blob: a930e5cedf4eebfb0b98b55c297041c1d72e3575 [file] [log] [blame]
// Copyright (c) 2014, the Dart project authors. Please see the AUTHORS file
// for details. All rights reserved. Use of this source code is governed by a
// BSD-style license that can be found in the LICENSE file.
#include "vm/regexp.h"
#include <memory>
#include "platform/splay-tree-inl.h"
#include "platform/unicode.h"
#include "unicode/uniset.h"
#include "vm/dart_entry.h"
#include "vm/regexp_assembler.h"
#include "vm/regexp_assembler_bytecode.h"
#include "vm/regexp_ast.h"
#include "vm/symbols.h"
#include "vm/thread.h"
#include "vm/unibrow-inl.h"
#if !defined(DART_PRECOMPILED_RUNTIME)
#include "vm/regexp_assembler_ir.h"
#endif // !defined(DART_PRECOMPILED_RUNTIME)
#define Z (zone())
namespace dart {
// Default to generating optimized regexp code.
static constexpr bool kRegexpOptimization = true;
// More makes code generation slower, less makes V8 benchmark score lower.
static constexpr intptr_t kMaxLookaheadForBoyerMoore = 8;
ContainedInLattice AddRange(ContainedInLattice containment,
const int32_t* ranges,
intptr_t ranges_length,
Interval new_range) {
ASSERT((ranges_length & 1) == 1);
ASSERT(ranges[ranges_length - 1] == Utf::kMaxCodePoint + 1);
if (containment == kLatticeUnknown) return containment;
bool inside = false;
int32_t last = 0;
for (intptr_t i = 0; i < ranges_length;
inside = !inside, last = ranges[i], i++) {
// Consider the range from last to ranges[i].
// We haven't got to the new range yet.
if (ranges[i] <= new_range.from()) continue;
// New range is wholly inside last-ranges[i]. Note that new_range.to() is
// inclusive, but the values in ranges are not.
if (last <= new_range.from() && new_range.to() < ranges[i]) {
return Combine(containment, inside ? kLatticeIn : kLatticeOut);
}
return kLatticeUnknown;
}
return containment;
}
// -------------------------------------------------------------------
// 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
// IR code that is subsequently compiled to native code.
// The Irregexp regexp engine is structured in three steps.
// 1) The parser generates an abstract syntax tree. See regexp_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 IR instructions 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 maintains some state as it runs. This consists of the
// following elements:
//
// * The capture registers. Used for string captures.
// * Other registers. Used for counters etc.
// * The current position.
// * The stack of backtracking information. Used when a matching node
// fails to find a match and needs to try an alternative.
//
// Conceptual regular expression execution model:
//
// There is a simple conceptual model of regular expression execution
// which will be presented first. The actual code generated is a more
// efficient simulation of the simple conceptual model:
//
// * Choice nodes are implemented as follows:
// For each choice except the last {
// push current position
// push backtrack code location
// <generate code to test for choice>
// backtrack code location:
// pop current position
// }
// <generate code to test for last choice>
//
// * Actions nodes are generated as follows
// <push affected registers on backtrack stack>
// <generate code to perform action>
// push backtrack code location
// <generate code to test for following nodes>
// backtrack code location:
// <pop affected registers to restore their state>
// <pop backtrack location from stack and go to it>
//
// * Matching nodes are generated as follows:
// if input string matches at current position
// update current position
// <generate code to test for following nodes>
// else
// <pop backtrack location from stack and go to it>
//
// Thus it can be seen that the current position is saved and restored
// by the choice nodes, whereas the registers are saved and restored by
// by the action nodes that manipulate them.
//
// The other interesting aspect of this model is that nodes are generated
// at the point where they are needed by a recursive call to Emit(). If
// the node has already been code generated then the Emit() call will
// generate a jump to the previously generated code instead. In order to
// limit recursion it is possible for the Emit() function to put the node
// on a work list for later generation and instead generate a jump. The
// destination of the jump is resolved later when the code is generated.
//
// Actual regular expression code generation.
//
// Code generation is actually more complicated than the above. In order
// to improve the efficiency of the generated code some optimizations are
// performed
//
// * Choice nodes have 1-character lookahead.
// A choice node looks at the following character and eliminates some of
// the choices immediately based on that character. This is not yet
// implemented.
// * Simple greedy loops store reduced backtracking information.
// A quantifier like /.*foo/m will greedily match the whole input. It will
// then need to backtrack to a point where it can match "foo". The naive
// implementation of this would push each character position onto the
// backtracking stack, then pop them off one by one. This would use space
// proportional to the length of the input string. However since the "."
// can only match in one way and always has a constant length (in this case
// of 1) it suffices to store the current position on the top of the stack
// once. Matching now becomes merely incrementing the current position and
// backtracking becomes decrementing the current position and checking the
// result against the stored current position. This is faster and saves
// space.
// * The current state is virtualized.
// This is used to defer expensive operations until it is clear that they
// are needed and to generate code for a node more than once, allowing
// specialized an efficient versions of the code to be created. This is
// explained in the section below.
//
// Execution state virtualization.
//
// Instead of emitting code, nodes that manipulate the state can record their
// manipulation in an object called the Trace. The Trace object can record a
// current position offset, an optional backtrack code location on the top of
// the virtualized backtrack stack and some register changes. When a node is
// to be emitted it can flush the Trace or update it. Flushing the Trace
// will emit code to bring the actual state into line with the virtual state.
// Avoiding flushing the state can postpone some work (e.g. updates of capture
// registers). Postponing work can save time when executing the regular
// expression since it may be found that the work never has to be done as a
// failure to match can occur. In addition it is much faster to jump to a
// known backtrack code location than it is to pop an unknown backtrack
// location from the stack and jump there.
//
// The virtual state found in the Trace affects code generation. For example
// the virtual state contains the difference between the actual current
// position and the virtual current position, and matching code needs to use
// this offset to attempt a match in the correct location of the input
// string. Therefore code generated for a non-trivial trace is specialized
// to that trace. The code generator therefore has the ability to generate
// code for each node several times. In order to limit the size of the
// generated code there is an arbitrary limit on how many specialized sets of
// code may be generated for a given node. If the limit is reached, the
// trace is flushed and a generic version of the code for a node is emitted.
// This is subsequently used for that node. The code emitted for non-generic
// trace is not recorded in the node and so it cannot currently be reused in
// the event that code generation is requested for an identical trace.
void RegExpTree::AppendToText(RegExpText* text) {
UNREACHABLE();
}
void RegExpAtom::AppendToText(RegExpText* text) {
text->AddElement(TextElement::Atom(this));
}
void RegExpCharacterClass::AppendToText(RegExpText* text) {
text->AddElement(TextElement::CharClass(this));
}
void RegExpText::AppendToText(RegExpText* text) {
for (intptr_t i = 0; i < elements()->length(); i++)
text->AddElement((*elements())[i]);
}
TextElement TextElement::Atom(RegExpAtom* atom) {
return TextElement(ATOM, atom);
}
TextElement TextElement::CharClass(RegExpCharacterClass* char_class) {
return TextElement(CHAR_CLASS, char_class);
}
intptr_t TextElement::length() const {
switch (text_type()) {
case ATOM:
return atom()->length();
case CHAR_CLASS:
return 1;
}
UNREACHABLE();
return 0;
}
class FrequencyCollator : public ValueObject {
public:
FrequencyCollator() : total_samples_(0) {
for (intptr_t i = 0; i < RegExpMacroAssembler::kTableSize; i++) {
frequencies_[i] = CharacterFrequency(i);
}
}
void CountCharacter(intptr_t character) {
intptr_t index = (character & RegExpMacroAssembler::kTableMask);
frequencies_[index].Increment();
total_samples_++;
}
// Does not measure in percent, but rather per-128 (the table size from the
// regexp macro assembler).
intptr_t Frequency(intptr_t in_character) {
ASSERT((in_character & RegExpMacroAssembler::kTableMask) == in_character);
if (total_samples_ < 1) return 1; // Division by zero.
intptr_t freq_in_per128 =
(frequencies_[in_character].counter() * 128) / total_samples_;
return freq_in_per128;
}
private:
class CharacterFrequency {
public:
CharacterFrequency() : counter_(0), character_(-1) {}
explicit CharacterFrequency(intptr_t character)
: counter_(0), character_(character) {}
void Increment() { counter_++; }
intptr_t counter() { return counter_; }
intptr_t character() { return character_; }
private:
intptr_t counter_;
intptr_t character_;
DISALLOW_ALLOCATION();
};
private:
CharacterFrequency frequencies_[RegExpMacroAssembler::kTableSize];
intptr_t total_samples_;
};
class RegExpCompiler : public ValueObject {
public:
RegExpCompiler(intptr_t capture_count, bool is_one_byte);
intptr_t AllocateRegister() { return next_register_++; }
// Lookarounds to match lone surrogates for unicode character class matches
// are never nested. We can therefore reuse registers.
intptr_t UnicodeLookaroundStackRegister() {
if (unicode_lookaround_stack_register_ == kNoRegister) {
unicode_lookaround_stack_register_ = AllocateRegister();
}
return unicode_lookaround_stack_register_;
}
intptr_t UnicodeLookaroundPositionRegister() {
if (unicode_lookaround_position_register_ == kNoRegister) {
unicode_lookaround_position_register_ = AllocateRegister();
}
return unicode_lookaround_position_register_;
}
#if !defined(DART_PRECOMPILED_RUNTIME)
RegExpEngine::CompilationResult Assemble(IRRegExpMacroAssembler* assembler,
RegExpNode* start,
intptr_t capture_count,
const String& pattern);
#endif
RegExpEngine::CompilationResult Assemble(
BytecodeRegExpMacroAssembler* assembler,
RegExpNode* start,
intptr_t capture_count,
const String& pattern);
inline void AddWork(RegExpNode* node) { work_list_->Add(node); }
static constexpr intptr_t kImplementationOffset = 0;
static constexpr intptr_t kNumberOfRegistersOffset = 0;
static constexpr intptr_t kCodeOffset = 1;
RegExpMacroAssembler* macro_assembler() { return macro_assembler_; }
EndNode* accept() { return accept_; }
static constexpr intptr_t kMaxRecursion = 100;
inline intptr_t recursion_depth() { return recursion_depth_; }
inline void IncrementRecursionDepth() { recursion_depth_++; }
inline void DecrementRecursionDepth() { recursion_depth_--; }
void SetRegExpTooBig() { reg_exp_too_big_ = true; }
inline bool one_byte() const { return is_one_byte_; }
bool read_backward() { return read_backward_; }
void set_read_backward(bool value) { read_backward_ = value; }
FrequencyCollator* frequency_collator() { return &frequency_collator_; }
intptr_t current_expansion_factor() { return current_expansion_factor_; }
void set_current_expansion_factor(intptr_t value) {
current_expansion_factor_ = value;
}
Zone* zone() const { return zone_; }
static constexpr intptr_t kNoRegister = -1;
private:
EndNode* accept_;
intptr_t next_register_;
intptr_t unicode_lookaround_stack_register_;
intptr_t unicode_lookaround_position_register_;
ZoneGrowableArray<RegExpNode*>* work_list_;
intptr_t recursion_depth_;
RegExpMacroAssembler* macro_assembler_;
bool is_one_byte_;
bool reg_exp_too_big_;
bool read_backward_;
intptr_t current_expansion_factor_;
FrequencyCollator frequency_collator_;
Zone* zone_;
};
class RecursionCheck : public ValueObject {
public:
explicit RecursionCheck(RegExpCompiler* compiler) : compiler_(compiler) {
compiler->IncrementRecursionDepth();
}
~RecursionCheck() { compiler_->DecrementRecursionDepth(); }
private:
RegExpCompiler* compiler_;
};
static RegExpEngine::CompilationResult IrregexpRegExpTooBig() {
return RegExpEngine::CompilationResult("RegExp too big");
}
// Attempts to compile the regexp using an Irregexp code generator. Returns
// a fixed array or a null handle depending on whether it succeeded.
RegExpCompiler::RegExpCompiler(intptr_t capture_count, bool is_one_byte)
: next_register_(2 * (capture_count + 1)),
unicode_lookaround_stack_register_(kNoRegister),
unicode_lookaround_position_register_(kNoRegister),
work_list_(nullptr),
recursion_depth_(0),
is_one_byte_(is_one_byte),
reg_exp_too_big_(false),
read_backward_(false),
current_expansion_factor_(1),
zone_(Thread::Current()->zone()) {
accept_ = new (Z) EndNode(EndNode::ACCEPT, Z);
}
#if !defined(DART_PRECOMPILED_RUNTIME)
RegExpEngine::CompilationResult RegExpCompiler::Assemble(
IRRegExpMacroAssembler* macro_assembler,
RegExpNode* start,
intptr_t capture_count,
const String& pattern) {
macro_assembler->set_slow_safe(false /* use_slow_safe_regexp_compiler */);
macro_assembler_ = macro_assembler;
ZoneGrowableArray<RegExpNode*> work_list(0);
work_list_ = &work_list;
BlockLabel fail;
macro_assembler_->PushBacktrack(&fail);
Trace new_trace;
start->Emit(this, &new_trace);
macro_assembler_->BindBlock(&fail);
macro_assembler_->Fail();
while (!work_list.is_empty()) {
work_list.RemoveLast()->Emit(this, &new_trace);
}
if (reg_exp_too_big_) return IrregexpRegExpTooBig();
macro_assembler->GenerateBacktrackBlock();
macro_assembler->FinalizeRegistersArray();
return RegExpEngine::CompilationResult(
macro_assembler->backtrack_goto(), macro_assembler->graph_entry(),
macro_assembler->num_blocks(), macro_assembler->num_stack_locals(),
next_register_);
}
#endif
RegExpEngine::CompilationResult RegExpCompiler::Assemble(
BytecodeRegExpMacroAssembler* macro_assembler,
RegExpNode* start,
intptr_t capture_count,
const String& pattern) {
macro_assembler->set_slow_safe(false /* use_slow_safe_regexp_compiler */);
macro_assembler_ = macro_assembler;
ZoneGrowableArray<RegExpNode*> work_list(0);
work_list_ = &work_list;
BlockLabel fail;
macro_assembler_->PushBacktrack(&fail);
Trace new_trace;
start->Emit(this, &new_trace);
macro_assembler_->BindBlock(&fail);
macro_assembler_->Fail();
while (!work_list.is_empty()) {
work_list.RemoveLast()->Emit(this, &new_trace);
}
if (reg_exp_too_big_) return IrregexpRegExpTooBig();
TypedData& bytecode = TypedData::ZoneHandle(macro_assembler->GetBytecode());
return RegExpEngine::CompilationResult(&bytecode, next_register_);
}
bool Trace::DeferredAction::Mentions(intptr_t that) {
if (action_type() == ActionNode::CLEAR_CAPTURES) {
Interval range = static_cast<DeferredClearCaptures*>(this)->range();
return range.Contains(that);
} else {
return reg() == that;
}
}
bool Trace::mentions_reg(intptr_t reg) {
for (DeferredAction* action = actions_; action != nullptr;
action = action->next()) {
if (action->Mentions(reg)) return true;
}
return false;
}
bool Trace::GetStoredPosition(intptr_t reg, intptr_t* cp_offset) {
ASSERT(*cp_offset == 0);
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;
}
// This is called as we come into a loop choice node and some other tricky
// nodes. It normalizes the state of the code generator to ensure we can
// generate generic code.
intptr_t Trace::FindAffectedRegisters(OutSet* affected_registers, Zone* zone) {
intptr_t 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 (intptr_t 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,
intptr_t max_register,
const OutSet& registers_to_pop,
const OutSet& registers_to_clear) {
for (intptr_t reg = max_register; reg >= 0; reg--) {
if (registers_to_pop.Get(reg)) {
assembler->PopRegister(reg);
} else if (registers_to_clear.Get(reg)) {
intptr_t clear_to = reg;
while (reg > 0 && registers_to_clear.Get(reg - 1)) {
reg--;
}
assembler->ClearRegisters(reg, clear_to);
}
}
}
void Trace::PerformDeferredActions(RegExpMacroAssembler* assembler,
intptr_t max_register,
const OutSet& affected_registers,
OutSet* registers_to_pop,
OutSet* registers_to_clear,
Zone* zone) {
for (intptr_t reg = 0; reg <= max_register; reg++) {
if (!affected_registers.Get(reg)) {
continue;
}
// The chronologically first deferred action in the trace
// is used to infer the action needed to restore a register
// to its previous state (or not, if it's safe to ignore it).
enum DeferredActionUndoType { ACTION_IGNORE, ACTION_RESTORE, ACTION_CLEAR };
DeferredActionUndoType undo_action = ACTION_IGNORE;
intptr_t value = 0;
bool absolute = false;
bool clear = false;
const intptr_t kNoStore = kMinInt32;
intptr_t 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 ACTION_IGNORE if we know there is no
// value to restore.
undo_action = ACTION_RESTORE;
ASSERT(store_position == kNoStore);
ASSERT(!clear);
break;
}
case ActionNode::INCREMENT_REGISTER:
if (!absolute) {
value++;
}
ASSERT(store_position == kNoStore);
ASSERT(!clear);
undo_action = 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 = ACTION_IGNORE;
} else {
undo_action = pc->is_capture() ? ACTION_CLEAR : ACTION_RESTORE;
}
ASSERT(!absolute);
ASSERT(value == 0);
break;
}
case ActionNode::CLEAR_CAPTURES: {
// Since we're scanning in reverse order, if we've already
// set the position we have to ignore historically earlier
// clearing operations.
if (store_position == kNoStore) {
clear = true;
}
undo_action = ACTION_RESTORE;
ASSERT(!absolute);
ASSERT(value == 0);
break;
}
default:
UNREACHABLE();
break;
}
}
}
// Prepare for the undo-action (e.g., push if it's going to be popped).
if (undo_action == ACTION_RESTORE) {
assembler->PushRegister(reg);
registers_to_pop->Set(reg, zone);
} else if (undo_action == 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();
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();
}
Zone* zone = successor->zone();
intptr_t max_register = FindAffectedRegisters(&affected_registers, zone);
OutSet registers_to_pop;
OutSet registers_to_clear;
PerformDeferredActions(assembler, max_register, affected_registers,
&registers_to_pop, &registers_to_clear, zone);
if (cp_offset_ != 0) {
assembler->AdvanceCurrentPosition(cp_offset_);
}
// Create a new trivial state and generate the node with that.
BlockLabel undo;
assembler->PushBacktrack(&undo);
Trace new_state;
successor->Emit(compiler, &new_state);
// On backtrack we need to restore state.
assembler->BindBlock(&undo);
RestoreAffectedRegisters(assembler, max_register, registers_to_pop,
registers_to_clear);
if (backtrack() == 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->BindBlock(label());
}
// Throw away everything on the backtrack stack since the start
// of the negative submatch and restore the character position.
assembler->ReadCurrentPositionFromRegister(current_position_register_);
assembler->ReadStackPointerFromRegister(stack_pointer_register_);
if (clear_capture_count_ > 0) {
// Clear any captures that might have been performed during the success
// of the body of the negative look-ahead.
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->BindBlock(label());
}
switch (action_) {
case ACCEPT:
assembler->Succeed();
return;
case BACKTRACK:
assembler->GoTo(trace->backtrack());
return;
case NEGATIVE_SUBMATCH_SUCCESS:
// This case is handled in a different virtual method.
UNREACHABLE();
}
UNIMPLEMENTED();
}
void GuardedAlternative::AddGuard(Guard* guard, Zone* zone) {
if (guards_ == nullptr) guards_ = new (zone) ZoneGrowableArray<Guard*>(1);
guards_->Add(guard);
}
ActionNode* ActionNode::SetRegister(intptr_t reg,
intptr_t 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(intptr_t 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(intptr_t 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(intptr_t stack_reg,
intptr_t 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(intptr_t stack_reg,
intptr_t position_reg,
intptr_t clear_register_count,
intptr_t 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(intptr_t start_register,
intptr_t repetition_register,
intptr_t 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:
ASSERT(!trace->mentions_reg(guard->reg()));
macro_assembler->IfRegisterGE(guard->reg(), guard->value(),
trace->backtrack());
break;
case Guard::GEQ:
ASSERT(!trace->mentions_reg(guard->reg()));
macro_assembler->IfRegisterLT(guard->reg(), guard->value(),
trace->backtrack());
break;
}
}
// Returns the number of characters in the equivalence class, omitting those
// that cannot occur in the source string because it is ASCII.
static intptr_t GetCaseIndependentLetters(uint16_t character,
bool one_byte_subject,
int32_t* letters) {
unibrow::Mapping<unibrow::Ecma262UnCanonicalize> jsregexp_uncanonicalize;
intptr_t length = jsregexp_uncanonicalize.get(character, '\0', letters);
// Unibrow returns 0 or 1 for characters where case independence is
// trivial.
if (length == 0) {
letters[0] = character;
length = 1;
}
if (!one_byte_subject || character <= Symbols::kMaxOneCharCodeSymbol) {
return length;
}
// The standard requires that non-ASCII characters cannot have ASCII
// character codes in their equivalence class.
// TODO(dcarney): issue 3550 this is not actually true for Latin1 anymore,
// is it? For example, \u00C5 is equivalent to \u212B.
return 0;
}
static inline bool EmitSimpleCharacter(Zone* zone,
RegExpCompiler* compiler,
uint16_t c,
BlockLabel* on_failure,
intptr_t cp_offset,
bool check,
bool preloaded) {
RegExpMacroAssembler* assembler = compiler->macro_assembler();
bool bound_checked = false;
if (!preloaded) {
assembler->LoadCurrentCharacter(cp_offset, on_failure, check);
bound_checked = true;
}
assembler->CheckNotCharacter(c, on_failure);
return bound_checked;
}
// Only emits non-letters (things that don't have case). Only used for case
// independent matches.
static inline bool EmitAtomNonLetter(Zone* zone,
RegExpCompiler* compiler,
uint16_t c,
BlockLabel* on_failure,
intptr_t cp_offset,
bool check,
bool preloaded) {
RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
bool one_byte = compiler->one_byte();
int32_t chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
intptr_t length = GetCaseIndependentLetters(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 > Symbols::kMaxOneCharCodeSymbol) {
// Can't match - see above.
return false; // Bounds not checked.
}
if (!preloaded) {
macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check);
checked = check;
}
macro_assembler->CheckNotCharacter(c, on_failure);
}
return checked;
}
static bool ShortCutEmitCharacterPair(RegExpMacroAssembler* macro_assembler,
bool one_byte,
uint16_t c1,
uint16_t c2,
BlockLabel* on_failure) {
uint16_t char_mask;
if (one_byte) {
char_mask = Symbols::kMaxOneCharCodeSymbol;
} else {
char_mask = Utf16::kMaxCodeUnit;
}
uint16_t exor = c1 ^ c2;
// Check whether exor has only one bit set.
if (((exor - 1) & exor) == 0) {
// If c1 and c2 differ only by one bit.
// Ecma262UnCanonicalize always gives the highest number last.
ASSERT(c2 > c1);
uint16_t mask = char_mask ^ exor;
macro_assembler->CheckNotCharacterAfterAnd(c1, mask, on_failure);
return true;
}
ASSERT(c2 > c1);
uint16_t diff = c2 - c1;
if (((diff - 1) & diff) == 0 && c1 >= diff) {
// If the characters differ by 2^n but don't differ by one bit then
// subtract the difference from the found character, then do the or
// trick. We avoid the theoretical case where negative numbers are
// involved in order to simplify code generation.
uint16_t mask = char_mask ^ diff;
macro_assembler->CheckNotCharacterAfterMinusAnd(c1 - diff, diff, mask,
on_failure);
return true;
}
return false;
}
typedef bool EmitCharacterFunction(Zone* zone,
RegExpCompiler* compiler,
uint16_t c,
BlockLabel* on_failure,
intptr_t cp_offset,
bool check,
bool preloaded);
// Only emits letters (things that have case). Only used for case independent
// matches.
static inline bool EmitAtomLetter(Zone* zone,
RegExpCompiler* compiler,
uint16_t c,
BlockLabel* on_failure,
intptr_t cp_offset,
bool check,
bool preloaded) {
RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
bool one_byte = compiler->one_byte();
int32_t chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
intptr_t length = GetCaseIndependentLetters(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);
}
BlockLabel ok;
ASSERT(unibrow::Ecma262UnCanonicalize::kMaxWidth == 4);
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->BindBlock(&ok);
}
break;
}
case 4:
macro_assembler->CheckCharacter(chars[3], &ok);
FALL_THROUGH;
case 3:
macro_assembler->CheckCharacter(chars[0], &ok);
macro_assembler->CheckCharacter(chars[1], &ok);
macro_assembler->CheckNotCharacter(chars[2], on_failure);
macro_assembler->BindBlock(&ok);
break;
default:
UNREACHABLE();
break;
}
return true;
}
static void EmitBoundaryTest(RegExpMacroAssembler* masm,
uint16_t border,
BlockLabel* fall_through,
BlockLabel* above_or_equal,
BlockLabel* below) {
if (below != fall_through) {
masm->CheckCharacterLT(border, below);
if (above_or_equal != fall_through) masm->GoTo(above_or_equal);
} else {
masm->CheckCharacterGT(border - 1, above_or_equal);
}
}
static void EmitDoubleBoundaryTest(RegExpMacroAssembler* masm,
uint16_t first,
uint16_t last,
BlockLabel* fall_through,
BlockLabel* in_range,
BlockLabel* out_of_range) {
if (in_range == fall_through) {
if (first == last) {
masm->CheckNotCharacter(first, out_of_range);
} else {
masm->CheckCharacterNotInRange(first, last, out_of_range);
}
} else {
if (first == last) {
masm->CheckCharacter(first, in_range);
} else {
masm->CheckCharacterInRange(first, last, in_range);
}
if (out_of_range != fall_through) masm->GoTo(out_of_range);
}
}
// even_label is for ranges[i] to ranges[i + 1] where i - start_index is even.
// odd_label is for ranges[i] to ranges[i + 1] where i - start_index is odd.
static void EmitUseLookupTable(RegExpMacroAssembler* masm,
ZoneGrowableArray<uint16_t>* ranges,
intptr_t start_index,
intptr_t end_index,
uint16_t min_char,
BlockLabel* fall_through,
BlockLabel* even_label,
BlockLabel* odd_label) {
const intptr_t kSize = RegExpMacroAssembler::kTableSize;
const intptr_t kMask = RegExpMacroAssembler::kTableMask;
intptr_t base = (min_char & ~kMask);
// Assert that everything is on one kTableSize page.
for (intptr_t i = start_index; i <= end_index; i++) {
ASSERT((ranges->At(i) & ~kMask) == base);
}
ASSERT(start_index == 0 || (ranges->At(start_index - 1) & ~kMask) <= base);
char templ[kSize];
BlockLabel* on_bit_set;
BlockLabel* on_bit_clear;
intptr_t bit;
if (even_label == fall_through) {
on_bit_set = odd_label;
on_bit_clear = even_label;
bit = 1;
} else {
on_bit_set = even_label;
on_bit_clear = odd_label;
bit = 0;
}
for (intptr_t i = 0; i < (ranges->At(start_index) & kMask) && i < kSize;
i++) {
templ[i] = bit;
}
intptr_t j = 0;
bit ^= 1;
for (intptr_t i = start_index; i < end_index; i++) {
for (j = (ranges->At(i) & kMask); j < (ranges->At(i + 1) & kMask); j++) {
templ[j] = bit;
}
bit ^= 1;
}
for (intptr_t i = j; i < kSize; i++) {
templ[i] = bit;
}
// TODO(erikcorry): Cache these.
const TypedData& ba = TypedData::ZoneHandle(
masm->zone(), TypedData::New(kTypedDataUint8ArrayCid, kSize, Heap::kOld));
for (intptr_t i = 0; i < kSize; i++) {
ba.SetUint8(i, templ[i]);
}
masm->CheckBitInTable(ba, on_bit_set);
if (on_bit_clear != fall_through) masm->GoTo(on_bit_clear);
}
static void CutOutRange(RegExpMacroAssembler* masm,
ZoneGrowableArray<uint16_t>* ranges,
intptr_t start_index,
intptr_t end_index,
intptr_t cut_index,
BlockLabel* even_label,
BlockLabel* odd_label) {
bool odd = (((cut_index - start_index) & 1) == 1);
BlockLabel* in_range_label = odd ? odd_label : even_label;
BlockLabel dummy;
EmitDoubleBoundaryTest(masm, ranges->At(cut_index),
ranges->At(cut_index + 1) - 1, &dummy, in_range_label,
&dummy);
ASSERT(!dummy.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 (intptr_t j = cut_index; j > start_index; j--) {
(*ranges)[j] = ranges->At(j - 1);
}
for (intptr_t j = cut_index + 1; j < end_index; j++) {
(*ranges)[j] = ranges->At(j + 1);
}
}
// Unicode case. Split the search space into kSize spaces that are handled
// with recursion.
static void SplitSearchSpace(ZoneGrowableArray<uint16_t>* ranges,
intptr_t start_index,
intptr_t end_index,
intptr_t* new_start_index,
intptr_t* new_end_index,
uint16_t* border) {
const intptr_t kSize = RegExpMacroAssembler::kTableSize;
const intptr_t kMask = RegExpMacroAssembler::kTableMask;
uint16_t first = ranges->At(start_index);
uint16_t last = ranges->At(end_index) - 1;
*new_start_index = start_index;
*border = (ranges->At(start_index) & ~kMask) + kSize;
while (*new_start_index < end_index) {
if (ranges->At(*new_start_index) > *border) break;
(*new_start_index)++;
}
// new_start_index is the index of the first edge that is beyond the
// current kSize space.
// For very large search spaces we do a binary chop search of the non-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).
intptr_t 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 > Symbols::kMaxOneCharCodeSymbol && // 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) {
intptr_t scan_forward_for_section_border = binary_chop_index;
intptr_t new_border = (ranges->At(binary_chop_index) | kMask) + 1;
while (scan_forward_for_section_border < end_index) {
if (ranges->At(scan_forward_for_section_border) > new_border) {
*new_start_index = scan_forward_for_section_border;
*border = new_border;
break;
}
scan_forward_for_section_border++;
}
}
ASSERT(*new_start_index > start_index);
*new_end_index = *new_start_index - 1;
if (ranges->At(*new_end_index) == *border) {
(*new_end_index)--;
}
if (*border >= ranges->At(end_index)) {
*border = ranges->At(end_index);
*new_start_index = end_index; // Won't be used.
*new_end_index = end_index - 1;
}
}
// Gets a series of segment boundaries representing a character class. If the
// character is in the range between an even and an odd boundary (counting from
// start_index) then go to even_label, otherwise go to odd_label. We already
// know that the character is in the range of min_char to max_char inclusive.
// Either label can be null indicating backtracking. Either label can also be
// equal to the fall_through label.
static void GenerateBranches(RegExpMacroAssembler* masm,
ZoneGrowableArray<uint16_t>* ranges,
intptr_t start_index,
intptr_t end_index,
uint16_t min_char,
uint16_t max_char,
BlockLabel* fall_through,
BlockLabel* even_label,
BlockLabel* odd_label) {
uint16_t first = ranges->At(start_index);
uint16_t last = ranges->At(end_index) - 1;
ASSERT(min_char < first);
// Just need to test if the character is before or on-or-after
// a particular character.
if (start_index == end_index) {
EmitBoundaryTest(masm, first, fall_through, even_label, odd_label);
return;
}
// Another almost trivial case: There is one interval in the middle that is
// different from the end intervals.
if (start_index + 1 == end_index) {
EmitDoubleBoundaryTest(masm, first, last, fall_through, even_label,
odd_label);
return;
}
// It's not worth using table lookup if there are very few intervals in the
// character class.
if (end_index - start_index <= 6) {
// It is faster to test for individual characters, so we look for those
// first, then try arbitrary ranges in the second round.
static intptr_t kNoCutIndex = -1;
intptr_t cut = kNoCutIndex;
for (intptr_t i = start_index; i < end_index; i++) {
if (ranges->At(i) == ranges->At(i + 1) - 1) {
cut = i;
break;
}
}
if (cut == kNoCutIndex) cut = start_index;
CutOutRange(masm, ranges, start_index, end_index, cut, even_label,
odd_label);
ASSERT(end_index - start_index >= 2);
GenerateBranches(masm, ranges, start_index + 1, end_index - 1, min_char,
max_char, fall_through, even_label, odd_label);
return;
}
// If there are a lot of intervals in the regexp, then we will use tables to
// determine whether the character is inside or outside the character class.
const intptr_t kBits = RegExpMacroAssembler::kTableSizeBits;
if ((max_char >> kBits) == (min_char >> kBits)) {
EmitUseLookupTable(masm, ranges, start_index, end_index, min_char,
fall_through, even_label, odd_label);
return;
}
if ((min_char >> kBits) != (first >> kBits)) {
masm->CheckCharacterLT(first, odd_label);
GenerateBranches(masm, ranges, start_index + 1, end_index, first, max_char,
fall_through, odd_label, even_label);
return;
}
intptr_t new_start_index = 0;
intptr_t new_end_index = 0;
uint16_t border = 0;
SplitSearchSpace(ranges, start_index, end_index, &new_start_index,
&new_end_index, &border);
BlockLabel handle_rest;
BlockLabel* above = &handle_rest;
if (border == last + 1) {
// We didn't find any section that started after the limit, so everything
// above the border is one of the terminal labels.
above = (end_index & 1) != (start_index & 1) ? odd_label : even_label;
ASSERT(new_end_index == end_index - 1);
}
ASSERT(start_index <= new_end_index);
ASSERT(new_start_index <= end_index);
ASSERT(start_index < new_start_index);
ASSERT(new_end_index < end_index);
ASSERT(new_end_index + 1 == new_start_index ||
(new_end_index + 2 == new_start_index &&
border == ranges->At(new_end_index + 1)));
ASSERT(min_char < border - 1);
ASSERT(border < max_char);
ASSERT(ranges->At(new_end_index) < border);
ASSERT(border < ranges->At(new_start_index) ||
(border == ranges->At(new_start_index) &&
new_start_index == end_index && new_end_index == end_index - 1 &&
border == last + 1));
ASSERT(new_start_index == 0 || border >= ranges->At(new_start_index - 1));
masm->CheckCharacterGT(border - 1, above);
BlockLabel dummy;
GenerateBranches(masm, ranges, start_index, new_end_index, min_char,
border - 1, &dummy, even_label, odd_label);
if (handle_rest.is_linked()) {
masm->BindBlock(&handle_rest);
bool flip = (new_start_index & 1) != (start_index & 1);
GenerateBranches(masm, ranges, new_start_index, end_index, border, max_char,
&dummy, flip ? odd_label : even_label,
flip ? even_label : odd_label);
}
}
static void EmitCharClass(RegExpMacroAssembler* macro_assembler,
RegExpCharacterClass* cc,
bool one_byte,
BlockLabel* on_failure,
intptr_t cp_offset,
bool check_offset,
bool preloaded,
Zone* zone) {
ZoneGrowableArray<CharacterRange>* ranges = cc->ranges();
if (!CharacterRange::IsCanonical(ranges)) {
CharacterRange::Canonicalize(ranges);
}
uint16_t max_char;
if (one_byte) {
max_char = Symbols::kMaxOneCharCodeSymbol;
} else {
max_char = Utf16::kMaxCodeUnit;
}
intptr_t range_count = ranges->length();
intptr_t last_valid_range = range_count - 1;
while (last_valid_range >= 0) {
const 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() && macro_assembler->CheckSpecialCharacterClass(
cc->standard_type(), on_failure)) {
return;
}
// A new list with ascending entries. Each entry is a code unit
// where there is a boundary between code units that are part of
// the class and code units that are not. Normally we insert an
// entry at zero which goes to the failure label, but if there
// was already one there we fall through for success on that entry.
// Subsequent entries have alternating meaning (success/failure).
ZoneGrowableArray<uint16_t>* range_boundaries =
new (zone) ZoneGrowableArray<uint16_t>(last_valid_range);
bool zeroth_entry_is_failure = !cc->is_negated();
for (intptr_t i = 0; i <= last_valid_range; i++) {
const CharacterRange& range = ranges->At(i);
if (range.from() == 0) {
ASSERT(i == 0);
zeroth_entry_is_failure = !zeroth_entry_is_failure;
} else {
range_boundaries->Add(range.from());
}
if (range.to() + 1 <= max_char) {
range_boundaries->Add(range.to() + 1);
}
}
intptr_t end_index = range_boundaries->length() - 1;
BlockLabel fall_through;
GenerateBranches(macro_assembler, range_boundaries,
0, // start_index.
end_index,
0, // min_char.
max_char, &fall_through,
zeroth_entry_is_failure ? &fall_through : on_failure,
zeroth_entry_is_failure ? on_failure : &fall_through);
macro_assembler->BindBlock(&fall_through);
}
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()) {
// We are being asked to generate a generic version, but that's already
// been done so just go to it.
macro_assembler->GoTo(&label_);
return DONE;
}
if (compiler->recursion_depth() >= RegExpCompiler::kMaxRecursion) {
// To avoid too deep recursion we push the node to the work queue and just
// generate a goto here.
compiler->AddWork(this);
macro_assembler->GoTo(&label_);
return DONE;
}
// Generate generic version of the node and bind the label for later use.
macro_assembler->BindBlock(&label_);
return CONTINUE;
}
// We are being asked to make a non-generic version. Keep track of how many
// non-generic versions we generate so as not to overdo it.
trace_count_++;
if (kRegexpOptimization && trace_count_ < kMaxCopiesCodeGenerated &&
compiler->recursion_depth() <= RegExpCompiler::kMaxRecursion) {
return CONTINUE;
}
// If we get here code has been generated for this node too many times or
// recursion is too deep. Time to switch to a generic version. The code for
// generic versions above can handle deep recursion properly.
trace->Flush(compiler, this);
return DONE;
}
intptr_t ActionNode::EatsAtLeast(intptr_t still_to_find,
intptr_t 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(intptr_t offset,
intptr_t budget,
BoyerMooreLookahead* bm,
bool not_at_start) {
if (action_type_ == BEGIN_SUBMATCH) {
bm->SetRest(offset);
} else if (action_type_ != POSITIVE_SUBMATCH_SUCCESS) {
on_success()->FillInBMInfo(offset, budget - 1, bm, not_at_start);
}
SaveBMInfo(bm, not_at_start, offset);
}
intptr_t AssertionNode::EatsAtLeast(intptr_t still_to_find,
intptr_t 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(intptr_t offset,
intptr_t budget,
BoyerMooreLookahead* bm,
bool not_at_start) {
// Match the behaviour of EatsAtLeast on this node.
if (assertion_type() == AT_START && not_at_start) return;
on_success()->FillInBMInfo(offset, budget - 1, bm, not_at_start);
SaveBMInfo(bm, not_at_start, offset);
}
intptr_t BackReferenceNode::EatsAtLeast(intptr_t still_to_find,
intptr_t 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);
}
intptr_t TextNode::EatsAtLeast(intptr_t still_to_find,
intptr_t budget,
bool not_at_start) {
if (read_backward()) return 0;
intptr_t answer = Length();
if (answer >= still_to_find) return answer;
if (budget <= 0) return answer;
// We are not at start after this node so we set the last argument to 'true'.
return answer +
on_success()->EatsAtLeast(still_to_find - answer, budget - 1, true);
}
intptr_t NegativeLookaroundChoiceNode::EatsAtLeast(intptr_t still_to_find,
intptr_t budget,
bool not_at_start) {
if (budget <= 0) return 0;
// Alternative 0 is the negative lookahead, alternative 1 is what comes
// afterwards.
RegExpNode* node = (*alternatives_)[1].node();
return node->EatsAtLeast(still_to_find, budget - 1, not_at_start);
}
void NegativeLookaroundChoiceNode::GetQuickCheckDetails(
QuickCheckDetails* details,
RegExpCompiler* compiler,
intptr_t filled_in,
bool not_at_start) {
// Alternative 0 is the negative lookahead, alternative 1 is what comes
// afterwards.
RegExpNode* node = (*alternatives_)[1].node();
return node->GetQuickCheckDetails(details, compiler, filled_in, not_at_start);
}
intptr_t ChoiceNode::EatsAtLeastHelper(intptr_t still_to_find,
intptr_t budget,
RegExpNode* ignore_this_node,
bool not_at_start) {
if (budget <= 0) return 0;
intptr_t min = 100;
intptr_t choice_count = alternatives_->length();
budget = (budget - 1) / choice_count;
for (intptr_t i = 0; i < choice_count; i++) {
RegExpNode* node = (*alternatives_)[i].node();
if (node == ignore_this_node) continue;
intptr_t 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;
}
intptr_t LoopChoiceNode::EatsAtLeast(intptr_t still_to_find,
intptr_t budget,
bool not_at_start) {
return EatsAtLeastHelper(still_to_find, budget - 1, loop_node_, not_at_start);
}
intptr_t ChoiceNode::EatsAtLeast(intptr_t still_to_find,
intptr_t 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 = Symbols::kMaxOneCharCodeSymbol;
} else {
char_mask = Utf16::kMaxCodeUnit;
}
mask_ = 0;
value_ = 0;
intptr_t char_shift = 0;
for (intptr_t i = 0; i < characters_; i++) {
Position* pos = &positions_[i];
if ((pos->mask & Symbols::kMaxOneCharCodeSymbol) != 0) {
found_useful_op = true;
}
mask_ |= (pos->mask & char_mask) << char_shift;
value_ |= (pos->value & char_mask) << char_shift;
char_shift += asc ? 8 : 16;
}
return found_useful_op;
}
bool RegExpNode::EmitQuickCheck(RegExpCompiler* compiler,
Trace* bounds_check_trace,
Trace* trace,
bool preload_has_checked_bounds,
BlockLabel* on_possible_success,
QuickCheckDetails* details,
bool fall_through_on_failure) {
if (details->characters() == 0) return false;
GetQuickCheckDetails(details, compiler, 0,
trace->at_start() == Trace::FALSE_VALUE);
if (details->cannot_match()) return false;
if (!details->Rationalize(compiler->one_byte())) return false;
ASSERT(details->characters() == 1 ||
compiler->macro_assembler()->CanReadUnaligned());
uint32_t mask = details->mask();
uint32_t value = details->value();
RegExpMacroAssembler* assembler = compiler->macro_assembler();
if (trace->characters_preloaded() != details->characters()) {
ASSERT(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 = Symbols::kMaxOneCharCodeSymbol;
} else {
char_mask = Utf16::kMaxCodeUnit;
}
if ((mask & char_mask) == char_mask) need_mask = false;
mask &= char_mask;
} else {
// For 2-character preloads in one-byte mode or 1-character preloads in
// two-byte mode we also use a 16 bit load with zero extend.
if (details->characters() == 2 && compiler->one_byte()) {
if ((mask & 0xffff) == 0xffff) need_mask = false;
} else if (details->characters() == 1 && !compiler->one_byte()) {
if ((mask & 0xffff) == 0xffff) need_mask = false;
} else {
if (mask == 0xffffffff) need_mask = false;
}
}
if (fall_through_on_failure) {
if (need_mask) {
assembler->CheckCharacterAfterAnd(value, mask, on_possible_success);
} else {
assembler->CheckCharacter(value, on_possible_success);
}
} else {
if (need_mask) {
assembler->CheckNotCharacterAfterAnd(value, mask, trace->backtrack());
} else {
assembler->CheckNotCharacter(value, trace->backtrack());
}
}
return true;
}
// Here is the meat of GetQuickCheckDetails (see also the comment on the
// super-class in the .h file).
//
// We iterate along the text object, building up for each character a
// mask and value that can be used to test for a quick failure to match.
// The masks and values for the positions will be combined into a single
// machine word for the current character width in order to be used in
// generating a quick check.
void TextNode::GetQuickCheckDetails(QuickCheckDetails* details,
RegExpCompiler* compiler,
intptr_t characters_filled_in,
bool not_at_start) {
#if defined(__GNUC__) && defined(__BYTE_ORDER__)
// TODO(zerny): Make the combination code byte-order independent.
ASSERT(details->characters() == 1 ||
(__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__));
#endif
// 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;
ASSERT(characters_filled_in < details->characters());
intptr_t characters = details->characters();
int32_t char_mask;
if (compiler->one_byte()) {
char_mask = Symbols::kMaxOneCharCodeSymbol;
} else {
char_mask = Utf16::kMaxCodeUnit;
}
for (intptr_t k = 0; k < elms_->length(); k++) {
TextElement elm = elms_->At(k);
if (elm.text_type() == TextElement::ATOM) {
ZoneGrowableArray<uint16_t>* quarks = elm.atom()->data();
for (intptr_t i = 0; i < characters && i < quarks->length(); i++) {
QuickCheckDetails::Position* pos =
details->positions(characters_filled_in);
uint16_t c = quarks->At(i);
if (c > char_mask) {
// If we expect a non-Latin1 character from an one-byte string,
// there is no way we can match. Not even case independent
// matching can turn an Latin1 character into non-Latin1 or
// vice versa.
// TODO(dcarney): issue 3550. Verify that this works as expected.
// For example, \u0178 is uppercase of \u00ff (y-umlaut).
details->set_cannot_match();
pos->determines_perfectly = false;
return;
}
if (elm.atom()->ignore_case()) {
int32_t chars[unibrow::Ecma262UnCanonicalize::kMaxWidth];
intptr_t length =
GetCaseIndependentLetters(c, compiler->one_byte(), chars);
ASSERT(length != 0); // Can only happen if c > char_mask (see above).
if (length == 1) {
// This letter has no case equivalents, so it's nice and simple
// and the mask-compare will determine definitely whether we have
// a match at this character position.
pos->mask = char_mask;
pos->value = c;
pos->determines_perfectly = true;
} else {
uint32_t common_bits = char_mask;
uint32_t bits = chars[0];
for (intptr_t j = 1; j < length; j++) {
uint32_t differing_bits = ((chars[j] & common_bits) ^ bits);
common_bits ^= differing_bits;
bits &= common_bits;
}
// If length is 2 and common bits has only one zero in it then
// our mask and compare instruction will determine definitely
// whether we have a match at this character position. Otherwise
// it can only be an approximate check.
uint32_t one_zero = (common_bits | ~char_mask);
if (length == 2 && ((~one_zero) & ((~one_zero) - 1)) == 0) {
pos->determines_perfectly = true;
}
pos->mask = common_bits;
pos->value = bits;
}
} else {
// Don't ignore case. Nice simple case where the mask-compare will
// determine definitely whether we have a match at this character
// position.
pos->mask = char_mask;
pos->value = c;
pos->determines_perfectly = true;
}
characters_filled_in++;
ASSERT(characters_filled_in <= details->characters());
if (characters_filled_in == details->characters()) {
return;
}
}
} else {
QuickCheckDetails::Position* pos =
details->positions(characters_filled_in);
RegExpCharacterClass* tree = elm.char_class();
ZoneGrowableArray<CharacterRange>* ranges = tree->ranges();
ASSERT(!ranges->is_empty());
if (!CharacterRange::IsCanonical(ranges)) {
CharacterRange::Canonicalize(ranges);
}
if (tree->is_negated()) {
// A quick check uses multi-character mask and compare. There is no
// useful way to incorporate a negative char class into this scheme
// so we just conservatively create a mask and value that will always
// succeed.
pos->mask = 0;
pos->value = 0;
} else {
intptr_t first_range = 0;
while (ranges->At(first_range).from() > char_mask) {
first_range++;
if (first_range == ranges->length()) {
details->set_cannot_match();
pos->determines_perfectly = false;
return;
}
}
CharacterRange range = ranges->At(first_range);
uint16_t from = range.from();
uint16_t to = range.to();
if (to > char_mask) {
to = char_mask;
}
uint32_t differing_bits = (from ^ to);
// A mask and compare is only perfect if the differing bits form a
// number like 00011111 with one single block of trailing 1s.
if ((differing_bits & (differing_bits + 1)) == 0 &&
from + differing_bits == to) {
pos->determines_perfectly = true;
}
uint32_t common_bits = ~SmearBitsRight(differing_bits);
uint32_t bits = (from & common_bits);
for (intptr_t i = first_range + 1; i < ranges->length(); i++) {
CharacterRange range = ranges->At(i);
uint16_t from = range.from();
uint16_t to = range.to();
if (from > char_mask) continue;
if (to > char_mask) to = char_mask;
// Here we are combining more ranges into the mask and compare
// value. With each new range the mask becomes more sparse and
// so the chances of a false positive rise. A character class
// with multiple ranges is assumed never to be equivalent to a
// mask and compare operation.
pos->determines_perfectly = false;
uint32_t new_common_bits = (from ^ to);
new_common_bits = ~SmearBitsRight(new_common_bits);
common_bits &= new_common_bits;
bits &= new_common_bits;
uint32_t differing_bits = (from & common_bits) ^ bits;
common_bits ^= differing_bits;
bits &= common_bits;
}
pos->mask = common_bits;
pos->value = bits;
}
characters_filled_in++;
ASSERT(characters_filled_in <= details->characters());
if (characters_filled_in == details->characters()) {
return;
}
}
}
ASSERT(characters_filled_in != details->characters());
if (!details->cannot_match()) {
on_success()->GetQuickCheckDetails(details, compiler, characters_filled_in,
true);
}
}
void 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(intptr_t by, bool one_byte) {
if (by >= characters_ || by < 0) {
// check that by < 0 => characters_ == 0
ASSERT(by >= 0 || characters_ == 0);
Clear();
return;
}
for (intptr_t i = 0; i < characters_ - by; i++) {
positions_[i] = positions_[by + i];
}
for (intptr_t 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, intptr_t from_index) {
ASSERT(characters_ == other->characters_);
if (other->cannot_match_) {
return;
}
if (cannot_match_) {
*this = *other;
return;
}
for (intptr_t 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;
uint16_t differing_bits = (pos->value ^ other_pos->value);
pos->mask &= ~differing_bits;
pos->value &= pos->mask;
}
}
class VisitMarker : public ValueObject {
public:
explicit VisitMarker(NodeInfo* info) : info_(info) {
ASSERT(!info->visited);
info->visited = true;
}
~VisitMarker() { info_->visited = false; }
private:
NodeInfo* info_;
};
RegExpNode* SeqRegExpNode::FilterOneByte(intptr_t depth) {
if (info()->replacement_calculated) return replacement();
if (depth < 0) return this;
ASSERT(!info()->visited);
VisitMarker marker(info());
return FilterSuccessor(depth - 1);
}
RegExpNode* SeqRegExpNode::FilterSuccessor(intptr_t 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(0x39c) || range.Contains(0x3bc) ||
range.Contains(0x178);
}
static bool RangesContainLatin1Equivalents(
ZoneGrowableArray<CharacterRange>* ranges) {
for (intptr_t i = 0; i < ranges->length(); i++) {
// TODO(dcarney): this could be a lot more efficient.
if (RangeContainsLatin1Equivalents(ranges->At(i))) return true;
}
return false;
}
static uint16_t ConvertNonLatin1ToLatin1(uint16_t c) {
ASSERT(c > Symbols::kMaxOneCharCodeSymbol);
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;
}
RegExpNode* TextNode::FilterOneByte(intptr_t depth) {
if (info()->replacement_calculated) return replacement();
if (depth < 0) return this;
ASSERT(!info()->visited);
VisitMarker marker(info());
intptr_t element_count = elms_->length();
for (intptr_t i = 0; i < element_count; i++) {
TextElement elm = elms_->At(i);
if (elm.text_type() == TextElement::ATOM) {
ZoneGrowableArray<uint16_t>* quarks = elm.atom()->data();
for (intptr_t j = 0; j < quarks->length(); j++) {
uint16_t c = quarks->At(j);
if (c <= Symbols::kMaxOneCharCodeSymbol) continue;
if (!elm.atom()->ignore_case()) 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 = ConvertNonLatin1ToLatin1(c);
// Character is outside Latin-1 completely
if (converted == 0) return set_replacement(nullptr);
// Convert quark to Latin-1 in place.
(*quarks)[0] = converted;
}
} else {
ASSERT(elm.text_type() == TextElement::CHAR_CLASS);
RegExpCharacterClass* cc = elm.char_class();
ZoneGrowableArray<CharacterRange>* ranges = cc->ranges();
if (!CharacterRange::IsCanonical(ranges)) {
CharacterRange::Canonicalize(ranges);
}
// Now they are in order so we only need to look at the first.
intptr_t range_count = ranges->length();
if (cc->is_negated()) {
if (range_count != 0 && ranges->At(0).from() == 0 &&
ranges->At(0).to() >= Symbols::kMaxOneCharCodeSymbol) {
// This will be handled in a later filter.
if (cc->flags().IgnoreCase() &&
RangesContainLatin1Equivalents(ranges)) {
continue;
}
return set_replacement(nullptr);
}
} else {
if (range_count == 0 ||
ranges->At(0).from() > Symbols::kMaxOneCharCodeSymbol) {
// This will be handled in a later filter.
if (cc->flags().IgnoreCase() &&
RangesContainLatin1Equivalents(ranges))
continue;
return set_replacement(nullptr);
}
}
}
}
return FilterSuccessor(depth - 1);
}
RegExpNode* LoopChoiceNode::FilterOneByte(intptr_t 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(intptr_t depth) {
if (info()->replacement_calculated) return replacement();
if (depth < 0) return this;
if (info()->visited) return this;
VisitMarker marker(info());
intptr_t choice_count = alternatives_->length();
for (intptr_t i = 0; i < choice_count; i++) {
GuardedAlternative alternative = alternatives_->At(i);
if (alternative.guards() != nullptr &&
alternative.guards()->length() != 0) {
set_replacement(this);
return this;
}
}
intptr_t surviving = 0;
RegExpNode* survivor = nullptr;
for (intptr_t i = 0; i < choice_count; i++) {
GuardedAlternative alternative = alternatives_->At(i);
RegExpNode* replacement = alternative.node()->FilterOneByte(depth - 1);
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.
ZoneGrowableArray<GuardedAlternative>* new_alternatives =
new (Z) ZoneGrowableArray<GuardedAlternative>(surviving);
for (intptr_t i = 0; i < choice_count; i++) {
RegExpNode* replacement =
(*alternatives_)[i].node()->FilterOneByte(depth - 1);
if (replacement != nullptr) {
(*alternatives_)[i].set_node(replacement);
new_alternatives->Add((*alternatives_)[i]);
}
}
alternatives_ = new_alternatives;
return this;
}
RegExpNode* NegativeLookaroundChoiceNode::FilterOneByte(intptr_t depth) {
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->FilterOneByte(depth - 1);
if (replacement == nullptr) return set_replacement(nullptr);
(*alternatives_)[1].set_node(replacement);
RegExpNode* neg_node = (*alternatives_)[0].node();
RegExpNode* neg_replacement = neg_node->FilterOneByte(depth - 1);
// 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);
}
void LoopChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details,
RegExpCompiler* compiler,
intptr_t characters_filled_in,
bool not_at_start) {
if (body_can_be_zero_length_ || info()->visited) return;
VisitMarker marker(info());
return ChoiceNode::GetQuickCheckDetails(details, compiler,
characters_filled_in, not_at_start);
}
void LoopChoiceNode::FillInBMInfo(intptr_t offset,
intptr_t budget,
BoyerMooreLookahead* bm,
bool not_at_start) {
if (body_can_be_zero_length_ || budget <= 0) {
bm->SetRest(offset);
SaveBMInfo(bm, not_at_start, offset);
return;
}
ChoiceNode::FillInBMInfo(offset, budget - 1, bm, not_at_start);
SaveBMInfo(bm, not_at_start, offset);
}
void ChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details,
RegExpCompiler* compiler,
intptr_t characters_filled_in,
bool not_at_start) {
not_at_start = (not_at_start || not_at_start_);
intptr_t choice_count = alternatives_->length();
ASSERT(choice_count > 0);
(*alternatives_)[0].node()->GetQuickCheckDetails(
details, compiler, characters_filled_in, not_at_start);
for (intptr_t 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);
}
}
// Check for [0-9A-Z_a-z].
static void EmitWordCheck(RegExpMacroAssembler* assembler,
BlockLabel* word,
BlockLabel* non_word,
bool fall_through_on_word) {
if (assembler->CheckSpecialCharacterClass(
fall_through_on_word ? 'w' : 'W',
fall_through_on_word ? non_word : word)) {
// Optimized implementation available.
return;
}
assembler->CheckCharacterGT('z', non_word);
assembler->CheckCharacterLT('0', non_word);
assembler->CheckCharacterGT('a' - 1, word);
assembler->CheckCharacterLT('9' + 1, word);
assembler->CheckCharacterLT('A', non_word);
assembler->CheckCharacterLT('Z' + 1, word);
if (fall_through_on_word) {
assembler->CheckNotCharacter('_', non_word);
} else {
assembler->CheckCharacter('_', word);
}
}
// Emit the code to check for a ^ in multiline mode (1-character lookbehind
// that matches newline or the start of input).
static void EmitHat(RegExpCompiler* compiler,
RegExpNode* on_success,
Trace* trace) {
RegExpMacroAssembler* assembler = compiler->macro_assembler();
// We will be loading the previous character into the current character
// register.
Trace new_trace(*trace);
new_trace.InvalidateCurrentCharacter();
BlockLabel ok;
if (new_trace.cp_offset() == 0) {
// The start of input counts as a newline in this context, so skip to
// ok if we are at the start.
assembler->CheckAtStart(&ok);
}
// We already checked that we are not at the start of input so it must be
// OK to load the previous character.
assembler->LoadCurrentCharacter(new_trace.cp_offset() - 1,
new_trace.backtrack(), false);
if (!assembler->CheckSpecialCharacterClass('n', new_trace.backtrack())) {
// Newline means \n, \r, 0x2028 or 0x2029.
if (!compiler->one_byte()) {
assembler->CheckCharacterAfterAnd(0x2028, 0xfffe, &ok);
}
assembler->CheckCharacter('\n', &ok);
assembler->CheckNotCharacter('\r', new_trace.backtrack());
}
assembler->BindBlock(&ok);
on_success->Emit(compiler, &new_trace);
}
// 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) {
intptr_t eats_at_least =
Utils::Minimum(kMaxLookaheadForBoyerMoore,
EatsAtLeast(kMaxLookaheadForBoyerMoore, kRecursionBudget,
not_at_start));
if (eats_at_least >= 1) {
BoyerMooreLookahead* bm =
new (Z) BoyerMooreLookahead(eats_at_least, compiler, Z);
FillInBMInfo(0, kRecursionBudget, bm, not_at_start);
if (bm->at(0)->is_non_word()) next_is_word_character = Trace::FALSE_VALUE;
if (bm->at(0)->is_word()) next_is_word_character = Trace::TRUE_VALUE;
}
} else {
if (lookahead->at(0)->is_non_word())
next_is_word_character = Trace::FALSE_VALUE;
if (lookahead->at(0)->is_word()) next_is_word_character = Trace::TRUE_VALUE;
}
bool at_boundary = (assertion_type_ == AssertionNode::AT_BOUNDARY);
if (next_is_word_character == Trace::UNKNOWN) {
BlockLabel before_non_word;
BlockLabel before_word;
if (trace->characters_preloaded() != 1) {
assembler->LoadCurrentCharacter(trace->cp_offset(), &before_non_word);
}
// Fall through on non-word.
EmitWordCheck(assembler, &before_word, &before_non_word, false);
// Next character is not a word character.
assembler->BindBlock(&before_non_word);
BlockLabel ok;
// Backtrack on \B (non-boundary check) if previous is a word,
// since we know next *is not* a word and this would be a boundary.
BacktrackIfPrevious(compiler, trace, at_boundary ? kIsNonWord : kIsWord);
if (!assembler->IsClosed()) {
assembler->GoTo(&ok);
}
assembler->BindBlock(&before_word);
BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord);
assembler->BindBlock(&ok);
} else if (next_is_word_character == Trace::TRUE_VALUE) {
BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord);
} else {
ASSERT(next_is_word_character == Trace::FALSE_VALUE);
BacktrackIfPrevious(compiler, trace, at_boundary ? kIsNonWord : kIsWord);
}
}
void AssertionNode::BacktrackIfPrevious(
RegExpCompiler* compiler,
Trace* trace,
AssertionNode::IfPrevious backtrack_if_previous) {
RegExpMacroAssembler* assembler = compiler->macro_assembler();
Trace new_trace(*trace);
new_trace.InvalidateCurrentCharacter();
BlockLabel fall_through, dummy;
BlockLabel* non_word = backtrack_if_previous == kIsNonWord
? new_trace.backtrack()
: &fall_through;
BlockLabel* word = backtrack_if_previous == kIsNonWord
? &fall_through
: new_trace.backtrack();
if (new_trace.cp_offset() == 0) {
// The start of input counts as a non-word character, so the question is
// decided if we are at the start.
assembler->CheckAtStart(non_word);
}
// We already checked that we are not at the start of input so it must be
// OK to load the previous character.
assembler->LoadCurrentCharacter(new_trace.cp_offset() - 1, &dummy, false);
EmitWordCheck(assembler, word, non_word, backtrack_if_previous == kIsNonWord);
assembler->BindBlock(&fall_through);
on_success()->Emit(compiler, &new_trace);
}
void AssertionNode::GetQuickCheckDetails(QuickCheckDetails* details,
RegExpCompiler* compiler,
intptr_t filled_in,
bool not_at_start) {
if (assertion_type_ == AT_START && not_at_start) {
details->set_cannot_match();
return;
}
return on_success()->GetQuickCheckDetails(details, compiler, filled_in,
not_at_start);
}
void AssertionNode::Emit(RegExpCompiler* compiler, Trace* trace) {
RegExpMacroAssembler* assembler = compiler->macro_assembler();
switch (assertion_type_) {
case AT_END: {
BlockLabel ok;
assembler->CheckPosition(trace->cp_offset(), &ok);
assembler->GoTo(trace->backtrack());
assembler->BindBlock(&ok);
break;
}
case AT_START: {
if (trace->at_start() == Trace::FALSE_VALUE) {
assembler->GoTo(trace->backtrack());
return;
}
if (trace->at_start() == Trace::UNKNOWN) {
assembler->CheckNotAtStart(trace->cp_offset(), trace->backtrack());
Trace at_start_trace = *trace;
at_start_trace.set_at_start(Trace::TRUE_VALUE);
on_success()->Emit(compiler, &at_start_trace);
return;
}
} break;
case AFTER_NEWLINE:
EmitHat(compiler, on_success(), trace);
return;
case AT_BOUNDARY:
case AT_NON_BOUNDARY: {
EmitBoundaryCheck(compiler, trace);
return;
}
}
on_success()->Emit(compiler, trace);
}
static bool DeterminedAlready(QuickCheckDetails* quick_check, intptr_t offset) {
if (quick_check == nullptr) return false;
if (offset >= quick_check->characters()) return false;
return quick_check->positions(offset)->determines_perfectly;
}
static void UpdateBoundsCheck(intptr_t index, intptr_t* checked_up_to) {
if (index > *checked_up_to) {
*checked_up_to = index;
}
}
// We call this repeatedly to generate code for each pass over the text node.
// The passes are in increasing order of difficulty because we hope one
// of the first passes will fail in which case we are saved the work of the
// later passes. for example for the case independent regexp /%[asdfghjkl]a/
// we will check the '%' in the first pass, the case independent 'a' in the
// second pass and the character class in the last pass.
//
// The passes are done from right to left, so for example to test for /bar/
// we will first test for an 'r' with offset 2, then an 'a' with offset 1
// and then a 'b' with offset 0. This means we can avoid the end-of-input
// bounds check most of the time. In the example we only need to check for
// end-of-input when loading the putative 'r'.
//
// A slight complication involves the fact that the first character may already
// be fetched into a register by the previous node. In this case we want to
// do the test for that character first. We do this in separate passes. The
// 'preloaded' argument indicates that we are doing such a 'pass'. If such a
// pass has been performed then subsequent passes will have true in
// first_element_checked to indicate that character does not need to be
// checked again.
//
// In addition to all this we are passed a Trace, which can
// contain an AlternativeGeneration object. In this AlternativeGeneration
// object we can see details of any quick check that was already passed in
// order to get to the code we are now generating. The quick check can involve
// loading characters, which means we do not need to recheck the bounds
// up to the limit the quick check already checked. In addition the quick
// check can have involved a mask and compare operation which may simplify
// or obviate the need for further checks at some character positions.
void TextNode::TextEmitPass(RegExpCompiler* compiler,
TextEmitPassType pass,
bool preloaded,
Trace* trace,
bool first_element_checked,
intptr_t* checked_up_to) {
RegExpMacroAssembler* assembler = compiler->macro_assembler();
bool one_byte = compiler->one_byte();
BlockLabel* backtrack = trace->backtrack();
QuickCheckDetails* quick_check = trace->quick_check_performed();
intptr_t element_count = elms_->length();
intptr_t backward_offset = read_backward() ? -Length() : 0;
for (intptr_t i = preloaded ? 0 : element_count - 1; i >= 0; i--) {
TextElement elm = elms_->At(i);
intptr_t cp_offset = trace->cp_offset() + elm.cp_offset() + backward_offset;
if (elm.text_type() == TextElement::ATOM) {
ZoneGrowableArray<uint16_t>* quarks = elm.atom()->data();
for (intptr_t j = preloaded ? 0 : quarks->length() - 1; j >= 0; j--) {
if (SkipPass(pass, elm.atom()->ignore_case())) continue;
if (first_element_checked && i == 0 && j == 0) continue;
if (DeterminedAlready(quick_check, elm.cp_offset() + j)) continue;
EmitCharacterFunction* emit_function = nullptr;
uint16_t quark = quarks->At(j);
if (elm.atom()->ignore_case()) {
// Everywhere else we assume that a non-Latin-1 character cannot match
// a Latin-1 character. Avoid the cases where this is assumption is
// invalid by using the Latin1 equivalent instead.
quark = Latin1::TryConvertToLatin1(quark);
}
switch (pass) {
case NON_LATIN1_MATCH:
ASSERT(one_byte);
if (quark > Symbols::kMaxOneCharCodeSymbol) {
assembler->GoTo(backtrack);
return;
}
break;
case NON_LETTER_CHARACTER_MATCH:
emit_function = &EmitAtomNonLetter;
break;
case SIMPLE_CHARACTER_MATCH:
emit_function = &EmitSimpleCharacter;
break;
case CASE_CHARACTER_MATCH:
emit_function = &EmitAtomLetter;
break;
default:
break;
}
if (emit_function != nullptr) {
const bool bounds_check =
*checked_up_to < (cp_offset + j) || read_backward();
bool bound_checked =
emit_function(Z, compiler, quarks->At(j), backtrack,
cp_offset + j, bounds_check, preloaded);
if (bound_checked) UpdateBoundsCheck(cp_offset + j, checked_up_to);
}
}
} else {
ASSERT(elm.text_type() == TextElement::CHAR_CLASS);
if (pass == CHARACTER_CLASS_MATCH) {
if (first_element_checked && i == 0) continue;
if (DeterminedAlready(quick_check, elm.cp_offset())) continue;
RegExpCharacterClass* cc = elm.char_class();
bool bounds_check = *checked_up_to < cp_offset || read_backward();
EmitCharClass(assembler, cc, one_byte, backtrack, cp_offset,
bounds_check, preloaded, Z);
UpdateBoundsCheck(cp_offset, checked_up_to);
}
}
}
}
intptr_t TextNode::Length() {
TextElement elm = elms_->Last();
ASSERT(elm.cp_offset() >= 0);
return elm.cp_offset() + elm.length();
}
bool TextNode::SkipPass(intptr_t intptr_t_pass, bool ignore_case) {
TextEmitPassType pass = static_cast<TextEmitPassType>(intptr_t_pass);
if (ignore_case) {
return pass == SIMPLE_CHARACTER_MATCH;
} else {
return pass == NON_LETTER_CHARACTER_MATCH || pass == CASE_CHARACTER_MATCH;
}
}
TextNode* TextNode::CreateForCharacterRanges(
ZoneGrowableArray<CharacterRange>* ranges,
bool read_backward,
RegExpNode* on_success,
RegExpFlags flags) {
ASSERT(ranges != nullptr);
ZoneGrowableArray<TextElement>* elms = new ZoneGrowableArray<TextElement>(1);
elms->Add(TextElement::CharClass(new RegExpCharacterClass(ranges, flags)));
return new TextNode(elms, read_backward, on_success);
}
TextNode* TextNode::CreateForSurrogatePair(CharacterRange lead,
CharacterRange trail,
bool read_backward,
RegExpNode* on_success,
RegExpFlags flags) {
auto lead_ranges = CharacterRange::List(on_success->zone(), lead);
auto trail_ranges = CharacterRange::List(on_success->zone(), trail);
auto elms = new ZoneGrowableArray<TextElement>(2);
elms->Add(
TextElement::CharClass(new RegExpCharacterClass(lead_ranges, flags)));
elms->Add(
TextElement::CharClass(new RegExpCharacterClass(trail_ranges, flags)));
return new TextNode(elms, read_backward, on_success);
}
// This generates the code to match a text node. A text node can contain
// straight character sequences (possibly to be matched in a case-independent
// way) and character classes. For efficiency we do not do this in a single
// pass from left to right. Instead we pass over the text node several times,
// emitting code for some character positions every time. See the comment on
// TextEmitPass for details.
void TextNode::Emit(RegExpCompiler* compiler, Trace* trace) {
LimitResult limit_result = LimitVersions(compiler, trace);
if (limit_result == DONE) return;
ASSERT(limit_result == CONTINUE);
if (trace->cp_offset() + Length() > RegExpMacroAssembler::kMaxCPOffset) {
compiler->SetRegExpTooBig();
return;
}
if (compiler->one_byte()) {
intptr_t dummy = 0;
TextEmitPass(compiler, NON_LATIN1_MATCH, false, trace, false, &dummy);
}
bool first_elt_done = false;
intptr_t bound_checked_to = trace->cp_offset() - 1;
bound_checked_to += trace->bound_checked_up_to();
// If a character is preloaded into the current character register then
// check that now.
if (trace->characters_preloaded() == 1) {
for (intptr_t pass = kFirstRealPass; pass <= kLastPass; pass++) {
TextEmitPass(compiler, static_cast<TextEmitPassType>(pass), true, trace,
false, &bound_checked_to);
}
first_elt_done = true;
}
for (intptr_t pass = kFirstRealPass; pass <= kLastPass; pass++) {
TextEmitPass(compiler, static_cast<TextEmitPassType>(pass), false, trace,
first_elt_done, &bound_checked_to);
}
Trace successor_trace(*trace);
// If we advance backward, we may end up at the start.
successor_trace.AdvanceCurrentPositionInTrace(
read_backward() ? -Length() : Length(), compiler);
successor_trace.set_at_start(read_backward() ? Trace::UNKNOWN
: Trace::FALSE_VALUE);
RecursionCheck rc(compiler);
on_success()->Emit(compiler, &successor_trace);
}
void Trace::InvalidateCurrentCharacter() {
characters_preloaded_ = 0;
}
void Trace::AdvanceCurrentPositionInTrace(intptr_t by,
RegExpCompiler* compiler) {
// 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->one_byte());
cp_offset_ += by;
if (cp_offset_ > RegExpMacroAssembler::kMaxCPOffset) {
compiler->SetRegExpTooBig();
cp_offset_ = 0;
}
bound_checked_up_to_ =
Utils::Maximum(static_cast<intptr_t>(0), bound_checked_up_to_ - by);
}
void TextNode::MakeCaseIndependent(bool is_one_byte) {
intptr_t element_count = elms_->length();
for (intptr_t i = 0; i < element_count; i++) {
TextElement elm = elms_->At(i);
if (elm.text_type() == TextElement::CHAR_CLASS) {
RegExpCharacterClass* cc = elm.char_class();
bool case_equivalents_already_added =
cc->flags().NeedsUnicodeCaseEquivalents();
if (cc->flags().IgnoreCase() && !case_equivalents_already_added) {
// None of the standard character classes is different in the case
// independent case and it slows us down if we don't know that.
if (cc->is_standard()) continue;
CharacterRange::AddCaseEquivalents(cc->ranges(), is_one_byte, Z);
}
}
}
}
intptr_t TextNode::GreedyLoopTextLength() {
TextElement elm = elms_->At(elms_->length() - 1);
return elm.cp_offset() + elm.length();
}
RegExpNode* TextNode::GetSuccessorOfOmnivorousTextNode(
RegExpCompiler* compiler) {
if (read_backward()) return nullptr;
if (elms_->length() != 1) return nullptr;
TextElement elm = elms_->At(0);
if (elm.text_type() != TextElement::CHAR_CLASS) return nullptr;
RegExpCharacterClass* node = elm.char_class();
ZoneGrowableArray<CharacterRange>* ranges = node->ranges();
if (!CharacterRange::IsCanonical(ranges)) {
CharacterRange::Canonicalize(ranges);
}
if (node->is_negated()) {
return ranges->length() == 0 ? on_success() : nullptr;
}
if (ranges->length() != 1) return nullptr;
uint32_t max_char;
if (compiler->one_byte()) {
max_char = Symbols::kMaxOneCharCodeSymbol;
} else {
max_char = Utf16::kMaxCodeUnit;
}
return ranges->At(0).IsEverything(max_char) ? on_success() : nullptr;
}
// Finds the fixed match length of a sequence of nodes that goes from
// this alternative and back to this choice node. If there are variable
// length nodes or other complications in the way then return a sentinel
// value indicating that a greedy loop cannot be constructed.
intptr_t ChoiceNode::GreedyLoopTextLengthForAlternative(
const GuardedAlternative* alternative) {
intptr_t length = 0;
RegExpNode* node = alternative->node();
// Later we will generate code for all these text nodes using recursion
// so we have to limit the max number.
intptr_t recursion_depth = 0;
while (node != this) {
if (recursion_depth++ > RegExpCompiler::kMaxRecursion) {
return kNodeIsTooComplexForGreedyLoops;
}
intptr_t node_length = node->GreedyLoopTextLength();
if (node_length == kNodeIsTooComplexForGreedyLoops) {
return kNodeIsTooComplexForGreedyLoops;
}
length += node_length;
SeqRegExpNode* seq_node = static_cast<SeqRegExpNode*>(node);
node = seq_node->on_success();
}
return read_backward() ? -length : length;
}
void LoopChoiceNode::AddLoopAlternative(GuardedAlternative alt) {
ASSERT(loop_node_ == nullptr);
AddAlternative(alt);
loop_node_ = alt.node();
}
void LoopChoiceNode::AddContinueAlternative(GuardedAlternative alt) {
ASSERT(continue_node_ == nullptr);
AddAlternative(alt);
continue_node_ = alt.node();
}
void LoopChoiceNode::Emit(RegExpCompiler* compiler, Trace* trace) {
RegExpMacroAssembler* macro_assembler = compiler->macro_assembler();
if (trace->stop_node() == this) {
// Back edge of greedy optimized loop node graph.
intptr_t text_length =
GreedyLoopTextLengthForAlternative(&alternatives_->At(0));
ASSERT(text_length != kNodeIsTooComplexForGreedyLoops);
// Update the counter-based backtracking info on the stack. This is an
// optimization for greedy loops (see below).
ASSERT(trace->cp_offset() == text_length);
macro_assembler->AdvanceCurrentPosition(text_length);
macro_assembler->GoTo(trace->loop_label());
return;
}
ASSERT(trace->stop_node() == nullptr);
if (!trace->is_trivial()) {
trace->Flush(compiler, this);
return;
}
ChoiceNode::Emit(compiler, trace);
}
intptr_t ChoiceNode::CalculatePreloadCharacters(RegExpCompiler* compiler,
intptr_t eats_at_least) {
intptr_t preload_characters =
Utils::Minimum(static_cast<intptr_t>(4), eats_at_least);
if (compiler->one_byte()) {
#if !defined(DART_COMPRESSED_POINTERS) && !defined(TARGET_ARCH_RISCV32)
if (preload_characters > 4) preload_characters = 4;
// We can't preload 3 characters because there is no machine instruction
// to do that. We can't just load 4 because we could be reading
// beyond the end of the string, which could cause a memory fault.
if (preload_characters == 3) preload_characters = 2;
#else
// Ensure LoadCodeUnitsInstr can always produce a Smi. See
// https://github.com/dart-lang/sdk/issues/29951
if (preload_characters > 2) preload_characters = 2;
#endif
} else {
#if !defined(DART_COMPRESSED_POINTERS) && !defined(TARGET_ARCH_RISCV32)
if (preload_characters > 2) preload_characters = 2;
#else
// Ensure LoadCodeUnitsInstr can always produce a Smi. See
// https://github.com/dart-lang/sdk/issues/29951
if (preload_characters > 1) preload_characters = 1;
#endif
}
if (!compiler->macro_assembler()->CanReadUnaligned()) {
if (preload_characters > 1) preload_characters = 1;
}
return preload_characters;
}
// This structure is used when generating the alternatives in a choice node. It
// records the way the alternative is being code generated.
struct AlternativeGeneration {
AlternativeGeneration()
: possible_success(),
expects_preload(false),
after(),
quick_check_details() {}
BlockLabel possible_success;
bool expects_preload;
BlockLabel after;
QuickCheckDetails quick_check_details;
};
// Creates a list of AlternativeGenerations. If the list has a reasonable
// size then it is on the stack, otherwise the excess is on the heap.
class AlternativeGenerationList {
public:
explicit AlternativeGenerationList(intptr_t count) : count_(count) {
ASSERT(count >= 0);
if (count > kAFew) {
excess_alt_gens_.reset(new AlternativeGeneration[count - kAFew]);
}
}
AlternativeGeneration* at(intptr_t i) {
ASSERT(0 <= i);
ASSERT(i < count_);
if (i < kAFew) {
return &a_few_alt_gens_[i];
}
return &excess_alt_gens_[i - kAFew];
}
private:
static constexpr intptr_t kAFew = 10;
intptr_t count_;
AlternativeGeneration a_few_alt_gens_[kAFew];
std::unique_ptr<AlternativeGeneration[]> excess_alt_gens_;
DISALLOW_ALLOCATION();
DISALLOW_COPY_AND_ASSIGN(AlternativeGenerationList);
};
static constexpr int32_t kRangeEndMarker = Utf::kMaxCodePoint + 1;
// The '2' variant is 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.
// 0x180E has been removed from Unicode's Zs category and thus
// from ECMAScript's WhiteSpace category as of Unicode 6.3.
static constexpr int32_t kSpaceRanges[] = {
'\t', '\r' + 1, ' ', ' ' + 1, 0x00A0, 0x00A1, 0x1680,
0x1681, 0x2000, 0x200B, 0x2028, 0x202A, 0x202F, 0x2030,
0x205F, 0x2060, 0x3000, 0x3001, 0xFEFF, 0xFF00, kRangeEndMarker};
static constexpr intptr_t kSpaceRangeCount = ARRAY_SIZE(kSpaceRanges);
static constexpr int32_t kWordRanges[] = {
'0', '9' + 1, 'A', 'Z' + 1, '_', '_' + 1, 'a', 'z' + 1, kRangeEndMarker};
static constexpr intptr_t kWordRangeCount = ARRAY_SIZE(kWordRanges);
static constexpr int32_t kDigitRanges[] = {'0', '9' + 1,