Home | History | Annotate | Download | only in re2
      1 // Copyright 2008 The RE2 Authors.  All Rights Reserved.
      2 // Use of this source code is governed by a BSD-style
      3 // license that can be found in the LICENSE file.
      4 
      5 // A DFA (deterministic finite automaton)-based regular expression search.
      6 //
      7 // The DFA search has two main parts: the construction of the automaton,
      8 // which is represented by a graph of State structures, and the execution
      9 // of the automaton over a given input string.
     10 //
     11 // The basic idea is that the State graph is constructed so that the
     12 // execution can simply start with a state s, and then for each byte c in
     13 // the input string, execute "s = s->next[c]", checking at each point whether
     14 // the current s represents a matching state.
     15 //
     16 // The simple explanation just given does convey the essence of this code,
     17 // but it omits the details of how the State graph gets constructed as well
     18 // as some performance-driven optimizations to the execution of the automaton.
     19 // All these details are explained in the comments for the code following
     20 // the definition of class DFA.
     21 //
     22 // See http://swtch.com/~rsc/regexp/ for a very bare-bones equivalent.
     23 
     24 #include "re2/prog.h"
     25 #include "re2/stringpiece.h"
     26 #include "util/atomicops.h"
     27 #include "util/flags.h"
     28 #include "util/sparse_set.h"
     29 
     30 DEFINE_bool(re2_dfa_bail_when_slow, true,
     31             "Whether the RE2 DFA should bail out early "
     32             "if the NFA would be faster (for testing).");
     33 
     34 namespace re2 {
     35 
     36 #if !defined(__linux__)  /* only Linux seems to have memrchr */
     37 static void* memrchr(const void* s, int c, size_t n) {
     38   const unsigned char* p = (const unsigned char*)s;
     39   for (p += n; n > 0; n--)
     40     if (*--p == c)
     41       return (void*)p;
     42 
     43   return NULL;
     44 }
     45 #endif
     46 
     47 // Changing this to true compiles in prints that trace execution of the DFA.
     48 // Generates a lot of output -- only useful for debugging.
     49 static const bool DebugDFA = false;
     50 
     51 // A DFA implementation of a regular expression program.
     52 // Since this is entirely a forward declaration mandated by C++,
     53 // some of the comments here are better understood after reading
     54 // the comments in the sections that follow the DFA definition.
     55 class DFA {
     56  public:
     57   DFA(Prog* prog, Prog::MatchKind kind, int64 max_mem);
     58   ~DFA();
     59   bool ok() const { return !init_failed_; }
     60   Prog::MatchKind kind() { return kind_; }
     61 
     62   // Searches for the regular expression in text, which is considered
     63   // as a subsection of context for the purposes of interpreting flags
     64   // like ^ and $ and \A and \z.
     65   // Returns whether a match was found.
     66   // If a match is found, sets *ep to the end point of the best match in text.
     67   // If "anchored", the match must begin at the start of text.
     68   // If "want_earliest_match", the match that ends first is used, not
     69   //   necessarily the best one.
     70   // If "run_forward" is true, the DFA runs from text.begin() to text.end().
     71   //   If it is false, the DFA runs from text.end() to text.begin(),
     72   //   returning the leftmost end of the match instead of the rightmost one.
     73   // If the DFA cannot complete the search (for example, if it is out of
     74   //   memory), it sets *failed and returns false.
     75   bool Search(const StringPiece& text, const StringPiece& context,
     76               bool anchored, bool want_earliest_match, bool run_forward,
     77               bool* failed, const char** ep, vector<int>* matches);
     78 
     79   // Builds out all states for the entire DFA.  FOR TESTING ONLY
     80   // Returns number of states.
     81   int BuildAllStates();
     82 
     83   // Computes min and max for matching strings.  Won't return strings
     84   // bigger than maxlen.
     85   bool PossibleMatchRange(string* min, string* max, int maxlen);
     86 
     87   // These data structures are logically private, but C++ makes it too
     88   // difficult to mark them as such.
     89   class Workq;
     90   class RWLocker;
     91   class StateSaver;
     92 
     93   // A single DFA state.  The DFA is represented as a graph of these
     94   // States, linked by the next_ pointers.  If in state s and reading
     95   // byte c, the next state should be s->next_[c].
     96   struct State {
     97     inline bool IsMatch() const { return flag_ & kFlagMatch; }
     98     void SaveMatch(vector<int>* v);
     99 
    100     int* inst_;         // Instruction pointers in the state.
    101     int ninst_;         // # of inst_ pointers.
    102     uint flag_;         // Empty string bitfield flags in effect on the way
    103                         // into this state, along with kFlagMatch if this
    104                         // is a matching state.
    105     State** next_;      // Outgoing arrows from State,
    106                         // one per input byte class
    107   };
    108 
    109   enum {
    110     kByteEndText = 256,         // imaginary byte at end of text
    111 
    112     kFlagEmptyMask = 0xFFF,     // State.flag_: bits holding kEmptyXXX flags
    113     kFlagMatch = 0x1000,        // State.flag_: this is a matching state
    114     kFlagLastWord = 0x2000,     // State.flag_: last byte was a word char
    115     kFlagNeedShift = 16,        // needed kEmpty bits are or'ed in shifted left
    116   };
    117 
    118 #ifndef STL_MSVC
    119   // STL function structures for use with unordered_set.
    120   struct StateEqual {
    121     bool operator()(const State* a, const State* b) const {
    122       if (a == b)
    123         return true;
    124       if (a == NULL || b == NULL)
    125         return false;
    126       if (a->ninst_ != b->ninst_)
    127         return false;
    128       if (a->flag_ != b->flag_)
    129         return false;
    130       for (int i = 0; i < a->ninst_; i++)
    131         if (a->inst_[i] != b->inst_[i])
    132           return false;
    133       return true;  // they're equal
    134     }
    135   };
    136 #endif  // STL_MSVC
    137   struct StateHash {
    138     size_t operator()(const State* a) const {
    139       if (a == NULL)
    140         return 0;
    141       const char* s = reinterpret_cast<const char*>(a->inst_);
    142       int len = a->ninst_ * sizeof a->inst_[0];
    143       if (sizeof(size_t) == sizeof(uint32))
    144         return Hash32StringWithSeed(s, len, a->flag_);
    145       else
    146         return Hash64StringWithSeed(s, len, a->flag_);
    147     }
    148 #ifdef STL_MSVC
    149     // Less than operator.
    150     bool operator()(const State* a, const State* b) const {
    151       if (a == b)
    152         return false;
    153       if (a == NULL || b == NULL)
    154         return a == NULL;
    155       if (a->ninst_ != b->ninst_)
    156         return a->ninst_ < b->ninst_;
    157       if (a->flag_ != b->flag_)
    158         return a->flag_ < b->flag_;
    159       for (int i = 0; i < a->ninst_; ++i)
    160         if (a->inst_[i] != b->inst_[i])
    161           return a->inst_[i] < b->inst_[i];
    162       return false;  // they're equal
    163     }
    164     // The two public members are required by msvc. 4 and 8 are default values.
    165     // Reference: http://msdn.microsoft.com/en-us/library/1s1byw77.aspx
    166     static const size_t bucket_size = 4;
    167     static const size_t min_buckets = 8;
    168 #endif  // STL_MSVC
    169   };
    170 
    171 #ifdef STL_MSVC
    172   typedef unordered_set<State*, StateHash> StateSet;
    173 #else  // !STL_MSVC
    174   typedef unordered_set<State*, StateHash, StateEqual> StateSet;
    175 #endif  // STL_MSVC
    176 
    177 
    178  private:
    179   // Special "firstbyte" values for a state.  (Values >= 0 denote actual bytes.)
    180   enum {
    181     kFbUnknown = -1,   // No analysis has been performed.
    182     kFbMany = -2,      // Many bytes will lead out of this state.
    183     kFbNone = -3,      // No bytes lead out of this state.
    184   };
    185 
    186   enum {
    187     // Indices into start_ for unanchored searches.
    188     // Add kStartAnchored for anchored searches.
    189     kStartBeginText = 0,          // text at beginning of context
    190     kStartBeginLine = 2,          // text at beginning of line
    191     kStartAfterWordChar = 4,      // text follows a word character
    192     kStartAfterNonWordChar = 6,   // text follows non-word character
    193     kMaxStart = 8,
    194 
    195     kStartAnchored = 1,
    196   };
    197 
    198   // Resets the DFA State cache, flushing all saved State* information.
    199   // Releases and reacquires cache_mutex_ via cache_lock, so any
    200   // State* existing before the call are not valid after the call.
    201   // Use a StateSaver to preserve important states across the call.
    202   // cache_mutex_.r <= L < mutex_
    203   // After: cache_mutex_.w <= L < mutex_
    204   void ResetCache(RWLocker* cache_lock);
    205 
    206   // Looks up and returns the State corresponding to a Workq.
    207   // L >= mutex_
    208   State* WorkqToCachedState(Workq* q, uint flag);
    209 
    210   // Looks up and returns a State matching the inst, ninst, and flag.
    211   // L >= mutex_
    212   State* CachedState(int* inst, int ninst, uint flag);
    213 
    214   // Clear the cache entirely.
    215   // Must hold cache_mutex_.w or be in destructor.
    216   void ClearCache();
    217 
    218   // Converts a State into a Workq: the opposite of WorkqToCachedState.
    219   // L >= mutex_
    220   static void StateToWorkq(State* s, Workq* q);
    221 
    222   // Runs a State on a given byte, returning the next state.
    223   State* RunStateOnByteUnlocked(State*, int);  // cache_mutex_.r <= L < mutex_
    224   State* RunStateOnByte(State*, int);          // L >= mutex_
    225 
    226   // Runs a Workq on a given byte followed by a set of empty-string flags,
    227   // producing a new Workq in nq.  If a match instruction is encountered,
    228   // sets *ismatch to true.
    229   // L >= mutex_
    230   void RunWorkqOnByte(Workq* q, Workq* nq,
    231                              int c, uint flag, bool* ismatch,
    232                              Prog::MatchKind kind,
    233                              int new_byte_loop);
    234 
    235   // Runs a Workq on a set of empty-string flags, producing a new Workq in nq.
    236   // L >= mutex_
    237   void RunWorkqOnEmptyString(Workq* q, Workq* nq, uint flag);
    238 
    239   // Adds the instruction id to the Workq, following empty arrows
    240   // according to flag.
    241   // L >= mutex_
    242   void AddToQueue(Workq* q, int id, uint flag);
    243 
    244   // For debugging, returns a text representation of State.
    245   static string DumpState(State* state);
    246 
    247   // For debugging, returns a text representation of a Workq.
    248   static string DumpWorkq(Workq* q);
    249 
    250   // Search parameters
    251   struct SearchParams {
    252     SearchParams(const StringPiece& text, const StringPiece& context,
    253                  RWLocker* cache_lock)
    254       : text(text), context(context),
    255         anchored(false),
    256         want_earliest_match(false),
    257         run_forward(false),
    258         start(NULL),
    259         firstbyte(kFbUnknown),
    260         cache_lock(cache_lock),
    261         failed(false),
    262         ep(NULL),
    263         matches(NULL) { }
    264 
    265     StringPiece text;
    266     StringPiece context;
    267     bool anchored;
    268     bool want_earliest_match;
    269     bool run_forward;
    270     State* start;
    271     int firstbyte;
    272     RWLocker *cache_lock;
    273     bool failed;     // "out" parameter: whether search gave up
    274     const char* ep;  // "out" parameter: end pointer for match
    275     vector<int>* matches;
    276 
    277    private:
    278     DISALLOW_EVIL_CONSTRUCTORS(SearchParams);
    279   };
    280 
    281   // Before each search, the parameters to Search are analyzed by
    282   // AnalyzeSearch to determine the state in which to start and the
    283   // "firstbyte" for that state, if any.
    284   struct StartInfo {
    285     StartInfo() : start(NULL), firstbyte(kFbUnknown) { }
    286     State* start;
    287     volatile int firstbyte;
    288   };
    289 
    290   // Fills in params->start and params->firstbyte using
    291   // the other search parameters.  Returns true on success,
    292   // false on failure.
    293   // cache_mutex_.r <= L < mutex_
    294   bool AnalyzeSearch(SearchParams* params);
    295   bool AnalyzeSearchHelper(SearchParams* params, StartInfo* info, uint flags);
    296 
    297   // The generic search loop, inlined to create specialized versions.
    298   // cache_mutex_.r <= L < mutex_
    299   // Might unlock and relock cache_mutex_ via params->cache_lock.
    300   inline bool InlinedSearchLoop(SearchParams* params,
    301                                 bool have_firstbyte,
    302                                 bool want_earliest_match,
    303                                 bool run_forward);
    304 
    305   // The specialized versions of InlinedSearchLoop.  The three letters
    306   // at the ends of the name denote the true/false values used as the
    307   // last three parameters of InlinedSearchLoop.
    308   // cache_mutex_.r <= L < mutex_
    309   // Might unlock and relock cache_mutex_ via params->cache_lock.
    310   bool SearchFFF(SearchParams* params);
    311   bool SearchFFT(SearchParams* params);
    312   bool SearchFTF(SearchParams* params);
    313   bool SearchFTT(SearchParams* params);
    314   bool SearchTFF(SearchParams* params);
    315   bool SearchTFT(SearchParams* params);
    316   bool SearchTTF(SearchParams* params);
    317   bool SearchTTT(SearchParams* params);
    318 
    319   // The main search loop: calls an appropriate specialized version of
    320   // InlinedSearchLoop.
    321   // cache_mutex_.r <= L < mutex_
    322   // Might unlock and relock cache_mutex_ via params->cache_lock.
    323   bool FastSearchLoop(SearchParams* params);
    324 
    325   // For debugging, a slow search loop that calls InlinedSearchLoop
    326   // directly -- because the booleans passed are not constants, the
    327   // loop is not specialized like the SearchFFF etc. versions, so it
    328   // runs much more slowly.  Useful only for debugging.
    329   // cache_mutex_.r <= L < mutex_
    330   // Might unlock and relock cache_mutex_ via params->cache_lock.
    331   bool SlowSearchLoop(SearchParams* params);
    332 
    333   // Looks up bytes in bytemap_ but handles case c == kByteEndText too.
    334   int ByteMap(int c) {
    335     if (c == kByteEndText)
    336       return prog_->bytemap_range();
    337     return prog_->bytemap()[c];
    338   }
    339 
    340   // Constant after initialization.
    341   Prog* prog_;              // The regular expression program to run.
    342   Prog::MatchKind kind_;    // The kind of DFA.
    343   int start_unanchored_;  // start of unanchored program
    344   bool init_failed_;        // initialization failed (out of memory)
    345 
    346   Mutex mutex_;  // mutex_ >= cache_mutex_.r
    347 
    348   // Scratch areas, protected by mutex_.
    349   Workq* q0_;             // Two pre-allocated work queues.
    350   Workq* q1_;
    351   int* astack_;         // Pre-allocated stack for AddToQueue
    352   int nastack_;
    353 
    354   // State* cache.  Many threads use and add to the cache simultaneously,
    355   // holding cache_mutex_ for reading and mutex_ (above) when adding.
    356   // If the cache fills and needs to be discarded, the discarding is done
    357   // while holding cache_mutex_ for writing, to avoid interrupting other
    358   // readers.  Any State* pointers are only valid while cache_mutex_
    359   // is held.
    360   Mutex cache_mutex_;
    361   int64 mem_budget_;       // Total memory budget for all States.
    362   int64 state_budget_;     // Amount of memory remaining for new States.
    363   StateSet state_cache_;   // All States computed so far.
    364   StartInfo start_[kMaxStart];
    365   bool cache_warned_;      // have printed to LOG(INFO) about the cache
    366 };
    367 
    368 // Shorthand for casting to uint8*.
    369 static inline const uint8* BytePtr(const void* v) {
    370   return reinterpret_cast<const uint8*>(v);
    371 }
    372 
    373 // Work queues
    374 
    375 // Marks separate thread groups of different priority
    376 // in the work queue when in leftmost-longest matching mode.
    377 #define Mark (-1)
    378 
    379 // Internally, the DFA uses a sparse array of
    380 // program instruction pointers as a work queue.
    381 // In leftmost longest mode, marks separate sections
    382 // of workq that started executing at different
    383 // locations in the string (earlier locations first).
    384 class DFA::Workq : public SparseSet {
    385  public:
    386   // Constructor: n is number of normal slots, maxmark number of mark slots.
    387   Workq(int n, int maxmark) :
    388     SparseSet(n+maxmark),
    389     n_(n),
    390     maxmark_(maxmark),
    391     nextmark_(n),
    392     last_was_mark_(true) {
    393   }
    394 
    395   bool is_mark(int i) { return i >= n_; }
    396 
    397   int maxmark() { return maxmark_; }
    398 
    399   void clear() {
    400     SparseSet::clear();
    401     nextmark_ = n_;
    402   }
    403 
    404   void mark() {
    405     if (last_was_mark_)
    406       return;
    407     last_was_mark_ = false;
    408     SparseSet::insert_new(nextmark_++);
    409   }
    410 
    411   int size() {
    412     return n_ + maxmark_;
    413   }
    414 
    415   void insert(int id) {
    416     if (contains(id))
    417       return;
    418     insert_new(id);
    419   }
    420 
    421   void insert_new(int id) {
    422     last_was_mark_ = false;
    423     SparseSet::insert_new(id);
    424   }
    425 
    426  private:
    427   int n_;                // size excluding marks
    428   int maxmark_;          // maximum number of marks
    429   int nextmark_;         // id of next mark
    430   bool last_was_mark_;   // last inserted was mark
    431   DISALLOW_EVIL_CONSTRUCTORS(Workq);
    432 };
    433 
    434 DFA::DFA(Prog* prog, Prog::MatchKind kind, int64 max_mem)
    435   : prog_(prog),
    436     kind_(kind),
    437     init_failed_(false),
    438     q0_(NULL),
    439     q1_(NULL),
    440     astack_(NULL),
    441     mem_budget_(max_mem),
    442     cache_warned_(false) {
    443   if (DebugDFA)
    444     fprintf(stderr, "\nkind %d\n%s\n", (int)kind_, prog_->DumpUnanchored().c_str());
    445   int nmark = 0;
    446   start_unanchored_ = 0;
    447   if (kind_ == Prog::kLongestMatch) {
    448     nmark = prog->size();
    449     start_unanchored_ = prog->start_unanchored();
    450   }
    451   nastack_ = 2 * prog->size() + nmark;
    452 
    453   // Account for space needed for DFA, q0, q1, astack.
    454   mem_budget_ -= sizeof(DFA);
    455   mem_budget_ -= (prog_->size() + nmark) *
    456                  (sizeof(int)+sizeof(int)) * 2;  // q0, q1
    457   mem_budget_ -= nastack_ * sizeof(int);  // astack
    458   if (mem_budget_ < 0) {
    459     LOG(INFO) << StringPrintf("DFA out of memory: prog size %lld mem %lld",
    460                               prog_->size(), max_mem);
    461     init_failed_ = true;
    462     return;
    463   }
    464 
    465   state_budget_ = mem_budget_;
    466 
    467   // Make sure there is a reasonable amount of working room left.
    468   // At minimum, the search requires room for two states in order
    469   // to limp along, restarting frequently.  We'll get better performance
    470   // if there is room for a larger number of states, say 20.
    471   int64 one_state = sizeof(State) + (prog_->size()+nmark)*sizeof(int) +
    472                     (prog_->bytemap_range()+1)*sizeof(State*);
    473   if (state_budget_ < 20*one_state) {
    474     LOG(INFO) << StringPrintf("DFA out of memory: prog size %lld mem %lld",
    475                               prog_->size(), max_mem);
    476     init_failed_ = true;
    477     return;
    478   }
    479 
    480   q0_ = new Workq(prog->size(), nmark);
    481   q1_ = new Workq(prog->size(), nmark);
    482   astack_ = new int[nastack_];
    483 }
    484 
    485 DFA::~DFA() {
    486   delete q0_;
    487   delete q1_;
    488   delete[] astack_;
    489   ClearCache();
    490 }
    491 
    492 // In the DFA state graph, s->next[c] == NULL means that the
    493 // state has not yet been computed and needs to be.  We need
    494 // a different special value to signal that s->next[c] is a
    495 // state that can never lead to a match (and thus the search
    496 // can be called off).  Hence DeadState.
    497 #define DeadState reinterpret_cast<State*>(1)
    498 
    499 // Signals that the rest of the string matches no matter what it is.
    500 #define FullMatchState reinterpret_cast<State*>(2)
    501 
    502 #define SpecialStateMax FullMatchState
    503 
    504 // Debugging printouts
    505 
    506 // For debugging, returns a string representation of the work queue.
    507 string DFA::DumpWorkq(Workq* q) {
    508   string s;
    509   const char* sep = "";
    510   for (DFA::Workq::iterator it = q->begin(); it != q->end(); ++it) {
    511     if (q->is_mark(*it)) {
    512       StringAppendF(&s, "|");
    513       sep = "";
    514     } else {
    515       StringAppendF(&s, "%s%d", sep, *it);
    516       sep = ",";
    517     }
    518   }
    519   return s;
    520 }
    521 
    522 // For debugging, returns a string representation of the state.
    523 string DFA::DumpState(State* state) {
    524   if (state == NULL)
    525     return "_";
    526   if (state == DeadState)
    527     return "X";
    528   if (state == FullMatchState)
    529     return "*";
    530   string s;
    531   const char* sep = "";
    532   StringAppendF(&s, "(%p)", state);
    533   for (int i = 0; i < state->ninst_; i++) {
    534     if (state->inst_[i] == Mark) {
    535       StringAppendF(&s, "|");
    536       sep = "";
    537     } else {
    538       StringAppendF(&s, "%s%d", sep, state->inst_[i]);
    539       sep = ",";
    540     }
    541   }
    542   StringAppendF(&s, " flag=%#x", state->flag_);
    543   return s;
    544 }
    545 
    546 //////////////////////////////////////////////////////////////////////
    547 //
    548 // DFA state graph construction.
    549 //
    550 // The DFA state graph is a heavily-linked collection of State* structures.
    551 // The state_cache_ is a set of all the State structures ever allocated,
    552 // so that if the same state is reached by two different paths,
    553 // the same State structure can be used.  This reduces allocation
    554 // requirements and also avoids duplication of effort across the two
    555 // identical states.
    556 //
    557 // A State is defined by an ordered list of instruction ids and a flag word.
    558 //
    559 // The choice of an ordered list of instructions differs from a typical
    560 // textbook DFA implementation, which would use an unordered set.
    561 // Textbook descriptions, however, only care about whether
    562 // the DFA matches, not where it matches in the text.  To decide where the
    563 // DFA matches, we need to mimic the behavior of the dominant backtracking
    564 // implementations like PCRE, which try one possible regular expression
    565 // execution, then another, then another, stopping when one of them succeeds.
    566 // The DFA execution tries these many executions in parallel, representing
    567 // each by an instruction id.  These pointers are ordered in the State.inst_
    568 // list in the same order that the executions would happen in a backtracking
    569 // search: if a match is found during execution of inst_[2], inst_[i] for i>=3
    570 // can be discarded.
    571 //
    572 // Textbooks also typically do not consider context-aware empty string operators
    573 // like ^ or $.  These are handled by the flag word, which specifies the set
    574 // of empty-string operators that should be matched when executing at the
    575 // current text position.  These flag bits are defined in prog.h.
    576 // The flag word also contains two DFA-specific bits: kFlagMatch if the state
    577 // is a matching state (one that reached a kInstMatch in the program)
    578 // and kFlagLastWord if the last processed byte was a word character, for the
    579 // implementation of \B and \b.
    580 //
    581 // The flag word also contains, shifted up 16 bits, the bits looked for by
    582 // any kInstEmptyWidth instructions in the state.  These provide a useful
    583 // summary indicating when new flags might be useful.
    584 //
    585 // The permanent representation of a State's instruction ids is just an array,
    586 // but while a state is being analyzed, these instruction ids are represented
    587 // as a Workq, which is an array that allows iteration in insertion order.
    588 
    589 // NOTE(rsc): The choice of State construction determines whether the DFA
    590 // mimics backtracking implementations (so-called leftmost first matching) or
    591 // traditional DFA implementations (so-called leftmost longest matching as
    592 // prescribed by POSIX).  This implementation chooses to mimic the
    593 // backtracking implementations, because we want to replace PCRE.  To get
    594 // POSIX behavior, the states would need to be considered not as a simple
    595 // ordered list of instruction ids, but as a list of unordered sets of instruction
    596 // ids.  A match by a state in one set would inhibit the running of sets
    597 // farther down the list but not other instruction ids in the same set.  Each
    598 // set would correspond to matches beginning at a given point in the string.
    599 // This is implemented by separating different sets with Mark pointers.
    600 
    601 // Looks in the State cache for a State matching q, flag.
    602 // If one is found, returns it.  If one is not found, allocates one,
    603 // inserts it in the cache, and returns it.
    604 DFA::State* DFA::WorkqToCachedState(Workq* q, uint flag) {
    605   if (DEBUG_MODE)
    606     mutex_.AssertHeld();
    607 
    608   // Construct array of instruction ids for the new state.
    609   // Only ByteRange, EmptyWidth, and Match instructions are useful to keep:
    610   // those are the only operators with any effect in
    611   // RunWorkqOnEmptyString or RunWorkqOnByte.
    612   int* inst = new int[q->size()];
    613   int n = 0;
    614   uint needflags = 0;     // flags needed by kInstEmptyWidth instructions
    615   bool sawmatch = false;  // whether queue contains guaranteed kInstMatch
    616   bool sawmark = false;  // whether queue contains a Mark
    617   if (DebugDFA)
    618     fprintf(stderr, "WorkqToCachedState %s [%#x]", DumpWorkq(q).c_str(), flag);
    619   for (Workq::iterator it = q->begin(); it != q->end(); ++it) {
    620     int id = *it;
    621     if (sawmatch && (kind_ == Prog::kFirstMatch || q->is_mark(id)))
    622       break;
    623     if (q->is_mark(id)) {
    624       if (n > 0 && inst[n-1] != Mark) {
    625         sawmark = true;
    626         inst[n++] = Mark;
    627       }
    628       continue;
    629     }
    630     Prog::Inst* ip = prog_->inst(id);
    631     switch (ip->opcode()) {
    632       case kInstAltMatch:
    633         // This state will continue to a match no matter what
    634         // the rest of the input is.  If it is the highest priority match
    635         // being considered, return the special FullMatchState
    636         // to indicate that it's all matches from here out.
    637         if (kind_ != Prog::kManyMatch &&
    638             (kind_ != Prog::kFirstMatch ||
    639              (it == q->begin() && ip->greedy(prog_))) &&
    640             (kind_ != Prog::kLongestMatch || !sawmark) &&
    641             (flag & kFlagMatch)) {
    642           delete[] inst;
    643           if (DebugDFA)
    644             fprintf(stderr, " -> FullMatchState\n");
    645           return FullMatchState;
    646         }
    647         // Fall through.
    648       case kInstByteRange:    // These are useful.
    649       case kInstEmptyWidth:
    650       case kInstMatch:
    651       case kInstAlt:          // Not useful, but necessary [*]
    652         inst[n++] = *it;
    653         if (ip->opcode() == kInstEmptyWidth)
    654           needflags |= ip->empty();
    655         if (ip->opcode() == kInstMatch && !prog_->anchor_end())
    656           sawmatch = true;
    657         break;
    658 
    659       default:                // The rest are not.
    660         break;
    661     }
    662 
    663     // [*] kInstAlt would seem useless to record in a state, since
    664     // we've already followed both its arrows and saved all the
    665     // interesting states we can reach from there.  The problem
    666     // is that one of the empty-width instructions might lead
    667     // back to the same kInstAlt (if an empty-width operator is starred),
    668     // producing a different evaluation order depending on whether
    669     // we keep the kInstAlt to begin with.  Sigh.
    670     // A specific case that this affects is /(^|a)+/ matching "a".
    671     // If we don't save the kInstAlt, we will match the whole "a" (0,1)
    672     // but in fact the correct leftmost-first match is the leading "" (0,0).
    673   }
    674   DCHECK_LE(n, q->size());
    675   if (n > 0 && inst[n-1] == Mark)
    676     n--;
    677 
    678   // If there are no empty-width instructions waiting to execute,
    679   // then the extra flag bits will not be used, so there is no
    680   // point in saving them.  (Discarding them reduces the number
    681   // of distinct states.)
    682   if (needflags == 0)
    683     flag &= kFlagMatch;
    684 
    685   // NOTE(rsc): The code above cannot do flag &= needflags,
    686   // because if the right flags were present to pass the current
    687   // kInstEmptyWidth instructions, new kInstEmptyWidth instructions
    688   // might be reached that in turn need different flags.
    689   // The only sure thing is that if there are no kInstEmptyWidth
    690   // instructions at all, no flags will be needed.
    691   // We could do the extra work to figure out the full set of
    692   // possibly needed flags by exploring past the kInstEmptyWidth
    693   // instructions, but the check above -- are any flags needed
    694   // at all? -- handles the most common case.  More fine-grained
    695   // analysis can only be justified by measurements showing that
    696   // too many redundant states are being allocated.
    697 
    698   // If there are no Insts in the list, it's a dead state,
    699   // which is useful to signal with a special pointer so that
    700   // the execution loop can stop early.  This is only okay
    701   // if the state is *not* a matching state.
    702   if (n == 0 && flag == 0) {
    703     delete[] inst;
    704     if (DebugDFA)
    705       fprintf(stderr, " -> DeadState\n");
    706     return DeadState;
    707   }
    708 
    709   // If we're in longest match mode, the state is a sequence of
    710   // unordered state sets separated by Marks.  Sort each set
    711   // to canonicalize, to reduce the number of distinct sets stored.
    712   if (kind_ == Prog::kLongestMatch) {
    713     int* ip = inst;
    714     int* ep = ip + n;
    715     while (ip < ep) {
    716       int* markp = ip;
    717       while (markp < ep && *markp != Mark)
    718         markp++;
    719       sort(ip, markp);
    720       if (markp < ep)
    721         markp++;
    722       ip = markp;
    723     }
    724   }
    725 
    726   // Save the needed empty-width flags in the top bits for use later.
    727   flag |= needflags << kFlagNeedShift;
    728 
    729   State* state = CachedState(inst, n, flag);
    730   delete[] inst;
    731   return state;
    732 }
    733 
    734 // Looks in the State cache for a State matching inst, ninst, flag.
    735 // If one is found, returns it.  If one is not found, allocates one,
    736 // inserts it in the cache, and returns it.
    737 DFA::State* DFA::CachedState(int* inst, int ninst, uint flag) {
    738   if (DEBUG_MODE)
    739     mutex_.AssertHeld();
    740 
    741   // Look in the cache for a pre-existing state.
    742   State state = { inst, ninst, flag, NULL };
    743   StateSet::iterator it = state_cache_.find(&state);
    744   if (it != state_cache_.end()) {
    745     if (DebugDFA)
    746       fprintf(stderr, " -cached-> %s\n", DumpState(*it).c_str());
    747     return *it;
    748   }
    749 
    750   // Must have enough memory for new state.
    751   // In addition to what we're going to allocate,
    752   // the state cache hash table seems to incur about 32 bytes per
    753   // State*, empirically.
    754   const int kStateCacheOverhead = 32;
    755   int nnext = prog_->bytemap_range() + 1;  // + 1 for kByteEndText slot
    756   int mem = sizeof(State) + nnext*sizeof(State*) + ninst*sizeof(int);
    757   if (mem_budget_ < mem + kStateCacheOverhead) {
    758     mem_budget_ = -1;
    759     return NULL;
    760   }
    761   mem_budget_ -= mem + kStateCacheOverhead;
    762 
    763   // Allocate new state, along with room for next and inst.
    764   char* space = new char[mem];
    765   State* s = reinterpret_cast<State*>(space);
    766   s->next_ = reinterpret_cast<State**>(s + 1);
    767   s->inst_ = reinterpret_cast<int*>(s->next_ + nnext);
    768   memset(s->next_, 0, nnext*sizeof s->next_[0]);
    769   memmove(s->inst_, inst, ninst*sizeof s->inst_[0]);
    770   s->ninst_ = ninst;
    771   s->flag_ = flag;
    772   if (DebugDFA)
    773     fprintf(stderr, " -> %s\n", DumpState(s).c_str());
    774 
    775   // Put state in cache and return it.
    776   state_cache_.insert(s);
    777   return s;
    778 }
    779 
    780 // Clear the cache.  Must hold cache_mutex_.w or be in destructor.
    781 void DFA::ClearCache() {
    782   // In case state_cache_ doesn't support deleting entries
    783   // during iteration, copy into a vector and then delete.
    784   vector<State*> v;
    785   v.reserve(state_cache_.size());
    786   for (StateSet::iterator it = state_cache_.begin();
    787        it != state_cache_.end(); ++it)
    788     v.push_back(*it);
    789   state_cache_.clear();
    790   for (int i = 0; i < v.size(); i++)
    791     delete[] reinterpret_cast<const char*>(v[i]);
    792 }
    793 
    794 // Copies insts in state s to the work queue q.
    795 void DFA::StateToWorkq(State* s, Workq* q) {
    796   q->clear();
    797   for (int i = 0; i < s->ninst_; i++) {
    798     if (s->inst_[i] == Mark)
    799       q->mark();
    800     else
    801       q->insert_new(s->inst_[i]);
    802   }
    803 }
    804 
    805 // Adds ip to the work queue, following empty arrows according to flag
    806 // and expanding kInstAlt instructions (two-target gotos).
    807 void DFA::AddToQueue(Workq* q, int id, uint flag) {
    808 
    809   // Use astack_ to hold our stack of states yet to process.
    810   // It is sized to have room for nastack_ == 2*prog->size() + nmark
    811   // instructions, which is enough: each instruction can be
    812   // processed by the switch below only once, and the processing
    813   // pushes at most two instructions plus maybe a mark.
    814   // (If we're using marks, nmark == prog->size(); otherwise nmark == 0.)
    815   int* stk = astack_;
    816   int nstk = 0;
    817 
    818   stk[nstk++] = id;
    819   while (nstk > 0) {
    820     DCHECK_LE(nstk, nastack_);
    821     id = stk[--nstk];
    822 
    823     if (id == Mark) {
    824       q->mark();
    825       continue;
    826     }
    827 
    828     if (id == 0)
    829       continue;
    830 
    831     // If ip is already on the queue, nothing to do.
    832     // Otherwise add it.  We don't actually keep all the ones
    833     // that get added -- for example, kInstAlt is ignored
    834     // when on a work queue -- but adding all ip's here
    835     // increases the likelihood of q->contains(id),
    836     // reducing the amount of duplicated work.
    837     if (q->contains(id))
    838       continue;
    839     q->insert_new(id);
    840 
    841     // Process instruction.
    842     Prog::Inst* ip = prog_->inst(id);
    843     switch (ip->opcode()) {
    844       case kInstFail:       // can't happen: discarded above
    845         break;
    846 
    847       case kInstByteRange:  // just save these on the queue
    848       case kInstMatch:
    849         break;
    850 
    851       case kInstCapture:    // DFA treats captures as no-ops.
    852       case kInstNop:
    853         stk[nstk++] = ip->out();
    854         break;
    855 
    856       case kInstAlt:        // two choices: expand both, in order
    857       case kInstAltMatch:
    858         // Want to visit out then out1, so push on stack in reverse order.
    859         // This instruction is the [00-FF]* loop at the beginning of
    860         // a leftmost-longest unanchored search, separate out from out1
    861         // with a Mark, so that out1's threads (which will start farther
    862         // to the right in the string being searched) are lower priority
    863         // than the current ones.
    864         stk[nstk++] = ip->out1();
    865         if (q->maxmark() > 0 &&
    866             id == prog_->start_unanchored() && id != prog_->start())
    867           stk[nstk++] = Mark;
    868         stk[nstk++] = ip->out();
    869         break;
    870 
    871       case kInstEmptyWidth:
    872         if ((ip->empty() & flag) == ip->empty())
    873           stk[nstk++] = ip->out();
    874         break;
    875     }
    876   }
    877 }
    878 
    879 // Running of work queues.  In the work queue, order matters:
    880 // the queue is sorted in priority order.  If instruction i comes before j,
    881 // then the instructions that i produces during the run must come before
    882 // the ones that j produces.  In order to keep this invariant, all the
    883 // work queue runners have to take an old queue to process and then
    884 // also a new queue to fill in.  It's not acceptable to add to the end of
    885 // an existing queue, because new instructions will not end up in the
    886 // correct position.
    887 
    888 // Runs the work queue, processing the empty strings indicated by flag.
    889 // For example, flag == kEmptyBeginLine|kEmptyEndLine means to match
    890 // both ^ and $.  It is important that callers pass all flags at once:
    891 // processing both ^ and $ is not the same as first processing only ^
    892 // and then processing only $.  Doing the two-step sequence won't match
    893 // ^$^$^$ but processing ^ and $ simultaneously will (and is the behavior
    894 // exhibited by existing implementations).
    895 void DFA::RunWorkqOnEmptyString(Workq* oldq, Workq* newq, uint flag) {
    896   newq->clear();
    897   for (Workq::iterator i = oldq->begin(); i != oldq->end(); ++i) {
    898     if (oldq->is_mark(*i))
    899       AddToQueue(newq, Mark, flag);
    900     else
    901       AddToQueue(newq, *i, flag);
    902   }
    903 }
    904 
    905 // Runs the work queue, processing the single byte c followed by any empty
    906 // strings indicated by flag.  For example, c == 'a' and flag == kEmptyEndLine,
    907 // means to match c$.  Sets the bool *ismatch to true if the end of the
    908 // regular expression program has been reached (the regexp has matched).
    909 void DFA::RunWorkqOnByte(Workq* oldq, Workq* newq,
    910                          int c, uint flag, bool* ismatch,
    911                          Prog::MatchKind kind,
    912                          int new_byte_loop) {
    913   if (DEBUG_MODE)
    914     mutex_.AssertHeld();
    915 
    916   newq->clear();
    917   for (Workq::iterator i = oldq->begin(); i != oldq->end(); ++i) {
    918     if (oldq->is_mark(*i)) {
    919       if (*ismatch)
    920         return;
    921       newq->mark();
    922       continue;
    923     }
    924     int id = *i;
    925     Prog::Inst* ip = prog_->inst(id);
    926     switch (ip->opcode()) {
    927       case kInstFail:        // never succeeds
    928       case kInstCapture:     // already followed
    929       case kInstNop:         // already followed
    930       case kInstAlt:         // already followed
    931       case kInstAltMatch:    // already followed
    932       case kInstEmptyWidth:  // already followed
    933         break;
    934 
    935       case kInstByteRange:   // can follow if c is in range
    936         if (ip->Matches(c))
    937           AddToQueue(newq, ip->out(), flag);
    938         break;
    939 
    940       case kInstMatch:
    941         if (prog_->anchor_end() && c != kByteEndText)
    942           break;
    943         *ismatch = true;
    944         if (kind == Prog::kFirstMatch) {
    945           // Can stop processing work queue since we found a match.
    946           return;
    947         }
    948         break;
    949     }
    950   }
    951 
    952   if (DebugDFA)
    953     fprintf(stderr, "%s on %d[%#x] -> %s [%d]\n", DumpWorkq(oldq).c_str(),
    954             c, flag, DumpWorkq(newq).c_str(), *ismatch);
    955 }
    956 
    957 // Processes input byte c in state, returning new state.
    958 // Caller does not hold mutex.
    959 DFA::State* DFA::RunStateOnByteUnlocked(State* state, int c) {
    960   // Keep only one RunStateOnByte going
    961   // even if the DFA is being run by multiple threads.
    962   MutexLock l(&mutex_);
    963   return RunStateOnByte(state, c);
    964 }
    965 
    966 // Processes input byte c in state, returning new state.
    967 DFA::State* DFA::RunStateOnByte(State* state, int c) {
    968   if (DEBUG_MODE)
    969     mutex_.AssertHeld();
    970   if (state <= SpecialStateMax) {
    971     if (state == FullMatchState) {
    972       // It is convenient for routines like PossibleMatchRange
    973       // if we implement RunStateOnByte for FullMatchState:
    974       // once you get into this state you never get out,
    975       // so it's pretty easy.
    976       return FullMatchState;
    977     }
    978     if (state == DeadState) {
    979       LOG(DFATAL) << "DeadState in RunStateOnByte";
    980       return NULL;
    981     }
    982     if (state == NULL) {
    983       LOG(DFATAL) << "NULL state in RunStateOnByte";
    984       return NULL;
    985     }
    986     LOG(DFATAL) << "Unexpected special state in RunStateOnByte";
    987     return NULL;
    988   }
    989 
    990   // If someone else already computed this, return it.
    991   MaybeReadMemoryBarrier(); // On alpha we need to ensure read ordering
    992   State* ns = state->next_[ByteMap(c)];
    993   ANNOTATE_HAPPENS_AFTER(ns);
    994   if (ns != NULL)
    995     return ns;
    996 
    997   // Convert state into Workq.
    998   StateToWorkq(state, q0_);
    999 
   1000   // Flags marking the kinds of empty-width things (^ $ etc)
   1001   // around this byte.  Before the byte we have the flags recorded
   1002   // in the State structure itself.  After the byte we have
   1003   // nothing yet (but that will change: read on).
   1004   uint needflag = state->flag_ >> kFlagNeedShift;
   1005   uint beforeflag = state->flag_ & kFlagEmptyMask;
   1006   uint oldbeforeflag = beforeflag;
   1007   uint afterflag = 0;
   1008 
   1009   if (c == '\n') {
   1010     // Insert implicit $ and ^ around \n
   1011     beforeflag |= kEmptyEndLine;
   1012     afterflag |= kEmptyBeginLine;
   1013   }
   1014 
   1015   if (c == kByteEndText) {
   1016     // Insert implicit $ and \z before the fake "end text" byte.
   1017     beforeflag |= kEmptyEndLine | kEmptyEndText;
   1018   }
   1019 
   1020   // The state flag kFlagLastWord says whether the last
   1021   // byte processed was a word character.  Use that info to
   1022   // insert empty-width (non-)word boundaries.
   1023   bool islastword = state->flag_ & kFlagLastWord;
   1024   bool isword = (c != kByteEndText && Prog::IsWordChar(c));
   1025   if (isword == islastword)
   1026     beforeflag |= kEmptyNonWordBoundary;
   1027   else
   1028     beforeflag |= kEmptyWordBoundary;
   1029 
   1030   // Okay, finally ready to run.
   1031   // Only useful to rerun on empty string if there are new, useful flags.
   1032   if (beforeflag & ~oldbeforeflag & needflag) {
   1033     RunWorkqOnEmptyString(q0_, q1_, beforeflag);
   1034     swap(q0_, q1_);
   1035   }
   1036   bool ismatch = false;
   1037   RunWorkqOnByte(q0_, q1_, c, afterflag, &ismatch, kind_, start_unanchored_);
   1038 
   1039   // Most of the time, we build the state from the output of
   1040   // RunWorkqOnByte, so swap q0_ and q1_ here.  However, so that
   1041   // RE2::Set can tell exactly which match instructions
   1042   // contributed to the match, don't swap if c is kByteEndText.
   1043   // The resulting state wouldn't be correct for further processing
   1044   // of the string, but we're at the end of the text so that's okay.
   1045   // Leaving q0_ alone preseves the match instructions that led to
   1046   // the current setting of ismatch.
   1047   if (c != kByteEndText || kind_ != Prog::kManyMatch)
   1048     swap(q0_, q1_);
   1049 
   1050   // Save afterflag along with ismatch and isword in new state.
   1051   uint flag = afterflag;
   1052   if (ismatch)
   1053     flag |= kFlagMatch;
   1054   if (isword)
   1055     flag |= kFlagLastWord;
   1056 
   1057   ns = WorkqToCachedState(q0_, flag);
   1058 
   1059   // Write barrier before updating state->next_ so that the
   1060   // main search loop can proceed without any locking, for speed.
   1061   // (Otherwise it would need one mutex operation per input byte.)
   1062   // The annotations below tell race detectors that:
   1063   //   a) the access to next_ should be ignored,
   1064   //   b) 'ns' is properly published.
   1065   WriteMemoryBarrier();  // Flush ns before linking to it.
   1066 
   1067   ANNOTATE_IGNORE_WRITES_BEGIN();
   1068   ANNOTATE_HAPPENS_BEFORE(ns);
   1069   state->next_[ByteMap(c)] = ns;
   1070   ANNOTATE_IGNORE_WRITES_END();
   1071   return ns;
   1072 }
   1073 
   1074 
   1075 //////////////////////////////////////////////////////////////////////
   1076 // DFA cache reset.
   1077 
   1078 // Reader-writer lock helper.
   1079 //
   1080 // The DFA uses a reader-writer mutex to protect the state graph itself.
   1081 // Traversing the state graph requires holding the mutex for reading,
   1082 // and discarding the state graph and starting over requires holding the
   1083 // lock for writing.  If a search needs to expand the graph but is out
   1084 // of memory, it will need to drop its read lock and then acquire the
   1085 // write lock.  Since it cannot then atomically downgrade from write lock
   1086 // to read lock, it runs the rest of the search holding the write lock.
   1087 // (This probably helps avoid repeated contention, but really the decision
   1088 // is forced by the Mutex interface.)  It's a bit complicated to keep
   1089 // track of whether the lock is held for reading or writing and thread
   1090 // that through the search, so instead we encapsulate it in the RWLocker
   1091 // and pass that around.
   1092 
   1093 class DFA::RWLocker {
   1094  public:
   1095   explicit RWLocker(Mutex* mu);
   1096   ~RWLocker();
   1097 
   1098   // If the lock is only held for reading right now,
   1099   // drop the read lock and re-acquire for writing.
   1100   // Subsequent calls to LockForWriting are no-ops.
   1101   // Notice that the lock is *released* temporarily.
   1102   void LockForWriting();
   1103 
   1104   // Returns whether the lock is already held for writing.
   1105   bool IsLockedForWriting() {
   1106     return writing_;
   1107   }
   1108 
   1109  private:
   1110   Mutex* mu_;
   1111   bool writing_;
   1112 
   1113   DISALLOW_EVIL_CONSTRUCTORS(RWLocker);
   1114 };
   1115 
   1116 DFA::RWLocker::RWLocker(Mutex* mu)
   1117   : mu_(mu), writing_(false) {
   1118 
   1119   mu_->ReaderLock();
   1120 }
   1121 
   1122 // This function is marked as NO_THREAD_SAFETY_ANALYSIS because the annotations
   1123 // does not support lock upgrade.
   1124 void DFA::RWLocker::LockForWriting() NO_THREAD_SAFETY_ANALYSIS {
   1125   if (!writing_) {
   1126     mu_->ReaderUnlock();
   1127     mu_->Lock();
   1128     writing_ = true;
   1129   }
   1130 }
   1131 
   1132 DFA::RWLocker::~RWLocker() {
   1133   if (writing_)
   1134     mu_->WriterUnlock();
   1135   else
   1136     mu_->ReaderUnlock();
   1137 }
   1138 
   1139 
   1140 // When the DFA's State cache fills, we discard all the states in the
   1141 // cache and start over.  Many threads can be using and adding to the
   1142 // cache at the same time, so we synchronize using the cache_mutex_
   1143 // to keep from stepping on other threads.  Specifically, all the
   1144 // threads using the current cache hold cache_mutex_ for reading.
   1145 // When a thread decides to flush the cache, it drops cache_mutex_
   1146 // and then re-acquires it for writing.  That ensures there are no
   1147 // other threads accessing the cache anymore.  The rest of the search
   1148 // runs holding cache_mutex_ for writing, avoiding any contention
   1149 // with or cache pollution caused by other threads.
   1150 
   1151 void DFA::ResetCache(RWLocker* cache_lock) {
   1152   // Re-acquire the cache_mutex_ for writing (exclusive use).
   1153   bool was_writing = cache_lock->IsLockedForWriting();
   1154   cache_lock->LockForWriting();
   1155 
   1156   // If we already held cache_mutex_ for writing, it means
   1157   // this invocation of Search() has already reset the
   1158   // cache once already.  That's a pretty clear indication
   1159   // that the cache is too small.  Warn about that, once.
   1160   // TODO(rsc): Only warn if state_cache_.size() < some threshold.
   1161   if (was_writing && !cache_warned_) {
   1162     LOG(INFO) << "DFA memory cache could be too small: "
   1163               << "only room for " << state_cache_.size() << " states.";
   1164     cache_warned_ = true;
   1165   }
   1166 
   1167   // Clear the cache, reset the memory budget.
   1168   for (int i = 0; i < kMaxStart; i++) {
   1169     start_[i].start = NULL;
   1170     start_[i].firstbyte = kFbUnknown;
   1171   }
   1172   ClearCache();
   1173   mem_budget_ = state_budget_;
   1174 }
   1175 
   1176 // Typically, a couple States do need to be preserved across a cache
   1177 // reset, like the State at the current point in the search.
   1178 // The StateSaver class helps keep States across cache resets.
   1179 // It makes a copy of the state's guts outside the cache (before the reset)
   1180 // and then can be asked, after the reset, to recreate the State
   1181 // in the new cache.  For example, in a DFA method ("this" is a DFA):
   1182 //
   1183 //   StateSaver saver(this, s);
   1184 //   ResetCache(cache_lock);
   1185 //   s = saver.Restore();
   1186 //
   1187 // The saver should always have room in the cache to re-create the state,
   1188 // because resetting the cache locks out all other threads, and the cache
   1189 // is known to have room for at least a couple states (otherwise the DFA
   1190 // constructor fails).
   1191 
   1192 class DFA::StateSaver {
   1193  public:
   1194   explicit StateSaver(DFA* dfa, State* state);
   1195   ~StateSaver();
   1196 
   1197   // Recreates and returns a state equivalent to the
   1198   // original state passed to the constructor.
   1199   // Returns NULL if the cache has filled, but
   1200   // since the DFA guarantees to have room in the cache
   1201   // for a couple states, should never return NULL
   1202   // if used right after ResetCache.
   1203   State* Restore();
   1204 
   1205  private:
   1206   DFA* dfa_;         // the DFA to use
   1207   int* inst_;        // saved info from State
   1208   int ninst_;
   1209   uint flag_;
   1210   bool is_special_;  // whether original state was special
   1211   State* special_;   // if is_special_, the original state
   1212 
   1213   DISALLOW_EVIL_CONSTRUCTORS(StateSaver);
   1214 };
   1215 
   1216 DFA::StateSaver::StateSaver(DFA* dfa, State* state) {
   1217   dfa_ = dfa;
   1218   if (state <= SpecialStateMax) {
   1219     inst_ = NULL;
   1220     ninst_ = 0;
   1221     flag_ = 0;
   1222     is_special_ = true;
   1223     special_ = state;
   1224     return;
   1225   }
   1226   is_special_ = false;
   1227   special_ = NULL;
   1228   flag_ = state->flag_;
   1229   ninst_ = state->ninst_;
   1230   inst_ = new int[ninst_];
   1231   memmove(inst_, state->inst_, ninst_*sizeof inst_[0]);
   1232 }
   1233 
   1234 DFA::StateSaver::~StateSaver() {
   1235   if (!is_special_)
   1236     delete[] inst_;
   1237 }
   1238 
   1239 DFA::State* DFA::StateSaver::Restore() {
   1240   if (is_special_)
   1241     return special_;
   1242   MutexLock l(&dfa_->mutex_);
   1243   State* s = dfa_->CachedState(inst_, ninst_, flag_);
   1244   if (s == NULL)
   1245     LOG(DFATAL) << "StateSaver failed to restore state.";
   1246   return s;
   1247 }
   1248 
   1249 
   1250 //////////////////////////////////////////////////////////////////////
   1251 //
   1252 // DFA execution.
   1253 //
   1254 // The basic search loop is easy: start in a state s and then for each
   1255 // byte c in the input, s = s->next[c].
   1256 //
   1257 // This simple description omits a few efficiency-driven complications.
   1258 //
   1259 // First, the State graph is constructed incrementally: it is possible
   1260 // that s->next[c] is null, indicating that that state has not been
   1261 // fully explored.  In this case, RunStateOnByte must be invoked to
   1262 // determine the next state, which is cached in s->next[c] to save
   1263 // future effort.  An alternative reason for s->next[c] to be null is
   1264 // that the DFA has reached a so-called "dead state", in which any match
   1265 // is no longer possible.  In this case RunStateOnByte will return NULL
   1266 // and the processing of the string can stop early.
   1267 //
   1268 // Second, a 256-element pointer array for s->next_ makes each State
   1269 // quite large (2kB on 64-bit machines).  Instead, dfa->bytemap_[]
   1270 // maps from bytes to "byte classes" and then next_ only needs to have
   1271 // as many pointers as there are byte classes.  A byte class is simply a
   1272 // range of bytes that the regexp never distinguishes between.
   1273 // A regexp looking for a[abc] would have four byte ranges -- 0 to 'a'-1,
   1274 // 'a', 'b' to 'c', and 'c' to 0xFF.  The bytemap slows us a little bit
   1275 // but in exchange we typically cut the size of a State (and thus our
   1276 // memory footprint) by about 5-10x.  The comments still refer to
   1277 // s->next[c] for simplicity, but code should refer to s->next_[bytemap_[c]].
   1278 //
   1279 // Third, it is common for a DFA for an unanchored match to begin in a
   1280 // state in which only one particular byte value can take the DFA to a
   1281 // different state.  That is, s->next[c] != s for only one c.  In this
   1282 // situation, the DFA can do better than executing the simple loop.
   1283 // Instead, it can call memchr to search very quickly for the byte c.
   1284 // Whether the start state has this property is determined during a
   1285 // pre-compilation pass, and if so, the byte b is passed to the search
   1286 // loop as the "firstbyte" argument, along with a boolean "have_firstbyte".
   1287 //
   1288 // Fourth, the desired behavior is to search for the leftmost-best match
   1289 // (approximately, the same one that Perl would find), which is not
   1290 // necessarily the match ending earliest in the string.  Each time a
   1291 // match is found, it must be noted, but the DFA must continue on in
   1292 // hope of finding a higher-priority match.  In some cases, the caller only
   1293 // cares whether there is any match at all, not which one is found.
   1294 // The "want_earliest_match" flag causes the search to stop at the first
   1295 // match found.
   1296 //
   1297 // Fifth, one algorithm that uses the DFA needs it to run over the
   1298 // input string backward, beginning at the end and ending at the beginning.
   1299 // Passing false for the "run_forward" flag causes the DFA to run backward.
   1300 //
   1301 // The checks for these last three cases, which in a naive implementation
   1302 // would be performed once per input byte, slow the general loop enough
   1303 // to merit specialized versions of the search loop for each of the
   1304 // eight possible settings of the three booleans.  Rather than write
   1305 // eight different functions, we write one general implementation and then
   1306 // inline it to create the specialized ones.
   1307 //
   1308 // Note that matches are delayed by one byte, to make it easier to
   1309 // accomodate match conditions depending on the next input byte (like $ and \b).
   1310 // When s->next[c]->IsMatch(), it means that there is a match ending just
   1311 // *before* byte c.
   1312 
   1313 // The generic search loop.  Searches text for a match, returning
   1314 // the pointer to the end of the chosen match, or NULL if no match.
   1315 // The bools are equal to the same-named variables in params, but
   1316 // making them function arguments lets the inliner specialize
   1317 // this function to each combination (see two paragraphs above).
   1318 inline bool DFA::InlinedSearchLoop(SearchParams* params,
   1319                                    bool have_firstbyte,
   1320                                    bool want_earliest_match,
   1321                                    bool run_forward) {
   1322   State* start = params->start;
   1323   const uint8* bp = BytePtr(params->text.begin());  // start of text
   1324   const uint8* p = bp;                              // text scanning point
   1325   const uint8* ep = BytePtr(params->text.end());    // end of text
   1326   const uint8* resetp = NULL;                       // p at last cache reset
   1327   if (!run_forward)
   1328     swap(p, ep);
   1329 
   1330   const uint8* bytemap = prog_->bytemap();
   1331   const uint8* lastmatch = NULL;   // most recent matching position in text
   1332   bool matched = false;
   1333   State* s = start;
   1334 
   1335   if (s->IsMatch()) {
   1336     matched = true;
   1337     lastmatch = p;
   1338     if (want_earliest_match) {
   1339       params->ep = reinterpret_cast<const char*>(lastmatch);
   1340       return true;
   1341     }
   1342   }
   1343 
   1344   while (p != ep) {
   1345     if (DebugDFA)
   1346       fprintf(stderr, "@%d: %s\n", static_cast<int>(p - bp),
   1347               DumpState(s).c_str());
   1348     if (have_firstbyte && s == start) {
   1349       // In start state, only way out is to find firstbyte,
   1350       // so use optimized assembly in memchr to skip ahead.
   1351       // If firstbyte isn't found, we can skip to the end
   1352       // of the string.
   1353       if (run_forward) {
   1354         if ((p = BytePtr(memchr(p, params->firstbyte, ep - p))) == NULL) {
   1355           p = ep;
   1356           break;
   1357         }
   1358       } else {
   1359         if ((p = BytePtr(memrchr(ep, params->firstbyte, p - ep))) == NULL) {
   1360           p = ep;
   1361           break;
   1362         }
   1363         p++;
   1364       }
   1365     }
   1366 
   1367     int c;
   1368     if (run_forward)
   1369       c = *p++;
   1370     else
   1371       c = *--p;
   1372 
   1373     // Note that multiple threads might be consulting
   1374     // s->next_[bytemap[c]] simultaneously.
   1375     // RunStateOnByte takes care of the appropriate locking,
   1376     // including a memory barrier so that the unlocked access
   1377     // (sometimes known as "double-checked locking") is safe.
   1378     // The alternative would be either one DFA per thread
   1379     // or one mutex operation per input byte.
   1380     //
   1381     // ns == DeadState means the state is known to be dead
   1382     // (no more matches are possible).
   1383     // ns == NULL means the state has not yet been computed
   1384     // (need to call RunStateOnByteUnlocked).
   1385     // RunStateOnByte returns ns == NULL if it is out of memory.
   1386     // ns == FullMatchState means the rest of the string matches.
   1387     //
   1388     // Okay to use bytemap[] not ByteMap() here, because
   1389     // c is known to be an actual byte and not kByteEndText.
   1390 
   1391     MaybeReadMemoryBarrier(); // On alpha we need to ensure read ordering
   1392     State* ns = s->next_[bytemap[c]];
   1393     ANNOTATE_HAPPENS_AFTER(ns);
   1394     if (ns == NULL) {
   1395       ns = RunStateOnByteUnlocked(s, c);
   1396       if (ns == NULL) {
   1397         // After we reset the cache, we hold cache_mutex exclusively,
   1398         // so if resetp != NULL, it means we filled the DFA state
   1399         // cache with this search alone (without any other threads).
   1400         // Benchmarks show that doing a state computation on every
   1401         // byte runs at about 0.2 MB/s, while the NFA (nfa.cc) can do the
   1402         // same at about 2 MB/s.  Unless we're processing an average
   1403         // of 10 bytes per state computation, fail so that RE2 can
   1404         // fall back to the NFA.
   1405         if (FLAGS_re2_dfa_bail_when_slow && resetp != NULL &&
   1406             (p - resetp) < 10*state_cache_.size()) {
   1407           params->failed = true;
   1408           return false;
   1409         }
   1410         resetp = p;
   1411 
   1412         // Prepare to save start and s across the reset.
   1413         StateSaver save_start(this, start);
   1414         StateSaver save_s(this, s);
   1415 
   1416         // Discard all the States in the cache.
   1417         ResetCache(params->cache_lock);
   1418 
   1419         // Restore start and s so we can continue.
   1420         if ((start = save_start.Restore()) == NULL ||
   1421             (s = save_s.Restore()) == NULL) {
   1422           // Restore already did LOG(DFATAL).
   1423           params->failed = true;
   1424           return false;
   1425         }
   1426         ns = RunStateOnByteUnlocked(s, c);
   1427         if (ns == NULL) {
   1428           LOG(DFATAL) << "RunStateOnByteUnlocked failed after ResetCache";
   1429           params->failed = true;
   1430           return false;
   1431         }
   1432       }
   1433     }
   1434     if (ns <= SpecialStateMax) {
   1435       if (ns == DeadState) {
   1436         params->ep = reinterpret_cast<const char*>(lastmatch);
   1437         return matched;
   1438       }
   1439       // FullMatchState
   1440       params->ep = reinterpret_cast<const char*>(ep);
   1441       return true;
   1442     }
   1443     s = ns;
   1444 
   1445     if (s->IsMatch()) {
   1446       matched = true;
   1447       // The DFA notices the match one byte late,
   1448       // so adjust p before using it in the match.
   1449       if (run_forward)
   1450         lastmatch = p - 1;
   1451       else
   1452         lastmatch = p + 1;
   1453       if (DebugDFA)
   1454         fprintf(stderr, "match @%d! [%s]\n",
   1455                 static_cast<int>(lastmatch - bp),
   1456                 DumpState(s).c_str());
   1457 
   1458       if (want_earliest_match) {
   1459         params->ep = reinterpret_cast<const char*>(lastmatch);
   1460         return true;
   1461       }
   1462     }
   1463   }
   1464 
   1465   // Process one more byte to see if it triggers a match.
   1466   // (Remember, matches are delayed one byte.)
   1467   int lastbyte;
   1468   if (run_forward) {
   1469     if (params->text.end() == params->context.end())
   1470       lastbyte = kByteEndText;
   1471     else
   1472       lastbyte = params->text.end()[0] & 0xFF;
   1473   } else {
   1474     if (params->text.begin() == params->context.begin())
   1475       lastbyte = kByteEndText;
   1476     else
   1477       lastbyte = params->text.begin()[-1] & 0xFF;
   1478   }
   1479 
   1480   MaybeReadMemoryBarrier(); // On alpha we need to ensure read ordering
   1481   State* ns = s->next_[ByteMap(lastbyte)];
   1482   ANNOTATE_HAPPENS_AFTER(ns);
   1483   if (ns == NULL) {
   1484     ns = RunStateOnByteUnlocked(s, lastbyte);
   1485     if (ns == NULL) {
   1486       StateSaver save_s(this, s);
   1487       ResetCache(params->cache_lock);
   1488       if ((s = save_s.Restore()) == NULL) {
   1489         params->failed = true;
   1490         return false;
   1491       }
   1492       ns = RunStateOnByteUnlocked(s, lastbyte);
   1493       if (ns == NULL) {
   1494         LOG(DFATAL) << "RunStateOnByteUnlocked failed after Reset";
   1495         params->failed = true;
   1496         return false;
   1497       }
   1498     }
   1499   }
   1500   s = ns;
   1501   if (DebugDFA)
   1502     fprintf(stderr, "@_: %s\n", DumpState(s).c_str());
   1503   if (s == FullMatchState) {
   1504     params->ep = reinterpret_cast<const char*>(ep);
   1505     return true;
   1506   }
   1507   if (s > SpecialStateMax && s->IsMatch()) {
   1508     matched = true;
   1509     lastmatch = p;
   1510     if (params->matches && kind_ == Prog::kManyMatch) {
   1511       vector<int>* v = params->matches;
   1512       v->clear();
   1513       for (int i = 0; i < s->ninst_; i++) {
   1514         Prog::Inst* ip = prog_->inst(s->inst_[i]);
   1515         if (ip->opcode() == kInstMatch)
   1516           v->push_back(ip->match_id());
   1517       }
   1518     }
   1519     if (DebugDFA)
   1520       fprintf(stderr, "match @%d! [%s]\n", static_cast<int>(lastmatch - bp),
   1521               DumpState(s).c_str());
   1522   }
   1523   params->ep = reinterpret_cast<const char*>(lastmatch);
   1524   return matched;
   1525 }
   1526 
   1527 // Inline specializations of the general loop.
   1528 bool DFA::SearchFFF(SearchParams* params) {
   1529   return InlinedSearchLoop(params, 0, 0, 0);
   1530 }
   1531 bool DFA::SearchFFT(SearchParams* params) {
   1532   return InlinedSearchLoop(params, 0, 0, 1);
   1533 }
   1534 bool DFA::SearchFTF(SearchParams* params) {
   1535   return InlinedSearchLoop(params, 0, 1, 0);
   1536 }
   1537 bool DFA::SearchFTT(SearchParams* params) {
   1538   return InlinedSearchLoop(params, 0, 1, 1);
   1539 }
   1540 bool DFA::SearchTFF(SearchParams* params) {
   1541   return InlinedSearchLoop(params, 1, 0, 0);
   1542 }
   1543 bool DFA::SearchTFT(SearchParams* params) {
   1544   return InlinedSearchLoop(params, 1, 0, 1);
   1545 }
   1546 bool DFA::SearchTTF(SearchParams* params) {
   1547   return InlinedSearchLoop(params, 1, 1, 0);
   1548 }
   1549 bool DFA::SearchTTT(SearchParams* params) {
   1550   return InlinedSearchLoop(params, 1, 1, 1);
   1551 }
   1552 
   1553 // For debugging, calls the general code directly.
   1554 bool DFA::SlowSearchLoop(SearchParams* params) {
   1555   return InlinedSearchLoop(params,
   1556                            params->firstbyte >= 0,
   1557                            params->want_earliest_match,
   1558                            params->run_forward);
   1559 }
   1560 
   1561 // For performance, calls the appropriate specialized version
   1562 // of InlinedSearchLoop.
   1563 bool DFA::FastSearchLoop(SearchParams* params) {
   1564   // Because the methods are private, the Searches array
   1565   // cannot be declared at top level.
   1566   static bool (DFA::*Searches[])(SearchParams*) = {
   1567     &DFA::SearchFFF,
   1568     &DFA::SearchFFT,
   1569     &DFA::SearchFTF,
   1570     &DFA::SearchFTT,
   1571     &DFA::SearchTFF,
   1572     &DFA::SearchTFT,
   1573     &DFA::SearchTTF,
   1574     &DFA::SearchTTT,
   1575   };
   1576 
   1577   bool have_firstbyte = (params->firstbyte >= 0);
   1578   int index = 4 * have_firstbyte +
   1579               2 * params->want_earliest_match +
   1580               1 * params->run_forward;
   1581   return (this->*Searches[index])(params);
   1582 }
   1583 
   1584 
   1585 // The discussion of DFA execution above ignored the question of how
   1586 // to determine the initial state for the search loop.  There are two
   1587 // factors that influence the choice of start state.
   1588 //
   1589 // The first factor is whether the search is anchored or not.
   1590 // The regexp program (Prog*) itself has
   1591 // two different entry points: one for anchored searches and one for
   1592 // unanchored searches.  (The unanchored version starts with a leading ".*?"
   1593 // and then jumps to the anchored one.)
   1594 //
   1595 // The second factor is where text appears in the larger context, which
   1596 // determines which empty-string operators can be matched at the beginning
   1597 // of execution.  If text is at the very beginning of context, \A and ^ match.
   1598 // Otherwise if text is at the beginning of a line, then ^ matches.
   1599 // Otherwise it matters whether the character before text is a word character
   1600 // or a non-word character.
   1601 //
   1602 // The two cases (unanchored vs not) and four cases (empty-string flags)
   1603 // combine to make the eight cases recorded in the DFA's begin_text_[2],
   1604 // begin_line_[2], after_wordchar_[2], and after_nonwordchar_[2] cached
   1605 // StartInfos.  The start state for each is filled in the first time it
   1606 // is used for an actual search.
   1607 
   1608 // Examines text, context, and anchored to determine the right start
   1609 // state for the DFA search loop.  Fills in params and returns true on success.
   1610 // Returns false on failure.
   1611 bool DFA::AnalyzeSearch(SearchParams* params) {
   1612   const StringPiece& text = params->text;
   1613   const StringPiece& context = params->context;
   1614 
   1615   // Sanity check: make sure that text lies within context.
   1616   if (text.begin() < context.begin() || text.end() > context.end()) {
   1617     LOG(DFATAL) << "Text is not inside context.";
   1618     params->start = DeadState;
   1619     return true;
   1620   }
   1621 
   1622   // Determine correct search type.
   1623   int start;
   1624   uint flags;
   1625   if (params->run_forward) {
   1626     if (text.begin() == context.begin()) {
   1627       start = kStartBeginText;
   1628       flags = kEmptyBeginText|kEmptyBeginLine;
   1629     } else if (text.begin()[-1] == '\n') {
   1630       start = kStartBeginLine;
   1631       flags = kEmptyBeginLine;
   1632     } else if (Prog::IsWordChar(text.begin()[-1] & 0xFF)) {
   1633       start = kStartAfterWordChar;
   1634       flags = kFlagLastWord;
   1635     } else {
   1636       start = kStartAfterNonWordChar;
   1637       flags = 0;
   1638     }
   1639   } else {
   1640     if (text.end() == context.end()) {
   1641       start = kStartBeginText;
   1642       flags = kEmptyBeginText|kEmptyBeginLine;
   1643     } else if (text.end()[0] == '\n') {
   1644       start = kStartBeginLine;
   1645       flags = kEmptyBeginLine;
   1646     } else if (Prog::IsWordChar(text.end()[0] & 0xFF)) {
   1647       start = kStartAfterWordChar;
   1648       flags = kFlagLastWord;
   1649     } else {
   1650       start = kStartAfterNonWordChar;
   1651       flags = 0;
   1652     }
   1653   }
   1654   if (params->anchored || prog_->anchor_start())
   1655     start |= kStartAnchored;
   1656   StartInfo* info = &start_[start];
   1657 
   1658   // Try once without cache_lock for writing.
   1659   // Try again after resetting the cache
   1660   // (ResetCache will relock cache_lock for writing).
   1661   if (!AnalyzeSearchHelper(params, info, flags)) {
   1662     ResetCache(params->cache_lock);
   1663     if (!AnalyzeSearchHelper(params, info, flags)) {
   1664       LOG(DFATAL) << "Failed to analyze start state.";
   1665       params->failed = true;
   1666       return false;
   1667     }
   1668   }
   1669 
   1670   if (DebugDFA)
   1671     fprintf(stderr, "anchored=%d fwd=%d flags=%#x state=%s firstbyte=%d\n",
   1672             params->anchored, params->run_forward, flags,
   1673             DumpState(info->start).c_str(), info->firstbyte);
   1674 
   1675   params->start = info->start;
   1676   params->firstbyte = ANNOTATE_UNPROTECTED_READ(info->firstbyte);
   1677 
   1678   return true;
   1679 }
   1680 
   1681 // Fills in info if needed.  Returns true on success, false on failure.
   1682 bool DFA::AnalyzeSearchHelper(SearchParams* params, StartInfo* info,
   1683                               uint flags) {
   1684   // Quick check; okay because of memory barriers below.
   1685   if (ANNOTATE_UNPROTECTED_READ(info->firstbyte) != kFbUnknown) {
   1686     ANNOTATE_HAPPENS_AFTER(&info->firstbyte);
   1687     return true;
   1688   }
   1689 
   1690   MutexLock l(&mutex_);
   1691   if (info->firstbyte != kFbUnknown) {
   1692     ANNOTATE_HAPPENS_AFTER(&info->firstbyte);
   1693     return true;
   1694   }
   1695 
   1696   q0_->clear();
   1697   AddToQueue(q0_,
   1698              params->anchored ? prog_->start() : prog_->start_unanchored(),
   1699              flags);
   1700   info->start = WorkqToCachedState(q0_, flags);
   1701   if (info->start == NULL)
   1702     return false;
   1703 
   1704   if (info->start == DeadState) {
   1705     ANNOTATE_HAPPENS_BEFORE(&info->firstbyte);
   1706     WriteMemoryBarrier();  // Synchronize with "quick check" above.
   1707     info->firstbyte = kFbNone;
   1708     return true;
   1709   }
   1710 
   1711   if (info->start == FullMatchState) {
   1712     ANNOTATE_HAPPENS_BEFORE(&info->firstbyte);
   1713     WriteMemoryBarrier();  // Synchronize with "quick check" above.
   1714     info->firstbyte = kFbNone;	// will be ignored
   1715     return true;
   1716   }
   1717 
   1718   // Compute info->firstbyte by running state on all
   1719   // possible byte values, looking for a single one that
   1720   // leads to a different state.
   1721   int firstbyte = kFbNone;
   1722   for (int i = 0; i < 256; i++) {
   1723     State* s = RunStateOnByte(info->start, i);
   1724     if (s == NULL) {
   1725       ANNOTATE_HAPPENS_BEFORE(&info->firstbyte);
   1726       WriteMemoryBarrier();  // Synchronize with "quick check" above.
   1727       info->firstbyte = firstbyte;
   1728       return false;
   1729     }
   1730     if (s == info->start)
   1731       continue;
   1732     // Goes to new state...
   1733     if (firstbyte == kFbNone) {
   1734       firstbyte = i;        // ... first one
   1735     } else {
   1736       firstbyte = kFbMany;  // ... too many
   1737       break;
   1738     }
   1739   }
   1740   ANNOTATE_HAPPENS_BEFORE(&info->firstbyte);
   1741   WriteMemoryBarrier();  // Synchronize with "quick check" above.
   1742   info->firstbyte = firstbyte;
   1743   return true;
   1744 }
   1745 
   1746 // The actual DFA search: calls AnalyzeSearch and then FastSearchLoop.
   1747 bool DFA::Search(const StringPiece& text,
   1748                  const StringPiece& context,
   1749                  bool anchored,
   1750                  bool want_earliest_match,
   1751                  bool run_forward,
   1752                  bool* failed,
   1753                  const char** epp,
   1754                  vector<int>* matches) {
   1755   *epp = NULL;
   1756   if (!ok()) {
   1757     *failed = true;
   1758     return false;
   1759   }
   1760   *failed = false;
   1761 
   1762   if (DebugDFA) {
   1763     fprintf(stderr, "\nprogram:\n%s\n", prog_->DumpUnanchored().c_str());
   1764     fprintf(stderr, "text %s anchored=%d earliest=%d fwd=%d kind %d\n",
   1765             text.as_string().c_str(), anchored, want_earliest_match,
   1766             run_forward, kind_);
   1767   }
   1768 
   1769   RWLocker l(&cache_mutex_);
   1770   SearchParams params(text, context, &l);
   1771   params.anchored = anchored;
   1772   params.want_earliest_match = want_earliest_match;
   1773   params.run_forward = run_forward;
   1774   params.matches = matches;
   1775 
   1776   if (!AnalyzeSearch(&params)) {
   1777     *failed = true;
   1778     return false;
   1779   }
   1780   if (params.start == DeadState)
   1781     return false;
   1782   if (params.start == FullMatchState) {
   1783     if (run_forward == want_earliest_match)
   1784       *epp = text.begin();
   1785     else
   1786       *epp = text.end();
   1787     return true;
   1788   }
   1789   if (DebugDFA)
   1790     fprintf(stderr, "start %s\n", DumpState(params.start).c_str());
   1791   bool ret = FastSearchLoop(&params);
   1792   if (params.failed) {
   1793     *failed = true;
   1794     return false;
   1795   }
   1796   *epp = params.ep;
   1797   return ret;
   1798 }
   1799 
   1800 // Deletes dfa.
   1801 //
   1802 // This is a separate function so that
   1803 // prog.h can be used without moving the definition of
   1804 // class DFA out of this file.  If you set
   1805 //   prog->dfa_ = dfa;
   1806 // then you also have to set
   1807 //   prog->delete_dfa_ = DeleteDFA;
   1808 // so that ~Prog can delete the dfa.
   1809 static void DeleteDFA(DFA* dfa) {
   1810   delete dfa;
   1811 }
   1812 
   1813 DFA* Prog::GetDFA(MatchKind kind) {
   1814   DFA*volatile* pdfa;
   1815   if (kind == kFirstMatch || kind == kManyMatch) {
   1816     pdfa = &dfa_first_;
   1817   } else {
   1818     kind = kLongestMatch;
   1819     pdfa = &dfa_longest_;
   1820   }
   1821 
   1822   // Quick check; okay because of memory barrier below.
   1823   DFA *dfa = ANNOTATE_UNPROTECTED_READ(*pdfa);
   1824   if (dfa != NULL) {
   1825     ANNOTATE_HAPPENS_AFTER(dfa);
   1826     return dfa;
   1827   }
   1828 
   1829   MutexLock l(&dfa_mutex_);
   1830   dfa = *pdfa;
   1831   if (dfa != NULL) {
   1832     ANNOTATE_HAPPENS_AFTER(dfa);
   1833     return dfa;
   1834   }
   1835 
   1836   // For a forward DFA, half the memory goes to each DFA.
   1837   // For a reverse DFA, all the memory goes to the
   1838   // "longest match" DFA, because RE2 never does reverse
   1839   // "first match" searches.
   1840   int64 m = dfa_mem_/2;
   1841   if (reversed_) {
   1842     if (kind == kLongestMatch || kind == kManyMatch)
   1843       m = dfa_mem_;
   1844     else
   1845       m = 0;
   1846   }
   1847   dfa = new DFA(this, kind, m);
   1848   delete_dfa_ = DeleteDFA;
   1849 
   1850   // Synchronize with "quick check" above.
   1851   ANNOTATE_HAPPENS_BEFORE(dfa);
   1852   WriteMemoryBarrier();
   1853   *pdfa = dfa;
   1854 
   1855   return dfa;
   1856 }
   1857 
   1858 
   1859 // Executes the regexp program to search in text,
   1860 // which itself is inside the larger context.  (As a convenience,
   1861 // passing a NULL context is equivalent to passing text.)
   1862 // Returns true if a match is found, false if not.
   1863 // If a match is found, fills in match0->end() to point at the end of the match
   1864 // and sets match0->begin() to text.begin(), since the DFA can't track
   1865 // where the match actually began.
   1866 //
   1867 // This is the only external interface (class DFA only exists in this file).
   1868 //
   1869 bool Prog::SearchDFA(const StringPiece& text, const StringPiece& const_context,
   1870                      Anchor anchor, MatchKind kind,
   1871                      StringPiece* match0, bool* failed, vector<int>* matches) {
   1872   *failed = false;
   1873 
   1874   StringPiece context = const_context;
   1875   if (context.begin() == NULL)
   1876     context = text;
   1877   bool carat = anchor_start();
   1878   bool dollar = anchor_end();
   1879   if (reversed_) {
   1880     bool t = carat;
   1881     carat = dollar;
   1882     dollar = t;
   1883   }
   1884   if (carat && context.begin() != text.begin())
   1885     return false;
   1886   if (dollar && context.end() != text.end())
   1887     return false;
   1888 
   1889   // Handle full match by running an anchored longest match
   1890   // and then checking if it covers all of text.
   1891   bool anchored = anchor == kAnchored || anchor_start() || kind == kFullMatch;
   1892   bool endmatch = false;
   1893   if (kind == kManyMatch) {
   1894     endmatch = true;
   1895   } else if (kind == kFullMatch || anchor_end()) {
   1896     endmatch = true;
   1897     kind = kLongestMatch;
   1898   }
   1899 
   1900   // If the caller doesn't care where the match is (just whether one exists),
   1901   // then we can stop at the very first match we find, the so-called
   1902   // "shortest match".
   1903   bool want_shortest_match = false;
   1904   if (match0 == NULL && !endmatch) {
   1905     want_shortest_match = true;
   1906     kind = kLongestMatch;
   1907   }
   1908 
   1909   DFA* dfa = GetDFA(kind);
   1910   const char* ep;
   1911   bool matched = dfa->Search(text, context, anchored,
   1912                              want_shortest_match, !reversed_,
   1913                              failed, &ep, matches);
   1914   if (*failed)
   1915     return false;
   1916   if (!matched)
   1917     return false;
   1918   if (endmatch && ep != (reversed_ ? text.begin() : text.end()))
   1919     return false;
   1920 
   1921   // If caller cares, record the boundary of the match.
   1922   // We only know where it ends, so use the boundary of text
   1923   // as the beginning.
   1924   if (match0) {
   1925     if (reversed_)
   1926       *match0 = StringPiece(ep, text.end() - ep);
   1927     else
   1928       *match0 = StringPiece(text.begin(), ep - text.begin());
   1929   }
   1930   return true;
   1931 }
   1932 
   1933 // Build out all states in DFA.  Returns number of states.
   1934 int DFA::BuildAllStates() {
   1935   if (!ok())
   1936     return 0;
   1937 
   1938   // Pick out start state for unanchored search
   1939   // at beginning of text.
   1940   RWLocker l(&cache_mutex_);
   1941   SearchParams params(NULL, NULL, &l);
   1942   params.anchored = false;
   1943   if (!AnalyzeSearch(&params) || params.start <= SpecialStateMax)
   1944     return 0;
   1945 
   1946   // Add start state to work queue.
   1947   StateSet queued;
   1948   vector<State*> q;
   1949   queued.insert(params.start);
   1950   q.push_back(params.start);
   1951 
   1952   // Flood to expand every state.
   1953   for (int i = 0; i < q.size(); i++) {
   1954     State* s = q[i];
   1955     for (int c = 0; c < 257; c++) {
   1956       State* ns = RunStateOnByteUnlocked(s, c);
   1957       if (ns > SpecialStateMax && queued.find(ns) == queued.end()) {
   1958         queued.insert(ns);
   1959         q.push_back(ns);
   1960       }
   1961     }
   1962   }
   1963 
   1964   return q.size();
   1965 }
   1966 
   1967 // Build out all states in DFA for kind.  Returns number of states.
   1968 int Prog::BuildEntireDFA(MatchKind kind) {
   1969   //LOG(ERROR) << "BuildEntireDFA is only for testing.";
   1970   return GetDFA(kind)->BuildAllStates();
   1971 }
   1972 
   1973 // Computes min and max for matching string.
   1974 // Won't return strings bigger than maxlen.
   1975 bool DFA::PossibleMatchRange(string* min, string* max, int maxlen) {
   1976   if (!ok())
   1977     return false;
   1978 
   1979   // NOTE: if future users of PossibleMatchRange want more precision when
   1980   // presented with infinitely repeated elements, consider making this a
   1981   // parameter to PossibleMatchRange.
   1982   static int kMaxEltRepetitions = 0;
   1983 
   1984   // Keep track of the number of times we've visited states previously. We only
   1985   // revisit a given state if it's part of a repeated group, so if the value
   1986   // portion of the map tuple exceeds kMaxEltRepetitions we bail out and set
   1987   // |*max| to |PrefixSuccessor(*max)|.
   1988   //
   1989   // Also note that previously_visited_states[UnseenStatePtr] will, in the STL
   1990   // tradition, implicitly insert a '0' value at first use. We take advantage
   1991   // of that property below.
   1992   map<State*, int> previously_visited_states;
   1993 
   1994   // Pick out start state for anchored search at beginning of text.
   1995   RWLocker l(&cache_mutex_);
   1996   SearchParams params(NULL, NULL, &l);
   1997   params.anchored = true;
   1998   if (!AnalyzeSearch(&params))
   1999     return false;
   2000   if (params.start == DeadState) {  // No matching strings
   2001     *min = "";
   2002     *max = "";
   2003     return true;
   2004   }
   2005   if (params.start == FullMatchState)  // Every string matches: no max
   2006     return false;
   2007 
   2008   // The DFA is essentially a big graph rooted at params.start,
   2009   // and paths in the graph correspond to accepted strings.
   2010   // Each node in the graph has potentially 256+1 arrows
   2011   // coming out, one for each byte plus the magic end of
   2012   // text character kByteEndText.
   2013 
   2014   // To find the smallest possible prefix of an accepted
   2015   // string, we just walk the graph preferring to follow
   2016   // arrows with the lowest bytes possible.  To find the
   2017   // largest possible prefix, we follow the largest bytes
   2018   // possible.
   2019 
   2020   // The test for whether there is an arrow from s on byte j is
   2021   //    ns = RunStateOnByteUnlocked(s, j);
   2022   //    if (ns == NULL)
   2023   //      return false;
   2024   //    if (ns != DeadState && ns->ninst > 0)
   2025   // The RunStateOnByteUnlocked call asks the DFA to build out the graph.
   2026   // It returns NULL only if the DFA has run out of memory,
   2027   // in which case we can't be sure of anything.
   2028   // The second check sees whether there was graph built
   2029   // and whether it is interesting graph.  Nodes might have
   2030   // ns->ninst == 0 if they exist only to represent the fact
   2031   // that a match was found on the previous byte.
   2032 
   2033   // Build minimum prefix.
   2034   State* s = params.start;
   2035   min->clear();
   2036   for (int i = 0; i < maxlen; i++) {
   2037     if (previously_visited_states[s] > kMaxEltRepetitions) {
   2038       VLOG(2) << "Hit kMaxEltRepetitions=" << kMaxEltRepetitions
   2039         << " for state s=" << s << " and min=" << CEscape(*min);
   2040       break;
   2041     }
   2042     previously_visited_states[s]++;
   2043 
   2044     // Stop if min is a match.
   2045     State* ns = RunStateOnByteUnlocked(s, kByteEndText);
   2046     if (ns == NULL)  // DFA out of memory
   2047       return false;
   2048     if (ns != DeadState && (ns == FullMatchState || ns->IsMatch()))
   2049       break;
   2050 
   2051     // Try to extend the string with low bytes.
   2052     bool extended = false;
   2053     for (int j = 0; j < 256; j++) {
   2054       ns = RunStateOnByteUnlocked(s, j);
   2055       if (ns == NULL)  // DFA out of memory
   2056         return false;
   2057       if (ns == FullMatchState ||
   2058           (ns > SpecialStateMax && ns->ninst_ > 0)) {
   2059         extended = true;
   2060         min->append(1, j);
   2061         s = ns;
   2062         break;
   2063       }
   2064     }
   2065     if (!extended)
   2066       break;
   2067   }
   2068 
   2069   // Build maximum prefix.
   2070   previously_visited_states.clear();
   2071   s = params.start;
   2072   max->clear();
   2073   for (int i = 0; i < maxlen; i++) {
   2074     if (previously_visited_states[s] > kMaxEltRepetitions) {
   2075       VLOG(2) << "Hit kMaxEltRepetitions=" << kMaxEltRepetitions
   2076         << " for state s=" << s << " and max=" << CEscape(*max);
   2077       break;
   2078     }
   2079     previously_visited_states[s] += 1;
   2080 
   2081     // Try to extend the string with high bytes.
   2082     bool extended = false;
   2083     for (int j = 255; j >= 0; j--) {
   2084       State* ns = RunStateOnByteUnlocked(s, j);
   2085       if (ns == NULL)
   2086         return false;
   2087       if (ns == FullMatchState ||
   2088           (ns > SpecialStateMax && ns->ninst_ > 0)) {
   2089         extended = true;
   2090         max->append(1, j);
   2091         s = ns;
   2092         break;
   2093       }
   2094     }
   2095     if (!extended) {
   2096       // Done, no need for PrefixSuccessor.
   2097       return true;
   2098     }
   2099   }
   2100 
   2101   // Stopped while still adding to *max - round aaaaaaaaaa... to aaaa...b
   2102   *max = PrefixSuccessor(*max);
   2103 
   2104   // If there are no bytes left, we have no way to say "there is no maximum
   2105   // string".  We could make the interface more complicated and be able to
   2106   // return "there is no maximum but here is a minimum", but that seems like
   2107   // overkill -- the most common no-max case is all possible strings, so not
   2108   // telling the caller that the empty string is the minimum match isn't a
   2109   // great loss.
   2110   if (max->empty())
   2111     return false;
   2112 
   2113   return true;
   2114 }
   2115 
   2116 // PossibleMatchRange for a Prog.
   2117 bool Prog::PossibleMatchRange(string* min, string* max, int maxlen) {
   2118   DFA* dfa = NULL;
   2119   {
   2120     MutexLock l(&dfa_mutex_);
   2121     // Have to use dfa_longest_ to get all strings for full matches.
   2122     // For example, (a|aa) never matches aa in first-match mode.
   2123     if (dfa_longest_ == NULL) {
   2124       dfa_longest_ = new DFA(this, Prog::kLongestMatch, dfa_mem_/2);
   2125       delete_dfa_ = DeleteDFA;
   2126     }
   2127     dfa = dfa_longest_;
   2128   }
   2129   return dfa->PossibleMatchRange(min, max, maxlen);
   2130 }
   2131 
   2132 }  // namespace re2
   2133