xref: /external/v8/src/jsregexp.cc

// Copyright 2011 the V8 project authors. All rights reserved.
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
//     * Redistributions of source code must retain the above copyright
//       notice, this list of conditions and the following disclaimer.
//     * Redistributions in binary form must reproduce the above
//       copyright notice, this list of conditions and the following
//       disclaimer in the documentation and/or other materials provided
//       with the distribution.
//     * Neither the name of Google Inc. nor the names of its
//       contributors may be used to endorse or promote products derived
//       from this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#include "v8.h"

#include "ast.h"
#include "compiler.h"
#include "execution.h"
#include "factory.h"
#include "jsregexp.h"
#include "platform.h"
#include "string-search.h"
#include "runtime.h"
#include "compilation-cache.h"
#include "string-stream.h"
#include "parser.h"
#include "regexp-macro-assembler.h"
#include "regexp-macro-assembler-tracer.h"
#include "regexp-macro-assembler-irregexp.h"
#include "regexp-stack.h"

#ifndef V8_INTERPRETED_REGEXP
#if V8_TARGET_ARCH_IA32
#include "ia32/regexp-macro-assembler-ia32.h"
#elif V8_TARGET_ARCH_X64
#include "x64/regexp-macro-assembler-x64.h"
#elif V8_TARGET_ARCH_ARM
#include "arm/regexp-macro-assembler-arm.h"
#elif V8_TARGET_ARCH_MIPS
#include "mips/regexp-macro-assembler-mips.h"
#else
#error Unsupported target architecture.
#endif
#endif

#include "interpreter-irregexp.h"


namespace v8 {
namespace internal {

Handle<Object> RegExpImpl::CreateRegExpLiteral(Handle<JSFunction> constructor,
                                               Handle<String> pattern,
                                               Handle<String> flags,
                                               bool* has_pending_exception) {
  // Call the construct code with 2 arguments.
  Handle<Object> argv[] = { pattern, flags };
  return Execution::New(constructor, ARRAY_SIZE(argv), argv,
                        has_pending_exception);
}


static JSRegExp::Flags RegExpFlagsFromString(Handle<String> str) {
  int flags = JSRegExp::NONE;
  for (int i = 0; i < str->length(); i++) {
    switch (str->Get(i)) {
      case 'i':
        flags |= JSRegExp::IGNORE_CASE;
        break;
      case 'g':
        flags |= JSRegExp::GLOBAL;
        break;
      case 'm':
        flags |= JSRegExp::MULTILINE;
        break;
    }
  }
  return JSRegExp::Flags(flags);
}


static inline void ThrowRegExpException(Handle<JSRegExp> re,
                                        Handle<String> pattern,
                                        Handle<String> error_text,
                                        const char* message) {
  Isolate* isolate = re->GetIsolate();
  Factory* factory = isolate->factory();
  Handle<FixedArray> elements = factory->NewFixedArray(2);
  elements->set(0, *pattern);
  elements->set(1, *error_text);
  Handle<JSArray> array = factory->NewJSArrayWithElements(elements);
  Handle<Object> regexp_err = factory->NewSyntaxError(message, array);
  isolate->Throw(*regexp_err);
}


// Generic RegExp methods. Dispatches to implementation specific methods.


Handle<Object> RegExpImpl::Compile(Handle<JSRegExp> re,
                                   Handle<String> pattern,
                                   Handle<String> flag_str) {
  Isolate* isolate = re->GetIsolate();
  JSRegExp::Flags flags = RegExpFlagsFromString(flag_str);
  CompilationCache* compilation_cache = isolate->compilation_cache();
  Handle<FixedArray> cached = compilation_cache->LookupRegExp(pattern, flags);
  bool in_cache = !cached.is_null();
  LOG(isolate, RegExpCompileEvent(re, in_cache));

  Handle<Object> result;
  if (in_cache) {
    re->set_data(*cached);
    return re;
  }
  pattern = FlattenGetString(pattern);
  ZoneScope zone_scope(isolate, DELETE_ON_EXIT);
  PostponeInterruptsScope postpone(isolate);
  RegExpCompileData parse_result;
  FlatStringReader reader(isolate, pattern);
  if (!RegExpParser::ParseRegExp(&reader, flags.is_multiline(),
                                 &parse_result)) {
    // Throw an exception if we fail to parse the pattern.
    ThrowRegExpException(re,
                         pattern,
                         parse_result.error,
                         "malformed_regexp");
    return Handle<Object>::null();
  }

  if (parse_result.simple && !flags.is_ignore_case()) {
    // Parse-tree is a single atom that is equal to the pattern.
    AtomCompile(re, pattern, flags, pattern);
  } else if (parse_result.tree->IsAtom() &&
      !flags.is_ignore_case() &&
      parse_result.capture_count == 0) {
    RegExpAtom* atom = parse_result.tree->AsAtom();
    Vector<const uc16> atom_pattern = atom->data();
    Handle<String> atom_string =
        isolate->factory()->NewStringFromTwoByte(atom_pattern);
    AtomCompile(re, pattern, flags, atom_string);
  } else {
    IrregexpInitialize(re, pattern, flags, parse_result.capture_count);
  }
  ASSERT(re->data()->IsFixedArray());
  // Compilation succeeded so the data is set on the regexp
  // and we can store it in the cache.
  Handle<FixedArray> data(FixedArray::cast(re->data()));
  compilation_cache->PutRegExp(pattern, flags, data);

  return re;
}


Handle<Object> RegExpImpl::Exec(Handle<JSRegExp> regexp,
                                Handle<String> subject,
                                int index,
                                Handle<JSArray> last_match_info) {
  switch (regexp->TypeTag()) {
    case JSRegExp::ATOM:
      return AtomExec(regexp, subject, index, last_match_info);
    case JSRegExp::IRREGEXP: {
      Handle<Object> result =
          IrregexpExec(regexp, subject, index, last_match_info);
      ASSERT(!result.is_null() ||
             regexp->GetIsolate()->has_pending_exception());
      return result;
    }
    default:
      UNREACHABLE();
      return Handle<Object>::null();
  }
}


// RegExp Atom implementation: Simple string search using indexOf.


void RegExpImpl::AtomCompile(Handle<JSRegExp> re,
                             Handle<String> pattern,
                             JSRegExp::Flags flags,
                             Handle<String> match_pattern) {
  re->GetIsolate()->factory()->SetRegExpAtomData(re,
                                                 JSRegExp::ATOM,
                                                 pattern,
                                                 flags,
                                                 match_pattern);
}


static void SetAtomLastCapture(FixedArray* array,
                               String* subject,
                               int from,
                               int to) {
  NoHandleAllocation no_handles;
  RegExpImpl::SetLastCaptureCount(array, 2);
  RegExpImpl::SetLastSubject(array, subject);
  RegExpImpl::SetLastInput(array, subject);
  RegExpImpl::SetCapture(array, 0, from);
  RegExpImpl::SetCapture(array, 1, to);
}


Handle<Object> RegExpImpl::AtomExec(Handle<JSRegExp> re,
                                    Handle<String> subject,
                                    int index,
                                    Handle<JSArray> last_match_info) {
  Isolate* isolate = re->GetIsolate();

  ASSERT(0 <= index);
  ASSERT(index <= subject->length());

  if (!subject->IsFlat()) FlattenString(subject);
  AssertNoAllocation no_heap_allocation;  // ensure vectors stay valid

  String* needle = String::cast(re->DataAt(JSRegExp::kAtomPatternIndex));
  int needle_len = needle->length();
  ASSERT(needle->IsFlat());

  if (needle_len != 0) {
    if (index + needle_len > subject->length()) {
      return isolate->factory()->null_value();
    }

    String::FlatContent needle_content = needle->GetFlatContent();
    String::FlatContent subject_content = subject->GetFlatContent();
    ASSERT(needle_content.IsFlat());
    ASSERT(subject_content.IsFlat());
    // dispatch on type of strings
    index = (needle_content.IsAscii()
             ? (subject_content.IsAscii()
                ? SearchString(isolate,
                               subject_content.ToAsciiVector(),
                               needle_content.ToAsciiVector(),
                               index)
                : SearchString(isolate,
                               subject_content.ToUC16Vector(),
                               needle_content.ToAsciiVector(),
                               index))
             : (subject_content.IsAscii()
                ? SearchString(isolate,
                               subject_content.ToAsciiVector(),
                               needle_content.ToUC16Vector(),
                               index)
                : SearchString(isolate,
                               subject_content.ToUC16Vector(),
                               needle_content.ToUC16Vector(),
                               index)));
    if (index == -1) return isolate->factory()->null_value();
  }
  ASSERT(last_match_info->HasFastElements());

  {
    NoHandleAllocation no_handles;
    FixedArray* array = FixedArray::cast(last_match_info->elements());
    SetAtomLastCapture(array, *subject, index, index + needle_len);
  }
  return last_match_info;
}


// Irregexp implementation.

// Ensures that the regexp object contains a compiled version of the
// source for either ASCII or non-ASCII strings.
// If the compiled version doesn't already exist, it is compiled
// from the source pattern.
// If compilation fails, an exception is thrown and this function
// returns false.
bool RegExpImpl::EnsureCompiledIrregexp(Handle<JSRegExp> re, bool is_ascii) {
  Object* compiled_code = re->DataAt(JSRegExp::code_index(is_ascii));
#ifdef V8_INTERPRETED_REGEXP
  if (compiled_code->IsByteArray()) return true;
#else  // V8_INTERPRETED_REGEXP (RegExp native code)
  if (compiled_code->IsCode()) return true;
#endif
  // We could potentially have marked this as flushable, but have kept
  // a saved version if we did not flush it yet.
  Object* saved_code = re->DataAt(JSRegExp::saved_code_index(is_ascii));
  if (saved_code->IsCode()) {
    // Reinstate the code in the original place.
    re->SetDataAt(JSRegExp::code_index(is_ascii), saved_code);
    ASSERT(compiled_code->IsSmi());
    return true;
  }
  return CompileIrregexp(re, is_ascii);
}


static bool CreateRegExpErrorObjectAndThrow(Handle<JSRegExp> re,
                                            bool is_ascii,
                                            Handle<String> error_message,
                                            Isolate* isolate) {
  Factory* factory = isolate->factory();
  Handle<FixedArray> elements = factory->NewFixedArray(2);
  elements->set(0, re->Pattern());
  elements->set(1, *error_message);
  Handle<JSArray> array = factory->NewJSArrayWithElements(elements);
  Handle<Object> regexp_err =
      factory->NewSyntaxError("malformed_regexp", array);
  isolate->Throw(*regexp_err);
  return false;
}


bool RegExpImpl::CompileIrregexp(Handle<JSRegExp> re, bool is_ascii) {
  // Compile the RegExp.
  Isolate* isolate = re->GetIsolate();
  ZoneScope zone_scope(isolate, DELETE_ON_EXIT);
  PostponeInterruptsScope postpone(isolate);
  // If we had a compilation error the last time this is saved at the
  // saved code index.
  Object* entry = re->DataAt(JSRegExp::code_index(is_ascii));
  // When arriving here entry can only be a smi, either representing an
  // uncompiled regexp, a previous compilation error, or code that has
  // been flushed.
  ASSERT(entry->IsSmi());
  int entry_value = Smi::cast(entry)->value();
  ASSERT(entry_value == JSRegExp::kUninitializedValue ||
         entry_value == JSRegExp::kCompilationErrorValue ||
         (entry_value < JSRegExp::kCodeAgeMask && entry_value >= 0));

  if (entry_value == JSRegExp::kCompilationErrorValue) {
    // A previous compilation failed and threw an error which we store in
    // the saved code index (we store the error message, not the actual
    // error). Recreate the error object and throw it.
    Object* error_string = re->DataAt(JSRegExp::saved_code_index(is_ascii));
    ASSERT(error_string->IsString());
    Handle<String> error_message(String::cast(error_string));
    CreateRegExpErrorObjectAndThrow(re, is_ascii, error_message, isolate);
    return false;
  }

  JSRegExp::Flags flags = re->GetFlags();

  Handle<String> pattern(re->Pattern());
  if (!pattern->IsFlat()) FlattenString(pattern);
  RegExpCompileData compile_data;
  FlatStringReader reader(isolate, pattern);
  if (!RegExpParser::ParseRegExp(&reader, flags.is_multiline(),
                                 &compile_data)) {
    // Throw an exception if we fail to parse the pattern.
    // THIS SHOULD NOT HAPPEN. We already pre-parsed it successfully once.
    ThrowRegExpException(re,
                         pattern,
                         compile_data.error,
                         "malformed_regexp");
    return false;
  }
  RegExpEngine::CompilationResult result =
      RegExpEngine::Compile(&compile_data,
                            flags.is_ignore_case(),
                            flags.is_multiline(),
                            pattern,
                            is_ascii);
  if (result.error_message != NULL) {
    // Unable to compile regexp.
    Handle<String> error_message =
        isolate->factory()->NewStringFromUtf8(CStrVector(result.error_message));
    CreateRegExpErrorObjectAndThrow(re, is_ascii, error_message, isolate);
    return false;
  }

  Handle<FixedArray> data = Handle<FixedArray>(FixedArray::cast(re->data()));
  data->set(JSRegExp::code_index(is_ascii), result.code);
  int register_max = IrregexpMaxRegisterCount(*data);
  if (result.num_registers > register_max) {
    SetIrregexpMaxRegisterCount(*data, result.num_registers);
  }

  return true;
}


int RegExpImpl::IrregexpMaxRegisterCount(FixedArray* re) {
  return Smi::cast(
      re->get(JSRegExp::kIrregexpMaxRegisterCountIndex))->value();
}


void RegExpImpl::SetIrregexpMaxRegisterCount(FixedArray* re, int value) {
  re->set(JSRegExp::kIrregexpMaxRegisterCountIndex, Smi::FromInt(value));
}


int RegExpImpl::IrregexpNumberOfCaptures(FixedArray* re) {
  return Smi::cast(re->get(JSRegExp::kIrregexpCaptureCountIndex))->value();
}


int RegExpImpl::IrregexpNumberOfRegisters(FixedArray* re) {
  return Smi::cast(re->get(JSRegExp::kIrregexpMaxRegisterCountIndex))->value();
}


ByteArray* RegExpImpl::IrregexpByteCode(FixedArray* re, bool is_ascii) {
  return ByteArray::cast(re->get(JSRegExp::code_index(is_ascii)));
}


Code* RegExpImpl::IrregexpNativeCode(FixedArray* re, bool is_ascii) {
  return Code::cast(re->get(JSRegExp::code_index(is_ascii)));
}


void RegExpImpl::IrregexpInitialize(Handle<JSRegExp> re,
                                    Handle<String> pattern,
                                    JSRegExp::Flags flags,
                                    int capture_count) {
  // Initialize compiled code entries to null.
  re->GetIsolate()->factory()->SetRegExpIrregexpData(re,
                                                     JSRegExp::IRREGEXP,
                                                     pattern,
                                                     flags,
                                                     capture_count);
}


int RegExpImpl::IrregexpPrepare(Handle<JSRegExp> regexp,
                                Handle<String> subject) {
  if (!subject->IsFlat()) FlattenString(subject);

  // Check the asciiness of the underlying storage.
  bool is_ascii = subject->IsAsciiRepresentationUnderneath();
  if (!EnsureCompiledIrregexp(regexp, is_ascii)) return -1;

#ifdef V8_INTERPRETED_REGEXP
  // Byte-code regexp needs space allocated for all its registers.
  return IrregexpNumberOfRegisters(FixedArray::cast(regexp->data()));
#else  // V8_INTERPRETED_REGEXP
  // Native regexp only needs room to output captures. Registers are handled
  // internally.
  return (IrregexpNumberOfCaptures(FixedArray::cast(regexp->data())) + 1) * 2;
#endif  // V8_INTERPRETED_REGEXP
}


RegExpImpl::IrregexpResult RegExpImpl::IrregexpExecOnce(
    Handle<JSRegExp> regexp,
    Handle<String> subject,
    int index,
    Vector<int> output) {
  Isolate* isolate = regexp->GetIsolate();

  Handle<FixedArray> irregexp(FixedArray::cast(regexp->data()), isolate);

  ASSERT(index >= 0);
  ASSERT(index <= subject->length());
  ASSERT(subject->IsFlat());

  bool is_ascii = subject->IsAsciiRepresentationUnderneath();

#ifndef V8_INTERPRETED_REGEXP
  ASSERT(output.length() >= (IrregexpNumberOfCaptures(*irregexp) + 1) * 2);
  do {
    EnsureCompiledIrregexp(regexp, is_ascii);
    Handle<Code> code(IrregexpNativeCode(*irregexp, is_ascii), isolate);
    NativeRegExpMacroAssembler::Result res =
        NativeRegExpMacroAssembler::Match(code,
                                          subject,
                                          output.start(),
                                          output.length(),
                                          index,
                                          isolate);
    if (res != NativeRegExpMacroAssembler::RETRY) {
      ASSERT(res != NativeRegExpMacroAssembler::EXCEPTION ||
             isolate->has_pending_exception());
      STATIC_ASSERT(
          static_cast<int>(NativeRegExpMacroAssembler::SUCCESS) == RE_SUCCESS);
      STATIC_ASSERT(
          static_cast<int>(NativeRegExpMacroAssembler::FAILURE) == RE_FAILURE);
      STATIC_ASSERT(static_cast<int>(NativeRegExpMacroAssembler::EXCEPTION)
                    == RE_EXCEPTION);
      return static_cast<IrregexpResult>(res);
    }
    // If result is RETRY, the string has changed representation, and we
    // must restart from scratch.
    // In this case, it means we must make sure we are prepared to handle
    // the, potentially, different subject (the string can switch between
    // being internal and external, and even between being ASCII and UC16,
    // but the characters are always the same).
    IrregexpPrepare(regexp, subject);
    is_ascii = subject->IsAsciiRepresentationUnderneath();
  } while (true);
  UNREACHABLE();
  return RE_EXCEPTION;
#else  // V8_INTERPRETED_REGEXP

  ASSERT(output.length() >= IrregexpNumberOfRegisters(*irregexp));
  // We must have done EnsureCompiledIrregexp, so we can get the number of
  // registers.
  int* register_vector = output.start();
  int number_of_capture_registers =
      (IrregexpNumberOfCaptures(*irregexp) + 1) * 2;
  for (int i = number_of_capture_registers - 1; i >= 0; i--) {
    register_vector[i] = -1;
  }
  Handle<ByteArray> byte_codes(IrregexpByteCode(*irregexp, is_ascii), isolate);

  IrregexpResult result = IrregexpInterpreter::Match(isolate,
                                                     byte_codes,
                                                     subject,
                                                     register_vector,
                                                     index);
  if (result == RE_EXCEPTION) {
    ASSERT(!isolate->has_pending_exception());
    isolate->StackOverflow();
  }
  return result;
#endif  // V8_INTERPRETED_REGEXP
}


Handle<Object> RegExpImpl::IrregexpExec(Handle<JSRegExp> jsregexp,
                                        Handle<String> subject,
                                        int previous_index,
                                        Handle<JSArray> last_match_info) {
  Isolate* isolate = jsregexp->GetIsolate();
  ASSERT_EQ(jsregexp->TypeTag(), JSRegExp::IRREGEXP);

  // Prepare space for the return values.
#ifdef V8_INTERPRETED_REGEXP
#ifdef DEBUG
  if (FLAG_trace_regexp_bytecodes) {
    String* pattern = jsregexp->Pattern();
    PrintF("\n\nRegexp match:   /%s/\n\n", *(pattern->ToCString()));
    PrintF("\n\nSubject string: '%s'\n\n", *(subject->ToCString()));
  }
#endif
#endif
  int required_registers = RegExpImpl::IrregexpPrepare(jsregexp, subject);
  if (required_registers < 0) {
    // Compiling failed with an exception.
    ASSERT(isolate->has_pending_exception());
    return Handle<Object>::null();
  }

  OffsetsVector registers(required_registers, isolate);

  IrregexpResult res = RegExpImpl::IrregexpExecOnce(
      jsregexp, subject, previous_index, Vector<int>(registers.vector(),
                                                     registers.length()));
  if (res == RE_SUCCESS) {
    int capture_register_count =
        (IrregexpNumberOfCaptures(FixedArray::cast(jsregexp->data())) + 1) * 2;
    last_match_info->EnsureSize(capture_register_count + kLastMatchOverhead);
    AssertNoAllocation no_gc;
    int* register_vector = registers.vector();
    FixedArray* array = FixedArray::cast(last_match_info->elements());
    for (int i = 0; i < capture_register_count; i += 2) {
      SetCapture(array, i, register_vector[i]);
      SetCapture(array, i + 1, register_vector[i + 1]);
    }
    SetLastCaptureCount(array, capture_register_count);
    SetLastSubject(array, *subject);
    SetLastInput(array, *subject);
    return last_match_info;
  }
  if (res == RE_EXCEPTION) {
    ASSERT(isolate->has_pending_exception());
    return Handle<Object>::null();
  }
  ASSERT(res == RE_FAILURE);
  return isolate->factory()->null_value();
}


// -------------------------------------------------------------------
// Implementation of the Irregexp regular expression engine.
//
// The Irregexp regular expression engine is intended to be a complete
// implementation of ECMAScript regular expressions.  It generates either
// bytecodes or native code.

//   The Irregexp regexp engine is structured in three steps.
//   1) The parser generates an abstract syntax tree.  See ast.cc.
//   2) From the AST a node network is created.  The nodes are all
//      subclasses of RegExpNode.  The nodes represent states when
//      executing a regular expression.  Several optimizations are
//      performed on the node network.
//   3) From the nodes we generate either byte codes or native code
//      that can actually execute the regular expression (perform
//      the search).  The code generation step is described in more
//      detail below.

// Code generation.
//
//   The nodes are divided into four main categories.
//   * Choice nodes
//        These represent places where the regular expression can
//        match in more than one way.  For example on entry to an
//        alternation (foo|bar) or a repetition (*, +, ? or {}).
//   * Action nodes
//        These represent places where some action should be
//        performed.  Examples include recording the current position
//        in the input string to a register (in order to implement
//        captures) or other actions on register for example in order
//        to implement the counters needed for {} repetitions.
//   * Matching nodes
//        These attempt to match some element part of the input string.
//        Examples of elements include character classes, plain strings
//        or back references.
//   * End nodes
//        These are used to implement the actions required on finding
//        a successful match or failing to find a match.
//
//   The code generated (whether as byte codes or native code) maintains
//   some state as it runs.  This consists of the following elements:
//
//   * The capture registers.  Used for string captures.
//   * Other registers.  Used for counters etc.
//   * The current position.
//   * The stack of backtracking information.  Used when a matching node
//     fails to find a match and needs to try an alternative.
//
// Conceptual regular expression execution model:
//
//   There is a simple conceptual model of regular expression execution
//   which will be presented first.  The actual code generated is a more
//   efficient simulation of the simple conceptual model:
//
//   * Choice nodes are implemented as follows:
//     For each choice except the last {
//       push current position
//       push backtrack code location
//       <generate code to test for choice>
//       backtrack code location:
//       pop current position
//     }
//     <generate code to test for last choice>
//
//   * Actions nodes are generated as follows
//     <push affected registers on backtrack stack>
//     <generate code to perform action>
//     push backtrack code location
//     <generate code to test for following nodes>
//     backtrack code location:
//     <pop affected registers to restore their state>
//     <pop backtrack location from stack and go to it>
//
//   * Matching nodes are generated as follows:
//     if input string matches at current position
//       update current position
//       <generate code to test for following nodes>
//     else
//       <pop backtrack location from stack and go to it>
//
//   Thus it can be seen that the current position is saved and restored
//   by the choice nodes, whereas the registers are saved and restored by
//   by the action nodes that manipulate them.
//
//   The other interesting aspect of this model is that nodes are generated
//   at the point where they are needed by a recursive call to Emit().  If
//   the node has already been code generated then the Emit() call will
//   generate a jump to the previously generated code instead.  In order to
//   limit recursion it is possible for the Emit() function to put the node
//   on a work list for later generation and instead generate a jump.  The
//   destination of the jump is resolved later when the code is generated.
//
// Actual regular expression code generation.
//
//   Code generation is actually more complicated than the above.  In order
//   to improve the efficiency of the generated code some optimizations are
//   performed
//
//   * Choice nodes have 1-character lookahead.
//     A choice node looks at the following character and eliminates some of
//     the choices immediately based on that character.  This is not yet
//     implemented.
//   * Simple greedy loops store reduced backtracking information.
//     A quantifier like /.*foo/m will greedily match the whole input.  It will
//     then need to backtrack to a point where it can match "foo".  The naive
//     implementation of this would push each character position onto the
//     backtracking stack, then pop them off one by one.  This would use space
//     proportional to the length of the input string.  However since the "."
//     can only match in one way and always has a constant length (in this case
//     of 1) it suffices to store the current position on the top of the stack
//     once.  Matching now becomes merely incrementing the current position and
//     backtracking becomes decrementing the current position and checking the
//     result against the stored current position.  This is faster and saves
//     space.
//   * The current state is virtualized.
//     This is used to defer expensive operations until it is clear that they
//     are needed and to generate code for a node more than once, allowing
//     specialized an efficient versions of the code to be created. This is
//     explained in the section below.
//
// Execution state virtualization.
//
//   Instead of emitting code, nodes that manipulate the state can record their
//   manipulation in an object called the Trace.  The Trace object can record a
//   current position offset, an optional backtrack code location on the top of
//   the virtualized backtrack stack and some register changes.  When a node is
//   to be emitted it can flush the Trace or update it.  Flushing the Trace
//   will emit code to bring the actual state into line with the virtual state.
//   Avoiding flushing the state can postpone some work (e.g. updates of capture
//   registers).  Postponing work can save time when executing the regular
//   expression since it may be found that the work never has to be done as a
//   failure to match can occur.  In addition it is much faster to jump to a
//   known backtrack code location than it is to pop an unknown backtrack
//   location from the stack and jump there.
//
//   The virtual state found in the Trace affects code generation.  For example
//   the virtual state contains the difference between the actual current
//   position and the virtual current position, and matching code needs to use
//   this offset to attempt a match in the correct location of the input
//   string.  Therefore code generated for a non-trivial trace is specialized
//   to that trace.  The code generator therefore has the ability to generate
//   code for each node several times.  In order to limit the size of the
//   generated code there is an arbitrary limit on how many specialized sets of
//   code may be generated for a given node.  If the limit is reached, the
//   trace is flushed and a generic version of the code for a node is emitted.
//   This is subsequently used for that node.  The code emitted for non-generic
//   trace is not recorded in the node and so it cannot currently be reused in
//   the event that code generation is requested for an identical trace.


void RegExpTree::AppendToText(RegExpText* text) {
  UNREACHABLE();
}


void RegExpAtom::AppendToText(RegExpText* text) {
  text->AddElement(TextElement::Atom(this));
}


void RegExpCharacterClass::AppendToText(RegExpText* text) {
  text->AddElement(TextElement::CharClass(this));
}


void RegExpText::AppendToText(RegExpText* text) {
  for (int i = 0; i < elements()->length(); i++)
    text->AddElement(elements()->at(i));
}


TextElement TextElement::Atom(RegExpAtom* atom) {
  TextElement result = TextElement(ATOM);
  result.data.u_atom = atom;
  return result;
}


TextElement TextElement::CharClass(
      RegExpCharacterClass* char_class) {
  TextElement result = TextElement(CHAR_CLASS);
  result.data.u_char_class = char_class;
  return result;
}


int TextElement::length() {
  if (type == ATOM) {
    return data.u_atom->length();
  } else {
    ASSERT(type == CHAR_CLASS);
    return 1;
  }
}


DispatchTable* ChoiceNode::GetTable(bool ignore_case) {
  if (table_ == NULL) {
    table_ = new DispatchTable();
    DispatchTableConstructor cons(table_, ignore_case);
    cons.BuildTable(this);
  }
  return table_;
}


class RegExpCompiler {
 public:
  RegExpCompiler(int capture_count, bool ignore_case, bool is_ascii);

  int AllocateRegister() {
    if (next_register_ >= RegExpMacroAssembler::kMaxRegister) {
      reg_exp_too_big_ = true;
      return next_register_;
    }
    return next_register_++;
  }

  RegExpEngine::CompilationResult Assemble(RegExpMacroAssembler* assembler,
                                           RegExpNode* start,
                                           int capture_count,
                                           Handle<String> pattern);

  inline void AddWork(RegExpNode* node) { work_list_->Add(node); }

  static const int kImplementationOffset = 0;
  static const int kNumberOfRegistersOffset = 0;
  static const int kCodeOffset = 1;

  RegExpMacroAssembler* macro_assembler() { return macro_assembler_; }
  EndNode* accept() { return accept_; }

  static const int kMaxRecursion = 100;
  inline int 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_; }

  int current_expansion_factor() { return current_expansion_factor_; }
  void set_current_expansion_factor(int value) {
    current_expansion_factor_ = value;
  }

  static const int kNoRegister = -1;

 private:
  EndNode* accept_;
  int next_register_;
  List<RegExpNode*>* work_list_;
  int recursion_depth_;
  RegExpMacroAssembler* macro_assembler_;
  bool ignore_case_;
  bool ascii_;
  bool reg_exp_too_big_;
  int current_expansion_factor_;
};


class RecursionCheck {
 public:
  explicit RecursionCheck(RegExpCompiler* compiler) : compiler_(compiler) {
    compiler->IncrementRecursionDepth();
  }
  ~RecursionCheck() { compiler_->DecrementRecursionDepth(); }
 private:
  RegExpCompiler* compiler_;
};


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(int capture_count, bool ignore_case, bool ascii)
    : next_register_(2 * (capture_count + 1)),
      work_list_(NULL),
      recursion_depth_(0),
      ignore_case_(ignore_case),
      ascii_(ascii),
      reg_exp_too_big_(false),
      current_expansion_factor_(1) {
  accept_ = new EndNode(EndNode::ACCEPT);
  ASSERT(next_register_ - 1 <= RegExpMacroAssembler::kMaxRegister);
}


RegExpEngine::CompilationResult RegExpCompiler::Assemble(
    RegExpMacroAssembler* macro_assembler,
    RegExpNode* start,
    int capture_count,
    Handle<String> pattern) {
  Heap* heap = pattern->GetHeap();

  bool use_slow_safe_regexp_compiler = false;
  if (heap->total_regexp_code_generated() >
          RegExpImpl::kRegWxpCompiledLimit &&
      heap->isolate()->memory_allocator()->SizeExecutable() >
          RegExpImpl::kRegExpExecutableMemoryLimit) {
    use_slow_safe_regexp_compiler = true;
  }

  macro_assembler->set_slow_safe(use_slow_safe_regexp_compiler);

#ifdef DEBUG
  if (FLAG_trace_regexp_assembler)
    macro_assembler_ = new RegExpMacroAssemblerTracer(macro_assembler);
  else
#endif
    macro_assembler_ = macro_assembler;

  List <RegExpNode*> work_list(0);
  work_list_ = &work_list;
  Label fail;
  macro_assembler_->PushBacktrack(&fail);
  Trace new_trace;
  start->Emit(this, &new_trace);
  macro_assembler_->Bind(&fail);
  macro_assembler_->Fail();
  while (!work_list.is_empty()) {
    work_list.RemoveLast()->Emit(this, &new_trace);
  }
  if (reg_exp_too_big_) return IrregexpRegExpTooBig();

  Handle<HeapObject> code = macro_assembler_->GetCode(pattern);
  heap->IncreaseTotalRegexpCodeGenerated(code->Size());
  work_list_ = NULL;
#ifdef DEBUG
  if (FLAG_print_code) {
    Handle<Code>::cast(code)->Disassemble(*pattern->ToCString());
  }
  if (FLAG_trace_regexp_assembler) {
    delete macro_assembler_;
  }
#endif
  return RegExpEngine::CompilationResult(*code, next_register_);
}


bool Trace::DeferredAction::Mentions(int that) {
  if (type() == ActionNode::CLEAR_CAPTURES) {
    Interval range = static_cast<DeferredClearCaptures*>(this)->range();
    return range.Contains(that);
  } else {
    return reg() == that;
  }
}


bool Trace::mentions_reg(int reg) {
  for (DeferredAction* action = actions_;
       action != NULL;
       action = action->next()) {
    if (action->Mentions(reg))
      return true;
  }
  return false;
}


bool Trace::GetStoredPosition(int reg, int* cp_offset) {
  ASSERT_EQ(0, *cp_offset);
  for (DeferredAction* action = actions_;
       action != NULL;
       action = action->next()) {
    if (action->Mentions(reg)) {
      if (action->type() == ActionNode::STORE_POSITION) {
        *cp_offset = static_cast<DeferredCapture*>(action)->cp_offset();
        return true;
      } else {
        return false;
      }
    }
  }
  return false;
}


int Trace::FindAffectedRegisters(OutSet* affected_registers) {
  int max_register = RegExpCompiler::kNoRegister;
  for (DeferredAction* action = actions_;
       action != NULL;
       action = action->next()) {
    if (action->type() == ActionNode::CLEAR_CAPTURES) {
      Interval range = static_cast<DeferredClearCaptures*>(action)->range();
      for (int i = range.from(); i <= range.to(); i++)
        affected_registers->Set(i);
      if (range.to() > max_register) max_register = range.to();
    } else {
      affected_registers->Set(action->reg());
      if (action->reg() > max_register) max_register = action->reg();
    }
  }
  return max_register;
}


void Trace::RestoreAffectedRegisters(RegExpMacroAssembler* assembler,
                                     int max_register,
                                     OutSet& registers_to_pop,
                                     OutSet& registers_to_clear) {
  for (int reg = max_register; reg >= 0; reg--) {
    if (registers_to_pop.Get(reg)) assembler->PopRegister(reg);
    else if (registers_to_clear.Get(reg)) {
      int clear_to = reg;
      while (reg > 0 && registers_to_clear.Get(reg - 1)) {
        reg--;
      }
      assembler->ClearRegisters(reg, clear_to);
    }
  }
}


void Trace::PerformDeferredActions(RegExpMacroAssembler* assembler,
                                   int max_register,
                                   OutSet& affected_registers,
                                   OutSet* registers_to_pop,
                                   OutSet* registers_to_clear) {
  // The "+1" is to avoid a push_limit of zero if stack_limit_slack() is 1.
  const int push_limit = (assembler->stack_limit_slack() + 1) / 2;

  // Count pushes performed to force a stack limit check occasionally.
  int pushes = 0;

  for (int 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;

    int value = 0;
    bool absolute = false;
    bool clear = false;
    int 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->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_EQ(store_position, -1);
            ASSERT(!clear);
            break;
          }
          case ActionNode::INCREMENT_REGISTER:
            if (!absolute) {
              value++;
            }
            ASSERT_EQ(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_EQ(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_EQ(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);
    } else if (undo_action == CLEAR) {
      registers_to_clear->Set(reg);
    }
    // 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);
    }
  }
}


// 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.
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();
  }

  int max_register = FindAffectedRegisters(&affected_registers);
  OutSet registers_to_pop;
  OutSet registers_to_clear;
  PerformDeferredActions(assembler,
                         max_register,
                         affected_registers,
                         &registers_to_pop,
                         &registers_to_clear);
  if (cp_offset_ != 0) {
    assembler->AdvanceCurrentPosition(cp_offset_);
  }

  // Create a new trivial state and generate the node with that.
  Label undo;
  assembler->PushBacktrack(&undo);
  Trace new_state;
  successor->Emit(compiler, &new_state);

  // On backtrack we need to restore state.
  assembler->Bind(&undo);
  RestoreAffectedRegisters(assembler,
                           max_register,
                           registers_to_pop,
                           registers_to_clear);
  if (backtrack() == NULL) {
    assembler->Backtrack();
  } else {
    assembler->PopCurrentPosition();
    assembler->GoTo(backtrack());
  }
}


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()->is_bound()) {
    // We are completely independent of the trace, since we ignore it,
    // so this code can be used as the generic version.
    assembler->Bind(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.
    int 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();
}


void EndNode::Emit(RegExpCompiler* compiler, Trace* trace) {
  if (!trace->is_trivial()) {
    trace->Flush(compiler, this);
    return;
  }
  RegExpMacroAssembler* assembler = compiler->macro_assembler();
  if (!label()->is_bound()) {
    assembler->Bind(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();
}


void GuardedAlternative::AddGuard(Guard* guard) {
  if (guards_ == NULL)
    guards_ = new ZoneList<Guard*>(1);
  guards_->Add(guard);
}


ActionNode* ActionNode::SetRegister(int reg,
                                    int val,
                                    RegExpNode* on_success) {
  ActionNode* result = new ActionNode(SET_REGISTER, on_success);
  result->data_.u_store_register.reg = reg;
  result->data_.u_store_register.value = val;
  return result;
}


ActionNode* ActionNode::IncrementRegister(int reg, RegExpNode* on_success) {
  ActionNode* result = new ActionNode(INCREMENT_REGISTER, on_success);
  result->data_.u_increment_register.reg = reg;
  return result;
}


ActionNode* ActionNode::StorePosition(int reg,
                                      bool is_capture,
                                      RegExpNode* on_success) {
  ActionNode* result = new ActionNode(STORE_POSITION, on_success);
  result->data_.u_position_register.reg = reg;
  result->data_.u_position_register.is_capture = is_capture;
  return result;
}


ActionNode* ActionNode::ClearCaptures(Interval range,
                                      RegExpNode* on_success) {
  ActionNode* result = new ActionNode(CLEAR_CAPTURES, on_success);
  result->data_.u_clear_captures.range_from = range.from();
  result->data_.u_clear_captures.range_to = range.to();
  return result;
}


ActionNode* ActionNode::BeginSubmatch(int stack_reg,
                                      int position_reg,
                                      RegExpNode* on_success) {
  ActionNode* result = new ActionNode(BEGIN_SUBMATCH, on_success);
  result->data_.u_submatch.stack_pointer_register = stack_reg;
  result->data_.u_submatch.current_position_register = position_reg;
  return result;
}


ActionNode* ActionNode::PositiveSubmatchSuccess(int stack_reg,
                                                int position_reg,
                                                int clear_register_count,
                                                int clear_register_from,
                                                RegExpNode* on_success) {
  ActionNode* result = new ActionNode(POSITIVE_SUBMATCH_SUCCESS, on_success);
  result->data_.u_submatch.stack_pointer_register = stack_reg;
  result->data_.u_submatch.current_position_register = position_reg;
  result->data_.u_submatch.clear_register_count = clear_register_count;
  result->data_.u_submatch.clear_register_from = clear_register_from;
  return result;
}


ActionNode* ActionNode::EmptyMatchCheck(int start_register,
                                        int repetition_register,
                                        int repetition_limit,
                                        RegExpNode* on_success) {
  ActionNode* result = new ActionNode(EMPTY_MATCH_CHECK, on_success);
  result->data_.u_empty_match_check.start_register = start_register;
  result->data_.u_empty_match_check.repetition_register = repetition_register;
  result->data_.u_empty_match_check.repetition_limit = repetition_limit;
  return result;
}


#define DEFINE_ACCEPT(Type)                                          \
  void Type##Node::Accept(NodeVisitor* visitor) {                    \
    visitor->Visit##Type(this);                                      \
  }
FOR_EACH_NODE_TYPE(DEFINE_ACCEPT)
#undef DEFINE_ACCEPT


void LoopChoiceNode::Accept(NodeVisitor* visitor) {
  visitor->VisitLoopChoice(this);
}


// -------------------------------------------------------------------
// Emit code.


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;
  }
}


// Returns the number of characters in the equivalence class, omitting those
// that cannot occur in the source string because it is ASCII.
static int GetCaseIndependentLetters(Isolate* isolate,
                                     uc16 character,
                                     bool ascii_subject,
                                     unibrow::uchar* letters) {
  int length =
      isolate->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 <= String::kMaxAsciiCharCode) {
    return length;
  }
  // The standard requires that non-ASCII characters cannot have ASCII
  // character codes in their equivalence class.
  return 0;
}


static inline bool EmitSimpleCharacter(Isolate* isolate,
                                       RegExpCompiler* compiler,
                                       uc16 c,
                                       Label* on_failure,
                                       int 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;
}


// Only emits non-letters (things that don't have case).  Only used for case
// independent matches.
static inline bool EmitAtomNonLetter(Isolate* isolate,
                                     RegExpCompiler* compiler,
                                     uc16 c,
                                     Label* on_failure,
                                     int cp_offset,
                                     bool check,
                                     bool preloaded) {
  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
  bool ascii = compiler->ascii();
  unibrow::uchar chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
  int length = GetCaseIndependentLetters(isolate, 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 > String::kMaxAsciiCharCodeU) {
      // 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 bool ShortCutEmitCharacterPair(RegExpMacroAssembler* macro_assembler,
                                      bool ascii,
                                      uc16 c1,
                                      uc16 c2,
                                      Label* on_failure) {
  uc16 char_mask;
  if (ascii) {
    char_mask = String::kMaxAsciiCharCode;
  } else {
    char_mask = String::kMaxUtf16CodeUnit;
  }
  uc16 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);
    uc16 mask = char_mask ^ exor;
    macro_assembler->CheckNotCharacterAfterAnd(c1, mask, on_failure);
    return true;
  }
  ASSERT(c2 > c1);
  uc16 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.
    uc16 mask = char_mask ^ diff;
    macro_assembler->CheckNotCharacterAfterMinusAnd(c1 - diff,
                                                    diff,
                                                    mask,
                                                    on_failure);
    return true;
  }
  return false;
}


typedef bool EmitCharacterFunction(Isolate* isolate,
                                   RegExpCompiler* compiler,
                                   uc16 c,
                                   Label* on_failure,
                                   int cp_offset,
                                   bool check,
                                   bool preloaded);

// Only emits letters (things that have case).  Only used for case independent
// matches.
static inline bool EmitAtomLetter(Isolate* isolate,
                                  RegExpCompiler* compiler,
                                  uc16 c,
                                  Label* on_failure,
                                  int cp_offset,
                                  bool check,
                                  bool preloaded) {
  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
  bool ascii = compiler->ascii();
  unibrow::uchar chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
  int length = GetCaseIndependentLetters(isolate, 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);
  }
  Label 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->Bind(&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->Bind(&ok);
      break;
    default:
      UNREACHABLE();
      break;
  }
  return true;
}


static void EmitCharClass(RegExpMacroAssembler* macro_assembler,
                          RegExpCharacterClass* cc,
                          bool ascii,
                          Label* on_failure,
                          int cp_offset,
                          bool check_offset,
                          bool preloaded) {
  ZoneList<CharacterRange>* ranges = cc->ranges();
  int max_char;
  if (ascii) {
    max_char = String::kMaxAsciiCharCode;
  } else {
    max_char = String::kMaxUtf16CodeUnit;
  }

  Label success;

  Label* char_is_in_class =
      cc->is_negated() ? on_failure : &success;

  int range_count = ranges->length();

  int last_valid_range = range_count - 1;
  while (last_valid_range >= 0) {
    CharacterRange& range = ranges->at(last_valid_range);
    if (range.from() <= max_char) {
      break;
    }
    last_valid_range--;
  }

  if (last_valid_range < 0) {
    if (!cc->is_negated()) {
      // TODO(plesner): We can remove this when the node level does our
      // ASCII optimizations for us.
      macro_assembler->GoTo(on_failure);
    }
    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;
  }

  for (int i = 0; i < last_valid_range; i++) {
    CharacterRange& range = ranges->at(i);
    Label next_range;
    uc16 from = range.from();
    uc16 to = range.to();
    if (from > max_char) {
      continue;
    }
    if (to > max_char) to = max_char;
    if (to == from) {
      macro_assembler->CheckCharacter(to, char_is_in_class);
    } else {
      if (from != 0) {
        macro_assembler->CheckCharacterLT(from, &next_range);
      }
      if (to != max_char) {
        macro_assembler->CheckCharacterLT(to + 1, char_is_in_class);
      } else {
        macro_assembler->GoTo(char_is_in_class);
      }
    }
    macro_assembler->Bind(&next_range);
  }

  CharacterRange& range = ranges->at(last_valid_range);
  uc16 from = range.from();
  uc16 to = range.to();

  if (to > max_char) to = max_char;
  ASSERT(to >= from);

  if (to == from) {
    if (cc->is_negated()) {
      macro_assembler->CheckCharacter(to, on_failure);
    } else {
      macro_assembler->CheckNotCharacter(to, on_failure);
    }
  } else {
    if (from != 0) {
      if (cc->is_negated()) {
        macro_assembler->CheckCharacterLT(from, &success);
      } else {
        macro_assembler->CheckCharacterLT(from, on_failure);
      }
    }
    if (to != String::kMaxUtf16CodeUnit) {
      if (cc->is_negated()) {
        macro_assembler->CheckCharacterLT(to + 1, on_failure);
      } else {
        macro_assembler->CheckCharacterGT(to, on_failure);
      }
    } else {
      if (cc->is_negated()) {
        macro_assembler->GoTo(on_failure);
      }
    }
  }
  macro_assembler->Bind(&success);
}


RegExpNode::~RegExpNode() {
}


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_.is_bound()) {
      // 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->Bind(&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 (FLAG_regexp_optimization &&
      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;
}


int ActionNode::EatsAtLeast(int still_to_find,
                            int recursion_depth,
                            bool not_at_start) {
  if (recursion_depth > RegExpCompiler::kMaxRecursion) return 0;
  if (type_ == POSITIVE_SUBMATCH_SUCCESS) return 0;  // Rewinds input!
  return on_success()->EatsAtLeast(still_to_find,
                                   recursion_depth + 1,
                                   not_at_start);
}


int AssertionNode::EatsAtLeast(int still_to_find,
                               int recursion_depth,
                               bool not_at_start) {
  if (recursion_depth > RegExpCompiler::kMaxRecursion) return 0;
  // If we know we are not at the start and we are asked "how many characters
  // will you match if you succeed?" then we can answer anything since false
  // implies false.  So lets just return the max answer (still_to_find) since
  // that won't prevent us from preloading a lot of characters for the other
  // branches in the node graph.
  if (type() == AT_START && not_at_start) return still_to_find;
  return on_success()->EatsAtLeast(still_to_find,
                                   recursion_depth + 1,
                                   not_at_start);
}


int BackReferenceNode::EatsAtLeast(int still_to_find,
                                   int recursion_depth,
                                   bool not_at_start) {
  if (recursion_depth > RegExpCompiler::kMaxRecursion) return 0;
  return on_success()->EatsAtLeast(still_to_find,
                                   recursion_depth + 1,
                                   not_at_start);
}


int TextNode::EatsAtLeast(int still_to_find,
                          int recursion_depth,
                          bool not_at_start) {
  int answer = Length();
  if (answer >= still_to_find) return answer;
  if (recursion_depth > RegExpCompiler::kMaxRecursion) 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,
                                            recursion_depth + 1,
                                            true);
}


int NegativeLookaheadChoiceNode::EatsAtLeast(int still_to_find,
                                             int recursion_depth,
                                             bool not_at_start) {
  if (recursion_depth > RegExpCompiler::kMaxRecursion) return 0;
  // Alternative 0 is the negative lookahead, alternative 1 is what comes
  // afterwards.
  RegExpNode* node = alternatives_->at(1).node();
  return node->EatsAtLeast(still_to_find, recursion_depth + 1, not_at_start);
}


void NegativeLookaheadChoiceNode::GetQuickCheckDetails(
    QuickCheckDetails* details,
    RegExpCompiler* compiler,
    int filled_in,
    bool not_at_start) {
  // Alternative 0 is the negative lookahead, alternative 1 is what comes
  // afterwards.
  RegExpNode* node = alternatives_->at(1).node();
  return node->GetQuickCheckDetails(details, compiler, filled_in, not_at_start);
}


int ChoiceNode::EatsAtLeastHelper(int still_to_find,
                                  int recursion_depth,
                                  RegExpNode* ignore_this_node,
                                  bool not_at_start) {
  if (recursion_depth > RegExpCompiler::kMaxRecursion) return 0;
  int min = 100;
  int choice_count = alternatives_->length();
  for (int i = 0; i < choice_count; i++) {
    RegExpNode* node = alternatives_->at(i).node();
    if (node == ignore_this_node) continue;
    int node_eats_at_least = node->EatsAtLeast(still_to_find,
                                               recursion_depth + 1,
                                               not_at_start);
    if (node_eats_at_least < min) min = node_eats_at_least;
  }
  return min;
}


int LoopChoiceNode::EatsAtLeast(int still_to_find,
                                int recursion_depth,
                                bool not_at_start) {
  return EatsAtLeastHelper(still_to_find,
                           recursion_depth,
                           loop_node_,
                           not_at_start);
}


int ChoiceNode::EatsAtLeast(int still_to_find,
                            int recursion_depth,
                            bool not_at_start) {
  return EatsAtLeastHelper(still_to_find,
                           recursion_depth,
                           NULL,
                           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;
}


bool QuickCheckDetails::Rationalize(bool asc) {
  bool found_useful_op = false;
  uint32_t char_mask;
  if (asc) {
    char_mask = String::kMaxAsciiCharCode;
  } else {
    char_mask = String::kMaxUtf16CodeUnit;
  }
  mask_ = 0;
  value_ = 0;
  int char_shift = 0;
  for (int i = 0; i < characters_; i++) {
    Position* pos = &positions_[i];
    if ((pos->mask & String::kMaxAsciiCharCode) != 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,
                                Label* 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);
  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 = String::kMaxAsciiCharCode;
    } else {
      char_mask = String::kMaxUtf16CodeUnit;
    }
    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 & 0x7f7f) == 0x7f7f) 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;
}


// 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,
                                    int characters_filled_in,
                                    bool not_at_start) {
  Isolate* isolate = Isolate::Current();
  ASSERT(characters_filled_in < details->characters());
  int characters = details->characters();
  int char_mask;
  if (compiler->ascii()) {
    char_mask = String::kMaxAsciiCharCode;
  } else {
    char_mask = String::kMaxUtf16CodeUnit;
  }
  for (int k = 0; k < elms_->length(); k++) {
    TextElement elm = elms_->at(k);
    if (elm.type == TextElement::ATOM) {
      Vector<const uc16> quarks = elm.data.u_atom->data();
      for (int i = 0; i < characters && i < quarks.length(); i++) {
        QuickCheckDetails::Position* pos =
            details->positions(characters_filled_in);
        uc16 c = quarks[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()) {
          unibrow::uchar chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
          int length = GetCaseIndependentLetters(isolate, 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 (int 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.data.u_char_class;
      ZoneList<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 {
        int 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);
        uc16 from = range.from();
        uc16 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 (int i = first_range + 1; i < ranges->length(); i++) {
          CharacterRange range = ranges->at(i);
          uc16 from = range.from();
          uc16 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());
  on_success()-> GetQuickCheckDetails(details,
                                      compiler,
                                      characters_filled_in,
                                      true);
}


void QuickCheckDetails::Clear() {
  for (int i = 0; i < characters_; i++) {
    positions_[i].mask = 0;
    positions_[i].value = 0;
    positions_[i].determines_perfectly = false;
  }
  characters_ = 0;
}


void QuickCheckDetails::Advance(int by, bool ascii) {
  ASSERT(by >= 0);
  if (by >= characters_) {
    Clear();
    return;
  }
  for (int i = 0; i < characters_ - by; i++) {
    positions_[i] = positions_[by + i];
  }
  for (int i = characters_ - by; i < characters_; i++) {
    positions_[i].mask = 0;
    positions_[i].value = 0;
    positions_[i].determines_perfectly = false;
  }
  characters_ -= by;
  // We could change mask_ and value_ here but we would never advance unless
  // they had already been used in a check and they won't be used again because
  // it would gain us nothing.  So there's no point.
}


void QuickCheckDetails::Merge(QuickCheckDetails* other, int from_index) {
  ASSERT(characters_ == other->characters_);
  if (other->cannot_match_) {
    return;
  }
  if (cannot_match_) {
    *this = *other;
    return;
  }
  for (int i = from_index; i < characters_; i++) {
    QuickCheckDetails::Position* pos = positions(i);
    QuickCheckDetails::Position* other_pos = other->positions(i);
    if (pos->mask != other_pos->mask ||
        pos->value != other_pos->value ||
        !other_pos->determines_perfectly) {
      // Our mask-compare operation will be approximate unless we have the
      // exact same operation on both sides of the alternation.
      pos->determines_perfectly = false;
    }
    pos->mask &= other_pos->mask;
    pos->value &= pos->mask;
    other_pos->value &= pos->mask;
    uc16 differing_bits = (pos->value ^ other_pos->value);
    pos->mask &= ~differing_bits;
    pos->value &= pos->mask;
  }
}


class VisitMarker {
 public:
  explicit VisitMarker(NodeInfo* info) : info_(info) {
    ASSERT(!info->visited);
    info->visited = true;
  }
  ~VisitMarker() {
    info_->visited = false;
  }
 private:
  NodeInfo* info_;
};


void LoopChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details,
                                          RegExpCompiler* compiler,
                                          int characters_filled_in,
                                          bool not_at_start) {
  if (body_can_be_zero_length_ || info()->visited) return;
  VisitMarker marker(info());
  return ChoiceNode::GetQuickCheckDetails(details,
                                          compiler,
                                          characters_filled_in,
                                          not_at_start);
}


void ChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details,
                                      RegExpCompiler* compiler,
                                      int characters_filled_in,
                                      bool not_at_start) {
  not_at_start = (not_at_start || not_at_start_);
  int choice_count = alternatives_->length();
  ASSERT(choice_count > 0);
  alternatives_->at(0).node()->GetQuickCheckDetails(details,
                                                    compiler,
                                                    characters_filled_in,
                                                    not_at_start);
  for (int i = 1; i < choice_count; i++) {
    QuickCheckDetails new_details(details->characters());
    RegExpNode* node = alternatives_->at(i).node();
    node->GetQuickCheckDetails(&new_details, compiler,
                               characters_filled_in,
                               not_at_start);
    // Here we merge the quick match details of the two branches.
    details->Merge(&new_details, characters_filled_in);
  }
}


// Check for [0-9A-Z_a-z].
static void EmitWordCheck(RegExpMacroAssembler* assembler,
                          Label* word,
                          Label* 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 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();

  Label 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->Bind(&ok);
  on_success->Emit(compiler, &new_trace);
}


// Emit the code to handle \b and \B (word-boundary or non-word-boundary)
// when we know whether the next character must be a word character or not.
static void EmitHalfBoundaryCheck(AssertionNode::AssertionNodeType type,
                                  RegExpCompiler* compiler,
                                  RegExpNode* on_success,
                                  Trace* trace) {
  RegExpMacroAssembler* assembler = compiler->macro_assembler();
  Label done;

  Trace new_trace(*trace);

  bool expect_word_character = (type == AssertionNode::AFTER_WORD_CHARACTER);
  Label* on_word = expect_word_character ? &done : new_trace.backtrack();
  Label* on_non_word = expect_word_character ? new_trace.backtrack() : &done;

  // Check whether previous character was a word character.
  switch (trace->at_start()) {
    case Trace::TRUE:
      if (expect_word_character) {
        assembler->GoTo(on_non_word);
      }
      break;
    case Trace::UNKNOWN:
      ASSERT_EQ(0, trace->cp_offset());
      assembler->CheckAtStart(on_non_word);
      // Fall through.
    case Trace::FALSE:
      int prev_char_offset = trace->cp_offset() - 1;
      assembler->LoadCurrentCharacter(prev_char_offset, NULL, false, 1);
      EmitWordCheck(assembler, on_word, on_non_word, expect_word_character);
      // We may or may not have loaded the previous character.
      new_trace.InvalidateCurrentCharacter();
  }

  assembler->Bind(&done);

  on_success->Emit(compiler, &new_trace);
}


// Emit the code to handle \b and \B (word-boundary or non-word-boundary).
static void EmitBoundaryCheck(AssertionNode::AssertionNodeType type,
                              RegExpCompiler* compiler,
                              RegExpNode* on_success,
                              Trace* trace) {
  RegExpMacroAssembler* assembler = compiler->macro_assembler();
  Label before_non_word;
  Label 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);

  // We will be loading the previous character into the current character
  // register.
  Trace new_trace(*trace);
  new_trace.InvalidateCurrentCharacter();

  Label ok;
  Label* boundary;
  Label* not_boundary;
  if (type == AssertionNode::AT_BOUNDARY) {
    boundary = &ok;
    not_boundary = new_trace.backtrack();
  } else {
    not_boundary = &ok;
    boundary = new_trace.backtrack();
  }

  // Next character is not a word character.
  assembler->Bind(&before_non_word);
  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(not_boundary);
  }
  // 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,
                                  &ok,  // Unused dummy label in this call.
                                  false);
  // Fall through on non-word.
  EmitWordCheck(assembler, boundary, not_boundary, false);
  assembler->GoTo(not_boundary);

  // Next character is a word character.
  assembler->Bind(&before_word);
  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(boundary);
  }
  // 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,
                                  &ok,  // Unused dummy label in this call.
                                  false);
  bool fall_through_on_word = (type == AssertionNode::AT_NON_BOUNDARY);
  EmitWordCheck(assembler, not_boundary, boundary, fall_through_on_word);

  assembler->Bind(&ok);

  on_success->Emit(compiler, &new_trace);
}


void AssertionNode::GetQuickCheckDetails(QuickCheckDetails* details,
                                         RegExpCompiler* compiler,
                                         int filled_in,
                                         bool not_at_start) {
  if (type_ == AT_START && not_at_start) {
    details->set_cannot_match();
    return;
  }
  return on_success()->GetQuickCheckDetails(details,
                                            compiler,
                                            filled_in,
                                            not_at_start);
}


void AssertionNode::Emit(RegExpCompiler* compiler, Trace* trace) {
  RegExpMacroAssembler* assembler = compiler->macro_assembler();
  switch (type_) {
    case AT_END: {
      Label ok;
      assembler->CheckPosition(trace->cp_offset(), &ok);
      assembler->GoTo(trace->backtrack());
      assembler->Bind(&ok);
      break;
    }
    case AT_START: {
      if (trace->at_start() == Trace::FALSE) {
        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(type_, compiler, on_success(), trace);
      return;
    }
    case AFTER_WORD_CHARACTER:
    case AFTER_NONWORD_CHARACTER: {
      EmitHalfBoundaryCheck(type_, compiler, on_success(), trace);
    }
  }
  on_success()->Emit(compiler, trace);
}


static bool DeterminedAlready(QuickCheckDetails* quick_check, int offset) {
  if (quick_check == NULL) return false;
  if (offset >= quick_check->characters()) return false;
  return quick_check->positions(offset)->determines_perfectly;
}


static void UpdateBoundsCheck(int index, int* checked_up_to) {
  if (index > *checked_up_to) {
    *checked_up_to = index;
  }
}


// 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
// bounds check most of the time.  In the example we only need to check for
// end-of-input when loading the putative 'r'.
//
// A slight complication involves the fact that the first character may already
// be fetched into a register by the previous node.  In this case we want to
// do the test for that character first.  We do this in separate passes.  The
// 'preloaded' argument indicates that we are doing such a 'pass'.  If such a
// pass has been performed then subsequent passes will have true in
// first_element_checked to indicate that that character does not need to be
// checked again.
//
// In addition to all this we are passed a Trace, which can
// contain an AlternativeGeneration object.  In this AlternativeGeneration
// object we can see details of any quick check that was already passed in
// order to get to the code we are now generating.  The quick check can involve
// 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,
                            int* checked_up_to) {
  Isolate* isolate = Isolate::Current();
  RegExpMacroAssembler* assembler = compiler->macro_assembler();
  bool ascii = compiler->ascii();
  Label* backtrack = trace->backtrack();
  QuickCheckDetails* quick_check = trace->quick_check_performed();
  int element_count = elms_->length();
  for (int i = preloaded ? 0 : element_count - 1; i >= 0; i--) {
    TextElement elm = elms_->at(i);
    int cp_offset = trace->cp_offset() + elm.cp_offset;
    if (elm.type == TextElement::ATOM) {
      Vector<const uc16> quarks = elm.data.u_atom->data();
      for (int 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[j] > String::kMaxAsciiCharCode) {
              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(isolate,
                                             compiler,
                                             quarks[j],
                                             backtrack,
                                             cp_offset + j,
                                             *checked_up_to < cp_offset + j,
                                             preloaded);
          if (bound_checked) UpdateBoundsCheck(cp_offset + j, checked_up_to);
        }
      }
    } else {
      ASSERT_EQ(elm.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.data.u_char_class;
        EmitCharClass(assembler,
                      cc,
                      ascii,
                      backtrack,
                      cp_offset,
                      *checked_up_to < cp_offset,
                      preloaded);
        UpdateBoundsCheck(cp_offset, checked_up_to);
      }
    }
  }
}


int TextNode::Length() {
  TextElement elm = elms_->last();
  ASSERT(elm.cp_offset >= 0);
  if (elm.type == TextElement::ATOM) {
    return elm.cp_offset + elm.data.u_atom->data().length();
  } else {
    return elm.cp_offset + 1;
  }
}


bool TextNode::SkipPass(int int_pass, bool ignore_case) {
  TextEmitPassType pass = static_cast<TextEmitPassType>(int_pass);
  if (ignore_case) {
    return pass == SIMPLE_CHARACTER_MATCH;
  } else {
    return pass == NON_LETTER_CHARACTER_MATCH || pass == CASE_CHARACTER_MATCH;
  }
}


// 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()) {
    int dummy = 0;
    TextEmitPass(compiler, NON_ASCII_MATCH, false, trace, false, &dummy);
  }

  bool first_elt_done = false;
  int 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 (int 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 (int 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 Trace::InvalidateCurrentCharacter() {
  characters_preloaded_ = 0;
}


void Trace::AdvanceCurrentPositionInTrace(int 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_ = Max(0, bound_checked_up_to_ - by);
}


void TextNode::MakeCaseIndependent(bool is_ascii) {
  int element_count = elms_->length();
  for (int i = 0; i < element_count; i++) {
    TextElement elm = elms_->at(i);
    if (elm.type == TextElement::CHAR_CLASS) {
      RegExpCharacterClass* cc = elm.data.u_char_class;
      // None of the standard character classes is different in the case
      // independent case and it slows us down if we don't know that.
      if (cc->is_standard()) continue;
      ZoneList<CharacterRange>* ranges = cc->ranges();
      int range_count = ranges->length();
      for (int j = 0; j < range_count; j++) {
        ranges->at(j).AddCaseEquivalents(ranges, is_ascii);
      }
    }
  }
}


int TextNode::GreedyLoopTextLength() {
  TextElement elm = elms_->at(elms_->length() - 1);
  if (elm.type == TextElement::CHAR_CLASS) {
    return elm.cp_offset + 1;
  } else {
    return elm.cp_offset + elm.data.u_atom->data().length();
  }
}


// Finds the fixed match length of a sequence of nodes that goes from
// this alternative and back to this choice node.  If there are variable
// length nodes or other complications in the way then return a sentinel
// value indicating that a greedy loop cannot be constructed.
int ChoiceNode::GreedyLoopTextLengthForAlternative(
    GuardedAlternative* alternative) {
  int length = 0;
  RegExpNode* node = alternative->node();
  // Later we will generate code for all these text nodes using recursion
  // so we have to limit the max number.
  int recursion_depth = 0;
  while (node != this) {
    if (recursion_depth++ > RegExpCompiler::kMaxRecursion) {
      return kNodeIsTooComplexForGreedyLoops;
    }
    int node_length = node->GreedyLoopTextLength();
    if (node_length == kNodeIsTooComplexForGreedyLoops) {
      return kNodeIsTooComplexForGreedyLoops;
    }
    length += node_length;
    SeqRegExpNode* seq_node = static_cast<SeqRegExpNode*>(node);
    node = seq_node->on_success();
  }
  return length;
}


void LoopChoiceNode::AddLoopAlternative(GuardedAlternative alt) {
  ASSERT_EQ(loop_node_, NULL);
  AddAlternative(alt);
  loop_node_ = alt.node();
}


void LoopChoiceNode::AddContinueAlternative(GuardedAlternative alt) {
  ASSERT_EQ(continue_node_, NULL);
  AddAlternative(alt);
  continue_node_ = alt.node();
}


void LoopChoiceNode::Emit(RegExpCompiler* compiler, Trace* trace) {
  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
  if (trace->stop_node() == this) {
    int text_length =
        GreedyLoopTextLengthForAlternative(&(alternatives_->at(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);
}


int ChoiceNode::CalculatePreloadCharacters(RegExpCompiler* compiler,
                                           bool not_at_start) {
  int preload_characters = EatsAtLeast(4, 0, not_at_start);
  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;
}


// 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 Malloced {
 public:
  AlternativeGeneration()
      : possible_success(),
        expects_preload(false),
        after(),
        quick_check_details() { }
  Label possible_success;
  bool expects_preload;
  Label 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(int count)
      : alt_gens_(count) {
    for (int i = 0; i < count && i < kAFew; i++) {
      alt_gens_.Add(a_few_alt_gens_ + i);
    }
    for (int i = kAFew; i < count; i++) {
      alt_gens_.Add(new AlternativeGeneration());
    }
  }
  ~AlternativeGenerationList() {
    for (int i = kAFew; i < alt_gens_.length(); i++) {
      delete alt_gens_[i];
      alt_gens_[i] = NULL;
    }
  }

  AlternativeGeneration* at(int i) {
    return alt_gens_[i];
  }

 private:
  static const int kAFew = 10;
  ZoneList<AlternativeGeneration*> alt_gens_;
  AlternativeGeneration a_few_alt_gens_[kAFew];
};


/* Code generation for choice nodes.
 *
 * We generate quick checks that do a mask and compare to eliminate a
 * choice.  If the quick check succeeds then it jumps to the continuation to
 * do slow checks and check subsequent nodes.  If it fails (the common case)
 * it falls through to the next choice.
 *
 * Here is the desired flow graph.  Nodes directly below each other imply
 * fallthrough.  Alternatives 1 and 2 have quick checks.  Alternative
 * 3 doesn't have a quick check so we have to call the slow check.
 * Nodes are marked Qn for quick checks and Sn for slow checks.  The entire
 * regexp continuation is generated directly after the Sn node, up to the
 * next GoTo if we decide to reuse some already generated code.  Some
 * nodes expect preload_characters to be preloaded into the current
 * character register.  R nodes do this preloading.  Vertices are marked
 * F for failures and S for success (possible success in the case of quick
 * nodes).  L, V, < and > are used as arrow heads.
 *
 * ----------> R
 *             |
 *             V
 *            Q1 -----> S1
 *             |   S   /
 *            F|      /
 *             |    F/
 *             |    /
 *             |   R
 *             |  /
 *             V L
 *            Q2 -----> S2
 *             |   S   /
 *            F|      /
 *             |    F/
 *             |    /
 *             |   R
 *             |  /
 *             V L
 *            S3
 *             |
 *            F|
 *             |
 *             R
 *             |
 * backtrack   V
 * <----------Q4
 *   \    F    |
 *    \        |S
 *     \   F   V
 *      \-----S4
 *
 * For greedy loops we reverse our expectation and expect to match rather
 * than fail. Therefore we want the loop code to look like this (U is the
 * unwind code that steps back in the greedy loop).  The following alternatives
 * look the same as above.
 *              _____
 *             /     \
 *             V     |
 * ----------> S1    |
 *            /|     |
 *           / |S    |
 *         F/  \_____/
 *         /
 *        |<-----------
 *        |            \
 *        V             \
 *        Q2 ---> S2     \
 *        |  S   /       |
 *       F|     /        |
 *        |   F/         |
 *        |   /          |
 *        |  R           |
 *        | /            |
 *   F    VL             |
 * <------U              |
 * back   |S             |
 *        \______________/
 */


void ChoiceNode::Emit(RegExpCompiler* compiler, Trace* trace) {
  RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
  int choice_count = alternatives_->length();
#ifdef DEBUG
  for (int i = 0; i < choice_count - 1; i++) {
    GuardedAlternative alternative = alternatives_->at(i);
    ZoneList<Guard*>* guards = alternative.guards();
    int guard_count = (guards == NULL) ? 0 : guards->length();
    for (int 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);

  int 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;

  int text_length = GreedyLoopTextLengthForAlternative(&(alternatives_->at(0)));
  bool greedy_loop = false;
  Label 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;
    Label 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);
    Label loop_label;
    macro_assembler->Bind(&loop_label);
    greedy_match_trace.set_stop_node(this);
    greedy_match_trace.set_loop_label(&loop_label);
    alternatives_->at(0).node()->Emit(compiler, &greedy_match_trace);
    macro_assembler->Bind(&greedy_match_failed);
  }

  Label second_choice;  // For use in greedy matches.
  macro_assembler->Bind(&second_choice);

  int first_normal_choice =