBT_API bool profiler_is_sampling_paused();
+
+// Is the current thread sleeping?
+MFBT_API bool profiler_thread_is_sleeping();
+
+// Get all the features supported by the profiler that are accepted by
+// profiler_start(). The result is the same whether the profiler is active or
+// not.
+MFBT_API uint32_t profiler_get_available_features();
+
+// Check if a profiler feature (specified via the ProfilerFeature type) is
+// active. Returns false if the profiler is inactive. Note: the return value
+// can become immediately out-of-date, much like the return value of
+// profiler_is_active().
+MFBT_API bool profiler_feature_active(uint32_t aFeature);
+
+// Get the params used to start the profiler. Returns 0 and an empty vector
+// (via outparams) if the profile is inactive. It's possible that the features
+// returned may be slightly different to those requested due to required
+// adjustments.
+MFBT_API void profiler_get_start_params(
+    int* aEntrySize, Maybe<double>* aDuration, double* aInterval,
+    uint32_t* aFeatures, Vector<const char*, 0, MallocAllocPolicy>* aFilters);
+
+// The number of milliseconds since the process started. Operates the same
+// whether the profiler is active or inactive.
+MFBT_API double profiler_time();
+
+// Get the current process's ID.
+MFBT_API int profiler_current_process_id();
+
+// Get the current thread's ID.
+MFBT_API int profiler_current_thread_id();
+
+// Statically initialized to 0, then set once from profiler_init(), which should
+// be called from the main thread before any other use of the profiler.
+extern MFBT_DATA int scProfilerMainThreadId;
+
+inline int profiler_main_thread_id() { return scProfilerMainThreadId; }
+
+inline bool profiler_is_main_thread() {
+  return profiler_current_thread_id() == profiler_main_thread_id();
+}
+
+// An object of this class is passed to profiler_suspend_and_sample_thread().
+// For each stack frame, one of the Collect methods will be called.
+class ProfilerStackCollector {
+ public:
+  // Some collectors need to worry about possibly overwriting previous
+  // generations of data. If that's not an issue, this can return Nothing,
+  // which is the default behaviour.
+  virtual Maybe<uint64_t> SamplePositionInBuffer() { return Nothing(); }
+  virtual Maybe<uint64_t> BufferRangeStart() { return Nothing(); }
+
+  // This method will be called once if the thread being suspended is the main
+  // thread. Default behaviour is to do nothing.
+  virtual void SetIsMainThread() {}
+
+  // WARNING: The target thread is suspended when the Collect methods are
+  // called. Do not try to allocate or acquire any locks, or you could
+  // deadlock. The target thread will have resumed by the time this function
+  // returns.
+
+  virtual void CollectNativeLeafAddr(void* aAddr) = 0;
+
+  virtual void CollectProfilingStackFrame(
+      const ProfilingStackFrame& aFrame) = 0;
+};
+
+// This method suspends the thread identified by aThreadId, samples its
+// profiling stack, JS stack, and (optionally) native stack, passing the
+// collected frames into aCollector. aFeatures dictates which compiler features
+// are used. |Leaf| is the only relevant one.
+MFBT_API void profiler_suspend_and_sample_thread(
+    int aThreadId, uint32_t aFeatures, ProfilerStackCollector& aCollector,
+    bool aSampleNative = true);
+
+struct ProfilerBacktraceDestructor {
+  MFBT_API void operator()(ProfilerBacktrace*);
+};
+
+using UniqueProfilerBacktrace =
+    UniquePtr<ProfilerBacktrace, ProfilerBacktraceDestructor>;
+
+// Immediately capture the current thread's call stack, store it in the provided
+// buffer (usually to avoid allocations if you can construct the buffer on the
+// stack). Returns false if unsuccessful, or if the profiler is inactive.
+MFBT_API bool profiler_capture_backtrace_into(
+    ProfileChunkedBuffer& aChunkedBuffer);
+
+// Immediately capture the current thread's call stack, and return it in a
+// ProfileChunkedBuffer (usually for later use in MarkerStack::TakeBacktrace()).
+// May be null if unsuccessful, or if the profiler is inactive.
+MFBT_API UniquePtr<ProfileChunkedBuffer> profiler_capture_backtrace();
+
+// Immediately capture the current thread's call stack, and return it in a
+// ProfilerBacktrace (usually for later use in marker function that take a
+// ProfilerBacktrace). May be null if unsuccessful, or if the profiler is
+// inactive.
+MFBT_API UniqueProfilerBacktrace profiler_get_backtrace();
+
+struct ProfilerStats {
+  unsigned n = 0;
+  double sum = 0;
+  double min = std::numeric_limits<double>::max();
+  double max = 0;
+  void Count(double v) {
+    ++n;
+    sum += v;
+    if (v < min) {
+      min = v;
+    }
+    if (v > max) {
+      max = v;
+    }
+  }
+};
+
+struct ProfilerBufferInfo {
+  // Index of the oldest entry.
+  uint64_t mRangeStart;
+  // Index of the newest entry.
+  uint64_t mRangeEnd;
+  // Buffer capacity in number of 8-byte entries.
+  uint32_t mEntryCount;
+  // Sampling stats: Interval (us) between successive samplings.
+  ProfilerStats mIntervalsUs;
+  // Sampling stats: Total duration (us) of each sampling. (Split detail below.)
+  ProfilerStats mOverheadsUs;
+  // Sampling stats: Time (us) to acquire the lock before sampling.
+  ProfilerStats mLockingsUs;
+  // Sampling stats: Time (us) to discard expired data.
+  ProfilerStats mCleaningsUs;
+  // Sampling stats: Time (us) to collect counter data.
+  ProfilerStats mCountersUs;
+  // Sampling stats: Time (us) to sample thread stacks.
+  ProfilerStats mThreadsUs;
+};
+
+// Get information about the current buffer status.
+// Returns Nothing() if the profiler is inactive.
+//
+// This information may be useful to a user-interface displaying the current
+// status of the profiler, allowing the user to get a sense for how fast the
+// buffer is being written to, and how much data is visible.
+MFBT_API Maybe<ProfilerBufferInfo> profiler_get_buffer_info();
+
+// Uncomment the following line to display profiler runtime statistics at
+// shutdown.
+// #  define PROFILER_RUNTIME_STATS
+
+#  ifdef PROFILER_RUNTIME_STATS
+// This class gathers durations and displays some basic stats when destroyed.
+// It is intended to be used as a static variable (see `AUTO_PROFILER_STATS`
+// below), to display stats at the end of the program.
+class StaticBaseProfilerStats {
+ public:
+  explicit StaticBaseProfilerStats(const char* aName) : mName(aName) {}
+
+  ~StaticBaseProfilerStats() {
+    // Using unsigned long long for computations and printfs.
+    using ULL = unsigned long long;
+    ULL n = static_cast<ULL>(mNumberDurations);
+    if (n != 0) {
+      ULL sumNs = static_cast<ULL>(mSumDurationsNs);
+      printf(
+          "[%d] Profiler stats `%s`: %llu ns / %llu = %llu ns, max %llu ns\n",
+          profiler_current_process_id(), mName, sumNs, n, sumNs / n,
+          static_cast<ULL>(mLongestDurationNs));
+    } else {
+      printf("[%d] Profiler stats `%s`: (nothing)\n",
+             profiler_current_process_id(), mName);
+    }
+  }
+
+  void AddDurationFrom(TimeStamp aStart) {
+    DurationNs duration = static_cast<DurationNs>(
+        (TimeStamp::NowUnfuzzed() - aStart).ToMicroseconds() * 1000 + 0.5);
+    mSumDurationsNs += duration;
+    ++mNumberDurations;
+    // Update mLongestDurationNs if this one is longer.
+    for (;;) {
+      DurationNs longest = mLongestDurationNs;
+      if (MOZ_LIKELY(longest >= duration)) {
+        // This duration is not the longest, nothing to do.
+        break;
+      }
+      if (MOZ_LIKELY(mLongestDurationNs.compareExchange(longest, duration))) {
+        // Successfully updated `mLongestDurationNs` with the new value.
+        break;
+      }
+      // Otherwise someone else just updated `mLongestDurationNs`, we need to
+      // try again by looping.
+    }
+  }
+
+ private:
+  using DurationNs = uint64_t;
+  using Count = uint32_t;
+
+  Atomic<DurationNs> mSumDurationsNs{0};
+  Atomic<DurationNs> mLongestDurationNs{0};
+  Atomic<Count> mNumberDurations{0};
+  const char* mName;
+};
+
+// RAII object that measure its scoped lifetime duration and reports it to a
+// `StaticBaseProfilerStats`.
+class MOZ_RAII AutoProfilerStats {
+ public:
+  explicit AutoProfilerStats(StaticBaseProfilerStats& aStats)
+      : mStats(aStats), mStart(TimeStamp::NowUnfuzzed()) {}
+
+  ~AutoProfilerStats() { mStats.AddDurationFrom(mStart); }
+
+ private:
+  StaticBaseProfilerStats& mStats;
+  TimeStamp mStart;
+};
+
+// Macro that should be used to collect basic statistics from measurements of
+// block durations, from where this macro is, until the end of its enclosing
+// scope. The name is used in the static variable name and when displaying stats
+// at the end of the program; Another location could use the same name but their
+// stats will not be combined, so use different name if these locations should
+// be distinguished.
+#    define AUTO_PROFILER_STATS(name)                                      \
+      static ::mozilla::baseprofiler::StaticBaseProfilerStats sStat##name( \
+          #name);                                                          \
+      ::mozilla::baseprofiler::AutoProfilerStats autoStat##name(sStat##name);
+
+#  else  // PROFILER_RUNTIME_STATS
+
+#    define AUTO_PROFILER_STATS(name)
+
+#  endif  // PROFILER_RUNTIME_STATS else
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+// BaseProfilerMarkers.h requires some stuff from this header.
+// TODO: Move common stuff to shared header, and move this #include to the top.
+#  include "mozilla/BaseProfilerMarkers.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+//---------------------------------------------------------------------------
+// Put profiling data into the profiler (labels and markers)
+//---------------------------------------------------------------------------
+
+// Insert an RAII object in this scope to enter a label stack frame. Any
+// samples collected in this scope will contain this label in their stack.
+// The label argument must be a static C string. It is usually of the
+// form "ClassName::FunctionName". (Ideally we'd use the compiler to provide
+// that for us, but __func__ gives us the function name without the class
+// name.) If the label applies to only part of a function, you can qualify it
+// like this: "ClassName::FunctionName:PartName".
+//
+// Use AUTO_BASE_PROFILER_LABEL_DYNAMIC_* if you want to add additional /
+// dynamic information to the label stack frame.
+#  define AUTO_BASE_PROFILER_LABEL(label, categoryPair)            \
+    ::mozilla::baseprofiler::AutoProfilerLabel BASE_PROFILER_RAII( \
+        label, nullptr,                                            \
+        ::mozilla::baseprofiler::ProfilingCategoryPair::categoryPair)
+
+// Similar to AUTO_BASE_PROFILER_LABEL, but with only one argument: the category
+// pair. The label string is taken from the category pair. This is convenient
+// for labels like
+// AUTO_BASE_PROFILER_LABEL_CATEGORY_PAIR(GRAPHICS_LayerBuilding) which would
+// otherwise just repeat the string.
+#  define AUTO_BASE_PROFILER_LABEL_CATEGORY_PAIR(categoryPair)         \
+    ::mozilla::baseprofiler::AutoProfilerLabel BASE_PROFILER_RAII(     \
+        "", nullptr,                                                   \
+        ::mozilla::baseprofiler::ProfilingCategoryPair::categoryPair,  \
+        uint32_t(::mozilla::baseprofiler::ProfilingStackFrame::Flags:: \
+                     LABEL_DETERMINED_BY_CATEGORY_PAIR))
+
+// Similar to AUTO_BASE_PROFILER_LABEL, but with an additional string. The
+// inserted RAII object stores the cStr pointer in a field; it does not copy the
+// string.
+//
+// WARNING: This means that the string you pass to this macro needs to live at
+// least until the end of the current scope. Be careful using this macro with
+// ns[C]String; the other AUTO_BASE_PROFILER_LABEL_DYNAMIC_* macros below are
+// preferred because they avoid this problem.
+//
+// If the profiler samples the current thread and walks the label stack while
+// this RAII object is on the stack, it will copy the supplied string into the
+// profile buffer. So there's one string copy operation, and it happens at
+// sample time.
+//
+// Compare this to the plain AUTO_BASE_PROFILER_LABEL macro, which only accepts
+// literal strings: When the label stack frames generated by
+// AUTO_BASE_PROFILER_LABEL are sampled, no string copy needs to be made because
+// the profile buffer can just store the raw pointers to the literal strings.
+// Consequently, AUTO_BASE_PROFILER_LABEL frames take up considerably less space
+// in the profile buffer than AUTO_BASE_PROFILER_LABEL_DYNAMIC_* frames.
+#  define AUTO_BASE_PROFILER_LABEL_DYNAMIC_CSTR(label, categoryPair, cStr) \
+    ::mozilla::baseprofiler::AutoProfilerLabel BASE_PROFILER_RAII(         \
+        label, cStr,                                                       \
+        ::mozilla::baseprofiler::ProfilingCategoryPair::categoryPair)
+
+// Similar to AUTO_BASE_PROFILER_LABEL_DYNAMIC_CSTR, but takes an std::string.
+//
+// Note: The use of the Maybe<>s ensures the scopes for the dynamic string and
+// the AutoProfilerLabel are appropriate, while also not incurring the runtime
+// cost of the string assignment unless the profiler is active. Therefore,
+// unlike AUTO_BASE_PROFILER_LABEL and AUTO_BASE_PROFILER_LABEL_DYNAMIC_CSTR,
+// this macro doesn't push/pop a label when the profiler is inactive.
+#  define AUTO_BASE_PROFILER_LABEL_DYNAMIC_STRING(label, categoryPair, str) \
+    Maybe<std::string> autoStr;                                             \
+    Maybe<::mozilla::baseprofiler::AutoProfilerLabel> raiiObjectString;     \
+    if (::mozilla::baseprofiler::profiler_is_active()) {                    \
+      autoStr.emplace(str);                                                 \
+      raiiObjectString.emplace(                                             \
+          label, autoStr->c_str(),                                          \
+          ::mozilla::baseprofiler::ProfilingCategoryPair::categoryPair);    \
+    }
+
+// Similar to AUTO_BASE_PROFILER_LABEL, but accepting a JSContext* parameter,
+// and a no-op if the profiler is disabled. Used to annotate functions for which
+// overhead in the range of nanoseconds is noticeable. It avoids overhead from
+// the TLS lookup because it can get the ProfilingStack from the JS context, and
+// avoids almost all overhead in the case where the profiler is disabled.
+#  define AUTO_BASE_PROFILER_LABEL_FAST(label, categoryPair, ctx)  \
+    ::mozilla::baseprofiler::AutoProfilerLabel BASE_PROFILER_RAII( \
+        ctx, label, nullptr,                                       \
+        ::mozilla::baseprofiler::ProfilingCategoryPair::categoryPair)
+
+// Similar to AUTO_BASE_PROFILER_LABEL_FAST, but also takes an extra string and
+// an additional set of flags. The flags parameter should carry values from the
+// ProfilingStackFrame::Flags enum.
+#  define AUTO_BASE_PROFILER_LABEL_DYNAMIC_FAST(label, dynamicString,     \
+                                                categoryPair, ctx, flags) \
+    ::mozilla::baseprofiler::AutoProfilerLabel BASE_PROFILER_RAII(        \
+        ctx, label, dynamicString,                                        \
+        ::mozilla::baseprofiler::ProfilingCategoryPair::categoryPair, flags)
+
+MFBT_API void profiler_add_js_marker(const char* aMarkerName,
+                                     const char* aMarkerText);
+
+// Returns true if any of the profiler mutexes are currently locked *on the
+// current thread*. This may be used by re-entrant code that may call profiler
+// functions while the same of a different profiler mutex is locked, which could
+// deadlock.
+bool profiler_is_locked_on_current_thread();
+
+//---------------------------------------------------------------------------
+// Output profiles
+//---------------------------------------------------------------------------
+
+// Set a user-friendly process name, used in JSON stream.
+MFBT_API void profiler_set_process_name(const std::string& aProcessName,
+                                        const std::string* aETLDplus1);
+
+// Get the profile encoded as a JSON string. A no-op (returning nullptr) if the
+// profiler is inactive.
+// If aIsShuttingDown is true, the current time is included as the process
+// shutdown time in the JSON's "meta" object.
+MFBT_API UniquePtr<char[]> profiler_get_profile(double aSinceTime = 0,
+                                                bool aIsShuttingDown = false,
+                                                bool aOnlyThreads = false);
+
+// Write the profile for this process (excluding subprocesses) into aWriter.
+// Returns false if the profiler is inactive.
+MFBT_API bool profiler_stream_json_for_this_process(
+    SpliceableJSONWriter& aWriter, double aSinceTime = 0,
+    bool aIsShuttingDown = false, bool aOnlyThreads = false);
+
+// Get the profile and write it into a file. A no-op if the profile is
+// inactive.
+MFBT_API void profiler_save_profile_to_file(const char* aFilename);
+
+//---------------------------------------------------------------------------
+// RAII classes
+//---------------------------------------------------------------------------
+
+class MOZ_RAII AutoProfilerInit {
+ public:
+  explicit AutoProfilerInit() { profiler_init(this); }
+
+  ~AutoProfilerInit() { profiler_shutdown(); }
+
+ private:
+};
+
+// Convenience class to register and unregister a thread with the profiler.
+// Needs to be the first object on the stack of the thread.
+class MOZ_RAII AutoProfilerRegisterThread final {
+ public:
+  explicit AutoProfilerRegisterThread(const char* aName) {
+    profiler_register_thread(aName, this);
+  }
+
+  ~AutoProfilerRegisterThread() { profiler_unregister_thread(); }
+
+ private:
+  AutoProfilerRegisterThread(const AutoProfilerRegisterThread&) = delete;
+  AutoProfilerRegisterThread& operator=(const AutoProfilerRegisterThread&) =
+      delete;
+};
+
+class MOZ_RAII AutoProfilerThreadSleep {
+ public:
+  explicit AutoProfilerThreadSleep() { profiler_thread_sleep(); }
+
+  ~AutoProfilerThreadSleep() { profiler_thread_wake(); }
+
+ private:
+};
+
+// Temporarily wake up the profiling of a thread while servicing events such as
+// Asynchronous Procedure Calls (APCs).
+class MOZ_RAII AutoProfilerThreadWake {
+ public:
+  explicit AutoProfilerThreadWake()
+      : mIssuedWake(profiler_thread_is_sleeping()) {
+    if (mIssuedWake) {
+      profiler_thread_wake();
+    }
+  }
+
+  ~AutoProfilerThreadWake() {
+    if (mIssuedWake) {
+      MOZ_ASSERT(!profiler_thread_is_sleeping());
+      profiler_thread_sleep();
+    }
+  }
+
+ private:
+  bool mIssuedWake;
+};
+
+// This class creates a non-owning ProfilingStack reference. Objects of this
+// class are stack-allocated, and so exist within a thread, and are thus bounded
+// by the lifetime of the thread, which ensures that the references held can't
+// be used after the ProfilingStack is destroyed.
+class MOZ_RAII AutoProfilerLabel {
+ public:
+  // This is the AUTO_BASE_PROFILER_LABEL and AUTO_BASE_PROFILER_LABEL_DYNAMIC
+  // variant.
+  AutoProfilerLabel(const char* aLabel, const char* aDynamicString,
+                    ProfilingCategoryPair aCategoryPair, uint32_t aFlags = 0) {
+    // Get the ProfilingStack from TLS.
+    Push(GetProfilingStack(), aLabel, aDynamicString, aCategoryPair, aFlags);
+  }
+
+  void Push(ProfilingStack* aProfilingStack, const char* aLabel,
+            const char* aDynamicString, ProfilingCategoryPair aCategoryPair,
+            uint32_t aFlags = 0) {
+    // This function runs both on and off the main thread.
+
+    mProfilingStack = aProfilingStack;
+    if (mProfilingStack) {
+      mProfilingStack->pushLabelFrame(aLabel, aDynamicString, this,
+                                      aCategoryPair, aFlags);
+    }
+  }
+
+  ~AutoProfilerLabel() {
+    // This function runs both on and off the main thread.
+
+    if (mProfilingStack) {
+      mProfilingStack->pop();
+    }
+  }
+
+  MFBT_API static ProfilingStack* GetProfilingStack();
+
+ private:
+  // We save a ProfilingStack pointer in the ctor so we don't have to redo the
+  // TLS lookup in the dtor.
+  ProfilingStack* mProfilingStack;
+
+ public:
+  // See the comment on the definition in platform.cpp for details about this.
+  static MOZ_THREAD_LOCAL(ProfilingStack*) sProfilingStack;
+};
+
+// Get the MOZ_PROFILER_STARTUP* environment variables that should be
+// supplied to a child process that is about to be launched, in order
+// to make that child process start with the same profiler settings as
+// in the current process.  The given function is invoked once for
+// each variable to be set.
+MFBT_API void GetProfilerEnvVarsForChildProcess(
+    std::function<void(const char* key, const char* value)>&& aSetEnv);
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif  // !MOZ_GECKO_PROFILER
+
+#endif  // BaseProfiler_h
diff --git a/mozglue/baseprofiler/public/BaseProfilerCounts.h b/mozglue/baseprofiler/public/BaseProfilerCounts.h
new file mode 100644
index 0000000000..fbcc713744
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilerCounts.h
@@ -0,0 +1,280 @@
+/* -*- Mode: C++; tab-width: 2; 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 BaseProfilerCounts_h
+#define BaseProfilerCounts_h
+
+#ifndef MOZ_GECKO_PROFILER
+
+#  define BASE_PROFILER_DEFINE_COUNT_TOTAL(label, category, description)
+#  define BASE_PROFILER_DEFINE_COUNT(label, category, description)
+#  define BASE_PROFILER_DEFINE_STATIC_COUNT_TOTAL(label, category, description)
+#  define AUTO_BASE_PROFILER_COUNT_TOTAL(label, count)
+#  define AUTO_BASE_PROFILER_COUNT(label)
+#  define AUTO_BASE_PROFILER_STATIC_COUNT(label, count)
+#  define AUTO_BASE_PROFILER_FORCE_ALLOCATION(label)
+
+#else
+
+#  include "mozilla/Atomics.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+class BaseProfilerCount;
+MFBT_API void profiler_add_sampled_counter(BaseProfilerCount* aCounter);
+MFBT_API void profiler_remove_sampled_counter(BaseProfilerCount* aCounter);
+
+typedef Atomic<int64_t, MemoryOrdering::Relaxed> ProfilerAtomicSigned;
+typedef Atomic<uint64_t, MemoryOrdering::Relaxed> ProfilerAtomicUnsigned;
+
+// Counter support
+// There are two types of counters:
+// 1) a simple counter which can be added to or subtracted from.  This could
+// track the number of objects of a type, the number of calls to something
+// (reflow, JIT, etc).
+// 2) a combined counter which has the above, plus a number-of-calls counter
+// that is incremented by 1 for each call to modify the count.  This provides
+// an optional source for a 'heatmap' of access.  This can be used (for
+// example) to track the amount of memory allocated, and provide a heatmap of
+// memory operations (allocs/frees).
+//
+// Counters are sampled by the profiler once per sample-period.  At this time,
+// all counters are global to the process.  In the future, there might be more
+// versions with per-thread or other discriminators.
+//
+// Typical usage:
+// There are two ways to use counters: With heap-created counter objects,
+// or using macros.  Note: the macros use statics, and will be slightly
+// faster/smaller, and you need to care about creating them before using
+// them.  They're similar to the use-pattern for the other AUTO_PROFILER*
+// macros, but they do need the PROFILER_DEFINE* to be use to instantiate
+// the statics.
+//
+// PROFILER_DEFINE_COUNT(mything, "JIT", "Some JIT byte count")
+// ...
+// void foo() { ... AUTO_PROFILER_COUNT(mything, number_of_bytes_used); ... }
+//
+// or (to also get a heatmap)
+//
+// PROFILER_DEFINE_COUNT_TOTAL(mything, "JIT", "Some JIT byte count")
+// ...
+// void foo() {
+//   ...
+//   AUTO_PROFILER_COUNT_TOTAL(mything, number_of_bytes_generated);
+//   ...
+// }
+//
+// To use without statics/macros:
+//
+// UniquePtr<ProfilerCounter> myCounter;
+// ...
+// myCounter =
+//   MakeUnique<ProfilerCounter>("mything", "JIT", "Some JIT byte count"));
+// ...
+// void foo() { ... myCounter->Add(number_of_bytes_generated0; ... }
+
+class BaseProfilerCount {
+ public:
+  BaseProfilerCount(const char* aLabel, ProfilerAtomicSigned* aCounter,
+                    ProfilerAtomicUnsigned* aNumber, const char* aCategory,
+                    const char* aDescription)
+      : mLabel(aLabel),
+        mCategory(aCategory),
+        mDescription(aDescription),
+        mCounter(aCounter),
+        mNumber(aNumber) {
+#  define COUNTER_CANARY 0xDEADBEEF
+#  ifdef DEBUG
+    mCanary = COUNTER_CANARY;
+    mPrevNumber = 0;
+#  endif
+    // Can't call profiler_* here since this may be non-xul-library
+  }
+#  ifdef DEBUG
+  ~BaseProfilerCount() { mCanary = 0; }
+#  endif
+
+  void Sample(int64_t& aCounter, uint64_t& aNumber) {
+    MOZ_ASSERT(mCanary == COUNTER_CANARY);
+
+    aCounter = *mCounter;
+    aNumber = mNumber ? *mNumber : 0;
+#  ifdef DEBUG
+    MOZ_ASSERT(aNumber >= mPrevNumber);
+    mPrevNumber = aNumber;
+#  endif
+  }
+
+  // We don't define ++ and Add() here, since the static defines directly
+  // increment the atomic counters, and the subclasses implement ++ and
+  // Add() directly.
+
+  // These typically are static strings (for example if you use the macros
+  // below)
+  const char* mLabel;
+  const char* mCategory;
+  const char* mDescription;
+  // We're ok with these being un-ordered in race conditions.  These are
+  // pointers because we want to be able to use statics and increment them
+  // directly.  Otherwise we could just have them inline, and not need the
+  // constructor args.
+  // These can be static globals (using the macros below), though they
+  // don't have to be - their lifetime must be longer than the use of them
+  // by the profiler (see profiler_add/remove_sampled_counter()).  If you're
+  // using a lot of these, they probably should be allocated at runtime (see
+  // class ProfilerCountOnly below).
+  ProfilerAtomicSigned* mCounter;
+  ProfilerAtomicUnsigned* mNumber;  // may be null
+
+#  ifdef DEBUG
+  uint32_t mCanary;
+  uint64_t mPrevNumber;  // value of number from the last Sample()
+#  endif
+};
+
+// Designed to be allocated dynamically, and simply incremented with obj++
+// or obj->Add(n)
+class ProfilerCounter final : public BaseProfilerCount {
+ public:
+  ProfilerCounter(const char* aLabel, const char* aCategory,
+                  const char* aDescription)
+      : BaseProfilerCount(aLabel, &mCounter, nullptr, aCategory, aDescription) {
+    // Assume we're in libxul
+    profiler_add_sampled_counter(this);
+  }
+
+  virtual ~ProfilerCounter() { profiler_remove_sampled_counter(this); }
+
+  BaseProfilerCount& operator++() {
+    Add(1);
+    return *this;
+  }
+
+  void Add(int64_t aNumber) { mCounter += aNumber; }
+
+  ProfilerAtomicSigned mCounter;
+};
+
+// Also keeps a heatmap (number of calls to ++/Add())
+class ProfilerCounterTotal final : public BaseProfilerCount {
+ public:
+  ProfilerCounterTotal(const char* aLabel, const char* aCategory,
+                       const char* aDescription)
+      : BaseProfilerCount(aLabel, &mCounter, &mNumber, aCategory,
+                          aDescription) {
+    // Assume we're in libxul
+    profiler_add_sampled_counter(this);
+  }
+
+  virtual ~ProfilerCounterTotal() { profiler_remove_sampled_counter(this); }
+
+  BaseProfilerCount& operator++() {
+    Add(1);
+    return *this;
+  }
+
+  void Add(int64_t aNumber) {
+    mCounter += aNumber;
+    mNumber++;
+  }
+
+  ProfilerAtomicSigned mCounter;
+  ProfilerAtomicUnsigned mNumber;
+};
+
+// Defines a counter that is sampled on each profiler tick, with a running
+// count (signed), and number-of-instances. Note that because these are two
+// independent Atomics, there is a possiblity that count will not include
+// the last call, but number of uses will.  I think this is not worth
+// worrying about
+#  define BASE_PROFILER_DEFINE_COUNT_TOTAL(label, category, description) \
+    ProfilerAtomicSigned profiler_count_##label(0);                      \
+    ProfilerAtomicUnsigned profiler_number_##label(0);                   \
+    const char profiler_category_##label[] = category;                   \
+    const char profiler_description_##label[] = description;             \
+    UniquePtr<::mozilla::baseprofiler::BaseProfilerCount> AutoCount_##label;
+
+// This counts, but doesn't keep track of the number of calls to
+// AUTO_PROFILER_COUNT()
+#  define BASE_PROFILER_DEFINE_COUNT(label, category, description) \
+    ProfilerAtomicSigned profiler_count_##label(0);                \
+    const char profiler_category_##label[] = category;             \
+    const char profiler_description_##label[] = description;       \
+    UniquePtr<::mozilla::baseprofiler::BaseProfilerCount> AutoCount_##label;
+
+// This will create a static initializer if used, but avoids a possible
+// allocation.
+#  define BASE_PROFILER_DEFINE_STATIC_COUNT_TOTAL(label, category,           \
+                                                  description)               \
+    ProfilerAtomicSigned profiler_count_##label(0);                          \
+    ProfilerAtomicUnsigned profiler_number_##label(0);                       \
+    ::mozilla::baseprofiler::BaseProfilerCount AutoCount_##label(            \
+        #label, &profiler_count_##label, &profiler_number_##label, category, \
+        description);
+
+// If we didn't care about static initializers, we could avoid the need for
+// a ptr to the BaseProfilerCount object.
+
+// XXX It would be better to do this without the if() and without the
+// theoretical race to set the UniquePtr (i.e. possible leak).
+#  define AUTO_BASE_PROFILER_COUNT_TOTAL(label, count)                      \
+    do {                                                                    \
+      profiler_number_##label++; /* do this first*/                         \
+      profiler_count_##label += count;                                      \
+      if (!AutoCount_##label) {                                             \
+        /* Ignore that we could call this twice in theory, and that we leak \
+         * them                                                             \
+         */                                                                 \
+        AutoCount_##label.reset(new BaseProfilerCount(                      \
+            #label, &profiler_count_##label, &profiler_number_##label,      \
+            profiler_category_##label, profiler_description_##label));      \
+        ::mozilla::baseprofiler::profiler_add_sampled_counter(              \
+            AutoCount_##label.get());                                       \
+      }                                                                     \
+    } while (0)
+
+#  define AUTO_BASE_PROFILER_COUNT(label, count)                            \
+    do {                                                                    \
+      profiler_count_##label += count; /* do this first*/                   \
+      if (!AutoCount_##label) {                                             \
+        /* Ignore that we could call this twice in theory, and that we leak \
+         * them                                                             \
+         */                                                                 \
+        AutoCount_##label.reset(new BaseProfilerCount(                      \
+            #label, nullptr, &profiler_number_##label,                      \
+            profiler_category_##label, profiler_description_##label));      \
+        ::mozilla::baseprofiler::profiler_add_sampled_counter(              \
+            AutoCount_##label.get());                                       \
+      }                                                                     \
+    } while (0)
+
+#  define AUTO_BASE_PROFILER_STATIC_COUNT(label, count) \
+    do {                                                \
+      profiler_number_##label++; /* do this first*/     \
+      profiler_count_##label += count;                  \
+    } while (0)
+
+// if we need to force the allocation
+#  define AUTO_BASE_PROFILER_FORCE_ALLOCATION(label)                        \
+    do {                                                                    \
+      if (!AutoCount_##label) {                                             \
+        /* Ignore that we could call this twice in theory, and that we leak \
+         * them                                                             \
+         */                                                                 \
+        AutoCount_##label.reset(                                            \
+            new ::mozilla::baseprofiler::BaseProfilerCount(                 \
+                #label, &profiler_count_##label, &profiler_number_##label,  \
+                profiler_category_##label, profiler_description_##label));  \
+      }                                                                     \
+    } while (0)
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif  // !MOZ_GECKO_PROFILER
+
+#endif  // BaseProfilerCounts_h
diff --git a/mozglue/baseprofiler/public/BaseProfilerDetail.h b/mozglue/baseprofiler/public/BaseProfilerDetail.h
new file mode 100644
index 0000000000..9027f32bc7
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilerDetail.h
@@ -0,0 +1,189 @@
+/* -*- Mode: C++; tab-width: 2; 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/. */
+
+// Internal Base Profiler utilities.
+
+#ifndef BaseProfilerDetail_h
+#define BaseProfilerDetail_h
+
+#include "mozilla/Atomics.h"
+#include "mozilla/Maybe.h"
+#include "mozilla/PlatformMutex.h"
+
+#ifndef MOZ_GECKO_PROFILER
+#  error Do not #include this header when MOZ_GECKO_PROFILER is not #defined.
+#endif
+
+namespace mozilla {
+namespace baseprofiler {
+
+// Implemented in platform.cpp
+MFBT_API int profiler_current_thread_id();
+
+namespace detail {
+
+// Thin shell around mozglue PlatformMutex, for Base Profiler internal use.
+class BaseProfilerMutex : private ::mozilla::detail::MutexImpl {
+ public:
+  BaseProfilerMutex() : ::mozilla::detail::MutexImpl() {}
+  explicit BaseProfilerMutex(const char* aName)
+      : ::mozilla::detail::MutexImpl(), mName(aName) {}
+
+  BaseProfilerMutex(const BaseProfilerMutex&) = delete;
+  BaseProfilerMutex& operator=(const BaseProfilerMutex&) = delete;
+  BaseProfilerMutex(BaseProfilerMutex&&) = delete;
+  BaseProfilerMutex& operator=(BaseProfilerMutex&&) = delete;
+
+#ifdef DEBUG
+  ~BaseProfilerMutex() { MOZ_ASSERT(mOwningThreadId == 0); }
+#endif  // DEBUG
+
+  [[nodiscard]] bool IsLockedOnCurrentThread() const {
+    return mOwningThreadId == baseprofiler::profiler_current_thread_id();
+  }
+
+  void AssertCurrentThreadOwns() const {
+    MOZ_ASSERT(IsLockedOnCurrentThread());
+  }
+
+  void Lock() {
+    const int tid = baseprofiler::profiler_current_thread_id();
+    MOZ_ASSERT(tid != 0);
+    MOZ_ASSERT(!IsLockedOnCurrentThread(), "Recursive locking");
+    ::mozilla::detail::MutexImpl::lock();
+    MOZ_ASSERT(mOwningThreadId == 0, "Not unlocked properly");
+    mOwningThreadId = tid;
+  }
+
+  [[nodiscard]] bool TryLock() {
+    const int tid = baseprofiler::profiler_current_thread_id();
+    MOZ_ASSERT(tid != 0);
+    MOZ_ASSERT(!IsLockedOnCurrentThread(), "Recursive locking");
+    if (!::mozilla::detail::MutexImpl::tryLock()) {
+      // Failed to lock, nothing more to do.
+      return false;
+    }
+    MOZ_ASSERT(mOwningThreadId == 0, "Not unlocked properly");
+    mOwningThreadId = tid;
+    return true;
+  }
+
+  void Unlock() {
+    MOZ_ASSERT(IsLockedOnCurrentThread(), "Unlocking when not locked here");
+    // We're still holding the mutex here, so it's safe to just reset
+    // `mOwningThreadId`.
+    mOwningThreadId = 0;
+    ::mozilla::detail::MutexImpl::unlock();
+  }
+
+  const char* GetName() const { return mName; }
+
+ private:
+  // Thread currently owning the lock, or 0.
+  // Atomic because it may be read at any time independent of the mutex.
+  // Relaxed because threads only need to know if they own it already, so:
+  // - If it's their id, only *they* wrote that value with a locked mutex.
+  // - If it's different from their thread id it doesn't matter what other
+  //   number it is (0 or another id) and that it can change again at any time.
+  Atomic<int, MemoryOrdering::Relaxed> mOwningThreadId{0};
+
+  const char* mName = nullptr;
+};
+
+// RAII class to lock a mutex.
+class MOZ_RAII BaseProfilerAutoLock {
+ public:
+  explicit BaseProfilerAutoLock(BaseProfilerMutex& aMutex) : mMutex(aMutex) {
+    mMutex.Lock();
+  }
+
+  BaseProfilerAutoLock(const BaseProfilerAutoLock&) = delete;
+  BaseProfilerAutoLock& operator=(const BaseProfilerAutoLock&) = delete;
+  BaseProfilerAutoLock(BaseProfilerAutoLock&&) = delete;
+  BaseProfilerAutoLock& operator=(BaseProfilerAutoLock&&) = delete;
+
+  ~BaseProfilerAutoLock() { mMutex.Unlock(); }
+
+ private:
+  BaseProfilerMutex& mMutex;
+};
+
+// Thin shell around mozglue PlatformMutex, for Base Profiler internal use.
+// Actual mutex may be disabled at construction time.
+class BaseProfilerMaybeMutex : private ::mozilla::detail::MutexImpl {
+ public:
+  explicit BaseProfilerMaybeMutex(bool aActivate) {
+    if (aActivate) {
+      mMaybeMutex.emplace();
+    }
+  }
+
+  BaseProfilerMaybeMutex(const BaseProfilerMaybeMutex&) = delete;
+  BaseProfilerMaybeMutex& operator=(const BaseProfilerMaybeMutex&) = delete;
+  BaseProfilerMaybeMutex(BaseProfilerMaybeMutex&&) = delete;
+  BaseProfilerMaybeMutex& operator=(BaseProfilerMaybeMutex&&) = delete;
+
+  ~BaseProfilerMaybeMutex() = default;
+
+  bool IsActivated() const { return mMaybeMutex.isSome(); }
+
+  [[nodiscard]] bool IsActivatedAndLockedOnCurrentThread() const {
+    if (!IsActivated()) {
+      // Not activated, so we can never be locked.
+      return false;
+    }
+    return mMaybeMutex->IsLockedOnCurrentThread();
+  }
+
+  void AssertCurrentThreadOwns() const {
+#ifdef DEBUG
+    if (IsActivated()) {
+      mMaybeMutex->AssertCurrentThreadOwns();
+    }
+#endif  // DEBUG
+  }
+
+  void Lock() {
+    if (IsActivated()) {
+      mMaybeMutex->Lock();
+    }
+  }
+
+  void Unlock() {
+    if (IsActivated()) {
+      mMaybeMutex->Unlock();
+    }
+  }
+
+ private:
+  Maybe<BaseProfilerMutex> mMaybeMutex;
+};
+
+// RAII class to lock a mutex.
+class MOZ_RAII BaseProfilerMaybeAutoLock {
+ public:
+  explicit BaseProfilerMaybeAutoLock(BaseProfilerMaybeMutex& aMaybeMutex)
+      : mMaybeMutex(aMaybeMutex) {
+    mMaybeMutex.Lock();
+  }
+
+  BaseProfilerMaybeAutoLock(const BaseProfilerMaybeAutoLock&) = delete;
+  BaseProfilerMaybeAutoLock& operator=(const BaseProfilerMaybeAutoLock&) =
+      delete;
+  BaseProfilerMaybeAutoLock(BaseProfilerMaybeAutoLock&&) = delete;
+  BaseProfilerMaybeAutoLock& operator=(BaseProfilerMaybeAutoLock&&) = delete;
+
+  ~BaseProfilerMaybeAutoLock() { mMaybeMutex.Unlock(); }
+
+ private:
+  BaseProfilerMaybeMutex& mMaybeMutex;
+};
+
+}  // namespace detail
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif  // BaseProfilerDetail_h
diff --git a/mozglue/baseprofiler/public/BaseProfilerMarkerTypes.h b/mozglue/baseprofiler/public/BaseProfilerMarkerTypes.h
new file mode 100644
index 0000000000..1556b7a272
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilerMarkerTypes.h
@@ -0,0 +1,69 @@
+/* -*- Mode: C++; tab-width: 2; 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 BaseProfilerMarkerTypes_h
+#define BaseProfilerMarkerTypes_h
+
+// This header contains common marker type definitions.
+//
+// It #include's "mozilla/BaseProfilerMarkers.h", see that file for how to
+// define other marker types, and how to add markers to the profiler buffers.
+//
+// If you don't need to use these common types, #include
+// "mozilla/BaseProfilerMarkers.h" instead.
+//
+// Types in this files can be defined without relying on xpcom.
+// Others are defined in "ProfilerMarkerTypes.h".
+
+// !!!                       /!\ WORK IN PROGRESS /!\                       !!!
+// This file contains draft marker definitions, but most are not used yet.
+// Further work is needed to complete these definitions, and use them to convert
+// existing PROFILER_ADD_MARKER calls. See meta bug 1661394.
+
+#include "mozilla/BaseProfilerMarkers.h"
+
+#ifdef MOZ_GECKO_PROFILER
+
+namespace mozilla::baseprofiler::markers {
+
+struct MediaSampleMarker {
+  static constexpr Span<const char> MarkerTypeName() {
+    return MakeStringSpan("MediaSample");
+  }
+  static void StreamJSONMarkerData(SpliceableJSONWriter& aWriter,
+                                   int64_t aSampleStartTimeUs,
+                                   int64_t aSampleEndTimeUs) {
+    aWriter.IntProperty("sampleStartTimeUs", aSampleStartTimeUs);
+    aWriter.IntProperty("sampleEndTimeUs", aSampleEndTimeUs);
+  }
+  static MarkerSchema MarkerTypeDisplay() {
+    using MS = MarkerSchema;
+    MS schema{MS::Location::markerChart, MS::Location::markerTable};
+    schema.AddKeyLabelFormat("sampleStartTimeUs", "Sample start time",
+                             MS::Format::microseconds);
+    schema.AddKeyLabelFormat("sampleEndTimeUs", "Sample end time",
+                             MS::Format::microseconds);
+    return schema;
+  }
+};
+
+struct ContentBuildMarker {
+  static constexpr Span<const char> MarkerTypeName() {
+    return MakeStringSpan("CONTENT_FULL_PAINT_TIME");
+  }
+  static void StreamJSONMarkerData(SpliceableJSONWriter& aWriter) {}
+  static MarkerSchema MarkerTypeDisplay() {
+    using MS = MarkerSchema;
+    MS schema{MS::Location::markerChart, MS::Location::markerTable};
+    return schema;
+  }
+};
+
+}  // namespace mozilla::baseprofiler::markers
+
+#endif  // MOZ_GECKO_PROFILER
+
+#endif  // BaseProfilerMarkerTypes_h
diff --git a/mozglue/baseprofiler/public/BaseProfilerMarkers.h b/mozglue/baseprofiler/public/BaseProfilerMarkers.h
new file mode 100644
index 0000000000..c63b018f95
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilerMarkers.h
@@ -0,0 +1,242 @@
+/* -*- Mode: C++; tab-width: 2; 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/. */
+
+// Markers are useful to delimit something important happening such as the first
+// paint. Unlike labels, which are only recorded in the profile buffer if a
+// sample is collected while the label is on the label stack, markers will
+// always be recorded in the profile buffer.
+//
+// This header contains basic definitions necessary to create marker types, and
+// to add markers to the profiler buffers.
+//
+// If basic marker types are needed, #include
+// "mozilla/BaseProfilerMarkerTypes.h" instead.
+//
+// But if you want to create your own marker type locally, you can #include this
+// header only; look at mozilla/BaseProfilerMarkerTypes.h for examples of how to
+// define types, and mozilla/BaseProfilerMarkerPrerequisites.h for some
+// supporting types.
+//
+// To then record markers:
+// - Use `baseprofiler::AddMarker(...)` from  mozglue or other libraries that
+//   are outside of xul, especially if they may happen outside of xpcom's
+//   lifetime (typically startup, shutdown, or tests).
+// - Otherwise #include "ProfilerMarkers.h" instead, and use
+//   `profiler_add_marker(...)`.
+// See these functions for more details.
+
+#ifndef BaseProfilerMarkers_h
+#define BaseProfilerMarkers_h
+
+#include "mozilla/BaseProfilerMarkersDetail.h"
+
+#ifndef MOZ_GECKO_PROFILER
+
+#  define BASE_PROFILER_MARKER_UNTYPED(markerName, categoryName, ...)
+#  define BASE_PROFILER_MARKER(markerName, categoryName, options, MarkerType, \
+                               ...)
+#  define BASE_PROFILER_MARKER_TEXT(markerName, categoryName, options, text)
+#  define AUTO_BASE_PROFILER_MARKER_TEXT(markerName, categoryName, options, \
+                                         text)
+
+#else  // ndef MOZ_GECKO_PROFILER
+
+#  include "mozilla/ProfileChunkedBuffer.h"
+#  include "mozilla/TimeStamp.h"
+#  include "mozilla/Unused.h"
+
+#  include <functional>
+#  include <string>
+#  include <utility>
+
+namespace mozilla::baseprofiler {
+
+// Add a marker to a given buffer. `AddMarker()` and related macros should be
+// used in most cases, see below for more information about them and the
+// parameters; This function may be useful when markers need to be recorded in a
+// local buffer outside of the main profiler buffer.
+template <typename MarkerType, typename... PayloadArguments>
+ProfileBufferBlockIndex AddMarkerToBuffer(
+    ProfileChunkedBuffer& aBuffer, const ProfilerString8View& aName,
+    const MarkerCategory& aCategory, MarkerOptions&& aOptions,
+    MarkerType aMarkerType, const PayloadArguments&... aPayloadArguments) {
+  Unused << aMarkerType;  // Only the empty object type is useful.
+  return base_profiler_markers_detail::AddMarkerToBuffer<MarkerType>(
+      aBuffer, aName, aCategory, std::move(aOptions),
+      ::mozilla::baseprofiler::profiler_capture_backtrace_into,
+      aPayloadArguments...);
+}
+
+// Add a marker (without payload) to a given buffer.
+inline ProfileBufferBlockIndex AddMarkerToBuffer(
+    ProfileChunkedBuffer& aBuffer, const ProfilerString8View& aName,
+    const MarkerCategory& aCategory, MarkerOptions&& aOptions = {}) {
+  return AddMarkerToBuffer(aBuffer, aName, aCategory, std::move(aOptions),
+                           markers::NoPayload{});
+}
+
+// Add a marker to the Base Profiler buffer.
+// - aName: Main name of this marker.
+// - aCategory: Category for this marker.
+// - aOptions: Optional settings (such as timing, inner window id,
+//   backtrace...), see `MarkerOptions` for details.
+// - aMarkerType: Empty object that specifies the type of marker.
+// - aPayloadArguments: Arguments expected by this marker type's
+// ` StreamJSONMarkerData` function.
+template <typename MarkerType, typename... PayloadArguments>
+ProfileBufferBlockIndex AddMarker(
+    const ProfilerString8View& aName, const MarkerCategory& aCategory,
+    MarkerOptions&& aOptions, MarkerType aMarkerType,
+    const PayloadArguments&... aPayloadArguments) {
+  if (!baseprofiler::profiler_can_accept_markers()) {
+    return {};
+  }
+  return ::mozilla::baseprofiler::AddMarkerToBuffer(
+      base_profiler_markers_detail::CachedBaseCoreBuffer(), aName, aCategory,
+      std::move(aOptions), aMarkerType, aPayloadArguments...);
+}
+
+// Add a marker (without payload) to the Base Profiler buffer.
+inline ProfileBufferBlockIndex AddMarker(const ProfilerString8View& aName,
+                                         const MarkerCategory& aCategory,
+                                         MarkerOptions&& aOptions = {}) {
+  return AddMarker(aName, aCategory, std::move(aOptions), markers::NoPayload{});
+}
+
+}  // namespace mozilla::baseprofiler
+
+// Same as `AddMarker()` (without payload). This macro is safe to use even if
+// MOZ_GECKO_PROFILER is not #defined.
+#  define BASE_PROFILER_MARKER_UNTYPED(markerName, categoryName, ...)  \
+    do {                                                               \
+      AUTO_PROFILER_STATS(BASE_PROFILER_MARKER_UNTYPED);               \
+      ::mozilla::baseprofiler::AddMarker(                              \
+          markerName, ::mozilla::baseprofiler::category::categoryName, \
+          ##__VA_ARGS__);                                              \
+    } while (false)
+
+// Same as `AddMarker()` (with payload). This macro is safe to use even if
+// MOZ_GECKO_PROFILER is not #defined.
+#  define BASE_PROFILER_MARKER(markerName, categoryName, options, MarkerType, \
+                               ...)                                           \
+    do {                                                                      \
+      AUTO_PROFILER_STATS(BASE_PROFILER_MARKER_with_##MarkerType);            \
+      ::mozilla::baseprofiler::AddMarker(                                     \
+          markerName, ::mozilla::baseprofiler::category::categoryName,        \
+          options, ::mozilla::baseprofiler::markers::MarkerType{},            \
+          ##__VA_ARGS__);                                                     \
+    } while (false)
+
+namespace mozilla::baseprofiler::markers {
+// Most common marker type. Others are in BaseProfilerMarkerTypes.h.
+struct TextMarker {
+  static constexpr Span<const char> MarkerTypeName() {
+    return MakeStringSpan("Text");
+  }
+  static void StreamJSONMarkerData(baseprofiler::SpliceableJSONWriter& aWriter,
+                                   const ProfilerString8View& aText) {
+    aWriter.StringProperty("name", aText);
+  }
+  static MarkerSchema MarkerTypeDisplay() {
+    using MS = MarkerSchema;
+    MS schema{MS::Location::markerChart, MS::Location::markerTable};
+    schema.SetChartLabel("{marker.data.name}");
+    schema.SetTableLabel("{marker.name} - {marker.data.name}");
+    schema.AddKeyLabelFormat("name", "Details", MarkerSchema::Format::string);
+    return schema;
+  }
+};
+
+struct Tracing {
+  static constexpr Span<const char> MarkerTypeName() {
+    return MakeStringSpan("tracing");
+  }
+  static void StreamJSONMarkerData(SpliceableJSONWriter& aWriter,
+                                   const ProfilerString8View& aCategory) {
+    if (aCategory.Length() != 0) {
+      aWriter.StringProperty("category", aCategory);
+    }
+  }
+  static MarkerSchema MarkerTypeDisplay() {
+    using MS = MarkerSchema;
+    MS schema{MS::Location::markerChart, MS::Location::markerTable,
+              MS::Location::timelineOverview};
+    schema.AddKeyLabelFormat("category", "Type", MS::Format::string);
+    return schema;
+  }
+};
+}  // namespace mozilla::baseprofiler::markers
+
+// Add a text marker. This macro is safe to use even if MOZ_GECKO_PROFILER is
+// not #defined.
+#  define BASE_PROFILER_MARKER_TEXT(markerName, categoryName, options, text) \
+    do {                                                                     \
+      AUTO_PROFILER_STATS(BASE_PROFILER_MARKER_TEXT);                        \
+      ::mozilla::baseprofiler::AddMarker(                                    \
+          markerName, ::mozilla::baseprofiler::category::categoryName,       \
+          options, ::mozilla::baseprofiler::markers::TextMarker{}, text);    \
+    } while (false)
+
+namespace mozilla::baseprofiler {
+
+// RAII object that adds a BASE_PROFILER_MARKER_TEXT when destroyed; the
+// marker's timing will be the interval from construction (unless an instant or
+// start time is already specified in the provided options) until destruction.
+class MOZ_RAII AutoProfilerTextMarker {
+ public:
+  AutoProfilerTextMarker(const char* aMarkerName,
+                         const MarkerCategory& aCategory,
+                         MarkerOptions&& aOptions, const std::string& aText)
+      : mMarkerName(aMarkerName),
+        mCategory(aCategory),
+        mOptions(std::move(aOptions)),
+        mText(aText) {
+    MOZ_ASSERT(mOptions.Timing().EndTime().IsNull(),
+               "AutoProfilerTextMarker options shouldn't have an end time");
+    if (mOptions.Timing().StartTime().IsNull()) {
+      mOptions.Set(MarkerTiming::InstantNow());
+    }
+  }
+
+  ~AutoProfilerTextMarker() {
+    mOptions.TimingRef().SetIntervalEnd();
+    AUTO_PROFILER_STATS(AUTO_BASE_PROFILER_MARKER_TEXT);
+    AddMarker(ProfilerString8View::WrapNullTerminatedString(mMarkerName),
+              mCategory, std::move(mOptions), markers::TextMarker{}, mText);
+  }
+
+ protected:
+  const char* mMarkerName;
+  MarkerCategory mCategory;
+  MarkerOptions mOptions;
+  std::string mText;
+};
+
+extern template MFBT_API ProfileBufferBlockIndex
+AddMarker(const ProfilerString8View&, const MarkerCategory&, MarkerOptions&&,
+          markers::TextMarker, const std::string&);
+
+extern template MFBT_API ProfileBufferBlockIndex
+AddMarkerToBuffer(ProfileChunkedBuffer&, const ProfilerString8View&,
+                  const MarkerCategory&, MarkerOptions&&, markers::NoPayload);
+
+extern template MFBT_API ProfileBufferBlockIndex AddMarkerToBuffer(
+    ProfileChunkedBuffer&, const ProfilerString8View&, const MarkerCategory&,
+    MarkerOptions&&, markers::TextMarker, const std::string&);
+
+}  // namespace mozilla::baseprofiler
+
+// Creates an AutoProfilerTextMarker RAII object.  This macro is safe to use
+// even if MOZ_GECKO_PROFILER is not #defined.
+#  define AUTO_BASE_PROFILER_MARKER_TEXT(markerName, categoryName, options,   \
+                                         text)                                \
+    ::mozilla::baseprofiler::AutoProfilerTextMarker BASE_PROFILER_RAII(       \
+        markerName, ::mozilla::baseprofiler::category::categoryName, options, \
+        text)
+
+#endif  // nfed MOZ_GECKO_PROFILER else
+
+#endif  // BaseProfilerMarkers_h
diff --git a/mozglue/baseprofiler/public/BaseProfilerMarkersDetail.h b/mozglue/baseprofiler/public/BaseProfilerMarkersDetail.h
new file mode 100644
index 0000000000..b5dbe27343
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilerMarkersDetail.h
@@ -0,0 +1,674 @@
+/* -*- Mode: C++; tab-width: 2; 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 BaseProfilerMarkersDetail_h
+#define BaseProfilerMarkersDetail_h
+
+#ifndef BaseProfilerMarkers_h
+#  error "This header should only be #included by BaseProfilerMarkers.h"
+#endif
+
+#include "mozilla/BaseProfilerMarkersPrerequisites.h"
+
+#ifdef MOZ_GECKO_PROFILER
+
+//                        ~~ HERE BE DRAGONS ~~
+//
+// Everything below is internal implementation detail, you shouldn't need to
+// look at it unless working on the profiler code.
+
+#  include "mozilla/BaseProfileJSONWriter.h"
+#  include "mozilla/ProfileBufferEntryKinds.h"
+
+#  include <limits>
+#  include <tuple>
+
+namespace mozilla::baseprofiler {
+// Implemented in platform.cpp
+MFBT_API ProfileChunkedBuffer& profiler_get_core_buffer();
+}  // namespace mozilla::baseprofiler
+
+namespace mozilla::base_profiler_markers_detail {
+
+// Get the core buffer from the profiler, and cache it in a
+// non-templated-function static reference.
+inline ProfileChunkedBuffer& CachedBaseCoreBuffer() {
+  static ProfileChunkedBuffer& coreBuffer =
+      baseprofiler::profiler_get_core_buffer();
+  return coreBuffer;
+}
+
+struct Streaming {
+  // A `MarkerDataDeserializer` is a free function that can read a serialized
+  // payload from an `EntryReader` and streams it as JSON object properties.
+  using MarkerDataDeserializer = void (*)(ProfileBufferEntryReader&,
+                                          baseprofiler::SpliceableJSONWriter&);
+
+  // A `MarkerTypeNameFunction` is a free function that returns the name of the
+  // marker type.
+  using MarkerTypeNameFunction = Span<const char> (*)();
+
+  // A `MarkerSchemaFunction` is a free function that returns a
+  // `MarkerSchema`, which contains all the information needed to stream
+  // the display schema associated with a marker type.
+  using MarkerSchemaFunction = MarkerSchema (*)();
+
+  struct MarkerTypeFunctions {
+    MarkerDataDeserializer mMarkerDataDeserializer = nullptr;
+    MarkerTypeNameFunction mMarkerTypeNameFunction = nullptr;
+    MarkerSchemaFunction mMarkerSchemaFunction = nullptr;
+  };
+
+  // A `DeserializerTag` will be added before the payload, to help select the
+  // correct deserializer when reading back the payload.
+  using DeserializerTag = uint8_t;
+
+  // Store a deserializer (and other marker-type-specific functions) and get its
+  // `DeserializerTag`.
+  // This is intended to be only used once per deserializer when a new marker
+  // type is used for the first time, so it should be called to initialize a
+  // `static const` tag that will be re-used by all markers of the corresponding
+  // payload type -- see use below.
+  MFBT_API static DeserializerTag TagForMarkerTypeFunctions(
+      MarkerDataDeserializer aDeserializer,
+      MarkerTypeNameFunction aMarkerTypeNameFunction,
+      MarkerSchemaFunction aMarkerSchemaFunction);
+
+  // Get the `MarkerDataDeserializer` for a given `DeserializerTag`.
+  MFBT_API static MarkerDataDeserializer DeserializerForTag(
+      DeserializerTag aTag);
+
+  // Retrieve all MarkerTypeFunctions's.
+  MFBT_API static Span<const MarkerTypeFunctions> MarkerTypeFunctionsArray();
+};
+
+// This helper will examine a marker type's `StreamJSONMarkerData` function, see
+// specialization below.
+template <typename T>
+struct StreamFunctionTypeHelper;
+
+// Helper specialization that takes the expected
+// `StreamJSONMarkerData(baseprofiler::SpliceableJSONWriter&, ...)` function and
+// provide information about the `...` parameters.
+template <typename R, typename... As>
+struct StreamFunctionTypeHelper<R(baseprofiler::SpliceableJSONWriter&, As...)> {
+  constexpr static size_t scArity = sizeof...(As);
+  using TupleType =
+      std::tuple<std::remove_cv_t<std::remove_reference_t<As>>...>;
+
+  // Serialization function that takes the exact same parameter types
+  // (const-ref'd) as `StreamJSONMarkerData`. This has to be inside the helper
+  // because only here can we access the raw parameter pack `As...`.
+  // And because we're using the same argument types through
+  // references-to-const, permitted implicit conversions can happen.
+  static ProfileBufferBlockIndex Serialize(
+      ProfileChunkedBuffer& aBuffer, const ProfilerString8View& aName,
+      const MarkerCategory& aCategory, MarkerOptions&& aOptions,
+      Streaming::DeserializerTag aDeserializerTag, const As&... aAs) {
+    // Note that options are first after the entry kind, because they contain
+    // the thread id, which is handled first to filter markers by threads.
+    return aBuffer.PutObjects(ProfileBufferEntryKind::Marker, aOptions, aName,
+                              aCategory, aDeserializerTag, aAs...);
+  }
+};
+
+// Helper for a marker type.
+// A marker type is defined in a `struct` with some expected static member
+// functions. See example in BaseProfilerMarkers.h.
+template <typename MarkerType>
+struct MarkerTypeSerialization {
+  // Definitions to access the expected
+  // `StreamJSONMarkerData(baseprofiler::SpliceableJSONWriter&, ...)` function
+  // and its parameters.
+  using StreamFunctionType =
+      StreamFunctionTypeHelper<decltype(MarkerType::StreamJSONMarkerData)>;
+  constexpr static size_t scStreamFunctionParameterCount =
+      StreamFunctionType::scArity;
+  using StreamFunctionUserParametersTuple =
+      typename StreamFunctionType::TupleType;
+  template <size_t i>
+  using StreamFunctionParameter =
+      std::tuple_element_t<i, StreamFunctionUserParametersTuple>;
+
+  template <typename... Ts>
+  static ProfileBufferBlockIndex Serialize(ProfileChunkedBuffer& aBuffer,
+                                           const ProfilerString8View& aName,
+                                           const MarkerCategory& aCategory,
+                                           MarkerOptions&& aOptions,
+                                           const Ts&... aTs) {
+    static_assert(!std::is_same_v<MarkerType,
+                                  ::mozilla::baseprofiler::markers::NoPayload>,
+                  "NoPayload should have been handled in the caller.");
+    // Register marker type functions, and get the tag for this deserializer.
+    // Note that the tag is stored in a function-static object, and this
+    // function is static in a templated struct, so there should only be one tag
+    // per MarkerType.
+    // Making the tag class-static may have been more efficient (to avoid a
+    // thread-safe init check at every call), but random global static
+    // initialization order would make it more complex to coordinate with
+    // `Streaming::TagForMarkerTypeFunctions()`, and also would add a (small)
+    // cost for everybody, even the majority of users not using the profiler.
+    static const Streaming::DeserializerTag tag =
+        Streaming::TagForMarkerTypeFunctions(Deserialize,
+                                             MarkerType::MarkerTypeName,
+                                             MarkerType::MarkerTypeDisplay);
+    return StreamFunctionType::Serialize(aBuffer, aName, aCategory,
+                                         std::move(aOptions), tag, aTs...);
+  }
+
+ private:
+  // This templated function will recursively deserialize each argument expected
+  // by `MarkerType::StreamJSONMarkerData()` on the stack, and call it at the
+  // end. E.g., for `StreamJSONMarkerData(int, char)`:
+  // - DeserializeArguments<0>(aER, aWriter) reads an int and calls:
+  // - DeserializeArguments<1>(aER, aWriter, const int&) reads a char and calls:
+  // - MarkerType::StreamJSONMarkerData(aWriter, const int&, const char&).
+  // Prototyping on godbolt showed that clang and gcc can flatten these
+  // recursive calls into one function with successive reads followed by the one
+  // stream call; tested up to 40 arguments: https://godbolt.org/z/5KeeM4
+  template <size_t i = 0, typename... Args>
+  static void DeserializeArguments(ProfileBufferEntryReader& aEntryReader,
+                                   baseprofiler::SpliceableJSONWriter& aWriter,
+                                   const Args&... aArgs) {
+    static_assert(sizeof...(Args) == i,
+                  "We should have collected `i` arguments so far");
+    if constexpr (i < scStreamFunctionParameterCount) {
+      // Deserialize the i-th argument on this stack.
+      auto argument = aEntryReader.ReadObject<StreamFunctionParameter<i>>();
+      // Add our local argument to the next recursive call.
+      DeserializeArguments<i + 1>(aEntryReader, aWriter, aArgs..., argument);
+    } else {
+      // We've read all the arguments, finally call the `StreamJSONMarkerData`
+      // function, which should write the appropriate JSON elements for this
+      // marker type. Note that the MarkerType-specific "type" element is
+      // already written.
+      MarkerType::StreamJSONMarkerData(aWriter, aArgs...);
+    }
+  }
+
+ public:
+  static void Deserialize(ProfileBufferEntryReader& aEntryReader,
+                          baseprofiler::SpliceableJSONWriter& aWriter) {
+    aWriter.StringProperty("type", MarkerType::MarkerTypeName());
+    DeserializeArguments(aEntryReader, aWriter);
+  }
+};
+
+template <>
+struct MarkerTypeSerialization<::mozilla::baseprofiler::markers::NoPayload> {
+  // Nothing! NoPayload has special handling avoiding payload work.
+};
+
+template <typename MarkerType, typename... Ts>
+static ProfileBufferBlockIndex AddMarkerWithOptionalStackToBuffer(
+    ProfileChunkedBuffer& aBuffer, const ProfilerString8View& aName,
+    const MarkerCategory& aCategory, MarkerOptions&& aOptions,
+    const Ts&... aTs) {
+  if constexpr (std::is_same_v<MarkerType,
+                               ::mozilla::baseprofiler::markers::NoPayload>) {
+    static_assert(sizeof...(Ts) == 0,
+                  "NoPayload does not accept any payload arguments.");
+    // Special case for NoPayload where there is a stack or inner window id:
+    // Because these options would be stored in the payload 'data' object, but
+    // there is no such object for NoPayload, we convert the marker to another
+    // type (without user fields in the 'data' object), so that the stack and/or
+    // inner window id are not lost.
+    // TODO: Remove this when bug 1646714 lands.
+    if (aOptions.Stack().GetChunkedBuffer() ||
+        !aOptions.InnerWindowId().IsUnspecified()) {
+      struct NoPayloadUserData {
+        static constexpr Span<const char> MarkerTypeName() {
+          return MakeStringSpan("NoPayloadUserData");
+        }
+        static void StreamJSONMarkerData(
+            baseprofiler::SpliceableJSONWriter& aWriter) {
+          // No user payload.
+        }
+        static mozilla::MarkerSchema MarkerTypeDisplay() {
+          using MS = mozilla::MarkerSchema;
+          MS schema{MS::Location::markerChart, MS::Location::markerTable};
+          // No user data to display.
+          return schema;
+        }
+      };
+      return MarkerTypeSerialization<NoPayloadUserData>::Serialize(
+          aBuffer, aName, aCategory, std::move(aOptions));
+    }
+
+    // Note that options are first after the entry kind, because they contain
+    // the thread id, which is handled first to filter markers by threads.
+    return aBuffer.PutObjects(
+        ProfileBufferEntryKind::Marker, std::move(aOptions), aName, aCategory,
+        base_profiler_markers_detail::Streaming::DeserializerTag(0));
+  } else {
+    return MarkerTypeSerialization<MarkerType>::Serialize(
+        aBuffer, aName, aCategory, std::move(aOptions), aTs...);
+  }
+}
+
+// Pointer to a function that can capture a backtrace into the provided
+// `ProfileChunkedBuffer`, and returns true when successful.
+using BacktraceCaptureFunction = bool (*)(ProfileChunkedBuffer&);
+
+// Add a marker with the given name, options, and arguments to the given buffer.
+// Because this may be called from either Base or Gecko Profiler functions, the
+// appropriate backtrace-capturing function must also be provided.
+template <typename MarkerType, typename... Ts>
+ProfileBufferBlockIndex AddMarkerToBuffer(
+    ProfileChunkedBuffer& aBuffer, const ProfilerString8View& aName,
+    const MarkerCategory& aCategory, MarkerOptions&& aOptions,
+    BacktraceCaptureFunction aBacktraceCaptureFunction, const Ts&... aTs) {
+  if (aOptions.ThreadId().IsUnspecified()) {
+    // If yet unspecified, set thread to this thread where the marker is added.
+    aOptions.Set(MarkerThreadId::CurrentThread());
+  }
+
+  if (aOptions.IsTimingUnspecified()) {
+    // If yet unspecified, set timing to this instant of adding the marker.
+    aOptions.Set(MarkerTiming::InstantNow());
+  }
+
+  if (aOptions.Stack().IsCaptureNeeded()) {
+    // A capture was requested, let's attempt to do it here&now. This avoids a
+    // lot of allocations that would be necessary if capturing a backtrace
+    // separately.
+    // TODO use a local on-stack byte buffer to remove last allocation.
+    // TODO reduce internal profiler stack levels, see bug 1659872.
+    ProfileBufferChunkManagerSingle chunkManager(
+        ProfileBufferChunkManager::scExpectedMaximumStackSize);
+    ProfileChunkedBuffer chunkedBuffer(
+        ProfileChunkedBuffer::ThreadSafety::WithoutMutex, chunkManager);
+    aOptions.StackRef().UseRequestedBacktrace(
+        aBacktraceCaptureFunction(chunkedBuffer) ? &chunkedBuffer : nullptr);
+    // This call must be made from here, while chunkedBuffer is in scope.
+    return AddMarkerWithOptionalStackToBuffer<MarkerType>(
+        aBuffer, aName, aCategory, std::move(aOptions), aTs...);
+  }
+
+  return AddMarkerWithOptionalStackToBuffer<MarkerType>(
+      aBuffer, aName, aCategory, std::move(aOptions), aTs...);
+}
+
+template <typename StackCallback>
+[[nodiscard]] bool DeserializeAfterKindAndStream(
+    ProfileBufferEntryReader& aEntryReader,
+    baseprofiler::SpliceableJSONWriter& aWriter, int aThreadIdOrZero,
+    StackCallback&& aStackCallback) {
+  // Each entry is made up of the following:
+  //   ProfileBufferEntry::Kind::Marker, <- already read by caller
+  //   options,                          <- next location in entries
+  //   name,
+  //   payload
+  const MarkerOptions options = aEntryReader.ReadObject<MarkerOptions>();
+  if (aThreadIdOrZero != 0 &&
+      options.ThreadId().ThreadId() != aThreadIdOrZero) {
+    // A specific thread is being read, we're not in it.
+    return false;
+  }
+  // Write the information to JSON with the following schema:
+  // [name, startTime, endTime, phase, category, data]
+  aWriter.StartArrayElement();
+  {
+    aWriter.UniqueStringElement(aEntryReader.ReadObject<ProfilerString8View>());
+
+    const double startTime = options.Timing().GetStartTime();
+    aWriter.DoubleElement(startTime);
+
+    const double endTime = options.Timing().GetEndTime();
+    aWriter.DoubleElement(endTime);
+
+    aWriter.IntElement(static_cast<int64_t>(options.Timing().MarkerPhase()));
+
+    MarkerCategory category = aEntryReader.ReadObject<MarkerCategory>();
+    aWriter.IntElement(static_cast<int64_t>(category.GetCategory()));
+
+    if (const auto tag =
+            aEntryReader.ReadObject<mozilla::base_profiler_markers_detail::
+                                        Streaming::DeserializerTag>();
+        tag != 0) {
+      aWriter.StartObjectElement(JSONWriter::SingleLineStyle);
+      {
+        // Stream "common props".
+
+        // TODO: Move this to top-level tuple, when frontend supports it.
+        if (!options.InnerWindowId().IsUnspecified()) {
+          // Here, we are converting uint64_t to double. Both Browsing Context
+          // and Inner Window IDs are created using
+          // `nsContentUtils::GenerateProcessSpecificId`, which is specifically
+          // designed to only use 53 of the 64 bits to be lossless when passed
+          // into and out of JS as a double.
+          aWriter.DoubleProperty(
+              "innerWindowID",
+              static_cast<double>(options.InnerWindowId().Id()));
+        }
+
+        // TODO: Move this to top-level tuple, when frontend supports it.
+        if (ProfileChunkedBuffer* chunkedBuffer =
+                options.Stack().GetChunkedBuffer();
+            chunkedBuffer) {
+          aWriter.StartObjectProperty("stack");
+          { std::forward<StackCallback>(aStackCallback)(*chunkedBuffer); }
+          aWriter.EndObject();
+        }
+
+        // Stream the payload, including the type.
+        mozilla::base_profiler_markers_detail::Streaming::MarkerDataDeserializer
+            deserializer = mozilla::base_profiler_markers_detail::Streaming::
+                DeserializerForTag(tag);
+        MOZ_RELEASE_ASSERT(deserializer);
+        deserializer(aEntryReader, aWriter);
+      }
+      aWriter.EndObject();
+    }
+  }
+  aWriter.EndArray();
+  return true;
+}
+
+}  // namespace mozilla::base_profiler_markers_detail
+
+namespace mozilla {
+
+// ----------------------------------------------------------------------------
+// Serializer, Deserializer: ProfilerStringView<CHAR>
+
+// The serialization starts with a ULEB128 number that encodes both whether the
+// ProfilerStringView is literal (Least Significant Bit = 0) or not (LSB = 1),
+// plus the string length (excluding null terminator) in bytes, shifted left by
+// 1 bit. Following that number:
+// - If literal, the string pointer value.
+// - If non-literal, the contents as bytes (excluding null terminator if any).
+template <typename CHAR>
+struct ProfileBufferEntryWriter::Serializer<ProfilerStringView<CHAR>> {
+  static Length Bytes(const ProfilerStringView<CHAR>& aString) {
+    MOZ_RELEASE_ASSERT(
+        aString.Length() < std::numeric_limits<Length>::max() / 2,
+        "Double the string length doesn't fit in Length type");
+    const Length stringLength = static_cast<Length>(aString.Length());
+    if (aString.IsLiteral()) {
+      // Literal -> Length shifted left and LSB=0, then pointer.
+      return ULEB128Size(stringLength << 1 | 0u) +
+             static_cast<ProfileChunkedBuffer::Length>(sizeof(const CHAR*));
+    }
+    // Non-literal -> Length shifted left and LSB=1, then string size in bytes.
+    return ULEB128Size((stringLength << 1) | 1u) + stringLength * sizeof(CHAR);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const ProfilerStringView<CHAR>& aString) {
+    MOZ_RELEASE_ASSERT(
+        aString.Length() < std::numeric_limits<Length>::max() / 2,
+        "Double the string length doesn't fit in Length type");
+    const Length stringLength = static_cast<Length>(aString.Length());
+    if (aString.IsLiteral()) {
+      // Literal -> Length shifted left and LSB=0, then pointer.
+      aEW.WriteULEB128(stringLength << 1 | 0u);
+      aEW.WriteObject(WrapProfileBufferRawPointer(aString.Data()));
+      return;
+    }
+    // Non-literal -> Length shifted left and LSB=1, then string size in bytes.
+    aEW.WriteULEB128(stringLength << 1 | 1u);
+    aEW.WriteBytes(aString.Data(), stringLength * sizeof(CHAR));
+  }
+};
+
+template <typename CHAR>
+struct ProfileBufferEntryReader::Deserializer<ProfilerStringView<CHAR>> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       ProfilerStringView<CHAR>& aString) {
+    const Length lengthAndIsLiteral = aER.ReadULEB128<Length>();
+    const Length stringLength = lengthAndIsLiteral >> 1;
+    if ((lengthAndIsLiteral & 1u) == 0u) {
+      // LSB==0 -> Literal string, read the string pointer.
+      aString.mStringView = std::basic_string_view<CHAR>(
+          aER.ReadObject<const CHAR*>(), stringLength);
+      aString.mOwnership = ProfilerStringView<CHAR>::Ownership::Literal;
+      return;
+    }
+    // LSB==1 -> Not a literal string, allocate a buffer to store the string
+    // (plus terminal, for safety), and give it to the ProfilerStringView; Note
+    // that this is a secret use of ProfilerStringView, which is intended to
+    // only be used between deserialization and JSON streaming.
+    CHAR* buffer = new CHAR[stringLength + 1];
+    aER.ReadBytes(buffer, stringLength * sizeof(CHAR));
+    buffer[stringLength] = CHAR(0);
+    aString.mStringView = std::basic_string_view<CHAR>(buffer, stringLength);
+    aString.mOwnership =
+        ProfilerStringView<CHAR>::Ownership::OwnedThroughStringView;
+  }
+
+  static ProfilerStringView<CHAR> Read(ProfileBufferEntryReader& aER) {
+    const Length lengthAndIsLiteral = aER.ReadULEB128<Length>();
+    const Length stringLength = lengthAndIsLiteral >> 1;
+    if ((lengthAndIsLiteral & 1u) == 0u) {
+      // LSB==0 -> Literal string, read the string pointer.
+      return ProfilerStringView<CHAR>(
+          aER.ReadObject<const CHAR*>(), stringLength,
+          ProfilerStringView<CHAR>::Ownership::Literal);
+    }
+    // LSB==1 -> Not a literal string, allocate a buffer to store the string
+    // (plus terminal, for safety), and give it to the ProfilerStringView; Note
+    // that this is a secret use of ProfilerStringView, which is intended to
+    // only be used between deserialization and JSON streaming.
+    CHAR* buffer = new CHAR[stringLength + 1];
+    aER.ReadBytes(buffer, stringLength * sizeof(CHAR));
+    buffer[stringLength] = CHAR(0);
+    return ProfilerStringView<CHAR>(
+        buffer, stringLength,
+        ProfilerStringView<CHAR>::Ownership::OwnedThroughStringView);
+  }
+};
+
+// Serializer, Deserializer: MarkerCategory
+
+// The serialization contains both category numbers encoded as ULEB128.
+template <>
+struct ProfileBufferEntryWriter::Serializer<MarkerCategory> {
+  static Length Bytes(const MarkerCategory& aCategory) {
+    return ULEB128Size(static_cast<uint32_t>(aCategory.CategoryPair()));
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const MarkerCategory& aCategory) {
+    aEW.WriteULEB128(static_cast<uint32_t>(aCategory.CategoryPair()));
+  }
+};
+
+template <>
+struct ProfileBufferEntryReader::Deserializer<MarkerCategory> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       MarkerCategory& aCategory) {
+    aCategory = Read(aER);
+  }
+
+  static MarkerCategory Read(ProfileBufferEntryReader& aER) {
+    return MarkerCategory(static_cast<baseprofiler::ProfilingCategoryPair>(
+        aER.ReadULEB128<uint32_t>()));
+  }
+};
+
+// ----------------------------------------------------------------------------
+// Serializer, Deserializer: MarkerTiming
+
+// The serialization starts with the marker phase, followed by one or two
+// timestamps as needed.
+template <>
+struct ProfileBufferEntryWriter::Serializer<MarkerTiming> {
+  static Length Bytes(const MarkerTiming& aTiming) {
+    MOZ_ASSERT(!aTiming.IsUnspecified());
+    const auto phase = aTiming.MarkerPhase();
+    switch (phase) {
+      case MarkerTiming::Phase::Instant:
+        return SumBytes(phase, aTiming.StartTime());
+      case MarkerTiming::Phase::Interval:
+        return SumBytes(phase, aTiming.StartTime(), aTiming.EndTime());
+      case MarkerTiming::Phase::IntervalStart:
+        return SumBytes(phase, aTiming.StartTime());
+      case MarkerTiming::Phase::IntervalEnd:
+        return SumBytes(phase, aTiming.EndTime());
+      default:
+        MOZ_RELEASE_ASSERT(phase == MarkerTiming::Phase::Instant ||
+                           phase == MarkerTiming::Phase::Interval ||
+                           phase == MarkerTiming::Phase::IntervalStart ||
+                           phase == MarkerTiming::Phase::IntervalEnd);
+        return 0;  // Only to avoid build errors.
+    }
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const MarkerTiming& aTiming) {
+    MOZ_ASSERT(!aTiming.IsUnspecified());
+    const auto phase = aTiming.MarkerPhase();
+    switch (phase) {
+      case MarkerTiming::Phase::Instant:
+        aEW.WriteObjects(phase, aTiming.StartTime());
+        return;
+      case MarkerTiming::Phase::Interval:
+        aEW.WriteObjects(phase, aTiming.StartTime(), aTiming.EndTime());
+        return;
+      case MarkerTiming::Phase::IntervalStart:
+        aEW.WriteObjects(phase, aTiming.StartTime());
+        return;
+      case MarkerTiming::Phase::IntervalEnd:
+        aEW.WriteObjects(phase, aTiming.EndTime());
+        return;
+      default:
+        MOZ_RELEASE_ASSERT(phase == MarkerTiming::Phase::Instant ||
+                           phase == MarkerTiming::Phase::Interval ||
+                           phase == MarkerTiming::Phase::IntervalStart ||
+                           phase == MarkerTiming::Phase::IntervalEnd);
+        return;
+    }
+  }
+};
+
+template <>
+struct ProfileBufferEntryReader::Deserializer<MarkerTiming> {
+  static void ReadInto(ProfileBufferEntryReader& aER, MarkerTiming& aTiming) {
+    aTiming.mPhase = aER.ReadObject<MarkerTiming::Phase>();
+    switch (aTiming.mPhase) {
+      case MarkerTiming::Phase::Instant:
+        aTiming.mStartTime = aER.ReadObject<TimeStamp>();
+        aTiming.mEndTime = TimeStamp{};
+        break;
+      case MarkerTiming::Phase::Interval:
+        aTiming.mStartTime = aER.ReadObject<TimeStamp>();
+        aTiming.mEndTime = aER.ReadObject<TimeStamp>();
+        break;
+      case MarkerTiming::Phase::IntervalStart:
+        aTiming.mStartTime = aER.ReadObject<TimeStamp>();
+        aTiming.mEndTime = TimeStamp{};
+        break;
+      case MarkerTiming::Phase::IntervalEnd:
+        aTiming.mStartTime = TimeStamp{};
+        aTiming.mEndTime = aER.ReadObject<TimeStamp>();
+        break;
+      default:
+        MOZ_RELEASE_ASSERT(aTiming.mPhase == MarkerTiming::Phase::Instant ||
+                           aTiming.mPhase == MarkerTiming::Phase::Interval ||
+                           aTiming.mPhase ==
+                               MarkerTiming::Phase::IntervalStart ||
+                           aTiming.mPhase == MarkerTiming::Phase::IntervalEnd);
+        break;
+    }
+  }
+
+  static MarkerTiming Read(ProfileBufferEntryReader& aER) {
+    TimeStamp start;
+    TimeStamp end;
+    auto phase = aER.ReadObject<MarkerTiming::Phase>();
+    switch (phase) {
+      case MarkerTiming::Phase::Instant:
+        start = aER.ReadObject<TimeStamp>();
+        break;
+      case MarkerTiming::Phase::Interval:
+        start = aER.ReadObject<TimeStamp>();
+        end = aER.ReadObject<TimeStamp>();
+        break;
+      case MarkerTiming::Phase::IntervalStart:
+        start = aER.ReadObject<TimeStamp>();
+        break;
+      case MarkerTiming::Phase::IntervalEnd:
+        end = aER.ReadObject<TimeStamp>();
+        break;
+      default:
+        MOZ_RELEASE_ASSERT(phase == MarkerTiming::Phase::Instant ||
+                           phase == MarkerTiming::Phase::Interval ||
+                           phase == MarkerTiming::Phase::IntervalStart ||
+                           phase == MarkerTiming::Phase::IntervalEnd);
+        break;
+    }
+    return MarkerTiming(start, end, phase);
+  }
+};
+
+// ----------------------------------------------------------------------------
+// Serializer, Deserializer: MarkerStack
+
+// The serialization only contains the `ProfileChunkedBuffer` from the
+// backtrace; if there is no backtrace or if it's empty, this will implicitly
+// store a nullptr (see
+// `ProfileBufferEntryWriter::Serializer<ProfilerChunkedBuffer*>`).
+template <>
+struct ProfileBufferEntryWriter::Serializer<MarkerStack> {
+  static Length Bytes(const MarkerStack& aStack) {
+    return SumBytes(aStack.GetChunkedBuffer());
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW, const MarkerStack& aStack) {
+    aEW.WriteObject(aStack.GetChunkedBuffer());
+  }
+};
+
+template <>
+struct ProfileBufferEntryReader::Deserializer<MarkerStack> {
+  static void ReadInto(ProfileBufferEntryReader& aER, MarkerStack& aStack) {
+    aStack = Read(aER);
+  }
+
+  static MarkerStack Read(ProfileBufferEntryReader& aER) {
+    return MarkerStack(aER.ReadObject<UniquePtr<ProfileChunkedBuffer>>());
+  }
+};
+
+// ----------------------------------------------------------------------------
+// Serializer, Deserializer: MarkerOptions
+
+// The serialization contains all members (either trivially-copyable, or they
+// provide their specialization above).
+template <>
+struct ProfileBufferEntryWriter::Serializer<MarkerOptions> {
+  static Length Bytes(const MarkerOptions& aOptions) {
+    return SumBytes(aOptions.ThreadId(), aOptions.Timing(), aOptions.Stack(),
+                    aOptions.InnerWindowId());
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const MarkerOptions& aOptions) {
+    aEW.WriteObjects(aOptions.ThreadId(), aOptions.Timing(), aOptions.Stack(),
+                     aOptions.InnerWindowId());
+  }
+};
+
+template <>
+struct ProfileBufferEntryReader::Deserializer<MarkerOptions> {
+  static void ReadInto(ProfileBufferEntryReader& aER, MarkerOptions& aOptions) {
+    aER.ReadIntoObjects(aOptions.mThreadId, aOptions.mTiming, aOptions.mStack,
+                        aOptions.mInnerWindowId);
+  }
+
+  static MarkerOptions Read(ProfileBufferEntryReader& aER) {
+    MarkerOptions options;
+    ReadInto(aER, options);
+    return options;
+  }
+};
+
+}  // namespace mozilla
+
+#endif  // MOZ_GECKO_PROFILER
+
+#endif  // BaseProfilerMarkersDetail_h
diff --git a/mozglue/baseprofiler/public/BaseProfilerMarkersPrerequisites.h b/mozglue/baseprofiler/public/BaseProfilerMarkersPrerequisites.h
new file mode 100644
index 0000000000..aa85b41896
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilerMarkersPrerequisites.h
@@ -0,0 +1,866 @@
+/* -*- Mode: C++; tab-width: 2; 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 header contains basic definitions required to create marker types, and
+// to add markers to the profiler buffers.
+//
+// In most cases, #include "mozilla/BaseProfilerMarkers.h" instead, or
+// #include "mozilla/BaseProfilerMarkerTypes.h" for common marker types.
+
+#ifndef BaseProfilerMarkersPrerequisites_h
+#define BaseProfilerMarkersPrerequisites_h
+
+#ifdef MOZ_GECKO_PROFILER
+
+#  include "BaseProfilingCategory.h"
+#  include "mozilla/Maybe.h"
+#  include "mozilla/ProfileChunkedBuffer.h"
+#  include "mozilla/TimeStamp.h"
+#  include "mozilla/UniquePtr.h"
+#  include "mozilla/Variant.h"
+
+#  include <initializer_list>
+#  include <string_view>
+#  include <string>
+#  include <type_traits>
+#  include <utility>
+#  include <vector>
+
+// TODO: Move common stuff to shared header instead.
+#  include "BaseProfiler.h"
+
+namespace mozilla {
+
+// Return a NotNull<const CHAR*> pointing at the literal empty string `""`.
+template <typename CHAR>
+constexpr const CHAR* LiteralEmptyStringPointer() {
+  static_assert(std::is_same_v<CHAR, char> || std::is_same_v<CHAR, char16_t>,
+                "Only char and char16_t are supported in Firefox");
+  if constexpr (std::is_same_v<CHAR, char>) {
+    return "";
+  }
+  if constexpr (std::is_same_v<CHAR, char16_t>) {
+    return u"";
+  }
+}
+
+// Return a string_view<CHAR> pointing at the literal empty string.
+template <typename CHAR>
+constexpr std::basic_string_view<CHAR> LiteralEmptyStringView() {
+  static_assert(std::is_same_v<CHAR, char> || std::is_same_v<CHAR, char16_t>,
+                "Only char and char16_t are supported in Firefox");
+  // Use `operator""sv()` from <string_view>.
+  using namespace std::literals::string_view_literals;
+  if constexpr (std::is_same_v<CHAR, char>) {
+    return ""sv;
+  }
+  if constexpr (std::is_same_v<CHAR, char16_t>) {
+    return u""sv;
+  }
+}
+
+// General string view, optimized for short on-stack life before serialization,
+// and between deserialization and JSON-streaming.
+template <typename CHAR>
+class MOZ_STACK_CLASS ProfilerStringView {
+ public:
+  // Default constructor points at "" (literal empty string).
+  constexpr ProfilerStringView() = default;
+
+  // Don't allow copy.
+  ProfilerStringView(const ProfilerStringView&) = delete;
+  ProfilerStringView& operator=(const ProfilerStringView&) = delete;
+
+  // Allow move. For consistency the moved-from string is always reset to "".
+  constexpr ProfilerStringView(ProfilerStringView&& aOther)
+      : mStringView(std::move(aOther.mStringView)),
+        mOwnership(aOther.mOwnership) {
+    if (mOwnership == Ownership::OwnedThroughStringView) {
+      // We now own the buffer, make the other point at the literal "".
+      aOther.mStringView = LiteralEmptyStringView<CHAR>();
+      aOther.mOwnership = Ownership::Literal;
+    }
+  }
+  constexpr ProfilerStringView& operator=(ProfilerStringView&& aOther) {
+    mStringView = std::move(aOther.mStringView);
+    mOwnership = aOther.mOwnership;
+    if (mOwnership == Ownership::OwnedThroughStringView) {
+      // We now own the buffer, make the other point at the literal "".
+      aOther.mStringView = LiteralEmptyStringView<CHAR>();
+      aOther.mOwnership = Ownership::Literal;
+    }
+    return *this;
+  }
+
+  ~ProfilerStringView() {
+    if (MOZ_UNLIKELY(mOwnership == Ownership::OwnedThroughStringView)) {
+      // We own the buffer pointed at by mStringView, destroy it.
+      // This is only used between deserialization and streaming.
+      delete mStringView.data();
+    }
+  }
+
+  // Implicit construction from nullptr, points at "" (literal empty string).
+  constexpr MOZ_IMPLICIT ProfilerStringView(decltype(nullptr)) {}
+
+  // Implicit constructor from a literal string.
+  template <size_t Np1>
+  constexpr MOZ_IMPLICIT ProfilerStringView(const CHAR (&aLiteralString)[Np1])
+      : ProfilerStringView(aLiteralString, Np1 - 1, Ownership::Literal) {}
+
+  // Constructor from a non-literal string.
+  constexpr ProfilerStringView(const CHAR* aString, size_t aLength)
+      : ProfilerStringView(aString, aLength, Ownership::Reference) {}
+
+  // Implicit constructor from a string_view.
+  constexpr MOZ_IMPLICIT ProfilerStringView(
+      const std::basic_string_view<CHAR>& aStringView)
+      : ProfilerStringView(aStringView.data(), aStringView.length(),
+                           Ownership::Reference) {}
+
+  // Implicit constructor from an expiring string_view. We assume that the
+  // pointed-at string will outlive this ProfilerStringView.
+  constexpr MOZ_IMPLICIT ProfilerStringView(
+      std::basic_string_view<CHAR>&& aStringView)
+      : ProfilerStringView(aStringView.data(), aStringView.length(),
+                           Ownership::Reference) {}
+
+  // Implicit constructor from std::string.
+  constexpr MOZ_IMPLICIT ProfilerStringView(
+      const std::basic_string<CHAR>& aString)
+      : ProfilerStringView(aString.data(), aString.length(),
+                           Ownership::Reference) {}
+
+  // Construction from a raw pointer to a null-terminated string.
+  // This is a named class-static function to make it more obvious where work is
+  // being done (to determine the string length), and encourage users to instead
+  // provide a length, if already known.
+  // TODO: Find callers and convert them to constructor instead if possible.
+  static constexpr ProfilerStringView WrapNullTerminatedString(
+      const CHAR* aString) {
+    return ProfilerStringView(
+        aString, aString ? std::char_traits<CHAR>::length(aString) : 0,
+        Ownership::Reference);
+  }
+
+  // Implicit constructor for an object with member functions `Data()`
+  // `Length()`, and `IsLiteral()`, common in xpcom strings.
+  template <
+      typename String,
+      typename DataReturnType = decltype(std::declval<const String>().Data()),
+      typename LengthReturnType =
+          decltype(std::declval<const String>().Length()),
+      typename IsLiteralReturnType =
+          decltype(std::declval<const String>().IsLiteral()),
+      typename =
+          std::enable_if_t<std::is_convertible_v<DataReturnType, const CHAR*> &&
+                           std::is_integral_v<LengthReturnType> &&
+                           std::is_same_v<IsLiteralReturnType, bool>>>
+  constexpr MOZ_IMPLICIT ProfilerStringView(const String& aString)
+      : ProfilerStringView(
+            static_cast<const CHAR*>(aString.Data()), aString.Length(),
+            aString.IsLiteral() ? Ownership::Literal : Ownership::Reference) {}
+
+  [[nodiscard]] constexpr const std::basic_string_view<CHAR>& StringView()
+      const {
+    return mStringView;
+  }
+
+  [[nodiscard]] constexpr const CHAR* Data() const {
+    return mStringView.data();
+  }
+
+  [[nodiscard]] constexpr size_t Length() const { return mStringView.length(); }
+
+  [[nodiscard]] constexpr bool IsLiteral() const {
+    return mOwnership == Ownership::Literal;
+  }
+  [[nodiscard]] constexpr bool IsReference() const {
+    return mOwnership == Ownership::Reference;
+  }
+  // No `IsOwned...()` because it's a secret, only used internally!
+
+  [[nodiscard]] operator Span<const CHAR>() const {
+    return Span<const CHAR>(Data(), Length());
+  }
+
+ private:
+  enum class Ownership { Literal, Reference, OwnedThroughStringView };
+
+  // Allow deserializer to store anything here.
+  friend ProfileBufferEntryReader::Deserializer<ProfilerStringView>;
+
+  constexpr ProfilerStringView(const CHAR* aString, size_t aLength,
+                               Ownership aOwnership)
+      : mStringView(aString ? std::basic_string_view<CHAR>(aString, aLength)
+                            : LiteralEmptyStringView<CHAR>()),
+        mOwnership(aString ? aOwnership : Ownership::Literal) {}
+
+  // String view to an outside string (literal or reference).
+  // We may actually own the pointed-at buffer, but it is only used internally
+  // between deserialization and JSON streaming.
+  std::basic_string_view<CHAR> mStringView = LiteralEmptyStringView<CHAR>();
+
+  Ownership mOwnership = Ownership::Literal;
+};
+
+using ProfilerString8View = ProfilerStringView<char>;
+using ProfilerString16View = ProfilerStringView<char16_t>;
+
+// This compulsory marker parameter contains the required category information.
+class MarkerCategory {
+ public:
+  // Constructor from category pair (includes both super- and sub-categories).
+  constexpr explicit MarkerCategory(
+      baseprofiler::ProfilingCategoryPair aCategoryPair)
+      : mCategoryPair(aCategoryPair) {}
+
+  // Returns the stored category pair.
+  constexpr baseprofiler::ProfilingCategoryPair CategoryPair() const {
+    return mCategoryPair;
+  }
+
+  // Returns the super-category from the stored category pair.
+  baseprofiler::ProfilingCategory GetCategory() const {
+    return GetProfilingCategoryPairInfo(mCategoryPair).mCategory;
+  }
+
+ private:
+  baseprofiler::ProfilingCategoryPair mCategoryPair =
+      baseprofiler::ProfilingCategoryPair::OTHER;
+};
+
+namespace baseprofiler::category {
+
+// Each category pair name constructs a MarkerCategory.
+// E.g.: mozilla::baseprofiler::category::OTHER_Profiling
+// Profiler macros will take the category name alone without namespace.
+// E.g.: `PROFILER_MARKER_UNTYPED("name", OTHER_Profiling)`
+#  define CATEGORY_ENUM_BEGIN_CATEGORY(name, labelAsString, color)
+#  define CATEGORY_ENUM_SUBCATEGORY(supercategory, name, labelAsString) \
+    static constexpr MarkerCategory name{ProfilingCategoryPair::name};
+#  define CATEGORY_ENUM_END_CATEGORY
+MOZ_PROFILING_CATEGORY_LIST(CATEGORY_ENUM_BEGIN_CATEGORY,
+                            CATEGORY_ENUM_SUBCATEGORY,
+                            CATEGORY_ENUM_END_CATEGORY)
+#  undef CATEGORY_ENUM_BEGIN_CATEGORY
+#  undef CATEGORY_ENUM_SUBCATEGORY
+#  undef CATEGORY_ENUM_END_CATEGORY
+
+// Import `MarkerCategory` into this namespace. This will allow using this type
+// dynamically in macros that prepend `::mozilla::baseprofiler::category::` to
+// the given category, e.g.:
+// `PROFILER_MARKER_UNTYPED("name", MarkerCategory(...))`
+using MarkerCategory = ::mozilla::MarkerCategory;
+
+}  // namespace baseprofiler::category
+
+// The classes below are all embedded in a `MarkerOptions` object.
+class MarkerOptions;
+
+// This marker option captures a given thread id.
+// If left unspecified (by default construction) during the add-marker call, the
+// current thread id will be used then.
+class MarkerThreadId {
+ public:
+  // Default constructor, keeps the thread id unspecified.
+  constexpr MarkerThreadId() = default;
+
+  // Constructor from a given thread id.
+  constexpr explicit MarkerThreadId(int aThreadId) : mThreadId(aThreadId) {}
+
+  // Use the current thread's id.
+  static MarkerThreadId CurrentThread() {
+    return MarkerThreadId(baseprofiler::profiler_current_thread_id());
+  }
+
+  // Use the main thread's id. This can be useful to record a marker from a
+  // possibly-unregistered thread, and display it in the main thread track.
+  static MarkerThreadId MainThread() {
+    return MarkerThreadId(baseprofiler::profiler_main_thread_id());
+  }
+
+  [[nodiscard]] constexpr int ThreadId() const { return mThreadId; }
+
+  [[nodiscard]] constexpr bool IsUnspecified() const { return mThreadId == 0; }
+
+ private:
+  int mThreadId = 0;
+};
+
+// This marker option contains marker timing information.
+// This class encapsulates the logic for correctly storing a marker based on its
+// Use the static methods to create the MarkerTiming. This is a transient object
+// that is being used to enforce the constraints of the combinations of the
+// data.
+class MarkerTiming {
+ public:
+  // The following static methods are used to create the MarkerTiming based on
+  // the type that it is.
+
+  static MarkerTiming InstantAt(const TimeStamp& aTime) {
+    MOZ_ASSERT(!aTime.IsNull(), "Time is null for an instant marker.");
+    return MarkerTiming{aTime, TimeStamp{}, MarkerTiming::Phase::Instant};
+  }
+
+  static MarkerTiming InstantNow() {
+    return InstantAt(TimeStamp::NowUnfuzzed());
+  }
+
+  static MarkerTiming Interval(const TimeStamp& aStartTime,
+                               const TimeStamp& aEndTime) {
+    MOZ_ASSERT(!aStartTime.IsNull(),
+               "Start time is null for an interval marker.");
+    MOZ_ASSERT(!aEndTime.IsNull(), "End time is null for an interval marker.");
+    return MarkerTiming{aStartTime, aEndTime, MarkerTiming::Phase::Interval};
+  }
+
+  static MarkerTiming IntervalUntilNowFrom(const TimeStamp& aStartTime) {
+    return Interval(aStartTime, TimeStamp::NowUnfuzzed());
+  }
+
+  static MarkerTiming IntervalStart(
+      const TimeStamp& aTime = TimeStamp::NowUnfuzzed()) {
+    MOZ_ASSERT(!aTime.IsNull(), "Time is null for an interval start marker.");
+    return MarkerTiming{aTime, TimeStamp{}, MarkerTiming::Phase::IntervalStart};
+  }
+
+  static MarkerTiming IntervalEnd(
+      const TimeStamp& aTime = TimeStamp::NowUnfuzzed()) {
+    MOZ_ASSERT(!aTime.IsNull(), "Time is null for an interval end marker.");
+    return MarkerTiming{TimeStamp{}, aTime, MarkerTiming::Phase::IntervalEnd};
+  }
+
+  // Set the interval end in this timing.
+  // If there was already a start time, this makes it a full interval.
+  void SetIntervalEnd(const TimeStamp& aTime = TimeStamp::NowUnfuzzed()) {
+    MOZ_ASSERT(!aTime.IsNull(), "Time is null for an interval end marker.");
+    mEndTime = aTime;
+    mPhase = mStartTime.IsNull() ? Phase::IntervalEnd : Phase::Interval;
+  }
+
+  [[nodiscard]] const TimeStamp& StartTime() const { return mStartTime; }
+  [[nodiscard]] const TimeStamp& EndTime() const { return mEndTime; }
+
+  enum class Phase : uint8_t {
+    Instant = 0,
+    Interval = 1,
+    IntervalStart = 2,
+    IntervalEnd = 3,
+  };
+
+  [[nodiscard]] Phase MarkerPhase() const {
+    MOZ_ASSERT(!IsUnspecified());
+    return mPhase;
+  }
+
+  // The following getter methods are used to put the value into the buffer for
+  // storage.
+  [[nodiscard]] double GetStartTime() const {
+    MOZ_ASSERT(!IsUnspecified());
+    // If mStartTime is null (e.g., for IntervalEnd), this will output 0.0 as
+    // expected.
+    return MarkerTiming::timeStampToDouble(mStartTime);
+  }
+
+  [[nodiscard]] double GetEndTime() const {
+    MOZ_ASSERT(!IsUnspecified());
+    // If mEndTime is null (e.g., for Instant or IntervalStart), this will
+    // output 0.0 as expected.
+    return MarkerTiming::timeStampToDouble(mEndTime);
+  }
+
+  [[nodiscard]] uint8_t GetPhase() const {
+    MOZ_ASSERT(!IsUnspecified());
+    return static_cast<uint8_t>(mPhase);
+  }
+
+ private:
+  friend ProfileBufferEntryWriter::Serializer<MarkerTiming>;
+  friend ProfileBufferEntryReader::Deserializer<MarkerTiming>;
+  friend MarkerOptions;
+
+  // Default timing leaves it internally "unspecified", serialization getters
+  // and add-marker functions will default to `InstantNow()`.
+  constexpr MarkerTiming() = default;
+
+  // This should only be used by internal profiler code.
+  [[nodiscard]] bool IsUnspecified() const {
+    return mStartTime.IsNull() && mEndTime.IsNull();
+  }
+
+  // Full constructor, used by static factory functions.
+  constexpr MarkerTiming(const TimeStamp& aStartTime, const TimeStamp& aEndTime,
+                         Phase aPhase)
+      : mStartTime(aStartTime), mEndTime(aEndTime), mPhase(aPhase) {}
+
+  static double timeStampToDouble(const TimeStamp& time) {
+    if (time.IsNull()) {
+      // The Phase lets us know not to use this value.
+      return 0;
+    }
+    return (time - TimeStamp::ProcessCreation()).ToMilliseconds();
+  }
+
+  TimeStamp mStartTime;
+  TimeStamp mEndTime;
+  Phase mPhase = Phase::Instant;
+};
+
+// This marker option allows three cases:
+// - By default, no stacks are captured.
+// - The caller can request a stack capture, and the add-marker code will take
+//   care of it in the most efficient way.
+// - The caller can still provide an existing backtrace, for cases where a
+//   marker reports something that happened elsewhere.
+class MarkerStack {
+ public:
+  // Default constructor, no capture.
+  constexpr MarkerStack() = default;
+
+  // Disallow copy.
+  MarkerStack(const MarkerStack&) = delete;
+  MarkerStack& operator=(const MarkerStack&) = delete;
+
+  // Allow move.
+  MarkerStack(MarkerStack&& aOther)
+      : mIsCaptureRequested(aOther.mIsCaptureRequested),
+        mOptionalChunkedBufferStorage(
+            std::move(aOther.mOptionalChunkedBufferStorage)),
+        mChunkedBuffer(aOther.mChunkedBuffer) {
+    AssertInvariants();
+    aOther.Clear();
+  }
+  MarkerStack& operator=(MarkerStack&& aOther) {
+    mIsCaptureRequested = aOther.mIsCaptureRequested;
+    mOptionalChunkedBufferStorage =
+        std::move(aOther.mOptionalChunkedBufferStorage);
+    mChunkedBuffer = aOther.mChunkedBuffer;
+    AssertInvariants();
+    aOther.Clear();
+    return *this;
+  }
+
+  // Take ownership of a backtrace. If null or empty, equivalent to NoStack().
+  explicit MarkerStack(UniquePtr<ProfileChunkedBuffer>&& aExternalChunkedBuffer)
+      : mIsCaptureRequested(false),
+        mOptionalChunkedBufferStorage(
+            (!aExternalChunkedBuffer || aExternalChunkedBuffer->IsEmpty())
+                ? nullptr
+                : std::move(aExternalChunkedBuffer)),
+        mChunkedBuffer(mOptionalChunkedBufferStorage.get()) {
+    AssertInvariants();
+  }
+
+  // Use an existing backtrace stored elsewhere, which the user must guarantee
+  // is alive during the add-marker call. If empty, equivalent to NoStack().
+  explicit MarkerStack(ProfileChunkedBuffer& aExternalChunkedBuffer)
+      : mIsCaptureRequested(false),
+        mChunkedBuffer(aExternalChunkedBuffer.IsEmpty()
+                           ? nullptr
+                           : &aExternalChunkedBuffer) {
+    AssertInvariants();
+  }
+
+  // Don't capture a stack in this marker.
+  static MarkerStack NoStack() { return MarkerStack(false); }
+
+  // Capture a stack when adding this marker.
+  static MarkerStack Capture() {
+    // Actual capture will be handled inside profiler_add_marker.
+    return MarkerStack(true);
+  }
+
+  // Optionally capture a stack, useful for avoiding long-winded ternaries.
+  static MarkerStack MaybeCapture(bool aDoCapture) {
+    return MarkerStack(aDoCapture);
+  }
+
+  // Use an existing backtrace stored elsewhere, which the user must guarantee
+  // is alive during the add-marker call. If empty, equivalent to NoStack().
+  static MarkerStack UseBacktrace(
+      ProfileChunkedBuffer& aExternalChunkedBuffer) {
+    return MarkerStack(aExternalChunkedBuffer);
+  }
+
+  // Take ownership of a backtrace previously captured with
+  // `profiler_capture_backtrace()`. If null, equivalent to NoStack().
+  static MarkerStack TakeBacktrace(
+      UniquePtr<ProfileChunkedBuffer>&& aExternalChunkedBuffer) {
+    return MarkerStack(std::move(aExternalChunkedBuffer));
+  }
+
+  [[nodiscard]] bool IsCaptureNeeded() const {
+    // If the chunked buffer already contains something, consider the capture
+    // request already fulfilled.
+    return mIsCaptureRequested;
+  }
+
+  ProfileChunkedBuffer* GetChunkedBuffer() const { return mChunkedBuffer; }
+
+  // Use backtrace after a request. If null, equivalent to NoStack().
+  void UseRequestedBacktrace(ProfileChunkedBuffer* aExternalChunkedBuffer) {
+    MOZ_RELEASE_ASSERT(IsCaptureNeeded());
+    mIsCaptureRequested = false;
+    if (aExternalChunkedBuffer && !aExternalChunkedBuffer->IsEmpty()) {
+      // We only need to use the provided buffer if it is not empty.
+      mChunkedBuffer = aExternalChunkedBuffer;
+    }
+    AssertInvariants();
+  }
+
+  void Clear() {
+    mIsCaptureRequested = false;
+    mOptionalChunkedBufferStorage.reset();
+    mChunkedBuffer = nullptr;
+    AssertInvariants();
+  }
+
+ private:
+  explicit MarkerStack(bool aIsCaptureRequested)
+      : mIsCaptureRequested(aIsCaptureRequested) {
+    AssertInvariants();
+  }
+
+  // This should be called after every constructor and non-const function.
+  void AssertInvariants() const {
+#  ifdef DEBUG
+    if (mIsCaptureRequested) {
+      MOZ_ASSERT(!mOptionalChunkedBufferStorage,
+                 "We should not hold a buffer when capture is requested");
+      MOZ_ASSERT(!mChunkedBuffer,
+                 "We should not point at a buffer when capture is requested");
+    } else {
+      if (mOptionalChunkedBufferStorage) {
+        MOZ_ASSERT(mChunkedBuffer == mOptionalChunkedBufferStorage.get(),
+                   "Non-null mOptionalChunkedBufferStorage must be pointed-at "
+                   "by mChunkedBuffer");
+      }
+      if (mChunkedBuffer) {
+        MOZ_ASSERT(!mChunkedBuffer->IsEmpty(),
+                   "Non-null mChunkedBuffer must not be empty");
+      }
+    }
+#  endif  // DEBUG
+  }
+
+  // True if a capture is requested when marker is added to the profile buffer.
+  bool mIsCaptureRequested = false;
+
+  // Optional storage for the backtrace, in case it was captured before the
+  // add-marker call.
+  UniquePtr<ProfileChunkedBuffer> mOptionalChunkedBufferStorage;
+
+  // If not null, this points to the backtrace. It may point to a backtrace
+  // temporarily stored on the stack, or to mOptionalChunkedBufferStorage.
+  ProfileChunkedBuffer* mChunkedBuffer = nullptr;
+};
+
+// This marker option captures a given inner window id.
+class MarkerInnerWindowId {
+ public:
+  // Default constructor, it leaves the id unspecified.
+  constexpr MarkerInnerWindowId() = default;
+
+  // Constructor with a specified inner window id.
+  constexpr explicit MarkerInnerWindowId(uint64_t i) : mInnerWindowId(i) {}
+
+  // Constructor with either specified inner window id or Nothing.
+  constexpr explicit MarkerInnerWindowId(const Maybe<uint64_t>& i)
+      : mInnerWindowId(i.valueOr(scNoId)) {}
+
+  // Explicit option with unspecified id.
+  constexpr static MarkerInnerWindowId NoId() { return MarkerInnerWindowId{}; }
+
+  [[nodiscard]] bool IsUnspecified() const { return mInnerWindowId == scNoId; }
+
+  [[nodiscard]] constexpr uint64_t Id() const { return mInnerWindowId; }
+
+ private:
+  static constexpr uint64_t scNoId = 0;
+  uint64_t mInnerWindowId = scNoId;
+};
+
+// This class combines each of the possible marker options above.
+class MarkerOptions {
+ public:
+  // Constructor from individual options (including none).
+  // Implicit to allow `{}` and one option type as-is.
+  // Options that are not provided here are defaulted. In particular, timing
+  // defaults to `MarkerTiming::InstantNow()` when the marker is recorded.
+  template <typename... Options>
+  MOZ_IMPLICIT MarkerOptions(Options&&... aOptions) {
+    (Set(std::forward<Options>(aOptions)), ...);
+  }
+
+  // Disallow copy.
+  MarkerOptions(const MarkerOptions&) = delete;
+  MarkerOptions& operator=(const MarkerOptions&) = delete;
+
+  // Allow move.
+  MarkerOptions(MarkerOptions&&) = default;
+  MarkerOptions& operator=(MarkerOptions&&) = default;
+
+  // The embedded `MarkerTiming` hasn't been specified yet.
+  [[nodiscard]] bool IsTimingUnspecified() const {
+    return mTiming.IsUnspecified();
+  }
+
+  // Each option may be added in a chain by e.g.:
+  // `options.Set(MarkerThreadId(123)).Set(MarkerTiming::IntervalEnd())`.
+  // When passed to an add-marker function, it must be an rvalue, either created
+  // on the spot, or `std::move`d from storage, e.g.:
+  // `PROFILER_MARKER_UNTYPED("...", std::move(options).Set(...))`;
+  //
+  // Options can be read by their name (without "Marker"), e.g.: `o.ThreadId()`.
+  // Add "Ref" for a non-const reference, e.g.: `o.ThreadIdRef() = ...;`
+#  define FUNCTIONS_ON_MEMBER(NAME)                      \
+    MarkerOptions& Set(Marker##NAME&& a##NAME)& {        \
+      m##NAME = std::move(a##NAME);                      \
+      return *this;                                      \
+    }                                                    \
+                                                         \
+    MarkerOptions&& Set(Marker##NAME&& a##NAME)&& {      \
+      m##NAME = std::move(a##NAME);                      \
+      return std::move(*this);                           \
+    }                                                    \
+                                                         \
+    const Marker##NAME& NAME() const { return m##NAME; } \
+                                                         \
+    Marker##NAME& NAME##Ref() { return m##NAME; }
+
+  FUNCTIONS_ON_MEMBER(ThreadId);
+  FUNCTIONS_ON_MEMBER(Timing);
+  FUNCTIONS_ON_MEMBER(Stack);
+  FUNCTIONS_ON_MEMBER(InnerWindowId);
+#  undef FUNCTIONS_ON_MEMBER
+
+ private:
+  friend ProfileBufferEntryReader::Deserializer<MarkerOptions>;
+
+  MarkerThreadId mThreadId;
+  MarkerTiming mTiming;
+  MarkerStack mStack;
+  MarkerInnerWindowId mInnerWindowId;
+};
+
+}  // namespace mozilla
+
+namespace mozilla::baseprofiler::markers {
+
+// Default marker payload types, with no extra information, not even a marker
+// type and payload. This is intended for label-only markers.
+struct NoPayload final {};
+
+}  // namespace mozilla::baseprofiler::markers
+
+namespace mozilla {
+
+class JSONWriter;
+
+// This class collects all the information necessary to stream the JSON schema
+// that informs the front-end how to display a type of markers.
+// It will be created and populated in `MarkerTypeDisplay()` functions in each
+// marker type definition, see Add/Set functions.
+class MarkerSchema {
+ public:
+  enum class Location : unsigned {
+    markerChart,
+    markerTable,
+    // This adds markers to the main marker timeline in the header.
+    timelineOverview,
+    // In the timeline, this is a section that breaks out markers that are
+    // related to memory. When memory counters are enabled, this is its own
+    // track, otherwise it is displayed with the main thread.
+    timelineMemory,
+    // This adds markers to the IPC timeline area in the header.
+    timelineIPC,
+    // This adds markers to the FileIO timeline area in the header.
+    timelineFileIO,
+    // TODO - This is not supported yet.
+    stackChart
+  };
+
+  // Used as constructor parameter, to explicitly specify that the location (and
+  // other display options) are handled as a special case in the front-end.
+  // In this case, *no* schema will be output for this type.
+  struct SpecialFrontendLocation {};
+
+  enum class Format {
+    // ----------------------------------------------------
+    // String types.
+
+    // Show the URL, and handle PII sanitization
+    url,
+    // Show the file path, and handle PII sanitization.
+    filePath,
+    // Important, do not put URL or file path information here, as it will not
+    // be sanitized. Please be careful with including other types of PII here as
+    // well.
+    // e.g. "Label: Some String"
+    string,
+
+    // ----------------------------------------------------
+    // Numeric types
+
+    // For time data that represents a duration of time.
+    // e.g. "Label: 5s, 5ms, 5μs"
+    duration,
+    // Data that happened at a specific time, relative to the start of the
+    // profile. e.g. "Label: 15.5s, 20.5ms, 30.5μs"
+    time,
+    // The following are alternatives to display a time only in a specific unit
+    // of time.
+    seconds,       // "Label: 5s"
+    milliseconds,  // "Label: 5ms"
+    microseconds,  // "Label: 5μs"
+    nanoseconds,   // "Label: 5ns"
+    // e.g. "Label: 5.55mb, 5 bytes, 312.5kb"
+    bytes,
+    // This should be a value between 0 and 1.
+    // "Label: 50%"
+    percentage,
+    // The integer should be used for generic representations of numbers.
+    // Do not use it for time information.
+    // "Label: 52, 5,323, 1,234,567"
+    integer,
+    // The decimal should be used for generic representations of numbers.
+    // Do not use it for time information.
+    // "Label: 52.23, 0.0054, 123,456.78"
+    decimal
+  };
+
+  enum class Searchable { notSearchable, searchable };
+
+  // Marker schema, with a non-empty list of locations where markers should be
+  // shown.
+  // Tech note: Even though `aLocations` are templated arguments, they are
+  // assigned to an `enum class` object, so they can only be of that enum type.
+  template <typename... Locations>
+  explicit MarkerSchema(Location aLocation, Locations... aLocations)
+      : mLocations{aLocation, aLocations...} {}
+
+  // Marker schema for types that have special frontend handling.
+  // Nothing else should be set in this case.
+  // Implicit to allow quick return from MarkerTypeDisplay functions.
+  MOZ_IMPLICIT MarkerSchema(SpecialFrontendLocation) {}
+
+  // Caller must specify location(s) or SpecialFrontendLocation above.
+  MarkerSchema() = delete;
+
+  // Optional labels in the marker chart, the chart tooltip, and the marker
+  // table. If not provided, the marker "name" will be used. The given string
+  // can contain element keys in braces to include data elements streamed by
+  // `StreamJSONMarkerData()`. E.g.: "This is {text}"
+
+#  define LABEL_SETTER(name)                       \
+    MarkerSchema& Set##name(std::string a##name) { \
+      m##name = std::move(a##name);                \
+      return *this;                                \
+    }
+
+  LABEL_SETTER(ChartLabel)
+  LABEL_SETTER(TooltipLabel)
+  LABEL_SETTER(TableLabel)
+
+#  undef LABEL_SETTER
+
+  MarkerSchema& SetAllLabels(std::string aText) {
+    // Here we set the same text in each label.
+    // TODO: Move to a single "label" field once the front-end allows it.
+    SetChartLabel(aText);
+    SetTooltipLabel(aText);
+    SetTableLabel(std::move(aText));
+    return *this;
+  }
+
+  // Each data element that is streamed by `StreamJSONMarkerData()` can be
+  // displayed as indicated by using one of the `Add...` function below.
+  // Each `Add...` will add a line in the full marker description. Parameters:
+  // - `aKey`: Element property name as streamed by `StreamJSONMarkerData()`.
+  // - `aLabel`: Optional prefix. Defaults to the key name.
+  // - `aFormat`: How to format the data element value, see `Format` above.
+  // - `aSearchable`: Optional, indicates if the value is used in searches,
+  //   defaults to false.
+
+  MarkerSchema& AddKeyFormat(std::string aKey, Format aFormat) {
+    mData.emplace_back(mozilla::VariantType<DynamicData>{},
+                       DynamicData{std::move(aKey), mozilla::Nothing{}, aFormat,
+                                   mozilla::Nothing{}});
+    return *this;
+  }
+
+  MarkerSchema& AddKeyLabelFormat(std::string aKey, std::string aLabel,
+                                  Format aFormat) {
+    mData.emplace_back(
+        mozilla::VariantType<DynamicData>{},
+        DynamicData{std::move(aKey), mozilla::Some(std::move(aLabel)), aFormat,
+                    mozilla::Nothing{}});
+    return *this;
+  }
+
+  MarkerSchema& AddKeyFormatSearchable(std::string aKey, Format aFormat,
+                                       Searchable aSearchable) {
+    mData.emplace_back(mozilla::VariantType<DynamicData>{},
+                       DynamicData{std::move(aKey), mozilla::Nothing{}, aFormat,
+                                   mozilla::Some(aSearchable)});
+    return *this;
+  }
+
+  MarkerSchema& AddKeyLabelFormatSearchable(std::string aKey,
+                                            std::string aLabel, Format aFormat,
+                                            Searchable aSearchable) {
+    mData.emplace_back(
+        mozilla::VariantType<DynamicData>{},
+        DynamicData{std::move(aKey), mozilla::Some(std::move(aLabel)), aFormat,
+                    mozilla::Some(aSearchable)});
+    return *this;
+  }
+
+  // The display may also include static rows.
+
+  MarkerSchema& AddStaticLabelValue(std::string aLabel, std::string aValue) {
+    mData.emplace_back(mozilla::VariantType<StaticData>{},
+                       StaticData{std::move(aLabel), std::move(aValue)});
+    return *this;
+  }
+
+  // Internal streaming function.
+  MFBT_API void Stream(JSONWriter& aWriter, const Span<const char>& aName) &&;
+
+ private:
+  MFBT_API static Span<const char> LocationToStringSpan(Location aLocation);
+  MFBT_API static Span<const char> FormatToStringSpan(Format aFormat);
+
+  // List of marker display locations. Empty for SpecialFrontendLocation.
+  std::vector<Location> mLocations;
+  // Labels for different places.
+  std::string mChartLabel;
+  std::string mTooltipLabel;
+  std::string mTableLabel;
+  // Main display, made of zero or more rows of key+label+format or label+value.
+ private:
+  struct DynamicData {
+    std::string mKey;
+    mozilla::Maybe<std::string> mLabel;
+    Format mFormat;
+    mozilla::Maybe<Searchable> mSearchable;
+  };
+  struct StaticData {
+    std::string mLabel;
+    std::string mValue;
+  };
+  using DataRow = mozilla::Variant<DynamicData, StaticData>;
+  using DataRowVector = std::vector<DataRow>;
+
+  DataRowVector mData;
+};
+
+}  // namespace mozilla
+
+#endif  // MOZ_GECKO_PROFILER
+
+#endif  // BaseProfilerMarkersPrerequisites_h
diff --git a/mozglue/baseprofiler/public/BaseProfilerSharedLibraries.h b/mozglue/baseprofiler/public/BaseProfilerSharedLibraries.h
new file mode 100644
index 0000000000..0a104193c3
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilerSharedLibraries.h
@@ -0,0 +1,146 @@
+/* -*- Mode: C++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim:set ts=2 sw=2 sts=2 et cindent: */
+/* 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 BASE_PROFILER_SHARED_LIBRARIES_H_
+#define BASE_PROFILER_SHARED_LIBRARIES_H_
+
+#include "BaseProfiler.h"
+
+#ifndef MOZ_GECKO_PROFILER
+#  error Do not #include this header when MOZ_GECKO_PROFILER is not #defined.
+#endif
+
+#include <algorithm>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string>
+#include <vector>
+
+class SharedLibrary {
+ public:
+  SharedLibrary(uintptr_t aStart, uintptr_t aEnd, uintptr_t aOffset,
+                const std::string& aBreakpadId, const std::string& aModuleName,
+                const std::string& aModulePath, const std::string& aDebugName,
+                const std::string& aDebugPath, const std::string& aVersion,
+                const char* aArch)
+      : mStart(aStart),
+        mEnd(aEnd),
+        mOffset(aOffset),
+        mBreakpadId(aBreakpadId),
+        mModuleName(aModuleName),
+        mModulePath(aModulePath),
+        mDebugName(aDebugName),
+        mDebugPath(aDebugPath),
+        mVersion(aVersion),
+        mArch(aArch) {}
+
+  SharedLibrary(const SharedLibrary& aEntry)
+      : mStart(aEntry.mStart),
+        mEnd(aEntry.mEnd),
+        mOffset(aEntry.mOffset),
+        mBreakpadId(aEntry.mBreakpadId),
+        mModuleName(aEntry.mModuleName),
+        mModulePath(aEntry.mModulePath),
+        mDebugName(aEntry.mDebugName),
+        mDebugPath(aEntry.mDebugPath),
+        mVersion(aEntry.mVersion),
+        mArch(aEntry.mArch) {}
+
+  SharedLibrary& operator=(const SharedLibrary& aEntry) {
+    // Gracefully handle self assignment
+    if (this == &aEntry) return *this;
+
+    mStart = aEntry.mStart;
+    mEnd = aEntry.mEnd;
+    mOffset = aEntry.mOffset;
+    mBreakpadId = aEntry.mBreakpadId;
+    mModuleName = aEntry.mModuleName;
+    mModulePath = aEntry.mModulePath;
+    mDebugName = aEntry.mDebugName;
+    mDebugPath = aEntry.mDebugPath;
+    mVersion = aEntry.mVersion;
+    mArch = aEntry.mArch;
+    return *this;
+  }
+
+  bool operator==(const SharedLibrary& other) const {
+    return (mStart == other.mStart) && (mEnd == other.mEnd) &&
+           (mOffset == other.mOffset) && (mModuleName == other.mModuleName) &&
+           (mModulePath == other.mModulePath) &&
+           (mDebugName == other.mDebugName) &&
+           (mDebugPath == other.mDebugPath) &&
+           (mBreakpadId == other.mBreakpadId) && (mVersion == other.mVersion) &&
+           (mArch == other.mArch);
+  }
+
+  uintptr_t GetStart() const { return mStart; }
+  uintptr_t GetEnd() const { return mEnd; }
+  uintptr_t GetOffset() const { return mOffset; }
+  const std::string& GetBreakpadId() const { return mBreakpadId; }
+  const std::string& GetModuleName() const { return mModuleName; }
+  const std::string& GetModulePath() const { return mModulePath; }
+  const std::string& GetDebugName() const { return mDebugName; }
+  const std::string& GetDebugPath() const { return mDebugPath; }
+  const std::string& GetVersion() const { return mVersion; }
+  const std::string& GetArch() const { return mArch; }
+
+ private:
+  SharedLibrary() : mStart{0}, mEnd{0}, mOffset{0} {}
+
+  uintptr_t mStart;
+  uintptr_t mEnd;
+  uintptr_t mOffset;
+  std::string mBreakpadId;
+  std::string mModuleName;
+  std::string mModulePath;
+  std::string mDebugName;
+  std::string mDebugPath;
+  std::string mVersion;
+  std::string mArch;
+};
+
+static bool CompareAddresses(const SharedLibrary& first,
+                             const SharedLibrary& second) {
+  return first.GetStart() < second.GetStart();
+}
+
+class SharedLibraryInfo {
+ public:
+  static SharedLibraryInfo GetInfoForSelf();
+  static void Initialize();
+
+  SharedLibraryInfo() {}
+
+  void AddSharedLibrary(SharedLibrary entry) { mEntries.push_back(entry); }
+
+  const SharedLibrary& GetEntry(size_t i) const { return mEntries[i]; }
+
+  SharedLibrary& GetMutableEntry(size_t i) { return mEntries[i]; }
+
+  // Removes items in the range [first, last)
+  // i.e. element at the "last" index is not removed
+  void RemoveEntries(size_t first, size_t last) {
+    mEntries.erase(mEntries.begin() + first, mEntries.begin() + last);
+  }
+
+  bool Contains(const SharedLibrary& searchItem) const {
+    return (mEntries.end() !=
+            std::find(mEntries.begin(), mEntries.end(), searchItem));
+  }
+
+  size_t GetSize() const { return mEntries.size(); }
+
+  void SortByAddress() {
+    std::sort(mEntries.begin(), mEntries.end(), CompareAddresses);
+  }
+
+  void Clear() { mEntries.clear(); }
+
+ private:
+  std::vector<SharedLibrary> mEntries;
+};
+
+#endif  // BASE_PROFILER_SHARED_LIBRARIES_H_
diff --git a/mozglue/baseprofiler/public/BaseProfilingCategory.h b/mozglue/baseprofiler/public/BaseProfilingCategory.h
new file mode 100644
index 0000000000..6892ec40f4
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilingCategory.h
@@ -0,0 +1,72 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 4 -*-
+ * vim: set ts=8 sts=4 et sw=4 tw=99:
+ * 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 BaseProfilingCategory_h
+#define BaseProfilingCategory_h
+
+#ifndef MOZ_GECKO_PROFILER
+#  error Do not #include this header when MOZ_GECKO_PROFILER is not #defined.
+#endif
+
+#include "mozilla/Types.h"
+
+#include <cstdint>
+
+#include "ProfilingCategoryList.h"
+
+namespace mozilla {
+namespace baseprofiler {
+
+// clang-format off
+
+// An enum that lists all possible category pairs in one list.
+// This is the enum that is used in profiler stack labels. Having one list that
+// includes subcategories from all categories in one list allows assigning the
+// category pair to a stack label with just one number.
+#define CATEGORY_ENUM_BEGIN_CATEGORY(name, labelAsString, color)
+#define CATEGORY_ENUM_SUBCATEGORY(supercategory, name, labelAsString) name,
+#define CATEGORY_ENUM_END_CATEGORY
+enum class ProfilingCategoryPair : uint32_t {
+  MOZ_PROFILING_CATEGORY_LIST(CATEGORY_ENUM_BEGIN_CATEGORY,
+                              CATEGORY_ENUM_SUBCATEGORY,
+                              CATEGORY_ENUM_END_CATEGORY)
+  COUNT,
+  LAST = COUNT - 1,
+};
+#undef CATEGORY_ENUM_BEGIN_CATEGORY
+#undef CATEGORY_ENUM_SUBCATEGORY
+#undef CATEGORY_ENUM_END_CATEGORY
+
+// An enum that lists just the categories without their subcategories.
+#define SUPERCATEGORY_ENUM_BEGIN_CATEGORY(name, labelAsString, color) name,
+#define SUPERCATEGORY_ENUM_SUBCATEGORY(supercategory, name, labelAsString)
+#define SUPERCATEGORY_ENUM_END_CATEGORY
+enum class ProfilingCategory : uint32_t {
+  MOZ_PROFILING_CATEGORY_LIST(SUPERCATEGORY_ENUM_BEGIN_CATEGORY,
+                              SUPERCATEGORY_ENUM_SUBCATEGORY,
+                              SUPERCATEGORY_ENUM_END_CATEGORY)
+  COUNT,
+  LAST = COUNT - 1,
+};
+#undef SUPERCATEGORY_ENUM_BEGIN_CATEGORY
+#undef SUPERCATEGORY_ENUM_SUBCATEGORY
+#undef SUPERCATEGORY_ENUM_END_CATEGORY
+
+// clang-format on
+
+struct ProfilingCategoryPairInfo {
+  ProfilingCategory mCategory;
+  uint32_t mSubcategoryIndex;
+  const char* mLabel;
+};
+
+MFBT_API const ProfilingCategoryPairInfo& GetProfilingCategoryPairInfo(
+    ProfilingCategoryPair aCategoryPair);
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif /* BaseProfilingCategory_h */
diff --git a/mozglue/baseprofiler/public/BaseProfilingStack.h b/mozglue/baseprofiler/public/BaseProfilingStack.h
new file mode 100644
index 0000000000..214fc1ebbf
--- /dev/null
+++ b/mozglue/baseprofiler/public/BaseProfilingStack.h
@@ -0,0 +1,520 @@
+/* -*- 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 BaseProfilingStack_h
+#define BaseProfilingStack_h
+
+#include "BaseProfilingCategory.h"
+
+#include "mozilla/Atomics.h"
+
+#include "BaseProfiler.h"
+
+#ifndef MOZ_GECKO_PROFILER
+#  error Do not #include this header when MOZ_GECKO_PROFILER is not #defined.
+#endif
+
+#include <algorithm>
+#include <stdint.h>
+
+// This file defines the classes ProfilingStack and ProfilingStackFrame.
+// The ProfilingStack manages an array of ProfilingStackFrames.
+// It keeps track of the "label stack" and the JS interpreter stack.
+// The two stack types are interleaved.
+//
+// Usage:
+//
+//  ProfilingStack* profilingStack = ...;
+//
+//  // For label frames:
+//  profilingStack->pushLabelFrame(...);
+//  // Execute some code. When finished, pop the frame:
+//  profilingStack->pop();
+//
+//  // For JS stack frames:
+//  profilingStack->pushJSFrame(...);
+//  // Execute some code. When finished, pop the frame:
+//  profilingStack->pop();
+//
+//
+// Concurrency considerations
+//
+// A thread's profiling stack (and the frames inside it) is only modified by
+// that thread. However, the profiling stack can be *read* by a different
+// thread, the sampler thread: Whenever the profiler wants to sample a given
+// thread A, the following happens:
+//  (1) Thread A is suspended.
+//  (2) The sampler thread (thread S) reads the ProfilingStack of thread A,
+//      including all ProfilingStackFrames that are currently in that stack
+//      (profilingStack->frames[0..profilingStack->stackSize()]).
+//  (3) Thread A is resumed.
+//
+// Thread suspension is achieved using platform-specific APIs; refer to each
+// platform's Sampler::SuspendAndSampleAndResumeThread implementation in
+// platform-*.cpp for details.
+//
+// When the thread is suspended, the values in profilingStack->stackPointer and
+// in the stack frame range
+// profilingStack->frames[0..profilingStack->stackPointer] need to be in a
+// consistent state, so that thread S does not read partially- constructed stack
+// frames. More specifically, we have two requirements:
+//  (1) When adding a new frame at the top of the stack, its ProfilingStackFrame
+//      data needs to be put in place *before* the stackPointer is incremented,
+//      and the compiler + CPU need to know that this order matters.
+//  (2) When popping an frame from the stack and then preparing the
+//      ProfilingStackFrame data for the next frame that is about to be pushed,
+//      the decrement of the stackPointer in pop() needs to happen *before* the
+//      ProfilingStackFrame for the new frame is being popuplated, and the
+//      compiler + CPU need to know that this order matters.
+//
+// We can express the relevance of these orderings in multiple ways.
+// Option A is to make stackPointer an atomic with SequentiallyConsistent
+// memory ordering. This would ensure that no writes in thread A would be
+// reordered across any writes to stackPointer, which satisfies requirements
+// (1) and (2) at the same time. Option A is the simplest.
+// Option B is to use ReleaseAcquire memory ordering both for writes to
+// stackPointer *and* for writes to ProfilingStackFrame fields. Release-stores
+// ensure that all writes that happened *before this write in program order* are
+// not reordered to happen after this write. ReleaseAcquire ordering places no
+// requirements on the ordering of writes that happen *after* this write in
+// program order.
+// Using release-stores for writes to stackPointer expresses requirement (1),
+// and using release-stores for writes to the ProfilingStackFrame fields
+// expresses requirement (2).
+//
+// Option B is more complicated than option A, but has much better performance
+// on x86/64: In a microbenchmark run on a Macbook Pro from 2017, switching
+// from option A to option B reduced the overhead of pushing+popping a
+// ProfilingStackFrame by 10 nanoseconds.
+// On x86/64, release-stores require no explicit hardware barriers or lock
+// instructions.
+// On ARM/64, option B may be slower than option A, because the compiler will
+// generate hardware barriers for every single release-store instead of just
+// for the writes to stackPointer. However, the actual performance impact of
+// this has not yet been measured on ARM, so we're currently using option B
+// everywhere. This is something that we may want to change in the future once
+// we've done measurements.
+
+namespace mozilla {
+namespace baseprofiler {
+
+// A call stack can be specified to the JS engine such that all JS entry/exits
+// to functions push/pop a stack frame to/from the specified stack.
+//
+// For more detailed information, see vm/GeckoProfiler.h.
+//
+class ProfilingStackFrame {
+  // A ProfilingStackFrame represents either a label frame or a JS frame.
+
+  // WARNING WARNING WARNING
+  //
+  // All the fields below are Atomic<...,ReleaseAcquire>. This is needed so
+  // that writes to these fields are release-writes, which ensures that
+  // earlier writes in this thread don't get reordered after the writes to
+  // these fields. In particular, the decrement of the stack pointer in
+  // ProfilingStack::pop() is a write that *must* happen before the values in
+  // this ProfilingStackFrame are changed. Otherwise, the sampler thread might
+  // see an inconsistent state where the stack pointer still points to a
+  // ProfilingStackFrame which has already been popped off the stack and whose
+  // fields have now been partially repopulated with new values.
+  // See the "Concurrency considerations" paragraph at the top of this file
+  // for more details.
+
+  // Descriptive label for this stack frame. Must be a static string! Can be
+  // an empty string, but not a null pointer.
+  Atomic<const char*, ReleaseAcquire> label_;
+
+  // An additional descriptive string of this frame which is combined with
+  // |label_| in profiler output. Need not be (and usually isn't) static. Can
+  // be null.
+  Atomic<const char*, ReleaseAcquire> dynamicString_;
+
+  // Stack pointer for non-JS stack frames, the script pointer otherwise.
+  Atomic<void*, ReleaseAcquire> spOrScript;
+
+  // ID of the JS Realm for JS stack frames.
+  // Must not be used on non-JS frames; it'll contain either the default 0,
+  // or a leftover value from a previous JS stack frame that was using this
+  // ProfilingStackFrame object.
+  mozilla::Atomic<uint64_t, mozilla::ReleaseAcquire> realmID_;
+
+  // The bytecode offset for JS stack frames.
+  // Must not be used on non-JS frames; it'll contain either the default 0,
+  // or a leftover value from a previous JS stack frame that was using this
+  // ProfilingStackFrame object.
+  Atomic<int32_t, ReleaseAcquire> pcOffsetIfJS_;
+
+  // Bits 0...8 hold the Flags. Bits 9...31 hold the category pair.
+  Atomic<uint32_t, ReleaseAcquire> flagsAndCategoryPair_;
+
+ public:
+  ProfilingStackFrame() = default;
+  ProfilingStackFrame& operator=(const ProfilingStackFrame& other) {
+    label_ = other.label();
+    dynamicString_ = other.dynamicString();
+    void* spScript = other.spOrScript;
+    spOrScript = spScript;
+    int32_t offsetIfJS = other.pcOffsetIfJS_;
+    pcOffsetIfJS_ = offsetIfJS;
+    int64_t realmID = other.realmID_;
+    realmID_ = realmID;
+    uint32_t flagsAndCategory = other.flagsAndCategoryPair_;
+    flagsAndCategoryPair_ = flagsAndCategory;
+    return *this;
+  }
+
+  // Reserve up to 16 bits for flags, and 16 for category pair.
+  enum class Flags : uint32_t {
+    // The first three flags describe the kind of the frame and are
+    // mutually exclusive. (We still give them individual bits for
+    // simplicity.)
+
+    // A regular label frame. These usually come from AutoProfilerLabel.
+    IS_LABEL_FRAME = 1 << 0,
+
+    // A special frame indicating the start of a run of JS profiling stack
+    // frames. IS_SP_MARKER_FRAME frames are ignored, except for the sp
+    // field. These frames are needed to get correct ordering between JS
+    // and LABEL frames because JS frames don't carry sp information.
+    // SP is short for "stack pointer".
+    IS_SP_MARKER_FRAME = 1 << 1,
+
+    // A JS frame.
+    IS_JS_FRAME = 1 << 2,
+
+    // An interpreter JS frame that has OSR-ed into baseline. IS_JS_FRAME
+    // frames can have this flag set and unset during their lifetime.
+    // JS_OSR frames are ignored.
+    JS_OSR = 1 << 3,
+
+    // The next three are mutually exclusive.
+    // By default, for profiling stack frames that have both a label and a
+    // dynamic string, the two strings are combined into one string of the
+    // form "<label> <dynamicString>" during JSON serialization. The
+    // following flags can be used to change this preset.
+    STRING_TEMPLATE_METHOD = 1 << 4,  // "<label>.<dynamicString>"
+    STRING_TEMPLATE_GETTER = 1 << 5,  // "get <label>.<dynamicString>"
+    STRING_TEMPLATE_SETTER = 1 << 6,  // "set <label>.<dynamicString>"
+
+    // If set, causes this stack frame to be marked as "relevantForJS" in
+    // the profile JSON, which will make it show up in the "JS only" call
+    // tree view.
+    RELEVANT_FOR_JS = 1 << 7,
+
+    // If set, causes the label on this ProfilingStackFrame to be ignored
+    // and to be replaced by the subcategory's label.
+    LABEL_DETERMINED_BY_CATEGORY_PAIR = 1 << 8,
+
+    // Frame dynamic string does not contain user data.
+    NONSENSITIVE = 1 << 9,
+
+    // A JS Baseline Interpreter frame.
+    IS_BLINTERP_FRAME = 1 << 10,
+
+    FLAGS_BITCOUNT = 16,
+    FLAGS_MASK = (1 << FLAGS_BITCOUNT) - 1
+  };
+
+  static_assert(
+      uint32_t(ProfilingCategoryPair::LAST) <=
+          (UINT32_MAX >> uint32_t(Flags::FLAGS_BITCOUNT)),
+      "Too many category pairs to fit into u32 with together with the "
+      "reserved bits for the flags");
+
+  bool isLabelFrame() const {
+    return uint32_t(flagsAndCategoryPair_) & uint32_t(Flags::IS_LABEL_FRAME);
+  }
+
+  bool isSpMarkerFrame() const {
+    return uint32_t(flagsAndCategoryPair_) &
+           uint32_t(Flags::IS_SP_MARKER_FRAME);
+  }
+
+  bool isJsFrame() const {
+    return uint32_t(flagsAndCategoryPair_) & uint32_t(Flags::IS_JS_FRAME);
+  }
+
+  bool isOSRFrame() const {
+    return uint32_t(flagsAndCategoryPair_) & uint32_t(Flags::JS_OSR);
+  }
+
+  void setIsOSRFrame(bool isOSR) {
+    if (isOSR) {
+      flagsAndCategoryPair_ =
+          uint32_t(flagsAndCategoryPair_) | uint32_t(Flags::JS_OSR);
+    } else {
+      flagsAndCategoryPair_ =
+          uint32_t(flagsAndCategoryPair_) & ~uint32_t(Flags::JS_OSR);
+    }
+  }
+
+  const char* label() const {
+    uint32_t flagsAndCategoryPair = flagsAndCategoryPair_;
+    if (flagsAndCategoryPair &
+        uint32_t(Flags::LABEL_DETERMINED_BY_CATEGORY_PAIR)) {
+      auto categoryPair = ProfilingCategoryPair(
+          flagsAndCategoryPair >> uint32_t(Flags::FLAGS_BITCOUNT));
+      return GetProfilingCategoryPairInfo(categoryPair).mLabel;
+    }
+    return label_;
+  }
+
+  const char* dynamicString() const { return dynamicString_; }
+
+  void initLabelFrame(const char* aLabel, const char* aDynamicString, void* sp,
+                      ProfilingCategoryPair aCategoryPair, uint32_t aFlags) {
+    label_ = aLabel;
+    dynamicString_ = aDynamicString;
+    spOrScript = sp;
+    // pcOffsetIfJS_ is not set and must not be used on label frames.
+    flagsAndCategoryPair_ =
+        uint32_t(Flags::IS_LABEL_FRAME) |
+        (uint32_t(aCategoryPair) << uint32_t(Flags::FLAGS_BITCOUNT)) | aFlags;
+    MOZ_ASSERT(isLabelFrame());
+  }
+
+  void initSpMarkerFrame(void* sp) {
+    label_ = "";
+    dynamicString_ = nullptr;
+    spOrScript = sp;
+    // pcOffsetIfJS_ is not set and must not be used on sp marker frames.
+    flagsAndCategoryPair_ = uint32_t(Flags::IS_SP_MARKER_FRAME) |
+                            (uint32_t(ProfilingCategoryPair::OTHER)
+                             << uint32_t(Flags::FLAGS_BITCOUNT));
+    MOZ_ASSERT(isSpMarkerFrame());
+  }
+
+  void initJsFrame(const char* aLabel, const char* aDynamicString,
+                   void* /* JSScript* */ aScript, int32_t aOffset,
+                   uint64_t aRealmID) {
+    label_ = aLabel;
+    dynamicString_ = aDynamicString;
+    spOrScript = aScript;
+    pcOffsetIfJS_ = aOffset;
+    realmID_ = aRealmID;
+    flagsAndCategoryPair_ =
+        uint32_t(Flags::IS_JS_FRAME) | (uint32_t(ProfilingCategoryPair::JS)
+                                        << uint32_t(Flags::FLAGS_BITCOUNT));
+    MOZ_ASSERT(isJsFrame());
+  }
+
+  uint32_t flags() const {
+    return uint32_t(flagsAndCategoryPair_) & uint32_t(Flags::FLAGS_MASK);
+  }
+
+  ProfilingCategoryPair categoryPair() const {
+    return ProfilingCategoryPair(flagsAndCategoryPair_ >>
+                                 uint32_t(Flags::FLAGS_BITCOUNT));
+  }
+
+  uint64_t realmID() const { return realmID_; }
+
+  void* stackAddress() const {
+    MOZ_ASSERT(!isJsFrame());
+    return spOrScript;
+  }
+
+  // Note that the pointer returned might be invalid.
+  void* rawScript() const {
+    MOZ_ASSERT(isJsFrame());
+    return spOrScript;
+  }
+  void setRawScript(void* aScript) {
+    MOZ_ASSERT(isJsFrame());
+    spOrScript = aScript;
+  }
+
+  int32_t pcOffset() const {
+    MOZ_ASSERT(isJsFrame());
+    return pcOffsetIfJS_;
+  }
+
+  void setPCOffset(int32_t aOffset) {
+    MOZ_ASSERT(isJsFrame());
+    pcOffsetIfJS_ = aOffset;
+  }
+
+  // The offset of a pc into a script's code can actually be 0, so to
+  // signify a nullptr pc, use a -1 index. This is checked against in
+  // pc() and setPC() to set/get the right pc.
+  static const int32_t NullPCOffset = -1;
+};
+
+// Each thread has its own ProfilingStack. That thread modifies the
+// ProfilingStack, pushing and popping elements as necessary.
+//
+// The ProfilingStack is also read periodically by the profiler's sampler
+// thread. This happens only when the thread that owns the ProfilingStack is
+// suspended. So there are no genuine parallel accesses.
+//
+// However, it is possible for pushing/popping to be interrupted by a periodic
+// sample. Because of this, we need pushing/popping to be effectively atomic.
+//
+// - When pushing a new frame, we increment the stack pointer -- making the new
+//   frame visible to the sampler thread -- only after the new frame has been
+//   fully written. The stack pointer is Atomic<uint32_t,ReleaseAcquire>, so
+//   the increment is a release-store, which ensures that this store is not
+//   reordered before the writes of the frame.
+//
+// - When popping an old frame, the only operation is the decrementing of the
+//   stack pointer, which is obviously atomic.
+//
+class ProfilingStack final {
+ public:
+  ProfilingStack() = default;
+
+  MFBT_API ~ProfilingStack();
+
+  void pushLabelFrame(const char* label, const char* dynamicString, void* sp,
+                      ProfilingCategoryPair categoryPair, uint32_t flags = 0) {
+    // This thread is the only one that ever changes the value of
+    // stackPointer.
+    // Store the value of the atomic in a non-atomic local variable so that
+    // the compiler won't generate two separate loads from the atomic for
+    // the size check and the frames[] array indexing operation.
+    uint32_t stackPointerVal = stackPointer;
+
+    if (MOZ_UNLIKELY(stackPointerVal >= capacity)) {
+      ensureCapacitySlow();
+    }
+    frames[stackPointerVal].initLabelFrame(label, dynamicString, sp,
+                                           categoryPair, flags);
+
+    // This must happen at the end! The compiler will not reorder this
+    // update because stackPointer is Atomic<..., ReleaseAcquire>, so any
+    // the writes above will not be reordered below the stackPointer store.
+    // Do the read and the write as two separate statements, in order to
+    // make it clear that we don't need an atomic increment, which would be
+    // more expensive on x86 than the separate operations done here.
+    // However, don't use stackPointerVal here; instead, allow the compiler
+    // to turn this store into a non-atomic increment instruction which
+    // takes up less code size.
+    stackPointer = stackPointer + 1;
+  }
+
+  void pushSpMarkerFrame(void* sp) {
+    uint32_t oldStackPointer = stackPointer;
+
+    if (MOZ_UNLIKELY(oldStackPointer >= capacity)) {
+      ensureCapacitySlow();
+    }
+    frames[oldStackPointer].initSpMarkerFrame(sp);
+
+    // This must happen at the end, see the comment in pushLabelFrame.
+    stackPointer = oldStackPointer + 1;
+  }
+
+  void pushJsOffsetFrame(const char* label, const char* dynamicString,
+                         void* script, int32_t offset, uint64_t aRealmID) {
+    // This thread is the only one that ever changes the value of
+    // stackPointer. Only load the atomic once.
+    uint32_t oldStackPointer = stackPointer;
+
+    if (MOZ_UNLIKELY(oldStackPointer >= capacity)) {
+      ensureCapacitySlow();
+    }
+    frames[oldStackPointer].initJsFrame(label, dynamicString, script, offset,
+                                        aRealmID);
+
+    // This must happen at the end, see the comment in pushLabelFrame.
+    stackPointer = stackPointer + 1;
+  }
+
+  void pop() {
+    MOZ_ASSERT(stackPointer > 0);
+    // Do the read and the write as two separate statements, in order to
+    // make it clear that we don't need an atomic decrement, which would be
+    // more expensive on x86 than the separate operations done here.
+    // This thread is the only one that ever changes the value of
+    // stackPointer.
+    uint32_t oldStackPointer = stackPointer;
+    stackPointer = oldStackPointer - 1;
+  }
+
+  uint32_t stackSize() const { return stackPointer; }
+  uint32_t stackCapacity() const { return capacity; }
+
+ private:
+  // Out of line path for expanding the buffer, since otherwise this would get
+  // inlined in every DOM WebIDL call.
+  MFBT_API MOZ_COLD void ensureCapacitySlow();
+
+  // No copying.
+  ProfilingStack(const ProfilingStack&) = delete;
+  void operator=(const ProfilingStack&) = delete;
+
+  // No moving either.
+  ProfilingStack(ProfilingStack&&) = delete;
+  void operator=(ProfilingStack&&) = delete;
+
+  uint32_t capacity = 0;
+
+ public:
+  // The pointer to the stack frames, this is read from the profiler thread and
+  // written from the current thread.
+  //
+  // This is effectively a unique pointer.
+  Atomic<ProfilingStackFrame*, SequentiallyConsistent> frames{nullptr};
+
+  // This may exceed the capacity, so instead use the stackSize() method to
+  // determine the number of valid frames in stackFrames. When this is less
+  // than stackCapacity(), it refers to the first free stackframe past the top
+  // of the in-use stack (i.e. frames[stackPointer - 1] is the top stack
+  // frame).
+  //
+  // WARNING WARNING WARNING
+  //
+  // This is an atomic variable that uses ReleaseAcquire memory ordering.
+  // See the "Concurrency considerations" paragraph at the top of this file
+  // for more details.
+  Atomic<uint32_t, ReleaseAcquire> stackPointer{0};
+};
+
+class AutoGeckoProfilerEntry;
+class GeckoProfilerEntryMarker;
+class GeckoProfilerBaselineOSRMarker;
+
+class GeckoProfilerThread {
+  friend class AutoGeckoProfilerEntry;
+  friend class GeckoProfilerEntryMarker;
+  friend class GeckoProfilerBaselineOSRMarker;
+
+  ProfilingStack* profilingStack_;
+
+  // Same as profilingStack_ if the profiler is currently active, otherwise
+  // null.
+  ProfilingStack* profilingStackIfEnabled_;
+
+ public:
+  MFBT_API GeckoProfilerThread();
+
+  uint32_t stackPointer() {
+    MOZ_ASSERT(infraInstalled());
+    return profilingStack_->stackPointer;
+  }
+  ProfilingStackFrame* stack() { return profilingStack_->frames; }
+  ProfilingStack* getProfilingStack() { return profilingStack_; }
+  ProfilingStack* getProfilingStackIfEnabled() {
+    return profilingStackIfEnabled_;
+  }
+
+  /*
+   * True if the profiler infrastructure is setup.  Should be true in builds
+   * that include profiler support except during early startup or late
+   * shutdown.  Unrelated to the presence of the Gecko Profiler addon.
+   */
+  bool infraInstalled() { return profilingStack_ != nullptr; }
+
+  MFBT_API void setProfilingStack(ProfilingStack* profilingStack, bool enabled);
+  void enable(bool enable) {
+    profilingStackIfEnabled_ = enable ? profilingStack_ : nullptr;
+  }
+};
+
+}  // namespace baseprofiler
+}  // namespace mozilla
+
+#endif /* BaseProfilingStack_h */
diff --git a/mozglue/baseprofiler/public/BlocksRingBuffer.h b/mozglue/baseprofiler/public/BlocksRingBuffer.h
new file mode 100644
index 0000000000..6948ab8cf4
--- /dev/null
+++ b/mozglue/baseprofiler/public/BlocksRingBuffer.h
@@ -0,0 +1,1000 @@
+/* -*- Mode: C++; tab-width: 2; 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 BlocksRingBuffer_h
+#define BlocksRingBuffer_h
+
+#include "mozilla/BaseProfilerDetail.h"
+#include "mozilla/ModuloBuffer.h"
+#include "mozilla/ProfileBufferIndex.h"
+#include "mozilla/ScopeExit.h"
+
+#include <functional>
+#include <string>
+#include <tuple>
+#include <utility>
+
+namespace mozilla {
+
+// Thread-safe Ring buffer that can store blocks of different sizes during
+// defined sessions.
+// Each *block* contains an *entry* and the entry size:
+// [ entry_size | entry ] [ entry_size | entry ] ...
+// *In-session* is a period of time during which `BlocksRingBuffer` allows
+// reading and writing. *Out-of-session*, the `BlocksRingBuffer` object is
+// still valid, but contains no data, and gracefully denies accesses.
+//
+// To write an entry, the buffer reserves a block of sufficient size (to contain
+// user data of predetermined size), writes the entry size, and lets the caller
+// fill the entry contents using ModuloBuffer::Iterator APIs and a few entry-
+// specific APIs. E.g.:
+// ```
+// BlockRingsBuffer brb(PowerOfTwo<BlockRingsBuffer::Length>(1024));
+// brb.ReserveAndPut([]() { return sizeof(123); },
+//                   [&](ProfileBufferEntryWriter& aEW) {
+//                     aEW.WriteObject(123);
+//                   });
+// ```
+// Other `Put...` functions may be used as shortcuts for simple entries.
+// The objects given to the caller's callbacks should only be used inside the
+// callbacks and not stored elsewhere, because they keep their own references to
+// the BlocksRingBuffer and therefore should not live longer.
+// Different type of objects may be serialized into an entry, see `Serializer`
+// for more information.
+//
+// When reading data, the buffer iterates over blocks (it knows how to read the
+// entry size, and therefore move to the next block), and lets the caller read
+// the entry inside of each block. E.g.:
+// ```
+// brb.Read([](BlocksRingBuffer::Reader aR) {}
+//   for (ProfileBufferEntryReader aER : aR) {
+//     /* Use ProfileBufferEntryReader functions to read serialized objects. */
+//     int n = aER.ReadObject<int>();
+//   }
+// });
+// ```
+// Different type of objects may be deserialized from an entry, see
+// `Deserializer` for more information.
+//
+// The caller may retrieve the `ProfileBufferBlockIndex` corresponding to an
+// entry (`ProfileBufferBlockIndex` is an opaque type preventing the user from
+// modifying it). That index may later be used to get back to that particular
+// entry if it still exists.
+class BlocksRingBuffer {
+ public:
+  // Using ModuloBuffer as underlying circular byte buffer.
+  using Buffer = ModuloBuffer<uint32_t, ProfileBufferIndex>;
+  using Byte = Buffer::Byte;
+
+  // Length type for total buffer (as PowerOfTwo<Length>) and each entry.
+  using Length = uint32_t;
+
+  enum class ThreadSafety { WithoutMutex, WithMutex };
+
+  // Default constructor starts out-of-session (nothing to read or write).
+  explicit BlocksRingBuffer(ThreadSafety aThreadSafety)
+      : mMutex(aThreadSafety != ThreadSafety::WithoutMutex) {}
+
+  // Create a buffer of the given length.
+  explicit BlocksRingBuffer(ThreadSafety aThreadSafety,
+                            PowerOfTwo<Length> aLength)
+      : mMutex(aThreadSafety != ThreadSafety::WithoutMutex),
+        mMaybeUnderlyingBuffer(Some(UnderlyingBuffer(aLength))) {}
+
+  // Take ownership of an existing buffer.
+  BlocksRingBuffer(ThreadSafety aThreadSafety,
+                   UniquePtr<Buffer::Byte[]> aExistingBuffer,
+                   PowerOfTwo<Length> aLength)
+      : mMutex(aThreadSafety != ThreadSafety::WithoutMutex),
+        mMaybeUnderlyingBuffer(
+            Some(UnderlyingBuffer(std::move(aExistingBuffer), aLength))) {}
+
+  // Use an externally-owned buffer.
+  BlocksRingBuffer(ThreadSafety aThreadSafety, Buffer::Byte* aExternalBuffer,
+                   PowerOfTwo<Length> aLength)
+      : mMutex(aThreadSafety != ThreadSafety::WithoutMutex),
+        mMaybeUnderlyingBuffer(
+            Some(UnderlyingBuffer(aExternalBuffer, aLength))) {}
+
+  // Destructor doesn't need to do anything special. (Clearing entries would
+  // only update indices and stats, which won't be accessible after the object
+  // is destroyed anyway.)
+  ~BlocksRingBuffer() = default;
+
+  // Remove underlying buffer, if any.
+  void Reset() {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    ResetUnderlyingBuffer();
+  }
+
+  // Create a buffer of the given length.
+  void Set(PowerOfTwo<Length> aLength) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    ResetUnderlyingBuffer();
+    mMaybeUnderlyingBuffer.emplace(aLength);
+  }
+
+  // Take ownership of an existing buffer.
+  void Set(UniquePtr<Buffer::Byte[]> aExistingBuffer,
+           PowerOfTwo<Length> aLength) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    ResetUnderlyingBuffer();
+    mMaybeUnderlyingBuffer.emplace(std::move(aExistingBuffer), aLength);
+  }
+
+  // Use an externally-owned buffer.
+  void Set(Buffer::Byte* aExternalBuffer, PowerOfTwo<Length> aLength) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    ResetUnderlyingBuffer();
+    mMaybeUnderlyingBuffer.emplace(aExternalBuffer, aLength);
+  }
+
+  // This cannot change during the lifetime of this buffer, so there's no need
+  // to lock.
+  bool IsThreadSafe() const { return mMutex.IsActivated(); }
+
+  [[nodiscard]] bool IsInSession() const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return !!mMaybeUnderlyingBuffer;
+  }
+
+  // Lock the buffer mutex and run the provided callback.
+  // This can be useful when the caller needs to explicitly lock down this
+  // buffer, but not do anything else with it.
+  template <typename Callback>
+  auto LockAndRun(Callback&& aCallback) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return std::forward<Callback>(aCallback)();
+  }
+
+  // Buffer length in bytes.
+  Maybe<PowerOfTwo<Length>> BufferLength() const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return mMaybeUnderlyingBuffer.map([](const UnderlyingBuffer& aBuffer) {
+      return aBuffer.mBuffer.BufferLength();
+    });
+    ;
+  }
+
+  // Size of external resources.
+  size_t SizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const {
+    if (!mMaybeUnderlyingBuffer) {
+      return 0;
+    }
+    return mMaybeUnderlyingBuffer->mBuffer.SizeOfExcludingThis(aMallocSizeOf);
+  }
+
+  size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+    return aMallocSizeOf(this) + SizeOfExcludingThis(aMallocSizeOf);
+  }
+
+  // Snapshot of the buffer state.
+  struct State {
+    // Index to the first block.
+    ProfileBufferBlockIndex mRangeStart;
+
+    // Index past the last block. Equals mRangeStart if empty.
+    ProfileBufferBlockIndex mRangeEnd;
+
+    // Number of blocks that have been pushed into this buffer.
+    uint64_t mPushedBlockCount = 0;
+
+    // Number of blocks that have been removed from this buffer.
+    // Note: Live entries = pushed - cleared.
+    uint64_t mClearedBlockCount = 0;
+  };
+
+  // Get a snapshot of the current state.
+  // When out-of-session, mFirstReadIndex==mNextWriteIndex, and
+  // mPushedBlockCount==mClearedBlockCount==0.
+  // Note that these may change right after this thread-safe call, so they
+  // should only be used for statistical purposes.
+  State GetState() const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return {
+        mFirstReadIndex, mNextWriteIndex,
+        mMaybeUnderlyingBuffer ? mMaybeUnderlyingBuffer->mPushedBlockCount : 0,
+        mMaybeUnderlyingBuffer ? mMaybeUnderlyingBuffer->mClearedBlockCount
+                               : 0};
+  }
+
+  class Reader;
+
+  // Class that can iterate through blocks and provide
+  // `ProfileBufferEntryReader`s.
+  // Created through `Reader`, lives within a lock guard lifetime.
+  class BlockIterator {
+   public:
+#ifdef DEBUG
+    ~BlockIterator() {
+      // No BlockIterator should live outside of a mutexed call.
+      mRing->mMutex.AssertCurrentThreadOwns();
+    }
+#endif  // DEBUG
+
+    // Comparison with other iterator, mostly used in range-for loops.
+    bool operator==(const BlockIterator aRhs) const {
+      MOZ_ASSERT(mRing == aRhs.mRing);
+      return mBlockIndex == aRhs.mBlockIndex;
+    }
+    bool operator!=(const BlockIterator aRhs) const {
+      MOZ_ASSERT(mRing == aRhs.mRing);
+      return mBlockIndex != aRhs.mBlockIndex;
+    }
+
+    // Advance to next BlockIterator.
+    BlockIterator& operator++() {
+      mBlockIndex = NextBlockIndex();
+      return *this;
+    }
+
+    // Dereferencing creates a `ProfileBufferEntryReader` for the entry inside
+    // this block.
+    ProfileBufferEntryReader operator*() const {
+      return mRing->ReaderInBlockAt(mBlockIndex);
+    }
+
+    // True if this iterator is just past the last entry.
+    bool IsAtEnd() const {
+      MOZ_ASSERT(mBlockIndex <= BufferRangeEnd());
+      return mBlockIndex == BufferRangeEnd();
+    }
+
+    // Can be used as reference to come back to this entry with `ReadAt()`.
+    ProfileBufferBlockIndex CurrentBlockIndex() const { return mBlockIndex; }
+
+    // Index past the end of this block, which is the start of the next block.
+    ProfileBufferBlockIndex NextBlockIndex() const {
+      MOZ_ASSERT(!IsAtEnd());
+      const Length entrySize =
+          mRing->ReaderInBlockAt(mBlockIndex).RemainingBytes();
+      return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          mBlockIndex.ConvertToProfileBufferIndex() + ULEB128Size(entrySize) +
+          entrySize);
+    }
+
+    // Index of the first block in the whole buffer.
+    ProfileBufferBlockIndex BufferRangeStart() const {
+      return mRing->mFirstReadIndex;
+    }
+
+    // Index past the last block in the whole buffer.
+    ProfileBufferBlockIndex BufferRangeEnd() const {
+      return mRing->mNextWriteIndex;
+    }
+
+   private:
+    // Only a Reader can instantiate a BlockIterator.
+    friend class Reader;
+
+    BlockIterator(const BlocksRingBuffer& aRing,
+                  ProfileBufferBlockIndex aBlockIndex)
+        : mRing(WrapNotNull(&aRing)), mBlockIndex(aBlockIndex) {
+      // No BlockIterator should live outside of a mutexed call.
+      mRing->mMutex.AssertCurrentThreadOwns();
+    }
+
+    // Using a non-null pointer instead of a reference, to allow copying.
+    // This BlockIterator should only live inside one of the thread-safe
+    // BlocksRingBuffer functions, for this reference to stay valid.
+    NotNull<const BlocksRingBuffer*> mRing;
+    ProfileBufferBlockIndex mBlockIndex;
+  };
+
+  // Class that can create `BlockIterator`s (e.g., for range-for), or just
+  // iterate through entries; lives within a lock guard lifetime.
+  class MOZ_RAII Reader {
+   public:
+    Reader(const Reader&) = delete;
+    Reader& operator=(const Reader&) = delete;
+    Reader(Reader&&) = delete;
+    Reader& operator=(Reader&&) = delete;
+
+#ifdef DEBUG
+    ~Reader() {
+      // No Reader should live outside of a mutexed call.
+      mRing.mMutex.AssertCurrentThreadOwns();
+    }
+#endif  // DEBUG
+
+    // Index of the first block in the whole buffer.
+    ProfileBufferBlockIndex BufferRangeStart() const {
+      return mRing.mFirstReadIndex;
+    }
+
+    // Index past the last block in the whole buffer.
+    ProfileBufferBlockIndex BufferRangeEnd() const {
+      return mRing.mNextWriteIndex;
+    }
+
+    // Iterators to the first and past-the-last blocks.
+    // Compatible with range-for (see `ForEach` below as example).
+    BlockIterator begin() const {
+      return BlockIterator(mRing, BufferRangeStart());
+    }
+    // Note that a `BlockIterator` at the `end()` should not be dereferenced, as
+    // there is no actual block there!
+    BlockIterator end() const { return BlockIterator(mRing, BufferRangeEnd()); }
+
+    // Get a `BlockIterator` at the given `ProfileBufferBlockIndex`, clamped to
+    // the stored range. Note that a `BlockIterator` at the `end()` should not
+    // be dereferenced, as there is no actual block there!
+    BlockIterator At(ProfileBufferBlockIndex aBlockIndex) const {
+      if (aBlockIndex < BufferRangeStart()) {
+        // Anything before the range (including null ProfileBufferBlockIndex) is
+        // clamped at the beginning.
+        return begin();
+      }
+      // Otherwise we at least expect the index to be valid (pointing exactly at
+      // a live block, or just past the end.)
+      mRing.AssertBlockIndexIsValidOrEnd(aBlockIndex);
+      return BlockIterator(mRing, aBlockIndex);
+    }
+
+    // Run `aCallback(ProfileBufferEntryReader&)` on each entry from first to
+    // last. Callback should not store `ProfileBufferEntryReader`, as it may
+    // become invalid after this thread-safe call.
+    template <typename Callback>
+    void ForEach(Callback&& aCallback) const {
+      for (ProfileBufferEntryReader reader : *this) {
+        aCallback(reader);
+      }
+    }
+
+   private:
+    friend class BlocksRingBuffer;
+
+    explicit Reader(const BlocksRingBuffer& aRing) : mRing(aRing) {
+      // No Reader should live outside of a mutexed call.
+      mRing.mMutex.AssertCurrentThreadOwns();
+    }
+
+    // This Reader should only live inside one of the thread-safe
+    // BlocksRingBuffer functions, for this reference to stay valid.
+    const BlocksRingBuffer& mRing;
+  };
+
+  // Call `aCallback(BlocksRingBuffer::Reader*)` (nullptr when out-of-session),
+  // and return whatever `aCallback` returns. Callback should not store
+  // `Reader`, because it may become invalid after this call.
+  template <typename Callback>
+  auto Read(Callback&& aCallback) const {
+    {
+      baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+      if (MOZ_LIKELY(mMaybeUnderlyingBuffer)) {
+        Reader reader(*this);
+        return std::forward<Callback>(aCallback)(&reader);
+      }
+    }
+    return std::forward<Callback>(aCallback)(nullptr);
+  }
+
+  // Call `aCallback(ProfileBufferEntryReader&)` on each item.
+  // Callback should not store `ProfileBufferEntryReader`, because it may become
+  // invalid after this call.
+  template <typename Callback>
+  void ReadEach(Callback&& aCallback) const {
+    Read([&](Reader* aReader) {
+      if (MOZ_LIKELY(aReader)) {
+        aReader->ForEach(aCallback);
+      }
+    });
+  }
+
+  // Call `aCallback(Maybe<ProfileBufferEntryReader>&&)` on the entry at
+  // the given ProfileBufferBlockIndex; The `Maybe` will be `Nothing` if
+  // out-of-session, or if that entry doesn't exist anymore, or if we've reached
+  // just past the last entry. Return whatever `aCallback` returns. Callback
+  // should not store `ProfileBufferEntryReader`, because it may become invalid
+  // after this call.
+  template <typename Callback>
+  auto ReadAt(ProfileBufferBlockIndex aBlockIndex, Callback&& aCallback) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    MOZ_ASSERT(aBlockIndex <= mNextWriteIndex);
+    Maybe<ProfileBufferEntryReader> maybeEntryReader;
+    if (MOZ_LIKELY(mMaybeUnderlyingBuffer) && aBlockIndex >= mFirstReadIndex &&
+        aBlockIndex < mNextWriteIndex) {
+      AssertBlockIndexIsValid(aBlockIndex);
+      maybeEntryReader.emplace(ReaderInBlockAt(aBlockIndex));
+    }
+    return std::forward<Callback>(aCallback)(std::move(maybeEntryReader));
+  }
+
+  // Main function to write entries.
+  // Reserve `aCallbackBytes()` bytes, call `aCallback()` with a pointer to an
+  // on-stack temporary ProfileBufferEntryWriter (nullptr when out-of-session),
+  // and return whatever `aCallback` returns. Callback should not store
+  // `ProfileBufferEntryWriter`, because it may become invalid after this
+  // thread-safe call. Note: `aCallbackBytes` is a callback instead of a simple
+  // value, to delay this potentially-expensive computation until after we're
+  // checked that we're in-session; use `Put(Length, Callback)` below if you
+  // know the size already.
+  template <typename CallbackBytes, typename Callback>
+  auto ReserveAndPut(CallbackBytes aCallbackBytes, Callback&& aCallback) {
+    Maybe<ProfileBufferEntryWriter> maybeEntryWriter;
+
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+
+    if (MOZ_LIKELY(mMaybeUnderlyingBuffer)) {
+      const Length entryBytes = std::forward<CallbackBytes>(aCallbackBytes)();
+      MOZ_RELEASE_ASSERT(entryBytes > 0);
+      const Length bufferBytes =
+          mMaybeUnderlyingBuffer->mBuffer.BufferLength().Value();
+      MOZ_RELEASE_ASSERT(entryBytes <= bufferBytes - ULEB128Size(entryBytes),
+                         "Entry would wrap and overwrite itself");
+      // Compute block size from the requested entry size.
+      const Length blockBytes = ULEB128Size(entryBytes) + entryBytes;
+      // We will put this new block at the end of the current buffer.
+      const ProfileBufferIndex blockIndex =
+          mNextWriteIndex.ConvertToProfileBufferIndex();
+      // Compute the end of this new block.
+      const ProfileBufferIndex blockEnd = blockIndex + blockBytes;
+      while (blockEnd >
+             mFirstReadIndex.ConvertToProfileBufferIndex() + bufferBytes) {
+        // About to trample on an old block.
+        ProfileBufferEntryReader reader = ReaderInBlockAt(mFirstReadIndex);
+        mMaybeUnderlyingBuffer->mClearedBlockCount += 1;
+        // Move the buffer reading start past this cleared block.
+        mFirstReadIndex = ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+            mFirstReadIndex.ConvertToProfileBufferIndex() +
+            ULEB128Size(reader.RemainingBytes()) + reader.RemainingBytes());
+      }
+      // Store the new end of buffer.
+      mNextWriteIndex =
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(blockEnd);
+      mMaybeUnderlyingBuffer->mPushedBlockCount += 1;
+      // Finally, let aCallback write into the entry.
+      mMaybeUnderlyingBuffer->mBuffer.EntryWriterFromTo(maybeEntryWriter,
+                                                        blockIndex, blockEnd);
+      MOZ_ASSERT(maybeEntryWriter.isSome(),
+                 "Non-empty entry should always create an EntryWriter");
+      maybeEntryWriter->WriteULEB128(entryBytes);
+      MOZ_ASSERT(maybeEntryWriter->RemainingBytes() == entryBytes);
+    }
+
+#ifdef DEBUG
+    auto checkAllWritten = MakeScopeExit([&]() {
+      MOZ_ASSERT(!maybeEntryWriter || maybeEntryWriter->RemainingBytes() == 0);
+    });
+#endif  // DEBUG
+    return std::forward<Callback>(aCallback)(maybeEntryWriter);
+  }
+
+  // Add a new entry of known size, call `aCallback` with a pointer to a
+  // temporary ProfileBufferEntryWriter (can be null when out-of-session), and
+  // return whatever `aCallback` returns. Callback should not store the
+  // `ProfileBufferEntryWriter`, as it may become invalid after this thread-safe
+  // call.
+  template <typename Callback>
+  auto Put(Length aBytes, Callback&& aCallback) {
+    return ReserveAndPut([aBytes]() { return aBytes; },
+                         std::forward<Callback>(aCallback));
+  }
+
+  // Add a new entry copied from the given buffer, return block index.
+  ProfileBufferBlockIndex PutFrom(const void* aSrc, Length aBytes) {
+    return ReserveAndPut([aBytes]() { return aBytes; },
+                         [&](Maybe<ProfileBufferEntryWriter>& aEntryWriter) {
+                           if (MOZ_UNLIKELY(aEntryWriter.isNothing())) {
+                             // Out-of-session, return "empty" index.
+                             return ProfileBufferBlockIndex{};
+                           }
+                           aEntryWriter->WriteBytes(aSrc, aBytes);
+                           return aEntryWriter->CurrentBlockIndex();
+                         });
+  }
+
+  // Add a new single entry with *all* given object (using a Serializer for
+  // each), return block index.
+  template <typename... Ts>
+  ProfileBufferBlockIndex PutObjects(const Ts&... aTs) {
+    static_assert(sizeof...(Ts) > 0,
+                  "PutObjects must be given at least one object.");
+    return ReserveAndPut(
+        [&]() { return ProfileBufferEntryWriter::SumBytes(aTs...); },
+        [&](Maybe<ProfileBufferEntryWriter>& aEntryWriter) {
+          if (MOZ_UNLIKELY(aEntryWriter.isNothing())) {
+            // Out-of-session, return "empty" index.
+            return ProfileBufferBlockIndex{};
+          }
+          aEntryWriter->WriteObjects(aTs...);
+          return aEntryWriter->CurrentBlockIndex();
+        });
+  }
+
+  // Add a new entry copied from the given object, return block index.
+  template <typename T>
+  ProfileBufferBlockIndex PutObject(const T& aOb) {
+    return PutObjects(aOb);
+  }
+
+  // Append the contents of another BlocksRingBuffer to this one.
+  ProfileBufferBlockIndex AppendContents(const BlocksRingBuffer& aSrc) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+
+    if (MOZ_UNLIKELY(!mMaybeUnderlyingBuffer)) {
+      // We are out-of-session, could not append contents.
+      return ProfileBufferBlockIndex{};
+    }
+
+    baseprofiler::detail::BaseProfilerMaybeAutoLock srcLock(aSrc.mMutex);
+
+    if (MOZ_UNLIKELY(!aSrc.mMaybeUnderlyingBuffer)) {
+      // The other BRB is out-of-session, nothing to copy, we're done.
+      return ProfileBufferBlockIndex{};
+    }
+
+    const ProfileBufferIndex srcStartIndex =
+        aSrc.mFirstReadIndex.ConvertToProfileBufferIndex();
+    const ProfileBufferIndex srcEndIndex =
+        aSrc.mNextWriteIndex.ConvertToProfileBufferIndex();
+    const Length bytesToCopy = static_cast<Length>(srcEndIndex - srcStartIndex);
+
+    if (MOZ_UNLIKELY(bytesToCopy == 0)) {
+      // The other BRB is empty, nothing to copy, we're done.
+      return ProfileBufferBlockIndex{};
+    }
+
+    const Length bufferBytes =
+        mMaybeUnderlyingBuffer->mBuffer.BufferLength().Value();
+
+    MOZ_RELEASE_ASSERT(bytesToCopy <= bufferBytes,
+                       "Entry would wrap and overwrite itself");
+
+    // We will put all copied blocks at the end of the current buffer.
+    const ProfileBufferIndex dstStartIndex =
+        mNextWriteIndex.ConvertToProfileBufferIndex();
+    // Compute where the copy will end...
+    const ProfileBufferIndex dstEndIndex = dstStartIndex + bytesToCopy;
+
+    while (dstEndIndex >
+           mFirstReadIndex.ConvertToProfileBufferIndex() + bufferBytes) {
+      // About to trample on an old block.
+      ProfileBufferEntryReader reader = ReaderInBlockAt(mFirstReadIndex);
+      mMaybeUnderlyingBuffer->mClearedBlockCount += 1;
+      // Move the buffer reading start past this cleared block.
+      mFirstReadIndex = ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          mFirstReadIndex.ConvertToProfileBufferIndex() +
+          ULEB128Size(reader.RemainingBytes()) + reader.RemainingBytes());
+    }
+
+    // Store the new end of buffer.
+    mNextWriteIndex =
+        ProfileBufferBlockIndex::CreateFromProfileBufferIndex(dstEndIndex);
+    // Update our pushed count with the number of live blocks we are copying.
+    mMaybeUnderlyingBuffer->mPushedBlockCount +=
+        aSrc.mMaybeUnderlyingBuffer->mPushedBlockCount -
+        aSrc.mMaybeUnderlyingBuffer->mClearedBlockCount;
+
+    auto reader = aSrc.mMaybeUnderlyingBuffer->mBuffer.EntryReaderFromTo(
+        srcStartIndex, srcEndIndex, nullptr, nullptr);
+    auto writer = mMaybeUnderlyingBuffer->mBuffer.EntryWriterFromTo(
+        dstStartIndex, dstEndIndex);
+    writer.WriteFromReader(reader, bytesToCopy);
+
+    return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(dstStartIndex);
+  }
+
+  // Clear all entries: Move read index to the end so that these entries cannot
+  // be read anymore.
+  void Clear() {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    ClearAllEntries();
+  }
+
+  // Clear all entries strictly before aBlockIndex, and move read index to the
+  // end so that these entries cannot be read anymore.
+  void ClearBefore(ProfileBufferBlockIndex aBlockIndex) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    if (!mMaybeUnderlyingBuffer) {
+      return;
+    }
+    // Don't accept a not-yet-written index. One-past-the-end is ok.
+    MOZ_ASSERT(aBlockIndex <= mNextWriteIndex);
+    if (aBlockIndex <= mFirstReadIndex) {
+      // Already cleared.
+      return;
+    }
+    if (aBlockIndex == mNextWriteIndex) {
+      // Right past the end, just clear everything.
+      ClearAllEntries();
+      return;
+    }
+    // Otherwise we need to clear a subset of entries.
+    AssertBlockIndexIsValid(aBlockIndex);
+    // Just count skipped entries.
+    Reader reader(*this);
+    BlockIterator it = reader.begin();
+    for (; it.CurrentBlockIndex() < aBlockIndex; ++it) {
+      MOZ_ASSERT(it.CurrentBlockIndex() < reader.end().CurrentBlockIndex());
+      mMaybeUnderlyingBuffer->mClearedBlockCount += 1;
+    }
+    MOZ_ASSERT(it.CurrentBlockIndex() == aBlockIndex);
+    // Move read index to given index, so there's effectively no more entries
+    // before.
+    mFirstReadIndex = aBlockIndex;
+  }
+
+#ifdef DEBUG
+  void Dump() const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    if (!mMaybeUnderlyingBuffer) {
+      printf("empty BlocksRingBuffer\n");
+      return;
+    }
+    using ULL = unsigned long long;
+    printf("start=%llu (%llu) end=%llu (%llu) - ",
+           ULL(mFirstReadIndex.ConvertToProfileBufferIndex()),
+           ULL(mFirstReadIndex.ConvertToProfileBufferIndex() &
+               (mMaybeUnderlyingBuffer->mBuffer.BufferLength().Value() - 1)),
+           ULL(mNextWriteIndex.ConvertToProfileBufferIndex()),
+           ULL(mNextWriteIndex.ConvertToProfileBufferIndex() &
+               (mMaybeUnderlyingBuffer->mBuffer.BufferLength().Value() - 1)));
+    mMaybeUnderlyingBuffer->mBuffer.Dump();
+  }
+#endif  // DEBUG
+
+ private:
+  // In DEBUG mode, assert that `aBlockIndex` is a valid index for a live block.
+  // (Not just in range, but points exactly at the start of a block.)
+  // Slow, so avoid it for internal checks; this is more to check what callers
+  // provide us.
+  void AssertBlockIndexIsValid(ProfileBufferBlockIndex aBlockIndex) const {
+#ifdef DEBUG
+    mMutex.AssertCurrentThreadOwns();
+    MOZ_ASSERT(aBlockIndex >= mFirstReadIndex);
+    MOZ_ASSERT(aBlockIndex < mNextWriteIndex);
+    // Quick check (default), or slow check (change '1' to '0') below:
+#  if 1
+    // Quick check that this looks like a valid block start.
+    // Read the entry size at the start of the block.
+    const Length entryBytes = ReaderInBlockAt(aBlockIndex).RemainingBytes();
+    MOZ_ASSERT(entryBytes > 0, "Empty entries are not allowed");
+    MOZ_ASSERT(
+        entryBytes < mMaybeUnderlyingBuffer->mBuffer.BufferLength().Value() -
+                         ULEB128Size(entryBytes),
+        "Entry would wrap and overwrite itself");
+    // The end of the block should be inside the live buffer range.
+    MOZ_ASSERT(aBlockIndex.ConvertToProfileBufferIndex() +
+                   ULEB128Size(entryBytes) + entryBytes <=
+               mNextWriteIndex.ConvertToProfileBufferIndex());
+#  else
+    // Slow check that the index is really the start of the block.
+    // This kills performances, as it reads from the first index until
+    // aBlockIndex. Only use to debug issues locally.
+    Reader reader(*this);
+    BlockIterator it = reader.begin();
+    for (; it.CurrentBlockIndex() < aBlockIndex; ++it) {
+      MOZ_ASSERT(it.CurrentBlockIndex() < reader.end().CurrentBlockIndex());
+    }
+    MOZ_ASSERT(it.CurrentBlockIndex() == aBlockIndex);
+#  endif
+#endif  // DEBUG
+  }
+
+  // In DEBUG mode, assert that `aBlockIndex` is a valid index for a live block,
+  // or is just past-the-end. (Not just in range, but points exactly at the
+  // start of a block.) Slow, so avoid it for internal checks; this is more to
+  // check what callers provide us.
+  void AssertBlockIndexIsValidOrEnd(ProfileBufferBlockIndex aBlockIndex) const {
+#ifdef DEBUG
+    mMutex.AssertCurrentThreadOwns();
+    if (aBlockIndex == mNextWriteIndex) {
+      return;
+    }
+    AssertBlockIndexIsValid(aBlockIndex);
+#endif  // DEBUG
+  }
+
+  // Create a reader for the block starting at aBlockIndex.
+  ProfileBufferEntryReader ReaderInBlockAt(
+      ProfileBufferBlockIndex aBlockIndex) const {
+    mMutex.AssertCurrentThreadOwns();
+    MOZ_ASSERT(mMaybeUnderlyingBuffer.isSome());
+    MOZ_ASSERT(aBlockIndex >= mFirstReadIndex);
+    MOZ_ASSERT(aBlockIndex < mNextWriteIndex);
+    // Create a reader from the given index until the end of the buffer.
+    ProfileBufferEntryReader reader =
+        mMaybeUnderlyingBuffer->mBuffer.EntryReaderFromTo(
+            aBlockIndex.ConvertToProfileBufferIndex(),
+            mNextWriteIndex.ConvertToProfileBufferIndex(), nullptr, nullptr);
+    // Read the block size at the beginning.
+    const Length entryBytes = reader.ReadULEB128<Length>();
+    // Make sure we don't overshoot the buffer.
+    MOZ_RELEASE_ASSERT(entryBytes <= reader.RemainingBytes());
+    ProfileBufferIndex nextBlockIndex =
+        aBlockIndex.ConvertToProfileBufferIndex() + ULEB128Size(entryBytes) +
+        entryBytes;
+    // And reduce the reader to the entry area. Only provide a next-block-index
+    // if it's not at the end of the buffer (i.e., there's an actual block
+    // there).
+    reader = mMaybeUnderlyingBuffer->mBuffer.EntryReaderFromTo(
+        aBlockIndex.ConvertToProfileBufferIndex() + ULEB128Size(entryBytes),
+        nextBlockIndex, aBlockIndex,
+        (nextBlockIndex < mNextWriteIndex.ConvertToProfileBufferIndex())
+            ? ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+                  nextBlockIndex)
+            : ProfileBufferBlockIndex{});
+    return reader;
+  }
+
+  ProfileBufferEntryReader FullBufferReader() const {
+    mMutex.AssertCurrentThreadOwns();
+    if (!mMaybeUnderlyingBuffer) {
+      return {};
+    }
+    return mMaybeUnderlyingBuffer->mBuffer.EntryReaderFromTo(
+        mFirstReadIndex.ConvertToProfileBufferIndex(),
+        mNextWriteIndex.ConvertToProfileBufferIndex(), nullptr, nullptr);
+  }
+
+  // Clear all entries: Move read index to the end so that these entries cannot
+  // be read anymore.
+  void ClearAllEntries() {
+    mMutex.AssertCurrentThreadOwns();
+    if (!mMaybeUnderlyingBuffer) {
+      return;
+    }
+    // Mark all entries pushed so far as cleared.
+    mMaybeUnderlyingBuffer->mClearedBlockCount =
+        mMaybeUnderlyingBuffer->mPushedBlockCount;
+    // Move read index to write index, so there's effectively no more entries
+    // that can be read. (Not setting both to 0, in case user is keeping
+    // `ProfileBufferBlockIndex`'es to old entries.)
+    mFirstReadIndex = mNextWriteIndex;
+  }
+
+  // If there is an underlying buffer, clear all entries, and discard the
+  // buffer. This BlocksRingBuffer will now gracefully reject all API calls, and
+  // is in a state where a new underlying buffer may be set.
+  void ResetUnderlyingBuffer() {
+    mMutex.AssertCurrentThreadOwns();
+    if (!mMaybeUnderlyingBuffer) {
+      return;
+    }
+    ClearAllEntries();
+    mMaybeUnderlyingBuffer.reset();
+  }
+
+  // Used to de/serialize a BlocksRingBuffer (e.g., containing a backtrace).
+  friend ProfileBufferEntryWriter::Serializer<BlocksRingBuffer>;
+  friend ProfileBufferEntryReader::Deserializer<BlocksRingBuffer>;
+  friend ProfileBufferEntryWriter::Serializer<UniquePtr<BlocksRingBuffer>>;
+  friend ProfileBufferEntryReader::Deserializer<UniquePtr<BlocksRingBuffer>>;
+
+  // Mutex guarding the following members.
+  mutable baseprofiler::detail::BaseProfilerMaybeMutex mMutex;
+
+  struct UnderlyingBuffer {
+    // Create a buffer of the given length.
+    explicit UnderlyingBuffer(PowerOfTwo<Length> aLength) : mBuffer(aLength) {
+      MOZ_ASSERT(aLength.Value() > ULEB128MaxSize<Length>(),
+                 "Buffer should be able to contain more than a block size");
+    }
+
+    // Take ownership of an existing buffer.
+    UnderlyingBuffer(UniquePtr<Buffer::Byte[]> aExistingBuffer,
+                     PowerOfTwo<Length> aLength)
+        : mBuffer(std::move(aExistingBuffer), aLength) {
+      MOZ_ASSERT(aLength.Value() > ULEB128MaxSize<Length>(),
+                 "Buffer should be able to contain more than a block size");
+    }
+
+    // Use an externally-owned buffer.
+    UnderlyingBuffer(Buffer::Byte* aExternalBuffer, PowerOfTwo<Length> aLength)
+        : mBuffer(aExternalBuffer, aLength) {
+      MOZ_ASSERT(aLength.Value() > ULEB128MaxSize<Length>(),
+                 "Buffer should be able to contain more than a block size");
+    }
+
+    // Only allow move-construction.
+    UnderlyingBuffer(UnderlyingBuffer&&) = default;
+
+    // Copies and move-assignment are explictly disallowed.
+    UnderlyingBuffer(const UnderlyingBuffer&) = delete;
+    UnderlyingBuffer& operator=(const UnderlyingBuffer&) = delete;
+    UnderlyingBuffer& operator=(UnderlyingBuffer&&) = delete;
+
+    // Underlying circular byte buffer.
+    Buffer mBuffer;
+
+    // Statistics.
+    uint64_t mPushedBlockCount = 0;
+    uint64_t mClearedBlockCount = 0;
+  };
+
+  // Underlying buffer, with stats.
+  // Only valid during in-session period.
+  Maybe<UnderlyingBuffer> mMaybeUnderlyingBuffer;
+
+  // Index to the first block to be read (or cleared). Initialized to 1 because
+  // 0 is reserved for the "empty" ProfileBufferBlockIndex value. Kept between
+  // sessions, so that stored indices from one session will be gracefully denied
+  // in future sessions.
+  ProfileBufferBlockIndex mFirstReadIndex =
+      ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          ProfileBufferIndex(1));
+  // Index where the next new block should be allocated. Initialized to 1
+  // because 0 is reserved for the "empty" ProfileBufferBlockIndex value. Kept
+  // between sessions, so that stored indices from one session will be
+  // gracefully denied in future sessions.
+  ProfileBufferBlockIndex mNextWriteIndex =
+      ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          ProfileBufferIndex(1));
+};
+
+// ----------------------------------------------------------------------------
+// BlocksRingBuffer serialization
+
+// A BlocksRingBuffer can hide another one!
+// This will be used to store marker backtraces; They can be read back into a
+// UniquePtr<BlocksRingBuffer>.
+// Format: len (ULEB128) | start | end | buffer (len bytes) | pushed | cleared
+// len==0 marks an out-of-session buffer, or empty buffer.
+template <>
+struct ProfileBufferEntryWriter::Serializer<BlocksRingBuffer> {
+  static Length Bytes(const BlocksRingBuffer& aBuffer) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(aBuffer.mMutex);
+    if (aBuffer.mMaybeUnderlyingBuffer.isNothing()) {
+      // Out-of-session, we only need 1 byte to store a length of 0.
+      return ULEB128Size<Length>(0);
+    }
+    const auto start = aBuffer.mFirstReadIndex.ConvertToProfileBufferIndex();
+    const auto end = aBuffer.mNextWriteIndex.ConvertToProfileBufferIndex();
+    const auto len = end - start;
+    if (len == 0) {
+      // In-session but empty, also store a length of 0.
+      return ULEB128Size<Length>(0);
+    }
+    return ULEB128Size(len) + sizeof(start) + sizeof(end) + len +
+           sizeof(aBuffer.mMaybeUnderlyingBuffer->mPushedBlockCount) +
+           sizeof(aBuffer.mMaybeUnderlyingBuffer->mClearedBlockCount);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const BlocksRingBuffer& aBuffer) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(aBuffer.mMutex);
+    if (aBuffer.mMaybeUnderlyingBuffer.isNothing()) {
+      // Out-of-session, only store a length of 0.
+      aEW.WriteULEB128<Length>(0);
+      return;
+    }
+    const auto start = aBuffer.mFirstReadIndex.ConvertToProfileBufferIndex();
+    const auto end = aBuffer.mNextWriteIndex.ConvertToProfileBufferIndex();
+    MOZ_ASSERT(end - start <= std::numeric_limits<Length>::max());
+    const auto len = static_cast<Length>(end - start);
+    if (len == 0) {
+      // In-session but empty, only store a length of 0.
+      aEW.WriteULEB128<Length>(0);
+      return;
+    }
+    // In-session.
+    // Store buffer length, start and end indices.
+    aEW.WriteULEB128<Length>(len);
+    aEW.WriteObject(start);
+    aEW.WriteObject(end);
+    // Write all the bytes.
+    auto reader = aBuffer.FullBufferReader();
+    aEW.WriteFromReader(reader, reader.RemainingBytes());
+    // And write stats.
+    aEW.WriteObject(aBuffer.mMaybeUnderlyingBuffer->mPushedBlockCount);
+    aEW.WriteObject(aBuffer.mMaybeUnderlyingBuffer->mClearedBlockCount);
+  }
+};
+
+// A serialized BlocksRingBuffer can be read into an empty buffer (either
+// out-of-session, or in-session with enough room).
+template <>
+struct ProfileBufferEntryReader::Deserializer<BlocksRingBuffer> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       BlocksRingBuffer& aBuffer) {
+    // Expect an empty buffer, as we're going to overwrite it.
+    MOZ_ASSERT(aBuffer.GetState().mRangeStart == aBuffer.GetState().mRangeEnd);
+    // Read the stored buffer length.
+    const auto len = aER.ReadULEB128<Length>();
+    if (len == 0) {
+      // 0-length means an "uninteresting" buffer, just return now.
+      return;
+    }
+    // We have a non-empty buffer to read.
+    if (aBuffer.BufferLength().isSome()) {
+      // Output buffer is in-session (i.e., it already has a memory buffer
+      // attached). Make sure the caller allocated enough space.
+      MOZ_RELEASE_ASSERT(aBuffer.BufferLength()->Value() >= len);
+    } else {
+      // Output buffer is out-of-session, attach a new memory buffer.
+      aBuffer.Set(PowerOfTwo<Length>(len));
+      MOZ_ASSERT(aBuffer.BufferLength()->Value() >= len);
+    }
+    // Read start and end indices.
+    const auto start = aER.ReadObject<ProfileBufferIndex>();
+    aBuffer.mFirstReadIndex =
+        ProfileBufferBlockIndex::CreateFromProfileBufferIndex(start);
+    const auto end = aER.ReadObject<ProfileBufferIndex>();
+    aBuffer.mNextWriteIndex =
+        ProfileBufferBlockIndex::CreateFromProfileBufferIndex(end);
+    MOZ_ASSERT(end - start == len);
+    // Copy bytes into the buffer.
+    auto writer =
+        aBuffer.mMaybeUnderlyingBuffer->mBuffer.EntryWriterFromTo(start, end);
+    writer.WriteFromReader(aER, end - start);
+    MOZ_ASSERT(writer.RemainingBytes() == 0);
+    // Finally copy stats.
+    aBuffer.mMaybeUnderlyingBuffer->mPushedBlockCount = aER.ReadObject<decltype(
+        aBuffer.mMaybeUnderlyingBuffer->mPushedBlockCount)>();
+    aBuffer.mMaybeUnderlyingBuffer->mClearedBlockCount =
+        aER.ReadObject<decltype(
+            aBuffer.mMaybeUnderlyingBuffer->mClearedBlockCount)>();
+  }
+
+  // We cannot output a BlocksRingBuffer object (not copyable), use `ReadInto()`
+  // or `aER.ReadObject<UniquePtr<BlocksRinbBuffer>>()` instead.
+  static BlocksRingBuffer Read(ProfileBufferEntryReader& aER) = delete;
+};
+
+// A BlocksRingBuffer is usually refererenced through a UniquePtr, for
+// convenience we support (de)serializing that UniquePtr directly.
+// This is compatible with the non-UniquePtr serialization above, with a null
+// pointer being treated like an out-of-session or empty buffer; and any of
+// these would be deserialized into a null pointer.
+template <>
+struct ProfileBufferEntryWriter::Serializer<UniquePtr<BlocksRingBuffer>> {
+  static Length Bytes(const UniquePtr<BlocksRingBuffer>& aBufferUPtr) {
+    if (!aBufferUPtr) {
+      // Null pointer, treat it like an empty buffer, i.e., write length of 0.
+      return ULEB128Size<Length>(0);
+    }
+    // Otherwise write the pointed-at BlocksRingBuffer (which could be
+    // out-of-session or empty.)
+    return SumBytes(*aBufferUPtr);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const UniquePtr<BlocksRingBuffer>& aBufferUPtr) {
+    if (!aBufferUPtr) {
+      // Null pointer, treat it like an empty buffer, i.e., write length of 0.
+      aEW.WriteULEB128<Length>(0);
+      return;
+    }
+    // Otherwise write the pointed-at BlocksRingBuffer (which could be
+    // out-of-session or empty.)
+    aEW.WriteObject(*aBufferUPtr);
+  }
+};
+
+template <>
+struct ProfileBufferEntryReader::Deserializer<UniquePtr<BlocksRingBuffer>> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       UniquePtr<BlocksRingBuffer>& aBuffer) {
+    aBuffer = Read(aER);
+  }
+
+  static UniquePtr<BlocksRingBuffer> Read(ProfileBufferEntryReader& aER) {
+    UniquePtr<BlocksRingBuffer> bufferUPtr;
+    // Keep a copy of the reader before reading the length, so we can restart
+    // from here below.
+    ProfileBufferEntryReader readerBeforeLen = aER;
+    // Read the stored buffer length.
+    const auto len = aER.ReadULEB128<Length>();
+    if (len == 0) {
+      // 0-length means an "uninteresting" buffer, just return nullptr.
+      return bufferUPtr;
+    }
+    // We have a non-empty buffer.
+    // allocate an empty BlocksRingBuffer without mutex.
+    bufferUPtr = MakeUnique<BlocksRingBuffer>(
+        BlocksRingBuffer::ThreadSafety::WithoutMutex);
+    // Rewind the reader before the length and deserialize the contents, using
+    // the non-UniquePtr Deserializer.
+    aER = readerBeforeLen;
+    aER.ReadIntoObject(*bufferUPtr);
+    return bufferUPtr;
+  }
+};
+
+}  // namespace mozilla
+
+#endif  // BlocksRingBuffer_h
diff --git a/mozglue/baseprofiler/public/ModuloBuffer.h b/mozglue/baseprofiler/public/ModuloBuffer.h
new file mode 100644
index 0000000000..80e765279e
--- /dev/null
+++ b/mozglue/baseprofiler/public/ModuloBuffer.h
@@ -0,0 +1,618 @@
+/* -*- Mode: C++; tab-width: 2; 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 ModuloBuffer_h
+#define ModuloBuffer_h
+
+#include "mozilla/leb128iterator.h"
+#include "mozilla/Maybe.h"
+#include "mozilla/MemoryReporting.h"
+#include "mozilla/NotNull.h"
+#include "mozilla/PowerOfTwo.h"
+#include "mozilla/ProfileBufferEntrySerialization.h"
+#include "mozilla/UniquePtr.h"
+
+#include <functional>
+#include <iterator>
+#include <limits>
+#include <type_traits>
+
+namespace mozilla {
+
+// The ModuloBuffer class is a circular buffer that holds raw byte values, with
+// data-read/write helpers.
+//
+// OffsetT: Type of the internal offset into the buffer of bytes, it should be
+// large enough to access all bytes of the buffer. It will also be used as
+// Length (in bytes) of the buffer and of any subset. Default uint32_t
+// IndexT: Type of the external index, it should be large enough that overflows
+// should not happen during the lifetime of the ModuloBuffer.
+//
+// The basic usage is to create an iterator-like object with `ReaderAt(Index)`
+// or `WriterAt(Index)`, and use it to read/write data blobs. Iterators
+// automatically manage the wrap-around (through "Modulo", which is effectively
+// an AND-masking with the PowerOfTwo buffer size.)
+//
+// There is zero safety: No thread safety, no checks that iterators may be
+// overwriting data that's still to be read, etc. It's up to the caller to add
+// adequate checks.
+// The intended use is as an underlying buffer for a safer container.
+template <typename OffsetT = uint32_t, typename IndexT = uint64_t>
+class ModuloBuffer {
+ public:
+  using Byte = uint8_t;
+  static_assert(sizeof(Byte) == 1, "ModuloBuffer::Byte must be 1 byte");
+  using Offset = OffsetT;
+  static_assert(!std::numeric_limits<Offset>::is_signed,
+                "ModuloBuffer::Offset must be an unsigned integral type");
+  using Length = Offset;
+  using Index = IndexT;
+  static_assert(!std::numeric_limits<Index>::is_signed,
+                "ModuloBuffer::Index must be an unsigned integral type");
+  static_assert(sizeof(Index) >= sizeof(Offset),
+                "ModuloBuffer::Index size must >= Offset");
+
+  // Create a buffer of the given length.
+  explicit ModuloBuffer(PowerOfTwo<Length> aLength)
+      : mMask(aLength.Mask()),
+        mBuffer(WrapNotNull(new Byte[aLength.Value()])),
+        mBufferDeleter([](Byte* aBuffer) { delete[] aBuffer; }) {}
+
+  // Take ownership of an existing buffer. Existing contents is ignored.
+  // Done by extracting the raw pointer from UniquePtr<Byte[]>, and adding
+  // an equivalent `delete[]` in `mBufferDeleter`.
+  ModuloBuffer(UniquePtr<Byte[]> aExistingBuffer, PowerOfTwo<Length> aLength)
+      : mMask(aLength.Mask()),
+        mBuffer(WrapNotNull(aExistingBuffer.release())),
+        mBufferDeleter([](Byte* aBuffer) { delete[] aBuffer; }) {}
+
+  // Use an externally-owned buffer. Existing contents is ignored.
+  ModuloBuffer(Byte* aExternalBuffer, PowerOfTwo<Length> aLength)
+      : mMask(aLength.Mask()), mBuffer(WrapNotNull(aExternalBuffer)) {}
+
+  // Disallow copying, as we may uniquely own the resource.
+  ModuloBuffer(const ModuloBuffer& aOther) = delete;
+  ModuloBuffer& operator=(const ModuloBuffer& aOther) = delete;
+
+  // Allow move-construction. Stealing ownership if the original had it.
+  // This effectively prevents copy construction, and all assignments; needed so
+  // that a ModuloBuffer may be initialized from a separate construction.
+  // The moved-from ModuloBuffer still points at the resource but doesn't own
+  // it, so it won't try to free it; but accesses are not guaranteed, so it
+  // should not be used anymore.
+  ModuloBuffer(ModuloBuffer&& aOther)
+      : mMask(std::move(aOther.mMask)),
+        mBuffer(std::move(aOther.mBuffer)),
+        mBufferDeleter(std::move(aOther.mBufferDeleter)) {
+    // The above move leaves `aOther.mBufferDeleter` in a valid state but with
+    // an unspecified value, so it could theoretically still contain the
+    // original function, which would be bad because we don't want aOther to
+    // delete the resource that `this` now owns.
+    if (aOther.mBufferDeleter) {
+      // `aOther` still had a non-empty deleter, reset it.
+      aOther.mBufferDeleter = nullptr;
+    }
+  }
+
+  // Disallow assignment, as we have some `const` members.
+  ModuloBuffer& operator=(ModuloBuffer&& aOther) = delete;
+
+  // Destructor, deletes the resource if we uniquely own it.
+  ~ModuloBuffer() {
+    if (mBufferDeleter) {
+      mBufferDeleter(mBuffer);
+    }
+  }
+
+  PowerOfTwo<Length> BufferLength() const {
+    return PowerOfTwo<Length>(mMask.MaskValue() + 1);
+  }
+
+  // Size of external resources.
+  // Note: `mBufferDeleter`'s potential external data (for its captures) is not
+  // included, as it's hidden in the `std::function` implementation.
+  size_t SizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const {
+    if (!mBufferDeleter) {
+      // If we don't have a buffer deleter, assume we don't own the data, so
+      // it's probably on the stack, or should be reported by its owner.
+      return 0;
+    }
+    return aMallocSizeOf(mBuffer);
+  }
+
+  size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+    return aMallocSizeOf(this) + SizeOfExcludingThis(aMallocSizeOf);
+  }
+
+  ProfileBufferEntryReader EntryReaderFromTo(
+      Index aStart, Index aEnd, ProfileBufferBlockIndex aBlockIndex,
+      ProfileBufferBlockIndex aNextBlockIndex) const {
+    using EntrySpan = Span<const ProfileBufferEntryReader::Byte>;
+    if (aStart == aEnd) {
+      return ProfileBufferEntryReader{};
+    }
+    // Don't allow over-wrapping.
+    MOZ_ASSERT(aEnd - aStart <= mMask.MaskValue() + 1);
+    // Start offset in 0 .. (buffer size - 1)
+    Offset start = static_cast<Offset>(aStart) & mMask;
+    // End offset in 1 .. (buffer size)
+    Offset end = (static_cast<Offset>(aEnd - 1) & mMask) + 1;
+    if (start < end) {
+      // Segment doesn't cross buffer threshold, one span is enough.
+      return ProfileBufferEntryReader{EntrySpan(&mBuffer[start], end - start),
+                                      aBlockIndex, aNextBlockIndex};
+    }
+    // Segment crosses buffer threshold, we need one span until the end and one
+    // span restarting at the beginning of the buffer.
+    return ProfileBufferEntryReader{
+        EntrySpan(&mBuffer[start], mMask.MaskValue() + 1 - start),
+        EntrySpan(&mBuffer[0], end), aBlockIndex, aNextBlockIndex};
+  }
+
+  // Return an entry writer for the given range.
+  ProfileBufferEntryWriter EntryWriterFromTo(Index aStart, Index aEnd) const {
+    using EntrySpan = Span<ProfileBufferEntryReader::Byte>;
+    if (aStart == aEnd) {
+      return ProfileBufferEntryWriter{};
+    }
+    MOZ_ASSERT(aEnd - aStart <= mMask.MaskValue() + 1);
+    // Start offset in 0 .. (buffer size - 1)
+    Offset start = static_cast<Offset>(aStart) & mMask;
+    // End offset in 1 .. (buffer size)
+    Offset end = (static_cast<Offset>(aEnd - 1) & mMask) + 1;
+    if (start < end) {
+      // Segment doesn't cross buffer threshold, one span is enough.
+      return ProfileBufferEntryWriter{
+          EntrySpan(&mBuffer[start], end - start),
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(aStart),
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(aEnd)};
+    }
+    // Segment crosses buffer threshold, we need one span until the end and one
+    // span restarting at the beginning of the buffer.
+    return ProfileBufferEntryWriter{
+        EntrySpan(&mBuffer[start], mMask.MaskValue() + 1 - start),
+        EntrySpan(&mBuffer[0], end),
+        ProfileBufferBlockIndex::CreateFromProfileBufferIndex(aStart),
+        ProfileBufferBlockIndex::CreateFromProfileBufferIndex(aEnd)};
+  }
+
+  // Emplace an entry writer into `aMaybeEntryWriter` for the given range.
+  void EntryWriterFromTo(Maybe<ProfileBufferEntryWriter>& aMaybeEntryWriter,
+                         Index aStart, Index aEnd) const {
+    MOZ_ASSERT(aMaybeEntryWriter.isNothing(),
+               "Reference entry writer should be Nothing.");
+    using EntrySpan = Span<ProfileBufferEntryReader::Byte>;
+    if (aStart == aEnd) {
+      return;
+    }
+    MOZ_ASSERT(aEnd - aStart <= mMask.MaskValue() + 1);
+    // Start offset in 0 .. (buffer size - 1)
+    Offset start = static_cast<Offset>(aStart) & mMask;
+    // End offset in 1 .. (buffer size)
+    Offset end = (static_cast<Offset>(aEnd - 1) & mMask) + 1;
+    if (start < end) {
+      // Segment doesn't cross buffer threshold, one span is enough.
+      aMaybeEntryWriter.emplace(
+          EntrySpan(&mBuffer[start], end - start),
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(aStart),
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(aEnd));
+    } else {
+      // Segment crosses buffer threshold, we need one span until the end and
+      // one span restarting at the beginning of the buffer.
+      aMaybeEntryWriter.emplace(
+          EntrySpan(&mBuffer[start], mMask.MaskValue() + 1 - start),
+          EntrySpan(&mBuffer[0], end),
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(aStart),
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(aEnd));
+    }
+  }
+
+  // All ModuloBuffer operations should be done through this iterator, which has
+  // an effectively infinite range. The underlying wrapping-around is hidden.
+  // Use `ReaderAt(Index)` or `WriterAt(Index)` to create it.
+  //
+  // `const Iterator<...>` means the iterator itself cannot change, i.e., it
+  // cannot move, and only its const methods are available. Note that these
+  // const methods may still be used to modify the buffer contents (e.g.:
+  // `operator*()`, `Poke()`).
+  //
+  // `Iterator</*IsBufferConst=*/true>` means the buffer contents cannot be
+  // modified, i.e., write operations are forbidden, but the iterator may still
+  // move if non-const itself.
+  template <bool IsBufferConst>
+  class Iterator {
+    // Alias to const- or mutable-`ModuloBuffer` depending on `IsBufferConst`.
+    using ConstOrMutableBuffer =
+        std::conditional_t<IsBufferConst, const ModuloBuffer, ModuloBuffer>;
+
+    // Implementation note about the strange enable-if's below:
+    //   `template <bool NotIBC = !IsBufferConst> enable_if_t<NotIBC>`
+    // which intuitively could be simplified to:
+    //   `enable_if_t<!IsBufferConst>`
+    // The former extra-templated syntax is in fact necessary to delay
+    // instantiation of these functions until they are actually needed.
+    //
+    // If we were just doing `enable_if_t<!IsBufferConst>`, this would only
+    // depend on the *class* (`ModuloBuffer<...>::Iterator`), which gets
+    // instantiated when a `ModuloBuffer` is created with some template
+    // arguments; at that point, all non-templated methods get instantiated, so
+    // there's no "SFINAE" happening, and `enable_if_t<...>` is actually doing
+    // `typename enable_if<...>::type` on the spot, but there is no `type` if
+    // `IsBufferConst` is true, so it just fails right away. E.g.:
+    // error: no type named 'type' in 'std::enable_if<false, void>';
+    //        'enable_if' cannot be used to disable this declaration
+    // note: in instantiation of template type alias 'enable_if_t'
+    // > std::enable_if_t<!IsBufferConst> WriteObject(const T& aObject) {
+    //       in instantiation of template class
+    //       'mozilla::ModuloBuffer<...>::Iterator<true>'
+    // > auto it = mb.ReaderAt(1);
+    //
+    // By adding another template level `template <bool NotIsBufferConst =
+    // !IsBufferConst>`, the instantiation is delayed until the function is
+    // actually invoked somewhere, e.g. `it.Poke(...);`.
+    // So at that invocation point, the compiler looks for a "Poke" name in it,
+    // and considers potential template instantiations that could work. The
+    // `enable_if_t` is *now* attempted, with `NotIsBufferConst` taking its
+    // value from `!IsBufferConst`:
+    // - If `IsBufferConst` is false, `NotIsBufferConst` is true,
+    // `enable_if<NotIsBufferConst>` does define a `type` (`void` by default),
+    // so `enable_if_t` happily becomes `void`, the function exists and may be
+    // called.
+    // - Otherwise if `IsBufferConst` is true, `NotIsBufferConst` is false,
+    // `enable_if<NotIsBufferConst>` does *not* define a `type`, therefore
+    // `enable_if_t` produces an error because there is no `type`. Now "SFINAE"
+    // happens: This "Substitution Failure Is Not An Error" (by itself)... But
+    // then, there are no other functions named "Poke" as requested in the
+    // `it.Poke(...);` call, so we are now getting an error (can't find
+    // function), as expected because `it` had `IsBufferConst`==true. (But at
+    // least the compiler waited until this invocation attempt before outputting
+    // an error.)
+    //
+    // C++ is fun!
+
+   public:
+    // These definitions are expected by std functions, to recognize this as an
+    // iterator. See https://en.cppreference.com/w/cpp/iterator/iterator_traits
+    using difference_type = Index;
+    using value_type = Byte;
+    using pointer = std::conditional_t<IsBufferConst, const Byte*, Byte*>;
+    using reference = std::conditional_t<IsBufferConst, const Byte&, Byte&>;
+    using iterator_category = std::random_access_iterator_tag;
+
+    // Can always copy/assign from the same kind of iterator.
+    Iterator(const Iterator& aRhs) = default;
+    Iterator& operator=(const Iterator& aRhs) = default;
+
+    // Can implicitly copy an Iterator-to-mutable (reader+writer) to
+    // Iterator-to-const (reader-only), but not the reverse.
+    template <bool IsRhsBufferConst,
+              typename = std::enable_if_t<(!IsRhsBufferConst) && IsBufferConst>>
+    MOZ_IMPLICIT Iterator(const Iterator<IsRhsBufferConst>& aRhs)
+        : mModuloBuffer(aRhs.mModuloBuffer), mIndex(aRhs.mIndex) {}
+
+    // Can implicitly assign from an Iterator-to-mutable (reader+writer) to
+    // Iterator-to-const (reader-only), but not the reverse.
+    template <bool IsRhsBufferConst,
+              typename = std::enable_if_t<(!IsRhsBufferConst) && IsBufferConst>>
+    Iterator& operator=(const Iterator<IsRhsBufferConst>& aRhs) {
+      mModuloBuffer = aRhs.mModuloBuffer;
+      mIndex = aRhs.mIndex;
+      return *this;
+    }
+
+    // Current location of the iterator in the `Index` range.
+    // Note that due to wrapping, multiple indices may effectively point at the
+    // same byte in the buffer.
+    Index CurrentIndex() const { return mIndex; }
+
+    // Location comparison in the `Index` range. I.e., two `Iterator`s may look
+    // unequal, but refer to the same buffer location.
+    // Must be on the same buffer.
+    bool operator==(const Iterator& aRhs) const {
+      MOZ_ASSERT(mModuloBuffer == aRhs.mModuloBuffer);
+      return mIndex == aRhs.mIndex;
+    }
+    bool operator!=(const Iterator& aRhs) const {
+      MOZ_ASSERT(mModuloBuffer == aRhs.mModuloBuffer);
+      return mIndex != aRhs.mIndex;
+    }
+    bool operator<(const Iterator& aRhs) const {
+      MOZ_ASSERT(mModuloBuffer == aRhs.mModuloBuffer);
+      return mIndex < aRhs.mIndex;
+    }
+    bool operator<=(const Iterator& aRhs) const {
+      MOZ_ASSERT(mModuloBuffer == aRhs.mModuloBuffer);
+      return mIndex <= aRhs.mIndex;
+    }
+    bool operator>(const Iterator& aRhs) const {
+      MOZ_ASSERT(mModuloBuffer == aRhs.mModuloBuffer);
+      return mIndex > aRhs.mIndex;
+    }
+    bool operator>=(const Iterator& aRhs) const {
+      MOZ_ASSERT(mModuloBuffer == aRhs.mModuloBuffer);
+      return mIndex >= aRhs.mIndex;
+    }
+
+    // Movement in the `Index` range.
+    Iterator& operator++() {
+      ++mIndex;
+      return *this;
+    }
+    Iterator operator++(int) {
+      Iterator here(*mModuloBuffer, mIndex);
+      ++mIndex;
+      return here;
+    }
+    Iterator& operator--() {
+      --mIndex;
+      return *this;
+    }
+    Iterator operator--(int) {
+      Iterator here(*mModuloBuffer, mIndex);
+      --mIndex;
+      return here;
+    }
+    Iterator& operator+=(Length aLength) {
+      mIndex += aLength;
+      return *this;
+    }
+    Iterator operator+(Length aLength) const {
+      return Iterator(*mModuloBuffer, mIndex + aLength);
+    }
+    friend Iterator operator+(Length aLength, const Iterator& aIt) {
+      return aIt + aLength;
+    }
+    Iterator& operator-=(Length aLength) {
+      mIndex -= aLength;
+      return *this;
+    }
+    Iterator operator-(Length aLength) const {
+      return Iterator(*mModuloBuffer, mIndex - aLength);
+    }
+
+    // Distance from `aRef` to here in the `Index` range.
+    // May be negative (as 2's complement) if `aRef > *this`.
+    Index operator-(const Iterator& aRef) const {
+      MOZ_ASSERT(mModuloBuffer == aRef.mModuloBuffer);
+      return mIndex - aRef.mIndex;
+    }
+
+    // Dereference a single byte (read-only if `IsBufferConst` is true).
+    reference operator*() const {
+      return mModuloBuffer->mBuffer[OffsetInBuffer()];
+    }
+
+    // Random-access dereference.
+    reference operator[](Length aLength) const { return *(*this + aLength); }
+
+    // Write data (if `IsBufferConst` is false) but don't move iterator.
+    template <bool NotIsBufferConst = !IsBufferConst>
+    std::enable_if_t<NotIsBufferConst> Poke(const void* aSrc,
+                                            Length aLength) const {
+      // Don't allow data larger than the buffer.
+      MOZ_ASSERT(aLength <= mModuloBuffer->BufferLength().Value());
+      // Offset inside the buffer (corresponding to our Index).
+      Offset offset = OffsetInBuffer();
+      // Compute remaining bytes between this offset and the end of the buffer.
+      Length remaining = mModuloBuffer->BufferLength().Value() - offset;
+      if (MOZ_LIKELY(remaining >= aLength)) {
+        // Enough space to write everything before the end.
+        memcpy(&mModuloBuffer->mBuffer[offset], aSrc, aLength);
+      } else {
+        // Not enough space. Write as much as possible before the end.
+        memcpy(&mModuloBuffer->mBuffer[offset], aSrc, remaining);
+        // And then continue from the beginning of the buffer.
+        memcpy(&mModuloBuffer->mBuffer[0],
+               static_cast<const Byte*>(aSrc) + remaining,
+               (aLength - remaining));
+      }
+    }
+
+    // Write object data (if `IsBufferConst` is false) but don't move iterator.
+    // Note that this copies bytes from the object, with the intent to read them
+    // back later. Restricted to trivially-copyable types, which support this
+    // without Undefined Behavior!
+    template <typename T, bool NotIsBufferConst = !IsBufferConst>
+    std::enable_if_t<NotIsBufferConst> PokeObject(const T& aObject) const {
+      static_assert(std::is_trivially_copyable<T>::value,
+                    "PokeObject<T> - T must be trivially copyable");
+      return Poke(&aObject, sizeof(T));
+    }
+
+    // Write data (if `IsBufferConst` is false) and move iterator ahead.
+    template <bool NotIsBufferConst = !IsBufferConst>
+    std::enable_if_t<NotIsBufferConst> Write(const void* aSrc, Length aLength) {
+      Poke(aSrc, aLength);
+      mIndex += aLength;
+    }
+
+    // Write object data (if `IsBufferConst` is false) and move iterator ahead.
+    // Note that this copies bytes from the object, with the intent to read them
+    // back later. Restricted to trivially-copyable types, which support this
+    // without Undefined Behavior!
+    template <typename T, bool NotIsBufferConst = !IsBufferConst>
+    std::enable_if_t<NotIsBufferConst> WriteObject(const T& aObject) {
+      static_assert(std::is_trivially_copyable<T>::value,
+                    "WriteObject<T> - T must be trivially copyable");
+      return Write(&aObject, sizeof(T));
+    }
+
+    // Number of bytes needed to represent `aValue` in unsigned LEB128.
+    template <typename T>
+    static unsigned ULEB128Size(T aValue) {
+      return ::mozilla::ULEB128Size(aValue);
+    }
+
+    // Write number as unsigned LEB128 (if `IsBufferConst` is false) and move
+    // iterator ahead.
+    template <typename T, bool NotIsBufferConst = !IsBufferConst>
+    std::enable_if_t<NotIsBufferConst> WriteULEB128(T aValue) {
+      ::mozilla::WriteULEB128(aValue, *this);
+    }
+
+    // Read data but don't move iterator.
+    void Peek(void* aDst, Length aLength) const {
+      // Don't allow data larger than the buffer.
+      MOZ_ASSERT(aLength <= mModuloBuffer->BufferLength().Value());
+      // Offset inside the buffer (corresponding to our Index).
+      Offset offset = OffsetInBuffer();
+      // Compute remaining bytes between this offset and the end of the buffer.
+      Length remaining = mModuloBuffer->BufferLength().Value() - offset;
+      if (MOZ_LIKELY(remaining >= aLength)) {
+        // Can read everything we need before the end of the buffer.
+        memcpy(aDst, &mModuloBuffer->mBuffer[offset], aLength);
+      } else {
+        // Read as much as possible before the end of the buffer.
+        memcpy(aDst, &mModuloBuffer->mBuffer[offset], remaining);
+        // And then continue from the beginning of the buffer.
+        memcpy(static_cast<Byte*>(aDst) + remaining, &mModuloBuffer->mBuffer[0],
+               (aLength - remaining));
+      }
+    }
+
+    // Read data into an object but don't move iterator.
+    // Note that this overwrites `aObject` with bytes from the buffer.
+    // Restricted to trivially-copyable types, which support this without
+    // Undefined Behavior!
+    template <typename T>
+    void PeekIntoObject(T& aObject) const {
+      static_assert(std::is_trivially_copyable<T>::value,
+                    "PeekIntoObject<T> - T must be trivially copyable");
+      Peek(&aObject, sizeof(T));
+    }
+
+    // Read data as an object but don't move iterator.
+    // Note that this creates an default `T` first, and then overwrites it with
+    // bytes from the buffer. Restricted to trivially-copyable types, which
+    // support this without Undefined Behavior!
+    template <typename T>
+    T PeekObject() const {
+      static_assert(std::is_trivially_copyable<T>::value,
+                    "PeekObject<T> - T must be trivially copyable");
+      T object;
+      PeekIntoObject(object);
+      return object;
+    }
+
+    // Read data and move iterator ahead.
+    void Read(void* aDst, Length aLength) {
+      Peek(aDst, aLength);
+      mIndex += aLength;
+    }
+
+    // Read data into a mutable iterator and move both iterators ahead.
+    void ReadInto(Iterator</* IsBufferConst */ false>& aDst, Length aLength) {
+      // Don't allow data larger than the buffer.
+      MOZ_ASSERT(aLength <= mModuloBuffer->BufferLength().Value());
+      MOZ_ASSERT(aLength <= aDst.mModuloBuffer->BufferLength().Value());
+      // Offset inside the buffer (corresponding to our Index).
+      Offset offset = OffsetInBuffer();
+      // Compute remaining bytes between this offset and the end of the buffer.
+      Length remaining = mModuloBuffer->BufferLength().Value() - offset;
+      if (MOZ_LIKELY(remaining >= aLength)) {
+        // Can read everything we need before the end of the buffer.
+        aDst.Write(&mModuloBuffer->mBuffer[offset], aLength);
+      } else {
+        // Read as much as possible before the end of the buffer.
+        aDst.Write(&mModuloBuffer->mBuffer[offset], remaining);
+        // And then continue from the beginning of the buffer.
+        aDst.Write(&mModuloBuffer->mBuffer[0], (aLength - remaining));
+      }
+      mIndex += aLength;
+    }
+
+    // Read data into an object and move iterator ahead.
+    // Note that this overwrites `aObject` with bytes from the buffer.
+    // Restricted to trivially-copyable types, which support this without
+    // Undefined Behavior!
+    template <typename T>
+    void ReadIntoObject(T& aObject) {
+      static_assert(std::is_trivially_copyable<T>::value,
+                    "ReadIntoObject<T> - T must be trivially copyable");
+      Read(&aObject, sizeof(T));
+    }
+
+    // Read data as an object and move iterator ahead.
+    // Note that this creates an default `T` first, and then overwrites it with
+    // bytes from the buffer. Restricted to trivially-copyable types, which
+    // support this without Undefined Behavior!
+    template <typename T>
+    T ReadObject() {
+      static_assert(std::is_trivially_copyable<T>::value,
+                    "ReadObject<T> - T must be trivially copyable");
+      T object;
+      ReadIntoObject(object);
+      return object;
+    }
+
+    // Read an unsigned LEB128 number and move iterator ahead.
+    template <typename T>
+    T ReadULEB128() {
+      return ::mozilla::ReadULEB128<T>(*this);
+    }
+
+   private:
+    // Only a ModuloBuffer can instantiate its iterator.
+    friend class ModuloBuffer;
+
+    Iterator(ConstOrMutableBuffer& aBuffer, Index aIndex)
+        : mModuloBuffer(WrapNotNull(&aBuffer)), mIndex(aIndex) {}
+
+    // Convert the Iterator's mIndex into an offset inside the byte buffer.
+    Offset OffsetInBuffer() const {
+      return static_cast<Offset>(mIndex) & mModuloBuffer->mMask;
+    }
+
+    // ModuloBuffer that this Iterator operates on.
+    // Using a non-null pointer instead of a reference, to allow re-assignment
+    // of an Iterator variable.
+    NotNull<ConstOrMutableBuffer*> mModuloBuffer;
+
+    // Position of this iterator in the wider `Index` range. (Will be wrapped
+    // around as needed when actually accessing bytes from the buffer.)
+    Index mIndex;
+  };
+
+  // Shortcut to iterator to const (read-only) data.
+  using Reader = Iterator<true>;
+  // Shortcut to iterator to non-const (read/write) data.
+  using Writer = Iterator<false>;
+
+  // Create an iterator to const data at the given index.
+  Reader ReaderAt(Index aIndex) const { return Reader(*this, aIndex); }
+
+  // Create an iterator to non-const data at the given index.
+  Writer WriterAt(Index aIndex) { return Writer(*this, aIndex); }
+
+#ifdef DEBUG
+  void Dump() const {
+    Length len = BufferLength().Value();
+    if (len > 128) {
+      len = 128;
+    }
+    for (Length i = 0; i < len; ++i) {
+      printf("%02x ", mBuffer[i]);
+    }
+    printf("\n");
+  }
+#endif  // DEBUG
+
+ private:
+  // Mask used to convert an index to an offset in `mBuffer`
+  const PowerOfTwoMask<Offset> mMask;
+
+  // Buffer data. `const NotNull<...>` shows that `mBuffer is `const`, and
+  // `Byte* const` shows that the pointer cannot be changed to point at
+  // something else, but the pointed-at `Byte`s are writable.
+  const NotNull<Byte* const> mBuffer;
+
+  // Function used to release the buffer resource (if needed).
+  std::function<void(Byte*)> mBufferDeleter;
+};
+
+}  // namespace mozilla
+
+#endif  // ModuloBuffer_h
diff --git a/mozglue/baseprofiler/public/PowerOfTwo.h b/mozglue/baseprofiler/public/PowerOfTwo.h
new file mode 100644
index 0000000000..7d396c15e6
--- /dev/null
+++ b/mozglue/baseprofiler/public/PowerOfTwo.h
@@ -0,0 +1,322 @@
+/* -*- Mode: C++; tab-width: 2; 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/. */
+
+// PowerOfTwo is a value type that always hold a power of 2.
+// It has the same size as their underlying unsigned type, but offer the
+// guarantee of being a power of 2, which permits some optimizations when
+// involved in modulo operations (using masking instead of actual modulo).
+//
+// PowerOfTwoMask contains a mask corresponding to a power of 2.
+// E.g., 2^8 is 256 or 0x100, the corresponding mask is 2^8-1 or 255 or 0xFF.
+// It should be used instead of PowerOfTwo in situations where most operations
+// would be modulo, this saves having to recompute the mask from the stored
+// power of 2.
+//
+// One common use would be for ring-buffer containers with a power-of-2 size,
+// where an index is usually converted to an in-buffer offset by `i % size`.
+// Instead, the container could store a PowerOfTwo or PowerOfTwoMask, and do
+// `i % p2` or `i & p2m`, which is more efficient than for arbitrary sizes.
+//
+// Shortcuts for common 32- and 64-bit values: PowerOfTwo32, etc.
+//
+// To create constexpr constants, use MakePowerOfTwo<Type, Value>(), etc.
+
+#ifndef PowerOfTwo_h
+#define PowerOfTwo_h
+
+#include "mozilla/MathAlgorithms.h"
+
+#include <limits>
+
+namespace mozilla {
+
+// Compute the smallest power of 2 greater than or equal to aInput, except if
+// that would overflow in which case the highest possible power of 2 if chosen.
+// 0->1, 1->1, 2->2, 3->4, ... 2^31->2^31, 2^31+1->2^31 (for uint32_t), etc.
+template <typename T>
+T FriendlyRoundUpPow2(T aInput) {
+  // This is the same code as `RoundUpPow2()`, except we handle any type (that
+  // CeilingLog2 supports) and allow the greater-than-max-power case.
+  constexpr T max = T(1) << (sizeof(T) * CHAR_BIT - 1);
+  if (aInput >= max) {
+    return max;
+  }
+  return T(1) << CeilingLog2(aInput);
+}
+
+namespace detail {
+// Same function name `CountLeadingZeroes` with uint32_t and uint64_t overloads.
+inline uint_fast8_t CountLeadingZeroes(uint32_t aValue) {
+  MOZ_ASSERT(aValue != 0);
+  return detail::CountLeadingZeroes32(aValue);
+}
+inline uint_fast8_t CountLeadingZeroes(uint64_t aValue) {
+  MOZ_ASSERT(aValue != 0);
+  return detail::CountLeadingZeroes64(aValue);
+}
+// Refuse anything else.
+template <typename T>
+inline uint_fast8_t CountLeadingZeroes(T aValue) = delete;
+}  // namespace detail
+
+// Compute the smallest 2^N-1 mask where aInput can fit.
+// I.e., `x & mask == x`, but `x & (mask >> 1) != x`.
+// Or looking at binary, we want a mask with as many leading zeroes as the
+// input, by right-shifting a full mask: (8-bit examples)
+// input:          00000000    00000001   00000010  00010110  01111111 10000000
+// N leading 0s:   ^^^^^^^^ 8  ^^^^^^^ 7  ^^^^^^ 6  ^^^ 3     ^ 1      0
+// full mask:      11111111    11111111   11111111  11111111  11111111 11111111
+// full mask >> N: 00000000    00000001   00000011  00011111  01111111 11111111
+template <typename T>
+T RoundUpPow2Mask(T aInput) {
+  // Special case, as CountLeadingZeroes(0) is undefined. (And even if that was
+  // defined, shifting by the full type size is also undefined!)
+  if (aInput == 0) {
+    return 0;
+  }
+  return T(-1) >> detail::CountLeadingZeroes(aInput);
+}
+
+template <typename T>
+class PowerOfTwoMask;
+
+template <typename T, T Mask>
+constexpr PowerOfTwoMask<T> MakePowerOfTwoMask();
+
+template <typename T>
+class PowerOfTwo;
+
+template <typename T, T Value>
+constexpr PowerOfTwo<T> MakePowerOfTwo();
+
+// PowerOfTwoMask will always contain a mask for a power of 2, which is useful
+// for power-of-2 modulo operations (e.g., to keep an index inside a power-of-2
+// container).
+// Use this instead of PowerOfTwo if masking is the primary use of the value.
+//
+// Note that this class can store a "full" mask where all bits are set, so it
+// works for mask corresponding to the power of 2 that would overflow `T`
+// (e.g., 2^32 for uint32_t gives a mask of 2^32-1, which fits in a uint32_t).
+// For this reason there is no API that computes the power of 2 corresponding to
+// the mask; But this can be done explicitly with `MaskValue() + 1`, which may
+// be useful for computing things like distance-to-the-end by doing
+// `MaskValue() + 1 - offset`, which works fine with unsigned number types.
+template <typename T>
+class PowerOfTwoMask {
+  static_assert(!std::numeric_limits<T>::is_signed,
+                "PowerOfTwoMask must use an unsigned type");
+
+ public:
+  // Construct a power of 2 mask where the given value can fit.
+  // Cannot be constexpr because of `RoundUpPow2Mask()`.
+  explicit PowerOfTwoMask(T aInput) : mMask(RoundUpPow2Mask(aInput)) {}
+
+  // Compute the mask corresponding to a PowerOfTwo.
+  // This saves having to compute the nearest 2^N-1.
+  // Not a conversion constructor, as that could be ambiguous whether we'd want
+  // the mask corresponding to the power of 2 (2^N -> 2^N-1), or the mask that
+  // can *contain* the PowerOfTwo value (2^N -> 2^(N+1)-1).
+  // Note: Not offering reverse PowerOfTwoMark-to-PowerOfTwo conversion, because
+  // that could result in an unexpected 0 result for the largest possible mask.
+  template <typename U>
+  static constexpr PowerOfTwoMask<U> MaskForPowerOfTwo(
+      const PowerOfTwo<U>& aP2) {
+    return PowerOfTwoMask(aP2);
+  }
+
+  // Allow smaller unsigned types as input.
+  // Bigger or signed types must be explicitly converted by the caller.
+  template <typename U>
+  explicit constexpr PowerOfTwoMask(U aInput)
+      : mMask(RoundUpPow2Mask(static_cast<T>(aInput))) {
+    static_assert(!std::numeric_limits<T>::is_signed,
+                  "PowerOfTwoMask does not accept signed types");
+    static_assert(sizeof(U) <= sizeof(T),
+                  "PowerOfTwoMask does not accept bigger types");
+  }
+
+  constexpr T MaskValue() const { return mMask; }
+
+  // `x & aPowerOfTwoMask` just works.
+  template <typename U>
+  friend U operator&(U aNumber, PowerOfTwoMask aP2M) {
+    return static_cast<U>(aNumber & aP2M.MaskValue());
+  }
+
+  // `aPowerOfTwoMask & x` just works.
+  template <typename U>
+  friend constexpr U operator&(PowerOfTwoMask aP2M, U aNumber) {
+    return static_cast<U>(aP2M.MaskValue() & aNumber);
+  }
+
+  // `x % aPowerOfTwoMask(2^N-1)` is equivalent to `x % 2^N` but is more
+  // optimal by doing `x & (2^N-1)`.
+  // Useful for templated code doing modulo with a template argument type.
+  template <typename U>
+  friend constexpr U operator%(U aNumerator, PowerOfTwoMask aDenominator) {
+    return aNumerator & aDenominator.MaskValue();
+  }
+
+  constexpr bool operator==(const PowerOfTwoMask& aRhs) const {
+    return mMask == aRhs.mMask;
+  }
+  constexpr bool operator!=(const PowerOfTwoMask& aRhs) const {
+    return mMask != aRhs.mMask;
+  }
+
+ private:
+  // Trust `PowerOfTwo` to call the private Trusted constructor below.
+  friend class PowerOfTwo<T>;
+
+  // Trust `MakePowerOfTwoMask()` to call the private Trusted constructor below.
+  template <typename U, U Mask>
+  friend constexpr PowerOfTwoMask<U> MakePowerOfTwoMask();
+
+  struct Trusted {
+    T mMask;
+  };
+  // Construct the mask corresponding to a PowerOfTwo.
+  // This saves having to compute the nearest 2^N-1.
+  // Note: Not a public PowerOfTwo->PowerOfTwoMask conversion constructor, as
+  // that could be ambiguous whether we'd want the mask corresponding to the
+  // power of 2 (2^N -> 2^N-1), or the mask that can *contain* the PowerOfTwo
+  // value (2^N -> 2^(N+1)-1).
+  explicit constexpr PowerOfTwoMask(const Trusted& aP2) : mMask(aP2.mMask) {}
+
+  T mMask = 0;
+};
+
+// Make a PowerOfTwoMask constant, statically-checked.
+template <typename T, T Mask>
+constexpr PowerOfTwoMask<T> MakePowerOfTwoMask() {
+  static_assert(Mask == T(-1) || IsPowerOfTwo(Mask + 1),
+                "MakePowerOfTwoMask<T, Mask>: Mask must be 2^N-1");
+  using Trusted = typename PowerOfTwoMask<T>::Trusted;
+  return PowerOfTwoMask<T>(Trusted{Mask});
+}
+
+// PowerOfTwo will always contain a power of 2.
+template <typename T>
+class PowerOfTwo {
+  static_assert(!std::numeric_limits<T>::is_signed,
+                "PowerOfTwo must use an unsigned type");
+
+ public:
+  // Construct a power of 2 that can fit the given value, or the highest power
+  // of 2 possible.
+  // Caller should explicitly check/assert `Value() <= aInput` if they want to.
+  // Cannot be constexpr because of `FriendlyRoundUpPow2()`.
+  explicit PowerOfTwo(T aInput) : mValue(FriendlyRoundUpPow2(aInput)) {}
+
+  // Allow smaller unsigned types as input.
+  // Bigger or signed types must be explicitly converted by the caller.
+  template <typename U>
+  explicit PowerOfTwo(U aInput)
+      : mValue(FriendlyRoundUpPow2(static_cast<T>(aInput))) {
+    static_assert(!std::numeric_limits<T>::is_signed,
+                  "PowerOfTwo does not accept signed types");
+    static_assert(sizeof(U) <= sizeof(T),
+                  "PowerOfTwo does not accept bigger types");
+  }
+
+  constexpr T Value() const { return mValue; }
+
+  // Binary mask corresponding to the power of 2, useful for modulo.
+  // E.g., `x & powerOfTwo(y).Mask()` == `x % powerOfTwo(y)`.
+  // Consider PowerOfTwoMask class instead of PowerOfTwo if masking is the
+  // primary use case.
+  constexpr T MaskValue() const { return mValue - 1; }
+
+  // PowerOfTwoMask corresponding to this power of 2, useful for modulo.
+  constexpr PowerOfTwoMask<T> Mask() const {
+    using Trusted = typename PowerOfTwoMask<T>::Trusted;
+    return PowerOfTwoMask<T>(Trusted{MaskValue()});
+  }
+
+  // `x % aPowerOfTwo` works optimally.
+  // Useful for templated code doing modulo with a template argument type.
+  // Use PowerOfTwoMask class instead if masking is the primary use case.
+  template <typename U>
+  friend constexpr U operator%(U aNumerator, PowerOfTwo aDenominator) {
+    return aNumerator & aDenominator.MaskValue();
+  }
+
+  constexpr bool operator==(const PowerOfTwo& aRhs) const {
+    return mValue == aRhs.mValue;
+  }
+  constexpr bool operator!=(const PowerOfTwo& aRhs) const {
+    return mValue != aRhs.mValue;
+  }
+  constexpr bool operator<(const PowerOfTwo& aRhs) const {
+    return mValue < aRhs.mValue;
+  }
+  constexpr bool operator<=(const PowerOfTwo& aRhs) const {
+    return mValue <= aRhs.mValue;
+  }
+  constexpr bool operator>(const PowerOfTwo& aRhs) const {
+    return mValue > aRhs.mValue;
+  }
+  constexpr bool operator>=(const PowerOfTwo& aRhs) const {
+    return mValue >= aRhs.mValue;
+  }
+
+ private:
+  // Trust `MakePowerOfTwo()` to call the private Trusted constructor below.
+  template <typename U, U Value>
+  friend constexpr PowerOfTwo<U> MakePowerOfTwo();
+
+  struct Trusted {
+    T mValue;
+  };
+  // Construct a PowerOfTwo with the given trusted value.
+  // This saves having to compute the nearest 2^N.
+  // Note: Not offering PowerOfTwoMark-to-PowerOfTwo conversion, because that
+  // could result in an unexpected 0 result for the largest possible mask.
+  explicit constexpr PowerOfTwo(const Trusted& aP2) : mValue(aP2.mValue) {}
+
+  // The smallest power of 2 is 2^0 == 1.
+  T mValue = 1;
+};
+
+// Make a PowerOfTwo constant, statically-checked.
+template <typename T, T Value>
+constexpr PowerOfTwo<T> MakePowerOfTwo() {
+  static_assert(IsPowerOfTwo(Value),
+                "MakePowerOfTwo<T, Value>: Value must be 2^N");
+  using Trusted = typename PowerOfTwo<T>::Trusted;
+  return PowerOfTwo<T>(Trusted{Value});
+}
+
+// Shortcuts for the most common types and functions.
+
+using PowerOfTwoMask32 = PowerOfTwoMask<uint32_t>;
+using PowerOfTwo32 = PowerOfTwo<uint32_t>;
+using PowerOfTwoMask64 = PowerOfTwoMask<uint64_t>;
+using PowerOfTwo64 = PowerOfTwo<uint64_t>;
+
+template <uint32_t Mask>
+constexpr PowerOfTwoMask32 MakePowerOfTwoMask32() {
+  return MakePowerOfTwoMask<uint32_t, Mask>();
+}
+
+template <uint32_t Value>
+constexpr PowerOfTwo32 MakePowerOfTwo32() {
+  return MakePowerOfTwo<uint32_t, Value>();
+}
+
+template <uint64_t Mask>
+constexpr PowerOfTwoMask64 MakePowerOfTwoMask64() {
+  return MakePowerOfTwoMask<uint64_t, Mask>();
+}
+
+template <uint64_t Value>
+constexpr PowerOfTwo64 MakePowerOfTwo64() {
+  return MakePowerOfTwo<uint64_t, Value>();
+}
+
+}  // namespace mozilla
+
+#endif  // PowerOfTwo_h
diff --git a/mozglue/baseprofiler/public/ProfileBufferChunk.h b/mozglue/baseprofiler/public/ProfileBufferChunk.h
new file mode 100644
index 0000000000..24a516bcaf
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileBufferChunk.h
@@ -0,0 +1,543 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferChunk_h
+#define ProfileBufferChunk_h
+
+#include "mozilla/MemoryReporting.h"
+#include "mozilla/ProfileBufferIndex.h"
+#include "mozilla/Span.h"
+#include "mozilla/TimeStamp.h"
+#include "mozilla/UniquePtr.h"
+
+#if defined(MOZ_MEMORY)
+#  include "mozmemory.h"
+#endif
+
+#include <algorithm>
+#include <limits>
+#include <type_traits>
+
+#ifdef DEBUG
+#  include <cstdio>
+#endif
+
+namespace mozilla {
+
+// Represents a single chunk of memory, with a link to the next chunk (or null).
+//
+// A chunk is made of an internal header (which contains a public part) followed
+// by user-accessible bytes.
+//
+// +---------------+---------+----------------------------------------------+
+// | public Header | private |         memory containing user blocks        |
+// +---------------+---------+----------------------------------------------+
+//                           <---------------BufferBytes()------------------>
+// <------------------------------ChunkBytes()------------------------------>
+//
+// The chunk can reserve "blocks", but doesn't know the internal contents of
+// each block, it only knows where the first one starts, and where the last one
+// ends (which is where the next one will begin, if not already out of range).
+// It is up to the user to add structure to each block so that they can be
+// distinguished when later read.
+//
+// +---------------+---------+----------------------------------------------+
+// | public Header | private |      [1st block]...[last full block]         |
+// +---------------+---------+----------------------------------------------+
+//  ChunkHeader().mOffsetFirstBlock ^                             ^
+//                           ChunkHeader().mOffsetPastLastBlock --'
+//
+// It is possible to attempt to reserve more than the remaining space, in which
+// case only what is available is returned. The caller is responsible for using
+// another chunk, reserving a block "tail" in it, and using both parts to
+// constitute a full block. (This initial tail may be empty in some chunks.)
+//
+// +---------------+---------+----------------------------------------------+
+// | public Header | private | tail][1st block]...[last full block][head... |
+// +---------------+---------+----------------------------------------------+
+//  ChunkHeader().mOffsetFirstBlock ^                                       ^
+//                                     ChunkHeader().mOffsetPastLastBlock --'
+//
+// Each Chunk has an internal state (checked in DEBUG builds) that directs how
+// to use it during creation, initialization, use, end of life, recycling, and
+// destruction. See `State` below for details.
+// In particular:
+// - `ReserveInitialBlockAsTail()` must be called before the first `Reserve()`
+//   after construction or recycling, even with a size of 0 (no actual tail),
+// - `MarkDone()` and `MarkRecycled()` must be called as appropriate.
+class ProfileBufferChunk {
+ public:
+  using Byte = uint8_t;
+  using Length = uint32_t;
+
+  using SpanOfBytes = Span<Byte>;
+
+  // Hint about the size of the metadata (public and private headers).
+  // `Create()` below takes the minimum *buffer* size, so the minimum total
+  // Chunk size is at least `SizeofChunkMetadata() + aMinBufferBytes`.
+  [[nodiscard]] static constexpr Length SizeofChunkMetadata() {
+    return static_cast<Length>(sizeof(InternalHeader));
+  }
+
+  // Allocate space for a chunk with a given minimum size, and construct it.
+  // The actual size may be higher, to match the actual space taken in the
+  // memory pool.
+  [[nodiscard]] static UniquePtr<ProfileBufferChunk> Create(
+      Length aMinBufferBytes) {
+    // We need at least one byte, to cover the always-present `mBuffer` byte.
+    aMinBufferBytes = std::max(aMinBufferBytes, Length(1));
+    // Trivial struct with the same alignment as `ProfileBufferChunk`, and size
+    // equal to that alignment, because typically the sizeof of an object is
+    // a multiple of its alignment.
+    struct alignas(alignof(InternalHeader)) ChunkStruct {
+      Byte c[alignof(InternalHeader)];
+    };
+    static_assert(std::is_trivial_v<ChunkStruct>,
+                  "ChunkStruct must be trivial to avoid any construction");
+    // Allocate an array of that struct, enough to contain the expected
+    // `ProfileBufferChunk` (with its header+buffer).
+    size_t count = (sizeof(InternalHeader) + aMinBufferBytes +
+                    (alignof(InternalHeader) - 1)) /
+                   alignof(InternalHeader);
+#if defined(MOZ_MEMORY)
+    // Potentially expand the array to use more of the effective allocation.
+    count = (malloc_good_size(count * sizeof(ChunkStruct)) +
+             (sizeof(ChunkStruct) - 1)) /
+            sizeof(ChunkStruct);
+#endif
+    auto chunkStorage = MakeUnique<ChunkStruct[]>(count);
+    MOZ_ASSERT(reinterpret_cast<uintptr_t>(chunkStorage.get()) %
+                   alignof(InternalHeader) ==
+               0);
+    // After the allocation, compute the actual chunk size (including header).
+    const size_t chunkBytes = count * sizeof(ChunkStruct);
+    MOZ_ASSERT(chunkBytes >= sizeof(ProfileBufferChunk),
+               "Not enough space to construct a ProfileBufferChunk");
+    MOZ_ASSERT(chunkBytes <=
+               static_cast<size_t>(std::numeric_limits<Length>::max()));
+    // Compute the size of the user-accessible buffer inside the chunk.
+    const Length bufferBytes =
+        static_cast<Length>(chunkBytes - sizeof(InternalHeader));
+    MOZ_ASSERT(bufferBytes >= aMinBufferBytes,
+               "Not enough space for minimum buffer size");
+    // Construct the header at the beginning of the allocated array, with the
+    // known buffer size.
+    new (chunkStorage.get()) ProfileBufferChunk(bufferBytes);
+    // We now have a proper `ProfileBufferChunk` object, create the appropriate
+    // UniquePtr for it.
+    UniquePtr<ProfileBufferChunk> chunk{
+        reinterpret_cast<ProfileBufferChunk*>(chunkStorage.release())};
+    MOZ_ASSERT(
+        size_t(reinterpret_cast<const char*>(
+                   &chunk.get()->BufferSpan()[bufferBytes - 1]) -
+               reinterpret_cast<const char*>(chunk.get())) == chunkBytes - 1,
+        "Buffer span spills out of chunk allocation");
+    return chunk;
+  }
+
+#ifdef DEBUG
+  ~ProfileBufferChunk() {
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::InUse);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Full);
+    MOZ_ASSERT(mInternalHeader.mState == InternalHeader::State::Created ||
+               mInternalHeader.mState == InternalHeader::State::Done ||
+               mInternalHeader.mState == InternalHeader::State::Recycled);
+  }
+#endif
+
+  // Must be called with the first block tail (may be empty), which will be
+  // skipped if the reader starts with this ProfileBufferChunk.
+  [[nodiscard]] SpanOfBytes ReserveInitialBlockAsTail(Length aTailSize) {
+#ifdef DEBUG
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::InUse);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Full);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Done);
+    MOZ_ASSERT(mInternalHeader.mState == InternalHeader::State::Created ||
+               mInternalHeader.mState == InternalHeader::State::Recycled);
+    mInternalHeader.mState = InternalHeader::State::InUse;
+#endif
+    mInternalHeader.mHeader.mOffsetFirstBlock = aTailSize;
+    mInternalHeader.mHeader.mOffsetPastLastBlock = aTailSize;
+    return SpanOfBytes(&mBuffer, aTailSize);
+  }
+
+  struct ReserveReturn {
+    SpanOfBytes mSpan;
+    ProfileBufferBlockIndex mBlockRangeIndex;
+  };
+
+  // Reserve a block of up to `aBlockSize` bytes, and return a Span to it, and
+  // its starting index. The actual size may be smaller, if the block cannot fit
+  // in the remaining space.
+  [[nodiscard]] ReserveReturn ReserveBlock(Length aBlockSize) {
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Created);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Full);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Done);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Recycled);
+    MOZ_ASSERT(mInternalHeader.mState == InternalHeader::State::InUse);
+    MOZ_ASSERT(RangeStart() != 0,
+               "Expected valid range start before first Reserve()");
+    const Length blockOffset = mInternalHeader.mHeader.mOffsetPastLastBlock;
+    Length reservedSize = aBlockSize;
+    if (MOZ_UNLIKELY(aBlockSize >= RemainingBytes())) {
+      reservedSize = RemainingBytes();
+#ifdef DEBUG
+      mInternalHeader.mState = InternalHeader::State::Full;
+#endif
+    }
+    mInternalHeader.mHeader.mOffsetPastLastBlock += reservedSize;
+    mInternalHeader.mHeader.mBlockCount += 1;
+    return {SpanOfBytes(&mBuffer + blockOffset, reservedSize),
+            ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+                mInternalHeader.mHeader.mRangeStart + blockOffset)};
+  }
+
+  // When a chunk will not be used to store more blocks (because it is full, or
+  // because the profiler will not add more data), it should be marked "done".
+  // Access to its content is still allowed.
+  void MarkDone() {
+#ifdef DEBUG
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Created);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Done);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Recycled);
+    MOZ_ASSERT(mInternalHeader.mState == InternalHeader::State::InUse ||
+               mInternalHeader.mState == InternalHeader::State::Full);
+    mInternalHeader.mState = InternalHeader::State::Done;
+#endif
+    mInternalHeader.mHeader.mDoneTimeStamp = TimeStamp::NowUnfuzzed();
+  }
+
+  // A "Done" chunk may be recycled, to avoid allocating a new one.
+  void MarkRecycled() {
+#ifdef DEBUG
+    // We also allow Created and already-Recycled chunks to be recycled, this
+    // way it's easier to recycle chunks when their state is not easily
+    // trackable.
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::InUse);
+    MOZ_ASSERT(mInternalHeader.mState != InternalHeader::State::Full);
+    MOZ_ASSERT(mInternalHeader.mState == InternalHeader::State::Created ||
+               mInternalHeader.mState == InternalHeader::State::Done ||
+               mInternalHeader.mState == InternalHeader::State::Recycled);
+    mInternalHeader.mState = InternalHeader::State::Recycled;
+#endif
+    // Reset all header fields, in case this recycled chunk gets read.
+    mInternalHeader.mHeader.Reset();
+  }
+
+  // Public header, meant to uniquely identify a chunk, it may be shared with
+  // other processes to coordinate global memory handling.
+  struct Header {
+    explicit Header(Length aBufferBytes) : mBufferBytes(aBufferBytes) {}
+
+    // Reset all members to their as-new values (apart from the buffer size,
+    // which cannot change), ready for re-use.
+    void Reset() {
+      mOffsetFirstBlock = 0;
+      mOffsetPastLastBlock = 0;
+      mDoneTimeStamp = TimeStamp{};
+      mBlockCount = 0;
+      mRangeStart = 0;
+      mProcessId = 0;
+    }
+
+    // Note: Part of the ordering of members below is to avoid unnecessary
+    // padding.
+
+    // Members managed by the ProfileBufferChunk.
+
+    // Offset of the first block (past the initial tail block, which may be 0).
+    Length mOffsetFirstBlock = 0;
+    // Offset past the last byte of the last reserved block
+    // It may be past mBufferBytes when last block continues in the next
+    // ProfileBufferChunk. It may be before mBufferBytes if ProfileBufferChunk
+    // is marked "Done" before the end is reached.
+    Length mOffsetPastLastBlock = 0;
+    // Timestamp when buffer is "Done" (which happens when the last block is
+    // written). This will be used to find and discard the oldest
+    // ProfileBufferChunk.
+    TimeStamp mDoneTimeStamp;
+    // Number of bytes in the buffer, set once at construction time.
+    const Length mBufferBytes;
+    // Number of reserved blocks (including final one even if partial, but
+    // excluding initial tail).
+    Length mBlockCount = 0;
+
+    // Meta-data set by the user.
+
+    // Index of the first byte of this ProfileBufferChunk, relative to all
+    // Chunks for this process. Index 0 is reserved as nullptr-like index,
+    // mRangeStart should be set to a non-0 value before the first `Reserve()`.
+    ProfileBufferIndex mRangeStart = 0;
+    // Process writing to this ProfileBufferChunk.
+    int mProcessId = 0;
+
+    // A bit of spare space (necessary here because of the alignment due to
+    // other members), may be later repurposed for extra data.
+    const int mPADDING = 0;
+  };
+
+  [[nodiscard]] const Header& ChunkHeader() const {
+    return mInternalHeader.mHeader;
+  }
+
+  [[nodiscard]] Length BufferBytes() const {
+    return ChunkHeader().mBufferBytes;
+  }
+
+  // Total size of the chunk (buffer + header).
+  [[nodiscard]] Length ChunkBytes() const {
+    return static_cast<Length>(sizeof(InternalHeader)) + BufferBytes();
+  }
+
+  // Size of external resources, in this case all the following chunks.
+  [[nodiscard]] size_t SizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const {
+    const ProfileBufferChunk* const next = GetNext();
+    return next ? next->SizeOfIncludingThis(aMallocSizeOf) : 0;
+  }
+
+  // Size of this chunk and all following ones.
+  [[nodiscard]] size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+    // Just in case `aMallocSizeOf` falls back on just `sizeof`, make sure we
+    // account for at least the actual Chunk requested allocation size.
+    return std::max<size_t>(aMallocSizeOf(this), ChunkBytes()) +
+           SizeOfExcludingThis(aMallocSizeOf);
+  }
+
+  [[nodiscard]] Length RemainingBytes() const {
+    return BufferBytes() - OffsetPastLastBlock();
+  }
+
+  [[nodiscard]] Length OffsetFirstBlock() const {
+    return ChunkHeader().mOffsetFirstBlock;
+  }
+
+  [[nodiscard]] Length OffsetPastLastBlock() const {
+    return ChunkHeader().mOffsetPastLastBlock;
+  }
+
+  [[nodiscard]] Length BlockCount() const { return ChunkHeader().mBlockCount; }
+
+  [[nodiscard]] int ProcessId() const { return ChunkHeader().mProcessId; }
+
+  void SetProcessId(int aProcessId) {
+    mInternalHeader.mHeader.mProcessId = aProcessId;
+  }
+
+  // Global range index at the start of this Chunk.
+  [[nodiscard]] ProfileBufferIndex RangeStart() const {
+    return ChunkHeader().mRangeStart;
+  }
+
+  void SetRangeStart(ProfileBufferIndex aRangeStart) {
+    mInternalHeader.mHeader.mRangeStart = aRangeStart;
+  }
+
+  // Get a read-only Span to the buffer. It is up to the caller to decypher the
+  // contents, based on known offsets and the internal block structure.
+  [[nodiscard]] Span<const Byte> BufferSpan() const {
+    return Span<const Byte>(&mBuffer, BufferBytes());
+  }
+
+  [[nodiscard]] Byte ByteAt(Length aOffset) const {
+    MOZ_ASSERT(aOffset < OffsetPastLastBlock());
+    return *(&mBuffer + aOffset);
+  }
+
+  [[nodiscard]] ProfileBufferChunk* GetNext() {
+    return mInternalHeader.mNext.get();
+  }
+  [[nodiscard]] const ProfileBufferChunk* GetNext() const {
+    return mInternalHeader.mNext.get();
+  }
+
+  [[nodiscard]] UniquePtr<ProfileBufferChunk> ReleaseNext() {
+    return std::move(mInternalHeader.mNext);
+  }
+
+  void InsertNext(UniquePtr<ProfileBufferChunk>&& aChunk) {
+    if (!aChunk) {
+      return;
+    }
+    aChunk->SetLast(ReleaseNext());
+    mInternalHeader.mNext = std::move(aChunk);
+  }
+
+  // Find the last chunk in this chain (it may be `this`).
+  [[nodiscard]] ProfileBufferChunk* Last() {
+    ProfileBufferChunk* chunk = this;
+    for (;;) {
+      ProfileBufferChunk* next = chunk->GetNext();
+      if (!next) {
+        return chunk;
+      }
+      chunk = next;
+    }
+  }
+  [[nodiscard]] const ProfileBufferChunk* Last() const {
+    const ProfileBufferChunk* chunk = this;
+    for (;;) {
+      const ProfileBufferChunk* next = chunk->GetNext();
+      if (!next) {
+        return chunk;
+      }
+      chunk = next;
+    }
+  }
+
+  void SetLast(UniquePtr<ProfileBufferChunk>&& aChunk) {
+    if (!aChunk) {
+      return;
+    }
+    Last()->mInternalHeader.mNext = std::move(aChunk);
+  }
+
+  // Join two possibly-null chunk lists.
+  [[nodiscard]] static UniquePtr<ProfileBufferChunk> Join(
+      UniquePtr<ProfileBufferChunk>&& aFirst,
+      UniquePtr<ProfileBufferChunk>&& aLast) {
+    if (aFirst) {
+      aFirst->SetLast(std::move(aLast));
+      return std::move(aFirst);
+    }
+    return std::move(aLast);
+  }
+
+#ifdef DEBUG
+  void Dump(std::FILE* aFile = stdout) const {
+    fprintf(aFile,
+            "Chunk[%p] chunkSize=%u bufferSize=%u state=%s rangeStart=%u "
+            "firstBlockOffset=%u offsetPastLastBlock=%u blockCount=%u",
+            this, unsigned(ChunkBytes()), unsigned(BufferBytes()),
+            mInternalHeader.StateString(), unsigned(RangeStart()),
+            unsigned(OffsetFirstBlock()), unsigned(OffsetPastLastBlock()),
+            unsigned(BlockCount()));
+    const auto len = OffsetPastLastBlock();
+    constexpr unsigned columns = 16;
+    unsigned char ascii[columns + 1];
+    ascii[columns] = '\0';
+    for (Length i = 0; i < len; ++i) {
+      if (i % columns == 0) {
+        fprintf(aFile, "\n  %4u=0x%03x:", unsigned(i), unsigned(i));
+        for (unsigned a = 0; a < columns; ++a) {
+          ascii[a] = ' ';
+        }
+      }
+      unsigned char sep = ' ';
+      if (i == OffsetFirstBlock()) {
+        if (i == OffsetPastLastBlock()) {
+          sep = '#';
+        } else {
+          sep = '[';
+        }
+      } else if (i == OffsetPastLastBlock()) {
+        sep = ']';
+      }
+      unsigned char c = *(&mBuffer + i);
+      fprintf(aFile, "%c%02x", sep, c);
+
+      if (i == len - 1) {
+        if (i + 1 == OffsetPastLastBlock()) {
+          // Special case when last block ends right at the end.
+          fprintf(aFile, "]");
+        } else {
+          fprintf(aFile, " ");
+        }
+      } else if (i % columns == columns - 1) {
+        fprintf(aFile, " ");
+      }
+
+      ascii[i % columns] = (c >= ' ' && c <= '~') ? c : '.';
+
+      if (i % columns == columns - 1) {
+        fprintf(aFile, " %s", ascii);
+      }
+    }
+
+    if (len % columns < columns - 1) {
+      for (Length i = len % columns; i < columns; ++i) {
+        fprintf(aFile, "   ");
+      }
+      fprintf(aFile, " %s", ascii);
+    }
+
+    fprintf(aFile, "\n");
+  }
+#endif  // DEBUG
+
+ private:
+  // ProfileBufferChunk constructor. Use static `Create()` to allocate and
+  // construct a ProfileBufferChunk.
+  explicit ProfileBufferChunk(Length aBufferBytes)
+      : mInternalHeader(aBufferBytes) {}
+
+  // This internal header starts with the public `Header`, and adds some data
+  // only necessary for local handling.
+  // This encapsulation is also necessary to perform placement-new in
+  // `Create()`.
+  struct InternalHeader {
+    explicit InternalHeader(Length aBufferBytes) : mHeader(aBufferBytes) {}
+
+    Header mHeader;
+    UniquePtr<ProfileBufferChunk> mNext;
+
+#ifdef DEBUG
+    enum class State {
+      Created,  // Self-set. Just constructed, waiting for initial block tail.
+      InUse,    // Ready to accept blocks.
+      Full,     // Self-set. Blocks reach the end (or further).
+      Done,     // Blocks won't be added anymore.
+      Recycled  // Still full of data, but expecting an initial block tail.
+    };
+
+    State mState = State::Created;
+    // Transition table: (X=unexpected)
+    // Method          \  State   Created  InUse    Full     Done     Recycled
+    // ReserveInitialBlockAsTail   InUse     X       X        X        InUse
+    // Reserve                       X   InUse/Full  X        X          X
+    // MarkDone                      X     Done     Done      X          X
+    // MarkRecycled                  X       X       X      Recycled     X
+    // destructor                    ok      X       X        ok         ok
+
+    const char* StateString() const {
+      switch (mState) {
+        case State::Created:
+          return "Created";
+        case State::InUse:
+          return "InUse";
+        case State::Full:
+          return "Full";
+        case State::Done:
+          return "Done";
+        case State::Recycled:
+          return "Recycled";
+        default:
+          return "?";
+      }
+    }
+#else  // DEBUG
+    const char* StateString() const { return "(non-DEBUG)"; }
+#endif
+  };
+
+  InternalHeader mInternalHeader;
+
+  // KEEP THIS LAST!
+  // First byte of the buffer. Note that ProfileBufferChunk::Create allocates a
+  // bigger block, such that `mBuffer` is the first of `mBufferBytes` available
+  // bytes.
+  // The initialization is not strictly needed, because bytes should only be
+  // read after they have been written and `mOffsetPastLastBlock` has been
+  // updated. However:
+  // - Reviewbot complains that it's not initialized.
+  // - It's cheap to initialize one byte.
+  // - In the worst case (reading does happen), zero is not a valid entry size
+  //   and should get caught in entry readers.
+  Byte mBuffer = '\0';
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileBufferChunk_h
diff --git a/mozglue/baseprofiler/public/ProfileBufferChunkManager.h b/mozglue/baseprofiler/public/ProfileBufferChunkManager.h
new file mode 100644
index 0000000000..e7f12bf21f
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileBufferChunkManager.h
@@ -0,0 +1,134 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferChunkManager_h
+#define ProfileBufferChunkManager_h
+
+#include "mozilla/ProfileBufferChunk.h"
+#include "mozilla/ScopeExit.h"
+
+#include <functional>
+
+namespace mozilla {
+
+// Manages the ProfileBufferChunks for this process.
+// The main user of this class is the buffer that needs chunks to store its
+// data.
+// The main ProfileBufferChunks responsibilities are:
+// - It can create new chunks, they are called "unreleased".
+// - Later these chunks are returned here, and become "released".
+// - The manager is free to destroy or recycle the oldest released chunks
+//   (usually to reclaim memory), and will inform the user through a provided
+//   callback.
+// - The user may access still-alive released chunks.
+class ProfileBufferChunkManager {
+ public:
+  virtual ~ProfileBufferChunkManager()
+#ifdef DEBUG
+  {
+    MOZ_ASSERT(!mUser, "Still registered when being destroyed");
+  }
+#else
+      = default;
+#endif
+
+  // Expected maximum size needed to store one stack sample.
+  // Most ChunkManager sub-classes will require chunk sizes, this can serve as
+  // a minimum recommendation to hold most backtraces.
+  constexpr static ProfileBufferChunk::Length scExpectedMaximumStackSize =
+      128 * 1024;
+
+  // Estimated maximum buffer size.
+  [[nodiscard]] virtual size_t MaxTotalSize() const = 0;
+
+  // Create or recycle a chunk right now. May return null in case of allocation
+  // failure.
+  // Note that the chunk-destroyed callback may be invoked during this call;
+  // user should be careful with reentrancy issues.
+  [[nodiscard]] virtual UniquePtr<ProfileBufferChunk> GetChunk() = 0;
+
+  // `aChunkReceiver` may be called with a new or recycled chunk, or nullptr.
+  // (See `FulfillChunkRequests()` regarding when the callback may happen.)
+  virtual void RequestChunk(
+      std::function<void(UniquePtr<ProfileBufferChunk>)>&& aChunkReceiver) = 0;
+
+  // This method may be invoked at any time on any thread (and not necessarily
+  // by the main user of this class), to do the work necessary to respond to a
+  // previous `RequestChunk()`.
+  // It is optional: If it is never called, or called too late, the user is
+  // responsible for directly calling `GetChunk()` when a chunk is really
+  // needed (or it should at least fail gracefully).
+  // The idea is to fulfill chunk request on a separate thread, and most
+  // importantly outside of profiler calls, to avoid doing expensive memory
+  // allocations during these calls.
+  virtual void FulfillChunkRequests() = 0;
+
+  // One chunk is released by the user, the ProfileBufferChunkManager should
+  // keep it as long as possible (depending on local or global memory/time
+  // limits). Note that the chunk-destroyed callback may be invoked during this
+  // call; user should be careful with reentrancy issues.
+  virtual void ReleaseChunk(UniquePtr<ProfileBufferChunk> aChunk) = 0;
+
+  // `aChunkDestroyedCallback` will be called whenever the contents of a
+  // previously-released chunk is about to be destroyed or recycled.
+  // Note that it may be called during other functions above, or at other times
+  // from the same or other threads; user should be careful with reentrancy
+  // issues.
+  virtual void SetChunkDestroyedCallback(
+      std::function<void(const ProfileBufferChunk&)>&&
+          aChunkDestroyedCallback) = 0;
+
+  // Give away all released chunks that have not yet been destroyed.
+  [[nodiscard]] virtual UniquePtr<ProfileBufferChunk>
+  GetExtantReleasedChunks() = 0;
+
+  // Let a callback see all released chunks that have not yet been destroyed, if
+  // any. Return whatever the callback returns.
+  template <typename Callback>
+  [[nodiscard]] auto PeekExtantReleasedChunks(Callback&& aCallback) {
+    const ProfileBufferChunk* chunks = PeekExtantReleasedChunksAndLock();
+    auto unlock =
+        MakeScopeExit([&]() { UnlockAfterPeekExtantReleasedChunks(); });
+    return std::forward<Callback>(aCallback)(chunks);
+  }
+
+  // Chunks that were still unreleased will never be released.
+  virtual void ForgetUnreleasedChunks() = 0;
+
+  [[nodiscard]] virtual size_t SizeOfExcludingThis(
+      MallocSizeOf aMallocSizeOf) const = 0;
+  [[nodiscard]] virtual size_t SizeOfIncludingThis(
+      MallocSizeOf aMallocSizeOf) const = 0;
+
+ protected:
+  // Derived classes to implement `PeekExtantReleasedChunks` through these:
+  virtual const ProfileBufferChunk* PeekExtantReleasedChunksAndLock() = 0;
+  virtual void UnlockAfterPeekExtantReleasedChunks() = 0;
+
+#ifdef DEBUG
+ public:
+  // DEBUG checks ensuring that this manager and its users avoid UAFs.
+  // Derived classes should assert that mUser is not null in their functions.
+
+  void RegisteredWith(const void* aUser) {
+    MOZ_ASSERT(!mUser);
+    MOZ_ASSERT(aUser);
+    mUser = aUser;
+  }
+
+  void DeregisteredFrom(const void* aUser) {
+    MOZ_ASSERT(mUser == aUser);
+    mUser = nullptr;
+  }
+
+ protected:
+  const void* mUser = nullptr;
+#endif  // DEBUG
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileBufferChunkManager_h
diff --git a/mozglue/baseprofiler/public/ProfileBufferChunkManagerSingle.h b/mozglue/baseprofiler/public/ProfileBufferChunkManagerSingle.h
new file mode 100644
index 0000000000..c91b38cbdb
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileBufferChunkManagerSingle.h
@@ -0,0 +1,172 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferChunkManagerSingle_h
+#define ProfileBufferChunkManagerSingle_h
+
+#include "mozilla/ProfileBufferChunkManager.h"
+
+#ifdef DEBUG
+#  include "mozilla/Atomics.h"
+#endif  // DEBUG
+
+namespace mozilla {
+
+// Manages only one Chunk.
+// The first call to `Get`/`RequestChunk()` will retrieve the one chunk, and all
+// subsequent calls will return nullptr. That chunk may still be released, but
+// it will never be destroyed or recycled.
+// Unlike others, this manager may be `Reset()`, to allow another round of
+// small-data gathering.
+// The main use is with short-lived ProfileChunkedBuffers that collect little
+// data that can fit in one chunk, e.g., capturing one stack.
+// It is not thread-safe.
+class ProfileBufferChunkManagerSingle final : public ProfileBufferChunkManager {
+ public:
+  using Length = ProfileBufferChunk::Length;
+
+  // Use a preallocated chunk. (Accepting null to gracefully handle OOM.)
+  explicit ProfileBufferChunkManagerSingle(UniquePtr<ProfileBufferChunk> aChunk)
+      : mInitialChunk(std::move(aChunk)),
+        mBufferBytes(mInitialChunk ? mInitialChunk->BufferBytes() : 0) {
+    MOZ_ASSERT(!mInitialChunk || !mInitialChunk->GetNext(),
+               "Expected at most one chunk");
+  }
+
+  // ChunkMinBufferBytes: Minimum number of user-available bytes in the Chunk.
+  // Note that Chunks use a bit more memory for their header.
+  explicit ProfileBufferChunkManagerSingle(Length aChunkMinBufferBytes)
+      : mInitialChunk(ProfileBufferChunk::Create(aChunkMinBufferBytes)),
+        mBufferBytes(mInitialChunk ? mInitialChunk->BufferBytes() : 0) {}
+
+#ifdef DEBUG
+  ~ProfileBufferChunkManagerSingle() { MOZ_ASSERT(mVirtuallyLocked == false); }
+#endif  // DEBUG
+
+  // Reset this manager, using the provided chunk (probably coming from the
+  // ProfileChunkedBuffer that just used it); if null, fallback on current or
+  // released chunk.
+  void Reset(UniquePtr<ProfileBufferChunk> aPossibleChunk) {
+    if (aPossibleChunk) {
+      mInitialChunk = std::move(aPossibleChunk);
+      mReleasedChunk = nullptr;
+    } else if (!mInitialChunk) {
+      MOZ_ASSERT(!!mReleasedChunk, "Can't reset properly!");
+      mInitialChunk = std::move(mReleasedChunk);
+    }
+
+    if (mInitialChunk) {
+      mInitialChunk->MarkRecycled();
+      mBufferBytes = mInitialChunk->BufferBytes();
+    } else {
+      mBufferBytes = 0;
+    }
+  }
+
+  [[nodiscard]] size_t MaxTotalSize() const final { return mBufferBytes; }
+
+  // One of `GetChunk` and `RequestChunk` will only work the very first time (if
+  // there's even a chunk).
+  [[nodiscard]] UniquePtr<ProfileBufferChunk> GetChunk() final {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    return std::move(mInitialChunk);
+  }
+
+  void RequestChunk(std::function<void(UniquePtr<ProfileBufferChunk>)>&&
+                        aChunkReceiver) final {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    // Simple retrieval.
+    std::move(aChunkReceiver)(GetChunk());
+  }
+
+  void FulfillChunkRequests() final {
+    // Nothing to do here.
+  }
+
+  void ReleaseChunk(UniquePtr<ProfileBufferChunk> aChunk) final {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    if (!aChunk) {
+      return;
+    }
+    MOZ_ASSERT(!mReleasedChunk, "Unexpected 2nd released chunk");
+    MOZ_ASSERT(!aChunk->GetNext(), "Only expected one released chunk");
+    mReleasedChunk = std::move(aChunk);
+  }
+
+  void SetChunkDestroyedCallback(
+      std::function<void(const ProfileBufferChunk&)>&& aChunkDestroyedCallback)
+      final {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    // The chunk-destroyed callback will never actually be called, but we keep
+    // the callback here in case the caller expects it to live as long as this
+    // manager.
+    mChunkDestroyedCallback = std::move(aChunkDestroyedCallback);
+  }
+
+  [[nodiscard]] UniquePtr<ProfileBufferChunk> GetExtantReleasedChunks() final {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    return std::move(mReleasedChunk);
+  }
+
+  void ForgetUnreleasedChunks() final {
+    MOZ_ASSERT(mUser, "Not registered yet");
+  }
+
+  [[nodiscard]] size_t SizeOfExcludingThis(
+      MallocSizeOf aMallocSizeOf) const final {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    size_t size = 0;
+    if (mInitialChunk) {
+      size += mInitialChunk->SizeOfIncludingThis(aMallocSizeOf);
+    }
+    if (mReleasedChunk) {
+      size += mReleasedChunk->SizeOfIncludingThis(aMallocSizeOf);
+    }
+    // Note: Missing size of std::function external resources (if any).
+    return size;
+  }
+
+  [[nodiscard]] size_t SizeOfIncludingThis(
+      MallocSizeOf aMallocSizeOf) const final {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    return aMallocSizeOf(this) + SizeOfExcludingThis(aMallocSizeOf);
+  }
+
+ protected:
+  // This manager is not thread-safe, so there's not actual locking needed.
+  const ProfileBufferChunk* PeekExtantReleasedChunksAndLock() final {
+    MOZ_ASSERT(mVirtuallyLocked.compareExchange(false, true));
+    MOZ_ASSERT(mUser, "Not registered yet");
+    return mReleasedChunk.get();
+  }
+  void UnlockAfterPeekExtantReleasedChunks() final {
+    MOZ_ASSERT(mVirtuallyLocked.compareExchange(true, false));
+  }
+
+ private:
+  // Initial chunk created with this manager, given away at first Get/Request.
+  UniquePtr<ProfileBufferChunk> mInitialChunk;
+
+  // Storage for the released chunk (which should probably not happen, as it
+  // means the chunk is full).
+  UniquePtr<ProfileBufferChunk> mReleasedChunk;
+
+  // Size of the one chunk we're managing. Stored here, because the chunk may
+  // be moved out and inaccessible from here.
+  Length mBufferBytes;
+
+  // The chunk-destroyed callback will never actually be called, but we keep it
+  // here in case the caller expects it to live as long as this manager.
+  std::function<void(const ProfileBufferChunk&)> mChunkDestroyedCallback;
+
+#ifdef DEBUG
+  mutable Atomic<bool> mVirtuallyLocked{false};
+#endif  // DEBUG
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileBufferChunkManagerSingle_h
diff --git a/mozglue/baseprofiler/public/ProfileBufferChunkManagerWithLocalLimit.h b/mozglue/baseprofiler/public/ProfileBufferChunkManagerWithLocalLimit.h
new file mode 100644
index 0000000000..5b1af6d66c
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileBufferChunkManagerWithLocalLimit.h
@@ -0,0 +1,428 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferChunkManagerWithLocalLimit_h
+#define ProfileBufferChunkManagerWithLocalLimit_h
+
+#include "BaseProfiler.h"
+#include "mozilla/BaseProfilerDetail.h"
+#include "mozilla/ProfileBufferChunkManager.h"
+#include "mozilla/ProfileBufferControlledChunkManager.h"
+
+#include <utility>
+
+namespace mozilla {
+
+// Manages the Chunks for this process in a thread-safe manner, with a maximum
+// size per process.
+//
+// "Unreleased" chunks are not owned here, only "released" chunks can be
+// destroyed or recycled when reaching the memory limit, so it is theoretically
+// possible to break that limit, if:
+// - The user of this class doesn't release their chunks, AND/OR
+// - The limit is too small (e.g., smaller than 2 or 3 chunks, which should be
+//   the usual number of unreleased chunks in flight).
+// In this case, it just means that we will use more memory than allowed,
+// potentially risking OOMs. Hopefully this shouldn't happen in real code,
+// assuming that the user is doing the right thing and releasing chunks ASAP,
+// and that the memory limit is reasonably large.
+class ProfileBufferChunkManagerWithLocalLimit final
+    : public ProfileBufferChunkManager,
+      public ProfileBufferControlledChunkManager {
+ public:
+  using Length = ProfileBufferChunk::Length;
+
+  // MaxTotalBytes: Maximum number of bytes allocated in all local Chunks.
+  // ChunkMinBufferBytes: Minimum number of user-available bytes in each Chunk.
+  // Note that Chunks use a bit more memory for their header.
+  explicit ProfileBufferChunkManagerWithLocalLimit(size_t aMaxTotalBytes,
+                                                   Length aChunkMinBufferBytes)
+      : mMaxTotalBytes(aMaxTotalBytes),
+        mChunkMinBufferBytes(aChunkMinBufferBytes) {}
+
+  ~ProfileBufferChunkManagerWithLocalLimit() {
+    if (mUpdateCallback) {
+      // Signal the end of this callback.
+      std::move(mUpdateCallback)(Update(nullptr));
+    }
+  }
+
+  [[nodiscard]] size_t MaxTotalSize() const final {
+    // `mMaxTotalBytes` is `const` so there is no need to lock the mutex.
+    return mMaxTotalBytes;
+  }
+
+  [[nodiscard]] UniquePtr<ProfileBufferChunk> GetChunk() final {
+    AUTO_PROFILER_STATS(Local_GetChunk);
+
+    ChunkAndUpdate chunkAndUpdate = [&]() {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+      return GetChunk(lock);
+    }();
+
+    baseprofiler::detail::BaseProfilerAutoLock lock(mUpdateCallbackMutex);
+    if (mUpdateCallback && !chunkAndUpdate.second.IsNotUpdate()) {
+      mUpdateCallback(std::move(chunkAndUpdate.second));
+    }
+
+    return std::move(chunkAndUpdate.first);
+  }
+
+  void RequestChunk(std::function<void(UniquePtr<ProfileBufferChunk>)>&&
+                        aChunkReceiver) final {
+    AUTO_PROFILER_STATS(Local_RequestChunk);
+    baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+    if (mChunkReceiver) {
+      // We already have a chunk receiver, meaning a request is pending.
+      return;
+    }
+    // Store the chunk receiver. This indicates that a request is pending, and
+    // it will be handled in the next `FulfillChunkRequests()` call.
+    mChunkReceiver = std::move(aChunkReceiver);
+  }
+
+  void FulfillChunkRequests() final {
+    AUTO_PROFILER_STATS(Local_FulfillChunkRequests);
+    std::function<void(UniquePtr<ProfileBufferChunk>)> chunkReceiver;
+    ChunkAndUpdate chunkAndUpdate = [&]() -> ChunkAndUpdate {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+      if (!mChunkReceiver) {
+        // No receiver means no pending request, we're done.
+        return {};
+      }
+      // Otherwise there is a request, extract the receiver to call below.
+      std::swap(chunkReceiver, mChunkReceiver);
+      MOZ_ASSERT(!mChunkReceiver, "mChunkReceiver should have been emptied");
+      // And allocate the requested chunk. This may fail, it's fine, we're
+      // letting the receiver know about it.
+      AUTO_PROFILER_STATS(Local_FulfillChunkRequests_GetChunk);
+      return GetChunk(lock);
+    }();
+
+    if (chunkReceiver) {
+      {
+        baseprofiler::detail::BaseProfilerAutoLock lock(mUpdateCallbackMutex);
+        if (mUpdateCallback && !chunkAndUpdate.second.IsNotUpdate()) {
+          mUpdateCallback(std::move(chunkAndUpdate.second));
+        }
+      }
+
+      // Invoke callback outside of lock, so that it can use other chunk manager
+      // functions if needed.
+      // Note that this means there could be a race, where another request
+      // happens now and even gets fulfilled before this one is! It should be
+      // rare, and shouldn't be a problem anyway, the user will still get their
+      // requested chunks, new/recycled chunks look the same so their order
+      // doesn't matter.
+      std::move(chunkReceiver)(std::move(chunkAndUpdate.first));
+    }
+  }
+
+  void ReleaseChunk(UniquePtr<ProfileBufferChunk> aChunk) final {
+    if (!aChunk) {
+      return;
+    }
+
+    MOZ_RELEASE_ASSERT(!aChunk->GetNext(), "ReleaseChunk only accepts 1 chunk");
+    MOZ_RELEASE_ASSERT(!aChunk->ChunkHeader().mDoneTimeStamp.IsNull(),
+                       "Released chunk should have a 'Done' timestamp");
+
+    Update update = [&]() {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+      MOZ_ASSERT(mUser, "Not registered yet");
+      // Keep a pointer to the first newly-released chunk, so we can use it to
+      // prepare an update (after `aChunk` is moved-from).
+      const ProfileBufferChunk* const newlyReleasedChunk = aChunk.get();
+      // Transfer the chunk size from the unreleased bucket to the released one.
+      mUnreleasedBufferBytes -= aChunk->BufferBytes();
+      mReleasedBufferBytes += aChunk->BufferBytes();
+      if (!mReleasedChunks) {
+        // No other released chunks at the moment, we're starting the list.
+        MOZ_ASSERT(mReleasedBufferBytes == aChunk->BufferBytes());
+        mReleasedChunks = std::move(aChunk);
+      } else {
+        // Insert aChunk in mReleasedChunks to keep done-timestamp order.
+        const TimeStamp& releasedChunkDoneTimeStamp =
+            aChunk->ChunkHeader().mDoneTimeStamp;
+        if (releasedChunkDoneTimeStamp <
+            mReleasedChunks->ChunkHeader().mDoneTimeStamp) {
+          // aChunk is the oldest -> Insert at the beginning.
+          aChunk->SetLast(std::move(mReleasedChunks));
+          mReleasedChunks = std::move(aChunk);
+        } else {
+          // Go through the already-released chunk list, and insert aChunk
+          // before the first younger released chunk, or at the end.
+          ProfileBufferChunk* chunk = mReleasedChunks.get();
+          for (;;) {
+            ProfileBufferChunk* const nextChunk = chunk->GetNext();
+            if (!nextChunk || releasedChunkDoneTimeStamp <
+                                  nextChunk->ChunkHeader().mDoneTimeStamp) {
+              // Either we're at the last released chunk, or the next released
+              // chunk is younger -> Insert right after this released chunk.
+              chunk->InsertNext(std::move(aChunk));
+              break;
+            }
+            chunk = nextChunk;
+          }
+        }
+      }
+
+      return Update(mUnreleasedBufferBytes, mReleasedBufferBytes,
+                    mReleasedChunks.get(), newlyReleasedChunk);
+    }();
+
+    baseprofiler::detail::BaseProfilerAutoLock lock(mUpdateCallbackMutex);
+    if (mUpdateCallback && !update.IsNotUpdate()) {
+      mUpdateCallback(std::move(update));
+    }
+  }
+
+  void SetChunkDestroyedCallback(
+      std::function<void(const ProfileBufferChunk&)>&& aChunkDestroyedCallback)
+      final {
+    baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+    MOZ_ASSERT(mUser, "Not registered yet");
+    mChunkDestroyedCallback = std::move(aChunkDestroyedCallback);
+  }
+
+  [[nodiscard]] UniquePtr<ProfileBufferChunk> GetExtantReleasedChunks() final {
+    UniquePtr<ProfileBufferChunk> chunks;
+    size_t unreleasedBufferBytes = [&]() {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+      MOZ_ASSERT(mUser, "Not registered yet");
+      mReleasedBufferBytes = 0;
+      chunks = std::move(mReleasedChunks);
+      return mUnreleasedBufferBytes;
+    }();
+
+    baseprofiler::detail::BaseProfilerAutoLock lock(mUpdateCallbackMutex);
+    if (mUpdateCallback) {
+      mUpdateCallback(Update(unreleasedBufferBytes, 0, nullptr, nullptr));
+    }
+
+    return chunks;
+  }
+
+  void ForgetUnreleasedChunks() final {
+    Update update = [&]() {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+      MOZ_ASSERT(mUser, "Not registered yet");
+      mUnreleasedBufferBytes = 0;
+      return Update(0, mReleasedBufferBytes, mReleasedChunks.get(), nullptr);
+    }();
+    baseprofiler::detail::BaseProfilerAutoLock lock(mUpdateCallbackMutex);
+    if (mUpdateCallback) {
+      mUpdateCallback(std::move(update));
+    }
+  }
+
+  [[nodiscard]] size_t SizeOfExcludingThis(
+      MallocSizeOf aMallocSizeOf) const final {
+    baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+    return SizeOfExcludingThis(aMallocSizeOf, lock);
+  }
+
+  [[nodiscard]] size_t SizeOfIncludingThis(
+      MallocSizeOf aMallocSizeOf) const final {
+    baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+    MOZ_ASSERT(mUser, "Not registered yet");
+    return aMallocSizeOf(this) + SizeOfExcludingThis(aMallocSizeOf, lock);
+  }
+
+  void SetUpdateCallback(UpdateCallback&& aUpdateCallback) final {
+    {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mUpdateCallbackMutex);
+      if (mUpdateCallback) {
+        // Signal the end of the previous callback.
+        std::move(mUpdateCallback)(Update(nullptr));
+        mUpdateCallback = nullptr;
+      }
+    }
+
+    if (aUpdateCallback) {
+      Update initialUpdate = [&]() {
+        baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+        return Update(mUnreleasedBufferBytes, mReleasedBufferBytes,
+                      mReleasedChunks.get(), nullptr);
+      }();
+
+      baseprofiler::detail::BaseProfilerAutoLock lock(mUpdateCallbackMutex);
+      MOZ_ASSERT(!mUpdateCallback, "Only one update callback allowed");
+      mUpdateCallback = std::move(aUpdateCallback);
+      mUpdateCallback(std::move(initialUpdate));
+    }
+  }
+
+  void DestroyChunksAtOrBefore(TimeStamp aDoneTimeStamp) final {
+    MOZ_ASSERT(!aDoneTimeStamp.IsNull());
+    baseprofiler::detail::BaseProfilerAutoLock lock(mMutex);
+    for (;;) {
+      if (!mReleasedChunks) {
+        // We don't own any released chunks (anymore), we're done.
+        break;
+      }
+      if (mReleasedChunks->ChunkHeader().mDoneTimeStamp > aDoneTimeStamp) {
+        // The current chunk is strictly after the given timestamp, we're done.
+        break;
+      }
+      // We've found a chunk at or before the timestamp, discard it.
+      DiscardOldestReleasedChunk(lock);
+    }
+  }
+
+ protected:
+  const ProfileBufferChunk* PeekExtantReleasedChunksAndLock() final {
+    mMutex.Lock();
+    MOZ_ASSERT(mUser, "Not registered yet");
+    return mReleasedChunks.get();
+  }
+  void UnlockAfterPeekExtantReleasedChunks() final { mMutex.Unlock(); }
+
+ private:
+  void MaybeRecycleChunk(
+      UniquePtr<ProfileBufferChunk>&& chunk,
+      const baseprofiler::detail::BaseProfilerAutoLock& aLock) {
+    // Try to recycle big-enough chunks. (All chunks should have the same size,
+    // but it's a cheap test and may allow future adjustments based on actual
+    // data rate.)
+    if (chunk->BufferBytes() >= mChunkMinBufferBytes) {
+      // We keep up to two recycled chunks at any time.
+      if (!mRecycledChunks) {
+        mRecycledChunks = std::move(chunk);
+      } else if (!mRecycledChunks->GetNext()) {
+        mRecycledChunks->InsertNext(std::move(chunk));
+      }
+    }
+  }
+
+  UniquePtr<ProfileBufferChunk> TakeRecycledChunk(
+      const baseprofiler::detail::BaseProfilerAutoLock& aLock) {
+    UniquePtr<ProfileBufferChunk> recycled;
+    if (mRecycledChunks) {
+      recycled = std::exchange(mRecycledChunks, mRecycledChunks->ReleaseNext());
+      recycled->MarkRecycled();
+    }
+    return recycled;
+  }
+
+  void DiscardOldestReleasedChunk(
+      const baseprofiler::detail::BaseProfilerAutoLock& aLock) {
+    MOZ_ASSERT(!!mReleasedChunks);
+    UniquePtr<ProfileBufferChunk> oldest =
+        std::exchange(mReleasedChunks, mReleasedChunks->ReleaseNext());
+    mReleasedBufferBytes -= oldest->BufferBytes();
+    if (mChunkDestroyedCallback) {
+      // Inform the user that we're going to destroy this chunk.
+      mChunkDestroyedCallback(*oldest);
+    }
+    MaybeRecycleChunk(std::move(oldest), aLock);
+  }
+
+  using ChunkAndUpdate = std::pair<UniquePtr<ProfileBufferChunk>, Update>;
+  [[nodiscard]] ChunkAndUpdate GetChunk(
+      const baseprofiler::detail::BaseProfilerAutoLock& aLock) {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    // After this function, the total memory consumption will be the sum of:
+    // - Bytes from released (i.e., full) chunks,
+    // - Bytes from unreleased (still in use) chunks,
+    // - Bytes from the chunk we want to create/recycle. (Note that we don't
+    //   count the extra bytes of chunk header, and of extra allocation ability,
+    //   for the new chunk, as it's assumed to be negligible compared to the
+    //   total memory limit.)
+    // If this total is higher than the local limit, we'll want to destroy
+    // the oldest released chunks until we're under the limit; if any, we may
+    // recycle one of them to avoid a deallocation followed by an allocation.
+    while (mReleasedBufferBytes + mUnreleasedBufferBytes +
+                   mChunkMinBufferBytes >=
+               mMaxTotalBytes &&
+           !!mReleasedChunks) {
+      // We have reached the local limit, discard the oldest released chunk.
+      DiscardOldestReleasedChunk(aLock);
+    }
+
+    // Extract the recycled chunk, if any.
+    ChunkAndUpdate chunkAndUpdate{TakeRecycledChunk(aLock), Update()};
+    UniquePtr<ProfileBufferChunk>& chunk = chunkAndUpdate.first;
+
+    if (!chunk) {
+      // No recycled chunk -> Create a chunk now. (This could still fail.)
+      chunk = ProfileBufferChunk::Create(mChunkMinBufferBytes);
+    }
+
+    if (chunk) {
+      // We do have a chunk (recycled or new), record its size as "unreleased".
+      mUnreleasedBufferBytes += chunk->BufferBytes();
+
+      chunkAndUpdate.second =
+          Update(mUnreleasedBufferBytes, mReleasedBufferBytes,
+                 mReleasedChunks.get(), nullptr);
+    }
+
+    return chunkAndUpdate;
+  }
+
+  [[nodiscard]] size_t SizeOfExcludingThis(
+      MallocSizeOf aMallocSizeOf,
+      const baseprofiler::detail::BaseProfilerAutoLock&) const {
+    MOZ_ASSERT(mUser, "Not registered yet");
+    size_t size = 0;
+    if (mReleasedChunks) {
+      size += mReleasedChunks->SizeOfIncludingThis(aMallocSizeOf);
+    }
+    if (mRecycledChunks) {
+      size += mRecycledChunks->SizeOfIncludingThis(aMallocSizeOf);
+    }
+    // Note: Missing size of std::function external resources (if any).
+    return size;
+  }
+
+  // Maxumum number of bytes that should be used by all unreleased and released
+  // chunks. Note that only released chunks can be destroyed here, so it is the
+  // responsibility of the user to properly release their chunks when possible.
+  const size_t mMaxTotalBytes;
+
+  // Minimum number of bytes that new chunks should be able to store.
+  // Used when calling `ProfileBufferChunk::Create()`.
+  const Length mChunkMinBufferBytes;
+
+  // Mutex guarding the following members.
+  mutable baseprofiler::detail::BaseProfilerMutex mMutex;
+
+  // Number of bytes currently held in chunks that have been given away (through
+  // `GetChunk` or `RequestChunk`) and not released yet.
+  size_t mUnreleasedBufferBytes = 0;
+
+  // Number of bytes currently held in chunks that have been released and stored
+  // in `mReleasedChunks` below.
+  size_t mReleasedBufferBytes = 0;
+
+  // List of all released chunks. The oldest one should be at the start of the
+  // list, and may be destroyed or recycled when the memory limit is reached.
+  UniquePtr<ProfileBufferChunk> mReleasedChunks;
+
+  // This may hold chunks that were released then slated for destruction, they
+  // will be reused next time an allocation would have been needed.
+  UniquePtr<ProfileBufferChunk> mRecycledChunks;
+
+  // Optional callback used to notify the user when a chunk is about to be
+  // destroyed or recycled. (The data content is always destroyed, but the chunk
+  // container may be reused.)
+  std::function<void(const ProfileBufferChunk&)> mChunkDestroyedCallback;
+
+  // Callback set from `RequestChunk()`, until it is serviced in
+  // `FulfillChunkRequests()`. There can only be one request in flight.
+  std::function<void(UniquePtr<ProfileBufferChunk>)> mChunkReceiver;
+
+  // Separate mutex guarding mUpdateCallback, so that it may be invoked outside
+  // of the main buffer `mMutex`.
+  mutable baseprofiler::detail::BaseProfilerMutex mUpdateCallbackMutex;
+
+  UpdateCallback mUpdateCallback;
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileBufferChunkManagerWithLocalLimit_h
diff --git a/mozglue/baseprofiler/public/ProfileBufferControlledChunkManager.h b/mozglue/baseprofiler/public/ProfileBufferControlledChunkManager.h
new file mode 100644
index 0000000000..45b39b163c
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileBufferControlledChunkManager.h
@@ -0,0 +1,203 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferControlledChunkManager_h
+#define ProfileBufferControlledChunkManager_h
+
+#include "mozilla/ProfileBufferChunk.h"
+
+#include <functional>
+#include <vector>
+
+namespace mozilla {
+
+// A "Controlled" chunk manager will provide updates about chunks that it
+// creates, releases, and destroys; and it can destroy released chunks as
+// requested.
+class ProfileBufferControlledChunkManager {
+ public:
+  using Length = ProfileBufferChunk::Length;
+
+  virtual ~ProfileBufferControlledChunkManager() = default;
+
+  // Minimum amount of chunk metadata to be transferred between processes.
+  struct ChunkMetadata {
+    // Timestamp when chunk was marked "done", which is used to:
+    // - determine its age, so the oldest one will be destroyed first,
+    // - uniquely identify this chunk in this process. (The parent process is
+    //   responsible for associating this timestamp to its process id.)
+    TimeStamp mDoneTimeStamp;
+    // Size of this chunk's buffer.
+    Length mBufferBytes;
+
+    ChunkMetadata(TimeStamp aDoneTimeStamp, Length aBufferBytes)
+        : mDoneTimeStamp(aDoneTimeStamp), mBufferBytes(aBufferBytes) {}
+  };
+
+  // Class collecting all information necessary to describe updates that
+  // happened in a chunk manager.
+  // An update can be folded into a previous update.
+  class Update {
+   public:
+    // Construct a "not-an-Update" object, which should only be used after a
+    // real update is folded into it.
+    Update() = default;
+
+    // Construct a "final" Update, which marks the end of all updates from a
+    // chunk manager.
+    explicit Update(decltype(nullptr)) : mUnreleasedBytes(FINAL) {}
+
+    // Construct an Update from the given data and released chunks.
+    // The chunk pointers may be null, and it doesn't matter if
+    // `aNewlyReleasedChunks` is already linked to `aExistingReleasedChunks` or
+    // not.
+    Update(size_t aUnreleasedBytes, size_t aReleasedBytes,
+           const ProfileBufferChunk* aExistingReleasedChunks,
+           const ProfileBufferChunk* aNewlyReleasedChunks)
+        : mUnreleasedBytes(aUnreleasedBytes),
+          mReleasedBytes(aReleasedBytes),
+          mOldestDoneTimeStamp(
+              aExistingReleasedChunks
+                  ? aExistingReleasedChunks->ChunkHeader().mDoneTimeStamp
+                  : TimeStamp{}) {
+      MOZ_RELEASE_ASSERT(
+          !IsNotUpdate(),
+          "Empty update should only be constructed with default constructor");
+      MOZ_RELEASE_ASSERT(
+          !IsFinal(),
+          "Final update should only be constructed with nullptr constructor");
+      for (const ProfileBufferChunk* chunk = aNewlyReleasedChunks; chunk;
+           chunk = chunk->GetNext()) {
+        mNewlyReleasedChunks.emplace_back(ChunkMetadata{
+            chunk->ChunkHeader().mDoneTimeStamp, chunk->BufferBytes()});
+      }
+    }
+
+    // Construct an Update from raw data.
+    // This may be used to re-construct an Update that was previously
+    // serialized.
+    Update(size_t aUnreleasedBytes, size_t aReleasedBytes,
+           TimeStamp aOldestDoneTimeStamp,
+           std::vector<ChunkMetadata>&& aNewlyReleasedChunks)
+        : mUnreleasedBytes(aUnreleasedBytes),
+          mReleasedBytes(aReleasedBytes),
+          mOldestDoneTimeStamp(aOldestDoneTimeStamp),
+          mNewlyReleasedChunks(std::move(aNewlyReleasedChunks)) {}
+
+    // Clear the Update completely and return it to a "not-an-Update" state.
+    void Clear() {
+      mUnreleasedBytes = NO_UPDATE;
+      mReleasedBytes = 0;
+      mOldestDoneTimeStamp = TimeStamp{};
+      mNewlyReleasedChunks.clear();
+    }
+
+    bool IsNotUpdate() const { return mUnreleasedBytes == NO_UPDATE; }
+
+    bool IsFinal() const { return mUnreleasedBytes == FINAL; }
+
+    size_t UnreleasedBytes() const {
+      MOZ_RELEASE_ASSERT(!IsNotUpdate(),
+                         "Cannot access UnreleasedBytes from empty update");
+      MOZ_RELEASE_ASSERT(!IsFinal(),
+                         "Cannot access UnreleasedBytes from final update");
+      return mUnreleasedBytes;
+    }
+
+    size_t ReleasedBytes() const {
+      MOZ_RELEASE_ASSERT(!IsNotUpdate(),
+                         "Cannot access ReleasedBytes from empty update");
+      MOZ_RELEASE_ASSERT(!IsFinal(),
+                         "Cannot access ReleasedBytes from final update");
+      return mReleasedBytes;
+    }
+
+    TimeStamp OldestDoneTimeStamp() const {
+      MOZ_RELEASE_ASSERT(!IsNotUpdate(),
+                         "Cannot access OldestDoneTimeStamp from empty update");
+      MOZ_RELEASE_ASSERT(!IsFinal(),
+                         "Cannot access OldestDoneTimeStamp from final update");
+      return mOldestDoneTimeStamp;
+    }
+
+    const std::vector<ChunkMetadata>& NewlyReleasedChunksRef() const {
+      MOZ_RELEASE_ASSERT(
+          !IsNotUpdate(),
+          "Cannot access NewlyReleasedChunksRef from empty update");
+      MOZ_RELEASE_ASSERT(
+          !IsFinal(), "Cannot access NewlyReleasedChunksRef from final update");
+      return mNewlyReleasedChunks;
+    }
+
+    // Fold a later update into this one.
+    void Fold(Update&& aNewUpdate) {
+      MOZ_ASSERT(
+          !IsFinal() || aNewUpdate.IsFinal(),
+          "There shouldn't be another non-final update after the final update");
+
+      if (IsNotUpdate() || aNewUpdate.IsFinal()) {
+        // We were empty, or the new update is the final update, we just switch
+        // to that new update.
+        *this = std::move(aNewUpdate);
+        return;
+      }
+
+      mUnreleasedBytes = aNewUpdate.mUnreleasedBytes;
+      mReleasedBytes = aNewUpdate.mReleasedBytes;
+      if (!aNewUpdate.mOldestDoneTimeStamp.IsNull()) {
+        MOZ_ASSERT(mOldestDoneTimeStamp.IsNull() ||
+                   mOldestDoneTimeStamp <= aNewUpdate.mOldestDoneTimeStamp);
+        mOldestDoneTimeStamp = aNewUpdate.mOldestDoneTimeStamp;
+        auto it = mNewlyReleasedChunks.begin();
+        while (it != mNewlyReleasedChunks.end() &&
+               it->mDoneTimeStamp < mOldestDoneTimeStamp) {
+          it = mNewlyReleasedChunks.erase(it);
+        }
+      }
+      if (!aNewUpdate.mNewlyReleasedChunks.empty()) {
+        mNewlyReleasedChunks.reserve(mNewlyReleasedChunks.size() +
+                                     aNewUpdate.mNewlyReleasedChunks.size());
+        mNewlyReleasedChunks.insert(mNewlyReleasedChunks.end(),
+                                    aNewUpdate.mNewlyReleasedChunks.begin(),
+                                    aNewUpdate.mNewlyReleasedChunks.end());
+      }
+    }
+
+   private:
+    static const size_t NO_UPDATE = size_t(-1);
+    static const size_t FINAL = size_t(-2);
+
+    size_t mUnreleasedBytes = NO_UPDATE;
+    size_t mReleasedBytes = 0;
+    TimeStamp mOldestDoneTimeStamp;
+    std::vector<ChunkMetadata> mNewlyReleasedChunks;
+  };
+
+  using UpdateCallback = std::function<void(Update&&)>;
+
+  // This *may* be set (or reset) by an object that needs to know about all
+  // chunk updates that happen in this manager. The main use will be to
+  // coordinate the global memory usage of Firefox.
+  // If a non-empty callback is given, it will be immediately invoked with the
+  // current state.
+  // When the callback is about to be destroyed (by overwriting it here, or in
+  // the class destructor), it will be invoked one last time with an empty
+  // update.
+  // Note that the callback (even the first current-state callback) will be
+  // invoked from inside a locked scope, so it should *not* call other functions
+  // of the chunk manager. A side benefit of this locking is that it guarantees
+  // that no two invocations can overlap.
+  virtual void SetUpdateCallback(UpdateCallback&& aUpdateCallback) = 0;
+
+  // This is a request to destroy all chunks before the given timestamp.
+  // This timestamp should be one that was given in a previous UpdateCallback
+  // call. Obviously, only released chunks can be destroyed.
+  virtual void DestroyChunksAtOrBefore(TimeStamp aDoneTimeStamp) = 0;
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileBufferControlledChunkManager_h
diff --git a/mozglue/baseprofiler/public/ProfileBufferEntryKinds.h b/mozglue/baseprofiler/public/ProfileBufferEntryKinds.h
new file mode 100644
index 0000000000..c8280a92d7
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileBufferEntryKinds.h
@@ -0,0 +1,94 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferEntryKinds_h
+#define ProfileBufferEntryKinds_h
+
+#include <cstdint>
+
+namespace mozilla {
+
+// This is equal to sizeof(double), which is the largest non-char variant in
+// |u|.
+static constexpr size_t ProfileBufferEntryNumChars = 8;
+
+// NOTE!  If you add entries, you need to verify if they need to be added to the
+// switch statement in DuplicateLastSample!
+// This will evaluate the MACRO with (KIND, TYPE, SIZE)
+#define FOR_EACH_PROFILE_BUFFER_ENTRY_KIND(MACRO)                 \
+  MACRO(CategoryPair, int, sizeof(int))                           \
+  MACRO(CollectionStart, double, sizeof(double))                  \
+  MACRO(CollectionEnd, double, sizeof(double))                    \
+  MACRO(Label, const char*, sizeof(const char*))                  \
+  MACRO(FrameFlags, uint64_t, sizeof(uint64_t))                   \
+  MACRO(DynamicStringFragment, char*, ProfileBufferEntryNumChars) \
+  MACRO(JitReturnAddr, void*, sizeof(void*))                      \
+  MACRO(InnerWindowID, uint64_t, sizeof(uint64_t))                \
+  MACRO(LineNumber, int, sizeof(int))                             \
+  MACRO(ColumnNumber, int, sizeof(int))                           \
+  MACRO(NativeLeafAddr, void*, sizeof(void*))                     \
+  MACRO(Pause, double, sizeof(double))                            \
+  MACRO(Resume, double, sizeof(double))                           \
+  MACRO(PauseSampling, double, sizeof(double))                    \
+  MACRO(ResumeSampling, double, sizeof(double))                   \
+  MACRO(Responsiveness, double, sizeof(double))                   \
+  MACRO(ThreadId, int, sizeof(int))                               \
+  MACRO(Time, double, sizeof(double))                             \
+  MACRO(TimeBeforeCompactStack, double, sizeof(double))           \
+  MACRO(CounterId, void*, sizeof(void*))                          \
+  MACRO(CounterKey, uint64_t, sizeof(uint64_t))                   \
+  MACRO(Number, uint64_t, sizeof(uint64_t))                       \
+  MACRO(Count, int64_t, sizeof(int64_t))                          \
+  MACRO(ProfilerOverheadTime, double, sizeof(double))             \
+  MACRO(ProfilerOverheadDuration, double, sizeof(double))
+
+// The `Kind` is a single byte identifying the type of data that is actually
+// stored in a `ProfileBufferEntry`, as per the list in
+// `FOR_EACH_PROFILE_BUFFER_ENTRY_KIND`.
+//
+// This byte is also used to identify entries in ProfileChunkedBuffer blocks,
+// for both "legacy" entries that do contain a `ProfileBufferEntry`, and for
+// new types of entries that may carry more data of different types.
+// TODO: Eventually each type of "legacy" entry should be replaced with newer,
+// more efficient kinds of entries (e.g., stack frames could be stored in one
+// bigger entry, instead of multiple `ProfileBufferEntry`s); then we could
+// discard `ProfileBufferEntry` and move this enum to a more appropriate spot.
+using ProfileBufferEntryKindUnderlyingType = uint8_t;
+
+enum class ProfileBufferEntryKind : ProfileBufferEntryKindUnderlyingType {
+  INVALID = 0,
+#define KIND(KIND, TYPE, SIZE) KIND,
+  FOR_EACH_PROFILE_BUFFER_ENTRY_KIND(KIND)
+#undef KIND
+
+  // Any value under `LEGACY_LIMIT` represents a `ProfileBufferEntry`.
+  LEGACY_LIMIT,
+
+  // Any value starting here does *not* represent a `ProfileBufferEntry` and
+  // requires separate decoding and handling.
+
+  // Markers and their data.
+  Marker = LEGACY_LIMIT,
+
+  // Entry with "running times", such as CPU usage measurements.
+  // Optional between TimeBeforeCompactStack and CompactStack.
+  RunningTimes,
+
+  // Optional between TimeBeforeCompactStack and CompactStack.
+  UnresponsiveDurationMs,
+
+  // Collection of legacy stack entries, must follow a ThreadId and
+  // TimeBeforeCompactStack (which are not included in the CompactStack;
+  // TimeBeforeCompactStack is equivalent to Time, but indicates that a
+  // CompactStack follows shortly afterwards).
+  CompactStack,
+
+  MODERN_LIMIT
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileBufferEntryKinds_h
diff --git a/mozglue/baseprofiler/public/ProfileBufferEntrySerialization.h b/mozglue/baseprofiler/public/ProfileBufferEntrySerialization.h
new file mode 100644
index 0000000000..267b99f10d
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileBufferEntrySerialization.h
@@ -0,0 +1,1166 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferEntrySerialization_h
+#define ProfileBufferEntrySerialization_h
+
+#include "mozilla/Assertions.h"
+#include "mozilla/leb128iterator.h"
+#include "mozilla/Likely.h"
+#include "mozilla/Maybe.h"
+#include "mozilla/ProfileBufferIndex.h"
+#include "mozilla/Span.h"
+#include "mozilla/Tuple.h"
+#include "mozilla/UniquePtrExtensions.h"
+#include "mozilla/Unused.h"
+#include "mozilla/Variant.h"
+
+#include <string>
+#include <tuple>
+
+namespace mozilla {
+
+class ProfileBufferEntryWriter;
+
+// Iterator-like class used to read from an entry.
+// An entry may be split in two memory segments (e.g., the ends of a ring
+// buffer, or two chunks of a chunked buffer); it doesn't deal with this
+// underlying buffer, but only with one or two spans pointing at the space
+// where the entry lives.
+class ProfileBufferEntryReader {
+ public:
+  using Byte = uint8_t;
+  using Length = uint32_t;
+
+  using SpanOfConstBytes = Span<const Byte>;
+
+  // Class to be specialized for types to be read from a profile buffer entry.
+  // See common specializations at the bottom of this header.
+  // The following static functions must be provided:
+  //   static void ReadInto(EntryReader aER&, T& aT)
+  //   {
+  //     /* Call `aER.ReadX(...)` function to deserialize into aT, be sure to
+  //        read exactly `Bytes(aT)`! */
+  //   }
+  //   static T Read(EntryReader& aER) {
+  //     /* Call `aER.ReadX(...)` function to deserialize and return a `T`, be
+  //        sure to read exactly `Bytes(returned value)`! */
+  //   }
+  template <typename T>
+  struct Deserializer;
+
+  ProfileBufferEntryReader() = default;
+
+  // Reader over one Span.
+  ProfileBufferEntryReader(SpanOfConstBytes aSpan,
+                           ProfileBufferBlockIndex aCurrentBlockIndex,
+                           ProfileBufferBlockIndex aNextBlockIndex)
+      : mCurrentSpan(aSpan),
+        mNextSpanOrEmpty(aSpan.Last(0)),
+        mCurrentBlockIndex(aCurrentBlockIndex),
+        mNextBlockIndex(aNextBlockIndex) {
+    // 2nd internal Span points at the end of the 1st internal Span, to enforce
+    // invariants.
+    CheckInvariants();
+  }
+
+  // Reader over two Spans, the second one must not be empty.
+  ProfileBufferEntryReader(SpanOfConstBytes aSpanHead,
+                           SpanOfConstBytes aSpanTail,
+                           ProfileBufferBlockIndex aCurrentBlockIndex,
+                           ProfileBufferBlockIndex aNextBlockIndex)
+      : mCurrentSpan(aSpanHead),
+        mNextSpanOrEmpty(aSpanTail),
+        mCurrentBlockIndex(aCurrentBlockIndex),
+        mNextBlockIndex(aNextBlockIndex) {
+    MOZ_RELEASE_ASSERT(!mNextSpanOrEmpty.IsEmpty());
+    if (MOZ_UNLIKELY(mCurrentSpan.IsEmpty())) {
+      // First span is already empty, skip it.
+      mCurrentSpan = mNextSpanOrEmpty;
+      mNextSpanOrEmpty = mNextSpanOrEmpty.Last(0);
+    }
+    CheckInvariants();
+  }
+
+  // Allow copying, which is needed when used as an iterator in some std
+  // functions (e.g., string assignment), and to occasionally backtrack.
+  // Be aware that the main profile buffer APIs give a reference to an entry
+  // reader, and expect that reader to advance to the end of the entry, so don't
+  // just advance copies!
+  ProfileBufferEntryReader(const ProfileBufferEntryReader&) = default;
+  ProfileBufferEntryReader& operator=(const ProfileBufferEntryReader&) =
+      default;
+
+  // Don't =default moving, as it doesn't bring any benefit in this class.
+
+  [[nodiscard]] Length RemainingBytes() const {
+    return mCurrentSpan.LengthBytes() + mNextSpanOrEmpty.LengthBytes();
+  }
+
+  void SetRemainingBytes(Length aBytes) {
+    MOZ_RELEASE_ASSERT(aBytes <= RemainingBytes());
+    if (aBytes <= mCurrentSpan.LengthBytes()) {
+      mCurrentSpan = mCurrentSpan.First(aBytes);
+      mNextSpanOrEmpty = mCurrentSpan.Last(0);
+    } else {
+      mNextSpanOrEmpty =
+          mNextSpanOrEmpty.First(aBytes - mCurrentSpan.LengthBytes());
+    }
+  }
+
+  [[nodiscard]] ProfileBufferBlockIndex CurrentBlockIndex() const {
+    return mCurrentBlockIndex;
+  }
+
+  [[nodiscard]] ProfileBufferBlockIndex NextBlockIndex() const {
+    return mNextBlockIndex;
+  }
+
+  // Create a reader of size zero, pointing at aOffset past the current position
+  // of this Reader, so it can be used as end iterator.
+  [[nodiscard]] ProfileBufferEntryReader EmptyIteratorAtOffset(
+      Length aOffset) const {
+    MOZ_RELEASE_ASSERT(aOffset <= RemainingBytes());
+    if (MOZ_LIKELY(aOffset < mCurrentSpan.LengthBytes())) {
+      // aOffset is before the end of mCurrentSpan.
+      return ProfileBufferEntryReader(mCurrentSpan.Subspan(aOffset, 0),
+                                      mCurrentBlockIndex, mNextBlockIndex);
+    }
+    // aOffset is right at the end of mCurrentSpan, or inside mNextSpanOrEmpty.
+    return ProfileBufferEntryReader(
+        mNextSpanOrEmpty.Subspan(aOffset - mCurrentSpan.LengthBytes(), 0),
+        mCurrentBlockIndex, mNextBlockIndex);
+  }
+
+  // Be like a limited input iterator, with only `*`, prefix-`++`, `==`, `!=`.
+  // These definitions are expected by std functions, to recognize this as an
+  // iterator. See https://en.cppreference.com/w/cpp/iterator/iterator_traits
+  using difference_type = std::make_signed_t<Length>;
+  using value_type = Byte;
+  using pointer = const Byte*;
+  using reference = const Byte&;
+  using iterator_category = std::input_iterator_tag;
+
+  [[nodiscard]] const Byte& operator*() {
+    // Assume the caller will read from the returned reference (and not just
+    // take the address).
+    MOZ_RELEASE_ASSERT(mCurrentSpan.LengthBytes() >= 1);
+    return *(mCurrentSpan.Elements());
+  }
+
+  ProfileBufferEntryReader& operator++() {
+    MOZ_RELEASE_ASSERT(mCurrentSpan.LengthBytes() >= 1);
+    if (MOZ_LIKELY(mCurrentSpan.LengthBytes() > 1)) {
+      // More than 1 byte left in mCurrentSpan, just eat it.
+      mCurrentSpan = mCurrentSpan.From(1);
+    } else {
+      // mCurrentSpan will be empty, move mNextSpanOrEmpty to mCurrentSpan.
+      mCurrentSpan = mNextSpanOrEmpty;
+      mNextSpanOrEmpty = mNextSpanOrEmpty.Last(0);
+    }
+    CheckInvariants();
+    return *this;
+  }
+
+  ProfileBufferEntryReader& operator+=(Length aBytes) {
+    MOZ_RELEASE_ASSERT(aBytes <= RemainingBytes());
+    if (MOZ_LIKELY(aBytes <= mCurrentSpan.LengthBytes())) {
+      // All bytes are in mCurrentSpan.
+      // Update mCurrentSpan past the read bytes.
+      mCurrentSpan = mCurrentSpan.From(aBytes);
+      if (mCurrentSpan.IsEmpty() && !mNextSpanOrEmpty.IsEmpty()) {
+        // Don't leave mCurrentSpan empty, move non-empty mNextSpanOrEmpty into
+        // mCurrentSpan.
+        mCurrentSpan = mNextSpanOrEmpty;
+        mNextSpanOrEmpty = mNextSpanOrEmpty.Last(0);
+      }
+    } else {
+      // mCurrentSpan does not hold enough bytes.
+      // This should only happen at most once: Only for double spans, and when
+      // data crosses the gap.
+      const Length tail =
+          aBytes - static_cast<Length>(mCurrentSpan.LengthBytes());
+      // Move mNextSpanOrEmpty to mCurrentSpan, past the data. So the next call
+      // will go back to the true case above.
+      mCurrentSpan = mNextSpanOrEmpty.From(tail);
+      mNextSpanOrEmpty = mNextSpanOrEmpty.Last(0);
+    }
+    CheckInvariants();
+    return *this;
+  }
+
+  [[nodiscard]] bool operator==(const ProfileBufferEntryReader& aOther) const {
+    return mCurrentSpan.Elements() == aOther.mCurrentSpan.Elements();
+  }
+  [[nodiscard]] bool operator!=(const ProfileBufferEntryReader& aOther) const {
+    return mCurrentSpan.Elements() != aOther.mCurrentSpan.Elements();
+  }
+
+  // Read an unsigned LEB128 number and move iterator ahead.
+  template <typename T>
+  [[nodiscard]] T ReadULEB128() {
+    return ::mozilla::ReadULEB128<T>(*this);
+  }
+
+  // Read a sequence of bytes, like memcpy.
+  void ReadBytes(void* aDest, Length aBytes) {
+    MOZ_RELEASE_ASSERT(aBytes <= RemainingBytes());
+    if (MOZ_LIKELY(aBytes <= mCurrentSpan.LengthBytes())) {
+      // All bytes are in mCurrentSpan.
+      memcpy(aDest, mCurrentSpan.Elements(), aBytes);
+      // Update mCurrentSpan past the read bytes.
+      mCurrentSpan = mCurrentSpan.From(aBytes);
+      if (mCurrentSpan.IsEmpty() && !mNextSpanOrEmpty.IsEmpty()) {
+        // Don't leave mCurrentSpan empty, move non-empty mNextSpanOrEmpty into
+        // mCurrentSpan.
+        mCurrentSpan = mNextSpanOrEmpty;
+        mNextSpanOrEmpty = mNextSpanOrEmpty.Last(0);
+      }
+    } else {
+      // mCurrentSpan does not hold enough bytes.
+      // This should only happen at most once: Only for double spans, and when
+      // data crosses the gap.
+      // Split data between the end of mCurrentSpan and the beginning of
+      // mNextSpanOrEmpty.
+      memcpy(aDest, mCurrentSpan.Elements(), mCurrentSpan.LengthBytes());
+      const Length tail =
+          aBytes - static_cast<Length>(mCurrentSpan.LengthBytes());
+      memcpy(reinterpret_cast<Byte*>(aDest) + mCurrentSpan.LengthBytes(),
+             mNextSpanOrEmpty.Elements(), tail);
+      // Move mNextSpanOrEmpty to mCurrentSpan, past the data. So the next call
+      // will go back to the true case above.
+      mCurrentSpan = mNextSpanOrEmpty.From(tail);
+      mNextSpanOrEmpty = mNextSpanOrEmpty.Last(0);
+    }
+    CheckInvariants();
+  }
+
+  template <typename T>
+  void ReadIntoObject(T& aObject) {
+    Deserializer<T>::ReadInto(*this, aObject);
+  }
+
+  // Read into one or more objects, sequentially.
+  // `EntryReader::ReadIntoObjects()` with nothing is implicitly allowed, this
+  // could be useful for generic programming.
+  template <typename... Ts>
+  void ReadIntoObjects(Ts&... aTs) {
+    (ReadIntoObject(aTs), ...);
+  }
+
+  // Read data as an object and move iterator ahead.
+  template <typename T>
+  [[nodiscard]] T ReadObject() {
+    T ob = Deserializer<T>::Read(*this);
+    return ob;
+  }
+
+ private:
+  friend class ProfileBufferEntryWriter;
+
+  // Invariants:
+  // - mCurrentSpan cannot be empty unless mNextSpanOrEmpty is also empty. So
+  //   mCurrentSpan always points at the next byte to read or the end.
+  // - If mNextSpanOrEmpty is empty, it points at the end of mCurrentSpan. So
+  //   when reaching the end of mCurrentSpan, we can blindly move
+  //   mNextSpanOrEmpty to mCurrentSpan and keep the invariants.
+  SpanOfConstBytes mCurrentSpan;
+  SpanOfConstBytes mNextSpanOrEmpty;
+  ProfileBufferBlockIndex mCurrentBlockIndex;
+  ProfileBufferBlockIndex mNextBlockIndex;
+
+  void CheckInvariants() const {
+    MOZ_ASSERT(!mCurrentSpan.IsEmpty() || mNextSpanOrEmpty.IsEmpty());
+    MOZ_ASSERT(!mNextSpanOrEmpty.IsEmpty() ||
+               (mNextSpanOrEmpty == mCurrentSpan.Last(0)));
+  }
+};
+
+// Iterator-like class used to write into an entry.
+// An entry may be split in two memory segments (e.g., the ends of a ring
+// buffer, or two chunks of a chunked buffer); it doesn't deal with this
+// underlying buffer, but only with one or two spans pointing at the space
+// reserved for the entry.
+class ProfileBufferEntryWriter {
+ public:
+  using Byte = uint8_t;
+  using Length = uint32_t;
+
+  using SpanOfBytes = Span<Byte>;
+
+  // Class to be specialized for types to be written in an entry.
+  // See common specializations at the bottom of this header.
+  // The following static functions must be provided:
+  //   static Length Bytes(const T& aT) {
+  //     /* Return number of bytes that will be written. */
+  //   }
+  //   static void Write(ProfileBufferEntryWriter& aEW,
+  //                     const T& aT) {
+  //     /* Call `aEW.WriteX(...)` functions to serialize aT, be sure to write
+  //        exactly `Bytes(aT)` bytes! */
+  //   }
+  template <typename T>
+  struct Serializer;
+
+  ProfileBufferEntryWriter() = default;
+
+  ProfileBufferEntryWriter(SpanOfBytes aSpan,
+                           ProfileBufferBlockIndex aCurrentBlockIndex,
+                           ProfileBufferBlockIndex aNextBlockIndex)
+      : mCurrentSpan(aSpan),
+        mCurrentBlockIndex(aCurrentBlockIndex),
+        mNextBlockIndex(aNextBlockIndex) {}
+
+  ProfileBufferEntryWriter(SpanOfBytes aSpanHead, SpanOfBytes aSpanTail,
+                           ProfileBufferBlockIndex aCurrentBlockIndex,
+                           ProfileBufferBlockIndex aNextBlockIndex)
+      : mCurrentSpan(aSpanHead),
+        mNextSpanOrEmpty(aSpanTail),
+        mCurrentBlockIndex(aCurrentBlockIndex),
+        mNextBlockIndex(aNextBlockIndex) {
+    // Either:
+    // - mCurrentSpan is not empty, OR
+    // - mNextSpanOrEmpty is empty if mNextSpanOrEmpty is empty as well.
+    MOZ_RELEASE_ASSERT(!mCurrentSpan.IsEmpty() || mNextSpanOrEmpty.IsEmpty());
+  }
+
+  // Disable copying and moving, so we can't have multiple writing heads.
+  ProfileBufferEntryWriter(const ProfileBufferEntryWriter&) = delete;
+  ProfileBufferEntryWriter& operator=(const ProfileBufferEntryWriter&) = delete;
+  ProfileBufferEntryWriter(ProfileBufferEntryWriter&&) = delete;
+  ProfileBufferEntryWriter& operator=(ProfileBufferEntryWriter&&) = delete;
+
+  void Set() {
+    mCurrentSpan = SpanOfBytes{};
+    mNextSpanOrEmpty = SpanOfBytes{};
+    mCurrentBlockIndex = nullptr;
+    mNextBlockIndex = nullptr;
+  }
+
+  void Set(SpanOfBytes aSpan, ProfileBufferBlockIndex aCurrentBlockIndex,
+           ProfileBufferBlockIndex aNextBlockIndex) {
+    mCurrentSpan = aSpan;
+    mNextSpanOrEmpty = SpanOfBytes{};
+    mCurrentBlockIndex = aCurrentBlockIndex;
+    mNextBlockIndex = aNextBlockIndex;
+  }
+
+  void Set(SpanOfBytes aSpan0, SpanOfBytes aSpan1,
+           ProfileBufferBlockIndex aCurrentBlockIndex,
+           ProfileBufferBlockIndex aNextBlockIndex) {
+    mCurrentSpan = aSpan0;
+    mNextSpanOrEmpty = aSpan1;
+    mCurrentBlockIndex = aCurrentBlockIndex;
+    mNextBlockIndex = aNextBlockIndex;
+    // Either:
+    // - mCurrentSpan is not empty, OR
+    // - mNextSpanOrEmpty is empty if mNextSpanOrEmpty is empty as well.
+    MOZ_RELEASE_ASSERT(!mCurrentSpan.IsEmpty() || mNextSpanOrEmpty.IsEmpty());
+  }
+
+  [[nodiscard]] Length RemainingBytes() const {
+    return mCurrentSpan.LengthBytes() + mNextSpanOrEmpty.LengthBytes();
+  }
+
+  [[nodiscard]] ProfileBufferBlockIndex CurrentBlockIndex() const {
+    return mCurrentBlockIndex;
+  }
+
+  [[nodiscard]] ProfileBufferBlockIndex NextBlockIndex() const {
+    return mNextBlockIndex;
+  }
+
+  // Be like a limited output iterator, with only `*` and prefix-`++`.
+  // These definitions are expected by std functions, to recognize this as an
+  // iterator. See https://en.cppreference.com/w/cpp/iterator/iterator_traits
+  using value_type = Byte;
+  using pointer = Byte*;
+  using reference = Byte&;
+  using iterator_category = std::output_iterator_tag;
+
+  [[nodiscard]] Byte& operator*() {
+    MOZ_RELEASE_ASSERT(RemainingBytes() >= 1);
+    return *(
+        (MOZ_LIKELY(!mCurrentSpan.IsEmpty()) ? mCurrentSpan : mNextSpanOrEmpty)
+            .Elements());
+  }
+
+  ProfileBufferEntryWriter& operator++() {
+    if (MOZ_LIKELY(mCurrentSpan.LengthBytes() >= 1)) {
+      // There is at least 1 byte in mCurrentSpan, eat it.
+      mCurrentSpan = mCurrentSpan.From(1);
+    } else {
+      // mCurrentSpan is empty, move mNextSpanOrEmpty (past the first byte) to
+      // mCurrentSpan.
+      MOZ_RELEASE_ASSERT(mNextSpanOrEmpty.LengthBytes() >= 1);
+      mCurrentSpan = mNextSpanOrEmpty.From(1);
+      mNextSpanOrEmpty = mNextSpanOrEmpty.First(0);
+    }
+    return *this;
+  }
+
+  ProfileBufferEntryWriter& operator+=(Length aBytes) {
+    // Note: This is a rare operation. The code below is a copy of `WriteBytes`
+    // but without the `memcpy`s.
+    MOZ_RELEASE_ASSERT(aBytes <= RemainingBytes());
+    if (MOZ_LIKELY(aBytes <= mCurrentSpan.LengthBytes())) {
+      // Data fits in mCurrentSpan.
+      // Update mCurrentSpan. It may become empty, so in case of a double span,
+      // the next call will go to the false case below.
+      mCurrentSpan = mCurrentSpan.From(aBytes);
+    } else {
+      // Data does not fully fit in mCurrentSpan.
+      // This should only happen at most once: Only for double spans, and when
+      // data crosses the gap or starts there.
+      const Length tail =
+          aBytes - static_cast<Length>(mCurrentSpan.LengthBytes());
+      // Move mNextSpanOrEmpty to mCurrentSpan, past the data. So the next call
+      // will go back to the true case above.
+      mCurrentSpan = mNextSpanOrEmpty.From(tail);
+      mNextSpanOrEmpty = mNextSpanOrEmpty.First(0);
+    }
+    return *this;
+  }
+
+  // Number of bytes needed to represent `aValue` in unsigned LEB128.
+  template <typename T>
+  [[nodiscard]] static unsigned ULEB128Size(T aValue) {
+    return ::mozilla::ULEB128Size(aValue);
+  }
+
+  // Write number as unsigned LEB128 and move iterator ahead.
+  template <typename T>
+  void WriteULEB128(T aValue) {
+    ::mozilla::WriteULEB128(aValue, *this);
+  }
+
+  // Number of bytes needed to serialize objects.
+  template <typename... Ts>
+  [[nodiscard]] static Length SumBytes(const Ts&... aTs) {
+    return (0 + ... + Serializer<Ts>::Bytes(aTs));
+  }
+
+  // Write a sequence of bytes, like memcpy.
+  void WriteBytes(const void* aSrc, Length aBytes) {
+    MOZ_RELEASE_ASSERT(aBytes <= RemainingBytes());
+    if (MOZ_LIKELY(aBytes <= mCurrentSpan.LengthBytes())) {
+      // Data fits in mCurrentSpan.
+      memcpy(mCurrentSpan.Elements(), aSrc, aBytes);
+      // Update mCurrentSpan. It may become empty, so in case of a double span,
+      // the next call will go to the false case below.
+      mCurrentSpan = mCurrentSpan.From(aBytes);
+    } else {
+      // Data does not fully fit in mCurrentSpan.
+      // This should only happen at most once: Only for double spans, and when
+      // data crosses the gap or starts there.
+      // Split data between the end of mCurrentSpan and the beginning of
+      // mNextSpanOrEmpty. (mCurrentSpan could be empty, it's ok to do a memcpy
+      // because Span::Elements() is never null.)
+      memcpy(mCurrentSpan.Elements(), aSrc, mCurrentSpan.LengthBytes());
+      const Length tail =
+          aBytes - static_cast<Length>(mCurrentSpan.LengthBytes());
+      memcpy(mNextSpanOrEmpty.Elements(),
+             reinterpret_cast<const Byte*>(aSrc) + mCurrentSpan.LengthBytes(),
+             tail);
+      // Move mNextSpanOrEmpty to mCurrentSpan, past the data. So the next call
+      // will go back to the true case above.
+      mCurrentSpan = mNextSpanOrEmpty.From(tail);
+      mNextSpanOrEmpty = mNextSpanOrEmpty.First(0);
+    }
+  }
+
+  void WriteFromReader(ProfileBufferEntryReader& aReader, Length aBytes) {
+    MOZ_RELEASE_ASSERT(aBytes <= RemainingBytes());
+    MOZ_RELEASE_ASSERT(aBytes <= aReader.RemainingBytes());
+    Length read0 = std::min(
+        aBytes, static_cast<Length>(aReader.mCurrentSpan.LengthBytes()));
+    if (read0 != 0) {
+      WriteBytes(aReader.mCurrentSpan.Elements(), read0);
+    }
+    Length read1 = aBytes - read0;
+    if (read1 != 0) {
+      WriteBytes(aReader.mNextSpanOrEmpty.Elements(), read1);
+    }
+    aReader += aBytes;
+  }
+
+  // Write a single object by using the appropriate Serializer.
+  template <typename T>
+  void WriteObject(const T& aObject) {
+    Serializer<T>::Write(*this, aObject);
+  }
+
+  // Write one or more objects, sequentially.
+  // Allow `EntryWrite::WriteObjects()` with nothing, this could be useful
+  // for generic programming.
+  template <typename... Ts>
+  void WriteObjects(const Ts&... aTs) {
+    (WriteObject(aTs), ...);
+  }
+
+ private:
+  // The two spans covering the memory still to be written.
+  SpanOfBytes mCurrentSpan;
+  SpanOfBytes mNextSpanOrEmpty;
+  ProfileBufferBlockIndex mCurrentBlockIndex;
+  ProfileBufferBlockIndex mNextBlockIndex;
+};
+
+// ============================================================================
+// Serializer and Deserializer ready-to-use specializations.
+
+// ----------------------------------------------------------------------------
+// Trivially-copyable types (default)
+
+// The default implementation works for all trivially-copyable types (e.g.,
+// PODs).
+//
+// Usage: `aEW.WriteObject(123);`.
+//
+// Raw pointers, though trivially-copyable, are explictly forbidden when writing
+// (to avoid unexpected leaks/UAFs), instead use one of
+// `WrapProfileBufferLiteralCStringPointer`, `WrapProfileBufferUnownedCString`,
+// or `WrapProfileBufferRawPointer` as needed.
+template <typename T>
+struct ProfileBufferEntryWriter::Serializer {
+  static_assert(std::is_trivially_copyable<T>::value,
+                "Serializer only works with trivially-copyable types by "
+                "default, use/add specialization for other types.");
+
+  static constexpr Length Bytes(const T&) { return sizeof(T); }
+
+  static void Write(ProfileBufferEntryWriter& aEW, const T& aT) {
+    static_assert(!std::is_pointer<T>::value,
+                  "Serializer won't write raw pointers by default, use "
+                  "WrapProfileBufferRawPointer or other.");
+    aEW.WriteBytes(&aT, sizeof(T));
+  }
+};
+
+// Usage: `aER.ReadObject<int>();` or `int x; aER.ReadIntoObject(x);`.
+template <typename T>
+struct ProfileBufferEntryReader::Deserializer {
+  static_assert(std::is_trivially_copyable<T>::value,
+                "Deserializer only works with trivially-copyable types by "
+                "default, use/add specialization for other types.");
+
+  static void ReadInto(ProfileBufferEntryReader& aER, T& aT) {
+    aER.ReadBytes(&aT, sizeof(T));
+  }
+
+  static T Read(ProfileBufferEntryReader& aER) {
+    // Note that this creates a default `T` first, and then overwrites it with
+    // bytes from the buffer. Trivially-copyable types support this without UB.
+    T ob;
+    ReadInto(aER, ob);
+    return ob;
+  }
+};
+
+// ----------------------------------------------------------------------------
+// Strip const/volatile/reference from types.
+
+// Automatically strip `const`.
+template <typename T>
+struct ProfileBufferEntryWriter::Serializer<const T>
+    : public ProfileBufferEntryWriter::Serializer<T> {};
+
+template <typename T>
+struct ProfileBufferEntryReader::Deserializer<const T>
+    : public ProfileBufferEntryReader::Deserializer<T> {};
+
+// Automatically strip `volatile`.
+template <typename T>
+struct ProfileBufferEntryWriter::Serializer<volatile T>
+    : public ProfileBufferEntryWriter::Serializer<T> {};
+
+template <typename T>
+struct ProfileBufferEntryReader::Deserializer<volatile T>
+    : public ProfileBufferEntryReader::Deserializer<T> {};
+
+// Automatically strip `lvalue-reference`.
+template <typename T>
+struct ProfileBufferEntryWriter::Serializer<T&>
+    : public ProfileBufferEntryWriter::Serializer<T> {};
+
+template <typename T>
+struct ProfileBufferEntryReader::Deserializer<T&>
+    : public ProfileBufferEntryReader::Deserializer<T> {};
+
+// Automatically strip `rvalue-reference`.
+template <typename T>
+struct ProfileBufferEntryWriter::Serializer<T&&>
+    : public ProfileBufferEntryWriter::Serializer<T> {};
+
+template <typename T>
+struct ProfileBufferEntryReader::Deserializer<T&&>
+    : public ProfileBufferEntryReader::Deserializer<T> {};
+
+// ----------------------------------------------------------------------------
+// ProfileBufferBlockIndex
+
+// ProfileBufferBlockIndex, serialized as the underlying value.
+template <>
+struct ProfileBufferEntryWriter::Serializer<ProfileBufferBlockIndex> {
+  static constexpr Length Bytes(const ProfileBufferBlockIndex& aBlockIndex) {
+    return sizeof(ProfileBufferBlockIndex);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const ProfileBufferBlockIndex& aBlockIndex) {
+    aEW.WriteBytes(&aBlockIndex, sizeof(aBlockIndex));
+  }
+};
+
+template <>
+struct ProfileBufferEntryReader::Deserializer<ProfileBufferBlockIndex> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       ProfileBufferBlockIndex& aBlockIndex) {
+    aER.ReadBytes(&aBlockIndex, sizeof(aBlockIndex));
+  }
+
+  static ProfileBufferBlockIndex Read(ProfileBufferEntryReader& aER) {
+    ProfileBufferBlockIndex blockIndex;
+    ReadInto(aER, blockIndex);
+    return blockIndex;
+  }
+};
+
+// ----------------------------------------------------------------------------
+// Literal C string pointer
+
+// Wrapper around a pointer to a literal C string.
+template <size_t NonTerminalCharacters>
+struct ProfileBufferLiteralCStringPointer {
+  const char* mCString;
+};
+
+// Wrap a pointer to a literal C string.
+template <size_t CharactersIncludingTerminal>
+ProfileBufferLiteralCStringPointer<CharactersIncludingTerminal - 1>
+WrapProfileBufferLiteralCStringPointer(
+    const char (&aCString)[CharactersIncludingTerminal]) {
+  return {aCString};
+}
+
+// Literal C strings, serialized as the raw pointer because it is unique and
+// valid for the whole program lifetime.
+//
+// Usage: `aEW.WriteObject(WrapProfileBufferLiteralCStringPointer("hi"));`.
+//
+// No deserializer is provided for this type, instead it must be deserialized as
+// a raw pointer: `aER.ReadObject<const char*>();`
+template <size_t CharactersIncludingTerminal>
+struct ProfileBufferEntryReader::Deserializer<
+    ProfileBufferLiteralCStringPointer<CharactersIncludingTerminal>> {
+  static constexpr Length Bytes(
+      const ProfileBufferLiteralCStringPointer<CharactersIncludingTerminal>&) {
+    // We're only storing a pointer, its size is independent from the pointer
+    // value.
+    return sizeof(const char*);
+  }
+
+  static void Write(
+      ProfileBufferEntryWriter& aEW,
+      const ProfileBufferLiteralCStringPointer<CharactersIncludingTerminal>&
+          aWrapper) {
+    // Write the pointer *value*, not the string contents.
+    aEW.WriteBytes(aWrapper.mCString, sizeof(aWrapper.mCString));
+  }
+};
+
+// ----------------------------------------------------------------------------
+// C string contents
+
+// Wrapper around a pointer to a C string whose contents will be serialized.
+struct ProfileBufferUnownedCString {
+  const char* mCString;
+};
+
+// Wrap a pointer to a C string whose contents will be serialized.
+inline ProfileBufferUnownedCString WrapProfileBufferUnownedCString(
+    const char* aCString) {
+  return {aCString};
+}
+
+// The contents of a (probably) unowned C string are serialized as the number of
+// characters (encoded as ULEB128) and all the characters in the string. The
+// terminal '\0' is omitted.
+//
+// Usage: `aEW.WriteObject(WrapProfileBufferUnownedCString(str.c_str()))`.
+//
+// No deserializer is provided for this pointer type, instead it must be
+// deserialized as one of the other string types that manages its contents,
+// e.g.: `aER.ReadObject<std::string>();`
+template <>
+struct ProfileBufferEntryWriter::Serializer<ProfileBufferUnownedCString> {
+  static Length Bytes(const ProfileBufferUnownedCString& aS) {
+    const auto len = strlen(aS.mCString);
+    return ULEB128Size(len) + len;
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const ProfileBufferUnownedCString& aS) {
+    const auto len = strlen(aS.mCString);
+    aEW.WriteULEB128(len);
+    aEW.WriteBytes(aS.mCString, len);
+  }
+};
+
+// ----------------------------------------------------------------------------
+// Raw pointers
+
+// Wrapper around a pointer to be serialized as the raw pointer value.
+template <typename T>
+struct ProfileBufferRawPointer {
+  T* mRawPointer;
+};
+
+// Wrap a pointer to be serialized as the raw pointer value.
+template <typename T>
+ProfileBufferRawPointer<T> WrapProfileBufferRawPointer(T* aRawPointer) {
+  return {aRawPointer};
+}
+
+// Raw pointers are serialized as the raw pointer value.
+//
+// Usage: `aEW.WriteObject(WrapProfileBufferRawPointer(ptr));`
+//
+// The wrapper is compulsory when writing pointers (to avoid unexpected
+// leaks/UAFs), but reading can be done straight into a raw pointer object,
+// e.g.: `aER.ReadObject<Foo*>;`.
+template <typename T>
+struct ProfileBufferEntryWriter::Serializer<ProfileBufferRawPointer<T>> {
+  template <typename U>
+  static constexpr Length Bytes(const U&) {
+    return sizeof(T*);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const ProfileBufferRawPointer<T>& aWrapper) {
+    aEW.WriteBytes(&aWrapper.mRawPointer, sizeof(aWrapper.mRawPointer));
+  }
+};
+
+// Usage: `aER.ReadObject<Foo*>;` or `Foo* p; aER.ReadIntoObject(p);`, no
+// wrapper necessary.
+template <typename T>
+struct ProfileBufferEntryReader::Deserializer<ProfileBufferRawPointer<T>> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       ProfileBufferRawPointer<T>& aPtr) {
+    aER.ReadBytes(&aPtr.mRawPointer, sizeof(aPtr));
+  }
+
+  static ProfileBufferRawPointer<T> Read(ProfileBufferEntryReader& aER) {
+    ProfileBufferRawPointer<T> rawPointer;
+    ReadInto(aER, rawPointer);
+    return rawPointer;
+  }
+};
+
+// ----------------------------------------------------------------------------
+// std::string contents
+
+// std::string contents are serialized as the number of characters (encoded as
+// ULEB128) and all the characters in the string. The terminal '\0' is omitted.
+//
+// Usage: `std::string s = ...; aEW.WriteObject(s);`
+template <typename CHAR>
+struct ProfileBufferEntryWriter::Serializer<std::basic_string<CHAR>> {
+  static Length Bytes(const std::basic_string<CHAR>& aS) {
+    const Length len = static_cast<Length>(aS.length());
+    return ULEB128Size(len) + len;
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const std::basic_string<CHAR>& aS) {
+    const Length len = static_cast<Length>(aS.length());
+    aEW.WriteULEB128(len);
+    aEW.WriteBytes(aS.c_str(), len * sizeof(CHAR));
+  }
+};
+
+// Usage: `std::string s = aEW.ReadObject<std::string>(s);` or
+// `std::string s; aER.ReadIntoObject(s);`
+template <typename CHAR>
+struct ProfileBufferEntryReader::Deserializer<std::basic_string<CHAR>> {
+  static void ReadCharsInto(ProfileBufferEntryReader& aER,
+                            std::basic_string<CHAR>& aS, size_t aLength) {
+    // Assign to `aS` by using iterators.
+    // (`aER+0` so we get the same iterator type as `aER+len`.)
+    aS.assign(aER, aER.EmptyIteratorAtOffset(aLength));
+    aER += aLength;
+  }
+
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       std::basic_string<CHAR>& aS) {
+    ReadCharsInto(
+        aER, aS,
+        aER.ReadULEB128<typename std::basic_string<CHAR>::size_type>());
+  }
+
+  static std::basic_string<CHAR> ReadChars(ProfileBufferEntryReader& aER,
+                                           size_t aLength) {
+    // Construct a string by using iterators.
+    // (`aER+0` so we get the same iterator type as `aER+len`.)
+    std::basic_string<CHAR> s(aER, aER.EmptyIteratorAtOffset(aLength));
+    aER += aLength;
+    return s;
+  }
+
+  static std::basic_string<CHAR> Read(ProfileBufferEntryReader& aER) {
+    return ReadChars(
+        aER, aER.ReadULEB128<typename std::basic_string<CHAR>::size_type>());
+  }
+};
+
+// ----------------------------------------------------------------------------
+// mozilla::UniqueFreePtr<CHAR>
+
+// UniqueFreePtr<CHAR>, which points at a string allocated with `malloc`
+// (typically generated by `strdup()`), is serialized as the number of
+// *bytes* (encoded as ULEB128) and all the characters in the string. The
+// null terminator is omitted.
+// `CHAR` can be any type that has a specialization for
+// `std::char_traits<CHAR>::length(const CHAR*)`.
+//
+// Note: A nullptr pointer will be serialized like an empty string, so when
+// deserializing it will result in an allocated buffer only containing a
+// single null terminator.
+template <typename CHAR>
+struct ProfileBufferEntryWriter::Serializer<UniqueFreePtr<CHAR>> {
+  static Length Bytes(const UniqueFreePtr<CHAR>& aS) {
+    if (!aS) {
+      // Null pointer, store it as if it was an empty string (so: 0 bytes).
+      return ULEB128Size(0u);
+    }
+    // Note that we store the size in *bytes*, not in number of characters.
+    const auto bytes = std::char_traits<CHAR>::length(aS.get()) * sizeof(CHAR);
+    return ULEB128Size(bytes) + bytes;
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const UniqueFreePtr<CHAR>& aS) {
+    if (!aS) {
+      // Null pointer, store it as if it was an empty string (so we write a
+      // length of 0 bytes).
+      aEW.WriteULEB128(0u);
+      return;
+    }
+    // Note that we store the size in *bytes*, not in number of characters.
+    const auto bytes = std::char_traits<CHAR>::length(aS.get()) * sizeof(CHAR);
+    aEW.WriteULEB128(bytes);
+    aEW.WriteBytes(aS.get(), bytes);
+  }
+};
+
+template <typename CHAR>
+struct ProfileBufferEntryReader::Deserializer<UniqueFreePtr<CHAR>> {
+  static void ReadInto(ProfileBufferEntryReader& aER, UniqueFreePtr<CHAR>& aS) {
+    aS = Read(aER);
+  }
+
+  static UniqueFreePtr<CHAR> Read(ProfileBufferEntryReader& aER) {
+    // Read the number of *bytes* that follow.
+    const auto bytes = aER.ReadULEB128<size_t>();
+    // We need a buffer of the non-const character type.
+    using NC_CHAR = std::remove_const_t<CHAR>;
+    // We allocate the required number of bytes, plus one extra character for
+    // the null terminator.
+    NC_CHAR* buffer = static_cast<NC_CHAR*>(malloc(bytes + sizeof(NC_CHAR)));
+    // Copy the characters into the buffer.
+    aER.ReadBytes(buffer, bytes);
+    // And append a null terminator.
+    buffer[bytes / sizeof(NC_CHAR)] = NC_CHAR(0);
+    return UniqueFreePtr<CHAR>(buffer);
+  }
+};
+
+// ----------------------------------------------------------------------------
+// std::tuple
+
+// std::tuple is serialized as a sequence of each recursively-serialized item.
+//
+// This is equivalent to manually serializing each item, so reading/writing
+// tuples is equivalent to reading/writing their elements in order, e.g.:
+// ```
+// std::tuple<int, std::string> is = ...;
+// aEW.WriteObject(is); // Write the tuple, equivalent to:
+// aEW.WriteObject(/* int */ std::get<0>(is), /* string */ std::get<1>(is));
+// ...
+// // Reading back can be done directly into a tuple:
+// auto is = aER.ReadObject<std::tuple<int, std::string>>();
+// // Or each item could be read separately:
+// auto i = aER.ReadObject<int>(); auto s = aER.ReadObject<std::string>();
+// ```
+template <typename... Ts>
+struct ProfileBufferEntryWriter::Serializer<std::tuple<Ts...>> {
+ private:
+  template <size_t... Is>
+  static Length TupleBytes(const std::tuple<Ts...>& aTuple,
+                           std::index_sequence<Is...>) {
+    return (0 + ... + SumBytes(std::get<Is>(aTuple)));
+  }
+
+  template <size_t... Is>
+  static void TupleWrite(ProfileBufferEntryWriter& aEW,
+                         const std::tuple<Ts...>& aTuple,
+                         std::index_sequence<Is...>) {
+    (aEW.WriteObject(std::get<Is>(aTuple)), ...);
+  }
+
+ public:
+  static Length Bytes(const std::tuple<Ts...>& aTuple) {
+    // Generate a 0..N-1 index pack, we'll add the sizes of each item.
+    return TupleBytes(aTuple, std::index_sequence_for<Ts...>());
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const std::tuple<Ts...>& aTuple) {
+    // Generate a 0..N-1 index pack, we'll write each item.
+    TupleWrite(aEW, aTuple, std::index_sequence_for<Ts...>());
+  }
+};
+
+template <typename... Ts>
+struct ProfileBufferEntryReader::Deserializer<std::tuple<Ts...>> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       std::tuple<Ts...>& aTuple) {
+    aER.ReadBytes(&aTuple, Bytes(aTuple));
+  }
+
+  static std::tuple<Ts...> Read(ProfileBufferEntryReader& aER) {
+    // Note that this creates default `Ts` first, and then overwrites them.
+    std::tuple<Ts...> ob;
+    ReadInto(aER, ob);
+    return ob;
+  }
+};
+
+// ----------------------------------------------------------------------------
+// mozilla::Tuple
+
+// Tuple is serialized as a sequence of each recursively-serialized
+// item.
+//
+// This is equivalent to manually serializing each item, so reading/writing
+// tuples is equivalent to reading/writing their elements in order, e.g.:
+// ```
+// Tuple<int, std::string> is = ...;
+// aEW.WriteObject(is); // Write the Tuple, equivalent to:
+// aEW.WriteObject(/* int */ std::get<0>(is), /* string */ std::get<1>(is));
+// ...
+// // Reading back can be done directly into a Tuple:
+// auto is = aER.ReadObject<Tuple<int, std::string>>();
+// // Or each item could be read separately:
+// auto i = aER.ReadObject<int>(); auto s = aER.ReadObject<std::string>();
+// ```
+template <typename... Ts>
+struct ProfileBufferEntryWriter::Serializer<Tuple<Ts...>> {
+ private:
+  template <size_t... Is>
+  static Length TupleBytes(const Tuple<Ts...>& aTuple,
+                           std::index_sequence<Is...>) {
+    return (0 + ... + SumBytes(Get<Is>(aTuple)));
+  }
+
+  template <size_t... Is>
+  static void TupleWrite(ProfileBufferEntryWriter& aEW,
+                         const Tuple<Ts...>& aTuple,
+                         std::index_sequence<Is...>) {
+    (aEW.WriteObject(Get<Is>(aTuple)), ...);
+  }
+
+ public:
+  static Length Bytes(const Tuple<Ts...>& aTuple) {
+    // Generate a 0..N-1 index pack, we'll add the sizes of each item.
+    return TupleBytes(aTuple, std::index_sequence_for<Ts...>());
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW, const Tuple<Ts...>& aTuple) {
+    // Generate a 0..N-1 index pack, we'll write each item.
+    TupleWrite(aEW, aTuple, std::index_sequence_for<Ts...>());
+  }
+};
+
+template <typename... Ts>
+struct ProfileBufferEntryReader::Deserializer<Tuple<Ts...>> {
+  static void ReadInto(ProfileBufferEntryReader& aER, Tuple<Ts...>& aTuple) {
+    aER.ReadBytes(&aTuple, Bytes(aTuple));
+  }
+
+  static Tuple<Ts...> Read(ProfileBufferEntryReader& aER) {
+    // Note that this creates default `Ts` first, and then overwrites them.
+    Tuple<Ts...> ob;
+    ReadInto(aER, ob);
+    return ob;
+  }
+};
+
+// ----------------------------------------------------------------------------
+// mozilla::Span
+
+// Span. All elements are serialized in sequence.
+// The caller is assumed to know the number of elements (they may manually
+// write&read it before the span if needed).
+// Similar to tuples, reading/writing spans is equivalent to reading/writing
+// their elements in order.
+template <class T, size_t N>
+struct ProfileBufferEntryWriter::Serializer<Span<T, N>> {
+  static Length Bytes(const Span<T, N>& aSpan) {
+    Length bytes = 0;
+    for (const T& element : aSpan) {
+      bytes += SumBytes(element);
+    }
+    return bytes;
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW, const Span<T, N>& aSpan) {
+    for (const T& element : aSpan) {
+      aEW.WriteObject(element);
+    }
+  }
+};
+
+template <class T, size_t N>
+struct ProfileBufferEntryReader::Deserializer<Span<T, N>> {
+  // Read elements back into span pointing at a pre-allocated buffer.
+  static void ReadInto(ProfileBufferEntryReader& aER, Span<T, N>& aSpan) {
+    for (T& element : aSpan) {
+      aER.ReadIntoObject(element);
+    }
+  }
+
+  // A Span does not own its data, this would probably leak so we forbid this.
+  static Span<T, N> Read(ProfileBufferEntryReader& aER) = delete;
+};
+
+// ----------------------------------------------------------------------------
+// mozilla::Maybe
+
+// Maybe<T> is serialized as one byte containing either 'm' (Nothing),
+// or 'M' followed by the recursively-serialized `T` object.
+template <typename T>
+struct ProfileBufferEntryWriter::Serializer<Maybe<T>> {
+  static Length Bytes(const Maybe<T>& aMaybe) {
+    // 1 byte to store nothing/something flag, then object size if present.
+    return aMaybe.isNothing() ? 1 : (1 + SumBytes(aMaybe.ref()));
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW, const Maybe<T>& aMaybe) {
+    // 'm'/'M' is just an arbitrary 1-byte value to distinguish states.
+    if (aMaybe.isNothing()) {
+      aEW.WriteObject<char>('m');
+    } else {
+      aEW.WriteObject<char>('M');
+      // Use the Serializer for the contained type.
+      aEW.WriteObject(aMaybe.ref());
+    }
+  }
+};
+
+template <typename T>
+struct ProfileBufferEntryReader::Deserializer<Maybe<T>> {
+  static void ReadInto(ProfileBufferEntryReader& aER, Maybe<T>& aMaybe) {
+    char c = aER.ReadObject<char>();
+    if (c == 'm') {
+      aMaybe.reset();
+    } else {
+      MOZ_ASSERT(c == 'M');
+      // If aMaybe is empty, create a default `T` first, to be overwritten.
+      // Otherwise we'll just overwrite whatever was already there.
+      if (aMaybe.isNothing()) {
+        aMaybe.emplace();
+      }
+      // Use the Deserializer for the contained type.
+      aER.ReadIntoObject(aMaybe.ref());
+    }
+  }
+
+  static Maybe<T> Read(ProfileBufferEntryReader& aER) {
+    Maybe<T> maybe;
+    char c = aER.ReadObject<char>();
+    MOZ_ASSERT(c == 'M' || c == 'm');
+    if (c == 'M') {
+      // Note that this creates a default `T` inside the Maybe first, and then
+      // overwrites it.
+      maybe = Some(T{});
+      // Use the Deserializer for the contained type.
+      aER.ReadIntoObject(maybe.ref());
+    }
+    return maybe;
+  }
+};
+
+// ----------------------------------------------------------------------------
+// mozilla::Variant
+
+// Variant is serialized as the tag (0-based index of the stored type, encoded
+// as ULEB128), and the recursively-serialized object.
+template <typename... Ts>
+struct ProfileBufferEntryWriter::Serializer<Variant<Ts...>> {
+ public:
+  static Length Bytes(const Variant<Ts...>& aVariantTs) {
+    return aVariantTs.match([](auto aIndex, const auto& aAlternative) {
+      return ULEB128Size(aIndex) + SumBytes(aAlternative);
+    });
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const Variant<Ts...>& aVariantTs) {
+    aVariantTs.match([&aEW](auto aIndex, const auto& aAlternative) {
+      aEW.WriteULEB128(aIndex);
+      aEW.WriteObject(aAlternative);
+    });
+  }
+};
+
+template <typename... Ts>
+struct ProfileBufferEntryReader::Deserializer<Variant<Ts...>> {
+ private:
+  // Called from the fold expression in `VariantReadInto()`, only the selected
+  // variant will deserialize the object.
+  template <size_t I>
+  static void VariantIReadInto(ProfileBufferEntryReader& aER,
+                               Variant<Ts...>& aVariantTs, unsigned aTag) {
+    if (I == aTag) {
+      // Ensure the variant contains the target type. Note that this may create
+      // a default object.
+      if (!aVariantTs.template is<I>()) {
+        aVariantTs = Variant<Ts...>(VariantIndex<I>{});
+      }
+      aER.ReadIntoObject(aVariantTs.template as<I>());
+    }
+  }
+
+  template <size_t... Is>
+  static void VariantReadInto(ProfileBufferEntryReader& aER,
+                              Variant<Ts...>& aVariantTs,
+                              std::index_sequence<Is...>) {
+    unsigned tag = aER.ReadULEB128<unsigned>();
+    (VariantIReadInto<Is>(aER, aVariantTs, tag), ...);
+  }
+
+ public:
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       Variant<Ts...>& aVariantTs) {
+    // Generate a 0..N-1 index pack, the selected variant will deserialize
+    // itself.
+    VariantReadInto(aER, aVariantTs, std::index_sequence_for<Ts...>());
+  }
+
+  static Variant<Ts...> Read(ProfileBufferEntryReader& aER) {
+    // Note that this creates a default `Variant` of the first type, and then
+    // overwrites it. Consider using `ReadInto` for more control if needed.
+    Variant<Ts...> variant(VariantIndex<0>{});
+    ReadInto(aER, variant);
+    return variant;
+  }
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileBufferEntrySerialization_h
diff --git a/mozglue/baseprofiler/public/ProfileBufferIndex.h b/mozglue/baseprofiler/public/ProfileBufferIndex.h
new file mode 100644
index 0000000000..5cda6bd89e
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileBufferIndex.h
@@ -0,0 +1,97 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileBufferIndex_h
+#define ProfileBufferIndex_h
+
+#include "mozilla/Attributes.h"
+
+#include <cstddef>
+#include <cstdint>
+
+namespace mozilla {
+
+// Generic index into a Profiler buffer, mostly for internal usage.
+// Intended to appear infinite (it should effectively never wrap).
+// 0 (zero) is reserved as nullptr-like value; it may indicate failure result,
+// or it may point at the earliest available block.
+using ProfileBufferIndex = uint64_t;
+
+// Externally-opaque class encapsulating a block index, i.e. a
+// ProfileBufferIndex that is guaranteed to point at the start of a Profile
+// buffer block (until it is destroyed, but then that index cannot be reused and
+// functions should gracefully handle expired blocks).
+// Users may get these from Profile buffer functions, to later access previous
+// blocks; they should avoid converting and operating on their value.
+class ProfileBufferBlockIndex {
+ public:
+  // Default constructor with internal 0 value, for which Profile buffers must
+  // guarantee that it is before any valid entries; All public APIs should
+  // fail gracefully, doing and/or returning Nothing.
+  ProfileBufferBlockIndex() : mBlockIndex(0) {}
+
+  // Implicit conversion from literal `nullptr` to internal 0 value, to allow
+  // convenient init/reset/comparison with 0 index.
+  MOZ_IMPLICIT ProfileBufferBlockIndex(std::nullptr_t) : mBlockIndex(0) {}
+
+  // Explicit conversion to bool, works in `if` and other tests.
+  // Only returns false for default `ProfileBufferBlockIndex{}` value.
+  explicit operator bool() const { return mBlockIndex != 0; }
+
+  // Comparison operators. Default `ProfileBufferBlockIndex{}` value is always
+  // the lowest.
+  [[nodiscard]] bool operator==(const ProfileBufferBlockIndex& aRhs) const {
+    return mBlockIndex == aRhs.mBlockIndex;
+  }
+  [[nodiscard]] bool operator!=(const ProfileBufferBlockIndex& aRhs) const {
+    return mBlockIndex != aRhs.mBlockIndex;
+  }
+  [[nodiscard]] bool operator<(const ProfileBufferBlockIndex& aRhs) const {
+    return mBlockIndex < aRhs.mBlockIndex;
+  }
+  [[nodiscard]] bool operator<=(const ProfileBufferBlockIndex& aRhs) const {
+    return mBlockIndex <= aRhs.mBlockIndex;
+  }
+  [[nodiscard]] bool operator>(const ProfileBufferBlockIndex& aRhs) const {
+    return mBlockIndex > aRhs.mBlockIndex;
+  }
+  [[nodiscard]] bool operator>=(const ProfileBufferBlockIndex& aRhs) const {
+    return mBlockIndex >= aRhs.mBlockIndex;
+  }
+
+  // Explicit conversion to ProfileBufferIndex, mostly used by internal Profile
+  // buffer code.
+  [[nodiscard]] ProfileBufferIndex ConvertToProfileBufferIndex() const {
+    return mBlockIndex;
+  }
+
+  // Explicit creation from ProfileBufferIndex, mostly used by internal
+  // Profile buffer code.
+  [[nodiscard]] static ProfileBufferBlockIndex CreateFromProfileBufferIndex(
+      ProfileBufferIndex aIndex) {
+    return ProfileBufferBlockIndex(aIndex);
+  }
+
+ private:
+  // Private to prevent easy construction from any value. Use
+  // `CreateFromProfileBufferIndex()` instead.
+  // The main reason for this indirection is to make it harder to create these
+  // objects, because only the profiler code should need to do it. Ideally, this
+  // class should be used wherever a block index should be stored, but there is
+  // so much code that uses `uint64_t` that it would be a big task to change
+  // them all. So for now we allow conversions to/from numbers, but it's as ugly
+  // as possible to make sure it doesn't get too common; and if one day we want
+  // to tackle a global change, it should be easy to find all these locations
+  // thanks to the explicit conversion functions.
+  explicit ProfileBufferBlockIndex(ProfileBufferIndex aBlockIndex)
+      : mBlockIndex(aBlockIndex) {}
+
+  ProfileBufferIndex mBlockIndex;
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileBufferIndex_h
diff --git a/mozglue/baseprofiler/public/ProfileChunkedBuffer.h b/mozglue/baseprofiler/public/ProfileChunkedBuffer.h
new file mode 100644
index 0000000000..d4d55eafcb
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfileChunkedBuffer.h
@@ -0,0 +1,1872 @@
+/* -*- Mode: C++; tab-width: 2; 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 ProfileChunkedBuffer_h
+#define ProfileChunkedBuffer_h
+
+#include "mozilla/BaseProfilerDetail.h"
+#include "mozilla/NotNull.h"
+#include "mozilla/ProfileBufferChunkManager.h"
+#include "mozilla/ProfileBufferChunkManagerSingle.h"
+#include "mozilla/ProfileBufferEntrySerialization.h"
+#include "mozilla/RefCounted.h"
+#include "mozilla/RefPtr.h"
+#include "mozilla/ScopeExit.h"
+#include "mozilla/Unused.h"
+
+#include <cstdio>
+#include <utility>
+
+namespace mozilla {
+
+namespace detail {
+
+// Internal accessor pointing at a position inside a chunk.
+// It can handle two groups of chunks (typically the extant chunks stored in
+// the store manager, and the current chunk).
+// The main operations are:
+// - ReadEntrySize() to read an entry size, 0 means failure.
+// - operator+=(Length) to skip a number of bytes.
+// - EntryReader() creates an entry reader at the current position for a given
+//   size (it may fail with an empty reader), and skips the entry.
+// Note that there is no "past-the-end" position -- as soon as InChunkPointer
+// reaches the end, it becomes effectively null.
+class InChunkPointer {
+ public:
+  using Byte = ProfileBufferChunk::Byte;
+  using Length = ProfileBufferChunk::Length;
+
+  // Nullptr-like InChunkPointer, may be used as end iterator.
+  InChunkPointer()
+      : mChunk(nullptr), mNextChunkGroup(nullptr), mOffsetInChunk(0) {}
+
+  // InChunkPointer over one or two chunk groups, pointing at the given
+  // block index (if still in range).
+  // This constructor should only be used with *trusted* block index values!
+  InChunkPointer(const ProfileBufferChunk* aChunk,
+                 const ProfileBufferChunk* aNextChunkGroup,
+                 ProfileBufferBlockIndex aBlockIndex)
+      : mChunk(aChunk), mNextChunkGroup(aNextChunkGroup) {
+    if (mChunk) {
+      mOffsetInChunk = mChunk->OffsetFirstBlock();
+      Adjust();
+    } else if (mNextChunkGroup) {
+      mChunk = mNextChunkGroup;
+      mNextChunkGroup = nullptr;
+      mOffsetInChunk = mChunk->OffsetFirstBlock();
+      Adjust();
+    } else {
+      mOffsetInChunk = 0;
+    }
+
+    // Try to advance to given position.
+    if (!AdvanceToGlobalRangePosition(aBlockIndex)) {
+      // Block does not exist anymore (or block doesn't look valid), reset the
+      // in-chunk pointer.
+      mChunk = nullptr;
+      mNextChunkGroup = nullptr;
+    }
+  }
+
+  // InChunkPointer over one or two chunk groups, will start at the first
+  // block (if any). This may be slow, so avoid using it too much.
+  InChunkPointer(const ProfileBufferChunk* aChunk,
+                 const ProfileBufferChunk* aNextChunkGroup,
+                 ProfileBufferIndex aIndex = ProfileBufferIndex(0))
+      : mChunk(aChunk), mNextChunkGroup(aNextChunkGroup) {
+    if (mChunk) {
+      mOffsetInChunk = mChunk->OffsetFirstBlock();
+      Adjust();
+    } else if (mNextChunkGroup) {
+      mChunk = mNextChunkGroup;
+      mNextChunkGroup = nullptr;
+      mOffsetInChunk = mChunk->OffsetFirstBlock();
+      Adjust();
+    } else {
+      mOffsetInChunk = 0;
+    }
+
+    // Try to advance to given position.
+    if (!AdvanceToGlobalRangePosition(aIndex)) {
+      // Block does not exist anymore, reset the in-chunk pointer.
+      mChunk = nullptr;
+      mNextChunkGroup = nullptr;
+    }
+  }
+
+  // Compute the current position in the global range.
+  // 0 if null (including if we're reached the end).
+  [[nodiscard]] ProfileBufferIndex GlobalRangePosition() const {
+    if (IsNull()) {
+      return 0;
+    }
+    return mChunk->RangeStart() + mOffsetInChunk;
+  }
+
+  // Move InChunkPointer forward to the block at the given global block
+  // position, which is assumed to be valid exactly -- but it may be obsolete.
+  // 0 stays where it is (if valid already).
+  // MOZ_ASSERTs if the index is invalid.
+  [[nodiscard]] bool AdvanceToGlobalRangePosition(
+      ProfileBufferBlockIndex aBlockIndex) {
+    if (IsNull()) {
+      // Pointer is null already. (Not asserting because it's acceptable.)
+      return false;
+    }
+    if (!aBlockIndex) {
+      // Special null position, just stay where we are.
+      return ShouldPointAtValidBlock();
+    }
+    if (aBlockIndex.ConvertToProfileBufferIndex() < GlobalRangePosition()) {
+      // Past the requested position, stay where we are (assuming the current
+      // position was valid).
+      return ShouldPointAtValidBlock();
+    }
+    for (;;) {
+      if (aBlockIndex.ConvertToProfileBufferIndex() <
+          mChunk->RangeStart() + mChunk->OffsetPastLastBlock()) {
+        // Target position is in this chunk's written space, move to it.
+        mOffsetInChunk =
+            aBlockIndex.ConvertToProfileBufferIndex() - mChunk->RangeStart();
+        return ShouldPointAtValidBlock();
+      }
+      // Position is after this chunk, try next chunk.
+      GoToNextChunk();
+      if (IsNull()) {
+        return false;
+      }
+      // Skip whatever block tail there is, we don't allow pointing in the
+      // middle of a block.
+      mOffsetInChunk = mChunk->OffsetFirstBlock();
+      if (aBlockIndex.ConvertToProfileBufferIndex() < GlobalRangePosition()) {
+        // Past the requested position, meaning that the given position was in-
+        // between blocks -> Failure.
+        MOZ_ASSERT(false, "AdvanceToGlobalRangePosition - In-between blocks");
+        return false;
+      }
+    }
+  }
+
+  // Move InChunkPointer forward to the block at or after the given global
+  // range position.
+  // 0 stays where it is (if valid already).
+  [[nodiscard]] bool AdvanceToGlobalRangePosition(
+      ProfileBufferIndex aPosition) {
+    if (aPosition == 0) {
+      // Special position '0', just stay where we are.
+      // Success if this position is already valid.
+      return !IsNull();
+    }
+    for (;;) {
+      ProfileBufferIndex currentPosition = GlobalRangePosition();
+      if (currentPosition == 0) {
+        // Pointer is null.
+        return false;
+      }
+      if (aPosition <= currentPosition) {
+        // At or past the requested position, stay where we are.
+        return true;
+      }
+      if (aPosition < mChunk->RangeStart() + mChunk->OffsetPastLastBlock()) {
+        // Target position is in this chunk's written space, move to it.
+        for (;;) {
+          // Skip the current block.
+          mOffsetInChunk += ReadEntrySize();
+          if (mOffsetInChunk >= mChunk->OffsetPastLastBlock()) {
+            // Reached the end of the chunk, this can happen for the last
+            // block, let's just continue to the next chunk.
+            break;
+          }
+          if (aPosition <= mChunk->RangeStart() + mOffsetInChunk) {
+            // We're at or after the position, return at this block position.
+            return true;
+          }
+        }
+      }
+      // Position is after this chunk, try next chunk.
+      GoToNextChunk();
+      if (IsNull()) {
+        return false;
+      }
+      // Skip whatever block tail there is, we don't allow pointing in the
+      // middle of a block.
+      mOffsetInChunk = mChunk->OffsetFirstBlock();
+    }
+  }
+
+  [[nodiscard]] Byte ReadByte() {
+    MOZ_ASSERT(!IsNull());
+    MOZ_ASSERT(mOffsetInChunk < mChunk->OffsetPastLastBlock());
+    Byte byte = mChunk->ByteAt(mOffsetInChunk);
+    if (MOZ_UNLIKELY(++mOffsetInChunk == mChunk->OffsetPastLastBlock())) {
+      Adjust();
+    }
+    return byte;
+  }
+
+  // Read and skip a ULEB128-encoded size.
+  // 0 means failure (0-byte entries are not allowed.)
+  // Note that this doesn't guarantee that there are actually that many bytes
+  // available to read! (EntryReader() below may gracefully fail.)
+  [[nodiscard]] Length ReadEntrySize() {
+    ULEB128Reader<Length> reader;
+    if (IsNull()) {
+      return 0;
+    }
+    for (;;) {
+      const bool isComplete = reader.FeedByteIsComplete(ReadByte());
+      if (MOZ_UNLIKELY(IsNull())) {
+        // End of chunks, so there's no actual entry after this anyway.
+        return 0;
+      }
+      if (MOZ_LIKELY(isComplete)) {
+        if (MOZ_UNLIKELY(reader.Value() > mChunk->BufferBytes())) {
+          // Don't allow entries larger than a chunk.
+          return 0;
+        }
+        return reader.Value();
+      }
+    }
+  }
+
+  InChunkPointer& operator+=(Length aLength) {
+    MOZ_ASSERT(!IsNull());
+    mOffsetInChunk += aLength;
+    Adjust();
+    return *this;
+  }
+
+  [[nodiscard]] ProfileBufferEntryReader EntryReader(Length aLength) {
+    if (IsNull() || aLength == 0) {
+      return ProfileBufferEntryReader();
+    }
+
+    MOZ_ASSERT(mOffsetInChunk < mChunk->OffsetPastLastBlock());
+
+    // We should be pointing at the entry, past the entry size.
+    const ProfileBufferIndex entryIndex = GlobalRangePosition();
+    // Verify that there's enough space before for the size (starting at index
+    // 1 at least).
+    MOZ_ASSERT(entryIndex >= 1u + ULEB128Size(aLength));
+
+    const Length remaining = mChunk->OffsetPastLastBlock() - mOffsetInChunk;
+    Span<const Byte> mem0 = mChunk->BufferSpan();
+    mem0 = mem0.From(mOffsetInChunk);
+    if (aLength <= remaining) {
+      // Move to the end of this block, which could make this null if we have
+      // reached the end of all buffers.
+      *this += aLength;
+      return ProfileBufferEntryReader(
+          mem0.To(aLength),
+          // Block starts before the entry size.
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+              entryIndex - ULEB128Size(aLength)),
+          // Block ends right after the entry (could be null for last entry).
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+              GlobalRangePosition()));
+    }
+
+    // We need to go to the next chunk for the 2nd part of this block.
+    GoToNextChunk();
+    if (IsNull()) {
+      return ProfileBufferEntryReader();
+    }
+
+    Span<const Byte> mem1 = mChunk->BufferSpan();
+    const Length tail = aLength - remaining;
+    MOZ_ASSERT(tail <= mChunk->BufferBytes());
+    MOZ_ASSERT(tail == mChunk->OffsetFirstBlock());
+    // We are in the correct chunk, move the offset to the end of the block.
+    mOffsetInChunk = tail;
+    // And adjust as needed, which could make this null if we have reached the
+    // end of all buffers.
+    Adjust();
+    return ProfileBufferEntryReader(
+        mem0, mem1.To(tail),
+        // Block starts before the entry size.
+        ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+            entryIndex - ULEB128Size(aLength)),
+        // Block ends right after the entry (could be null for last entry).
+        ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+            GlobalRangePosition()));
+  }
+
+  [[nodiscard]] bool IsNull() const { return !mChunk; }
+
+  [[nodiscard]] bool operator==(const InChunkPointer& aOther) const {
+    if (IsNull() || aOther.IsNull()) {
+      return IsNull() && aOther.IsNull();
+    }
+    return mChunk == aOther.mChunk && mOffsetInChunk == aOther.mOffsetInChunk;
+  }
+
+  [[nodiscard]] bool operator!=(const InChunkPointer& aOther) const {
+    return !(*this == aOther);
+  }
+
+  [[nodiscard]] Byte operator*() const {
+    MOZ_ASSERT(!IsNull());
+    MOZ_ASSERT(mOffsetInChunk < mChunk->OffsetPastLastBlock());
+    return mChunk->ByteAt(mOffsetInChunk);
+  }
+
+  InChunkPointer& operator++() {
+    MOZ_ASSERT(!IsNull());
+    MOZ_ASSERT(mOffsetInChunk < mChunk->OffsetPastLastBlock());
+    if (MOZ_UNLIKELY(++mOffsetInChunk == mChunk->OffsetPastLastBlock())) {
+      mOffsetInChunk = 0;
+      GoToNextChunk();
+      Adjust();
+    }
+    return *this;
+  }
+
+ private:
+  void GoToNextChunk() {
+    MOZ_ASSERT(!IsNull());
+    const ProfileBufferIndex expectedNextRangeStart =
+        mChunk->RangeStart() + mChunk->BufferBytes();
+
+    mChunk = mChunk->GetNext();
+    if (!mChunk) {
+      // Reached the end of the current chunk group, try the next one (which
+      // may be null too, especially on the 2nd try).
+      mChunk = mNextChunkGroup;
+      mNextChunkGroup = nullptr;
+    }
+
+    if (mChunk && mChunk->RangeStart() == 0) {
+      // Reached a chunk without a valid (non-null) range start, assume there
+      // are only unused chunks from here on.
+      mChunk = nullptr;
+    }
+
+    MOZ_ASSERT(!mChunk || mChunk->RangeStart() == expectedNextRangeStart,
+               "We don't handle discontinuous buffers (yet)");
+    // Non-DEBUG fallback: Stop reading past discontinuities.
+    // (They should be rare, only happening on temporary OOMs.)
+    // TODO: Handle discontinuities (by skipping over incomplete blocks).
+    if (mChunk && mChunk->RangeStart() != expectedNextRangeStart) {
+      mChunk = nullptr;
+    }
+  }
+
+  // We want `InChunkPointer` to always point at a valid byte (or be null).
+  // After some operations, `mOffsetInChunk` may point past the end of the
+  // current `mChunk`, in which case we need to adjust our position to be inside
+  // the appropriate chunk. E.g., if we're 10 bytes after the end of the current
+  // chunk, we should end up at offset 10 in the next chunk.
+  // Note that we may "fall off" the last chunk and make this `InChunkPointer`
+  // effectively null.
+  void Adjust() {
+    while (mChunk && mOffsetInChunk >= mChunk->OffsetPastLastBlock()) {
+      // TODO: Try to adjust offset between chunks relative to mRangeStart
+      // differences. But we don't handle discontinuities yet.
+      if (mOffsetInChunk < mChunk->BufferBytes()) {
+        mOffsetInChunk -= mChunk->BufferBytes();
+      } else {
+        mOffsetInChunk -= mChunk->OffsetPastLastBlock();
+      }
+      GoToNextChunk();
+    }
+  }
+
+  // Check if the current position is likely to point at a valid block.
+  // (Size should be reasonable, and block should fully fit inside buffer.)
+  // MOZ_ASSERTs on failure, to catch incorrect uses of block indices (which
+  // should only point at valid blocks if still in range). Non-asserting build
+  // fallback should still be handled.
+  [[nodiscard]] bool ShouldPointAtValidBlock() const {
+    if (IsNull()) {
+      // Pointer is null, no blocks here.
+      MOZ_ASSERT(false, "ShouldPointAtValidBlock - null pointer");
+      return false;
+    }
+    // Use a copy, so we don't modify `*this`.
+    InChunkPointer pointer = *this;
+    // Try to read the entry size.
+    Length entrySize = pointer.ReadEntrySize();
+    if (entrySize == 0) {
+      // Entry size of zero means we read 0 or a way-too-big value.
+      MOZ_ASSERT(false, "ShouldPointAtValidBlock - invalid size");
+      return false;
+    }
+    // See if the last byte of the entry is still inside the buffer.
+    pointer += entrySize - 1;
+    MOZ_ASSERT(!IsNull(), "ShouldPointAtValidBlock - past end of buffer");
+    return !IsNull();
+  }
+
+  const ProfileBufferChunk* mChunk;
+  const ProfileBufferChunk* mNextChunkGroup;
+  Length mOffsetInChunk;
+};
+
+}  // namespace detail
+
+// Thread-safe buffer that can store blocks of different sizes during defined
+// sessions, using Chunks (from a ChunkManager) as storage.
+//
+// Each *block* contains an *entry* and the entry size:
+// [ entry_size | entry ] [ entry_size | entry ] ...
+//
+// *In-session* is a period of time during which `ProfileChunkedBuffer` allows
+// reading and writing.
+// *Out-of-session*, the `ProfileChunkedBuffer` object is still valid, but
+// contains no data, and gracefully denies accesses.
+//
+// To write an entry, the buffer reserves a block of sufficient size (to contain
+// user data of predetermined size), writes the entry size, and lets the caller
+// fill the entry contents using a ProfileBufferEntryWriter. E.g.:
+// ```
+// ProfileChunkedBuffer cb(...);
+// cb.ReserveAndPut([]() { return sizeof(123); },
+//                  [&](Maybe<ProfileBufferEntryWriter>& aEW) {
+//                    if (aEW) { aEW->WriteObject(123); }
+//                  });
+// ```
+// Other `Put...` functions may be used as shortcuts for simple entries.
+// The objects given to the caller's callbacks should only be used inside the
+// callbacks and not stored elsewhere, because they keep their own references to
+// chunk memory and therefore should not live longer.
+// Different type of objects may be serialized into an entry, see
+// `ProfileBufferEntryWriter::Serializer` for more information.
+//
+// When reading data, the buffer iterates over blocks (it knows how to read the
+// entry size, and therefore move to the next block), and lets the caller read
+// the entry inside of each block. E.g.:
+// ```
+// cb.ReadEach([](ProfileBufferEntryReader& aER) {
+//   /* Use ProfileBufferEntryReader functions to read serialized objects. */
+//   int n = aER.ReadObject<int>();
+// });
+// ```
+// Different type of objects may be deserialized from an entry, see
+// `ProfileBufferEntryReader::Deserializer` for more information.
+//
+// Writers may retrieve the block index corresponding to an entry
+// (`ProfileBufferBlockIndex` is an opaque type preventing the user from easily
+// modifying it). That index may later be used with `ReadAt` to get back to the
+// entry in that particular block -- if it still exists.
+class ProfileChunkedBuffer {
+ public:
+  using Byte = ProfileBufferChunk::Byte;
+  using Length = ProfileBufferChunk::Length;
+
+  enum class ThreadSafety { WithoutMutex, WithMutex };
+
+  // Default constructor starts out-of-session (nothing to read or write).
+  explicit ProfileChunkedBuffer(ThreadSafety aThreadSafety)
+      : mMutex(aThreadSafety != ThreadSafety::WithoutMutex) {}
+
+  // Start in-session with external chunk manager.
+  ProfileChunkedBuffer(ThreadSafety aThreadSafety,
+                       ProfileBufferChunkManager& aChunkManager)
+      : mMutex(aThreadSafety != ThreadSafety::WithoutMutex) {
+    SetChunkManager(aChunkManager);
+  }
+
+  // Start in-session with owned chunk manager.
+  ProfileChunkedBuffer(ThreadSafety aThreadSafety,
+                       UniquePtr<ProfileBufferChunkManager>&& aChunkManager)
+      : mMutex(aThreadSafety != ThreadSafety::WithoutMutex) {
+    SetChunkManager(std::move(aChunkManager));
+  }
+
+  ~ProfileChunkedBuffer() {
+    // Do proper clean-up by resetting the chunk manager.
+    ResetChunkManager();
+  }
+
+  // This cannot change during the lifetime of this buffer, so there's no need
+  // to lock.
+  [[nodiscard]] bool IsThreadSafe() const { return mMutex.IsActivated(); }
+
+  [[nodiscard]] bool IsInSession() const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return !!mChunkManager;
+  }
+
+  // Stop using the current chunk manager.
+  // If we own the current chunk manager, it will be destroyed.
+  // This will always clear currently-held chunks, if any.
+  void ResetChunkManager() {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    Unused << ResetChunkManager(lock);
+  }
+
+  // Set the current chunk manager.
+  // The caller is responsible for keeping the chunk manager alive as along as
+  // it's used here (until the next (Re)SetChunkManager, or
+  // ~ProfileChunkedBuffer).
+  void SetChunkManager(ProfileBufferChunkManager& aChunkManager) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    Unused << ResetChunkManager(lock);
+    SetChunkManager(aChunkManager, lock);
+  }
+
+  // Set the current chunk manager, and keep ownership of it.
+  void SetChunkManager(UniquePtr<ProfileBufferChunkManager>&& aChunkManager) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    Unused << ResetChunkManager(lock);
+    mOwnedChunkManager = std::move(aChunkManager);
+    if (mOwnedChunkManager) {
+      SetChunkManager(*mOwnedChunkManager, lock);
+    }
+  }
+
+  // Stop using the current chunk manager, and return it if owned here.
+  [[nodiscard]] UniquePtr<ProfileBufferChunkManager> ExtractChunkManager() {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return ResetChunkManager(lock);
+  }
+
+  // Clear the contents of this buffer, ready to receive new chunks.
+  // Note that memory is not freed: No chunks are destroyed, they are all
+  // receycled.
+  // Also the range doesn't reset, instead it continues at some point after the
+  // previous range. This may be useful if the caller may be keeping indexes
+  // into old chunks that have now been cleared, using these indexes will fail
+  // gracefully (instead of potentially pointing into new data).
+  void Clear() {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    if (MOZ_UNLIKELY(!mChunkManager)) {
+      // Out-of-session.
+      return;
+    }
+
+    mRangeStart = mRangeEnd = mNextChunkRangeStart;
+    mPushedBlockCount = 0;
+    mClearedBlockCount = 0;
+    mFailedPutBytes = 0;
+
+    // Recycle all released chunks as "next" chunks. This will reduce the number
+    // of future allocations. Also, when using ProfileBufferChunkManagerSingle,
+    // this retrieves the one chunk if it was released.
+    UniquePtr<ProfileBufferChunk> releasedChunks =
+        mChunkManager->GetExtantReleasedChunks();
+    if (releasedChunks) {
+      // Released chunks should be in the "Done" state, they need to be marked
+      // "recycled" before they can be reused.
+      for (ProfileBufferChunk* chunk = releasedChunks.get(); chunk;
+           chunk = chunk->GetNext()) {
+        chunk->MarkRecycled();
+      }
+      mNextChunks = ProfileBufferChunk::Join(std::move(mNextChunks),
+                                             std::move(releasedChunks));
+    }
+
+    if (mCurrentChunk) {
+      // We already have a current chunk (empty or in-use), mark it "done" and
+      // then "recycled", ready to be reused.
+      mCurrentChunk->MarkDone();
+      mCurrentChunk->MarkRecycled();
+    } else {
+      if (!mNextChunks) {
+        // No current chunk, and no next chunks to recycle, nothing more to do.
+        // The next "Put" operation will try to allocate a chunk as needed.
+        return;
+      }
+
+      // No current chunk, take a next chunk.
+      mCurrentChunk = std::exchange(mNextChunks, mNextChunks->ReleaseNext());
+    }
+
+    // Here, there was already a current chunk, or one has just been taken.
+    // Make sure it's ready to receive new entries.
+    InitializeCurrentChunk(lock);
+  }
+
+  // Buffer maximum length in bytes.
+  Maybe<size_t> BufferLength() const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    if (!mChunkManager) {
+      return Nothing{};
+    }
+    return Some(mChunkManager->MaxTotalSize());
+  }
+
+  [[nodiscard]] size_t SizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return SizeOfExcludingThis(aMallocSizeOf, lock);
+  }
+
+  [[nodiscard]] size_t SizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return aMallocSizeOf(this) + SizeOfExcludingThis(aMallocSizeOf, lock);
+  }
+
+  // Snapshot of the buffer state.
+  struct State {
+    // Index to/before the first block.
+    ProfileBufferIndex mRangeStart = 1;
+
+    // Index past the last block. Equals mRangeStart if empty.
+    ProfileBufferIndex mRangeEnd = 1;
+
+    // Number of blocks that have been pushed into this buffer.
+    uint64_t mPushedBlockCount = 0;
+
+    // Number of blocks that have been removed from this buffer.
+    // Note: Live entries = pushed - cleared.
+    uint64_t mClearedBlockCount = 0;
+
+    // Number of bytes that could not be put into this buffer.
+    uint64_t mFailedPutBytes = 0;
+  };
+
+  // Get a snapshot of the current state.
+  // When out-of-session, mFirstReadIndex==mNextWriteIndex, and
+  // mPushedBlockCount==mClearedBlockCount==0.
+  // Note that these may change right after this thread-safe call, so they
+  // should only be used for statistical purposes.
+  [[nodiscard]] State GetState() const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return {mRangeStart, mRangeEnd, mPushedBlockCount, mClearedBlockCount,
+            mFailedPutBytes};
+  }
+
+  [[nodiscard]] bool IsEmpty() const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return mRangeStart == mRangeEnd;
+  }
+
+  // True if this buffer is already locked on this thread.
+  // This should be used if some functions may call an already-locked buffer,
+  // e.g.: Put -> memory hook -> profiler_add_native_allocation_marker -> Put.
+  [[nodiscard]] bool IsThreadSafeAndLockedOnCurrentThread() const {
+    return mMutex.IsActivatedAndLockedOnCurrentThread();
+  }
+
+  // Lock the buffer mutex and run the provided callback.
+  // This can be useful when the caller needs to explicitly lock down this
+  // buffer, but not do anything else with it.
+  template <typename Callback>
+  auto LockAndRun(Callback&& aCallback) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return std::forward<Callback>(aCallback)();
+  }
+
+  // Reserve a block that can hold an entry of the given `aCallbackEntryBytes()`
+  // size, write the entry size (ULEB128-encoded), and invoke and return
+  // `aCallback(Maybe<ProfileBufferEntryWriter>&)`.
+  // Note: `aCallbackEntryBytes` is a callback instead of a simple value, to
+  // delay this potentially-expensive computation until after we're checked that
+  // we're in-session; use `Put(Length, Callback)` below if you know the size
+  // already.
+  template <typename CallbackEntryBytes, typename Callback>
+  auto ReserveAndPut(CallbackEntryBytes&& aCallbackEntryBytes,
+                     Callback&& aCallback)
+      -> decltype(std::forward<Callback>(aCallback)(
+          std::declval<Maybe<ProfileBufferEntryWriter>&>())) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+
+    // This can only be read in the 2nd lambda below after it has been written
+    // by the first lambda.
+    Length entryBytes;
+
+    return ReserveAndPutRaw(
+        [&]() {
+          entryBytes = std::forward<CallbackEntryBytes>(aCallbackEntryBytes)();
+          MOZ_ASSERT(entryBytes != 0, "Empty entries are not allowed");
+          return ULEB128Size(entryBytes) + entryBytes;
+        },
+        [&](Maybe<ProfileBufferEntryWriter>& aMaybeEntryWriter) {
+          if (aMaybeEntryWriter.isSome()) {
+            aMaybeEntryWriter->WriteULEB128(entryBytes);
+            MOZ_ASSERT(aMaybeEntryWriter->RemainingBytes() == entryBytes);
+          }
+          return std::forward<Callback>(aCallback)(aMaybeEntryWriter);
+        },
+        lock);
+  }
+
+  template <typename Callback>
+  auto Put(Length aEntryBytes, Callback&& aCallback) {
+    return ReserveAndPut([aEntryBytes]() { return aEntryBytes; },
+                         std::forward<Callback>(aCallback));
+  }
+
+  // Add a new entry copied from the given buffer, return block index.
+  ProfileBufferBlockIndex PutFrom(const void* aSrc, Length aBytes) {
+    return ReserveAndPut(
+        [aBytes]() { return aBytes; },
+        [aSrc, aBytes](Maybe<ProfileBufferEntryWriter>& aMaybeEntryWriter) {
+          if (aMaybeEntryWriter.isNothing()) {
+            return ProfileBufferBlockIndex{};
+          }
+          aMaybeEntryWriter->WriteBytes(aSrc, aBytes);
+          return aMaybeEntryWriter->CurrentBlockIndex();
+        });
+  }
+
+  // Add a new single entry with *all* given object (using a Serializer for
+  // each), return block index.
+  template <typename... Ts>
+  ProfileBufferBlockIndex PutObjects(const Ts&... aTs) {
+    static_assert(sizeof...(Ts) > 0,
+                  "PutObjects must be given at least one object.");
+    return ReserveAndPut(
+        [&]() { return ProfileBufferEntryWriter::SumBytes(aTs...); },
+        [&](Maybe<ProfileBufferEntryWriter>& aMaybeEntryWriter) {
+          if (aMaybeEntryWriter.isNothing()) {
+            return ProfileBufferBlockIndex{};
+          }
+          aMaybeEntryWriter->WriteObjects(aTs...);
+          return aMaybeEntryWriter->CurrentBlockIndex();
+        });
+  }
+
+  // Add a new entry copied from the given object, return block index.
+  template <typename T>
+  ProfileBufferBlockIndex PutObject(const T& aOb) {
+    return PutObjects(aOb);
+  }
+
+  // Get *all* chunks related to this buffer, including extant chunks in its
+  // ChunkManager, and yet-unused new/recycled chunks.
+  // We don't expect this buffer to be used again, though it's still possible
+  // and will allocate the first buffer when needed.
+  [[nodiscard]] UniquePtr<ProfileBufferChunk> GetAllChunks() {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    if (MOZ_UNLIKELY(!mChunkManager)) {
+      // Out-of-session.
+      return nullptr;
+    }
+    UniquePtr<ProfileBufferChunk> chunks =
+        mChunkManager->GetExtantReleasedChunks();
+    Unused << HandleRequestedChunk_IsPending(lock);
+    if (MOZ_LIKELY(!!mCurrentChunk)) {
+      mCurrentChunk->MarkDone();
+      chunks =
+          ProfileBufferChunk::Join(std::move(chunks), std::move(mCurrentChunk));
+    }
+    chunks =
+        ProfileBufferChunk::Join(std::move(chunks), std::move(mNextChunks));
+    mChunkManager->ForgetUnreleasedChunks();
+    mRangeStart = mRangeEnd = mNextChunkRangeStart;
+    return chunks;
+  }
+
+  class Reader;
+
+  // Class that can iterate through blocks and provide
+  // `ProfileBufferEntryReader`s.
+  // Created through `Reader`, lives within a lock guard lifetime.
+  class BlockIterator {
+   public:
+#ifdef DEBUG
+    ~BlockIterator() {
+      // No BlockIterator should live outside of a mutexed call.
+      mBuffer->mMutex.AssertCurrentThreadOwns();
+    }
+#endif  // DEBUG
+
+    // Comparison with other iterator, mostly used in range-for loops.
+    [[nodiscard]] bool operator==(const BlockIterator& aRhs) const {
+      MOZ_ASSERT(mBuffer == aRhs.mBuffer);
+      return mCurrentBlockIndex == aRhs.mCurrentBlockIndex;
+    }
+    [[nodiscard]] bool operator!=(const BlockIterator& aRhs) const {
+      MOZ_ASSERT(mBuffer == aRhs.mBuffer);
+      return mCurrentBlockIndex != aRhs.mCurrentBlockIndex;
+    }
+
+    // Advance to next BlockIterator.
+    BlockIterator& operator++() {
+      mBuffer->mMutex.AssertCurrentThreadOwns();
+      mCurrentBlockIndex =
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+              mNextBlockPointer.GlobalRangePosition());
+      mCurrentEntry =
+          mNextBlockPointer.EntryReader(mNextBlockPointer.ReadEntrySize());
+      return *this;
+    }
+
+    // Dereferencing creates a `ProfileBufferEntryReader` object for the entry
+    // inside this block.
+    // (Note: It would be possible to return a `const
+    // ProfileBufferEntryReader&`, but not useful in practice, because in most
+    // case the user will want to read, which is non-const.)
+    [[nodiscard]] ProfileBufferEntryReader operator*() const {
+      return mCurrentEntry;
+    }
+
+    // True if this iterator is just past the last entry.
+    [[nodiscard]] bool IsAtEnd() const {
+      return mCurrentEntry.RemainingBytes() == 0;
+    }
+
+    // Can be used as reference to come back to this entry with `GetEntryAt()`.
+    [[nodiscard]] ProfileBufferBlockIndex CurrentBlockIndex() const {
+      return mCurrentBlockIndex;
+    }
+
+    // Index past the end of this block, which is the start of the next block.
+    [[nodiscard]] ProfileBufferBlockIndex NextBlockIndex() const {
+      MOZ_ASSERT(!IsAtEnd());
+      return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          mNextBlockPointer.GlobalRangePosition());
+    }
+
+    // Index of the first block in the whole buffer.
+    [[nodiscard]] ProfileBufferBlockIndex BufferRangeStart() const {
+      mBuffer->mMutex.AssertCurrentThreadOwns();
+      return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          mBuffer->mRangeStart);
+    }
+
+    // Index past the last block in the whole buffer.
+    [[nodiscard]] ProfileBufferBlockIndex BufferRangeEnd() const {
+      mBuffer->mMutex.AssertCurrentThreadOwns();
+      return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          mBuffer->mRangeEnd);
+    }
+
+   private:
+    // Only a Reader can instantiate a BlockIterator.
+    friend class Reader;
+
+    BlockIterator(const ProfileChunkedBuffer& aBuffer,
+                  const ProfileBufferChunk* aChunks0,
+                  const ProfileBufferChunk* aChunks1,
+                  ProfileBufferBlockIndex aBlockIndex)
+        : mNextBlockPointer(aChunks0, aChunks1, aBlockIndex),
+          mCurrentBlockIndex(
+              ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+                  mNextBlockPointer.GlobalRangePosition())),
+          mCurrentEntry(
+              mNextBlockPointer.EntryReader(mNextBlockPointer.ReadEntrySize())),
+          mBuffer(WrapNotNull(&aBuffer)) {
+      // No BlockIterator should live outside of a mutexed call.
+      mBuffer->mMutex.AssertCurrentThreadOwns();
+    }
+
+    detail::InChunkPointer mNextBlockPointer;
+
+    ProfileBufferBlockIndex mCurrentBlockIndex;
+
+    ProfileBufferEntryReader mCurrentEntry;
+
+    // Using a non-null pointer instead of a reference, to allow copying.
+    // This BlockIterator should only live inside one of the thread-safe
+    // ProfileChunkedBuffer functions, for this reference to stay valid.
+    NotNull<const ProfileChunkedBuffer*> mBuffer;
+  };
+
+  // Class that can create `BlockIterator`s (e.g., for range-for), or just
+  // iterate through entries; lives within a lock guard lifetime.
+  class MOZ_RAII Reader {
+   public:
+    Reader(const Reader&) = delete;
+    Reader& operator=(const Reader&) = delete;
+    Reader(Reader&&) = delete;
+    Reader& operator=(Reader&&) = delete;
+
+#ifdef DEBUG
+    ~Reader() {
+      // No Reader should live outside of a mutexed call.
+      mBuffer.mMutex.AssertCurrentThreadOwns();
+    }
+#endif  // DEBUG
+
+    // Index of the first block in the whole buffer.
+    [[nodiscard]] ProfileBufferBlockIndex BufferRangeStart() const {
+      mBuffer.mMutex.AssertCurrentThreadOwns();
+      return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          mBuffer.mRangeStart);
+    }
+
+    // Index past the last block in the whole buffer.
+    [[nodiscard]] ProfileBufferBlockIndex BufferRangeEnd() const {
+      mBuffer.mMutex.AssertCurrentThreadOwns();
+      return ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+          mBuffer.mRangeEnd);
+    }
+
+    // Iterators to the first and past-the-last blocks.
+    // Compatible with range-for (see `ForEach` below as example).
+    [[nodiscard]] BlockIterator begin() const {
+      return BlockIterator(mBuffer, mChunks0, mChunks1, nullptr);
+    }
+    // Note that a `BlockIterator` at the `end()` should not be dereferenced, as
+    // there is no actual block there!
+    [[nodiscard]] BlockIterator end() const {
+      return BlockIterator(mBuffer, nullptr, nullptr, nullptr);
+    }
+
+    // Get a `BlockIterator` at the given `ProfileBufferBlockIndex`, clamped to
+    // the stored range. Note that a `BlockIterator` at the `end()` should not
+    // be dereferenced, as there is no actual block there!
+    [[nodiscard]] BlockIterator At(ProfileBufferBlockIndex aBlockIndex) const {
+      if (aBlockIndex < BufferRangeStart()) {
+        // Anything before the range (including null ProfileBufferBlockIndex) is
+        // clamped at the beginning.
+        return begin();
+      }
+      // Otherwise we at least expect the index to be valid (pointing exactly at
+      // a live block, or just past the end.)
+      return BlockIterator(mBuffer, mChunks0, mChunks1, aBlockIndex);
+    }
+
+    // Run `aCallback(ProfileBufferEntryReader&)` on each entry from first to
+    // last. Callback should not store `ProfileBufferEntryReader`, as it may
+    // become invalid after this thread-safe call.
+    template <typename Callback>
+    void ForEach(Callback&& aCallback) const {
+      for (ProfileBufferEntryReader reader : *this) {
+        aCallback(reader);
+      }
+    }
+
+    // If this reader only points at one chunk with some data, this data will be
+    // exposed as a single entry.
+    [[nodiscard]] ProfileBufferEntryReader SingleChunkDataAsEntry() {
+      const ProfileBufferChunk* onlyNonEmptyChunk = nullptr;
+      for (const ProfileBufferChunk* chunkList : {mChunks0, mChunks1}) {
+        for (const ProfileBufferChunk* chunk = chunkList; chunk;
+             chunk = chunk->GetNext()) {
+          if (chunk->OffsetFirstBlock() != chunk->OffsetPastLastBlock()) {
+            if (onlyNonEmptyChunk) {
+              // More than one non-empty chunk.
+              return ProfileBufferEntryReader();
+            }
+            onlyNonEmptyChunk = chunk;
+          }
+        }
+      }
+      if (!onlyNonEmptyChunk) {
+        // No non-empty chunks.
+        return ProfileBufferEntryReader();
+      }
+      // Here, we have found one chunk that had some data.
+      return ProfileBufferEntryReader(
+          onlyNonEmptyChunk->BufferSpan().FromTo(
+              onlyNonEmptyChunk->OffsetFirstBlock(),
+              onlyNonEmptyChunk->OffsetPastLastBlock()),
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+              onlyNonEmptyChunk->RangeStart()),
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+              onlyNonEmptyChunk->RangeStart() +
+              (onlyNonEmptyChunk->OffsetPastLastBlock() -
+               onlyNonEmptyChunk->OffsetFirstBlock())));
+    }
+
+   private:
+    friend class ProfileChunkedBuffer;
+
+    explicit Reader(const ProfileChunkedBuffer& aBuffer,
+                    const ProfileBufferChunk* aChunks0,
+                    const ProfileBufferChunk* aChunks1)
+        : mBuffer(aBuffer), mChunks0(aChunks0), mChunks1(aChunks1) {
+      // No Reader should live outside of a mutexed call.
+      mBuffer.mMutex.AssertCurrentThreadOwns();
+    }
+
+    // This Reader should only live inside one of the thread-safe
+    // ProfileChunkedBuffer functions, for this reference to stay valid.
+    const ProfileChunkedBuffer& mBuffer;
+    const ProfileBufferChunk* mChunks0;
+    const ProfileBufferChunk* mChunks1;
+  };
+
+  // In in-session, call `aCallback(ProfileChunkedBuffer::Reader&)` and return
+  // true. Callback should not store `Reader`, because it may become invalid
+  // after this call.
+  // If out-of-session, return false (callback is not invoked).
+  template <typename Callback>
+  [[nodiscard]] auto Read(Callback&& aCallback) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    if (MOZ_UNLIKELY(!mChunkManager)) {
+      // Out-of-session.
+      return std::forward<Callback>(aCallback)(static_cast<Reader*>(nullptr));
+    }
+    return mChunkManager->PeekExtantReleasedChunks(
+        [&](const ProfileBufferChunk* aOldestChunk) {
+          Reader reader(*this, aOldestChunk, mCurrentChunk.get());
+          return std::forward<Callback>(aCallback)(&reader);
+        });
+  }
+
+  // Invoke `aCallback(ProfileBufferEntryReader& [, ProfileBufferBlockIndex])`
+  // on each entry, it must read or at least skip everything. Either/both chunk
+  // pointers may be null.
+  template <typename Callback>
+  static void ReadEach(const ProfileBufferChunk* aChunks0,
+                       const ProfileBufferChunk* aChunks1,
+                       Callback&& aCallback) {
+    static_assert(std::is_invocable_v<Callback, ProfileBufferEntryReader&> ||
+                      std::is_invocable_v<Callback, ProfileBufferEntryReader&,
+                                          ProfileBufferBlockIndex>,
+                  "ReadEach callback must take ProfileBufferEntryReader& and "
+                  "optionally a ProfileBufferBlockIndex");
+    detail::InChunkPointer p{aChunks0, aChunks1};
+    while (!p.IsNull()) {
+      // The position right before an entry size *is* a block index.
+      const ProfileBufferBlockIndex blockIndex =
+          ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+              p.GlobalRangePosition());
+      Length entrySize = p.ReadEntrySize();
+      if (entrySize == 0) {
+        return;
+      }
+      ProfileBufferEntryReader entryReader = p.EntryReader(entrySize);
+      if (entryReader.RemainingBytes() == 0) {
+        return;
+      }
+      MOZ_ASSERT(entryReader.RemainingBytes() == entrySize);
+      if constexpr (std::is_invocable_v<Callback, ProfileBufferEntryReader&,
+                                        ProfileBufferBlockIndex>) {
+        aCallback(entryReader, blockIndex);
+      } else {
+        Unused << blockIndex;
+        aCallback(entryReader);
+      }
+      MOZ_ASSERT(entryReader.RemainingBytes() == 0);
+    }
+  }
+
+  // Invoke `aCallback(ProfileBufferEntryReader& [, ProfileBufferBlockIndex])`
+  // on each entry, it must read or at least skip everything.
+  template <typename Callback>
+  void ReadEach(Callback&& aCallback) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    if (MOZ_UNLIKELY(!mChunkManager)) {
+      // Out-of-session.
+      return;
+    }
+    mChunkManager->PeekExtantReleasedChunks(
+        [&](const ProfileBufferChunk* aOldestChunk) {
+          ReadEach(aOldestChunk, mCurrentChunk.get(),
+                   std::forward<Callback>(aCallback));
+        });
+  }
+
+  // Call `aCallback(Maybe<ProfileBufferEntryReader>&&)` on the entry at
+  // the given ProfileBufferBlockIndex; The `Maybe` will be `Nothing` if
+  // out-of-session, or if that entry doesn't exist anymore, or if we've reached
+  // just past the last entry. Return whatever `aCallback` returns. Callback
+  // should not store `ProfileBufferEntryReader`, because it may become invalid
+  // after this call.
+  // Either/both chunk pointers may be null.
+  template <typename Callback>
+  [[nodiscard]] static auto ReadAt(ProfileBufferBlockIndex aMinimumBlockIndex,
+                                   const ProfileBufferChunk* aChunks0,
+                                   const ProfileBufferChunk* aChunks1,
+                                   Callback&& aCallback) {
+    static_assert(
+        std::is_invocable_v<Callback, Maybe<ProfileBufferEntryReader>&&>,
+        "ReadAt callback must take a Maybe<ProfileBufferEntryReader>&&");
+    Maybe<ProfileBufferEntryReader> maybeEntryReader;
+    if (detail::InChunkPointer p{aChunks0, aChunks1}; !p.IsNull()) {
+      // If the pointer position is before the given position, try to advance.
+      if (p.GlobalRangePosition() >=
+              aMinimumBlockIndex.ConvertToProfileBufferIndex() ||
+          p.AdvanceToGlobalRangePosition(
+              aMinimumBlockIndex.ConvertToProfileBufferIndex())) {
+        MOZ_ASSERT(p.GlobalRangePosition() >=
+                   aMinimumBlockIndex.ConvertToProfileBufferIndex());
+        // Here we're pointing at the start of a block, try to read the entry
+        // size. (Entries cannot be empty, so 0 means failure.)
+        if (Length entrySize = p.ReadEntrySize(); entrySize != 0) {
+          maybeEntryReader.emplace(p.EntryReader(entrySize));
+          if (maybeEntryReader->RemainingBytes() == 0) {
+            // An empty entry reader means there was no complete block at the
+            // given index.
+            maybeEntryReader.reset();
+          } else {
+            MOZ_ASSERT(maybeEntryReader->RemainingBytes() == entrySize);
+          }
+        }
+      }
+    }
+#ifdef DEBUG
+    auto assertAllRead = MakeScopeExit([&]() {
+      MOZ_ASSERT(!maybeEntryReader || maybeEntryReader->RemainingBytes() == 0);
+    });
+#endif  // DEBUG
+    return std::forward<Callback>(aCallback)(std::move(maybeEntryReader));
+  }
+
+  // Call `aCallback(Maybe<ProfileBufferEntryReader>&&)` on the entry at
+  // the given ProfileBufferBlockIndex; The `Maybe` will be `Nothing` if
+  // out-of-session, or if that entry doesn't exist anymore, or if we've reached
+  // just past the last entry. Return whatever `aCallback` returns. Callback
+  // should not store `ProfileBufferEntryReader`, because it may become invalid
+  // after this call.
+  template <typename Callback>
+  [[nodiscard]] auto ReadAt(ProfileBufferBlockIndex aBlockIndex,
+                            Callback&& aCallback) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    if (MOZ_UNLIKELY(!mChunkManager)) {
+      // Out-of-session.
+      return std::forward<Callback>(aCallback)(Nothing{});
+    }
+    return mChunkManager->PeekExtantReleasedChunks(
+        [&](const ProfileBufferChunk* aOldestChunk) {
+          return ReadAt(aBlockIndex, aOldestChunk, mCurrentChunk.get(),
+                        std::forward<Callback>(aCallback));
+        });
+  }
+
+  // Append the contents of another ProfileChunkedBuffer to this one.
+  ProfileBufferBlockIndex AppendContents(const ProfileChunkedBuffer& aSrc) {
+    ProfileBufferBlockIndex firstBlockIndex;
+    // If we start failing, we'll stop writing.
+    bool failed = false;
+    aSrc.ReadEach([&](ProfileBufferEntryReader& aER) {
+      if (failed) {
+        return;
+      }
+      failed =
+          !Put(aER.RemainingBytes(), [&](Maybe<ProfileBufferEntryWriter>& aEW) {
+            if (aEW.isNothing()) {
+              return false;
+            }
+            if (!firstBlockIndex) {
+              firstBlockIndex = aEW->CurrentBlockIndex();
+            }
+            aEW->WriteFromReader(aER, aER.RemainingBytes());
+            return true;
+          });
+    });
+    return failed ? nullptr : firstBlockIndex;
+  }
+
+#ifdef DEBUG
+  void Dump(std::FILE* aFile = stdout) const {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    fprintf(aFile,
+            "ProfileChunkedBuffer[%p] State: range %u-%u pushed=%u cleared=%u "
+            "(live=%u) failed-puts=%u bytes",
+            this, unsigned(mRangeStart), unsigned(mRangeEnd),
+            unsigned(mPushedBlockCount), unsigned(mClearedBlockCount),
+            unsigned(mPushedBlockCount) - unsigned(mClearedBlockCount),
+            unsigned(mFailedPutBytes));
+    if (MOZ_UNLIKELY(!mChunkManager)) {
+      fprintf(aFile, " - Out-of-session\n");
+      return;
+    }
+    fprintf(aFile, " - chunks:\n");
+    bool hasChunks = false;
+    mChunkManager->PeekExtantReleasedChunks(
+        [&](const ProfileBufferChunk* aOldestChunk) {
+          for (const ProfileBufferChunk* chunk = aOldestChunk; chunk;
+               chunk = chunk->GetNext()) {
+            fprintf(aFile, "R ");
+            chunk->Dump(aFile);
+            hasChunks = true;
+          }
+        });
+    if (mCurrentChunk) {
+      fprintf(aFile, "C ");
+      mCurrentChunk->Dump(aFile);
+      hasChunks = true;
+    }
+    for (const ProfileBufferChunk* chunk = mNextChunks.get(); chunk;
+         chunk = chunk->GetNext()) {
+      fprintf(aFile, "N ");
+      chunk->Dump(aFile);
+      hasChunks = true;
+    }
+    switch (mRequestedChunkHolder->GetState()) {
+      case RequestedChunkRefCountedHolder::State::Unused:
+        fprintf(aFile, " - No request pending.\n");
+        break;
+      case RequestedChunkRefCountedHolder::State::Requested:
+        fprintf(aFile, " - Request pending.\n");
+        break;
+      case RequestedChunkRefCountedHolder::State::Fulfilled:
+        fprintf(aFile, " - Request fulfilled.\n");
+        break;
+    }
+    if (!hasChunks) {
+      fprintf(aFile, " No chunks.\n");
+    }
+  }
+#endif  // DEBUG
+
+ private:
+  // Used to de/serialize a ProfileChunkedBuffer (e.g., containing a backtrace).
+  friend ProfileBufferEntryWriter::Serializer<ProfileChunkedBuffer>;
+  friend ProfileBufferEntryReader::Deserializer<ProfileChunkedBuffer>;
+  friend ProfileBufferEntryWriter::Serializer<UniquePtr<ProfileChunkedBuffer>>;
+  friend ProfileBufferEntryReader::Deserializer<
+      UniquePtr<ProfileChunkedBuffer>>;
+
+  [[nodiscard]] UniquePtr<ProfileBufferChunkManager> ResetChunkManager(
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock&) {
+    UniquePtr<ProfileBufferChunkManager> chunkManager;
+    if (mChunkManager) {
+      mRequestedChunkHolder = nullptr;
+      mChunkManager->ForgetUnreleasedChunks();
+#ifdef DEBUG
+      mChunkManager->DeregisteredFrom(this);
+#endif
+      mChunkManager = nullptr;
+      chunkManager = std::move(mOwnedChunkManager);
+      if (mCurrentChunk) {
+        mCurrentChunk->MarkDone();
+        mCurrentChunk = nullptr;
+      }
+      mNextChunks = nullptr;
+      mNextChunkRangeStart = mRangeEnd;
+      mRangeStart = mRangeEnd;
+      mPushedBlockCount = 0;
+      mClearedBlockCount = 0;
+      mFailedPutBytes = 0;
+    }
+    return chunkManager;
+  }
+
+  void SetChunkManager(
+      ProfileBufferChunkManager& aChunkManager,
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
+    MOZ_ASSERT(!mChunkManager);
+    mChunkManager = &aChunkManager;
+#ifdef DEBUG
+    mChunkManager->RegisteredWith(this);
+#endif
+
+    mChunkManager->SetChunkDestroyedCallback(
+        [this](const ProfileBufferChunk& aChunk) {
+          for (;;) {
+            ProfileBufferIndex rangeStart = mRangeStart;
+            if (MOZ_LIKELY(rangeStart <= aChunk.RangeStart())) {
+              if (MOZ_LIKELY(mRangeStart.compareExchange(
+                      rangeStart,
+                      aChunk.RangeStart() + aChunk.BufferBytes()))) {
+                break;
+              }
+            }
+          }
+          mClearedBlockCount += aChunk.BlockCount();
+        });
+
+    // We start with one chunk right away, and request a following one now
+    // so it should be available before the current chunk is full.
+    SetAndInitializeCurrentChunk(mChunkManager->GetChunk(), aLock);
+    mRequestedChunkHolder = MakeRefPtr<RequestedChunkRefCountedHolder>();
+    RequestChunk(aLock);
+  }
+
+  [[nodiscard]] size_t SizeOfExcludingThis(
+      MallocSizeOf aMallocSizeOf,
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock&) const {
+    if (MOZ_UNLIKELY(!mChunkManager)) {
+      // Out-of-session.
+      return 0;
+    }
+    size_t size = mChunkManager->SizeOfIncludingThis(aMallocSizeOf);
+    if (mCurrentChunk) {
+      size += mCurrentChunk->SizeOfIncludingThis(aMallocSizeOf);
+    }
+    if (mNextChunks) {
+      size += mNextChunks->SizeOfIncludingThis(aMallocSizeOf);
+    }
+    return size;
+  }
+
+  void InitializeCurrentChunk(
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock&) {
+    MOZ_ASSERT(!!mCurrentChunk);
+    mCurrentChunk->SetRangeStart(mNextChunkRangeStart);
+    mNextChunkRangeStart += mCurrentChunk->BufferBytes();
+    Unused << mCurrentChunk->ReserveInitialBlockAsTail(0);
+  }
+
+  void SetAndInitializeCurrentChunk(
+      UniquePtr<ProfileBufferChunk>&& aChunk,
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
+    mCurrentChunk = std::move(aChunk);
+    if (mCurrentChunk) {
+      InitializeCurrentChunk(aLock);
+    }
+  }
+
+  void RequestChunk(
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
+    if (HandleRequestedChunk_IsPending(aLock)) {
+      // There is already a pending request, don't start a new one.
+      return;
+    }
+
+    // Ensure the `RequestedChunkHolder` knows we're starting a request.
+    mRequestedChunkHolder->StartRequest();
+
+    // Request a chunk, the callback carries a `RefPtr` of the
+    // `RequestedChunkHolder`, so it's guaranteed to live until it's invoked,
+    // even if this `ProfileChunkedBuffer` changes its `ChunkManager` or is
+    // destroyed.
+    mChunkManager->RequestChunk(
+        [requestedChunkHolder = RefPtr<RequestedChunkRefCountedHolder>(
+             mRequestedChunkHolder)](UniquePtr<ProfileBufferChunk> aChunk) {
+          requestedChunkHolder->AddRequestedChunk(std::move(aChunk));
+        });
+  }
+
+  [[nodiscard]] bool HandleRequestedChunk_IsPending(
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
+    MOZ_ASSERT(!!mChunkManager);
+    MOZ_ASSERT(!!mRequestedChunkHolder);
+
+    if (mRequestedChunkHolder->GetState() ==
+        RequestedChunkRefCountedHolder::State::Unused) {
+      return false;
+    }
+
+    // A request is either in-flight or fulfilled.
+    Maybe<UniquePtr<ProfileBufferChunk>> maybeChunk =
+        mRequestedChunkHolder->GetChunkIfFulfilled();
+    if (maybeChunk.isNothing()) {
+      // Request is still pending.
+      return true;
+    }
+
+    // Since we extracted the provided chunk, the holder should now be unused.
+    MOZ_ASSERT(mRequestedChunkHolder->GetState() ==
+               RequestedChunkRefCountedHolder::State::Unused);
+
+    // Request has been fulfilled.
+    UniquePtr<ProfileBufferChunk>& chunk = *maybeChunk;
+    if (chunk) {
+      // Try to use as current chunk if needed.
+      if (!mCurrentChunk) {
+        SetAndInitializeCurrentChunk(std::move(chunk), aLock);
+        // We've just received a chunk and made it current, request a next chunk
+        // for later.
+        MOZ_ASSERT(!mNextChunks);
+        RequestChunk(aLock);
+        return true;
+      }
+
+      if (!mNextChunks) {
+        mNextChunks = std::move(chunk);
+      } else {
+        mNextChunks->InsertNext(std::move(chunk));
+      }
+    }
+
+    return false;
+  }
+
+  // Get a pointer to the next chunk available
+  [[nodiscard]] ProfileBufferChunk* GetOrCreateCurrentChunk(
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
+    ProfileBufferChunk* current = mCurrentChunk.get();
+    if (MOZ_UNLIKELY(!current)) {
+      // No current chunk ready.
+      MOZ_ASSERT(!mNextChunks,
+                 "There shouldn't be next chunks when there is no current one");
+      // See if a request has recently been fulfilled, ignore pending status.
+      Unused << HandleRequestedChunk_IsPending(aLock);
+      current = mCurrentChunk.get();
+      if (MOZ_UNLIKELY(!current)) {
+        // There was no pending chunk, try to get one right now.
+        // This may still fail, but we can't do anything else about it, the
+        // caller must handle the nullptr case.
+        // Attempt a request for later.
+        SetAndInitializeCurrentChunk(mChunkManager->GetChunk(), aLock);
+        current = mCurrentChunk.get();
+      }
+    }
+    return current;
+  }
+
+  // Get a pointer to the next chunk available
+  [[nodiscard]] ProfileBufferChunk* GetOrCreateNextChunk(
+      const baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock) {
+    MOZ_ASSERT(!!mCurrentChunk,
+               "Why ask for a next chunk when there isn't even a current one?");
+    ProfileBufferChunk* next = mNextChunks.get();
+    if (MOZ_UNLIKELY(!next)) {
+      // No next chunk ready, see if a request has recently been fulfilled,
+      // ignore pending status.
+      Unused << HandleRequestedChunk_IsPending(aLock);
+      next = mNextChunks.get();
+      if (MOZ_UNLIKELY(!next)) {
+        // There was no pending chunk, try to get one right now.
+        mNextChunks = mChunkManager->GetChunk();
+        next = mNextChunks.get();
+        // This may still fail, but we can't do anything else about it, the
+        // caller must handle the nullptr case.
+        if (MOZ_UNLIKELY(!next)) {
+          // Attempt a request for later.
+          RequestChunk(aLock);
+        }
+      }
+    }
+    return next;
+  }
+
+  // Reserve a block of `aCallbackBlockBytes()` size, and invoke and return
+  // `aCallback(Maybe<ProfileBufferEntryWriter>&)`. Note that this is the "raw"
+  // version that doesn't write the entry size at the beginning of the block.
+  // Note: `aCallbackBlockBytes` is a callback instead of a simple value, to
+  // delay this potentially-expensive computation until after we're checked that
+  // we're in-session; use `Put(Length, Callback)` below if you know the size
+  // already.
+  template <typename CallbackBlockBytes, typename Callback>
+  auto ReserveAndPutRaw(CallbackBlockBytes&& aCallbackBlockBytes,
+                        Callback&& aCallback,
+                        baseprofiler::detail::BaseProfilerMaybeAutoLock& aLock,
+                        uint64_t aBlockCount = 1) {
+    // The entry writer that will point into one or two chunks to write
+    // into, empty by default (failure).
+    Maybe<ProfileBufferEntryWriter> maybeEntryWriter;
+
+    // The current chunk will be filled if we need to write more than its
+    // remaining space.
+    bool currentChunkFilled = false;
+
+    // If the current chunk gets filled, we may or may not initialize the next
+    // chunk!
+    bool nextChunkInitialized = false;
+
+    if (MOZ_LIKELY(mChunkManager)) {
+      // In-session.
+
+      const Length blockBytes =
+          std::forward<CallbackBlockBytes>(aCallbackBlockBytes)();
+
+      if (ProfileBufferChunk* current = GetOrCreateCurrentChunk(aLock);
+          MOZ_LIKELY(current)) {
+        if (blockBytes <= current->RemainingBytes()) {
+          // Block fits in current chunk with only one span.
+          currentChunkFilled = blockBytes == current->RemainingBytes();
+          const auto [mem0, blockIndex] = current->ReserveBlock(blockBytes);
+          MOZ_ASSERT(mem0.LengthBytes() == blockBytes);
+          maybeEntryWriter.emplace(
+              mem0, blockIndex,
+              ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+                  blockIndex.ConvertToProfileBufferIndex() + blockBytes));
+          MOZ_ASSERT(maybeEntryWriter->RemainingBytes() == blockBytes);
+          mRangeEnd += blockBytes;
+          mPushedBlockCount += aBlockCount;
+        } else {
+          // Block doesn't fit fully in current chunk, it needs to overflow into
+          // the next one.
+          // Whether or not we can write this entry, the current chunk is now
+          // considered full, so it will be released. (Otherwise we could refuse
+          // this entry, but later accept a smaller entry into this chunk, which
+          // would be somewhat inconsistent.)
+          currentChunkFilled = true;
+          // Make sure the next chunk is available (from a previous request),
+          // otherwise create one on the spot.
+          if (ProfileBufferChunk* next = GetOrCreateNextChunk(aLock);
+              MOZ_LIKELY(next)) {
+            // Here, we know we have a current and a next chunk.
+            // Reserve head of block at the end of the current chunk.
+            const auto [mem0, blockIndex] =
+                current->ReserveBlock(current->RemainingBytes());
+            MOZ_ASSERT(mem0.LengthBytes() < blockBytes);
+            MOZ_ASSERT(current->RemainingBytes() == 0);
+            // Set the next chunk range, and reserve the needed space for the
+            // tail of the block.
+            next->SetRangeStart(mNextChunkRangeStart);
+            mNextChunkRangeStart += next->BufferBytes();
+            const auto mem1 = next->ReserveInitialBlockAsTail(
+                blockBytes - mem0.LengthBytes());
+            MOZ_ASSERT(next->RemainingBytes() != 0);
+            nextChunkInitialized = true;
+            // Block is split in two spans.
+            maybeEntryWriter.emplace(
+                mem0, mem1, blockIndex,
+                ProfileBufferBlockIndex::CreateFromProfileBufferIndex(
+                    blockIndex.ConvertToProfileBufferIndex() + blockBytes));
+            MOZ_ASSERT(maybeEntryWriter->RemainingBytes() == blockBytes);
+            mRangeEnd += blockBytes;
+            mPushedBlockCount += aBlockCount;
+          } else {
+            // Cannot get a new chunk. Record put failure.
+            mFailedPutBytes += blockBytes;
+          }
+        }
+      } else {
+        // Cannot get a current chunk. Record put failure.
+        mFailedPutBytes += blockBytes;
+      }
+    }  // end of `if (MOZ_LIKELY(mChunkManager))`
+
+    // Here, we either have a `Nothing` (failure), or a non-empty entry writer
+    // pointing at the start of the block.
+
+    // After we invoke the callback and return, we may need to handle the
+    // current chunk being filled.
+    auto handleFilledChunk = MakeScopeExit([&]() {
+      // If the entry writer was not already empty, the callback *must* have
+      // filled the full entry.
+      MOZ_ASSERT(!maybeEntryWriter || maybeEntryWriter->RemainingBytes() == 0);
+
+      if (currentChunkFilled) {
+        // Extract current (now filled) chunk.
+        UniquePtr<ProfileBufferChunk> filled = std::move(mCurrentChunk);
+
+        if (mNextChunks) {
+          // Cycle to the next chunk.
+          mCurrentChunk =
+              std::exchange(mNextChunks, mNextChunks->ReleaseNext());
+
+          // Make sure it is initialized (it is now the current chunk).
+          if (!nextChunkInitialized) {
+            InitializeCurrentChunk(aLock);
+          }
+        }
+
+        // And finally mark filled chunk done and release it.
+        filled->MarkDone();
+        mChunkManager->ReleaseChunk(std::move(filled));
+
+        // Request another chunk if needed.
+        // In most cases, here we should have one current chunk and no next
+        // chunk, so we want to do a request so there hopefully will be a next
+        // chunk available when the current one gets filled.
+        // But we also for a request if we don't even have a current chunk (if
+        // it's too late, it's ok because the next `ReserveAndPutRaw` wil just
+        // allocate one on the spot.)
+        // And if we already have a next chunk, there's no need for more now.
+        if (!mCurrentChunk || !mNextChunks) {
+          RequestChunk(aLock);
+        }
+      }
+    });
+
+    return std::forward<Callback>(aCallback)(maybeEntryWriter);
+  }
+
+  // Reserve a block of `aBlockBytes` size, and invoke and return
+  // `aCallback(Maybe<ProfileBufferEntryWriter>&)`. Note that this is the "raw"
+  // version that doesn't write the entry size at the beginning of the block.
+  template <typename Callback>
+  auto ReserveAndPutRaw(Length aBlockBytes, Callback&& aCallback,
+                        uint64_t aBlockCount) {
+    baseprofiler::detail::BaseProfilerMaybeAutoLock lock(mMutex);
+    return ReserveAndPutRaw([aBlockBytes]() { return aBlockBytes; },
+                            std::forward<Callback>(aCallback), lock,
+                            aBlockCount);
+  }
+
+  // Mutex guarding the following members.
+  mutable baseprofiler::detail::BaseProfilerMaybeMutex mMutex;
+
+  // Pointer to the current Chunk Manager (or null when out-of-session.)
+  // It may be owned locally (see below) or externally.
+  ProfileBufferChunkManager* mChunkManager = nullptr;
+
+  // Only non-null when we own the current Chunk Manager.
+  UniquePtr<ProfileBufferChunkManager> mOwnedChunkManager;
+
+  UniquePtr<ProfileBufferChunk> mCurrentChunk;
+
+  UniquePtr<ProfileBufferChunk> mNextChunks;
+
+  // Class used to transfer requested chunks from a `ChunkManager` to a
+  // `ProfileChunkedBuffer`.
+  // It needs to be ref-counted because the request may be fulfilled
+  // asynchronously, and either side may be destroyed during the request.
+  // It cannot use the `ProfileChunkedBuffer` mutex, because that buffer and its
+  // mutex could be destroyed during the request.
+  class RequestedChunkRefCountedHolder
+      : public external::AtomicRefCounted<RequestedChunkRefCountedHolder> {
+   public:
+    MOZ_DECLARE_REFCOUNTED_TYPENAME(RequestedChunkRefCountedHolder)
+
+    enum class State { Unused, Requested, Fulfilled };
+
+    // Get the current state. Note that it may change after the function
+    // returns, so it should be used carefully, e.g., `ProfileChunkedBuffer` can
+    // see if a request is pending or fulfilled, to avoid starting another
+    // request.
+    [[nodiscard]] State GetState() const {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
+      return mState;
+    }
+
+    // Must be called by `ProfileChunkedBuffer` when it requests a chunk.
+    // There cannot be more than one request in-flight.
+    void StartRequest() {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
+      MOZ_ASSERT(mState == State::Unused, "Already requested or fulfilled");
+      mState = State::Requested;
+    }
+
+    // Must be called by the `ChunkManager` with a chunk.
+    // If the `ChunkManager` cannot provide a chunk (because of memory limits,
+    // or it gets destroyed), it must call this anyway with a nullptr.
+    void AddRequestedChunk(UniquePtr<ProfileBufferChunk>&& aChunk) {
+      baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
+      MOZ_ASSERT(mState == State::Requested);
+      mState = State::Fulfilled;
+      mRequestedChunk = std::move(aChunk);
+    }
+
+    // The `ProfileChunkedBuffer` can try to extract the provided chunk after a
+    // request:
+    // - Nothing -> Request is not fulfilled yet.
+    // - Some(nullptr) -> The `ChunkManager` was not able to provide a chunk.
+    // - Some(chunk) -> Requested chunk.
+    [[nodiscard]] Maybe<UniquePtr<ProfileBufferChunk>> GetChunkIfFulfilled() {
+      Maybe<UniquePtr<ProfileBufferChunk>> maybeChunk;
+      baseprofiler::detail::BaseProfilerAutoLock lock(mRequestMutex);
+      MOZ_ASSERT(mState == State::Requested || mState == State::Fulfilled);
+      if (mState == State::Fulfilled) {
+        mState = State::Unused;
+        maybeChunk.emplace(std::move(mRequestedChunk));
+      }
+      return maybeChunk;
+    }
+
+   private:
+    // Mutex guarding the following members.
+    mutable baseprofiler::detail::BaseProfilerMutex mRequestMutex;
+    State mState = State::Unused;
+    UniquePtr<ProfileBufferChunk> mRequestedChunk;
+  };
+
+  // Requested-chunk holder, kept alive when in-session, but may also live
+  // longer if a request is in-flight.
+  RefPtr<RequestedChunkRefCountedHolder> mRequestedChunkHolder;
+
+  // Range start of the next chunk to become current. Starting at 1 because
+  // 0 is a reserved index similar to nullptr.
+  ProfileBufferIndex mNextChunkRangeStart = 1;
+
+  // Index to the first block.
+  // Atomic because it may be increased when a Chunk is destroyed, and the
+  // callback may be invoked from anywhere, including from inside one of our
+  // locked section, so we cannot protect it with a mutex.
+  Atomic<ProfileBufferIndex, MemoryOrdering::ReleaseAcquire> mRangeStart{1};
+
+  // Index past the last block. Equals mRangeStart if empty.
+  ProfileBufferIndex mRangeEnd = 1;
+
+  // Number of blocks that have been pushed into this buffer.
+  uint64_t mPushedBlockCount = 0;
+
+  // Number of blocks that have been removed from this buffer.
+  // Note: Live entries = pushed - cleared.
+  // Atomic because it may be updated when a Chunk is destroyed, and the
+  // callback may be invoked from anywhere, including from inside one of our
+  // locked section, so we cannot protect it with a mutex.
+  Atomic<uint64_t, MemoryOrdering::ReleaseAcquire> mClearedBlockCount{0};
+
+  // Number of bytes that could not be put into this buffer.
+  uint64_t mFailedPutBytes = 0;
+};
+
+// ----------------------------------------------------------------------------
+// ProfileChunkedBuffer serialization
+
+// A ProfileChunkedBuffer can hide another one!
+// This will be used to store marker backtraces; They can be read back into a
+// UniquePtr<ProfileChunkedBuffer>.
+// Format: len (ULEB128) | start | end | buffer (len bytes) | pushed | cleared
+// len==0 marks an out-of-session buffer, or empty buffer.
+template <>
+struct ProfileBufferEntryWriter::Serializer<ProfileChunkedBuffer> {
+  static Length Bytes(const ProfileChunkedBuffer& aBuffer) {
+    return aBuffer.Read([&](ProfileChunkedBuffer::Reader* aReader) {
+      if (!aReader) {
+        // Out-of-session, we only need 1 byte to store a length of 0.
+        return ULEB128Size<Length>(0);
+      }
+      ProfileBufferEntryReader reader = aReader->SingleChunkDataAsEntry();
+      const ProfileBufferIndex start =
+          reader.CurrentBlockIndex().ConvertToProfileBufferIndex();
+      const ProfileBufferIndex end =
+          reader.NextBlockIndex().ConvertToProfileBufferIndex();
+      MOZ_ASSERT(end - start <= std::numeric_limits<Length>::max());
+      const Length len = static_cast<Length>(end - start);
+      if (len == 0) {
+        // In-session but empty, also store a length of 0.
+        return ULEB128Size<Length>(0);
+      }
+      // In-session.
+      return static_cast<Length>(ULEB128Size(len) + sizeof(start) + len +
+                                 sizeof(aBuffer.mPushedBlockCount) +
+                                 sizeof(aBuffer.mClearedBlockCount));
+    });
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const ProfileChunkedBuffer& aBuffer) {
+    aBuffer.Read([&](ProfileChunkedBuffer::Reader* aReader) {
+      if (!aReader) {
+        // Out-of-session, only store a length of 0.
+        aEW.WriteULEB128<Length>(0);
+        return;
+      }
+      ProfileBufferEntryReader reader = aReader->SingleChunkDataAsEntry();
+      const ProfileBufferIndex start =
+          reader.CurrentBlockIndex().ConvertToProfileBufferIndex();
+      const ProfileBufferIndex end =
+          reader.NextBlockIndex().ConvertToProfileBufferIndex();
+      MOZ_ASSERT(end - start <= std::numeric_limits<Length>::max());
+      const Length len = static_cast<Length>(end - start);
+      MOZ_ASSERT(len <= aEW.RemainingBytes());
+      if (len == 0) {
+        // In-session but empty, only store a length of 0.
+        aEW.WriteULEB128<Length>(0);
+        return;
+      }
+      // In-session.
+      // Store buffer length, and start index.
+      aEW.WriteULEB128(len);
+      aEW.WriteObject(start);
+      // Write all the bytes.
+      aEW.WriteFromReader(reader, reader.RemainingBytes());
+      // And write stats.
+      aEW.WriteObject(static_cast<uint64_t>(aBuffer.mPushedBlockCount));
+      aEW.WriteObject(static_cast<uint64_t>(aBuffer.mClearedBlockCount));
+      // Note: Failed pushes are not important to serialize.
+    });
+  }
+};
+
+// A serialized ProfileChunkedBuffer can be read into an empty buffer (either
+// out-of-session, or in-session with enough room).
+template <>
+struct ProfileBufferEntryReader::Deserializer<ProfileChunkedBuffer> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       ProfileChunkedBuffer& aBuffer) {
+    // Expect an empty buffer, as we're going to overwrite it.
+    MOZ_ASSERT(aBuffer.GetState().mRangeStart == aBuffer.GetState().mRangeEnd);
+    // Read the stored buffer length.
+    const auto len = aER.ReadULEB128<ProfileChunkedBuffer::Length>();
+    if (len == 0) {
+      // 0-length means an "uninteresting" buffer, just return now.
+      return;
+    }
+    // We have a non-empty buffer to read.
+
+    // Read start and end indices.
+    const auto start = aER.ReadObject<ProfileBufferIndex>();
+    aBuffer.mRangeStart = start;
+    // For now, set the end to be the start (the buffer is still empty). It will
+    // be updated in `ReserveAndPutRaw()` below.
+    aBuffer.mRangeEnd = start;
+
+    if (aBuffer.IsInSession()) {
+      // Output buffer is in-session (i.e., it already has a memory buffer
+      // attached). Make sure the caller allocated enough space.
+      MOZ_RELEASE_ASSERT(aBuffer.BufferLength().value() >= len);
+    } else {
+      // Output buffer is out-of-session, set a new chunk manager that will
+      // provide a single chunk of just the right size.
+      aBuffer.SetChunkManager(MakeUnique<ProfileBufferChunkManagerSingle>(len));
+      MOZ_ASSERT(aBuffer.BufferLength().value() >= len);
+    }
+
+    // Copy bytes into the buffer.
+    aBuffer.ReserveAndPutRaw(
+        len,
+        [&](Maybe<ProfileBufferEntryWriter>& aEW) {
+          MOZ_RELEASE_ASSERT(aEW.isSome());
+          aEW->WriteFromReader(aER, len);
+        },
+        0);
+    // Finally copy stats.
+    aBuffer.mPushedBlockCount = aER.ReadObject<uint64_t>();
+    aBuffer.mClearedBlockCount = aER.ReadObject<uint64_t>();
+    // Failed puts are not important to keep.
+    aBuffer.mFailedPutBytes = 0;
+  }
+
+  // We cannot output a ProfileChunkedBuffer object (not copyable), use
+  // `ReadInto()` or `aER.ReadObject<UniquePtr<BlocksRinbBuffer>>()` instead.
+  static ProfileChunkedBuffer Read(ProfileBufferEntryReader& aER) = delete;
+};
+
+// A ProfileChunkedBuffer is usually refererenced through a UniquePtr, for
+// convenience we support (de)serializing that UniquePtr directly.
+// This is compatible with the non-UniquePtr serialization above, with a null
+// pointer being treated like an out-of-session or empty buffer; and any of
+// these would be deserialized into a null pointer.
+template <>
+struct ProfileBufferEntryWriter::Serializer<UniquePtr<ProfileChunkedBuffer>> {
+  static Length Bytes(const UniquePtr<ProfileChunkedBuffer>& aBufferUPtr) {
+    if (!aBufferUPtr) {
+      // Null pointer, treat it like an empty buffer, i.e., write length of 0.
+      return ULEB128Size<Length>(0);
+    }
+    // Otherwise write the pointed-at ProfileChunkedBuffer (which could be
+    // out-of-session or empty.)
+    return SumBytes(*aBufferUPtr);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    const UniquePtr<ProfileChunkedBuffer>& aBufferUPtr) {
+    if (!aBufferUPtr) {
+      // Null pointer, treat it like an empty buffer, i.e., write length of 0.
+      aEW.WriteULEB128<Length>(0);
+      return;
+    }
+    // Otherwise write the pointed-at ProfileChunkedBuffer (which could be
+    // out-of-session or empty.)
+    aEW.WriteObject(*aBufferUPtr);
+  }
+};
+
+// Serialization of a raw pointer to ProfileChunkedBuffer.
+// Use Deserializer<UniquePtr<ProfileChunkedBuffer>> to read it back.
+template <>
+struct ProfileBufferEntryWriter::Serializer<ProfileChunkedBuffer*> {
+  static Length Bytes(ProfileChunkedBuffer* aBufferUPtr) {
+    if (!aBufferUPtr) {
+      // Null pointer, treat it like an empty buffer, i.e., write length of 0.
+      return ULEB128Size<Length>(0);
+    }
+    // Otherwise write the pointed-at ProfileChunkedBuffer (which could be
+    // out-of-session or empty.)
+    return SumBytes(*aBufferUPtr);
+  }
+
+  static void Write(ProfileBufferEntryWriter& aEW,
+                    ProfileChunkedBuffer* aBufferUPtr) {
+    if (!aBufferUPtr) {
+      // Null pointer, treat it like an empty buffer, i.e., write length of 0.
+      aEW.WriteULEB128<Length>(0);
+      return;
+    }
+    // Otherwise write the pointed-at ProfileChunkedBuffer (which could be
+    // out-of-session or empty.)
+    aEW.WriteObject(*aBufferUPtr);
+  }
+};
+
+template <>
+struct ProfileBufferEntryReader::Deserializer<UniquePtr<ProfileChunkedBuffer>> {
+  static void ReadInto(ProfileBufferEntryReader& aER,
+                       UniquePtr<ProfileChunkedBuffer>& aBuffer) {
+    aBuffer = Read(aER);
+  }
+
+  static UniquePtr<ProfileChunkedBuffer> Read(ProfileBufferEntryReader& aER) {
+    UniquePtr<ProfileChunkedBuffer> bufferUPtr;
+    // Keep a copy of the reader before reading the length, so we can restart
+    // from here below.
+    ProfileBufferEntryReader readerBeforeLen = aER;
+    // Read the stored buffer length.
+    const auto len = aER.ReadULEB128<ProfileChunkedBuffer::Length>();
+    if (len == 0) {
+      // 0-length means an "uninteresting" buffer, just return nullptr.
+      return bufferUPtr;
+    }
+    // We have a non-empty buffer.
+    // allocate an empty ProfileChunkedBuffer without mutex.
+    bufferUPtr = MakeUnique<ProfileChunkedBuffer>(
+        ProfileChunkedBuffer::ThreadSafety::WithoutMutex);
+    // Rewind the reader before the length and deserialize the contents, using
+    // the non-UniquePtr Deserializer.
+    aER = readerBeforeLen;
+    aER.ReadIntoObject(*bufferUPtr);
+    return bufferUPtr;
+  }
+};
+
+}  // namespace mozilla
+
+#endif  // ProfileChunkedBuffer_h
diff --git a/mozglue/baseprofiler/public/ProfilingCategoryList.h b/mozglue/baseprofiler/public/ProfilingCategoryList.h
new file mode 100644
index 0000000000..437f24aaa1
--- /dev/null
+++ b/mozglue/baseprofiler/public/ProfilingCategoryList.h
@@ -0,0 +1,122 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: set ts=8 sts=2 et sw=2 tw=99:
+ * 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 baseprofiler_ProfilingCategoryList_h
+#define baseprofiler_ProfilingCategoryList_h
+
+// Profiler sub-categories are applied to each sampled stack to describe the
+// type of workload that the CPU is busy with. Only one sub-category can be
+// assigned so be mindful that these are non-overlapping. The active category is
+// set by pushing a label to the profiling stack, or by the unwinder in cases
+// such as JITs. A profile sample in arbitrary C++/Rust will typically be
+// categorized based on the top of the label stack.
+//
+// The list of available color names for categories is:
+//    transparent
+//    blue
+//    green
+//    grey
+//    lightblue
+//    magenta
+//    orange
+//    purple
+//    yellow
+
+// clang-format off
+
+#define MOZ_PROFILING_CATEGORY_LIST(BEGIN_CATEGORY, SUBCATEGORY, END_CATEGORY) \
+  BEGIN_CATEGORY(IDLE, "Idle", "transparent") \
+    SUBCATEGORY(IDLE, IDLE, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(OTHER, "Other", "grey") \
+    SUBCATEGORY(OTHER, OTHER, "Other") \
+    SUBCATEGORY(OTHER, OTHER_PreferenceRead, "Preference Read") \
+    SUBCATEGORY(OTHER, OTHER_Profiling, "Profiling") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(TEST, "Test", "darkgray") \
+    SUBCATEGORY(TEST, TEST, "Test") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(LAYOUT, "Layout", "purple") \
+    SUBCATEGORY(LAYOUT, LAYOUT, "Other") \
+    SUBCATEGORY(LAYOUT, LAYOUT_FrameConstruction, "Frame construction") \
+    SUBCATEGORY(LAYOUT, LAYOUT_Reflow, "Reflow") \
+    SUBCATEGORY(LAYOUT, LAYOUT_CSSParsing, "CSS parsing") \
+    SUBCATEGORY(LAYOUT, LAYOUT_SelectorQuery, "Selector query") \
+    SUBCATEGORY(LAYOUT, LAYOUT_StyleComputation, "Style computation") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(JS, "JavaScript", "yellow") \
+    SUBCATEGORY(JS, JS, "Other") \
+    SUBCATEGORY(JS, JS_Parsing, "Parsing") \
+    SUBCATEGORY(JS, JS_BaselineCompilation, "JIT Compile (baseline)") \
+    SUBCATEGORY(JS, JS_IonCompilation, "JIT Compile (ion)") \
+    SUBCATEGORY(JS, JS_Interpreter, "Interpreter") \
+    SUBCATEGORY(JS, JS_BaselineInterpret, "JIT (baseline-interpreter)") \
+    SUBCATEGORY(JS, JS_Baseline, "JIT (baseline)") \
+    SUBCATEGORY(JS, JS_IonMonkey, "JIT (ion)") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(GCCC, "GC / CC", "orange") \
+    SUBCATEGORY(GCCC, GCCC, "Other") \
+    SUBCATEGORY(GCCC, GCCC_MinorGC, "Minor GC") \
+    SUBCATEGORY(GCCC, GCCC_MajorGC, "Major GC (Other)") \
+    SUBCATEGORY(GCCC, GCCC_MajorGC_Mark, "Major GC (Mark)") \
+    SUBCATEGORY(GCCC, GCCC_MajorGC_Sweep, "Major GC (Sweep)") \
+    SUBCATEGORY(GCCC, GCCC_MajorGC_Compact, "Major GC (Compact)") \
+    SUBCATEGORY(GCCC, GCCC_UnmarkGray, "Unmark Gray") \
+    SUBCATEGORY(GCCC, GCCC_Barrier, "Barrier") \
+    SUBCATEGORY(GCCC, GCCC_FreeSnowWhite, "CC (Free Snow White)") \
+    SUBCATEGORY(GCCC, GCCC_BuildGraph, "CC (Build Graph)") \
+    SUBCATEGORY(GCCC, GCCC_ScanRoots, "CC (Scan Roots)") \
+    SUBCATEGORY(GCCC, GCCC_CollectWhite, "CC (Collect White)") \
+    SUBCATEGORY(GCCC, GCCC_Finalize, "CC (Finalize)") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(NETWORK, "Network", "lightblue") \
+    SUBCATEGORY(NETWORK, NETWORK, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(GRAPHICS, "Graphics", "green") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS, "Other") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS_DisplayListBuilding, "DisplayList building") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS_DisplayListMerging, "DisplayList merging") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS_LayerBuilding, "Layer building") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS_TileAllocation, "Tile allocation") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS_WRDisplayList, "WebRender display list") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS_Rasterization, "Rasterization") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS_FlushingAsyncPaints, "Flushing async paints") \
+    SUBCATEGORY(GRAPHICS, GRAPHICS_ImageDecoding, "Image decoding") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(DOM, "DOM", "blue") \
+    SUBCATEGORY(DOM, DOM, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(JAVA_ANDROID, "Android", "yellow") \
+    SUBCATEGORY(JAVA_ANDROID, JAVA_ANDROID, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(JAVA_ANDROIDX, "AndroidX", "orange") \
+    SUBCATEGORY(JAVA_ANDROIDX, JAVA_ANDROIDX, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(JAVA_LANGUAGE, "Java", "blue") \
+    SUBCATEGORY(JAVA_LANGUAGE, JAVA_LANGUAGE, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(JAVA_MOZILLA, "Mozilla", "green") \
+    SUBCATEGORY(JAVA_MOZILLA, JAVA_MOZILLA, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(JAVA_KOTLIN, "Kotlin", "purple") \
+    SUBCATEGORY(JAVA_KOTLIN, JAVA_KOTLIN, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(JAVA_BLOCKED, "Blocked", "lightblue") \
+    SUBCATEGORY(JAVA_BLOCKED, JAVA_BLOCKED, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(IPC, "IPC", "lightgreen") \
+    SUBCATEGORY(IPC, IPC, "Other") \
+  END_CATEGORY \
+  BEGIN_CATEGORY(MEDIA, "Media", "orange") \
+    SUBCATEGORY(MEDIA, MEDIA, "Other") \
+    SUBCATEGORY(MEDIA, MEDIA_CUBEB, "Cubeb") \
+    SUBCATEGORY(MEDIA, MEDIA_PLAYBACK, "Playback") \
+    SUBCATEGORY(MEDIA, MEDIA_RT, "Real-time rendering") \
+  END_CATEGORY
+
+// clang-format on
+
+#endif  // baseprofiler_ProfilingCategoryList_h
diff --git a/mozglue/baseprofiler/public/leb128iterator.h b/mozglue/baseprofiler/public/leb128iterator.h
new file mode 100644
index 0000000000..636baf916f
--- /dev/null
+++ b/mozglue/baseprofiler/public/leb128iterator.h
@@ -0,0 +1,207 @@
+/* -*- Mode: C++; tab-width: 2; 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/. */
+
+// LEB128 utilities that can read/write unsigned LEB128 numbers from/to
+// iterators.
+//
+// LEB128 = Little Endian Base 128, where small numbers take few bytes, but
+// large numbers are still allowed, which is ideal when serializing numbers that
+// are likely to be small.
+// Each byte contains 7 bits from the number, starting at the "little end", the
+// top bit is 0 for the last byte, 1 otherwise.
+// Numbers 0-127 only take 1 byte. 128-16383 take 2 bytes. Etc.
+//
+// Iterators only need to provide:
+// - `*it` to return a reference to the next byte to be read from or written to.
+// - `++it` to advance the iterator after a byte is written.
+//
+// The caller must always provide sufficient space to write any number, by:
+// - pre-allocating a large enough buffer, or
+// - allocating more space when `++it` reaches the end and/or `*it` is invoked
+//   after the end, or
+// - moving the underlying pointer to an appropriate location (e.g., wrapping
+//   around a circular buffer).
+// The caller must also provide enough bytes to read a full value (i.e., at
+// least one byte should have its top bit unset), and a type large enough to
+// hold the stored value.
+//
+// Note: There are insufficient checks for validity! These functions are
+// intended to be used together, i.e., the user should only `ReadULEB128()` from
+// a sufficiently-large buffer that the same user filled with `WriteULEB128()`.
+// Using with externally-sourced data (e.g., DWARF) is *not* recommended.
+//
+// https://en.wikipedia.org/wiki/LEB128
+
+#ifndef leb128iterator_h
+#define leb128iterator_h
+
+#include "mozilla/Assertions.h"
+#include "mozilla/Likely.h"
+
+#include <climits>
+#include <cstdint>
+#include <limits>
+#include <type_traits>
+
+namespace mozilla {
+
+// Number of bytes needed to represent `aValue`.
+template <typename T>
+constexpr uint_fast8_t ULEB128Size(T aValue) {
+  static_assert(!std::numeric_limits<T>::is_signed,
+                "ULEB128Size only takes unsigned types");
+  // We need one output byte per 7 bits of non-zero value. So we just remove
+  // 7 least significant bits at a time until the value becomes zero.
+  // Note the special case of 0, which still needs 1 output byte; this is done
+  // by starting the first loop before we check for 0.
+  uint_fast8_t size = 0;
+  for (;;) {
+    size += 1;
+    aValue >>= 7;
+    // Expecting small values, so it should be more likely that `aValue == 0`.
+    if (MOZ_LIKELY(aValue == 0)) {
+      return size;
+    }
+  }
+}
+
+// Maximum number of bytes needed to represent any value of type `T`.
+template <typename T>
+constexpr uint_fast8_t ULEB128MaxSize() {
+  return ULEB128Size<T>(std::numeric_limits<T>::max());
+}
+
+// Write `aValue` in LEB128 to `aIterator`.
+// The iterator will be moved past the last byte.
+template <typename T, typename It>
+void WriteULEB128(T aValue, It& aIterator) {
+  static_assert(!std::numeric_limits<T>::is_signed,
+                "WriteULEB128 only takes unsigned types");
+  using IteratorValue = std::remove_reference_t<decltype(*aIterator)>;
+  static_assert(sizeof(IteratorValue) == 1,
+                "WriteULEB128 expects an iterator to single bytes");
+  // 0. Don't test for 0 yet, as we want to output one byte for it.
+  for (;;) {
+    // 1. Extract the 7 least significant bits.
+    const uint_fast8_t byte = aValue & 0x7Fu;
+    // 2. Remove them from `aValue`.
+    aValue >>= 7;
+    // 3. Write the 7 bits, and set the 8th bit if `aValue` is not 0 yet
+    // (meaning there will be more bytes after this one.)
+    // Expecting small values, so it should be more likely that `aValue == 0`.
+    // Note: No absolute need to force-cast to IteratorValue, because we have
+    // only changed the bottom 8 bits above. However the compiler could warn
+    // about a narrowing conversion from potentially-multibyte uint_fast8_t down
+    // to whatever single-byte type `*iterator* expects, so we make it explicit.
+    *aIterator = static_cast<IteratorValue>(
+        MOZ_LIKELY(aValue == 0) ? byte : (byte | 0x80u));
+    // 4. Always advance the iterator to the next byte.
+    ++aIterator;
+    // 5. We're done if `aValue` is 0.
+    // Expecting small values, so it should be more likely that `aValue == 0`.
+    if (MOZ_LIKELY(aValue == 0)) {
+      return;
+    }
+  }
+}
+
+// Read an LEB128 value from `aIterator`.
+// The iterator will be moved past the last byte.
+template <typename T, typename It>
+T ReadULEB128(It& aIterator) {
+  static_assert(!std::numeric_limits<T>::is_signed,
+                "ReadULEB128 must return an unsigned type");
+  using IteratorValue = std::remove_reference_t<decltype(*aIterator)>;
+  static_assert(sizeof(IteratorValue) == 1,
+                "ReadULEB128 expects an iterator to single bytes");
+  // Incoming bits will be added to `result`...
+  T result = 0;
+  // ... starting with the least significant bits.
+  uint_fast8_t shift = 0;
+  for (;;) {
+    // 1. Read one byte from the iterator.
+    // `static_cast` just in case IteratorValue is not implicitly convertible to
+    // uint_fast8_t. It wouldn't matter if the sign was extended, we're only
+    // dealing with the bottom 8 bits below.
+    const uint_fast8_t byte = static_cast<uint_fast8_t>(*aIterator);
+    // 2. Always advance the iterator.
+    ++aIterator;
+    // 3. Extract the 7 bits of value, and shift them in place into `result`.
+    result |= static_cast<T>(byte & 0x7fu) << shift;
+    // 4. If the 8th bit is *not* set, this was the last byte.
+    // Expecting small values, so it should be more likely that the bit is off.
+    if (MOZ_LIKELY((byte & 0x80u) == 0)) {
+      return result;
+    }
+    // There are more bytes to read.
+    // 5. Next byte will contain more significant bits above the past 7.
+    shift += 7;
+    // Safety check that we're not going to shift by >= than the type size,
+    // which is Undefined Behavior in C++.
+    MOZ_ASSERT(shift < CHAR_BIT * sizeof(T));
+  }
+}
+
+// constexpr ULEB128 reader class.
+// Mostly useful when dealing with non-trivial byte feeds.
+template <typename T>
+class ULEB128Reader {
+  static_assert(!std::numeric_limits<T>::is_signed,
+                "ULEB128Reader must handle an unsigned type");
+
+ public:
+  constexpr ULEB128Reader() = default;
+
+  // Don't allow copy/assignment, it doesn't make sense for a stateful parser.
+  constexpr ULEB128Reader(const ULEB128Reader&) = delete;
+  constexpr ULEB128Reader& operator=(const ULEB128Reader&) = delete;
+
+  // Feed a byte into the parser.
+  // Returns true if this was the last byte.
+  [[nodiscard]] constexpr bool FeedByteIsComplete(unsigned aByte) {
+    MOZ_ASSERT(!IsComplete());
+    // Extract the 7 bits of value, and shift them in place into the value.
+    mValue |= static_cast<T>(aByte & 0x7fu) << mShift;
+    // If the 8th bit is *not* set, this was the last byte.
+    // Expecting small values, so it should be more likely that the bit is off.
+    if (MOZ_LIKELY((aByte & 0x80u) == 0)) {
+      mShift = mCompleteShift;
+      return true;
+    }
+    // There are more bytes to read.
+    // Next byte will contain more significant bits above the past 7.
+    mShift += 7;
+    // Safety check that we're not going to shift by >= than the type size,
+    // which is Undefined Behavior in C++.
+    MOZ_ASSERT(mShift < CHAR_BIT * sizeof(T));
+    return false;
+  }
+
+  constexpr void Reset() {
+    mValue = 0;
+    mShift = 0;
+  }
+
+  [[nodiscard]] constexpr bool IsComplete() const {
+    return mShift == mCompleteShift;
+  }
+
+  [[nodiscard]] constexpr T Value() const {
+    MOZ_ASSERT(IsComplete());
+    return mValue;
+  }
+
+ private:
+  // Special value of `mShift` indicating that parsing is complete.
+  constexpr static unsigned mCompleteShift = 0x10000u;
+
+  T mValue = 0;
+  unsigned mShift = 0;
+};
+
+}  // namespace mozilla
+
+#endif  // leb128iterator_h