re2/re2/re2.cc

Go to the documentation of this file.

// Copyright 2003-2009 The RE2 Authors.  All Rights Reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
 
// Regular expression interface RE2.
//
// Originally the PCRE C++ wrapper, but adapted to use
// the new automata-based regular expression engines.
 
#include "re2/re2.h"
 
#include <assert.h>
#include <ctype.h>
#include <errno.h>
#ifdef _MSC_VER
#include <intrin.h>
#endif
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include <algorithm>
#include <atomic>
#include <iterator>
#include <mutex>
#include <string>
#include <utility>
#include <vector>
 
#include "util/util.h"
#include "util/logging.h"
#include "util/strutil.h"
#include "util/utf.h"
#include "re2/prog.h"
#include "re2/regexp.h"
#include "re2/sparse_array.h"
 
namespace re2 {
 
// Maximum number of args we can set
static const int kMaxArgs = 16;
static const int kVecSize = 1+kMaxArgs;
 
const int RE2::Options::kDefaultMaxMem;  // initialized in re2.h
 
RE2::Options::Options(RE2::CannedOptions opt)
  : encoding_(opt == RE2::Latin1 ? EncodingLatin1 : EncodingUTF8),
    posix_syntax_(opt == RE2::POSIX),
    longest_match_(opt == RE2::POSIX),
    log_errors_(opt != RE2::Quiet),
    max_mem_(kDefaultMaxMem),
    literal_(false),
    never_nl_(false),
    dot_nl_(false),
    never_capture_(false),
    case_sensitive_(true),
    perl_classes_(false),
    word_boundary_(false),
    one_line_(false) {
}
 
// static empty objects for use as const references.
// To avoid global constructors, allocated in RE2::Init().
static const std::string* empty_string;
static const std::map<std::string, int>* empty_named_groups;
static const std::map<int, std::string>* empty_group_names;
 
// Converts from Regexp error code to RE2 error code.
// Maybe some day they will diverge.  In any event, this
// hides the existence of Regexp from RE2 users.
static RE2::ErrorCode RegexpErrorToRE2(re2::RegexpStatusCode code) {
  switch (code) {
    case re2::kRegexpSuccess:
      return RE2::NoError;
    case re2::kRegexpInternalError:
      return RE2::ErrorInternal;
    case re2::kRegexpBadEscape:
      return RE2::ErrorBadEscape;
    case re2::kRegexpBadCharClass:
      return RE2::ErrorBadCharClass;
    case re2::kRegexpBadCharRange:
      return RE2::ErrorBadCharRange;
    case re2::kRegexpMissingBracket:
      return RE2::ErrorMissingBracket;
    case re2::kRegexpMissingParen:
      return RE2::ErrorMissingParen;
    case re2::kRegexpUnexpectedParen:
      return RE2::ErrorUnexpectedParen;
    case re2::kRegexpTrailingBackslash:
      return RE2::ErrorTrailingBackslash;
    case re2::kRegexpRepeatArgument:
      return RE2::ErrorRepeatArgument;
    case re2::kRegexpRepeatSize:
      return RE2::ErrorRepeatSize;
    case re2::kRegexpRepeatOp:
      return RE2::ErrorRepeatOp;
    case re2::kRegexpBadPerlOp:
      return RE2::ErrorBadPerlOp;
    case re2::kRegexpBadUTF8:
      return RE2::ErrorBadUTF8;
    case re2::kRegexpBadNamedCapture:
      return RE2::ErrorBadNamedCapture;
  }
  return RE2::ErrorInternal;
}
 
static std::string trunc(const StringPiece& pattern) {
  if (pattern.size() < 100)
    return std::string(pattern);
  return std::string(pattern.substr(0, 100)) + "...";
}
 
 
RE2::RE2(const char* pattern) {
  Init(pattern, DefaultOptions);
}
 
RE2::RE2(const std::string& pattern) {
  Init(pattern, DefaultOptions);
}
 
RE2::RE2(const StringPiece& pattern) {
  Init(pattern, DefaultOptions);
}
 
RE2::RE2(const StringPiece& pattern, const Options& options) {
  Init(pattern, options);
}
 
int RE2::Options::ParseFlags() const {
  int flags = Regexp::ClassNL;
  switch (encoding()) {
    default:
      if (log_errors())
        LOG(ERROR) << "Unknown encoding " << encoding();
      break;
    case RE2::Options::EncodingUTF8:
      break;
    case RE2::Options::EncodingLatin1:
      flags |= Regexp::Latin1;
      break;
  }
 
  if (!posix_syntax())
    flags |= Regexp::LikePerl;
 
  if (literal())
    flags |= Regexp::Literal;
 
  if (never_nl())
    flags |= Regexp::NeverNL;
 
  if (dot_nl())
    flags |= Regexp::DotNL;
 
  if (never_capture())
    flags |= Regexp::NeverCapture;
 
  if (!case_sensitive())
    flags |= Regexp::FoldCase;
 
  if (perl_classes())
    flags |= Regexp::PerlClasses;
 
  if (word_boundary())
    flags |= Regexp::PerlB;
 
  if (one_line())
    flags |= Regexp::OneLine;
 
  return flags;
}
 
void RE2::Init(const StringPiece& pattern, const Options& options) {
  static std::once_flag empty_once;
  std::call_once(empty_once, []() {
    empty_string = new std::string;
    empty_named_groups = new std::map<std::string, int>;
    empty_group_names = new std::map<int, std::string>;
  });
 
  pattern_.assign(pattern.data(), pattern.size());
  options_.Copy(options);
  entire_regexp_ = NULL;
  error_ = empty_string;
  error_code_ = NoError;
  error_arg_.clear();
  prefix_.clear();
  prefix_foldcase_ = false;
  suffix_regexp_ = NULL;
  prog_ = NULL;
  num_captures_ = -1;
  is_one_pass_ = false;
 
  rprog_ = NULL;
  named_groups_ = NULL;
  group_names_ = NULL;
 
  RegexpStatus status;
  entire_regexp_ = Regexp::Parse(
    pattern_,
    static_cast<Regexp::ParseFlags>(options_.ParseFlags()),
    &status);
  if (entire_regexp_ == NULL) {
    if (options_.log_errors()) {
      LOG(ERROR) << "Error parsing '" << trunc(pattern_) << "': "
                 << status.Text();
    }
    error_ = new std::string(status.Text());
    error_code_ = RegexpErrorToRE2(status.code());
    error_arg_ = std::string(status.error_arg());
    return;
  }
 
  re2::Regexp* suffix;
  if (entire_regexp_->RequiredPrefix(&prefix_, &prefix_foldcase_, &suffix))
    suffix_regexp_ = suffix;
  else
    suffix_regexp_ = entire_regexp_->Incref();
 
  // Two thirds of the memory goes to the forward Prog,
  // one third to the reverse prog, because the forward
  // Prog has two DFAs but the reverse prog has one.
  prog_ = suffix_regexp_->CompileToProg(options_.max_mem()*2/3);
  if (prog_ == NULL) {
    if (options_.log_errors())
      LOG(ERROR) << "Error compiling '" << trunc(pattern_) << "'";
    error_ = new std::string("pattern too large - compile failed");
    error_code_ = RE2::ErrorPatternTooLarge;
    return;
  }
 
  // We used to compute this lazily, but it's used during the
  // typical control flow for a match call, so we now compute
  // it eagerly, which avoids the overhead of std::once_flag.
  num_captures_ = suffix_regexp_->NumCaptures();
 
  // Could delay this until the first match call that
  // cares about submatch information, but the one-pass
  // machine's memory gets cut from the DFA memory budget,
  // and that is harder to do if the DFA has already
  // been built.
  is_one_pass_ = prog_->IsOnePass();
}
 
// Returns rprog_, computing it if needed.
re2::Prog* RE2::ReverseProg() const {
  std::call_once(rprog_once_, [](const RE2* re) {
    re->rprog_ =
        re->suffix_regexp_->CompileToReverseProg(re->options_.max_mem() / 3);
    if (re->rprog_ == NULL) {
      if (re->options_.log_errors())
        LOG(ERROR) << "Error reverse compiling '" << trunc(re->pattern_) << "'";
      // We no longer touch error_ and error_code_ because failing to compile
      // the reverse Prog is not a showstopper: falling back to NFA execution
      // is fine. More importantly, an RE2 object is supposed to be logically
      // immutable: whatever ok() would have returned after Init() completed,
      // it should continue to return that no matter what ReverseProg() does.
    }
  }, this);
  return rprog_;
}
 
RE2::~RE2() {
  if (suffix_regexp_)
    suffix_regexp_->Decref();
  if (entire_regexp_)
    entire_regexp_->Decref();
  delete prog_;
  delete rprog_;
  if (error_ != empty_string)
    delete error_;
  if (named_groups_ != NULL && named_groups_ != empty_named_groups)
    delete named_groups_;
  if (group_names_ != NULL &&  group_names_ != empty_group_names)
    delete group_names_;
}
 
int RE2::ProgramSize() const {
  if (prog_ == NULL)
    return -1;
  return prog_->size();
}
 
int RE2::ReverseProgramSize() const {
  if (prog_ == NULL)
    return -1;
  Prog* prog = ReverseProg();
  if (prog == NULL)
    return -1;
  return prog->size();
}
 
// Finds the most significant non-zero bit in n.
static int FindMSBSet(uint32_t n) {
  DCHECK_NE(n, 0);
#if defined(__GNUC__)
  return 31 ^ __builtin_clz(n);
#elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
  unsigned long c;
  _BitScanReverse(&c, n);
  return static_cast<int>(c);
#else
  int c = 0;
  for (int shift = 1 << 4; shift != 0; shift >>= 1) {
    uint32_t word = n >> shift;
    if (word != 0) {
      n = word;
      c += shift;
    }
  }
  return c;
#endif
}
 
static int Fanout(Prog* prog, std::vector<int>* histogram) {
  SparseArray<int> fanout(prog->size());
  prog->Fanout(&fanout);
  int data[32] = {};
  int size = 0;
  for (SparseArray<int>::iterator i = fanout.begin(); i != fanout.end(); ++i) {
    if (i->value() == 0)
      continue;
    uint32_t value = i->value();
    int bucket = FindMSBSet(value);
    bucket += value & (value-1) ? 1 : 0;
    ++data[bucket];
    size = std::max(size, bucket+1);
  }
  if (histogram != NULL)
    histogram->assign(data, data+size);
  return size-1;
}
 
int RE2::ProgramFanout(std::vector<int>* histogram) const {
  if (prog_ == NULL)
    return -1;
  return Fanout(prog_, histogram);
}
 
int RE2::ReverseProgramFanout(std::vector<int>* histogram) const {
  if (prog_ == NULL)
    return -1;
  Prog* prog = ReverseProg();
  if (prog == NULL)
    return -1;
  return Fanout(prog, histogram);
}
 
// Returns named_groups_, computing it if needed.
const std::map<std::string, int>& RE2::NamedCapturingGroups() const {
  std::call_once(named_groups_once_, [](const RE2* re) {
    if (re->suffix_regexp_ != NULL)
      re->named_groups_ = re->suffix_regexp_->NamedCaptures();
    if (re->named_groups_ == NULL)
      re->named_groups_ = empty_named_groups;
  }, this);
  return *named_groups_;
}
 
// Returns group_names_, computing it if needed.
const std::map<int, std::string>& RE2::CapturingGroupNames() const {
  std::call_once(group_names_once_, [](const RE2* re) {
    if (re->suffix_regexp_ != NULL)
      re->group_names_ = re->suffix_regexp_->CaptureNames();
    if (re->group_names_ == NULL)
      re->group_names_ = empty_group_names;
  }, this);
  return *group_names_;
}
 
/***** Convenience interfaces *****/
 
bool RE2::FullMatchN(const StringPiece& text, const RE2& re,
                     const Arg* const args[], int n) {
  return re.DoMatch(text, ANCHOR_BOTH, NULL, args, n);
}
 
bool RE2::PartialMatchN(const StringPiece& text, const RE2& re,
                        const Arg* const args[], int n) {
  return re.DoMatch(text, UNANCHORED, NULL, args, n);
}
 
bool RE2::ConsumeN(StringPiece* input, const RE2& re,
                   const Arg* const args[], int n) {
  size_t consumed;
  if (re.DoMatch(*input, ANCHOR_START, &consumed, args, n)) {
    input->remove_prefix(consumed);
    return true;
  } else {
    return false;
  }
}
 
bool RE2::FindAndConsumeN(StringPiece* input, const RE2& re,
                          const Arg* const args[], int n) {
  size_t consumed;
  if (re.DoMatch(*input, UNANCHORED, &consumed, args, n)) {
    input->remove_prefix(consumed);
    return true;
  } else {
    return false;
  }
}
 
bool RE2::Replace(std::string* str,
                  const RE2& re,
                  const StringPiece& rewrite) {
  StringPiece vec[kVecSize];
  int nvec = 1 + MaxSubmatch(rewrite);
  if (nvec > 1 + re.NumberOfCapturingGroups())
    return false;
  if (nvec > static_cast<int>(arraysize(vec)))
    return false;
  if (!re.Match(*str, 0, str->size(), UNANCHORED, vec, nvec))
    return false;
 
  std::string s;
  if (!re.Rewrite(&s, rewrite, vec, nvec))
    return false;
 
  assert(vec[0].data() >= str->data());
  assert(vec[0].data() + vec[0].size() <= str->data() + str->size());
  str->replace(vec[0].data() - str->data(), vec[0].size(), s);
  return true;
}
 
int RE2::GlobalReplace(std::string* str,
                       const RE2& re,
                       const StringPiece& rewrite) {
  StringPiece vec[kVecSize];
  int nvec = 1 + MaxSubmatch(rewrite);
  if (nvec > 1 + re.NumberOfCapturingGroups())
    return false;
  if (nvec > static_cast<int>(arraysize(vec)))
    return false;
 
  const char* p = str->data();
  const char* ep = p + str->size();
  const char* lastend = NULL;
  std::string out;
  int count = 0;
#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
  // Iterate just once when fuzzing. Otherwise, we easily get bogged down
  // and coverage is unlikely to improve despite significant expense.
  while (p == str->data()) {
#else
  while (p <= ep) {
#endif
    if (!re.Match(*str, static_cast<size_t>(p - str->data()),
                  str->size(), UNANCHORED, vec, nvec))
      break;
    if (p < vec[0].data())
      out.append(p, vec[0].data() - p);
    if (vec[0].data() == lastend && vec[0].empty()) {
      // Disallow empty match at end of last match: skip ahead.
      //
      // fullrune() takes int, not ptrdiff_t. However, it just looks
      // at the leading byte and treats any length >= 4 the same.
      if (re.options().encoding() == RE2::Options::EncodingUTF8 &&
          fullrune(p, static_cast<int>(std::min(ptrdiff_t{4}, ep - p)))) {
        // re is in UTF-8 mode and there is enough left of str
        // to allow us to advance by up to UTFmax bytes.
        Rune r;
        int n = chartorune(&r, p);
        // Some copies of chartorune have a bug that accepts
        // encodings of values in (10FFFF, 1FFFFF] as valid.
        if (r > Runemax) {
          n = 1;
          r = Runeerror;
        }
        if (!(n == 1 && r == Runeerror)) {  // no decoding error
          out.append(p, n);
          p += n;
          continue;
        }
      }
      // Most likely, re is in Latin-1 mode. If it is in UTF-8 mode,
      // we fell through from above and the GIGO principle applies.
      if (p < ep)
        out.append(p, 1);
      p++;
      continue;
    }
    re.Rewrite(&out, rewrite, vec, nvec);
    p = vec[0].data() + vec[0].size();
    lastend = p;
    count++;
  }
 
  if (count == 0)
    return 0;
 
  if (p < ep)
    out.append(p, ep - p);
  using std::swap;
  swap(out, *str);
  return count;
}
 
bool RE2::Extract(const StringPiece& text,
                  const RE2& re,
                  const StringPiece& rewrite,
                  std::string* out) {
  StringPiece vec[kVecSize];
  int nvec = 1 + MaxSubmatch(rewrite);
  if (nvec > 1 + re.NumberOfCapturingGroups())
    return false;
  if (nvec > static_cast<int>(arraysize(vec)))
    return false;
  if (!re.Match(text, 0, text.size(), UNANCHORED, vec, nvec))
    return false;
 
  out->clear();
  return re.Rewrite(out, rewrite, vec, nvec);
}
 
std::string RE2::QuoteMeta(const StringPiece& unquoted) {
  std::string result;
  result.reserve(unquoted.size() << 1);
 
  // Escape any ascii character not in [A-Za-z_0-9].
  //
  // Note that it's legal to escape a character even if it has no
  // special meaning in a regular expression -- so this function does
  // that.  (This also makes it identical to the perl function of the
  // same name except for the null-character special case;
  // see `perldoc -f quotemeta`.)
  for (size_t ii = 0; ii < unquoted.size(); ++ii) {
    // Note that using 'isalnum' here raises the benchmark time from
    // 32ns to 58ns:
    if ((unquoted[ii] < 'a' || unquoted[ii] > 'z') &&
        (unquoted[ii] < 'A' || unquoted[ii] > 'Z') &&
        (unquoted[ii] < '0' || unquoted[ii] > '9') &&
        unquoted[ii] != '_' &&
        // If this is the part of a UTF8 or Latin1 character, we need
        // to copy this byte without escaping.  Experimentally this is
        // what works correctly with the regexp library.
        !(unquoted[ii] & 128)) {
      if (unquoted[ii] == '\0') {  // Special handling for null chars.
        // Note that this special handling is not strictly required for RE2,
        // but this quoting is required for other regexp libraries such as
        // PCRE.
        // Can't use "\\0" since the next character might be a digit.
        result += "\\x00";
        continue;
      }
      result += '\\';
    }
    result += unquoted[ii];
  }
 
  return result;
}
 
bool RE2::PossibleMatchRange(std::string* min, std::string* max,
                             int maxlen) const {
  if (prog_ == NULL)
    return false;
 
  int n = static_cast<int>(prefix_.size());
  if (n > maxlen)
    n = maxlen;
 
  // Determine initial min max from prefix_ literal.
  *min = prefix_.substr(0, n);
  *max = prefix_.substr(0, n);
  if (prefix_foldcase_) {
    // prefix is ASCII lowercase; change *min to uppercase.
    for (int i = 0; i < n; i++) {
      char& c = (*min)[i];
      if ('a' <= c && c <= 'z')
        c += 'A' - 'a';
    }
  }
 
  // Add to prefix min max using PossibleMatchRange on regexp.
  std::string dmin, dmax;
  maxlen -= n;
  if (maxlen > 0 && prog_->PossibleMatchRange(&dmin, &dmax, maxlen)) {
    min->append(dmin);
    max->append(dmax);
  } else if (!max->empty()) {
    // prog_->PossibleMatchRange has failed us,
    // but we still have useful information from prefix_.
    // Round up *max to allow any possible suffix.
    PrefixSuccessor(max);
  } else {
    // Nothing useful.
    *min = "";
    *max = "";
    return false;
  }
 
  return true;
}
 
// Avoid possible locale nonsense in standard strcasecmp.
// The string a is known to be all lowercase.
static int ascii_strcasecmp(const char* a, const char* b, size_t len) {
  const char* ae = a + len;
 
  for (; a < ae; a++, b++) {
    uint8_t x = *a;
    uint8_t y = *b;
    if ('A' <= y && y <= 'Z')
      y += 'a' - 'A';
    if (x != y)
      return x - y;
  }
  return 0;
}
 
 
/***** Actual matching and rewriting code *****/
 
bool RE2::Match(const StringPiece& text,
                size_t startpos,
                size_t endpos,
                Anchor re_anchor,
                StringPiece* submatch,
                int nsubmatch) const {
  if (!ok()) {
    if (options_.log_errors())
      LOG(ERROR) << "Invalid RE2: " << *error_;
    return false;
  }
 
  if (startpos > endpos || endpos > text.size()) {
    if (options_.log_errors())
      LOG(ERROR) << "RE2: invalid startpos, endpos pair. ["
                 << "startpos: " << startpos << ", "
                 << "endpos: " << endpos << ", "
                 << "text size: " << text.size() << "]";
    return false;
  }
 
  StringPiece subtext = text;
  subtext.remove_prefix(startpos);
  subtext.remove_suffix(text.size() - endpos);
 
  // Use DFAs to find exact location of match, filter out non-matches.
 
  // Don't ask for the location if we won't use it.
  // SearchDFA can do extra optimizations in that case.
  StringPiece match;
  StringPiece* matchp = &match;
  if (nsubmatch == 0)
    matchp = NULL;
 
  int ncap = 1 + NumberOfCapturingGroups();
  if (ncap > nsubmatch)
    ncap = nsubmatch;
 
  // If the regexp is anchored explicitly, must not be in middle of text.
  if (prog_->anchor_start() && startpos != 0)
    return false;
  if (prog_->anchor_end() && endpos != text.size())
    return false;
 
  // If the regexp is anchored explicitly, update re_anchor
  // so that we can potentially fall into a faster case below.
  if (prog_->anchor_start() && prog_->anchor_end())
    re_anchor = ANCHOR_BOTH;
  else if (prog_->anchor_start() && re_anchor != ANCHOR_BOTH)
    re_anchor = ANCHOR_START;
 
  // Check for the required prefix, if any.
  size_t prefixlen = 0;
  if (!prefix_.empty()) {
    if (startpos != 0)
      return false;
    prefixlen = prefix_.size();
    if (prefixlen > subtext.size())
      return false;
    if (prefix_foldcase_) {
      if (ascii_strcasecmp(&prefix_[0], subtext.data(), prefixlen) != 0)
        return false;
    } else {
      if (memcmp(&prefix_[0], subtext.data(), prefixlen) != 0)
        return false;
    }
    subtext.remove_prefix(prefixlen);
    // If there is a required prefix, the anchor must be at least ANCHOR_START.
    if (re_anchor != ANCHOR_BOTH)
      re_anchor = ANCHOR_START;
  }
 
  Prog::Anchor anchor = Prog::kUnanchored;
  Prog::MatchKind kind = Prog::kFirstMatch;
  if (options_.longest_match())
    kind = Prog::kLongestMatch;
 
  bool can_one_pass = (is_one_pass_ && ncap <= Prog::kMaxOnePassCapture);
 
  // BitState allocates a bitmap of size prog_->list_count() * text.size().
  // It also allocates a stack of 3-word structures which could potentially
  // grow as large as prog_->list_count() * text.size(), but in practice is
  // much smaller.
  const int kMaxBitStateBitmapSize = 256*1024;  // bitmap size <= max (bits)
  bool can_bit_state = prog_->CanBitState();
  size_t bit_state_text_max = kMaxBitStateBitmapSize / prog_->list_count();
 
#ifdef RE2_HAVE_THREAD_LOCAL
  hooks::context = this;
#endif
  bool dfa_failed = false;
  bool skipped_test = false;
  switch (re_anchor) {
    default:
      LOG(DFATAL) << "Unexpected re_anchor value: " << re_anchor;
      return false;
 
    case UNANCHORED: {
      if (prog_->anchor_end()) {
        // This is a very special case: we don't need the forward DFA because
        // we already know where the match must end! Instead, the reverse DFA
        // can say whether there is a match and (optionally) where it starts.
        Prog* prog = ReverseProg();
        if (prog == NULL) {
          // Fall back to NFA below.
          skipped_test = true;
          break;
        }
        if (!prog->SearchDFA(subtext, text, Prog::kAnchored,
                             Prog::kLongestMatch, matchp, &dfa_failed, NULL)) {
          if (dfa_failed) {
            if (options_.log_errors())
              LOG(ERROR) << "DFA out of memory: "
                         << "pattern length " << pattern_.size() << ", "
                         << "program size " << prog->size() << ", "
                         << "list count " << prog->list_count() << ", "
                         << "bytemap range " << prog->bytemap_range();
            // Fall back to NFA below.
            skipped_test = true;
            break;
          }
          return false;
        }
        if (matchp == NULL)  // Matched.  Don't care where.
          return true;
        break;
      }
 
      if (!prog_->SearchDFA(subtext, text, anchor, kind,
                            matchp, &dfa_failed, NULL)) {
        if (dfa_failed) {
          if (options_.log_errors())
            LOG(ERROR) << "DFA out of memory: "
                       << "pattern length " << pattern_.size() << ", "
                       << "program size " << prog_->size() << ", "
                       << "list count " << prog_->list_count() << ", "
                       << "bytemap range " << prog_->bytemap_range();
          // Fall back to NFA below.
          skipped_test = true;
          break;
        }
        return false;
      }
      if (matchp == NULL)  // Matched.  Don't care where.
        return true;
      // SearchDFA set match.end() but didn't know where the
      // match started.  Run the regexp backward from match.end()
      // to find the longest possible match -- that's where it started.
      Prog* prog = ReverseProg();
      if (prog == NULL) {
        // Fall back to NFA below.
        skipped_test = true;
        break;
      }
      if (!prog->SearchDFA(match, text, Prog::kAnchored,
                           Prog::kLongestMatch, &match, &dfa_failed, NULL)) {
        if (dfa_failed) {
          if (options_.log_errors())
            LOG(ERROR) << "DFA out of memory: "
                       << "pattern length " << pattern_.size() << ", "
                       << "program size " << prog->size() << ", "
                       << "list count " << prog->list_count() << ", "
                       << "bytemap range " << prog->bytemap_range();
          // Fall back to NFA below.
          skipped_test = true;
          break;
        }
        if (options_.log_errors())
          LOG(ERROR) << "SearchDFA inconsistency";
        return false;
      }
      break;
    }
 
    case ANCHOR_BOTH:
    case ANCHOR_START:
      if (re_anchor == ANCHOR_BOTH)
        kind = Prog::kFullMatch;
      anchor = Prog::kAnchored;
 
      // If only a small amount of text and need submatch
      // information anyway and we're going to use OnePass or BitState
      // to get it, we might as well not even bother with the DFA:
      // OnePass or BitState will be fast enough.
      // On tiny texts, OnePass outruns even the DFA, and
      // it doesn't have the shared state and occasional mutex that
      // the DFA does.
      if (can_one_pass && text.size() <= 4096 &&
          (ncap > 1 || text.size() <= 16)) {
        skipped_test = true;
        break;
      }
      if (can_bit_state && text.size() <= bit_state_text_max && ncap > 1) {
        skipped_test = true;
        break;
      }
      if (!prog_->SearchDFA(subtext, text, anchor, kind,
                            &match, &dfa_failed, NULL)) {
        if (dfa_failed) {
          if (options_.log_errors())
            LOG(ERROR) << "DFA out of memory: "
                       << "pattern length " << pattern_.size() << ", "
                       << "program size " << prog_->size() << ", "
                       << "list count " << prog_->list_count() << ", "
                       << "bytemap range " << prog_->bytemap_range();
          // Fall back to NFA below.
          skipped_test = true;
          break;
        }
        return false;
      }
      break;
  }
 
  if (!skipped_test && ncap <= 1) {
    // We know exactly where it matches.  That's enough.
    if (ncap == 1)
      submatch[0] = match;
  } else {
    StringPiece subtext1;
    if (skipped_test) {
      // DFA ran out of memory or was skipped:
      // need to search in entire original text.
      subtext1 = subtext;
    } else {
      // DFA found the exact match location:
      // let NFA run an anchored, full match search
      // to find submatch locations.
      subtext1 = match;
      anchor = Prog::kAnchored;
      kind = Prog::kFullMatch;
    }
 
    if (can_one_pass && anchor != Prog::kUnanchored) {
      if (!prog_->SearchOnePass(subtext1, text, anchor, kind, submatch, ncap)) {
        if (!skipped_test && options_.log_errors())
          LOG(ERROR) << "SearchOnePass inconsistency";
        return false;
      }
    } else if (can_bit_state && subtext1.size() <= bit_state_text_max) {
      if (!prog_->SearchBitState(subtext1, text, anchor,
                                 kind, submatch, ncap)) {
        if (!skipped_test && options_.log_errors())
          LOG(ERROR) << "SearchBitState inconsistency";
        return false;
      }
    } else {
      if (!prog_->SearchNFA(subtext1, text, anchor, kind, submatch, ncap)) {
        if (!skipped_test && options_.log_errors())
          LOG(ERROR) << "SearchNFA inconsistency";
        return false;
      }
    }
  }
 
  // Adjust overall match for required prefix that we stripped off.
  if (prefixlen > 0 && nsubmatch > 0)
    submatch[0] = StringPiece(submatch[0].data() - prefixlen,
                              submatch[0].size() + prefixlen);
 
  // Zero submatches that don't exist in the regexp.
  for (int i = ncap; i < nsubmatch; i++)
    submatch[i] = StringPiece();
  return true;
}
 
// Internal matcher - like Match() but takes Args not StringPieces.
bool RE2::DoMatch(const StringPiece& text,
                  Anchor re_anchor,
                  size_t* consumed,
                  const Arg* const* args,
                  int n) const {
  if (!ok()) {
    if (options_.log_errors())
      LOG(ERROR) << "Invalid RE2: " << *error_;
    return false;
  }
 
  if (NumberOfCapturingGroups() < n) {
    // RE has fewer capturing groups than number of Arg pointers passed in.
    return false;
  }
 
  // Count number of capture groups needed.
  int nvec;
  if (n == 0 && consumed == NULL)
    nvec = 0;
  else
    nvec = n+1;
 
  StringPiece* vec;
  StringPiece stkvec[kVecSize];
  StringPiece* heapvec = NULL;
 
  if (nvec <= static_cast<int>(arraysize(stkvec))) {
    vec = stkvec;
  } else {
    vec = new StringPiece[nvec];
    heapvec = vec;
  }
 
  if (!Match(text, 0, text.size(), re_anchor, vec, nvec)) {
    delete[] heapvec;
    return false;
  }
 
  if (consumed != NULL)
    *consumed = static_cast<size_t>(vec[0].end() - text.begin());
 
  if (n == 0 || args == NULL) {
    // We are not interested in results
    delete[] heapvec;
    return true;
  }
 
  // If we got here, we must have matched the whole pattern.
  for (int i = 0; i < n; i++) {
    const StringPiece& s = vec[i+1];
    if (!args[i]->Parse(s.data(), s.size())) {
      // TODO: Should we indicate what the error was?
      delete[] heapvec;
      return false;
    }
  }
 
  delete[] heapvec;
  return true;
}
 
// Checks that the rewrite string is well-formed with respect to this
// regular expression.
bool RE2::CheckRewriteString(const StringPiece& rewrite,
                             std::string* error) const {
  int max_token = -1;
  for (const char *s = rewrite.data(), *end = s + rewrite.size();
       s < end; s++) {
    int c = *s;
    if (c != '\\') {
      continue;
    }
    if (++s == end) {
      *error = "Rewrite schema error: '\\' not allowed at end.";
      return false;
    }
    c = *s;
    if (c == '\\') {
      continue;
    }
    if (!isdigit(c)) {
      *error = "Rewrite schema error: "
               "'\\' must be followed by a digit or '\\'.";
      return false;
    }
    int n = (c - '0');
    if (max_token < n) {
      max_token = n;
    }
  }
 
  if (max_token > NumberOfCapturingGroups()) {
    *error = StringPrintf(
        "Rewrite schema requests %d matches, but the regexp only has %d "
        "parenthesized subexpressions.",
        max_token, NumberOfCapturingGroups());
    return false;
  }
  return true;
}
 
// Returns the maximum submatch needed for the rewrite to be done by Replace().
// E.g. if rewrite == "foo \\2,\\1", returns 2.
int RE2::MaxSubmatch(const StringPiece& rewrite) {
  int max = 0;
  for (const char *s = rewrite.data(), *end = s + rewrite.size();
       s < end; s++) {
    if (*s == '\\') {
      s++;
      int c = (s < end) ? *s : -1;
      if (isdigit(c)) {
        int n = (c - '0');
        if (n > max)
          max = n;
      }
    }
  }
  return max;
}
 
// Append the "rewrite" string, with backslash subsitutions from "vec",
// to string "out".
bool RE2::Rewrite(std::string* out,
                  const StringPiece& rewrite,
                  const StringPiece* vec,
                  int veclen) const {
  for (const char *s = rewrite.data(), *end = s + rewrite.size();
       s < end; s++) {
    if (*s != '\\') {
      out->push_back(*s);
      continue;
    }
    s++;
    int c = (s < end) ? *s : -1;
    if (isdigit(c)) {
      int n = (c - '0');
      if (n >= veclen) {
        if (options_.log_errors()) {
          LOG(ERROR) << "invalid substitution \\" << n
                     << " from " << veclen << " groups";
        }
        return false;
      }
      StringPiece snip = vec[n];
      if (!snip.empty())
        out->append(snip.data(), snip.size());
    } else if (c == '\\') {
      out->push_back('\\');
    } else {
      if (options_.log_errors())
        LOG(ERROR) << "invalid rewrite pattern: " << rewrite.data();
      return false;
    }
  }
  return true;
}
 
/***** Parsers for various types *****/
 
namespace re2_internal {
 
template <>
bool Parse(const char* str, size_t n, void* dest) {
  // We fail if somebody asked us to store into a non-NULL void* pointer
  return (dest == NULL);
}
 
template <>
bool Parse(const char* str, size_t n, std::string* dest) {
  if (dest == NULL) return true;
  dest->assign(str, n);
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, StringPiece* dest) {
  if (dest == NULL) return true;
  *dest = StringPiece(str, n);
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, char* dest) {
  if (n != 1) return false;
  if (dest == NULL) return true;
  *dest = str[0];
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, signed char* dest) {
  if (n != 1) return false;
  if (dest == NULL) return true;
  *dest = str[0];
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, unsigned char* dest) {
  if (n != 1) return false;
  if (dest == NULL) return true;
  *dest = str[0];
  return true;
}
 
// Largest number spec that we are willing to parse
static const int kMaxNumberLength = 32;
 
// REQUIRES "buf" must have length at least nbuf.
// Copies "str" into "buf" and null-terminates.
// Overwrites *np with the new length.
static const char* TerminateNumber(char* buf, size_t nbuf, const char* str,
                                   size_t* np, bool accept_spaces) {
  size_t n = *np;
  if (n == 0) return "";
  if (n > 0 && isspace(*str)) {
    // We are less forgiving than the strtoxxx() routines and do not
    // allow leading spaces. We do allow leading spaces for floats.
    if (!accept_spaces) {
      return "";
    }
    while (n > 0 && isspace(*str)) {
      n--;
      str++;
    }
  }
 
  // Although buf has a fixed maximum size, we can still handle
  // arbitrarily large integers correctly by omitting leading zeros.
  // (Numbers that are still too long will be out of range.)
  // Before deciding whether str is too long,
  // remove leading zeros with s/000+/00/.
  // Leaving the leading two zeros in place means that
  // we don't change 0000x123 (invalid) into 0x123 (valid).
  // Skip over leading - before replacing.
  bool neg = false;
  if (n >= 1 && str[0] == '-') {
    neg = true;
    n--;
    str++;
  }
 
  if (n >= 3 && str[0] == '0' && str[1] == '0') {
    while (n >= 3 && str[2] == '0') {
      n--;
      str++;
    }
  }
 
  if (neg) {  // make room in buf for -
    n++;
    str--;
  }
 
  if (n > nbuf-1) return "";
 
  memmove(buf, str, n);
  if (neg) {
    buf[0] = '-';
  }
  buf[n] = '\0';
  *np = n;
  return buf;
}
 
template <>
bool Parse(const char* str, size_t n, float* dest) {
  if (n == 0) return false;
  static const int kMaxLength = 200;
  char buf[kMaxLength+1];
  str = TerminateNumber(buf, sizeof buf, str, &n, true);
  char* end;
  errno = 0;
  float r = strtof(str, &end);
  if (end != str + n) return false;   // Leftover junk
  if (errno) return false;
  if (dest == NULL) return true;
  *dest = r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, double* dest) {
  if (n == 0) return false;
  static const int kMaxLength = 200;
  char buf[kMaxLength+1];
  str = TerminateNumber(buf, sizeof buf, str, &n, true);
  char* end;
  errno = 0;
  double r = strtod(str, &end);
  if (end != str + n) return false;   // Leftover junk
  if (errno) return false;
  if (dest == NULL) return true;
  *dest = r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, long* dest, int radix) {
  if (n == 0) return false;
  char buf[kMaxNumberLength+1];
  str = TerminateNumber(buf, sizeof buf, str, &n, false);
  char* end;
  errno = 0;
  long r = strtol(str, &end, radix);
  if (end != str + n) return false;   // Leftover junk
  if (errno) return false;
  if (dest == NULL) return true;
  *dest = r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, unsigned long* dest, int radix) {
  if (n == 0) return false;
  char buf[kMaxNumberLength+1];
  str = TerminateNumber(buf, sizeof buf, str, &n, false);
  if (str[0] == '-') {
    // strtoul() will silently accept negative numbers and parse
    // them.  This module is more strict and treats them as errors.
    return false;
  }
 
  char* end;
  errno = 0;
  unsigned long r = strtoul(str, &end, radix);
  if (end != str + n) return false;   // Leftover junk
  if (errno) return false;
  if (dest == NULL) return true;
  *dest = r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, short* dest, int radix) {
  long r;
  if (!Parse(str, n, &r, radix)) return false;  // Could not parse
  if ((short)r != r) return false;              // Out of range
  if (dest == NULL) return true;
  *dest = (short)r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, unsigned short* dest, int radix) {
  unsigned long r;
  if (!Parse(str, n, &r, radix)) return false;  // Could not parse
  if ((unsigned short)r != r) return false;     // Out of range
  if (dest == NULL) return true;
  *dest = (unsigned short)r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, int* dest, int radix) {
  long r;
  if (!Parse(str, n, &r, radix)) return false;  // Could not parse
  if ((int)r != r) return false;                // Out of range
  if (dest == NULL) return true;
  *dest = (int)r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, unsigned int* dest, int radix) {
  unsigned long r;
  if (!Parse(str, n, &r, radix)) return false;  // Could not parse
  if ((unsigned int)r != r) return false;       // Out of range
  if (dest == NULL) return true;
  *dest = (unsigned int)r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, long long* dest, int radix) {
  if (n == 0) return false;
  char buf[kMaxNumberLength+1];
  str = TerminateNumber(buf, sizeof buf, str, &n, false);
  char* end;
  errno = 0;
  long long r = strtoll(str, &end, radix);
  if (end != str + n) return false;   // Leftover junk
  if (errno) return false;
  if (dest == NULL) return true;
  *dest = r;
  return true;
}
 
template <>
bool Parse(const char* str, size_t n, unsigned long long* dest, int radix) {
  if (n == 0) return false;
  char buf[kMaxNumberLength+1];
  str = TerminateNumber(buf, sizeof buf, str, &n, false);
  if (str[0] == '-') {
    // strtoull() will silently accept negative numbers and parse
    // them.  This module is more strict and treats them as errors.
    return false;
  }
  char* end;
  errno = 0;
  unsigned long long r = strtoull(str, &end, radix);
  if (end != str + n) return false;   // Leftover junk
  if (errno) return false;
  if (dest == NULL) return true;
  *dest = r;
  return true;
}
 
}  // namespace re2_internal
 
namespace hooks {
 
#ifdef RE2_HAVE_THREAD_LOCAL
thread_local const RE2* context = NULL;
#endif
 
template <typename T>
union Hook {
  void Store(T* cb) { cb_.store(cb, std::memory_order_release); }
  T* Load() const { return cb_.load(std::memory_order_acquire); }
 
#if !defined(__clang__) && defined(_MSC_VER)
  // Citing https://github.com/protocolbuffers/protobuf/pull/4777 as precedent,
  // this is a gross hack to make std::atomic<T*> constant-initialized on MSVC.
  static_assert(ATOMIC_POINTER_LOCK_FREE == 2,
                "std::atomic<T*> must be always lock-free");
  T* cb_for_constinit_;
#endif
 
  std::atomic<T*> cb_;
};
 
template <typename T>
static void DoNothing(const T&) {}
 
#define DEFINE_HOOK(type, name)                                       \
  static Hook<type##Callback> name##_hook = {{&DoNothing<type>}};     \
  void Set##type##Hook(type##Callback* cb) { name##_hook.Store(cb); } \
  type##Callback* Get##type##Hook() { return name##_hook.Load(); }
 
DEFINE_HOOK(DFAStateCacheReset, dfa_state_cache_reset)
DEFINE_HOOK(DFASearchFailure, dfa_search_failure)
 
#undef DEFINE_HOOK
 
}  // namespace hooks
 
}  // namespace re2