Google Git
Sign in
cobalt / cobalt / ac9ac065c5083565917b15e7a672a79599f2f00b / . / src / v8 / src / regexp / jsregexp.cc
blob: a26a1d77ce7fbfe22ef2b533c891eb1a48869ab6 [file] [log] [blame]
// Copyright 2012 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "src/regexp/jsregexp.h"
#include <memory>
#include <vector>
#include "src/base/platform/platform.h"
#include "src/compilation-cache.h"
#include "src/elements.h"
#include "src/execution.h"
#include "src/factory.h"
#include "src/isolate-inl.h"
#include "src/messages.h"
#include "src/ostreams.h"
#include "src/regexp/interpreter-irregexp.h"
#include "src/regexp/jsregexp-inl.h"
#include "src/regexp/regexp-macro-assembler-irregexp.h"
#include "src/regexp/regexp-macro-assembler-tracer.h"
#include "src/regexp/regexp-macro-assembler.h"
#include "src/regexp/regexp-parser.h"
#include "src/regexp/regexp-stack.h"
#include "src/runtime/runtime.h"
#include "src/splay-tree-inl.h"
#include "src/string-search.h"
#include "src/unicode-decoder.h"
#include "src/unicode-inl.h"
#ifdef V8_INTL_SUPPORT
#include "unicode/uniset.h"
#include "unicode/utypes.h"
#endif // V8_INTL_SUPPORT
#ifndef V8_INTERPRETED_REGEXP
#if V8_TARGET_ARCH_IA32
#include "src/regexp/ia32/regexp-macro-assembler-ia32.h"
#elif V8_TARGET_ARCH_X64
#include "src/regexp/x64/regexp-macro-assembler-x64.h"
#elif V8_TARGET_ARCH_ARM64
#include "src/regexp/arm64/regexp-macro-assembler-arm64.h"
#elif V8_TARGET_ARCH_ARM
#include "src/regexp/arm/regexp-macro-assembler-arm.h"
#elif V8_TARGET_ARCH_PPC
#include "src/regexp/ppc/regexp-macro-assembler-ppc.h"
#elif V8_TARGET_ARCH_S390
#include "src/regexp/s390/regexp-macro-assembler-s390.h"
#elif V8_TARGET_ARCH_MIPS
#include "src/regexp/mips/regexp-macro-assembler-mips.h"
#elif V8_TARGET_ARCH_MIPS64
#include "src/regexp/mips64/regexp-macro-assembler-mips64.h"
#else
#error Unsupported target architecture.
#endif
#endif
namespace v8 {
namespace internal {
MUST_USE_RESULT
static inline MaybeHandle<Object> ThrowRegExpException(
Handle<JSRegExp> re, Handle<String> pattern, Handle<String> error_text) {
Isolate* isolate = re->GetIsolate();
THROW_NEW_ERROR(isolate, NewSyntaxError(MessageTemplate::kMalformedRegExp,
pattern, error_text),
Object);
}
inline void ThrowRegExpException(Handle<JSRegExp> re,
Handle<String> error_text) {
USE(ThrowRegExpException(re, Handle<String>(re->Pattern()), error_text));
}
ContainedInLattice AddRange(ContainedInLattice containment,
const int* ranges,
int ranges_length,
Interval new_range) {
DCHECK_EQ(1, ranges_length & 1);
DCHECK_EQ(String::kMaxCodePoint + 1, ranges[ranges_length - 1]);
if (containment == kLatticeUnknown) return containment;
bool inside = false;
int last = 0;
for (int i = 0; i < ranges_length; inside = !inside, last = ranges[i], i++) {
// Consider the range from last to ranges[i].
// We haven't got to the new range yet.
if (ranges[i] <= new_range.from()) continue;
// New range is wholly inside last-ranges[i]. Note that new_range.to() is
// inclusive, but the values in ranges are not.
if (last <= new_range.from() && new_range.to() < ranges[i]) {
return Combine(containment, inside ? kLatticeIn : kLatticeOut);
}
return kLatticeUnknown;
}
return containment;
}
// More makes code generation slower, less makes V8 benchmark score lower.
const int kMaxLookaheadForBoyerMoore = 8;
// In a 3-character pattern you can maximally step forwards 3 characters
// at a time, which is not always enough to pay for the extra logic.
const int kPatternTooShortForBoyerMoore = 2;
// Identifies the sort of regexps where the regexp engine is faster
// than the code used for atom matches.
static bool HasFewDifferentCharacters(Handle<String> pattern) {
int length = Min(kMaxLookaheadForBoyerMoore, pattern->length());
if (length <= kPatternTooShortForBoyerMoore) return false;
const int kMod = 128;
bool character_found[kMod];
int different = 0;
memset(&character_found[0], 0, sizeof(character_found));
for (int i = 0; i < length; i++) {
int ch = (pattern->Get(i) & (kMod - 1));
if (!character_found[ch]) {
character_found[ch] = true;
different++;
// We declare a regexp low-alphabet if it has at least 3 times as many
// characters as it has different characters.
if (different * 3 > length) return false;
}
}
return true;
}
// Generic RegExp methods. Dispatches to implementation specific methods.
MaybeHandle<Object> RegExpImpl::Compile(Handle<JSRegExp> re,
Handle<String> pattern,
JSRegExp::Flags flags) {
DCHECK(pattern->IsFlat());
Isolate* isolate = re->GetIsolate();
Zone zone(isolate->allocator(), ZONE_NAME);
CompilationCache* compilation_cache = isolate->compilation_cache();
MaybeHandle<FixedArray> maybe_cached =
compilation_cache->LookupRegExp(pattern, flags);
Handle<FixedArray> cached;
if (maybe_cached.ToHandle(&cached)) {
re->set_data(*cached);
return re;
}
PostponeInterruptsScope postpone(isolate);
RegExpCompileData parse_result;
FlatStringReader reader(isolate, pattern);
DCHECK(!isolate->has_pending_exception());
if (!RegExpParser::ParseRegExp(isolate, &zone, &reader, flags,
&parse_result)) {
// Throw an exception if we fail to parse the pattern.
return ThrowRegExpException(re, pattern, parse_result.error);
}
bool has_been_compiled = false;
if (parse_result.simple && !IgnoreCase(flags) && !IsSticky(flags) &&
!HasFewDifferentCharacters(pattern)) {
// Parse-tree is a single atom that is equal to the pattern.
AtomCompile(re, pattern, flags, pattern);
has_been_compiled = true;
} else if (parse_result.tree->IsAtom() && !IsSticky(flags) &&
parse_result.capture_count == 0) {
RegExpAtom* atom = parse_result.tree->AsAtom();
Vector<const uc16> atom_pattern = atom->data();
Handle<String> atom_string;
ASSIGN_RETURN_ON_EXCEPTION(
isolate, atom_string,
isolate->factory()->NewStringFromTwoByte(atom_pattern), Object);
if (!IgnoreCase(atom->flags()) && !HasFewDifferentCharacters(atom_string)) {
AtomCompile(re, pattern, flags, atom_string);
has_been_compiled = true;
}
}
if (!has_been_compiled) {
IrregexpInitialize(re, pattern, flags, parse_result.capture_count);
}
DCHECK(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;
}
MaybeHandle<Object> RegExpImpl::Exec(Handle<JSRegExp> regexp,
Handle<String> subject, int index,
Handle<RegExpMatchInfo> last_match_info) {
switch (regexp->TypeTag()) {
case JSRegExp::ATOM:
return AtomExec(regexp, subject, index, last_match_info);
case JSRegExp::IRREGEXP: {
return IrregexpExec(regexp, subject, index, last_match_info);
}
default:
UNREACHABLE();
}
}
// 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(Handle<RegExpMatchInfo> last_match_info,
String* subject, int from, int to) {
SealHandleScope shs(last_match_info->GetIsolate());
last_match_info->SetNumberOfCaptureRegisters(2);
last_match_info->SetLastSubject(subject);
last_match_info->SetLastInput(subject);
last_match_info->SetCapture(0, from);
last_match_info->SetCapture(1, to);
}
int RegExpImpl::AtomExecRaw(Handle<JSRegExp> regexp,
Handle<String> subject,
int index,
int32_t* output,
int output_size) {
Isolate* isolate = regexp->GetIsolate();
DCHECK_LE(0, index);
DCHECK_LE(index, subject->length());
subject = String::Flatten(subject);
DisallowHeapAllocation no_gc; // ensure vectors stay valid
String* needle = String::cast(regexp->DataAt(JSRegExp::kAtomPatternIndex));
int needle_len = needle->length();
DCHECK(needle->IsFlat());
DCHECK_LT(0, needle_len);
if (index + needle_len > subject->length()) {
return RegExpImpl::RE_FAILURE;
}
for (int i = 0; i < output_size; i += 2) {
String::FlatContent needle_content = needle->GetFlatContent();
String::FlatContent subject_content = subject->GetFlatContent();
DCHECK(needle_content.IsFlat());
DCHECK(subject_content.IsFlat());
// dispatch on type of strings
index =
(needle_content.IsOneByte()
? (subject_content.IsOneByte()
? SearchString(isolate, subject_content.ToOneByteVector(),
needle_content.ToOneByteVector(), index)
: SearchString(isolate, subject_content.ToUC16Vector(),
needle_content.ToOneByteVector(), index))
: (subject_content.IsOneByte()
? SearchString(isolate, subject_content.ToOneByteVector(),
needle_content.ToUC16Vector(), index)
: SearchString(isolate, subject_content.ToUC16Vector(),
needle_content.ToUC16Vector(), index)));
if (index == -1) {
return i / 2; // Return number of matches.
} else {
output[i] = index;
output[i+1] = index + needle_len;
index += needle_len;
}
}
return output_size / 2;
}
Handle<Object> RegExpImpl::AtomExec(Handle<JSRegExp> re, Handle<String> subject,
int index,
Handle<RegExpMatchInfo> last_match_info) {
Isolate* isolate = re->GetIsolate();
static const int kNumRegisters = 2;
STATIC_ASSERT(kNumRegisters <= Isolate::kJSRegexpStaticOffsetsVectorSize);
int32_t* output_registers = isolate->jsregexp_static_offsets_vector();
int res = AtomExecRaw(re, subject, index, output_registers, kNumRegisters);
if (res == RegExpImpl::RE_FAILURE) return isolate->factory()->null_value();
DCHECK_EQ(res, RegExpImpl::RE_SUCCESS);
SealHandleScope shs(isolate);
SetAtomLastCapture(last_match_info, *subject, output_registers[0],
output_registers[1]);
return last_match_info;
}
// Irregexp implementation.
// Ensures that the regexp object contains a compiled version of the
// source for either one-byte or two-byte subject 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,
Handle<String> sample_subject,
bool is_one_byte) {
Object* compiled_code = re->DataAt(JSRegExp::code_index(is_one_byte));
#ifdef V8_INTERPRETED_REGEXP
if (compiled_code->IsByteArray()) return true;
#else // V8_INTERPRETED_REGEXP (RegExp native code)
if (compiled_code->IsCode()) return true;
#endif
return CompileIrregexp(re, sample_subject, is_one_byte);
}
bool RegExpImpl::CompileIrregexp(Handle<JSRegExp> re,
Handle<String> sample_subject,
bool is_one_byte) {
// Compile the RegExp.
Isolate* isolate = re->GetIsolate();
Zone zone(isolate->allocator(), ZONE_NAME);
PostponeInterruptsScope postpone(isolate);
#ifdef DEBUG
Object* entry = re->DataAt(JSRegExp::code_index(is_one_byte));
// When arriving here entry can only be a smi representing an uncompiled
// regexp.
DCHECK(entry->IsSmi());
int entry_value = Smi::ToInt(entry);
DCHECK_EQ(JSRegExp::kUninitializedValue, entry_value);
#endif
JSRegExp::Flags flags = re->GetFlags();
Handle<String> pattern(re->Pattern());
pattern = String::Flatten(pattern);
RegExpCompileData compile_data;
FlatStringReader reader(isolate, pattern);
if (!RegExpParser::ParseRegExp(isolate, &zone, &reader, flags,
&compile_data)) {
// Throw an exception if we fail to parse the pattern.
// THIS SHOULD NOT HAPPEN. We already pre-parsed it successfully once.
USE(ThrowRegExpException(re, pattern, compile_data.error));
return false;
}
RegExpEngine::CompilationResult result =
RegExpEngine::Compile(isolate, &zone, &compile_data, flags, pattern,
sample_subject, is_one_byte);
if (result.error_message != nullptr) {
// Unable to compile regexp.
if (FLAG_abort_on_stack_or_string_length_overflow &&
strncmp(result.error_message, "Stack overflow", 15) == 0) {
FATAL("Aborting on stack overflow");
}
Handle<String> error_message = isolate->factory()->NewStringFromUtf8(
CStrVector(result.error_message)).ToHandleChecked();
ThrowRegExpException(re, error_message);
return false;
}
Handle<FixedArray> data = Handle<FixedArray>(FixedArray::cast(re->data()));
data->set(JSRegExp::code_index(is_one_byte), result.code);
SetIrregexpCaptureNameMap(*data, compile_data.capture_name_map);
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));
}
void RegExpImpl::SetIrregexpCaptureNameMap(FixedArray* re,
Handle<FixedArray> value) {
if (value.is_null()) {
re->set(JSRegExp::kIrregexpCaptureNameMapIndex, Smi::kZero);
} else {
re->set(JSRegExp::kIrregexpCaptureNameMapIndex, *value);
}
}
int RegExpImpl::IrregexpNumberOfCaptures(FixedArray* re) {
return Smi::ToInt(re->get(JSRegExp::kIrregexpCaptureCountIndex));
}
int RegExpImpl::IrregexpNumberOfRegisters(FixedArray* re) {
return Smi::ToInt(re->get(JSRegExp::kIrregexpMaxRegisterCountIndex));
}
ByteArray* RegExpImpl::IrregexpByteCode(FixedArray* re, bool is_one_byte) {
return ByteArray::cast(re->get(JSRegExp::code_index(is_one_byte)));
}
Code* RegExpImpl::IrregexpNativeCode(FixedArray* re, bool is_one_byte) {
return Code::cast(re->get(JSRegExp::code_index(is_one_byte)));
}
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) {
DCHECK(subject->IsFlat());
// Check representation of the underlying storage.
bool is_one_byte = subject->IsOneByteRepresentationUnderneath();
if (!EnsureCompiledIrregexp(regexp, subject, is_one_byte)) return -1;
#ifdef V8_INTERPRETED_REGEXP
// Byte-code regexp needs space allocated for all its registers.
// The result captures are copied to the start of the registers array
// if the match succeeds. This way those registers are not clobbered
// when we set the last match info from last successful match.
return IrregexpNumberOfRegisters(FixedArray::cast(regexp->data())) +
(IrregexpNumberOfCaptures(FixedArray::cast(regexp->data())) + 1) * 2;
#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
}
int RegExpImpl::IrregexpExecRaw(Handle<JSRegExp> regexp,
Handle<String> subject,
int index,
int32_t* output,
int output_size) {
Isolate* isolate = regexp->GetIsolate();
Handle<FixedArray> irregexp(FixedArray::cast(regexp->data()), isolate);
DCHECK_LE(0, index);
DCHECK_LE(index, subject->length());
DCHECK(subject->IsFlat());
bool is_one_byte = subject->IsOneByteRepresentationUnderneath();
#ifndef V8_INTERPRETED_REGEXP
DCHECK(output_size >= (IrregexpNumberOfCaptures(*irregexp) + 1) * 2);
do {
EnsureCompiledIrregexp(regexp, subject, is_one_byte);
Handle<Code> code(IrregexpNativeCode(*irregexp, is_one_byte), isolate);
// The stack is used to allocate registers for the compiled regexp code.
// This means that in case of failure, the output registers array is left
// untouched and contains the capture results from the previous successful
// match. We can use that to set the last match info lazily.
NativeRegExpMacroAssembler::Result res =
NativeRegExpMacroAssembler::Match(code,
subject,
output,
output_size,
index,
isolate);
if (res != NativeRegExpMacroAssembler::RETRY) {
DCHECK(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 Latin1 and UC16,
// but the characters are always the same).
IrregexpPrepare(regexp, subject);
is_one_byte = subject->IsOneByteRepresentationUnderneath();
} while (true);
UNREACHABLE();
#else // V8_INTERPRETED_REGEXP
DCHECK(output_size >= IrregexpNumberOfRegisters(*irregexp));
// We must have done EnsureCompiledIrregexp, so we can get the number of
// registers.
int number_of_capture_registers =
(IrregexpNumberOfCaptures(*irregexp) + 1) * 2;
int32_t* raw_output = &output[number_of_capture_registers];
// We do not touch the actual capture result registers until we know there
// has been a match so that we can use those capture results to set the
// last match info.
for (int i = number_of_capture_registers - 1; i >= 0; i--) {
raw_output[i] = -1;
}
Handle<ByteArray> byte_codes(IrregexpByteCode(*irregexp, is_one_byte),
isolate);
IrregexpResult result = IrregexpInterpreter::Match(isolate,
byte_codes,
subject,
raw_output,
index);
if (result == RE_SUCCESS) {
// Copy capture results to the start of the registers array.
MemCopy(output, raw_output, number_of_capture_registers * sizeof(int32_t));
}
if (result == RE_EXCEPTION) {
DCHECK(!isolate->has_pending_exception());
isolate->StackOverflow();
}
return result;
#endif // V8_INTERPRETED_REGEXP
}
MaybeHandle<Object> RegExpImpl::IrregexpExec(
Handle<JSRegExp> regexp, Handle<String> subject, int previous_index,
Handle<RegExpMatchInfo> last_match_info) {
Isolate* isolate = regexp->GetIsolate();
DCHECK_EQ(regexp->TypeTag(), JSRegExp::IRREGEXP);
subject = String::Flatten(subject);
// Prepare space for the return values.
#if defined(V8_INTERPRETED_REGEXP) && defined(DEBUG)
if (FLAG_trace_regexp_bytecodes) {
String* pattern = regexp->Pattern();
PrintF("\n\nRegexp match: /%s/\n\n", pattern->ToCString().get());
PrintF("\n\nSubject string: '%s'\n\n", subject->ToCString().get());
}
#endif
int required_registers = RegExpImpl::IrregexpPrepare(regexp, subject);
if (required_registers < 0) {
// Compiling failed with an exception.
DCHECK(isolate->has_pending_exception());
return MaybeHandle<Object>();
}
int32_t* output_registers = nullptr;
if (required_registers > Isolate::kJSRegexpStaticOffsetsVectorSize) {
output_registers = NewArray<int32_t>(required_registers);
}
std::unique_ptr<int32_t[]> auto_release(output_registers);
if (output_registers == nullptr) {
output_registers = isolate->jsregexp_static_offsets_vector();
}
int res = RegExpImpl::IrregexpExecRaw(
regexp, subject, previous_index, output_registers, required_registers);
if (res == RE_SUCCESS) {
int capture_count =
IrregexpNumberOfCaptures(FixedArray::cast(regexp->data()));
return SetLastMatchInfo(
last_match_info, subject, capture_count, output_registers);
}
if (res == RE_EXCEPTION) {
DCHECK(isolate->has_pending_exception());
return MaybeHandle<Object>();
}
DCHECK(res == RE_FAILURE);
return isolate->factory()->null_value();
}
Handle<RegExpMatchInfo> RegExpImpl::SetLastMatchInfo(
Handle<RegExpMatchInfo> last_match_info, Handle<String> subject,
int capture_count, int32_t* match) {
// This is the only place where match infos can grow. If, after executing the
// regexp, RegExpExecStub finds that the match info is too small, it restarts
// execution in RegExpImpl::Exec, which finally grows the match info right
// here.
int capture_register_count = (capture_count + 1) * 2;
Handle<RegExpMatchInfo> result =
RegExpMatchInfo::ReserveCaptures(last_match_info, capture_register_count);
result->SetNumberOfCaptureRegisters(capture_register_count);
if (*result != *last_match_info) {
// The match info has been reallocated, update the corresponding reference
// on the native context.
Isolate* isolate = last_match_info->GetIsolate();
if (*last_match_info == *isolate->regexp_last_match_info()) {
isolate->native_context()->set_regexp_last_match_info(*result);
} else if (*last_match_info == *isolate->regexp_internal_match_info()) {
isolate->native_context()->set_regexp_internal_match_info(*result);
}
}
DisallowHeapAllocation no_allocation;
if (match != nullptr) {
for (int i = 0; i < capture_register_count; i += 2) {
result->SetCapture(i, match[i]);
result->SetCapture(i + 1, match[i + 1]);
}
}
result->SetLastSubject(*subject);
result->SetLastInput(*subject);
return result;
}
RegExpImpl::GlobalCache::GlobalCache(Handle<JSRegExp> regexp,
Handle<String> subject, Isolate* isolate)
: register_array_(nullptr),
register_array_size_(0),
regexp_(regexp),
subject_(subject) {
#ifdef V8_INTERPRETED_REGEXP
bool interpreted = true;
#else
bool interpreted = false;
#endif // V8_INTERPRETED_REGEXP
if (regexp_->TypeTag() == JSRegExp::ATOM) {
static const int kAtomRegistersPerMatch = 2;
registers_per_match_ = kAtomRegistersPerMatch;
// There is no distinction between interpreted and native for atom regexps.
interpreted = false;
} else {
registers_per_match_ = RegExpImpl::IrregexpPrepare(regexp_, subject_);
if (registers_per_match_ < 0) {
num_matches_ = -1; // Signal exception.
return;
}
}
DCHECK(IsGlobal(regexp->GetFlags()));
if (!interpreted) {
register_array_size_ =
Max(registers_per_match_, Isolate::kJSRegexpStaticOffsetsVectorSize);
max_matches_ = register_array_size_ / registers_per_match_;
} else {
// Global loop in interpreted regexp is not implemented. We choose
// the size of the offsets vector so that it can only store one match.
register_array_size_ = registers_per_match_;
max_matches_ = 1;
}
if (register_array_size_ > Isolate::kJSRegexpStaticOffsetsVectorSize) {
register_array_ = NewArray<int32_t>(register_array_size_);
} else {
register_array_ = isolate->jsregexp_static_offsets_vector();
}
// Set state so that fetching the results the first time triggers a call
// to the compiled regexp.
current_match_index_ = max_matches_ - 1;
num_matches_ = max_matches_;
DCHECK_LE(2, registers_per_match_); // Each match has at least one capture.
DCHECK_GE(register_array_size_, registers_per_match_);
int32_t* last_match =
&register_array_[current_match_index_ * registers_per_match_];
last_match[0] = -1;
last_match[1] = 0;
}
int RegExpImpl::GlobalCache::AdvanceZeroLength(int last_index) {
if (IsUnicode(regexp_->GetFlags()) && last_index + 1 < subject_->length() &&
unibrow::Utf16::IsLeadSurrogate(subject_->Get(last_index)) &&
unibrow::Utf16::IsTrailSurrogate(subject_->Get(last_index + 1))) {
// Advance over the surrogate pair.
return last_index + 2;
}
return last_index + 1;
}
// -------------------------------------------------------------------
// 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, Zone* zone) {
UNREACHABLE();
}
void RegExpAtom::AppendToText(RegExpText* text, Zone* zone) {
text->AddElement(TextElement::Atom(this), zone);
}
void RegExpCharacterClass::AppendToText(RegExpText* text, Zone* zone) {
text->AddElement(TextElement::CharClass(this), zone);
}
void RegExpText::AppendToText(RegExpText* text, Zone* zone) {
for (int i = 0; i < elements()->length(); i++)
text->AddElement(elements()->at(i), zone);
}
TextElement TextElement::Atom(RegExpAtom* atom) {
return TextElement(ATOM, atom);
}
TextElement TextElement::CharClass(RegExpCharacterClass* char_class) {
return TextElement(CHAR_CLASS, char_class);
}
int TextElement::length() const {
switch (text_type()) {
case ATOM:
return atom()->length();
case CHAR_CLASS:
return 1;
}
UNREACHABLE();
}
DispatchTable* ChoiceNode::GetTable(bool ignore_case) {
if (table_ == nullptr) {
table_ = new(zone()) DispatchTable(zone());
DispatchTableConstructor cons(table_, ignore_case, zone());
cons.BuildTable(this);
}
return table_;
}
class FrequencyCollator {
public:
FrequencyCollator() : total_samples_(0) {
for (int i = 0; i < RegExpMacroAssembler::kTableSize; i++) {
frequencies_[i] = CharacterFrequency(i);
}
}
void CountCharacter(int character) {
int index = (character & RegExpMacroAssembler::kTableMask);
frequencies_[index].Increment();
total_samples_++;
}
// Does not measure in percent, but rather per-128 (the table size from the
// regexp macro assembler).
int Frequency(int in_character) {
DCHECK((in_character & RegExpMacroAssembler::kTableMask) == in_character);
if (total_samples_ < 1) return 1; // Division by zero.
int freq_in_per128 =
(frequencies_[in_character].counter() * 128) / total_samples_;
return freq_in_per128;
}
private:
class CharacterFrequency {
public:
CharacterFrequency() : counter_(0), character_(-1) { }
explicit CharacterFrequency(int character)
: counter_(0), character_(character) { }
void Increment() { counter_++; }
int counter() { return counter_; }
int character() { return character_; }
private:
int counter_;
int character_;
};
private:
CharacterFrequency frequencies_[RegExpMacroAssembler::kTableSize];
int total_samples_;
};
class RegExpCompiler {
public:
RegExpCompiler(Isolate* isolate, Zone* zone, int capture_count,
bool is_one_byte);
int AllocateRegister() {
if (next_register_ >= RegExpMacroAssembler::kMaxRegister) {
reg_exp_too_big_ = true;
return next_register_;
}
return next_register_++;
}
// Lookarounds to match lone surrogates for unicode character class matches
// are never nested. We can therefore reuse registers.
int UnicodeLookaroundStackRegister() {
if (unicode_lookaround_stack_register_ == kNoRegister) {
unicode_lookaround_stack_register_ = AllocateRegister();
}
return unicode_lookaround_stack_register_;
}
int UnicodeLookaroundPositionRegister() {
if (unicode_lookaround_position_register_ == kNoRegister) {
unicode_lookaround_position_register_ = AllocateRegister();
}
return unicode_lookaround_position_register_;
}
RegExpEngine::CompilationResult Assemble(RegExpMacroAssembler* assembler,
RegExpNode* start,
int capture_count,
Handle<String> pattern);
inline void AddWork(RegExpNode* node) {
if (!node->on_work_list() && !node->label()->is_bound()) {
node->set_on_work_list(true);
work_list_->push_back(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 one_byte() { return one_byte_; }
inline bool optimize() { return optimize_; }
inline void set_optimize(bool value) { optimize_ = value; }
inline bool limiting_recursion() { return limiting_recursion_; }
inline void set_limiting_recursion(bool value) {
limiting_recursion_ = value;
}
bool read_backward() { return read_backward_; }
void set_read_backward(bool value) { read_backward_ = value; }
FrequencyCollator* frequency_collator() { return &frequency_collator_; }
int current_expansion_factor() { return current_expansion_factor_; }
void set_current_expansion_factor(int value) {
current_expansion_factor_ = value;
}
Isolate* isolate() const { return isolate_; }
Zone* zone() const { return zone_; }
static const int kNoRegister = -1;
private:
EndNode* accept_;
int next_register_;
int unicode_lookaround_stack_register_;
int unicode_lookaround_position_register_;
std::vector<RegExpNode*>* work_list_;
int recursion_depth_;
RegExpMacroAssembler* macro_assembler_;
bool one_byte_;
bool reg_exp_too_big_;
bool limiting_recursion_;
bool optimize_;
bool read_backward_;
int current_expansion_factor_;
FrequencyCollator frequency_collator_;
Isolate* isolate_;
Zone* zone_;
};
class RecursionCheck {
public:
explicit RecursionCheck(RegExpCompiler* compiler) : compiler_(compiler) {
compiler->IncrementRecursionDepth();
}
~RecursionCheck() { compiler_->DecrementRecursionDepth(); }
private:
RegExpCompiler* compiler_;
};
static RegExpEngine::CompilationResult IrregexpRegExpTooBig(Isolate* isolate) {
return RegExpEngine::CompilationResult(isolate, "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(Isolate* isolate, Zone* zone, int capture_count,
bool one_byte)
: next_register_(2 * (capture_count + 1)),
unicode_lookaround_stack_register_(kNoRegister),
unicode_lookaround_position_register_(kNoRegister),
work_list_(nullptr),
recursion_depth_(0),
one_byte_(one_byte),
reg_exp_too_big_(false),
limiting_recursion_(false),
optimize_(FLAG_regexp_optimization),
read_backward_(false),
current_expansion_factor_(1),
frequency_collator_(),
isolate_(isolate),
zone_(zone) {
accept_ = new(zone) EndNode(EndNode::ACCEPT, zone);
DCHECK_GE(RegExpMacroAssembler::kMaxRegister, next_register_ - 1);
}
RegExpEngine::CompilationResult RegExpCompiler::Assemble(
RegExpMacroAssembler* macro_assembler,
RegExpNode* start,
int capture_count,
Handle<String> pattern) {
Isolate* isolate = pattern->GetHeap()->isolate();
#ifdef DEBUG
if (FLAG_trace_regexp_assembler)
macro_assembler_ = new RegExpMacroAssemblerTracer(isolate, macro_assembler);
else
#endif
macro_assembler_ = macro_assembler;
std::vector<RegExpNode*> work_list;
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.empty()) {
RegExpNode* node = work_list.back();
work_list.pop_back();
node->set_on_work_list(false);
if (!node->label()->is_bound()) node->Emit(this, &new_trace);
}
if (reg_exp_too_big_) {
macro_assembler_->AbortedCodeGeneration();
return IrregexpRegExpTooBig(isolate_);
}
Handle<HeapObject> code = macro_assembler_->GetCode(pattern);
isolate->IncreaseTotalRegexpCodeGenerated(code->Size());
work_list_ = nullptr;
#if defined(ENABLE_DISASSEMBLER) && !defined(V8_INTERPRETED_REGEXP)
if (FLAG_print_code) {
CodeTracer::Scope trace_scope(isolate->GetCodeTracer());
OFStream os(trace_scope.file());
Handle<Code>::cast(code)->Disassemble(pattern->ToCString().get(), os);
}
#endif
#ifdef DEBUG
if (FLAG_trace_regexp_assembler) {
delete macro_assembler_;
}
#endif
return RegExpEngine::CompilationResult(*code, next_register_);
}
bool Trace::DeferredAction::Mentions(int that) {
if (action_type() == ActionNode::CLEAR_CAPTURES) {
Interval range = static_cast<DeferredClearCaptures*>(this)->range();
return range.Contains(that);
} else {
return reg() == that;
}
}
bool Trace::mentions_reg(int reg) {
for (DeferredAction* action = actions_; action != nullptr;
action = action->next()) {
if (action->Mentions(reg))
return true;
}
return false;
}
bool Trace::GetStoredPosition(int reg, int* cp_offset) {
DCHECK_EQ(0, *cp_offset);
for (DeferredAction* action = actions_; action != nullptr;
action = action->next()) {
if (action->Mentions(reg)) {
if (action->action_type() == ActionNode::STORE_POSITION) {
*cp_offset = static_cast<DeferredCapture*>(action)->cp_offset();
return true;
} else {
return false;
}
}
}
return false;
}
int Trace::FindAffectedRegisters(OutSet* affected_registers,
Zone* zone) {
int max_register = RegExpCompiler::kNoRegister;
for (DeferredAction* action = actions_; action != nullptr;
action = action->next()) {
if (action->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, zone);
if (range.to() > max_register) max_register = range.to();
} else {
affected_registers->Set(action->reg(), zone);
if (action->reg() > max_register) max_register = action->reg();
}
}
return max_register;
}
void Trace::RestoreAffectedRegisters(RegExpMacroAssembler* assembler,
int max_register,
const OutSet& registers_to_pop,
const 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,
const OutSet& affected_registers,
OutSet* registers_to_pop,
OutSet* registers_to_clear,
Zone* zone) {
// 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;
static const int kNoStore = kMinInt;
int store_position = kNoStore;
// This is a little tricky because we are scanning the actions in reverse
// historical order (newest first).
for (DeferredAction* action = actions_; action != nullptr;
action = action->next()) {
if (action->Mentions(reg)) {
switch (action->action_type()) {
case ActionNode::SET_REGISTER: {
Trace::DeferredSetRegister* psr =
static_cast<Trace::DeferredSetRegister*>(action);
if (!absolute) {
value += psr->value();
absolute = true;
}
// SET_REGISTER is currently only used for newly introduced loop
// counters. They can have a significant previous value if they
// occur 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;
DCHECK_EQ(store_position, kNoStore);
DCHECK(!clear);
break;
}
case ActionNode::INCREMENT_REGISTER:
if (!absolute) {
value++;
}
DCHECK_EQ(store_position, kNoStore);
DCHECK(!clear);
undo_action = RESTORE;
break;
case ActionNode::STORE_POSITION: {
Trace::DeferredCapture* pc =
static_cast<Trace::DeferredCapture*>(action);
if (!clear && store_position == kNoStore) {
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;
}
DCHECK(!absolute);
DCHECK_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 == kNoStore) {
clear = true;
}
undo_action = RESTORE;
DCHECK(!absolute);
DCHECK_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, zone);
} else if (undo_action == CLEAR) {
registers_to_clear->Set(reg, zone);
}
// Perform the chronologically last action (or accumulated increment)
// for the register.
if (store_position != kNoStore) {
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();
DCHECK(!is_trivial());
if (actions_ == nullptr && backtrack() == nullptr) {
// 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() != nullptr) {
// 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,
compiler->zone());
OutSet registers_to_pop;
OutSet registers_to_clear;
PerformDeferredActions(assembler,
max_register,
affected_registers,
&registers_to_pop,
&registers_to_clear,
compiler->zone());
if (cp_offset_ != 0) {
assembler->AdvanceCurrentPosition(cp_offset_);
}
// Create a new trivial state and generate the node with that.
Label undo;
assembler->PushBacktrack(&undo);
if (successor->KeepRecursing(compiler)) {
Trace new_state;
successor->Emit(compiler, &new_state);
} else {
compiler->AddWork(successor);
assembler->GoTo(successor->label());
}
// On backtrack we need to restore state.
assembler->Bind(&undo);
RestoreAffectedRegisters(assembler,
max_register,
registers_to_pop,
registers_to_clear);
if (backtrack() == nullptr) {
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, Zone* zone) {
if (guards_ == nullptr) guards_ = new (zone) ZoneList<Guard*>(1, zone);
guards_->Add(guard, zone);
}
ActionNode* ActionNode::SetRegister(int reg,
int val,
RegExpNode* on_success) {
ActionNode* result =
new(on_success->zone()) 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(on_success->zone()) 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(on_success->zone()) 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(on_success->zone()) 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(on_success->zone()) 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(on_success->zone()) 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(on_success->zone()) 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:
DCHECK(!trace->mentions_reg(guard->reg()));
macro_assembler->IfRegisterGE(guard->reg(),
guard->value(),
trace->backtrack());
break;
case Guard::GEQ:
DCHECK(!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 Latin1.
static int GetCaseIndependentLetters(Isolate* isolate, uc16 character,
bool one_byte_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 (one_byte_subject) {
int new_length = 0;
for (int i = 0; i < length; i++) {
if (letters[i] <= String::kMaxOneByteCharCode) {
letters[new_length++] = letters[i];
}
}
length = new_length;
}
return length;
}
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 one_byte = compiler->one_byte();
unibrow::uchar chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
int length = GetCaseIndependentLetters(isolate, c, one_byte, chars);
if (length < 1) {
// This can't match. Must be an one-byte subject and a non-one-byte
// character. We do not need to do anything since the one-byte 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 (one_byte && c > String::kMaxOneByteCharCodeU) {
// 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 one_byte, uc16 c1, uc16 c2,
Label* on_failure) {
uc16 char_mask;
if (one_byte) {
char_mask = String::kMaxOneByteCharCode;
} 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.
DCHECK(c2 > c1);
uc16 mask = char_mask ^ exor;
macro_assembler->CheckNotCharacterAfterAnd(c1, mask, on_failure);
return true;
}
DCHECK(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 one_byte = compiler->one_byte();
unibrow::uchar chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
int length = GetCaseIndependentLetters(isolate, c, one_byte, 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;
DCHECK_EQ(4, unibrow::Ecma262UnCanonicalize::kMaxWidth);
switch (length) {
case 2: {
if (ShortCutEmitCharacterPair(macro_assembler, one_byte, 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 EmitBoundaryTest(RegExpMacroAssembler* masm,
int border,
Label* fall_through,
Label* above_or_equal,
Label* below) {
if (below != fall_through) {
masm->CheckCharacterLT(border, below);
if (above_or_equal != fall_through) masm->GoTo(above_or_equal);
} else {
masm->CheckCharacterGT(border - 1, above_or_equal);
}
}
static void EmitDoubleBoundaryTest(RegExpMacroAssembler* masm,
int first,
int last,
Label* fall_through,
Label* in_range,
Label* out_of_range) {
if (in_range == fall_through) {
if (first == last) {
masm->CheckNotCharacter(first, out_of_range);
} else {
masm->CheckCharacterNotInRange(first, last, out_of_range);
}
} else {
if (first == last) {
masm->CheckCharacter(first, in_range);
} else {
masm->CheckCharacterInRange(first, last, in_range);
}
if (out_of_range != fall_through) masm->GoTo(out_of_range);
}
}
// even_label is for ranges[i] to ranges[i + 1] where i - start_index is even.
// odd_label is for ranges[i] to ranges[i + 1] where i - start_index is odd.
static void EmitUseLookupTable(
RegExpMacroAssembler* masm,
ZoneList<int>* ranges,
int start_index,
int end_index,
int min_char,
Label* fall_through,
Label* even_label,
Label* odd_label) {
static const int kSize = RegExpMacroAssembler::kTableSize;
static const int kMask = RegExpMacroAssembler::kTableMask;
int base = (min_char & ~kMask);
USE(base);
// Assert that everything is on one kTableSize page.
for (int i = start_index; i <= end_index; i++) {
DCHECK_EQ(ranges->at(i) & ~kMask, base);
}
DCHECK(start_index == 0 || (ranges->at(start_index - 1) & ~kMask) <= base);
char templ[kSize];
Label* on_bit_set;
Label* on_bit_clear;
int bit;
if (even_label == fall_through) {
on_bit_set = odd_label;
on_bit_clear = even_label;
bit = 1;
} else {
on_bit_set = even_label;
on_bit_clear = odd_label;
bit = 0;
}
for (int i = 0; i < (ranges->at(start_index) & kMask) && i < kSize; i++) {
templ[i] = bit;
}
int j = 0;
bit ^= 1;
for (int i = start_index; i < end_index; i++) {
for (j = (ranges->at(i) & kMask); j < (ranges->at(i + 1) & kMask); j++) {
templ[j] = bit;
}
bit ^= 1;
}
for (int i = j; i < kSize; i++) {
templ[i] = bit;
}
Factory* factory = masm->isolate()->factory();
// TODO(erikcorry): Cache these.
Handle<ByteArray> ba = factory->NewByteArray(kSize, TENURED);
for (int i = 0; i < kSize; i++) {
ba->set(i, templ[i]);
}
masm->CheckBitInTable(ba, on_bit_set);
if (on_bit_clear != fall_through) masm->GoTo(on_bit_clear);
}
static void CutOutRange(RegExpMacroAssembler* masm,
ZoneList<int>* ranges,
int start_index,
int end_index,
int cut_index,
Label* even_label,
Label* odd_label) {
bool odd = (((cut_index - start_index) & 1) == 1);
Label* in_range_label = odd ? odd_label : even_label;
Label dummy;
EmitDoubleBoundaryTest(masm,
ranges->at(cut_index),
ranges->at(cut_index + 1) - 1,
&dummy,
in_range_label,
&dummy);
DCHECK(!dummy.is_linked());
// Cut out the single range by rewriting the array. This creates a new
// range that is a merger of the two ranges on either side of the one we
// are cutting out. The oddity of the labels is preserved.
for (int j = cut_index; j > start_index; j--) {
ranges->at(j) = ranges->at(j - 1);
}
for (int j = cut_index + 1; j < end_index; j++) {
ranges->at(j) = ranges->at(j + 1);
}
}
// Unicode case. Split the search space into kSize spaces that are handled
// with recursion.
static void SplitSearchSpace(ZoneList<int>* ranges,
int start_index,
int end_index,
int* new_start_index,
int* new_end_index,
int* border) {
static const int kSize = RegExpMacroAssembler::kTableSize;
static const int kMask = RegExpMacroAssembler::kTableMask;
int first = ranges->at(start_index);
int last = ranges->at(end_index) - 1;
*new_start_index = start_index;
*border = (ranges->at(start_index) & ~kMask) + kSize;
while (*new_start_index < end_index) {
if (ranges->at(*new_start_index) > *border) break;
(*new_start_index)++;
}
// new_start_index is the index of the first edge that is beyond the
// current kSize space.
// For very large search spaces we do a binary chop search of the non-Latin1
// space instead of just going to the end of the current kSize space. The
// heuristics are complicated a little by the fact that any 128-character
// encoding space can be quickly tested with a table lookup, so we don't
// wish to do binary chop search at a smaller granularity than that. A
// 128-character space can take up a lot of space in the ranges array if,
// for example, we only want to match every second character (eg. the lower
// case characters on some Unicode pages).
int binary_chop_index = (end_index + start_index) / 2;
// The first test ensures that we get to the code that handles the Latin1
// range with a single not-taken branch, speeding up this important
// character range (even non-Latin1 charset-based text has spaces and
// punctuation).
if (*border - 1 > String::kMaxOneByteCharCode && // Latin1 case.
end_index - start_index > (*new_start_index - start_index) * 2 &&
last - first > kSize * 2 && binary_chop_index > *new_start_index &&
ranges->at(binary_chop_index) >= first + 2 * kSize) {
int scan_forward_for_section_border = binary_chop_index;;
int new_border = (ranges->at(binary_chop_index) | kMask) + 1;
while (scan_forward_for_section_border < end_index) {
if (ranges->at(scan_forward_for_section_border) > new_border) {
*new_start_index = scan_forward_for_section_border;
*border = new_border;
break;
}
scan_forward_for_section_border++;
}
}
DCHECK(*new_start_index > start_index);
*new_end_index = *new_start_index - 1;
if (ranges->at(*new_end_index) == *border) {
(*new_end_index)--;
}
if (*border >= ranges->at(end_index)) {
*border = ranges->at(end_index);
*new_start_index = end_index; // Won't be used.
*new_end_index = end_index - 1;
}
}
// Gets a series of segment boundaries representing a character class. If the
// character is in the range between an even and an odd boundary (counting from
// start_index) then go to even_label, otherwise go to odd_label. We already
// know that the character is in the range of min_char to max_char inclusive.
// Either label can be nullptr indicating backtracking. Either label can also
// be equal to the fall_through label.
static void GenerateBranches(RegExpMacroAssembler* masm, ZoneList<int>* ranges,
int start_index, int end_index, uc32 min_char,
uc32 max_char, Label* fall_through,
Label* even_label, Label* odd_label) {
DCHECK_LE(min_char, String::kMaxUtf16CodeUnit);
DCHECK_LE(max_char, String::kMaxUtf16CodeUnit);
int first = ranges->at(start_index);
int last = ranges->at(end_index) - 1;
DCHECK_LT(min_char, first);
// Just need to test if the character is before or on-or-after
// a particular character.
if (start_index == end_index) {
EmitBoundaryTest(masm, first, fall_through, even_label, odd_label);
return;
}
// Another almost trivial case: There is one interval in the middle that is
// different from the end intervals.
if (start_index + 1 == end_index) {
EmitDoubleBoundaryTest(
masm, first, last, fall_through, even_label, odd_label);
return;
}
// It's not worth using table lookup if there are very few intervals in the
// character class.
if (end_index - start_index <= 6) {
// It is faster to test for individual characters, so we look for those
// first, then try arbitrary ranges in the second round.
static int kNoCutIndex = -1;
int cut = kNoCutIndex;
for (int i = start_index; i < end_index; i++) {
if (ranges->at(i) == ranges->at(i + 1) - 1) {
cut = i;
break;
}
}
if (cut == kNoCutIndex) cut = start_index;
CutOutRange(
masm, ranges, start_index, end_index, cut, even_label, odd_label);
DCHECK_GE(end_index - start_index, 2);
GenerateBranches(masm,
ranges,
start_index + 1,
end_index - 1,
min_char,
max_char,
fall_through,
even_label,
odd_label);
return;
}
// If there are a lot of intervals in the regexp, then we will use tables to
// determine whether the character is inside or outside the character class.
static const int kBits = RegExpMacroAssembler::kTableSizeBits;
if ((max_char >> kBits) == (min_char >> kBits)) {
EmitUseLookupTable(masm,
ranges,
start_index,
end_index,
min_char,
fall_through,
even_label,
odd_label);
return;
}
if ((min_char >> kBits) != (first >> kBits)) {
masm->CheckCharacterLT(first, odd_label);
GenerateBranches(masm,
ranges,
start_index + 1,
end_index,
first,
max_char,
fall_through,
odd_label,
even_label);
return;
}
int new_start_index = 0;
int new_end_index = 0;
int border = 0;
SplitSearchSpace(ranges,
start_index,
end_index,
&new_start_index,
&new_end_index,
&border);
Label handle_rest;
Label* above = &handle_rest;
if (border == last + 1) {
// We didn't find any section that started after the limit, so everything
// above the border is one of the terminal labels.
above = (end_index & 1) != (start_index & 1) ? odd_label : even_label;
DCHECK(new_end_index == end_index - 1);
}
DCHECK_LE(start_index, new_end_index);
DCHECK_LE(new_start_index, end_index);
DCHECK_LT(start_index, new_start_index);
DCHECK_LT(new_end_index, end_index);
DCHECK(new_end_index + 1 == new_start_index ||
(new_end_index + 2 == new_start_index &&
border == ranges->at(new_end_index + 1)));
DCHECK_LT(min_char, border - 1);
DCHECK_LT(border, max_char);
DCHECK_LT(ranges->at(new_end_index), border);
DCHECK(border < ranges->at(new_start_index) ||
(border == ranges->at(new_start_index) &&
new_start_index == end_index &&
new_end_index == end_index - 1 &&
border == last + 1));
DCHECK(new_start_index == 0 || border >= ranges->at(new_start_index - 1));
masm->CheckCharacterGT(border - 1, above);
Label dummy;
GenerateBranches(masm,
ranges,
start_index,
new_end_index,
min_char,
border - 1,
&dummy,
even_label,
odd_label);
if (handle_rest.is_linked()) {
masm->Bind(&handle_rest);
bool flip = (new_start_index & 1) != (start_index & 1);
GenerateBranches(masm,
ranges,
new_start_index,
end_index,
border,
max_char,
&dummy,
flip ? odd_label : even_label,
flip ? even_label : odd_label);
}
}
static void EmitCharClass(RegExpMacroAssembler* macro_assembler,
RegExpCharacterClass* cc, bool one_byte,
Label* on_failure, int cp_offset, bool check_offset,
bool preloaded, Zone* zone) {
ZoneList<CharacterRange>* ranges = cc->ranges(zone);
CharacterRange::Canonicalize(ranges);
int max_char;
if (one_byte) {
max_char = String::kMaxOneByteCharCode;
} else {
max_char = String::kMaxUtf16CodeUnit;
}
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()) {
macro_assembler->GoTo(on_failure);
}
if (check_offset) {
macro_assembler->CheckPosition(cp_offset, on_failure);
}
return;
}
if (last_valid_range == 0 &&
ranges->at(0).IsEverything(max_char)) {
if (cc->is_negated()) {
macro_assembler->GoTo(on_failure);
} else {
// This is a common case hit by non-anchored expressions.
if (check_offset) {
macro_assembler->CheckPosition(cp_offset, on_failure);
}
}
return;
}
if (!preloaded) {
macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check_offset);
}
if (cc->is_standard(zone) &&
macro_assembler->CheckSpecialCharacterClass(cc->standard_type(),
on_failure)) {
return;
}
// A new list with ascending entries. Each entry is a code unit
// where there is a boundary between code units that are part of
// the class and code units that are not. Normally we insert an
// entry at zero which goes to the failure label, but if there
// was already one there we fall through for success on that entry.
// Subsequent entries have alternating meaning (success/failure).
ZoneList<int>* range_boundaries =
new(zone) ZoneList<int>(last_valid_range, zone);
bool zeroth_entry_is_failure = !cc->is_negated();
for (int i = 0; i <= last_valid_range; i++) {
CharacterRange& range = ranges->at(i);
if (range.from() == 0) {
DCHECK_EQ(i, 0);
zeroth_entry_is_failure = !zeroth_entry_is_failure;
} else {
range_boundaries->Add(range.from(), zone);
}
range_boundaries->Add(range.to() + 1, zone);
}
int end_index = range_boundaries->length() - 1;
if (range_boundaries->at(end_index) > max_char) {
end_index--;
}
Label fall_through;
GenerateBranches(macro_assembler,
range_boundaries,
0, // start_index.
end_index,
0, // min_char.
max_char,
&fall_through,
zeroth_entry_is_failure ? &fall_through : on_failure,
zeroth_entry_is_failure ? on_failure : &fall_through);
macro_assembler->Bind(&fall_through);
}
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() != nullptr) {
return CONTINUE;
}
RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
if (trace->is_trivial()) {
if (label_.is_bound() || on_work_list() || !KeepRecursing(compiler)) {
// If a generic version is already scheduled to be generated or we have
// recursed too deeply then just generate a jump to that code.
macro_assembler->GoTo(&label_);
// This will queue it up for generation of a generic version if it hasn't
// already been queued.
compiler->AddWork(this);
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 (KeepRecursing(compiler) && compiler->optimize() &&
trace_count_ < kMaxCopiesCodeGenerated) {
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.
bool was_limiting = compiler->limiting_recursion();
compiler->set_limiting_recursion(true);
trace->Flush(compiler, this);
compiler->set_limiting_recursion(was_limiting);
return DONE;
}
bool RegExpNode::KeepRecursing(RegExpCompiler* compiler) {
return !compiler->limiting_recursion() &&
compiler->recursion_depth() <= RegExpCompiler::kMaxRecursion;
}
int ActionNode::EatsAtLeast(int still_to_find,
int budget,
bool not_at_start) {
if (budget <= 0) return 0;
if (action_type_ == POSITIVE_SUBMATCH_SUCCESS) return 0; // Rewinds input!
return on_success()->EatsAtLeast(still_to_find,
budget - 1,
not_at_start);
}
void ActionNode::FillInBMInfo(Isolate* isolate, int offset, int budget,
BoyerMooreLookahead* bm, bool not_at_start) {
if (action_type_ != POSITIVE_SUBMATCH_SUCCESS) {
on_success()->FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start);
}
SaveBMInfo(bm, not_at_start, offset);
}
int AssertionNode::EatsAtLeast(int still_to_find,
int budget,
bool not_at_start) {
if (budget <= 0) 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 (assertion_type() == AT_START && not_at_start) return still_to_find;
return on_success()->EatsAtLeast(still_to_find,
budget - 1,
not_at_start);
}
void AssertionNode::FillInBMInfo(Isolate* isolate, int offset, int budget,
BoyerMooreLookahead* bm, bool not_at_start) {
// Match the behaviour of EatsAtLeast on this node.
if (assertion_type() == AT_START && not_at_start) return;
on_success()->FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start);
SaveBMInfo(bm, not_at_start, offset);
}
int BackReferenceNode::EatsAtLeast(int still_to_find,
int budget,
bool not_at_start) {
if (read_backward()) return 0;
if (budget <= 0) return 0;
return on_success()->EatsAtLeast(still_to_find,
budget - 1,
not_at_start);
}
int TextNode::EatsAtLeast(int still_to_find,
int budget,
bool not_at_start) {
if (read_backward()) return 0;
int answer = Length();
if (answer >= still_to_find) return answer;
if (budget <= 0) return answer;
// We are not at start after this node so we set the last argument to 'true'.
return answer + on_success()->EatsAtLeast(still_to_find - answer,
budget - 1,
true);
}
int NegativeLookaroundChoiceNode::EatsAtLeast(int still_to_find, int budget,
bool not_at_start) {
if (budget <= 0) return 0;
// Alternative 0 is the negative lookahead, alternative 1 is what comes
// afterwards.
RegExpNode* node = alternatives_->at(1).node();
return node->EatsAtLeast(still_to_find, budget - 1, not_at_start);
}
void NegativeLookaroundChoiceNode::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 budget,
RegExpNode* ignore_this_node,
bool not_at_start) {
if (budget <= 0) return 0;
int min = 100;
int choice_count = alternatives_->length();
budget = (budget - 1) / choice_count;
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, budget, not_at_start);
if (node_eats_at_least < min) min = node_eats_at_least;
if (min == 0) return 0;
}
return min;
}
int LoopChoiceNode::EatsAtLeast(int still_to_find,
int budget,
bool not_at_start) {
return EatsAtLeastHelper(still_to_find,
budget - 1,
loop_node_,
not_at_start);
}
int ChoiceNode::EatsAtLeast(int still_to_find,
int budget,
bool not_at_start) {
return EatsAtLeastHelper(still_to_find, budget, nullptr, 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::kMaxOneByteCharCode;
} 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::kMaxOneByteCharCode) != 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* bounds_check_trace,
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_VALUE);
if (details->cannot_match()) return false;
if (!details->Rationalize(compiler->one_byte())) return false;
DCHECK(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()) {
DCHECK(trace->cp_offset() == bounds_check_trace->cp_offset());
// We are attempting to preload the minimum number of characters
// any choice would eat, so if the bounds check fails, then none of the
// choices can succeed, so we can just immediately backtrack, rather
// than go to the next choice.
assembler->LoadCurrentCharacter(trace->cp_offset(),
bounds_check_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->one_byte()) {
char_mask = String::kMaxOneByteCharCode;
} else {
char_mask = String::kMaxUtf16CodeUnit;
}
if ((mask & char_mask) == char_mask) need_mask = false;
mask &= char_mask;
} else {
// For 2-character preloads in one-byte mode or 1-character preloads in
// two-byte mode we also use a 16 bit load with zero extend.
static const uint32_t kTwoByteMask = 0xFFFF;
static const uint32_t kFourByteMask = 0xFFFFFFFF;
if (details->characters() == 2 && compiler->one_byte()) {
if ((mask & kTwoByteMask) == kTwoByteMask) need_mask = false;
} else if (details->characters() == 1 && !compiler->one_byte()) {
if ((mask & kTwoByteMask) == kTwoByteMask) need_mask = false;
} else {
if (mask == kFourByteMask) 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) {
// Do not collect any quick check details if the text node reads backward,
// since it reads in the opposite direction than we use for quick checks.
if (read_backward()) return;
Isolate* isolate = compiler->macro_assembler()->isolate();
DCHECK(characters_filled_in < details->characters());
int characters = details->characters();
int char_mask;
if (compiler->one_byte()) {
char_mask = String::kMaxOneByteCharCode;
} else {
char_mask = String::kMaxUtf16CodeUnit;
}
for (int k = 0; k < elements()->length(); k++) {
TextElement elm = elements()->at(k);
if (elm.text_type() == TextElement::ATOM) {
Vector<const uc16> quarks = elm.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 (elm.atom()->ignore_case()) {
unibrow::uchar chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
int length = GetCaseIndependentLetters(isolate, c,
compiler->one_byte(), chars);
if (length == 0) {
// This can happen because all case variants are non-Latin1, but we
// know the input is Latin1.
details->set_cannot_match();
pos->determines_perfectly = false;
return;
}
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.
if (c > char_mask) {
details->set_cannot_match();
pos->determines_perfectly = false;
return;
}
pos->mask = char_mask;
pos->value = c;
pos->determines_perfectly = true;
}
characters_filled_in++;
DCHECK(characters_filled_in <= details->characters());
if (characters_filled_in == details->characters()) {
return;
}
}
} else {
QuickCheckDetails::Position* pos =
details->positions(characters_filled_in);
RegExpCharacterClass* tree = elm.char_class();
ZoneList<CharacterRange>* ranges = tree->ranges(zone());
DCHECK(!ranges->is_empty());
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++;
DCHECK(characters_filled_in <= details->characters());
if (characters_filled_in == details->characters()) {
return;
}
}
}
DCHECK(characters_filled_in != details->characters());
if (!details->cannot_match()) {
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 one_byte) {
if (by >= characters_ || by < 0) {
DCHECK_IMPLIES(by < 0, characters_ == 0);
Clear();
return;
}
DCHECK_LE(characters_ - by, 4);
DCHECK_LE(characters_, 4);
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) {
DCHECK(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) {
DCHECK(!info->visited);
info->visited = true;
}
~VisitMarker() {
info_->visited = false;
}
private:
NodeInfo* info_;
};
RegExpNode* SeqRegExpNode::FilterOneByte(int depth) {
if (info()->replacement_calculated) return replacement();
if (depth < 0) return this;
DCHECK(!info()->visited);
VisitMarker marker(info());
return FilterSuccessor(depth - 1);
}
RegExpNode* SeqRegExpNode::FilterSuccessor(int depth) {
RegExpNode* next = on_success_->FilterOneByte(depth - 1);
if (next == nullptr) return set_replacement(nullptr);
on_success_ = next;
return set_replacement(this);
}
// We need to check for the following characters: 0x39C 0x3BC 0x178.
static inline bool RangeContainsLatin1Equivalents(CharacterRange range) {
// TODO(dcarney): this could be a lot more efficient.
return range.Contains(0x039C) || range.Contains(0x03BC) ||
range.Contains(0x0178);
}
static bool RangesContainLatin1Equivalents(ZoneList<CharacterRange>* ranges) {
for (int i = 0; i < ranges->length(); i++) {
// TODO(dcarney): this could be a lot more efficient.
if (RangeContainsLatin1Equivalents(ranges->at(i))) return true;
}
return false;
}
RegExpNode* TextNode::FilterOneByte(int depth) {
if (info()->replacement_calculated) return replacement();
if (depth < 0) return this;
DCHECK(!info()->visited);
VisitMarker marker(info());
int element_count = elements()->length();
for (int i = 0; i < element_count; i++) {
TextElement elm = elements()->at(i);
if (elm.text_type() == TextElement::ATOM) {
Vector<const uc16> quarks = elm.atom()->data();
for (int j = 0; j < quarks.length(); j++) {
uint16_t c = quarks[j];
if (c <= String::kMaxOneByteCharCode) continue;
if (!IgnoreCase(elm.atom()->flags())) return set_replacement(nullptr);
// Here, we need to check for characters whose upper and lower cases
// are outside the Latin-1 range.
uint16_t converted = unibrow::Latin1::ConvertNonLatin1ToLatin1(c);
// Character is outside Latin-1 completely
if (converted == 0) return set_replacement(nullptr);
// Convert quark to Latin-1 in place.
uint16_t* copy = const_cast<uint16_t*>(quarks.start());
copy[j] = converted;
}
} else {
DCHECK(elm.text_type() == TextElement::CHAR_CLASS);
RegExpCharacterClass* cc = elm.char_class();
ZoneList<CharacterRange>* ranges = cc->ranges(zone());
CharacterRange::Canonicalize(ranges);
// Now they are in order so we only need to look at the first.
int range_count = ranges->length();
if (cc->is_negated()) {
if (range_count != 0 &&
ranges->at(0).from() == 0 &&
ranges->at(0).to() >= String::kMaxOneByteCharCode) {
// This will be handled in a later filter.
if (IgnoreCase(cc->flags()) && RangesContainLatin1Equivalents(ranges))
continue;
return set_replacement(nullptr);
}
} else {
if (range_count == 0 ||
ranges->at(0).from() > String::kMaxOneByteCharCode) {
// This will be handled in a later filter.
if (IgnoreCase(cc->flags()) && RangesContainLatin1Equivalents(ranges))
continue;
return set_replacement(nullptr);
}
}
}
}
return FilterSuccessor(depth - 1);
}
RegExpNode* LoopChoiceNode::FilterOneByte(int depth) {
if (info()->replacement_calculated) return replacement();
if (depth < 0) return this;
if (info()->visited) return this;
{
VisitMarker marker(info());
RegExpNode* continue_replacement = continue_node_->FilterOneByte(depth - 1);
// If we can't continue after the loop then there is no sense in doing the
// loop.
if (continue_replacement == nullptr) return set_replacement(nullptr);
}
return ChoiceNode::FilterOneByte(depth - 1);
}
RegExpNode* ChoiceNode::FilterOneByte(int depth) {
if (info()->replacement_calculated) return replacement();
if (depth < 0) return this;
if (info()->visited) return this;
VisitMarker marker(info());
int choice_count = alternatives_->length();
for (int i = 0; i < choice_count; i++) {
GuardedAlternative alternative = alternatives_->at(i);
if (alternative.guards() != nullptr &&
alternative.guards()->length() != 0) {
set_replacement(this);
return this;
}
}
int surviving = 0;
RegExpNode* survivor = nullptr;
for (int i = 0; i < choice_count; i++) {
GuardedAlternative alternative = alternatives_->at(i);
RegExpNode* replacement = alternative.node()->FilterOneByte(depth - 1);
DCHECK(replacement != this); // No missing EMPTY_MATCH_CHECK.
if (replacement != nullptr) {
alternatives_->at(i).set_node(replacement);
surviving++;
survivor = replacement;
}
}
if (surviving <