build / android / gyp / create_unwind_table.py [blame]

#!/usr/bin/env python3
# Copyright 2021 The Chromium Authors
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.
"""Creates a table of unwind information in Android Chrome's bespoke format."""

import abc
import argparse
import collections
import enum
import json
import logging
import re
import struct
import subprocess
import sys
from typing import (Dict, Iterable, List, NamedTuple, Sequence, TextIO, Tuple,
                    Union)

from util import build_utils

_STACK_CFI_INIT_REGEX = re.compile(
    r'^STACK CFI INIT ([0-9a-f]+) ([0-9a-f]+) (.+)$')
_STACK_CFI_REGEX = re.compile(r'^STACK CFI ([0-9a-f]+) (.+)$')


class AddressCfi(NamedTuple):
  """Record representing CFI for an address within a function.

  Represents the Call Frame Information required to unwind from an address in a
  function.

  Attributes:
      address: The address.
      unwind_instructions: The unwind instructions for the address.

  """
  address: int
  unwind_instructions: str


class FunctionCfi(NamedTuple):
  """Record representing CFI for a function.

  Note: address_cfi[0].address is the start address of the function.

  Attributes:
      size: The function size in bytes.
      address_cfi: The CFI at each address in the function.

  """
  size: int
  address_cfi: Tuple[AddressCfi, ...]


def FilterToNonTombstoneCfi(stream: TextIO) -> Iterable[str]:
  """Generates non-tombstone STACK CFI lines from the stream.

  STACK CFI functions with address 0 correspond are a 'tombstone' record
  associated with dead code and can be ignored. See
  https://bugs.llvm.org/show_bug.cgi?id=47148#c2.

  Args:
      stream: A file object.

  Returns:
      An iterable over the non-tombstone STACK CFI lines in the stream.
  """
  in_tombstone_function = False
  for line in stream:
    if not line.startswith('STACK CFI '):
      continue

    if line.startswith('STACK CFI INIT 0 '):
      in_tombstone_function = True
    elif line.startswith('STACK CFI INIT '):
      in_tombstone_function = False

    if not in_tombstone_function:
      yield line


def ReadFunctionCfi(stream: TextIO) -> Iterable[FunctionCfi]:
  """Generates FunctionCfi records from the stream.

  Args:
      stream: A file object.

  Returns:
      An iterable over FunctionCfi corresponding to the non-tombstone STACK CFI
      lines in the stream.
  """
  current_function_address = None
  current_function_size = None
  current_function_address_cfi = []
  for line in FilterToNonTombstoneCfi(stream):
    cfi_init_match = _STACK_CFI_INIT_REGEX.search(line)
    if cfi_init_match:
      # Function CFI with address 0 are tombstone entries per
      # https://bugs.llvm.org/show_bug.cgi?id=47148#c2 and should have been
      # filtered in `FilterToNonTombstoneCfi`.
      assert current_function_address != 0
      if (current_function_address is not None
          and current_function_size is not None):
        yield FunctionCfi(current_function_size,
                          tuple(current_function_address_cfi))
      current_function_address = int(cfi_init_match.group(1), 16)
      current_function_size = int(cfi_init_match.group(2), 16)
      current_function_address_cfi = [
          AddressCfi(int(cfi_init_match.group(1), 16), cfi_init_match.group(3))
      ]
    else:
      cfi_match = _STACK_CFI_REGEX.search(line)
      assert cfi_match
      current_function_address_cfi.append(
          AddressCfi(int(cfi_match.group(1), 16), cfi_match.group(2)))

  assert current_function_address is not None
  assert current_function_size is not None
  yield FunctionCfi(current_function_size, tuple(current_function_address_cfi))


def EncodeAsBytes(*values: int) -> bytes:
  """Encodes the argument ints as bytes.

  This function validates that the inputs are within the range that can be
  represented as bytes.

  Args:
    values: Integers in range [0, 255].

  Returns:
    The values encoded as bytes.
  """
  for i, value in enumerate(values):
    if not 0 <= value <= 255:
      raise ValueError('value = %d out of bounds at byte %d' % (value, i))
  return bytes(values)


def Uleb128Encode(value: int) -> bytes:
  """Encodes the argument int to ULEB128 format.

  Args:
    value: Unsigned integer.

  Returns:
    The values encoded as ULEB128 bytes.
  """
  if value < 0:
    raise ValueError(f'Cannot uleb128 encode negative value ({value}).')

  uleb128_bytes = []
  done = False
  while not done:
    value, lowest_seven_bits = divmod(value, 0x80)
    done = value == 0
    uleb128_bytes.append(lowest_seven_bits | (0x80 if not done else 0x00))
  return EncodeAsBytes(*uleb128_bytes)


def EncodeStackPointerUpdate(offset: int) -> bytes:
  """Encodes a stack pointer update as arm unwind instructions.

  Args:
    offset: Offset to apply on stack pointer. Should be in range [-0x204, inf).

  Returns:
    A list of arm unwind instructions as bytes.
  """
  assert offset % 4 == 0

  abs_offset = abs(offset)
  instruction_code = 0b01000000 if offset < 0 else 0b00000000
  if 0x04 <= abs_offset <= 0x200:
    instructions = [
        # vsp = vsp + (xxxxxx << 2) + 4. Covers range 0x04-0x100 inclusive.
        instruction_code | ((min(abs_offset, 0x100) - 4) >> 2)
    ]
    # For vsp increments of 0x104-0x200 we use 00xxxxxx twice.
    if abs_offset >= 0x104:
      instructions.append(instruction_code | ((abs_offset - 0x100 - 4) >> 2))
    try:
      return EncodeAsBytes(*instructions)
    except ValueError as e:
      raise RuntimeError('offset = %d produced out of range value' %
                         offset) from e
  else:
    # This only encodes positive sp movement.
    assert offset > 0, offset
    return EncodeAsBytes(0b10110010  # vsp = vsp + 0x204 + (uleb128 << 2)
                         ) + Uleb128Encode((offset - 0x204) >> 2)


def EncodePop(registers: Sequence[int]) -> bytes:
  """Encodes popping of a sequence of registers as arm unwind instructions.

  Args:
    registers: Collection of target registers to accept values popped from
      stack. Register value order in the sequence does not matter. Values are
      popped based on register index order.

  Returns:
    A list of arm unwind instructions as bytes.
  """
  assert all(
      r in range(4, 16)
      for r in registers), f'Can only pop r4 ~ r15. Registers:\n{registers}.'
  assert len(registers) > 0, 'Register sequence cannot be empty.'

  instructions: List[int] = []

  # Check if the pushed registers are continuous set starting from r4 (and
  # ending prior to r12). This scenario has its own encoding.
  pop_lr = 14 in registers
  non_lr_registers = [r for r in registers if r != 14]
  non_lr_registers_continuous_from_r4 = \
    sorted(non_lr_registers) == list(range(4, 4 + len(non_lr_registers)))

  if (pop_lr and 0 < len(non_lr_registers) <= 8
      and non_lr_registers_continuous_from_r4):
    instructions.append(0b10101000
                        | (len(non_lr_registers) - 1)  # Pop r4-r[4+nnn], r14.
                        )
  else:
    register_bits = 0
    for register in registers:
      register_bits |= 1 << register
    register_bits = register_bits >> 4  # Skip r0 ~ r3.
    instructions.extend([
        # Pop up to 12 integer registers under masks {r15-r12}, {r11-r4}.
        0b10000000 | (register_bits >> 8),
        register_bits & 0xff
    ])

  return EncodeAsBytes(*instructions)


class UnwindType(enum.Enum):
  """
  The type of unwind action to perform.
  """

  # Use lr as the return address.
  RETURN_TO_LR = 1

  # Increment or decrement the stack pointer and/or pop registers (r4 ~ r15).
  # If both, the increment/decrement occurs first.
  UPDATE_SP_AND_OR_POP_REGISTERS = 2

  # Restore the stack pointer from a register then increment/decrement the stack
  # pointer.
  RESTORE_SP_FROM_REGISTER = 3

  # No action necessary. Used for floating point register pops.
  NO_ACTION = 4


class AddressUnwind(NamedTuple):
  """Record representing unwind information for an address within a function.

  Attributes:
      address_offset: The offset of the address from the start of the function.
      unwind_type: The type of unwind to perform from the address.
      sp_offset: The offset to apply to the stack pointer.
      registers: The registers involved in the unwind.
  """
  address_offset: int
  unwind_type: UnwindType
  sp_offset: int
  registers: Tuple[int, ...]


class FunctionUnwind(NamedTuple):
  """Record representing unwind information for a function.

  Attributes:
      address: The address of the function.
      size: The function size in bytes.
      address_unwind_info: The unwind info at each address in the function.
  """

  address: int
  size: int
  address_unwinds: Tuple[AddressUnwind, ...]


def EncodeAddressUnwind(address_unwind: AddressUnwind) -> bytes:
  """Encodes an `AddressUnwind` object as arm unwind instructions.

  Args:
    address_unwind: Record representing unwind information for an address within
      a function.

  Returns:
    A list of arm unwind instructions as bytes.
  """
  if address_unwind.unwind_type == UnwindType.RETURN_TO_LR:
    return EncodeAsBytes(0b10110000)  # Finish.
  if address_unwind.unwind_type == UnwindType.UPDATE_SP_AND_OR_POP_REGISTERS:
    return ((EncodeStackPointerUpdate(address_unwind.sp_offset)
             if address_unwind.sp_offset else b'') +
            (EncodePop(address_unwind.registers)
             if address_unwind.registers else b''))

  if address_unwind.unwind_type == UnwindType.RESTORE_SP_FROM_REGISTER:
    assert len(address_unwind.registers) == 1
    return (EncodeAsBytes(0b10010000
                          | address_unwind.registers[0]  # Set vsp = r[nnnn].
                          ) +
            (EncodeStackPointerUpdate(address_unwind.sp_offset)
             if address_unwind.sp_offset else b''))

  if address_unwind.unwind_type == UnwindType.NO_ACTION:
    return b''

  assert False, 'unknown unwind type'
  return b''


class UnwindInstructionsParser(abc.ABC):
  """Base class for parsers of breakpad unwind instruction sequences.

  Provides regexes matching breakpad instruction sequences understood by the
  parser, and parsing of the sequences from the regex match.
  """

  @abc.abstractmethod
  def GetBreakpadInstructionsRegex(self) -> re.Pattern:
    pass

  @abc.abstractmethod
  def ParseFromMatch(self, address_offset: int, cfa_sp_offset: int,
                     match: re.Match) -> Tuple[AddressUnwind, int]:
    """Returns the regex matching the breakpad instructions.

    Args:
      address_offset: Offset from function start address.
      cfa_sp_offset: CFA stack pointer offset.

    Returns:
      The unwind info for the address plus the new cfa_sp_offset.
    """


class NullParser(UnwindInstructionsParser):
  """Translates the state before any instruction has been executed."""

  regex = re.compile(r'^\.cfa: sp 0 \+ \.ra: lr$')

  def GetBreakpadInstructionsRegex(self) -> re.Pattern:
    return self.regex

  def ParseFromMatch(self, address_offset: int, cfa_sp_offset: int,
                     match: re.Match) -> Tuple[AddressUnwind, int]:
    return AddressUnwind(address_offset, UnwindType.RETURN_TO_LR, 0, ()), 0


class PushOrSubSpParser(UnwindInstructionsParser):
  """Translates unwinds from push or sub sp, #constant instructions."""

  # We expect at least one of the three outer groups to be non-empty. Cases:
  #
  # Standard prologue pushes.
  #   Match the first two and optionally the third.
  #
  # Standard prologue sub sp, #constant.
  #   Match only the first.
  #
  # Pushes in dynamic stack allocation functions after saving sp.
  #   Match only the third since they don't alter the stack pointer or store the
  #   return address.
  #
  # Leaf functions that use callee-save registers.
  #   Match the first and third but not the second.
  regex = re.compile(r'^(?:\.cfa: sp (\d+) \+ ?)?'
                     r'(?:\.ra: \.cfa (-\d+) \+ \^ ?)?'
                     r'((?:r\d+: \.cfa -\d+ \+ \^ ?)*)$')

  # 'r' followed by digits, with 'r' matched via positive lookbehind so only the
  # number appears in the match.
  register_regex = re.compile('(?<=r)(\d+)')

  def GetBreakpadInstructionsRegex(self) -> re.Pattern:
    return self.regex

  def ParseFromMatch(self, address_offset: int, cfa_sp_offset: int,
                     match: re.Match) -> Tuple[AddressUnwind, int]:
    # The group will be None if the outer non-capturing groups for the(\d+) and
    # (-\d+) expressions are not matched.
    new_cfa_sp_offset, ra_cfa_offset = (int(group) if group else None
                                        for group in match.groups()[:2])

    # Registers are pushed in reverse order by register number so are popped in
    # order. Sort them to ensure the proper order.
    registers = sorted([
        int(register)
        for register in self.register_regex.findall(match.group(3))
        # `UpdateSpAndOrPopRegisters` only supports popping of register
        # r4 ~ r15. The ignored registers are translated to sp increments by
        # the following calculation on `sp_offset`.
        if int(register) in range(4, 16)
    ] +
                       # Also pop lr (ra in breakpad terms) if it was stored.
                       ([14] if ra_cfa_offset is not None else []))

    sp_offset = 0
    if new_cfa_sp_offset is not None:
      sp_offset = new_cfa_sp_offset - cfa_sp_offset
      assert sp_offset % 4 == 0
      if sp_offset >= len(registers) * 4:
        # Handles the sub sp, #constant case, and push instructions that push
        # caller-save registers r0-r3 which don't get encoded in the unwind
        # instructions. In the latter case we need to move the stack pointer up
        # to the first pushed register.
        sp_offset -= len(registers) * 4

    return AddressUnwind(address_offset,
                         UnwindType.UPDATE_SP_AND_OR_POP_REGISTERS, sp_offset,
                         tuple(registers)), new_cfa_sp_offset or cfa_sp_offset


class VPushParser(UnwindInstructionsParser):
  # VPushes that occur in dynamic stack allocation functions after storing the
  # stack pointer don't change the stack pointer or push any register that we
  # care about. The first group will not match in those cases.
  #
  # Breakpad doesn't seem to understand how to name the floating point
  # registers so calls them unnamed_register.
  regex = re.compile(r'^(?:\.cfa: sp (\d+) \+ )?'
                     r'(?:unnamed_register\d+: \.cfa -\d+ \+ \^ ?)+$')

  def GetBreakpadInstructionsRegex(self) -> re.Pattern:
    return self.regex

  def ParseFromMatch(self, address_offset: int, cfa_sp_offset: int,
                     match: re.Match) -> Tuple[AddressUnwind, int]:
    # `match.group(1)`, which corresponds to the (\d+) expression, will be None
    # if the first outer non-capturing group is not matched.
    new_cfa_sp_offset = int(match.group(1)) if match.group(1) else None
    if new_cfa_sp_offset is None:
      return (AddressUnwind(address_offset, UnwindType.NO_ACTION, 0,
                            ()), cfa_sp_offset)

    sp_offset = new_cfa_sp_offset - cfa_sp_offset
    assert sp_offset % 4 == 0
    return AddressUnwind(address_offset,
                         UnwindType.UPDATE_SP_AND_OR_POP_REGISTERS, sp_offset,
                         ()), new_cfa_sp_offset


class StoreSpParser(UnwindInstructionsParser):
  regex = re.compile(r'^\.cfa: r(\d+) (\d+) \+$')

  def GetBreakpadInstructionsRegex(self) -> re.Pattern:
    return self.regex

  def ParseFromMatch(self, address_offset: int, cfa_sp_offset: int,
                     match: re.Match) -> Tuple[AddressUnwind, int]:
    register = int(match.group(1))
    new_cfa_sp_offset = int(match.group(2))
    sp_offset = new_cfa_sp_offset - cfa_sp_offset
    assert sp_offset % 4 == 0
    return AddressUnwind(address_offset, UnwindType.RESTORE_SP_FROM_REGISTER,
                         sp_offset, (register, )), new_cfa_sp_offset


def EncodeUnwindInstructionTable(complete_instruction_sequences: Iterable[bytes]
                                 ) -> Tuple[bytes, Dict[bytes, int]]:
  """Encodes the unwind instruction table.

  Deduplicates the encoded unwind instruction sequences. Generates the table and
  a dictionary mapping a function to its starting index in the table.

  The instruction table is used by the unwinder to provide the sequence of
  unwind instructions to execute for each function, separated by offset
  into the function.

  Args:
    complete_instruction_sequences: An iterable of encoded unwind instruction
      sequences. The sequences represent the series of unwind instructions to
      execute corresponding to offsets within each function.

  Returns:
    A tuple containing:
    - The unwind instruction table as bytes.
    - The mapping from the instruction sequence to the offset in the unwind
      instruction table. This mapping is used to construct the function offset
      table, which references entries in the unwind instruction table.
  """
  # As the function offset table uses variable length number encoding (uleb128),
  # which means smaller number uses fewer bytes to represent, we should sort
  # the unwind instruction table by number of references from the function
  # offset table in order to minimize the size of the function offset table.
  ref_counts: Dict[bytes, int] = collections.defaultdict(int)
  for sequence in complete_instruction_sequences:
    ref_counts[sequence] += 1

  def ComputeScore(sequence):
    """ Score for each sequence is computed as  ref_count / size_of_sequence.

    According to greedy algorithm, items with higher value / space cost ratio
    should be prioritized. Here value is bytes saved in the function offset
    table, represetned by ref_count. Space cost is the space taken in the
    unwind instruction table, represented by size_of_sequence.

    Note: In order to ensure build-time determinism, `sequence` is also returned
    to resolve sorting order when scores are the same.
    """
    return ref_counts[sequence] / len(sequence), sequence

  ordered_sequences = sorted(ref_counts.keys(), key=ComputeScore, reverse=True)
  offsets: Dict[bytes, int] = {}
  current_offset = 0
  for sequence in ordered_sequences:
    offsets[sequence] = current_offset
    current_offset += len(sequence)
  return b''.join(ordered_sequences), offsets


class EncodedAddressUnwind(NamedTuple):
  """Record representing unwind information for an address within a function.

  This structure represents the same concept as `AddressUnwind`. The only
  difference is that how to unwind from the address is represented as
  encoded ARM unwind instructions.

  Attributes:
    address_offset: The offset of the address from the start address of the
      function.
    complete_instruction_sequence: The full ARM unwind instruction sequence to
      unwind from the `address_offset`.
  """
  address_offset: int
  complete_instruction_sequence: bytes


def EncodeAddressUnwinds(address_unwinds: Tuple[AddressUnwind, ...]
                         ) -> Tuple[EncodedAddressUnwind, ...]:
  """Encodes the unwind instructions and offset for the addresses within a
  function.

  Args:
    address_unwinds: A tuple of unwind state for addresses within a function.

  Returns:
    The encoded unwind instructions and offsets for the addresses within a
    function, ordered by decreasing offset.
  """
  sorted_address_unwinds: List[AddressUnwind] = sorted(
      address_unwinds,
      key=lambda address_unwind: address_unwind.address_offset,
      reverse=True)
  unwind_instructions: List[bytes] = [
      EncodeAddressUnwind(address_unwind)
      for address_unwind in sorted_address_unwinds
  ]

  # A complete instruction sequence contains all the unwind instructions
  # necessary to unwind from an offset within a function. For a given offset
  # this includes the offset's instructions plus the instructions for all
  # earlier offsets. The offsets are stored in reverse order, hence the i:
  # range rather than :i+1.
  complete_instruction_sequences = [
      b''.join(unwind_instructions[i:]) for i in range(len(unwind_instructions))
  ]

  encoded_unwinds: List[EncodedAddressUnwind] = []
  for address_unwind, sequence in zip(sorted_address_unwinds,
                                      complete_instruction_sequences):
    encoded_unwinds.append(
        EncodedAddressUnwind(address_unwind.address_offset, sequence))
  return tuple(encoded_unwinds)


class EncodedFunctionUnwind(NamedTuple):
  """Record representing unwind information for a function.

  This structure represents the same concept as `FunctionUnwind`, but with
  some differences:
  - Attribute `address` is split into 2 attributes: `page_number` and
    `page_offset`.
  - Attribute `size` is dropped.
  - Attribute `address_unwinds` becomes a collection of `EncodedAddressUnwind`s,
    instead of a collection of `AddressUnwind`s.

  Attributes:
    page_number: The upper bits (17 ~ 31bits) of byte offset from text section
      start.
    page_offset: The lower bits (1 ~ 16bits) of instruction offset from text
      section start.
    address_unwinds: A collection of `EncodedAddressUnwind`s.

  """

  page_number: int
  page_offset: int
  address_unwinds: Tuple[EncodedAddressUnwind, ...]


# The trivial unwind is defined as a single `RETURN_TO_LR` instruction
# at the start of the function.
TRIVIAL_UNWIND: Tuple[EncodedAddressUnwind, ...] = EncodeAddressUnwinds(
    (AddressUnwind(address_offset=0,
                   unwind_type=UnwindType.RETURN_TO_LR,
                   sp_offset=0,
                   registers=()), ))

# The refuse to unwind filler unwind is used to fill the invalid space
# before the first function in the first page and after the last function
# in the last page.
REFUSE_TO_UNWIND: Tuple[EncodedAddressUnwind, ...] = (EncodedAddressUnwind(
    address_offset=0,
    complete_instruction_sequence=bytes([0b10000000, 0b00000000])), )


def EncodeFunctionUnwinds(function_unwinds: Iterable[FunctionUnwind],
                          text_section_start_address: int
                          ) -> Iterable[EncodedFunctionUnwind]:
  """Encodes the unwind state for all functions defined in the binary.

  This function
  - sorts the collection of `FunctionUnwind`s by address.
  - fills in gaps between functions with trivial unwind.
  - fills the space in the last page after last function with refuse to unwind.
  - fills the space in the first page before the first function with refuse
    to unwind.

  Args:
    function_unwinds: An iterable of function unwind states.
    text_section_start_address: The address of .text section in ELF file.

  Returns:
    The encoded function unwind states with no gaps between functions, ordered
    by ascending address.
  """

  def GetPageNumber(address: int) -> int:
    """Calculates the page number.

    Page number is calculated as byte_offset_from_text_section_start >> 17,
    i.e. the upper bits (17 ~ 31bits) of byte offset from text section start.
    """
    return (address - text_section_start_address) >> 17

  def GetPageOffset(address: int) -> int:
    """Calculates the page offset.

    Page offset is calculated as (byte_offset_from_text_section_start >> 1)
    & 0xffff, i.e. the lower bits (1 ~ 16bits) of instruction offset from
    text section start.
    """
    return ((address - text_section_start_address) >> 1) & 0xffff

  sorted_function_unwinds: List[FunctionUnwind] = sorted(
      function_unwinds, key=lambda function_unwind: function_unwind.address)

  if sorted_function_unwinds[0].address > text_section_start_address:
    yield EncodedFunctionUnwind(page_number=0,
                                page_offset=0,
                                address_unwinds=REFUSE_TO_UNWIND)

  prev_func_end_address: int = sorted_function_unwinds[0].address

  gaps = 0
  for unwind in sorted_function_unwinds:
    assert prev_func_end_address <= unwind.address, (
        'Detected overlap between functions.')

    if prev_func_end_address < unwind.address:
      # Gaps between functions are typically filled by regions of thunks which
      # do not alter the stack pointer. Filling these gaps with TRIVIAL_UNWIND
      # is the appropriate unwind strategy.
      gaps += 1
      yield EncodedFunctionUnwind(GetPageNumber(prev_func_end_address),
                                  GetPageOffset(prev_func_end_address),
                                  TRIVIAL_UNWIND)

    yield EncodedFunctionUnwind(GetPageNumber(unwind.address),
                                GetPageOffset(unwind.address),
                                EncodeAddressUnwinds(unwind.address_unwinds))

    prev_func_end_address = unwind.address + unwind.size

  if GetPageOffset(prev_func_end_address) != 0:
    yield EncodedFunctionUnwind(GetPageNumber(prev_func_end_address),
                                GetPageOffset(prev_func_end_address),
                                REFUSE_TO_UNWIND)

  logging.info('%d/%d gaps between functions filled with trivial unwind.', gaps,
               len(sorted_function_unwinds))


def EncodeFunctionOffsetTable(
    encoded_address_unwind_sequences: Iterable[
        Tuple[EncodedAddressUnwind, ...]],
    unwind_instruction_table_offsets: Dict[bytes, int]
) -> Tuple[bytes, Dict[Tuple[EncodedAddressUnwind, ...], int]]:
  """Encodes the function offset table.

  The function offset table maps local instruction offset from function
  start to the location in the unwind instruction table.

  Args:
    encoded_address_unwind_sequences: An iterable of encoded address unwind
      sequences.
    unwind_instruction_table_offsets: The offset mapping returned from
      `EncodeUnwindInstructionTable`.

  Returns:
    A tuple containing:
    - The function offset table as bytes.
    - The mapping from the `EncodedAddressUnwind`s to the offset in the function
      offset table. This mapping is used to construct the function table, which
      references entries in the function offset table.
  """
  function_offset_table = bytearray()
  offsets: Dict[Tuple[EncodedAddressUnwind, ...], int] = {}

  for sequence in encoded_address_unwind_sequences:
    if sequence in offsets:
      continue

    offsets[sequence] = len(function_offset_table)
    for address_offset, complete_instruction_sequence in sequence:
      # Note: address_offset is the number of bytes from one address to another,
      # while the instruction_offset is the number of 2-byte instructions
      # from one address to another.
      instruction_offset = address_offset >> 1
      function_offset_table += (
          Uleb128Encode(instruction_offset) + Uleb128Encode(
              unwind_instruction_table_offsets[complete_instruction_sequence]))

  return bytes(function_offset_table), offsets


def EncodePageTableAndFunctionTable(
    function_unwinds: Iterable[EncodedFunctionUnwind],
    function_offset_table_offsets: Dict[Tuple[EncodedAddressUnwind, ...], int]
) -> Tuple[bytes, bytes]:
  """Encode page table and function table as bytes.

  Page table:
  A table that contains the mapping from page_number to the location of the
  entry for the first function on the page in the function table.

  Function table:
  A table that contains the mapping from page_offset to the location of an entry
  in the function offset table.

  Args:
    function_unwinds: All encoded function unwinds in the module.
    function_offset_table_offsets: The offset mapping returned from
      `EncodeFunctionOffsetTable`.

  Returns:
    A tuple containing:
    - The page table as bytes.
    - The function table as bytes.
  """
  page_function_unwinds: Dict[
      int, List[EncodedFunctionUnwind]] = collections.defaultdict(list)
  for function_unwind in function_unwinds:
    page_function_unwinds[function_unwind.page_number].append(function_unwind)

  raw_page_table: List[int] = []
  function_table = bytearray()

  for page_number, same_page_function_unwinds in sorted(
      page_function_unwinds.items(), key=lambda item: item[0]):
    # Pad empty pages.
    # Empty pages can occur when a function spans over multiple pages.
    # Example:
    # A page table with a starting function that spans 3 over pages.
    # page_table:
    # [0, 1, 1, 1]
    # function_table:
    # [
    #   # Page 0
    #   (0, 20) # This function spans from page 0 offset 0 to page 3 offset 5.
    #   # Page 1 is empty.
    #   # Page 2 is empty.
    #   # Page 3
    #   (6, 70)
    # ]
    assert page_number > len(raw_page_table) - 1
    number_of_empty_pages = page_number - len(raw_page_table)
    # The function table is represented as `base::FunctionTableEntry[]`,
    # where `base::FunctionTableEntry` is 4 bytes.
    function_table_index = len(function_table) // 4
    raw_page_table.extend([function_table_index] * (number_of_empty_pages + 1))
    assert page_number == len(raw_page_table) - 1

    for function_unwind in sorted(
        same_page_function_unwinds,
        key=lambda function_unwind: function_unwind.page_offset):
      function_table += struct.pack(
          'HH', function_unwind.page_offset,
          function_offset_table_offsets[function_unwind.address_unwinds])

  page_table = struct.pack(f'{len(raw_page_table)}I', *raw_page_table)

  return page_table, bytes(function_table)


ALL_PARSERS: Tuple[UnwindInstructionsParser, ...] = (
    NullParser(),
    PushOrSubSpParser(),
    StoreSpParser(),
    VPushParser(),
)


def ParseAddressCfi(address_cfi: AddressCfi, function_start_address: int,
                    parsers: Tuple[UnwindInstructionsParser, ...],
                    prev_cfa_sp_offset: int
                    ) -> Tuple[Union[AddressUnwind, None], bool, int]:
  """Parses address CFI with given parsers.

  Args:
    address_cfi: The CFI for an address in the function.
    function_start_address: The start address of the function.
    parsers: Available parsers to try on CFI data.
    prev_cfa_sp_offset: Previous CFA stack pointer offset.

  Returns:
    A tuple containing:
    - An `AddressUnwind` object when the parse is successful, None otherwise.
    - Whether the address is in function epilogue.
    - The new cfa_sp_offset.
  """
  for parser in parsers:
    match = parser.GetBreakpadInstructionsRegex().search(
        address_cfi.unwind_instructions)
    if not match:
      continue

    address_unwind, cfa_sp_offset = parser.ParseFromMatch(
        address_cfi.address - function_start_address, prev_cfa_sp_offset, match)

    in_epilogue = (
        prev_cfa_sp_offset > cfa_sp_offset
        and address_unwind.unwind_type != UnwindType.RESTORE_SP_FROM_REGISTER)

    return (address_unwind if not in_epilogue else None, in_epilogue,
            cfa_sp_offset)

  return None, False, prev_cfa_sp_offset


def GenerateUnwinds(function_cfis: Iterable[FunctionCfi],
                    parsers: Tuple[UnwindInstructionsParser, ...]
                    ) -> Iterable[FunctionUnwind]:
  """Generates parsed function unwind states from breakpad CFI data.

  This function parses `FunctionCfi`s to `FunctionUnwind`s using
  `UnwindInstructionParser`.

  Args:
    function_cfis: An iterable of function CFI data.
    parsers: Available parsers to try on CFI address data.

  Returns:
    An iterable of parsed function unwind states.
  """
  functions = 0
  addresses = 0
  handled_addresses = 0
  epilogues_seen = 0

  for function_cfi in function_cfis:
    functions += 1
    address_unwinds: List[AddressUnwind] = []
    cfa_sp_offset = 0
    for address_cfi in function_cfi.address_cfi:
      addresses += 1

      address_unwind, in_epilogue, cfa_sp_offset = ParseAddressCfi(
          address_cfi, function_cfi.address_cfi[0].address, parsers,
          cfa_sp_offset)

      if address_unwind:
        handled_addresses += 1
        address_unwinds.append(address_unwind)
        continue

      if in_epilogue:
        epilogues_seen += 1
        break

      logging.info('unrecognized CFI: %x %s.', address_cfi.address,
                   address_cfi.unwind_instructions)

    if address_unwinds:
      # We expect that the unwind information for every function starts with a
      # trivial unwind (RETURN_TO_LR) prior to the execution of any code in the
      # function. This is required by the arm calling convention which involves
      # setting lr to the return address on calling into a function.
      assert address_unwinds[0].address_offset == 0
      assert address_unwinds[0].unwind_type == UnwindType.RETURN_TO_LR

      yield FunctionUnwind(function_cfi.address_cfi[0].address,
                           function_cfi.size, tuple(address_unwinds))

  logging.info('%d functions.', functions)
  logging.info('%d/%d addresses handled.', handled_addresses, addresses)
  logging.info('epilogues_seen: %d.', epilogues_seen)


def EncodeUnwindInfo(page_table: bytes, function_table: bytes,
                     function_offset_table: bytes,
                     unwind_instruction_table: bytes) -> bytes:
  """Encodes all unwind tables as a single binary.

  Concats all unwind table binaries together and attach a header at the start
  with a offset-size pair for each table.

  offset: The offset to the target table from the start of the unwind info
    binary in bytes.
  size: The declared size of the target table.

  Both offset and size are represented as 32bit integers.
  See `base::ChromeUnwindInfoHeaderAndroid` for more details.

  Args:
    page_table: The page table as bytes.
    function_table: The function table as bytes.
    function_offset_table: The function offset table as bytes.
    unwind_instruction_table: The unwind instruction table as bytes.

  Returns:
    A single binary containing
    - A header that points to the location of each table.
    - All unwind tables.
  """
  unwind_info_header = bytearray()
  # Each table is represented as (offset, size) pair, both offset and size
  # are represented as 4 byte integer.
  unwind_info_header_size = 4 * 2 * 4
  unwind_info_body = bytearray()

  # Both the page_table and the function table need to be aligned because their
  # contents are interpreted as multi-byte integers. However, the byte size of
  # the header, the page table, the function table are all multiples of 4 and
  # the resource will be memory mapped at a 4 byte boundary, so no extra care
  # is required to align the page table and the function table.
  #
  # The function offset table and the unwind instruction table are accessed
  # byte by byte, so they only need 1 byte alignment.

  assert len(page_table) % 4 == 0, (
      'Each entry in the page table should be 4-byte integer.')
  assert len(function_table) % 4 == 0, (
      'Each entry in the function table should be a pair of 2 2-byte integers.')

  for table in page_table, function_table:
    offset = unwind_info_header_size + len(unwind_info_body)
    # For the page table and the function_table, declared size is the number of
    # entries in each table. The tables will be aligned to a 4 byte boundary
    # because the resource will be memory mapped at a 4 byte boundary and the
    # header is a multiple of 4 bytes.
    declared_size = len(table) // 4
    unwind_info_header += struct.pack('II', offset, declared_size)
    unwind_info_body += table

  for table in function_offset_table, unwind_instruction_table:
    offset = unwind_info_header_size + len(unwind_info_body)
    # Because both the function offset table and the unwind instruction table
    # contain variable length encoded numbers, the declared size is simply the
    # number of bytes in each table. The tables only require 1 byte alignment.
    declared_size = len(table)
    unwind_info_header += struct.pack('II', offset, declared_size)
    unwind_info_body += table

  return bytes(unwind_info_header + unwind_info_body)


def GenerateUnwindTables(
    encoded_function_unwinds_iterable: Iterable[EncodedFunctionUnwind]
) -> Tuple[bytes, bytes, bytes, bytes]:
  """Generates all unwind tables as bytes.

  Args:
    encoded_function_unwinds_iterable: Encoded function unwinds for all
      functions in the ELF binary.

  Returns:
    A tuple containing:
    - The page table as bytes.
    - The function table as bytes.
    - The function offset table as bytes.
    - The unwind instruction table as bytes.
  """
  encoded_function_unwinds: List[EncodedFunctionUnwind] = list(
      encoded_function_unwinds_iterable)
  complete_instruction_sequences: List[bytes] = []
  encoded_address_unwind_sequences: List[Tuple[EncodedAddressUnwind, ...]] = []

  for encoded_function_unwind in encoded_function_unwinds:
    encoded_address_unwind_sequences.append(
        encoded_function_unwind.address_unwinds)
    for address_unwind in encoded_function_unwind.address_unwinds:
      complete_instruction_sequences.append(
          address_unwind.complete_instruction_sequence)

  unwind_instruction_table, unwind_instruction_table_offsets = (
      EncodeUnwindInstructionTable(complete_instruction_sequences))

  function_offset_table, function_offset_table_offsets = (
      EncodeFunctionOffsetTable(encoded_address_unwind_sequences,
                                unwind_instruction_table_offsets))

  page_table, function_table = EncodePageTableAndFunctionTable(
      encoded_function_unwinds, function_offset_table_offsets)

  return (page_table, function_table, function_offset_table,
          unwind_instruction_table)


def ReadTextSectionStartAddress(readobj_path: str, libchrome_path: str) -> int:
  """Reads the .text section start address of libchrome ELF.

  Arguments:
    readobj_path: Path to llvm-obj binary.
    libchrome_path: Path to libchrome binary.

  Returns:
    The text section start address as a number.
  """
  def GetSectionName(section) -> str:
    # See crbug.com/1426287 for context on different JSON names.
    if 'Name' in section['Section']['Name']:
      return section['Section']['Name']['Name']
    return section['Section']['Name']['Value']

  proc = subprocess.Popen(
      [readobj_path, '--sections', '--elf-output-style=JSON', libchrome_path],
      stdout=subprocess.PIPE,
      encoding='ascii')

  elfs = json.loads(proc.stdout.read())[0]
  sections = elfs['Sections']

  return next(s['Section']['Address'] for s in sections
              if GetSectionName(s) == '.text')


def main():
  build_utils.InitLogging('CREATE_UNWIND_TABLE_DEBUG')
  parser = argparse.ArgumentParser(description=__doc__)
  parser.add_argument('--input_path',
                      help='Path to the unstripped binary.',
                      required=True,
                      metavar='FILE')
  parser.add_argument('--output_path',
                      help='Path to unwind info binary output.',
                      required=True,
                      metavar='FILE')
  parser.add_argument('--dump_syms_path',
                      required=True,
                      help='The path of the dump_syms binary.',
                      metavar='FILE')
  parser.add_argument('--readobj_path',
                      required=True,
                      help='The path of the llvm-readobj binary.',
                      metavar='FILE')

  args = parser.parse_args()
  proc = subprocess.Popen(['./' + args.dump_syms_path, args.input_path, '-v'],
                          stdout=subprocess.PIPE,
                          encoding='ascii')

  function_cfis = ReadFunctionCfi(proc.stdout)
  function_unwinds = GenerateUnwinds(function_cfis, parsers=ALL_PARSERS)
  encoded_function_unwinds = EncodeFunctionUnwinds(
      function_unwinds,
      ReadTextSectionStartAddress(args.readobj_path, args.input_path))
  (page_table, function_table, function_offset_table,
   unwind_instruction_table) = GenerateUnwindTables(encoded_function_unwinds)
  unwind_info: bytes = EncodeUnwindInfo(page_table, function_table,
                                        function_offset_table,
                                        unwind_instruction_table)

  if proc.wait():
    logging.critical('dump_syms exited with return code %d', proc.returncode)
    sys.exit(proc.returncode)

  with open(args.output_path, 'wb') as f:
    f.write(unwind_info)

  return 0


if __name__ == '__main__':
  sys.exit(main())