Add debug fission support.

This added debug fission support.
It tries to find the dwp file from the debug dir /usr/lib/debug/*/debug
and read symbols from them.

Most of this patch comes from
https://critique.corp.google.com/#review/52048295
and some fixes after that.
The elf_reader.cc comes from TOT google code. I just
removed some google dependency.

Current problems from this patch
1: Some type mismatch: from uint8_t * to char *.
2: Some hack to find the .dwp file. (replace .debug with .dwp)

BUG=chromium:604440
R=dehao@google.com, ivanpe@chromium.org

Review URL: https://codereview.chromium.org/1884283002 .

7 files changed, 2177 insertions, 103 deletions

diff --git a/src/common/dwarf/bytereader.cc b/src/common/dwarf/bytereader.cc
index 3ccbcad6..14b43adb 100644
--- a/src/common/dwarf/bytereader.cc
+++ b/src/common/dwarf/bytereader.cc
@@ -243,4 +243,8 @@ uint64 ByteReader::ReadEncodedPointer(const uint8_t *buffer,
   return pointer;
 }
 
+Endianness ByteReader::GetEndianness() const {
+  return endian_;
+}
+
 }  // namespace dwarf2reader
diff --git a/src/common/dwarf/bytereader.h b/src/common/dwarf/bytereader.h
index cf583094..59d43034 100644
--- a/src/common/dwarf/bytereader.h
+++ b/src/common/dwarf/bytereader.h
@@ -280,6 +280,7 @@ class ByteReader {
                             DwarfPointerEncoding encoding,
                             size_t *len) const;
 
+  Endianness GetEndianness() const;
  private:
 
   // Function pointer type for our address and offset readers.
diff --git a/src/common/dwarf/dwarf2enums.h b/src/common/dwarf/dwarf2enums.h
index 5565d66e..6b8a7245 100644
--- a/src/common/dwarf/dwarf2enums.h
+++ b/src/common/dwarf/dwarf2enums.h
@@ -149,7 +149,10 @@ enum DwarfForm {
   DW_FORM_sec_offset = 0x17,
   DW_FORM_exprloc = 0x18,
   DW_FORM_flag_present = 0x19,
-  DW_FORM_ref_sig8 = 0x20
+  DW_FORM_ref_sig8 = 0x20,
+  // Extensions for Fission.  See http://gcc.gnu.org/wiki/DebugFission.
+  DW_FORM_GNU_addr_index = 0x1f01,
+  DW_FORM_GNU_str_index = 0x1f02
 };
 
 // Attribute names and codes
@@ -264,6 +267,13 @@ enum DwarfAttribute {
   DW_AT_body_begin = 0x2105,
   DW_AT_body_end   = 0x2106,
   DW_AT_GNU_vector = 0x2107,
+  // Extensions for Fission.  See http://gcc.gnu.org/wiki/DebugFission.
+  DW_AT_GNU_dwo_name = 0x2130,
+  DW_AT_GNU_dwo_id = 0x2131,
+  DW_AT_GNU_ranges_base = 0x2132,
+  DW_AT_GNU_addr_base = 0x2133,
+  DW_AT_GNU_pubnames = 0x2134,
+  DW_AT_GNU_pubtypes = 0x2135,
   // VMS extensions.
   DW_AT_VMS_rtnbeg_pd_address = 0x2201,
   // UPC extension.
@@ -491,7 +501,22 @@ enum DwarfOpcode {
   DW_OP_lo_user                      =0xe0,
   DW_OP_hi_user                      =0xff,  
   // GNU extensions
-  DW_OP_GNU_push_tls_address         =0xe0
+  DW_OP_GNU_push_tls_address         =0xe0,
+  // Extensions for Fission.  See http://gcc.gnu.org/wiki/DebugFission.
+  DW_OP_GNU_addr_index               =0xfb,
+  DW_OP_GNU_const_index              =0xfc
+};
+
+// Section identifiers for DWP files
+enum DwarfSectionId {
+  DW_SECT_INFO = 1,
+  DW_SECT_TYPES = 2,
+  DW_SECT_ABBREV = 3,
+  DW_SECT_LINE = 4,
+  DW_SECT_LOC = 5,
+  DW_SECT_STR_OFFSETS = 6,
+  DW_SECT_MACINFO = 7,
+  DW_SECT_MACRO = 8
 };
 
 // Source languages.  These are values for DW_AT_language.
diff --git a/src/common/dwarf/dwarf2reader.cc b/src/common/dwarf/dwarf2reader.cc
index fcd29b8c..003ed7b1 100644
--- a/src/common/dwarf/dwarf2reader.cc
+++ b/src/common/dwarf/dwarf2reader.cc
@@ -44,6 +44,8 @@
 #include <string>
 #include <utility>
 
+#include <sys/stat.h>
+
 #include "common/dwarf/bytereader-inl.h"
 #include "common/dwarf/bytereader.h"
 #include "common/dwarf/line_state_machine.h"
@@ -51,11 +53,38 @@
 
 namespace dwarf2reader {
 
-CompilationUnit::CompilationUnit(const SectionMap& sections, uint64 offset,
+CompilationUnit::CompilationUnit(const string& path,
+                                 const SectionMap& sections, uint64 offset,
                                  ByteReader* reader, Dwarf2Handler* handler)
-    : offset_from_section_start_(offset), reader_(reader),
-      sections_(sections), handler_(handler), abbrevs_(NULL),
-      string_buffer_(NULL), string_buffer_length_(0) {}
+    : path_(path), offset_from_section_start_(offset), reader_(reader),
+      sections_(sections), handler_(handler), abbrevs_(),
+      string_buffer_(NULL), string_buffer_length_(0),
+      str_offsets_buffer_(NULL), str_offsets_buffer_length_(0),
+      addr_buffer_(NULL), addr_buffer_length_(0),
+      is_split_dwarf_(false), dwo_id_(0), dwo_name_(),
+      skeleton_dwo_id_(0), ranges_base_(0), addr_base_(0),
+      have_checked_for_dwp_(false), dwp_path_(),
+      dwp_byte_reader_(), dwp_reader_() {}
+
+// Initialize a compilation unit from a .dwo or .dwp file.
+// In this case, we need the .debug_addr section from the
+// executable file that contains the corresponding skeleton
+// compilation unit.  We also inherit the Dwarf2Handler from
+// the executable file, and call it as if we were still
+// processing the original compilation unit.
+
+void CompilationUnit::SetSplitDwarf(const uint8_t* addr_buffer,
+                                    uint64 addr_buffer_length,
+                                    uint64 addr_base,
+                                    uint64 ranges_base,
+                                    uint64 dwo_id) {
+  is_split_dwarf_ = true;
+  addr_buffer_ = addr_buffer;
+  addr_buffer_length_ = addr_buffer_length;
+  addr_base_ = addr_base;
+  ranges_base_ = ranges_base;
+  skeleton_dwo_id_ = dwo_id;
+}
 
 // Read a DWARF2/3 abbreviation section.
 // Each abbrev consists of a abbreviation number, a tag, a byte
@@ -174,6 +203,8 @@ const uint8_t *CompilationUnit::SkipAttribute(const uint8_t *start,
       return start + strlen(reinterpret_cast<const char *>(start)) + 1;
     case DW_FORM_udata:
     case DW_FORM_ref_udata:
+    case DW_FORM_GNU_str_index:
+    case DW_FORM_GNU_addr_index:
       reader_->ReadUnsignedLEB128(start, &len);
       return start + len;
 
@@ -296,9 +327,31 @@ uint64 CompilationUnit::Start() {
     string_buffer_length_ = iter->second.second;
   }
 
+  // Set the string offsets section if we have one.
+  iter = sections_.find(".debug_str_offsets");
+  if (iter != sections_.end()) {
+    str_offsets_buffer_ = iter->second.first;
+    str_offsets_buffer_length_ = iter->second.second;
+  }
+
+  // Set the address section if we have one.
+  iter = sections_.find(".debug_addr");
+  if (iter != sections_.end()) {
+    addr_buffer_ = iter->second.first;
+    addr_buffer_length_ = iter->second.second;
+  }
+
   // Now that we have our abbreviations, start processing DIE's.
   ProcessDIEs();
 
+  // If this is a skeleton compilation unit generated with split DWARF,
+  // and the client needs the full debug info, we need to find the full
+  // compilation unit in a .dwo or .dwp file.
+  if (!is_split_dwarf_
+      && dwo_name_ != NULL
+      && handler_->NeedSplitDebugInfo())
+    ProcessSplitDwarf();
+
   return ourlength;
 }
 
@@ -320,48 +373,46 @@ const uint8_t *CompilationUnit::ProcessAttribute(
       return ProcessAttribute(dieoffset, start, attr, form);
 
     case DW_FORM_flag_present:
-      handler_->ProcessAttributeUnsigned(dieoffset, attr, form, 1);
+      ProcessAttributeUnsigned(dieoffset, attr, form, 1);
       return start;
     case DW_FORM_data1:
     case DW_FORM_flag:
-      handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
-                                         reader_->ReadOneByte(start));
+      ProcessAttributeUnsigned(dieoffset, attr, form,
+                               reader_->ReadOneByte(start));
       return start + 1;
     case DW_FORM_data2:
-      handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
-                                         reader_->ReadTwoBytes(start));
+      ProcessAttributeUnsigned(dieoffset, attr, form,
+                               reader_->ReadTwoBytes(start));
       return start + 2;
     case DW_FORM_data4:
-      handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
-                                         reader_->ReadFourBytes(start));
+      ProcessAttributeUnsigned(dieoffset, attr, form,
+                               reader_->ReadFourBytes(start));
       return start + 4;
     case DW_FORM_data8:
-      handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
-                                         reader_->ReadEightBytes(start));
+      ProcessAttributeUnsigned(dieoffset, attr, form,
+                               reader_->ReadEightBytes(start));
       return start + 8;
     case DW_FORM_string: {
       const char *str = reinterpret_cast<const char *>(start);
-      handler_->ProcessAttributeString(dieoffset, attr, form,
-                                       str);
+      ProcessAttributeString(dieoffset, attr, form, str);
       return start + strlen(str) + 1;
     }
     case DW_FORM_udata:
-      handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
-                                         reader_->ReadUnsignedLEB128(start,
-                                                                     &len));
+      ProcessAttributeUnsigned(dieoffset, attr, form,
+                               reader_->ReadUnsignedLEB128(start, &len));
       return start + len;
 
     case DW_FORM_sdata:
-      handler_->ProcessAttributeSigned(dieoffset, attr, form,
-                                      reader_->ReadSignedLEB128(start, &len));
+      ProcessAttributeSigned(dieoffset, attr, form,
+                             reader_->ReadSignedLEB128(start, &len));
       return start + len;
     case DW_FORM_addr:
-      handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
-                                         reader_->ReadAddress(start));
+      ProcessAttributeUnsigned(dieoffset, attr, form,
+                               reader_->ReadAddress(start));
       return start + reader_->AddressSize();
     case DW_FORM_sec_offset:
-      handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
-                                         reader_->ReadOffset(start));
+      ProcessAttributeUnsigned(dieoffset, attr, form,
+                               reader_->ReadOffset(start));
       return start + reader_->OffsetSize();
 
     case DW_FORM_ref1:
@@ -441,10 +492,32 @@ const uint8_t *CompilationUnit::ProcessAttribute(
       assert(string_buffer_ + offset < string_buffer_ + string_buffer_length_);
 
       const char *str = reinterpret_cast<const char *>(string_buffer_ + offset);
-      handler_->ProcessAttributeString(dieoffset, attr, form,
-                                       str);
+      ProcessAttributeString(dieoffset, attr, form, str);
       return start + reader_->OffsetSize();
     }
+
+    case DW_FORM_GNU_str_index: {
+      uint64 str_index = reader_->ReadUnsignedLEB128(start, &len);
+      const uint8_t* offset_ptr =
+          str_offsets_buffer_ + str_index * reader_->OffsetSize();
+      const uint64 offset = reader_->ReadOffset(offset_ptr);
+      if (offset >= string_buffer_length_) {
+        return NULL;
+      }
+
+      const char* str = reinterpret_cast<const char *>(string_buffer_) + offset;
+      ProcessAttributeString(dieoffset, attr, form, str);
+      return start + len;
+      break;
+    }
+    case DW_FORM_GNU_addr_index: {
+      uint64 addr_index = reader_->ReadUnsignedLEB128(start, &len);
+      const uint8_t* addr_ptr =
+          addr_buffer_ + addr_base_ + addr_index * reader_->AddressSize();
+      ProcessAttributeUnsigned(dieoffset, attr, form,
+                               reader_->ReadAddress(addr_ptr));
+      return start + len;
+    }
   }
   fprintf(stderr, "Unhandled form type\n");
   return NULL;
@@ -458,6 +531,16 @@ const uint8_t *CompilationUnit::ProcessDIE(uint64 dieoffset,
        i++)  {
     start = ProcessAttribute(dieoffset, start, i->first, i->second);
   }
+
+  // If this is a compilation unit in a split DWARF object, verify that
+  // the dwo_id matches. If it does not match, we will ignore this
+  // compilation unit.
+  if (abbrev.tag == DW_TAG_compile_unit
+      && is_split_dwarf_
+      && dwo_id_ != skeleton_dwo_id_) {
+    return NULL;
+  }
+
   return start;
 }
 
@@ -515,6 +598,307 @@ void CompilationUnit::ProcessDIEs() {
   }
 }
 
+// Check for a valid ELF file and return the Address size.
+// Returns 0 if not a valid ELF file.
+inline int GetElfWidth(const ElfReader& elf) {
+  if (elf.IsElf32File())
+    return 4;
+  if (elf.IsElf64File())
+    return 8;
+  return 0;
+}
+
+void CompilationUnit::ProcessSplitDwarf() {
+  struct stat statbuf;
+  if (!have_checked_for_dwp_) {
+    // Look for a .dwp file in the same directory as the executable.
+    have_checked_for_dwp_ = true;
+    string dwp_suffix(".dwp");
+    dwp_path_ = path_ + dwp_suffix;
+    if (stat(dwp_path_.c_str(), &statbuf) != 0) {
+      // Fall back to a split .debug file in the same directory.
+      string debug_suffix(".debug");
+      dwp_path_ = path_;
+      size_t found = path_.rfind(debug_suffix);
+      if (found + debug_suffix.length() == path_.length())
+        dwp_path_ = dwp_path_.replace(found, debug_suffix.length(), dwp_suffix);
+    }
+    if (stat(dwp_path_.c_str(), &statbuf) == 0) {
+      ElfReader* elf = new ElfReader(dwp_path_);
+      int width = GetElfWidth(*elf);
+      if (width != 0) {
+        dwp_byte_reader_.reset(new ByteReader(reader_->GetEndianness()));
+        dwp_byte_reader_->SetAddressSize(width);
+        dwp_reader_.reset(new DwpReader(*dwp_byte_reader_, elf));
+        dwp_reader_->Initialize();
+      } else {
+        delete elf;
+      }
+    }
+  }
+  bool found_in_dwp = false;
+  if (dwp_reader_ != NULL) {
+    // If we have a .dwp file, read the debug sections for the requested CU.
+    SectionMap sections;
+    dwp_reader_->ReadDebugSectionsForCU(dwo_id_, &sections);
+    if (!sections.empty()) {
+      found_in_dwp = true;
+      CompilationUnit dwp_comp_unit(dwp_path_, sections, 0,
+                                    dwp_byte_reader_.get(), handler_);
+      dwp_comp_unit.SetSplitDwarf(addr_buffer_, addr_buffer_length_, addr_base_,
+                                  ranges_base_, dwo_id_);
+      dwp_comp_unit.Start();
+    }
+  }
+  if (!found_in_dwp) {
+    // If no .dwp file, try to open the .dwo file.
+    if (stat(dwo_name_, &statbuf) == 0) {
+      ElfReader elf(dwo_name_);
+      int width = GetElfWidth(elf);
+      if (width != 0) {
+        ByteReader reader(ENDIANNESS_LITTLE);
+        reader.SetAddressSize(width);
+        SectionMap sections;
+        ReadDebugSectionsFromDwo(&elf, &sections);
+        CompilationUnit dwo_comp_unit(dwo_name_, sections, 0, &reader,
+                                      handler_);
+        dwo_comp_unit.SetSplitDwarf(addr_buffer_, addr_buffer_length_,
+                                    addr_base_, ranges_base_, dwo_id_);
+        dwo_comp_unit.Start();
+      }
+    }
+  }
+}
+
+void CompilationUnit::ReadDebugSectionsFromDwo(ElfReader* elf_reader,
+                                               SectionMap* sections) {
+  static const char* const section_names[] = {
+    ".debug_abbrev",
+    ".debug_info",
+    ".debug_str_offsets",
+    ".debug_str"
+  };
+  for (unsigned int i = 0u;
+       i < sizeof(section_names)/sizeof(*(section_names)); ++i) {
+    string base_name = section_names[i];
+    string dwo_name = base_name + ".dwo";
+    size_t section_size;
+    const char* section_data = elf_reader->GetSectionByName(dwo_name,
+                                                            &section_size);
+    if (section_data != NULL)
+      sections->insert(std::make_pair(
+          base_name, std::make_pair(
+             reinterpret_cast<const uint8_t *>(section_data),
+             section_size)));
+  }
+}
+
+DwpReader::DwpReader(const ByteReader& byte_reader, ElfReader* elf_reader)
+    : elf_reader_(elf_reader), byte_reader_(byte_reader),
+      cu_index_(NULL), cu_index_size_(0), string_buffer_(NULL),
+      string_buffer_size_(0), version_(0), ncolumns_(0), nunits_(0),
+      nslots_(0), phash_(NULL), pindex_(NULL), shndx_pool_(NULL),
+      offset_table_(NULL), size_table_(NULL), abbrev_data_(NULL),
+      abbrev_size_(0), info_data_(NULL), info_size_(0),
+      str_offsets_data_(NULL), str_offsets_size_(0) {}
+
+DwpReader::~DwpReader() {
+  if (elf_reader_) delete elf_reader_;
+}
+
+void DwpReader::Initialize() {
+  cu_index_ = elf_reader_->GetSectionByName(".debug_cu_index",
+                                            &cu_index_size_);
+  if (cu_index_ == NULL) {
+    return;
+  }
+  // The .debug_str.dwo section is shared by all CUs in the file.
+  string_buffer_ = elf_reader_->GetSectionByName(".debug_str.dwo",
+                                                 &string_buffer_size_);
+
+  version_ = byte_reader_.ReadFourBytes(
+      reinterpret_cast<const uint8_t *>(cu_index_));
+
+  if (version_ == 1) {
+    nslots_ = byte_reader_.ReadFourBytes(
+        reinterpret_cast<const uint8_t *>(cu_index_)
+        + 3 * sizeof(uint32));
+    phash_ = cu_index_ + 4 * sizeof(uint32);
+    pindex_ = phash_ + nslots_ * sizeof(uint64);
+    shndx_pool_ = pindex_ + nslots_ * sizeof(uint32);
+    if (shndx_pool_ >= cu_index_ + cu_index_size_) {
+      version_ = 0;
+    }
+  } else if (version_ == 2) {
+    ncolumns_ = byte_reader_.ReadFourBytes(
+        reinterpret_cast<const uint8_t *>(cu_index_) + sizeof(uint32));
+    nunits_ = byte_reader_.ReadFourBytes(
+        reinterpret_cast<const uint8_t *>(cu_index_) + 2 * sizeof(uint32));
+    nslots_ = byte_reader_.ReadFourBytes(
+        reinterpret_cast<const uint8_t *>(cu_index_) + 3 * sizeof(uint32));
+    phash_ = cu_index_ + 4 * sizeof(uint32);
+    pindex_ = phash_ + nslots_ * sizeof(uint64);
+    offset_table_ = pindex_ + nslots_ * sizeof(uint32);
+    size_table_ = offset_table_ + ncolumns_ * (nunits_ + 1) * sizeof(uint32);
+    abbrev_data_ = elf_reader_->GetSectionByName(".debug_abbrev.dwo",
+                                                 &abbrev_size_);
+    info_data_ = elf_reader_->GetSectionByName(".debug_info.dwo", &info_size_);
+    str_offsets_data_ = elf_reader_->GetSectionByName(".debug_str_offsets.dwo",
+                                                      &str_offsets_size_);
+    if (size_table_ >= cu_index_ + cu_index_size_) {
+      version_ = 0;
+    }
+  }
+}
+
+void DwpReader::ReadDebugSectionsForCU(uint64 dwo_id,
+                                       SectionMap* sections) {
+  if (version_ == 1) {
+    int slot = LookupCU(dwo_id);
+    if (slot == -1) {
+      return;
+    }
+
+    // The index table points to the section index pool, where we
+    // can read a list of section indexes for the debug sections
+    // for the CU whose dwo_id we are looking for.
+    int index = byte_reader_.ReadFourBytes(
+        reinterpret_cast<const uint8_t *>(pindex_)
+        + slot * sizeof(uint32));
+    const char* shndx_list = shndx_pool_ + index * sizeof(uint32);
+    for (;;) {
+      if (shndx_list >= cu_index_ + cu_index_size_) {
+        version_ = 0;
+        return;
+      }
+      unsigned int shndx = byte_reader_.ReadFourBytes(
+          reinterpret_cast<const uint8_t *>(shndx_list));
+      shndx_list += sizeof(uint32);
+      if (shndx == 0)
+        break;
+      const char* section_name = elf_reader_->GetSectionName(shndx);
+      size_t section_size;
+      const char* section_data;
+      // We're only interested in these four debug sections.
+      // The section names in the .dwo file end with ".dwo", but we
+      // add them to the sections table with their normal names.
+      if (!strncmp(section_name, ".debug_abbrev", strlen(".debug_abbrev"))) {
+        section_data = elf_reader_->GetSectionByIndex(shndx, &section_size);
+        sections->insert(std::make_pair(
+            ".debug_abbrev",
+            std::make_pair(reinterpret_cast<const uint8_t *> (section_data),
+                                                              section_size)));
+      } else if (!strncmp(section_name, ".debug_info", strlen(".debug_info"))) {
+        section_data = elf_reader_->GetSectionByIndex(shndx, &section_size);
+        sections->insert(std::make_pair(
+            ".debug_info",
+            std::make_pair(reinterpret_cast<const uint8_t *> (section_data),
+                           section_size)));
+      } else if (!strncmp(section_name, ".debug_str_offsets",
+                          strlen(".debug_str_offsets"))) {
+        section_data = elf_reader_->GetSectionByIndex(shndx, &section_size);
+        sections->insert(std::make_pair(
+            ".debug_str_offsets",
+            std::make_pair(reinterpret_cast<const uint8_t *> (section_data),
+                           section_size)));
+      }
+    }
+    sections->insert(std::make_pair(
+        ".debug_str",
+        std::make_pair(reinterpret_cast<const uint8_t *> (string_buffer_),
+                       string_buffer_size_)));
+  } else if (version_ == 2) {
+    uint32 index = LookupCUv2(dwo_id);
+    if (index == 0) {
+      return;
+    }
+
+    // The index points to a row in each of the section offsets table
+    // and the section size table, where we can read the offsets and sizes
+    // of the contributions to each debug section from the CU whose dwo_id
+    // we are looking for. Row 0 of the section offsets table has the
+    // section ids for each column of the table. The size table begins
+    // with row 1.
+    const char* id_row = offset_table_;
+    const char* offset_row = offset_table_
+                             + index * ncolumns_ * sizeof(uint32);
+    const char* size_row =
+        size_table_ + (index - 1) * ncolumns_ * sizeof(uint32);
+    if (size_row + ncolumns_ * sizeof(uint32) > cu_index_ + cu_index_size_) {
+      version_ = 0;
+      return;
+    }
+    for (unsigned int col = 0u; col < ncolumns_; ++col) {
+      uint32 section_id =
+          byte_reader_.ReadFourBytes(reinterpret_cast<const uint8_t *>(id_row)
+                                     + col * sizeof(uint32));
+      uint32 offset = byte_reader_.ReadFourBytes(
+          reinterpret_cast<const uint8_t *>(offset_row)
+          + col * sizeof(uint32));
+      uint32 size = byte_reader_.ReadFourBytes(
+          reinterpret_cast<const uint8_t *>(size_row) + col * sizeof(uint32));
+      if (section_id == DW_SECT_ABBREV) {
+        sections->insert(std::make_pair(
+            ".debug_abbrev",
+            std::make_pair(reinterpret_cast<const uint8_t *> (abbrev_data_)
+                           + offset, size)));
+      } else if (section_id == DW_SECT_INFO) {
+        sections->insert(std::make_pair(
+            ".debug_info",
+            std::make_pair(reinterpret_cast<const uint8_t *> (info_data_)
+                           + offset, size)));
+      } else if (section_id == DW_SECT_STR_OFFSETS) {
+        sections->insert(std::make_pair(
+            ".debug_str_offsets",
+            std::make_pair(reinterpret_cast<const uint8_t *> (str_offsets_data_)
+                           + offset, size)));
+      }
+    }
+    sections->insert(std::make_pair(
+        ".debug_str",
+        std::make_pair(reinterpret_cast<const uint8_t *> (string_buffer_),
+                       string_buffer_size_)));
+  }
+}
+
+int DwpReader::LookupCU(uint64 dwo_id) {
+  uint32 slot = static_cast<uint32>(dwo_id) & (nslots_ - 1);
+  uint64 probe = byte_reader_.ReadEightBytes(
+      reinterpret_cast<const uint8_t *>(phash_) + slot * sizeof(uint64));
+  if (probe != 0 && probe != dwo_id) {
+    uint32 secondary_hash =
+        (static_cast<uint32>(dwo_id >> 32) & (nslots_ - 1)) | 1;
+    do {
+      slot = (slot + secondary_hash) & (nslots_ - 1);
+      probe = byte_reader_.ReadEightBytes(
+          reinterpret_cast<const uint8_t *>(phash_) + slot * sizeof(uint64));
+    } while (probe != 0 && probe != dwo_id);
+  }
+  if (probe == 0)
+    return -1;
+  return slot;
+}
+
+uint32 DwpReader::LookupCUv2(uint64 dwo_id) {
+  uint32 slot = static_cast<uint32>(dwo_id) & (nslots_ - 1);
+  uint64 probe = byte_reader_.ReadEightBytes(
+      reinterpret_cast<const uint8_t *>(phash_) + slot * sizeof(uint64));
+  uint32 index = byte_reader_.ReadFourBytes(
+      reinterpret_cast<const uint8_t *>(pindex_) + slot * sizeof(uint32));
+  if (index != 0 && probe != dwo_id) {
+    uint32 secondary_hash =
+        (static_cast<uint32>(dwo_id >> 32) & (nslots_ - 1)) | 1;
+    do {
+      slot = (slot + secondary_hash) & (nslots_ - 1);
+      probe = byte_reader_.ReadEightBytes(
+          reinterpret_cast<const uint8_t *>(phash_) + slot * sizeof(uint64));
+      index = byte_reader_.ReadFourBytes(
+          reinterpret_cast<const uint8_t *>(pindex_) + slot * sizeof(uint32));
+    } while (index != 0 && probe != dwo_id);
+  }
+  return index;
+}
+
 LineInfo::LineInfo(const uint8_t *buffer, uint64 buffer_length,
                    ByteReader* reader, LineInfoHandler* handler):
     handler_(handler), reader_(reader), buffer_(buffer) {
diff --git a/src/common/dwarf/dwarf2reader.h b/src/common/dwarf/dwarf2reader.h
index 1f5b37a0..064c42bc 100644
--- a/src/common/dwarf/dwarf2reader.h
+++ b/src/common/dwarf/dwarf2reader.h
@@ -47,16 +47,19 @@
 #include <string>
 #include <utility>
 #include <vector>
+#include <memory>
 
 #include "common/dwarf/bytereader.h"
 #include "common/dwarf/dwarf2enums.h"
 #include "common/dwarf/types.h"
 #include "common/using_std_string.h"
+#include "common/dwarf/elf_reader.h"
 
 namespace dwarf2reader {
 struct LineStateMachine;
 class Dwarf2Handler;
 class LineInfoHandler;
+class DwpReader;
 
 // This maps from a string naming a section to a pair containing a
 // the data for the section, and the size of the section.
@@ -184,6 +187,106 @@ class LineInfoHandler {
                        uint32 file_num, uint32 line_num, uint32 column_num) { }
 };
 
+// This class is the main interface between the reader and the
+// client.  The virtual functions inside this get called for
+// interesting events that happen during DWARF2 reading.
+// The default implementation skips everything.
+class Dwarf2Handler {
+ public:
+  Dwarf2Handler() { }
+
+  virtual ~Dwarf2Handler() { }
+
+  // Start to process a compilation unit at OFFSET from the beginning of the
+  // .debug_info section. Return false if you would like to skip this
+  // compilation unit.
+  virtual bool StartCompilationUnit(uint64 offset, uint8 address_size,
+                                    uint8 offset_size, uint64 cu_length,
+                                    uint8 dwarf_version) { return false; }
+
+  // When processing a skeleton compilation unit, resulting from a split
+  // DWARF compilation, once the skeleton debug info has been read,
+  // the reader will call this function to ask the client if it needs
+  // the full debug info from the .dwo or .dwp file.  Return true if
+  // you need it, or false to skip processing the split debug info.
+  virtual bool NeedSplitDebugInfo() { return true; }
+
+  // Start to process a split compilation unit at OFFSET from the beginning of
+  // the debug_info section in the .dwp/.dwo file.  Return false if you would
+  // like to skip this compilation unit.
+  virtual bool StartSplitCompilationUnit(uint64 offset,
+                                         uint64 cu_length) { return false; }
+
+  // Start to process a DIE at OFFSET from the beginning of the .debug_info
+  // section. Return false if you would like to skip this DIE.
+  virtual bool StartDIE(uint64 offset, enum DwarfTag tag) { return false; }
+
+  // Called when we have an attribute with unsigned data to give to our
+  // handler. The attribute is for the DIE at OFFSET from the beginning of the
+  // .debug_info section. Its name is ATTR, its form is FORM, and its value is
+  // DATA.
+  virtual void ProcessAttributeUnsigned(uint64 offset,
+                                        enum DwarfAttribute attr,
+                                        enum DwarfForm form,
+                                        uint64 data) { }
+
+  // Called when we have an attribute with signed data to give to our handler.
+  // The attribute is for the DIE at OFFSET from the beginning of the
+  // .debug_info section. Its name is ATTR, its form is FORM, and its value is
+  // DATA.
+  virtual void ProcessAttributeSigned(uint64 offset,
+                                      enum DwarfAttribute attr,
+                                      enum DwarfForm form,
+                                      int64 data) { }
+
+  // Called when we have an attribute whose value is a reference to
+  // another DIE. The attribute belongs to the DIE at OFFSET from the
+  // beginning of the .debug_info section. Its name is ATTR, its form
+  // is FORM, and the offset of the DIE being referred to from the
+  // beginning of the .debug_info section is DATA.
+  virtual void ProcessAttributeReference(uint64 offset,
+                                         enum DwarfAttribute attr,
+                                         enum DwarfForm form,
+                                         uint64 data) { }
+
+  // Called when we have an attribute with a buffer of data to give to our
+  // handler. The attribute is for the DIE at OFFSET from the beginning of the
+  // .debug_info section. Its name is ATTR, its form is FORM, DATA points to
+  // the buffer's contents, and its length in bytes is LENGTH. The buffer is
+  // owned by the caller, not the callee, and may not persist for very long.
+  // If you want the data to be available later, it needs to be copied.
+  virtual void ProcessAttributeBuffer(uint64 offset,
+                                      enum DwarfAttribute attr,
+                                      enum DwarfForm form,
+                                      const uint8_t *data,
+                                      uint64 len) { }
+
+  // Called when we have an attribute with string data to give to our handler.
+  // The attribute is for the DIE at OFFSET from the beginning of the
+  // .debug_info section. Its name is ATTR, its form is FORM, and its value is
+  // DATA.
+  virtual void ProcessAttributeString(uint64 offset,
+                                      enum DwarfAttribute attr,
+                                      enum DwarfForm form,
+                                      const string& data) { }
+
+  // Called when we have an attribute whose value is the 64-bit signature
+  // of a type unit in the .debug_types section. OFFSET is the offset of
+  // the DIE whose attribute we're reporting. ATTR and FORM are the
+  // attribute's name and form. SIGNATURE is the type unit's signature.
+  virtual void ProcessAttributeSignature(uint64 offset,
+                                         enum DwarfAttribute attr,
+                                         enum DwarfForm form,
+                                         uint64 signature) { }
+
+  // Called when finished processing the DIE at OFFSET.
+  // Because DWARF2/3 specifies a tree of DIEs, you may get starts
+  // before ends of the previous DIE, as we process children before
+  // ending the parent.
+  virtual void EndDIE(uint64 offset) { }
+
+};
+
 // The base of DWARF2/3 debug info is a DIE (Debugging Information
 // Entry.
 // DWARF groups DIE's into a tree and calls the root of this tree a
@@ -225,12 +328,21 @@ class CompilationUnit {
   // Initialize a compilation unit.  This requires a map of sections,
   // the offset of this compilation unit in the .debug_info section, a
   // ByteReader, and a Dwarf2Handler class to call callbacks in.
-  CompilationUnit(const SectionMap& sections, uint64 offset,
+  CompilationUnit(const string& path, const SectionMap& sections, uint64 offset,
                   ByteReader* reader, Dwarf2Handler* handler);
   virtual ~CompilationUnit() {
     if (abbrevs_) delete abbrevs_;
   }
 
+  // Initialize a compilation unit from a .dwo or .dwp file.
+  // In this case, we need the .debug_addr section from the
+  // executable file that contains the corresponding skeleton
+  // compilation unit.  We also inherit the Dwarf2Handler from
+  // the executable file, and call it as if we were still
+  // processing the original compilation unit.
+  void SetSplitDwarf(const uint8_t* addr_buffer, uint64 addr_buffer_length,
+                     uint64 addr_base, uint64 ranges_base, uint64 dwo_id);
+
   // Begin reading a Dwarf2 compilation unit, and calling the
   // callbacks in the Dwarf2Handler
 
@@ -281,6 +393,73 @@ class CompilationUnit {
                                   enum DwarfAttribute attr,
                                   enum DwarfForm form);
 
+  // Called when we have an attribute with unsigned data to give to
+  // our handler.  The attribute is for the DIE at OFFSET from the
+  // beginning of compilation unit, has a name of ATTR, a form of
+  // FORM, and the actual data of the attribute is in DATA.
+  // If we see a DW_AT_GNU_dwo_id attribute, save the value so that
+  // we can find the debug info in a .dwo or .dwp file.
+  void ProcessAttributeUnsigned(uint64 offset,
+                                enum DwarfAttribute attr,
+                                enum DwarfForm form,
+                                uint64 data) {
+    if (attr == DW_AT_GNU_dwo_id) {
+      dwo_id_ = data;
+    }
+    else if (attr == DW_AT_GNU_addr_base) {
+      addr_base_ = data;
+    }
+    else if (attr == DW_AT_GNU_ranges_base) {
+      ranges_base_ = data;
+    }
+    // TODO(yunlian): When we add DW_AT_ranges_base from DWARF-5,
+    // that base will apply to DW_AT_ranges attributes in the
+    // skeleton CU as well as in the .dwo/.dwp files.
+    else if (attr == DW_AT_ranges && is_split_dwarf_) {
+      data += ranges_base_;
+    }
+    handler_->ProcessAttributeUnsigned(offset, attr, form, data);
+  }
+
+  // Called when we have an attribute with signed data to give to
+  // our handler.  The attribute is for the DIE at OFFSET from the
+  // beginning of compilation unit, has a name of ATTR, a form of
+  // FORM, and the actual data of the attribute is in DATA.
+  void ProcessAttributeSigned(uint64 offset,
+                              enum DwarfAttribute attr,
+                              enum DwarfForm form,
+                              int64 data) {
+    handler_->ProcessAttributeSigned(offset, attr, form, data);
+  }
+
+  // Called when we have an attribute with a buffer of data to give to
+  // our handler.  The attribute is for the DIE at OFFSET from the
+  // beginning of compilation unit, has a name of ATTR, a form of
+  // FORM, and the actual data of the attribute is in DATA, and the
+  // length of the buffer is LENGTH.
+  void ProcessAttributeBuffer(uint64 offset,
+                              enum DwarfAttribute attr,
+                              enum DwarfForm form,
+                              const uint8_t* data,
+                              uint64 len) {
+    handler_->ProcessAttributeBuffer(offset, attr, form, data, len);
+  }
+
+  // Called when we have an attribute with string data to give to
+  // our handler.  The attribute is for the DIE at OFFSET from the
+  // beginning of compilation unit, has a name of ATTR, a form of
+  // FORM, and the actual data of the attribute is in DATA.
+  // If we see a DW_AT_GNU_dwo_name attribute, save the value so
+  // that we can find the debug info in a .dwo or .dwp file.
+  void ProcessAttributeString(uint64 offset,
+                              enum DwarfAttribute attr,
+                              enum DwarfForm form,
+                              const char* data) {
+    if (attr == DW_AT_GNU_dwo_name)
+      dwo_name_ = data;
+    handler_->ProcessAttributeString(offset, attr, form, data);
+  }
+
   // Processes all DIEs for this compilation unit
   void ProcessDIEs();
 
@@ -292,6 +471,16 @@ class CompilationUnit {
   // new place to position the stream to.
   const uint8_t *SkipAttribute(const uint8_t *start, enum DwarfForm form);
 
+  // Process the actual debug information in a split DWARF file.
+  void ProcessSplitDwarf();
+
+  // Read the debug sections from a .dwo file.
+  void ReadDebugSectionsFromDwo(ElfReader* elf_reader,
+                                SectionMap* sections);
+
+  // Path of the file containing the debug information.
+  const string path_;
+
   // Offset from section start is the offset of this compilation unit
   // from the beginning of the .debug_info section.
   uint64 offset_from_section_start_;
@@ -322,94 +511,141 @@ class CompilationUnit {
   // ProcessAttribute, which is in the hot path for DWARF2 reading.
   const uint8_t *string_buffer_;
   uint64 string_buffer_length_;
-};
 
-// This class is the main interface between the reader and the
-// client.  The virtual functions inside this get called for
-// interesting events that happen during DWARF2 reading.
-// The default implementation skips everything.
+  // String offsets section buffer and length, if we have a string offsets
+  // section (.debug_str_offsets or .debug_str_offsets.dwo).
+  const uint8_t* str_offsets_buffer_;
+  uint64 str_offsets_buffer_length_;
 
-class Dwarf2Handler {
+  // Address section buffer and length, if we have an address section
+  // (.debug_addr).
+  const uint8_t* addr_buffer_;
+  uint64 addr_buffer_length_;
+
+  // Flag indicating whether this compilation unit is part of a .dwo
+  // or .dwp file.  If true, we are reading this unit because a
+  // skeleton compilation unit in an executable file had a
+  // DW_AT_GNU_dwo_name or DW_AT_GNU_dwo_id attribute.
+  // In a .dwo file, we expect the string offsets section to
+  // have a ".dwo" suffix, and we will use the ".debug_addr" section
+  // associated with the skeleton compilation unit.
+  bool is_split_dwarf_;
+
+  // The value of the DW_AT_GNU_dwo_id attribute, if any.
+  uint64 dwo_id_;
+
+  // The value of the DW_AT_GNU_dwo_name attribute, if any.
+  const char* dwo_name_;
+
+  // If this is a split DWARF CU, the value of the DW_AT_GNU_dwo_id attribute
+  // from the skeleton CU.
+  uint64 skeleton_dwo_id_;
+
+  // The value of the DW_AT_GNU_ranges_base attribute, if any.
+  uint64 ranges_base_;
+
+  // The value of the DW_AT_GNU_addr_base attribute, if any.
+  uint64 addr_base_;
+
+  // True if we have already looked for a .dwp file.
+  bool have_checked_for_dwp_;
+
+  // Path to the .dwp file.
+  string dwp_path_;
+
+  // ByteReader for the DWP file.
+  std::unique_ptr<ByteReader> dwp_byte_reader_;
+
+  // DWP reader.
+   std::unique_ptr<DwpReader> dwp_reader_;
+};
+
+// A Reader for a .dwp file.  Supports the fetching of DWARF debug
+// info for a given dwo_id.
+//
+// There are two versions of .dwp files.  In both versions, the
+// .dwp file is an ELF file containing only debug sections.
+// In Version 1, the file contains many copies of each debug
+// section, one for each .dwo file that is packaged in the .dwp
+// file, and the .debug_cu_index section maps from the dwo_id
+// to a set of section indexes.  In Version 2, the file contains
+// one of each debug section, and the .debug_cu_index section
+// maps from the dwo_id to a set of offsets and lengths that
+// identify each .dwo file's contribution to the larger sections.
+
+class DwpReader {
  public:
-  Dwarf2Handler() { }
+  DwpReader(const ByteReader& byte_reader, ElfReader* elf_reader);
 
-  virtual ~Dwarf2Handler() { }
+  ~DwpReader();
 
-  // Start to process a compilation unit at OFFSET from the beginning of the
-  // .debug_info section. Return false if you would like to skip this
-  // compilation unit.
-  virtual bool StartCompilationUnit(uint64 offset, uint8 address_size,
-                                    uint8 offset_size, uint64 cu_length,
-                                    uint8 dwarf_version) { return false; }
+  // Read the CU index and initialize data members.
+  void Initialize();
 
-  // Start to process a DIE at OFFSET from the beginning of the .debug_info
-  // section. Return false if you would like to skip this DIE.
-  virtual bool StartDIE(uint64 offset, enum DwarfTag tag) { return false; }
+  // Read the debug sections for the given dwo_id.
+  void ReadDebugSectionsForCU(uint64 dwo_id, SectionMap* sections);
 
-  // Called when we have an attribute with unsigned data to give to our
-  // handler. The attribute is for the DIE at OFFSET from the beginning of the
-  // .debug_info section. Its name is ATTR, its form is FORM, and its value is
-  // DATA.
-  virtual void ProcessAttributeUnsigned(uint64 offset,
-                                        enum DwarfAttribute attr,
-                                        enum DwarfForm form,
-                                        uint64 data) { }
+ private:
+  // Search a v1 hash table for "dwo_id".  Returns the slot index
+  // where the dwo_id was found, or -1 if it was not found.
+  int LookupCU(uint64 dwo_id);
 
-  // Called when we have an attribute with signed data to give to our handler.
-  // The attribute is for the DIE at OFFSET from the beginning of the
-  // .debug_info section. Its name is ATTR, its form is FORM, and its value is
-  // DATA.
-  virtual void ProcessAttributeSigned(uint64 offset,
-                                      enum DwarfAttribute attr,
-                                      enum DwarfForm form,
-                                      int64 data) { }
+  // Search a v2 hash table for "dwo_id".  Returns the row index
+  // in the offsets and sizes tables, or 0 if it was not found.
+  uint32 LookupCUv2(uint64 dwo_id);
 
-  // Called when we have an attribute whose value is a reference to
-  // another DIE. The attribute belongs to the DIE at OFFSET from the
-  // beginning of the .debug_info section. Its name is ATTR, its form
-  // is FORM, and the offset of the DIE being referred to from the
-  // beginning of the .debug_info section is DATA.
-  virtual void ProcessAttributeReference(uint64 offset,
-                                         enum DwarfAttribute attr,
-                                         enum DwarfForm form,
-                                         uint64 data) { }
+  // The ELF reader for the .dwp file.
+  ElfReader* elf_reader_;
 
-  // Called when we have an attribute with a buffer of data to give to our
-  // handler. The attribute is for the DIE at OFFSET from the beginning of the
-  // .debug_info section. Its name is ATTR, its form is FORM, DATA points to
-  // the buffer's contents, and its length in bytes is LENGTH. The buffer is
-  // owned by the caller, not the callee, and may not persist for very long.
-  // If you want the data to be available later, it needs to be copied.
-  virtual void ProcessAttributeBuffer(uint64 offset,
-                                      enum DwarfAttribute attr,
-                                      enum DwarfForm form,
-                                      const uint8_t *data,
-                                      uint64 len) { }
+  // The ByteReader for the .dwp file.
+  const ByteReader& byte_reader_;
 
-  // Called when we have an attribute with string data to give to our handler.
-  // The attribute is for the DIE at OFFSET from the beginning of the
-  // .debug_info section. Its name is ATTR, its form is FORM, and its value is
-  // DATA.
-  virtual void ProcessAttributeString(uint64 offset,
-                                      enum DwarfAttribute attr,
-                                      enum DwarfForm form,
-                                      const string& data) { }
+  // Pointer to the .debug_cu_index section.
+  const char* cu_index_;
 
-  // Called when we have an attribute whose value is the 64-bit signature
-  // of a type unit in the .debug_types section. OFFSET is the offset of
-  // the DIE whose attribute we're reporting. ATTR and FORM are the
-  // attribute's name and form. SIGNATURE is the type unit's signature.
-  virtual void ProcessAttributeSignature(uint64 offset,
-                                         enum DwarfAttribute attr,
-                                         enum DwarfForm form,
-                                         uint64 signature) { }
+  // Size of the .debug_cu_index section.
+  size_t cu_index_size_;
 
-  // Called when finished processing the DIE at OFFSET.
-  // Because DWARF2/3 specifies a tree of DIEs, you may get starts
-  // before ends of the previous DIE, as we process children before
-  // ending the parent.
-  virtual void EndDIE(uint64 offset) { }
+  // Pointer to the .debug_str.dwo section.
+  const char* string_buffer_;
+
+  // Size of the .debug_str.dwo section.
+  size_t string_buffer_size_;
+
+  // Version of the .dwp file.  We support versions 1 and 2 currently.
+  int version_;
+
+  // Number of columns in the section tables (version 2).
+  unsigned int ncolumns_;
+
+  // Number of units in the section tables (version 2).
+  unsigned int nunits_;
+
+  // Number of slots in the hash table.
+  unsigned int nslots_;
+
+  // Pointer to the beginning of the hash table.
+  const char* phash_;
+
+  // Pointer to the beginning of the index table.
+  const char* pindex_;
+
+  // Pointer to the beginning of the section index pool (version 1).
+  const char* shndx_pool_;
+
+  // Pointer to the beginning of the section offset table (version 2).
+  const char* offset_table_;
+
+  // Pointer to the beginning of the section size table (version 2).
+  const char* size_table_;
 
+  // Contents of the sections of interest (version 2).
+  const char* abbrev_data_;
+  size_t abbrev_size_;
+  const char* info_data_;
+  size_t info_size_;
+  const char* str_offsets_data_;
+  size_t str_offsets_size_;
 };
 
 // This class is a reader for DWARF's Call Frame Information.  CFI
diff --git a/src/common/dwarf/elf_reader.cc b/src/common/dwarf/elf_reader.cc
new file mode 100644
index 00000000..81683141
--- /dev/null
+++ b/src/common/dwarf/elf_reader.cc
@@ -0,0 +1,1258 @@
+// Copyright 2005 Google Inc. All Rights Reserved.
+// Author: chatham@google.com (Andrew Chatham)
+// Author: satorux@google.com (Satoru Takabayashi)
+//
+// Code for reading in ELF files.
+//
+// For information on the ELF format, see
+// http://www.x86.org/ftp/manuals/tools/elf.pdf
+//
+// I also liked:
+// http://www.caldera.com/developers/gabi/1998-04-29/contents.html
+//
+// A note about types: When dealing with the file format, we use types
+// like Elf32_Word, but in the public interfaces we treat all
+// addresses as uint64. As a result, we should be able to symbolize
+// 64-bit binaries from a 32-bit process (which we don't do,
+// anyway). size_t should therefore be avoided, except where required
+// by things like mmap().
+//
+// Although most of this code can deal with arbitrary ELF files of
+// either word size, the public ElfReader interface only examines
+// files loaded into the current address space, which must all match
+// __WORDSIZE. This code cannot handle ELF files with a non-native
+// byte ordering.
+//
+// TODO(chatham): It would be nice if we could accomplish this task
+// without using malloc(), so we could use it as the process is dying.
+
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE  // needed for pread()
+#endif
+
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <elf.h>
+#include <string.h>
+
+#include <algorithm>
+#include <map>
+#include <string>
+#include <vector>
+#include "zlib.h"
+
+#include "elf_reader.h"
+//#include "using_std_string.h"
+// EM_AARCH64 is not defined by elf.h of GRTE v3 on x86.
+// TODO(dougkwan): Remove this when v17 is retired.
+#if !defined(EM_AARCH64)
+#define EM_AARCH64      183             /* ARM AARCH64 */
+#endif
+
+// TODO(dthomson): Can be removed once all Java code is using the Google3
+// launcher. We need to avoid processing PLT functions as it causes memory
+// fragmentation in malloc, which is fixed in tcmalloc - and if the Google3
+// launcher is used the JVM will then use tcmalloc. b/13735638
+//DEFINE_bool(elfreader_process_dynsyms, true,
+//            "Activate PLT function processing");
+
+using std::string;
+using std::vector;
+
+namespace {
+
+// The lowest bit of an ARM symbol value is used to indicate a Thumb address.
+const int kARMThumbBitOffset = 0;
+
+// Converts an ARM Thumb symbol value to a true aligned address value.
+template <typename T>
+T AdjustARMThumbSymbolValue(const T& symbol_table_value) {
+  return symbol_table_value & ~(1 << kARMThumbBitOffset);
+}
+
+// Names of PLT-related sections.
+const char kElfPLTRelSectionName[] = ".rel.plt";      // Use Rel struct.
+const char kElfPLTRelaSectionName[] = ".rela.plt";    // Use Rela struct.
+const char kElfPLTSectionName[] = ".plt";
+const char kElfDynSymSectionName[] = ".dynsym";
+
+const int kX86PLTCodeSize = 0x10;  // Size of one x86 PLT function in bytes.
+const int kARMPLTCodeSize = 0xc;
+const int kAARCH64PLTCodeSize = 0x10;
+
+const int kX86PLT0Size = 0x10;  // Size of the special PLT0 entry.
+const int kARMPLT0Size = 0x14;
+const int kAARCH64PLT0Size = 0x20;
+
+// Suffix for PLT functions when it needs to be explicitly identified as such.
+const char kPLTFunctionSuffix[] = "@plt";
+
+}  // namespace
+
+namespace dwarf2reader {
+
+template <class ElfArch> class ElfReaderImpl;
+
+// 32-bit and 64-bit ELF files are processed exactly the same, except
+// for various field sizes. Elf32 and Elf64 encompass all of the
+// differences between the two formats, and all format-specific code
+// in this file is templated on one of them.
+class Elf32 {
+ public:
+  typedef Elf32_Ehdr Ehdr;
+  typedef Elf32_Shdr Shdr;
+  typedef Elf32_Phdr Phdr;
+  typedef Elf32_Word Word;
+  typedef Elf32_Sym Sym;
+  typedef Elf32_Rel Rel;
+  typedef Elf32_Rela Rela;
+
+  // What should be in the EI_CLASS header.
+  static const int kElfClass = ELFCLASS32;
+
+  // Given a symbol pointer, return the binding type (eg STB_WEAK).
+  static char Bind(const Elf32_Sym *sym) {
+    return ELF32_ST_BIND(sym->st_info);
+  }
+  // Given a symbol pointer, return the symbol type (eg STT_FUNC).
+  static char Type(const Elf32_Sym *sym) {
+    return ELF32_ST_TYPE(sym->st_info);
+  }
+
+  // Extract the symbol index from the r_info field of a relocation.
+  static int r_sym(const Elf32_Word r_info) {
+    return ELF32_R_SYM(r_info);
+  }
+};
+
+
+class Elf64 {
+ public:
+  typedef Elf64_Ehdr Ehdr;
+  typedef Elf64_Shdr Shdr;
+  typedef Elf64_Phdr Phdr;
+  typedef Elf64_Word Word;
+  typedef Elf64_Sym Sym;
+  typedef Elf64_Rel Rel;
+  typedef Elf64_Rela Rela;
+
+  // What should be in the EI_CLASS header.
+  static const int kElfClass = ELFCLASS64;
+
+  static char Bind(const Elf64_Sym *sym) {
+    return ELF64_ST_BIND(sym->st_info);
+  }
+  static char Type(const Elf64_Sym *sym) {
+    return ELF64_ST_TYPE(sym->st_info);
+  }
+  static int r_sym(const Elf64_Xword r_info) {
+    return ELF64_R_SYM(r_info);
+  }
+};
+
+
+// ElfSectionReader mmaps a section of an ELF file ("section" is ELF
+// terminology). The ElfReaderImpl object providing the section header
+// must exist for the lifetime of this object.
+//
+// The motivation for mmaping individual sections of the file is that
+// many Google executables are large enough when unstripped that we
+// have to worry about running out of virtual address space.
+//
+// For compressed sections we have no choice but to allocate memory.
+template<class ElfArch>
+class ElfSectionReader {
+ public:
+  ElfSectionReader(const char *name, const string &path, int fd,
+                   const typename ElfArch::Shdr &section_header)
+      : contents_aligned_(NULL),
+        contents_(NULL),
+        header_(section_header) {
+    // Back up to the beginning of the page we're interested in.
+    const size_t additional = header_.sh_offset % getpagesize();
+    const size_t offset_aligned = header_.sh_offset - additional;
+    section_size_ = header_.sh_size;
+    size_aligned_ = section_size_ + additional;
+    // If the section has been stripped or is empty, do not attempt
+    // to process its contents.
+    if (header_.sh_type == SHT_NOBITS || header_.sh_size == 0)
+      return;
+    contents_aligned_ = mmap(NULL, size_aligned_, PROT_READ, MAP_SHARED,
+                             fd, offset_aligned);
+    // Set where the offset really should begin.
+    contents_ = reinterpret_cast<char *>(contents_aligned_) +
+                (header_.sh_offset - offset_aligned);
+
+    // Check for and handle any compressed contents.
+    //if (strncmp(name, ".zdebug_", strlen(".zdebug_")) == 0)
+    //  DecompressZlibContents();
+    // TODO(saugustine): Add support for proposed elf-section flag
+    // "SHF_COMPRESS".
+  }
+
+  ~ElfSectionReader() {
+    if (contents_aligned_ != NULL)
+      munmap(contents_aligned_, size_aligned_);
+    else
+      delete[] contents_;
+  }
+
+  // Return the section header for this section.
+  typename ElfArch::Shdr const &header() const { return header_; }
+
+  // Return memory at the given offset within this section.
+  const char *GetOffset(typename ElfArch::Word bytes) const {
+    return contents_ + bytes;
+  }
+
+  const char *contents() const { return contents_; }
+  size_t section_size() const { return section_size_; }
+
+ private:
+  // page-aligned file contents
+  void *contents_aligned_;
+  // contents as usable by the client. For non-compressed sections,
+  // pointer within contents_aligned_ to where the section data
+  // begins; for compressed sections, pointer to the decompressed
+  // data.
+  char *contents_;
+  // size of contents_aligned_
+  size_t size_aligned_;
+  // size of contents.
+  size_t section_size_;
+  const typename ElfArch::Shdr header_;
+};
+
+// An iterator over symbols in a given section. It handles walking
+// through the entries in the specified section and mapping symbol
+// entries to their names in the appropriate string table (in
+// another section).
+template<class ElfArch>
+class SymbolIterator {
+ public:
+  SymbolIterator(ElfReaderImpl<ElfArch> *reader,
+                 typename ElfArch::Word section_type)
+      : symbol_section_(reader->GetSectionByType(section_type)),
+        string_section_(NULL),
+        num_symbols_in_section_(0),
+        symbol_within_section_(0) {
+
+    // If this section type doesn't exist, leave
+    // num_symbols_in_section_ as zero, so this iterator is already
+    // done().
+    if (symbol_section_ != NULL) {
+      num_symbols_in_section_ = symbol_section_->header().sh_size /
+                                symbol_section_->header().sh_entsize;
+
+      // Symbol sections have sh_link set to the section number of
+      // the string section containing the symbol names.
+      string_section_ = reader->GetSection(symbol_section_->header().sh_link);
+    }
+  }
+
+  // Return true iff we have passed all symbols in this section.
+  bool done() const {
+    return symbol_within_section_ >= num_symbols_in_section_;
+  }
+
+  // Advance to the next symbol in this section.
+  // REQUIRES: !done()
+  void Next() { ++symbol_within_section_; }
+
+  // Return a pointer to the current symbol.
+  // REQUIRES: !done()
+  const typename ElfArch::Sym *GetSymbol() const {
+    return reinterpret_cast<const typename ElfArch::Sym*>(
+        symbol_section_->GetOffset(symbol_within_section_ *
+                                   symbol_section_->header().sh_entsize));
+  }
+
+  // Return the name of the current symbol, NULL if it has none.
+  // REQUIRES: !done()
+  const char *GetSymbolName() const {
+    int name_offset = GetSymbol()->st_name;
+    if (name_offset == 0)
+      return NULL;
+    return string_section_->GetOffset(name_offset);
+  }
+
+  int GetCurrentSymbolIndex() const {
+    return symbol_within_section_;
+  }
+
+ private:
+  const ElfSectionReader<ElfArch> *const symbol_section_;
+  const ElfSectionReader<ElfArch> *string_section_;
+  int num_symbols_in_section_;
+  int symbol_within_section_;
+};
+
+
+// Copied from strings/strutil.h.  Per chatham,
+// this library should not depend on strings.
+
+static inline bool MyHasSuffixString(const string& str, const string& suffix) {
+  int len = str.length();
+  int suflen = suffix.length();
+  return (suflen <= len) && (str.compare(len-suflen, suflen, suffix) == 0);
+}
+
+
+// ElfReader loads an ELF binary and can provide information about its
+// contents. It is most useful for matching addresses to function
+// names. It does not understand debugging formats (eg dwarf2), so it
+// can't print line numbers. It takes a path to an elf file and a
+// readable file descriptor for that file, which it does not assume
+// ownership of.
+template<class ElfArch>
+class ElfReaderImpl {
+ public:
+  explicit ElfReaderImpl(const string &path, int fd)
+      : path_(path),
+        fd_(fd),
+        section_headers_(NULL),
+        program_headers_(NULL),
+        opd_section_(NULL),
+        base_for_text_(0),
+        plts_supported_(false),
+        plt_code_size_(0),
+        plt0_size_(0),
+        visited_relocation_entries_(false) {
+    string error;
+    is_dwp_ = MyHasSuffixString(path, ".dwp");
+    ParseHeaders(fd, path);
+    // Currently we need some extra information for PowerPC64 binaries
+    // including a way to read the .opd section for function descriptors and a
+    // way to find the linked base for function symbols.
+    if (header_.e_machine == EM_PPC64) {
+      // "opd_section_" must always be checked for NULL before use.
+      opd_section_ = GetSectionInfoByName(".opd", &opd_info_);
+      for (unsigned int k = 0u; k < GetNumSections(); ++k) {
+        const char *name = GetSectionName(section_headers_[k].sh_name);
+        if (strncmp(name, ".text", strlen(".text")) == 0) {
+          base_for_text_ =
+              section_headers_[k].sh_addr - section_headers_[k].sh_offset;
+          break;
+        }
+      }
+    }
+    // Turn on PLTs.
+    if (header_.e_machine == EM_386 || header_.e_machine == EM_X86_64) {
+      plt_code_size_ = kX86PLTCodeSize;
+      plt0_size_ = kX86PLT0Size;
+      plts_supported_ = true;
+    } else if (header_.e_machine == EM_ARM) {
+      plt_code_size_ = kARMPLTCodeSize;
+      plt0_size_ = kARMPLT0Size;
+      plts_supported_ = true;
+    } else if (header_.e_machine == EM_AARCH64) {
+      plt_code_size_ = kAARCH64PLTCodeSize;
+      plt0_size_ = kAARCH64PLT0Size;
+      plts_supported_ = true;
+    }
+  }
+
+  ~ElfReaderImpl() {
+    for (unsigned int i = 0u; i < sections_.size(); ++i)
+      delete sections_[i];
+    delete [] section_headers_;
+    delete [] program_headers_;
+  }
+
+  // Examine the headers of the file and return whether the file looks
+  // like an ELF file for this architecture. Takes an already-open
+  // file descriptor for the candidate file, reading in the prologue
+  // to see if the ELF file appears to match the current
+  // architecture. If error is non-NULL, it will be set with a reason
+  // in case of failure.
+  static bool IsArchElfFile(int fd, string *error) {
+    unsigned char header[EI_NIDENT];
+    if (pread(fd, header, sizeof(header), 0) != sizeof(header)) {
+      if (error != NULL) *error = "Could not read header";
+      return false;
+    }
+
+    if (memcmp(header, ELFMAG, SELFMAG) != 0) {
+      if (error != NULL) *error = "Missing ELF magic";
+      return false;
+    }
+
+    if (header[EI_CLASS] != ElfArch::kElfClass) {
+      if (error != NULL) *error = "Different word size";
+      return false;
+    }
+
+    int endian = 0;
+    if (header[EI_DATA] == ELFDATA2LSB)
+      endian = __LITTLE_ENDIAN;
+    else if (header[EI_DATA] == ELFDATA2MSB)
+      endian = __BIG_ENDIAN;
+    if (endian != __BYTE_ORDER) {
+      if (error != NULL) *error = "Different byte order";
+      return false;
+    }
+
+    return true;
+  }
+
+  // Return true if we can use this symbol in Address-to-Symbol map.
+  bool CanUseSymbol(const char *name, const typename ElfArch::Sym *sym) {
+    // For now we only save FUNC and NOTYPE symbols. For now we just
+    // care about functions, but some functions written in assembler
+    // don't have a proper ELF type attached to them, so we store
+    // NOTYPE symbols as well. The remaining significant type is
+    // OBJECT (eg global variables), which represent about 25% of
+    // the symbols in a typical google3 binary.
+    if (ElfArch::Type(sym) != STT_FUNC &&
+        ElfArch::Type(sym) != STT_NOTYPE) {
+      return false;
+    }
+
+    // Target specific filtering.
+    switch (header_.e_machine) {
+    case EM_AARCH64:
+    case EM_ARM:
+      // Filter out '$x' special local symbols used by tools
+      return name[0] != '$' || ElfArch::Bind(sym) != STB_LOCAL;
+    case EM_X86_64:
+      // Filter out read-only constants like .LC123.
+      return name[0] != '.' || ElfArch::Bind(sym) != STB_LOCAL;
+    default:
+      return true;
+    }
+  }
+
+  // Iterate over the symbols in a section, either SHT_DYNSYM or
+  // SHT_SYMTAB. Add all symbols to the given SymbolMap.
+  /*
+  void GetSymbolPositions(SymbolMap *symbols,
+                          typename ElfArch::Word section_type,
+                          uint64 mem_offset,
+                          uint64 file_offset) {
+    // This map is used to filter out "nested" functions.
+    // See comment below.
+    AddrToSymMap addr_to_sym_map;
+    for (SymbolIterator<ElfArch> it(this, section_type);
+         !it.done(); it.Next()) {
+      const char *name = it.GetSymbolName();
+      if (name == NULL)
+        continue;
+      const typename ElfArch::Sym *sym = it.GetSymbol();
+      if (CanUseSymbol(name, sym)) {
+        const int sec = sym->st_shndx;
+
+        // We don't support special section indices. The most common
+        // is SHN_ABS, for absolute symbols used deep in the bowels of
+        // glibc. Also ignore any undefined symbols.
+        if (sec == SHN_UNDEF ||
+            (sec >= SHN_LORESERVE && sec <= SHN_HIRESERVE)) {
+          continue;
+        }
+
+        const typename ElfArch::Shdr& hdr = section_headers_[sec];
+
+        // Adjust for difference between where we expected to mmap
+        // this section, and where it was actually mmapped.
+        const int64 expected_base = hdr.sh_addr - hdr.sh_offset;
+        const int64 real_base = mem_offset - file_offset;
+        const int64 adjust = real_base - expected_base;
+
+        uint64 start = sym->st_value + adjust;
+
+        // Adjust function symbols for PowerPC64 by dereferencing and adjusting
+        // the function descriptor to get the function address.
+        if (header_.e_machine == EM_PPC64 && ElfArch::Type(sym) == STT_FUNC) {
+          const uint64 opd_addr =
+              AdjustPPC64FunctionDescriptorSymbolValue(sym->st_value);
+          // Only adjust the returned value if the function address was found.
+          if (opd_addr != sym->st_value) {
+            const int64 adjust_function_symbols =
+                real_base - base_for_text_;
+            start = opd_addr + adjust_function_symbols;
+          }
+        }
+
+        addr_to_sym_map.push_back(std::make_pair(start, sym));
+      }
+    }
+    std::sort(addr_to_sym_map.begin(), addr_to_sym_map.end(), &AddrToSymSorter);
+    addr_to_sym_map.erase(std::unique(addr_to_sym_map.begin(),
+                                      addr_to_sym_map.end(), &AddrToSymEquals),
+                          addr_to_sym_map.end());
+
+    // Squeeze out any "nested functions".
+    // Nested functions are not allowed in C, but libc plays tricks.
+    //
+    // For example, here is disassembly of /lib64/tls/libc-2.3.5.so:
+    //   0x00000000000aa380 <read+0>:             cmpl   $0x0,0x2781b9(%rip)
+    //   0x00000000000aa387 <read+7>:             jne    0xaa39b <read+27>
+    //   0x00000000000aa389 <__read_nocancel+0>:  mov    $0x0,%rax
+    //   0x00000000000aa390 <__read_nocancel+7>:  syscall
+    //   0x00000000000aa392 <__read_nocancel+9>:  cmp $0xfffffffffffff001,%rax
+    //   0x00000000000aa398 <__read_nocancel+15>: jae    0xaa3ef <read+111>
+    //   0x00000000000aa39a <__read_nocancel+17>: retq
+    //   0x00000000000aa39b <read+27>:            sub    $0x28,%rsp
+    //   0x00000000000aa39f <read+31>:            mov    %rdi,0x8(%rsp)
+    //   ...
+    // Without removing __read_nocancel, symbolizer will return NULL
+    // given e.g. 0xaa39f (because the lower bound is __read_nocancel,
+    // but 0xaa39f is beyond its end.
+    if (addr_to_sym_map.empty()) {
+      return;
+    }
+    const ElfSectionReader<ElfArch> *const symbol_section =
+        this->GetSectionByType(section_type);
+    const ElfSectionReader<ElfArch> *const string_section =
+        this->GetSection(symbol_section->header().sh_link);
+
+    typename AddrToSymMap::iterator curr = addr_to_sym_map.begin();
+    // Always insert the first symbol.
+    symbols->AddSymbol(string_section->GetOffset(curr->second->st_name),
+                       curr->first, curr->second->st_size);
+    typename AddrToSymMap::iterator prev = curr++;
+    for (; curr != addr_to_sym_map.end(); ++curr) {
+      const uint64 prev_addr = prev->first;
+      const uint64 curr_addr = curr->first;
+      const typename ElfArch::Sym *const prev_sym = prev->second;
+      const typename ElfArch::Sym *const curr_sym = curr->second;
+      if (prev_addr + prev_sym->st_size <= curr_addr ||
+          // The next condition is true if two symbols overlap like this:
+          //
+          //   Previous symbol  |----------------------------|
+          //   Current symbol     |-------------------------------|
+          //
+          // These symbols are not found in google3 codebase, but in
+          // jdk1.6.0_01_gg1/jre/lib/i386/server/libjvm.so.
+          //
+          // 0619e040 00000046 t CardTableModRefBS::write_region_work()
+          // 0619e070 00000046 t CardTableModRefBS::write_ref_array_work()
+          //
+          // We allow overlapped symbols rather than ignore these.
+          // Due to the way SymbolMap::GetSymbolAtPosition() works,
+          // lookup for any address in [curr_addr, curr_addr + its size)
+          // (e.g. 0619e071) will produce the current symbol,
+          // which is the desired outcome.
+          prev_addr + prev_sym->st_size < curr_addr + curr_sym->st_size) {
+        const char *name = string_section->GetOffset(curr_sym->st_name);
+        symbols->AddSymbol(name, curr_addr, curr_sym->st_size);
+        prev = curr;
+      } else {
+        // Current symbol is "nested" inside previous one like this:
+        //
+        //   Previous symbol  |----------------------------|
+        //   Current symbol     |---------------------|
+        //
+        // This happens within glibc, e.g. __read_nocancel is nested
+        // "inside" __read. Ignore "inner" symbol.
+        //DCHECK_LE(curr_addr + curr_sym->st_size,
+        //          prev_addr + prev_sym->st_size);
+        ;
+      }
+    }
+  }
+*/
+
+  void VisitSymbols(typename ElfArch::Word section_type,
+                    ElfReader::SymbolSink *sink) {
+    VisitSymbols(section_type, sink, -1, -1, false);
+  }
+
+  void VisitSymbols(typename ElfArch::Word section_type,
+                    ElfReader::SymbolSink *sink,
+                    int symbol_binding,
+                    int symbol_type,
+                    bool get_raw_symbol_values) {
+    for (SymbolIterator<ElfArch> it(this, section_type);
+         !it.done(); it.Next()) {
+      const char *name = it.GetSymbolName();
+      if (!name) continue;
+      const typename ElfArch::Sym *sym = it.GetSymbol();
+      if ((symbol_binding < 0 || ElfArch::Bind(sym) == symbol_binding) &&
+          (symbol_type < 0 || ElfArch::Type(sym) == symbol_type)) {
+        typename ElfArch::Sym symbol = *sym;
+        // Add a PLT symbol in addition to the main undefined symbol.
+        // Only do this for SHT_DYNSYM, because PLT symbols are dynamic.
+        int symbol_index = it.GetCurrentSymbolIndex();
+        // TODO(dthomson): Can be removed once all Java code is using the
+        // Google3 launcher.
+        if (section_type == SHT_DYNSYM &&
+            static_cast<unsigned int>(symbol_index) < symbols_plt_offsets_.size() &&
+            symbols_plt_offsets_[symbol_index] != 0) {
+          string plt_name = string(name) + kPLTFunctionSuffix;
+          if (plt_function_names_[symbol_index].empty()) {
+            plt_function_names_[symbol_index] = plt_name;
+          } else if (plt_function_names_[symbol_index] != plt_name) {
+		;
+          }
+          sink->AddSymbol(plt_function_names_[symbol_index].c_str(),
+                          symbols_plt_offsets_[it.GetCurrentSymbolIndex()],
+                          plt_code_size_);
+        }
+        if (!get_raw_symbol_values)
+          AdjustSymbolValue(&symbol);
+        sink->AddSymbol(name, symbol.st_value, symbol.st_size);
+      }
+    }
+  }
+
+  void VisitRelocationEntries() {
+    if (visited_relocation_entries_) {
+      return;
+    }
+    visited_relocation_entries_ = true;
+
+    if (!plts_supported_) {
+      return;
+    }
+    // First determine if PLTs exist. If not, then there is nothing to do.
+    ElfReader::SectionInfo plt_section_info;
+    const char* plt_section =
+        GetSectionInfoByName(kElfPLTSectionName, &plt_section_info);
+    if (!plt_section) {
+      return;
+    }
+    if (plt_section_info.size == 0) {
+      return;
+    }
+
+    // The PLTs could be referenced by either a Rel or Rela (Rel with Addend)
+    // section.
+    ElfReader::SectionInfo rel_section_info;
+    ElfReader::SectionInfo rela_section_info;
+    const char* rel_section =
+        GetSectionInfoByName(kElfPLTRelSectionName, &rel_section_info);
+    const char* rela_section =
+        GetSectionInfoByName(kElfPLTRelaSectionName, &rela_section_info);
+
+    const typename ElfArch::Rel* rel =
+        reinterpret_cast<const typename ElfArch::Rel*>(rel_section);
+    const typename ElfArch::Rela* rela =
+        reinterpret_cast<const typename ElfArch::Rela*>(rela_section);
+
+    if (!rel_section && !rela_section) {
+      return;
+    }
+
+    // Use either Rel or Rela section, depending on which one exists.
+    size_t section_size = rel_section ? rel_section_info.size
+                                      : rela_section_info.size;
+    size_t entry_size = rel_section ? sizeof(typename ElfArch::Rel)
+                                    : sizeof(typename ElfArch::Rela);
+
+    // Determine the number of entries in the dynamic symbol table.
+    ElfReader::SectionInfo dynsym_section_info;
+    const char* dynsym_section =
+        GetSectionInfoByName(kElfDynSymSectionName, &dynsym_section_info);
+    // The dynsym section might not exist, or it might be empty. In either case
+    // there is nothing to be done so return.
+    if (!dynsym_section || dynsym_section_info.size == 0) {
+      return;
+    }
+    size_t num_dynamic_symbols =
+        dynsym_section_info.size / dynsym_section_info.entsize;
+    symbols_plt_offsets_.resize(num_dynamic_symbols, 0);
+
+    // TODO(dthomson): Can be removed once all Java code is using the
+    // Google3 launcher.
+    // Make storage room for PLT function name strings.
+    plt_function_names_.resize(num_dynamic_symbols);
+
+    for (size_t i = 0; i < section_size / entry_size; ++i) {
+      // Determine symbol index from the |r_info| field.
+      int sym_index = ElfArch::r_sym(rel_section ? rel[i].r_info
+                                                 : rela[i].r_info);
+      if (static_cast<unsigned int>(sym_index) >= symbols_plt_offsets_.size()) {
+        continue;
+      }
+      symbols_plt_offsets_[sym_index] =
+          plt_section_info.addr + plt0_size_ + i * plt_code_size_;
+    }
+  }
+
+  // Return an ElfSectionReader for the first section of the given
+  // type by iterating through all section headers. Returns NULL if
+  // the section type is not found.
+  const ElfSectionReader<ElfArch> *GetSectionByType(
+      typename ElfArch::Word section_type) {
+    for (unsigned int k = 0u; k < GetNumSections(); ++k) {
+      if (section_headers_[k].sh_type == section_type) {
+        return GetSection(k);
+      }
+    }
+    return NULL;
+  }
+
+  // Return the name of section "shndx".  Returns NULL if the section
+  // is not found.
+  const char *GetSectionNameByIndex(int shndx) {
+    return GetSectionName(section_headers_[shndx].sh_name);
+  }
+
+  // Return a pointer to section "shndx", and store the size in
+  // "size".  Returns NULL if the section is not found.
+  const char *GetSectionContentsByIndex(int shndx, size_t *size) {
+    const ElfSectionReader<ElfArch> *section = GetSection(shndx);
+    if (section != NULL) {
+      *size = section->section_size();
+      return section->contents();
+    }
+    return NULL;
+  }
+
+  // Return a pointer to the first section of the given name by
+  // iterating through all section headers, and store the size in
+  // "size".  Returns NULL if the section name is not found.
+  const char *GetSectionContentsByName(const string &section_name,
+                                       size_t *size) {
+    for (unsigned int k = 0u; k < GetNumSections(); ++k) {
+      // When searching for sections in a .dwp file, the sections
+      // we're looking for will always be at the end of the section
+      // table, so reverse the direction of iteration.
+      int shndx = is_dwp_ ? GetNumSections() - k - 1 : k;
+      const char *name = GetSectionName(section_headers_[shndx].sh_name);
+      if (name != NULL && ElfReader::SectionNamesMatch(section_name, name)) {
+        const ElfSectionReader<ElfArch> *section = GetSection(shndx);
+        if (section == NULL) {
+          return NULL;
+        } else {
+          *size = section->section_size();
+          return section->contents();
+        }
+      }
+    }
+    return NULL;
+  }
+
+  // This is like GetSectionContentsByName() but it returns a lot of extra
+  // information about the section.
+  const char *GetSectionInfoByName(const string &section_name,
+                                   ElfReader::SectionInfo *info) {
+    for (unsigned int k = 0u; k < GetNumSections(); ++k) {
+      // When searching for sections in a .dwp file, the sections
+      // we're looking for will always be at the end of the section
+      // table, so reverse the direction of iteration.
+      int shndx = is_dwp_ ? GetNumSections() - k - 1 : k;
+      const char *name = GetSectionName(section_headers_[shndx].sh_name);
+      if (name != NULL && ElfReader::SectionNamesMatch(section_name, name)) {
+        const ElfSectionReader<ElfArch> *section = GetSection(shndx);
+        if (section == NULL) {
+          return NULL;
+        } else {
+          info->type = section->header().sh_type;
+          info->flags = section->header().sh_flags;
+          info->addr = section->header().sh_addr;
+          info->offset = section->header().sh_offset;
+          info->size = section->header().sh_size;
+          info->link = section->header().sh_link;
+          info->info = section->header().sh_info;
+          info->addralign = section->header().sh_addralign;
+          info->entsize = section->header().sh_entsize;
+          return section->contents();
+        }
+      }
+    }
+    return NULL;
+  }
+
+  // p_vaddr of the first PT_LOAD segment (if any), or 0 if no PT_LOAD
+  // segments are present. This is the address an ELF image was linked
+  // (by static linker) to be loaded at. Usually (but not always) 0 for
+  // shared libraries and position-independent executables.
+  uint64 VaddrOfFirstLoadSegment() const {
+    // Relocatable objects (of type ET_REL) do not have LOAD segments.
+    if (header_.e_type == ET_REL) {
+      return 0;
+    }
+    for (int i = 0; i < GetNumProgramHeaders(); ++i) {
+      if (program_headers_[i].p_type == PT_LOAD) {
+        return program_headers_[i].p_vaddr;
+      }
+    }
+    return 0;
+  }
+
+  // According to the LSB ("ELF special sections"), sections with debug
+  // info are prefixed by ".debug".  The names are not specified, but they
+  // look like ".debug_line", ".debug_info", etc.
+  bool HasDebugSections() {
+    // Debug sections are likely to be near the end, so reverse the
+    // direction of iteration.
+    for (int k = GetNumSections() - 1; k >= 0; --k) {
+      const char *name = GetSectionName(section_headers_[k].sh_name);
+      if (strncmp(name, ".debug", strlen(".debug")) == 0) return true;
+      if (strncmp(name, ".zdebug", strlen(".zdebug")) == 0) return true;
+    }
+    return false;
+  }
+
+  bool IsDynamicSharedObject() const {
+    return header_.e_type == ET_DYN;
+  }
+
+  // Return the number of sections.
+  uint64_t GetNumSections() const {
+    if (HasManySections())
+      return first_section_header_.sh_size;
+    return header_.e_shnum;
+  }
+
+ private:
+  typedef vector<pair<uint64, const typename ElfArch::Sym *> > AddrToSymMap;
+
+  static bool AddrToSymSorter(const typename AddrToSymMap::value_type& lhs,
+                              const typename AddrToSymMap::value_type& rhs) {
+    return lhs.first < rhs.first;
+  }
+
+  static bool AddrToSymEquals(const typename AddrToSymMap::value_type& lhs,
+                              const typename AddrToSymMap::value_type& rhs) {
+    return lhs.first == rhs.first;
+  }
+
+  // Does this ELF file have too many sections to fit in the program header?
+  bool HasManySections() const {
+    return header_.e_shnum == SHN_UNDEF;
+  }
+
+  // Return the number of program headers.
+  int GetNumProgramHeaders() const {
+    if (HasManySections() && header_.e_phnum == 0xffff &&
+        first_section_header_.sh_info != 0)
+      return first_section_header_.sh_info;
+    return header_.e_phnum;
+  }
+
+  // Return the index of the string table.
+  int GetStringTableIndex() const {
+    if (HasManySections()) {
+      if (header_.e_shstrndx == 0xffff)
+        return first_section_header_.sh_link;
+      else if (header_.e_shstrndx >= GetNumSections())
+        return 0;
+    }
+    return header_.e_shstrndx;
+  }
+
+  // Given an offset into the section header string table, return the
+  // section name.
+  const char *GetSectionName(typename ElfArch::Word sh_name) {
+    const ElfSectionReader<ElfArch> *shstrtab =
+        GetSection(GetStringTableIndex());
+    if (shstrtab != NULL) {
+      return shstrtab->GetOffset(sh_name);
+    }
+    return NULL;
+  }
+
+  // Return an ElfSectionReader for the given section. The reader will
+  // be freed when this object is destroyed.
+  const ElfSectionReader<ElfArch> *GetSection(int num) {
+    const char *name;
+    // Hard-coding the name for the section-name string table prevents
+    // infinite recursion.
+    if (num == GetStringTableIndex())
+      name = ".shstrtab";
+    else
+      name = GetSectionNameByIndex(num);
+    ElfSectionReader<ElfArch> *& reader = sections_[num];
+    if (reader == NULL)
+      reader = new ElfSectionReader<ElfArch>(name, path_, fd_,
+                                             section_headers_[num]);
+    return reader;
+  }
+
+  // Parse out the overall header information from the file and assert
+  // that it looks sane. This contains information like the magic
+  // number and target architecture.
+  bool ParseHeaders(int fd, const string &path) {
+    // Read in the global ELF header.
+    if (pread(fd, &header_, sizeof(header_), 0) != sizeof(header_)) {
+      return false;
+    }
+
+    // Must be an executable, dynamic shared object or relocatable object
+    if (header_.e_type != ET_EXEC &&
+        header_.e_type != ET_DYN &&
+        header_.e_type != ET_REL) {
+      return false;
+    }
+    // Need a section header.
+    if (header_.e_shoff == 0) {
+      return false;
+    }
+
+    if (header_.e_shnum == SHN_UNDEF) {
+      // The number of sections in the program header is only a 16-bit value. In
+      // the event of overflow (greater than SHN_LORESERVE sections), e_shnum
+      // will read SHN_UNDEF and the true number of section header table entries
+      // is found in the sh_size field of the first section header.
+      // See: http://www.sco.com/developers/gabi/2003-12-17/ch4.sheader.html
+      if (pread(fd, &first_section_header_, sizeof(first_section_header_),
+                header_.e_shoff) != sizeof(first_section_header_)) {
+        return false;
+      }
+    }
+
+    // Dynamically allocate enough space to store the section headers
+    // and read them out of the file.
+    const int section_headers_size =
+        GetNumSections() * sizeof(*section_headers_);
+    section_headers_ = new typename ElfArch::Shdr[section_headers_size];
+    if (pread(fd, section_headers_, section_headers_size, header_.e_shoff) !=
+        section_headers_size) {
+      return false;
+    }
+
+    // Dynamically allocate enough space to store the program headers
+    // and read them out of the file.
+    //const int program_headers_size =
+    //    GetNumProgramHeaders() * sizeof(*program_headers_);
+    program_headers_ = new typename ElfArch::Phdr[GetNumProgramHeaders()];
+
+    // Presize the sections array for efficiency.
+    sections_.resize(GetNumSections(), NULL);
+    return true;
+  }
+
+  // Given the "value" of a function descriptor return the address of the
+  // function (i.e. the dereferenced value). Otherwise return "value".
+  uint64 AdjustPPC64FunctionDescriptorSymbolValue(uint64 value) {
+    if (opd_section_ != NULL &&
+        opd_info_.addr <= value &&
+        value < opd_info_.addr + opd_info_.size) {
+      uint64 offset = value - opd_info_.addr;
+      return (*reinterpret_cast<const uint64*>(opd_section_ + offset));
+    }
+    return value;
+  }
+
+  void AdjustSymbolValue(typename ElfArch::Sym* sym) {
+    switch (header_.e_machine) {
+    case EM_ARM:
+      // For ARM architecture, if the LSB of the function symbol offset is set,
+      // it indicates a Thumb function.  This bit should not be taken literally.
+      // Clear it.
+      if (ElfArch::Type(sym) == STT_FUNC)
+        sym->st_value = AdjustARMThumbSymbolValue(sym->st_value);
+      break;
+    case EM_386:
+      // No adjustment needed for Intel x86 architecture.  However, explicitly
+      // define this case as we use it quite often.
+      break;
+    case EM_PPC64:
+      // PowerPC64 currently has function descriptors as part of the ABI.
+      // Function symbols need to be adjusted accordingly.
+      if (ElfArch::Type(sym) == STT_FUNC)
+        sym->st_value = AdjustPPC64FunctionDescriptorSymbolValue(sym->st_value);
+      break;
+    default:
+      break;
+    }
+  }
+
+  friend class SymbolIterator<ElfArch>;
+
+  // The file we're reading.
+  const string path_;
+  // Open file descriptor for path_. Not owned by this object.
+  const int fd_;
+
+  // The global header of the ELF file.
+  typename ElfArch::Ehdr header_;
+
+  // The header of the first section. This may be used to supplement the ELF
+  // file header.
+  typename ElfArch::Shdr first_section_header_;
+
+  // Array of GetNumSections() section headers, allocated when we read
+  // in the global header.
+  typename ElfArch::Shdr *section_headers_;
+
+  // Array of GetNumProgramHeaders() program headers, allocated when we read
+  // in the global header.
+  typename ElfArch::Phdr *program_headers_;
+
+  // An array of pointers to ElfSectionReaders. Sections are
+  // mmaped as they're needed and not released until this object is
+  // destroyed.
+  vector<ElfSectionReader<ElfArch>*> sections_;
+
+  // For PowerPC64 we need to keep track of function descriptors when looking up
+  // values for funtion symbols values. Function descriptors are kept in the
+  // .opd section and are dereferenced to find the function address.
+  ElfReader::SectionInfo opd_info_;
+  const char *opd_section_;  // Must be checked for NULL before use.
+  int64 base_for_text_;
+
+  // Read PLT-related sections for the current architecture.
+  bool plts_supported_;
+  // Code size of each PLT function for the current architecture.
+  size_t plt_code_size_;
+  // Size of the special first entry in the .plt section that calls the runtime
+  // loader resolution routine, and that all other entries jump to when doing
+  // lazy symbol binding.
+  size_t plt0_size_;
+
+  // Maps a dynamic symbol index to a PLT offset.
+  // The vector entry index is the dynamic symbol index.
+  std::vector<uint64> symbols_plt_offsets_;
+
+  // Container for PLT function name strings. These strings are passed by
+  // reference to SymbolSink::AddSymbol() so they need to be stored somewhere.
+  std::vector<string> plt_function_names_;
+
+  bool visited_relocation_entries_;
+
+  // True if this is a .dwp file.
+  bool is_dwp_;
+};
+
+ElfReader::ElfReader(const string &path)
+    : path_(path), fd_(-1), impl32_(NULL), impl64_(NULL) {
+  // linux 2.6.XX kernel can show deleted files like this:
+  //   /var/run/nscd/dbYLJYaE (deleted)
+  // and the kernel-supplied vdso and vsyscall mappings like this:
+  //   [vdso]
+  //   [vsyscall]
+  if (MyHasSuffixString(path, " (deleted)"))
+    return;
+  if (path == "[vdso]")
+    return;
+  if (path == "[vsyscall]")
+    return;
+
+  fd_ = open(path.c_str(), O_RDONLY);
+}
+
+ElfReader::~ElfReader() {
+  if (fd_ != -1)
+    close(fd_);
+  if (impl32_ != NULL)
+    delete impl32_;
+  if (impl64_ != NULL)
+    delete impl64_;
+}
+
+
+// The only word-size specific part of this file is IsNativeElfFile().
+#if __WORDSIZE == 32
+#define NATIVE_ELF_ARCH Elf32
+#elif __WORDSIZE == 64
+#define NATIVE_ELF_ARCH Elf64
+#else
+#error "Invalid word size"
+#endif
+
+template <typename ElfArch>
+static bool IsElfFile(const int fd, const string &path) {
+  if (fd < 0)
+    return false;
+  if (!ElfReaderImpl<ElfArch>::IsArchElfFile(fd, NULL)) {
+    // No error message here.  IsElfFile gets called many times.
+    return false;
+  }
+  return true;
+}
+
+bool ElfReader::IsNativeElfFile() const {
+  return IsElfFile<NATIVE_ELF_ARCH>(fd_, path_);
+}
+
+bool ElfReader::IsElf32File() const {
+  return IsElfFile<Elf32>(fd_, path_);
+}
+
+bool ElfReader::IsElf64File() const {
+  return IsElfFile<Elf64>(fd_, path_);
+}
+
+/*
+void ElfReader::AddSymbols(SymbolMap *symbols,
+                           uint64 mem_offset, uint64 file_offset,
+                           uint64 length) {
+  if (fd_ < 0)
+    return;
+  // TODO(chatham): Actually use the information about file offset and
+  // the length of the mapped section. On some machines the data
+  // section gets mapped as executable, and we'll end up reading the
+  // file twice and getting some of the offsets wrong.
+  if (IsElf32File()) {
+    GetImpl32()->GetSymbolPositions(symbols, SHT_SYMTAB,
+                                    mem_offset, file_offset);
+    GetImpl32()->GetSymbolPositions(symbols, SHT_DYNSYM,
+                                    mem_offset, file_offset);
+  } else if (IsElf64File()) {
+    GetImpl64()->GetSymbolPositions(symbols, SHT_SYMTAB,
+                                    mem_offset, file_offset);
+    GetImpl64()->GetSymbolPositions(symbols, SHT_DYNSYM,
+                                    mem_offset, file_offset);
+  }
+}
+*/
+
+void ElfReader::VisitSymbols(ElfReader::SymbolSink *sink) {
+  VisitSymbols(sink, -1, -1);
+}
+
+void ElfReader::VisitSymbols(ElfReader::SymbolSink *sink,
+                             int symbol_binding,
+                             int symbol_type) {
+  VisitSymbols(sink, symbol_binding, symbol_type, false);
+}
+
+void ElfReader::VisitSymbols(ElfReader::SymbolSink *sink,
+                             int symbol_binding,
+                             int symbol_type,
+                             bool get_raw_symbol_values) {
+  if (IsElf32File()) {
+    GetImpl32()->VisitRelocationEntries();
+    GetImpl32()->VisitSymbols(SHT_SYMTAB, sink, symbol_binding, symbol_type,
+                              get_raw_symbol_values);
+    GetImpl32()->VisitSymbols(SHT_DYNSYM, sink, symbol_binding, symbol_type,
+                              get_raw_symbol_values);
+  } else if (IsElf64File()) {
+    GetImpl64()->VisitRelocationEntries();
+    GetImpl64()->VisitSymbols(SHT_SYMTAB, sink, symbol_binding, symbol_type,
+                              get_raw_symbol_values);
+    GetImpl64()->VisitSymbols(SHT_DYNSYM, sink, symbol_binding, symbol_type,
+                              get_raw_symbol_values);
+  }
+}
+
+uint64 ElfReader::VaddrOfFirstLoadSegment() {
+  if (IsElf32File()) {
+    return GetImpl32()->VaddrOfFirstLoadSegment();
+  } else if (IsElf64File()) {
+    return GetImpl64()->VaddrOfFirstLoadSegment();
+  } else {
+    return 0;
+  }
+}
+
+const char *ElfReader::GetSectionName(int shndx) {
+  if (shndx < 0 || static_cast<unsigned int>(shndx) >= GetNumSections()) return NULL;
+  if (IsElf32File()) {
+    return GetImpl32()->GetSectionNameByIndex(shndx);
+  } else if (IsElf64File()) {
+    return GetImpl64()->GetSectionNameByIndex(shndx);
+  } else {
+    return NULL;
+  }
+}
+
+uint64 ElfReader::GetNumSections() {
+  if (IsElf32File()) {
+    return GetImpl32()->GetNumSections();
+  } else if (IsElf64File()) {
+    return GetImpl64()->GetNumSections();
+  } else {
+    return 0;
+  }
+}
+
+const char *ElfReader::GetSectionByIndex(int shndx, size_t *size) {
+  if (IsElf32File()) {
+    return GetImpl32()->GetSectionContentsByIndex(shndx, size);
+  } else if (IsElf64File()) {
+    return GetImpl64()->GetSectionContentsByIndex(shndx, size);
+  } else {
+    return NULL;
+  }
+}
+
+const char *ElfReader::GetSectionByName(const string &section_name,
+                                        size_t *size) {
+  if (IsElf32File()) {
+    return GetImpl32()->GetSectionContentsByName(section_name, size);
+  } else if (IsElf64File()) {
+    return GetImpl64()->GetSectionContentsByName(section_name, size);
+  } else {
+    return NULL;
+  }
+}
+
+const char *ElfReader::GetSectionInfoByName(const string &section_name,
+                                            SectionInfo *info) {
+  if (IsElf32File()) {
+    return GetImpl32()->GetSectionInfoByName(section_name, info);
+  } else if (IsElf64File()) {
+    return GetImpl64()->GetSectionInfoByName(section_name, info);
+  } else {
+    return NULL;
+  }
+}
+
+bool ElfReader::SectionNamesMatch(const string &name, const string &sh_name) {
+  if ((name.find(".debug_", 0) == 0) && (sh_name.find(".zdebug_", 0) == 0)) {
+    const string name_suffix(name, strlen(".debug_"));
+    const string sh_name_suffix(sh_name, strlen(".zdebug_"));
+    return name_suffix == sh_name_suffix;
+  }
+  return name == sh_name;
+}
+
+bool ElfReader::IsDynamicSharedObject() {
+  if (IsElf32File()) {
+    return GetImpl32()->IsDynamicSharedObject();
+  } else if (IsElf64File()) {
+    return GetImpl64()->IsDynamicSharedObject();
+  } else {
+    return false;
+  }
+}
+
+ElfReaderImpl<Elf32> *ElfReader::GetImpl32() {
+  if (impl32_ == NULL) {
+    impl32_ = new ElfReaderImpl<Elf32>(path_, fd_);
+  }
+  return impl32_;
+}
+
+ElfReaderImpl<Elf64> *ElfReader::GetImpl64() {
+  if (impl64_ == NULL) {
+    impl64_ = new ElfReaderImpl<Elf64>(path_, fd_);
+  }
+  return impl64_;
+}
+
+// Return true if file is an ELF binary of ElfArch, with unstripped
+// debug info (debug_only=true) or symbol table (debug_only=false).
+// Otherwise, return false.
+template <typename ElfArch>
+static bool IsNonStrippedELFBinaryImpl(const string &path, const int fd,
+                                       bool debug_only) {
+  if (!ElfReaderImpl<ElfArch>::IsArchElfFile(fd, NULL)) return false;
+  ElfReaderImpl<ElfArch> elf_reader(path, fd);
+  return debug_only ?
+      elf_reader.HasDebugSections()
+      : (elf_reader.GetSectionByType(SHT_SYMTAB) != NULL);
+}
+
+// Helper for the IsNon[Debug]StrippedELFBinary functions.
+static bool IsNonStrippedELFBinaryHelper(const string &path,
+                                         bool debug_only) {
+  const int fd = open(path.c_str(), O_RDONLY);
+  if (fd == -1) {
+    return false;
+  }
+
+  if (IsNonStrippedELFBinaryImpl<Elf32>(path, fd, debug_only) ||
+      IsNonStrippedELFBinaryImpl<Elf64>(path, fd, debug_only)) {
+    close(fd);
+    return true;
+  }
+  close(fd);
+  return false;
+}
+
+bool ElfReader::IsNonStrippedELFBinary(const string &path) {
+  return IsNonStrippedELFBinaryHelper(path, false);
+}
+
+bool ElfReader::IsNonDebugStrippedELFBinary(const string &path) {
+  return IsNonStrippedELFBinaryHelper(path, true);
+}
+}  // namespace dwarf2reader
diff --git a/src/common/dwarf/elf_reader.h b/src/common/dwarf/elf_reader.h
new file mode 100644
index 00000000..07477341
--- /dev/null
+++ b/src/common/dwarf/elf_reader.h
@@ -0,0 +1,166 @@
+// Copyright 2005 Google Inc. All Rights Reserved.
+// Author: chatham@google.com (Andrew Chatham)
+// Author: satorux@google.com (Satoru Takabayashi)
+//
+// ElfReader handles reading in ELF. It can extract symbols from the
+// current process, which may be used to symbolize stack traces
+// without having to make a potentially dangerous call to fork().
+//
+// ElfReader dynamically allocates memory, so it is not appropriate to
+// use once the address space might be corrupted, such as during
+// process death.
+//
+// ElfReader supports both 32-bit and 64-bit ELF binaries.
+
+#ifndef COMMON_DWARF_ELF_READER_H__
+#define COMMON_DWARF_ELF_READER_H__
+
+#include <string>
+#include <vector>
+
+#include "common/dwarf/types.h"
+
+using std::string;
+using std::vector;
+using std::pair;
+
+namespace dwarf2reader {
+
+class SymbolMap;
+class Elf32;
+class Elf64;
+template<typename ElfArch>
+class ElfReaderImpl;
+
+class ElfReader {
+ public:
+  explicit ElfReader(const string &path);
+  ~ElfReader();
+
+  // Parse the ELF prologue of this file and return whether it was
+  // successfully parsed and matches the word size and byte order of
+  // the current process.
+  bool IsNativeElfFile() const;
+
+  // Similar to IsNativeElfFile but checks if it's a 32-bit ELF file.
+  bool IsElf32File() const;
+
+  // Similar to IsNativeElfFile but checks if it's a 64-bit ELF file.
+  bool IsElf64File() const;
+
+  // Checks if it's an ELF file of type ET_DYN (shared object file).
+  bool IsDynamicSharedObject();
+
+  // Add symbols in the given ELF file into the provided SymbolMap,
+  // assuming that the file has been loaded into the specified
+  // offset.
+  //
+  // The remaining arguments are typically taken from a
+  // ProcMapsIterator (base/sysinfo.h) and describe which portions of
+  // the ELF file are mapped into which parts of memory:
+  //
+  // mem_offset - position at which the segment is mapped into memory
+  // file_offset - offset in the file where the mapping begins
+  // length - length of the mapped segment
+  void AddSymbols(SymbolMap *symbols,
+                  uint64 mem_offset, uint64 file_offset,
+                  uint64 length);
+
+  class SymbolSink {
+   public:
+    virtual ~SymbolSink() {}
+    virtual void AddSymbol(const char *name, uint64 address, uint64 size) = 0;
+  };
+
+  // Like AddSymbols above, but with no address correction.
+  // Processes any SHT_SYMTAB section, followed by any SHT_DYNSYM section.
+  void VisitSymbols(SymbolSink *sink);
+
+  // Like VisitSymbols above, but for a specific symbol binding/type.
+  // A negative value for the binding and type parameters means any
+  // binding or type.
+  void VisitSymbols(SymbolSink *sink, int symbol_binding, int symbol_type);
+
+  // Like VisitSymbols above but can optionally export raw symbol values instead
+  // of adjusted ones.
+  void VisitSymbols(SymbolSink *sink, int symbol_binding, int symbol_type,
+                    bool get_raw_symbol_values);
+
+  // p_vaddr of the first PT_LOAD segment (if any), or 0 if no PT_LOAD
+  // segments are present. This is the address an ELF image was linked
+  // (by static linker) to be loaded at. Usually (but not always) 0 for
+  // shared libraries and position-independent executables.
+  uint64 VaddrOfFirstLoadSegment();
+
+  // Return the name of section "shndx".  Returns NULL if the section
+  // is not found.
+  const char *GetSectionName(int shndx);
+
+  // Return the number of sections in the given ELF file.
+  uint64 GetNumSections();
+
+  // Get section "shndx" from the given ELF file.  On success, return
+  // the pointer to the section and store the size in "size".
+  // On error, return NULL.  The returned section data is only valid
+  // until the ElfReader gets destroyed.
+  const char *GetSectionByIndex(int shndx, size_t *size);
+
+  // Get section with "section_name" (ex. ".text", ".symtab") in the
+  // given ELF file.  On success, return the pointer to the section
+  // and store the size in "size".  On error, return NULL.  The
+  // returned section data is only valid until the ElfReader gets
+  // destroyed.
+  const char *GetSectionByName(const string &section_name, size_t *size);
+
+  // This is like GetSectionByName() but it returns a lot of extra information
+  // about the section. The SectionInfo structure is almost identical to
+  // the typedef struct Elf64_Shdr defined in <elf.h>, but is redefined
+  // here so that the many short macro names in <elf.h> don't have to be
+  // added to our already cluttered namespace.
+  struct SectionInfo {
+    uint32 type;                // Section type (SHT_xxx constant from elf.h).
+    uint64 flags;               // Section flags (SHF_xxx constants from elf.h).
+    uint64 addr;                // Section virtual address at execution.
+    uint64 offset;              // Section file offset.
+    uint64 size;                // Section size in bytes.
+    uint32 link;                // Link to another section.
+    uint32 info;                // Additional section information.
+    uint64 addralign;           // Section alignment.
+    uint64 entsize;             // Entry size if section holds a table.
+  };
+  const char *GetSectionInfoByName(const string &section_name,
+                                   SectionInfo *info);
+
+  // Check if "path" is an ELF binary that has not been stripped of symbol
+  // tables.  This function supports both 32-bit and 64-bit ELF binaries.
+  static bool IsNonStrippedELFBinary(const string &path);
+
+  // Check if "path" is an ELF binary that has not been stripped of debug
+  // info. Unlike IsNonStrippedELFBinary, this function will return
+  // false for binaries passed through "strip -S".
+  static bool IsNonDebugStrippedELFBinary(const string &path);
+
+  // Match a requested section name with the section name as it
+  // appears in the elf-file, adjusting for compressed debug section
+  // names.  For example, returns true if name == ".debug_abbrev" and
+  // sh_name == ".zdebug_abbrev"
+  static bool SectionNamesMatch(const string &name, const string &sh_name);
+
+ private:
+  // Lazily initialize impl32_ and return it.
+  ElfReaderImpl<Elf32> *GetImpl32();
+  // Ditto for impl64_.
+  ElfReaderImpl<Elf64> *GetImpl64();
+
+  // Path of the file we're reading.
+  const string path_;
+  // Read-only file descriptor for the file. May be -1 if there was an
+  // error during open.
+  int fd_;
+  ElfReaderImpl<Elf32> *impl32_;
+  ElfReaderImpl<Elf64> *impl64_;
+};
+
+}  // namespace dwarf2reader
+
+#endif  // COMMON_DWARF_ELF_READER_H__