pdb_source_line_writer.cc

来自「SumatraPDF是一款小型开源的pdf阅读工具。虽然玲珑小巧（只有800多K」· CC 代码 · 共 801 行 · 第 1/2 页

  }

  CComBSTR name;
  int stack_param_size;
  if (!GetSymbolFunctionName(symbol, &name, &stack_param_size)) {
    return false;
  }

  fprintf(output_, "PUBLIC %x %x %ws\n", rva,
          stack_param_size > 0 ? stack_param_size : 0, name);
  return true;
}

bool PDBSourceLineWriter::PrintPDBInfo() {
  PDBModuleInfo info;
  if (!GetModuleInfo(&info)) {
    return false;
  }

  // Hard-code "windows" for the OS because that's the only thing that makes
  // sense for PDB files.  (This might not be strictly correct for Windows CE
  // support, but we don't care about that at the moment.)
  fprintf(output_, "MODULE windows %ws %ws %ws\n",
          info.cpu.c_str(), info.debug_identifier.c_str(),
          info.debug_file.c_str());

  return true;
}

// wcstol_positive_strict is sort of like wcstol, but much stricter.  string
// should be a buffer pointing to a null-terminated string containing only
// decimal digits.  If the entire string can be converted to an integer
// without overflowing, and there are no non-digit characters before the
// result is set to the value and this function returns true.  Otherwise,
// this function returns false.  This is an alternative to the strtol, atoi,
// and scanf families, which are not as strict about input and in some cases
// don't provide a good way for the caller to determine if a conversion was
// successful.
static bool wcstol_positive_strict(wchar_t *string, int *result) {
  int value = 0;
  for (wchar_t *c = string; *c != '\0'; ++c) {
    int last_value = value;
    value *= 10;
    // Detect overflow.
    if (value / 10 != last_value || value < 0) {
      return false;
    }
    if (*c < '0' || *c > '9') {
      return false;
    }
    unsigned int c_value = *c - '0';
    last_value = value;
    value += c_value;
    // Detect overflow.
    if (value < last_value) {
      return false;
    }
    // Forbid leading zeroes unless the string is just "0".
    if (value == 0 && *(c+1) != '\0') {
      return false;
    }
  }
  *result = value;
  return true;
}

// static
bool PDBSourceLineWriter::GetSymbolFunctionName(IDiaSymbol *function,
                                                BSTR *name,
                                                int *stack_param_size) {
  *stack_param_size = -1;
  const DWORD undecorate_options = UNDNAME_NO_MS_KEYWORDS |
                                   UNDNAME_NO_FUNCTION_RETURNS |
                                   UNDNAME_NO_ALLOCATION_MODEL |
                                   UNDNAME_NO_ALLOCATION_LANGUAGE |
                                   UNDNAME_NO_THISTYPE |
                                   UNDNAME_NO_ACCESS_SPECIFIERS |
                                   UNDNAME_NO_THROW_SIGNATURES |
                                   UNDNAME_NO_MEMBER_TYPE |
                                   UNDNAME_NO_RETURN_UDT_MODEL |
                                   UNDNAME_NO_ECSU;

  // Use get_undecoratedNameEx to get readable C++ names with arguments.
  if (function->get_undecoratedNameEx(undecorate_options, name) != S_OK) {
    if (function->get_name(name) != S_OK) {
      fprintf(stderr, "failed to get function name\n");
      return false;
    }
    // If a name comes from get_name because no undecorated form existed,
    // it's already formatted properly to be used as output.  Don't do any
    // additional processing.
    //
    // MSVC7's DIA seems to not undecorate names in as many cases as MSVC8's.
    // This will result in calling get_name for some C++ symbols, so
    // all of the parameter and return type information may not be included in
    // the name string.
  } else {
    // C++ uses a bogus "void" argument for functions and methods that don't
    // take any parameters.  Take it out of the undecorated name because it's
    // ugly and unnecessary.
    const wchar_t *replace_string = L"(void)";
    const size_t replace_length = wcslen(replace_string);
    const wchar_t *replacement_string = L"()";
    size_t length = wcslen(*name);
    if (length >= replace_length) {
      wchar_t *name_end = *name + length - replace_length;
      if (wcscmp(name_end, replace_string) == 0) {
        WindowsStringUtils::safe_wcscpy(name_end, replace_length,
                                        replacement_string);
        length = wcslen(*name);
      }
    }

    // Undecorate names used for stdcall and fastcall.  These names prefix
    // the identifier with '_' (stdcall) or '@' (fastcall) and suffix it
    // with '@' followed by the number of bytes of parameters, in decimal.
    // If such a name is found, take note of the size and undecorate it.
    // Only do this for names that aren't C++, which is determined based on
    // whether the undecorated name contains any ':' or '(' characters.
    if (!wcschr(*name, ':') && !wcschr(*name, '(') &&
        (*name[0] == '_' || *name[0] == '@')) {
      wchar_t *last_at = wcsrchr(*name + 1, '@');
      if (last_at && wcstol_positive_strict(last_at + 1, stack_param_size)) {
        // If this function adheres to the fastcall convention, it accepts up
        // to the first 8 bytes of parameters in registers (%ecx and %edx).
        // We're only interested in the stack space used for parameters, so
        // so subtract 8 and don't let the size go below 0.
        if (*name[0] == '@') {
          if (*stack_param_size > 8) {
            *stack_param_size -= 8;
          } else {
            *stack_param_size = 0;
          }
        }

        // Undecorate the name by moving it one character to the left in its
        // buffer, and terminating it where the last '@' had been.
        WindowsStringUtils::safe_wcsncpy(*name, length,
                                         *name + 1, last_at - *name - 1);
     } else if (*name[0] == '_') {
        // This symbol's name is encoded according to the cdecl rules.  The
        // name doesn't end in a '@' character followed by a decimal positive
        // integer, so it's not a stdcall name.  Strip off the leading
        // underscore.
        WindowsStringUtils::safe_wcsncpy(*name, length, *name + 1, length);
      }
    }
  }

  return true;
}

// static
int PDBSourceLineWriter::GetFunctionStackParamSize(IDiaSymbol *function) {
  // This implementation is highly x86-specific.

  // Gather the symbols corresponding to data.
  CComPtr<IDiaEnumSymbols> data_children;
  if (FAILED(function->findChildren(SymTagData, NULL, nsNone,
                                    &data_children))) {
    return 0;
  }

  // lowest_base is the lowest %ebp-relative byte offset used for a parameter.
  // highest_end is one greater than the highest offset (i.e. base + length).
  // Stack parameters are assumed to be contiguous, because in reality, they
  // are.
  int lowest_base = INT_MAX;
  int highest_end = INT_MIN;

  CComPtr<IDiaSymbol> child;
  DWORD count;
  while (SUCCEEDED(data_children->Next(1, &child, &count)) && count == 1) {
    // If any operation fails at this point, just proceed to the next child.
    // Use the next_child label instead of continue because child needs to
    // be released before it's reused.  Declare constructable/destructable
    // types early to avoid gotos that cross initializations.
    CComPtr<IDiaSymbol> child_type;

    // DataIsObjectPtr is only used for |this|.  Because |this| can be passed
    // as a stack parameter, look for it in addition to traditional
    // parameters.
    DWORD child_kind;
    if (FAILED(child->get_dataKind(&child_kind)) ||
        (child_kind != DataIsParam && child_kind != DataIsObjectPtr)) {
      goto next_child;
    }

    // Only concentrate on register-relative parameters.  Parameters may also
    // be enregistered (passed directly in a register), but those don't
    // consume any stack space, so they're not of interest.
    DWORD child_location_type;
    if (FAILED(child->get_locationType(&child_location_type)) ||
        child_location_type != LocIsRegRel) {
      goto next_child;
    }

    // Of register-relative parameters, the only ones that make any sense are
    // %ebp- or %esp-relative.  Note that MSVC's debugging information always
    // gives parameters as %ebp-relative even when a function doesn't use a
    // traditional frame pointer and stack parameters are accessed relative to
    // %esp, so just look for %ebp-relative parameters.  If you wanted to
    // access parameters, you'd probably want to treat these %ebp-relative
    // offsets as if they were relative to %esp before a function's prolog
    // executed.
    DWORD child_register;
    if (FAILED(child->get_registerId(&child_register)) ||
        child_register != CV_REG_EBP) {
      goto next_child;
    }

    LONG child_register_offset;
    if (FAILED(child->get_offset(&child_register_offset))) {
      goto next_child;
    }

    // IDiaSymbol::get_type can succeed but still pass back a NULL value.
    if (FAILED(child->get_type(&child_type)) || !child_type) {
      goto next_child;
    }

    ULONGLONG child_length;
    if (FAILED(child_type->get_length(&child_length))) {
      goto next_child;
    }

    int child_end = child_register_offset + static_cast<ULONG>(child_length);
    if (child_register_offset < lowest_base) {
      lowest_base = child_register_offset;
    }
    if (child_end > highest_end) {
      highest_end = child_end;
    }

next_child:
    child.Release();
  }

  int param_size = 0;
  // Make sure lowest_base isn't less than 4, because [%esp+4] is the lowest
  // possible address to find a stack parameter before executing a function's
  // prolog (see above).  Some optimizations cause parameter offsets to be
  // lower than 4, but we're not concerned with those because we're only
  // looking for parameters contained in addresses higher than where the
  // return address is stored.
  if (lowest_base < 4) {
    lowest_base = 4;
  }
  if (highest_end > lowest_base) {
    // All stack parameters are pushed as at least 4-byte quantities.  If the
    // last type was narrower than 4 bytes, promote it.  This assumes that all
    // parameters' offsets are 4-byte-aligned, which is always the case.  Only
    // worry about the last type, because we're not summing the type sizes,
    // just looking at the lowest and highest offsets.
    int remainder = highest_end % 4;
    if (remainder) {
      highest_end += 4 - remainder;
    }

    param_size = highest_end - lowest_base;
  }

  return param_size;
}

bool PDBSourceLineWriter::WriteMap(FILE *map_file) {
  output_ = map_file;

  bool ret = PrintPDBInfo() &&
             PrintSourceFiles() && 
             PrintFunctions() &&
             PrintFrameData();

  output_ = NULL;
  return ret;
}

void PDBSourceLineWriter::Close() {
  session_.Release();
}

bool PDBSourceLineWriter::GetModuleInfo(PDBModuleInfo *info) {
  if (!info) {
    return false;
  }

  info->debug_file.clear();
  info->debug_identifier.clear();
  info->cpu.clear();

  CComPtr<IDiaSymbol> global;
  if (FAILED(session_->get_globalScope(&global))) {
    return false;
  }

  // All CPUs in CV_CPU_TYPE_e (cvconst.h) below 0x10 are x86.  There's no
  // single specific constant to use.
  DWORD platform;
  if (SUCCEEDED(global->get_platform(&platform)) && platform < 0x10) {
    info->cpu = L"x86";
  } else {
    // Unexpected, but handle gracefully.
    info->cpu = L"unknown";
  }

  // DWORD* and int* are not compatible.  This is clean and avoids a cast.
  DWORD age;
  if (FAILED(global->get_age(&age))) {
    return false;
  }

  bool uses_guid;
  if (!UsesGUID(&uses_guid)) {
    return false;
  }

  if (uses_guid) {
    GUID guid;
    if (FAILED(global->get_guid(&guid))) {
      return false;
    }

    // Use the same format that the MS symbol server uses in filesystem
    // hierarchies.
    wchar_t age_string[9];
    swprintf(age_string, sizeof(age_string) / sizeof(age_string[0]),
             L"%x", age);

    // remove when VC++7.1 is no longer supported
    age_string[sizeof(age_string) / sizeof(age_string[0]) - 1] = L'\0';

    info->debug_identifier = GUIDString::GUIDToSymbolServerWString(&guid);
    info->debug_identifier.append(age_string);
  } else {
    DWORD signature;
    if (FAILED(global->get_signature(&signature))) {
      return false;
    }

    // Use the same format that the MS symbol server uses in filesystem
    // hierarchies.
    wchar_t identifier_string[17];
    swprintf(identifier_string,
             sizeof(identifier_string) / sizeof(identifier_string[0]),
             L"%08X%x", signature, age);

    // remove when VC++7.1 is no longer supported
    identifier_string[sizeof(identifier_string) /
                      sizeof(identifier_string[0]) - 1] = L'\0';

    info->debug_identifier = identifier_string;
  }

  CComBSTR debug_file_string;
  if (FAILED(global->get_symbolsFileName(&debug_file_string))) {
    return false;
  }
  info->debug_file =
      WindowsStringUtils::GetBaseName(wstring(debug_file_string));

  return true;
}

bool PDBSourceLineWriter::UsesGUID(bool *uses_guid) {
  if (!uses_guid)
    return false;

  CComPtr<IDiaSymbol> global;
  if (FAILED(session_->get_globalScope(&global)))
    return false;

  GUID guid;
  if (FAILED(global->get_guid(&guid)))
    return false;

  DWORD signature;
  if (FAILED(global->get_signature(&signature)))
    return false;

  // There are two possibilities for guid: either it's a real 128-bit GUID
  // as identified in a code module by a new-style CodeView record, or it's
  // a 32-bit signature (timestamp) as identified by an old-style record.
  // See MDCVInfoPDB70 and MDCVInfoPDB20 in minidump_format.h.
  //
  // Because DIA doesn't provide a way to directly determine whether a module
  // uses a GUID or a 32-bit signature, this code checks whether the first 32
  // bits of guid are the same as the signature, and if the rest of guid is
  // zero.  If so, then with a pretty high degree of certainty, there's an
  // old-style CodeView record in use.  This method will only falsely find an
  // an old-style CodeView record if a real 128-bit GUID has its first 32
  // bits set the same as the module's signature (timestamp) and the rest of
  // the GUID is set to 0.  This is highly unlikely.

  GUID signature_guid = {signature};  // 0-initializes other members
  *uses_guid = !IsEqualGUID(guid, signature_guid);
  return true;
}

}  // namespace google_breakpad