Google Git
Sign in
cobalt / cobalt / da5c4f09a295755b8923113581bebdc12153e6dd / . / src / third_party / mozjs-45 / js / src / irregexp / RegExpEngine.cpp
blob: b84accb4786a52385f6463b538c6c6c7c08eddd6 [file] [log] [blame]
/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 4 -*-
* vim: set ts=8 sts=4 et sw=4 tw=99: */
// Copyright 2012 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 "irregexp/RegExpEngine.h"
#include "irregexp/NativeRegExpMacroAssembler.h"
#include "irregexp/RegExpMacroAssembler.h"
#include "jit/JitCommon.h"
using namespace js;
using namespace js::irregexp;
using mozilla::ArrayLength;
using mozilla::DebugOnly;
using mozilla::Maybe;
#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);
}
static const int kMaxLookaheadForBoyerMoore = 8;
RegExpNode::RegExpNode(LifoAlloc* alloc)
: replacement_(nullptr), trace_count_(0), alloc_(alloc)
{
bm_info_[0] = bm_info_[1] = nullptr;
}
// -------------------------------------------------------------------
// CharacterRange
// The '2' variant has inclusive from and exclusive to.
// This covers \s as defined in ECMA-262 5.1, 15.10.2.12,
// which include WhiteSpace (7.2) or LineTerminator (7.3) values.
static const int kSpaceRanges[] = { '\t', '\r' + 1, ' ', ' ' + 1,
0x00A0, 0x00A1, 0x1680, 0x1681, 0x180E, 0x180F, 0x2000, 0x200B,
0x2028, 0x202A, 0x202F, 0x2030, 0x205F, 0x2060, 0x3000, 0x3001,
0xFEFF, 0xFF00, 0x10000 };
static const int kSpaceRangeCount = ArrayLength(kSpaceRanges);
static const int kWordRanges[] = {
'0', '9' + 1, 'A', 'Z' + 1, '_', '_' + 1, 'a', 'z' + 1, 0x10000 };
static const int kWordRangeCount = ArrayLength(kWordRanges);
static const int kDigitRanges[] = { '0', '9' + 1, 0x10000 };
static const int kDigitRangeCount = ArrayLength(kDigitRanges);
static const int kSurrogateRanges[] = { 0xd800, 0xe000, 0x10000 };
static const int kSurrogateRangeCount = ArrayLength(kSurrogateRanges);
static const int kLineTerminatorRanges[] = { 0x000A, 0x000B, 0x000D, 0x000E,
0x2028, 0x202A, 0x10000 };
static const int kLineTerminatorRangeCount = ArrayLength(kLineTerminatorRanges);
static const int kMaxOneByteCharCode = 0xff;
static const int kMaxUtf16CodeUnit = 0xffff;
static char16_t
MaximumCharacter(bool ascii)
{
return ascii ? kMaxOneByteCharCode : kMaxUtf16CodeUnit;
}
static void
AddClass(const int* elmv, int elmc,
CharacterRangeVector* ranges)
{
elmc--;
MOZ_ASSERT(elmv[elmc] == 0x10000);
for (int i = 0; i < elmc; i += 2) {
MOZ_ASSERT(elmv[i] < elmv[i + 1]);
ranges->append(CharacterRange(elmv[i], elmv[i + 1] - 1));
}
}
static void
AddClassNegated(const int* elmv,
int elmc,
CharacterRangeVector* ranges)
{
elmc--;
MOZ_ASSERT(elmv[elmc] == 0x10000);
MOZ_ASSERT(elmv[0] != 0x0000);
MOZ_ASSERT(elmv[elmc-1] != kMaxUtf16CodeUnit);
char16_t last = 0x0000;
for (int i = 0; i < elmc; i += 2) {
MOZ_ASSERT(last <= elmv[i] - 1);
MOZ_ASSERT(elmv[i] < elmv[i + 1]);
ranges->append(CharacterRange(last, elmv[i] - 1));
last = elmv[i + 1];
}
ranges->append(CharacterRange(last, kMaxUtf16CodeUnit));
}
void
CharacterRange::AddClassEscape(LifoAlloc* alloc, char16_t type,
CharacterRangeVector* ranges)
{
switch (type) {
case 's':
AddClass(kSpaceRanges, kSpaceRangeCount, ranges);
break;
case 'S':
AddClassNegated(kSpaceRanges, kSpaceRangeCount, ranges);
break;
case 'w':
AddClass(kWordRanges, kWordRangeCount, ranges);
break;
case 'W':
AddClassNegated(kWordRanges, kWordRangeCount, ranges);
break;
case 'd':
AddClass(kDigitRanges, kDigitRangeCount, ranges);
break;
case 'D':
AddClassNegated(kDigitRanges, kDigitRangeCount, ranges);
break;
case '.':
AddClassNegated(kLineTerminatorRanges, kLineTerminatorRangeCount, ranges);
break;
// This is not a character range as defined by the spec but a
// convenient shorthand for a character class that matches any
// character.
case '*':
ranges->append(CharacterRange::Everything());
break;
// This is the set of characters matched by the $ and ^ symbols
// in multiline mode.
case 'n':
AddClass(kLineTerminatorRanges, kLineTerminatorRangeCount, ranges);
break;
default:
MOZ_CRASH("Bad character class escape");
}
}
// 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(0x39c) || range.Contains(0x3bc) || range.Contains(0x178);
}
static bool
RangesContainLatin1Equivalents(const CharacterRangeVector& ranges)
{
for (size_t i = 0; i < ranges.length(); i++) {
// TODO(dcarney): this could be a lot more efficient.
if (RangeContainsLatin1Equivalents(ranges[i]))
return true;
}
return false;
}
static const size_t kEcma262UnCanonicalizeMaxWidth = 4;
// Returns the number of characters in the equivalence class, omitting those
// that cannot occur in the source string if it is a one byte string.
static int
GetCaseIndependentLetters(char16_t character,
bool ascii_subject,
char16_t* letters)
{
const char16_t choices[] = {
character,
unicode::ToLowerCase(character),
unicode::ToUpperCase(character)
};
size_t count = 0;
for (size_t i = 0; i < ArrayLength(choices); i++) {
char16_t c = choices[i];
// The standard requires that non-ASCII characters cannot have ASCII
// character codes in their equivalence class, even though this
// situation occurs multiple times in the unicode tables.
static const unsigned kMaxAsciiCharCode = 127;
if (character > kMaxAsciiCharCode && c <= kMaxAsciiCharCode)
continue;
// Skip characters that can't appear in one byte strings.
if (ascii_subject && c > kMaxOneByteCharCode)
continue;
// Watch for duplicates.
bool found = false;
for (size_t j = 0; j < count; j++) {
if (letters[j] == c) {
found = true;
break;
}
}
if (found)
continue;
letters[count++] = c;
}
return count;
}
static char16_t
ConvertNonLatin1ToLatin1(char16_t c)
{
MOZ_ASSERT(c > kMaxOneByteCharCode);
switch (c) {
// This are equivalent characters in unicode.
case 0x39c:
case 0x3bc:
return 0xb5;
// This is an uppercase of a Latin-1 character
// outside of Latin-1.
case 0x178:
return 0xff;
}
return 0;
}
void
CharacterRange::AddCaseEquivalents(bool is_ascii, CharacterRangeVector* ranges)
{
char16_t bottom = from();
char16_t top = to();
if (is_ascii && !RangeContainsLatin1Equivalents(*this)) {
if (bottom > kMaxOneByteCharCode)
return;
if (top > kMaxOneByteCharCode)
top = kMaxOneByteCharCode;
}
for (char16_t c = bottom;; c++) {
char16_t chars[kEcma262UnCanonicalizeMaxWidth];
size_t length = GetCaseIndependentLetters(c, is_ascii, chars);
for (size_t i = 0; i < length; i++) {
char16_t other = chars[i];
if (other == c)
continue;
// Try to combine with an existing range.
bool found = false;
for (size_t i = 0; i < ranges->length(); i++) {
CharacterRange& range = (*ranges)[i];
if (range.Contains(other)) {
found = true;
break;
} else if (other == range.from() - 1) {
range.set_from(other);
found = true;
break;
} else if (other == range.to() + 1) {
range.set_to(other);
found = true;
break;
}
}
if (!found)
ranges->append(CharacterRange::Singleton(other));
}
if (c == top)
break;
}
}
static bool
CompareInverseRanges(const CharacterRangeVector& ranges, const int* special_class, size_t length)
{
length--; // Remove final 0x10000.
MOZ_ASSERT(special_class[length] == 0x10000);
MOZ_ASSERT(ranges.length() != 0);
MOZ_ASSERT(length != 0);
MOZ_ASSERT(special_class[0] != 0);
if (ranges.length() != (length >> 1) + 1)
return false;
CharacterRange range = ranges[0];
if (range.from() != 0)
return false;
for (size_t i = 0; i < length; i += 2) {
if (special_class[i] != (range.to() + 1))
return false;
range = ranges[(i >> 1) + 1];
if (special_class[i+1] != range.from())
return false;
}
if (range.to() != 0xffff)
return false;
return true;
}
static bool
CompareRanges(const CharacterRangeVector& ranges, const int* special_class, size_t length)
{
length--; // Remove final 0x10000.
MOZ_ASSERT(special_class[length] == 0x10000);
if (ranges.length() * 2 != length)
return false;
for (size_t i = 0; i < length; i += 2) {
CharacterRange range = ranges[i >> 1];
if (range.from() != special_class[i] || range.to() != special_class[i + 1] - 1)
return false;
}
return true;
}
bool
RegExpCharacterClass::is_standard(LifoAlloc* alloc)
{
// TODO(lrn): Remove need for this function, by not throwing away information
// along the way.
if (is_negated_)
return false;
if (set_.is_standard())
return true;
if (CompareRanges(set_.ranges(alloc), kSpaceRanges, kSpaceRangeCount)) {
set_.set_standard_set_type('s');
return true;
}
if (CompareInverseRanges(set_.ranges(alloc), kSpaceRanges, kSpaceRangeCount)) {
set_.set_standard_set_type('S');
return true;
}
if (CompareInverseRanges(set_.ranges(alloc),
kLineTerminatorRanges,
kLineTerminatorRangeCount)) {
set_.set_standard_set_type('.');
return true;
}
if (CompareRanges(set_.ranges(alloc),
kLineTerminatorRanges,
kLineTerminatorRangeCount)) {
set_.set_standard_set_type('n');
return true;
}
if (CompareRanges(set_.ranges(alloc), kWordRanges, kWordRangeCount)) {
set_.set_standard_set_type('w');
return true;
}
if (CompareInverseRanges(set_.ranges(alloc), kWordRanges, kWordRangeCount)) {
set_.set_standard_set_type('W');
return true;
}
return false;
}
bool
CharacterRange::IsCanonical(const CharacterRangeVector& ranges)
{
int n = ranges.length();
if (n <= 1)
return true;
int max = ranges[0].to();
for (int i = 1; i < n; i++) {
CharacterRange next_range = ranges[i];
if (next_range.from() <= max + 1)
return false;
max = next_range.to();
}
return true;
}
// Move a number of elements in a zonelist to another position
// in the same list. Handles overlapping source and target areas.
static
void MoveRanges(CharacterRangeVector& list, int from, int to, int count)
{
// Ranges are potentially overlapping.
if (from < to) {
for (int i = count - 1; i >= 0; i--)
list[to + i] = list[from + i];
} else {
for (int i = 0; i < count; i++)
list[to + i] = list[from + i];
}
}
static int
InsertRangeInCanonicalList(CharacterRangeVector& list,
int count,
CharacterRange insert)
{
// Inserts a range into list[0..count[, which must be sorted
// by from value and non-overlapping and non-adjacent, using at most
// list[0..count] for the result. Returns the number of resulting
// canonicalized ranges. Inserting a range may collapse existing ranges into
// fewer ranges, so the return value can be anything in the range 1..count+1.
char16_t from = insert.from();
char16_t to = insert.to();
int start_pos = 0;
int end_pos = count;
for (int i = count - 1; i >= 0; i--) {
CharacterRange current = list[i];
if (current.from() > to + 1) {
end_pos = i;
} else if (current.to() + 1 < from) {
start_pos = i + 1;
break;
}
}
// Inserted range overlaps, or is adjacent to, ranges at positions
// [start_pos..end_pos[. Ranges before start_pos or at or after end_pos are
// not affected by the insertion.
// If start_pos == end_pos, the range must be inserted before start_pos.
// if start_pos < end_pos, the entire range from start_pos to end_pos
// must be merged with the insert range.
if (start_pos == end_pos) {
// Insert between existing ranges at position start_pos.
if (start_pos < count) {
MoveRanges(list, start_pos, start_pos + 1, count - start_pos);
}
list[start_pos] = insert;
return count + 1;
}
if (start_pos + 1 == end_pos) {
// Replace single existing range at position start_pos.
CharacterRange to_replace = list[start_pos];
int new_from = Min(to_replace.from(), from);
int new_to = Max(to_replace.to(), to);
list[start_pos] = CharacterRange(new_from, new_to);
return count;
}
// Replace a number of existing ranges from start_pos to end_pos - 1.
// Move the remaining ranges down.
int new_from = Min(list[start_pos].from(), from);
int new_to = Max(list[end_pos - 1].to(), to);
if (end_pos < count) {
MoveRanges(list, end_pos, start_pos + 1, count - end_pos);
}
list[start_pos] = CharacterRange(new_from, new_to);
return count - (end_pos - start_pos) + 1;
}
void
CharacterRange::Canonicalize(CharacterRangeVector& character_ranges)
{
if (character_ranges.length() <= 1) return;
// Check whether ranges are already canonical (increasing, non-overlapping,
// non-adjacent).
int n = character_ranges.length();
int max = character_ranges[0].to();
int i = 1;
while (i < n) {
CharacterRange current = character_ranges[i];
if (current.from() <= max + 1) {
break;
}
max = current.to();
i++;
}
// Canonical until the i'th range. If that's all of them, we are done.
if (i == n) return;
// The ranges at index i and forward are not canonicalized. Make them so by
// doing the equivalent of insertion sort (inserting each into the previous
// list, in order).
// Notice that inserting a range can reduce the number of ranges in the
// result due to combining of adjacent and overlapping ranges.
int read = i; // Range to insert.
size_t num_canonical = i; // Length of canonicalized part of list.
do {
num_canonical = InsertRangeInCanonicalList(character_ranges,
num_canonical,
character_ranges[read]);
read++;
} while (read < n);
while (character_ranges.length() > num_canonical)
character_ranges.popBack();
MOZ_ASSERT(CharacterRange::IsCanonical(character_ranges));
}
// -------------------------------------------------------------------
// SeqRegExpNode
class VisitMarker
{
public:
explicit VisitMarker(NodeInfo* info)
: info_(info)
{
MOZ_ASSERT(!info->visited);
info->visited = true;
}
~VisitMarker() {
info_->visited = false;
}
private:
NodeInfo* info_;
};
bool
SeqRegExpNode::FillInBMInfo(int offset,
int budget,
BoyerMooreLookahead* bm,
bool not_at_start)
{
if (!bm->CheckOverRecursed())
return false;
if (!on_success_->FillInBMInfo(offset, budget - 1, bm, not_at_start))
return false;
if (offset == 0)
set_bm_info(not_at_start, bm);
return true;
}
RegExpNode*
SeqRegExpNode::FilterASCII(int depth, bool ignore_case)
{
if (info()->replacement_calculated)
return replacement();
if (depth < 0)
return this;
MOZ_ASSERT(!info()->visited);
VisitMarker marker(info());
return FilterSuccessor(depth - 1, ignore_case);
}
RegExpNode*
SeqRegExpNode::FilterSuccessor(int depth, bool ignore_case)
{
RegExpNode* next = on_success_->FilterASCII(depth - 1, ignore_case);
if (next == nullptr)
return set_replacement(nullptr);
on_success_ = next;
return set_replacement(this);
}
// -------------------------------------------------------------------
// ActionNode
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);
}
bool
ActionNode::FillInBMInfo(int offset,
int budget,
BoyerMooreLookahead* bm,
bool not_at_start)
{
if (!bm->CheckOverRecursed())
return false;
if (action_type_ == BEGIN_SUBMATCH) {
bm->SetRest(offset);
} else if (action_type_ != POSITIVE_SUBMATCH_SUCCESS) {
if (!on_success()->FillInBMInfo(offset, budget - 1, bm, not_at_start))
return false;
}
SaveBMInfo(bm, not_at_start, offset);
return true;
}
/* static */ ActionNode*
ActionNode::SetRegister(int reg,
int val,
RegExpNode* on_success)
{
ActionNode* result = on_success->alloc()->newInfallible<ActionNode>(SET_REGISTER, on_success);
result->data_.u_store_register.reg = reg;
result->data_.u_store_register.value = val;
return result;
}
/* static */ ActionNode*
ActionNode::IncrementRegister(int reg, RegExpNode* on_success)
{
ActionNode* result = on_success->alloc()->newInfallible<ActionNode>(INCREMENT_REGISTER, on_success);
result->data_.u_increment_register.reg = reg;
return result;
}
/* static */ ActionNode*
ActionNode::StorePosition(int reg, bool is_capture, RegExpNode* on_success)
{
ActionNode* result = on_success->alloc()->newInfallible<ActionNode>(STORE_POSITION, on_success);
result->data_.u_position_register.reg = reg;
result->data_.u_position_register.is_capture = is_capture;
return result;
}
/* static */ ActionNode*
ActionNode::ClearCaptures(Interval range, RegExpNode* on_success)
{
ActionNode* result = on_success->alloc()->newInfallible<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;
}
/* static */ ActionNode*
ActionNode::BeginSubmatch(int stack_pointer_reg, int position_reg, RegExpNode* on_success)
{
ActionNode* result = on_success->alloc()->newInfallible<ActionNode>(BEGIN_SUBMATCH, on_success);
result->data_.u_submatch.stack_pointer_register = stack_pointer_reg;
result->data_.u_submatch.current_position_register = position_reg;
return result;
}
/* static */ ActionNode*
ActionNode::PositiveSubmatchSuccess(int stack_pointer_reg,
int restore_reg,
int clear_capture_count,
int clear_capture_from,
RegExpNode* on_success)
{
ActionNode* result = on_success->alloc()->newInfallible<ActionNode>(POSITIVE_SUBMATCH_SUCCESS, on_success);
result->data_.u_submatch.stack_pointer_register = stack_pointer_reg;
result->data_.u_submatch.current_position_register = restore_reg;
result->data_.u_submatch.clear_register_count = clear_capture_count;
result->data_.u_submatch.clear_register_from = clear_capture_from;
return result;
}
/* static */ ActionNode*
ActionNode::EmptyMatchCheck(int start_register,
int repetition_register,
int repetition_limit,
RegExpNode* on_success)
{
ActionNode* result = on_success->alloc()->newInfallible<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;
}
// -------------------------------------------------------------------
// TextNode
int
TextNode::EatsAtLeast(int still_to_find, int budget, bool not_at_start)
{
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
TextNode::GreedyLoopTextLength()
{
TextElement elm = elements()[elements().length() - 1];
return elm.cp_offset() + elm.length();
}
RegExpNode*
TextNode::FilterASCII(int depth, bool ignore_case)
{
if (info()->replacement_calculated)
return replacement();
if (depth < 0)
return this;
MOZ_ASSERT(!info()->visited);
VisitMarker marker(info());
int element_count = elements().length();
for (int i = 0; i < element_count; i++) {
TextElement elm = elements()[i];
if (elm.text_type() == TextElement::ATOM) {
CharacterVector& quarks = const_cast<CharacterVector&>(elm.atom()->data());
for (size_t j = 0; j < quarks.length(); j++) {
uint16_t c = quarks[j];
if (c <= kMaxOneByteCharCode)
continue;
if (!ignore_case)
return set_replacement(nullptr);
// Here, we need to check for characters whose upper and lower cases
// are outside the Latin-1 range.
char16_t converted = ConvertNonLatin1ToLatin1(c);
if (converted == 0) {
// Character is outside Latin-1 completely
return set_replacement(nullptr);
}
// Convert quark to Latin-1 in place.
quarks[j] = converted;
}
} else {
MOZ_ASSERT(elm.text_type() == TextElement::CHAR_CLASS);
RegExpCharacterClass* cc = elm.char_class();
CharacterRangeVector& ranges = cc->ranges(alloc());
if (!CharacterRange::IsCanonical(ranges))
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[0].from() == 0 &&
ranges[0].to() >= kMaxOneByteCharCode)
{
// This will be handled in a later filter.
if (ignore_case && RangesContainLatin1Equivalents(ranges))
continue;
return set_replacement(nullptr);
}
} else {
if (range_count == 0 ||
ranges[0].from() > kMaxOneByteCharCode)
{
// This will be handled in a later filter.
if (ignore_case && RangesContainLatin1Equivalents(ranges))
continue;
return set_replacement(nullptr);
}
}
}
}
return FilterSuccessor(depth - 1, ignore_case);
}
void
TextNode::CalculateOffsets()
{
int element_count = elements().length();
// Set up the offsets of the elements relative to the start. This is a fixed
// quantity since a TextNode can only contain fixed-width things.
int cp_offset = 0;
for (int i = 0; i < element_count; i++) {
TextElement& elm = elements()[i];
elm.set_cp_offset(cp_offset);
cp_offset += elm.length();
}
}
void TextNode::MakeCaseIndependent(bool is_ascii)
{
int element_count = elements().length();
for (int i = 0; i < element_count; i++) {
TextElement elm = elements()[i];
if (elm.text_type() == TextElement::CHAR_CLASS) {
RegExpCharacterClass* cc = elm.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(alloc()))
continue;
CharacterRangeVector& ranges = cc->ranges(alloc());
int range_count = ranges.length();
for (int j = 0; j < range_count; j++)
ranges[j].AddCaseEquivalents(is_ascii, &ranges);
}
}
}
// -------------------------------------------------------------------
// AssertionNode
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);
}
bool
AssertionNode::FillInBMInfo(int offset, int budget, BoyerMooreLookahead* bm, bool not_at_start)
{
if (!bm->CheckOverRecursed())
return false;
// Match the behaviour of EatsAtLeast on this node.
if (assertion_type() == AT_START && not_at_start)
return true;
if (!on_success()->FillInBMInfo(offset, budget - 1, bm, not_at_start))
return false;
SaveBMInfo(bm, not_at_start, offset);
return true;
}
// -------------------------------------------------------------------
// BackReferenceNode
int
BackReferenceNode::EatsAtLeast(int still_to_find, int budget, bool not_at_start)
{
if (budget <= 0)
return 0;
return on_success()->EatsAtLeast(still_to_find, budget - 1, not_at_start);
}
bool
BackReferenceNode::FillInBMInfo(int offset, int budget, BoyerMooreLookahead* bm, bool not_at_start)
{
// Working out the set of characters that a backreference can match is too
// hard, so we just say that any character can match.
bm->SetRest(offset);
SaveBMInfo(bm, not_at_start, offset);
return true;
}
// -------------------------------------------------------------------
// ChoiceNode
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;
size_t choice_count = alternatives().length();
budget = (budget - 1) / choice_count;
for (size_t i = 0; i < choice_count; i++) {
RegExpNode* node = alternatives()[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
ChoiceNode::EatsAtLeast(int still_to_find, int budget, bool not_at_start)
{
return EatsAtLeastHelper(still_to_find,
budget,
nullptr,
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();
MOZ_ASSERT(choice_count > 0);
alternatives()[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()[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);
}
}
bool
ChoiceNode::FillInBMInfo(int offset,
int budget,
BoyerMooreLookahead* bm,
bool not_at_start)
{
if (!bm->CheckOverRecursed())
return false;
const GuardedAlternativeVector& alts = alternatives();
budget = (budget - 1) / alts.length();
for (size_t i = 0; i < alts.length(); i++) {
const GuardedAlternative& alt = alts[i];
if (alt.guards() != nullptr && alt.guards()->length() != 0) {
bm->SetRest(offset); // Give up trying to fill in info.
SaveBMInfo(bm, not_at_start, offset);
return true;
}
if (!alt.node()->FillInBMInfo(offset, budget, bm, not_at_start))
return false;
}
SaveBMInfo(bm, not_at_start, offset);
return true;
}
RegExpNode*
ChoiceNode::FilterASCII(int depth, bool ignore_case)
{
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++) {
const GuardedAlternative alternative = alternatives()[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()[i];
RegExpNode* replacement =
alternative.node()->FilterASCII(depth - 1, ignore_case);
MOZ_ASSERT(replacement != this); // No missing EMPTY_MATCH_CHECK.
if (replacement != nullptr) {
alternatives()[i].set_node(replacement);
surviving++;
survivor = replacement;
}
}
if (surviving < 2)
return set_replacement(survivor);
set_replacement(this);
if (surviving == choice_count)
return this;
// Only some of the nodes survived the filtering. We need to rebuild the
// alternatives list.
GuardedAlternativeVector new_alternatives(*alloc());
new_alternatives.reserve(surviving);
for (int i = 0; i < choice_count; i++) {
RegExpNode* replacement =
alternatives()[i].node()->FilterASCII(depth - 1, ignore_case);
if (replacement != nullptr) {
alternatives()[i].set_node(replacement);
AutoEnterOOMUnsafeRegion oomUnsafe;
if (!new_alternatives.append(alternatives()[i]))
oomUnsafe.crash("ChoiceNode::FilterASCII");
}
}
alternatives_ = Move(new_alternatives);
return this;
}
// -------------------------------------------------------------------
// NegativeLookaheadChoiceNode
bool
NegativeLookaheadChoiceNode::FillInBMInfo(int offset,
int budget,
BoyerMooreLookahead* bm,
bool not_at_start)
{
if (!bm->CheckOverRecursed())
return false;
if (!alternatives()[1].node()->FillInBMInfo(offset, budget - 1, bm, not_at_start))
return false;
if (offset == 0)
set_bm_info(not_at_start, bm);
return true;
}
int
NegativeLookaheadChoiceNode::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()[1].node();
return node->EatsAtLeast(still_to_find, budget - 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()[1].node();
return node->GetQuickCheckDetails(details, compiler, filled_in, not_at_start);
}
RegExpNode*
NegativeLookaheadChoiceNode::FilterASCII(int depth, bool ignore_case)
{
if (info()->replacement_calculated)
return replacement();
if (depth < 0)
return this;
if (info()->visited)
return this;
VisitMarker marker(info());
// Alternative 0 is the negative lookahead, alternative 1 is what comes
// afterwards.
RegExpNode* node = alternatives()[1].node();
RegExpNode* replacement = node->FilterASCII(depth - 1, ignore_case);
if (replacement == nullptr)
return set_replacement(nullptr);
alternatives()[1].set_node(replacement);
RegExpNode* neg_node = alternatives()[0].node();
RegExpNode* neg_replacement = neg_node->FilterASCII(depth - 1, ignore_case);
// If the negative lookahead is always going to fail then
// we don't need to check it.
if (neg_replacement == nullptr)
return set_replacement(replacement);
alternatives()[0].set_node(neg_replacement);
return set_replacement(this);
}
// -------------------------------------------------------------------
// LoopChoiceNode
void
GuardedAlternative::AddGuard(LifoAlloc* alloc, Guard* guard)
{
if (guards_ == nullptr)
guards_ = alloc->newInfallible<GuardVector>(*alloc);
guards_->append(guard);
}
void
LoopChoiceNode::AddLoopAlternative(GuardedAlternative alt)
{
MOZ_ASSERT(loop_node_ == nullptr);
AddAlternative(alt);
loop_node_ = alt.node();
}
void
LoopChoiceNode::AddContinueAlternative(GuardedAlternative alt)
{
MOZ_ASSERT(continue_node_ == nullptr);
AddAlternative(alt);
continue_node_ = alt.node();
}
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);
}
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);
}
bool
LoopChoiceNode::FillInBMInfo(int offset,
int budget,
BoyerMooreLookahead* bm,
bool not_at_start)
{
if (body_can_be_zero_length_ || budget <= 0) {
bm->SetRest(offset);
SaveBMInfo(bm, not_at_start, offset);
return true;
}
if (!ChoiceNode::FillInBMInfo(offset, budget - 1, bm, not_at_start))
return false;
SaveBMInfo(bm, not_at_start, offset);
return true;
}
RegExpNode*
LoopChoiceNode::FilterASCII(int depth, bool ignore_case)
{
if (info()->replacement_calculated)
return replacement();
if (depth < 0)
return this;
if (info()->visited)
return this;
{
VisitMarker marker(info());
RegExpNode* continue_replacement =
continue_node_->FilterASCII(depth - 1, ignore_case);
// 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::FilterASCII(depth - 1, ignore_case);
}
// -------------------------------------------------------------------
// Analysis
void
Analysis::EnsureAnalyzed(RegExpNode* that)
{
JS_CHECK_RECURSION(cx, fail("Stack overflow"); return);
if (that->info()->been_analyzed || that->info()->being_analyzed)
return;
that->info()->being_analyzed = true;
that->Accept(this);
that->info()->being_analyzed = false;
that->info()->been_analyzed = true;
}
void
Analysis::VisitEnd(EndNode* that)
{
// nothing to do
}
void
Analysis::VisitText(TextNode* that)
{
if (ignore_case_)
that->MakeCaseIndependent(is_ascii_);
EnsureAnalyzed(that->on_success());
if (!has_failed()) {
that->CalculateOffsets();
}
}
void
Analysis::VisitAction(ActionNode* that)
{
RegExpNode* target = that->on_success();
EnsureAnalyzed(target);
if (!has_failed()) {
// If the next node is interested in what it follows then this node
// has to be interested too so it can pass the information on.
that->info()->AddFromFollowing(target->info());
}
}
void
Analysis::VisitChoice(ChoiceNode* that)
{
NodeInfo* info = that->info();
for (size_t i = 0; i < that->alternatives().length(); i++) {
RegExpNode* node = that->alternatives()[i].node();
EnsureAnalyzed(node);
if (has_failed()) return;
// Anything the following nodes need to know has to be known by
// this node also, so it can pass it on.
info->AddFromFollowing(node->info());
}
}
void
Analysis::VisitLoopChoice(LoopChoiceNode* that)
{
NodeInfo* info = that->info();
for (size_t i = 0; i < that->alternatives().length(); i++) {
RegExpNode* node = that->alternatives()[i].node();
if (node != that->loop_node()) {
EnsureAnalyzed(node);
if (has_failed()) return;
info->AddFromFollowing(node->info());
}
}
// Check the loop last since it may need the value of this node
// to get a correct result.
EnsureAnalyzed(that->loop_node());
if (!has_failed())
info->AddFromFollowing(that->loop_node()->info());
}
void
Analysis::VisitBackReference(BackReferenceNode* that)
{
EnsureAnalyzed(that->on_success());
}
void
Analysis::VisitAssertion(AssertionNode* that)
{
EnsureAnalyzed(that->on_success());
}
// -------------------------------------------------------------------
// 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 RegExpAST.cpp.
// 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.
/* static */ TextElement
TextElement::Atom(RegExpAtom* atom)
{
return TextElement(ATOM, atom);
}
/* static */ 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;
}
MOZ_CRASH("Bad text type");
}
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) {
MOZ_ASSERT((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 irregexp::RegExpCompiler
{
public:
RegExpCompiler(JSContext* cx, LifoAlloc* alloc, int capture_count,
bool ignore_case, bool is_ascii, bool match_only);
int AllocateRegister() {
if (next_register_ >= RegExpMacroAssembler::kMaxRegister) {
reg_exp_too_big_ = true;
return next_register_;
}
return next_register_++;
}
RegExpCode Assemble(JSContext* cx,
RegExpMacroAssembler* assembler,
RegExpNode* start,
int capture_count);
inline void AddWork(RegExpNode* node) {
AutoEnterOOMUnsafeRegion oomUnsafe;
if (!work_list_.append(node))
oomUnsafe.crash("AddWork");
}
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_; }
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;
}
JSContext* cx() const { return cx_; }
LifoAlloc* alloc() const { return alloc_; }
static const int kNoRegister = -1;
private:
EndNode* accept_;
int next_register_;
Vector<RegExpNode*, 4, SystemAllocPolicy> work_list_;
int recursion_depth_;
RegExpMacroAssembler* macro_assembler_;
bool ignore_case_;
bool ascii_;
bool match_only_;
bool reg_exp_too_big_;
int current_expansion_factor_;
FrequencyCollator frequency_collator_;
JSContext* cx_;
LifoAlloc* alloc_;
};
class RecursionCheck
{
public:
explicit RecursionCheck(RegExpCompiler* compiler) : compiler_(compiler) {
compiler->IncrementRecursionDepth();
}
~RecursionCheck() { compiler_->DecrementRecursionDepth(); }
private:
RegExpCompiler* compiler_;
};
// 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(JSContext* cx, LifoAlloc* alloc, int capture_count,
bool ignore_case, bool ascii, bool match_only)
: next_register_(2 * (capture_count + 1)),
recursion_depth_(0),
ignore_case_(ignore_case),
ascii_(ascii),
match_only_(match_only),
reg_exp_too_big_(false),
current_expansion_factor_(1),
frequency_collator_(),
cx_(cx),
alloc_(alloc)
{
accept_ = alloc->newInfallible<EndNode>(alloc, EndNode::ACCEPT);
MOZ_ASSERT(next_register_ - 1 <= RegExpMacroAssembler::kMaxRegister);
}
RegExpCode
RegExpCompiler::Assemble(JSContext* cx,
RegExpMacroAssembler* assembler,
RegExpNode* start,
int capture_count)
{
macro_assembler_ = assembler;
macro_assembler_->set_slow_safe(false);
jit::Label fail;
macro_assembler_->PushBacktrack(&fail);
Trace new_trace;
start->Emit(this, &new_trace);
macro_assembler_->BindBacktrack(&fail);
macro_assembler_->Fail();
while (!work_list_.empty())
work_list_.popCopy()->Emit(this, &new_trace);
RegExpCode code = macro_assembler_->GenerateCode(cx, match_only_);
if (code.empty())
return RegExpCode();
if (reg_exp_too_big_) {
code.destroy();
JS_ReportError(cx, "regexp too big");
return RegExpCode();
}
return code;
}
template <typename CharT>
static void
SampleChars(FrequencyCollator* collator, const CharT* chars, size_t length)
{
// Sample some characters from the middle of the string.
static const int kSampleSize = 128;
int chars_sampled = 0;
int half_way = (int(length) - kSampleSize) / 2;
for (size_t i = Max(0, half_way);
i < length && chars_sampled < kSampleSize;
i++, chars_sampled++)
{
collator->CountCharacter(chars[i]);
}
}
static bool
IsNativeRegExpEnabled(JSContext* cx)
{
#if defined(JS_CODEGEN_NONE) || defined(COBALT_DISABLE_JIT)
return false;
#else
return cx->runtime()->options().nativeRegExp();
#endif
}
RegExpCode
irregexp::CompilePattern(JSContext* cx, RegExpShared* shared, RegExpCompileData* data,
HandleLinearString sample, bool is_global, bool ignore_case,
bool is_ascii, bool match_only, bool force_bytecode, bool sticky)
{
if ((data->capture_count + 1) * 2 - 1 > RegExpMacroAssembler::kMaxRegister) {
JS_ReportError(cx, "regexp too big");
return RegExpCode();
}
LifoAlloc& alloc = cx->tempLifoAlloc();
RegExpCompiler compiler(cx, &alloc, data->capture_count, ignore_case, is_ascii, match_only);
// Sample some characters from the middle of the string.
if (sample->hasLatin1Chars()) {
JS::AutoCheckCannotGC nogc;
SampleChars(compiler.frequency_collator(), sample->latin1Chars(nogc), sample->length());
} else {
JS::AutoCheckCannotGC nogc;
SampleChars(compiler.frequency_collator(), sample->twoByteChars(nogc), sample->length());
}
// Wrap the body of the regexp in capture #0.
RegExpNode* captured_body = RegExpCapture::ToNode(data->tree,
0,
&compiler,
compiler.accept());
RegExpNode* node = captured_body;
bool is_end_anchored = data->tree->IsAnchoredAtEnd();
bool is_start_anchored = sticky || data->tree->IsAnchoredAtStart();
int max_length = data->tree->max_match();
if (!is_start_anchored) {
// Add a .*? at the beginning, outside the body capture, unless
// this expression is anchored at the beginning.
RegExpNode* loop_node =
RegExpQuantifier::ToNode(0,
RegExpTree::kInfinity,
false,
alloc.newInfallible<RegExpCharacterClass>('*'),
&compiler,
captured_body,
data->contains_anchor);
if (data->contains_anchor) {
// Unroll loop once, to take care of the case that might start
// at the start of input.
ChoiceNode* first_step_node = alloc.newInfallible<ChoiceNode>(&alloc, 2);
RegExpNode* char_class =
alloc.newInfallible<TextNode>(alloc.newInfallible<RegExpCharacterClass>('*'), loop_node);
first_step_node->AddAlternative(GuardedAlternative(captured_body));
first_step_node->AddAlternative(GuardedAlternative(char_class));
node = first_step_node;
} else {
node = loop_node;
}
}
if (is_ascii) {
node = node->FilterASCII(RegExpCompiler::kMaxRecursion, ignore_case);
// Do it again to propagate the new nodes to places where they were not
// put because they had not been calculated yet.
if (node != nullptr) {
node = node->FilterASCII(RegExpCompiler::kMaxRecursion, ignore_case);
}
}
if (node == nullptr)
node = alloc.newInfallible<EndNode>(&alloc, EndNode::BACKTRACK);
Analysis analysis(cx, ignore_case, is_ascii);
analysis.EnsureAnalyzed(node);
if (analysis.has_failed()) {
JS_ReportError(cx, analysis.errorMessage());
return RegExpCode();
}
Maybe<jit::JitContext> ctx;
Maybe<NativeRegExpMacroAssembler> native_assembler;
Maybe<InterpretedRegExpMacroAssembler> interpreted_assembler;
RegExpMacroAssembler* assembler;
if (IsNativeRegExpEnabled(cx) && !force_bytecode) {
NativeRegExpMacroAssembler::Mode mode =
is_ascii ? NativeRegExpMacroAssembler::ASCII
: NativeRegExpMacroAssembler::CHAR16;
ctx.emplace(cx, (jit::TempAllocator*) nullptr);
native_assembler.emplace(&alloc, shared, cx->runtime(), mode, (data->capture_count + 1) * 2);
assembler = native_assembler.ptr();
} else {
interpreted_assembler.emplace(&alloc, shared, (data->capture_count + 1) * 2);
assembler = interpreted_assembler.ptr();
}
// Inserted here, instead of in Assembler, because it depends on information
// in the AST that isn't replicated in the Node structure.
static const int kMaxBacksearchLimit = 1024;
if (is_end_anchored &&
!is_start_anchored &&
max_length < kMaxBacksearchLimit) {
assembler->SetCurrentPositionFromEnd(max_length);
}
if (is_global) {
assembler->set_global_mode((data->tree->min_match() > 0)
? RegExpMacroAssembler::GLOBAL_NO_ZERO_LENGTH_CHECK
: RegExpMacroAssembler::GLOBAL);
}
return compiler.Assemble(cx, assembler, node, data->capture_count);
}
template <typename CharT>
RegExpRunStatus
irregexp::ExecuteCode(JSContext* cx, jit::JitCode* codeBlock, const CharT* chars, size_t start,
size_t length, MatchPairs* matches)
{
typedef void (*RegExpCodeSignature)(InputOutputData*);
InputOutputData data(chars, chars + length, start, matches);
RegExpCodeSignature function = reinterpret_cast<RegExpCodeSignature>(codeBlock->raw());
{
JS::AutoSuppressGCAnalysis nogc;
CALL_GENERATED_1(function, &data);
}
return (RegExpRunStatus) data.result;
}
template RegExpRunStatus
irregexp::ExecuteCode(JSContext* cx, jit::JitCode* codeBlock, const Latin1Char* chars, size_t start,
size_t length, MatchPairs* matches);
template RegExpRunStatus
irregexp::ExecuteCode(JSContext* cx, jit::JitCode* codeBlock, const char16_t* chars, size_t start,
size_t length, MatchPairs* matches);
// -------------------------------------------------------------------
// Tree to graph conversion
RegExpNode*
RegExpAtom::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
TextElementVector* elms =
compiler->alloc()->newInfallible<TextElementVector>(*compiler->alloc());
elms->append(TextElement::Atom(this));
return compiler->alloc()->newInfallible<TextNode>(elms, on_success);
}
RegExpNode*
RegExpText::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
return compiler->alloc()->newInfallible<TextNode>(&elements_, on_success);
}
RegExpNode*
RegExpCharacterClass::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
return compiler->alloc()->newInfallible<TextNode>(this, on_success);
}
RegExpNode*
RegExpDisjunction::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
const RegExpTreeVector& alternatives = this->alternatives();
size_t length = alternatives.length();
ChoiceNode* result = compiler->alloc()->newInfallible<ChoiceNode>(compiler->alloc(), length);
for (size_t i = 0; i < length; i++) {
GuardedAlternative alternative(alternatives[i]->ToNode(compiler, on_success));
result->AddAlternative(alternative);
}
return result;
}
RegExpNode*
RegExpQuantifier::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
return ToNode(min(),
max(),
is_greedy(),
body(),
compiler,
on_success);
}
// Scoped object to keep track of how much we unroll quantifier loops in the
// regexp graph generator.
class RegExpExpansionLimiter
{
public:
static const int kMaxExpansionFactor = 6;
RegExpExpansionLimiter(RegExpCompiler* compiler, int factor)
: compiler_(compiler),
saved_expansion_factor_(compiler->current_expansion_factor()),
ok_to_expand_(saved_expansion_factor_ <= kMaxExpansionFactor)
{
MOZ_ASSERT(factor > 0);
if (ok_to_expand_) {
if (factor > kMaxExpansionFactor) {
// Avoid integer overflow of the current expansion factor.
ok_to_expand_ = false;
compiler->set_current_expansion_factor(kMaxExpansionFactor + 1);
} else {
int new_factor = saved_expansion_factor_ * factor;
ok_to_expand_ = (new_factor <= kMaxExpansionFactor);
compiler->set_current_expansion_factor(new_factor);
}
}
}
~RegExpExpansionLimiter() {
compiler_->set_current_expansion_factor(saved_expansion_factor_);
}
bool ok_to_expand() { return ok_to_expand_; }
private:
RegExpCompiler* compiler_;
int saved_expansion_factor_;
bool ok_to_expand_;
};
/* static */ RegExpNode*
RegExpQuantifier::ToNode(int min,
int max,
bool is_greedy,
RegExpTree* body,
RegExpCompiler* compiler,
RegExpNode* on_success,
bool not_at_start /* = false */)
{
// x{f, t} becomes this:
//
// (r++)<-.
// | `
// | (x)
// v ^
// (r=0)-->(?)---/ [if r < t]
// |
// [if r >= f] \----> ...
//
// 15.10.2.5 RepeatMatcher algorithm.
// The parser has already eliminated the case where max is 0. In the case
// where max_match is zero the parser has removed the quantifier if min was
// > 0 and removed the atom if min was 0. See AddQuantifierToAtom.
// If we know that we cannot match zero length then things are a little
// simpler since we don't need to make the special zero length match check
// from step 2.1. If the min and max are small we can unroll a little in
// this case.
static const int kMaxUnrolledMinMatches = 3; // Unroll (foo)+ and (foo){3,}
static const int kMaxUnrolledMaxMatches = 3; // Unroll (foo)? and (foo){x,3}
if (max == 0)
return on_success; // This can happen due to recursion.
bool body_can_be_empty = (body->min_match() == 0);
int body_start_reg = RegExpCompiler::kNoRegister;
Interval capture_registers = body->CaptureRegisters();
bool needs_capture_clearing = !capture_registers.is_empty();
LifoAlloc* alloc = compiler->alloc();
if (body_can_be_empty) {
body_start_reg = compiler->AllocateRegister();
} else if (!needs_capture_clearing) {
// Only unroll if there are no captures and the body can't be
// empty.
{
RegExpExpansionLimiter limiter(compiler, min + ((max != min) ? 1 : 0));
if (min > 0 && min <= kMaxUnrolledMinMatches && limiter.ok_to_expand()) {
int new_max = (max == kInfinity) ? max : max - min;
// Recurse once to get the loop or optional matches after the fixed
// ones.
RegExpNode* answer = ToNode(0, new_max, is_greedy, body, compiler, on_success, true);
// Unroll the forced matches from 0 to min. This can cause chains of
// TextNodes (which the parser does not generate). These should be
// combined if it turns out they hinder good code generation.
for (int i = 0; i < min; i++)
answer = body->ToNode(compiler, answer);
return answer;
}
}
if (max <= kMaxUnrolledMaxMatches && min == 0) {
MOZ_ASSERT(max > 0); // Due to the 'if' above.
RegExpExpansionLimiter limiter(compiler, max);
if (limiter.ok_to_expand()) {
// Unroll the optional matches up to max.
RegExpNode* answer = on_success;
for (int i = 0; i < max; i++) {
ChoiceNode* alternation = alloc->newInfallible<ChoiceNode>(alloc, 2);
if (is_greedy) {
alternation->AddAlternative(GuardedAlternative(body->ToNode(compiler, answer)));
alternation->AddAlternative(GuardedAlternative(on_success));
} else {
alternation->AddAlternative(GuardedAlternative(on_success));
alternation->AddAlternative(GuardedAlternative(body->ToNode(compiler, answer)));
}
answer = alternation;
if (not_at_start) alternation->set_not_at_start();
}
return answer;
}
}
}
bool has_min = min > 0;
bool has_max = max < RegExpTree::kInfinity;
bool needs_counter = has_min || has_max;
int reg_ctr = needs_counter
? compiler->AllocateRegister()
: RegExpCompiler::kNoRegister;
LoopChoiceNode* center = alloc->newInfallible<LoopChoiceNode>(alloc, body->min_match() == 0);
if (not_at_start)
center->set_not_at_start();
RegExpNode* loop_return = needs_counter
? static_cast<RegExpNode*>(ActionNode::IncrementRegister(reg_ctr, center))
: static_cast<RegExpNode*>(center);
if (body_can_be_empty) {
// If the body can be empty we need to check if it was and then
// backtrack.
loop_return = ActionNode::EmptyMatchCheck(body_start_reg,
reg_ctr,
min,
loop_return);
}
RegExpNode* body_node = body->ToNode(compiler, loop_return);
if (body_can_be_empty) {
// If the body can be empty we need to store the start position
// so we can bail out if it was empty.
body_node = ActionNode::StorePosition(body_start_reg, false, body_node);
}
if (needs_capture_clearing) {
// Before entering the body of this loop we need to clear captures.
body_node = ActionNode::ClearCaptures(capture_registers, body_node);
}
GuardedAlternative body_alt(body_node);
if (has_max) {
Guard* body_guard = alloc->newInfallible<Guard>(reg_ctr, Guard::LT, max);
body_alt.AddGuard(alloc, body_guard);
}
GuardedAlternative rest_alt(on_success);
if (has_min) {
Guard* rest_guard = alloc->newInfallible<Guard>(reg_ctr, Guard::GEQ, min);
rest_alt.AddGuard(alloc, rest_guard);
}
if (is_greedy) {
center->AddLoopAlternative(body_alt);
center->AddContinueAlternative(rest_alt);
} else {
center->AddContinueAlternative(rest_alt);
center->AddLoopAlternative(body_alt);
}
if (needs_counter)
return ActionNode::SetRegister(reg_ctr, 0, center);
return center;
}
RegExpNode*
RegExpAssertion::ToNode(RegExpCompiler* compiler,
RegExpNode* on_success)
{
NodeInfo info;
LifoAlloc* alloc = compiler->alloc();
switch (assertion_type()) {
case START_OF_LINE:
return AssertionNode::AfterNewline(on_success);
case START_OF_INPUT:
return AssertionNode::AtStart(on_success);
case BOUNDARY:
return AssertionNode::AtBoundary(on_success);
case NON_BOUNDARY:
return AssertionNode::AtNonBoundary(on_success);
case END_OF_INPUT:
return AssertionNode::AtEnd(on_success);
case END_OF_LINE: {
// Compile $ in multiline regexps as an alternation with a positive
// lookahead in one side and an end-of-input on the other side.
// We need two registers for the lookahead.
int stack_pointer_register = compiler->AllocateRegister();
int position_register = compiler->AllocateRegister();
// The ChoiceNode to distinguish between a newline and end-of-input.
ChoiceNode* result = alloc->newInfallible<ChoiceNode>(alloc, 2);
// Create a newline atom.
CharacterRangeVector* newline_ranges = alloc->newInfallible<CharacterRangeVector>(*alloc);
CharacterRange::AddClassEscape(alloc, 'n', newline_ranges);
RegExpCharacterClass* newline_atom = alloc->newInfallible<RegExpCharacterClass>('n');
TextNode* newline_matcher =
alloc->newInfallible<TextNode>(newline_atom,
ActionNode::PositiveSubmatchSuccess(stack_pointer_register,
position_register,
0, // No captures inside.
-1, // Ignored if no captures.
on_success));
// Create an end-of-input matcher.
RegExpNode* end_of_line =
ActionNode::BeginSubmatch(stack_pointer_register, position_register, newline_matcher);
// Add the two alternatives to the ChoiceNode.
GuardedAlternative eol_alternative(end_of_line);
result->AddAlternative(eol_alternative);
GuardedAlternative end_alternative(AssertionNode::AtEnd(on_success));
result->AddAlternative(end_alternative);
return result;
}
default:
MOZ_CRASH("Bad assertion type");
}
return on_success;
}
RegExpNode*
RegExpBackReference::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
return compiler->alloc()->newInfallible<BackReferenceNode>(RegExpCapture::StartRegister(index()),
RegExpCapture::EndRegister(index()),
on_success);
}
RegExpNode*
RegExpEmpty::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
return on_success;
}
RegExpNode*
RegExpLookahead::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
int stack_pointer_register = compiler->AllocateRegister();
int position_register = compiler->AllocateRegister();
const int registers_per_capture = 2;
const int register_of_first_capture = 2;
int register_count = capture_count_ * registers_per_capture;
int register_start =
register_of_first_capture + capture_from_ * registers_per_capture;
if (is_positive()) {
RegExpNode* bodyNode =
body()->ToNode(compiler,
ActionNode::PositiveSubmatchSuccess(stack_pointer_register,
position_register,
register_count,
register_start,
on_success));
return ActionNode::BeginSubmatch(stack_pointer_register,
position_register,
bodyNode);
}
// We use a ChoiceNode for a negative lookahead because it has most of
// the characteristics we need. It has the body of the lookahead as its
// first alternative and the expression after the lookahead of the second
// alternative. If the first alternative succeeds then the
// NegativeSubmatchSuccess will unwind the stack including everything the
// choice node set up and backtrack. If the first alternative fails then
// the second alternative is tried, which is exactly the desired result
// for a negative lookahead. The NegativeLookaheadChoiceNode is a special
// ChoiceNode that knows to ignore the first exit when calculating quick
// checks.
LifoAlloc* alloc = compiler->alloc();
RegExpNode* success =
alloc->newInfallible<NegativeSubmatchSuccess>(alloc,
stack_pointer_register,
position_register,
register_count,
register_start);
GuardedAlternative body_alt(body()->ToNode(compiler, success));
ChoiceNode* choice_node =
alloc->newInfallible<NegativeLookaheadChoiceNode>(alloc, body_alt, GuardedAlternative(on_success));
return ActionNode::BeginSubmatch(stack_pointer_register,
position_register,
choice_node);
}
RegExpNode*
RegExpCapture::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
return ToNode(body(), index(), compiler, on_success);
}
/* static */ RegExpNode*
RegExpCapture::ToNode(RegExpTree* body,
int index,
RegExpCompiler* compiler,
RegExpNode* on_success)
{
int start_reg = RegExpCapture::StartRegister(index);
int end_reg = RegExpCapture::EndRegister(index);
RegExpNode* store_end = ActionNode::StorePosition(end_reg, true, on_success);
RegExpNode* body_node = body->ToNode(compiler, store_end);
return ActionNode::StorePosition(start_reg, true, body_node);
}
RegExpNode*
RegExpAlternative::ToNode(RegExpCompiler* compiler, RegExpNode* on_success)
{
const RegExpTreeVector& children = nodes();
RegExpNode* current = on_success;
for (int i = children.length() - 1; i >= 0; i--)
current = children[i]->ToNode(compiler, current);
return current;
}
// -------------------------------------------------------------------
// BoyerMooreLookahead
ContainedInLattice
irregexp::AddRange(ContainedInLattice containment,
const int* ranges,
int ranges_length,
Interval new_range)
{
MOZ_ASSERT((ranges_length & 1) == 1);
MOZ_ASSERT(ranges[ranges_length - 1] == kMaxUtf16CodeUnit + 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;
}
void
BoyerMoorePositionInfo::Set(int character)
{
SetInterval(Interval(character, character));
}
void
BoyerMoorePositionInfo::SetInterval(const Interval& interval)
{
s_ = AddRange(s_, kSpaceRanges, kSpaceRangeCount, interval);
w_ = AddRange(w_, kWordRanges, kWordRangeCount, interval);
d_ = AddRange(d_, kDigitRanges, kDigitRangeCount, interval);
surrogate_ =
AddRange(surrogate_, kSurrogateRanges, kSurrogateRangeCount, interval);
if (interval.to() - interval.from() >= kMapSize - 1) {
if (map_count_ != kMapSize) {
map_count_ = kMapSize;
for (int i = 0; i < kMapSize; i++)
map_[i] = true;
}
return;
}
for (int i = interval.from(); i <= interval.to(); i++) {
int mod_character = (i & kMask);
if (!map_[mod_character]) {
map_count_++;
map_[mod_character] = true;
}
if (map_count_ == kMapSize)
return;
}
}
void
BoyerMoorePositionInfo::SetAll()
{
s_ = w_ = d_ = kLatticeUnknown;
if (map_count_ != kMapSize) {
map_count_ = kMapSize;
for (int i = 0; i < kMapSize; i++)
map_[i] = true;
}
}
BoyerMooreLookahead::BoyerMooreLookahead(LifoAlloc* alloc, size_t length, RegExpCompiler* compiler)
: length_(length), compiler_(compiler), bitmaps_(*alloc)
{
max_char_ = MaximumCharacter(compiler->ascii());
bitmaps_.reserve(length);
for (size_t i = 0; i < length; i++)
bitmaps_.append(alloc->newInfallible<BoyerMoorePositionInfo>(alloc));
}
// Find the longest range of lookahead that has the fewest number of different
// characters that can occur at a given position. Since we are optimizing two
// different parameters at once this is a tradeoff.
bool BoyerMooreLookahead::FindWorthwhileInterval(int* from, int* to) {
int biggest_points = 0;
// If more than 32 characters out of 128 can occur it is unlikely that we can
// be lucky enough to step forwards much of the time.
const int kMaxMax = 32;
for (int max_number_of_chars = 4;
max_number_of_chars < kMaxMax;
max_number_of_chars *= 2) {
biggest_points =
FindBestInterval(max_number_of_chars, biggest_points, from, to);
}
if (biggest_points == 0) return false;
return true;
}
// Find the highest-points range between 0 and length_ where the character
// information is not too vague. 'Too vague' means that there are more than
// max_number_of_chars that can occur at this position. Calculates the number
// of points as the product of width-of-the-range and
// probability-of-finding-one-of-the-characters, where the probability is
// calculated using the frequency distribution of the sample subject string.
int
BoyerMooreLookahead::FindBestInterval(int max_number_of_chars, int old_biggest_points,
int* from, int* to)
{
int biggest_points = old_biggest_points;
static const int kSize = RegExpMacroAssembler::kTableSize;
for (int i = 0; i < length_; ) {
while (i < length_ && Count(i) > max_number_of_chars) i++;
if (i == length_) break;
int remembered_from = i;
bool union_map[kSize];
for (int j = 0; j < kSize; j++) union_map[j] = false;
while (i < length_ && Count(i) <= max_number_of_chars) {
BoyerMoorePositionInfo* map = bitmaps_[i];
for (int j = 0; j < kSize; j++) union_map[j] |= map->at(j);
i++;
}
int frequency = 0;
for (int j = 0; j < kSize; j++) {
if (union_map[j]) {
// Add 1 to the frequency to give a small per-character boost for
// the cases where our sampling is not good enough and many
// characters have a frequency of zero. This means the frequency
// can theoretically be up to 2*kSize though we treat it mostly as
// a fraction of kSize.
frequency += compiler_->frequency_collator()->Frequency(j) + 1;
}
}
// We use the probability of skipping times the distance we are skipping to
// judge the effectiveness of this. Actually we have a cut-off: By
// dividing by 2 we switch off the skipping if the probability of skipping
// is less than 50%. This is because the multibyte mask-and-compare
// skipping in quickcheck is more likely to do well on this case.
bool in_quickcheck_range = ((i - remembered_from < 4) ||
(compiler_->ascii() ? remembered_from <= 4 : remembered_from <= 2));
// Called 'probability' but it is only a rough estimate and can actually
// be outside the 0-kSize range.
int probability = (in_quickcheck_range ? kSize / 2 : kSize) - frequency;
int points = (i - remembered_from) * probability;
if (points > biggest_points) {
*from = remembered_from;
*to = i - 1;
biggest_points = points;
}
}
return biggest_points;
}
// Take all the characters that will not prevent a successful match if they
// occur in the subject string in the range between min_lookahead and
// max_lookahead (inclusive) measured from the current position. If the
// character at max_lookahead offset is not one of these characters, then we
// can safely skip forwards by the number of characters in the range.
int BoyerMooreLookahead::GetSkipTable(int min_lookahead,
int max_lookahead,
uint8_t* boolean_skip_table)
{
const int kSize = RegExpMacroAssembler::kTableSize;
const int kSkipArrayEntry = 0;
const int kDontSkipArrayEntry = 1;
for (int i = 0; i < kSize; i++)
boolean_skip_table[i] = kSkipArrayEntry;
int skip = max_lookahead + 1 - min_lookahead;
for (int i = max_lookahead; i >= min_lookahead; i--) {
BoyerMoorePositionInfo* map = bitmaps_[i];
for (int j = 0; j < kSize; j++) {
if (map->at(j))
boolean_skip_table[j] = kDontSkipArrayEntry;
}
}
return skip;
}
// See comment on the implementation of GetSkipTable.
bool
BoyerMooreLookahead::EmitSkipInstructions(RegExpMacroAssembler* masm)
{
const int kSize = RegExpMacroAssembler::kTableSize;
int min_lookahead = 0;
int max_lookahead = 0;
if (!FindWorthwhileInterval(&min_lookahead, &max_lookahead))
return false;
bool found_single_character = false;
int single_character = 0;
for (int i = max_lookahead; i >= min_lookahead; i--) {
BoyerMoorePositionInfo* map = bitmaps_[i];
if (map->map_count() > 1 ||
(found_single_character && map->map_count() != 0)) {
found_single_character = false;
break;
}
for (int j = 0; j < kSize; j++) {
if (map->at(j)) {
found_single_character = true;
single_character = j;
break;
}
}
}
int lookahead_width = max_lookahead + 1 - min_lookahead;
if (found_single_character && lookahead_width == 1 && max_lookahead < 3) {
// The mask-compare can probably handle this better.
return false;
}
if (found_single_character) {
jit::Label cont, again;
masm->Bind(&again);
masm->LoadCurrentCharacter(max_lookahead, &cont, true);
if (max_char_ > kSize) {
masm->CheckCharacterAfterAnd(single_character,
RegExpMacroAssembler::kTableMask,
&cont);
} else {
masm->CheckCharacter(single_character, &cont);
}
masm->AdvanceCurrentPosition(lookahead_width);
masm->JumpOrBacktrack(&again);
masm->Bind(&cont);
return true;
}
uint8_t* boolean_skip_table;
{
AutoEnterOOMUnsafeRegion oomUnsafe;
boolean_skip_table = static_cast<uint8_t*>(js_malloc(kSize));
if (!boolean_skip_table || !masm->shared->addTable(boolean_skip_table))
oomUnsafe.crash("Table malloc");
}
int skip_distance = GetSkipTable(min_lookahead, max_lookahead, boolean_skip_table);
MOZ_ASSERT(skip_distance != 0);
jit::Label cont, again;
masm->Bind(&again);
masm->LoadCurrentCharacter(max_lookahead, &cont, true);
masm->CheckBitInTable(boolean_skip_table, &cont);
masm->AdvanceCurrentPosition(skip_distance);
masm->JumpOrBacktrack(&again);
masm->Bind(&cont);
return true;
}
bool
BoyerMooreLookahead::CheckOverRecursed()
{
JS_CHECK_RECURSION(compiler()->cx(), compiler()->SetRegExpTooBig(); return false);
return true;
}
// -------------------------------------------------------------------
// Trace
bool Trace::DeferredAction::Mentions(int that)
{
if (action_type() == ActionNode::CLEAR_CAPTURES) {
Interval range = static_cast<DeferredClearCaptures*>(this)->range();
return range.Contains(that);
}
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)
{
MOZ_ASSERT(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;
}
return false;
}
}
return false;
}
int
Trace::FindAffectedRegisters(LifoAlloc* alloc, OutSet* affected_registers)
{
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(alloc, i);
if (range.to() > max_register) max_register = range.to();
} else {
affected_registers->Set(alloc, 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);
}
}
}
enum DeferredActionUndoType {
DEFER_IGNORE,
DEFER_RESTORE,
DEFER_CLEAR
};
void
Trace::PerformDeferredActions(LifoAlloc* alloc,
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).
DeferredActionUndoType undo_action = DEFER_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 != 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
// 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 = DEFER_RESTORE;
MOZ_ASSERT(store_position == -1);
MOZ_ASSERT(!clear);
break;
}
case ActionNode::INCREMENT_REGISTER:
if (!absolute) {
value++;
}
MOZ_ASSERT(store_position == -1);
MOZ_ASSERT(!clear);
undo_action = DEFER_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 = DEFER_IGNORE;
} else {
undo_action = pc->is_capture() ? DEFER_CLEAR : DEFER_RESTORE;
}
MOZ_ASSERT(!absolute);
MOZ_ASSERT(value == 0);
break;
}
case ActionNode::CLEAR_CAPTURES: {
// Since we're scanning in reverse order, if we've already
// set the position we have to ignore historically earlier
// clearing operations.
if (store_position == -1) {
clear = true;
}
undo_action = DEFER_RESTORE;
MOZ_ASSERT(!absolute);
MOZ_ASSERT(value == 0);
break;
}
default:
MOZ_CRASH("Bad action");
}
}
}
// Prepare for the undo-action (e.g., push if it's going to be popped).
if (undo_action == DEFER_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(alloc, reg);
} else if (undo_action == DEFER_CLEAR) {
registers_to_clear->Set(alloc, 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();
MOZ_ASSERT(!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(compiler->alloc(), &affected_registers);
OutSet registers_to_pop;
OutSet registers_to_clear;
PerformDeferredActions(compiler->alloc(),
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.
jit::Label undo;
assembler->PushBacktrack(&undo);
Trace new_state;
successor->Emit(compiler, &new_state);
// On backtrack we need to restore state.
assembler->BindBacktrack(&undo);
RestoreAffectedRegisters(assembler,
max_register,
registers_to_pop,
registers_to_clear);
if (backtrack() == nullptr) {
assembler->Backtrack();
} else {
assembler->PopCurrentPosition();
assembler->JumpOrBacktrack(backtrack());
}
}
void
Trace::InvalidateCurrentCharacter()
{
characters_preloaded_ = 0;
}
void
Trace::AdvanceCurrentPositionInTrace(int by, RegExpCompiler* compiler)
{
MOZ_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
OutSet::Set(LifoAlloc* alloc, unsigned value)
{
if (value < kFirstLimit) {
first_ |= (1 << value);
} else {
if (remaining_ == nullptr)
remaining_ = alloc->newInfallible<RemainingVector>(*alloc);
for (size_t i = 0; i < remaining().length(); i++) {
if (remaining()[i] == value)
return;
}
remaining().append(value);
}
}
bool
OutSet::Get(unsigned value)
{
if (value < kFirstLimit)
return (first_ & (1 << value)) != 0;
if (remaining_ == nullptr)
return false;
for (size_t i = 0; i < remaining().length(); i++) {
if (remaining()[i] == value)
return true;
}
return false;
}
// -------------------------------------------------------------------
// Graph emitting
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()->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->ReadBacktrackStackPointerFromRegister(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()->bound()) {
assembler->Bind(label());
}
switch (action_) {
case ACCEPT:
assembler->Succeed();
return;
case BACKTRACK:
assembler->JumpOrBacktrack(trace->backtrack());
return;
case NEGATIVE_SUBMATCH_SUCCESS:
// This case is handled in a different virtual method.
MOZ_CRASH("Bad action: NEGATIVE_SUBMATCH_SUCCESS");
}
MOZ_CRASH("Bad action");
}
// 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();
jit::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);
}
// Check for [0-9A-Z_a-z].
static void
EmitWordCheck(RegExpMacroAssembler* assembler,
jit::Label* word, jit::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 handle \b and \B (word-boundary or non-word-boundary).
void
AssertionNode::EmitBoundaryCheck(RegExpCompiler* compiler, Trace* trace)
{
RegExpMacroAssembler* assembler = compiler->macro_assembler();
Trace::TriBool next_is_word_character = Trace::UNKNOWN;
bool not_at_start = (trace->at_start() == Trace::FALSE_VALUE);
BoyerMooreLookahead* lookahead = bm_info(not_at_start);
if (lookahead == nullptr) {
int eats_at_least =
Min(kMaxLookaheadForBoyerMoore, EatsAtLeast(kMaxLookaheadForBoyerMoore,
kRecursionBudget,
not_at_start));
if (eats_at_least >= 1) {
BoyerMooreLookahead* bm =
alloc()->newInfallible<BoyerMooreLookahead>(alloc(), eats_at_least, compiler);
FillInBMInfo(0, kRecursionBudget, bm, not_at_start);
if (bm->at(0)->is_non_word())
next_is_word_character = Trace::FALSE_VALUE;
if (bm->at(0)->is_word()) next_is_word_character = Trace::TRUE_VALUE;
}
} else {
if (lookahead->at(0)->is_non_word())
next_is_word_character = Trace::FALSE_VALUE;
if (lookahead->at(0)->is_word())
next_is_word_character = Trace::TRUE_VALUE;
}
bool at_boundary = (assertion_type_ == AssertionNode::AT_BOUNDARY);
if (next_is_word_character == Trace::UNKNOWN) {
jit::Label before_non_word;
jit::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);
// Next character is not a word character.
assembler->Bind(&before_non_word);
jit::Label ok;
BacktrackIfPrevious(compiler, trace, at_boundary ? kIsNonWord : kIsWord);
assembler->JumpOrBacktrack(&ok);
assembler->Bind(&before_word);
BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord);
assembler->Bind(&ok);
} else if (next_is_word_character == Trace::TRUE_VALUE) {
BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord);
} else {
MOZ_ASSERT(next_is_word_character == Trace::FALSE_VALUE);