Bug 938157 - Lightweight CFI/EXIDX unwinding library for SPS. r=n.nethercote, nfroyd, bgirard.

--HG--
extra : rebase_source : bd427749667ddd6641eff414879c3706a5cb5f5e

tools/profiler/AutoObjectMapper.cpp

@@ -0,0 +1,207 @@
+/* -*- 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/NullPtr.h"
+
+#include "PlatformMacros.h"
+#include "AutoObjectMapper.h"
+
+#if defined(SPS_OS_android) && !defined(MOZ_WIDGET_GONK)
+# include <dlfcn.h>
+# include "mozilla/Types.h"
+  // FIXME move these out of mozglue/linker/ElfLoader.h into their
+  // own header, so as to avoid conflicts arising from two definitions
+  // of Array
+  extern "C" {
+    MFBT_API size_t
+    __dl_get_mappable_length(void *handle);
+    MFBT_API void *
+    __dl_mmap(void *handle, void *addr, size_t length, off_t offset);
+    MFBT_API void
+    __dl_munmap(void *handle, void *addr, size_t length);
+  }
+  // The following are for get_installation_lib_dir()
+# include "nsString.h"
+# include "nsDirectoryServiceUtils.h"
+# include "nsDirectoryServiceDefs.h"
+#endif
+
+
+// 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];
+  snprintf(buf, sizeof(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;
+}
+
+
+#if defined(SPS_OS_android) && !defined(MOZ_WIDGET_GONK)
+// A helper function for AutoObjectMapperFaultyLib::Map.  Finds out
+// where the installation's lib directory is, since we'll have to look
+// in there to get hold of libmozglue.so.  Returned C string is heap
+// allocated and the caller must deallocate it.
+static char*
+get_installation_lib_dir()
+{
+  nsCOMPtr<nsIProperties>
+    directoryService(do_GetService(NS_DIRECTORY_SERVICE_CONTRACTID));
+  if (!directoryService) {
+    return nullptr;
+  }
+  nsCOMPtr<nsIFile> greDir;
+  nsresult rv = directoryService->Get(NS_GRE_DIR, NS_GET_IID(nsIFile),
+                                      getter_AddRefs(greDir));
+  if (NS_FAILED(rv)) return nullptr;
+  nsCString path;
+  rv = greDir->GetNativePath(path);
+  if (NS_FAILED(rv)) {
+    return nullptr;
+  }
+  return strdup(path.get());
+}
+
+AutoObjectMapperFaultyLib::AutoObjectMapperFaultyLib(void(*aLog)(const char*))
+  : AutoObjectMapperPOSIX(aLog)
+  , mHdl(nullptr)
+{}
+
+AutoObjectMapperFaultyLib::~AutoObjectMapperFaultyLib() {
+  if (mHdl) {
+    // We've got an object mapped by faulty.lib.  Unmap it via faulty.lib.
+    MOZ_ASSERT(mSize > 0);
+    // Assert on the basis that no valid mapping would start at page zero.
+    MOZ_ASSERT(mImage);
+    __dl_munmap(mHdl, mImage, mSize);
+    dlclose(mHdl);
+    // Stop assertions in ~AutoObjectMapperPOSIX from failing.
+    mImage = nullptr;
+    mSize  = 0;
+  }
+  // At this point the parent class destructor, ~AutoObjectMapperPOSIX,
+  // gets called.  If that has something mapped in the normal way, it
+  // will unmap it in the normal way.  Unfortunately there's no
+  // obvious way to enforce the requirement that the object is mapped
+  // either by faulty.lib or by the parent class, but not by both.
+}
+
+bool AutoObjectMapperFaultyLib::Map(/*OUT*/void** start, /*OUT*/size_t* length,
+                                    std::string fileName)
+{
+  MOZ_ASSERT(!mHdl);
+
+  if (fileName == "libmozglue.so") {
+
+    // Do (2) in the comment above.
+    char* libdir = get_installation_lib_dir();
+    if (libdir) {
+      fileName = std::string(libdir) + "/lib/" + fileName;
+      free(libdir);
+    }
+    // Hand the problem off to the standard mapper.
+    return AutoObjectMapperPOSIX::Map(start, length, fileName);
+
+  } else {
+
+    // Do cases (1) and (3) in the comment above.  We have to
+    // grapple with faulty.lib directly.
+    void* hdl = dlopen(fileName.c_str(), RTLD_GLOBAL | RTLD_LAZY);
+    if (!hdl) {
+      failedToMessage(mLog, "get handle for ELF file", fileName);
+      return false;
+    }
+
+    size_t sz = __dl_get_mappable_length(hdl);
+    if (sz == 0) {
+      dlclose(hdl);
+      failedToMessage(mLog, "get size for ELF file", fileName);
+      return false;
+    }
+
+    void* image = __dl_mmap(hdl, nullptr, sz, 0);
+    if (image == MAP_FAILED) {
+      dlclose(hdl);
+      failedToMessage(mLog, "mmap ELF file", fileName);
+      return false;
+    }
+
+    mHdl   = hdl;
+    mImage = *start  = image;
+    mSize  = *length = sz;
+    return true;
+  }
+}
+
+#endif // defined(SPS_OS_android) && !defined(MOZ_WIDGET_GONK)

tools/profiler/AutoObjectMapper.h

@@ -0,0 +1,115 @@
+/* -*- 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|.
+  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*);
+};
+
+
+#if defined(SPS_OS_android) && !defined(MOZ_WIDGET_GONK)
+// This is a variant of AutoObjectMapperPOSIX suitable for use in
+// conjunction with faulty.lib on Android.  How it behaves depends on
+// the name of the file to be mapped.  There are three possible cases:
+//
+// (1) /foo/bar/xyzzy/blah.apk!/libwurble.so
+//     We hand it as-is to faulty.lib and let it fish the relevant
+//     bits out of the APK.
+//
+// (2) libmozglue.so
+//     This is part of the Fennec installation, but is not in the
+//     APK.  Instead we have to figure out the installation path
+//     and look for it there.  Because of faulty.lib limitations,
+//     we have to use regular open/mmap instead of faulty.lib.
+//
+// (3) libanythingelse.so
+//     faulty.lib assumes this is a system library, and prepends
+//     "/system/lib/" to the path.  So as in (1), we can give it
+//     as-is to faulty.lib.
+//
+// Hence (1) and (3) require special-casing here.  Case (2) simply
+// hands the problem to the parent class.
+
+class MOZ_STACK_CLASS AutoObjectMapperFaultyLib : public AutoObjectMapperPOSIX {
+public:
+  AutoObjectMapperFaultyLib(void(*aLog)(const char*));
+
+  ~AutoObjectMapperFaultyLib();
+
+  bool Map(/*OUT*/void** start, /*OUT*/size_t* length, std::string fileName);
+
+private:
+  // faulty.lib requires us to maintain an abstract handle that can be
+  // used later to unmap the area.  If this is non-NULL, it is assumed
+  // that unmapping is to be done by faulty.lib.  Otherwise it goes
+  // via the normal mechanism.
+  void* mHdl;
+
+  // Disable copying and assignment.
+  AutoObjectMapperFaultyLib(const AutoObjectMapperFaultyLib&);
+  AutoObjectMapperFaultyLib& operator=(const AutoObjectMapperFaultyLib&);
+  // 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 // defined(SPS_OS_android) && !defined(MOZ_WIDGET_GONK)
+
+#endif // AutoObjectMapper_h

tools/profiler/LulCommon.cpp

@@ -0,0 +1,110 @@
+/* -*- 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>
+
+
+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:
+  UniqueString(string str) { str_ = strdup(str.c_str()); }
+  ~UniqueString() { free(reinterpret_cast<void*>(const_cast<char*>(str_))); }
+  const char* str_;
+};
+
+class UniqueStringUniverse {
+public:
+  UniqueStringUniverse() {};
+  const UniqueString* FindOrCopy(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;
+    }
+  }
+private:
+  std::map<string, UniqueString*> map_;
+};
+
+
+const UniqueString* ToUniqueString(string str) {
+  // For the initialisation of sUniverse to be threadsafe,
+  // this relies on C++11's semantics.
+  static UniqueStringUniverse sUniverse;
+  return sUniverse.FindOrCopy(str);
+}
+
+const char* const FromUniqueString(const UniqueString* ustr)
+{
+  return ustr->str_;
+}
+
+}  // namespace lul

tools/profiler/LulCommonExt.h

@@ -0,0 +1,582 @@
+/* -*- 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"
+#include "mozilla/NullPtr.h"
+
+namespace lul {
+
+////////////////////////////////////////////////////////////////
+// UniqueString
+//
+
+// Abstract type
+class UniqueString;
+
+// Unique-ify a string.  |ToUniqueString| can never return nullptr.
+const UniqueString* ToUniqueString(std::string);
+
+// Get the contained C string (debugging only)
+const char* const FromUniqueString(const UniqueString*);
+
+// Some handy pre-uniqified strings.  Z is an escape character:
+//   ZS        '$'
+//   ZD        '.'
+//   Zeq       '='
+//   Zplus     '+'
+//   Zstar     '*'
+//   Zslash    '/'
+//   Zpercent  '%'
+//   Zat       '@'
+//   Zcaret    '^'
+
+// Note that ustr__empty and (UniqueString*)nullptr are considered
+// to be different.
+//
+// Unfortunately these have to be written as functions so as to
+// make them safe to use in static initialisers.
+
+// ""
+inline static const UniqueString* ustr__empty() {
+  static const UniqueString* us = nullptr;
+  if (!us) us = ToUniqueString("");
+  return us;
+}
+
+// ".cfa"
+inline static const UniqueString* ustr__ZDcfa() {
+  static const UniqueString* us = nullptr;
+  if (!us) us = ToUniqueString(".cfa");
+  return us;
+}
+
+// ".ra"
+inline static const UniqueString* ustr__ZDra() {
+  static const UniqueString* us = nullptr;
+  if (!us) us = ToUniqueString(".ra");
+  return us;
+}
+
+
+////////////////////////////////////////////////////////////////
+// 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() {
+    typedef char type_must_be_complete[sizeof(T)];
+    delete ptr;
+  }
+
+  void reset(T* p = 0) {
+    typedef char type_must_be_complete[sizeof(T)];
+
+    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() {
+    typedef char type_must_be_complete[sizeof(T)];
+    delete[] ptr;
+  }
+
+  void reset(T* p = 0) {
+    typedef char type_must_be_complete[sizeof(T)];
+
+    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() {
+    typedef char type_must_be_complete[sizeof(T)];
+    free_((void*) ptr);
+  }
+
+  void reset(T* p = 0) {
+    typedef char type_must_be_complete[sizeof(T)];
+
+    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 (ident == ustr__empty()) {
+        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

tools/profiler/LulDwarfInt.h

@@ -0,0 +1,127 @@
+/* -*- 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'
+};
+
+} // namespace lul
+
+#endif // LulDwarfInt_h

tools/profiler/LulDwarfSummariser.cpp

@@ -0,0 +1,242 @@
+/* -*- 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"
+
+// Set this to 1 for verbose logging
+#define DEBUG_SUMMARISER 0
+
+namespace lul {
+
+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];
+    snprintf(buf, sizeof(buf), "LUL Entry(%llx, %llu)\n",
+             (unsigned long long int)aAddress,
+             (unsigned long long int)aLength);
+    buf[sizeof(buf)-1] = 0;
+    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, int aOldReg, intptr_t aOffset, bool aDeref)
+{
+  aAddress += mTextBias;
+  if (DEBUG_SUMMARISER) {
+    char buf[100];
+    snprintf(buf, sizeof(buf),
+             "LUL  0x%llx  old-r%d = %sr%d + %ld%s\n",
+             (unsigned long long int)aAddress, aNewReg,
+             aDeref ? "*(" : "", aOldReg, (long)aOffset, aDeref ? ")" : "");
+    buf[sizeof(buf)-1] = 0;
+    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;
+  }
+
+  // FIXME: factor out common parts of the arch-dependent summarisers.
+
+#if defined(LUL_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 (aDeref) {
+        goto cant_summarise;
+      }
+      switch (aOldReg) {
+        case DW_REG_ARM_R7:  case DW_REG_ARM_R11:
+        case DW_REG_ARM_R12: case DW_REG_ARM_R13:
+          break;
+        default:
+          goto cant_summarise;
+      }
+      mCurrRules.mCfaExpr = LExpr(LExpr::NODEREF, aOldReg, aOffset);
+      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: {
+      // Check the aOldReg is valid.
+      switch (aOldReg) {
+        case DW_REG_CFA:
+        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:
+          break;
+        default:
+          goto cant_summarise;
+      }
+      // This is a new rule for one of r{7,11,12,13,14,15} and has a
+      // representable offset.  In particular the new value of r15 is
+      // going to be the return address.
+      LExpr expr = LExpr(aDeref ? LExpr::DEREF : LExpr::NODEREF,
+                         aOldReg, aOffset);
+      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:
+      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 == LExpr::UNKNOWN) {
+    mCurrRules.mR7expr = LExpr(LExpr::NODEREF, DW_REG_ARM_R7, 0);
+  }
+  if (mCurrRules.mR11expr.mHow == LExpr::UNKNOWN) {
+    mCurrRules.mR11expr = LExpr(LExpr::NODEREF, DW_REG_ARM_R11, 0);
+  }
+  if (mCurrRules.mR12expr.mHow == LExpr::UNKNOWN) {
+    mCurrRules.mR12expr = LExpr(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(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 == LExpr::UNKNOWN) {
+    mCurrRules.mR15expr = LExpr(LExpr::NODEREF, DW_REG_ARM_R14, 0);
+  }
+
+#elif defined(LUL_ARCH_x64) || defined(LUL_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.  In the 64 bit case
+  // we need to check that aOffset will fit into an int32_t.  In the
+  // 32 bit case it is expected that the compiler will fold out the
+  // test since it always succeeds.
+  if (aNewReg == DW_REG_CFA) {
+    // This is a rule that defines the CFA.  The only forms we can
+    // represent are: = SP+offset or = FP+offset.
+    if (!aDeref && aOffset == (intptr_t)(int32_t)aOffset &&
+        (aOldReg == DW_REG_INTEL_XSP || aOldReg == DW_REG_INTEL_XBP)) {
+      mCurrRules.mCfaExpr = LExpr(LExpr::NODEREF, aOldReg, aOffset);
+    } else {
+      goto cant_summarise;
+    }
+  }
+  else
+  if ((aNewReg == DW_REG_INTEL_XSP ||
+       aNewReg == DW_REG_INTEL_XBP || aNewReg == DW_REG_INTEL_XIP) &&
+      (aOldReg == DW_REG_CFA ||
+       aOldReg == DW_REG_INTEL_XSP ||
+       aOldReg == DW_REG_INTEL_XBP || aOldReg == DW_REG_INTEL_XIP) &&
+      aOffset == (intptr_t)(int32_t)aOffset) {
+    // This is a new rule for SP, BP or the return address
+    // respectively, and has a representable offset.
+    LExpr expr = LExpr(aDeref ? LExpr::DEREF : LExpr::NODEREF,
+                       aOldReg, aOffset);
+    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");
+    }
+  }
+  else {
+    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 == LExpr::UNKNOWN) {
+    mCurrRules.mXspExpr = LExpr(LExpr::NODEREF, DW_REG_CFA, 0);
+  }
+
+  // Also, gcc says "Undef" for BP when it is unchanged.
+  if (mCurrRules.mXbpExpr.mHow == LExpr::UNKNOWN) {
+    mCurrRules.mXbpExpr = LExpr(LExpr::NODEREF, DW_REG_INTEL_XBP, 0);
+  }
+
+#else
+
+# error "Unsupported arch"
+#endif
+
+  return;
+ cant_summarise:
+  if (0) {
+    mLog("LUL  can't summarise\n");
+  }
+}
+
+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

tools/profiler/LulDwarfSummariser.h

@@ -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/. */
+
+#ifndef LulDwarfSummariser_h
+#define LulDwarfSummariser_h
+
+#include "LulMainInt.h"
+
+namespace lul {
+
+class Summariser
+{
+public:
+  Summariser(SecMap* aSecMap, uintptr_t aTextBias, void(*aLog)(const char*));
+
+  void Entry(uintptr_t aAddress, uintptr_t aLength);
+  void End();
+  void Rule(uintptr_t aAddress,
+            int aNewReg, int aOldReg, intptr_t aOffset, bool aDeref);
+
+private:
+  // The SecMap in which we park the finished summaries (RuleSets).
+  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

tools/profiler/LulElfExt.h

@@ -0,0 +1,66 @@
+/* -*- 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 {
+
+// 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,
+                    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,
+                            void (*log)(const char*));
+
+}  // namespace lul
+
+#endif // LulElfExt_h

tools/profiler/LulElfInt.h

@@ -0,0 +1,234 @@
+/* -*- 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 "LulPlatformMacros.h"
+
+
+// (derived from)
+// elfutils.h: Utilities for dealing with ELF files.
+//
+
+#if defined(LUL_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
+
+#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
+
+//FIXME
+extern "C" {
+  extern char*  basename(const char*  path);
+};
+#else
+
+# include <link.h>
+#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

tools/profiler/LulExidx.cpp

@@ -0,0 +1,690 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+
+/* libunwind - a platform-independent unwind library
+   Copyright 2011 Linaro Limited
+
+This file is part of libunwind.
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
+
+
+// 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.
+
+
+// Derived from libunwind, with extensive modifications.
+
+// This file translates EXIDX unwind information into the same format
+// that LUL uses for CFI information.  Hence LUL's CFI unwinding
+// abilities also become usable for EXIDX.
+//
+// See: "Exception Handling ABI for the ARM Architecture", ARM IHI 0038A
+// http://infocenter.arm.com/help/topic/com.arm.doc.ihi0038a/IHI0038A_ehabi.pdf
+
+// EXIDX data is presented in two parts:
+//
+// * an index table.  This contains two words per routine,
+//   the first of which identifies the routine, and the second
+//   of which is a reference to the unwind bytecode.  If the
+//   bytecode is very compact -- 3 bytes or less -- it can be
+//   stored directly in the second word.
+//
+// * an area containing the unwind bytecodes.
+//
+// General flow is: ExceptionTableInfo::Start iterates over all
+// of the index table entries (pairs).  For each entry, it:
+//
+// * calls ExceptionTableInfo::ExtabEntryExtract to copy the bytecode
+//   out into an intermediate buffer.
+
+// * uses ExceptionTableInfo::ExtabEntryDecode to parse the intermediate
+//   buffer.  Each bytecode instruction is bundled into a
+//   arm_ex_to_module::extab_data structure, and handed to ..
+//
+// * .. ARMExToModule::ImproveStackFrame, which in turn hands it to
+//   ARMExToModule::TranslateCmd, and that generates the pseudo-CFI
+//   records that Breakpad stores.
+
+// This file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/arm_ex_to_module.cc
+//   src/common/arm_ex_reader.cc
+
+#include "mozilla/Assertions.h"
+#include "mozilla/NullPtr.h"
+
+#include "LulExidxExt.h"
+
+
+#define ARM_EXBUF_START(x) (((x) >> 4) & 0x0f)
+#define ARM_EXBUF_COUNT(x) ((x) & 0x0f)
+#define ARM_EXBUF_END(x)   (ARM_EXBUF_START(x) + ARM_EXBUF_COUNT(x))
+
+namespace lul {
+
+// Translate command from extab_data to command for Module.
+int ARMExToModule::TranslateCmd(const struct extab_data* edata,
+                                LExpr& vsp) {
+  int ret = 0;
+  switch (edata->cmd) {
+    case ARM_EXIDX_CMD_FINISH:
+      /* Copy LR to PC if there isn't currently a rule for PC in force. */
+      if (curr_rules_.mR15expr.mHow == LExpr::UNKNOWN) {
+        if (curr_rules_.mR14expr.mHow == LExpr::UNKNOWN) {
+          curr_rules_.mR15expr = LExpr(LExpr::NODEREF, DW_REG_ARM_R14, 0);
+        } else {
+          curr_rules_.mR15expr = curr_rules_.mR14expr;
+        }
+      }
+      break;
+    case ARM_EXIDX_CMD_SUB_FROM_VSP:
+      vsp = vsp.add_delta(- static_cast<long>(edata->data));
+      break;
+    case ARM_EXIDX_CMD_ADD_TO_VSP:
+      vsp = vsp.add_delta(static_cast<long>(edata->data));
+      break;
+    case ARM_EXIDX_CMD_REG_POP:
+      for (unsigned int i = 0; i < 16; i++) {
+        if (edata->data & (1 << i)) {
+          // See if we're summarising for int register |i|.  If so,
+          // describe how to pull it off the stack.  The cast of |i| is
+          // a bit of a kludge but works because DW_REG_ARM_Rn has the
+          // value |n|, for 0 <= |n| <= 15 -- that is, for the ARM 
+          // general-purpose registers.
+          LExpr* regI_exprP = curr_rules_.ExprForRegno((DW_REG_NUMBER)i);
+          if (regI_exprP) {
+            *regI_exprP = vsp.deref();
+          }
+          vsp = vsp.add_delta(4);
+        }
+      }
+      /* Set cfa in case the SP got popped. */
+      if (edata->data & (1 << 13)) {
+        vsp = curr_rules_.mR13expr;
+      }
+      break;
+    case ARM_EXIDX_CMD_REG_TO_SP: {
+      MOZ_ASSERT (edata->data < 16);
+      int    reg_no    = edata->data;
+      // Same comment as above, re the casting of |reg_no|, applies.
+      LExpr* reg_exprP = curr_rules_.ExprForRegno((DW_REG_NUMBER)reg_no);
+      if (reg_exprP) {
+        if (reg_exprP->mHow == LExpr::UNKNOWN) {
+          curr_rules_.mR13expr = LExpr(LExpr::NODEREF, reg_no, 0);
+        } else {
+          curr_rules_.mR13expr = *reg_exprP;
+        }
+        vsp = curr_rules_.mR13expr;
+      }
+      break;
+    }
+    case ARM_EXIDX_CMD_VFP_POP:
+      /* Don't recover VFP registers, but be sure to adjust the stack
+         pointer. */
+      for (unsigned int i = ARM_EXBUF_START(edata->data);
+           i <= ARM_EXBUF_END(edata->data); i++) {
+        vsp = vsp.add_delta(8);
+      }
+      if (!(edata->data & ARM_EXIDX_VFP_FSTMD)) {
+        vsp = vsp.add_delta(4);
+      }
+      break;
+    case ARM_EXIDX_CMD_WREG_POP:
+      for (unsigned int i = ARM_EXBUF_START(edata->data);
+           i <= ARM_EXBUF_END(edata->data); i++) {
+        vsp = vsp.add_delta(8);
+      }
+      break;
+    case ARM_EXIDX_CMD_WCGR_POP:
+      // Pop wCGR registers under mask {wCGR3,2,1,0}, hence "i < 4"
+      for (unsigned int i = 0; i < 4; i++) {
+        if (edata->data & (1 << i)) {
+          vsp = vsp.add_delta(4);
+        }
+      }
+      break;
+    case ARM_EXIDX_CMD_REFUSED:
+    case ARM_EXIDX_CMD_RESERVED:
+      ret = -1;
+      break;
+  }
+  return ret;
+}
+
+void ARMExToModule::AddStackFrame(uintptr_t addr, size_t size) {
+  // Here we are effectively reinitialising the EXIDX summariser for a
+  // new code address range.  smap_ stays unchanged.  All other fields
+  // are reinitialised.
+  vsp_ = LExpr(LExpr::NODEREF, DW_REG_ARM_R13, 0);
+  (void) new (&curr_rules_) RuleSet();
+  curr_rules_.mAddr = (uintptr_t)addr;
+  curr_rules_.mLen  = (uintptr_t)size;
+  if (0) {
+    char buf[100];
+    sprintf(buf, "  AddStackFrame    %llx .. %llx",
+            (uint64_t)addr, (uint64_t)(addr + size - 1));
+    log_(buf);
+  }
+}
+
+int ARMExToModule::ImproveStackFrame(const struct extab_data* edata) {
+  return TranslateCmd(edata, vsp_) ;
+}
+
+void ARMExToModule::DeleteStackFrame() {
+}
+
+void ARMExToModule::SubmitStackFrame() {
+  // JRS: I'm really not sure what this means, or if it is necessary
+  // return address always winds up in pc
+  //stack_frame_entry_->initial_rules[ustr__ZDra()] // ".ra"
+  //  = stack_frame_entry_->initial_rules[ustr__pc()];
+  // maybe don't need to do anything here?
+
+  // the final value of vsp is the new value of sp
+  curr_rules_.mR13expr = vsp_;
+
+  // Finally, add the completed RuleSet to the SecMap
+  if (curr_rules_.mLen > 0) {
+
+    // Futz with the rules for r4 .. r11 in the same way as happens
+    // with the CFI summariser:
+    /* 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 (curr_rules_.mR7expr.mHow == LExpr::UNKNOWN) {
+      curr_rules_.mR7expr = LExpr(LExpr::NODEREF, DW_REG_ARM_R7, 0);
+    }
+    if (curr_rules_.mR11expr.mHow == LExpr::UNKNOWN) {
+      curr_rules_.mR11expr = LExpr(LExpr::NODEREF, DW_REG_ARM_R11, 0);
+    }
+    if (curr_rules_.mR12expr.mHow == LExpr::UNKNOWN) {
+      curr_rules_.mR12expr = LExpr(LExpr::NODEREF, DW_REG_ARM_R12, 0);
+    }
+    if (curr_rules_.mR14expr.mHow == LExpr::UNKNOWN) {
+      curr_rules_.mR14expr = LExpr(LExpr::NODEREF, DW_REG_ARM_R14, 0);
+    }
+
+    // And add them
+    smap_->AddRuleSet(&curr_rules_);
+
+    if (0) {
+      curr_rules_.Print(log_);
+    }
+    if (0) {
+      char buf[100];
+      sprintf(buf, "  SubmitStackFrame %llx .. %llx",
+              (uint64_t)curr_rules_.mAddr,
+              (uint64_t)(curr_rules_.mAddr + curr_rules_.mLen - 1));
+      log_(buf);
+    }
+  }
+}
+
+
+#define ARM_EXIDX_CANT_UNWIND 0x00000001
+#define ARM_EXIDX_COMPACT     0x80000000
+#define ARM_EXTBL_OP_FINISH   0xb0
+#define ARM_EXIDX_TABLE_LIMIT (255*4)
+
+using lul::ARM_EXIDX_CMD_FINISH;
+using lul::ARM_EXIDX_CMD_SUB_FROM_VSP;
+using lul::ARM_EXIDX_CMD_ADD_TO_VSP;
+using lul::ARM_EXIDX_CMD_REG_POP;
+using lul::ARM_EXIDX_CMD_REG_TO_SP;
+using lul::ARM_EXIDX_CMD_VFP_POP;
+using lul::ARM_EXIDX_CMD_WREG_POP;
+using lul::ARM_EXIDX_CMD_WCGR_POP;
+using lul::ARM_EXIDX_CMD_RESERVED;
+using lul::ARM_EXIDX_CMD_REFUSED;
+using lul::exidx_entry;
+using lul::ARM_EXIDX_VFP_SHIFT_16;
+using lul::ARM_EXIDX_VFP_FSTMD;
+using lul::MemoryRange;
+
+
+static void* Prel31ToAddr(const void* addr)
+{
+  uint32_t offset32 = *reinterpret_cast<const uint32_t*>(addr);
+  // sign extend offset32[30:0] to 64 bits -- copy bit 30 to positions
+  // 63:31 inclusive.
+  uint64_t offset64 = offset32;
+  if (offset64 & (1ULL << 30))
+    offset64 |= 0xFFFFFFFF80000000ULL;
+  else
+    offset64 &= 0x000000007FFFFFFFULL;
+  return ((char*)addr) + (uintptr_t)offset64;
+}
+
+
+// Extract unwind bytecode for the function denoted by |entry| into |buf|,
+// and return the number of bytes of |buf| written, along with a code
+// indicating the outcome.
+
+ExceptionTableInfo::ExExtractResult
+ExceptionTableInfo::ExtabEntryExtract(const struct exidx_entry* entry,
+                                      uint8_t* buf, size_t buf_size,
+                                      /*OUT*/size_t* buf_used)
+{
+  MemoryRange mr_out(buf, buf_size);
+
+  *buf_used = 0;
+
+# define PUT_BUF_U8(_byte) \
+  do { if (!mr_out.Covers(*buf_used, 1)) return ExOutBufOverflow; \
+       buf[(*buf_used)++] = (_byte); } while (0)
+
+# define GET_EX_U32(_lval, _addr, _sec_mr) \
+  do { if (!(_sec_mr).Covers(reinterpret_cast<const uint8_t*>(_addr) \
+                             - (_sec_mr).data(), 4)) \
+         return ExInBufOverflow; \
+       (_lval) = *(reinterpret_cast<const uint32_t*>(_addr)); } while (0)
+
+# define GET_EXIDX_U32(_lval, _addr) \
+            GET_EX_U32(_lval, _addr, mr_exidx_)
+# define GET_EXTAB_U32(_lval, _addr) \
+            GET_EX_U32(_lval, _addr, mr_extab_)
+
+  uint32_t data;
+  GET_EXIDX_U32(data, &entry->data);
+
+  // A function can be marked CANT_UNWIND if (eg) it is known to be
+  // at the bottom of the stack.
+  if (data == ARM_EXIDX_CANT_UNWIND)
+    return ExCantUnwind;
+
+  uint32_t  pers;          // personality number
+  uint32_t  extra;         // number of extra data words required
+  uint32_t  extra_allowed; // number of extra data words allowed
+  uint32_t* extbl_data;    // the handler entry, if not inlined
+
+  if (data & ARM_EXIDX_COMPACT) {
+    // The handler table entry has been inlined into the index table entry.
+    // In this case it can only be an ARM-defined compact model, since
+    // bit 31 is 1.  Only personalities 0, 1 and 2 are defined for the
+    // ARM compact model, but 1 and 2 are "Long format" and may require
+    // extra data words.  Hence the allowable personalities here are:
+    //   personality 0, in which case 'extra' has no meaning
+    //   personality 1, with zero extra words
+    //   personality 2, with zero extra words
+    extbl_data = nullptr;
+    pers  = (data >> 24) & 0x0F;
+    extra = (data >> 16) & 0xFF;
+    extra_allowed = 0;
+  }
+  else {
+    // The index table entry is a pointer to the handler entry.  Note
+    // that Prel31ToAddr will read the given address, but we already
+    // range-checked above.
+    extbl_data = reinterpret_cast<uint32_t*>(Prel31ToAddr(&entry->data));
+    GET_EXTAB_U32(data, extbl_data);
+    if (!(data & ARM_EXIDX_COMPACT)) {
+      // This denotes a "generic model" handler.  That will involve
+      // executing arbitary machine code, which is something we
+      // can't represent here; hence reject it.
+      return ExCantRepresent;
+    }
+    // So we have a compact model representation.  Again, 3 possible
+    // personalities, but this time up to 255 allowable extra words.
+    pers  = (data >> 24) & 0x0F;
+    extra = (data >> 16) & 0xFF;
+    extra_allowed = 255;
+    extbl_data++;
+  }
+
+  // Now look at the the handler table entry.  The first word is
+  // |data| and subsequent words start at |*extbl_data|.  The number
+  // of extra words to use is |extra|, provided that the personality
+  // allows extra words.  Even if it does, none may be available --
+  // extra_allowed is the maximum number of extra words allowed. */
+  if (pers == 0) {
+    // "Su16" in the documentation -- 3 unwinding insn bytes
+    // |extra| has no meaning here; instead that byte is an unwind-info byte
+    PUT_BUF_U8(data >> 16);
+    PUT_BUF_U8(data >> 8);
+    PUT_BUF_U8(data);
+  }
+  else if ((pers == 1 || pers == 2) && extra <= extra_allowed) {
+    // "Lu16" or "Lu32" respectively -- 2 unwinding insn bytes,
+    // and up to 255 extra words.
+    PUT_BUF_U8(data >> 8);
+    PUT_BUF_U8(data);
+    for (uint32_t j = 0; j < extra; j++) {
+      GET_EXTAB_U32(data, extbl_data);
+      extbl_data++;
+      PUT_BUF_U8(data >> 24);
+      PUT_BUF_U8(data >> 16);
+      PUT_BUF_U8(data >> 8);
+      PUT_BUF_U8(data >> 0);
+    }
+  }
+  else {
+    // The entry is invalid.
+    return ExInvalid;
+  }
+
+  // Make sure the entry is terminated with "FINISH"
+  if (*buf_used > 0 && buf[(*buf_used) - 1] != ARM_EXTBL_OP_FINISH)
+    PUT_BUF_U8(ARM_EXTBL_OP_FINISH);
+
+  return ExSuccess;
+
+# undef GET_EXTAB_U32
+# undef GET_EXIDX_U32
+# undef GET_U32
+# undef PUT_BUF_U8
+}
+
+
+// Take the unwind information extracted by ExtabEntryExtract
+// and parse it into frame-unwind instructions.  These are as
+// specified in "Table 4, ARM-defined frame-unwinding instructions"
+// in the specification document detailed in comments at the top
+// of this file.
+//
+// This reads from |buf[0, +data_size)|.  It checks for overruns of
+// the input buffer and returns a negative value if that happens, or
+// for any other failure cases.  It returns zero in case of success.
+int ExceptionTableInfo::ExtabEntryDecode(const uint8_t* buf, size_t buf_size)
+{
+  if (buf == nullptr || buf_size == 0)
+    return -1;
+
+  MemoryRange mr_in(buf, buf_size);
+  const uint8_t* buf_initially = buf;
+
+# define GET_BUF_U8(_lval) \
+  do { if (!mr_in.Covers(buf - buf_initially, 1)) return -1; \
+       (_lval) = *(buf++); } while (0)
+
+  const uint8_t* end = buf + buf_size;
+
+  while (buf < end) {
+    struct lul::extab_data edata;
+    memset(&edata, 0, sizeof(edata));
+
+    uint8_t op;
+    GET_BUF_U8(op);
+    if ((op & 0xc0) == 0x00) {
+      // vsp = vsp + (xxxxxx << 2) + 4
+      edata.cmd = ARM_EXIDX_CMD_ADD_TO_VSP;
+      edata.data = (((int)op & 0x3f) << 2) + 4;
+    }
+    else if ((op & 0xc0) == 0x40) {
+      // vsp = vsp - (xxxxxx << 2) - 4
+      edata.cmd = ARM_EXIDX_CMD_SUB_FROM_VSP;
+      edata.data = (((int)op & 0x3f) << 2) + 4;
+    }
+    else if ((op & 0xf0) == 0x80) {
+      uint8_t op2;
+      GET_BUF_U8(op2);
+      if (op == 0x80 && op2 == 0x00) {
+        // Refuse to unwind
+        edata.cmd = ARM_EXIDX_CMD_REFUSED;
+      } else {
+        // Pop up to 12 integer registers under masks {r15-r12},{r11-r4}
+        edata.cmd = ARM_EXIDX_CMD_REG_POP;
+        edata.data = ((op & 0xf) << 8) | op2;
+        edata.data = edata.data << 4;
+      }
+    }
+    else if ((op & 0xf0) == 0x90) {
+      if (op == 0x9d || op == 0x9f) {
+        // 9d: Reserved as prefix for ARM register to register moves
+        // 9f: Reserved as perfix for Intel Wireless MMX reg to reg moves
+        edata.cmd = ARM_EXIDX_CMD_RESERVED;
+      } else {
+        // Set vsp = r[nnnn]
+        edata.cmd = ARM_EXIDX_CMD_REG_TO_SP;
+        edata.data = op & 0x0f;
+      }
+    }
+    else if ((op & 0xf0) == 0xa0) {
+      // Pop r4 to r[4+nnn],          or
+      // Pop r4 to r[4+nnn] and r14   or
+      unsigned end = (op & 0x07);
+      edata.data = (1 << (end + 1)) - 1;
+      edata.data = edata.data << 4;
+      if (op & 0x08) edata.data |= 1 << 14;
+      edata.cmd = ARM_EXIDX_CMD_REG_POP;
+    }
+    else if (op == ARM_EXTBL_OP_FINISH) {
+      // Finish
+      edata.cmd = ARM_EXIDX_CMD_FINISH;
+      buf = end;
+    }
+    else if (op == 0xb1) {
+      uint8_t op2;
+      GET_BUF_U8(op2);
+      if (op2 == 0 || (op2 & 0xf0)) {
+        // Spare
+        edata.cmd = ARM_EXIDX_CMD_RESERVED;
+      } else {
+        // Pop integer registers under mask {r3,r2,r1,r0}
+        edata.cmd = ARM_EXIDX_CMD_REG_POP;
+        edata.data = op2 & 0x0f;
+      }
+    }
+    else if (op == 0xb2) {
+      // vsp = vsp + 0x204 + (uleb128 << 2)
+      uint64_t offset = 0;
+      uint8_t byte, shift = 0;
+      do {
+        GET_BUF_U8(byte);
+        offset |= (byte & 0x7f) << shift;
+        shift += 7;
+      } while ((byte & 0x80) && buf < end);
+      edata.data = offset * 4 + 0x204;
+      edata.cmd = ARM_EXIDX_CMD_ADD_TO_VSP;
+    }
+    else if (op == 0xb3 || op == 0xc8 || op == 0xc9) {
+      // b3: Pop VFP regs D[ssss]    to D[ssss+cccc],    FSTMFDX-ishly
+      // c8: Pop VFP regs D[16+ssss] to D[16+ssss+cccc], FSTMFDD-ishly
+      // c9: Pop VFP regs D[ssss]    to D[ssss+cccc],    FSTMFDD-ishly
+      edata.cmd = ARM_EXIDX_CMD_VFP_POP;
+      GET_BUF_U8(edata.data);
+      if (op == 0xc8) edata.data |= ARM_EXIDX_VFP_SHIFT_16;
+      if (op != 0xb3) edata.data |= ARM_EXIDX_VFP_FSTMD;
+    }
+    else if ((op & 0xf8) == 0xb8 || (op & 0xf8) == 0xd0) {
+      // b8: Pop VFP regs D[8] to D[8+nnn], FSTMFDX-ishly
+      // d0: Pop VFP regs D[8] to D[8+nnn], FSTMFDD-ishly
+      edata.cmd = ARM_EXIDX_CMD_VFP_POP;
+      edata.data = 0x80 | (op & 0x07);
+      if ((op & 0xf8) == 0xd0) edata.data |= ARM_EXIDX_VFP_FSTMD;
+    }
+    else if (op >= 0xc0 && op <= 0xc5) {
+      // Intel Wireless MMX pop wR[10]-wr[10+nnn], nnn != 6,7
+      edata.cmd = ARM_EXIDX_CMD_WREG_POP;
+      edata.data = 0xa0 | (op & 0x07);
+    }
+    else if (op == 0xc6) {
+      // Intel Wireless MMX pop wR[ssss] to wR[ssss+cccc]
+      edata.cmd = ARM_EXIDX_CMD_WREG_POP;
+      GET_BUF_U8(edata.data);
+    }
+    else if (op == 0xc7) {
+      uint8_t op2;
+      GET_BUF_U8(op2);
+      if (op2 == 0 || (op2 & 0xf0)) {
+        // Spare
+        edata.cmd = ARM_EXIDX_CMD_RESERVED;
+      } else {
+        // Intel Wireless MMX pop wCGR registers under mask {wCGR3,2,1,0}
+        edata.cmd = ARM_EXIDX_CMD_WCGR_POP;
+        edata.data = op2 & 0x0f;
+      }
+    }
+    else {
+      // Spare
+      edata.cmd = ARM_EXIDX_CMD_RESERVED;
+    }
+
+    int ret = handler_->ImproveStackFrame(&edata);
+    if (ret < 0) return ret;
+  }
+  return 0;
+
+# undef GET_BUF_U8
+}
+
+void ExceptionTableInfo::Start()
+{
+  const struct exidx_entry* start
+    = reinterpret_cast<const struct exidx_entry*>(mr_exidx_.data());
+  const struct exidx_entry* end
+    = reinterpret_cast<const struct exidx_entry*>(mr_exidx_.data()
+                                                  + mr_exidx_.length());
+
+  // Iterate over each of the EXIDX entries (pairs of 32-bit words).
+  // These occupy the entire .exidx section.
+  for (const struct exidx_entry* entry = start; entry < end; ++entry) {
+
+    // Figure out the code address range that this table entry is
+    // associated with.
+    //
+    // I don't claim to understand the biasing here.  It appears that
+    //   (Prel31ToAddr(&entry->addr))
+    //    - mapping_addr_ + loading_addr_) & 0x7fffffff
+    // produces a SVMA.  Adding the text_bias_ gives plausible AVMAs.
+    uint32_t svma = (reinterpret_cast<char*>(Prel31ToAddr(&entry->addr))
+                     - mapping_addr_ + loading_addr_) & 0x7fffffff;
+    uint32_t next_svma;
+    if (entry < end - 1) {
+      next_svma = (reinterpret_cast<char*>(Prel31ToAddr(&((entry + 1)->addr)))
+                   - mapping_addr_ + loading_addr_) & 0x7fffffff;
+    } else {
+      // This is the last EXIDX entry in the sequence, so we don't
+      // have an address for the start of the next function, to limit
+      // this one.  Instead use the address of the last byte of the
+      // text section associated with this .exidx section, that we
+      // have been given.  So as to avoid junking up the CFI unwind
+      // tables with absurdly large address ranges in the case where
+      // text_last_svma_ is wrong, only use the value if it is nonzero
+      // and within one page of |svma|.  Otherwise assume a length of 1.
+      //
+      // In some cases, gcc has been observed to finish the exidx
+      // section with an entry of length 1 marked CANT_UNWIND,
+      // presumably exactly for the purpose of giving a definite
+      // length for the last real entry, without having to look at
+      // text segment boundaries.
+      bool plausible = false;
+      next_svma = svma + 1;
+      if (text_last_svma_ != 0) {
+        uint32_t maybe_next_svma = text_last_svma_ + 1;
+        if (maybe_next_svma > svma && maybe_next_svma - svma <= 4096) {
+          next_svma = maybe_next_svma;
+          plausible = true;
+        }
+      }
+      if (!plausible) {
+        char buf[100];
+        snprintf(buf, sizeof(buf),
+                 "ExceptionTableInfo: implausible EXIDX last entry size %d"
+                 "; using 1 instead.", (int32_t)(text_last_svma_ - svma));
+        buf[sizeof(buf)-1] = 0;
+        log_(buf);
+      }
+    }
+
+    // Extract the unwind info into |buf|.  This might fail for
+    // various reasons.  It involves reading both the .exidx and
+    // .extab sections.  All accesses to those sections are
+    // bounds-checked.
+    uint8_t buf[ARM_EXIDX_TABLE_LIMIT];
+    size_t buf_used = 0;
+    ExExtractResult res = ExtabEntryExtract(entry, buf, sizeof(buf), &buf_used);
+    if (res != ExSuccess) {
+      // Couldn't extract the unwind info, for some reason.  Move on.
+      switch (res) {
+        case ExInBufOverflow:
+          log_("ExtabEntryExtract: .exidx/.extab section overrun");
+          break;
+        case ExOutBufOverflow:
+          log_("ExtabEntryExtract: bytecode buffer overflow");
+          break;
+        case ExCantUnwind:
+          log_("ExtabEntryExtract: function is marked CANT_UNWIND");
+          break;
+        case ExCantRepresent:
+          log_("ExtabEntryExtract: bytecode can't be represented");
+          break;
+        case ExInvalid:
+          log_("ExtabEntryExtract: index table entry is invalid");
+          break;
+        default: {
+          char buf[100];
+          snprintf(buf, sizeof(buf),
+                   "ExtabEntryExtract: unknown error: %d", (int)res);
+          buf[sizeof(buf)-1] = 0;
+          log_(buf);
+          break;
+        }
+      }
+      continue;
+    }
+
+    // Finally, work through the unwind instructions in |buf| and
+    // create CFI entries that Breakpad can use.  This can also fail.
+    // First, add a new stack frame entry, into which ExtabEntryDecode
+    // will write the CFI entries.
+    handler_->AddStackFrame(svma + text_bias_, next_svma - svma);
+    int ret = ExtabEntryDecode(buf, buf_used);
+    if (ret < 0) {
+      handler_->DeleteStackFrame();
+      char buf[100];
+      snprintf(buf, sizeof(buf),
+               "ExtabEntryDecode: failed with error code: %d", ret);
+      buf[sizeof(buf)-1] = 0;
+      log_(buf);
+      continue;
+    }
+    handler_->SubmitStackFrame();
+  } /* iterating over .exidx */
+}
+
+} // namespace lul

tools/profiler/LulExidxExt.h

@@ -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: */
+
+/* libunwind - a platform-independent unwind library
+   Copyright 2011 Linaro Limited
+
+This file is part of libunwind.
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
+
+
+// Copyright (c) 2010, 2011 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.
+
+
+// Derived from libunwind, with extensive modifications.
+// This file is derived from the following files in
+// toolkit/crashreporter/google-breakpad:
+//   src/common/arm_ex_to_module.h
+//   src/common/memory_range.h
+//   src/common/arm_ex_reader.h
+
+#ifndef LulExidxExt_h
+#define LulExidxExt_h
+
+#include "LulMainInt.h"
+
+using lul::LExpr;
+using lul::RuleSet;
+using lul::SecMap;
+
+namespace lul {
+
+typedef enum extab_cmd {
+  ARM_EXIDX_CMD_FINISH,
+  ARM_EXIDX_CMD_SUB_FROM_VSP,
+  ARM_EXIDX_CMD_ADD_TO_VSP,
+  ARM_EXIDX_CMD_REG_POP,
+  ARM_EXIDX_CMD_REG_TO_SP,
+  ARM_EXIDX_CMD_VFP_POP,
+  ARM_EXIDX_CMD_WREG_POP,
+  ARM_EXIDX_CMD_WCGR_POP,
+  ARM_EXIDX_CMD_RESERVED,
+  ARM_EXIDX_CMD_REFUSED,
+} extab_cmd_t;
+
+struct exidx_entry {
+  uint32_t addr;
+  uint32_t data;
+};
+
+struct extab_data {
+  extab_cmd_t cmd;
+  uint32_t data;
+};
+
+enum extab_cmd_flags {
+  ARM_EXIDX_VFP_SHIFT_16 = 1 << 16,
+  ARM_EXIDX_VFP_FSTMD = 1 << 17, // distinguishes FSTMxxD from FSTMxxX
+};
+
+// Receives information from arm_ex_reader::ExceptionTableInfo
+// and adds it to the SecMap object
+// This is in effect the EXIDX summariser.
+class ARMExToModule {
+ public:
+   ARMExToModule(SecMap* smap, void(*log)(const char*)) : smap_(smap)
+                                                        , log_(log) { }
+  ~ARMExToModule() { }
+  void AddStackFrame(uintptr_t addr, size_t size);
+  int  ImproveStackFrame(const struct extab_data* edata);
+  void DeleteStackFrame();
+  void SubmitStackFrame();
+ private:
+  SecMap* smap_;
+  LExpr   vsp_;        // Always appears to be of the form "sp + offset"
+  RuleSet curr_rules_; // Also holds the address range being summarised
+  // debugging message sink
+  void (*log_)(const char*);
+  int TranslateCmd(const struct extab_data* edata, LExpr& vsp);
+};
+
+
+// (derived from)
+// memory_range.h: Define the google_breakpad::MemoryRange class, which
+// is a lightweight wrapper with a pointer and a length to encapsulate
+// a contiguous range of memory.
+
+// 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(const void* data, size_t length) {
+    Set(data, length);
+  }
+
+  // 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 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_;
+};
+
+
+// This class is a reader for ARM unwind information
+// from .ARM.exidx and .ARM.extab sections.
+class ExceptionTableInfo {
+ public:
+  ExceptionTableInfo(const char* exidx, size_t exidx_size,
+                     const char* extab, size_t extab_size,
+                     uint32_t text_last_svma,
+                     lul::ARMExToModule* handler,
+                     const char* mapping_addr,
+                     uint32_t loading_addr,
+                     uintptr_t text_bias,
+                     void (*log)(const char*))
+      : mr_exidx_(lul::MemoryRange(exidx, exidx_size)),
+        mr_extab_(lul::MemoryRange(extab, extab_size)),
+        text_last_svma_(text_last_svma),
+        handler_(handler), mapping_addr_(mapping_addr),
+        loading_addr_(loading_addr),
+        text_bias_(text_bias),
+        log_(log) { }
+
+  ~ExceptionTableInfo() { }
+
+  // Parses the entries in .ARM.exidx and possibly
+  // in .ARM.extab tables, reports what we find to
+  // arm_ex_to_module::ARMExToModule.
+  void Start();
+
+ private:
+  lul::MemoryRange mr_exidx_;
+  lul::MemoryRange mr_extab_;
+  uint32_t text_last_svma_;
+  lul::ARMExToModule* handler_;
+  const char* mapping_addr_;
+  uint32_t loading_addr_;
+  uintptr_t text_bias_;
+  // debugging message sink
+  void (*log_)(const char*);
+  enum ExExtractResult {
+    ExSuccess,        // success
+    ExInBufOverflow,  // out-of-range while reading .exidx
+    ExOutBufOverflow, // output buffer is too small
+    ExCantUnwind,     // this function is marked CANT_UNWIND
+    ExCantRepresent,  // entry valid, but we can't represent it
+    ExInvalid         // entry is invalid
+  };
+  ExExtractResult
+    ExtabEntryExtract(const struct lul::exidx_entry* entry,
+                      uint8_t* buf, size_t buf_size,
+                      /*OUT*/size_t* buf_used);
+
+  int ExtabEntryDecode(const uint8_t* buf, size_t buf_size);
+};
+
+} // namespace lul
+
+#endif // LulExidxExt_h

tools/profiler/LulMain.h

@@ -0,0 +1,311 @@
+/* -*- 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 <pthread.h>   // pthread_t
+
+#include <map>
+
+#include "LulPlatformMacros.h"
+#include "LulRWLock.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:
+  // Construct a valid one.
+  TaggedUWord(uintptr_t w)
+    : mValue(w)
+    , mValid(true)
+  {}
+
+  // Construct an invalid one.
+  TaggedUWord()
+    : mValue(0)
+    , mValid(false)
+  {}
+
+  // Add in a second one.
+  void Add(TaggedUWord other) {
+    if (mValid && other.Valid()) {
+      mValue += other.Value();
+    } else {
+      mValue = 0;
+      mValid = false;
+    }
+  }
+
+  // Is it word-aligned?
+  bool IsAligned() const {
+    return mValid && (mValue & (sizeof(uintptr_t)-1)) == 0;
+  }
+
+  uintptr_t Value() const { return mValue; }
+  bool      Valid() const { return mValid; }
+
+private:
+  uintptr_t mValue;
+  bool mValid;
+};
+
+
+// The registers, with validity tags, that will be unwound.
+
+struct UnwindRegs {
+#if defined(LUL_ARCH_arm)
+  TaggedUWord r7;
+  TaggedUWord r11;
+  TaggedUWord r12;
+  TaggedUWord r13;
+  TaggedUWord r14;
+  TaggedUWord r15;
+#elif defined(LUL_ARCH_x64) || defined(LUL_ARCH_x86)
+  TaggedUWord xbp;
+  TaggedUWord xsp;
+  TaggedUWord xip;
+#else
+# error "Unknown plat"
+#endif
+};
+
+
+// The maximum number of bytes in a stack snapshot.  This can be
+// increased if necessary, but larger values cost performance, since a
+// stack snapshot needs to be copied between sampling and worker
+// threads for each snapshot.  In practice 32k seems to be enough
+// to get good backtraces.
+static const size_t N_STACK_BYTES = 32768;
+
+// 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];
+};
+
+
+// The core unwinder library class.  Just one of these is needed, and
+// it can be shared by multiple unwinder threads.
+//
+// Access to the library is mediated by a single reader-writer lock.
+// All attempts to change the library's internal shared state -- that
+// is, loading or unloading unwind info -- are forced single-threaded
+// by causing the called routine to acquire a write-lock.  Unwind
+// requests do not change the library's internal shared state and
+// therefore require only a read-lock.  Hence multiple threads can
+// unwind in parallel.
+//
+// The library needs to maintain state which is private to each
+// unwinder thread -- the CFI (Dwarf Call Frame Information) fast
+// cache.  Hence unwinder threads first need to register with the
+// library, so their identities are known.  Also, for maximum
+// effectiveness of the CFI caching, it is preferable to have a small
+// number of very-busy unwinder threads rather than a large number of
+// mostly-idle unwinder threads.
+//
+// None of the methods may be safely called from within a signal
+// handler, since this risks deadlock.  In particular this means
+// a thread may not unwind itself from within a signal handler
+// frame.  It might be safe to call Unwind() on its own stack
+// from not-inside a signal frame, although even that cannot be
+// guaranteed deadlock free.
+
+class PriMap;
+class SegArray;
+class CFICache;
+
+class LUL {
+public:
+  // Create; supply a logging sink.  Initialises the rw-lock.
+  LUL(void (*aLog)(const char*));
+
+  // Destroy.  This acquires mRWlock for writing.  By doing that, waits
+  // for all unwinder threads to finish any Unwind() calls they may be
+  // in.  All resources are freed and all registered unwinder threads
+  // are deregistered.
+  ~LUL();
+
+  // 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.
+  //
+  // Acquires mRWlock for writing.  This must be called only after the
+  // code area in question really has been mapped.
+  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.  Acquires mRWlock for
+  // writing.  This must be called only after the code area in
+  // question really has been mapped.
+  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.
+  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.
+  void NotifyBeforeUnmapAll() {
+    NotifyBeforeUnmap(0, UINTPTR_MAX);
+  }
+
+  // Returns the number of mappings currently registered.  Acquires
+  // mRWlock for writing.
+  size_t CountMappings();
+
+  // Register the calling thread for unwinding.  Acquires mRWlock for
+  // writing.
+  void RegisterUnwinderThread();
+
+  // 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.
+  //
+  // Unwinding may optionally use stack scanning.  The maximum number
+  // of frames that may be recovered by stack scanning is
+  // |aScannedFramesAllowed| and the actual number recovered is
+  // written into *aScannedFramesAcquired.  |aScannedFramesAllowed|
+  // must be less than or equal to |aFramesAvail|.
+  //
+  // 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.
+  //
+  // Acquires mRWlock for reading.  Hence multiple threads may unwind
+  // at once, but no thread may be unwinding whilst the library loads
+  // or discards unwind information.  Returns false if the calling
+  // thread is not registered for unwinding.
+  //
+  // Up to aScannedFramesAllowed stack-scanned frames may be recovered.
+  //
+  // The calling thread must previously have registered itself via
+  // RegisterUnwinderThread.
+  void Unwind(/*OUT*/uintptr_t* aFramePCs,
+              /*OUT*/uintptr_t* aFrameSPs,
+              /*OUT*/size_t* aFramesUsed,
+              /*OUT*/size_t* aScannedFramesAcquired,
+              size_t aFramesAvail,
+              size_t aScannedFramesAllowed,
+              UnwindRegs* aStartRegs, StackImage* aStackImg);
+
+  // The logging sink.  Call to send debug strings to the caller-
+  // specified destination.
+  void (*mLog)(const char*);
+
+private:
+  // Invalidate the caches.  Requires mRWlock to be held for writing;
+  // does not acquire it itself.
+  void InvalidateCFICaches();
+
+  // The one-and-only lock, a reader-writer lock, for the library.
+  LulRWLock* mRWlock;
+
+  // The top level mapping from code address ranges to postprocessed
+  // unwind info.  Basically a sorted array of (addr, len, info)
+  // records.  Threads wishing to query this field must hold mRWlock
+  // for reading.  Threads wishing to modify this field must hold
+  // mRWlock for writing.  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.  Threads
+  // wishing to query this field must hold mRWlock for reading.
+  // Threads wishing to modify this field must hold mRWlock for
+  // writing.
+  SegArray* mSegArray;
+
+  // The thread-local data: a mapping from threads to CFI-fast-caches.
+  // Threads wishing to query this field must hold mRWlock for
+  // reading.  Threads wishing to modify this field must hold mRWlock
+  // for writing.
+  //
+  // The CFICaches themselves are thread-local and can be both read
+  // and written when mRWlock is held for reading.  It would probably
+  // be faster to use the pthread_{set,get}specific functions, but
+  // also more difficult.  This map is queried once per unwind, in
+  // order to get hold of the CFI cache for a given thread.
+  std::map<pthread_t, CFICache*> mCaches;
+};
+
+
+// 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

tools/profiler/LulMainInt.h

@@ -0,0 +1,266 @@
+/* -*- 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 "LulPlatformMacros.h"
+
+#include <vector>
+
+#include "mozilla/Assertions.h"
+
+// 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 {
+
+////////////////////////////////////////////////////////////////
+// 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(LUL_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(LUL_ARCH_x64)
+  // 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(LUL_ARCH_x86)
+  DW_REG_INTEL_XBP = 5,
+  DW_REG_INTEL_XSP = 4,
+  DW_REG_INTEL_XIP = 8,
+#else
+# error "Unknown arch"
+#endif
+};
+
+
+////////////////////////////////////////////////////////////////
+// LExpr                                                      //
+////////////////////////////////////////////////////////////////
+
+// An expression -- very primitive.  Denotes either "register +
+// offset" or a dereferenced version of the same.  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.
+
+struct LExpr {
+  // Denotes an expression with no value.
+  LExpr()
+    : mHow(UNKNOWN)
+    , mReg(0)
+    , mOffset(0)
+  {}
+
+  // Denotes any expressible expression.
+  LExpr(uint8_t how, int16_t reg, int32_t offset)
+    : mHow(how)
+    , mReg(reg)
+    , mOffset(offset)
+  {}
+
+  // 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
+  }
+
+  // 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.
+
+  enum { UNKNOWN=0, // This LExpr denotes no value.
+         NODEREF,   // Value is  (mReg + mOffset).
+         DEREF };   // Value is *(mReg + mOffset).
+
+  uint8_t mHow;    // UNKNOWN, NODEREF or DEREF
+  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*));
+
+  // 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(LUL_ARCH_x64) || defined(LUL_ARCH_x86)
+  LExpr  mXipExpr; // return address
+  LExpr  mXspExpr;
+  LExpr  mXbpExpr;
+#elif defined(LUL_ARCH_arm)
+  LExpr  mR15expr; // return address
+  LExpr  mR14expr;
+  LExpr  mR13expr;
+  LExpr  mR12expr;
+  LExpr  mR11expr;
+  LExpr  mR7expr;
+#else
+#   error "Unknown arch"
+#endif
+};
+
+
+////////////////////////////////////////////////////////////////
+// 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.
+
+  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(RuleSet* rs);
+
+  // 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(); }
+
+  // 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()).
+  std::vector<RuleSet> mRuleSets;
+
+  // A logging sink, for debugging.
+  void (*mLog)(const char*);
+};
+
+} // namespace lul
+
+#endif // ndef LulMainInt_h

tools/profiler/LulPlatformMacros.h

@@ -0,0 +1,53 @@
+/* -*- 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 LulPlatformMacros_h
+#define LulPlatformMacros_h
+
+#include <stdint.h>
+#include <stdlib.h>
+
+// Define platform selection macros in a consistent way.  The primary
+// factorisation is on (ARCH,OS) pairs ("PLATforms") but ARCH_ and OS_
+// macros are defined too, since they are sometimes convenient.
+
+#undef LUL_PLAT_x64_linux
+#undef LUL_PLAT_x86_linux
+#undef LUL_PLAT_arm_android
+#undef LUL_PLAT_x86_android
+
+#undef LUL_ARCH_arm
+#undef LUL_ARCH_x86
+#undef LUL_ARCH_x64
+
+#undef LUL_OS_android
+#undef LUL_OS_linux
+
+#if defined(__linux__) && defined(__x86_64__)
+# define LUL_PLAT_x64_linux 1
+# define LUL_ARCH_x64 1
+# define LUL_OS_linux 1
+
+#elif defined(__linux__) && defined(__i386__) && !defined(__ANDROID__)
+# define LUL_PLAT_x86_linux 1
+# define LUL_ARCH_x86 1
+# define LUL_OS_linux 1
+
+#elif defined(__ANDROID__) && defined(__arm__)
+# define LUL_PLAT_arm_android 1
+# define LUL_ARCH_arm 1
+# define LUL_OS_android 1
+
+#elif defined(__ANDROID__) && defined(__i386__)
+# define LUL_PLAT_x86_android 1
+# define LUL_ARCH_x86 1
+# define LUL_OS_android 1
+
+#else
+# error "Unsupported platform"
+#endif
+
+#endif // LulPlatformMacros_h

tools/profiler/LulRWLock.cpp

@@ -0,0 +1,93 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "mozilla/Assertions.h"
+#include "mozilla/DebugOnly.h"
+#include "mozilla/NullPtr.h"
+
+#include "LulRWLock.h"
+
+
+namespace lul {
+
+// An implementation for targets where libpthread does provide
+// pthread_rwlock_t.  These are straight wrappers around the
+// equivalent pthread functions.
+
+#if defined(LUL_OS_linux)
+
+LulRWLock::LulRWLock() {
+  mozilla::DebugOnly<int> r = pthread_rwlock_init(&mLock, nullptr);
+  MOZ_ASSERT(!r);
+}
+
+LulRWLock::~LulRWLock() {
+  mozilla::DebugOnly<int>r = pthread_rwlock_destroy(&mLock);
+  MOZ_ASSERT(!r);
+}
+
+void
+LulRWLock::WrLock() {
+  mozilla::DebugOnly<int>r = pthread_rwlock_wrlock(&mLock);
+  MOZ_ASSERT(!r);
+}
+
+void
+LulRWLock::RdLock() {
+  mozilla::DebugOnly<int>r = pthread_rwlock_rdlock(&mLock);
+  MOZ_ASSERT(!r);
+}
+
+void
+LulRWLock::Unlock() {
+  mozilla::DebugOnly<int>r = pthread_rwlock_unlock(&mLock);
+  MOZ_ASSERT(!r);
+}
+
+
+// An implementation for cases where libpthread does not provide
+// pthread_rwlock_t.  Currently this is a kludge in that it uses
+// normal mutexes, resulting in the following limitations: (1) at most
+// one reader is allowed at once, and (2) any thread that tries to
+// read-acquire the lock more than once will deadlock.  (2) could be
+// avoided if it were possible to use recursive pthread_mutex_t's.
+
+#elif defined(LUL_OS_android)
+
+LulRWLock::LulRWLock() {
+  mozilla::DebugOnly<int> r = pthread_mutex_init(&mLock, nullptr);
+  MOZ_ASSERT(!r);
+}
+
+LulRWLock::~LulRWLock() {
+  mozilla::DebugOnly<int>r = pthread_mutex_destroy(&mLock);
+  MOZ_ASSERT(!r);
+}
+
+void
+LulRWLock::WrLock() {
+  mozilla::DebugOnly<int>r = pthread_mutex_lock(&mLock);
+  MOZ_ASSERT(!r);
+}
+
+void
+LulRWLock::RdLock() {
+  mozilla::DebugOnly<int>r = pthread_mutex_lock(&mLock);
+  MOZ_ASSERT(!r);
+}
+
+void
+LulRWLock::Unlock() {
+  mozilla::DebugOnly<int>r = pthread_mutex_unlock(&mLock);
+  MOZ_ASSERT(!r);
+}
+
+
+#else
+# error "Unsupported OS"
+#endif
+
+} // namespace lul

tools/profiler/LulRWLock.h

@@ -0,0 +1,42 @@
+/* -*- 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 LulRWLock_h
+#define LulRWLock_h
+
+#include "LulPlatformMacros.h"
+#include <pthread.h>
+
+// This class provides a simple wrapper around reader-writer locks.
+// This is necessary because Android 2.2's libpthread does not
+// provide such functionality, so there has to be a way to provide
+// an alternative with an equivalent interface.
+
+namespace lul {
+
+class LulRWLock {
+
+public:
+  LulRWLock();
+  ~LulRWLock();
+  void WrLock();
+  void RdLock();
+  void Unlock();
+
+private:
+#if defined(LUL_OS_android)
+  pthread_mutex_t mLock;
+#elif defined(LUL_OS_linux)
+  pthread_rwlock_t mLock;
+#else
+# error "Unsupported OS"
+#endif
+
+};
+
+} // namespace lul
+
+#endif // LulRWLock_h

tools/profiler/UnwinderThread2.h

@@ -9,6 +9,11 @@
 #include "GeckoProfilerImpl.h"
 #include "ProfileEntry.h"
 
+#include "PlatformMacros.h"
+#if defined(SPS_OS_android) || defined(SPS_OS_linux)
+# include "LulMain.h"
+#endif
+
 /* Top level exports of UnwinderThread.cpp. */
 
 // Abstract type.  A buffer which is used to transfer information between
@@ -89,4 +94,11 @@ void utb__release_sync_buffer(LinkedUWTBuffer* utb);
 // This typedef must match uwt__release_full_buffer and uwt__finish_sync_buffer
 typedef void (*UTB_RELEASE_FUNC)(ThreadProfile*,UnwinderThreadBuffer*,void*);
 
+#if defined(SPS_OS_android) || defined(SPS_OS_linux)
+// Notify |aLUL| of the objects in the current process, so as to get
+// it to read unwind data for them.  This has to be externally visible
+// so it can be used in LUL unit tests, tools/profiler/tests/gtest/LulTest.cpp.
+void read_procmaps(lul::LUL* aLUL);
+#endif
+
 #endif /* ndef MOZ_UNWINDER_THREAD_2_H */

tools/profiler/moz.build

@@ -45,12 +45,24 @@ if CONFIG['MOZ_ENABLE_PROFILER_SPS']:
     ]
 
     if CONFIG['OS_TARGET'] in ('Android', 'Linux'):
-	# These files cannot be built in unified mode because of name clashes with mozglue headers on Android.
+        UNIFIED_SOURCES += [
+            'AutoObjectMapper.cpp',
+            'LulCommon.cpp',
+            'LulDwarf.cpp',
+            'LulDwarfSummariser.cpp',
+            'LulElf.cpp',
+            'LulMain.cpp',
+            'LulRWLock.cpp',
+        ]
+        # These files cannot be built in unified mode because of name clashes with mozglue headers on Android.
         SOURCES += [
             'platform-linux.cc',
             'shared-libraries-linux.cc',
         ]
         if CONFIG['CPU_ARCH'] == 'arm':
+            UNIFIED_SOURCES += [
+                'LulExidx.cpp',
+            ]
             SOURCES += [
                 'EHABIStackWalk.cpp',
             ]
@@ -82,6 +94,9 @@ if CONFIG['MOZ_ENABLE_PROFILER_SPS']:
     if CONFIG['ANDROID_CPU_ARCH'] == 'armeabi':
         DEFINES['ARCH_ARMV6'] = True
 
+    if CONFIG['ENABLE_TESTS']:
+        DIRS += ['tests/gtest']
+
     FINAL_LIBRARY = 'xul'
 
 EXPORTS += [

tools/profiler/platform.cpp

@@ -237,7 +237,6 @@ bool sps_version2()
   return version == 2;
 }
 
-#if !defined(ANDROID)
 /* Has MOZ_PROFILER_VERBOSE been set? */
 bool moz_profiler_verbose()
 {
@@ -253,7 +252,6 @@ bool moz_profiler_verbose()
 
   return status == 2;
 }
-#endif
 
 static inline const char* name_UnwMode(UnwMode m)
 {
@@ -367,6 +365,7 @@ void read_profiler_env_vars()
     LOGF("SPS: UnwindStackScan   = %d (max dubious frames per unwind).",
         (int)sUnwindStackScan);
     LOG( "SPS: Use env var MOZ_PROFILER_MODE=help for further information.");
+    LOG( "SPS: Note that MOZ_PROFILER_MODE=help sets all values to defaults.");
     LOG( "SPS:");
   }
 }
@@ -394,7 +393,11 @@ void profiler_usage() {
   LOG( "SPS:   The number of dubious (stack-scanned) frames allowed");
   LOG( "SPS: ");
   LOG( "SPS:   MOZ_PROFILER_NEW");
-  LOG( "SPS:   Needs to be set to use Breakpad-based unwinding.");
+  LOG( "SPS:   Needs to be set to use LUL-based unwinding.");
+  LOG( "SPS: ");
+  LOG( "SPS:   MOZ_PROFILER_LUL_TEST");
+  LOG( "SPS:   If set to any value, runs LUL unit tests at startup of");
+  LOG( "SPS:   the unwinder thread, and prints a short summary of results.");
   LOG( "SPS: ");
   LOGF("SPS:   This platform %s native unwinding.",
        is_native_unwinding_avail() ? "supports" : "does not support");
@@ -415,6 +418,7 @@ void profiler_usage() {
   LOGF("SPS: UnwindStackScan   = %d (max dubious frames per unwind).",
        (int)sUnwindStackScan);
   LOG( "SPS: Use env var MOZ_PROFILER_MODE=help for further information.");
+  LOG( "SPS: Note that MOZ_PROFILER_MODE=help sets all values to defaults.");
   LOG( "SPS:");
 
   return;
@@ -614,6 +618,8 @@ void mozilla_sampler_start(int aProfileEntries, double aInterval,
                            const char** aThreadNameFilters, uint32_t aFilterCount)
 
 {
+  LOG("BEGIN mozilla_sampler_start");
+
   if (!stack_key_initialized)
     profiler_init(nullptr);
 
@@ -686,15 +692,20 @@ void mozilla_sampler_start(int aProfileEntries, double aInterval,
   nsCOMPtr<nsIObserverService> os = mozilla::services::GetObserverService();
   if (os)
     os->NotifyObservers(nullptr, "profiler-started", nullptr);
+
+  LOG("END   mozilla_sampler_start");
 }
 
 void mozilla_sampler_stop()
 {
+  LOG("BEGIN mozilla_sampler_stop");
+
   if (!stack_key_initialized)
     profiler_init(nullptr);
 
   TableTicker *t = tlsTicker.get();
   if (!t) {
+    LOG("END   mozilla_sampler_stop-early");
     return;
   }
 
@@ -732,6 +743,8 @@ void mozilla_sampler_stop()
   nsCOMPtr<nsIObserverService> os = mozilla::services::GetObserverService();
   if (os)
     os->NotifyObservers(nullptr, "profiler-stopped", nullptr);
+
+  LOG("END   mozilla_sampler_stop");
 }
 
 bool mozilla_sampler_is_paused() {

tools/profiler/platform.h

@@ -54,18 +54,24 @@
 
 #define ASSERT(a) MOZ_ASSERT(a)
 
+bool moz_profiler_verbose();
+
 #ifdef ANDROID
 # if defined(__arm__) || defined(__thumb__)
 #  define ENABLE_SPS_LEAF_DATA
 #  define ENABLE_ARM_LR_SAVING
 # endif
 # define LOG(text) \
-    __android_log_write(ANDROID_LOG_ERROR, "Profiler", text)
+    do { if (moz_profiler_verbose()) \
+           __android_log_write(ANDROID_LOG_ERROR, "Profiler", text); \
+    } while (0)
 # define LOGF(format, ...) \
-    __android_log_print(ANDROID_LOG_ERROR, "Profiler", format, __VA_ARGS__)
+    do { if (moz_profiler_verbose()) \
+           __android_log_print(ANDROID_LOG_ERROR, "Profiler", format, \
+                               __VA_ARGS__); \
+    } while (0)
 
 #else
-  extern bool moz_profiler_verbose();
 # define LOG(text) \
     do { if (moz_profiler_verbose()) fprintf(stderr, "Profiler: %s\n", text); \
     } while (0)

tools/profiler/tests/gtest/LulTest.cpp

@@ -0,0 +1,50 @@
+/* -*- 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 "gtest/gtest.h"
+#include "LulMain.h"
+#include "GeckoProfiler.h"     // for TracingMetadata
+#include "UnwinderThread2.h"   // for read_procmaps
+
+// Set this to 0 to make LUL be completely silent during tests.
+// Set it to 1 to get logging output from LUL, presumably for
+// the purpose of debugging it.
+#define DEBUG_LUL_TEST 0
+
+// LUL needs a callback for its logging sink.
+static void
+logging_sink_for_LUL(const char* str) {
+  if (DEBUG_LUL_TEST == 0) {
+    return;
+  }
+  // Ignore any trailing \n, since LOG will add one anyway.
+  size_t n = strlen(str);
+  if (n > 0 && str[n-1] == '\n') {
+    char* tmp = strdup(str);
+    tmp[n-1] = 0;
+    fprintf(stderr, "LUL-in-gtest: %s\n", tmp);
+    free(tmp);
+  } else {
+    fprintf(stderr, "LUL-in-gtest: %s\n", str);
+  }
+}
+
+TEST(LUL, unwind_consistency) {
+  // Set up LUL and get it to read unwind info for libxul.so, which is
+  // all we care about here, plus (incidentally) practically every
+  // other object in the process too.
+  lul::LUL* lul = new lul::LUL(logging_sink_for_LUL);
+  lul->RegisterUnwinderThread();
+  read_procmaps(lul);
+
+  // Run unwind tests and receive information about how many there
+  // were and how many were successful.
+  int nTests = 0, nTestsPassed = 0;
+  RunLulUnitTests(&nTests, &nTestsPassed, lul);
+  EXPECT_TRUE(nTests == 6) << "Unexpected number of tests";
+  EXPECT_TRUE(nTestsPassed == nTests) << "Not all tests passed";
+
+  delete lul;
+}

tools/profiler/tests/gtest/moz.build

@@ -0,0 +1,20 @@
+# -*- Mode: python; c-basic-offset: 4; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, you can obtain one at http://mozilla.org/MPL/2.0/.
+
+LIBRARY_NAME = 'sps_test'
+
+if CONFIG['OS_TARGET'] in ('Android', 'Linux'):
+    UNIFIED_SOURCES += [
+        'LulTest.cpp',
+    ]
+
+EXPORT_LIBRARY = True
+
+LOCAL_INCLUDES += [
+    '/tools/profiler',
+]
+
+FINAL_LIBRARY = 'xul-gtest'