Port and integrate the irregexp engine from V8

runtime/vm/regexp.cc

 // Copyright (c) 2014, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/regexp.h"
 
+#include "vm/dart_entry.h"
+#include "vm/regexp_assembler.h"
 #include "vm/regexp_ast.h"
+#include "vm/unibrow-inl.h"
 #include "vm/unicode.h"
 #include "vm/symbols.h"
 
 #define I isolate()
 #define CI compiler->isolate()
 
 namespace dart {
 
+DECLARE_FLAG(bool, trace_irregexp);
 
 #define DEFINE_ACCEPT(Type)                                          \
   void Type##Node::Accept(NodeVisitor* visitor) {                    \
     visitor->Visit##Type(this);                                      \
   }
 FOR_EACH_NODE_TYPE(DEFINE_ACCEPT)
 #undef DEFINE_ACCEPT
 
 
 // Default to generating optimized regexp code.
@@ skipping 12 matching lines @@
 static const intptr_t kSpaceRanges[] = { '\t', '\r' + 1, ' ', ' ' + 1,
     0x00A0, 0x00A1, 0x1680, 0x1681, 0x180E, 0x180F, 0x2000, 0x200B,
     0x2028, 0x202A, 0x202F, 0x2030, 0x205F, 0x2060, 0x3000, 0x3001,
     0xFEFF, 0xFF00, 0x10000 };
 static const intptr_t kSpaceRangeCount = ARRAY_SIZE(kSpaceRanges);
 static const intptr_t kWordRanges[] = {
     '0', '9' + 1, 'A', 'Z' + 1, '_', '_' + 1, 'a', 'z' + 1, 0x10000 };
 static const intptr_t kWordRangeCount = ARRAY_SIZE(kWordRanges);
 static const intptr_t kDigitRanges[] = { '0', '9' + 1, 0x10000 };
 static const intptr_t kDigitRangeCount = ARRAY_SIZE(kDigitRanges);
+static const intptr_t kSurrogateRanges[] = { 0xd800, 0xe000, 0x10000 };
+static const intptr_t kSurrogateRangeCount = ARRAY_SIZE(kSurrogateRanges);
 static const intptr_t kLineTerminatorRanges[] = {
     0x000A, 0x000B, 0x000D, 0x000E, 0x2028, 0x202A, 0x10000 };
 static const intptr_t kLineTerminatorRangeCount =
     ARRAY_SIZE(kLineTerminatorRanges);
 
 
 static inline void PrintUtf16(uint16_t c) {
   const char* format = (0x20 <= c && c <= 0x7F) ?
         "%c" : (c <= 0xff) ? "\\x%02x" : "\\u%04x";
   OS::Print(format, c);
@@ skipping 40 matching lines @@
     info->visited = true;
   }
   ~VisitMarker() {
     info_->visited = false;
   }
  private:
   NodeInfo* info_;
 };
 
 
+class FrequencyCollator : public ValueObject {
+ public:
+  FrequencyCollator() : total_samples_(0) {
+    for (intptr_t i = 0; i < RegExpMacroAssembler::kTableSize; i++) {
+      frequencies_[i] = CharacterFrequency(i);
+    }
+  }
+
+  void CountCharacter(intptr_t character) {
+    intptr_t index = (character & RegExpMacroAssembler::kTableMask);
+    frequencies_[index].Increment();
+    total_samples_++;
+  }
+
+  // Does not measure in percent, but rather per-128 (the table size from the
+  // regexp macro assembler).
+  intptr_t Frequency(intptr_t in_character) {
+    ASSERT((in_character & RegExpMacroAssembler::kTableMask) == in_character);
+    if (total_samples_ < 1) return 1;  // Division by zero.
+    intptr_t freq_in_per128 =
+        (frequencies_[in_character].counter() * 128) / total_samples_;
+    return freq_in_per128;
+  }
+
+ private:
+  class CharacterFrequency {
+   public:
+    CharacterFrequency() : counter_(0), character_(-1) { }
+    explicit CharacterFrequency(intptr_t character)
+        : counter_(0), character_(character) { }
+
+    void Increment() { counter_++; }
+    intptr_t counter() { return counter_; }
+    intptr_t character() { return character_; }
+
+   private:
+    intptr_t counter_;
+    intptr_t character_;
+  };
+
+
+ private:
+  CharacterFrequency frequencies_[RegExpMacroAssembler::kTableSize];
+  intptr_t total_samples_;
+};
+
+
 class RegExpCompiler {
  public:
-  RegExpCompiler(intptr_t capture_count, bool ignore_case, bool is_ascii);
+  RegExpCompiler(intptr_t capture_count,
+                 bool ignore_case,
+                 intptr_t specialization_cid);
 
   intptr_t AllocateRegister() {
     return next_register_++;
   }
 
-  RegExpEngine::CompilationResult Assemble(RegExpMacroAssembler* assembler,
+  RegExpEngine::CompilationResult Assemble(IRRegExpMacroAssembler* assembler,
                                            RegExpNode* start,
                                            intptr_t capture_count,
                                            const String& pattern);
 
+  inline void AddWork(RegExpNode* node) { work_list_->Add(node); }
+
   static const intptr_t kImplementationOffset = 0;
   static const intptr_t kNumberOfRegistersOffset = 0;
   static const intptr_t kCodeOffset = 1;
 
-  RegExpMacroAssembler* macro_assembler() { return macro_assembler_; }
+  IRRegExpMacroAssembler* macro_assembler() { return macro_assembler_; }
   EndNode* accept() { return accept_; }
 
   static const intptr_t kMaxRecursion = 100;
   inline intptr_t recursion_depth() { return recursion_depth_; }
   inline void IncrementRecursionDepth() { recursion_depth_++; }
   inline void DecrementRecursionDepth() { recursion_depth_--; }
 
   void SetRegExpTooBig() { reg_exp_too_big_ = true; }
 
   inline bool ignore_case() { return ignore_case_; }
-  inline bool ascii() { return ascii_; }
+  inline bool ascii() const {
+    return (specialization_cid_ == kOneByteStringCid ||
+            specialization_cid_ == kExternalOneByteStringCid);
+  }
+  inline intptr_t specialization_cid() { return specialization_cid_; }
+  FrequencyCollator* frequency_collator() { return &frequency_collator_; }
 
   intptr_t current_expansion_factor() { return current_expansion_factor_; }
   void set_current_expansion_factor(intptr_t value) {
     current_expansion_factor_ = value;
   }
 
   Isolate* isolate() const { return isolate_; }
 
   static const intptr_t kNoRegister = -1;
 
  private:
   EndNode* accept_;
   intptr_t next_register_;
+  ZoneGrowableArray<RegExpNode*>* work_list_;
   intptr_t recursion_depth_;
-  RegExpMacroAssembler* macro_assembler_;
+  IRRegExpMacroAssembler* macro_assembler_;
   bool ignore_case_;
-  bool ascii_;
+  intptr_t specialization_cid_;
   bool reg_exp_too_big_;
   intptr_t current_expansion_factor_;
+  FrequencyCollator frequency_collator_;
   Isolate* isolate_;
 };
 
 
+class RecursionCheck : public ValueObject {
+ public:
+  explicit RecursionCheck(RegExpCompiler* compiler) : compiler_(compiler) {
+    compiler->IncrementRecursionDepth();
+  }
+  ~RecursionCheck() { compiler_->DecrementRecursionDepth(); }
+ private:
+  RegExpCompiler* compiler_;
+};
+
+
 // Scoped object to keep track of how much we unroll quantifier loops in the
 // regexp graph generator.
-class RegExpExpansionLimiter {
+class RegExpExpansionLimiter : public ValueObject {
  public:
   static const intptr_t kMaxExpansionFactor = 6;
   RegExpExpansionLimiter(RegExpCompiler* compiler, intptr_t factor)
       : compiler_(compiler),
         saved_expansion_factor_(compiler->current_expansion_factor()),
         ok_to_expand_(saved_expansion_factor_ <= kMaxExpansionFactor) {
     ASSERT(factor > 0);
     if (ok_to_expand_) {
       if (factor > kMaxExpansionFactor) {
         // Avoid integer overflow of the current expansion factor.
@@ skipping 531 matching lines @@
   // we don't need to check it.
   if (neg_replacement == NULL) return set_replacement(replacement);
   (*alternatives_)[0].set_node(neg_replacement);
   return set_replacement(this);
 }
 
 
 // Code emission ---------------------------------------------------------------
 
 
-#define DEFINE_EMIT(Type)                                           \
-  void Type##Node::Emit(RegExpCompiler* compiler, Trace* trace) {   \
-    UNIMPLEMENTED();                                                  \
-  }
-FOR_EACH_NODE_TYPE(DEFINE_EMIT)
-DEFINE_EMIT(LoopChoice)
-void NegativeSubmatchSuccess::Emit(RegExpCompiler* compiler,
-                                   Trace* trace) {
+void ChoiceNode::GenerateGuard(RegExpMacroAssembler* macro_assembler,
+                               Guard* guard,
+                               Trace* trace) {
+  switch (guard->op()) {
+    case Guard::LT:
+      ASSERT(!trace->mentions_reg(guard->reg()));
+      macro_assembler->IfRegisterGE(guard->reg(),
+                                    guard->value(),
+                                    trace->backtrack());
+      break;
+    case Guard::GEQ:
+      ASSERT(!trace->mentions_reg(guard->reg()));
+      macro_assembler->IfRegisterLT(guard->reg(),
+                                    guard->value(),
+                                    trace->backtrack());
+      break;
+  }
+}
+
+
+void NegativeSubmatchSuccess::Emit(RegExpCompiler* compiler, Trace* trace) {
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+
+  // Omit flushing the trace. We discard the entire stack frame anyway.
+
+  if (!label()->IsBound()) {
+    // We are completely independent of the trace, since we ignore it,
+    // so this code can be used as the generic version.
+    assembler->BindBlock(label());
+  }
+
+  // Throw away everything on the backtrack stack since the start
+  // of the negative submatch and restore the character position.
+  assembler->ReadCurrentPositionFromRegister(current_position_register_);
+  assembler->ReadStackPointerFromRegister(stack_pointer_register_);
+  if (clear_capture_count_ > 0) {
+    // Clear any captures that might have been performed during the success
+    // of the body of the negative look-ahead.
+    intptr_t clear_capture_end =
+        clear_capture_start_ + clear_capture_count_ - 1;
+    assembler->ClearRegisters(clear_capture_start_, clear_capture_end);
+  }
+  // Now that we have unwound the stack we find at the top of the stack the
+  // backtrack that the BeginSubmatch node got.
+  assembler->Backtrack();
+}
+
+
+bool Trace::GetStoredPosition(intptr_t reg, intptr_t* cp_offset) {
+  ASSERT(*cp_offset == 0);
+  for (DeferredAction* action = actions_;
+       action != NULL;
+       action = action->next()) {
+    if (action->Mentions(reg)) {
+      if (action->action_type() == ActionNode::STORE_POSITION) {
+        *cp_offset = static_cast<DeferredCapture*>(action)->cp_offset();
+        return true;
+      } else {
+        return false;
+      }
+    }
+  }
+  return false;
+}
+
+
+// This is called as we come into a loop choice node and some other tricky
+// nodes.  It normalizes the state of the code generator to ensure we can
+// generate generic code.
+intptr_t Trace::FindAffectedRegisters(OutSet* affected_registers,
+                                      Isolate* isolate) {
+  intptr_t max_register = RegExpCompiler::kNoRegister;
+  for (DeferredAction* action = actions_;
+       action != NULL;
+       action = action->next()) {
+    if (action->action_type() == ActionNode::CLEAR_CAPTURES) {
+      Interval range = static_cast<DeferredClearCaptures*>(action)->range();
+      for (intptr_t i = range.from(); i <= range.to(); i++)
+        affected_registers->Set(i, isolate);
+      if (range.to() > max_register) max_register = range.to();
+    } else {
+      affected_registers->Set(action->reg(), isolate);
+      if (action->reg() > max_register) max_register = action->reg();
+    }
+  }
+  return max_register;
+}
+
+
+bool Trace::DeferredAction::Mentions(intptr_t that) {
+  if (action_type() == ActionNode::CLEAR_CAPTURES) {
+    Interval range = static_cast<DeferredClearCaptures*>(this)->range();
+    return range.Contains(that);
+  } else {
+    return reg() == that;
+  }
+}
+
+
+bool Trace::mentions_reg(intptr_t reg) {
+  for (DeferredAction* action = actions_;
+       action != NULL;
+       action = action->next()) {
+    if (action->Mentions(reg))
+      return true;
+  }
+  return false;
+}
+
+
+void Trace::RestoreAffectedRegisters(RegExpMacroAssembler* assembler,
+                                     intptr_t max_register,
+                                     const OutSet& registers_to_pop,
+                                     const OutSet& registers_to_clear) {
+  for (intptr_t reg = max_register; reg >= 0; reg--) {
+    if (registers_to_pop.Get(reg)) {
+      assembler->PopRegister(reg);
+    } else if (registers_to_clear.Get(reg)) {
+      intptr_t clear_to = reg;
+      while (reg > 0 && registers_to_clear.Get(reg - 1)) {
+        reg--;
+      }
+      assembler->ClearRegisters(reg, clear_to);
+    }
+  }
+}
+
+
+void Trace::PerformDeferredActions(RegExpMacroAssembler* assembler,
+                                   intptr_t max_register,
+                                   const OutSet& affected_registers,
+                                   OutSet* registers_to_pop,
+                                   OutSet* registers_to_clear,
+                                   Isolate* isolate) {
+  // The "+1" is to avoid a push_limit of zero if stack_limit_slack() is 1.
+  const intptr_t push_limit = (assembler->stack_limit_slack() + 1) / 2;
+
+  // Count pushes performed to force a stack limit check occasionally.
+  intptr_t pushes = 0;
+
+  for (intptr_t reg = 0; reg <= max_register; reg++) {
+    if (!affected_registers.Get(reg)) {
+      continue;
+    }
+
+    // The chronologically first deferred action in the trace
+    // is used to infer the action needed to restore a register
+    // to its previous state (or not, if it's safe to ignore it).
+    enum DeferredActionUndoType { IGNORE, RESTORE, CLEAR };
+    DeferredActionUndoType undo_action = IGNORE;
+
+    intptr_t value = 0;
+    bool absolute = false;
+    bool clear = false;
+    intptr_t store_position = -1;
+    // This is a little tricky because we are scanning the actions in reverse
+    // historical order (newest first).
+    for (DeferredAction* action = actions_;
+         action != NULL;
+         action = action->next()) {
+      if (action->Mentions(reg)) {
+        switch (action->action_type()) {
+          case ActionNode::SET_REGISTER: {
+            Trace::DeferredSetRegister* psr =
+                static_cast<Trace::DeferredSetRegister*>(action);
+            if (!absolute) {
+              value += psr->value();
+              absolute = true;
+            }
+            // SET_REGISTER is currently only used for newly introduced loop
+            // counters. They can have a significant previous value if they
+            // occour in a loop. TODO(lrn): Propagate this information, so
+            // we can set undo_action to IGNORE if we know there is no value to
+            // restore.
+            undo_action = RESTORE;
+            ASSERT(store_position == -1);
+            ASSERT(!clear);
+            break;
+          }
+          case ActionNode::INCREMENT_REGISTER:
+            if (!absolute) {
+              value++;
+            }
+            ASSERT(store_position == -1);
+            ASSERT(!clear);
+            undo_action = RESTORE;
+            break;
+          case ActionNode::STORE_POSITION: {
+            Trace::DeferredCapture* pc =
+                static_cast<Trace::DeferredCapture*>(action);
+            if (!clear && store_position == -1) {
+              store_position = pc->cp_offset();
+            }
+
+            // For captures we know that stores and clears alternate.
+            // Other register, are never cleared, and if the occur
+            // inside a loop, they might be assigned more than once.
+            if (reg <= 1) {
+              // Registers zero and one, aka "capture zero", is
+              // always set correctly if we succeed. There is no
+              // need to undo a setting on backtrack, because we
+              // will set it again or fail.
+              undo_action = IGNORE;
+            } else {
+              undo_action = pc->is_capture() ? CLEAR : RESTORE;
+            }
+            ASSERT(!absolute);
+            ASSERT(value == 0);
+            break;
+          }
+          case ActionNode::CLEAR_CAPTURES: {
+            // Since we're scanning in reverse order, if we've already
+            // set the position we have to ignore historically earlier
+            // clearing operations.
+            if (store_position == -1) {
+              clear = true;
+            }
+            undo_action = RESTORE;
+            ASSERT(!absolute);
+            ASSERT(value == 0);
+            break;
+          }
+          default:
+            UNREACHABLE();
+            break;
+        }
+      }
+    }
+    // Prepare for the undo-action (e.g., push if it's going to be popped).
+    if (undo_action == RESTORE) {
+      pushes++;
+      RegExpMacroAssembler::StackCheckFlag stack_check =
+          RegExpMacroAssembler::kNoStackLimitCheck;
+      if (pushes == push_limit) {
+        stack_check = RegExpMacroAssembler::kCheckStackLimit;
+        pushes = 0;
+      }
+
+      assembler->PushRegister(reg, stack_check);
+      registers_to_pop->Set(reg, isolate);
+    } else if (undo_action == CLEAR) {
+      registers_to_clear->Set(reg, isolate);
+    }
+    // Perform the chronologically last action (or accumulated increment)
+    // for the register.
+    if (store_position != -1) {
+      assembler->WriteCurrentPositionToRegister(reg, store_position);
+    } else if (clear) {
+      assembler->ClearRegisters(reg, reg);
+    } else if (absolute) {
+      assembler->SetRegister(reg, value);
+    } else if (value != 0) {
+      assembler->AdvanceRegister(reg, value);
+    }
+  }
+}
+
+
+void Trace::Flush(RegExpCompiler* compiler, RegExpNode* successor) {
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+
+  ASSERT(!is_trivial());
+
+  if (actions_ == NULL && backtrack() == NULL) {
+    // Here we just have some deferred cp advances to fix and we are back to
+    // a normal situation.  We may also have to forget some information gained
+    // through a quick check that was already performed.
+    if (cp_offset_ != 0) assembler->AdvanceCurrentPosition(cp_offset_);
+    // Create a new trivial state and generate the node with that.
+    Trace new_state;
+    successor->Emit(compiler, &new_state);
+    return;
+  }
+
+  // Generate deferred actions here along with code to undo them again.
+  OutSet affected_registers;
+
+  if (backtrack() != NULL) {
+    // Here we have a concrete backtrack location.  These are set up by choice
+    // nodes and so they indicate that we have a deferred save of the current
+    // position which we may need to emit here.
+    assembler->PushCurrentPosition();
+  }
+
+  intptr_t max_register = FindAffectedRegisters(&affected_registers, CI);
+  OutSet registers_to_pop;
+  OutSet registers_to_clear;
+  PerformDeferredActions(assembler,
+                         max_register,
+                         affected_registers,
+                         &registers_to_pop,
+                         &registers_to_clear,
+                         CI);
+  if (cp_offset_ != 0) {
+    assembler->AdvanceCurrentPosition(cp_offset_);
+  }
+
+  // Create a new trivial state and generate the node with that.
+  BlockLabel undo;
+  assembler->PushBacktrack(&undo);
+  Trace new_state;
+  successor->Emit(compiler, &new_state);
+
+  // On backtrack we need to restore state.
+  assembler->BindBlock(&undo);
+  RestoreAffectedRegisters(assembler,
+                           max_register,
+                           registers_to_pop,
+                           registers_to_clear);
+  if (backtrack() == NULL) {
+    assembler->Backtrack();
+  } else {
+    assembler->PopCurrentPosition();
+    assembler->GoTo(backtrack());
+  }
+}
+
+
+void Trace::InvalidateCurrentCharacter() {
+  characters_preloaded_ = 0;
+}
+
+
+void Trace::AdvanceCurrentPositionInTrace(intptr_t by,
+                                          RegExpCompiler* compiler) {
+  ASSERT(by > 0);
+  // We don't have an instruction for shifting the current character register
+  // down or for using a shifted value for anything so lets just forget that
+  // we preloaded any characters into it.
+  characters_preloaded_ = 0;
+  // Adjust the offsets of the quick check performed information.  This
+  // information is used to find out what we already determined about the
+  // characters by means of mask and compare.
+  quick_check_performed_.Advance(by, compiler->ascii());
+  cp_offset_ += by;
+  if (cp_offset_ > RegExpMacroAssembler::kMaxCPOffset) {
+    compiler->SetRegExpTooBig();
+    cp_offset_ = 0;
+  }
+  bound_checked_up_to_ = Utils::Maximum(static_cast<intptr_t>(0),
+                                        bound_checked_up_to_ - by);
+}
+
+
+void EndNode::Emit(RegExpCompiler* compiler, Trace* trace) {
+  if (!trace->is_trivial()) {
+    trace->Flush(compiler, this);
+    return;
+  }
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  if (!label()->IsBound()) {
+    assembler->BindBlock(label());
+  }
+  switch (action_) {
+    case ACCEPT:
+      assembler->Succeed();
+      return;
+    case BACKTRACK:
+      assembler->GoTo(trace->backtrack());
+      return;
+    case NEGATIVE_SUBMATCH_SUCCESS:
+      // This case is handled in a different virtual method.
+      UNREACHABLE();
+  }
   UNIMPLEMENTED();
 }
-#undef DEFINE_EMIT
-
-
-#define DEFINE_BMINFO(Type)                                         \
-  void Type##Node::FillInBMInfo(intptr_t initial_offset,            \
-                                intptr_t budget,                    \
-                                BoyerMooreLookahead* bm,            \
-                                bool not_at_start) {                \
-    UNIMPLEMENTED();                                                \
-  }
-FOR_EACH_NODE_TYPE(DEFINE_BMINFO)
-DEFINE_BMINFO(LoopChoice)
-#undef DEFINE_BMINFO
+
+
+bool QuickCheckDetails::Rationalize(bool asc) {
+  bool found_useful_op = false;
+  uint32_t char_mask;
+  if (asc) {
+    char_mask = Symbols::kMaxOneCharCodeSymbol;
+  } else {
+    char_mask = Utf16::kMaxCodeUnit;
+  }
+  mask_ = 0;
+  value_ = 0;
+  intptr_t char_shift = 0;
+  for (intptr_t i = 0; i < characters_; i++) {
+    Position* pos = &positions_[i];
+    if ((pos->mask & Symbols::kMaxOneCharCodeSymbol) != 0) {
+      found_useful_op = true;
+    }
+    mask_ |= (pos->mask & char_mask) << char_shift;
+    value_ |= (pos->value & char_mask) << char_shift;
+    char_shift += asc ? 8 : 16;
+  }
+  return found_useful_op;
+}
+
+
+bool RegExpNode::EmitQuickCheck(RegExpCompiler* compiler,
+                                Trace* trace,
+                                bool preload_has_checked_bounds,
+                                BlockLabel* on_possible_success,
+                                QuickCheckDetails* details,
+                                bool fall_through_on_failure) {
+  if (details->characters() == 0) return false;
+  GetQuickCheckDetails(
+      details, compiler, 0, trace->at_start() == Trace::FALSE_VALUE);
+  if (details->cannot_match()) return false;
+  if (!details->Rationalize(compiler->ascii())) return false;
+  ASSERT(details->characters() == 1 ||
+         compiler->macro_assembler()->CanReadUnaligned());
+  uint32_t mask = details->mask();
+  uint32_t value = details->value();
+
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+
+  if (trace->characters_preloaded() != details->characters()) {
+    assembler->LoadCurrentCharacter(trace->cp_offset(),
+                                    trace->backtrack(),
+                                    !preload_has_checked_bounds,
+                                    details->characters());
+  }
+
+
+  bool need_mask = true;
+
+  if (details->characters() == 1) {
+    // If number of characters preloaded is 1 then we used a byte or 16 bit
+    // load so the value is already masked down.
+    uint32_t char_mask;
+    if (compiler->ascii()) {
+      char_mask = Symbols::kMaxOneCharCodeSymbol;
+    } else {
+      char_mask = Utf16::kMaxCodeUnit;
+    }
+    if ((mask & char_mask) == char_mask) need_mask = false;
+    mask &= char_mask;
+  } else {
+    // For 2-character preloads in ASCII mode or 1-character preloads in
+    // TWO_BYTE mode we also use a 16 bit load with zero extend.
+    if (details->characters() == 2 && compiler->ascii()) {
+      if ((mask & 0xffff) == 0xffff) need_mask = false;
+    } else if (details->characters() == 1 && !compiler->ascii()) {
+      if ((mask & 0xffff) == 0xffff) need_mask = false;
+    } else {
+      if (mask == 0xffffffff) need_mask = false;
+    }
+  }
+
+  if (fall_through_on_failure) {
+    if (need_mask) {
+      assembler->CheckCharacterAfterAnd(value, mask, on_possible_success);
+    } else {
+      assembler->CheckCharacter(value, on_possible_success);
+    }
+  } else {
+    if (need_mask) {
+      assembler->CheckNotCharacterAfterAnd(value, mask, trace->backtrack());
+    } else {
+      assembler->CheckNotCharacter(value, trace->backtrack());
+    }
+  }
+  return true;
+}
+
+
+// Emit the code to check for a ^ in multiline mode (1-character lookbehind
+// that matches newline or the start of input).
+static void EmitHat(RegExpCompiler* compiler,
+                    RegExpNode* on_success,
+                    Trace* trace) {
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  // We will be loading the previous character into the current character
+  // register.
+  Trace new_trace(*trace);
+  new_trace.InvalidateCurrentCharacter();
+
+  BlockLabel ok;
+  if (new_trace.cp_offset() == 0) {
+    // The start of input counts as a newline in this context, so skip to
+    // ok if we are at the start.
+    assembler->CheckAtStart(&ok);
+  }
+  // We already checked that we are not at the start of input so it must be
+  // OK to load the previous character.
+  assembler->LoadCurrentCharacter(new_trace.cp_offset() -1,
+                                  new_trace.backtrack(),
+                                  false);
+  if (!assembler->CheckSpecialCharacterClass('n',
+                                             new_trace.backtrack())) {
+    // Newline means \n, \r, 0x2028 or 0x2029.
+    if (!compiler->ascii()) {
+      assembler->CheckCharacterAfterAnd(0x2028, 0xfffe, &ok);
+    }
+    assembler->CheckCharacter('\n', &ok);
+    assembler->CheckNotCharacter('\r', new_trace.backtrack());
+  }
+  assembler->BindBlock(&ok);
+  on_success->Emit(compiler, &new_trace);
+}
+
+
+void AssertionNode::Emit(RegExpCompiler* compiler, Trace* trace) {
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  switch (assertion_type_) {
+    case AT_END: {
+      BlockLabel ok;
+      assembler->CheckPosition(trace->cp_offset(), &ok);
+      assembler->GoTo(trace->backtrack());
+      assembler->BindBlock(&ok);
+      break;
+    }
+    case AT_START: {
+      if (trace->at_start() == Trace::FALSE_VALUE) {
+        assembler->GoTo(trace->backtrack());
+        return;
+      }
+      if (trace->at_start() == Trace::UNKNOWN) {
+        assembler->CheckNotAtStart(trace->backtrack());
+        Trace at_start_trace = *trace;
+        at_start_trace.set_at_start(true);
+        on_success()->Emit(compiler, &at_start_trace);
+        return;
+      }
+    }
+    break;
+    case AFTER_NEWLINE:
+      EmitHat(compiler, on_success(), trace);
+      return;
+    case AT_BOUNDARY:
+    case AT_NON_BOUNDARY: {
+      EmitBoundaryCheck(compiler, trace);
+      return;
+    }
+  }
+  on_success()->Emit(compiler, trace);
+}
+
+
+RegExpNode::LimitResult RegExpNode::LimitVersions(RegExpCompiler* compiler,
+                                                  Trace* trace) {
+  // If we are generating a greedy loop then don't stop and don't reuse code.
+  if (trace->stop_node() != NULL) {
+    return CONTINUE;
+  }
+
+  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
+  if (trace->is_trivial()) {
+    if (label_.IsBound()) {
+      // We are being asked to generate a generic version, but that's already
+      // been done so just go to it.
+      macro_assembler->GoTo(&label_);
+      return DONE;
+    }
+    if (compiler->recursion_depth() >= RegExpCompiler::kMaxRecursion) {
+      // To avoid too deep recursion we push the node to the work queue and just
+      // generate a goto here.
+      compiler->AddWork(this);
+      macro_assembler->GoTo(&label_);
+      return DONE;
+    }
+    // Generate generic version of the node and bind the label for later use.
+    macro_assembler->BindBlock(&label_);
+    return CONTINUE;
+  }
+
+  // We are being asked to make a non-generic version.  Keep track of how many
+  // non-generic versions we generate so as not to overdo it.
+  trace_count_++;
+  if (kRegexpOptimization &&
+      trace_count_ < kMaxCopiesCodeGenerated &&
+      compiler->recursion_depth() <= RegExpCompiler::kMaxRecursion) {
+    return CONTINUE;
+  }
+
+  // If we get here code has been generated for this node too many times or
+  // recursion is too deep.  Time to switch to a generic version.  The code for
+  // generic versions above can handle deep recursion properly.
+  trace->Flush(compiler, this);
+  return DONE;
+}
+
+
+// This generates the code to match a text node.  A text node can contain
+// straight character sequences (possibly to be matched in a case-independent
+// way) and character classes.  For efficiency we do not do this in a single
+// pass from left to right.  Instead we pass over the text node several times,
+// emitting code for some character positions every time.  See the comment on
+// TextEmitPass for details.
+void TextNode::Emit(RegExpCompiler* compiler, Trace* trace) {
+  LimitResult limit_result = LimitVersions(compiler, trace);
+  if (limit_result == DONE) return;
+  ASSERT(limit_result == CONTINUE);
+
+  if (trace->cp_offset() + Length() > RegExpMacroAssembler::kMaxCPOffset) {
+    compiler->SetRegExpTooBig();
+    return;
+  }
+
+  if (compiler->ascii()) {
+    intptr_t dummy = 0;
+    TextEmitPass(compiler, NON_ASCII_MATCH, false, trace, false, &dummy);
+  }
+
+  bool first_elt_done = false;
+  intptr_t bound_checked_to = trace->cp_offset() - 1;
+  bound_checked_to += trace->bound_checked_up_to();
+
+  // If a character is preloaded into the current character register then
+  // check that now.
+  if (trace->characters_preloaded() == 1) {
+    for (intptr_t pass = kFirstRealPass; pass <= kLastPass; pass++) {
+      if (!SkipPass(pass, compiler->ignore_case())) {
+        TextEmitPass(compiler,
+                     static_cast<TextEmitPassType>(pass),
+                     true,
+                     trace,
+                     false,
+                     &bound_checked_to);
+      }
+    }
+    first_elt_done = true;
+  }
+
+  for (intptr_t pass = kFirstRealPass; pass <= kLastPass; pass++) {
+    if (!SkipPass(pass, compiler->ignore_case())) {
+      TextEmitPass(compiler,
+                   static_cast<TextEmitPassType>(pass),
+                   false,
+                   trace,
+                   first_elt_done,
+                   &bound_checked_to);
+    }
+  }
+
+  Trace successor_trace(*trace);
+  successor_trace.set_at_start(false);
+  successor_trace.AdvanceCurrentPositionInTrace(Length(), compiler);
+  RecursionCheck rc(compiler);
+  on_success()->Emit(compiler, &successor_trace);
+}
+
+
+void LoopChoiceNode::Emit(RegExpCompiler* compiler, Trace* trace) {
+  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
+  if (trace->stop_node() == this) {
+    intptr_t text_length =
+        GreedyLoopTextLengthForAlternative(&((*alternatives_)[0]));
+    ASSERT(text_length != kNodeIsTooComplexForGreedyLoops);
+    // Update the counter-based backtracking info on the stack.  This is an
+    // optimization for greedy loops (see below).
+    ASSERT(trace->cp_offset() == text_length);
+    macro_assembler->AdvanceCurrentPosition(text_length);
+    macro_assembler->GoTo(trace->loop_label());
+    return;
+  }
+  ASSERT(trace->stop_node() == NULL);
+  if (!trace->is_trivial()) {
+    trace->Flush(compiler, this);
+    return;
+  }
+  ChoiceNode::Emit(compiler, trace);
+}
+
+
+// This class is used when generating the alternatives in a choice node.  It
+// records the way the alternative is being code generated.
+class AlternativeGeneration {
+ public:
+  AlternativeGeneration()
+      : possible_success(),
+        expects_preload(false),
+        after(),
+        quick_check_details() { }
+  BlockLabel possible_success;
+  bool expects_preload;
+  BlockLabel after;
+  QuickCheckDetails quick_check_details;
+};
+
+
+// Creates a list of AlternativeGenerations.  If the list has a reasonable
+// size then it is on the stack, otherwise the excess is on the heap.
+class AlternativeGenerationList {
+ public:
+  explicit AlternativeGenerationList(intptr_t count)
+      : alt_gens_(count) {
+    for (intptr_t i = 0; i < count && i < kAFew; i++) {
+      alt_gens_.Add(a_few_alt_gens_ + i);
+    }
+    for (intptr_t i = kAFew; i < count; i++) {
+      alt_gens_.Add(new AlternativeGeneration());
+    }
+  }
+  ~AlternativeGenerationList() {
+    for (intptr_t i = kAFew; i < alt_gens_.length(); i++) {
+      delete alt_gens_[i];
+      alt_gens_[i] = NULL;
+    }
+  }
+
+  AlternativeGeneration* at(intptr_t i) {
+    return alt_gens_[i];
+  }
+
+ private:
+  static const intptr_t kAFew = 10;
+  GrowableArray<AlternativeGeneration*> alt_gens_;
+  AlternativeGeneration a_few_alt_gens_[kAFew];
+};
+
+
+void ChoiceNode::Emit(RegExpCompiler* compiler, Trace* trace) {
+  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
+  intptr_t choice_count = alternatives_->length();
+
+#ifdef DEBUG
+  for (intptr_t i = 0; i < choice_count - 1; i++) {
+    GuardedAlternative alternative = alternatives_->At(i);
+    ZoneGrowableArray<Guard*>* guards = alternative.guards();
+    intptr_t guard_count = (guards == NULL) ? 0 : guards->length();
+    for (intptr_t j = 0; j < guard_count; j++) {
+      ASSERT(!trace->mentions_reg(guards->At(j)->reg()));
+    }
+  }
+#endif
+
+  LimitResult limit_result = LimitVersions(compiler, trace);
+  if (limit_result == DONE) return;
+  ASSERT(limit_result == CONTINUE);
+
+  intptr_t new_flush_budget = trace->flush_budget() / choice_count;
+  if (trace->flush_budget() == 0 && trace->actions() != NULL) {
+    trace->Flush(compiler, this);
+    return;
+  }
+
+  RecursionCheck rc(compiler);
+
+  Trace* current_trace = trace;
+
+  intptr_t text_length =
+      GreedyLoopTextLengthForAlternative(&((*alternatives_)[0]));
+  bool greedy_loop = false;
+  BlockLabel greedy_loop_label;
+  Trace counter_backtrack_trace;
+  counter_backtrack_trace.set_backtrack(&greedy_loop_label);
+  if (not_at_start()) counter_backtrack_trace.set_at_start(false);
+
+  if (choice_count > 1 && text_length != kNodeIsTooComplexForGreedyLoops) {
+    // Here we have special handling for greedy loops containing only text nodes
+    // and other simple nodes.  These are handled by pushing the current
+    // position on the stack and then incrementing the current position each
+    // time around the switch.  On backtrack we decrement the current position
+    // and check it against the pushed value.  This avoids pushing backtrack
+    // information for each iteration of the loop, which could take up a lot of
+    // space.
+    greedy_loop = true;
+    ASSERT(trace->stop_node() == NULL);
+    macro_assembler->PushCurrentPosition();
+    current_trace = &counter_backtrack_trace;
+    BlockLabel greedy_match_failed;
+    Trace greedy_match_trace;
+    if (not_at_start()) greedy_match_trace.set_at_start(false);
+    greedy_match_trace.set_backtrack(&greedy_match_failed);
+    BlockLabel loop_label;
+    macro_assembler->BindBlock(&loop_label);
+    greedy_match_trace.set_stop_node(this);
+    greedy_match_trace.set_loop_label(&loop_label);
+    (*alternatives_)[0].node()->Emit(compiler, &greedy_match_trace);
+    macro_assembler->BindBlock(&greedy_match_failed);
+  }
+
+  BlockLabel second_choice;  // For use in greedy matches.
+  macro_assembler->BindBlock(&second_choice);
+
+  intptr_t first_normal_choice = greedy_loop ? 1 : 0;
+
+  bool not_at_start = current_trace->at_start() == Trace::FALSE_VALUE;
+  const intptr_t kEatsAtLeastNotYetInitialized = -1;
+  intptr_t eats_at_least = kEatsAtLeastNotYetInitialized;
+
+  bool skip_was_emitted = false;
+
+  if (!greedy_loop && choice_count == 2) {
+    GuardedAlternative alt1 = (*alternatives_)[1];
+    if (alt1.guards() == NULL || alt1.guards()->length() == 0) {
+      RegExpNode* eats_anything_node = alt1.node();
+      if (eats_anything_node->GetSuccessorOfOmnivorousTextNode(compiler) ==
+          this) {
+        // At this point we know that we are at a non-greedy loop that will eat
+        // any character one at a time.  Any non-anchored regexp has such a
+        // loop prepended to it in order to find where it starts.  We look for
+        // a pattern of the form ...abc... where we can look 6 characters ahead
+        // and step forwards 3 if the character is not one of abc.  Abc need
+        // not be atoms, they can be any reasonably limited character class or
+        // small alternation.
+        ASSERT(trace->is_trivial());  // This is the case on LoopChoiceNodes.
+        BoyerMooreLookahead* lookahead = bm_info(not_at_start);
+        if (lookahead == NULL) {
+          eats_at_least = Utils::Minimum(kMaxLookaheadForBoyerMoore,
+                              EatsAtLeast(kMaxLookaheadForBoyerMoore,
+                                          kRecursionBudget,
+                                          not_at_start));
+          if (eats_at_least >= 1) {
+            BoyerMooreLookahead* bm =
+                new(I) BoyerMooreLookahead(eats_at_least, compiler, I);
+            GuardedAlternative alt0 = alternatives_->At(0);
+            alt0.node()->FillInBMInfo(0, kRecursionBudget, bm, not_at_start);
+            skip_was_emitted = bm->EmitSkipInstructions(macro_assembler);
+          }
+        } else {
+          skip_was_emitted = lookahead->EmitSkipInstructions(macro_assembler);
+        }
+      }
+    }
+  }
+
+  if (eats_at_least == kEatsAtLeastNotYetInitialized) {
+    // Save some time by looking at most one machine word ahead.
+    eats_at_least =
+        EatsAtLeast(compiler->ascii() ? 4 : 2, kRecursionBudget, not_at_start);
+  }
+  intptr_t preload_characters =
+      CalculatePreloadCharacters(compiler, eats_at_least);
+
+  bool preload_is_current = !skip_was_emitted &&
+      (current_trace->characters_preloaded() == preload_characters);
+  bool preload_has_checked_bounds = preload_is_current;
+
+  AlternativeGenerationList alt_gens(choice_count);
+
+  // For now we just call all choices one after the other.  The idea ultimately
+  // is to use the Dispatch table to try only the relevant ones.
+  for (intptr_t i = first_normal_choice; i < choice_count; i++) {
+    GuardedAlternative alternative = alternatives_->At(i);
+    AlternativeGeneration* alt_gen = alt_gens.at(i);
+    alt_gen->quick_check_details.set_characters(preload_characters);
+    ZoneGrowableArray<Guard*>* guards = alternative.guards();
+    intptr_t guard_count = (guards == NULL) ? 0 : guards->length();
+    Trace new_trace(*current_trace);
+    new_trace.set_characters_preloaded(preload_is_current ?
+                                         preload_characters :
+                                         0);
+    if (preload_has_checked_bounds) {
+      new_trace.set_bound_checked_up_to(preload_characters);
+    }
+    new_trace.quick_check_performed()->Clear();
+    if (not_at_start_) new_trace.set_at_start(Trace::FALSE_VALUE);
+    alt_gen->expects_preload = preload_is_current;
+    bool generate_full_check_inline = false;
+    if (kRegexpOptimization &&
+        try_to_emit_quick_check_for_alternative(i) &&
+        alternative.node()->EmitQuickCheck(compiler,
+                                           &new_trace,
+                                           preload_has_checked_bounds,
+                                           &alt_gen->possible_success,
+                                           &alt_gen->quick_check_details,
+                                           i < choice_count - 1)) {
+      // Quick check was generated for this choice.
+      preload_is_current = true;
+      preload_has_checked_bounds = true;
+      // On the last choice in the ChoiceNode we generated the quick
+      // check to fall through on possible success.  So now we need to
+      // generate the full check inline.
+      if (i == choice_count - 1) {
+        macro_assembler->BindBlock(&alt_gen->possible_success);
+        new_trace.set_quick_check_performed(&alt_gen->quick_check_details);
+        new_trace.set_characters_preloaded(preload_characters);
+        new_trace.set_bound_checked_up_to(preload_characters);
+        generate_full_check_inline = true;
+      }
+    } else if (alt_gen->quick_check_details.cannot_match()) {
+      if (i == choice_count - 1 && !greedy_loop) {
+        macro_assembler->GoTo(trace->backtrack());
+      }
+      continue;
+    } else {
+      // No quick check was generated.  Put the full code here.
+      // If this is not the first choice then there could be slow checks from
+      // previous cases that go here when they fail.  There's no reason to
+      // insist that they preload characters since the slow check we are about
+      // to generate probably can't use it.
+      if (i != first_normal_choice) {
+        alt_gen->expects_preload = false;
+        new_trace.InvalidateCurrentCharacter();
+      }
+      if (i < choice_count - 1) {
+        new_trace.set_backtrack(&alt_gen->after);
+      }
+      generate_full_check_inline = true;
+    }
+    if (generate_full_check_inline) {
+      if (new_trace.actions() != NULL) {
+        new_trace.set_flush_budget(new_flush_budget);
+      }
+      for (intptr_t j = 0; j < guard_count; j++) {
+        GenerateGuard(macro_assembler, guards->At(j), &new_trace);
+      }
+      alternative.node()->Emit(compiler, &new_trace);
+      preload_is_current = false;
+    }
+    macro_assembler->BindBlock(&alt_gen->after);
+  }
+  if (greedy_loop) {
+    macro_assembler->BindBlock(&greedy_loop_label);
+    // If we have unwound to the bottom then backtrack.
+    macro_assembler->CheckGreedyLoop(trace->backtrack());
+    // Otherwise try the second priority at an earlier position.
+    macro_assembler->AdvanceCurrentPosition(-text_length);
+    macro_assembler->GoTo(&second_choice);
+  }
+
+  // At this point we need to generate slow checks for the alternatives where
+  // the quick check was inlined.  We can recognize these because the associated
+  // label was bound.
+  for (intptr_t i = first_normal_choice; i < choice_count - 1; i++) {
+    AlternativeGeneration* alt_gen = alt_gens.at(i);
+    Trace new_trace(*current_trace);
+    // If there are actions to be flushed we have to limit how many times
+    // they are flushed.  Take the budget of the parent trace and distribute
+    // it fairly amongst the children.
+    if (new_trace.actions() != NULL) {
+      new_trace.set_flush_budget(new_flush_budget);
+    }
+    EmitOutOfLineContinuation(compiler,
+                              &new_trace,
+                              alternatives_->At(i),
+                              alt_gen,
+                              preload_characters,
+                              alt_gens.at(i + 1)->expects_preload);
+  }
+}
+
+
+void ChoiceNode::EmitOutOfLineContinuation(RegExpCompiler* compiler,
+                                           Trace* trace,
+                                           GuardedAlternative alternative,
+                                           AlternativeGeneration* alt_gen,
+                                           intptr_t preload_characters,
+                                           bool next_expects_preload) {
+  if (!alt_gen->possible_success.IsLinked()) return;
+
+  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
+  macro_assembler->BindBlock(&alt_gen->possible_success);
+  Trace out_of_line_trace(*trace);
+  out_of_line_trace.set_characters_preloaded(preload_characters);
+  out_of_line_trace.set_quick_check_performed(&alt_gen->quick_check_details);
+  if (not_at_start_) out_of_line_trace.set_at_start(Trace::FALSE_VALUE);
+  ZoneGrowableArray<Guard*>* guards = alternative.guards();
+  intptr_t guard_count = (guards == NULL) ? 0 : guards->length();
+  if (next_expects_preload) {
+    BlockLabel reload_current_char;
+    out_of_line_trace.set_backtrack(&reload_current_char);
+    for (intptr_t j = 0; j < guard_count; j++) {
+      GenerateGuard(macro_assembler, guards->At(j), &out_of_line_trace);
+    }
+    alternative.node()->Emit(compiler, &out_of_line_trace);
+    macro_assembler->BindBlock(&reload_current_char);
+    // Reload the current character, since the next quick check expects that.
+    // We don't need to check bounds here because we only get into this
+    // code through a quick check which already did the checked load.
+    macro_assembler->LoadCurrentCharacter(trace->cp_offset(),
+                                          NULL,
+                                          false,
+                                          preload_characters);
+    macro_assembler->GoTo(&(alt_gen->after));
+  } else {
+    out_of_line_trace.set_backtrack(&(alt_gen->after));
+    for (intptr_t j = 0; j < guard_count; j++) {
+      GenerateGuard(macro_assembler, guards->At(j), &out_of_line_trace);
+    }
+    alternative.node()->Emit(compiler, &out_of_line_trace);
+  }
+}
+
+
+void ActionNode::Emit(RegExpCompiler* compiler, Trace* trace) {
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  LimitResult limit_result = LimitVersions(compiler, trace);
+  if (limit_result == DONE) return;
+  ASSERT(limit_result == CONTINUE);
+
+  RecursionCheck rc(compiler);
+
+  switch (action_type_) {
+    case STORE_POSITION: {
+      Trace::DeferredCapture
+          new_capture(data_.u_position_register.reg,
+                      data_.u_position_register.is_capture,
+                      trace);
+      Trace new_trace = *trace;
+      new_trace.add_action(&new_capture);
+      on_success()->Emit(compiler, &new_trace);
+      break;
+    }
+    case INCREMENT_REGISTER: {
+      Trace::DeferredIncrementRegister
+          new_increment(data_.u_increment_register.reg);
+      Trace new_trace = *trace;
+      new_trace.add_action(&new_increment);
+      on_success()->Emit(compiler, &new_trace);
+      break;
+    }
+    case SET_REGISTER: {
+      Trace::DeferredSetRegister
+          new_set(data_.u_store_register.reg, data_.u_store_register.value);
+      Trace new_trace = *trace;
+      new_trace.add_action(&new_set);
+      on_success()->Emit(compiler, &new_trace);
+      break;
+    }
+    case CLEAR_CAPTURES: {
+      Trace::DeferredClearCaptures
+        new_capture(Interval(data_.u_clear_captures.range_from,
+                             data_.u_clear_captures.range_to));
+      Trace new_trace = *trace;
+      new_trace.add_action(&new_capture);
+      on_success()->Emit(compiler, &new_trace);
+      break;
+    }
+    case BEGIN_SUBMATCH:
+      if (!trace->is_trivial()) {
+        trace->Flush(compiler, this);
+      } else {
+        assembler->WriteCurrentPositionToRegister(
+            data_.u_submatch.current_position_register, 0);
+        assembler->WriteStackPointerToRegister(
+            data_.u_submatch.stack_pointer_register);
+        on_success()->Emit(compiler, trace);
+      }
+      break;
+    case EMPTY_MATCH_CHECK: {
+      intptr_t start_pos_reg = data_.u_empty_match_check.start_register;
+      intptr_t stored_pos = 0;
+      intptr_t rep_reg = data_.u_empty_match_check.repetition_register;
+      bool has_minimum = (rep_reg != RegExpCompiler::kNoRegister);
+      bool know_dist = trace->GetStoredPosition(start_pos_reg, &stored_pos);
+      if (know_dist && !has_minimum && stored_pos == trace->cp_offset()) {
+        // If we know we haven't advanced and there is no minimum we
+        // can just backtrack immediately.
+        assembler->GoTo(trace->backtrack());
+      } else if (know_dist && stored_pos < trace->cp_offset()) {
+        // If we know we've advanced we can generate the continuation
+        // immediately.
+        on_success()->Emit(compiler, trace);
+      } else if (!trace->is_trivial()) {
+        trace->Flush(compiler, this);
+      } else {
+        BlockLabel skip_empty_check;
+        // If we have a minimum number of repetitions we check the current
+        // number first and skip the empty check if it's not enough.
+        if (has_minimum) {
+          intptr_t limit = data_.u_empty_match_check.repetition_limit;
+          assembler->IfRegisterLT(rep_reg, limit, &skip_empty_check);
+        }
+        // If the match is empty we bail out, otherwise we fall through
+        // to the on-success continuation.
+        assembler->IfRegisterEqPos(data_.u_empty_match_check.start_register,
+                                   trace->backtrack());
+        assembler->BindBlock(&skip_empty_check);
+        on_success()->Emit(compiler, trace);
+      }
+      break;
+    }
+    case POSITIVE_SUBMATCH_SUCCESS: {
+      if (!trace->is_trivial()) {
+        trace->Flush(compiler, this);
+        return;
+      }
+      assembler->ReadCurrentPositionFromRegister(
+          data_.u_submatch.current_position_register);
+      assembler->ReadStackPointerFromRegister(
+          data_.u_submatch.stack_pointer_register);
+      intptr_t clear_register_count = data_.u_submatch.clear_register_count;
+      if (clear_register_count == 0) {
+        on_success()->Emit(compiler, trace);
+        return;
+      }
+      intptr_t clear_registers_from = data_.u_submatch.clear_register_from;
+      BlockLabel clear_registers_backtrack;
+      Trace new_trace = *trace;
+      new_trace.set_backtrack(&clear_registers_backtrack);
+      on_success()->Emit(compiler, &new_trace);
+
+      assembler->BindBlock(&clear_registers_backtrack);
+      intptr_t clear_registers_to =
+          clear_registers_from + clear_register_count - 1;
+      assembler->ClearRegisters(clear_registers_from, clear_registers_to);
+
+      ASSERT(trace->backtrack() == NULL);
+      assembler->Backtrack();
+      return;
+    }
+    default:
+      UNREACHABLE();
+  }
+}
+
+
+void BackReferenceNode::Emit(RegExpCompiler* compiler, Trace* trace) {
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  if (!trace->is_trivial()) {
+    trace->Flush(compiler, this);
+    return;
+  }
+
+  LimitResult limit_result = LimitVersions(compiler, trace);
+  if (limit_result == DONE) return;
+  ASSERT(limit_result == CONTINUE);
+
+  RecursionCheck rc(compiler);
+
+  ASSERT(start_reg_ + 1 == end_reg_);
+  if (compiler->ignore_case()) {
+    assembler->CheckNotBackReferenceIgnoreCase(start_reg_,
+                                               trace->backtrack());
+  } else {
+    assembler->CheckNotBackReference(start_reg_, trace->backtrack());
+  }
+  on_success()->Emit(compiler, trace);
+}
+
+
+void ActionNode::FillInBMInfo(intptr_t offset,
+                              intptr_t budget,
+                              BoyerMooreLookahead* bm,
+                              bool not_at_start) {
+  if (action_type_ == BEGIN_SUBMATCH) {
+    bm->SetRest(offset);
+  } else if (action_type_ != POSITIVE_SUBMATCH_SUCCESS) {
+    on_success()->FillInBMInfo(offset, budget - 1, bm, not_at_start);
+  }
+  SaveBMInfo(bm, not_at_start, offset);
+}
+
+
+void AssertionNode::FillInBMInfo(intptr_t offset,
+                                 intptr_t budget,
+                                 BoyerMooreLookahead* bm,
+                                 bool not_at_start) {
+  // Match the behaviour of EatsAtLeast on this node.
+  if (assertion_type() == AT_START && not_at_start) return;
+  on_success()->FillInBMInfo(offset, budget - 1, bm, not_at_start);
+  SaveBMInfo(bm, not_at_start, offset);
+}
+
+
+void BackReferenceNode::FillInBMInfo(intptr_t offset,
+                                     intptr_t budget,
+                                     BoyerMooreLookahead* bm,
+                                     bool not_at_start) {
+  // Working out the set of characters that a backreference can match is too
+  // hard, so we just say that any character can match.
+  bm->SetRest(offset);
+  SaveBMInfo(bm, not_at_start, offset);
+}
+
+
+// Returns the number of characters in the equivalence class, omitting those
+// that cannot occur in the source string because it is ASCII.
+static intptr_t GetCaseIndependentLetters(uint16_t character,
+                                          bool ascii_subject,
+                                          int32_t* letters) {
+  unibrow::Mapping<unibrow::Ecma262UnCanonicalize> jsregexp_uncanonicalize;
+  intptr_t length = jsregexp_uncanonicalize.get(character, '\0', letters);
+  // Unibrow returns 0 or 1 for characters where case independence is
+  // trivial.
+  if (length == 0) {
+    letters[0] = character;
+    length = 1;
+  }
+  if (!ascii_subject || character <= Symbols::kMaxOneCharCodeSymbol) {
+    return length;
+  }
+  // The standard requires that non-ASCII characters cannot have ASCII
+  // character codes in their equivalence class.
+  return 0;
+}
+
+
+void ChoiceNode::FillInBMInfo(intptr_t offset,
+                              intptr_t budget,
+                              BoyerMooreLookahead* bm,
+                              bool not_at_start) {
+  ZoneGrowableArray<GuardedAlternative>* alts = alternatives();
+  budget = (budget - 1) / alts->length();
+  for (intptr_t i = 0; i < alts->length(); i++) {
+    GuardedAlternative& alt = (*alts)[i];
+    if (alt.guards() != NULL && alt.guards()->length() != 0) {
+      bm->SetRest(offset);  // Give up trying to fill in info.
+      SaveBMInfo(bm, not_at_start, offset);
+      return;
+    }
+    alt.node()->FillInBMInfo(offset, budget, bm, not_at_start);
+  }
+  SaveBMInfo(bm, not_at_start, offset);
+}
+
+
+void EndNode::FillInBMInfo(intptr_t offset,
+                           intptr_t budget,
+                           BoyerMooreLookahead* bm,
+                           bool not_at_start) {
+  // Returning 0 from EatsAtLeast should ensure we never get here.
+  UNREACHABLE();
+}
+
+
+void LoopChoiceNode::FillInBMInfo(intptr_t offset,
+                                  intptr_t budget,
+                                  BoyerMooreLookahead* bm,
+                                  bool not_at_start) {
+  if (body_can_be_zero_length_ || budget <= 0) {
+    bm->SetRest(offset);
+    SaveBMInfo(bm, not_at_start, offset);
+    return;
+  }
+  ChoiceNode::FillInBMInfo(offset, budget - 1, bm, not_at_start);
+  SaveBMInfo(bm, not_at_start, offset);
+}
+
+
+void TextNode::FillInBMInfo(intptr_t initial_offset,
+                            intptr_t budget,
+                            BoyerMooreLookahead* bm,
+                            bool not_at_start) {
+  if (initial_offset >= bm->length()) return;
+  intptr_t offset = initial_offset;
+  intptr_t max_char = bm->max_char();
+  for (intptr_t i = 0; i < elements()->length(); i++) {
+    if (offset >= bm->length()) {
+      if (initial_offset == 0) set_bm_info(not_at_start, bm);
+      return;
+    }
+    TextElement text = elements()->At(i);
+    if (text.text_type() == TextElement::ATOM) {
+      RegExpAtom* atom = text.atom();
+      for (intptr_t j = 0; j < atom->length(); j++, offset++) {
+        if (offset >= bm->length()) {
+          if (initial_offset == 0) set_bm_info(not_at_start, bm);
+          return;
+        }
+        uint16_t character = atom->data()->At(j);
+        if (bm->compiler()->ignore_case()) {
+          int32_t chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
+          intptr_t length = GetCaseIndependentLetters(
+              character,
+              bm->max_char() == Symbols::kMaxOneCharCodeSymbol,
+              chars);
+          for (intptr_t j = 0; j < length; j++) {
+            bm->Set(offset, chars[j]);
+          }
+        } else {
+          if (character <= max_char) bm->Set(offset, character);
+        }
+      }
+    } else {
+      ASSERT(text.text_type() == TextElement::CHAR_CLASS);
+      RegExpCharacterClass* char_class = text.char_class();
+      ZoneGrowableArray<CharacterRange>* ranges = char_class->ranges();
+      if (char_class->is_negated()) {
+        bm->SetAll(offset);
+      } else {
+        for (intptr_t k = 0; k < ranges->length(); k++) {
+          CharacterRange& range = (*ranges)[k];
+          if (range.from() > max_char) continue;
+          intptr_t to = Utils::Minimum(max_char,
+                                       static_cast<intptr_t>(range.to()));
+          bm->SetInterval(offset, Interval(range.from(), to));
+        }
+      }
+      offset++;
+    }
+  }
+  if (offset >= bm->length()) {
+    if (initial_offset == 0) set_bm_info(not_at_start, bm);
+    return;
+  }
+  on_success()->FillInBMInfo(offset,
+                             budget - 1,
+                             bm,
+                             true);  // Not at start after a text node.
+  if (initial_offset == 0) set_bm_info(not_at_start, bm);
+}
+
+
+// Check for [0-9A-Z_a-z].
+static void EmitWordCheck(RegExpMacroAssembler* assembler,
+                          BlockLabel* word,
+                          BlockLabel* non_word,
+                          bool fall_through_on_word) {
+  if (assembler->CheckSpecialCharacterClass(
+          fall_through_on_word ? 'w' : 'W',
+          fall_through_on_word ? non_word : word)) {
+    // Optimized implementation available.
+    return;
+  }
+  assembler->CheckCharacterGT('z', non_word);
+  assembler->CheckCharacterLT('0', non_word);
+  assembler->CheckCharacterGT('a' - 1, word);
+  assembler->CheckCharacterLT('9' + 1, word);
+  assembler->CheckCharacterLT('A', non_word);
+  assembler->CheckCharacterLT('Z' + 1, word);
+  if (fall_through_on_word) {
+    assembler->CheckNotCharacter('_', non_word);
+  } else {
+    assembler->CheckCharacter('_', word);
+  }
+}
 
 
 // Emit the code to handle \b and \B (word-boundary or non-word-boundary).
 void AssertionNode::EmitBoundaryCheck(RegExpCompiler* compiler, Trace* trace) {
-  UNIMPLEMENTED();
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  Trace::TriBool next_is_word_character = Trace::UNKNOWN;
+  bool not_at_start = (trace->at_start() == Trace::FALSE_VALUE);
+  BoyerMooreLookahead* lookahead = bm_info(not_at_start);
+  if (lookahead == NULL) {
+    intptr_t eats_at_least =
+        Utils::Minimum(kMaxLookaheadForBoyerMoore,
+                       EatsAtLeast(kMaxLookaheadForBoyerMoore,
+                                   kRecursionBudget,
+                                   not_at_start));
+    if (eats_at_least >= 1) {
+      BoyerMooreLookahead* bm =
+          new(I) BoyerMooreLookahead(eats_at_least, compiler, I);
+      FillInBMInfo(0, kRecursionBudget, bm, not_at_start);
+      if (bm->at(0)->is_non_word())
+        next_is_word_character = Trace::FALSE_VALUE;
+      if (bm->at(0)->is_word()) next_is_word_character = Trace::TRUE_VALUE;
+    }
+  } else {
+    if (lookahead->at(0)->is_non_word())
+      next_is_word_character = Trace::FALSE_VALUE;
+    if (lookahead->at(0)->is_word())
+      next_is_word_character = Trace::TRUE_VALUE;
+  }
+  bool at_boundary = (assertion_type_ == AssertionNode::AT_BOUNDARY);
+  if (next_is_word_character == Trace::UNKNOWN) {
+    BlockLabel before_non_word;
+    BlockLabel before_word;
+    if (trace->characters_preloaded() != 1) {
+      assembler->LoadCurrentCharacter(trace->cp_offset(), &before_non_word);
+    }
+    // Fall through on non-word.
+    EmitWordCheck(assembler, &before_word, &before_non_word, false);
+    // Next character is not a word character.
+    assembler->BindBlock(&before_non_word);
+    BlockLabel ok;
+    // Backtrack on \B (non-boundary check) if previous is a word,
+    // since we know next *is not* a word and this would be a boundary.
+    BacktrackIfPrevious(compiler, trace, at_boundary ? kIsNonWord : kIsWord);
+
+    if (!assembler->IsClosed()) {
+      assembler->GoTo(&ok);
+    }
+
+    assembler->BindBlock(&before_word);
+    BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord);
+    assembler->BindBlock(&ok);
+  } else if (next_is_word_character == Trace::TRUE_VALUE) {
+    BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord);
+  } else {
+    ASSERT(next_is_word_character == Trace::FALSE_VALUE);
+    BacktrackIfPrevious(compiler, trace, at_boundary ? kIsNonWord : kIsWord);
+  }
 }
 
 
 void AssertionNode::BacktrackIfPrevious(
     RegExpCompiler* compiler,
     Trace* trace,
     AssertionNode::IfPrevious backtrack_if_previous) {
-  UNIMPLEMENTED();
-}
-
-
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  Trace new_trace(*trace);
+  new_trace.InvalidateCurrentCharacter();
+
+  BlockLabel fall_through, dummy;
+
+  BlockLabel* non_word = backtrack_if_previous == kIsNonWord ?
+                         new_trace.backtrack() :
+                         &fall_through;
+  BlockLabel* word = backtrack_if_previous == kIsNonWord ?
+                     &fall_through :
+                     new_trace.backtrack();
+
+  if (new_trace.cp_offset() == 0) {
+    // The start of input counts as a non-word character, so the question is
+    // decided if we are at the start.
+    assembler->CheckAtStart(non_word);
+  }
+  // We already checked that we are not at the start of input so it must be
+  // OK to load the previous character.
+  assembler->LoadCurrentCharacter(new_trace.cp_offset() - 1, &dummy, false);
+  EmitWordCheck(assembler, word, non_word, backtrack_if_previous == kIsNonWord);
+
+  assembler->BindBlock(&fall_through);
+  on_success()->Emit(compiler, &new_trace);
+}
+
+
+static bool DeterminedAlready(QuickCheckDetails* quick_check, intptr_t offset) {
+  if (quick_check == NULL) return false;
+  if (offset >= quick_check->characters()) return false;
+  return quick_check->positions(offset)->determines_perfectly;
+}
+
+
+static void UpdateBoundsCheck(intptr_t index, intptr_t* checked_up_to) {
+  if (index > *checked_up_to) {
+    *checked_up_to = index;
+  }
+}
+
+
+typedef bool EmitCharacterFunction(Isolate* isolate,
+                                   RegExpCompiler* compiler,
+                                   uint16_t c,
+                                   BlockLabel* on_failure,
+                                   intptr_t cp_offset,
+                                   bool check,
+                                   bool preloaded);
+
+
+static inline bool EmitSimpleCharacter(Isolate* isolate,
+                                       RegExpCompiler* compiler,
+                                       uint16_t c,
+                                       BlockLabel* on_failure,
+                                       intptr_t cp_offset,
+                                       bool check,
+                                       bool preloaded) {
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  bool bound_checked = false;
+  if (!preloaded) {
+    assembler->LoadCurrentCharacter(
+        cp_offset,
+        on_failure,
+        check);
+    bound_checked = true;
+  }
+  assembler->CheckNotCharacter(c, on_failure);
+  return bound_checked;
+}
+
+
+static bool ShortCutEmitCharacterPair(RegExpMacroAssembler* macro_assembler,
+                                      bool ascii,
+                                      uint16_t c1,
+                                      uint16_t c2,
+                                      BlockLabel* on_failure) {
+  uint16_t char_mask;
+  if (ascii) {
+    char_mask = Symbols::kMaxOneCharCodeSymbol;
+  } else {
+    char_mask = Utf16::kMaxCodeUnit;
+  }
+  uint16_t exor = c1 ^ c2;
+  // Check whether exor has only one bit set.
+  if (((exor - 1) & exor) == 0) {
+    // If c1 and c2 differ only by one bit.
+    // Ecma262UnCanonicalize always gives the highest number last.
+    ASSERT(c2 > c1);
+    uint16_t mask = char_mask ^ exor;
+    macro_assembler->CheckNotCharacterAfterAnd(c1, mask, on_failure);
+    return true;
+  }
+  ASSERT(c2 > c1);
+  uint16_t diff = c2 - c1;
+  if (((diff - 1) & diff) == 0 && c1 >= diff) {
+    // If the characters differ by 2^n but don't differ by one bit then
+    // subtract the difference from the found character, then do the or
+    // trick.  We avoid the theoretical case where negative numbers are
+    // involved in order to simplify code generation.
+    uint16_t mask = char_mask ^ diff;
+    macro_assembler->CheckNotCharacterAfterMinusAnd(c1 - diff,
+                                                    diff,
+                                                    mask,
+                                                    on_failure);
+    return true;
+  }
+  return false;
+}
+
+// Only emits letters (things that have case).  Only used for case independent
+// matches.
+static inline bool EmitAtomLetter(Isolate* isolate,
+                                  RegExpCompiler* compiler,
+                                  uint16_t c,
+                                  BlockLabel* on_failure,
+                                  intptr_t cp_offset,
+                                  bool check,
+                                  bool preloaded) {
+  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
+  bool ascii = compiler->ascii();
+  int32_t chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
+  intptr_t length = GetCaseIndependentLetters(c, ascii, chars);
+  if (length <= 1) return false;
+  // We may not need to check against the end of the input string
+  // if this character lies before a character that matched.
+  if (!preloaded) {
+    macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check);
+  }
+  BlockLabel ok;
+  ASSERT(unibrow::Ecma262UnCanonicalize::kMaxWidth == 4);
+  switch (length) {
+    case 2: {
+      if (ShortCutEmitCharacterPair(macro_assembler,
+                                    ascii,
+                                    chars[0],
+                                    chars[1],
+                                    on_failure)) {
+      } else {
+        macro_assembler->CheckCharacter(chars[0], &ok);
+        macro_assembler->CheckNotCharacter(chars[1], on_failure);
+        macro_assembler->BindBlock(&ok);
+      }
+      break;
+    }
+    case 4:
+      macro_assembler->CheckCharacter(chars[3], &ok);
+      // Fall through!
+    case 3:
+      macro_assembler->CheckCharacter(chars[0], &ok);
+      macro_assembler->CheckCharacter(chars[1], &ok);
+      macro_assembler->CheckNotCharacter(chars[2], on_failure);
+      macro_assembler->BindBlock(&ok);
+      break;
+    default:
+      UNREACHABLE();
+      break;
+  }
+  return true;
+}
+
+
+// Only emits non-letters (things that don't have case).  Only used for case
+// independent matches.
+static inline bool EmitAtomNonLetter(Isolate* isolate,
+                                     RegExpCompiler* compiler,
+                                     uint16_t c,
+                                     BlockLabel* on_failure,
+                                     intptr_t cp_offset,
+                                     bool check,
+                                     bool preloaded) {
+  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
+  bool ascii = compiler->ascii();
+  int32_t chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
+  intptr_t length = GetCaseIndependentLetters(c, ascii, chars);
+  if (length < 1) {
+    // This can't match.  Must be an ASCII subject and a non-ASCII character.
+    // We do not need to do anything since the ASCII pass already handled this.
+    return false;  // Bounds not checked.
+  }
+  bool checked = false;
+  // We handle the length > 1 case in a later pass.
+  if (length == 1) {
+    if (ascii && c > Symbols::kMaxOneCharCodeSymbol) {
+      // Can't match - see above.
+      return false;  // Bounds not checked.
+    }
+    if (!preloaded) {
+      macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check);
+      checked = check;
+    }
+    macro_assembler->CheckNotCharacter(c, on_failure);
+  }
+  return checked;
+}
+
+
+static void EmitBoundaryTest(RegExpMacroAssembler* masm,
+                             intptr_t border,
+                             BlockLabel* fall_through,
+                             BlockLabel* above_or_equal,
+                             BlockLabel* below) {
+  if (below != fall_through) {
+    masm->CheckCharacterLT(border, below);
+    if (above_or_equal != fall_through) masm->GoTo(above_or_equal);
+  } else {
+    masm->CheckCharacterGT(border - 1, above_or_equal);
+  }
+}
+
+
+static void EmitDoubleBoundaryTest(RegExpMacroAssembler* masm,
+                                   intptr_t first,
+                                   intptr_t last,
+                                   BlockLabel* fall_through,
+                                   BlockLabel* in_range,
+                                   BlockLabel* out_of_range) {
+  if (in_range == fall_through) {
+    if (first == last) {
+      masm->CheckNotCharacter(first, out_of_range);
+    } else {
+      masm->CheckCharacterNotInRange(first, last, out_of_range);
+    }
+  } else {
+    if (first == last) {
+      masm->CheckCharacter(first, in_range);
+    } else {
+      masm->CheckCharacterInRange(first, last, in_range);
+    }
+    if (out_of_range != fall_through) masm->GoTo(out_of_range);
+  }
+}
+
+
+static void CutOutRange(RegExpMacroAssembler* masm,
+                        ZoneGrowableArray<int>* ranges,
+                        intptr_t start_index,
+                        intptr_t end_index,
+                        intptr_t cut_index,
+                        BlockLabel* even_label,
+                        BlockLabel* odd_label) {
+  bool odd = (((cut_index - start_index) & 1) == 1);
+  BlockLabel* in_range_label = odd ? odd_label : even_label;
+  BlockLabel dummy;
+  EmitDoubleBoundaryTest(masm,
+                         ranges->At(cut_index),
+                         ranges->At(cut_index + 1) - 1,
+                         &dummy,
+                         in_range_label,
+                         &dummy);
+  ASSERT(!dummy.IsLinked());
+  // Cut out the single range by rewriting the array.  This creates a new
+  // range that is a merger of the two ranges on either side of the one we
+  // are cutting out.  The oddity of the labels is preserved.
+  for (intptr_t j = cut_index; j > start_index; j--) {
+    (*ranges)[j] = ranges->At(j - 1);
+  }
+  for (intptr_t j = cut_index + 1; j < end_index; j++) {
+    (*ranges)[j] = ranges->At(j + 1);
+  }
+}
+
+
+// even_label is for ranges[i] to ranges[i + 1] where i - start_index is even.
+// odd_label is for ranges[i] to ranges[i + 1] where i - start_index is odd.
+static void EmitUseLookupTable(
+    RegExpMacroAssembler* masm,
+    ZoneGrowableArray<int>* ranges,
+    intptr_t start_index,
+    intptr_t end_index,
+    intptr_t min_char,
+    BlockLabel* fall_through,
+    BlockLabel* even_label,
+    BlockLabel* odd_label) {
+  static const intptr_t kSize = RegExpMacroAssembler::kTableSize;
+  static const intptr_t kMask = RegExpMacroAssembler::kTableMask;
+
+  intptr_t base = (min_char & ~kMask);
+
+  // Assert that everything is on one kTableSize page.
+  for (intptr_t i = start_index; i <= end_index; i++) {
+    ASSERT((ranges->At(i) & ~kMask) == base);
+  }
+  ASSERT(start_index == 0 || (ranges->At(start_index - 1) & ~kMask) <= base);
+
+  char templ[kSize];
+  BlockLabel* on_bit_set;
+  BlockLabel* on_bit_clear;
+  intptr_t bit;
+  if (even_label == fall_through) {
+    on_bit_set = odd_label;
+    on_bit_clear = even_label;
+    bit = 1;
+  } else {
+    on_bit_set = even_label;
+    on_bit_clear = odd_label;
+    bit = 0;
+  }
+  for (intptr_t i = 0; i < (ranges->At(start_index) & kMask) && i < kSize;
+       i++) {
+    templ[i] = bit;
+  }
+  intptr_t j = 0;
+  bit ^= 1;
+  for (intptr_t i = start_index; i < end_index; i++) {
+    for (j = (ranges->At(i) & kMask); j < (ranges->At(i + 1) & kMask); j++) {
+      templ[j] = bit;
+    }
+    bit ^= 1;
+  }
+  for (intptr_t i = j; i < kSize; i++) {
+    templ[i] = bit;
+  }
+  // TODO(erikcorry): Cache these.
+  TypedData& ba = TypedData::ZoneHandle(

[Florian Schneider, 2014/10/01 17:04:14]

const TypedData&

[jgruber1, 2014/10/03 18:59:52]

Done.

+        masm->isolate(),
+        TypedData::New(kTypedDataUint8ArrayCid, kSize, Heap::kOld));
+  for (intptr_t i = 0; i < kSize; i++) {
+    ba.SetUint8(i, templ[i]);
+  }
+  masm->CheckBitInTable(ba, on_bit_set);
+  if (on_bit_clear != fall_through) masm->GoTo(on_bit_clear);
+}
+
+
+// Unicode case.  Split the search space into kSize spaces that are handled
+// with recursion.
+static void SplitSearchSpace(ZoneGrowableArray<int>* ranges,
+                             intptr_t start_index,
+                             intptr_t end_index,
+                             intptr_t* new_start_index,
+                             intptr_t* new_end_index,
+                             intptr_t* border) {
+  static const intptr_t kSize = RegExpMacroAssembler::kTableSize;
+  static const intptr_t kMask = RegExpMacroAssembler::kTableMask;
+
+  intptr_t first = ranges->At(start_index);
+  intptr_t last = ranges->At(end_index) - 1;
+
+  *new_start_index = start_index;
+  *border = (ranges->At(start_index) & ~kMask) + kSize;
+  while (*new_start_index < end_index) {
+    if (ranges->At(*new_start_index) > *border) break;
+    (*new_start_index)++;
+  }
+  // new_start_index is the index of the first edge that is beyond the
+  // current kSize space.
+
+  // For very large search spaces we do a binary chop search of the non-ASCII
+  // space instead of just going to the end of the current kSize space.  The
+  // heuristics are complicated a little by the fact that any 128-character
+  // encoding space can be quickly tested with a table lookup, so we don't
+  // wish to do binary chop search at a smaller granularity than that.  A
+  // 128-character space can take up a lot of space in the ranges array if,
+  // for example, we only want to match every second character (eg. the lower
+  // case characters on some Unicode pages).
+  intptr_t binary_chop_index = (end_index + start_index) / 2;
+  // The first test ensures that we get to the code that handles the ASCII
+  // range with a single not-taken branch, speeding up this important
+  // character range (even non-ASCII charset-based text has spaces and
+  // punctuation).
+  if (*border - 1 > Symbols::kMaxOneCharCodeSymbol &&  // ASCII case.
+      end_index - start_index > (*new_start_index - start_index) * 2 &&
+      last - first > kSize * 2 &&
+      binary_chop_index > *new_start_index &&
+      ranges->At(binary_chop_index) >= first + 2 * kSize) {
+    intptr_t scan_forward_for_section_border = binary_chop_index;;
+    intptr_t new_border = (ranges->At(binary_chop_index) | kMask) + 1;
+
+    while (scan_forward_for_section_border < end_index) {
+      if (ranges->At(scan_forward_for_section_border) > new_border) {
+        *new_start_index = scan_forward_for_section_border;
+        *border = new_border;
+        break;
+      }
+      scan_forward_for_section_border++;
+    }
+  }
+
+  ASSERT(*new_start_index > start_index);
+  *new_end_index = *new_start_index - 1;
+  if (ranges->At(*new_end_index) == *border) {
+    (*new_end_index)--;
+  }
+  if (*border >= ranges->At(end_index)) {
+    *border = ranges->At(end_index);
+    *new_start_index = end_index;  // Won't be used.
+    *new_end_index = end_index - 1;
+  }
+}
+
+
+// Gets a series of segment boundaries representing a character class.  If the
+// character is in the range between an even and an odd boundary (counting from
+// start_index) then go to even_label, otherwise go to odd_label.  We already
+// know that the character is in the range of min_char to max_char inclusive.
+// Either label can be NULL indicating backtracking.  Either label can also be
+// equal to the fall_through label.
+static void GenerateBranches(RegExpMacroAssembler* masm,
+                             ZoneGrowableArray<int>* ranges,
+                             intptr_t start_index,
+                             intptr_t end_index,
+                             uint16_t min_char,
+                             uint16_t max_char,
+                             BlockLabel* fall_through,
+                             BlockLabel* even_label,
+                             BlockLabel* odd_label) {
+  intptr_t first = ranges->At(start_index);
+  intptr_t last = ranges->At(end_index) - 1;
+
+  ASSERT(min_char < first);
+
+  // Just need to test if the character is before or on-or-after
+  // a particular character.
+  if (start_index == end_index) {
+    EmitBoundaryTest(masm, first, fall_through, even_label, odd_label);
+    return;
+  }
+
+  // Another almost trivial case:  There is one interval in the middle that is
+  // different from the end intervals.
+  if (start_index + 1 == end_index) {
+    EmitDoubleBoundaryTest(
+        masm, first, last, fall_through, even_label, odd_label);
+    return;
+  }
+
+  // It's not worth using table lookup if there are very few intervals in the
+  // character class.
+  if (end_index - start_index <= 6) {
+    // It is faster to test for individual characters, so we look for those
+    // first, then try arbitrary ranges in the second round.
+    static intptr_t kNoCutIndex = -1;
+    intptr_t cut = kNoCutIndex;
+    for (intptr_t i = start_index; i < end_index; i++) {
+      if (ranges->At(i) == ranges->At(i + 1) - 1) {
+        cut = i;
+        break;
+      }
+    }
+    if (cut == kNoCutIndex) cut = start_index;
+    CutOutRange(
+        masm, ranges, start_index, end_index, cut, even_label, odd_label);
+    ASSERT(end_index - start_index >= 2);
+    GenerateBranches(masm,
+                     ranges,
+                     start_index + 1,
+                     end_index - 1,
+                     min_char,
+                     max_char,
+                     fall_through,
+                     even_label,
+                     odd_label);
+    return;
+  }
+
+  // If there are a lot of intervals in the regexp, then we will use tables to
+  // determine whether the character is inside or outside the character class.
+  static const intptr_t kBits = RegExpMacroAssembler::kTableSizeBits;
+
+  if ((max_char >> kBits) == (min_char >> kBits)) {
+    EmitUseLookupTable(masm,
+                       ranges,
+                       start_index,
+                       end_index,
+                       min_char,
+                       fall_through,
+                       even_label,
+                       odd_label);
+    return;
+  }
+
+  if ((min_char >> kBits) != (first >> kBits)) {
+    masm->CheckCharacterLT(first, odd_label);
+    GenerateBranches(masm,
+                     ranges,
+                     start_index + 1,
+                     end_index,
+                     first,
+                     max_char,
+                     fall_through,
+                     odd_label,
+                     even_label);
+    return;
+  }
+
+  intptr_t new_start_index = 0;
+  intptr_t new_end_index = 0;
+  intptr_t border = 0;
+
+  SplitSearchSpace(ranges,
+                   start_index,
+                   end_index,
+                   &new_start_index,
+                   &new_end_index,
+                   &border);
+
+  BlockLabel handle_rest;
+  BlockLabel* above = &handle_rest;
+  if (border == last + 1) {
+    // We didn't find any section that started after the limit, so everything
+    // above the border is one of the terminal labels.
+    above = (end_index & 1) != (start_index & 1) ? odd_label : even_label;
+    ASSERT(new_end_index == end_index - 1);
+  }
+
+  ASSERT(start_index <= new_end_index);
+  ASSERT(new_start_index <= end_index);
+  ASSERT(start_index < new_start_index);
+  ASSERT(new_end_index < end_index);
+  ASSERT(new_end_index + 1 == new_start_index ||
+         (new_end_index + 2 == new_start_index &&
+          border == ranges->At(new_end_index + 1)));
+  ASSERT(min_char < border - 1);
+  ASSERT(border < max_char);
+  ASSERT(ranges->At(new_end_index) < border);
+  ASSERT(border < ranges->At(new_start_index) ||
+         (border == ranges->At(new_start_index) &&
+          new_start_index == end_index &&
+          new_end_index == end_index - 1 &&
+          border == last + 1));
+  ASSERT(new_start_index == 0 || border >= ranges->At(new_start_index - 1));
+
+  masm->CheckCharacterGT(border - 1, above);
+  BlockLabel dummy;
+  GenerateBranches(masm,
+                   ranges,
+                   start_index,
+                   new_end_index,
+                   min_char,
+                   border - 1,
+                   &dummy,
+                   even_label,
+                   odd_label);
+
+  if (handle_rest.IsLinked()) {
+    masm->BindBlock(&handle_rest);
+    bool flip = (new_start_index & 1) != (start_index & 1);
+    GenerateBranches(masm,
+                     ranges,
+                     new_start_index,
+                     end_index,
+                     border,
+                     max_char,
+                     &dummy,
+                     flip ? odd_label : even_label,
+                     flip ? even_label : odd_label);
+  }
+}
+
+
+static void EmitCharClass(RegExpMacroAssembler* macro_assembler,
+                          RegExpCharacterClass* cc,
+                          bool ascii,
+                          BlockLabel* on_failure,
+                          intptr_t cp_offset,
+                          bool check_offset,
+                          bool preloaded,
+                          Isolate* isolate) {
+  ZoneGrowableArray<CharacterRange>* ranges = cc->ranges();
+  if (!CharacterRange::IsCanonical(ranges)) {
+    CharacterRange::Canonicalize(ranges);
+  }
+
+  intptr_t max_char;
+  if (ascii) {
+    max_char = Symbols::kMaxOneCharCodeSymbol;
+  } else {
+    max_char = Utf16::kMaxCodeUnit;
+  }
+
+  intptr_t range_count = ranges->length();
+
+  intptr_t last_valid_range = range_count - 1;
+  while (last_valid_range >= 0) {
+    CharacterRange& range = (*ranges)[last_valid_range];
+    if (range.from() <= max_char) {
+      break;
+    }
+    last_valid_range--;
+  }
+
+  if (last_valid_range < 0) {
+    if (!cc->is_negated()) {
+      macro_assembler->GoTo(on_failure);
+    }
+    if (check_offset) {
+      macro_assembler->CheckPosition(cp_offset, on_failure);
+    }
+    return;
+  }
+
+  if (last_valid_range == 0 &&
+      ranges->At(0).IsEverything(max_char)) {
+    if (cc->is_negated()) {
+      macro_assembler->GoTo(on_failure);
+    } else {
+      // This is a common case hit by non-anchored expressions.
+      if (check_offset) {
+        macro_assembler->CheckPosition(cp_offset, on_failure);
+      }
+    }
+    return;
+  }
+  if (last_valid_range == 0 &&
+      !cc->is_negated() &&
+      ranges->At(0).IsEverything(max_char)) {
+    // This is a common case hit by non-anchored expressions.
+    if (check_offset) {
+      macro_assembler->CheckPosition(cp_offset, on_failure);
+    }
+    return;
+  }
+
+  if (!preloaded) {
+    macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check_offset);
+  }
+
+  if (cc->is_standard() &&
+        macro_assembler->CheckSpecialCharacterClass(cc->standard_type(),
+                                                    on_failure)) {
+      return;
+  }
+
+
+  // A new list with ascending entries.  Each entry is a code unit
+  // where there is a boundary between code units that are part of
+  // the class and code units that are not.  Normally we insert an
+  // entry at zero which goes to the failure label, but if there
+  // was already one there we fall through for success on that entry.
+  // Subsequent entries have alternating meaning (success/failure).
+  ZoneGrowableArray<int>* range_boundaries =
+      new(isolate) ZoneGrowableArray<int>(last_valid_range);
+
+  bool zeroth_entry_is_failure = !cc->is_negated();
+
+  for (intptr_t i = 0; i <= last_valid_range; i++) {
+    CharacterRange& range = (*ranges)[i];
+    if (range.from() == 0) {
+      ASSERT(i == 0);
+      zeroth_entry_is_failure = !zeroth_entry_is_failure;
+    } else {
+      range_boundaries->Add(range.from());
+    }
+    range_boundaries->Add(range.to() + 1);
+  }
+  intptr_t end_index = range_boundaries->length() - 1;
+  if (range_boundaries->At(end_index) > max_char) {
+    end_index--;
+  }
+
+  BlockLabel fall_through;
+  GenerateBranches(macro_assembler,
+                   range_boundaries,
+                   0,  // start_index.
+                   end_index,
+                   0,  // min_char.
+                   max_char,
+                   &fall_through,
+                   zeroth_entry_is_failure ? &fall_through : on_failure,
+                   zeroth_entry_is_failure ? on_failure : &fall_through);
+  macro_assembler->BindBlock(&fall_through);
+}
+
+
 // We call this repeatedly to generate code for each pass over the text node.
 // The passes are in increasing order of difficulty because we hope one
 // of the first passes will fail in which case we are saved the work of the
 // later passes.  for example for the case independent regexp /%[asdfghjkl]a/
 // we will check the '%' in the first pass, the case independent 'a' in the
 // second pass and the character class in the last pass.
 //
 // The passes are done from right to left, so for example to test for /bar/
 // we will first test for an 'r' with offset 2, then an 'a' with offset 1
 // and then a 'b' with offset 0.  This means we can avoid the end-of-input
@@ skipping 15 matching lines @@
 // loading characters, which means we do not need to recheck the bounds
 // up to the limit the quick check already checked.  In addition the quick
 // check can have involved a mask and compare operation which may simplify
 // or obviate the need for further checks at some character positions.
 void TextNode::TextEmitPass(RegExpCompiler* compiler,
                             TextEmitPassType pass,
                             bool preloaded,
                             Trace* trace,
                             bool first_element_checked,
                             intptr_t* checked_up_to) {
-  UNIMPLEMENTED();
+  RegExpMacroAssembler* assembler = compiler->macro_assembler();
+  bool ascii = compiler->ascii();
+  BlockLabel* backtrack = trace->backtrack();
+  QuickCheckDetails* quick_check = trace->quick_check_performed();
+  intptr_t element_count = elms_->length();
+  for (intptr_t i = preloaded ? 0 : element_count - 1; i >= 0; i--) {
+    TextElement elm = elms_->At(i);
+    intptr_t cp_offset = trace->cp_offset() + elm.cp_offset();
+    if (elm.text_type() == TextElement::ATOM) {
+      ZoneGrowableArray<uint16_t>* quarks = elm.atom()->data();
+      for (intptr_t j = preloaded ? 0 : quarks->length() - 1; j >= 0; j--) {
+        if (first_element_checked && i == 0 && j == 0) continue;
+        if (DeterminedAlready(quick_check, elm.cp_offset() + j)) continue;
+        EmitCharacterFunction* emit_function = NULL;
+        switch (pass) {
+          case NON_ASCII_MATCH:
+            ASSERT(ascii);
+            if (quarks->At(j) > Symbols::kMaxOneCharCodeSymbol) {
+              assembler->GoTo(backtrack);
+              return;
+            }
+            break;
+          case NON_LETTER_CHARACTER_MATCH:
+            emit_function = &EmitAtomNonLetter;
+            break;
+          case SIMPLE_CHARACTER_MATCH:
+            emit_function = &EmitSimpleCharacter;
+            break;
+          case CASE_CHARACTER_MATCH:
+            emit_function = &EmitAtomLetter;
+            break;
+          default:
+            break;
+        }
+        if (emit_function != NULL) {
+          bool bound_checked = emit_function(I,
+                                             compiler,
+                                             quarks->At(j),
+                                             backtrack,
+                                             cp_offset + j,
+                                             *checked_up_to < cp_offset + j,
+                                             preloaded);
+          if (bound_checked) UpdateBoundsCheck(cp_offset + j, checked_up_to);
+        }
+      }
+    } else {
+      ASSERT(elm.text_type() == TextElement::CHAR_CLASS);
+      if (pass == CHARACTER_CLASS_MATCH) {
+        if (first_element_checked && i == 0) continue;
+        if (DeterminedAlready(quick_check, elm.cp_offset())) continue;
+        RegExpCharacterClass* cc = elm.char_class();
+        EmitCharClass(assembler,
+                      cc,
+                      ascii,
+                      backtrack,
+                      cp_offset,
+                      *checked_up_to < cp_offset,
+                      preloaded,
+                      I);
+        UpdateBoundsCheck(cp_offset, checked_up_to);
+      }
+    }
+  }
 }
 
 
 intptr_t ChoiceNode::CalculatePreloadCharacters(RegExpCompiler* compiler,
-                                           intptr_t eats_at_least) {
-  UNIMPLEMENTED();
-  return NULL;
+                                                intptr_t eats_at_least) {
+  intptr_t preload_characters = Utils::Minimum(static_cast<intptr_t>(4),
+                                               eats_at_least);
+  if (compiler->macro_assembler()->CanReadUnaligned()) {
+    bool ascii = compiler->ascii();
+    if (ascii) {
+      if (preload_characters > 4) preload_characters = 4;
+      // We can't preload 3 characters because there is no machine instruction
+      // to do that.  We can't just load 4 because we could be reading
+      // beyond the end of the string, which could cause a memory fault.
+      if (preload_characters == 3) preload_characters = 2;
+    } else {
+      if (preload_characters > 2) preload_characters = 2;
+    }
+  } else {
+    if (preload_characters > 1) preload_characters = 1;
+  }
+  return preload_characters;
 }
 
 
-void ChoiceNode::EmitOutOfLineContinuation(RegExpCompiler* compiler,
-                                           Trace* trace,
-                                           GuardedAlternative alternative,
-                                           AlternativeGeneration* alt_gen,
-                                           intptr_t preload_characters,
-                                           bool next_expects_preload) {
-  UNIMPLEMENTED();
-}
-
-
 intptr_t TextNode::EatsAtLeast(intptr_t still_to_find,
                                intptr_t budget,
                                bool not_at_start) {
   intptr_t answer = Length();
   if (answer >= still_to_find) return answer;
   if (budget <= 0) return answer;
   // We are not at start after this node so we set the last argument to 'true'.
   return answer + on_success()->EatsAtLeast(still_to_find - answer,
                                             budget - 1,
                                             true);
@@ skipping 71 matching lines @@
                                                   intptr_t budget,
                                                   bool not_at_start) {
   if (budget <= 0) return 0;
   // Alternative 0 is the negative lookahead, alternative 1 is what comes
   // afterwards.
   RegExpNode* node = (*alternatives_)[1].node();
   return node->EatsAtLeast(still_to_find, budget - 1, not_at_start);
 }
 
 
+// Takes the left-most 1-bit and smears it out, setting all bits to its right.
+static inline uint32_t SmearBitsRight(uint32_t v) {
+  v |= v >> 1;
+  v |= v >> 2;
+  v |= v >> 4;
+  v |= v >> 8;
+  v |= v >> 16;
+  return v;
+}
+
+
 // Here is the meat of GetQuickCheckDetails (see also the comment on the
 // super-class in the .h file).
 //
 // We iterate along the text object, building up for each character a
 // mask and value that can be used to test for a quick failure to match.
 // The masks and values for the positions will be combined into a single
 // machine word for the current character width in order to be used in
 // generating a quick check.
 void TextNode::GetQuickCheckDetails(QuickCheckDetails* details,
                                     RegExpCompiler* compiler,
                                     intptr_t characters_filled_in,
                                     bool not_at_start) {
-  UNIMPLEMENTED();
+  ASSERT(characters_filled_in < details->characters());
+  intptr_t characters = details->characters();
+  intptr_t char_mask;
+  if (compiler->ascii()) {
+    char_mask = Symbols::kMaxOneCharCodeSymbol;
+  } else {
+    char_mask = Utf16::kMaxCodeUnit;
+  }
+  for (intptr_t k = 0; k < elms_->length(); k++) {
+    TextElement elm = elms_->At(k);
+    if (elm.text_type() == TextElement::ATOM) {
+      ZoneGrowableArray<uint16_t>* quarks = elm.atom()->data();
+      for (intptr_t i = 0; i < characters && i < quarks->length(); i++) {
+        QuickCheckDetails::Position* pos =
+            details->positions(characters_filled_in);
+        uint16_t c = quarks->At(i);
+        if (c > char_mask) {
+          // If we expect a non-ASCII character from an ASCII string,
+          // there is no way we can match. Not even case independent
+          // matching can turn an ASCII character into non-ASCII or
+          // vice versa.
+          details->set_cannot_match();
+          pos->determines_perfectly = false;
+          return;
+        }
+        if (compiler->ignore_case()) {
+          int32_t chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
+          intptr_t length =
+              GetCaseIndependentLetters(c, compiler->ascii(), chars);
+          ASSERT(length != 0);  // Can only happen if c > char_mask (see above).
+          if (length == 1) {
+            // This letter has no case equivalents, so it's nice and simple
+            // and the mask-compare will determine definitely whether we have
+            // a match at this character position.
+            pos->mask = char_mask;
+            pos->value = c;
+            pos->determines_perfectly = true;
+          } else {
+            uint32_t common_bits = char_mask;
+            uint32_t bits = chars[0];
+            for (intptr_t j = 1; j < length; j++) {
+              uint32_t differing_bits = ((chars[j] & common_bits) ^ bits);
+              common_bits ^= differing_bits;
+              bits &= common_bits;
+            }
+            // If length is 2 and common bits has only one zero in it then
+            // our mask and compare instruction will determine definitely
+            // whether we have a match at this character position.  Otherwise
+            // it can only be an approximate check.
+            uint32_t one_zero = (common_bits | ~char_mask);
+            if (length == 2 && ((~one_zero) & ((~one_zero) - 1)) == 0) {
+              pos->determines_perfectly = true;
+            }
+            pos->mask = common_bits;
+            pos->value = bits;
+          }
+        } else {
+          // Don't ignore case.  Nice simple case where the mask-compare will
+          // determine definitely whether we have a match at this character
+          // position.
+          pos->mask = char_mask;
+          pos->value = c;
+          pos->determines_perfectly = true;
+        }
+        characters_filled_in++;
+        ASSERT(characters_filled_in <= details->characters());
+        if (characters_filled_in == details->characters()) {
+          return;
+        }
+      }
+    } else {
+      QuickCheckDetails::Position* pos =
+          details->positions(characters_filled_in);
+      RegExpCharacterClass* tree = elm.char_class();
+      ZoneGrowableArray<CharacterRange>* ranges = tree->ranges();
+      if (tree->is_negated()) {
+        // A quick check uses multi-character mask and compare.  There is no
+        // useful way to incorporate a negative char class into this scheme
+        // so we just conservatively create a mask and value that will always
+        // succeed.
+        pos->mask = 0;
+        pos->value = 0;
+      } else {
+        intptr_t first_range = 0;
+        while (ranges->At(first_range).from() > char_mask) {
+          first_range++;
+          if (first_range == ranges->length()) {
+            details->set_cannot_match();
+            pos->determines_perfectly = false;
+            return;
+          }
+        }
+        CharacterRange range = ranges->At(first_range);
+        uint16_t from = range.from();
+        uint16_t to = range.to();
+        if (to > char_mask) {
+          to = char_mask;
+        }
+        uint32_t differing_bits = (from ^ to);
+        // A mask and compare is only perfect if the differing bits form a
+        // number like 00011111 with one single block of trailing 1s.
+        if ((differing_bits & (differing_bits + 1)) == 0 &&
+             from + differing_bits == to) {
+          pos->determines_perfectly = true;
+        }
+        uint32_t common_bits = ~SmearBitsRight(differing_bits);
+        uint32_t bits = (from & common_bits);
+        for (intptr_t i = first_range + 1; i < ranges->length(); i++) {
+          CharacterRange range = ranges->At(i);
+          uint16_t from = range.from();
+          uint16_t to = range.to();
+          if (from > char_mask) continue;
+          if (to > char_mask) to = char_mask;
+          // Here we are combining more ranges into the mask and compare
+          // value.  With each new range the mask becomes more sparse and
+          // so the chances of a false positive rise.  A character class
+          // with multiple ranges is assumed never to be equivalent to a
+          // mask and compare operation.
+          pos->determines_perfectly = false;
+          uint32_t new_common_bits = (from ^ to);
+          new_common_bits = ~SmearBitsRight(new_common_bits);
+          common_bits &= new_common_bits;
+          bits &= new_common_bits;
+          uint32_t differing_bits = (from & common_bits) ^ bits;
+          common_bits ^= differing_bits;
+          bits &= common_bits;
+        }
+        pos->mask = common_bits;
+        pos->value = bits;
+      }
+      characters_filled_in++;
+      ASSERT(characters_filled_in <= details->characters());
+      if (characters_filled_in == details->characters()) {
+        return;
+      }
+    }
+  }
+  ASSERT(characters_filled_in != details->characters());
+  if (!details->cannot_match()) {
+    on_success()-> GetQuickCheckDetails(details,
+                                        compiler,
+                                        characters_filled_in,
+                                        true);
+  }
 }
 
 
 void LoopChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details,
                                           RegExpCompiler* compiler,
                                           intptr_t characters_filled_in,
                                           bool not_at_start) {
-  UNIMPLEMENTED();
+  if (body_can_be_zero_length_ || info()->visited) return;
+  VisitMarker marker(info());
+  return ChoiceNode::GetQuickCheckDetails(details,
+                                          compiler,
+                                          characters_filled_in,
+                                          not_at_start);
 }
 
 
 intptr_t TextNode::Length() {
   TextElement elm = elms_->Last();
   ASSERT(elm.cp_offset() >= 0);
   return elm.cp_offset() + elm.length();
 }
 
 
@@ skipping 518 matching lines @@
                         Isolate* isolate) {
   uint16_t bottom = from();
   uint16_t top = to();
   if (is_ascii && !RangeContainsLatin1Equivalents(*this)) {
     if (bottom > Symbols::kMaxOneCharCodeSymbol) return;
     if (top > Symbols::kMaxOneCharCodeSymbol) {
       top = Symbols::kMaxOneCharCodeSymbol;
     }
   }
 
-  UNIMPLEMENTED();
+  unibrow::Mapping<unibrow::Ecma262UnCanonicalize> jsregexp_uncanonicalize;
+  unibrow::Mapping<unibrow::CanonicalizationRange> jsregexp_canonrange;
+  int32_t chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
+  if (top == bottom) {
+    // If this is a singleton we just expand the one character.
+    intptr_t length = jsregexp_uncanonicalize.get(bottom, '\0', chars); // NOLINT
+    for (intptr_t i = 0; i < length; i++) {
+      uint32_t chr = chars[i];
+      if (chr != bottom) {
+        ranges->Add(CharacterRange::Singleton(chars[i]));
+      }
+    }
+  } else {
+    // If this is a range we expand the characters block by block,
+    // expanding contiguous subranges (blocks) one at a time.
+    // The approach is as follows.  For a given start character we
+    // look up the remainder of the block that contains it (represented
+    // by the end point), for instance we find 'z' if the character
+    // is 'c'.  A block is characterized by the property
+    // that all characters uncanonicalize in the same way, except that
+    // each entry in the result is incremented by the distance from the first
+    // element.  So a-z is a block because 'a' uncanonicalizes to ['a', 'A'] and // NOLINT
+    // the k'th letter uncanonicalizes to ['a' + k, 'A' + k].
+    // Once we've found the end point we look up its uncanonicalization
+    // and produce a range for each element.  For instance for [c-f]
+    // we look up ['z', 'Z'] and produce [c-f] and [C-F].  We then only
+    // add a range if it is not already contained in the input, so [c-f]
+    // will be skipped but [C-F] will be added.  If this range is not
+    // completely contained in a block we do this for all the blocks
+    // covered by the range (handling characters that is not in a block
+    // as a "singleton block").
+    int32_t range[unibrow::Ecma262UnCanonicalize::kMaxWidth];
+    intptr_t pos = bottom;
+    while (pos <= top) {
+      intptr_t length = jsregexp_canonrange.get(pos, '\0', range);
+      uint16_t block_end;
+      if (length == 0) {
+        block_end = pos;
+      } else {
+        ASSERT(length == 1);
+        block_end = range[0];
+      }
+      intptr_t end = (block_end > top) ? top : block_end;
+      length = jsregexp_uncanonicalize.get(block_end, '\0', range); // NOLINT
+      for (intptr_t i = 0; i < length; i++) {
+        uint32_t c = range[i];
+        uint16_t range_from = c - (block_end - pos);
+        uint16_t range_to = c - (block_end - end);
+        if (!(bottom <= range_from && range_to <= top)) {
+          ranges->Add(CharacterRange(range_from, range_to));
+        }
+      }
+      pos = end + 1;
+    }
+  }
+}
 
-//  unibrow::uchar chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
-//  if (top == bottom) {
-//    // If this is a singleton we just expand the one character.
-//    intptr_t length = isolate->jsregexp_uncanonicalize()->get(bottom, '\0', chars); // NOLINT
-//    for (intptr_t i = 0; i < length; i++) {
-//      uint32_t chr = chars[i];
-//      if (chr != bottom) {
-//        ranges->Add(CharacterRange::Singleton(chars[i]), zone);
-//      }
-//    }
-//  } else {
-//    // If this is a range we expand the characters block by block,
-//    // expanding contiguous subranges (blocks) one at a time.
-//    // The approach is as follows.  For a given start character we
-//    // look up the remainder of the block that contains it (represented
-//    // by the end point), for instance we find 'z' if the character
-//    // is 'c'.  A block is characterized by the property
-//    // that all characters uncanonicalize in the same way, except that
-//    // each entry in the result is incremented by the distance from the first
-//    // element.  So a-z is a block because 'a' uncanonicalizes to ['a', 'A'] and // NOLINT
-//    // the k'th letter uncanonicalizes to ['a' + k, 'A' + k].
-//    // Once we've found the end point we look up its uncanonicalization
-//    // and produce a range for each element.  For instance for [c-f]
-//    // we look up ['z', 'Z'] and produce [c-f] and [C-F].  We then only
-//    // add a range if it is not already contained in the input, so [c-f]
-//    // will be skipped but [C-F] will be added.  If this range is not
-//    // completely contained in a block we do this for all the blocks
-//    // covered by the range (handling characters that is not in a block
-//    // as a "singleton block").
-//    unibrow::uchar range[unibrow::Ecma262UnCanonicalize::kMaxWidth];
-//    intptr_t pos = bottom;
-//    while (pos <= top) {
-//      intptr_t length = isolate->jsregexp_canonrange()->get(pos, '\0', range);
-//      uint16_t block_end;
-//      if (length == 0) {
-//        block_end = pos;
-//      } else {
-//        ASSERT(length == 1);
-//        block_end = range[0];
-//      }
-//      intptr_t end = (block_end > top) ? top : block_end;
-//      length = isolate->jsregexp_uncanonicalize()->get(block_end, '\0', range); // NOLINT
-//      for (intptr_t i = 0; i < length; i++) {
-//        uint32_t c = range[i];
-//        uint16_t range_from = c - (block_end - pos);
-//        uint16_t range_to = c - (block_end - end);
-//        if (!(bottom <= range_from && range_to <= top)) {
-//          ranges->Add(CharacterRange(range_from, range_to));
-//        }
-//      }
-//      pos = end + 1;
-//    }
-//  }
+
+// -------------------------------------------------------------------
+// Splay tree
+
+
+// Workaround for the fact that ZoneGrowableArray does not have contains().
+static bool ArrayContains(ZoneGrowableArray<unsigned>* array,
+                          unsigned value) {
+  for (intptr_t i = 0; i < array->length(); i++) {
+    if (array->At(i) == value) {
+      return true;
+    }
+  }
+  return false;
+}
+
+
+void OutSet::Set(unsigned value, Isolate* isolate) {
+  if (value < kFirstLimit) {
+    first_ |= (1 << value);
+  } else {
+    if (remaining_ == NULL)
+      remaining_ = new(isolate) ZoneGrowableArray<unsigned>(1);
+
+    bool remaining_contains_value = ArrayContains(remaining_, value);
+    if (remaining_->is_empty() || !remaining_contains_value) {
+      remaining_->Add(value);
+    }
+  }
+}
+
+
+bool OutSet::Get(unsigned value) const {
+  if (value < kFirstLimit) {
+    return (first_ & (1 << value)) != 0;
+  } else if (remaining_ == NULL) {
+    return false;
+  } else {
+    return ArrayContains(remaining_, value);
+  }
 }
 
 void AssertionNode::GetQuickCheckDetails(QuickCheckDetails* details,
                                          RegExpCompiler* compiler,
                                          intptr_t filled_in,
                                          bool not_at_start) {
   if (assertion_type_ == AT_START && not_at_start) {
     details->set_cannot_match();
     return;
   }
@@ skipping 33 matching lines @@
 intptr_t LoopChoiceNode::EatsAtLeast(intptr_t still_to_find,
                                      intptr_t budget,
                                      bool not_at_start) {
   return EatsAtLeastHelper(still_to_find,
                            budget - 1,
                            loop_node_,
                            not_at_start);
 }
 
 
-class RecursionCheck {
- public:
-  explicit RecursionCheck(RegExpCompiler* compiler) : compiler_(compiler) {
-    compiler->IncrementRecursionDepth();
-  }
-  ~RecursionCheck() { compiler_->DecrementRecursionDepth(); }
- private:
-  RegExpCompiler* compiler_;
-};
-
-
-DispatchTable* ChoiceNode::GetTable(bool ignore_case) {
-  UNIMPLEMENTED();
-  return NULL;
-}
-
-
 void ChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details,
                                       RegExpCompiler* compiler,
                                       intptr_t characters_filled_in,
                                       bool not_at_start) {
   not_at_start = (not_at_start || not_at_start_);
   intptr_t choice_count = alternatives_->length();
   ASSERT(choice_count > 0);
   (*alternatives_)[0].node()->GetQuickCheckDetails(details,
                                                     compiler,
                                                     characters_filled_in,
@@ skipping 102 matching lines @@
     RegExpCompiler* compiler,
     intptr_t filled_in,
     bool not_at_start) {
   // Alternative 0 is the negative lookahead, alternative 1 is what comes
   // afterwards.
   RegExpNode* node = (*alternatives_)[1].node();
   return node->GetQuickCheckDetails(details, compiler, filled_in, not_at_start);
 }
 
 
+static RegExpEngine::CompilationResult IrregexpRegExpTooBig() {
+  return RegExpEngine::CompilationResult("RegExp too big");
+}
+
+
 // Attempts to compile the regexp using an Irregexp code generator.  Returns
 // a fixed array or a null handle depending on whether it succeeded.
 RegExpCompiler::RegExpCompiler(intptr_t capture_count, bool ignore_case,
-                               bool ascii)
+                               intptr_t specialization_cid)
     : next_register_(2 * (capture_count + 1)),
+      work_list_(NULL),
       recursion_depth_(0),
       ignore_case_(ignore_case),
-      ascii_(ascii),
+      specialization_cid_(specialization_cid),
       reg_exp_too_big_(false),
       current_expansion_factor_(1),
       isolate_(Isolate::Current()) {
   accept_ = new(I) EndNode(EndNode::ACCEPT, I);
 }
 
 
 RegExpEngine::CompilationResult RegExpCompiler::Assemble(
-    RegExpMacroAssembler* macro_assembler,
+    IRRegExpMacroAssembler* macro_assembler,
     RegExpNode* start,
     intptr_t capture_count,
     const String& pattern) {
-  UNIMPLEMENTED();
-  return RegExpEngine::CompilationResult("");
+  static const bool use_slow_safe_regexp_compiler = false;
+
+  macro_assembler->set_slow_safe(use_slow_safe_regexp_compiler);
+  macro_assembler_ = macro_assembler;
+
+  ZoneGrowableArray<RegExpNode*> work_list(0);
+  work_list_ = &work_list;
+  BlockLabel fail;
+  macro_assembler_->PushBacktrack(&fail);
+  Trace new_trace;
+  start->Emit(this, &new_trace);
+  macro_assembler_->BindBlock(&fail);
+  macro_assembler_->Fail();
+  while (!work_list.is_empty()) {
+    work_list.RemoveLast()->Emit(this, &new_trace);
+  }
+  if (reg_exp_too_big_) return IrregexpRegExpTooBig();
+
+  macro_assembler->FinalizeIndirectGotos();
+
+  return RegExpEngine::CompilationResult(macro_assembler,
+                                         macro_assembler->graph_entry(),
+                                         macro_assembler->num_blocks(),
+                                         macro_assembler->num_stack_locals());
 }
 
 
 RegExpEngine::CompilationResult RegExpEngine::Compile(
     RegExpCompileData* data,
-    bool ignore_case,
-    bool is_global,
-    bool is_multiline,
-    const String& pattern,
-    const String& sample_subject,
-    bool is_ascii) {
+    const ParsedFunction* parsed_function,
+    ZoneGrowableArray<const ICData*>* ic_data_array) {
   Isolate* isolate = Isolate::Current();
 
-  RegExpCompiler compiler(data->capture_count, ignore_case, is_ascii);
+  const Function& function = parsed_function->function();
+  const intptr_t specialization_cid = function.regexp_specialization();
+  const bool is_ascii = (specialization_cid == kOneByteStringCid ||
+                         specialization_cid == kExternalOneByteStringCid);
+  JSRegExp& regexp = JSRegExp::Handle(isolate, function.regexp());
+  const String& pattern = String::Handle(isolate, regexp.pattern());
+  const String& sample_subject =
+      String::Handle(isolate, regexp.sample_subject(specialization_cid));
 
-  // TODO(jgruber): Character frequency sampling.
+  ASSERT(!regexp.IsNull());
+  ASSERT(!pattern.IsNull());
+  ASSERT(!sample_subject.IsNull());
+
+  const bool ignore_case = regexp.is_ignore_case();
+  const bool is_global = regexp.is_global();
+
+  RegExpCompiler compiler(data->capture_count, ignore_case, specialization_cid);
+
+  // Sample some characters from the middle of the string.
+  static const intptr_t kSampleSize = 128;
+
+  intptr_t chars_sampled = 0;
+  intptr_t half_way = (sample_subject.Length() - kSampleSize) / 2;
+  for (intptr_t i = Utils::Maximum(static_cast<intptr_t>(0), half_way);
+       i < sample_subject.Length() && chars_sampled < kSampleSize;
+       i++, chars_sampled++) {
+    compiler.frequency_collator()->CountCharacter(sample_subject.CharAt(i));
+  }
 
   // Wrap the body of the regexp in capture #0.
   RegExpNode* captured_body = RegExpCapture::ToNode(data->tree,
                                                     0,
                                                     &compiler,
                                                     compiler.accept());
 
   RegExpNode* node = captured_body;
   bool is_start_anchored = data->tree->IsAnchoredAtStart();
+  bool is_end_anchored = data->tree->IsAnchoredAtEnd();
+  intptr_t max_length = data->tree->max_match();
   if (!is_start_anchored) {
     // Add a .*? at the beginning, outside the body capture, unless
     // this expression is anchored at the beginning.
     RegExpNode* loop_node =
         RegExpQuantifier::ToNode(0,
                                  RegExpTree::kInfinity,
                                  false,
                                  new(isolate) RegExpCharacterClass('*'),
                                  &compiler,
                                  captured_body,
@@ skipping 23 matching lines @@
 
   if (node == NULL) node = new(isolate) EndNode(EndNode::BACKTRACK, isolate);
   data->node = node;
   Analysis analysis(ignore_case, is_ascii);
   analysis.EnsureAnalyzed(node);
   if (analysis.has_failed()) {
     const char* error_message = analysis.error_message();
     return CompilationResult(error_message);
   }
 
-  // TODO(jgruber): Assemble native code.
-
-  return RegExpEngine::CompilationResult("");
-}
-
-
+  // Native regexp implementation.
+
+  IRRegExpMacroAssembler* macro_assembler =
+      new(isolate) IRRegExpMacroAssembler(specialization_cid,
+                                          data->capture_count,
+                                          parsed_function,
+                                          ic_data_array,
+                                          isolate);
+
+  // Inserted here, instead of in Assembler, because it depends on information
+  // in the AST that isn't replicated in the Node structure.
+  static const intptr_t kMaxBacksearchLimit = 1024;
+  if (is_end_anchored &&
+      !is_start_anchored &&
+      max_length < kMaxBacksearchLimit) {
+    macro_assembler->SetCurrentPositionFromEnd(max_length);
+  }
+
+  if (is_global) {
+    macro_assembler->set_global_mode(
+        (data->tree->min_match() > 0)
+            ? RegExpMacroAssembler::GLOBAL_NO_ZERO_LENGTH_CHECK
+            : RegExpMacroAssembler::GLOBAL);
+  }
+
+  RegExpEngine::CompilationResult result =
+      compiler.Assemble(macro_assembler,
+                        node,
+                        data->capture_count,
+                        pattern);
+
+  if (FLAG_trace_irregexp) {
+    macro_assembler->PrintBlocks();
+  }
+
+  return result;
+}
+
+
+static void CreateSpecializedFunction(Isolate* isolate,
+                                      JSRegExp* regexp,
+                                      intptr_t specialization_cid,
+                                      const Object& owner) {
+  const intptr_t kParamCount = RegExpMacroAssembler::kParamCount;
+
+  Function& fn = Function::Handle(isolate,
+        Function::New(String::Handle(isolate, Symbols::New("RegExp")),
+                      RawFunction::kIrregexpFunction,
+                      true,   // Static.
+                      false,  // Not const.
+                      false,  // Not abstract.
+                      false,  // Not external.
+                      false,  // Not native.
+                      owner,
+                      0));  // Requires a non-negative token position.
+
+  fn.set_num_fixed_parameters(kParamCount);
+  fn.set_parameter_types(
+        Array::Handle(isolate, Array::New(kParamCount, Heap::kOld)));
+  fn.set_parameter_names(
+        Array::Handle(isolate, Array::New(kParamCount, Heap::kOld)));
+  fn.SetParameterTypeAt(0, Type::Handle(isolate, Type::DynamicType()));
+  fn.SetParameterNameAt(0, Symbols::string_param_());
+  fn.SetParameterTypeAt(1, Type::Handle(isolate, Type::DynamicType()));
+  fn.SetParameterNameAt(1, Symbols::start_index_param_());
+  fn.set_result_type(Type::Handle(isolate, Type::ArrayType()));
+
+  // Cache the result.
+  regexp->set_function(specialization_cid, fn);
+
+  fn.set_regexp(*regexp);
+  fn.set_regexp_specialization(specialization_cid);
+
+  // The function is compiled lazily during the first call.
+}
+
+
+RawJSRegExp* RegExpEngine::New(Isolate* isolate,
+                               const String& pattern,
+                               bool multi_line,
+                               bool ignore_case) {
+  JSRegExp& regexp = JSRegExp::Handle(JSRegExp::New(0));
+
+  regexp.set_pattern(pattern);
+
+  if (multi_line) regexp.set_is_multi_line();
+  if (ignore_case) regexp.set_is_ignore_case();
+
+  // TODO(jgruber): We might want to use normal string searching algorithms
+  // for simple patterns.
+  regexp.set_is_complex();
+  regexp.set_is_global();   // All dart regexps are global.
+
+  const Library& lib = Library::Handle(isolate, Library::CoreLibrary());
+  const Class& owner = Class::Handle(isolate,
+        lib.LookupClass(String::Handle(isolate, Symbols::New("RegExp"))));
+
+  CreateSpecializedFunction(isolate, &regexp, kOneByteStringCid, owner);
+  CreateSpecializedFunction(isolate, &regexp, kTwoByteStringCid, owner);
+  CreateSpecializedFunction(isolate, &regexp, kExternalOneByteStringCid, owner);
+  CreateSpecializedFunction(isolate, &regexp, kExternalTwoByteStringCid, owner);
+
+  return regexp.raw();
+}
+
+
+void BoyerMoorePositionInfo::Set(intptr_t character) {
+  SetInterval(Interval(character, character));
+}
+
+
+ContainedInLattice AddRange(ContainedInLattice containment,
+                            const intptr_t* ranges,
+                            intptr_t ranges_length,
+                            Interval new_range) {
+  ASSERT((ranges_length & 1) == 1);
+  ASSERT(ranges[ranges_length - 1] == Utf16::kMaxCodeUnit + 1);
+  if (containment == kLatticeUnknown) return containment;
+  bool inside = false;
+  intptr_t last = 0;
+  for (intptr_t i = 0; i < ranges_length;
+       inside = !inside, last = ranges[i], i++) {
+    // Consider the range from last to ranges[i].
+    // We haven't got to the new range yet.
+    if (ranges[i] <= new_range.from()) continue;
+    // New range is wholly inside last-ranges[i].  Note that new_range.to() is
+    // inclusive, but the values in ranges are not.
+    if (last <= new_range.from() && new_range.to() < ranges[i]) {
+      return Combine(containment, inside ? kLatticeIn : kLatticeOut);
+    }
+    return kLatticeUnknown;
+  }
+  return containment;
+}
+
+
+void BoyerMoorePositionInfo::SetInterval(const Interval& interval) {
+  s_ = AddRange(s_, kSpaceRanges, kSpaceRangeCount, interval);
+  w_ = AddRange(w_, kWordRanges, kWordRangeCount, interval);
+  d_ = AddRange(d_, kDigitRanges, kDigitRangeCount, interval);
+  surrogate_ =
+      AddRange(surrogate_, kSurrogateRanges, kSurrogateRangeCount, interval);
+  if (interval.to() - interval.from() >= kMapSize - 1) {
+    if (map_count_ != kMapSize) {
+      map_count_ = kMapSize;
+      for (intptr_t i = 0; i < kMapSize; i++) (*map_)[i] = true;
+    }
+    return;
+  }
+  for (intptr_t i = interval.from(); i <= interval.to(); i++) {
+    intptr_t mod_character = (i & kMask);
+    if (!map_->At(mod_character)) {
+      map_count_++;
+      (*map_)[mod_character] = true;
+    }
+    if (map_count_ == kMapSize) return;
+  }
+}
+
+
+void BoyerMoorePositionInfo::SetAll() {
+  s_ = w_ = d_ = kLatticeUnknown;
+  if (map_count_ != kMapSize) {
+    map_count_ = kMapSize;
+    for (intptr_t i = 0; i < kMapSize; i++) (*map_)[i] = true;
+  }
+}
+
+
+BoyerMooreLookahead::BoyerMooreLookahead(
+    intptr_t length, RegExpCompiler* compiler, Isolate* isolate)
+    : length_(length),
+      compiler_(compiler) {
+  if (compiler->ascii()) {
+    max_char_ = Symbols::kMaxOneCharCodeSymbol;
+  } else {
+    max_char_ = Utf16::kMaxCodeUnit;
+  }
+  bitmaps_ = new(isolate) ZoneGrowableArray<BoyerMoorePositionInfo*>(length);
+  for (intptr_t i = 0; i < length; i++) {
+    bitmaps_->Add(new(isolate) BoyerMoorePositionInfo(isolate));
+  }
+}
+
+
+// Take all the characters that will not prevent a successful match if they
+// occur in the subject string in the range between min_lookahead and
+// max_lookahead (inclusive) measured from the current position.  If the
+// character at max_lookahead offset is not one of these characters, then we
+// can safely skip forwards by the number of characters in the range.
+intptr_t BoyerMooreLookahead::GetSkipTable(intptr_t min_lookahead,
+                                           intptr_t max_lookahead,
+                                           TypedData* boolean_skip_table) {

[Florian Schneider, 2014/10/01 17:04:14]

const TypedData& since this will never be NULL.

[jgruber1, 2014/10/03 18:59:52]

Done.

+  const intptr_t kSize = RegExpMacroAssembler::kTableSize;
+
+  const intptr_t kSkipArrayEntry = 0;
+  const intptr_t kDontSkipArrayEntry = 1;
+
+  for (intptr_t i = 0; i < kSize; i++) {
+    boolean_skip_table->SetUint8(i, kSkipArrayEntry);
+  }
+  intptr_t skip = max_lookahead + 1 - min_lookahead;
+
+  for (intptr_t i = max_lookahead; i >= min_lookahead; i--) {
+    BoyerMoorePositionInfo* map = bitmaps_->At(i);
+    for (intptr_t j = 0; j < kSize; j++) {
+      if (map->at(j)) {
+        boolean_skip_table->SetUint8(j, kDontSkipArrayEntry);
+      }
+    }
+  }
+
+  return skip;
+}
+
+
+// Find the longest range of lookahead that has the fewest number of different
+// characters that can occur at a given position.  Since we are optimizing two
+// different parameters at once this is a tradeoff.
+bool BoyerMooreLookahead::FindWorthwhileInterval(intptr_t* from, intptr_t* to) {
+  intptr_t biggest_points = 0;
+  // If more than 32 characters out of 128 can occur it is unlikely that we can
+  // be lucky enough to step forwards much of the time.
+  const intptr_t kMaxMax = 32;
+  for (intptr_t max_number_of_chars = 4;
+       max_number_of_chars < kMaxMax;
+       max_number_of_chars *= 2) {
+    biggest_points =
+        FindBestInterval(max_number_of_chars, biggest_points, from, to);
+  }
+  if (biggest_points == 0) return false;
+  return true;
+}
+
+
+// Find the highest-points range between 0 and length_ where the character
+// information is not too vague.  'Too vague' means that there are more than
+// max_number_of_chars that can occur at this position.  Calculates the number
+// of points as the product of width-of-the-range and
+// probability-of-finding-one-of-the-characters, where the probability is
+// calculated using the frequency distribution of the sample subject string.
+intptr_t BoyerMooreLookahead::FindBestInterval(
+    intptr_t max_number_of_chars,
+    intptr_t old_biggest_points,
+    intptr_t* from,
+    intptr_t* to) {
+  intptr_t biggest_points = old_biggest_points;
+  static const intptr_t kSize = RegExpMacroAssembler::kTableSize;
+  for (intptr_t i = 0; i < length_; ) {
+    while (i < length_ && Count(i) > max_number_of_chars) i++;
+    if (i == length_) break;
+    intptr_t remembered_from = i;
+    bool union_map[kSize];
+    for (intptr_t j = 0; j < kSize; j++) union_map[j] = false;
+    while (i < length_ && Count(i) <= max_number_of_chars) {
+      BoyerMoorePositionInfo* map = bitmaps_->At(i);
+      for (intptr_t j = 0; j < kSize; j++) union_map[j] |= map->at(j);
+      i++;
+    }
+    intptr_t frequency = 0;
+    for (intptr_t j = 0; j < kSize; j++) {
+      if (union_map[j]) {
+        // Add 1 to the frequency to give a small per-character boost for
+        // the cases where our sampling is not good enough and many
+        // characters have a frequency of zero.  This means the frequency
+        // can theoretically be up to 2*kSize though we treat it mostly as
+        // a fraction of kSize.
+        frequency += compiler_->frequency_collator()->Frequency(j) + 1;
+      }
+    }
+    // We use the probability of skipping times the distance we are skipping to
+    // judge the effectiveness of this.  Actually we have a cut-off:  By
+    // dividing by 2 we switch off the skipping if the probability of skipping
+    // is less than 50%.  This is because the multibyte mask-and-compare
+    // skipping in quickcheck is more likely to do well on this case.
+    bool in_quickcheck_range = ((i - remembered_from < 4) ||
+        (compiler_->ascii() ? remembered_from <= 4 : remembered_from <= 2));
+    // Called 'probability' but it is only a rough estimate and can actually
+    // be outside the 0-kSize range.
+    intptr_t probability =
+        (in_quickcheck_range ? kSize / 2 : kSize) - frequency;
+    intptr_t points = (i - remembered_from) * probability;
+    if (points > biggest_points) {
+      *from = remembered_from;
+      *to = i - 1;
+      biggest_points = points;
+    }
+  }
+  return biggest_points;
+}
+
+
+// See comment above on the implementation of GetSkipTable.
+bool BoyerMooreLookahead::EmitSkipInstructions(RegExpMacroAssembler* masm) {
+  const intptr_t kSize = RegExpMacroAssembler::kTableSize;
+
+  intptr_t min_lookahead = 0;
+  intptr_t max_lookahead = 0;
+
+  if (!FindWorthwhileInterval(&min_lookahead, &max_lookahead)) return false;
+
+  bool found_single_character = false;
+  intptr_t single_character = 0;
+  for (intptr_t i = max_lookahead; i >= min_lookahead; i--) {
+    BoyerMoorePositionInfo* map = bitmaps_->At(i);
+    if (map->map_count() > 1 ||
+        (found_single_character && map->map_count() != 0)) {
+      found_single_character = false;
+      break;
+    }
+    for (intptr_t j = 0; j < kSize; j++) {
+      if (map->at(j)) {
+        found_single_character = true;
+        single_character = j;
+        break;
+      }
+    }
+  }
+
+  intptr_t lookahead_width = max_lookahead + 1 - min_lookahead;
+
+  if (found_single_character && lookahead_width == 1 && max_lookahead < 3) {
+    // The mask-compare can probably handle this better.
+    return false;
+  }
+
+  if (found_single_character) {
+    BlockLabel cont, again;
+    masm->BindBlock(&again);
+    masm->LoadCurrentCharacter(max_lookahead, &cont, true);
+    if (max_char_ > kSize) {
+      masm->CheckCharacterAfterAnd(single_character,
+                                   RegExpMacroAssembler::kTableMask,
+                                   &cont);
+    } else {
+      masm->CheckCharacter(single_character, &cont);
+    }
+    masm->AdvanceCurrentPosition(lookahead_width);
+    masm->GoTo(&again);
+    masm->BindBlock(&cont);
+    return true;
+  }
+
+  TypedData& boolean_skip_table = TypedData::ZoneHandle(

[Florian Schneider, 2014/10/01 17:04:14]

const TypedData&

[jgruber1, 2014/10/03 18:59:52]

Done.

+        compiler_->isolate(),
+        TypedData::New(kTypedDataUint8ArrayCid, kSize, Heap::kOld));
+  intptr_t skip_distance = GetSkipTable(
+      min_lookahead, max_lookahead, &boolean_skip_table);
+  ASSERT(skip_distance != 0);
+
+  BlockLabel cont, again;
+
+  masm->BindBlock(&again);
+  masm->LoadCurrentCharacter(max_lookahead, &cont, true);
+  masm->CheckBitInTable(boolean_skip_table, &cont);
+  masm->AdvanceCurrentPosition(skip_distance);
+  masm->GoTo(&again);
+  masm->BindBlock(&cont);
+
+  return true;
+}
+
+
 // -------------------------------------------------------------------
 // Analysis
 
 
 void Analysis::EnsureAnalyzed(RegExpNode* that) {
   if (that->info()->been_analyzed || that->info()->being_analyzed)
     return;
   that->info()->being_analyzed = true;
   that->Accept(this);
   that->info()->being_analyzed = false;
@@ skipping 168 matching lines @@
 
 
 void DotPrinter::PrintAttributes(RegExpNode* that) {
   OS::Print("  a%p [shape=Mrecord, color=grey, fontcolor=grey, "
             "margin=0.1, fontsize=10, label=\"{", that);
   AttributePrinter printer;
   NodeInfo* info = that->info();
   printer.PrintBit("NI", info->follows_newline_interest);
   printer.PrintBit("WI", info->follows_word_interest);
   printer.PrintBit("SI", info->follows_start_interest);
-  Label* label = that->label();
+  BlockLabel* label = that->label();
   if (label->IsBound())
     printer.PrintPositive("@", label->Position());
   OS::Print("}\"];\n"
             "  a%p -> n%p [style=dashed, color=grey, arrowhead=none];\n",
             that, that);
 }
 
 
 static const bool kPrintDispatchTable = false;
 void DotPrinter::VisitChoice(ChoiceNode* that) {
@@ skipping 141 matching lines @@
                             RegExpNode* node,
                             bool ignore_case) {
   DotPrinter printer(ignore_case);
   printer.PrintNode(label, node);
 }
 
 
 #endif  // DEBUG
 
 }  // namespace dart