From def9aa5d0ac0b77c6a6f592fe6c8512480926c5c Mon Sep 17 00:00:00 2001 From: Jakob Gruber Date: Mon, 17 Jun 2019 09:22:43 +0200 Subject: [PATCH] [regexp] Extract more parts of the regexp compiler Bug: v8:9359 Change-Id: I06a4ccc53abff25237a1113774a0b17bdf861c86 Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1658157 Reviewed-by: Peter Marshall Commit-Queue: Jakob Gruber Cr-Commit-Position: refs/heads/master@{#62198} --- BUILD.gn | 8 + src/regexp/jsregexp-inl.h | 5 - src/regexp/jsregexp.cc | 4418 +------------------------------ src/regexp/jsregexp.h | 1288 +-------- src/regexp/regexp-compiler.cc | 3714 ++++++++++++++++++++++++++ src/regexp/regexp-compiler.h | 554 +++- src/regexp/regexp-dotprinter.cc | 339 +++ src/regexp/regexp-dotprinter.h | 23 + src/regexp/regexp-nodes.h | 678 +++++ src/regexp/regexp-parser.cc | 1 + 10 files changed, 5338 insertions(+), 5690 deletions(-) create mode 100644 src/regexp/regexp-compiler.cc create mode 100644 src/regexp/regexp-dotprinter.cc create mode 100644 src/regexp/regexp-dotprinter.h create mode 100644 src/regexp/regexp-nodes.h diff --git a/BUILD.gn b/BUILD.gn index 8b5c97df79..4b67c8734a 100644 --- a/BUILD.gn +++ b/BUILD.gn @@ -2686,7 +2686,10 @@ v8_source_set("v8_base_without_compiler") { "src/regexp/regexp-ast.h", "src/regexp/regexp-bytecodes.h", "src/regexp/regexp-compiler-tonode.cc", + "src/regexp/regexp-compiler.cc", "src/regexp/regexp-compiler.h", + "src/regexp/regexp-dotprinter.cc", + "src/regexp/regexp-dotprinter.h", "src/regexp/regexp-interpreter.cc", "src/regexp/regexp-interpreter.h", "src/regexp/regexp-macro-assembler-arch.h", @@ -2697,6 +2700,7 @@ v8_source_set("v8_base_without_compiler") { "src/regexp/regexp-macro-assembler-tracer.h", "src/regexp/regexp-macro-assembler.cc", "src/regexp/regexp-macro-assembler.h", + "src/regexp/regexp-nodes.h", "src/regexp/regexp-parser.cc", "src/regexp/regexp-parser.h", "src/regexp/regexp-stack.cc", @@ -2949,6 +2953,10 @@ v8_source_set("v8_base_without_compiler") { "src/objects/elements.cc", "src/objects/objects.cc", "src/parsing/parser.cc", + + # Explicit template instantiation clash (these files are also very large). + "src/regexp/regexp-compiler-tonode.cc", + "src/regexp/regexp-compiler.cc", ] if (v8_current_cpu == "x86") { diff --git a/src/regexp/jsregexp-inl.h b/src/regexp/jsregexp-inl.h index b542add17b..6358174949 100644 --- a/src/regexp/jsregexp-inl.h +++ b/src/regexp/jsregexp-inl.h @@ -75,11 +75,6 @@ int32_t* RegExpImpl::GlobalCache::LastSuccessfulMatch() { return ®ister_array_[index]; } -RegExpEngine::CompilationResult::CompilationResult(Isolate* isolate, - const char* error_message) - : error_message(error_message), - code(ReadOnlyRoots(isolate).the_hole_value()) {} - } // namespace internal } // namespace v8 diff --git a/src/regexp/jsregexp.cc b/src/regexp/jsregexp.cc index a38b2d05a6..400b6695c7 100644 --- a/src/regexp/jsregexp.cc +++ b/src/regexp/jsregexp.cc @@ -4,39 +4,16 @@ #include "src/regexp/jsregexp.h" -#include -#include - -#include "src/base/platform/platform.h" #include "src/codegen/compilation-cache.h" -#include "src/diagnostics/code-tracer.h" -#include "src/execution/execution.h" -#include "src/execution/isolate-inl.h" -#include "src/execution/message-template.h" -#include "src/heap/factory.h" #include "src/heap/heap-inl.h" -#include "src/objects/elements.h" #include "src/regexp/jsregexp-inl.h" #include "src/regexp/regexp-compiler.h" +#include "src/regexp/regexp-dotprinter.h" #include "src/regexp/regexp-interpreter.h" #include "src/regexp/regexp-macro-assembler-arch.h" #include "src/regexp/regexp-macro-assembler-irregexp.h" -#include "src/regexp/regexp-macro-assembler-tracer.h" #include "src/regexp/regexp-parser.h" -#include "src/regexp/regexp-stack.h" -#include "src/runtime/runtime.h" #include "src/strings/string-search.h" -#include "src/strings/unicode-decoder.h" -#include "src/strings/unicode-inl.h" -#include "src/utils/ostreams.h" -#include "src/utils/splay-tree-inl.h" -#include "src/zone/zone-list-inl.h" - -#ifdef V8_INTL_SUPPORT -#include "unicode/locid.h" -#include "unicode/uniset.h" -#include "unicode/utypes.h" -#endif // V8_INTL_SUPPORT namespace v8 { namespace internal { @@ -82,12 +59,6 @@ ContainedInLattice AddRange(ContainedInLattice containment, return containment; } -// More makes code generation slower, less makes V8 benchmark score lower. -const int kMaxLookaheadForBoyerMoore = 8; -// In a 3-character pattern you can maximally step forwards 3 characters -// at a time, which is not always enough to pay for the extra logic. -const int kPatternTooShortForBoyerMoore = 2; - // Identifies the sort of regexps where the regexp engine is faster // than the code used for atom matches. static bool HasFewDifferentCharacters(Handle pattern) { @@ -644,4370 +615,17 @@ int RegExpImpl::GlobalCache::AdvanceZeroLength(int last_index) { return last_index + 1; } -// ------------------------------------------------------------------- -// Implementation of the Irregexp regular expression engine. -// -// The Irregexp regular expression engine is intended to be a complete -// implementation of ECMAScript regular expressions. It generates either -// bytecodes or native code. - -// The Irregexp regexp engine is structured in three steps. -// 1) The parser generates an abstract syntax tree. See ast.cc. -// 2) From the AST a node network is created. The nodes are all -// subclasses of RegExpNode. The nodes represent states when -// executing a regular expression. Several optimizations are -// performed on the node network. -// 3) From the nodes we generate either byte codes or native code -// that can actually execute the regular expression (perform -// the search). The code generation step is described in more -// detail below. - -// Code generation. -// -// The nodes are divided into four main categories. -// * Choice nodes -// These represent places where the regular expression can -// match in more than one way. For example on entry to an -// alternation (foo|bar) or a repetition (*, +, ? or {}). -// * Action nodes -// These represent places where some action should be -// performed. Examples include recording the current position -// in the input string to a register (in order to implement -// captures) or other actions on register for example in order -// to implement the counters needed for {} repetitions. -// * Matching nodes -// These attempt to match some element part of the input string. -// Examples of elements include character classes, plain strings -// or back references. -// * End nodes -// These are used to implement the actions required on finding -// a successful match or failing to find a match. -// -// The code generated (whether as byte codes or native code) maintains -// some state as it runs. This consists of the following elements: -// -// * The capture registers. Used for string captures. -// * Other registers. Used for counters etc. -// * The current position. -// * The stack of backtracking information. Used when a matching node -// fails to find a match and needs to try an alternative. -// -// Conceptual regular expression execution model: -// -// There is a simple conceptual model of regular expression execution -// which will be presented first. The actual code generated is a more -// efficient simulation of the simple conceptual model: -// -// * Choice nodes are implemented as follows: -// For each choice except the last { -// push current position -// push backtrack code location -// -// backtrack code location: -// pop current position -// } -// -// -// * Actions nodes are generated as follows -// -// -// push backtrack code location -// -// backtrack code location: -// -// -// -// * Matching nodes are generated as follows: -// if input string matches at current position -// update current position -// -// else -// -// -// Thus it can be seen that the current position is saved and restored -// by the choice nodes, whereas the registers are saved and restored by -// by the action nodes that manipulate them. -// -// The other interesting aspect of this model is that nodes are generated -// at the point where they are needed by a recursive call to Emit(). If -// the node has already been code generated then the Emit() call will -// generate a jump to the previously generated code instead. In order to -// limit recursion it is possible for the Emit() function to put the node -// on a work list for later generation and instead generate a jump. The -// destination of the jump is resolved later when the code is generated. -// -// Actual regular expression code generation. -// -// Code generation is actually more complicated than the above. In order -// to improve the efficiency of the generated code some optimizations are -// performed -// -// * Choice nodes have 1-character lookahead. -// A choice node looks at the following character and eliminates some of -// the choices immediately based on that character. This is not yet -// implemented. -// * Simple greedy loops store reduced backtracking information. -// A quantifier like /.*foo/m will greedily match the whole input. It will -// then need to backtrack to a point where it can match "foo". The naive -// implementation of this would push each character position onto the -// backtracking stack, then pop them off one by one. This would use space -// proportional to the length of the input string. However since the "." -// can only match in one way and always has a constant length (in this case -// of 1) it suffices to store the current position on the top of the stack -// once. Matching now becomes merely incrementing the current position and -// backtracking becomes decrementing the current position and checking the -// result against the stored current position. This is faster and saves -// space. -// * The current state is virtualized. -// This is used to defer expensive operations until it is clear that they -// are needed and to generate code for a node more than once, allowing -// specialized an efficient versions of the code to be created. This is -// explained in the section below. -// -// Execution state virtualization. -// -// Instead of emitting code, nodes that manipulate the state can record their -// manipulation in an object called the Trace. The Trace object can record a -// current position offset, an optional backtrack code location on the top of -// the virtualized backtrack stack and some register changes. When a node is -// to be emitted it can flush the Trace or update it. Flushing the Trace -// will emit code to bring the actual state into line with the virtual state. -// Avoiding flushing the state can postpone some work (e.g. updates of capture -// registers). Postponing work can save time when executing the regular -// expression since it may be found that the work never has to be done as a -// failure to match can occur. In addition it is much faster to jump to a -// known backtrack code location than it is to pop an unknown backtrack -// location from the stack and jump there. -// -// The virtual state found in the Trace affects code generation. For example -// the virtual state contains the difference between the actual current -// position and the virtual current position, and matching code needs to use -// this offset to attempt a match in the correct location of the input -// string. Therefore code generated for a non-trivial trace is specialized -// to that trace. The code generator therefore has the ability to generate -// code for each node several times. In order to limit the size of the -// generated code there is an arbitrary limit on how many specialized sets of -// code may be generated for a given node. If the limit is reached, the -// trace is flushed and a generic version of the code for a node is emitted. -// This is subsequently used for that node. The code emitted for non-generic -// trace is not recorded in the node and so it cannot currently be reused in -// the event that code generation is requested for an identical trace. - - -void RegExpTree::AppendToText(RegExpText* text, Zone* zone) { - UNREACHABLE(); -} - - -void RegExpAtom::AppendToText(RegExpText* text, Zone* zone) { - text->AddElement(TextElement::Atom(this), zone); -} - - -void RegExpCharacterClass::AppendToText(RegExpText* text, Zone* zone) { - text->AddElement(TextElement::CharClass(this), zone); -} - - -void RegExpText::AppendToText(RegExpText* text, Zone* zone) { - for (int i = 0; i < elements()->length(); i++) - text->AddElement(elements()->at(i), zone); -} - - -TextElement TextElement::Atom(RegExpAtom* atom) { - return TextElement(ATOM, atom); -} - - -TextElement TextElement::CharClass(RegExpCharacterClass* char_class) { - return TextElement(CHAR_CLASS, char_class); -} - - -int TextElement::length() const { - switch (text_type()) { - case ATOM: - return atom()->length(); - - case CHAR_CLASS: - return 1; - } - UNREACHABLE(); -} - - -DispatchTable* ChoiceNode::GetTable(bool ignore_case) { - if (table_ == nullptr) { - table_ = new(zone()) DispatchTable(zone()); - DispatchTableConstructor cons(table_, ignore_case, zone()); - cons.BuildTable(this); - } - return table_; -} - - -class RecursionCheck { - public: - explicit RecursionCheck(RegExpCompiler* compiler) : compiler_(compiler) { - compiler->IncrementRecursionDepth(); - } - ~RecursionCheck() { compiler_->DecrementRecursionDepth(); } - private: - RegExpCompiler* compiler_; -}; - - -static RegExpEngine::CompilationResult IrregexpRegExpTooBig(Isolate* isolate) { - return RegExpEngine::CompilationResult(isolate, "RegExp too big"); -} - - -// Attempts to compile the regexp using an Irregexp code generator. Returns -// a fixed array or a null handle depending on whether it succeeded. -RegExpCompiler::RegExpCompiler(Isolate* isolate, Zone* zone, int capture_count, - bool one_byte) - : next_register_(2 * (capture_count + 1)), - unicode_lookaround_stack_register_(kNoRegister), - unicode_lookaround_position_register_(kNoRegister), - work_list_(nullptr), - recursion_depth_(0), - one_byte_(one_byte), - reg_exp_too_big_(false), - limiting_recursion_(false), - optimize_(FLAG_regexp_optimization), - read_backward_(false), - current_expansion_factor_(1), - frequency_collator_(), - isolate_(isolate), - zone_(zone) { - accept_ = new(zone) EndNode(EndNode::ACCEPT, zone); - DCHECK_GE(RegExpMacroAssembler::kMaxRegister, next_register_ - 1); -} - -RegExpEngine::CompilationResult RegExpCompiler::Assemble( - Isolate* isolate, RegExpMacroAssembler* macro_assembler, RegExpNode* start, - int capture_count, Handle pattern) { -#ifdef DEBUG - if (FLAG_trace_regexp_assembler) - macro_assembler_ = new RegExpMacroAssemblerTracer(isolate, macro_assembler); - else -#endif - macro_assembler_ = macro_assembler; - - std::vector work_list; - work_list_ = &work_list; - Label fail; - macro_assembler_->PushBacktrack(&fail); - Trace new_trace; - start->Emit(this, &new_trace); - macro_assembler_->Bind(&fail); - macro_assembler_->Fail(); - while (!work_list.empty()) { - RegExpNode* node = work_list.back(); - work_list.pop_back(); - node->set_on_work_list(false); - if (!node->label()->is_bound()) node->Emit(this, &new_trace); - } - if (reg_exp_too_big_) { - macro_assembler_->AbortedCodeGeneration(); - return IrregexpRegExpTooBig(isolate_); - } - - Handle code = macro_assembler_->GetCode(pattern); - isolate->IncreaseTotalRegexpCodeGenerated(code->Size()); - work_list_ = nullptr; -#ifdef ENABLE_DISASSEMBLER - if (FLAG_print_code && !FLAG_regexp_interpret_all) { - CodeTracer::Scope trace_scope(isolate->GetCodeTracer()); - OFStream os(trace_scope.file()); - Handle::cast(code)->Disassemble(pattern->ToCString().get(), os); - } -#endif -#ifdef DEBUG - if (FLAG_trace_regexp_assembler) { - delete macro_assembler_; - } -#endif - return RegExpEngine::CompilationResult(*code, next_register_); -} - - -bool Trace::DeferredAction::Mentions(int that) { - if (action_type() == ActionNode::CLEAR_CAPTURES) { - Interval range = static_cast(this)->range(); - return range.Contains(that); - } else { - return reg() == that; - } -} - - -bool Trace::mentions_reg(int reg) { - for (DeferredAction* action = actions_; action != nullptr; - action = action->next()) { - if (action->Mentions(reg)) - return true; - } - return false; -} - - -bool Trace::GetStoredPosition(int reg, int* cp_offset) { - DCHECK_EQ(0, *cp_offset); - for (DeferredAction* action = actions_; action != nullptr; - action = action->next()) { - if (action->Mentions(reg)) { - if (action->action_type() == ActionNode::STORE_POSITION) { - *cp_offset = static_cast(action)->cp_offset(); - return true; - } else { - return false; - } - } - } - return false; -} - - -int Trace::FindAffectedRegisters(OutSet* affected_registers, - Zone* zone) { - int max_register = RegExpCompiler::kNoRegister; - for (DeferredAction* action = actions_; action != nullptr; - action = action->next()) { - if (action->action_type() == ActionNode::CLEAR_CAPTURES) { - Interval range = static_cast(action)->range(); - for (int i = range.from(); i <= range.to(); i++) - affected_registers->Set(i, zone); - if (range.to() > max_register) max_register = range.to(); - } else { - affected_registers->Set(action->reg(), zone); - if (action->reg() > max_register) max_register = action->reg(); - } - } - return max_register; -} - - -void Trace::RestoreAffectedRegisters(RegExpMacroAssembler* assembler, - int max_register, - const OutSet& registers_to_pop, - const OutSet& registers_to_clear) { - for (int reg = max_register; reg >= 0; reg--) { - if (registers_to_pop.Get(reg)) { - assembler->PopRegister(reg); - } else if (registers_to_clear.Get(reg)) { - int clear_to = reg; - while (reg > 0 && registers_to_clear.Get(reg - 1)) { - reg--; - } - assembler->ClearRegisters(reg, clear_to); - } - } -} - - -void Trace::PerformDeferredActions(RegExpMacroAssembler* assembler, - int max_register, - const OutSet& affected_registers, - OutSet* registers_to_pop, - OutSet* registers_to_clear, - Zone* zone) { - // The "+1" is to avoid a push_limit of zero if stack_limit_slack() is 1. - const int push_limit = (assembler->stack_limit_slack() + 1) / 2; - - // Count pushes performed to force a stack limit check occasionally. - int pushes = 0; - - for (int reg = 0; reg <= max_register; reg++) { - if (!affected_registers.Get(reg)) { - continue; - } - - // The chronologically first deferred action in the trace - // is used to infer the action needed to restore a register - // to its previous state (or not, if it's safe to ignore it). - enum DeferredActionUndoType { IGNORE, RESTORE, CLEAR }; - DeferredActionUndoType undo_action = IGNORE; - - int value = 0; - bool absolute = false; - bool clear = false; - static const int kNoStore = kMinInt; - int store_position = kNoStore; - // This is a little tricky because we are scanning the actions in reverse - // historical order (newest first). - for (DeferredAction* action = actions_; action != nullptr; - action = action->next()) { - if (action->Mentions(reg)) { - switch (action->action_type()) { - case ActionNode::SET_REGISTER: { - Trace::DeferredSetRegister* psr = - static_cast(action); - if (!absolute) { - value += psr->value(); - absolute = true; - } - // SET_REGISTER is currently only used for newly introduced loop - // counters. They can have a significant previous value if they - // occur in a loop. TODO(lrn): Propagate this information, so - // we can set undo_action to IGNORE if we know there is no value to - // restore. - undo_action = RESTORE; - DCHECK_EQ(store_position, kNoStore); - DCHECK(!clear); - break; - } - case ActionNode::INCREMENT_REGISTER: - if (!absolute) { - value++; - } - DCHECK_EQ(store_position, kNoStore); - DCHECK(!clear); - undo_action = RESTORE; - break; - case ActionNode::STORE_POSITION: { - Trace::DeferredCapture* pc = - static_cast(action); - if (!clear && store_position == kNoStore) { - store_position = pc->cp_offset(); - } - - // For captures we know that stores and clears alternate. - // Other register, are never cleared, and if the occur - // inside a loop, they might be assigned more than once. - if (reg <= 1) { - // Registers zero and one, aka "capture zero", is - // always set correctly if we succeed. There is no - // need to undo a setting on backtrack, because we - // will set it again or fail. - undo_action = IGNORE; - } else { - undo_action = pc->is_capture() ? CLEAR : RESTORE; - } - DCHECK(!absolute); - DCHECK_EQ(value, 0); - break; - } - case ActionNode::CLEAR_CAPTURES: { - // Since we're scanning in reverse order, if we've already - // set the position we have to ignore historically earlier - // clearing operations. - if (store_position == kNoStore) { - clear = true; - } - undo_action = RESTORE; - DCHECK(!absolute); - DCHECK_EQ(value, 0); - break; - } - default: - UNREACHABLE(); - break; - } - } - } - // Prepare for the undo-action (e.g., push if it's going to be popped). - if (undo_action == RESTORE) { - pushes++; - RegExpMacroAssembler::StackCheckFlag stack_check = - RegExpMacroAssembler::kNoStackLimitCheck; - if (pushes == push_limit) { - stack_check = RegExpMacroAssembler::kCheckStackLimit; - pushes = 0; - } - - assembler->PushRegister(reg, stack_check); - registers_to_pop->Set(reg, zone); - } else if (undo_action == CLEAR) { - registers_to_clear->Set(reg, zone); - } - // Perform the chronologically last action (or accumulated increment) - // for the register. - if (store_position != kNoStore) { - assembler->WriteCurrentPositionToRegister(reg, store_position); - } else if (clear) { - assembler->ClearRegisters(reg, reg); - } else if (absolute) { - assembler->SetRegister(reg, value); - } else if (value != 0) { - assembler->AdvanceRegister(reg, value); - } - } -} - - -// This is called as we come into a loop choice node and some other tricky -// nodes. It normalizes the state of the code generator to ensure we can -// generate generic code. -void Trace::Flush(RegExpCompiler* compiler, RegExpNode* successor) { - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - - DCHECK(!is_trivial()); - - if (actions_ == nullptr && backtrack() == nullptr) { - // Here we just have some deferred cp advances to fix and we are back to - // a normal situation. We may also have to forget some information gained - // through a quick check that was already performed. - if (cp_offset_ != 0) assembler->AdvanceCurrentPosition(cp_offset_); - // Create a new trivial state and generate the node with that. - Trace new_state; - successor->Emit(compiler, &new_state); - return; - } - - // Generate deferred actions here along with code to undo them again. - OutSet affected_registers; - - if (backtrack() != nullptr) { - // Here we have a concrete backtrack location. These are set up by choice - // nodes and so they indicate that we have a deferred save of the current - // position which we may need to emit here. - assembler->PushCurrentPosition(); - } - - int max_register = FindAffectedRegisters(&affected_registers, - compiler->zone()); - OutSet registers_to_pop; - OutSet registers_to_clear; - PerformDeferredActions(assembler, - max_register, - affected_registers, - ®isters_to_pop, - ®isters_to_clear, - compiler->zone()); - if (cp_offset_ != 0) { - assembler->AdvanceCurrentPosition(cp_offset_); - } - - // Create a new trivial state and generate the node with that. - Label undo; - assembler->PushBacktrack(&undo); - if (successor->KeepRecursing(compiler)) { - Trace new_state; - successor->Emit(compiler, &new_state); - } else { - compiler->AddWork(successor); - assembler->GoTo(successor->label()); - } - - // On backtrack we need to restore state. - assembler->Bind(&undo); - RestoreAffectedRegisters(assembler, - max_register, - registers_to_pop, - registers_to_clear); - if (backtrack() == nullptr) { - assembler->Backtrack(); - } else { - assembler->PopCurrentPosition(); - assembler->GoTo(backtrack()); - } -} - - -void NegativeSubmatchSuccess::Emit(RegExpCompiler* compiler, Trace* trace) { - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - - // Omit flushing the trace. We discard the entire stack frame anyway. - - if (!label()->is_bound()) { - // We are completely independent of the trace, since we ignore it, - // so this code can be used as the generic version. - assembler->Bind(label()); - } - - // Throw away everything on the backtrack stack since the start - // of the negative submatch and restore the character position. - assembler->ReadCurrentPositionFromRegister(current_position_register_); - assembler->ReadStackPointerFromRegister(stack_pointer_register_); - if (clear_capture_count_ > 0) { - // Clear any captures that might have been performed during the success - // of the body of the negative look-ahead. - int clear_capture_end = clear_capture_start_ + clear_capture_count_ - 1; - assembler->ClearRegisters(clear_capture_start_, clear_capture_end); - } - // Now that we have unwound the stack we find at the top of the stack the - // backtrack that the BeginSubmatch node got. - assembler->Backtrack(); -} - - -void EndNode::Emit(RegExpCompiler* compiler, Trace* trace) { - if (!trace->is_trivial()) { - trace->Flush(compiler, this); - return; - } - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - if (!label()->is_bound()) { - assembler->Bind(label()); - } - switch (action_) { - case ACCEPT: - assembler->Succeed(); - return; - case BACKTRACK: - assembler->GoTo(trace->backtrack()); - return; - case NEGATIVE_SUBMATCH_SUCCESS: - // This case is handled in a different virtual method. - UNREACHABLE(); - } - UNIMPLEMENTED(); -} - - -void GuardedAlternative::AddGuard(Guard* guard, Zone* zone) { - if (guards_ == nullptr) guards_ = new (zone) ZoneList(1, zone); - guards_->Add(guard, zone); -} - - -ActionNode* ActionNode::SetRegister(int reg, - int val, - RegExpNode* on_success) { - ActionNode* result = - new(on_success->zone()) ActionNode(SET_REGISTER, on_success); - result->data_.u_store_register.reg = reg; - result->data_.u_store_register.value = val; - return result; -} - - -ActionNode* ActionNode::IncrementRegister(int reg, RegExpNode* on_success) { - ActionNode* result = - new(on_success->zone()) ActionNode(INCREMENT_REGISTER, on_success); - result->data_.u_increment_register.reg = reg; - return result; -} - - -ActionNode* ActionNode::StorePosition(int reg, - bool is_capture, - RegExpNode* on_success) { - ActionNode* result = - new(on_success->zone()) ActionNode(STORE_POSITION, on_success); - result->data_.u_position_register.reg = reg; - result->data_.u_position_register.is_capture = is_capture; - return result; -} - - -ActionNode* ActionNode::ClearCaptures(Interval range, - RegExpNode* on_success) { - ActionNode* result = - new(on_success->zone()) ActionNode(CLEAR_CAPTURES, on_success); - result->data_.u_clear_captures.range_from = range.from(); - result->data_.u_clear_captures.range_to = range.to(); - return result; -} - - -ActionNode* ActionNode::BeginSubmatch(int stack_reg, - int position_reg, - RegExpNode* on_success) { - ActionNode* result = - new(on_success->zone()) ActionNode(BEGIN_SUBMATCH, on_success); - result->data_.u_submatch.stack_pointer_register = stack_reg; - result->data_.u_submatch.current_position_register = position_reg; - return result; -} - - -ActionNode* ActionNode::PositiveSubmatchSuccess(int stack_reg, - int position_reg, - int clear_register_count, - int clear_register_from, - RegExpNode* on_success) { - ActionNode* result = - new(on_success->zone()) ActionNode(POSITIVE_SUBMATCH_SUCCESS, on_success); - result->data_.u_submatch.stack_pointer_register = stack_reg; - result->data_.u_submatch.current_position_register = position_reg; - result->data_.u_submatch.clear_register_count = clear_register_count; - result->data_.u_submatch.clear_register_from = clear_register_from; - return result; -} - - -ActionNode* ActionNode::EmptyMatchCheck(int start_register, - int repetition_register, - int repetition_limit, - RegExpNode* on_success) { - ActionNode* result = - new(on_success->zone()) ActionNode(EMPTY_MATCH_CHECK, on_success); - result->data_.u_empty_match_check.start_register = start_register; - result->data_.u_empty_match_check.repetition_register = repetition_register; - result->data_.u_empty_match_check.repetition_limit = repetition_limit; - return result; -} - - -#define DEFINE_ACCEPT(Type) \ - void Type##Node::Accept(NodeVisitor* visitor) { \ - visitor->Visit##Type(this); \ - } -FOR_EACH_NODE_TYPE(DEFINE_ACCEPT) -#undef DEFINE_ACCEPT - - -void LoopChoiceNode::Accept(NodeVisitor* visitor) { - visitor->VisitLoopChoice(this); -} - - -// ------------------------------------------------------------------- -// Emit code. - - -void ChoiceNode::GenerateGuard(RegExpMacroAssembler* macro_assembler, - Guard* guard, - Trace* trace) { - switch (guard->op()) { - case Guard::LT: - DCHECK(!trace->mentions_reg(guard->reg())); - macro_assembler->IfRegisterGE(guard->reg(), - guard->value(), - trace->backtrack()); - break; - case Guard::GEQ: - DCHECK(!trace->mentions_reg(guard->reg())); - macro_assembler->IfRegisterLT(guard->reg(), - guard->value(), - trace->backtrack()); - break; - } -} - - -// Returns the number of characters in the equivalence class, omitting those -// that cannot occur in the source string because it is Latin1. -static int GetCaseIndependentLetters(Isolate* isolate, uc16 character, - bool one_byte_subject, - unibrow::uchar* letters, - int letter_length) { -#ifdef V8_INTL_SUPPORT - icu::UnicodeSet set; - set.add(character); - set = set.closeOver(USET_CASE_INSENSITIVE); - int32_t range_count = set.getRangeCount(); - int items = 0; - for (int32_t i = 0; i < range_count; i++) { - UChar32 start = set.getRangeStart(i); - UChar32 end = set.getRangeEnd(i); - CHECK(end - start + items <= letter_length); - while (start <= end) { - if (one_byte_subject && start > String::kMaxOneByteCharCode) break; - letters[items++] = (unibrow::uchar)(start); - start++; - } - } - return items; -#else - int length = - isolate->jsregexp_uncanonicalize()->get(character, '\0', letters); - // Unibrow returns 0 or 1 for characters where case independence is - // trivial. - if (length == 0) { - letters[0] = character; - length = 1; - } - - if (one_byte_subject) { - int new_length = 0; - for (int i = 0; i < length; i++) { - if (letters[i] <= String::kMaxOneByteCharCode) { - letters[new_length++] = letters[i]; - } - } - length = new_length; - } - - return length; -#endif // V8_INTL_SUPPORT -} - -static inline bool EmitSimpleCharacter(Isolate* isolate, - RegExpCompiler* compiler, - uc16 c, - Label* on_failure, - int cp_offset, - bool check, - bool preloaded) { - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - bool bound_checked = false; - if (!preloaded) { - assembler->LoadCurrentCharacter( - cp_offset, - on_failure, - check); - bound_checked = true; - } - assembler->CheckNotCharacter(c, on_failure); - return bound_checked; -} - - -// Only emits non-letters (things that don't have case). Only used for case -// independent matches. -static inline bool EmitAtomNonLetter(Isolate* isolate, - RegExpCompiler* compiler, - uc16 c, - Label* on_failure, - int cp_offset, - bool check, - bool preloaded) { - RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); - bool one_byte = compiler->one_byte(); - unibrow::uchar chars[4]; - int length = GetCaseIndependentLetters(isolate, c, one_byte, chars, 4); - if (length < 1) { - // This can't match. Must be an one-byte subject and a non-one-byte - // character. We do not need to do anything since the one-byte pass - // already handled this. - return false; // Bounds not checked. - } - bool checked = false; - // We handle the length > 1 case in a later pass. - if (length == 1) { - if (one_byte && c > String::kMaxOneByteCharCodeU) { - // Can't match - see above. - return false; // Bounds not checked. - } - if (!preloaded) { - macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check); - checked = check; - } - macro_assembler->CheckNotCharacter(c, on_failure); - } - return checked; -} - - -static bool ShortCutEmitCharacterPair(RegExpMacroAssembler* macro_assembler, - bool one_byte, uc16 c1, uc16 c2, - Label* on_failure) { - uc16 char_mask; - if (one_byte) { - char_mask = String::kMaxOneByteCharCode; - } else { - char_mask = String::kMaxUtf16CodeUnit; - } - uc16 exor = c1 ^ c2; - // Check whether exor has only one bit set. - if (((exor - 1) & exor) == 0) { - // If c1 and c2 differ only by one bit. - // Ecma262UnCanonicalize always gives the highest number last. - DCHECK(c2 > c1); - uc16 mask = char_mask ^ exor; - macro_assembler->CheckNotCharacterAfterAnd(c1, mask, on_failure); - return true; - } - DCHECK(c2 > c1); - uc16 diff = c2 - c1; - if (((diff - 1) & diff) == 0 && c1 >= diff) { - // If the characters differ by 2^n but don't differ by one bit then - // subtract the difference from the found character, then do the or - // trick. We avoid the theoretical case where negative numbers are - // involved in order to simplify code generation. - uc16 mask = char_mask ^ diff; - macro_assembler->CheckNotCharacterAfterMinusAnd(c1 - diff, - diff, - mask, - on_failure); - return true; - } - return false; -} - -using EmitCharacterFunction = bool(Isolate* isolate, RegExpCompiler* compiler, - uc16 c, Label* on_failure, int cp_offset, - bool check, bool preloaded); - -// Only emits letters (things that have case). Only used for case independent -// matches. -static inline bool EmitAtomLetter(Isolate* isolate, - RegExpCompiler* compiler, - uc16 c, - Label* on_failure, - int cp_offset, - bool check, - bool preloaded) { - RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); - bool one_byte = compiler->one_byte(); - unibrow::uchar chars[4]; - int length = GetCaseIndependentLetters(isolate, c, one_byte, chars, 4); - if (length <= 1) return false; - // We may not need to check against the end of the input string - // if this character lies before a character that matched. - if (!preloaded) { - macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check); - } - Label ok; - switch (length) { - case 2: { - if (ShortCutEmitCharacterPair(macro_assembler, one_byte, chars[0], - chars[1], on_failure)) { - } else { - macro_assembler->CheckCharacter(chars[0], &ok); - macro_assembler->CheckNotCharacter(chars[1], on_failure); - macro_assembler->Bind(&ok); - } - break; - } - case 4: - macro_assembler->CheckCharacter(chars[3], &ok); - V8_FALLTHROUGH; - case 3: - macro_assembler->CheckCharacter(chars[0], &ok); - macro_assembler->CheckCharacter(chars[1], &ok); - macro_assembler->CheckNotCharacter(chars[2], on_failure); - macro_assembler->Bind(&ok); - break; - default: - UNREACHABLE(); - } - return true; -} - - -static void EmitBoundaryTest(RegExpMacroAssembler* masm, - int border, - Label* fall_through, - Label* above_or_equal, - Label* below) { - if (below != fall_through) { - masm->CheckCharacterLT(border, below); - if (above_or_equal != fall_through) masm->GoTo(above_or_equal); - } else { - masm->CheckCharacterGT(border - 1, above_or_equal); - } -} - - -static void EmitDoubleBoundaryTest(RegExpMacroAssembler* masm, - int first, - int last, - Label* fall_through, - Label* in_range, - Label* out_of_range) { - if (in_range == fall_through) { - if (first == last) { - masm->CheckNotCharacter(first, out_of_range); - } else { - masm->CheckCharacterNotInRange(first, last, out_of_range); - } - } else { - if (first == last) { - masm->CheckCharacter(first, in_range); - } else { - masm->CheckCharacterInRange(first, last, in_range); - } - if (out_of_range != fall_through) masm->GoTo(out_of_range); - } -} - - -// even_label is for ranges[i] to ranges[i + 1] where i - start_index is even. -// odd_label is for ranges[i] to ranges[i + 1] where i - start_index is odd. -static void EmitUseLookupTable( - RegExpMacroAssembler* masm, - ZoneList* ranges, - int start_index, - int end_index, - int min_char, - Label* fall_through, - Label* even_label, - Label* odd_label) { - static const int kSize = RegExpMacroAssembler::kTableSize; - static const int kMask = RegExpMacroAssembler::kTableMask; - - int base = (min_char & ~kMask); - USE(base); - - // Assert that everything is on one kTableSize page. - for (int i = start_index; i <= end_index; i++) { - DCHECK_EQ(ranges->at(i) & ~kMask, base); - } - DCHECK(start_index == 0 || (ranges->at(start_index - 1) & ~kMask) <= base); - - char templ[kSize]; - Label* on_bit_set; - Label* on_bit_clear; - int bit; - if (even_label == fall_through) { - on_bit_set = odd_label; - on_bit_clear = even_label; - bit = 1; - } else { - on_bit_set = even_label; - on_bit_clear = odd_label; - bit = 0; - } - for (int i = 0; i < (ranges->at(start_index) & kMask) && i < kSize; i++) { - templ[i] = bit; - } - int j = 0; - bit ^= 1; - for (int i = start_index; i < end_index; i++) { - for (j = (ranges->at(i) & kMask); j < (ranges->at(i + 1) & kMask); j++) { - templ[j] = bit; - } - bit ^= 1; - } - for (int i = j; i < kSize; i++) { - templ[i] = bit; - } - Factory* factory = masm->isolate()->factory(); - // TODO(erikcorry): Cache these. - Handle ba = factory->NewByteArray(kSize, AllocationType::kOld); - for (int i = 0; i < kSize; i++) { - ba->set(i, templ[i]); - } - masm->CheckBitInTable(ba, on_bit_set); - if (on_bit_clear != fall_through) masm->GoTo(on_bit_clear); -} - - -static void CutOutRange(RegExpMacroAssembler* masm, - ZoneList* ranges, - int start_index, - int end_index, - int cut_index, - Label* even_label, - Label* odd_label) { - bool odd = (((cut_index - start_index) & 1) == 1); - Label* in_range_label = odd ? odd_label : even_label; - Label dummy; - EmitDoubleBoundaryTest(masm, - ranges->at(cut_index), - ranges->at(cut_index + 1) - 1, - &dummy, - in_range_label, - &dummy); - DCHECK(!dummy.is_linked()); - // Cut out the single range by rewriting the array. This creates a new - // range that is a merger of the two ranges on either side of the one we - // are cutting out. The oddity of the labels is preserved. - for (int j = cut_index; j > start_index; j--) { - ranges->at(j) = ranges->at(j - 1); - } - for (int j = cut_index + 1; j < end_index; j++) { - ranges->at(j) = ranges->at(j + 1); - } -} - - -// Unicode case. Split the search space into kSize spaces that are handled -// with recursion. -static void SplitSearchSpace(ZoneList* ranges, - int start_index, - int end_index, - int* new_start_index, - int* new_end_index, - int* border) { - static const int kSize = RegExpMacroAssembler::kTableSize; - static const int kMask = RegExpMacroAssembler::kTableMask; - - int first = ranges->at(start_index); - int last = ranges->at(end_index) - 1; - - *new_start_index = start_index; - *border = (ranges->at(start_index) & ~kMask) + kSize; - while (*new_start_index < end_index) { - if (ranges->at(*new_start_index) > *border) break; - (*new_start_index)++; - } - // new_start_index is the index of the first edge that is beyond the - // current kSize space. - - // For very large search spaces we do a binary chop search of the non-Latin1 - // space instead of just going to the end of the current kSize space. The - // heuristics are complicated a little by the fact that any 128-character - // encoding space can be quickly tested with a table lookup, so we don't - // wish to do binary chop search at a smaller granularity than that. A - // 128-character space can take up a lot of space in the ranges array if, - // for example, we only want to match every second character (eg. the lower - // case characters on some Unicode pages). - int binary_chop_index = (end_index + start_index) / 2; - // The first test ensures that we get to the code that handles the Latin1 - // range with a single not-taken branch, speeding up this important - // character range (even non-Latin1 charset-based text has spaces and - // punctuation). - if (*border - 1 > String::kMaxOneByteCharCode && // Latin1 case. - end_index - start_index > (*new_start_index - start_index) * 2 && - last - first > kSize * 2 && binary_chop_index > *new_start_index && - ranges->at(binary_chop_index) >= first + 2 * kSize) { - int scan_forward_for_section_border = binary_chop_index;; - int new_border = (ranges->at(binary_chop_index) | kMask) + 1; - - while (scan_forward_for_section_border < end_index) { - if (ranges->at(scan_forward_for_section_border) > new_border) { - *new_start_index = scan_forward_for_section_border; - *border = new_border; - break; - } - scan_forward_for_section_border++; - } - } - - DCHECK(*new_start_index > start_index); - *new_end_index = *new_start_index - 1; - if (ranges->at(*new_end_index) == *border) { - (*new_end_index)--; - } - if (*border >= ranges->at(end_index)) { - *border = ranges->at(end_index); - *new_start_index = end_index; // Won't be used. - *new_end_index = end_index - 1; - } -} - -// Gets a series of segment boundaries representing a character class. If the -// character is in the range between an even and an odd boundary (counting from -// start_index) then go to even_label, otherwise go to odd_label. We already -// know that the character is in the range of min_char to max_char inclusive. -// Either label can be nullptr indicating backtracking. Either label can also -// be equal to the fall_through label. -static void GenerateBranches(RegExpMacroAssembler* masm, ZoneList* ranges, - int start_index, int end_index, uc32 min_char, - uc32 max_char, Label* fall_through, - Label* even_label, Label* odd_label) { - DCHECK_LE(min_char, String::kMaxUtf16CodeUnit); - DCHECK_LE(max_char, String::kMaxUtf16CodeUnit); - - int first = ranges->at(start_index); - int last = ranges->at(end_index) - 1; - - DCHECK_LT(min_char, first); - - // Just need to test if the character is before or on-or-after - // a particular character. - if (start_index == end_index) { - EmitBoundaryTest(masm, first, fall_through, even_label, odd_label); - return; - } - - // Another almost trivial case: There is one interval in the middle that is - // different from the end intervals. - if (start_index + 1 == end_index) { - EmitDoubleBoundaryTest( - masm, first, last, fall_through, even_label, odd_label); - return; - } - - // It's not worth using table lookup if there are very few intervals in the - // character class. - if (end_index - start_index <= 6) { - // It is faster to test for individual characters, so we look for those - // first, then try arbitrary ranges in the second round. - static int kNoCutIndex = -1; - int cut = kNoCutIndex; - for (int i = start_index; i < end_index; i++) { - if (ranges->at(i) == ranges->at(i + 1) - 1) { - cut = i; - break; - } - } - if (cut == kNoCutIndex) cut = start_index; - CutOutRange( - masm, ranges, start_index, end_index, cut, even_label, odd_label); - DCHECK_GE(end_index - start_index, 2); - GenerateBranches(masm, - ranges, - start_index + 1, - end_index - 1, - min_char, - max_char, - fall_through, - even_label, - odd_label); - return; - } - - // If there are a lot of intervals in the regexp, then we will use tables to - // determine whether the character is inside or outside the character class. - static const int kBits = RegExpMacroAssembler::kTableSizeBits; - - if ((max_char >> kBits) == (min_char >> kBits)) { - EmitUseLookupTable(masm, - ranges, - start_index, - end_index, - min_char, - fall_through, - even_label, - odd_label); - return; - } - - if ((min_char >> kBits) != (first >> kBits)) { - masm->CheckCharacterLT(first, odd_label); - GenerateBranches(masm, - ranges, - start_index + 1, - end_index, - first, - max_char, - fall_through, - odd_label, - even_label); - return; - } - - int new_start_index = 0; - int new_end_index = 0; - int border = 0; - - SplitSearchSpace(ranges, - start_index, - end_index, - &new_start_index, - &new_end_index, - &border); - - Label handle_rest; - Label* above = &handle_rest; - if (border == last + 1) { - // We didn't find any section that started after the limit, so everything - // above the border is one of the terminal labels. - above = (end_index & 1) != (start_index & 1) ? odd_label : even_label; - DCHECK(new_end_index == end_index - 1); - } - - DCHECK_LE(start_index, new_end_index); - DCHECK_LE(new_start_index, end_index); - DCHECK_LT(start_index, new_start_index); - DCHECK_LT(new_end_index, end_index); - DCHECK(new_end_index + 1 == new_start_index || - (new_end_index + 2 == new_start_index && - border == ranges->at(new_end_index + 1))); - DCHECK_LT(min_char, border - 1); - DCHECK_LT(border, max_char); - DCHECK_LT(ranges->at(new_end_index), border); - DCHECK(border < ranges->at(new_start_index) || - (border == ranges->at(new_start_index) && - new_start_index == end_index && - new_end_index == end_index - 1 && - border == last + 1)); - DCHECK(new_start_index == 0 || border >= ranges->at(new_start_index - 1)); - - masm->CheckCharacterGT(border - 1, above); - Label dummy; - GenerateBranches(masm, - ranges, - start_index, - new_end_index, - min_char, - border - 1, - &dummy, - even_label, - odd_label); - if (handle_rest.is_linked()) { - masm->Bind(&handle_rest); - bool flip = (new_start_index & 1) != (start_index & 1); - GenerateBranches(masm, - ranges, - new_start_index, - end_index, - border, - max_char, - &dummy, - flip ? odd_label : even_label, - flip ? even_label : odd_label); - } -} - - -static void EmitCharClass(RegExpMacroAssembler* macro_assembler, - RegExpCharacterClass* cc, bool one_byte, - Label* on_failure, int cp_offset, bool check_offset, - bool preloaded, Zone* zone) { - ZoneList* ranges = cc->ranges(zone); - CharacterRange::Canonicalize(ranges); - - int max_char; - if (one_byte) { - max_char = String::kMaxOneByteCharCode; - } else { - max_char = String::kMaxUtf16CodeUnit; - } - - int range_count = ranges->length(); - - int last_valid_range = range_count - 1; - while (last_valid_range >= 0) { - CharacterRange& range = ranges->at(last_valid_range); - if (range.from() <= max_char) { - break; - } - last_valid_range--; - } - - if (last_valid_range < 0) { - if (!cc->is_negated()) { - macro_assembler->GoTo(on_failure); - } - if (check_offset) { - macro_assembler->CheckPosition(cp_offset, on_failure); - } - return; - } - - if (last_valid_range == 0 && - ranges->at(0).IsEverything(max_char)) { - if (cc->is_negated()) { - macro_assembler->GoTo(on_failure); - } else { - // This is a common case hit by non-anchored expressions. - if (check_offset) { - macro_assembler->CheckPosition(cp_offset, on_failure); - } - } - return; - } - - if (!preloaded) { - macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check_offset); - } - - if (cc->is_standard(zone) && - macro_assembler->CheckSpecialCharacterClass(cc->standard_type(), - on_failure)) { - return; - } - - - // A new list with ascending entries. Each entry is a code unit - // where there is a boundary between code units that are part of - // the class and code units that are not. Normally we insert an - // entry at zero which goes to the failure label, but if there - // was already one there we fall through for success on that entry. - // Subsequent entries have alternating meaning (success/failure). - ZoneList* range_boundaries = - new(zone) ZoneList(last_valid_range, zone); - - bool zeroth_entry_is_failure = !cc->is_negated(); - - for (int i = 0; i <= last_valid_range; i++) { - CharacterRange& range = ranges->at(i); - if (range.from() == 0) { - DCHECK_EQ(i, 0); - zeroth_entry_is_failure = !zeroth_entry_is_failure; - } else { - range_boundaries->Add(range.from(), zone); - } - range_boundaries->Add(range.to() + 1, zone); - } - int end_index = range_boundaries->length() - 1; - if (range_boundaries->at(end_index) > max_char) { - end_index--; - } - - Label fall_through; - GenerateBranches(macro_assembler, - range_boundaries, - 0, // start_index. - end_index, - 0, // min_char. - max_char, - &fall_through, - zeroth_entry_is_failure ? &fall_through : on_failure, - zeroth_entry_is_failure ? on_failure : &fall_through); - macro_assembler->Bind(&fall_through); -} - -RegExpNode::~RegExpNode() = default; - -RegExpNode::LimitResult RegExpNode::LimitVersions(RegExpCompiler* compiler, - Trace* trace) { - // If we are generating a greedy loop then don't stop and don't reuse code. - if (trace->stop_node() != nullptr) { - return CONTINUE; - } - - RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); - if (trace->is_trivial()) { - if (label_.is_bound() || on_work_list() || !KeepRecursing(compiler)) { - // If a generic version is already scheduled to be generated or we have - // recursed too deeply then just generate a jump to that code. - macro_assembler->GoTo(&label_); - // This will queue it up for generation of a generic version if it hasn't - // already been queued. - compiler->AddWork(this); - return DONE; - } - // Generate generic version of the node and bind the label for later use. - macro_assembler->Bind(&label_); - return CONTINUE; - } - - // We are being asked to make a non-generic version. Keep track of how many - // non-generic versions we generate so as not to overdo it. - trace_count_++; - if (KeepRecursing(compiler) && compiler->optimize() && - trace_count_ < kMaxCopiesCodeGenerated) { - return CONTINUE; - } - - // If we get here code has been generated for this node too many times or - // recursion is too deep. Time to switch to a generic version. The code for - // generic versions above can handle deep recursion properly. - bool was_limiting = compiler->limiting_recursion(); - compiler->set_limiting_recursion(true); - trace->Flush(compiler, this); - compiler->set_limiting_recursion(was_limiting); - return DONE; -} - - -bool RegExpNode::KeepRecursing(RegExpCompiler* compiler) { - return !compiler->limiting_recursion() && - compiler->recursion_depth() <= RegExpCompiler::kMaxRecursion; -} - - -int ActionNode::EatsAtLeast(int still_to_find, - int budget, - bool not_at_start) { - if (budget <= 0) return 0; - if (action_type_ == POSITIVE_SUBMATCH_SUCCESS) return 0; // Rewinds input! - return on_success()->EatsAtLeast(still_to_find, - budget - 1, - not_at_start); -} - - -void ActionNode::FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) { - if (action_type_ != POSITIVE_SUBMATCH_SUCCESS) { - on_success()->FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start); - } - SaveBMInfo(bm, not_at_start, offset); -} - - -int AssertionNode::EatsAtLeast(int still_to_find, - int budget, - bool not_at_start) { - if (budget <= 0) return 0; - // If we know we are not at the start and we are asked "how many characters - // will you match if you succeed?" then we can answer anything since false - // implies false. So lets just return the max answer (still_to_find) since - // that won't prevent us from preloading a lot of characters for the other - // branches in the node graph. - if (assertion_type() == AT_START && not_at_start) return still_to_find; - return on_success()->EatsAtLeast(still_to_find, - budget - 1, - not_at_start); -} - - -void AssertionNode::FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) { - // Match the behaviour of EatsAtLeast on this node. - if (assertion_type() == AT_START && not_at_start) return; - on_success()->FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start); - SaveBMInfo(bm, not_at_start, offset); -} - - -int BackReferenceNode::EatsAtLeast(int still_to_find, - int budget, - bool not_at_start) { - if (read_backward()) return 0; - if (budget <= 0) return 0; - return on_success()->EatsAtLeast(still_to_find, - budget - 1, - not_at_start); -} - - -int TextNode::EatsAtLeast(int still_to_find, - int budget, - bool not_at_start) { - if (read_backward()) return 0; - int answer = Length(); - if (answer >= still_to_find) return answer; - if (budget <= 0) return answer; - // We are not at start after this node so we set the last argument to 'true'. - return answer + on_success()->EatsAtLeast(still_to_find - answer, - budget - 1, - true); -} - - -int NegativeLookaroundChoiceNode::EatsAtLeast(int still_to_find, int budget, - bool not_at_start) { - if (budget <= 0) return 0; - // Alternative 0 is the negative lookahead, alternative 1 is what comes - // afterwards. - RegExpNode* node = alternatives_->at(1).node(); - return node->EatsAtLeast(still_to_find, budget - 1, not_at_start); -} - - -void NegativeLookaroundChoiceNode::GetQuickCheckDetails( - QuickCheckDetails* details, RegExpCompiler* compiler, int filled_in, - bool not_at_start) { - // Alternative 0 is the negative lookahead, alternative 1 is what comes - // afterwards. - RegExpNode* node = alternatives_->at(1).node(); - return node->GetQuickCheckDetails(details, compiler, filled_in, not_at_start); -} - - -int ChoiceNode::EatsAtLeastHelper(int still_to_find, - int budget, - RegExpNode* ignore_this_node, - bool not_at_start) { - if (budget <= 0) return 0; - int min = 100; - int choice_count = alternatives_->length(); - budget = (budget - 1) / choice_count; - for (int i = 0; i < choice_count; i++) { - RegExpNode* node = alternatives_->at(i).node(); - if (node == ignore_this_node) continue; - int node_eats_at_least = - node->EatsAtLeast(still_to_find, budget, not_at_start); - if (node_eats_at_least < min) min = node_eats_at_least; - if (min == 0) return 0; - } - return min; -} - - -int LoopChoiceNode::EatsAtLeast(int still_to_find, - int budget, - bool not_at_start) { - return EatsAtLeastHelper(still_to_find, - budget - 1, - loop_node_, - not_at_start); -} - - -int ChoiceNode::EatsAtLeast(int still_to_find, - int budget, - bool not_at_start) { - return EatsAtLeastHelper(still_to_find, budget, nullptr, not_at_start); -} - - -// Takes the left-most 1-bit and smears it out, setting all bits to its right. -static inline uint32_t SmearBitsRight(uint32_t v) { - v |= v >> 1; - v |= v >> 2; - v |= v >> 4; - v |= v >> 8; - v |= v >> 16; - return v; -} - - -bool QuickCheckDetails::Rationalize(bool asc) { - bool found_useful_op = false; - uint32_t char_mask; - if (asc) { - char_mask = String::kMaxOneByteCharCode; - } else { - char_mask = String::kMaxUtf16CodeUnit; - } - mask_ = 0; - value_ = 0; - int char_shift = 0; - for (int i = 0; i < characters_; i++) { - Position* pos = &positions_[i]; - if ((pos->mask & String::kMaxOneByteCharCode) != 0) { - found_useful_op = true; - } - mask_ |= (pos->mask & char_mask) << char_shift; - value_ |= (pos->value & char_mask) << char_shift; - char_shift += asc ? 8 : 16; - } - return found_useful_op; -} - - -bool RegExpNode::EmitQuickCheck(RegExpCompiler* compiler, - Trace* bounds_check_trace, - Trace* trace, - bool preload_has_checked_bounds, - Label* on_possible_success, - QuickCheckDetails* details, - bool fall_through_on_failure) { - if (details->characters() == 0) return false; - GetQuickCheckDetails( - details, compiler, 0, trace->at_start() == Trace::FALSE_VALUE); - if (details->cannot_match()) return false; - if (!details->Rationalize(compiler->one_byte())) return false; - DCHECK(details->characters() == 1 || - compiler->macro_assembler()->CanReadUnaligned()); - uint32_t mask = details->mask(); - uint32_t value = details->value(); - - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - - if (trace->characters_preloaded() != details->characters()) { - DCHECK(trace->cp_offset() == bounds_check_trace->cp_offset()); - // We are attempting to preload the minimum number of characters - // any choice would eat, so if the bounds check fails, then none of the - // choices can succeed, so we can just immediately backtrack, rather - // than go to the next choice. - assembler->LoadCurrentCharacter(trace->cp_offset(), - bounds_check_trace->backtrack(), - !preload_has_checked_bounds, - details->characters()); - } - - - bool need_mask = true; - - if (details->characters() == 1) { - // If number of characters preloaded is 1 then we used a byte or 16 bit - // load so the value is already masked down. - uint32_t char_mask; - if (compiler->one_byte()) { - char_mask = String::kMaxOneByteCharCode; - } else { - char_mask = String::kMaxUtf16CodeUnit; - } - if ((mask & char_mask) == char_mask) need_mask = false; - mask &= char_mask; - } else { - // For 2-character preloads in one-byte mode or 1-character preloads in - // two-byte mode we also use a 16 bit load with zero extend. - static const uint32_t kTwoByteMask = 0xFFFF; - static const uint32_t kFourByteMask = 0xFFFFFFFF; - if (details->characters() == 2 && compiler->one_byte()) { - if ((mask & kTwoByteMask) == kTwoByteMask) need_mask = false; - } else if (details->characters() == 1 && !compiler->one_byte()) { - if ((mask & kTwoByteMask) == kTwoByteMask) need_mask = false; - } else { - if (mask == kFourByteMask) need_mask = false; - } - } - - if (fall_through_on_failure) { - if (need_mask) { - assembler->CheckCharacterAfterAnd(value, mask, on_possible_success); - } else { - assembler->CheckCharacter(value, on_possible_success); - } - } else { - if (need_mask) { - assembler->CheckNotCharacterAfterAnd(value, mask, trace->backtrack()); - } else { - assembler->CheckNotCharacter(value, trace->backtrack()); - } - } - return true; -} - - -// Here is the meat of GetQuickCheckDetails (see also the comment on the -// super-class in the .h file). -// -// We iterate along the text object, building up for each character a -// mask and value that can be used to test for a quick failure to match. -// The masks and values for the positions will be combined into a single -// machine word for the current character width in order to be used in -// generating a quick check. -void TextNode::GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, - int characters_filled_in, - bool not_at_start) { - // Do not collect any quick check details if the text node reads backward, - // since it reads in the opposite direction than we use for quick checks. - if (read_backward()) return; - Isolate* isolate = compiler->macro_assembler()->isolate(); - DCHECK(characters_filled_in < details->characters()); - int characters = details->characters(); - int char_mask; - if (compiler->one_byte()) { - char_mask = String::kMaxOneByteCharCode; - } else { - char_mask = String::kMaxUtf16CodeUnit; - } - for (int k = 0; k < elements()->length(); k++) { - TextElement elm = elements()->at(k); - if (elm.text_type() == TextElement::ATOM) { - Vector quarks = elm.atom()->data(); - for (int i = 0; i < characters && i < quarks.length(); i++) { - QuickCheckDetails::Position* pos = - details->positions(characters_filled_in); - uc16 c = quarks[i]; - if (elm.atom()->ignore_case()) { - unibrow::uchar chars[4]; - int length = GetCaseIndependentLetters( - isolate, c, compiler->one_byte(), chars, 4); - if (length == 0) { - // This can happen because all case variants are non-Latin1, but we - // know the input is Latin1. - details->set_cannot_match(); - pos->determines_perfectly = false; - return; - } - if (length == 1) { - // This letter has no case equivalents, so it's nice and simple - // and the mask-compare will determine definitely whether we have - // a match at this character position. - pos->mask = char_mask; - pos->value = c; - pos->determines_perfectly = true; - } else { - uint32_t common_bits = char_mask; - uint32_t bits = chars[0]; - for (int j = 1; j < length; j++) { - uint32_t differing_bits = ((chars[j] & common_bits) ^ bits); - common_bits ^= differing_bits; - bits &= common_bits; - } - // If length is 2 and common bits has only one zero in it then - // our mask and compare instruction will determine definitely - // whether we have a match at this character position. Otherwise - // it can only be an approximate check. - uint32_t one_zero = (common_bits | ~char_mask); - if (length == 2 && ((~one_zero) & ((~one_zero) - 1)) == 0) { - pos->determines_perfectly = true; - } - pos->mask = common_bits; - pos->value = bits; - } - } else { - // Don't ignore case. Nice simple case where the mask-compare will - // determine definitely whether we have a match at this character - // position. - if (c > char_mask) { - details->set_cannot_match(); - pos->determines_perfectly = false; - return; - } - pos->mask = char_mask; - pos->value = c; - pos->determines_perfectly = true; - } - characters_filled_in++; - DCHECK(characters_filled_in <= details->characters()); - if (characters_filled_in == details->characters()) { - return; - } - } - } else { - QuickCheckDetails::Position* pos = - details->positions(characters_filled_in); - RegExpCharacterClass* tree = elm.char_class(); - ZoneList* ranges = tree->ranges(zone()); - DCHECK(!ranges->is_empty()); - if (tree->is_negated()) { - // A quick check uses multi-character mask and compare. There is no - // useful way to incorporate a negative char class into this scheme - // so we just conservatively create a mask and value that will always - // succeed. - pos->mask = 0; - pos->value = 0; - } else { - int first_range = 0; - while (ranges->at(first_range).from() > char_mask) { - first_range++; - if (first_range == ranges->length()) { - details->set_cannot_match(); - pos->determines_perfectly = false; - return; - } - } - CharacterRange range = ranges->at(first_range); - uc16 from = range.from(); - uc16 to = range.to(); - if (to > char_mask) { - to = char_mask; - } - uint32_t differing_bits = (from ^ to); - // A mask and compare is only perfect if the differing bits form a - // number like 00011111 with one single block of trailing 1s. - if ((differing_bits & (differing_bits + 1)) == 0 && - from + differing_bits == to) { - pos->determines_perfectly = true; - } - uint32_t common_bits = ~SmearBitsRight(differing_bits); - uint32_t bits = (from & common_bits); - for (int i = first_range + 1; i < ranges->length(); i++) { - CharacterRange range = ranges->at(i); - uc16 from = range.from(); - uc16 to = range.to(); - if (from > char_mask) continue; - if (to > char_mask) to = char_mask; - // Here we are combining more ranges into the mask and compare - // value. With each new range the mask becomes more sparse and - // so the chances of a false positive rise. A character class - // with multiple ranges is assumed never to be equivalent to a - // mask and compare operation. - pos->determines_perfectly = false; - uint32_t new_common_bits = (from ^ to); - new_common_bits = ~SmearBitsRight(new_common_bits); - common_bits &= new_common_bits; - bits &= new_common_bits; - uint32_t differing_bits = (from & common_bits) ^ bits; - common_bits ^= differing_bits; - bits &= common_bits; - } - pos->mask = common_bits; - pos->value = bits; - } - characters_filled_in++; - DCHECK(characters_filled_in <= details->characters()); - if (characters_filled_in == details->characters()) { - return; - } - } - } - DCHECK(characters_filled_in != details->characters()); - if (!details->cannot_match()) { - on_success()-> GetQuickCheckDetails(details, - compiler, - characters_filled_in, - true); - } -} - - -void QuickCheckDetails::Clear() { - for (int i = 0; i < characters_; i++) { - positions_[i].mask = 0; - positions_[i].value = 0; - positions_[i].determines_perfectly = false; - } - characters_ = 0; -} - - -void QuickCheckDetails::Advance(int by, bool one_byte) { - if (by >= characters_ || by < 0) { - DCHECK_IMPLIES(by < 0, characters_ == 0); - Clear(); - return; - } - DCHECK_LE(characters_ - by, 4); - DCHECK_LE(characters_, 4); - for (int i = 0; i < characters_ - by; i++) { - positions_[i] = positions_[by + i]; - } - for (int i = characters_ - by; i < characters_; i++) { - positions_[i].mask = 0; - positions_[i].value = 0; - positions_[i].determines_perfectly = false; - } - characters_ -= by; - // We could change mask_ and value_ here but we would never advance unless - // they had already been used in a check and they won't be used again because - // it would gain us nothing. So there's no point. -} - - -void QuickCheckDetails::Merge(QuickCheckDetails* other, int from_index) { - DCHECK(characters_ == other->characters_); - if (other->cannot_match_) { - return; - } - if (cannot_match_) { - *this = *other; - return; - } - for (int i = from_index; i < characters_; i++) { - QuickCheckDetails::Position* pos = positions(i); - QuickCheckDetails::Position* other_pos = other->positions(i); - if (pos->mask != other_pos->mask || - pos->value != other_pos->value || - !other_pos->determines_perfectly) { - // Our mask-compare operation will be approximate unless we have the - // exact same operation on both sides of the alternation. - pos->determines_perfectly = false; - } - pos->mask &= other_pos->mask; - pos->value &= pos->mask; - other_pos->value &= pos->mask; - uc16 differing_bits = (pos->value ^ other_pos->value); - pos->mask &= ~differing_bits; - pos->value &= pos->mask; - } -} - - -class VisitMarker { - public: - explicit VisitMarker(NodeInfo* info) : info_(info) { - DCHECK(!info->visited); - info->visited = true; - } - ~VisitMarker() { - info_->visited = false; - } - private: - NodeInfo* info_; -}; - -RegExpNode* SeqRegExpNode::FilterOneByte(int depth) { - if (info()->replacement_calculated) return replacement(); - if (depth < 0) return this; - DCHECK(!info()->visited); - VisitMarker marker(info()); - return FilterSuccessor(depth - 1); -} - -RegExpNode* SeqRegExpNode::FilterSuccessor(int depth) { - RegExpNode* next = on_success_->FilterOneByte(depth - 1); - if (next == nullptr) return set_replacement(nullptr); - on_success_ = next; - return set_replacement(this); -} - - -static bool RangesContainLatin1Equivalents(ZoneList* ranges) { - for (int i = 0; i < ranges->length(); i++) { - // TODO(dcarney): this could be a lot more efficient. - if (RangeContainsLatin1Equivalents(ranges->at(i))) return true; - } - return false; -} - -RegExpNode* TextNode::FilterOneByte(int depth) { - if (info()->replacement_calculated) return replacement(); - if (depth < 0) return this; - DCHECK(!info()->visited); - VisitMarker marker(info()); - int element_count = elements()->length(); - for (int i = 0; i < element_count; i++) { - TextElement elm = elements()->at(i); - if (elm.text_type() == TextElement::ATOM) { - Vector quarks = elm.atom()->data(); - for (int j = 0; j < quarks.length(); j++) { - uint16_t c = quarks[j]; - if (elm.atom()->ignore_case()) { - c = unibrow::Latin1::TryConvertToLatin1(c); - } - if (c > unibrow::Latin1::kMaxChar) return set_replacement(nullptr); - // Replace quark in case we converted to Latin-1. - uint16_t* writable_quarks = const_cast(quarks.begin()); - writable_quarks[j] = c; - } - } else { - DCHECK(elm.text_type() == TextElement::CHAR_CLASS); - RegExpCharacterClass* cc = elm.char_class(); - ZoneList* ranges = cc->ranges(zone()); - CharacterRange::Canonicalize(ranges); - // Now they are in order so we only need to look at the first. - int range_count = ranges->length(); - if (cc->is_negated()) { - if (range_count != 0 && - ranges->at(0).from() == 0 && - ranges->at(0).to() >= String::kMaxOneByteCharCode) { - // This will be handled in a later filter. - if (IgnoreCase(cc->flags()) && RangesContainLatin1Equivalents(ranges)) - continue; - return set_replacement(nullptr); - } - } else { - if (range_count == 0 || - ranges->at(0).from() > String::kMaxOneByteCharCode) { - // This will be handled in a later filter. - if (IgnoreCase(cc->flags()) && RangesContainLatin1Equivalents(ranges)) - continue; - return set_replacement(nullptr); - } - } - } - } - return FilterSuccessor(depth - 1); -} - -RegExpNode* LoopChoiceNode::FilterOneByte(int depth) { - if (info()->replacement_calculated) return replacement(); - if (depth < 0) return this; - if (info()->visited) return this; - { - VisitMarker marker(info()); - - RegExpNode* continue_replacement = continue_node_->FilterOneByte(depth - 1); - // If we can't continue after the loop then there is no sense in doing the - // loop. - if (continue_replacement == nullptr) return set_replacement(nullptr); - } - - return ChoiceNode::FilterOneByte(depth - 1); -} - -RegExpNode* ChoiceNode::FilterOneByte(int depth) { - if (info()->replacement_calculated) return replacement(); - if (depth < 0) return this; - if (info()->visited) return this; - VisitMarker marker(info()); - int choice_count = alternatives_->length(); - - for (int i = 0; i < choice_count; i++) { - GuardedAlternative alternative = alternatives_->at(i); - if (alternative.guards() != nullptr && - alternative.guards()->length() != 0) { - set_replacement(this); - return this; - } - } - - int surviving = 0; - RegExpNode* survivor = nullptr; - for (int i = 0; i < choice_count; i++) { - GuardedAlternative alternative = alternatives_->at(i); - RegExpNode* replacement = alternative.node()->FilterOneByte(depth - 1); - DCHECK(replacement != this); // No missing EMPTY_MATCH_CHECK. - if (replacement != nullptr) { - alternatives_->at(i).set_node(replacement); - surviving++; - survivor = replacement; - } - } - if (surviving < 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. - ZoneList* new_alternatives = - new(zone()) ZoneList(surviving, zone()); - for (int i = 0; i < choice_count; i++) { - RegExpNode* replacement = - alternatives_->at(i).node()->FilterOneByte(depth - 1); - if (replacement != nullptr) { - alternatives_->at(i).set_node(replacement); - new_alternatives->Add(alternatives_->at(i), zone()); - } - } - alternatives_ = new_alternatives; - return this; -} - -RegExpNode* NegativeLookaroundChoiceNode::FilterOneByte(int 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_->at(1).node(); - RegExpNode* replacement = node->FilterOneByte(depth - 1); - if (replacement == nullptr) return set_replacement(nullptr); - alternatives_->at(1).set_node(replacement); - - RegExpNode* neg_node = alternatives_->at(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_->at(0).set_node(neg_replacement); - return set_replacement(this); -} - - -void LoopChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, - int characters_filled_in, - bool not_at_start) { - if (body_can_be_zero_length_ || info()->visited) return; - VisitMarker marker(info()); - return ChoiceNode::GetQuickCheckDetails(details, - compiler, - characters_filled_in, - not_at_start); -} - - -void LoopChoiceNode::FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) { - if (body_can_be_zero_length_ || budget <= 0) { - bm->SetRest(offset); - SaveBMInfo(bm, not_at_start, offset); - return; - } - ChoiceNode::FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start); - SaveBMInfo(bm, not_at_start, offset); -} - - -void ChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, - int characters_filled_in, - bool not_at_start) { - not_at_start = (not_at_start || not_at_start_); - int choice_count = alternatives_->length(); - DCHECK_LT(0, choice_count); - alternatives_->at(0).node()->GetQuickCheckDetails(details, - compiler, - characters_filled_in, - not_at_start); - for (int i = 1; i < choice_count; i++) { - QuickCheckDetails new_details(details->characters()); - RegExpNode* node = alternatives_->at(i).node(); - node->GetQuickCheckDetails(&new_details, compiler, - characters_filled_in, - not_at_start); - // Here we merge the quick match details of the two branches. - details->Merge(&new_details, characters_filled_in); - } -} - - -// Check for [0-9A-Z_a-z]. -static void EmitWordCheck(RegExpMacroAssembler* assembler, - Label* word, - Label* non_word, - bool fall_through_on_word) { - if (assembler->CheckSpecialCharacterClass( - fall_through_on_word ? 'w' : 'W', - fall_through_on_word ? non_word : word)) { - // Optimized implementation available. - return; - } - assembler->CheckCharacterGT('z', non_word); - assembler->CheckCharacterLT('0', non_word); - assembler->CheckCharacterGT('a' - 1, word); - assembler->CheckCharacterLT('9' + 1, word); - assembler->CheckCharacterLT('A', non_word); - assembler->CheckCharacterLT('Z' + 1, word); - if (fall_through_on_word) { - assembler->CheckNotCharacter('_', non_word); - } else { - assembler->CheckCharacter('_', word); - } -} - - -// Emit the code to check for a ^ in multiline mode (1-character lookbehind -// that matches newline or the start of input). -static void EmitHat(RegExpCompiler* compiler, - RegExpNode* on_success, - Trace* trace) { - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - // We will be loading the previous character into the current character - // register. - Trace new_trace(*trace); - new_trace.InvalidateCurrentCharacter(); - - Label ok; - if (new_trace.cp_offset() == 0) { - // The start of input counts as a newline in this context, so skip to - // ok if we are at the start. - assembler->CheckAtStart(&ok); - } - // We already checked that we are not at the start of input so it must be - // OK to load the previous character. - assembler->LoadCurrentCharacter(new_trace.cp_offset() -1, - new_trace.backtrack(), - false); - if (!assembler->CheckSpecialCharacterClass('n', - new_trace.backtrack())) { - // Newline means \n, \r, 0x2028 or 0x2029. - if (!compiler->one_byte()) { - assembler->CheckCharacterAfterAnd(0x2028, 0xFFFE, &ok); - } - assembler->CheckCharacter('\n', &ok); - assembler->CheckNotCharacter('\r', new_trace.backtrack()); - } - assembler->Bind(&ok); - on_success->Emit(compiler, &new_trace); -} - - -// Emit the code to handle \b and \B (word-boundary or non-word-boundary). -void AssertionNode::EmitBoundaryCheck(RegExpCompiler* compiler, Trace* trace) { - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - Isolate* isolate = assembler->isolate(); - Trace::TriBool next_is_word_character = Trace::UNKNOWN; - bool not_at_start = (trace->at_start() == Trace::FALSE_VALUE); - BoyerMooreLookahead* lookahead = bm_info(not_at_start); - if (lookahead == nullptr) { - int eats_at_least = - Min(kMaxLookaheadForBoyerMoore, EatsAtLeast(kMaxLookaheadForBoyerMoore, - kRecursionBudget, - not_at_start)); - if (eats_at_least >= 1) { - BoyerMooreLookahead* bm = - new(zone()) BoyerMooreLookahead(eats_at_least, compiler, zone()); - FillInBMInfo(isolate, 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) { - Label before_non_word; - Label before_word; - if (trace->characters_preloaded() != 1) { - assembler->LoadCurrentCharacter(trace->cp_offset(), &before_non_word); - } - // Fall through on non-word. - EmitWordCheck(assembler, &before_word, &before_non_word, false); - // Next character is not a word character. - assembler->Bind(&before_non_word); - Label ok; - BacktrackIfPrevious(compiler, trace, at_boundary ? kIsNonWord : kIsWord); - assembler->GoTo(&ok); - - assembler->Bind(&before_word); - BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord); - assembler->Bind(&ok); - } else if (next_is_word_character == Trace::TRUE_VALUE) { - BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord); - } else { - DCHECK(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(); - - Label fall_through, dummy; - - Label* non_word = backtrack_if_previous == kIsNonWord ? - new_trace.backtrack() : - &fall_through; - Label* 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->Bind(&fall_through); - on_success()->Emit(compiler, &new_trace); -} - - -void AssertionNode::GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, - int 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: { - Label ok; - assembler->CheckPosition(trace->cp_offset(), &ok); - assembler->GoTo(trace->backtrack()); - assembler->Bind(&ok); - break; - } - case AT_START: { - if (trace->at_start() == Trace::FALSE_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, int offset) { - if (quick_check == nullptr) return false; - if (offset >= quick_check->characters()) return false; - return quick_check->positions(offset)->determines_perfectly; -} - - -static void UpdateBoundsCheck(int index, int* checked_up_to) { - if (index > *checked_up_to) { - *checked_up_to = index; - } -} - - -// We call this repeatedly to generate code for each pass over the text node. -// The passes are in increasing order of difficulty because we hope one -// of the first passes will fail in which case we are saved the work of the -// later passes. for example for the case independent regexp /%[asdfghjkl]a/ -// we will check the '%' in the first pass, the case independent 'a' in the -// second pass and the character class in the last pass. -// -// The passes are done from right to left, so for example to test for /bar/ -// we will first test for an 'r' with offset 2, then an 'a' with offset 1 -// and then a 'b' with offset 0. This means we can avoid the end-of-input -// bounds check most of the time. In the example we only need to check for -// end-of-input when loading the putative 'r'. -// -// A slight complication involves the fact that the first character may already -// be fetched into a register by the previous node. In this case we want to -// do the test for that character first. We do this in separate passes. The -// 'preloaded' argument indicates that we are doing such a 'pass'. If such a -// pass has been performed then subsequent passes will have true in -// first_element_checked to indicate that that character does not need to be -// checked again. -// -// In addition to all this we are passed a Trace, which can -// contain an AlternativeGeneration object. In this AlternativeGeneration -// object we can see details of any quick check that was already passed in -// order to get to the code we are now generating. The quick check can involve -// loading characters, which means we do not need to recheck the bounds -// up to the limit the quick check already checked. In addition the quick -// check can have involved a mask and compare operation which may simplify -// or obviate the need for further checks at some character positions. -void TextNode::TextEmitPass(RegExpCompiler* compiler, - TextEmitPassType pass, - bool preloaded, - Trace* trace, - bool first_element_checked, - int* checked_up_to) { - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - Isolate* isolate = assembler->isolate(); - bool one_byte = compiler->one_byte(); - Label* backtrack = trace->backtrack(); - QuickCheckDetails* quick_check = trace->quick_check_performed(); - int element_count = elements()->length(); - int backward_offset = read_backward() ? -Length() : 0; - for (int i = preloaded ? 0 : element_count - 1; i >= 0; i--) { - TextElement elm = elements()->at(i); - int cp_offset = trace->cp_offset() + elm.cp_offset() + backward_offset; - if (elm.text_type() == TextElement::ATOM) { - if (SkipPass(pass, elm.atom()->ignore_case())) continue; - Vector quarks = elm.atom()->data(); - for (int j = preloaded ? 0 : quarks.length() - 1; j >= 0; j--) { - if (first_element_checked && i == 0 && j == 0) continue; - if (DeterminedAlready(quick_check, elm.cp_offset() + j)) continue; - EmitCharacterFunction* emit_function = nullptr; - uc16 quark = quarks[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 = unibrow::Latin1::TryConvertToLatin1(quark); - } - switch (pass) { - case NON_LATIN1_MATCH: - DCHECK(one_byte); - if (quark > String::kMaxOneByteCharCode) { - 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) { - bool bounds_check = *checked_up_to < cp_offset + j || read_backward(); - bool bound_checked = - emit_function(isolate, compiler, quark, backtrack, cp_offset + j, - bounds_check, preloaded); - if (bound_checked) UpdateBoundsCheck(cp_offset + j, checked_up_to); - } - } - } else { - DCHECK_EQ(TextElement::CHAR_CLASS, elm.text_type()); - 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, zone()); - UpdateBoundsCheck(cp_offset, checked_up_to); - } - } - } -} - - -int TextNode::Length() { - TextElement elm = elements()->last(); - DCHECK_LE(0, elm.cp_offset()); - return elm.cp_offset() + elm.length(); -} - -bool TextNode::SkipPass(TextEmitPassType pass, bool ignore_case) { - if (ignore_case) { - return pass == SIMPLE_CHARACTER_MATCH; - } else { - return pass == NON_LETTER_CHARACTER_MATCH || pass == CASE_CHARACTER_MATCH; - } -} - -TextNode* TextNode::CreateForCharacterRanges(Zone* zone, - ZoneList* ranges, - bool read_backward, - RegExpNode* on_success, - JSRegExp::Flags flags) { - DCHECK_NOT_NULL(ranges); - ZoneList* elms = new (zone) ZoneList(1, zone); - elms->Add(TextElement::CharClass( - new (zone) RegExpCharacterClass(zone, ranges, flags)), - zone); - return new (zone) TextNode(elms, read_backward, on_success); -} - -TextNode* TextNode::CreateForSurrogatePair(Zone* zone, CharacterRange lead, - CharacterRange trail, - bool read_backward, - RegExpNode* on_success, - JSRegExp::Flags flags) { - ZoneList* lead_ranges = CharacterRange::List(zone, lead); - ZoneList* trail_ranges = CharacterRange::List(zone, trail); - ZoneList* elms = new (zone) ZoneList(2, zone); - elms->Add(TextElement::CharClass( - new (zone) RegExpCharacterClass(zone, lead_ranges, flags)), - zone); - elms->Add(TextElement::CharClass( - new (zone) RegExpCharacterClass(zone, trail_ranges, flags)), - zone); - return new (zone) 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; - DCHECK(limit_result == CONTINUE); - - if (trace->cp_offset() + Length() > RegExpMacroAssembler::kMaxCPOffset) { - compiler->SetRegExpTooBig(); - return; - } - - if (compiler->one_byte()) { - int dummy = 0; - TextEmitPass(compiler, NON_LATIN1_MATCH, false, trace, false, &dummy); - } - - bool first_elt_done = false; - int bound_checked_to = trace->cp_offset() - 1; - bound_checked_to += trace->bound_checked_up_to(); - - // If a character is preloaded into the current character register then - // check that now. - if (trace->characters_preloaded() == 1) { - for (int pass = kFirstRealPass; pass <= kLastPass; pass++) { - TextEmitPass(compiler, static_cast(pass), true, trace, - false, &bound_checked_to); - } - first_elt_done = true; - } - - for (int pass = kFirstRealPass; pass <= kLastPass; pass++) { - TextEmitPass(compiler, static_cast(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(int 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_ = Max(0, bound_checked_up_to_ - by); -} - - -void TextNode::MakeCaseIndependent(Isolate* isolate, bool is_one_byte) { - int element_count = elements()->length(); - for (int i = 0; i < element_count; i++) { - TextElement elm = elements()->at(i); - if (elm.text_type() == TextElement::CHAR_CLASS) { - RegExpCharacterClass* cc = elm.char_class(); -#ifdef V8_INTL_SUPPORT - bool case_equivalents_already_added = - NeedsUnicodeCaseEquivalents(cc->flags()); -#else - bool case_equivalents_already_added = false; -#endif - if (IgnoreCase(cc->flags()) && !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(zone())) continue; - ZoneList* ranges = cc->ranges(zone()); - CharacterRange::AddCaseEquivalents(isolate, zone(), ranges, - is_one_byte); - } - } - } -} - - -int TextNode::GreedyLoopTextLength() { return Length(); } - - -RegExpNode* TextNode::GetSuccessorOfOmnivorousTextNode( - RegExpCompiler* compiler) { - if (read_backward()) return nullptr; - if (elements()->length() != 1) return nullptr; - TextElement elm = elements()->at(0); - if (elm.text_type() != TextElement::CHAR_CLASS) return nullptr; - RegExpCharacterClass* node = elm.char_class(); - ZoneList* ranges = node->ranges(zone()); - 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 = String::kMaxOneByteCharCode; - } else { - max_char = String::kMaxUtf16CodeUnit; - } - 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. -int ChoiceNode::GreedyLoopTextLengthForAlternative( - GuardedAlternative* alternative) { - int length = 0; - RegExpNode* node = alternative->node(); - // Later we will generate code for all these text nodes using recursion - // so we have to limit the max number. - int recursion_depth = 0; - while (node != this) { - if (recursion_depth++ > RegExpCompiler::kMaxRecursion) { - return kNodeIsTooComplexForGreedyLoops; - } - int node_length = node->GreedyLoopTextLength(); - if (node_length == kNodeIsTooComplexForGreedyLoops) { - return kNodeIsTooComplexForGreedyLoops; - } - length += node_length; - SeqRegExpNode* seq_node = static_cast(node); - node = seq_node->on_success(); - } - return read_backward() ? -length : length; -} - - -void LoopChoiceNode::AddLoopAlternative(GuardedAlternative alt) { - DCHECK_NULL(loop_node_); - AddAlternative(alt); - loop_node_ = alt.node(); -} - - -void LoopChoiceNode::AddContinueAlternative(GuardedAlternative alt) { - DCHECK_NULL(continue_node_); - 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. - int text_length = - GreedyLoopTextLengthForAlternative(&(alternatives_->at(0))); - DCHECK_NE(kNodeIsTooComplexForGreedyLoops, text_length); - // Update the counter-based backtracking info on the stack. This is an - // optimization for greedy loops (see below). - DCHECK(trace->cp_offset() == text_length); - macro_assembler->AdvanceCurrentPosition(text_length); - macro_assembler->GoTo(trace->loop_label()); - return; - } - DCHECK_NULL(trace->stop_node()); - if (!trace->is_trivial()) { - trace->Flush(compiler, this); - return; - } - ChoiceNode::Emit(compiler, trace); -} - - -int ChoiceNode::CalculatePreloadCharacters(RegExpCompiler* compiler, - int eats_at_least) { - int preload_characters = Min(4, eats_at_least); - DCHECK_LE(preload_characters, 4); - if (compiler->macro_assembler()->CanReadUnaligned()) { - bool one_byte = compiler->one_byte(); - if (one_byte) { - // We can't preload 3 characters because there is no machine instruction - // to do that. We can't just load 4 because we could be reading - // beyond the end of the string, which could cause a memory fault. - if (preload_characters == 3) preload_characters = 2; - } else { - if (preload_characters > 2) preload_characters = 2; - } - } else { - if (preload_characters > 1) preload_characters = 1; - } - return preload_characters; -} - - -// This class is used when generating the alternatives in a choice node. It -// records the way the alternative is being code generated. -class AlternativeGeneration: public Malloced { - public: - AlternativeGeneration() - : possible_success(), - expects_preload(false), - after(), - quick_check_details() { } - Label possible_success; - bool expects_preload; - Label after; - QuickCheckDetails quick_check_details; -}; - - -// Creates a list of AlternativeGenerations. If the list has a reasonable -// size then it is on the stack, otherwise the excess is on the heap. -class AlternativeGenerationList { - public: - AlternativeGenerationList(int count, Zone* zone) - : alt_gens_(count, zone) { - for (int i = 0; i < count && i < kAFew; i++) { - alt_gens_.Add(a_few_alt_gens_ + i, zone); - } - for (int i = kAFew; i < count; i++) { - alt_gens_.Add(new AlternativeGeneration(), zone); - } - } - ~AlternativeGenerationList() { - for (int i = kAFew; i < alt_gens_.length(); i++) { - delete alt_gens_[i]; - alt_gens_[i] = nullptr; - } - } - - AlternativeGeneration* at(int i) { - return alt_gens_[i]; - } - - private: - static const int kAFew = 10; - ZoneList alt_gens_; - AlternativeGeneration a_few_alt_gens_[kAFew]; -}; - -void BoyerMoorePositionInfo::Set(int character) { - SetInterval(Interval(character, character)); -} - - -void BoyerMoorePositionInfo::SetInterval(const Interval& interval) { - s_ = AddRange(s_, kSpaceRanges, kSpaceRangeCount, interval); - w_ = AddRange(w_, kWordRanges, kWordRangeCount, interval); - d_ = AddRange(d_, kDigitRanges, kDigitRangeCount, interval); - surrogate_ = - AddRange(surrogate_, kSurrogateRanges, kSurrogateRangeCount, interval); - if (interval.to() - interval.from() >= kMapSize - 1) { - if (map_count_ != kMapSize) { - map_count_ = kMapSize; - for (int i = 0; i < kMapSize; i++) map_->at(i) = true; - } - return; - } - for (int i = interval.from(); i <= interval.to(); i++) { - int mod_character = (i & kMask); - if (!map_->at(mod_character)) { - map_count_++; - map_->at(mod_character) = true; - } - if (map_count_ == kMapSize) return; - } -} - - -void BoyerMoorePositionInfo::SetAll() { - s_ = w_ = d_ = kLatticeUnknown; - if (map_count_ != kMapSize) { - map_count_ = kMapSize; - for (int i = 0; i < kMapSize; i++) map_->at(i) = true; - } -} - - -BoyerMooreLookahead::BoyerMooreLookahead( - int length, RegExpCompiler* compiler, Zone* zone) - : length_(length), - compiler_(compiler) { - if (compiler->one_byte()) { - max_char_ = String::kMaxOneByteCharCode; - } else { - max_char_ = String::kMaxUtf16CodeUnit; - } - bitmaps_ = new(zone) ZoneList(length, zone); - for (int i = 0; i < length; i++) { - bitmaps_->Add(new(zone) BoyerMoorePositionInfo(zone), zone); - } -} - - -// Find the longest range of lookahead that has the fewest number of different -// characters that can occur at a given position. Since we are optimizing two -// different parameters at once this is a tradeoff. -bool BoyerMooreLookahead::FindWorthwhileInterval(int* from, int* to) { - int biggest_points = 0; - // If more than 32 characters out of 128 can occur it is unlikely that we can - // be lucky enough to step forwards much of the time. - const int kMaxMax = 32; - for (int max_number_of_chars = 4; - max_number_of_chars < kMaxMax; - max_number_of_chars *= 2) { - biggest_points = - FindBestInterval(max_number_of_chars, biggest_points, from, to); - } - if (biggest_points == 0) return false; - return true; -} - - -// Find the highest-points range between 0 and length_ where the character -// information is not too vague. 'Too vague' means that there are more than -// max_number_of_chars that can occur at this position. Calculates the number -// of points as the product of width-of-the-range and -// probability-of-finding-one-of-the-characters, where the probability is -// calculated using the frequency distribution of the sample subject string. -int BoyerMooreLookahead::FindBestInterval( - int max_number_of_chars, int old_biggest_points, int* from, int* to) { - int biggest_points = old_biggest_points; - static const int kSize = RegExpMacroAssembler::kTableSize; - for (int i = 0; i < length_; ) { - while (i < length_ && Count(i) > max_number_of_chars) i++; - if (i == length_) break; - int remembered_from = i; - bool union_map[kSize]; - for (int j = 0; j < kSize; j++) union_map[j] = false; - while (i < length_ && Count(i) <= max_number_of_chars) { - BoyerMoorePositionInfo* map = bitmaps_->at(i); - for (int j = 0; j < kSize; j++) union_map[j] |= map->at(j); - i++; - } - int frequency = 0; - for (int j = 0; j < kSize; j++) { - if (union_map[j]) { - // Add 1 to the frequency to give a small per-character boost for - // the cases where our sampling is not good enough and many - // characters have a frequency of zero. This means the frequency - // can theoretically be up to 2*kSize though we treat it mostly as - // a fraction of kSize. - frequency += compiler_->frequency_collator()->Frequency(j) + 1; - } - } - // We use the probability of skipping times the distance we are skipping to - // judge the effectiveness of this. Actually we have a cut-off: By - // dividing by 2 we switch off the skipping if the probability of skipping - // is less than 50%. This is because the multibyte mask-and-compare - // skipping in quickcheck is more likely to do well on this case. - bool in_quickcheck_range = - ((i - remembered_from < 4) || - (compiler_->one_byte() ? remembered_from <= 4 : remembered_from <= 2)); - // Called 'probability' but it is only a rough estimate and can actually - // be outside the 0-kSize range. - int probability = (in_quickcheck_range ? kSize / 2 : kSize) - frequency; - int points = (i - remembered_from) * probability; - if (points > biggest_points) { - *from = remembered_from; - *to = i - 1; - biggest_points = points; - } - } - return biggest_points; -} - - -// Take all the characters that will not prevent a successful match if they -// occur in the subject string in the range between min_lookahead and -// max_lookahead (inclusive) measured from the current position. If the -// character at max_lookahead offset is not one of these characters, then we -// can safely skip forwards by the number of characters in the range. -int BoyerMooreLookahead::GetSkipTable(int min_lookahead, - int max_lookahead, - Handle boolean_skip_table) { - const int kSize = RegExpMacroAssembler::kTableSize; - - const int kSkipArrayEntry = 0; - const int kDontSkipArrayEntry = 1; - - for (int i = 0; i < kSize; i++) { - boolean_skip_table->set(i, kSkipArrayEntry); - } - int skip = max_lookahead + 1 - min_lookahead; - - for (int i = max_lookahead; i >= min_lookahead; i--) { - BoyerMoorePositionInfo* map = bitmaps_->at(i); - for (int j = 0; j < kSize; j++) { - if (map->at(j)) { - boolean_skip_table->set(j, kDontSkipArrayEntry); - } - } - } - - return skip; -} - - -// See comment above on the implementation of GetSkipTable. -void BoyerMooreLookahead::EmitSkipInstructions(RegExpMacroAssembler* masm) { - const int kSize = RegExpMacroAssembler::kTableSize; - - int min_lookahead = 0; - int max_lookahead = 0; - - if (!FindWorthwhileInterval(&min_lookahead, &max_lookahead)) return; - - bool found_single_character = false; - int single_character = 0; - for (int i = max_lookahead; i >= min_lookahead; i--) { - BoyerMoorePositionInfo* map = bitmaps_->at(i); - if (map->map_count() > 1 || - (found_single_character && map->map_count() != 0)) { - found_single_character = false; - break; - } - for (int j = 0; j < kSize; j++) { - if (map->at(j)) { - found_single_character = true; - single_character = j; - break; - } - } - } - - int lookahead_width = max_lookahead + 1 - min_lookahead; - - if (found_single_character && lookahead_width == 1 && max_lookahead < 3) { - // The mask-compare can probably handle this better. - return; - } - - if (found_single_character) { - Label cont, again; - masm->Bind(&again); - masm->LoadCurrentCharacter(max_lookahead, &cont, true); - if (max_char_ > kSize) { - masm->CheckCharacterAfterAnd(single_character, - RegExpMacroAssembler::kTableMask, - &cont); - } else { - masm->CheckCharacter(single_character, &cont); - } - masm->AdvanceCurrentPosition(lookahead_width); - masm->GoTo(&again); - masm->Bind(&cont); - return; - } - - Factory* factory = masm->isolate()->factory(); - Handle boolean_skip_table = - factory->NewByteArray(kSize, AllocationType::kOld); - int skip_distance = GetSkipTable( - min_lookahead, max_lookahead, boolean_skip_table); - DCHECK_NE(0, skip_distance); - - Label cont, again; - masm->Bind(&again); - masm->LoadCurrentCharacter(max_lookahead, &cont, true); - masm->CheckBitInTable(boolean_skip_table, &cont); - masm->AdvanceCurrentPosition(skip_distance); - masm->GoTo(&again); - masm->Bind(&cont); -} - - -/* Code generation for choice nodes. - * - * We generate quick checks that do a mask and compare to eliminate a - * choice. If the quick check succeeds then it jumps to the continuation to - * do slow checks and check subsequent nodes. If it fails (the common case) - * it falls through to the next choice. - * - * Here is the desired flow graph. Nodes directly below each other imply - * fallthrough. Alternatives 1 and 2 have quick checks. Alternative - * 3 doesn't have a quick check so we have to call the slow check. - * Nodes are marked Qn for quick checks and Sn for slow checks. The entire - * regexp continuation is generated directly after the Sn node, up to the - * next GoTo if we decide to reuse some already generated code. Some - * nodes expect preload_characters to be preloaded into the current - * character register. R nodes do this preloading. Vertices are marked - * F for failures and S for success (possible success in the case of quick - * nodes). L, V, < and > are used as arrow heads. - * - * ----------> R - * | - * V - * Q1 -----> S1 - * | S / - * F| / - * | F/ - * | / - * | R - * | / - * V L - * Q2 -----> S2 - * | S / - * F| / - * | F/ - * | / - * | R - * | / - * V L - * S3 - * | - * F| - * | - * R - * | - * backtrack V - * <----------Q4 - * \ F | - * \ |S - * \ F V - * \-----S4 - * - * For greedy loops we push the current position, then generate the code that - * eats the input specially in EmitGreedyLoop. The other choice (the - * continuation) is generated by the normal code in EmitChoices, and steps back - * in the input to the starting position when it fails to match. The loop code - * looks like this (U is the unwind code that steps back in the greedy loop). - * - * _____ - * / \ - * V | - * ----------> S1 | - * /| | - * / |S | - * F/ \_____/ - * / - * |<----- - * | \ - * V |S - * Q2 ---> U----->backtrack - * | F / - * S| / - * V F / - * S2--/ - */ - -GreedyLoopState::GreedyLoopState(bool not_at_start) { - counter_backtrack_trace_.set_backtrack(&label_); - if (not_at_start) counter_backtrack_trace_.set_at_start(Trace::FALSE_VALUE); -} - - -void ChoiceNode::AssertGuardsMentionRegisters(Trace* trace) { -#ifdef DEBUG - int choice_count = alternatives_->length(); - for (int i = 0; i < choice_count - 1; i++) { - GuardedAlternative alternative = alternatives_->at(i); - ZoneList* guards = alternative.guards(); - int guard_count = (guards == nullptr) ? 0 : guards->length(); - for (int j = 0; j < guard_count; j++) { - DCHECK(!trace->mentions_reg(guards->at(j)->reg())); - } - } -#endif -} - - -void ChoiceNode::SetUpPreLoad(RegExpCompiler* compiler, - Trace* current_trace, - PreloadState* state) { - if (state->eats_at_least_ == PreloadState::kEatsAtLeastNotYetInitialized) { - // Save some time by looking at most one machine word ahead. - state->eats_at_least_ = - EatsAtLeast(compiler->one_byte() ? 4 : 2, kRecursionBudget, - current_trace->at_start() == Trace::FALSE_VALUE); - } - state->preload_characters_ = - CalculatePreloadCharacters(compiler, state->eats_at_least_); - - state->preload_is_current_ = - (current_trace->characters_preloaded() == state->preload_characters_); - state->preload_has_checked_bounds_ = state->preload_is_current_; -} - - -void ChoiceNode::Emit(RegExpCompiler* compiler, Trace* trace) { - int choice_count = alternatives_->length(); - - if (choice_count == 1 && alternatives_->at(0).guards() == nullptr) { - alternatives_->at(0).node()->Emit(compiler, trace); - return; - } - - AssertGuardsMentionRegisters(trace); - - LimitResult limit_result = LimitVersions(compiler, trace); - if (limit_result == DONE) return; - DCHECK(limit_result == CONTINUE); - - // For loop nodes we already flushed (see LoopChoiceNode::Emit), but for - // other choice nodes we only flush if we are out of code size budget. - if (trace->flush_budget() == 0 && trace->actions() != nullptr) { - trace->Flush(compiler, this); - return; - } - - RecursionCheck rc(compiler); - - PreloadState preload; - preload.init(); - GreedyLoopState greedy_loop_state(not_at_start()); - - int text_length = GreedyLoopTextLengthForAlternative(&alternatives_->at(0)); - AlternativeGenerationList alt_gens(choice_count, zone()); - - if (choice_count > 1 && text_length != kNodeIsTooComplexForGreedyLoops) { - trace = EmitGreedyLoop(compiler, - trace, - &alt_gens, - &preload, - &greedy_loop_state, - text_length); - } else { - // TODO(erikcorry): Delete this. We don't need this label, but it makes us - // match the traces produced pre-cleanup. - Label second_choice; - compiler->macro_assembler()->Bind(&second_choice); - - preload.eats_at_least_ = EmitOptimizedUnanchoredSearch(compiler, trace); - - EmitChoices(compiler, - &alt_gens, - 0, - trace, - &preload); - } - - // At this point we need to generate slow checks for the alternatives where - // the quick check was inlined. We can recognize these because the associated - // label was bound. - int new_flush_budget = trace->flush_budget() / choice_count; - for (int i = 0; i < choice_count; i++) { - AlternativeGeneration* alt_gen = alt_gens.at(i); - Trace new_trace(*trace); - // If there are actions to be flushed we have to limit how many times - // they are flushed. Take the budget of the parent trace and distribute - // it fairly amongst the children. - if (new_trace.actions() != nullptr) { - new_trace.set_flush_budget(new_flush_budget); - } - bool next_expects_preload = - i == choice_count - 1 ? false : alt_gens.at(i + 1)->expects_preload; - EmitOutOfLineContinuation(compiler, - &new_trace, - alternatives_->at(i), - alt_gen, - preload.preload_characters_, - next_expects_preload); - } -} - - -Trace* ChoiceNode::EmitGreedyLoop(RegExpCompiler* compiler, - Trace* trace, - AlternativeGenerationList* alt_gens, - PreloadState* preload, - GreedyLoopState* greedy_loop_state, - int text_length) { - RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); - // Here we have special handling for greedy loops containing only text nodes - // and other simple nodes. These are handled by pushing the current - // position on the stack and then incrementing the current position each - // time around the switch. On backtrack we decrement the current position - // and check it against the pushed value. This avoids pushing backtrack - // information for each iteration of the loop, which could take up a lot of - // space. - DCHECK(trace->stop_node() == nullptr); - macro_assembler->PushCurrentPosition(); - Label greedy_match_failed; - Trace greedy_match_trace; - if (not_at_start()) greedy_match_trace.set_at_start(Trace::FALSE_VALUE); - greedy_match_trace.set_backtrack(&greedy_match_failed); - Label loop_label; - macro_assembler->Bind(&loop_label); - greedy_match_trace.set_stop_node(this); - greedy_match_trace.set_loop_label(&loop_label); - alternatives_->at(0).node()->Emit(compiler, &greedy_match_trace); - macro_assembler->Bind(&greedy_match_failed); - - Label second_choice; // For use in greedy matches. - macro_assembler->Bind(&second_choice); - - Trace* new_trace = greedy_loop_state->counter_backtrack_trace(); - - EmitChoices(compiler, - alt_gens, - 1, - new_trace, - preload); - - macro_assembler->Bind(greedy_loop_state->label()); - // If we have unwound to the bottom then backtrack. - macro_assembler->CheckGreedyLoop(trace->backtrack()); - // Otherwise try the second priority at an earlier position. - macro_assembler->AdvanceCurrentPosition(-text_length); - macro_assembler->GoTo(&second_choice); - return new_trace; -} - -int ChoiceNode::EmitOptimizedUnanchoredSearch(RegExpCompiler* compiler, - Trace* trace) { - int eats_at_least = PreloadState::kEatsAtLeastNotYetInitialized; - if (alternatives_->length() != 2) return eats_at_least; - - GuardedAlternative alt1 = alternatives_->at(1); - if (alt1.guards() != nullptr && alt1.guards()->length() != 0) { - return eats_at_least; - } - RegExpNode* eats_anything_node = alt1.node(); - if (eats_anything_node->GetSuccessorOfOmnivorousTextNode(compiler) != this) { - return eats_at_least; - } - - // Really we should be creating a new trace when we execute this function, - // but there is no need, because the code it generates cannot backtrack, and - // we always arrive here with a trivial trace (since it's the entry to a - // loop. That also implies that there are no preloaded characters, which is - // good, because it means we won't be violating any assumptions by - // overwriting those characters with new load instructions. - DCHECK(trace->is_trivial()); - - RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); - Isolate* isolate = macro_assembler->isolate(); - // At this point we know that we are at a non-greedy loop that will eat - // any character one at a time. Any non-anchored regexp has such a - // loop prepended to it in order to find where it starts. We look for - // a pattern of the form ...abc... where we can look 6 characters ahead - // and step forwards 3 if the character is not one of abc. Abc need - // not be atoms, they can be any reasonably limited character class or - // small alternation. - BoyerMooreLookahead* bm = bm_info(false); - if (bm == nullptr) { - eats_at_least = Min(kMaxLookaheadForBoyerMoore, - EatsAtLeast(kMaxLookaheadForBoyerMoore, - kRecursionBudget, - false)); - if (eats_at_least >= 1) { - bm = new(zone()) BoyerMooreLookahead(eats_at_least, - compiler, - zone()); - GuardedAlternative alt0 = alternatives_->at(0); - alt0.node()->FillInBMInfo(isolate, 0, kRecursionBudget, bm, false); - } - } - if (bm != nullptr) { - bm->EmitSkipInstructions(macro_assembler); - } - return eats_at_least; -} - - -void ChoiceNode::EmitChoices(RegExpCompiler* compiler, - AlternativeGenerationList* alt_gens, - int first_choice, - Trace* trace, - PreloadState* preload) { - RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); - SetUpPreLoad(compiler, trace, preload); - - // For now we just call all choices one after the other. The idea ultimately - // is to use the Dispatch table to try only the relevant ones. - int choice_count = alternatives_->length(); - - int new_flush_budget = trace->flush_budget() / choice_count; - - for (int i = first_choice; i < choice_count; i++) { - bool is_last = i == choice_count - 1; - bool fall_through_on_failure = !is_last; - GuardedAlternative alternative = alternatives_->at(i); - AlternativeGeneration* alt_gen = alt_gens->at(i); - alt_gen->quick_check_details.set_characters(preload->preload_characters_); - ZoneList* guards = alternative.guards(); - int guard_count = (guards == nullptr) ? 0 : guards->length(); - Trace new_trace(*trace); - new_trace.set_characters_preloaded(preload->preload_is_current_ ? - preload->preload_characters_ : - 0); - if (preload->preload_has_checked_bounds_) { - new_trace.set_bound_checked_up_to(preload->preload_characters_); - } - new_trace.quick_check_performed()->Clear(); - if (not_at_start_) new_trace.set_at_start(Trace::FALSE_VALUE); - if (!is_last) { - new_trace.set_backtrack(&alt_gen->after); - } - alt_gen->expects_preload = preload->preload_is_current_; - bool generate_full_check_inline = false; - if (compiler->optimize() && - try_to_emit_quick_check_for_alternative(i == 0) && - alternative.node()->EmitQuickCheck( - compiler, trace, &new_trace, preload->preload_has_checked_bounds_, - &alt_gen->possible_success, &alt_gen->quick_check_details, - fall_through_on_failure)) { - // Quick check was generated for this choice. - preload->preload_is_current_ = true; - preload->preload_has_checked_bounds_ = true; - // If we generated the quick check to fall through on possible success, - // we now need to generate the full check inline. - if (!fall_through_on_failure) { - macro_assembler->Bind(&alt_gen->possible_success); - new_trace.set_quick_check_performed(&alt_gen->quick_check_details); - new_trace.set_characters_preloaded(preload->preload_characters_); - new_trace.set_bound_checked_up_to(preload->preload_characters_); - generate_full_check_inline = true; - } - } else if (alt_gen->quick_check_details.cannot_match()) { - if (!fall_through_on_failure) { - macro_assembler->GoTo(trace->backtrack()); - } - continue; - } else { - // No quick check was generated. Put the full code here. - // If this is not the first choice then there could be slow checks from - // previous cases that go here when they fail. There's no reason to - // insist that they preload characters since the slow check we are about - // to generate probably can't use it. - if (i != first_choice) { - alt_gen->expects_preload = false; - new_trace.InvalidateCurrentCharacter(); - } - generate_full_check_inline = true; - } - if (generate_full_check_inline) { - if (new_trace.actions() != nullptr) { - new_trace.set_flush_budget(new_flush_budget); - } - for (int j = 0; j < guard_count; j++) { - GenerateGuard(macro_assembler, guards->at(j), &new_trace); - } - alternative.node()->Emit(compiler, &new_trace); - preload->preload_is_current_ = false; - } - macro_assembler->Bind(&alt_gen->after); - } -} - - -void ChoiceNode::EmitOutOfLineContinuation(RegExpCompiler* compiler, - Trace* trace, - GuardedAlternative alternative, - AlternativeGeneration* alt_gen, - int preload_characters, - bool next_expects_preload) { - if (!alt_gen->possible_success.is_linked()) return; - - RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); - macro_assembler->Bind(&alt_gen->possible_success); - Trace out_of_line_trace(*trace); - out_of_line_trace.set_characters_preloaded(preload_characters); - out_of_line_trace.set_quick_check_performed(&alt_gen->quick_check_details); - if (not_at_start_) out_of_line_trace.set_at_start(Trace::FALSE_VALUE); - ZoneList* guards = alternative.guards(); - int guard_count = (guards == nullptr) ? 0 : guards->length(); - if (next_expects_preload) { - Label reload_current_char; - out_of_line_trace.set_backtrack(&reload_current_char); - for (int j = 0; j < guard_count; j++) { - GenerateGuard(macro_assembler, guards->at(j), &out_of_line_trace); - } - alternative.node()->Emit(compiler, &out_of_line_trace); - macro_assembler->Bind(&reload_current_char); - // Reload the current character, since the next quick check expects that. - // We don't need to check bounds here because we only get into this - // code through a quick check which already did the checked load. - macro_assembler->LoadCurrentCharacter(trace->cp_offset(), nullptr, false, - preload_characters); - macro_assembler->GoTo(&(alt_gen->after)); - } else { - out_of_line_trace.set_backtrack(&(alt_gen->after)); - for (int j = 0; j < guard_count; j++) { - GenerateGuard(macro_assembler, guards->at(j), &out_of_line_trace); - } - alternative.node()->Emit(compiler, &out_of_line_trace); - } -} - - -void ActionNode::Emit(RegExpCompiler* compiler, Trace* trace) { - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - LimitResult limit_result = LimitVersions(compiler, trace); - if (limit_result == DONE) return; - DCHECK(limit_result == CONTINUE); - - RecursionCheck rc(compiler); - - switch (action_type_) { - case STORE_POSITION: { - Trace::DeferredCapture - new_capture(data_.u_position_register.reg, - data_.u_position_register.is_capture, - trace); - Trace new_trace = *trace; - new_trace.add_action(&new_capture); - on_success()->Emit(compiler, &new_trace); - break; - } - case INCREMENT_REGISTER: { - Trace::DeferredIncrementRegister - new_increment(data_.u_increment_register.reg); - Trace new_trace = *trace; - new_trace.add_action(&new_increment); - on_success()->Emit(compiler, &new_trace); - break; - } - case SET_REGISTER: { - Trace::DeferredSetRegister - new_set(data_.u_store_register.reg, data_.u_store_register.value); - Trace new_trace = *trace; - new_trace.add_action(&new_set); - on_success()->Emit(compiler, &new_trace); - break; - } - case CLEAR_CAPTURES: { - Trace::DeferredClearCaptures - new_capture(Interval(data_.u_clear_captures.range_from, - data_.u_clear_captures.range_to)); - Trace new_trace = *trace; - new_trace.add_action(&new_capture); - on_success()->Emit(compiler, &new_trace); - break; - } - case BEGIN_SUBMATCH: - if (!trace->is_trivial()) { - trace->Flush(compiler, this); - } else { - assembler->WriteCurrentPositionToRegister( - data_.u_submatch.current_position_register, 0); - assembler->WriteStackPointerToRegister( - data_.u_submatch.stack_pointer_register); - on_success()->Emit(compiler, trace); - } - break; - case EMPTY_MATCH_CHECK: { - int start_pos_reg = data_.u_empty_match_check.start_register; - int stored_pos = 0; - int rep_reg = data_.u_empty_match_check.repetition_register; - bool has_minimum = (rep_reg != RegExpCompiler::kNoRegister); - bool know_dist = trace->GetStoredPosition(start_pos_reg, &stored_pos); - if (know_dist && !has_minimum && stored_pos == trace->cp_offset()) { - // If we know we haven't advanced and there is no minimum we - // can just backtrack immediately. - assembler->GoTo(trace->backtrack()); - } else if (know_dist && stored_pos < trace->cp_offset()) { - // If we know we've advanced we can generate the continuation - // immediately. - on_success()->Emit(compiler, trace); - } else if (!trace->is_trivial()) { - trace->Flush(compiler, this); - } else { - Label skip_empty_check; - // If we have a minimum number of repetitions we check the current - // number first and skip the empty check if it's not enough. - if (has_minimum) { - int limit = data_.u_empty_match_check.repetition_limit; - assembler->IfRegisterLT(rep_reg, limit, &skip_empty_check); - } - // If the match is empty we bail out, otherwise we fall through - // to the on-success continuation. - assembler->IfRegisterEqPos(data_.u_empty_match_check.start_register, - trace->backtrack()); - assembler->Bind(&skip_empty_check); - on_success()->Emit(compiler, trace); - } - break; - } - case POSITIVE_SUBMATCH_SUCCESS: { - if (!trace->is_trivial()) { - trace->Flush(compiler, this); - return; - } - assembler->ReadCurrentPositionFromRegister( - data_.u_submatch.current_position_register); - assembler->ReadStackPointerFromRegister( - data_.u_submatch.stack_pointer_register); - int clear_register_count = data_.u_submatch.clear_register_count; - if (clear_register_count == 0) { - on_success()->Emit(compiler, trace); - return; - } - int clear_registers_from = data_.u_submatch.clear_register_from; - Label clear_registers_backtrack; - Trace new_trace = *trace; - new_trace.set_backtrack(&clear_registers_backtrack); - on_success()->Emit(compiler, &new_trace); - - assembler->Bind(&clear_registers_backtrack); - int clear_registers_to = clear_registers_from + clear_register_count - 1; - assembler->ClearRegisters(clear_registers_from, clear_registers_to); - - DCHECK(trace->backtrack() == nullptr); - assembler->Backtrack(); - return; - } - default: - UNREACHABLE(); - } -} - - -void BackReferenceNode::Emit(RegExpCompiler* compiler, Trace* trace) { - RegExpMacroAssembler* assembler = compiler->macro_assembler(); - if (!trace->is_trivial()) { - trace->Flush(compiler, this); - return; - } - - LimitResult limit_result = LimitVersions(compiler, trace); - if (limit_result == DONE) return; - DCHECK(limit_result == CONTINUE); - - RecursionCheck rc(compiler); - - DCHECK_EQ(start_reg_ + 1, end_reg_); - if (IgnoreCase(flags_)) { - assembler->CheckNotBackReferenceIgnoreCase( - start_reg_, read_backward(), IsUnicode(flags_), trace->backtrack()); - } else { - assembler->CheckNotBackReference(start_reg_, read_backward(), - trace->backtrack()); - } - // We are going to advance backward, so we may end up at the start. - if (read_backward()) trace->set_at_start(Trace::UNKNOWN); - - // Check that the back reference does not end inside a surrogate pair. - if (IsUnicode(flags_) && !compiler->one_byte()) { - assembler->CheckNotInSurrogatePair(trace->cp_offset(), trace->backtrack()); - } - on_success()->Emit(compiler, trace); -} - - -// ------------------------------------------------------------------- -// Dot/dotty output - - -#ifdef DEBUG - - -class DotPrinter: public NodeVisitor { - public: - DotPrinter(std::ostream& os, bool ignore_case) // NOLINT - : os_(os), - ignore_case_(ignore_case) {} - void PrintNode(const char* label, RegExpNode* node); - void Visit(RegExpNode* node); - void PrintAttributes(RegExpNode* from); - void PrintOnFailure(RegExpNode* from, RegExpNode* to); -#define DECLARE_VISIT(Type) \ - virtual void Visit##Type(Type##Node* that); -FOR_EACH_NODE_TYPE(DECLARE_VISIT) -#undef DECLARE_VISIT - private: - std::ostream& os_; - bool ignore_case_; -}; - - -void DotPrinter::PrintNode(const char* label, RegExpNode* node) { - os_ << "digraph G {\n graph [label=\""; - for (int i = 0; label[i]; i++) { - switch (label[i]) { - case '\\': - os_ << "\\\\"; - break; - case '"': - os_ << "\""; - break; - default: - os_ << label[i]; - break; - } - } - os_ << "\"];\n"; - Visit(node); - os_ << "}" << std::endl; -} - - -void DotPrinter::Visit(RegExpNode* node) { - if (node->info()->visited) return; - node->info()->visited = true; - node->Accept(this); -} - - -void DotPrinter::PrintOnFailure(RegExpNode* from, RegExpNode* on_failure) { - os_ << " n" << from << " -> n" << on_failure << " [style=dotted];\n"; - Visit(on_failure); -} - - -class TableEntryBodyPrinter { - public: - TableEntryBodyPrinter(std::ostream& os, ChoiceNode* choice) // NOLINT - : os_(os), - choice_(choice) {} - void Call(uc16 from, DispatchTable::Entry entry) { - OutSet* out_set = entry.out_set(); - for (unsigned i = 0; i < OutSet::kFirstLimit; i++) { - if (out_set->Get(i)) { - os_ << " n" << choice() << ":s" << from << "o" << i << " -> n" - << choice()->alternatives()->at(i).node() << ";\n"; - } - } - } - private: - ChoiceNode* choice() { return choice_; } - std::ostream& os_; - ChoiceNode* choice_; -}; - - -class TableEntryHeaderPrinter { - public: - explicit TableEntryHeaderPrinter(std::ostream& os) // NOLINT - : first_(true), - os_(os) {} - void Call(uc16 from, DispatchTable::Entry entry) { - if (first_) { - first_ = false; - } else { - os_ << "|"; - } - os_ << "{\\" << AsUC16(from) << "-\\" << AsUC16(entry.to()) << "|{"; - OutSet* out_set = entry.out_set(); - int priority = 0; - for (unsigned i = 0; i < OutSet::kFirstLimit; i++) { - if (out_set->Get(i)) { - if (priority > 0) os_ << "|"; - os_ << " " << priority; - priority++; - } - } - os_ << "}}"; - } - - private: - bool first_; - std::ostream& os_; -}; - - -class AttributePrinter { - public: - explicit AttributePrinter(std::ostream& os) // NOLINT - : os_(os), - first_(true) {} - void PrintSeparator() { - if (first_) { - first_ = false; - } else { - os_ << "|"; - } - } - void PrintBit(const char* name, bool value) { - if (!value) return; - PrintSeparator(); - os_ << "{" << name << "}"; - } - void PrintPositive(const char* name, int value) { - if (value < 0) return; - PrintSeparator(); - os_ << "{" << name << "|" << value << "}"; - } - - private: - std::ostream& os_; - bool first_; -}; - - -void DotPrinter::PrintAttributes(RegExpNode* that) { - os_ << " a" << that << " [shape=Mrecord, color=grey, fontcolor=grey, " - << "margin=0.1, fontsize=10, label=\"{"; - AttributePrinter printer(os_); - NodeInfo* info = that->info(); - printer.PrintBit("NI", info->follows_newline_interest); - printer.PrintBit("WI", info->follows_word_interest); - printer.PrintBit("SI", info->follows_start_interest); - Label* label = that->label(); - if (label->is_bound()) - printer.PrintPositive("@", label->pos()); - os_ << "}\"];\n" - << " a" << that << " -> n" << that - << " [style=dashed, color=grey, arrowhead=none];\n"; -} - - -static const bool kPrintDispatchTable = false; -void DotPrinter::VisitChoice(ChoiceNode* that) { - if (kPrintDispatchTable) { - os_ << " n" << that << " [shape=Mrecord, label=\""; - TableEntryHeaderPrinter header_printer(os_); - that->GetTable(ignore_case_)->ForEach(&header_printer); - os_ << "\"]\n"; - PrintAttributes(that); - TableEntryBodyPrinter body_printer(os_, that); - that->GetTable(ignore_case_)->ForEach(&body_printer); - } else { - os_ << " n" << that << " [shape=Mrecord, label=\"?\"];\n"; - for (int i = 0; i < that->alternatives()->length(); i++) { - GuardedAlternative alt = that->alternatives()->at(i); - os_ << " n" << that << " -> n" << alt.node(); - } - } - for (int i = 0; i < that->alternatives()->length(); i++) { - GuardedAlternative alt = that->alternatives()->at(i); - alt.node()->Accept(this); - } -} - - -void DotPrinter::VisitText(TextNode* that) { - Zone* zone = that->zone(); - os_ << " n" << that << " [label=\""; - for (int i = 0; i < that->elements()->length(); i++) { - if (i > 0) os_ << " "; - TextElement elm = that->elements()->at(i); - switch (elm.text_type()) { - case TextElement::ATOM: { - Vector data = elm.atom()->data(); - for (int i = 0; i < data.length(); i++) { - os_ << static_cast(data[i]); - } - break; - } - case TextElement::CHAR_CLASS: { - RegExpCharacterClass* node = elm.char_class(); - os_ << "["; - if (node->is_negated()) os_ << "^"; - for (int j = 0; j < node->ranges(zone)->length(); j++) { - CharacterRange range = node->ranges(zone)->at(j); - os_ << AsUC16(range.from()) << "-" << AsUC16(range.to()); - } - os_ << "]"; - break; - } - default: - UNREACHABLE(); - } - } - os_ << "\", shape=box, peripheries=2];\n"; - PrintAttributes(that); - os_ << " n" << that << " -> n" << that->on_success() << ";\n"; - Visit(that->on_success()); -} - - -void DotPrinter::VisitBackReference(BackReferenceNode* that) { - os_ << " n" << that << " [label=\"$" << that->start_register() << "..$" - << that->end_register() << "\", shape=doubleoctagon];\n"; - PrintAttributes(that); - os_ << " n" << that << " -> n" << that->on_success() << ";\n"; - Visit(that->on_success()); -} - - -void DotPrinter::VisitEnd(EndNode* that) { - os_ << " n" << that << " [style=bold, shape=point];\n"; - PrintAttributes(that); -} - - -void DotPrinter::VisitAssertion(AssertionNode* that) { - os_ << " n" << that << " ["; - switch (that->assertion_type()) { - case AssertionNode::AT_END: - os_ << "label=\"$\", shape=septagon"; - break; - case AssertionNode::AT_START: - os_ << "label=\"^\", shape=septagon"; - break; - case AssertionNode::AT_BOUNDARY: - os_ << "label=\"\\b\", shape=septagon"; - break; - case AssertionNode::AT_NON_BOUNDARY: - os_ << "label=\"\\B\", shape=septagon"; - break; - case AssertionNode::AFTER_NEWLINE: - os_ << "label=\"(?<=\\n)\", shape=septagon"; - break; - } - os_ << "];\n"; - PrintAttributes(that); - RegExpNode* successor = that->on_success(); - os_ << " n" << that << " -> n" << successor << ";\n"; - Visit(successor); -} - - -void DotPrinter::VisitAction(ActionNode* that) { - os_ << " n" << that << " ["; - switch (that->action_type_) { - case ActionNode::SET_REGISTER: - os_ << "label=\"$" << that->data_.u_store_register.reg - << ":=" << that->data_.u_store_register.value << "\", shape=octagon"; - break; - case ActionNode::INCREMENT_REGISTER: - os_ << "label=\"$" << that->data_.u_increment_register.reg - << "++\", shape=octagon"; - break; - case ActionNode::STORE_POSITION: - os_ << "label=\"$" << that->data_.u_position_register.reg - << ":=$pos\", shape=octagon"; - break; - case ActionNode::BEGIN_SUBMATCH: - os_ << "label=\"$" << that->data_.u_submatch.current_position_register - << ":=$pos,begin\", shape=septagon"; - break; - case ActionNode::POSITIVE_SUBMATCH_SUCCESS: - os_ << "label=\"escape\", shape=septagon"; - break; - case ActionNode::EMPTY_MATCH_CHECK: - os_ << "label=\"$" << that->data_.u_empty_match_check.start_register - << "=$pos?,$" << that->data_.u_empty_match_check.repetition_register - << "<" << that->data_.u_empty_match_check.repetition_limit - << "?\", shape=septagon"; - break; - case ActionNode::CLEAR_CAPTURES: { - os_ << "label=\"clear $" << that->data_.u_clear_captures.range_from - << " to $" << that->data_.u_clear_captures.range_to - << "\", shape=septagon"; - break; - } - } - os_ << "];\n"; - PrintAttributes(that); - RegExpNode* successor = that->on_success(); - os_ << " n" << that << " -> n" << successor << ";\n"; - Visit(successor); -} - - -class DispatchTableDumper { - public: - explicit DispatchTableDumper(std::ostream& os) : os_(os) {} - void Call(uc16 key, DispatchTable::Entry entry); - private: - std::ostream& os_; -}; - - -void DispatchTableDumper::Call(uc16 key, DispatchTable::Entry entry) { - os_ << "[" << AsUC16(key) << "-" << AsUC16(entry.to()) << "]: {"; - OutSet* set = entry.out_set(); - bool first = true; - for (unsigned i = 0; i < OutSet::kFirstLimit; i++) { - if (set->Get(i)) { - if (first) { - first = false; - } else { - os_ << ", "; - } - os_ << i; - } - } - os_ << "}\n"; -} - - -void DispatchTable::Dump() { - OFStream os(stderr); - DispatchTableDumper dumper(os); - tree()->ForEach(&dumper); -} - - -void RegExpEngine::DotPrint(const char* label, - RegExpNode* node, +void RegExpEngine::DotPrint(const char* label, RegExpNode* node, bool ignore_case) { - StdoutStream os; - DotPrinter printer(os, ignore_case); - printer.PrintNode(label, node); + DotPrinter::DotPrint(label, node, ignore_case); } - -#endif // DEBUG - -// ------------------------------------------------------------------- -// Splay tree - - -OutSet* OutSet::Extend(unsigned value, Zone* zone) { - if (Get(value)) - return this; - if (successors(zone) != nullptr) { - for (int i = 0; i < successors(zone)->length(); i++) { - OutSet* successor = successors(zone)->at(i); - if (successor->Get(value)) - return successor; - } - } else { - successors_ = new(zone) ZoneList(2, zone); - } - OutSet* result = new(zone) OutSet(first_, remaining_); - result->Set(value, zone); - successors(zone)->Add(result, zone); - return result; -} - - -void OutSet::Set(unsigned value, Zone *zone) { - if (value < kFirstLimit) { - first_ |= (1 << value); - } else { - if (remaining_ == nullptr) - remaining_ = new(zone) ZoneList(1, zone); - if (remaining_->is_empty() || !remaining_->Contains(value)) - remaining_->Add(value, zone); - } -} - - -bool OutSet::Get(unsigned value) const { - if (value < kFirstLimit) { - return (first_ & (1 << value)) != 0; - } else if (remaining_ == nullptr) { - return false; - } else { - return remaining_->Contains(value); - } -} - - -const uc32 DispatchTable::Config::kNoKey = unibrow::Utf8::kBadChar; - - -void DispatchTable::AddRange(CharacterRange full_range, int value, - Zone* zone) { - CharacterRange current = full_range; - if (tree()->is_empty()) { - // If this is the first range we just insert into the table. - ZoneSplayTree::Locator loc; - bool inserted = tree()->Insert(current.from(), &loc); - DCHECK(inserted); - USE(inserted); - loc.set_value(Entry(current.from(), current.to(), - empty()->Extend(value, zone))); - return; - } - // First see if there is a range to the left of this one that - // overlaps. - ZoneSplayTree::Locator loc; - if (tree()->FindGreatestLessThan(current.from(), &loc)) { - Entry* entry = &loc.value(); - // If we've found a range that overlaps with this one, and it - // starts strictly to the left of this one, we have to fix it - // because the following code only handles ranges that start on - // or after the start point of the range we're adding. - if (entry->from() < current.from() && entry->to() >= current.from()) { - // Snap the overlapping range in half around the start point of - // the range we're adding. - CharacterRange left = - CharacterRange::Range(entry->from(), current.from() - 1); - CharacterRange right = CharacterRange::Range(current.from(), entry->to()); - // The left part of the overlapping range doesn't overlap. - // Truncate the whole entry to be just the left part. - entry->set_to(left.to()); - // The right part is the one that overlaps. We add this part - // to the map and let the next step deal with merging it with - // the range we're adding. - ZoneSplayTree::Locator loc; - bool inserted = tree()->Insert(right.from(), &loc); - DCHECK(inserted); - USE(inserted); - loc.set_value(Entry(right.from(), - right.to(), - entry->out_set())); - } - } - while (current.is_valid()) { - if (tree()->FindLeastGreaterThan(current.from(), &loc) && - (loc.value().from() <= current.to()) && - (loc.value().to() >= current.from())) { - Entry* entry = &loc.value(); - // We have overlap. If there is space between the start point of - // the range we're adding and where the overlapping range starts - // then we have to add a range covering just that space. - if (current.from() < entry->from()) { - ZoneSplayTree::Locator ins; - bool inserted = tree()->Insert(current.from(), &ins); - DCHECK(inserted); - USE(inserted); - ins.set_value(Entry(current.from(), - entry->from() - 1, - empty()->Extend(value, zone))); - current.set_from(entry->from()); - } - DCHECK_EQ(current.from(), entry->from()); - // If the overlapping range extends beyond the one we want to add - // we have to snap the right part off and add it separately. - if (entry->to() > current.to()) { - ZoneSplayTree::Locator ins; - bool inserted = tree()->Insert(current.to() + 1, &ins); - DCHECK(inserted); - USE(inserted); - ins.set_value(Entry(current.to() + 1, - entry->to(), - entry->out_set())); - entry->set_to(current.to()); - } - DCHECK(entry->to() <= current.to()); - // The overlapping range is now completely contained by the range - // we're adding so we can just update it and move the start point - // of the range we're adding just past it. - entry->AddValue(value, zone); - DCHECK(entry->to() + 1 > current.from()); - current.set_from(entry->to() + 1); - } else { - // There is no overlap so we can just add the range - ZoneSplayTree::Locator ins; - bool inserted = tree()->Insert(current.from(), &ins); - DCHECK(inserted); - USE(inserted); - ins.set_value(Entry(current.from(), - current.to(), - empty()->Extend(value, zone))); - break; - } - } -} - - -OutSet* DispatchTable::Get(uc32 value) { - ZoneSplayTree::Locator loc; - if (!tree()->FindGreatestLessThan(value, &loc)) - return empty(); - Entry* entry = &loc.value(); - if (value <= entry->to()) - return entry->out_set(); - else - return empty(); -} - - -// ------------------------------------------------------------------- -// Analysis - - -void Analysis::EnsureAnalyzed(RegExpNode* that) { - StackLimitCheck check(isolate()); - if (check.HasOverflowed()) { - fail("Stack overflow"); - return; - } - if (that->info()->been_analyzed || that->info()->being_analyzed) - return; - that->info()->being_analyzed = true; - that->Accept(this); - that->info()->being_analyzed = false; - that->info()->been_analyzed = true; -} - - -void Analysis::VisitEnd(EndNode* that) { - // nothing to do -} - - -void TextNode::CalculateOffsets() { - int element_count = elements()->length(); - // Set up the offsets of the elements relative to the start. This is a fixed - // quantity since a TextNode can only contain fixed-width things. - int cp_offset = 0; - for (int i = 0; i < element_count; i++) { - TextElement& elm = elements()->at(i); - elm.set_cp_offset(cp_offset); - cp_offset += elm.length(); - } -} - - -void Analysis::VisitText(TextNode* that) { - that->MakeCaseIndependent(isolate(), is_one_byte_); - EnsureAnalyzed(that->on_success()); - if (!has_failed()) { - that->CalculateOffsets(); - } -} - - -void Analysis::VisitAction(ActionNode* that) { - RegExpNode* target = that->on_success(); - EnsureAnalyzed(target); - if (!has_failed()) { - // If the next node is interested in what it follows then this node - // has to be interested too so it can pass the information on. - that->info()->AddFromFollowing(target->info()); - } -} - - -void Analysis::VisitChoice(ChoiceNode* that) { - NodeInfo* info = that->info(); - for (int i = 0; i < that->alternatives()->length(); i++) { - RegExpNode* node = that->alternatives()->at(i).node(); - EnsureAnalyzed(node); - if (has_failed()) return; - // Anything the following nodes need to know has to be known by - // this node also, so it can pass it on. - info->AddFromFollowing(node->info()); - } -} - - -void Analysis::VisitLoopChoice(LoopChoiceNode* that) { - NodeInfo* info = that->info(); - for (int i = 0; i < that->alternatives()->length(); i++) { - RegExpNode* node = that->alternatives()->at(i).node(); - if (node != that->loop_node()) { - EnsureAnalyzed(node); - if (has_failed()) return; - info->AddFromFollowing(node->info()); - } - } - // Check the loop last since it may need the value of this node - // to get a correct result. - EnsureAnalyzed(that->loop_node()); - if (!has_failed()) { - info->AddFromFollowing(that->loop_node()->info()); - } -} - - -void Analysis::VisitBackReference(BackReferenceNode* that) { - EnsureAnalyzed(that->on_success()); -} - - -void Analysis::VisitAssertion(AssertionNode* that) { - EnsureAnalyzed(that->on_success()); -} - - -void BackReferenceNode::FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, - bool not_at_start) { - // Working out the set of characters that a backreference can match is too - // hard, so we just say that any character can match. - bm->SetRest(offset); - SaveBMInfo(bm, not_at_start, offset); -} - - -STATIC_ASSERT(BoyerMoorePositionInfo::kMapSize == - RegExpMacroAssembler::kTableSize); - - -void ChoiceNode::FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) { - ZoneList* alts = alternatives(); - budget = (budget - 1) / alts->length(); - for (int i = 0; i < alts->length(); i++) { - GuardedAlternative& alt = alts->at(i); - if (alt.guards() != nullptr && alt.guards()->length() != 0) { - bm->SetRest(offset); // Give up trying to fill in info. - SaveBMInfo(bm, not_at_start, offset); - return; - } - alt.node()->FillInBMInfo(isolate, offset, budget, bm, not_at_start); - } - SaveBMInfo(bm, not_at_start, offset); -} - - -void TextNode::FillInBMInfo(Isolate* isolate, int initial_offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) { - if (initial_offset >= bm->length()) return; - int offset = initial_offset; - int max_char = bm->max_char(); - for (int i = 0; i < elements()->length(); i++) { - if (offset >= bm->length()) { - if (initial_offset == 0) set_bm_info(not_at_start, bm); - return; - } - TextElement text = elements()->at(i); - if (text.text_type() == TextElement::ATOM) { - RegExpAtom* atom = text.atom(); - for (int j = 0; j < atom->length(); j++, offset++) { - if (offset >= bm->length()) { - if (initial_offset == 0) set_bm_info(not_at_start, bm); - return; - } - uc16 character = atom->data()[j]; - if (IgnoreCase(atom->flags())) { - unibrow::uchar chars[4]; - int length = GetCaseIndependentLetters( - isolate, character, bm->max_char() == String::kMaxOneByteCharCode, - chars, 4); - for (int j = 0; j < length; j++) { - bm->Set(offset, chars[j]); - } - } else { - if (character <= max_char) bm->Set(offset, character); - } - } - } else { - DCHECK_EQ(TextElement::CHAR_CLASS, text.text_type()); - RegExpCharacterClass* char_class = text.char_class(); - ZoneList* ranges = char_class->ranges(zone()); - if (char_class->is_negated()) { - bm->SetAll(offset); - } else { - for (int k = 0; k < ranges->length(); k++) { - CharacterRange& range = ranges->at(k); - if (range.from() > max_char) continue; - int to = Min(max_char, static_cast(range.to())); - bm->SetInterval(offset, Interval(range.from(), to)); - } - } - offset++; - } - } - if (offset >= bm->length()) { - if (initial_offset == 0) set_bm_info(not_at_start, bm); - return; - } - on_success()->FillInBMInfo(isolate, offset, budget - 1, bm, - true); // Not at start after a text node. - if (initial_offset == 0) set_bm_info(not_at_start, bm); -} - - -// ------------------------------------------------------------------- -// Dispatch table construction - - -void DispatchTableConstructor::VisitEnd(EndNode* that) { - AddRange(CharacterRange::Everything()); -} - - -void DispatchTableConstructor::BuildTable(ChoiceNode* node) { - node->set_being_calculated(true); - ZoneList* alternatives = node->alternatives(); - for (int i = 0; i < alternatives->length(); i++) { - set_choice_index(i); - alternatives->at(i).node()->Accept(this); - } - node->set_being_calculated(false); -} - - -class AddDispatchRange { - public: - explicit AddDispatchRange(DispatchTableConstructor* constructor) - : constructor_(constructor) { } - void Call(uc32 from, DispatchTable::Entry entry); - private: - DispatchTableConstructor* constructor_; -}; - - -void AddDispatchRange::Call(uc32 from, DispatchTable::Entry entry) { - constructor_->AddRange(CharacterRange::Range(from, entry.to())); -} - - -void DispatchTableConstructor::VisitChoice(ChoiceNode* node) { - if (node->being_calculated()) - return; - DispatchTable* table = node->GetTable(ignore_case_); - AddDispatchRange adder(this); - table->ForEach(&adder); -} - - -void DispatchTableConstructor::VisitBackReference(BackReferenceNode* that) { - // TODO(160): Find the node that we refer back to and propagate its start - // set back to here. For now we just accept anything. - AddRange(CharacterRange::Everything()); -} - - -void DispatchTableConstructor::VisitAssertion(AssertionNode* that) { - RegExpNode* target = that->on_success(); - target->Accept(this); -} - - -static int CompareRangeByFrom(const CharacterRange* a, - const CharacterRange* b) { - return Compare(a->from(), b->from()); -} - - -void DispatchTableConstructor::AddInverse(ZoneList* ranges) { - ranges->Sort(CompareRangeByFrom); - uc16 last = 0; - for (int i = 0; i < ranges->length(); i++) { - CharacterRange range = ranges->at(i); - if (last < range.from()) - AddRange(CharacterRange::Range(last, range.from() - 1)); - if (range.to() >= last) { - if (range.to() == String::kMaxCodePoint) { - return; - } else { - last = range.to() + 1; - } - } - } - AddRange(CharacterRange::Range(last, String::kMaxCodePoint)); -} - - -void DispatchTableConstructor::VisitText(TextNode* that) { - TextElement elm = that->elements()->at(0); - switch (elm.text_type()) { - case TextElement::ATOM: { - uc16 c = elm.atom()->data()[0]; - AddRange(CharacterRange::Range(c, c)); - break; - } - case TextElement::CHAR_CLASS: { - RegExpCharacterClass* tree = elm.char_class(); - ZoneList* ranges = tree->ranges(that->zone()); - if (tree->is_negated()) { - AddInverse(ranges); - } else { - for (int i = 0; i < ranges->length(); i++) - AddRange(ranges->at(i)); - } - break; - } - default: { - UNIMPLEMENTED(); - } - } -} - - -void DispatchTableConstructor::VisitAction(ActionNode* that) { - RegExpNode* target = that->on_success(); - target->Accept(this); -} - -RegExpNode* OptionallyStepBackToLeadSurrogate(RegExpCompiler* compiler, - RegExpNode* on_success, - JSRegExp::Flags flags) { - // If the regexp matching starts within a surrogate pair, step back - // to the lead surrogate and start matching from there. - DCHECK(!compiler->read_backward()); - Zone* zone = compiler->zone(); - ZoneList* lead_surrogates = CharacterRange::List( - zone, CharacterRange::Range(kLeadSurrogateStart, kLeadSurrogateEnd)); - ZoneList* trail_surrogates = CharacterRange::List( - zone, CharacterRange::Range(kTrailSurrogateStart, kTrailSurrogateEnd)); - - ChoiceNode* optional_step_back = new (zone) ChoiceNode(2, zone); - - int stack_register = compiler->UnicodeLookaroundStackRegister(); - int position_register = compiler->UnicodeLookaroundPositionRegister(); - RegExpNode* step_back = TextNode::CreateForCharacterRanges( - zone, lead_surrogates, true, on_success, flags); - RegExpLookaround::Builder builder(true, step_back, stack_register, - position_register); - RegExpNode* match_trail = TextNode::CreateForCharacterRanges( - zone, trail_surrogates, false, builder.on_match_success(), flags); - - optional_step_back->AddAlternative( - GuardedAlternative(builder.ForMatch(match_trail))); - optional_step_back->AddAlternative(GuardedAlternative(on_success)); - - return optional_step_back; -} - - RegExpEngine::CompilationResult RegExpEngine::Compile( Isolate* isolate, Zone* zone, RegExpCompileData* data, JSRegExp::Flags flags, Handle pattern, Handle sample_subject, bool is_one_byte) { if ((data->capture_count + 1) * 2 - 1 > RegExpMacroAssembler::kMaxRegister) { - return IrregexpRegExpTooBig(isolate); + return CompilationResult::RegExpTooBig(); } bool is_sticky = IsSticky(flags); bool is_global = IsGlobal(flags); @@ -5030,10 +648,8 @@ RegExpEngine::CompilationResult RegExpEngine::Compile( } // Wrap the body of the regexp in capture #0. - RegExpNode* captured_body = RegExpCapture::ToNode(data->tree, - 0, - &compiler, - compiler.accept()); + RegExpNode* captured_body = + RegExpCapture::ToNode(data->tree, 0, &compiler, compiler.accept()); RegExpNode* node = captured_body; bool is_end_anchored = data->tree->IsAnchoredAtEnd(); bool is_start_anchored = data->tree->IsAnchoredAtStart(); @@ -5050,7 +666,7 @@ RegExpEngine::CompilationResult RegExpEngine::Compile( if (data->contains_anchor) { // Unroll loop once, to take care of the case that might start // at the start of input. - ChoiceNode* first_step_node = new(zone) ChoiceNode(2, zone); + ChoiceNode* first_step_node = new (zone) ChoiceNode(2, zone); first_step_node->AddAlternative(GuardedAlternative(captured_body)); first_step_node->AddAlternative(GuardedAlternative(new (zone) TextNode( new (zone) RegExpCharacterClass('*', default_flags), false, @@ -5068,7 +684,8 @@ RegExpEngine::CompilationResult RegExpEngine::Compile( node = node->FilterOneByte(RegExpCompiler::kMaxRecursion); } } else if (is_unicode && (is_global || is_sticky)) { - node = OptionallyStepBackToLeadSurrogate(&compiler, node, flags); + node = RegExpCompiler::OptionallyStepBackToLeadSurrogate(&compiler, node, + flags); } if (node == nullptr) node = new (zone) EndNode(EndNode::BACKTRACK, zone); @@ -5076,8 +693,7 @@ RegExpEngine::CompilationResult RegExpEngine::Compile( Analysis analysis(isolate, is_one_byte); analysis.EnsureAnalyzed(node); if (analysis.has_failed()) { - const char* error_message = analysis.error_message(); - return CompilationResult(isolate, error_message); + return CompilationResult(analysis.error_message()); } // Create the correct assembler for the architecture. @@ -5144,8 +760,11 @@ RegExpEngine::CompilationResult RegExpEngine::Compile( macro_assembler->set_global_mode(mode); } - return compiler.Assemble(isolate, macro_assembler.get(), node, - data->capture_count, pattern); + RegExpCompiler::CompilationResult result = compiler.Assemble( + isolate, macro_assembler.get(), node, data->capture_count, pattern); + + return RegExpEngine::CompilationResult(result.error_message, result.code, + result.num_registers); } bool RegExpEngine::TooMuchRegExpCode(Isolate* isolate, Handle pattern) { @@ -5258,12 +877,5 @@ void RegExpResultsCache::Clear(FixedArray cache) { } } -// We need to check for the following characters: 0x39C 0x3BC 0x178. -bool RangeContainsLatin1Equivalents(CharacterRange range) { - // TODO(dcarney): this could be a lot more efficient. - return range.Contains(0x039C) || range.Contains(0x03BC) || - range.Contains(0x0178); -} - } // namespace internal } // namespace v8 diff --git a/src/regexp/jsregexp.h b/src/regexp/jsregexp.h index 3ae88db5d0..a4fe0ca8d4 100644 --- a/src/regexp/jsregexp.h +++ b/src/regexp/jsregexp.h @@ -5,22 +5,13 @@ #ifndef V8_REGEXP_JSREGEXP_H_ #define V8_REGEXP_JSREGEXP_H_ -#include "src/execution/isolate.h" #include "src/objects/js-regexp.h" -#include "src/regexp/regexp-ast.h" -#include "src/regexp/regexp-macro-assembler.h" -#include "src/utils/allocation.h" -#include "src/zone/zone-splay-tree.h" namespace v8 { namespace internal { -class NodeVisitor; -class RegExpCompiler; -class RegExpMacroAssembler; class RegExpNode; class RegExpTree; -class BoyerMooreLookahead; inline bool IgnoreCase(JSRegExp::Flags flags) { return (flags & JSRegExp::kIgnoreCase) != 0; @@ -52,7 +43,7 @@ inline bool NeedsUnicodeCaseEquivalents(JSRegExp::Flags flags) { return IsUnicode(flags) && IgnoreCase(flags); } -class RegExpImpl { +class RegExpImpl final { public: // Whether the irregexp engine generates native code or interpreter bytecode. static bool UsesNativeRegExp() { return !FLAG_regexp_interpret_all; } @@ -190,1265 +181,6 @@ class RegExpImpl { bool is_one_byte); }; - -// Represents the location of one element relative to the intersection of -// two sets. Corresponds to the four areas of a Venn diagram. -enum ElementInSetsRelation { - kInsideNone = 0, - kInsideFirst = 1, - kInsideSecond = 2, - kInsideBoth = 3 -}; - - -// A set of unsigned integers that behaves especially well on small -// integers (< 32). May do zone-allocation. -class OutSet: public ZoneObject { - public: - OutSet() : first_(0), remaining_(nullptr), successors_(nullptr) {} - OutSet* Extend(unsigned value, Zone* zone); - V8_EXPORT_PRIVATE bool Get(unsigned value) const; - static const unsigned kFirstLimit = 32; - - private: - // Destructively set a value in this set. In most cases you want - // to use Extend instead to ensure that only one instance exists - // that contains the same values. - void Set(unsigned value, Zone* zone); - - // The successors are a list of sets that contain the same values - // as this set and the one more value that is not present in this - // set. - ZoneList* successors(Zone* zone) { return successors_; } - - OutSet(uint32_t first, ZoneList* remaining) - : first_(first), remaining_(remaining), successors_(nullptr) {} - uint32_t first_; - ZoneList* remaining_; - ZoneList* successors_; - friend class Trace; -}; - - -// A mapping from integers, specified as ranges, to a set of integers. -// Used for mapping character ranges to choices. -class DispatchTable : public ZoneObject { - public: - explicit DispatchTable(Zone* zone) : tree_(zone) { } - - class Entry { - public: - Entry() : from_(0), to_(0), out_set_(nullptr) {} - Entry(uc32 from, uc32 to, OutSet* out_set) - : from_(from), to_(to), out_set_(out_set) { - DCHECK(from <= to); - } - uc32 from() { return from_; } - uc32 to() { return to_; } - void set_to(uc32 value) { to_ = value; } - void AddValue(int value, Zone* zone) { - out_set_ = out_set_->Extend(value, zone); - } - OutSet* out_set() { return out_set_; } - private: - uc32 from_; - uc32 to_; - OutSet* out_set_; - }; - - class Config { - public: - using Key = uc32; - using Value = Entry; - static const uc32 kNoKey; - static const Entry NoValue() { return Value(); } - static inline int Compare(uc32 a, uc32 b) { - if (a == b) - return 0; - else if (a < b) - return -1; - else - return 1; - } - }; - - V8_EXPORT_PRIVATE void AddRange(CharacterRange range, int value, Zone* zone); - V8_EXPORT_PRIVATE OutSet* Get(uc32 value); - void Dump(); - - template - void ForEach(Callback* callback) { - return tree()->ForEach(callback); - } - - private: - // There can't be a static empty set since it allocates its - // successors in a zone and caches them. - OutSet* empty() { return &empty_; } - OutSet empty_; - ZoneSplayTree* tree() { return &tree_; } - ZoneSplayTree tree_; -}; - -#define FOR_EACH_NODE_TYPE(VISIT) \ - VISIT(End) \ - VISIT(Action) \ - VISIT(Choice) \ - VISIT(BackReference) \ - VISIT(Assertion) \ - VISIT(Text) - - -class Trace; -struct PreloadState; -class GreedyLoopState; -class AlternativeGenerationList; - -struct NodeInfo { - NodeInfo() - : being_analyzed(false), - been_analyzed(false), - follows_word_interest(false), - follows_newline_interest(false), - follows_start_interest(false), - at_end(false), - visited(false), - replacement_calculated(false) { } - - // Returns true if the interests and assumptions of this node - // matches the given one. - bool Matches(NodeInfo* that) { - return (at_end == that->at_end) && - (follows_word_interest == that->follows_word_interest) && - (follows_newline_interest == that->follows_newline_interest) && - (follows_start_interest == that->follows_start_interest); - } - - // Updates the interests of this node given the interests of the - // node preceding it. - void AddFromPreceding(NodeInfo* that) { - at_end |= that->at_end; - follows_word_interest |= that->follows_word_interest; - follows_newline_interest |= that->follows_newline_interest; - follows_start_interest |= that->follows_start_interest; - } - - bool HasLookbehind() { - return follows_word_interest || - follows_newline_interest || - follows_start_interest; - } - - // Sets the interests of this node to include the interests of the - // following node. - void AddFromFollowing(NodeInfo* that) { - follows_word_interest |= that->follows_word_interest; - follows_newline_interest |= that->follows_newline_interest; - follows_start_interest |= that->follows_start_interest; - } - - void ResetCompilationState() { - being_analyzed = false; - been_analyzed = false; - } - - bool being_analyzed: 1; - bool been_analyzed: 1; - - // These bits are set of this node has to know what the preceding - // character was. - bool follows_word_interest: 1; - bool follows_newline_interest: 1; - bool follows_start_interest: 1; - - bool at_end: 1; - bool visited: 1; - bool replacement_calculated: 1; -}; - - -// Details of a quick mask-compare check that can look ahead in the -// input stream. -class QuickCheckDetails { - public: - QuickCheckDetails() - : characters_(0), - mask_(0), - value_(0), - cannot_match_(false) { } - explicit QuickCheckDetails(int characters) - : characters_(characters), - mask_(0), - value_(0), - cannot_match_(false) { } - bool Rationalize(bool one_byte); - // Merge in the information from another branch of an alternation. - void Merge(QuickCheckDetails* other, int from_index); - // Advance the current position by some amount. - void Advance(int by, bool one_byte); - void Clear(); - bool cannot_match() { return cannot_match_; } - void set_cannot_match() { cannot_match_ = true; } - struct Position { - Position() : mask(0), value(0), determines_perfectly(false) { } - uc16 mask; - uc16 value; - bool determines_perfectly; - }; - int characters() { return characters_; } - void set_characters(int characters) { characters_ = characters; } - Position* positions(int index) { - DCHECK_LE(0, index); - DCHECK_GT(characters_, index); - return positions_ + index; - } - uint32_t mask() { return mask_; } - uint32_t value() { return value_; } - - private: - // How many characters do we have quick check information from. This is - // the same for all branches of a choice node. - int characters_; - Position positions_[4]; - // These values are the condensate of the above array after Rationalize(). - uint32_t mask_; - uint32_t value_; - // If set to true, there is no way this quick check can match at all. - // E.g., if it requires to be at the start of the input, and isn't. - bool cannot_match_; -}; - - -extern int kUninitializedRegExpNodePlaceHolder; - - -class RegExpNode: public ZoneObject { - public: - explicit RegExpNode(Zone* zone) - : replacement_(nullptr), - on_work_list_(false), - trace_count_(0), - zone_(zone) { - bm_info_[0] = bm_info_[1] = nullptr; - } - virtual ~RegExpNode(); - virtual void Accept(NodeVisitor* visitor) = 0; - // Generates a goto to this node or actually generates the code at this point. - virtual void Emit(RegExpCompiler* compiler, Trace* trace) = 0; - // How many characters must this node consume at a minimum in order to - // succeed. If we have found at least 'still_to_find' characters that - // must be consumed there is no need to ask any following nodes whether - // they are sure to eat any more characters. The not_at_start argument is - // used to indicate that we know we are not at the start of the input. In - // this case anchored branches will always fail and can be ignored when - // determining how many characters are consumed on success. - virtual int EatsAtLeast(int still_to_find, int budget, bool not_at_start) = 0; - // Emits some quick code that checks whether the preloaded characters match. - // Falls through on certain failure, jumps to the label on possible success. - // If the node cannot make a quick check it does nothing and returns false. - bool EmitQuickCheck(RegExpCompiler* compiler, - Trace* bounds_check_trace, - Trace* trace, - bool preload_has_checked_bounds, - Label* on_possible_success, - QuickCheckDetails* details_return, - bool fall_through_on_failure); - // For a given number of characters this returns a mask and a value. The - // next n characters are anded with the mask and compared with the value. - // A comparison failure indicates the node cannot match the next n characters. - // A comparison success indicates the node may match. - virtual void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, - int characters_filled_in, - bool not_at_start) = 0; - static const int kNodeIsTooComplexForGreedyLoops = kMinInt; - virtual int GreedyLoopTextLength() { return kNodeIsTooComplexForGreedyLoops; } - // Only returns the successor for a text node of length 1 that matches any - // character and that has no guards on it. - virtual RegExpNode* GetSuccessorOfOmnivorousTextNode( - RegExpCompiler* compiler) { - return nullptr; - } - - // Collects information on the possible code units (mod 128) that can match if - // we look forward. This is used for a Boyer-Moore-like string searching - // implementation. TODO(erikcorry): This should share more code with - // EatsAtLeast, GetQuickCheckDetails. The budget argument is used to limit - // the number of nodes we are willing to look at in order to create this data. - static const int kRecursionBudget = 200; - bool KeepRecursing(RegExpCompiler* compiler); - virtual void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) { - UNREACHABLE(); - } - - // If we know that the input is one-byte then there are some nodes that can - // never match. This method returns a node that can be substituted for - // itself, or nullptr if the node can never match. - virtual RegExpNode* FilterOneByte(int depth) { return this; } - // Helper for FilterOneByte. - RegExpNode* replacement() { - DCHECK(info()->replacement_calculated); - return replacement_; - } - RegExpNode* set_replacement(RegExpNode* replacement) { - info()->replacement_calculated = true; - replacement_ = replacement; - return replacement; // For convenience. - } - - // We want to avoid recalculating the lookahead info, so we store it on the - // node. Only info that is for this node is stored. We can tell that the - // info is for this node when offset == 0, so the information is calculated - // relative to this node. - void SaveBMInfo(BoyerMooreLookahead* bm, bool not_at_start, int offset) { - if (offset == 0) set_bm_info(not_at_start, bm); - } - - Label* label() { return &label_; } - // If non-generic code is generated for a node (i.e. the node is not at the - // start of the trace) then it cannot be reused. This variable sets a limit - // on how often we allow that to happen before we insist on starting a new - // trace and generating generic code for a node that can be reused by flushing - // the deferred actions in the current trace and generating a goto. - static const int kMaxCopiesCodeGenerated = 10; - - bool on_work_list() { return on_work_list_; } - void set_on_work_list(bool value) { on_work_list_ = value; } - - NodeInfo* info() { return &info_; } - - BoyerMooreLookahead* bm_info(bool not_at_start) { - return bm_info_[not_at_start ? 1 : 0]; - } - - Zone* zone() const { return zone_; } - - protected: - enum LimitResult { DONE, CONTINUE }; - RegExpNode* replacement_; - - LimitResult LimitVersions(RegExpCompiler* compiler, Trace* trace); - - void set_bm_info(bool not_at_start, BoyerMooreLookahead* bm) { - bm_info_[not_at_start ? 1 : 0] = bm; - } - - private: - static const int kFirstCharBudget = 10; - Label label_; - bool on_work_list_; - NodeInfo info_; - // This variable keeps track of how many times code has been generated for - // this node (in different traces). We don't keep track of where the - // generated code is located unless the code is generated at the start of - // a trace, in which case it is generic and can be reused by flushing the - // deferred operations in the current trace and generating a goto. - int trace_count_; - BoyerMooreLookahead* bm_info_[2]; - - Zone* zone_; -}; - - -class SeqRegExpNode: public RegExpNode { - public: - explicit SeqRegExpNode(RegExpNode* on_success) - : RegExpNode(on_success->zone()), on_success_(on_success) { } - RegExpNode* on_success() { return on_success_; } - void set_on_success(RegExpNode* node) { on_success_ = node; } - RegExpNode* FilterOneByte(int depth) override; - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override { - on_success_->FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start); - if (offset == 0) set_bm_info(not_at_start, bm); - } - - protected: - RegExpNode* FilterSuccessor(int depth); - - private: - RegExpNode* on_success_; -}; - - -class ActionNode: public SeqRegExpNode { - public: - enum ActionType { - SET_REGISTER, - INCREMENT_REGISTER, - STORE_POSITION, - BEGIN_SUBMATCH, - POSITIVE_SUBMATCH_SUCCESS, - EMPTY_MATCH_CHECK, - CLEAR_CAPTURES - }; - static ActionNode* SetRegister(int reg, int val, RegExpNode* on_success); - static ActionNode* IncrementRegister(int reg, RegExpNode* on_success); - static ActionNode* StorePosition(int reg, - bool is_capture, - RegExpNode* on_success); - static ActionNode* ClearCaptures(Interval range, RegExpNode* on_success); - static ActionNode* BeginSubmatch(int stack_pointer_reg, - int position_reg, - RegExpNode* on_success); - static ActionNode* PositiveSubmatchSuccess(int stack_pointer_reg, - int restore_reg, - int clear_capture_count, - int clear_capture_from, - RegExpNode* on_success); - static ActionNode* EmptyMatchCheck(int start_register, - int repetition_register, - int repetition_limit, - RegExpNode* on_success); - void Accept(NodeVisitor* visitor) override; - void Emit(RegExpCompiler* compiler, Trace* trace) override; - int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; - void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, int filled_in, - bool not_at_start) override { - return on_success()->GetQuickCheckDetails( - details, compiler, filled_in, not_at_start); - } - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override; - ActionType action_type() { return action_type_; } - // TODO(erikcorry): We should allow some action nodes in greedy loops. - int GreedyLoopTextLength() override { - return kNodeIsTooComplexForGreedyLoops; - } - - private: - union { - struct { - int reg; - int value; - } u_store_register; - struct { - int reg; - } u_increment_register; - struct { - int reg; - bool is_capture; - } u_position_register; - struct { - int stack_pointer_register; - int current_position_register; - int clear_register_count; - int clear_register_from; - } u_submatch; - struct { - int start_register; - int repetition_register; - int repetition_limit; - } u_empty_match_check; - struct { - int range_from; - int range_to; - } u_clear_captures; - } data_; - ActionNode(ActionType action_type, RegExpNode* on_success) - : SeqRegExpNode(on_success), - action_type_(action_type) { } - ActionType action_type_; - friend class DotPrinter; -}; - - -class TextNode: public SeqRegExpNode { - public: - TextNode(ZoneList* elms, bool read_backward, - RegExpNode* on_success) - : SeqRegExpNode(on_success), elms_(elms), read_backward_(read_backward) {} - TextNode(RegExpCharacterClass* that, bool read_backward, - RegExpNode* on_success) - : SeqRegExpNode(on_success), - elms_(new (zone()) ZoneList(1, zone())), - read_backward_(read_backward) { - elms_->Add(TextElement::CharClass(that), zone()); - } - // Create TextNode for a single character class for the given ranges. - static TextNode* CreateForCharacterRanges(Zone* zone, - ZoneList* ranges, - bool read_backward, - RegExpNode* on_success, - JSRegExp::Flags flags); - // Create TextNode for a surrogate pair with a range given for the - // lead and the trail surrogate each. - static TextNode* CreateForSurrogatePair(Zone* zone, CharacterRange lead, - CharacterRange trail, - bool read_backward, - RegExpNode* on_success, - JSRegExp::Flags flags); - void Accept(NodeVisitor* visitor) override; - void Emit(RegExpCompiler* compiler, Trace* trace) override; - int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; - void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, int characters_filled_in, - bool not_at_start) override; - ZoneList* elements() { return elms_; } - bool read_backward() { return read_backward_; } - void MakeCaseIndependent(Isolate* isolate, bool is_one_byte); - int GreedyLoopTextLength() override; - RegExpNode* GetSuccessorOfOmnivorousTextNode( - RegExpCompiler* compiler) override; - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override; - void CalculateOffsets(); - RegExpNode* FilterOneByte(int depth) override; - - private: - enum TextEmitPassType { - NON_LATIN1_MATCH, // Check for characters that can't match. - SIMPLE_CHARACTER_MATCH, // Case-dependent single character check. - NON_LETTER_CHARACTER_MATCH, // Check characters that have no case equivs. - CASE_CHARACTER_MATCH, // Case-independent single character check. - CHARACTER_CLASS_MATCH // Character class. - }; - static bool SkipPass(TextEmitPassType pass, bool ignore_case); - static const int kFirstRealPass = SIMPLE_CHARACTER_MATCH; - static const int kLastPass = CHARACTER_CLASS_MATCH; - void TextEmitPass(RegExpCompiler* compiler, - TextEmitPassType pass, - bool preloaded, - Trace* trace, - bool first_element_checked, - int* checked_up_to); - int Length(); - ZoneList* elms_; - bool read_backward_; -}; - - -class AssertionNode: public SeqRegExpNode { - public: - enum AssertionType { - AT_END, - AT_START, - AT_BOUNDARY, - AT_NON_BOUNDARY, - AFTER_NEWLINE - }; - static AssertionNode* AtEnd(RegExpNode* on_success) { - return new(on_success->zone()) AssertionNode(AT_END, on_success); - } - static AssertionNode* AtStart(RegExpNode* on_success) { - return new(on_success->zone()) AssertionNode(AT_START, on_success); - } - static AssertionNode* AtBoundary(RegExpNode* on_success) { - return new(on_success->zone()) AssertionNode(AT_BOUNDARY, on_success); - } - static AssertionNode* AtNonBoundary(RegExpNode* on_success) { - return new(on_success->zone()) AssertionNode(AT_NON_BOUNDARY, on_success); - } - static AssertionNode* AfterNewline(RegExpNode* on_success) { - return new(on_success->zone()) AssertionNode(AFTER_NEWLINE, on_success); - } - void Accept(NodeVisitor* visitor) override; - void Emit(RegExpCompiler* compiler, Trace* trace) override; - int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; - void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, int filled_in, - bool not_at_start) override; - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override; - AssertionType assertion_type() { return assertion_type_; } - - private: - void EmitBoundaryCheck(RegExpCompiler* compiler, Trace* trace); - enum IfPrevious { kIsNonWord, kIsWord }; - void BacktrackIfPrevious(RegExpCompiler* compiler, - Trace* trace, - IfPrevious backtrack_if_previous); - AssertionNode(AssertionType t, RegExpNode* on_success) - : SeqRegExpNode(on_success), assertion_type_(t) { } - AssertionType assertion_type_; -}; - - -class BackReferenceNode: public SeqRegExpNode { - public: - BackReferenceNode(int start_reg, int end_reg, JSRegExp::Flags flags, - bool read_backward, RegExpNode* on_success) - : SeqRegExpNode(on_success), - start_reg_(start_reg), - end_reg_(end_reg), - flags_(flags), - read_backward_(read_backward) {} - void Accept(NodeVisitor* visitor) override; - int start_register() { return start_reg_; } - int end_register() { return end_reg_; } - bool read_backward() { return read_backward_; } - void Emit(RegExpCompiler* compiler, Trace* trace) override; - int EatsAtLeast(int still_to_find, int recursion_depth, - bool not_at_start) override; - void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, int characters_filled_in, - bool not_at_start) override { - return; - } - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override; - - private: - int start_reg_; - int end_reg_; - JSRegExp::Flags flags_; - bool read_backward_; -}; - - -class EndNode: public RegExpNode { - public: - enum Action { ACCEPT, BACKTRACK, NEGATIVE_SUBMATCH_SUCCESS }; - EndNode(Action action, Zone* zone) : RegExpNode(zone), action_(action) {} - void Accept(NodeVisitor* visitor) override; - void Emit(RegExpCompiler* compiler, Trace* trace) override; - int EatsAtLeast(int still_to_find, int recursion_depth, - bool not_at_start) override { - return 0; - } - void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, int characters_filled_in, - bool not_at_start) override { - // Returning 0 from EatsAtLeast should ensure we never get here. - UNREACHABLE(); - } - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override { - // Returning 0 from EatsAtLeast should ensure we never get here. - UNREACHABLE(); - } - - private: - Action action_; -}; - - -class NegativeSubmatchSuccess: public EndNode { - public: - NegativeSubmatchSuccess(int stack_pointer_reg, - int position_reg, - int clear_capture_count, - int clear_capture_start, - Zone* zone) - : EndNode(NEGATIVE_SUBMATCH_SUCCESS, zone), - stack_pointer_register_(stack_pointer_reg), - current_position_register_(position_reg), - clear_capture_count_(clear_capture_count), - clear_capture_start_(clear_capture_start) { } - void Emit(RegExpCompiler* compiler, Trace* trace) override; - - private: - int stack_pointer_register_; - int current_position_register_; - int clear_capture_count_; - int clear_capture_start_; -}; - - -class Guard: public ZoneObject { - public: - enum Relation { LT, GEQ }; - Guard(int reg, Relation op, int value) - : reg_(reg), - op_(op), - value_(value) { } - int reg() { return reg_; } - Relation op() { return op_; } - int value() { return value_; } - - private: - int reg_; - Relation op_; - int value_; -}; - - -class GuardedAlternative { - public: - explicit GuardedAlternative(RegExpNode* node) - : node_(node), guards_(nullptr) {} - void AddGuard(Guard* guard, Zone* zone); - RegExpNode* node() { return node_; } - void set_node(RegExpNode* node) { node_ = node; } - ZoneList* guards() { return guards_; } - - private: - RegExpNode* node_; - ZoneList* guards_; -}; - - -class AlternativeGeneration; - - -class ChoiceNode: public RegExpNode { - public: - explicit ChoiceNode(int expected_size, Zone* zone) - : RegExpNode(zone), - alternatives_(new (zone) - ZoneList(expected_size, zone)), - table_(nullptr), - not_at_start_(false), - being_calculated_(false) {} - void Accept(NodeVisitor* visitor) override; - void AddAlternative(GuardedAlternative node) { - alternatives()->Add(node, zone()); - } - ZoneList* alternatives() { return alternatives_; } - DispatchTable* GetTable(bool ignore_case); - void Emit(RegExpCompiler* compiler, Trace* trace) override; - int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; - int EatsAtLeastHelper(int still_to_find, - int budget, - RegExpNode* ignore_this_node, - bool not_at_start); - void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, int characters_filled_in, - bool not_at_start) override; - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override; - - bool being_calculated() { return being_calculated_; } - bool not_at_start() { return not_at_start_; } - void set_not_at_start() { not_at_start_ = true; } - void set_being_calculated(bool b) { being_calculated_ = b; } - virtual bool try_to_emit_quick_check_for_alternative(bool is_first) { - return true; - } - RegExpNode* FilterOneByte(int depth) override; - virtual bool read_backward() { return false; } - - protected: - int GreedyLoopTextLengthForAlternative(GuardedAlternative* alternative); - ZoneList* alternatives_; - - private: - friend class DispatchTableConstructor; - friend class Analysis; - void GenerateGuard(RegExpMacroAssembler* macro_assembler, - Guard* guard, - Trace* trace); - int CalculatePreloadCharacters(RegExpCompiler* compiler, int eats_at_least); - void EmitOutOfLineContinuation(RegExpCompiler* compiler, - Trace* trace, - GuardedAlternative alternative, - AlternativeGeneration* alt_gen, - int preload_characters, - bool next_expects_preload); - void SetUpPreLoad(RegExpCompiler* compiler, - Trace* current_trace, - PreloadState* preloads); - void AssertGuardsMentionRegisters(Trace* trace); - int EmitOptimizedUnanchoredSearch(RegExpCompiler* compiler, Trace* trace); - Trace* EmitGreedyLoop(RegExpCompiler* compiler, - Trace* trace, - AlternativeGenerationList* alt_gens, - PreloadState* preloads, - GreedyLoopState* greedy_loop_state, - int text_length); - void EmitChoices(RegExpCompiler* compiler, - AlternativeGenerationList* alt_gens, - int first_choice, - Trace* trace, - PreloadState* preloads); - DispatchTable* table_; - // If true, this node is never checked at the start of the input. - // Allows a new trace to start with at_start() set to false. - bool not_at_start_; - bool being_calculated_; -}; - - -class NegativeLookaroundChoiceNode : public ChoiceNode { - public: - explicit NegativeLookaroundChoiceNode(GuardedAlternative this_must_fail, - GuardedAlternative then_do_this, - Zone* zone) - : ChoiceNode(2, zone) { - AddAlternative(this_must_fail); - AddAlternative(then_do_this); - } - int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; - void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, int characters_filled_in, - bool not_at_start) override; - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override { - alternatives_->at(1).node()->FillInBMInfo(isolate, offset, budget - 1, bm, - not_at_start); - if (offset == 0) set_bm_info(not_at_start, bm); - } - // For a negative lookahead we don't emit the quick check for the - // alternative that is expected to fail. This is because quick check code - // starts by loading enough characters for the alternative that takes fewest - // characters, but on a negative lookahead the negative branch did not take - // part in that calculation (EatsAtLeast) so the assumptions don't hold. - bool try_to_emit_quick_check_for_alternative(bool is_first) override { - return !is_first; - } - RegExpNode* FilterOneByte(int depth) override; -}; - - -class LoopChoiceNode: public ChoiceNode { - public: - LoopChoiceNode(bool body_can_be_zero_length, bool read_backward, Zone* zone) - : ChoiceNode(2, zone), - loop_node_(nullptr), - continue_node_(nullptr), - body_can_be_zero_length_(body_can_be_zero_length), - read_backward_(read_backward) {} - void AddLoopAlternative(GuardedAlternative alt); - void AddContinueAlternative(GuardedAlternative alt); - void Emit(RegExpCompiler* compiler, Trace* trace) override; - int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; - void GetQuickCheckDetails(QuickCheckDetails* details, - RegExpCompiler* compiler, int characters_filled_in, - bool not_at_start) override; - void FillInBMInfo(Isolate* isolate, int offset, int budget, - BoyerMooreLookahead* bm, bool not_at_start) override; - RegExpNode* loop_node() { return loop_node_; } - RegExpNode* continue_node() { return continue_node_; } - bool body_can_be_zero_length() { return body_can_be_zero_length_; } - bool read_backward() override { return read_backward_; } - void Accept(NodeVisitor* visitor) override; - RegExpNode* FilterOneByte(int depth) override; - - private: - // AddAlternative is made private for loop nodes because alternatives - // should not be added freely, we need to keep track of which node - // goes back to the node itself. - void AddAlternative(GuardedAlternative node) { - ChoiceNode::AddAlternative(node); - } - - RegExpNode* loop_node_; - RegExpNode* continue_node_; - bool body_can_be_zero_length_; - bool read_backward_; -}; - - -// Improve the speed that we scan for an initial point where a non-anchored -// regexp can match by using a Boyer-Moore-like table. This is done by -// identifying non-greedy non-capturing loops in the nodes that eat any -// character one at a time. For example in the middle of the regexp -// /foo[\s\S]*?bar/ we find such a loop. There is also such a loop implicitly -// inserted at the start of any non-anchored regexp. -// -// When we have found such a loop we look ahead in the nodes to find the set of -// characters that can come at given distances. For example for the regexp -// /.?foo/ we know that there are at least 3 characters ahead of us, and the -// sets of characters that can occur are [any, [f, o], [o]]. We find a range in -// the lookahead info where the set of characters is reasonably constrained. In -// our example this is from index 1 to 2 (0 is not constrained). We can now -// look 3 characters ahead and if we don't find one of [f, o] (the union of -// [f, o] and [o]) then we can skip forwards by the range size (in this case 2). -// -// For Unicode input strings we do the same, but modulo 128. -// -// We also look at the first string fed to the regexp and use that to get a hint -// of the character frequencies in the inputs. This affects the assessment of -// whether the set of characters is 'reasonably constrained'. -// -// We also have another lookahead mechanism (called quick check in the code), -// which uses a wide load of multiple characters followed by a mask and compare -// to determine whether a match is possible at this point. -enum ContainedInLattice { - kNotYet = 0, - kLatticeIn = 1, - kLatticeOut = 2, - kLatticeUnknown = 3 // Can also mean both in and out. -}; - - -inline ContainedInLattice Combine(ContainedInLattice a, ContainedInLattice b) { - return static_cast(a | b); -} - - -ContainedInLattice AddRange(ContainedInLattice a, - const int* ranges, - int ranges_size, - Interval new_range); - - -class BoyerMoorePositionInfo : public ZoneObject { - public: - explicit BoyerMoorePositionInfo(Zone* zone) - : map_(new(zone) ZoneList(kMapSize, zone)), - map_count_(0), - w_(kNotYet), - s_(kNotYet), - d_(kNotYet), - surrogate_(kNotYet) { - for (int i = 0; i < kMapSize; i++) { - map_->Add(false, zone); - } - } - - bool& at(int i) { return map_->at(i); } - - static const int kMapSize = 128; - static const int kMask = kMapSize - 1; - - int map_count() const { return map_count_; } - - void Set(int character); - void SetInterval(const Interval& interval); - void SetAll(); - bool is_non_word() { return w_ == kLatticeOut; } - bool is_word() { return w_ == kLatticeIn; } - - private: - ZoneList* map_; - int map_count_; // Number of set bits in the map. - ContainedInLattice w_; // The \w character class. - ContainedInLattice s_; // The \s character class. - ContainedInLattice d_; // The \d character class. - ContainedInLattice surrogate_; // Surrogate UTF-16 code units. -}; - - -class BoyerMooreLookahead : public ZoneObject { - public: - BoyerMooreLookahead(int length, RegExpCompiler* compiler, Zone* zone); - - int length() { return length_; } - int max_char() { return max_char_; } - RegExpCompiler* compiler() { return compiler_; } - - int Count(int map_number) { - return bitmaps_->at(map_number)->map_count(); - } - - BoyerMoorePositionInfo* at(int i) { return bitmaps_->at(i); } - - void Set(int map_number, int character) { - if (character > max_char_) return; - BoyerMoorePositionInfo* info = bitmaps_->at(map_number); - info->Set(character); - } - - void SetInterval(int map_number, const Interval& interval) { - if (interval.from() > max_char_) return; - BoyerMoorePositionInfo* info = bitmaps_->at(map_number); - if (interval.to() > max_char_) { - info->SetInterval(Interval(interval.from(), max_char_)); - } else { - info->SetInterval(interval); - } - } - - void SetAll(int map_number) { - bitmaps_->at(map_number)->SetAll(); - } - - void SetRest(int from_map) { - for (int i = from_map; i < length_; i++) SetAll(i); - } - void EmitSkipInstructions(RegExpMacroAssembler* masm); - - private: - // This is the value obtained by EatsAtLeast. If we do not have at least this - // many characters left in the sample string then the match is bound to fail. - // Therefore it is OK to read a character this far ahead of the current match - // point. - int length_; - RegExpCompiler* compiler_; - // 0xff for Latin1, 0xffff for UTF-16. - int max_char_; - ZoneList* bitmaps_; - - int GetSkipTable(int min_lookahead, - int max_lookahead, - Handle boolean_skip_table); - bool FindWorthwhileInterval(int* from, int* to); - int FindBestInterval( - int max_number_of_chars, int old_biggest_points, int* from, int* to); -}; - - -// There are many ways to generate code for a node. This class encapsulates -// the current way we should be generating. In other words it encapsulates -// the current state of the code generator. The effect of this is that we -// generate code for paths that the matcher can take through the regular -// expression. A given node in the regexp can be code-generated several times -// as it can be part of several traces. For example for the regexp: -// /foo(bar|ip)baz/ the code to match baz will be generated twice, once as part -// of the foo-bar-baz trace and once as part of the foo-ip-baz trace. The code -// to match foo is generated only once (the traces have a common prefix). The -// code to store the capture is deferred and generated (twice) after the places -// where baz has been matched. -class Trace { - public: - // A value for a property that is either known to be true, know to be false, - // or not known. - enum TriBool { - UNKNOWN = -1, FALSE_VALUE = 0, TRUE_VALUE = 1 - }; - - class DeferredAction { - public: - DeferredAction(ActionNode::ActionType action_type, int reg) - : action_type_(action_type), reg_(reg), next_(nullptr) {} - DeferredAction* next() { return next_; } - bool Mentions(int reg); - int reg() { return reg_; } - ActionNode::ActionType action_type() { return action_type_; } - private: - ActionNode::ActionType action_type_; - int reg_; - DeferredAction* next_; - friend class Trace; - }; - - class DeferredCapture : public DeferredAction { - public: - DeferredCapture(int reg, bool is_capture, Trace* trace) - : DeferredAction(ActionNode::STORE_POSITION, reg), - cp_offset_(trace->cp_offset()), - is_capture_(is_capture) { } - int cp_offset() { return cp_offset_; } - bool is_capture() { return is_capture_; } - private: - int cp_offset_; - bool is_capture_; - void set_cp_offset(int cp_offset) { cp_offset_ = cp_offset; } - }; - - class DeferredSetRegister : public DeferredAction { - public: - DeferredSetRegister(int reg, int value) - : DeferredAction(ActionNode::SET_REGISTER, reg), - value_(value) { } - int value() { return value_; } - private: - int value_; - }; - - class DeferredClearCaptures : public DeferredAction { - public: - explicit DeferredClearCaptures(Interval range) - : DeferredAction(ActionNode::CLEAR_CAPTURES, -1), - range_(range) { } - Interval range() { return range_; } - private: - Interval range_; - }; - - class DeferredIncrementRegister : public DeferredAction { - public: - explicit DeferredIncrementRegister(int reg) - : DeferredAction(ActionNode::INCREMENT_REGISTER, reg) { } - }; - - Trace() - : cp_offset_(0), - actions_(nullptr), - backtrack_(nullptr), - stop_node_(nullptr), - loop_label_(nullptr), - characters_preloaded_(0), - bound_checked_up_to_(0), - flush_budget_(100), - at_start_(UNKNOWN) {} - - // End the trace. This involves flushing the deferred actions in the trace - // and pushing a backtrack location onto the backtrack stack. Once this is - // done we can start a new trace or go to one that has already been - // generated. - void Flush(RegExpCompiler* compiler, RegExpNode* successor); - int cp_offset() { return cp_offset_; } - DeferredAction* actions() { return actions_; } - // A trivial trace is one that has no deferred actions or other state that - // affects the assumptions used when generating code. There is no recorded - // backtrack location in a trivial trace, so with a trivial trace we will - // generate code that, on a failure to match, gets the backtrack location - // from the backtrack stack rather than using a direct jump instruction. We - // always start code generation with a trivial trace and non-trivial traces - // are created as we emit code for nodes or add to the list of deferred - // actions in the trace. The location of the code generated for a node using - // a trivial trace is recorded in a label in the node so that gotos can be - // generated to that code. - bool is_trivial() { - return backtrack_ == nullptr && actions_ == nullptr && cp_offset_ == 0 && - characters_preloaded_ == 0 && bound_checked_up_to_ == 0 && - quick_check_performed_.characters() == 0 && at_start_ == UNKNOWN; - } - TriBool at_start() { return at_start_; } - void set_at_start(TriBool at_start) { at_start_ = at_start; } - Label* backtrack() { return backtrack_; } - Label* loop_label() { return loop_label_; } - RegExpNode* stop_node() { return stop_node_; } - int characters_preloaded() { return characters_preloaded_; } - int bound_checked_up_to() { return bound_checked_up_to_; } - int flush_budget() { return flush_budget_; } - QuickCheckDetails* quick_check_performed() { return &quick_check_performed_; } - bool mentions_reg(int reg); - // Returns true if a deferred position store exists to the specified - // register and stores the offset in the out-parameter. Otherwise - // returns false. - bool GetStoredPosition(int reg, int* cp_offset); - // These set methods and AdvanceCurrentPositionInTrace should be used only on - // new traces - the intention is that traces are immutable after creation. - void add_action(DeferredAction* new_action) { - DCHECK(new_action->next_ == nullptr); - new_action->next_ = actions_; - actions_ = new_action; - } - void set_backtrack(Label* backtrack) { backtrack_ = backtrack; } - void set_stop_node(RegExpNode* node) { stop_node_ = node; } - void set_loop_label(Label* label) { loop_label_ = label; } - void set_characters_preloaded(int count) { characters_preloaded_ = count; } - void set_bound_checked_up_to(int to) { bound_checked_up_to_ = to; } - void set_flush_budget(int to) { flush_budget_ = to; } - void set_quick_check_performed(QuickCheckDetails* d) { - quick_check_performed_ = *d; - } - void InvalidateCurrentCharacter(); - void AdvanceCurrentPositionInTrace(int by, RegExpCompiler* compiler); - - private: - int FindAffectedRegisters(OutSet* affected_registers, Zone* zone); - void PerformDeferredActions(RegExpMacroAssembler* macro, - int max_register, - const OutSet& affected_registers, - OutSet* registers_to_pop, - OutSet* registers_to_clear, - Zone* zone); - void RestoreAffectedRegisters(RegExpMacroAssembler* macro, - int max_register, - const OutSet& registers_to_pop, - const OutSet& registers_to_clear); - int cp_offset_; - DeferredAction* actions_; - Label* backtrack_; - RegExpNode* stop_node_; - Label* loop_label_; - int characters_preloaded_; - int bound_checked_up_to_; - QuickCheckDetails quick_check_performed_; - int flush_budget_; - TriBool at_start_; -}; - - -class GreedyLoopState { - public: - explicit GreedyLoopState(bool not_at_start); - - Label* label() { return &label_; } - Trace* counter_backtrack_trace() { return &counter_backtrack_trace_; } - - private: - Label label_; - Trace counter_backtrack_trace_; -}; - - -struct PreloadState { - static const int kEatsAtLeastNotYetInitialized = -1; - bool preload_is_current_; - bool preload_has_checked_bounds_; - int preload_characters_; - int eats_at_least_; - void init() { - eats_at_least_ = kEatsAtLeastNotYetInitialized; - } -}; - - -class NodeVisitor { - public: - virtual ~NodeVisitor() = default; -#define DECLARE_VISIT(Type) \ - virtual void Visit##Type(Type##Node* that) = 0; -FOR_EACH_NODE_TYPE(DECLARE_VISIT) -#undef DECLARE_VISIT - virtual void VisitLoopChoice(LoopChoiceNode* that) { VisitChoice(that); } -}; - - -// Node visitor used to add the start set of the alternatives to the -// dispatch table of a choice node. -class V8_EXPORT_PRIVATE DispatchTableConstructor : public NodeVisitor { - public: - DispatchTableConstructor(DispatchTable* table, bool ignore_case, - Zone* zone) - : table_(table), - choice_index_(-1), - ignore_case_(ignore_case), - zone_(zone) { } - - void BuildTable(ChoiceNode* node); - - void AddRange(CharacterRange range) { - table()->AddRange(range, choice_index_, zone_); - } - - void AddInverse(ZoneList* ranges); - -#define DECLARE_VISIT(Type) \ - virtual void Visit##Type(Type##Node* that); -FOR_EACH_NODE_TYPE(DECLARE_VISIT) -#undef DECLARE_VISIT - - DispatchTable* table() { return table_; } - void set_choice_index(int value) { choice_index_ = value; } - - protected: - DispatchTable* table_; - int choice_index_; - bool ignore_case_; - Zone* zone_; -}; - -// Assertion propagation moves information about assertions such as -// \b to the affected nodes. For instance, in /.\b./ information must -// be propagated to the first '.' that whatever follows needs to know -// if it matched a word or a non-word, and to the second '.' that it -// has to check if it succeeds a word or non-word. In this case the -// result will be something like: -// -// +-------+ +------------+ -// | . | | . | -// +-------+ ---> +------------+ -// | word? | | check word | -// +-------+ +------------+ -class Analysis: public NodeVisitor { - public: - Analysis(Isolate* isolate, bool is_one_byte) - : isolate_(isolate), is_one_byte_(is_one_byte), error_message_(nullptr) {} - void EnsureAnalyzed(RegExpNode* node); - -#define DECLARE_VISIT(Type) void Visit##Type(Type##Node* that) override; - FOR_EACH_NODE_TYPE(DECLARE_VISIT) -#undef DECLARE_VISIT - void VisitLoopChoice(LoopChoiceNode* that) override; - - bool has_failed() { return error_message_ != nullptr; } - const char* error_message() { - DCHECK(error_message_ != nullptr); - return error_message_; - } - void fail(const char* error_message) { - error_message_ = error_message; - } - - Isolate* isolate() const { return isolate_; } - - private: - Isolate* isolate_; - bool is_one_byte_; - const char* error_message_; - - DISALLOW_IMPLICIT_CONSTRUCTORS(Analysis); -}; - - struct RegExpCompileData { RegExpCompileData() : tree(nullptr), @@ -1465,13 +197,18 @@ struct RegExpCompileData { int capture_count; }; - -class RegExpEngine: public AllStatic { +class RegExpEngine final : public AllStatic { public: struct CompilationResult { - inline CompilationResult(Isolate* isolate, const char* error_message); - CompilationResult(Object code, int registers) - : code(code), num_registers(registers) {} + explicit CompilationResult(const char* error_message) + : error_message(error_message) {} + CompilationResult(const char* error_message, Object code, int registers) + : error_message(error_message), code(code), num_registers(registers) {} + + static CompilationResult RegExpTooBig() { + return CompilationResult("RegExp too big"); + } + const char* const error_message = nullptr; Object const code; int const num_registers = 0; @@ -1488,8 +225,7 @@ class RegExpEngine: public AllStatic { bool ignore_case); }; - -class RegExpResultsCache : public AllStatic { +class RegExpResultsCache final : public AllStatic { public: enum ResultsCacheType { REGEXP_MULTIPLE_INDICES, STRING_SPLIT_SUBSTRINGS }; diff --git a/src/regexp/regexp-compiler.cc b/src/regexp/regexp-compiler.cc new file mode 100644 index 0000000000..54cedcb455 --- /dev/null +++ b/src/regexp/regexp-compiler.cc @@ -0,0 +1,3714 @@ +// Copyright 2019 the V8 project authors. All rights reserved. +// Use of this source code is governed by a BSD-style license that can be +// found in the LICENSE file. + +#include "src/regexp/regexp-compiler.h" + +#include "src/diagnostics/code-tracer.h" +#include "src/execution/isolate.h" +#include "src/objects/objects-inl.h" +#include "src/regexp/jsregexp.h" // TODO(jgruber): Only needed for IgnoreCase. +#include "src/regexp/regexp-macro-assembler-arch.h" +#include "src/regexp/regexp-macro-assembler-tracer.h" +#include "src/utils/ostreams.h" +#include "src/utils/splay-tree-inl.h" +#include "src/zone/zone-list-inl.h" + +#ifdef V8_INTL_SUPPORT +#include "unicode/locid.h" +#include "unicode/uniset.h" +#include "unicode/utypes.h" +#endif // V8_INTL_SUPPORT + +namespace v8 { +namespace internal { + +using namespace regexp_compiler_constants; // NOLINT(build/namespaces) + +// Explicit template instantiations. +template class ZoneSplayTree; +template void DispatchTable::ForEach( + UnicodeRangeSplitter*); + +// ------------------------------------------------------------------- +// Implementation of the Irregexp regular expression engine. +// +// The Irregexp regular expression engine is intended to be a complete +// implementation of ECMAScript regular expressions. It generates either +// bytecodes or native code. + +// The Irregexp regexp engine is structured in three steps. +// 1) The parser generates an abstract syntax tree. See ast.cc. +// 2) From the AST a node network is created. The nodes are all +// subclasses of RegExpNode. The nodes represent states when +// executing a regular expression. Several optimizations are +// performed on the node network. +// 3) From the nodes we generate either byte codes or native code +// that can actually execute the regular expression (perform +// the search). The code generation step is described in more +// detail below. + +// Code generation. +// +// The nodes are divided into four main categories. +// * Choice nodes +// These represent places where the regular expression can +// match in more than one way. For example on entry to an +// alternation (foo|bar) or a repetition (*, +, ? or {}). +// * Action nodes +// These represent places where some action should be +// performed. Examples include recording the current position +// in the input string to a register (in order to implement +// captures) or other actions on register for example in order +// to implement the counters needed for {} repetitions. +// * Matching nodes +// These attempt to match some element part of the input string. +// Examples of elements include character classes, plain strings +// or back references. +// * End nodes +// These are used to implement the actions required on finding +// a successful match or failing to find a match. +// +// The code generated (whether as byte codes or native code) maintains +// some state as it runs. This consists of the following elements: +// +// * The capture registers. Used for string captures. +// * Other registers. Used for counters etc. +// * The current position. +// * The stack of backtracking information. Used when a matching node +// fails to find a match and needs to try an alternative. +// +// Conceptual regular expression execution model: +// +// There is a simple conceptual model of regular expression execution +// which will be presented first. The actual code generated is a more +// efficient simulation of the simple conceptual model: +// +// * Choice nodes are implemented as follows: +// For each choice except the last { +// push current position +// push backtrack code location +// +// backtrack code location: +// pop current position +// } +// +// +// * Actions nodes are generated as follows +// +// +// push backtrack code location +// +// backtrack code location: +// +// +// +// * Matching nodes are generated as follows: +// if input string matches at current position +// update current position +// +// else +// +// +// Thus it can be seen that the current position is saved and restored +// by the choice nodes, whereas the registers are saved and restored by +// by the action nodes that manipulate them. +// +// The other interesting aspect of this model is that nodes are generated +// at the point where they are needed by a recursive call to Emit(). If +// the node has already been code generated then the Emit() call will +// generate a jump to the previously generated code instead. In order to +// limit recursion it is possible for the Emit() function to put the node +// on a work list for later generation and instead generate a jump. The +// destination of the jump is resolved later when the code is generated. +// +// Actual regular expression code generation. +// +// Code generation is actually more complicated than the above. In order +// to improve the efficiency of the generated code some optimizations are +// performed +// +// * Choice nodes have 1-character lookahead. +// A choice node looks at the following character and eliminates some of +// the choices immediately based on that character. This is not yet +// implemented. +// * Simple greedy loops store reduced backtracking information. +// A quantifier like /.*foo/m will greedily match the whole input. It will +// then need to backtrack to a point where it can match "foo". The naive +// implementation of this would push each character position onto the +// backtracking stack, then pop them off one by one. This would use space +// proportional to the length of the input string. However since the "." +// can only match in one way and always has a constant length (in this case +// of 1) it suffices to store the current position on the top of the stack +// once. Matching now becomes merely incrementing the current position and +// backtracking becomes decrementing the current position and checking the +// result against the stored current position. This is faster and saves +// space. +// * The current state is virtualized. +// This is used to defer expensive operations until it is clear that they +// are needed and to generate code for a node more than once, allowing +// specialized an efficient versions of the code to be created. This is +// explained in the section below. +// +// Execution state virtualization. +// +// Instead of emitting code, nodes that manipulate the state can record their +// manipulation in an object called the Trace. The Trace object can record a +// current position offset, an optional backtrack code location on the top of +// the virtualized backtrack stack and some register changes. When a node is +// to be emitted it can flush the Trace or update it. Flushing the Trace +// will emit code to bring the actual state into line with the virtual state. +// Avoiding flushing the state can postpone some work (e.g. updates of capture +// registers). Postponing work can save time when executing the regular +// expression since it may be found that the work never has to be done as a +// failure to match can occur. In addition it is much faster to jump to a +// known backtrack code location than it is to pop an unknown backtrack +// location from the stack and jump there. +// +// The virtual state found in the Trace affects code generation. For example +// the virtual state contains the difference between the actual current +// position and the virtual current position, and matching code needs to use +// this offset to attempt a match in the correct location of the input +// string. Therefore code generated for a non-trivial trace is specialized +// to that trace. The code generator therefore has the ability to generate +// code for each node several times. In order to limit the size of the +// generated code there is an arbitrary limit on how many specialized sets of +// code may be generated for a given node. If the limit is reached, the +// trace is flushed and a generic version of the code for a node is emitted. +// This is subsequently used for that node. The code emitted for non-generic +// trace is not recorded in the node and so it cannot currently be reused in +// the event that code generation is requested for an identical trace. + +void RegExpTree::AppendToText(RegExpText* text, Zone* zone) { UNREACHABLE(); } + +void RegExpAtom::AppendToText(RegExpText* text, Zone* zone) { + text->AddElement(TextElement::Atom(this), zone); +} + +void RegExpCharacterClass::AppendToText(RegExpText* text, Zone* zone) { + text->AddElement(TextElement::CharClass(this), zone); +} + +void RegExpText::AppendToText(RegExpText* text, Zone* zone) { + for (int i = 0; i < elements()->length(); i++) + text->AddElement(elements()->at(i), zone); +} + +TextElement TextElement::Atom(RegExpAtom* atom) { + return TextElement(ATOM, atom); +} + +TextElement TextElement::CharClass(RegExpCharacterClass* char_class) { + return TextElement(CHAR_CLASS, char_class); +} + +int TextElement::length() const { + switch (text_type()) { + case ATOM: + return atom()->length(); + + case CHAR_CLASS: + return 1; + } + UNREACHABLE(); +} + +DispatchTable* ChoiceNode::GetTable(bool ignore_case) { + if (table_ == nullptr) { + table_ = new (zone()) DispatchTable(zone()); + DispatchTableConstructor cons(table_, ignore_case, zone()); + cons.BuildTable(this); + } + return table_; +} + +class RecursionCheck { + public: + explicit RecursionCheck(RegExpCompiler* compiler) : compiler_(compiler) { + compiler->IncrementRecursionDepth(); + } + ~RecursionCheck() { compiler_->DecrementRecursionDepth(); } + + private: + RegExpCompiler* compiler_; +}; + +// Attempts to compile the regexp using an Irregexp code generator. Returns +// a fixed array or a null handle depending on whether it succeeded. +RegExpCompiler::RegExpCompiler(Isolate* isolate, Zone* zone, int capture_count, + bool one_byte) + : next_register_(2 * (capture_count + 1)), + unicode_lookaround_stack_register_(kNoRegister), + unicode_lookaround_position_register_(kNoRegister), + work_list_(nullptr), + recursion_depth_(0), + one_byte_(one_byte), + reg_exp_too_big_(false), + limiting_recursion_(false), + optimize_(FLAG_regexp_optimization), + read_backward_(false), + current_expansion_factor_(1), + frequency_collator_(), + isolate_(isolate), + zone_(zone) { + accept_ = new (zone) EndNode(EndNode::ACCEPT, zone); + DCHECK_GE(RegExpMacroAssembler::kMaxRegister, next_register_ - 1); +} + +RegExpCompiler::CompilationResult RegExpCompiler::Assemble( + Isolate* isolate, RegExpMacroAssembler* macro_assembler, RegExpNode* start, + int capture_count, Handle pattern) { +#ifdef DEBUG + if (FLAG_trace_regexp_assembler) + macro_assembler_ = new RegExpMacroAssemblerTracer(isolate, macro_assembler); + else +#endif + macro_assembler_ = macro_assembler; + + std::vector work_list; + work_list_ = &work_list; + Label fail; + macro_assembler_->PushBacktrack(&fail); + Trace new_trace; + start->Emit(this, &new_trace); + macro_assembler_->Bind(&fail); + macro_assembler_->Fail(); + while (!work_list.empty()) { + RegExpNode* node = work_list.back(); + work_list.pop_back(); + node->set_on_work_list(false); + if (!node->label()->is_bound()) node->Emit(this, &new_trace); + } + if (reg_exp_too_big_) { + macro_assembler_->AbortedCodeGeneration(); + return CompilationResult::RegExpTooBig(); + } + + Handle code = macro_assembler_->GetCode(pattern); + isolate->IncreaseTotalRegexpCodeGenerated(code->Size()); + work_list_ = nullptr; +#ifdef ENABLE_DISASSEMBLER + if (FLAG_print_code && !FLAG_regexp_interpret_all) { + CodeTracer::Scope trace_scope(isolate->GetCodeTracer()); + OFStream os(trace_scope.file()); + Handle::cast(code)->Disassemble(pattern->ToCString().get(), os); + } +#endif +#ifdef DEBUG + if (FLAG_trace_regexp_assembler) { + delete macro_assembler_; + } +#endif + return {*code, next_register_}; +} + +bool Trace::DeferredAction::Mentions(int that) { + if (action_type() == ActionNode::CLEAR_CAPTURES) { + Interval range = static_cast(this)->range(); + return range.Contains(that); + } else { + return reg() == that; + } +} + +bool Trace::mentions_reg(int reg) { + for (DeferredAction* action = actions_; action != nullptr; + action = action->next()) { + if (action->Mentions(reg)) return true; + } + return false; +} + +bool Trace::GetStoredPosition(int reg, int* cp_offset) { + DCHECK_EQ(0, *cp_offset); + for (DeferredAction* action = actions_; action != nullptr; + action = action->next()) { + if (action->Mentions(reg)) { + if (action->action_type() == ActionNode::STORE_POSITION) { + *cp_offset = static_cast(action)->cp_offset(); + return true; + } else { + return false; + } + } + } + return false; +} + +int Trace::FindAffectedRegisters(OutSet* affected_registers, Zone* zone) { + int max_register = RegExpCompiler::kNoRegister; + for (DeferredAction* action = actions_; action != nullptr; + action = action->next()) { + if (action->action_type() == ActionNode::CLEAR_CAPTURES) { + Interval range = static_cast(action)->range(); + for (int i = range.from(); i <= range.to(); i++) + affected_registers->Set(i, zone); + if (range.to() > max_register) max_register = range.to(); + } else { + affected_registers->Set(action->reg(), zone); + if (action->reg() > max_register) max_register = action->reg(); + } + } + return max_register; +} + +void Trace::RestoreAffectedRegisters(RegExpMacroAssembler* assembler, + int max_register, + const OutSet& registers_to_pop, + const OutSet& registers_to_clear) { + for (int reg = max_register; reg >= 0; reg--) { + if (registers_to_pop.Get(reg)) { + assembler->PopRegister(reg); + } else if (registers_to_clear.Get(reg)) { + int clear_to = reg; + while (reg > 0 && registers_to_clear.Get(reg - 1)) { + reg--; + } + assembler->ClearRegisters(reg, clear_to); + } + } +} + +void Trace::PerformDeferredActions(RegExpMacroAssembler* assembler, + int max_register, + const OutSet& affected_registers, + OutSet* registers_to_pop, + OutSet* registers_to_clear, Zone* zone) { + // The "+1" is to avoid a push_limit of zero if stack_limit_slack() is 1. + const int push_limit = (assembler->stack_limit_slack() + 1) / 2; + + // Count pushes performed to force a stack limit check occasionally. + int pushes = 0; + + for (int reg = 0; reg <= max_register; reg++) { + if (!affected_registers.Get(reg)) { + continue; + } + + // The chronologically first deferred action in the trace + // is used to infer the action needed to restore a register + // to its previous state (or not, if it's safe to ignore it). + enum DeferredActionUndoType { IGNORE, RESTORE, CLEAR }; + DeferredActionUndoType undo_action = IGNORE; + + int value = 0; + bool absolute = false; + bool clear = false; + static const int kNoStore = kMinInt; + int store_position = kNoStore; + // This is a little tricky because we are scanning the actions in reverse + // historical order (newest first). + for (DeferredAction* action = actions_; action != nullptr; + action = action->next()) { + if (action->Mentions(reg)) { + switch (action->action_type()) { + case ActionNode::SET_REGISTER: { + Trace::DeferredSetRegister* psr = + static_cast(action); + if (!absolute) { + value += psr->value(); + absolute = true; + } + // SET_REGISTER is currently only used for newly introduced loop + // counters. They can have a significant previous value if they + // occur in a loop. TODO(lrn): Propagate this information, so + // we can set undo_action to IGNORE if we know there is no value to + // restore. + undo_action = RESTORE; + DCHECK_EQ(store_position, kNoStore); + DCHECK(!clear); + break; + } + case ActionNode::INCREMENT_REGISTER: + if (!absolute) { + value++; + } + DCHECK_EQ(store_position, kNoStore); + DCHECK(!clear); + undo_action = RESTORE; + break; + case ActionNode::STORE_POSITION: { + Trace::DeferredCapture* pc = + static_cast(action); + if (!clear && store_position == kNoStore) { + store_position = pc->cp_offset(); + } + + // For captures we know that stores and clears alternate. + // Other register, are never cleared, and if the occur + // inside a loop, they might be assigned more than once. + if (reg <= 1) { + // Registers zero and one, aka "capture zero", is + // always set correctly if we succeed. There is no + // need to undo a setting on backtrack, because we + // will set it again or fail. + undo_action = IGNORE; + } else { + undo_action = pc->is_capture() ? CLEAR : RESTORE; + } + DCHECK(!absolute); + DCHECK_EQ(value, 0); + break; + } + case ActionNode::CLEAR_CAPTURES: { + // Since we're scanning in reverse order, if we've already + // set the position we have to ignore historically earlier + // clearing operations. + if (store_position == kNoStore) { + clear = true; + } + undo_action = RESTORE; + DCHECK(!absolute); + DCHECK_EQ(value, 0); + break; + } + default: + UNREACHABLE(); + break; + } + } + } + // Prepare for the undo-action (e.g., push if it's going to be popped). + if (undo_action == RESTORE) { + pushes++; + RegExpMacroAssembler::StackCheckFlag stack_check = + RegExpMacroAssembler::kNoStackLimitCheck; + if (pushes == push_limit) { + stack_check = RegExpMacroAssembler::kCheckStackLimit; + pushes = 0; + } + + assembler->PushRegister(reg, stack_check); + registers_to_pop->Set(reg, zone); + } else if (undo_action == CLEAR) { + registers_to_clear->Set(reg, zone); + } + // Perform the chronologically last action (or accumulated increment) + // for the register. + if (store_position != kNoStore) { + assembler->WriteCurrentPositionToRegister(reg, store_position); + } else if (clear) { + assembler->ClearRegisters(reg, reg); + } else if (absolute) { + assembler->SetRegister(reg, value); + } else if (value != 0) { + assembler->AdvanceRegister(reg, value); + } + } +} + +// This is called as we come into a loop choice node and some other tricky +// nodes. It normalizes the state of the code generator to ensure we can +// generate generic code. +void Trace::Flush(RegExpCompiler* compiler, RegExpNode* successor) { + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + + DCHECK(!is_trivial()); + + if (actions_ == nullptr && backtrack() == nullptr) { + // Here we just have some deferred cp advances to fix and we are back to + // a normal situation. We may also have to forget some information gained + // through a quick check that was already performed. + if (cp_offset_ != 0) assembler->AdvanceCurrentPosition(cp_offset_); + // Create a new trivial state and generate the node with that. + Trace new_state; + successor->Emit(compiler, &new_state); + return; + } + + // Generate deferred actions here along with code to undo them again. + OutSet affected_registers; + + if (backtrack() != nullptr) { + // Here we have a concrete backtrack location. These are set up by choice + // nodes and so they indicate that we have a deferred save of the current + // position which we may need to emit here. + assembler->PushCurrentPosition(); + } + + int max_register = + FindAffectedRegisters(&affected_registers, compiler->zone()); + OutSet registers_to_pop; + OutSet registers_to_clear; + PerformDeferredActions(assembler, max_register, affected_registers, + ®isters_to_pop, ®isters_to_clear, + compiler->zone()); + if (cp_offset_ != 0) { + assembler->AdvanceCurrentPosition(cp_offset_); + } + + // Create a new trivial state and generate the node with that. + Label undo; + assembler->PushBacktrack(&undo); + if (successor->KeepRecursing(compiler)) { + Trace new_state; + successor->Emit(compiler, &new_state); + } else { + compiler->AddWork(successor); + assembler->GoTo(successor->label()); + } + + // On backtrack we need to restore state. + assembler->Bind(&undo); + RestoreAffectedRegisters(assembler, max_register, registers_to_pop, + registers_to_clear); + if (backtrack() == nullptr) { + assembler->Backtrack(); + } else { + assembler->PopCurrentPosition(); + assembler->GoTo(backtrack()); + } +} + +void NegativeSubmatchSuccess::Emit(RegExpCompiler* compiler, Trace* trace) { + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + + // Omit flushing the trace. We discard the entire stack frame anyway. + + if (!label()->is_bound()) { + // We are completely independent of the trace, since we ignore it, + // so this code can be used as the generic version. + assembler->Bind(label()); + } + + // Throw away everything on the backtrack stack since the start + // of the negative submatch and restore the character position. + assembler->ReadCurrentPositionFromRegister(current_position_register_); + assembler->ReadStackPointerFromRegister(stack_pointer_register_); + if (clear_capture_count_ > 0) { + // Clear any captures that might have been performed during the success + // of the body of the negative look-ahead. + int clear_capture_end = clear_capture_start_ + clear_capture_count_ - 1; + assembler->ClearRegisters(clear_capture_start_, clear_capture_end); + } + // Now that we have unwound the stack we find at the top of the stack the + // backtrack that the BeginSubmatch node got. + assembler->Backtrack(); +} + +void EndNode::Emit(RegExpCompiler* compiler, Trace* trace) { + if (!trace->is_trivial()) { + trace->Flush(compiler, this); + return; + } + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + if (!label()->is_bound()) { + assembler->Bind(label()); + } + switch (action_) { + case ACCEPT: + assembler->Succeed(); + return; + case BACKTRACK: + assembler->GoTo(trace->backtrack()); + return; + case NEGATIVE_SUBMATCH_SUCCESS: + // This case is handled in a different virtual method. + UNREACHABLE(); + } + UNIMPLEMENTED(); +} + +void GuardedAlternative::AddGuard(Guard* guard, Zone* zone) { + if (guards_ == nullptr) guards_ = new (zone) ZoneList(1, zone); + guards_->Add(guard, zone); +} + +ActionNode* ActionNode::SetRegister(int reg, int val, RegExpNode* on_success) { + ActionNode* result = + new (on_success->zone()) ActionNode(SET_REGISTER, on_success); + result->data_.u_store_register.reg = reg; + result->data_.u_store_register.value = val; + return result; +} + +ActionNode* ActionNode::IncrementRegister(int reg, RegExpNode* on_success) { + ActionNode* result = + new (on_success->zone()) ActionNode(INCREMENT_REGISTER, on_success); + result->data_.u_increment_register.reg = reg; + return result; +} + +ActionNode* ActionNode::StorePosition(int reg, bool is_capture, + RegExpNode* on_success) { + ActionNode* result = + new (on_success->zone()) ActionNode(STORE_POSITION, on_success); + result->data_.u_position_register.reg = reg; + result->data_.u_position_register.is_capture = is_capture; + return result; +} + +ActionNode* ActionNode::ClearCaptures(Interval range, RegExpNode* on_success) { + ActionNode* result = + new (on_success->zone()) ActionNode(CLEAR_CAPTURES, on_success); + result->data_.u_clear_captures.range_from = range.from(); + result->data_.u_clear_captures.range_to = range.to(); + return result; +} + +ActionNode* ActionNode::BeginSubmatch(int stack_reg, int position_reg, + RegExpNode* on_success) { + ActionNode* result = + new (on_success->zone()) ActionNode(BEGIN_SUBMATCH, on_success); + result->data_.u_submatch.stack_pointer_register = stack_reg; + result->data_.u_submatch.current_position_register = position_reg; + return result; +} + +ActionNode* ActionNode::PositiveSubmatchSuccess(int stack_reg, int position_reg, + int clear_register_count, + int clear_register_from, + RegExpNode* on_success) { + ActionNode* result = new (on_success->zone()) + ActionNode(POSITIVE_SUBMATCH_SUCCESS, on_success); + result->data_.u_submatch.stack_pointer_register = stack_reg; + result->data_.u_submatch.current_position_register = position_reg; + result->data_.u_submatch.clear_register_count = clear_register_count; + result->data_.u_submatch.clear_register_from = clear_register_from; + return result; +} + +ActionNode* ActionNode::EmptyMatchCheck(int start_register, + int repetition_register, + int repetition_limit, + RegExpNode* on_success) { + ActionNode* result = + new (on_success->zone()) ActionNode(EMPTY_MATCH_CHECK, on_success); + result->data_.u_empty_match_check.start_register = start_register; + result->data_.u_empty_match_check.repetition_register = repetition_register; + result->data_.u_empty_match_check.repetition_limit = repetition_limit; + return result; +} + +#define DEFINE_ACCEPT(Type) \ + void Type##Node::Accept(NodeVisitor* visitor) { visitor->Visit##Type(this); } +FOR_EACH_NODE_TYPE(DEFINE_ACCEPT) +#undef DEFINE_ACCEPT + +void LoopChoiceNode::Accept(NodeVisitor* visitor) { + visitor->VisitLoopChoice(this); +} + +// ------------------------------------------------------------------- +// Emit code. + +void ChoiceNode::GenerateGuard(RegExpMacroAssembler* macro_assembler, + Guard* guard, Trace* trace) { + switch (guard->op()) { + case Guard::LT: + DCHECK(!trace->mentions_reg(guard->reg())); + macro_assembler->IfRegisterGE(guard->reg(), guard->value(), + trace->backtrack()); + break; + case Guard::GEQ: + DCHECK(!trace->mentions_reg(guard->reg())); + macro_assembler->IfRegisterLT(guard->reg(), guard->value(), + trace->backtrack()); + break; + } +} + +// Returns the number of characters in the equivalence class, omitting those +// that cannot occur in the source string because it is Latin1. +static int GetCaseIndependentLetters(Isolate* isolate, uc16 character, + bool one_byte_subject, + unibrow::uchar* letters, + int letter_length) { +#ifdef V8_INTL_SUPPORT + icu::UnicodeSet set; + set.add(character); + set = set.closeOver(USET_CASE_INSENSITIVE); + int32_t range_count = set.getRangeCount(); + int items = 0; + for (int32_t i = 0; i < range_count; i++) { + UChar32 start = set.getRangeStart(i); + UChar32 end = set.getRangeEnd(i); + CHECK(end - start + items <= letter_length); + while (start <= end) { + if (one_byte_subject && start > String::kMaxOneByteCharCode) break; + letters[items++] = (unibrow::uchar)(start); + start++; + } + } + return items; +#else + int length = + isolate->jsregexp_uncanonicalize()->get(character, '\0', letters); + // Unibrow returns 0 or 1 for characters where case independence is + // trivial. + if (length == 0) { + letters[0] = character; + length = 1; + } + + if (one_byte_subject) { + int new_length = 0; + for (int i = 0; i < length; i++) { + if (letters[i] <= String::kMaxOneByteCharCode) { + letters[new_length++] = letters[i]; + } + } + length = new_length; + } + + return length; +#endif // V8_INTL_SUPPORT +} + +static inline bool EmitSimpleCharacter(Isolate* isolate, + RegExpCompiler* compiler, uc16 c, + Label* on_failure, int cp_offset, + bool check, bool preloaded) { + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + bool bound_checked = false; + if (!preloaded) { + assembler->LoadCurrentCharacter(cp_offset, on_failure, check); + bound_checked = true; + } + assembler->CheckNotCharacter(c, on_failure); + return bound_checked; +} + +// Only emits non-letters (things that don't have case). Only used for case +// independent matches. +static inline bool EmitAtomNonLetter(Isolate* isolate, RegExpCompiler* compiler, + uc16 c, Label* on_failure, int cp_offset, + bool check, bool preloaded) { + RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); + bool one_byte = compiler->one_byte(); + unibrow::uchar chars[4]; + int length = GetCaseIndependentLetters(isolate, c, one_byte, chars, 4); + if (length < 1) { + // This can't match. Must be an one-byte subject and a non-one-byte + // character. We do not need to do anything since the one-byte pass + // already handled this. + return false; // Bounds not checked. + } + bool checked = false; + // We handle the length > 1 case in a later pass. + if (length == 1) { + if (one_byte && c > String::kMaxOneByteCharCodeU) { + // Can't match - see above. + return false; // Bounds not checked. + } + if (!preloaded) { + macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check); + checked = check; + } + macro_assembler->CheckNotCharacter(c, on_failure); + } + return checked; +} + +static bool ShortCutEmitCharacterPair(RegExpMacroAssembler* macro_assembler, + bool one_byte, uc16 c1, uc16 c2, + Label* on_failure) { + uc16 char_mask; + if (one_byte) { + char_mask = String::kMaxOneByteCharCode; + } else { + char_mask = String::kMaxUtf16CodeUnit; + } + uc16 exor = c1 ^ c2; + // Check whether exor has only one bit set. + if (((exor - 1) & exor) == 0) { + // If c1 and c2 differ only by one bit. + // Ecma262UnCanonicalize always gives the highest number last. + DCHECK(c2 > c1); + uc16 mask = char_mask ^ exor; + macro_assembler->CheckNotCharacterAfterAnd(c1, mask, on_failure); + return true; + } + DCHECK(c2 > c1); + uc16 diff = c2 - c1; + if (((diff - 1) & diff) == 0 && c1 >= diff) { + // If the characters differ by 2^n but don't differ by one bit then + // subtract the difference from the found character, then do the or + // trick. We avoid the theoretical case where negative numbers are + // involved in order to simplify code generation. + uc16 mask = char_mask ^ diff; + macro_assembler->CheckNotCharacterAfterMinusAnd(c1 - diff, diff, mask, + on_failure); + return true; + } + return false; +} + +using EmitCharacterFunction = bool(Isolate* isolate, RegExpCompiler* compiler, + uc16 c, Label* on_failure, int cp_offset, + bool check, bool preloaded); + +// Only emits letters (things that have case). Only used for case independent +// matches. +static inline bool EmitAtomLetter(Isolate* isolate, RegExpCompiler* compiler, + uc16 c, Label* on_failure, int cp_offset, + bool check, bool preloaded) { + RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); + bool one_byte = compiler->one_byte(); + unibrow::uchar chars[4]; + int length = GetCaseIndependentLetters(isolate, c, one_byte, chars, 4); + if (length <= 1) return false; + // We may not need to check against the end of the input string + // if this character lies before a character that matched. + if (!preloaded) { + macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check); + } + Label ok; + switch (length) { + case 2: { + if (ShortCutEmitCharacterPair(macro_assembler, one_byte, chars[0], + chars[1], on_failure)) { + } else { + macro_assembler->CheckCharacter(chars[0], &ok); + macro_assembler->CheckNotCharacter(chars[1], on_failure); + macro_assembler->Bind(&ok); + } + break; + } + case 4: + macro_assembler->CheckCharacter(chars[3], &ok); + V8_FALLTHROUGH; + case 3: + macro_assembler->CheckCharacter(chars[0], &ok); + macro_assembler->CheckCharacter(chars[1], &ok); + macro_assembler->CheckNotCharacter(chars[2], on_failure); + macro_assembler->Bind(&ok); + break; + default: + UNREACHABLE(); + } + return true; +} + +static void EmitBoundaryTest(RegExpMacroAssembler* masm, int border, + Label* fall_through, Label* above_or_equal, + Label* below) { + if (below != fall_through) { + masm->CheckCharacterLT(border, below); + if (above_or_equal != fall_through) masm->GoTo(above_or_equal); + } else { + masm->CheckCharacterGT(border - 1, above_or_equal); + } +} + +static void EmitDoubleBoundaryTest(RegExpMacroAssembler* masm, int first, + int last, Label* fall_through, + Label* in_range, Label* out_of_range) { + if (in_range == fall_through) { + if (first == last) { + masm->CheckNotCharacter(first, out_of_range); + } else { + masm->CheckCharacterNotInRange(first, last, out_of_range); + } + } else { + if (first == last) { + masm->CheckCharacter(first, in_range); + } else { + masm->CheckCharacterInRange(first, last, in_range); + } + if (out_of_range != fall_through) masm->GoTo(out_of_range); + } +} + +// even_label is for ranges[i] to ranges[i + 1] where i - start_index is even. +// odd_label is for ranges[i] to ranges[i + 1] where i - start_index is odd. +static void EmitUseLookupTable(RegExpMacroAssembler* masm, + ZoneList* ranges, int start_index, + int end_index, int min_char, Label* fall_through, + Label* even_label, Label* odd_label) { + static const int kSize = RegExpMacroAssembler::kTableSize; + static const int kMask = RegExpMacroAssembler::kTableMask; + + int base = (min_char & ~kMask); + USE(base); + + // Assert that everything is on one kTableSize page. + for (int i = start_index; i <= end_index; i++) { + DCHECK_EQ(ranges->at(i) & ~kMask, base); + } + DCHECK(start_index == 0 || (ranges->at(start_index - 1) & ~kMask) <= base); + + char templ[kSize]; + Label* on_bit_set; + Label* on_bit_clear; + int bit; + if (even_label == fall_through) { + on_bit_set = odd_label; + on_bit_clear = even_label; + bit = 1; + } else { + on_bit_set = even_label; + on_bit_clear = odd_label; + bit = 0; + } + for (int i = 0; i < (ranges->at(start_index) & kMask) && i < kSize; i++) { + templ[i] = bit; + } + int j = 0; + bit ^= 1; + for (int i = start_index; i < end_index; i++) { + for (j = (ranges->at(i) & kMask); j < (ranges->at(i + 1) & kMask); j++) { + templ[j] = bit; + } + bit ^= 1; + } + for (int i = j; i < kSize; i++) { + templ[i] = bit; + } + Factory* factory = masm->isolate()->factory(); + // TODO(erikcorry): Cache these. + Handle ba = factory->NewByteArray(kSize, AllocationType::kOld); + for (int i = 0; i < kSize; i++) { + ba->set(i, templ[i]); + } + masm->CheckBitInTable(ba, on_bit_set); + if (on_bit_clear != fall_through) masm->GoTo(on_bit_clear); +} + +static void CutOutRange(RegExpMacroAssembler* masm, ZoneList* ranges, + int start_index, int end_index, int cut_index, + Label* even_label, Label* odd_label) { + bool odd = (((cut_index - start_index) & 1) == 1); + Label* in_range_label = odd ? odd_label : even_label; + Label dummy; + EmitDoubleBoundaryTest(masm, ranges->at(cut_index), + ranges->at(cut_index + 1) - 1, &dummy, in_range_label, + &dummy); + DCHECK(!dummy.is_linked()); + // Cut out the single range by rewriting the array. This creates a new + // range that is a merger of the two ranges on either side of the one we + // are cutting out. The oddity of the labels is preserved. + for (int j = cut_index; j > start_index; j--) { + ranges->at(j) = ranges->at(j - 1); + } + for (int j = cut_index + 1; j < end_index; j++) { + ranges->at(j) = ranges->at(j + 1); + } +} + +// Unicode case. Split the search space into kSize spaces that are handled +// with recursion. +static void SplitSearchSpace(ZoneList* ranges, int start_index, + int end_index, int* new_start_index, + int* new_end_index, int* border) { + static const int kSize = RegExpMacroAssembler::kTableSize; + static const int kMask = RegExpMacroAssembler::kTableMask; + + int first = ranges->at(start_index); + int last = ranges->at(end_index) - 1; + + *new_start_index = start_index; + *border = (ranges->at(start_index) & ~kMask) + kSize; + while (*new_start_index < end_index) { + if (ranges->at(*new_start_index) > *border) break; + (*new_start_index)++; + } + // new_start_index is the index of the first edge that is beyond the + // current kSize space. + + // For very large search spaces we do a binary chop search of the non-Latin1 + // space instead of just going to the end of the current kSize space. The + // heuristics are complicated a little by the fact that any 128-character + // encoding space can be quickly tested with a table lookup, so we don't + // wish to do binary chop search at a smaller granularity than that. A + // 128-character space can take up a lot of space in the ranges array if, + // for example, we only want to match every second character (eg. the lower + // case characters on some Unicode pages). + int binary_chop_index = (end_index + start_index) / 2; + // The first test ensures that we get to the code that handles the Latin1 + // range with a single not-taken branch, speeding up this important + // character range (even non-Latin1 charset-based text has spaces and + // punctuation). + if (*border - 1 > String::kMaxOneByteCharCode && // Latin1 case. + end_index - start_index > (*new_start_index - start_index) * 2 && + last - first > kSize * 2 && binary_chop_index > *new_start_index && + ranges->at(binary_chop_index) >= first + 2 * kSize) { + int scan_forward_for_section_border = binary_chop_index; + int new_border = (ranges->at(binary_chop_index) | kMask) + 1; + + while (scan_forward_for_section_border < end_index) { + if (ranges->at(scan_forward_for_section_border) > new_border) { + *new_start_index = scan_forward_for_section_border; + *border = new_border; + break; + } + scan_forward_for_section_border++; + } + } + + DCHECK(*new_start_index > start_index); + *new_end_index = *new_start_index - 1; + if (ranges->at(*new_end_index) == *border) { + (*new_end_index)--; + } + if (*border >= ranges->at(end_index)) { + *border = ranges->at(end_index); + *new_start_index = end_index; // Won't be used. + *new_end_index = end_index - 1; + } +} + +// Gets a series of segment boundaries representing a character class. If the +// character is in the range between an even and an odd boundary (counting from +// start_index) then go to even_label, otherwise go to odd_label. We already +// know that the character is in the range of min_char to max_char inclusive. +// Either label can be nullptr indicating backtracking. Either label can also +// be equal to the fall_through label. +static void GenerateBranches(RegExpMacroAssembler* masm, ZoneList* ranges, + int start_index, int end_index, uc32 min_char, + uc32 max_char, Label* fall_through, + Label* even_label, Label* odd_label) { + DCHECK_LE(min_char, String::kMaxUtf16CodeUnit); + DCHECK_LE(max_char, String::kMaxUtf16CodeUnit); + + int first = ranges->at(start_index); + int last = ranges->at(end_index) - 1; + + DCHECK_LT(min_char, first); + + // Just need to test if the character is before or on-or-after + // a particular character. + if (start_index == end_index) { + EmitBoundaryTest(masm, first, fall_through, even_label, odd_label); + return; + } + + // Another almost trivial case: There is one interval in the middle that is + // different from the end intervals. + if (start_index + 1 == end_index) { + EmitDoubleBoundaryTest(masm, first, last, fall_through, even_label, + odd_label); + return; + } + + // It's not worth using table lookup if there are very few intervals in the + // character class. + if (end_index - start_index <= 6) { + // It is faster to test for individual characters, so we look for those + // first, then try arbitrary ranges in the second round. + static int kNoCutIndex = -1; + int cut = kNoCutIndex; + for (int i = start_index; i < end_index; i++) { + if (ranges->at(i) == ranges->at(i + 1) - 1) { + cut = i; + break; + } + } + if (cut == kNoCutIndex) cut = start_index; + CutOutRange(masm, ranges, start_index, end_index, cut, even_label, + odd_label); + DCHECK_GE(end_index - start_index, 2); + GenerateBranches(masm, ranges, start_index + 1, end_index - 1, min_char, + max_char, fall_through, even_label, odd_label); + return; + } + + // If there are a lot of intervals in the regexp, then we will use tables to + // determine whether the character is inside or outside the character class. + static const int kBits = RegExpMacroAssembler::kTableSizeBits; + + if ((max_char >> kBits) == (min_char >> kBits)) { + EmitUseLookupTable(masm, ranges, start_index, end_index, min_char, + fall_through, even_label, odd_label); + return; + } + + if ((min_char >> kBits) != (first >> kBits)) { + masm->CheckCharacterLT(first, odd_label); + GenerateBranches(masm, ranges, start_index + 1, end_index, first, max_char, + fall_through, odd_label, even_label); + return; + } + + int new_start_index = 0; + int new_end_index = 0; + int border = 0; + + SplitSearchSpace(ranges, start_index, end_index, &new_start_index, + &new_end_index, &border); + + Label handle_rest; + Label* above = &handle_rest; + if (border == last + 1) { + // We didn't find any section that started after the limit, so everything + // above the border is one of the terminal labels. + above = (end_index & 1) != (start_index & 1) ? odd_label : even_label; + DCHECK(new_end_index == end_index - 1); + } + + DCHECK_LE(start_index, new_end_index); + DCHECK_LE(new_start_index, end_index); + DCHECK_LT(start_index, new_start_index); + DCHECK_LT(new_end_index, end_index); + DCHECK(new_end_index + 1 == new_start_index || + (new_end_index + 2 == new_start_index && + border == ranges->at(new_end_index + 1))); + DCHECK_LT(min_char, border - 1); + DCHECK_LT(border, max_char); + DCHECK_LT(ranges->at(new_end_index), border); + DCHECK(border < ranges->at(new_start_index) || + (border == ranges->at(new_start_index) && + new_start_index == end_index && new_end_index == end_index - 1 && + border == last + 1)); + DCHECK(new_start_index == 0 || border >= ranges->at(new_start_index - 1)); + + masm->CheckCharacterGT(border - 1, above); + Label dummy; + GenerateBranches(masm, ranges, start_index, new_end_index, min_char, + border - 1, &dummy, even_label, odd_label); + if (handle_rest.is_linked()) { + masm->Bind(&handle_rest); + bool flip = (new_start_index & 1) != (start_index & 1); + GenerateBranches(masm, ranges, new_start_index, end_index, border, max_char, + &dummy, flip ? odd_label : even_label, + flip ? even_label : odd_label); + } +} + +static void EmitCharClass(RegExpMacroAssembler* macro_assembler, + RegExpCharacterClass* cc, bool one_byte, + Label* on_failure, int cp_offset, bool check_offset, + bool preloaded, Zone* zone) { + ZoneList* ranges = cc->ranges(zone); + CharacterRange::Canonicalize(ranges); + + int max_char; + if (one_byte) { + max_char = String::kMaxOneByteCharCode; + } else { + max_char = String::kMaxUtf16CodeUnit; + } + + int range_count = ranges->length(); + + int last_valid_range = range_count - 1; + while (last_valid_range >= 0) { + CharacterRange& range = ranges->at(last_valid_range); + if (range.from() <= max_char) { + break; + } + last_valid_range--; + } + + if (last_valid_range < 0) { + if (!cc->is_negated()) { + macro_assembler->GoTo(on_failure); + } + if (check_offset) { + macro_assembler->CheckPosition(cp_offset, on_failure); + } + return; + } + + if (last_valid_range == 0 && ranges->at(0).IsEverything(max_char)) { + if (cc->is_negated()) { + macro_assembler->GoTo(on_failure); + } else { + // This is a common case hit by non-anchored expressions. + if (check_offset) { + macro_assembler->CheckPosition(cp_offset, on_failure); + } + } + return; + } + + if (!preloaded) { + macro_assembler->LoadCurrentCharacter(cp_offset, on_failure, check_offset); + } + + if (cc->is_standard(zone) && macro_assembler->CheckSpecialCharacterClass( + cc->standard_type(), on_failure)) { + return; + } + + // A new list with ascending entries. Each entry is a code unit + // where there is a boundary between code units that are part of + // the class and code units that are not. Normally we insert an + // entry at zero which goes to the failure label, but if there + // was already one there we fall through for success on that entry. + // Subsequent entries have alternating meaning (success/failure). + ZoneList* range_boundaries = + new (zone) ZoneList(last_valid_range, zone); + + bool zeroth_entry_is_failure = !cc->is_negated(); + + for (int i = 0; i <= last_valid_range; i++) { + CharacterRange& range = ranges->at(i); + if (range.from() == 0) { + DCHECK_EQ(i, 0); + zeroth_entry_is_failure = !zeroth_entry_is_failure; + } else { + range_boundaries->Add(range.from(), zone); + } + range_boundaries->Add(range.to() + 1, zone); + } + int end_index = range_boundaries->length() - 1; + if (range_boundaries->at(end_index) > max_char) { + end_index--; + } + + Label fall_through; + GenerateBranches(macro_assembler, range_boundaries, + 0, // start_index. + end_index, + 0, // min_char. + max_char, &fall_through, + zeroth_entry_is_failure ? &fall_through : on_failure, + zeroth_entry_is_failure ? on_failure : &fall_through); + macro_assembler->Bind(&fall_through); +} + +RegExpNode::~RegExpNode() = default; + +RegExpNode::LimitResult RegExpNode::LimitVersions(RegExpCompiler* compiler, + Trace* trace) { + // If we are generating a greedy loop then don't stop and don't reuse code. + if (trace->stop_node() != nullptr) { + return CONTINUE; + } + + RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); + if (trace->is_trivial()) { + if (label_.is_bound() || on_work_list() || !KeepRecursing(compiler)) { + // If a generic version is already scheduled to be generated or we have + // recursed too deeply then just generate a jump to that code. + macro_assembler->GoTo(&label_); + // This will queue it up for generation of a generic version if it hasn't + // already been queued. + compiler->AddWork(this); + return DONE; + } + // Generate generic version of the node and bind the label for later use. + macro_assembler->Bind(&label_); + return CONTINUE; + } + + // We are being asked to make a non-generic version. Keep track of how many + // non-generic versions we generate so as not to overdo it. + trace_count_++; + if (KeepRecursing(compiler) && compiler->optimize() && + trace_count_ < kMaxCopiesCodeGenerated) { + return CONTINUE; + } + + // If we get here code has been generated for this node too many times or + // recursion is too deep. Time to switch to a generic version. The code for + // generic versions above can handle deep recursion properly. + bool was_limiting = compiler->limiting_recursion(); + compiler->set_limiting_recursion(true); + trace->Flush(compiler, this); + compiler->set_limiting_recursion(was_limiting); + return DONE; +} + +bool RegExpNode::KeepRecursing(RegExpCompiler* compiler) { + return !compiler->limiting_recursion() && + compiler->recursion_depth() <= RegExpCompiler::kMaxRecursion; +} + +int ActionNode::EatsAtLeast(int still_to_find, int budget, bool not_at_start) { + if (budget <= 0) return 0; + if (action_type_ == POSITIVE_SUBMATCH_SUCCESS) return 0; // Rewinds input! + return on_success()->EatsAtLeast(still_to_find, budget - 1, not_at_start); +} + +void ActionNode::FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) { + if (action_type_ != POSITIVE_SUBMATCH_SUCCESS) { + on_success()->FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start); + } + SaveBMInfo(bm, not_at_start, offset); +} + +int AssertionNode::EatsAtLeast(int still_to_find, int budget, + bool not_at_start) { + if (budget <= 0) return 0; + // If we know we are not at the start and we are asked "how many characters + // will you match if you succeed?" then we can answer anything since false + // implies false. So lets just return the max answer (still_to_find) since + // that won't prevent us from preloading a lot of characters for the other + // branches in the node graph. + if (assertion_type() == AT_START && not_at_start) return still_to_find; + return on_success()->EatsAtLeast(still_to_find, budget - 1, not_at_start); +} + +void AssertionNode::FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) { + // Match the behaviour of EatsAtLeast on this node. + if (assertion_type() == AT_START && not_at_start) return; + on_success()->FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start); + SaveBMInfo(bm, not_at_start, offset); +} + +int BackReferenceNode::EatsAtLeast(int still_to_find, int budget, + bool not_at_start) { + if (read_backward()) return 0; + if (budget <= 0) return 0; + return on_success()->EatsAtLeast(still_to_find, budget - 1, not_at_start); +} + +int TextNode::EatsAtLeast(int still_to_find, int budget, bool not_at_start) { + if (read_backward()) return 0; + int answer = Length(); + if (answer >= still_to_find) return answer; + if (budget <= 0) return answer; + // We are not at start after this node so we set the last argument to 'true'. + return answer + + on_success()->EatsAtLeast(still_to_find - answer, budget - 1, true); +} + +int NegativeLookaroundChoiceNode::EatsAtLeast(int still_to_find, int budget, + bool not_at_start) { + if (budget <= 0) return 0; + // Alternative 0 is the negative lookahead, alternative 1 is what comes + // afterwards. + RegExpNode* node = alternatives_->at(1).node(); + return node->EatsAtLeast(still_to_find, budget - 1, not_at_start); +} + +void NegativeLookaroundChoiceNode::GetQuickCheckDetails( + QuickCheckDetails* details, RegExpCompiler* compiler, int filled_in, + bool not_at_start) { + // Alternative 0 is the negative lookahead, alternative 1 is what comes + // afterwards. + RegExpNode* node = alternatives_->at(1).node(); + return node->GetQuickCheckDetails(details, compiler, filled_in, not_at_start); +} + +int ChoiceNode::EatsAtLeastHelper(int still_to_find, int budget, + RegExpNode* ignore_this_node, + bool not_at_start) { + if (budget <= 0) return 0; + int min = 100; + int choice_count = alternatives_->length(); + budget = (budget - 1) / choice_count; + for (int i = 0; i < choice_count; i++) { + RegExpNode* node = alternatives_->at(i).node(); + if (node == ignore_this_node) continue; + int node_eats_at_least = + node->EatsAtLeast(still_to_find, budget, not_at_start); + if (node_eats_at_least < min) min = node_eats_at_least; + if (min == 0) return 0; + } + return min; +} + +int LoopChoiceNode::EatsAtLeast(int still_to_find, int budget, + bool not_at_start) { + return EatsAtLeastHelper(still_to_find, budget - 1, loop_node_, not_at_start); +} + +int ChoiceNode::EatsAtLeast(int still_to_find, int budget, bool not_at_start) { + return EatsAtLeastHelper(still_to_find, budget, nullptr, not_at_start); +} + +// Takes the left-most 1-bit and smears it out, setting all bits to its right. +static inline uint32_t SmearBitsRight(uint32_t v) { + v |= v >> 1; + v |= v >> 2; + v |= v >> 4; + v |= v >> 8; + v |= v >> 16; + return v; +} + +bool QuickCheckDetails::Rationalize(bool asc) { + bool found_useful_op = false; + uint32_t char_mask; + if (asc) { + char_mask = String::kMaxOneByteCharCode; + } else { + char_mask = String::kMaxUtf16CodeUnit; + } + mask_ = 0; + value_ = 0; + int char_shift = 0; + for (int i = 0; i < characters_; i++) { + Position* pos = &positions_[i]; + if ((pos->mask & String::kMaxOneByteCharCode) != 0) { + found_useful_op = true; + } + mask_ |= (pos->mask & char_mask) << char_shift; + value_ |= (pos->value & char_mask) << char_shift; + char_shift += asc ? 8 : 16; + } + return found_useful_op; +} + +bool RegExpNode::EmitQuickCheck(RegExpCompiler* compiler, + Trace* bounds_check_trace, Trace* trace, + bool preload_has_checked_bounds, + Label* on_possible_success, + QuickCheckDetails* details, + bool fall_through_on_failure) { + if (details->characters() == 0) return false; + GetQuickCheckDetails(details, compiler, 0, + trace->at_start() == Trace::FALSE_VALUE); + if (details->cannot_match()) return false; + if (!details->Rationalize(compiler->one_byte())) return false; + DCHECK(details->characters() == 1 || + compiler->macro_assembler()->CanReadUnaligned()); + uint32_t mask = details->mask(); + uint32_t value = details->value(); + + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + + if (trace->characters_preloaded() != details->characters()) { + DCHECK(trace->cp_offset() == bounds_check_trace->cp_offset()); + // We are attempting to preload the minimum number of characters + // any choice would eat, so if the bounds check fails, then none of the + // choices can succeed, so we can just immediately backtrack, rather + // than go to the next choice. + assembler->LoadCurrentCharacter( + trace->cp_offset(), bounds_check_trace->backtrack(), + !preload_has_checked_bounds, details->characters()); + } + + bool need_mask = true; + + if (details->characters() == 1) { + // If number of characters preloaded is 1 then we used a byte or 16 bit + // load so the value is already masked down. + uint32_t char_mask; + if (compiler->one_byte()) { + char_mask = String::kMaxOneByteCharCode; + } else { + char_mask = String::kMaxUtf16CodeUnit; + } + if ((mask & char_mask) == char_mask) need_mask = false; + mask &= char_mask; + } else { + // For 2-character preloads in one-byte mode or 1-character preloads in + // two-byte mode we also use a 16 bit load with zero extend. + static const uint32_t kTwoByteMask = 0xFFFF; + static const uint32_t kFourByteMask = 0xFFFFFFFF; + if (details->characters() == 2 && compiler->one_byte()) { + if ((mask & kTwoByteMask) == kTwoByteMask) need_mask = false; + } else if (details->characters() == 1 && !compiler->one_byte()) { + if ((mask & kTwoByteMask) == kTwoByteMask) need_mask = false; + } else { + if (mask == kFourByteMask) need_mask = false; + } + } + + if (fall_through_on_failure) { + if (need_mask) { + assembler->CheckCharacterAfterAnd(value, mask, on_possible_success); + } else { + assembler->CheckCharacter(value, on_possible_success); + } + } else { + if (need_mask) { + assembler->CheckNotCharacterAfterAnd(value, mask, trace->backtrack()); + } else { + assembler->CheckNotCharacter(value, trace->backtrack()); + } + } + return true; +} + +// Here is the meat of GetQuickCheckDetails (see also the comment on the +// super-class in the .h file). +// +// We iterate along the text object, building up for each character a +// mask and value that can be used to test for a quick failure to match. +// The masks and values for the positions will be combined into a single +// machine word for the current character width in order to be used in +// generating a quick check. +void TextNode::GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, + int characters_filled_in, + bool not_at_start) { + // Do not collect any quick check details if the text node reads backward, + // since it reads in the opposite direction than we use for quick checks. + if (read_backward()) return; + Isolate* isolate = compiler->macro_assembler()->isolate(); + DCHECK(characters_filled_in < details->characters()); + int characters = details->characters(); + int char_mask; + if (compiler->one_byte()) { + char_mask = String::kMaxOneByteCharCode; + } else { + char_mask = String::kMaxUtf16CodeUnit; + } + for (int k = 0; k < elements()->length(); k++) { + TextElement elm = elements()->at(k); + if (elm.text_type() == TextElement::ATOM) { + Vector quarks = elm.atom()->data(); + for (int i = 0; i < characters && i < quarks.length(); i++) { + QuickCheckDetails::Position* pos = + details->positions(characters_filled_in); + uc16 c = quarks[i]; + if (elm.atom()->ignore_case()) { + unibrow::uchar chars[4]; + int length = GetCaseIndependentLetters( + isolate, c, compiler->one_byte(), chars, 4); + if (length == 0) { + // This can happen because all case variants are non-Latin1, but we + // know the input is Latin1. + details->set_cannot_match(); + pos->determines_perfectly = false; + return; + } + if (length == 1) { + // This letter has no case equivalents, so it's nice and simple + // and the mask-compare will determine definitely whether we have + // a match at this character position. + pos->mask = char_mask; + pos->value = c; + pos->determines_perfectly = true; + } else { + uint32_t common_bits = char_mask; + uint32_t bits = chars[0]; + for (int j = 1; j < length; j++) { + uint32_t differing_bits = ((chars[j] & common_bits) ^ bits); + common_bits ^= differing_bits; + bits &= common_bits; + } + // If length is 2 and common bits has only one zero in it then + // our mask and compare instruction will determine definitely + // whether we have a match at this character position. Otherwise + // it can only be an approximate check. + uint32_t one_zero = (common_bits | ~char_mask); + if (length == 2 && ((~one_zero) & ((~one_zero) - 1)) == 0) { + pos->determines_perfectly = true; + } + pos->mask = common_bits; + pos->value = bits; + } + } else { + // Don't ignore case. Nice simple case where the mask-compare will + // determine definitely whether we have a match at this character + // position. + if (c > char_mask) { + details->set_cannot_match(); + pos->determines_perfectly = false; + return; + } + pos->mask = char_mask; + pos->value = c; + pos->determines_perfectly = true; + } + characters_filled_in++; + DCHECK(characters_filled_in <= details->characters()); + if (characters_filled_in == details->characters()) { + return; + } + } + } else { + QuickCheckDetails::Position* pos = + details->positions(characters_filled_in); + RegExpCharacterClass* tree = elm.char_class(); + ZoneList* ranges = tree->ranges(zone()); + DCHECK(!ranges->is_empty()); + if (tree->is_negated()) { + // A quick check uses multi-character mask and compare. There is no + // useful way to incorporate a negative char class into this scheme + // so we just conservatively create a mask and value that will always + // succeed. + pos->mask = 0; + pos->value = 0; + } else { + int first_range = 0; + while (ranges->at(first_range).from() > char_mask) { + first_range++; + if (first_range == ranges->length()) { + details->set_cannot_match(); + pos->determines_perfectly = false; + return; + } + } + CharacterRange range = ranges->at(first_range); + uc16 from = range.from(); + uc16 to = range.to(); + if (to > char_mask) { + to = char_mask; + } + uint32_t differing_bits = (from ^ to); + // A mask and compare is only perfect if the differing bits form a + // number like 00011111 with one single block of trailing 1s. + if ((differing_bits & (differing_bits + 1)) == 0 && + from + differing_bits == to) { + pos->determines_perfectly = true; + } + uint32_t common_bits = ~SmearBitsRight(differing_bits); + uint32_t bits = (from & common_bits); + for (int i = first_range + 1; i < ranges->length(); i++) { + CharacterRange range = ranges->at(i); + uc16 from = range.from(); + uc16 to = range.to(); + if (from > char_mask) continue; + if (to > char_mask) to = char_mask; + // Here we are combining more ranges into the mask and compare + // value. With each new range the mask becomes more sparse and + // so the chances of a false positive rise. A character class + // with multiple ranges is assumed never to be equivalent to a + // mask and compare operation. + pos->determines_perfectly = false; + uint32_t new_common_bits = (from ^ to); + new_common_bits = ~SmearBitsRight(new_common_bits); + common_bits &= new_common_bits; + bits &= new_common_bits; + uint32_t differing_bits = (from & common_bits) ^ bits; + common_bits ^= differing_bits; + bits &= common_bits; + } + pos->mask = common_bits; + pos->value = bits; + } + characters_filled_in++; + DCHECK(characters_filled_in <= details->characters()); + if (characters_filled_in == details->characters()) { + return; + } + } + } + DCHECK(characters_filled_in != details->characters()); + if (!details->cannot_match()) { + on_success()->GetQuickCheckDetails(details, compiler, characters_filled_in, + true); + } +} + +void QuickCheckDetails::Clear() { + for (int i = 0; i < characters_; i++) { + positions_[i].mask = 0; + positions_[i].value = 0; + positions_[i].determines_perfectly = false; + } + characters_ = 0; +} + +void QuickCheckDetails::Advance(int by, bool one_byte) { + if (by >= characters_ || by < 0) { + DCHECK_IMPLIES(by < 0, characters_ == 0); + Clear(); + return; + } + DCHECK_LE(characters_ - by, 4); + DCHECK_LE(characters_, 4); + for (int i = 0; i < characters_ - by; i++) { + positions_[i] = positions_[by + i]; + } + for (int i = characters_ - by; i < characters_; i++) { + positions_[i].mask = 0; + positions_[i].value = 0; + positions_[i].determines_perfectly = false; + } + characters_ -= by; + // We could change mask_ and value_ here but we would never advance unless + // they had already been used in a check and they won't be used again because + // it would gain us nothing. So there's no point. +} + +void QuickCheckDetails::Merge(QuickCheckDetails* other, int from_index) { + DCHECK(characters_ == other->characters_); + if (other->cannot_match_) { + return; + } + if (cannot_match_) { + *this = *other; + return; + } + for (int i = from_index; i < characters_; i++) { + QuickCheckDetails::Position* pos = positions(i); + QuickCheckDetails::Position* other_pos = other->positions(i); + if (pos->mask != other_pos->mask || pos->value != other_pos->value || + !other_pos->determines_perfectly) { + // Our mask-compare operation will be approximate unless we have the + // exact same operation on both sides of the alternation. + pos->determines_perfectly = false; + } + pos->mask &= other_pos->mask; + pos->value &= pos->mask; + other_pos->value &= pos->mask; + uc16 differing_bits = (pos->value ^ other_pos->value); + pos->mask &= ~differing_bits; + pos->value &= pos->mask; + } +} + +class VisitMarker { + public: + explicit VisitMarker(NodeInfo* info) : info_(info) { + DCHECK(!info->visited); + info->visited = true; + } + ~VisitMarker() { info_->visited = false; } + + private: + NodeInfo* info_; +}; + +RegExpNode* SeqRegExpNode::FilterOneByte(int depth) { + if (info()->replacement_calculated) return replacement(); + if (depth < 0) return this; + DCHECK(!info()->visited); + VisitMarker marker(info()); + return FilterSuccessor(depth - 1); +} + +RegExpNode* SeqRegExpNode::FilterSuccessor(int depth) { + RegExpNode* next = on_success_->FilterOneByte(depth - 1); + if (next == nullptr) return set_replacement(nullptr); + on_success_ = next; + return set_replacement(this); +} + +// We need to check for the following characters: 0x39C 0x3BC 0x178. +bool RangeContainsLatin1Equivalents(CharacterRange range) { + // TODO(dcarney): this could be a lot more efficient. + return range.Contains(0x039C) || range.Contains(0x03BC) || + range.Contains(0x0178); +} + +static bool RangesContainLatin1Equivalents(ZoneList* ranges) { + for (int i = 0; i < ranges->length(); i++) { + // TODO(dcarney): this could be a lot more efficient. + if (RangeContainsLatin1Equivalents(ranges->at(i))) return true; + } + return false; +} + +RegExpNode* TextNode::FilterOneByte(int depth) { + if (info()->replacement_calculated) return replacement(); + if (depth < 0) return this; + DCHECK(!info()->visited); + VisitMarker marker(info()); + int element_count = elements()->length(); + for (int i = 0; i < element_count; i++) { + TextElement elm = elements()->at(i); + if (elm.text_type() == TextElement::ATOM) { + Vector quarks = elm.atom()->data(); + for (int j = 0; j < quarks.length(); j++) { + uint16_t c = quarks[j]; + if (elm.atom()->ignore_case()) { + c = unibrow::Latin1::TryConvertToLatin1(c); + } + if (c > unibrow::Latin1::kMaxChar) return set_replacement(nullptr); + // Replace quark in case we converted to Latin-1. + uint16_t* writable_quarks = const_cast(quarks.begin()); + writable_quarks[j] = c; + } + } else { + DCHECK(elm.text_type() == TextElement::CHAR_CLASS); + RegExpCharacterClass* cc = elm.char_class(); + ZoneList* ranges = cc->ranges(zone()); + CharacterRange::Canonicalize(ranges); + // Now they are in order so we only need to look at the first. + int range_count = ranges->length(); + if (cc->is_negated()) { + if (range_count != 0 && ranges->at(0).from() == 0 && + ranges->at(0).to() >= String::kMaxOneByteCharCode) { + // This will be handled in a later filter. + if (IgnoreCase(cc->flags()) && RangesContainLatin1Equivalents(ranges)) + continue; + return set_replacement(nullptr); + } + } else { + if (range_count == 0 || + ranges->at(0).from() > String::kMaxOneByteCharCode) { + // This will be handled in a later filter. + if (IgnoreCase(cc->flags()) && RangesContainLatin1Equivalents(ranges)) + continue; + return set_replacement(nullptr); + } + } + } + } + return FilterSuccessor(depth - 1); +} + +RegExpNode* LoopChoiceNode::FilterOneByte(int depth) { + if (info()->replacement_calculated) return replacement(); + if (depth < 0) return this; + if (info()->visited) return this; + { + VisitMarker marker(info()); + + RegExpNode* continue_replacement = continue_node_->FilterOneByte(depth - 1); + // If we can't continue after the loop then there is no sense in doing the + // loop. + if (continue_replacement == nullptr) return set_replacement(nullptr); + } + + return ChoiceNode::FilterOneByte(depth - 1); +} + +RegExpNode* ChoiceNode::FilterOneByte(int depth) { + if (info()->replacement_calculated) return replacement(); + if (depth < 0) return this; + if (info()->visited) return this; + VisitMarker marker(info()); + int choice_count = alternatives_->length(); + + for (int i = 0; i < choice_count; i++) { + GuardedAlternative alternative = alternatives_->at(i); + if (alternative.guards() != nullptr && + alternative.guards()->length() != 0) { + set_replacement(this); + return this; + } + } + + int surviving = 0; + RegExpNode* survivor = nullptr; + for (int i = 0; i < choice_count; i++) { + GuardedAlternative alternative = alternatives_->at(i); + RegExpNode* replacement = alternative.node()->FilterOneByte(depth - 1); + DCHECK(replacement != this); // No missing EMPTY_MATCH_CHECK. + if (replacement != nullptr) { + alternatives_->at(i).set_node(replacement); + surviving++; + survivor = replacement; + } + } + if (surviving < 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. + ZoneList* new_alternatives = + new (zone()) ZoneList(surviving, zone()); + for (int i = 0; i < choice_count; i++) { + RegExpNode* replacement = + alternatives_->at(i).node()->FilterOneByte(depth - 1); + if (replacement != nullptr) { + alternatives_->at(i).set_node(replacement); + new_alternatives->Add(alternatives_->at(i), zone()); + } + } + alternatives_ = new_alternatives; + return this; +} + +RegExpNode* NegativeLookaroundChoiceNode::FilterOneByte(int 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_->at(1).node(); + RegExpNode* replacement = node->FilterOneByte(depth - 1); + if (replacement == nullptr) return set_replacement(nullptr); + alternatives_->at(1).set_node(replacement); + + RegExpNode* neg_node = alternatives_->at(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_->at(0).set_node(neg_replacement); + return set_replacement(this); +} + +void LoopChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, + int characters_filled_in, + bool not_at_start) { + if (body_can_be_zero_length_ || info()->visited) return; + VisitMarker marker(info()); + return ChoiceNode::GetQuickCheckDetails(details, compiler, + characters_filled_in, not_at_start); +} + +void LoopChoiceNode::FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) { + if (body_can_be_zero_length_ || budget <= 0) { + bm->SetRest(offset); + SaveBMInfo(bm, not_at_start, offset); + return; + } + ChoiceNode::FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start); + SaveBMInfo(bm, not_at_start, offset); +} + +void ChoiceNode::GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, + int characters_filled_in, + bool not_at_start) { + not_at_start = (not_at_start || not_at_start_); + int choice_count = alternatives_->length(); + DCHECK_LT(0, choice_count); + alternatives_->at(0).node()->GetQuickCheckDetails( + details, compiler, characters_filled_in, not_at_start); + for (int i = 1; i < choice_count; i++) { + QuickCheckDetails new_details(details->characters()); + RegExpNode* node = alternatives_->at(i).node(); + node->GetQuickCheckDetails(&new_details, compiler, characters_filled_in, + not_at_start); + // Here we merge the quick match details of the two branches. + details->Merge(&new_details, characters_filled_in); + } +} + +// Check for [0-9A-Z_a-z]. +static void EmitWordCheck(RegExpMacroAssembler* assembler, Label* word, + Label* non_word, bool fall_through_on_word) { + if (assembler->CheckSpecialCharacterClass( + fall_through_on_word ? 'w' : 'W', + fall_through_on_word ? non_word : word)) { + // Optimized implementation available. + return; + } + assembler->CheckCharacterGT('z', non_word); + assembler->CheckCharacterLT('0', non_word); + assembler->CheckCharacterGT('a' - 1, word); + assembler->CheckCharacterLT('9' + 1, word); + assembler->CheckCharacterLT('A', non_word); + assembler->CheckCharacterLT('Z' + 1, word); + if (fall_through_on_word) { + assembler->CheckNotCharacter('_', non_word); + } else { + assembler->CheckCharacter('_', word); + } +} + +// Emit the code to check for a ^ in multiline mode (1-character lookbehind +// that matches newline or the start of input). +static void EmitHat(RegExpCompiler* compiler, RegExpNode* on_success, + Trace* trace) { + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + // We will be loading the previous character into the current character + // register. + Trace new_trace(*trace); + new_trace.InvalidateCurrentCharacter(); + + Label ok; + if (new_trace.cp_offset() == 0) { + // The start of input counts as a newline in this context, so skip to + // ok if we are at the start. + assembler->CheckAtStart(&ok); + } + // We already checked that we are not at the start of input so it must be + // OK to load the previous character. + assembler->LoadCurrentCharacter(new_trace.cp_offset() - 1, + new_trace.backtrack(), false); + if (!assembler->CheckSpecialCharacterClass('n', new_trace.backtrack())) { + // Newline means \n, \r, 0x2028 or 0x2029. + if (!compiler->one_byte()) { + assembler->CheckCharacterAfterAnd(0x2028, 0xFFFE, &ok); + } + assembler->CheckCharacter('\n', &ok); + assembler->CheckNotCharacter('\r', new_trace.backtrack()); + } + assembler->Bind(&ok); + on_success->Emit(compiler, &new_trace); +} + +// Emit the code to handle \b and \B (word-boundary or non-word-boundary). +void AssertionNode::EmitBoundaryCheck(RegExpCompiler* compiler, Trace* trace) { + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + Isolate* isolate = assembler->isolate(); + Trace::TriBool next_is_word_character = Trace::UNKNOWN; + bool not_at_start = (trace->at_start() == Trace::FALSE_VALUE); + BoyerMooreLookahead* lookahead = bm_info(not_at_start); + if (lookahead == nullptr) { + int eats_at_least = Min(kMaxLookaheadForBoyerMoore, + EatsAtLeast(kMaxLookaheadForBoyerMoore, + kRecursionBudget, not_at_start)); + if (eats_at_least >= 1) { + BoyerMooreLookahead* bm = + new (zone()) BoyerMooreLookahead(eats_at_least, compiler, zone()); + FillInBMInfo(isolate, 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) { + Label before_non_word; + Label before_word; + if (trace->characters_preloaded() != 1) { + assembler->LoadCurrentCharacter(trace->cp_offset(), &before_non_word); + } + // Fall through on non-word. + EmitWordCheck(assembler, &before_word, &before_non_word, false); + // Next character is not a word character. + assembler->Bind(&before_non_word); + Label ok; + BacktrackIfPrevious(compiler, trace, at_boundary ? kIsNonWord : kIsWord); + assembler->GoTo(&ok); + + assembler->Bind(&before_word); + BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord); + assembler->Bind(&ok); + } else if (next_is_word_character == Trace::TRUE_VALUE) { + BacktrackIfPrevious(compiler, trace, at_boundary ? kIsWord : kIsNonWord); + } else { + DCHECK(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(); + + Label fall_through, dummy; + + Label* non_word = backtrack_if_previous == kIsNonWord ? new_trace.backtrack() + : &fall_through; + Label* 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->Bind(&fall_through); + on_success()->Emit(compiler, &new_trace); +} + +void AssertionNode::GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, + int 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: { + Label ok; + assembler->CheckPosition(trace->cp_offset(), &ok); + assembler->GoTo(trace->backtrack()); + assembler->Bind(&ok); + break; + } + case AT_START: { + if (trace->at_start() == Trace::FALSE_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, int offset) { + if (quick_check == nullptr) return false; + if (offset >= quick_check->characters()) return false; + return quick_check->positions(offset)->determines_perfectly; +} + +static void UpdateBoundsCheck(int index, int* checked_up_to) { + if (index > *checked_up_to) { + *checked_up_to = index; + } +} + +// We call this repeatedly to generate code for each pass over the text node. +// The passes are in increasing order of difficulty because we hope one +// of the first passes will fail in which case we are saved the work of the +// later passes. for example for the case independent regexp /%[asdfghjkl]a/ +// we will check the '%' in the first pass, the case independent 'a' in the +// second pass and the character class in the last pass. +// +// The passes are done from right to left, so for example to test for /bar/ +// we will first test for an 'r' with offset 2, then an 'a' with offset 1 +// and then a 'b' with offset 0. This means we can avoid the end-of-input +// bounds check most of the time. In the example we only need to check for +// end-of-input when loading the putative 'r'. +// +// A slight complication involves the fact that the first character may already +// be fetched into a register by the previous node. In this case we want to +// do the test for that character first. We do this in separate passes. The +// 'preloaded' argument indicates that we are doing such a 'pass'. If such a +// pass has been performed then subsequent passes will have true in +// first_element_checked to indicate that that character does not need to be +// checked again. +// +// In addition to all this we are passed a Trace, which can +// contain an AlternativeGeneration object. In this AlternativeGeneration +// object we can see details of any quick check that was already passed in +// order to get to the code we are now generating. The quick check can involve +// loading characters, which means we do not need to recheck the bounds +// up to the limit the quick check already checked. In addition the quick +// check can have involved a mask and compare operation which may simplify +// or obviate the need for further checks at some character positions. +void TextNode::TextEmitPass(RegExpCompiler* compiler, TextEmitPassType pass, + bool preloaded, Trace* trace, + bool first_element_checked, int* checked_up_to) { + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + Isolate* isolate = assembler->isolate(); + bool one_byte = compiler->one_byte(); + Label* backtrack = trace->backtrack(); + QuickCheckDetails* quick_check = trace->quick_check_performed(); + int element_count = elements()->length(); + int backward_offset = read_backward() ? -Length() : 0; + for (int i = preloaded ? 0 : element_count - 1; i >= 0; i--) { + TextElement elm = elements()->at(i); + int cp_offset = trace->cp_offset() + elm.cp_offset() + backward_offset; + if (elm.text_type() == TextElement::ATOM) { + if (SkipPass(pass, elm.atom()->ignore_case())) continue; + Vector quarks = elm.atom()->data(); + for (int j = preloaded ? 0 : quarks.length() - 1; j >= 0; j--) { + if (first_element_checked && i == 0 && j == 0) continue; + if (DeterminedAlready(quick_check, elm.cp_offset() + j)) continue; + EmitCharacterFunction* emit_function = nullptr; + uc16 quark = quarks[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 = unibrow::Latin1::TryConvertToLatin1(quark); + } + switch (pass) { + case NON_LATIN1_MATCH: + DCHECK(one_byte); + if (quark > String::kMaxOneByteCharCode) { + 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) { + bool bounds_check = *checked_up_to < cp_offset + j || read_backward(); + bool bound_checked = + emit_function(isolate, compiler, quark, backtrack, cp_offset + j, + bounds_check, preloaded); + if (bound_checked) UpdateBoundsCheck(cp_offset + j, checked_up_to); + } + } + } else { + DCHECK_EQ(TextElement::CHAR_CLASS, elm.text_type()); + 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, zone()); + UpdateBoundsCheck(cp_offset, checked_up_to); + } + } + } +} + +int TextNode::Length() { + TextElement elm = elements()->last(); + DCHECK_LE(0, elm.cp_offset()); + return elm.cp_offset() + elm.length(); +} + +bool TextNode::SkipPass(TextEmitPassType pass, bool ignore_case) { + if (ignore_case) { + return pass == SIMPLE_CHARACTER_MATCH; + } else { + return pass == NON_LETTER_CHARACTER_MATCH || pass == CASE_CHARACTER_MATCH; + } +} + +TextNode* TextNode::CreateForCharacterRanges(Zone* zone, + ZoneList* ranges, + bool read_backward, + RegExpNode* on_success, + JSRegExp::Flags flags) { + DCHECK_NOT_NULL(ranges); + ZoneList* elms = new (zone) ZoneList(1, zone); + elms->Add(TextElement::CharClass( + new (zone) RegExpCharacterClass(zone, ranges, flags)), + zone); + return new (zone) TextNode(elms, read_backward, on_success); +} + +TextNode* TextNode::CreateForSurrogatePair(Zone* zone, CharacterRange lead, + CharacterRange trail, + bool read_backward, + RegExpNode* on_success, + JSRegExp::Flags flags) { + ZoneList* lead_ranges = CharacterRange::List(zone, lead); + ZoneList* trail_ranges = CharacterRange::List(zone, trail); + ZoneList* elms = new (zone) ZoneList(2, zone); + elms->Add(TextElement::CharClass( + new (zone) RegExpCharacterClass(zone, lead_ranges, flags)), + zone); + elms->Add(TextElement::CharClass( + new (zone) RegExpCharacterClass(zone, trail_ranges, flags)), + zone); + return new (zone) 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; + DCHECK(limit_result == CONTINUE); + + if (trace->cp_offset() + Length() > RegExpMacroAssembler::kMaxCPOffset) { + compiler->SetRegExpTooBig(); + return; + } + + if (compiler->one_byte()) { + int dummy = 0; + TextEmitPass(compiler, NON_LATIN1_MATCH, false, trace, false, &dummy); + } + + bool first_elt_done = false; + int bound_checked_to = trace->cp_offset() - 1; + bound_checked_to += trace->bound_checked_up_to(); + + // If a character is preloaded into the current character register then + // check that now. + if (trace->characters_preloaded() == 1) { + for (int pass = kFirstRealPass; pass <= kLastPass; pass++) { + TextEmitPass(compiler, static_cast(pass), true, trace, + false, &bound_checked_to); + } + first_elt_done = true; + } + + for (int pass = kFirstRealPass; pass <= kLastPass; pass++) { + TextEmitPass(compiler, static_cast(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(int 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_ = Max(0, bound_checked_up_to_ - by); +} + +void TextNode::MakeCaseIndependent(Isolate* isolate, bool is_one_byte) { + int element_count = elements()->length(); + for (int i = 0; i < element_count; i++) { + TextElement elm = elements()->at(i); + if (elm.text_type() == TextElement::CHAR_CLASS) { + RegExpCharacterClass* cc = elm.char_class(); +#ifdef V8_INTL_SUPPORT + bool case_equivalents_already_added = + NeedsUnicodeCaseEquivalents(cc->flags()); +#else + bool case_equivalents_already_added = false; +#endif + if (IgnoreCase(cc->flags()) && !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(zone())) continue; + ZoneList* ranges = cc->ranges(zone()); + CharacterRange::AddCaseEquivalents(isolate, zone(), ranges, + is_one_byte); + } + } + } +} + +int TextNode::GreedyLoopTextLength() { return Length(); } + +RegExpNode* TextNode::GetSuccessorOfOmnivorousTextNode( + RegExpCompiler* compiler) { + if (read_backward()) return nullptr; + if (elements()->length() != 1) return nullptr; + TextElement elm = elements()->at(0); + if (elm.text_type() != TextElement::CHAR_CLASS) return nullptr; + RegExpCharacterClass* node = elm.char_class(); + ZoneList* ranges = node->ranges(zone()); + 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 = String::kMaxOneByteCharCode; + } else { + max_char = String::kMaxUtf16CodeUnit; + } + 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. +int ChoiceNode::GreedyLoopTextLengthForAlternative( + GuardedAlternative* alternative) { + int length = 0; + RegExpNode* node = alternative->node(); + // Later we will generate code for all these text nodes using recursion + // so we have to limit the max number. + int recursion_depth = 0; + while (node != this) { + if (recursion_depth++ > RegExpCompiler::kMaxRecursion) { + return kNodeIsTooComplexForGreedyLoops; + } + int node_length = node->GreedyLoopTextLength(); + if (node_length == kNodeIsTooComplexForGreedyLoops) { + return kNodeIsTooComplexForGreedyLoops; + } + length += node_length; + SeqRegExpNode* seq_node = static_cast(node); + node = seq_node->on_success(); + } + return read_backward() ? -length : length; +} + +void LoopChoiceNode::AddLoopAlternative(GuardedAlternative alt) { + DCHECK_NULL(loop_node_); + AddAlternative(alt); + loop_node_ = alt.node(); +} + +void LoopChoiceNode::AddContinueAlternative(GuardedAlternative alt) { + DCHECK_NULL(continue_node_); + 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. + int text_length = + GreedyLoopTextLengthForAlternative(&(alternatives_->at(0))); + DCHECK_NE(kNodeIsTooComplexForGreedyLoops, text_length); + // Update the counter-based backtracking info on the stack. This is an + // optimization for greedy loops (see below). + DCHECK(trace->cp_offset() == text_length); + macro_assembler->AdvanceCurrentPosition(text_length); + macro_assembler->GoTo(trace->loop_label()); + return; + } + DCHECK_NULL(trace->stop_node()); + if (!trace->is_trivial()) { + trace->Flush(compiler, this); + return; + } + ChoiceNode::Emit(compiler, trace); +} + +int ChoiceNode::CalculatePreloadCharacters(RegExpCompiler* compiler, + int eats_at_least) { + int preload_characters = Min(4, eats_at_least); + DCHECK_LE(preload_characters, 4); + if (compiler->macro_assembler()->CanReadUnaligned()) { + bool one_byte = compiler->one_byte(); + if (one_byte) { + // We can't preload 3 characters because there is no machine instruction + // to do that. We can't just load 4 because we could be reading + // beyond the end of the string, which could cause a memory fault. + if (preload_characters == 3) preload_characters = 2; + } else { + if (preload_characters > 2) preload_characters = 2; + } + } else { + if (preload_characters > 1) preload_characters = 1; + } + return preload_characters; +} + +// This class is used when generating the alternatives in a choice node. It +// records the way the alternative is being code generated. +class AlternativeGeneration : public Malloced { + public: + AlternativeGeneration() + : possible_success(), + expects_preload(false), + after(), + quick_check_details() {} + Label possible_success; + bool expects_preload; + Label after; + QuickCheckDetails quick_check_details; +}; + +// Creates a list of AlternativeGenerations. If the list has a reasonable +// size then it is on the stack, otherwise the excess is on the heap. +class AlternativeGenerationList { + public: + AlternativeGenerationList(int count, Zone* zone) : alt_gens_(count, zone) { + for (int i = 0; i < count && i < kAFew; i++) { + alt_gens_.Add(a_few_alt_gens_ + i, zone); + } + for (int i = kAFew; i < count; i++) { + alt_gens_.Add(new AlternativeGeneration(), zone); + } + } + ~AlternativeGenerationList() { + for (int i = kAFew; i < alt_gens_.length(); i++) { + delete alt_gens_[i]; + alt_gens_[i] = nullptr; + } + } + + AlternativeGeneration* at(int i) { return alt_gens_[i]; } + + private: + static const int kAFew = 10; + ZoneList alt_gens_; + AlternativeGeneration a_few_alt_gens_[kAFew]; +}; + +void BoyerMoorePositionInfo::Set(int character) { + SetInterval(Interval(character, character)); +} + +void BoyerMoorePositionInfo::SetInterval(const Interval& interval) { + s_ = AddRange(s_, kSpaceRanges, kSpaceRangeCount, interval); + w_ = AddRange(w_, kWordRanges, kWordRangeCount, interval); + d_ = AddRange(d_, kDigitRanges, kDigitRangeCount, interval); + surrogate_ = + AddRange(surrogate_, kSurrogateRanges, kSurrogateRangeCount, interval); + if (interval.to() - interval.from() >= kMapSize - 1) { + if (map_count_ != kMapSize) { + map_count_ = kMapSize; + for (int i = 0; i < kMapSize; i++) map_->at(i) = true; + } + return; + } + for (int i = interval.from(); i <= interval.to(); i++) { + int mod_character = (i & kMask); + if (!map_->at(mod_character)) { + map_count_++; + map_->at(mod_character) = true; + } + if (map_count_ == kMapSize) return; + } +} + +void BoyerMoorePositionInfo::SetAll() { + s_ = w_ = d_ = kLatticeUnknown; + if (map_count_ != kMapSize) { + map_count_ = kMapSize; + for (int i = 0; i < kMapSize; i++) map_->at(i) = true; + } +} + +BoyerMooreLookahead::BoyerMooreLookahead(int length, RegExpCompiler* compiler, + Zone* zone) + : length_(length), compiler_(compiler) { + if (compiler->one_byte()) { + max_char_ = String::kMaxOneByteCharCode; + } else { + max_char_ = String::kMaxUtf16CodeUnit; + } + bitmaps_ = new (zone) ZoneList(length, zone); + for (int i = 0; i < length; i++) { + bitmaps_->Add(new (zone) BoyerMoorePositionInfo(zone), zone); + } +} + +// Find the longest range of lookahead that has the fewest number of different +// characters that can occur at a given position. Since we are optimizing two +// different parameters at once this is a tradeoff. +bool BoyerMooreLookahead::FindWorthwhileInterval(int* from, int* to) { + int biggest_points = 0; + // If more than 32 characters out of 128 can occur it is unlikely that we can + // be lucky enough to step forwards much of the time. + const int kMaxMax = 32; + for (int max_number_of_chars = 4; max_number_of_chars < kMaxMax; + max_number_of_chars *= 2) { + biggest_points = + FindBestInterval(max_number_of_chars, biggest_points, from, to); + } + if (biggest_points == 0) return false; + return true; +} + +// Find the highest-points range between 0 and length_ where the character +// information is not too vague. 'Too vague' means that there are more than +// max_number_of_chars that can occur at this position. Calculates the number +// of points as the product of width-of-the-range and +// probability-of-finding-one-of-the-characters, where the probability is +// calculated using the frequency distribution of the sample subject string. +int BoyerMooreLookahead::FindBestInterval(int max_number_of_chars, + int old_biggest_points, int* from, + int* to) { + int biggest_points = old_biggest_points; + static const int kSize = RegExpMacroAssembler::kTableSize; + for (int i = 0; i < length_;) { + while (i < length_ && Count(i) > max_number_of_chars) i++; + if (i == length_) break; + int remembered_from = i; + bool union_map[kSize]; + for (int j = 0; j < kSize; j++) union_map[j] = false; + while (i < length_ && Count(i) <= max_number_of_chars) { + BoyerMoorePositionInfo* map = bitmaps_->at(i); + for (int j = 0; j < kSize; j++) union_map[j] |= map->at(j); + i++; + } + int frequency = 0; + for (int j = 0; j < kSize; j++) { + if (union_map[j]) { + // Add 1 to the frequency to give a small per-character boost for + // the cases where our sampling is not good enough and many + // characters have a frequency of zero. This means the frequency + // can theoretically be up to 2*kSize though we treat it mostly as + // a fraction of kSize. + frequency += compiler_->frequency_collator()->Frequency(j) + 1; + } + } + // We use the probability of skipping times the distance we are skipping to + // judge the effectiveness of this. Actually we have a cut-off: By + // dividing by 2 we switch off the skipping if the probability of skipping + // is less than 50%. This is because the multibyte mask-and-compare + // skipping in quickcheck is more likely to do well on this case. + bool in_quickcheck_range = + ((i - remembered_from < 4) || + (compiler_->one_byte() ? remembered_from <= 4 : remembered_from <= 2)); + // Called 'probability' but it is only a rough estimate and can actually + // be outside the 0-kSize range. + int probability = (in_quickcheck_range ? kSize / 2 : kSize) - frequency; + int points = (i - remembered_from) * probability; + if (points > biggest_points) { + *from = remembered_from; + *to = i - 1; + biggest_points = points; + } + } + return biggest_points; +} + +// Take all the characters that will not prevent a successful match if they +// occur in the subject string in the range between min_lookahead and +// max_lookahead (inclusive) measured from the current position. If the +// character at max_lookahead offset is not one of these characters, then we +// can safely skip forwards by the number of characters in the range. +int BoyerMooreLookahead::GetSkipTable(int min_lookahead, int max_lookahead, + Handle boolean_skip_table) { + const int kSize = RegExpMacroAssembler::kTableSize; + + const int kSkipArrayEntry = 0; + const int kDontSkipArrayEntry = 1; + + for (int i = 0; i < kSize; i++) { + boolean_skip_table->set(i, kSkipArrayEntry); + } + int skip = max_lookahead + 1 - min_lookahead; + + for (int i = max_lookahead; i >= min_lookahead; i--) { + BoyerMoorePositionInfo* map = bitmaps_->at(i); + for (int j = 0; j < kSize; j++) { + if (map->at(j)) { + boolean_skip_table->set(j, kDontSkipArrayEntry); + } + } + } + + return skip; +} + +// See comment above on the implementation of GetSkipTable. +void BoyerMooreLookahead::EmitSkipInstructions(RegExpMacroAssembler* masm) { + const int kSize = RegExpMacroAssembler::kTableSize; + + int min_lookahead = 0; + int max_lookahead = 0; + + if (!FindWorthwhileInterval(&min_lookahead, &max_lookahead)) return; + + bool found_single_character = false; + int single_character = 0; + for (int i = max_lookahead; i >= min_lookahead; i--) { + BoyerMoorePositionInfo* map = bitmaps_->at(i); + if (map->map_count() > 1 || + (found_single_character && map->map_count() != 0)) { + found_single_character = false; + break; + } + for (int j = 0; j < kSize; j++) { + if (map->at(j)) { + found_single_character = true; + single_character = j; + break; + } + } + } + + int lookahead_width = max_lookahead + 1 - min_lookahead; + + if (found_single_character && lookahead_width == 1 && max_lookahead < 3) { + // The mask-compare can probably handle this better. + return; + } + + if (found_single_character) { + Label cont, again; + masm->Bind(&again); + masm->LoadCurrentCharacter(max_lookahead, &cont, true); + if (max_char_ > kSize) { + masm->CheckCharacterAfterAnd(single_character, + RegExpMacroAssembler::kTableMask, &cont); + } else { + masm->CheckCharacter(single_character, &cont); + } + masm->AdvanceCurrentPosition(lookahead_width); + masm->GoTo(&again); + masm->Bind(&cont); + return; + } + + Factory* factory = masm->isolate()->factory(); + Handle boolean_skip_table = + factory->NewByteArray(kSize, AllocationType::kOld); + int skip_distance = + GetSkipTable(min_lookahead, max_lookahead, boolean_skip_table); + DCHECK_NE(0, skip_distance); + + Label cont, again; + masm->Bind(&again); + masm->LoadCurrentCharacter(max_lookahead, &cont, true); + masm->CheckBitInTable(boolean_skip_table, &cont); + masm->AdvanceCurrentPosition(skip_distance); + masm->GoTo(&again); + masm->Bind(&cont); +} + +/* Code generation for choice nodes. + * + * We generate quick checks that do a mask and compare to eliminate a + * choice. If the quick check succeeds then it jumps to the continuation to + * do slow checks and check subsequent nodes. If it fails (the common case) + * it falls through to the next choice. + * + * Here is the desired flow graph. Nodes directly below each other imply + * fallthrough. Alternatives 1 and 2 have quick checks. Alternative + * 3 doesn't have a quick check so we have to call the slow check. + * Nodes are marked Qn for quick checks and Sn for slow checks. The entire + * regexp continuation is generated directly after the Sn node, up to the + * next GoTo if we decide to reuse some already generated code. Some + * nodes expect preload_characters to be preloaded into the current + * character register. R nodes do this preloading. Vertices are marked + * F for failures and S for success (possible success in the case of quick + * nodes). L, V, < and > are used as arrow heads. + * + * ----------> R + * | + * V + * Q1 -----> S1 + * | S / + * F| / + * | F/ + * | / + * | R + * | / + * V L + * Q2 -----> S2 + * | S / + * F| / + * | F/ + * | / + * | R + * | / + * V L + * S3 + * | + * F| + * | + * R + * | + * backtrack V + * <----------Q4 + * \ F | + * \ |S + * \ F V + * \-----S4 + * + * For greedy loops we push the current position, then generate the code that + * eats the input specially in EmitGreedyLoop. The other choice (the + * continuation) is generated by the normal code in EmitChoices, and steps back + * in the input to the starting position when it fails to match. The loop code + * looks like this (U is the unwind code that steps back in the greedy loop). + * + * _____ + * / \ + * V | + * ----------> S1 | + * /| | + * / |S | + * F/ \_____/ + * / + * |<----- + * | \ + * V |S + * Q2 ---> U----->backtrack + * | F / + * S| / + * V F / + * S2--/ + */ + +GreedyLoopState::GreedyLoopState(bool not_at_start) { + counter_backtrack_trace_.set_backtrack(&label_); + if (not_at_start) counter_backtrack_trace_.set_at_start(Trace::FALSE_VALUE); +} + +void ChoiceNode::AssertGuardsMentionRegisters(Trace* trace) { +#ifdef DEBUG + int choice_count = alternatives_->length(); + for (int i = 0; i < choice_count - 1; i++) { + GuardedAlternative alternative = alternatives_->at(i); + ZoneList* guards = alternative.guards(); + int guard_count = (guards == nullptr) ? 0 : guards->length(); + for (int j = 0; j < guard_count; j++) { + DCHECK(!trace->mentions_reg(guards->at(j)->reg())); + } + } +#endif +} + +void ChoiceNode::SetUpPreLoad(RegExpCompiler* compiler, Trace* current_trace, + PreloadState* state) { + if (state->eats_at_least_ == PreloadState::kEatsAtLeastNotYetInitialized) { + // Save some time by looking at most one machine word ahead. + state->eats_at_least_ = + EatsAtLeast(compiler->one_byte() ? 4 : 2, kRecursionBudget, + current_trace->at_start() == Trace::FALSE_VALUE); + } + state->preload_characters_ = + CalculatePreloadCharacters(compiler, state->eats_at_least_); + + state->preload_is_current_ = + (current_trace->characters_preloaded() == state->preload_characters_); + state->preload_has_checked_bounds_ = state->preload_is_current_; +} + +void ChoiceNode::Emit(RegExpCompiler* compiler, Trace* trace) { + int choice_count = alternatives_->length(); + + if (choice_count == 1 && alternatives_->at(0).guards() == nullptr) { + alternatives_->at(0).node()->Emit(compiler, trace); + return; + } + + AssertGuardsMentionRegisters(trace); + + LimitResult limit_result = LimitVersions(compiler, trace); + if (limit_result == DONE) return; + DCHECK(limit_result == CONTINUE); + + // For loop nodes we already flushed (see LoopChoiceNode::Emit), but for + // other choice nodes we only flush if we are out of code size budget. + if (trace->flush_budget() == 0 && trace->actions() != nullptr) { + trace->Flush(compiler, this); + return; + } + + RecursionCheck rc(compiler); + + PreloadState preload; + preload.init(); + GreedyLoopState greedy_loop_state(not_at_start()); + + int text_length = GreedyLoopTextLengthForAlternative(&alternatives_->at(0)); + AlternativeGenerationList alt_gens(choice_count, zone()); + + if (choice_count > 1 && text_length != kNodeIsTooComplexForGreedyLoops) { + trace = EmitGreedyLoop(compiler, trace, &alt_gens, &preload, + &greedy_loop_state, text_length); + } else { + // TODO(erikcorry): Delete this. We don't need this label, but it makes us + // match the traces produced pre-cleanup. + Label second_choice; + compiler->macro_assembler()->Bind(&second_choice); + + preload.eats_at_least_ = EmitOptimizedUnanchoredSearch(compiler, trace); + + EmitChoices(compiler, &alt_gens, 0, trace, &preload); + } + + // At this point we need to generate slow checks for the alternatives where + // the quick check was inlined. We can recognize these because the associated + // label was bound. + int new_flush_budget = trace->flush_budget() / choice_count; + for (int i = 0; i < choice_count; i++) { + AlternativeGeneration* alt_gen = alt_gens.at(i); + Trace new_trace(*trace); + // If there are actions to be flushed we have to limit how many times + // they are flushed. Take the budget of the parent trace and distribute + // it fairly amongst the children. + if (new_trace.actions() != nullptr) { + new_trace.set_flush_budget(new_flush_budget); + } + bool next_expects_preload = + i == choice_count - 1 ? false : alt_gens.at(i + 1)->expects_preload; + EmitOutOfLineContinuation(compiler, &new_trace, alternatives_->at(i), + alt_gen, preload.preload_characters_, + next_expects_preload); + } +} + +Trace* ChoiceNode::EmitGreedyLoop(RegExpCompiler* compiler, Trace* trace, + AlternativeGenerationList* alt_gens, + PreloadState* preload, + GreedyLoopState* greedy_loop_state, + int text_length) { + RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); + // Here we have special handling for greedy loops containing only text nodes + // and other simple nodes. These are handled by pushing the current + // position on the stack and then incrementing the current position each + // time around the switch. On backtrack we decrement the current position + // and check it against the pushed value. This avoids pushing backtrack + // information for each iteration of the loop, which could take up a lot of + // space. + DCHECK(trace->stop_node() == nullptr); + macro_assembler->PushCurrentPosition(); + Label greedy_match_failed; + Trace greedy_match_trace; + if (not_at_start()) greedy_match_trace.set_at_start(Trace::FALSE_VALUE); + greedy_match_trace.set_backtrack(&greedy_match_failed); + Label loop_label; + macro_assembler->Bind(&loop_label); + greedy_match_trace.set_stop_node(this); + greedy_match_trace.set_loop_label(&loop_label); + alternatives_->at(0).node()->Emit(compiler, &greedy_match_trace); + macro_assembler->Bind(&greedy_match_failed); + + Label second_choice; // For use in greedy matches. + macro_assembler->Bind(&second_choice); + + Trace* new_trace = greedy_loop_state->counter_backtrack_trace(); + + EmitChoices(compiler, alt_gens, 1, new_trace, preload); + + macro_assembler->Bind(greedy_loop_state->label()); + // If we have unwound to the bottom then backtrack. + macro_assembler->CheckGreedyLoop(trace->backtrack()); + // Otherwise try the second priority at an earlier position. + macro_assembler->AdvanceCurrentPosition(-text_length); + macro_assembler->GoTo(&second_choice); + return new_trace; +} + +int ChoiceNode::EmitOptimizedUnanchoredSearch(RegExpCompiler* compiler, + Trace* trace) { + int eats_at_least = PreloadState::kEatsAtLeastNotYetInitialized; + if (alternatives_->length() != 2) return eats_at_least; + + GuardedAlternative alt1 = alternatives_->at(1); + if (alt1.guards() != nullptr && alt1.guards()->length() != 0) { + return eats_at_least; + } + RegExpNode* eats_anything_node = alt1.node(); + if (eats_anything_node->GetSuccessorOfOmnivorousTextNode(compiler) != this) { + return eats_at_least; + } + + // Really we should be creating a new trace when we execute this function, + // but there is no need, because the code it generates cannot backtrack, and + // we always arrive here with a trivial trace (since it's the entry to a + // loop. That also implies that there are no preloaded characters, which is + // good, because it means we won't be violating any assumptions by + // overwriting those characters with new load instructions. + DCHECK(trace->is_trivial()); + + RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); + Isolate* isolate = macro_assembler->isolate(); + // At this point we know that we are at a non-greedy loop that will eat + // any character one at a time. Any non-anchored regexp has such a + // loop prepended to it in order to find where it starts. We look for + // a pattern of the form ...abc... where we can look 6 characters ahead + // and step forwards 3 if the character is not one of abc. Abc need + // not be atoms, they can be any reasonably limited character class or + // small alternation. + BoyerMooreLookahead* bm = bm_info(false); + if (bm == nullptr) { + eats_at_least = + Min(kMaxLookaheadForBoyerMoore, + EatsAtLeast(kMaxLookaheadForBoyerMoore, kRecursionBudget, false)); + if (eats_at_least >= 1) { + bm = new (zone()) BoyerMooreLookahead(eats_at_least, compiler, zone()); + GuardedAlternative alt0 = alternatives_->at(0); + alt0.node()->FillInBMInfo(isolate, 0, kRecursionBudget, bm, false); + } + } + if (bm != nullptr) { + bm->EmitSkipInstructions(macro_assembler); + } + return eats_at_least; +} + +void ChoiceNode::EmitChoices(RegExpCompiler* compiler, + AlternativeGenerationList* alt_gens, + int first_choice, Trace* trace, + PreloadState* preload) { + RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); + SetUpPreLoad(compiler, trace, preload); + + // For now we just call all choices one after the other. The idea ultimately + // is to use the Dispatch table to try only the relevant ones. + int choice_count = alternatives_->length(); + + int new_flush_budget = trace->flush_budget() / choice_count; + + for (int i = first_choice; i < choice_count; i++) { + bool is_last = i == choice_count - 1; + bool fall_through_on_failure = !is_last; + GuardedAlternative alternative = alternatives_->at(i); + AlternativeGeneration* alt_gen = alt_gens->at(i); + alt_gen->quick_check_details.set_characters(preload->preload_characters_); + ZoneList* guards = alternative.guards(); + int guard_count = (guards == nullptr) ? 0 : guards->length(); + Trace new_trace(*trace); + new_trace.set_characters_preloaded( + preload->preload_is_current_ ? preload->preload_characters_ : 0); + if (preload->preload_has_checked_bounds_) { + new_trace.set_bound_checked_up_to(preload->preload_characters_); + } + new_trace.quick_check_performed()->Clear(); + if (not_at_start_) new_trace.set_at_start(Trace::FALSE_VALUE); + if (!is_last) { + new_trace.set_backtrack(&alt_gen->after); + } + alt_gen->expects_preload = preload->preload_is_current_; + bool generate_full_check_inline = false; + if (compiler->optimize() && + try_to_emit_quick_check_for_alternative(i == 0) && + alternative.node()->EmitQuickCheck( + compiler, trace, &new_trace, preload->preload_has_checked_bounds_, + &alt_gen->possible_success, &alt_gen->quick_check_details, + fall_through_on_failure)) { + // Quick check was generated for this choice. + preload->preload_is_current_ = true; + preload->preload_has_checked_bounds_ = true; + // If we generated the quick check to fall through on possible success, + // we now need to generate the full check inline. + if (!fall_through_on_failure) { + macro_assembler->Bind(&alt_gen->possible_success); + new_trace.set_quick_check_performed(&alt_gen->quick_check_details); + new_trace.set_characters_preloaded(preload->preload_characters_); + new_trace.set_bound_checked_up_to(preload->preload_characters_); + generate_full_check_inline = true; + } + } else if (alt_gen->quick_check_details.cannot_match()) { + if (!fall_through_on_failure) { + macro_assembler->GoTo(trace->backtrack()); + } + continue; + } else { + // No quick check was generated. Put the full code here. + // If this is not the first choice then there could be slow checks from + // previous cases that go here when they fail. There's no reason to + // insist that they preload characters since the slow check we are about + // to generate probably can't use it. + if (i != first_choice) { + alt_gen->expects_preload = false; + new_trace.InvalidateCurrentCharacter(); + } + generate_full_check_inline = true; + } + if (generate_full_check_inline) { + if (new_trace.actions() != nullptr) { + new_trace.set_flush_budget(new_flush_budget); + } + for (int j = 0; j < guard_count; j++) { + GenerateGuard(macro_assembler, guards->at(j), &new_trace); + } + alternative.node()->Emit(compiler, &new_trace); + preload->preload_is_current_ = false; + } + macro_assembler->Bind(&alt_gen->after); + } +} + +void ChoiceNode::EmitOutOfLineContinuation(RegExpCompiler* compiler, + Trace* trace, + GuardedAlternative alternative, + AlternativeGeneration* alt_gen, + int preload_characters, + bool next_expects_preload) { + if (!alt_gen->possible_success.is_linked()) return; + + RegExpMacroAssembler* macro_assembler = compiler->macro_assembler(); + macro_assembler->Bind(&alt_gen->possible_success); + Trace out_of_line_trace(*trace); + out_of_line_trace.set_characters_preloaded(preload_characters); + out_of_line_trace.set_quick_check_performed(&alt_gen->quick_check_details); + if (not_at_start_) out_of_line_trace.set_at_start(Trace::FALSE_VALUE); + ZoneList* guards = alternative.guards(); + int guard_count = (guards == nullptr) ? 0 : guards->length(); + if (next_expects_preload) { + Label reload_current_char; + out_of_line_trace.set_backtrack(&reload_current_char); + for (int j = 0; j < guard_count; j++) { + GenerateGuard(macro_assembler, guards->at(j), &out_of_line_trace); + } + alternative.node()->Emit(compiler, &out_of_line_trace); + macro_assembler->Bind(&reload_current_char); + // Reload the current character, since the next quick check expects that. + // We don't need to check bounds here because we only get into this + // code through a quick check which already did the checked load. + macro_assembler->LoadCurrentCharacter(trace->cp_offset(), nullptr, false, + preload_characters); + macro_assembler->GoTo(&(alt_gen->after)); + } else { + out_of_line_trace.set_backtrack(&(alt_gen->after)); + for (int j = 0; j < guard_count; j++) { + GenerateGuard(macro_assembler, guards->at(j), &out_of_line_trace); + } + alternative.node()->Emit(compiler, &out_of_line_trace); + } +} + +void ActionNode::Emit(RegExpCompiler* compiler, Trace* trace) { + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + LimitResult limit_result = LimitVersions(compiler, trace); + if (limit_result == DONE) return; + DCHECK(limit_result == CONTINUE); + + RecursionCheck rc(compiler); + + switch (action_type_) { + case STORE_POSITION: { + Trace::DeferredCapture new_capture(data_.u_position_register.reg, + data_.u_position_register.is_capture, + trace); + Trace new_trace = *trace; + new_trace.add_action(&new_capture); + on_success()->Emit(compiler, &new_trace); + break; + } + case INCREMENT_REGISTER: { + Trace::DeferredIncrementRegister new_increment( + data_.u_increment_register.reg); + Trace new_trace = *trace; + new_trace.add_action(&new_increment); + on_success()->Emit(compiler, &new_trace); + break; + } + case SET_REGISTER: { + Trace::DeferredSetRegister new_set(data_.u_store_register.reg, + data_.u_store_register.value); + Trace new_trace = *trace; + new_trace.add_action(&new_set); + on_success()->Emit(compiler, &new_trace); + break; + } + case CLEAR_CAPTURES: { + Trace::DeferredClearCaptures new_capture(Interval( + data_.u_clear_captures.range_from, data_.u_clear_captures.range_to)); + Trace new_trace = *trace; + new_trace.add_action(&new_capture); + on_success()->Emit(compiler, &new_trace); + break; + } + case BEGIN_SUBMATCH: + if (!trace->is_trivial()) { + trace->Flush(compiler, this); + } else { + assembler->WriteCurrentPositionToRegister( + data_.u_submatch.current_position_register, 0); + assembler->WriteStackPointerToRegister( + data_.u_submatch.stack_pointer_register); + on_success()->Emit(compiler, trace); + } + break; + case EMPTY_MATCH_CHECK: { + int start_pos_reg = data_.u_empty_match_check.start_register; + int stored_pos = 0; + int rep_reg = data_.u_empty_match_check.repetition_register; + bool has_minimum = (rep_reg != RegExpCompiler::kNoRegister); + bool know_dist = trace->GetStoredPosition(start_pos_reg, &stored_pos); + if (know_dist && !has_minimum && stored_pos == trace->cp_offset()) { + // If we know we haven't advanced and there is no minimum we + // can just backtrack immediately. + assembler->GoTo(trace->backtrack()); + } else if (know_dist && stored_pos < trace->cp_offset()) { + // If we know we've advanced we can generate the continuation + // immediately. + on_success()->Emit(compiler, trace); + } else if (!trace->is_trivial()) { + trace->Flush(compiler, this); + } else { + Label skip_empty_check; + // If we have a minimum number of repetitions we check the current + // number first and skip the empty check if it's not enough. + if (has_minimum) { + int limit = data_.u_empty_match_check.repetition_limit; + assembler->IfRegisterLT(rep_reg, limit, &skip_empty_check); + } + // If the match is empty we bail out, otherwise we fall through + // to the on-success continuation. + assembler->IfRegisterEqPos(data_.u_empty_match_check.start_register, + trace->backtrack()); + assembler->Bind(&skip_empty_check); + on_success()->Emit(compiler, trace); + } + break; + } + case POSITIVE_SUBMATCH_SUCCESS: { + if (!trace->is_trivial()) { + trace->Flush(compiler, this); + return; + } + assembler->ReadCurrentPositionFromRegister( + data_.u_submatch.current_position_register); + assembler->ReadStackPointerFromRegister( + data_.u_submatch.stack_pointer_register); + int clear_register_count = data_.u_submatch.clear_register_count; + if (clear_register_count == 0) { + on_success()->Emit(compiler, trace); + return; + } + int clear_registers_from = data_.u_submatch.clear_register_from; + Label clear_registers_backtrack; + Trace new_trace = *trace; + new_trace.set_backtrack(&clear_registers_backtrack); + on_success()->Emit(compiler, &new_trace); + + assembler->Bind(&clear_registers_backtrack); + int clear_registers_to = clear_registers_from + clear_register_count - 1; + assembler->ClearRegisters(clear_registers_from, clear_registers_to); + + DCHECK(trace->backtrack() == nullptr); + assembler->Backtrack(); + return; + } + default: + UNREACHABLE(); + } +} + +void BackReferenceNode::Emit(RegExpCompiler* compiler, Trace* trace) { + RegExpMacroAssembler* assembler = compiler->macro_assembler(); + if (!trace->is_trivial()) { + trace->Flush(compiler, this); + return; + } + + LimitResult limit_result = LimitVersions(compiler, trace); + if (limit_result == DONE) return; + DCHECK(limit_result == CONTINUE); + + RecursionCheck rc(compiler); + + DCHECK_EQ(start_reg_ + 1, end_reg_); + if (IgnoreCase(flags_)) { + assembler->CheckNotBackReferenceIgnoreCase( + start_reg_, read_backward(), IsUnicode(flags_), trace->backtrack()); + } else { + assembler->CheckNotBackReference(start_reg_, read_backward(), + trace->backtrack()); + } + // We are going to advance backward, so we may end up at the start. + if (read_backward()) trace->set_at_start(Trace::UNKNOWN); + + // Check that the back reference does not end inside a surrogate pair. + if (IsUnicode(flags_) && !compiler->one_byte()) { + assembler->CheckNotInSurrogatePair(trace->cp_offset(), trace->backtrack()); + } + on_success()->Emit(compiler, trace); +} + +// ------------------------------------------------------------------- +// Splay tree + +OutSet* OutSet::Extend(unsigned value, Zone* zone) { + if (Get(value)) return this; + if (successors(zone) != nullptr) { + for (int i = 0; i < successors(zone)->length(); i++) { + OutSet* successor = successors(zone)->at(i); + if (successor->Get(value)) return successor; + } + } else { + successors_ = new (zone) ZoneList(2, zone); + } + OutSet* result = new (zone) OutSet(first_, remaining_); + result->Set(value, zone); + successors(zone)->Add(result, zone); + return result; +} + +void OutSet::Set(unsigned value, Zone* zone) { + if (value < kFirstLimit) { + first_ |= (1 << value); + } else { + if (remaining_ == nullptr) + remaining_ = new (zone) ZoneList(1, zone); + if (remaining_->is_empty() || !remaining_->Contains(value)) + remaining_->Add(value, zone); + } +} + +bool OutSet::Get(unsigned value) const { + if (value < kFirstLimit) { + return (first_ & (1 << value)) != 0; + } else if (remaining_ == nullptr) { + return false; + } else { + return remaining_->Contains(value); + } +} + +const uc32 DispatchTable::Config::kNoKey = unibrow::Utf8::kBadChar; + +void DispatchTable::AddRange(CharacterRange full_range, int value, Zone* zone) { + CharacterRange current = full_range; + if (tree()->is_empty()) { + // If this is the first range we just insert into the table. + ZoneSplayTree::Locator loc; + bool inserted = tree()->Insert(current.from(), &loc); + DCHECK(inserted); + USE(inserted); + loc.set_value( + Entry(current.from(), current.to(), empty()->Extend(value, zone))); + return; + } + // First see if there is a range to the left of this one that + // overlaps. + ZoneSplayTree::Locator loc; + if (tree()->FindGreatestLessThan(current.from(), &loc)) { + Entry* entry = &loc.value(); + // If we've found a range that overlaps with this one, and it + // starts strictly to the left of this one, we have to fix it + // because the following code only handles ranges that start on + // or after the start point of the range we're adding. + if (entry->from() < current.from() && entry->to() >= current.from()) { + // Snap the overlapping range in half around the start point of + // the range we're adding. + CharacterRange left = + CharacterRange::Range(entry->from(), current.from() - 1); + CharacterRange right = CharacterRange::Range(current.from(), entry->to()); + // The left part of the overlapping range doesn't overlap. + // Truncate the whole entry to be just the left part. + entry->set_to(left.to()); + // The right part is the one that overlaps. We add this part + // to the map and let the next step deal with merging it with + // the range we're adding. + ZoneSplayTree::Locator loc; + bool inserted = tree()->Insert(right.from(), &loc); + DCHECK(inserted); + USE(inserted); + loc.set_value(Entry(right.from(), right.to(), entry->out_set())); + } + } + while (current.is_valid()) { + if (tree()->FindLeastGreaterThan(current.from(), &loc) && + (loc.value().from() <= current.to()) && + (loc.value().to() >= current.from())) { + Entry* entry = &loc.value(); + // We have overlap. If there is space between the start point of + // the range we're adding and where the overlapping range starts + // then we have to add a range covering just that space. + if (current.from() < entry->from()) { + ZoneSplayTree::Locator ins; + bool inserted = tree()->Insert(current.from(), &ins); + DCHECK(inserted); + USE(inserted); + ins.set_value(Entry(current.from(), entry->from() - 1, + empty()->Extend(value, zone))); + current.set_from(entry->from()); + } + DCHECK_EQ(current.from(), entry->from()); + // If the overlapping range extends beyond the one we want to add + // we have to snap the right part off and add it separately. + if (entry->to() > current.to()) { + ZoneSplayTree::Locator ins; + bool inserted = tree()->Insert(current.to() + 1, &ins); + DCHECK(inserted); + USE(inserted); + ins.set_value(Entry(current.to() + 1, entry->to(), entry->out_set())); + entry->set_to(current.to()); + } + DCHECK(entry->to() <= current.to()); + // The overlapping range is now completely contained by the range + // we're adding so we can just update it and move the start point + // of the range we're adding just past it. + entry->AddValue(value, zone); + DCHECK(entry->to() + 1 > current.from()); + current.set_from(entry->to() + 1); + } else { + // There is no overlap so we can just add the range + ZoneSplayTree::Locator ins; + bool inserted = tree()->Insert(current.from(), &ins); + DCHECK(inserted); + USE(inserted); + ins.set_value( + Entry(current.from(), current.to(), empty()->Extend(value, zone))); + break; + } + } +} + +OutSet* DispatchTable::Get(uc32 value) { + ZoneSplayTree::Locator loc; + if (!tree()->FindGreatestLessThan(value, &loc)) return empty(); + Entry* entry = &loc.value(); + if (value <= entry->to()) + return entry->out_set(); + else + return empty(); +} + +// ------------------------------------------------------------------- +// Analysis + +void Analysis::EnsureAnalyzed(RegExpNode* that) { + StackLimitCheck check(isolate()); + if (check.HasOverflowed()) { + fail("Stack overflow"); + return; + } + if (that->info()->been_analyzed || that->info()->being_analyzed) return; + that->info()->being_analyzed = true; + that->Accept(this); + that->info()->being_analyzed = false; + that->info()->been_analyzed = true; +} + +void Analysis::VisitEnd(EndNode* that) { + // nothing to do +} + +void TextNode::CalculateOffsets() { + int element_count = elements()->length(); + // Set up the offsets of the elements relative to the start. This is a fixed + // quantity since a TextNode can only contain fixed-width things. + int cp_offset = 0; + for (int i = 0; i < element_count; i++) { + TextElement& elm = elements()->at(i); + elm.set_cp_offset(cp_offset); + cp_offset += elm.length(); + } +} + +void Analysis::VisitText(TextNode* that) { + that->MakeCaseIndependent(isolate(), is_one_byte_); + EnsureAnalyzed(that->on_success()); + if (!has_failed()) { + that->CalculateOffsets(); + } +} + +void Analysis::VisitAction(ActionNode* that) { + RegExpNode* target = that->on_success(); + EnsureAnalyzed(target); + if (!has_failed()) { + // If the next node is interested in what it follows then this node + // has to be interested too so it can pass the information on. + that->info()->AddFromFollowing(target->info()); + } +} + +void Analysis::VisitChoice(ChoiceNode* that) { + NodeInfo* info = that->info(); + for (int i = 0; i < that->alternatives()->length(); i++) { + RegExpNode* node = that->alternatives()->at(i).node(); + EnsureAnalyzed(node); + if (has_failed()) return; + // Anything the following nodes need to know has to be known by + // this node also, so it can pass it on. + info->AddFromFollowing(node->info()); + } +} + +void Analysis::VisitLoopChoice(LoopChoiceNode* that) { + NodeInfo* info = that->info(); + for (int i = 0; i < that->alternatives()->length(); i++) { + RegExpNode* node = that->alternatives()->at(i).node(); + if (node != that->loop_node()) { + EnsureAnalyzed(node); + if (has_failed()) return; + info->AddFromFollowing(node->info()); + } + } + // Check the loop last since it may need the value of this node + // to get a correct result. + EnsureAnalyzed(that->loop_node()); + if (!has_failed()) { + info->AddFromFollowing(that->loop_node()->info()); + } +} + +void Analysis::VisitBackReference(BackReferenceNode* that) { + EnsureAnalyzed(that->on_success()); +} + +void Analysis::VisitAssertion(AssertionNode* that) { + EnsureAnalyzed(that->on_success()); +} + +void BackReferenceNode::FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, + bool not_at_start) { + // Working out the set of characters that a backreference can match is too + // hard, so we just say that any character can match. + bm->SetRest(offset); + SaveBMInfo(bm, not_at_start, offset); +} + +STATIC_ASSERT(BoyerMoorePositionInfo::kMapSize == + RegExpMacroAssembler::kTableSize); + +void ChoiceNode::FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) { + ZoneList* alts = alternatives(); + budget = (budget - 1) / alts->length(); + for (int i = 0; i < alts->length(); i++) { + GuardedAlternative& alt = alts->at(i); + if (alt.guards() != nullptr && alt.guards()->length() != 0) { + bm->SetRest(offset); // Give up trying to fill in info. + SaveBMInfo(bm, not_at_start, offset); + return; + } + alt.node()->FillInBMInfo(isolate, offset, budget, bm, not_at_start); + } + SaveBMInfo(bm, not_at_start, offset); +} + +void TextNode::FillInBMInfo(Isolate* isolate, int initial_offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) { + if (initial_offset >= bm->length()) return; + int offset = initial_offset; + int max_char = bm->max_char(); + for (int i = 0; i < elements()->length(); i++) { + if (offset >= bm->length()) { + if (initial_offset == 0) set_bm_info(not_at_start, bm); + return; + } + TextElement text = elements()->at(i); + if (text.text_type() == TextElement::ATOM) { + RegExpAtom* atom = text.atom(); + for (int j = 0; j < atom->length(); j++, offset++) { + if (offset >= bm->length()) { + if (initial_offset == 0) set_bm_info(not_at_start, bm); + return; + } + uc16 character = atom->data()[j]; + if (IgnoreCase(atom->flags())) { + unibrow::uchar chars[4]; + int length = GetCaseIndependentLetters( + isolate, character, bm->max_char() == String::kMaxOneByteCharCode, + chars, 4); + for (int j = 0; j < length; j++) { + bm->Set(offset, chars[j]); + } + } else { + if (character <= max_char) bm->Set(offset, character); + } + } + } else { + DCHECK_EQ(TextElement::CHAR_CLASS, text.text_type()); + RegExpCharacterClass* char_class = text.char_class(); + ZoneList* ranges = char_class->ranges(zone()); + if (char_class->is_negated()) { + bm->SetAll(offset); + } else { + for (int k = 0; k < ranges->length(); k++) { + CharacterRange& range = ranges->at(k); + if (range.from() > max_char) continue; + int to = Min(max_char, static_cast(range.to())); + bm->SetInterval(offset, Interval(range.from(), to)); + } + } + offset++; + } + } + if (offset >= bm->length()) { + if (initial_offset == 0) set_bm_info(not_at_start, bm); + return; + } + on_success()->FillInBMInfo(isolate, offset, budget - 1, bm, + true); // Not at start after a text node. + if (initial_offset == 0) set_bm_info(not_at_start, bm); +} + +// ------------------------------------------------------------------- +// Dispatch table construction + +void DispatchTableConstructor::VisitEnd(EndNode* that) { + AddRange(CharacterRange::Everything()); +} + +void DispatchTableConstructor::BuildTable(ChoiceNode* node) { + node->set_being_calculated(true); + ZoneList* alternatives = node->alternatives(); + for (int i = 0; i < alternatives->length(); i++) { + set_choice_index(i); + alternatives->at(i).node()->Accept(this); + } + node->set_being_calculated(false); +} + +class AddDispatchRange { + public: + explicit AddDispatchRange(DispatchTableConstructor* constructor) + : constructor_(constructor) {} + void Call(uc32 from, DispatchTable::Entry entry); + + private: + DispatchTableConstructor* constructor_; +}; + +void AddDispatchRange::Call(uc32 from, DispatchTable::Entry entry) { + constructor_->AddRange(CharacterRange::Range(from, entry.to())); +} + +void DispatchTableConstructor::VisitChoice(ChoiceNode* node) { + if (node->being_calculated()) return; + DispatchTable* table = node->GetTable(ignore_case_); + AddDispatchRange adder(this); + table->ForEach(&adder); +} + +void DispatchTableConstructor::VisitBackReference(BackReferenceNode* that) { + // TODO(160): Find the node that we refer back to and propagate its start + // set back to here. For now we just accept anything. + AddRange(CharacterRange::Everything()); +} + +void DispatchTableConstructor::VisitAssertion(AssertionNode* that) { + RegExpNode* target = that->on_success(); + target->Accept(this); +} + +static int CompareRangeByFrom(const CharacterRange* a, + const CharacterRange* b) { + return Compare(a->from(), b->from()); +} + +void DispatchTableConstructor::AddInverse(ZoneList* ranges) { + ranges->Sort(CompareRangeByFrom); + uc16 last = 0; + for (int i = 0; i < ranges->length(); i++) { + CharacterRange range = ranges->at(i); + if (last < range.from()) + AddRange(CharacterRange::Range(last, range.from() - 1)); + if (range.to() >= last) { + if (range.to() == String::kMaxCodePoint) { + return; + } else { + last = range.to() + 1; + } + } + } + AddRange(CharacterRange::Range(last, String::kMaxCodePoint)); +} + +void DispatchTableConstructor::VisitText(TextNode* that) { + TextElement elm = that->elements()->at(0); + switch (elm.text_type()) { + case TextElement::ATOM: { + uc16 c = elm.atom()->data()[0]; + AddRange(CharacterRange::Range(c, c)); + break; + } + case TextElement::CHAR_CLASS: { + RegExpCharacterClass* tree = elm.char_class(); + ZoneList* ranges = tree->ranges(that->zone()); + if (tree->is_negated()) { + AddInverse(ranges); + } else { + for (int i = 0; i < ranges->length(); i++) AddRange(ranges->at(i)); + } + break; + } + default: { + UNIMPLEMENTED(); + } + } +} + +void DispatchTableConstructor::VisitAction(ActionNode* that) { + RegExpNode* target = that->on_success(); + target->Accept(this); +} + +// static +RegExpNode* RegExpCompiler::OptionallyStepBackToLeadSurrogate( + RegExpCompiler* compiler, RegExpNode* on_success, JSRegExp::Flags flags) { + DCHECK(!compiler->read_backward()); + Zone* zone = compiler->zone(); + ZoneList* lead_surrogates = CharacterRange::List( + zone, CharacterRange::Range(kLeadSurrogateStart, kLeadSurrogateEnd)); + ZoneList* trail_surrogates = CharacterRange::List( + zone, CharacterRange::Range(kTrailSurrogateStart, kTrailSurrogateEnd)); + + ChoiceNode* optional_step_back = new (zone) ChoiceNode(2, zone); + + int stack_register = compiler->UnicodeLookaroundStackRegister(); + int position_register = compiler->UnicodeLookaroundPositionRegister(); + RegExpNode* step_back = TextNode::CreateForCharacterRanges( + zone, lead_surrogates, true, on_success, flags); + RegExpLookaround::Builder builder(true, step_back, stack_register, + position_register); + RegExpNode* match_trail = TextNode::CreateForCharacterRanges( + zone, trail_surrogates, false, builder.on_match_success(), flags); + + optional_step_back->AddAlternative( + GuardedAlternative(builder.ForMatch(match_trail))); + optional_step_back->AddAlternative(GuardedAlternative(on_success)); + + return optional_step_back; +} + +} // namespace internal +} // namespace v8 diff --git a/src/regexp/regexp-compiler.h b/src/regexp/regexp-compiler.h index e52c488c7f..31d640e7f1 100644 --- a/src/regexp/regexp-compiler.h +++ b/src/regexp/regexp-compiler.h @@ -5,8 +5,8 @@ #ifndef V8_REGEXP_REGEXP_COMPILER_H_ #define V8_REGEXP_REGEXP_COMPILER_H_ -#include "src/regexp/jsregexp.h" // TODO(jgruber): Remove if possible. -#include "src/regexp/regexp-macro-assembler-arch.h" +#include "src/regexp/regexp-nodes.h" +#include "src/zone/zone-splay-tree.h" namespace v8 { namespace internal { @@ -37,8 +37,530 @@ constexpr int kLineTerminatorRanges[] = {0x000A, 0x000B, 0x000D, 0x000E, 0x2028, 0x202A, kRangeEndMarker}; constexpr int kLineTerminatorRangeCount = arraysize(kLineTerminatorRanges); +// More makes code generation slower, less makes V8 benchmark score lower. +constexpr int kMaxLookaheadForBoyerMoore = 8; +// In a 3-character pattern you can maximally step forwards 3 characters +// at a time, which is not always enough to pay for the extra logic. +constexpr int kPatternTooShortForBoyerMoore = 2; + } // namespace regexp_compiler_constants +// A set of unsigned integers that behaves especially well on small +// integers (< 32). May do zone-allocation. +class OutSet : public ZoneObject { + public: + OutSet() : first_(0), remaining_(nullptr), successors_(nullptr) {} + OutSet* Extend(unsigned value, Zone* zone); + V8_EXPORT_PRIVATE bool Get(unsigned value) const; + static const unsigned kFirstLimit = 32; + + private: + // Destructively set a value in this set. In most cases you want + // to use Extend instead to ensure that only one instance exists + // that contains the same values. + void Set(unsigned value, Zone* zone); + + // The successors are a list of sets that contain the same values + // as this set and the one more value that is not present in this + // set. + ZoneList* successors(Zone* zone) { return successors_; } + + OutSet(uint32_t first, ZoneList* remaining) + : first_(first), remaining_(remaining), successors_(nullptr) {} + uint32_t first_; + ZoneList* remaining_; + ZoneList* successors_; + friend class Trace; +}; + +// A mapping from integers, specified as ranges, to a set of integers. +// Used for mapping character ranges to choices. +class DispatchTable : public ZoneObject { + public: + explicit DispatchTable(Zone* zone) : tree_(zone) {} + + class Entry { + public: + Entry() : from_(0), to_(0), out_set_(nullptr) {} + Entry(uc32 from, uc32 to, OutSet* out_set) + : from_(from), to_(to), out_set_(out_set) { + DCHECK(from <= to); + } + uc32 from() { return from_; } + uc32 to() { return to_; } + void set_to(uc32 value) { to_ = value; } + void AddValue(int value, Zone* zone) { + out_set_ = out_set_->Extend(value, zone); + } + OutSet* out_set() { return out_set_; } + + private: + uc32 from_; + uc32 to_; + OutSet* out_set_; + }; + + class Config { + public: + using Key = uc32; + using Value = Entry; + static const uc32 kNoKey; + static const Entry NoValue() { return Value(); } + static inline int Compare(uc32 a, uc32 b) { + if (a == b) + return 0; + else if (a < b) + return -1; + else + return 1; + } + }; + + V8_EXPORT_PRIVATE void AddRange(CharacterRange range, int value, Zone* zone); + V8_EXPORT_PRIVATE OutSet* Get(uc32 value); + void Dump(); + + template + void ForEach(Callback* callback) { + return tree()->ForEach(callback); + } + + private: + // There can't be a static empty set since it allocates its + // successors in a zone and caches them. + OutSet* empty() { return &empty_; } + OutSet empty_; + ZoneSplayTree* tree() { return &tree_; } + ZoneSplayTree tree_; +}; + +// Node visitor used to add the start set of the alternatives to the +// dispatch table of a choice node. +class V8_EXPORT_PRIVATE DispatchTableConstructor : public NodeVisitor { + public: + DispatchTableConstructor(DispatchTable* table, bool ignore_case, Zone* zone) + : table_(table), + choice_index_(-1), + ignore_case_(ignore_case), + zone_(zone) {} + + void BuildTable(ChoiceNode* node); + + void AddRange(CharacterRange range) { + table()->AddRange(range, choice_index_, zone_); + } + + void AddInverse(ZoneList* ranges); + +#define DECLARE_VISIT(Type) virtual void Visit##Type(Type##Node* that); + FOR_EACH_NODE_TYPE(DECLARE_VISIT) +#undef DECLARE_VISIT + + DispatchTable* table() { return table_; } + void set_choice_index(int value) { choice_index_ = value; } + + protected: + DispatchTable* table_; + int choice_index_; + bool ignore_case_; + Zone* zone_; +}; + +// Details of a quick mask-compare check that can look ahead in the +// input stream. +class QuickCheckDetails { + public: + QuickCheckDetails() + : characters_(0), mask_(0), value_(0), cannot_match_(false) {} + explicit QuickCheckDetails(int characters) + : characters_(characters), mask_(0), value_(0), cannot_match_(false) {} + bool Rationalize(bool one_byte); + // Merge in the information from another branch of an alternation. + void Merge(QuickCheckDetails* other, int from_index); + // Advance the current position by some amount. + void Advance(int by, bool one_byte); + void Clear(); + bool cannot_match() { return cannot_match_; } + void set_cannot_match() { cannot_match_ = true; } + struct Position { + Position() : mask(0), value(0), determines_perfectly(false) {} + uc16 mask; + uc16 value; + bool determines_perfectly; + }; + int characters() { return characters_; } + void set_characters(int characters) { characters_ = characters; } + Position* positions(int index) { + DCHECK_LE(0, index); + DCHECK_GT(characters_, index); + return positions_ + index; + } + uint32_t mask() { return mask_; } + uint32_t value() { return value_; } + + private: + // How many characters do we have quick check information from. This is + // the same for all branches of a choice node. + int characters_; + Position positions_[4]; + // These values are the condensate of the above array after Rationalize(). + uint32_t mask_; + uint32_t value_; + // If set to true, there is no way this quick check can match at all. + // E.g., if it requires to be at the start of the input, and isn't. + bool cannot_match_; +}; + +// Improve the speed that we scan for an initial point where a non-anchored +// regexp can match by using a Boyer-Moore-like table. This is done by +// identifying non-greedy non-capturing loops in the nodes that eat any +// character one at a time. For example in the middle of the regexp +// /foo[\s\S]*?bar/ we find such a loop. There is also such a loop implicitly +// inserted at the start of any non-anchored regexp. +// +// When we have found such a loop we look ahead in the nodes to find the set of +// characters that can come at given distances. For example for the regexp +// /.?foo/ we know that there are at least 3 characters ahead of us, and the +// sets of characters that can occur are [any, [f, o], [o]]. We find a range in +// the lookahead info where the set of characters is reasonably constrained. In +// our example this is from index 1 to 2 (0 is not constrained). We can now +// look 3 characters ahead and if we don't find one of [f, o] (the union of +// [f, o] and [o]) then we can skip forwards by the range size (in this case 2). +// +// For Unicode input strings we do the same, but modulo 128. +// +// We also look at the first string fed to the regexp and use that to get a hint +// of the character frequencies in the inputs. This affects the assessment of +// whether the set of characters is 'reasonably constrained'. +// +// We also have another lookahead mechanism (called quick check in the code), +// which uses a wide load of multiple characters followed by a mask and compare +// to determine whether a match is possible at this point. +enum ContainedInLattice { + kNotYet = 0, + kLatticeIn = 1, + kLatticeOut = 2, + kLatticeUnknown = 3 // Can also mean both in and out. +}; + +inline ContainedInLattice Combine(ContainedInLattice a, ContainedInLattice b) { + return static_cast(a | b); +} + +ContainedInLattice AddRange(ContainedInLattice a, const int* ranges, + int ranges_size, Interval new_range); + +class BoyerMoorePositionInfo : public ZoneObject { + public: + explicit BoyerMoorePositionInfo(Zone* zone) + : map_(new (zone) ZoneList(kMapSize, zone)), + map_count_(0), + w_(kNotYet), + s_(kNotYet), + d_(kNotYet), + surrogate_(kNotYet) { + for (int i = 0; i < kMapSize; i++) { + map_->Add(false, zone); + } + } + + bool& at(int i) { return map_->at(i); } + + static const int kMapSize = 128; + static const int kMask = kMapSize - 1; + + int map_count() const { return map_count_; } + + void Set(int character); + void SetInterval(const Interval& interval); + void SetAll(); + bool is_non_word() { return w_ == kLatticeOut; } + bool is_word() { return w_ == kLatticeIn; } + + private: + ZoneList* map_; + int map_count_; // Number of set bits in the map. + ContainedInLattice w_; // The \w character class. + ContainedInLattice s_; // The \s character class. + ContainedInLattice d_; // The \d character class. + ContainedInLattice surrogate_; // Surrogate UTF-16 code units. +}; + +class BoyerMooreLookahead : public ZoneObject { + public: + BoyerMooreLookahead(int length, RegExpCompiler* compiler, Zone* zone); + + int length() { return length_; } + int max_char() { return max_char_; } + RegExpCompiler* compiler() { return compiler_; } + + int Count(int map_number) { return bitmaps_->at(map_number)->map_count(); } + + BoyerMoorePositionInfo* at(int i) { return bitmaps_->at(i); } + + void Set(int map_number, int character) { + if (character > max_char_) return; + BoyerMoorePositionInfo* info = bitmaps_->at(map_number); + info->Set(character); + } + + void SetInterval(int map_number, const Interval& interval) { + if (interval.from() > max_char_) return; + BoyerMoorePositionInfo* info = bitmaps_->at(map_number); + if (interval.to() > max_char_) { + info->SetInterval(Interval(interval.from(), max_char_)); + } else { + info->SetInterval(interval); + } + } + + void SetAll(int map_number) { bitmaps_->at(map_number)->SetAll(); } + + void SetRest(int from_map) { + for (int i = from_map; i < length_; i++) SetAll(i); + } + void EmitSkipInstructions(RegExpMacroAssembler* masm); + + private: + // This is the value obtained by EatsAtLeast. If we do not have at least this + // many characters left in the sample string then the match is bound to fail. + // Therefore it is OK to read a character this far ahead of the current match + // point. + int length_; + RegExpCompiler* compiler_; + // 0xff for Latin1, 0xffff for UTF-16. + int max_char_; + ZoneList* bitmaps_; + + int GetSkipTable(int min_lookahead, int max_lookahead, + Handle boolean_skip_table); + bool FindWorthwhileInterval(int* from, int* to); + int FindBestInterval(int max_number_of_chars, int old_biggest_points, + int* from, int* to); +}; + +// There are many ways to generate code for a node. This class encapsulates +// the current way we should be generating. In other words it encapsulates +// the current state of the code generator. The effect of this is that we +// generate code for paths that the matcher can take through the regular +// expression. A given node in the regexp can be code-generated several times +// as it can be part of several traces. For example for the regexp: +// /foo(bar|ip)baz/ the code to match baz will be generated twice, once as part +// of the foo-bar-baz trace and once as part of the foo-ip-baz trace. The code +// to match foo is generated only once (the traces have a common prefix). The +// code to store the capture is deferred and generated (twice) after the places +// where baz has been matched. +class Trace { + public: + // A value for a property that is either known to be true, know to be false, + // or not known. + enum TriBool { UNKNOWN = -1, FALSE_VALUE = 0, TRUE_VALUE = 1 }; + + class DeferredAction { + public: + DeferredAction(ActionNode::ActionType action_type, int reg) + : action_type_(action_type), reg_(reg), next_(nullptr) {} + DeferredAction* next() { return next_; } + bool Mentions(int reg); + int reg() { return reg_; } + ActionNode::ActionType action_type() { return action_type_; } + + private: + ActionNode::ActionType action_type_; + int reg_; + DeferredAction* next_; + friend class Trace; + }; + + class DeferredCapture : public DeferredAction { + public: + DeferredCapture(int reg, bool is_capture, Trace* trace) + : DeferredAction(ActionNode::STORE_POSITION, reg), + cp_offset_(trace->cp_offset()), + is_capture_(is_capture) {} + int cp_offset() { return cp_offset_; } + bool is_capture() { return is_capture_; } + + private: + int cp_offset_; + bool is_capture_; + void set_cp_offset(int cp_offset) { cp_offset_ = cp_offset; } + }; + + class DeferredSetRegister : public DeferredAction { + public: + DeferredSetRegister(int reg, int value) + : DeferredAction(ActionNode::SET_REGISTER, reg), value_(value) {} + int value() { return value_; } + + private: + int value_; + }; + + class DeferredClearCaptures : public DeferredAction { + public: + explicit DeferredClearCaptures(Interval range) + : DeferredAction(ActionNode::CLEAR_CAPTURES, -1), range_(range) {} + Interval range() { return range_; } + + private: + Interval range_; + }; + + class DeferredIncrementRegister : public DeferredAction { + public: + explicit DeferredIncrementRegister(int reg) + : DeferredAction(ActionNode::INCREMENT_REGISTER, reg) {} + }; + + Trace() + : cp_offset_(0), + actions_(nullptr), + backtrack_(nullptr), + stop_node_(nullptr), + loop_label_(nullptr), + characters_preloaded_(0), + bound_checked_up_to_(0), + flush_budget_(100), + at_start_(UNKNOWN) {} + + // End the trace. This involves flushing the deferred actions in the trace + // and pushing a backtrack location onto the backtrack stack. Once this is + // done we can start a new trace or go to one that has already been + // generated. + void Flush(RegExpCompiler* compiler, RegExpNode* successor); + int cp_offset() { return cp_offset_; } + DeferredAction* actions() { return actions_; } + // A trivial trace is one that has no deferred actions or other state that + // affects the assumptions used when generating code. There is no recorded + // backtrack location in a trivial trace, so with a trivial trace we will + // generate code that, on a failure to match, gets the backtrack location + // from the backtrack stack rather than using a direct jump instruction. We + // always start code generation with a trivial trace and non-trivial traces + // are created as we emit code for nodes or add to the list of deferred + // actions in the trace. The location of the code generated for a node using + // a trivial trace is recorded in a label in the node so that gotos can be + // generated to that code. + bool is_trivial() { + return backtrack_ == nullptr && actions_ == nullptr && cp_offset_ == 0 && + characters_preloaded_ == 0 && bound_checked_up_to_ == 0 && + quick_check_performed_.characters() == 0 && at_start_ == UNKNOWN; + } + TriBool at_start() { return at_start_; } + void set_at_start(TriBool at_start) { at_start_ = at_start; } + Label* backtrack() { return backtrack_; } + Label* loop_label() { return loop_label_; } + RegExpNode* stop_node() { return stop_node_; } + int characters_preloaded() { return characters_preloaded_; } + int bound_checked_up_to() { return bound_checked_up_to_; } + int flush_budget() { return flush_budget_; } + QuickCheckDetails* quick_check_performed() { return &quick_check_performed_; } + bool mentions_reg(int reg); + // Returns true if a deferred position store exists to the specified + // register and stores the offset in the out-parameter. Otherwise + // returns false. + bool GetStoredPosition(int reg, int* cp_offset); + // These set methods and AdvanceCurrentPositionInTrace should be used only on + // new traces - the intention is that traces are immutable after creation. + void add_action(DeferredAction* new_action) { + DCHECK(new_action->next_ == nullptr); + new_action->next_ = actions_; + actions_ = new_action; + } + void set_backtrack(Label* backtrack) { backtrack_ = backtrack; } + void set_stop_node(RegExpNode* node) { stop_node_ = node; } + void set_loop_label(Label* label) { loop_label_ = label; } + void set_characters_preloaded(int count) { characters_preloaded_ = count; } + void set_bound_checked_up_to(int to) { bound_checked_up_to_ = to; } + void set_flush_budget(int to) { flush_budget_ = to; } + void set_quick_check_performed(QuickCheckDetails* d) { + quick_check_performed_ = *d; + } + void InvalidateCurrentCharacter(); + void AdvanceCurrentPositionInTrace(int by, RegExpCompiler* compiler); + + private: + int FindAffectedRegisters(OutSet* affected_registers, Zone* zone); + void PerformDeferredActions(RegExpMacroAssembler* macro, int max_register, + const OutSet& affected_registers, + OutSet* registers_to_pop, + OutSet* registers_to_clear, Zone* zone); + void RestoreAffectedRegisters(RegExpMacroAssembler* macro, int max_register, + const OutSet& registers_to_pop, + const OutSet& registers_to_clear); + int cp_offset_; + DeferredAction* actions_; + Label* backtrack_; + RegExpNode* stop_node_; + Label* loop_label_; + int characters_preloaded_; + int bound_checked_up_to_; + QuickCheckDetails quick_check_performed_; + int flush_budget_; + TriBool at_start_; +}; + +class GreedyLoopState { + public: + explicit GreedyLoopState(bool not_at_start); + + Label* label() { return &label_; } + Trace* counter_backtrack_trace() { return &counter_backtrack_trace_; } + + private: + Label label_; + Trace counter_backtrack_trace_; +}; + +struct PreloadState { + static const int kEatsAtLeastNotYetInitialized = -1; + bool preload_is_current_; + bool preload_has_checked_bounds_; + int preload_characters_; + int eats_at_least_; + void init() { eats_at_least_ = kEatsAtLeastNotYetInitialized; } +}; + +// Assertion propagation moves information about assertions such as +// \b to the affected nodes. For instance, in /.\b./ information must +// be propagated to the first '.' that whatever follows needs to know +// if it matched a word or a non-word, and to the second '.' that it +// has to check if it succeeds a word or non-word. In this case the +// result will be something like: +// +// +-------+ +------------+ +// | . | | . | +// +-------+ ---> +------------+ +// | word? | | check word | +// +-------+ +------------+ +class Analysis : public NodeVisitor { + public: + Analysis(Isolate* isolate, bool is_one_byte) + : isolate_(isolate), is_one_byte_(is_one_byte), error_message_(nullptr) {} + void EnsureAnalyzed(RegExpNode* node); + +#define DECLARE_VISIT(Type) void Visit##Type(Type##Node* that) override; + FOR_EACH_NODE_TYPE(DECLARE_VISIT) +#undef DECLARE_VISIT + void VisitLoopChoice(LoopChoiceNode* that) override; + + bool has_failed() { return error_message_ != nullptr; } + const char* error_message() { + DCHECK(error_message_ != nullptr); + return error_message_; + } + void fail(const char* error_message) { error_message_ = error_message; } + + Isolate* isolate() const { return isolate_; } + + private: + Isolate* isolate_; + bool is_one_byte_; + const char* error_message_; + + DISALLOW_IMPLICIT_CONSTRUCTORS(Analysis); +}; + class FrequencyCollator { public: FrequencyCollator() : total_samples_(0) { @@ -113,10 +635,30 @@ class RegExpCompiler { return unicode_lookaround_position_register_; } - RegExpEngine::CompilationResult Assemble(Isolate* isolate, - RegExpMacroAssembler* assembler, - RegExpNode* start, int capture_count, - Handle pattern); + struct CompilationResult final { + explicit CompilationResult(const char* error_message) + : error_message(error_message) {} + CompilationResult(Object code, int registers) + : code(code), num_registers(registers) {} + + static CompilationResult RegExpTooBig() { + return CompilationResult("RegExp too big"); + } + + const char* const error_message = nullptr; + Object code; + int num_registers = 0; + }; + + CompilationResult Assemble(Isolate* isolate, RegExpMacroAssembler* assembler, + RegExpNode* start, int capture_count, + Handle pattern); + + // If the regexp matching starts within a surrogate pair, step back to the + // lead surrogate and start matching from there. + static RegExpNode* OptionallyStepBackToLeadSurrogate(RegExpCompiler* compiler, + RegExpNode* on_success, + JSRegExp::Flags flags); inline void AddWork(RegExpNode* node) { if (!node->on_work_list() && !node->label()->is_bound()) { diff --git a/src/regexp/regexp-dotprinter.cc b/src/regexp/regexp-dotprinter.cc new file mode 100644 index 0000000000..2026a46bc3 --- /dev/null +++ b/src/regexp/regexp-dotprinter.cc @@ -0,0 +1,339 @@ +// Copyright 2019 the V8 project authors. All rights reserved. +// Use of this source code is governed by a BSD-style license that can be +// found in the LICENSE file. + +#include "src/regexp/regexp-dotprinter.h" + +#include "src/regexp/regexp-compiler.h" +#include "src/utils/ostreams.h" +#include "src/utils/splay-tree-inl.h" + +namespace v8 { +namespace internal { + +// ------------------------------------------------------------------- +// Dot/dotty output + +#ifdef DEBUG + +class DotPrinterImpl : public NodeVisitor { + public: + DotPrinterImpl(std::ostream& os, bool ignore_case) // NOLINT + : os_(os), ignore_case_(ignore_case) {} + void PrintNode(const char* label, RegExpNode* node); + void Visit(RegExpNode* node); + void PrintAttributes(RegExpNode* from); + void PrintOnFailure(RegExpNode* from, RegExpNode* to); +#define DECLARE_VISIT(Type) virtual void Visit##Type(Type##Node* that); + FOR_EACH_NODE_TYPE(DECLARE_VISIT) +#undef DECLARE_VISIT + private: + std::ostream& os_; + bool ignore_case_; +}; + +void DotPrinterImpl::PrintNode(const char* label, RegExpNode* node) { + os_ << "digraph G {\n graph [label=\""; + for (int i = 0; label[i]; i++) { + switch (label[i]) { + case '\\': + os_ << "\\\\"; + break; + case '"': + os_ << "\""; + break; + default: + os_ << label[i]; + break; + } + } + os_ << "\"];\n"; + Visit(node); + os_ << "}" << std::endl; +} + +void DotPrinterImpl::Visit(RegExpNode* node) { + if (node->info()->visited) return; + node->info()->visited = true; + node->Accept(this); +} + +void DotPrinterImpl::PrintOnFailure(RegExpNode* from, RegExpNode* on_failure) { + os_ << " n" << from << " -> n" << on_failure << " [style=dotted];\n"; + Visit(on_failure); +} + +class TableEntryBodyPrinter { + public: + TableEntryBodyPrinter(std::ostream& os, ChoiceNode* choice) // NOLINT + : os_(os), choice_(choice) {} + void Call(uc16 from, DispatchTable::Entry entry) { + OutSet* out_set = entry.out_set(); + for (unsigned i = 0; i < OutSet::kFirstLimit; i++) { + if (out_set->Get(i)) { + os_ << " n" << choice() << ":s" << from << "o" << i << " -> n" + << choice()->alternatives()->at(i).node() << ";\n"; + } + } + } + + private: + ChoiceNode* choice() { return choice_; } + std::ostream& os_; + ChoiceNode* choice_; +}; + +class TableEntryHeaderPrinter { + public: + explicit TableEntryHeaderPrinter(std::ostream& os) // NOLINT + : first_(true), os_(os) {} + void Call(uc16 from, DispatchTable::Entry entry) { + if (first_) { + first_ = false; + } else { + os_ << "|"; + } + os_ << "{\\" << AsUC16(from) << "-\\" << AsUC16(entry.to()) << "|{"; + OutSet* out_set = entry.out_set(); + int priority = 0; + for (unsigned i = 0; i < OutSet::kFirstLimit; i++) { + if (out_set->Get(i)) { + if (priority > 0) os_ << "|"; + os_ << " " << priority; + priority++; + } + } + os_ << "}}"; + } + + private: + bool first_; + std::ostream& os_; +}; + +class AttributePrinter { + public: + explicit AttributePrinter(std::ostream& os) // NOLINT + : os_(os), first_(true) {} + void PrintSeparator() { + if (first_) { + first_ = false; + } else { + os_ << "|"; + } + } + void PrintBit(const char* name, bool value) { + if (!value) return; + PrintSeparator(); + os_ << "{" << name << "}"; + } + void PrintPositive(const char* name, int value) { + if (value < 0) return; + PrintSeparator(); + os_ << "{" << name << "|" << value << "}"; + } + + private: + std::ostream& os_; + bool first_; +}; + +void DotPrinterImpl::PrintAttributes(RegExpNode* that) { + os_ << " a" << that << " [shape=Mrecord, color=grey, fontcolor=grey, " + << "margin=0.1, fontsize=10, label=\"{"; + AttributePrinter printer(os_); + NodeInfo* info = that->info(); + printer.PrintBit("NI", info->follows_newline_interest); + printer.PrintBit("WI", info->follows_word_interest); + printer.PrintBit("SI", info->follows_start_interest); + Label* label = that->label(); + if (label->is_bound()) printer.PrintPositive("@", label->pos()); + os_ << "}\"];\n" + << " a" << that << " -> n" << that + << " [style=dashed, color=grey, arrowhead=none];\n"; +} + +static const bool kPrintDispatchTable = false; +void DotPrinterImpl::VisitChoice(ChoiceNode* that) { + if (kPrintDispatchTable) { + os_ << " n" << that << " [shape=Mrecord, label=\""; + TableEntryHeaderPrinter header_printer(os_); + that->GetTable(ignore_case_)->ForEach(&header_printer); + os_ << "\"]\n"; + PrintAttributes(that); + TableEntryBodyPrinter body_printer(os_, that); + that->GetTable(ignore_case_)->ForEach(&body_printer); + } else { + os_ << " n" << that << " [shape=Mrecord, label=\"?\"];\n"; + for (int i = 0; i < that->alternatives()->length(); i++) { + GuardedAlternative alt = that->alternatives()->at(i); + os_ << " n" << that << " -> n" << alt.node(); + } + } + for (int i = 0; i < that->alternatives()->length(); i++) { + GuardedAlternative alt = that->alternatives()->at(i); + alt.node()->Accept(this); + } +} + +void DotPrinterImpl::VisitText(TextNode* that) { + Zone* zone = that->zone(); + os_ << " n" << that << " [label=\""; + for (int i = 0; i < that->elements()->length(); i++) { + if (i > 0) os_ << " "; + TextElement elm = that->elements()->at(i); + switch (elm.text_type()) { + case TextElement::ATOM: { + Vector data = elm.atom()->data(); + for (int i = 0; i < data.length(); i++) { + os_ << static_cast(data[i]); + } + break; + } + case TextElement::CHAR_CLASS: { + RegExpCharacterClass* node = elm.char_class(); + os_ << "["; + if (node->is_negated()) os_ << "^"; + for (int j = 0; j < node->ranges(zone)->length(); j++) { + CharacterRange range = node->ranges(zone)->at(j); + os_ << AsUC16(range.from()) << "-" << AsUC16(range.to()); + } + os_ << "]"; + break; + } + default: + UNREACHABLE(); + } + } + os_ << "\", shape=box, peripheries=2];\n"; + PrintAttributes(that); + os_ << " n" << that << " -> n" << that->on_success() << ";\n"; + Visit(that->on_success()); +} + +void DotPrinterImpl::VisitBackReference(BackReferenceNode* that) { + os_ << " n" << that << " [label=\"$" << that->start_register() << "..$" + << that->end_register() << "\", shape=doubleoctagon];\n"; + PrintAttributes(that); + os_ << " n" << that << " -> n" << that->on_success() << ";\n"; + Visit(that->on_success()); +} + +void DotPrinterImpl::VisitEnd(EndNode* that) { + os_ << " n" << that << " [style=bold, shape=point];\n"; + PrintAttributes(that); +} + +void DotPrinterImpl::VisitAssertion(AssertionNode* that) { + os_ << " n" << that << " ["; + switch (that->assertion_type()) { + case AssertionNode::AT_END: + os_ << "label=\"$\", shape=septagon"; + break; + case AssertionNode::AT_START: + os_ << "label=\"^\", shape=septagon"; + break; + case AssertionNode::AT_BOUNDARY: + os_ << "label=\"\\b\", shape=septagon"; + break; + case AssertionNode::AT_NON_BOUNDARY: + os_ << "label=\"\\B\", shape=septagon"; + break; + case AssertionNode::AFTER_NEWLINE: + os_ << "label=\"(?<=\\n)\", shape=septagon"; + break; + } + os_ << "];\n"; + PrintAttributes(that); + RegExpNode* successor = that->on_success(); + os_ << " n" << that << " -> n" << successor << ";\n"; + Visit(successor); +} + +void DotPrinterImpl::VisitAction(ActionNode* that) { + os_ << " n" << that << " ["; + switch (that->action_type_) { + case ActionNode::SET_REGISTER: + os_ << "label=\"$" << that->data_.u_store_register.reg + << ":=" << that->data_.u_store_register.value << "\", shape=octagon"; + break; + case ActionNode::INCREMENT_REGISTER: + os_ << "label=\"$" << that->data_.u_increment_register.reg + << "++\", shape=octagon"; + break; + case ActionNode::STORE_POSITION: + os_ << "label=\"$" << that->data_.u_position_register.reg + << ":=$pos\", shape=octagon"; + break; + case ActionNode::BEGIN_SUBMATCH: + os_ << "label=\"$" << that->data_.u_submatch.current_position_register + << ":=$pos,begin\", shape=septagon"; + break; + case ActionNode::POSITIVE_SUBMATCH_SUCCESS: + os_ << "label=\"escape\", shape=septagon"; + break; + case ActionNode::EMPTY_MATCH_CHECK: + os_ << "label=\"$" << that->data_.u_empty_match_check.start_register + << "=$pos?,$" << that->data_.u_empty_match_check.repetition_register + << "<" << that->data_.u_empty_match_check.repetition_limit + << "?\", shape=septagon"; + break; + case ActionNode::CLEAR_CAPTURES: { + os_ << "label=\"clear $" << that->data_.u_clear_captures.range_from + << " to $" << that->data_.u_clear_captures.range_to + << "\", shape=septagon"; + break; + } + } + os_ << "];\n"; + PrintAttributes(that); + RegExpNode* successor = that->on_success(); + os_ << " n" << that << " -> n" << successor << ";\n"; + Visit(successor); +} + +class DispatchTableDumper { + public: + explicit DispatchTableDumper(std::ostream& os) : os_(os) {} + void Call(uc16 key, DispatchTable::Entry entry); + + private: + std::ostream& os_; +}; + +void DispatchTableDumper::Call(uc16 key, DispatchTable::Entry entry) { + os_ << "[" << AsUC16(key) << "-" << AsUC16(entry.to()) << "]: {"; + OutSet* set = entry.out_set(); + bool first = true; + for (unsigned i = 0; i < OutSet::kFirstLimit; i++) { + if (set->Get(i)) { + if (first) { + first = false; + } else { + os_ << ", "; + } + os_ << i; + } + } + os_ << "}\n"; +} + +void DispatchTable::Dump() { + OFStream os(stderr); + DispatchTableDumper dumper(os); + tree()->ForEach(&dumper); +} + +#endif // DEBUG + +void DotPrinter::DotPrint(const char* label, RegExpNode* node, + bool ignore_case) { +#ifdef DEBUG + StdoutStream os; + DotPrinterImpl printer(os, ignore_case); + printer.PrintNode(label, node); +#endif // DEBUG +} + +} // namespace internal +} // namespace v8 diff --git a/src/regexp/regexp-dotprinter.h b/src/regexp/regexp-dotprinter.h new file mode 100644 index 0000000000..6ec4aefd9b --- /dev/null +++ b/src/regexp/regexp-dotprinter.h @@ -0,0 +1,23 @@ +// Copyright 2019 the V8 project authors. All rights reserved. +// Use of this source code is governed by a BSD-style license that can be +// found in the LICENSE file. + +#ifndef V8_REGEXP_REGEXP_DOTPRINTER_H_ +#define V8_REGEXP_REGEXP_DOTPRINTER_H_ + +#include "src/common/globals.h" + +namespace v8 { +namespace internal { + +class RegExpNode; + +class DotPrinter final : public AllStatic { + public: + static void DotPrint(const char* label, RegExpNode* node, bool ignore_case); +}; + +} // namespace internal +} // namespace v8 + +#endif // V8_REGEXP_REGEXP_DOTPRINTER_H_ diff --git a/src/regexp/regexp-nodes.h b/src/regexp/regexp-nodes.h new file mode 100644 index 0000000000..dd55189776 --- /dev/null +++ b/src/regexp/regexp-nodes.h @@ -0,0 +1,678 @@ +// Copyright 2019 the V8 project authors. All rights reserved. +// Use of this source code is governed by a BSD-style license that can be +// found in the LICENSE file. + +#ifndef V8_REGEXP_REGEXP_NODES_H_ +#define V8_REGEXP_REGEXP_NODES_H_ + +#include "src/regexp/regexp-macro-assembler.h" +#include "src/zone/zone.h" + +namespace v8 { +namespace internal { + +class AlternativeGenerationList; +class BoyerMooreLookahead; +class DispatchTable; +class GreedyLoopState; +class Label; +class NodeVisitor; +class QuickCheckDetails; +class RegExpCompiler; +class Trace; +struct PreloadState; + +#define FOR_EACH_NODE_TYPE(VISIT) \ + VISIT(End) \ + VISIT(Action) \ + VISIT(Choice) \ + VISIT(BackReference) \ + VISIT(Assertion) \ + VISIT(Text) + +struct NodeInfo final { + NodeInfo() + : being_analyzed(false), + been_analyzed(false), + follows_word_interest(false), + follows_newline_interest(false), + follows_start_interest(false), + at_end(false), + visited(false), + replacement_calculated(false) {} + + // Returns true if the interests and assumptions of this node + // matches the given one. + bool Matches(NodeInfo* that) { + return (at_end == that->at_end) && + (follows_word_interest == that->follows_word_interest) && + (follows_newline_interest == that->follows_newline_interest) && + (follows_start_interest == that->follows_start_interest); + } + + // Updates the interests of this node given the interests of the + // node preceding it. + void AddFromPreceding(NodeInfo* that) { + at_end |= that->at_end; + follows_word_interest |= that->follows_word_interest; + follows_newline_interest |= that->follows_newline_interest; + follows_start_interest |= that->follows_start_interest; + } + + bool HasLookbehind() { + return follows_word_interest || follows_newline_interest || + follows_start_interest; + } + + // Sets the interests of this node to include the interests of the + // following node. + void AddFromFollowing(NodeInfo* that) { + follows_word_interest |= that->follows_word_interest; + follows_newline_interest |= that->follows_newline_interest; + follows_start_interest |= that->follows_start_interest; + } + + void ResetCompilationState() { + being_analyzed = false; + been_analyzed = false; + } + + bool being_analyzed : 1; + bool been_analyzed : 1; + + // These bits are set of this node has to know what the preceding + // character was. + bool follows_word_interest : 1; + bool follows_newline_interest : 1; + bool follows_start_interest : 1; + + bool at_end : 1; + bool visited : 1; + bool replacement_calculated : 1; +}; + +class RegExpNode : public ZoneObject { + public: + explicit RegExpNode(Zone* zone) + : replacement_(nullptr), + on_work_list_(false), + trace_count_(0), + zone_(zone) { + bm_info_[0] = bm_info_[1] = nullptr; + } + virtual ~RegExpNode(); + virtual void Accept(NodeVisitor* visitor) = 0; + // Generates a goto to this node or actually generates the code at this point. + virtual void Emit(RegExpCompiler* compiler, Trace* trace) = 0; + // How many characters must this node consume at a minimum in order to + // succeed. If we have found at least 'still_to_find' characters that + // must be consumed there is no need to ask any following nodes whether + // they are sure to eat any more characters. The not_at_start argument is + // used to indicate that we know we are not at the start of the input. In + // this case anchored branches will always fail and can be ignored when + // determining how many characters are consumed on success. + virtual int EatsAtLeast(int still_to_find, int budget, bool not_at_start) = 0; + // Emits some quick code that checks whether the preloaded characters match. + // Falls through on certain failure, jumps to the label on possible success. + // If the node cannot make a quick check it does nothing and returns false. + bool EmitQuickCheck(RegExpCompiler* compiler, Trace* bounds_check_trace, + Trace* trace, bool preload_has_checked_bounds, + Label* on_possible_success, + QuickCheckDetails* details_return, + bool fall_through_on_failure); + // For a given number of characters this returns a mask and a value. The + // next n characters are anded with the mask and compared with the value. + // A comparison failure indicates the node cannot match the next n characters. + // A comparison success indicates the node may match. + virtual void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, + int characters_filled_in, + bool not_at_start) = 0; + static const int kNodeIsTooComplexForGreedyLoops = kMinInt; + virtual int GreedyLoopTextLength() { return kNodeIsTooComplexForGreedyLoops; } + // Only returns the successor for a text node of length 1 that matches any + // character and that has no guards on it. + virtual RegExpNode* GetSuccessorOfOmnivorousTextNode( + RegExpCompiler* compiler) { + return nullptr; + } + + // Collects information on the possible code units (mod 128) that can match if + // we look forward. This is used for a Boyer-Moore-like string searching + // implementation. TODO(erikcorry): This should share more code with + // EatsAtLeast, GetQuickCheckDetails. The budget argument is used to limit + // the number of nodes we are willing to look at in order to create this data. + static const int kRecursionBudget = 200; + bool KeepRecursing(RegExpCompiler* compiler); + virtual void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) { + UNREACHABLE(); + } + + // If we know that the input is one-byte then there are some nodes that can + // never match. This method returns a node that can be substituted for + // itself, or nullptr if the node can never match. + virtual RegExpNode* FilterOneByte(int depth) { return this; } + // Helper for FilterOneByte. + RegExpNode* replacement() { + DCHECK(info()->replacement_calculated); + return replacement_; + } + RegExpNode* set_replacement(RegExpNode* replacement) { + info()->replacement_calculated = true; + replacement_ = replacement; + return replacement; // For convenience. + } + + // We want to avoid recalculating the lookahead info, so we store it on the + // node. Only info that is for this node is stored. We can tell that the + // info is for this node when offset == 0, so the information is calculated + // relative to this node. + void SaveBMInfo(BoyerMooreLookahead* bm, bool not_at_start, int offset) { + if (offset == 0) set_bm_info(not_at_start, bm); + } + + Label* label() { return &label_; } + // If non-generic code is generated for a node (i.e. the node is not at the + // start of the trace) then it cannot be reused. This variable sets a limit + // on how often we allow that to happen before we insist on starting a new + // trace and generating generic code for a node that can be reused by flushing + // the deferred actions in the current trace and generating a goto. + static const int kMaxCopiesCodeGenerated = 10; + + bool on_work_list() { return on_work_list_; } + void set_on_work_list(bool value) { on_work_list_ = value; } + + NodeInfo* info() { return &info_; } + + BoyerMooreLookahead* bm_info(bool not_at_start) { + return bm_info_[not_at_start ? 1 : 0]; + } + + Zone* zone() const { return zone_; } + + protected: + enum LimitResult { DONE, CONTINUE }; + RegExpNode* replacement_; + + LimitResult LimitVersions(RegExpCompiler* compiler, Trace* trace); + + void set_bm_info(bool not_at_start, BoyerMooreLookahead* bm) { + bm_info_[not_at_start ? 1 : 0] = bm; + } + + private: + static const int kFirstCharBudget = 10; + Label label_; + bool on_work_list_; + NodeInfo info_; + // This variable keeps track of how many times code has been generated for + // this node (in different traces). We don't keep track of where the + // generated code is located unless the code is generated at the start of + // a trace, in which case it is generic and can be reused by flushing the + // deferred operations in the current trace and generating a goto. + int trace_count_; + BoyerMooreLookahead* bm_info_[2]; + + Zone* zone_; +}; + +class SeqRegExpNode : public RegExpNode { + public: + explicit SeqRegExpNode(RegExpNode* on_success) + : RegExpNode(on_success->zone()), on_success_(on_success) {} + RegExpNode* on_success() { return on_success_; } + void set_on_success(RegExpNode* node) { on_success_ = node; } + RegExpNode* FilterOneByte(int depth) override; + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override { + on_success_->FillInBMInfo(isolate, offset, budget - 1, bm, not_at_start); + if (offset == 0) set_bm_info(not_at_start, bm); + } + + protected: + RegExpNode* FilterSuccessor(int depth); + + private: + RegExpNode* on_success_; +}; + +class ActionNode : public SeqRegExpNode { + public: + enum ActionType { + SET_REGISTER, + INCREMENT_REGISTER, + STORE_POSITION, + BEGIN_SUBMATCH, + POSITIVE_SUBMATCH_SUCCESS, + EMPTY_MATCH_CHECK, + CLEAR_CAPTURES + }; + static ActionNode* SetRegister(int reg, int val, RegExpNode* on_success); + static ActionNode* IncrementRegister(int reg, RegExpNode* on_success); + static ActionNode* StorePosition(int reg, bool is_capture, + RegExpNode* on_success); + static ActionNode* ClearCaptures(Interval range, RegExpNode* on_success); + static ActionNode* BeginSubmatch(int stack_pointer_reg, int position_reg, + RegExpNode* on_success); + static ActionNode* PositiveSubmatchSuccess(int stack_pointer_reg, + int restore_reg, + int clear_capture_count, + int clear_capture_from, + RegExpNode* on_success); + static ActionNode* EmptyMatchCheck(int start_register, + int repetition_register, + int repetition_limit, + RegExpNode* on_success); + void Accept(NodeVisitor* visitor) override; + void Emit(RegExpCompiler* compiler, Trace* trace) override; + int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; + void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, int filled_in, + bool not_at_start) override { + return on_success()->GetQuickCheckDetails(details, compiler, filled_in, + not_at_start); + } + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override; + ActionType action_type() { return action_type_; } + // TODO(erikcorry): We should allow some action nodes in greedy loops. + int GreedyLoopTextLength() override { + return kNodeIsTooComplexForGreedyLoops; + } + + private: + union { + struct { + int reg; + int value; + } u_store_register; + struct { + int reg; + } u_increment_register; + struct { + int reg; + bool is_capture; + } u_position_register; + struct { + int stack_pointer_register; + int current_position_register; + int clear_register_count; + int clear_register_from; + } u_submatch; + struct { + int start_register; + int repetition_register; + int repetition_limit; + } u_empty_match_check; + struct { + int range_from; + int range_to; + } u_clear_captures; + } data_; + ActionNode(ActionType action_type, RegExpNode* on_success) + : SeqRegExpNode(on_success), action_type_(action_type) {} + ActionType action_type_; + friend class DotPrinterImpl; +}; + +class TextNode : public SeqRegExpNode { + public: + TextNode(ZoneList* elms, bool read_backward, + RegExpNode* on_success) + : SeqRegExpNode(on_success), elms_(elms), read_backward_(read_backward) {} + TextNode(RegExpCharacterClass* that, bool read_backward, + RegExpNode* on_success) + : SeqRegExpNode(on_success), + elms_(new (zone()) ZoneList(1, zone())), + read_backward_(read_backward) { + elms_->Add(TextElement::CharClass(that), zone()); + } + // Create TextNode for a single character class for the given ranges. + static TextNode* CreateForCharacterRanges(Zone* zone, + ZoneList* ranges, + bool read_backward, + RegExpNode* on_success, + JSRegExp::Flags flags); + // Create TextNode for a surrogate pair with a range given for the + // lead and the trail surrogate each. + static TextNode* CreateForSurrogatePair(Zone* zone, CharacterRange lead, + CharacterRange trail, + bool read_backward, + RegExpNode* on_success, + JSRegExp::Flags flags); + void Accept(NodeVisitor* visitor) override; + void Emit(RegExpCompiler* compiler, Trace* trace) override; + int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; + void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, int characters_filled_in, + bool not_at_start) override; + ZoneList* elements() { return elms_; } + bool read_backward() { return read_backward_; } + void MakeCaseIndependent(Isolate* isolate, bool is_one_byte); + int GreedyLoopTextLength() override; + RegExpNode* GetSuccessorOfOmnivorousTextNode( + RegExpCompiler* compiler) override; + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override; + void CalculateOffsets(); + RegExpNode* FilterOneByte(int depth) override; + + private: + enum TextEmitPassType { + NON_LATIN1_MATCH, // Check for characters that can't match. + SIMPLE_CHARACTER_MATCH, // Case-dependent single character check. + NON_LETTER_CHARACTER_MATCH, // Check characters that have no case equivs. + CASE_CHARACTER_MATCH, // Case-independent single character check. + CHARACTER_CLASS_MATCH // Character class. + }; + static bool SkipPass(TextEmitPassType pass, bool ignore_case); + static const int kFirstRealPass = SIMPLE_CHARACTER_MATCH; + static const int kLastPass = CHARACTER_CLASS_MATCH; + void TextEmitPass(RegExpCompiler* compiler, TextEmitPassType pass, + bool preloaded, Trace* trace, bool first_element_checked, + int* checked_up_to); + int Length(); + ZoneList* elms_; + bool read_backward_; +}; + +class AssertionNode : public SeqRegExpNode { + public: + enum AssertionType { + AT_END, + AT_START, + AT_BOUNDARY, + AT_NON_BOUNDARY, + AFTER_NEWLINE + }; + static AssertionNode* AtEnd(RegExpNode* on_success) { + return new (on_success->zone()) AssertionNode(AT_END, on_success); + } + static AssertionNode* AtStart(RegExpNode* on_success) { + return new (on_success->zone()) AssertionNode(AT_START, on_success); + } + static AssertionNode* AtBoundary(RegExpNode* on_success) { + return new (on_success->zone()) AssertionNode(AT_BOUNDARY, on_success); + } + static AssertionNode* AtNonBoundary(RegExpNode* on_success) { + return new (on_success->zone()) AssertionNode(AT_NON_BOUNDARY, on_success); + } + static AssertionNode* AfterNewline(RegExpNode* on_success) { + return new (on_success->zone()) AssertionNode(AFTER_NEWLINE, on_success); + } + void Accept(NodeVisitor* visitor) override; + void Emit(RegExpCompiler* compiler, Trace* trace) override; + int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; + void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, int filled_in, + bool not_at_start) override; + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override; + AssertionType assertion_type() { return assertion_type_; } + + private: + void EmitBoundaryCheck(RegExpCompiler* compiler, Trace* trace); + enum IfPrevious { kIsNonWord, kIsWord }; + void BacktrackIfPrevious(RegExpCompiler* compiler, Trace* trace, + IfPrevious backtrack_if_previous); + AssertionNode(AssertionType t, RegExpNode* on_success) + : SeqRegExpNode(on_success), assertion_type_(t) {} + AssertionType assertion_type_; +}; + +class BackReferenceNode : public SeqRegExpNode { + public: + BackReferenceNode(int start_reg, int end_reg, JSRegExp::Flags flags, + bool read_backward, RegExpNode* on_success) + : SeqRegExpNode(on_success), + start_reg_(start_reg), + end_reg_(end_reg), + flags_(flags), + read_backward_(read_backward) {} + void Accept(NodeVisitor* visitor) override; + int start_register() { return start_reg_; } + int end_register() { return end_reg_; } + bool read_backward() { return read_backward_; } + void Emit(RegExpCompiler* compiler, Trace* trace) override; + int EatsAtLeast(int still_to_find, int recursion_depth, + bool not_at_start) override; + void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, int characters_filled_in, + bool not_at_start) override { + return; + } + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override; + + private: + int start_reg_; + int end_reg_; + JSRegExp::Flags flags_; + bool read_backward_; +}; + +class EndNode : public RegExpNode { + public: + enum Action { ACCEPT, BACKTRACK, NEGATIVE_SUBMATCH_SUCCESS }; + EndNode(Action action, Zone* zone) : RegExpNode(zone), action_(action) {} + void Accept(NodeVisitor* visitor) override; + void Emit(RegExpCompiler* compiler, Trace* trace) override; + int EatsAtLeast(int still_to_find, int recursion_depth, + bool not_at_start) override { + return 0; + } + void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, int characters_filled_in, + bool not_at_start) override { + // Returning 0 from EatsAtLeast should ensure we never get here. + UNREACHABLE(); + } + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override { + // Returning 0 from EatsAtLeast should ensure we never get here. + UNREACHABLE(); + } + + private: + Action action_; +}; + +class NegativeSubmatchSuccess : public EndNode { + public: + NegativeSubmatchSuccess(int stack_pointer_reg, int position_reg, + int clear_capture_count, int clear_capture_start, + Zone* zone) + : EndNode(NEGATIVE_SUBMATCH_SUCCESS, zone), + stack_pointer_register_(stack_pointer_reg), + current_position_register_(position_reg), + clear_capture_count_(clear_capture_count), + clear_capture_start_(clear_capture_start) {} + void Emit(RegExpCompiler* compiler, Trace* trace) override; + + private: + int stack_pointer_register_; + int current_position_register_; + int clear_capture_count_; + int clear_capture_start_; +}; + +class Guard : public ZoneObject { + public: + enum Relation { LT, GEQ }; + Guard(int reg, Relation op, int value) : reg_(reg), op_(op), value_(value) {} + int reg() { return reg_; } + Relation op() { return op_; } + int value() { return value_; } + + private: + int reg_; + Relation op_; + int value_; +}; + +class GuardedAlternative { + public: + explicit GuardedAlternative(RegExpNode* node) + : node_(node), guards_(nullptr) {} + void AddGuard(Guard* guard, Zone* zone); + RegExpNode* node() { return node_; } + void set_node(RegExpNode* node) { node_ = node; } + ZoneList* guards() { return guards_; } + + private: + RegExpNode* node_; + ZoneList* guards_; +}; + +class AlternativeGeneration; + +class ChoiceNode : public RegExpNode { + public: + explicit ChoiceNode(int expected_size, Zone* zone) + : RegExpNode(zone), + alternatives_(new (zone) + ZoneList(expected_size, zone)), + table_(nullptr), + not_at_start_(false), + being_calculated_(false) {} + void Accept(NodeVisitor* visitor) override; + void AddAlternative(GuardedAlternative node) { + alternatives()->Add(node, zone()); + } + ZoneList* alternatives() { return alternatives_; } + DispatchTable* GetTable(bool ignore_case); + void Emit(RegExpCompiler* compiler, Trace* trace) override; + int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; + int EatsAtLeastHelper(int still_to_find, int budget, + RegExpNode* ignore_this_node, bool not_at_start); + void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, int characters_filled_in, + bool not_at_start) override; + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override; + + bool being_calculated() { return being_calculated_; } + bool not_at_start() { return not_at_start_; } + void set_not_at_start() { not_at_start_ = true; } + void set_being_calculated(bool b) { being_calculated_ = b; } + virtual bool try_to_emit_quick_check_for_alternative(bool is_first) { + return true; + } + RegExpNode* FilterOneByte(int depth) override; + virtual bool read_backward() { return false; } + + protected: + int GreedyLoopTextLengthForAlternative(GuardedAlternative* alternative); + ZoneList* alternatives_; + + private: + friend class DispatchTableConstructor; + friend class Analysis; + void GenerateGuard(RegExpMacroAssembler* macro_assembler, Guard* guard, + Trace* trace); + int CalculatePreloadCharacters(RegExpCompiler* compiler, int eats_at_least); + void EmitOutOfLineContinuation(RegExpCompiler* compiler, Trace* trace, + GuardedAlternative alternative, + AlternativeGeneration* alt_gen, + int preload_characters, + bool next_expects_preload); + void SetUpPreLoad(RegExpCompiler* compiler, Trace* current_trace, + PreloadState* preloads); + void AssertGuardsMentionRegisters(Trace* trace); + int EmitOptimizedUnanchoredSearch(RegExpCompiler* compiler, Trace* trace); + Trace* EmitGreedyLoop(RegExpCompiler* compiler, Trace* trace, + AlternativeGenerationList* alt_gens, + PreloadState* preloads, + GreedyLoopState* greedy_loop_state, int text_length); + void EmitChoices(RegExpCompiler* compiler, + AlternativeGenerationList* alt_gens, int first_choice, + Trace* trace, PreloadState* preloads); + DispatchTable* table_; + // If true, this node is never checked at the start of the input. + // Allows a new trace to start with at_start() set to false. + bool not_at_start_; + bool being_calculated_; +}; + +class NegativeLookaroundChoiceNode : public ChoiceNode { + public: + explicit NegativeLookaroundChoiceNode(GuardedAlternative this_must_fail, + GuardedAlternative then_do_this, + Zone* zone) + : ChoiceNode(2, zone) { + AddAlternative(this_must_fail); + AddAlternative(then_do_this); + } + int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; + void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, int characters_filled_in, + bool not_at_start) override; + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override { + alternatives_->at(1).node()->FillInBMInfo(isolate, offset, budget - 1, bm, + not_at_start); + if (offset == 0) set_bm_info(not_at_start, bm); + } + // For a negative lookahead we don't emit the quick check for the + // alternative that is expected to fail. This is because quick check code + // starts by loading enough characters for the alternative that takes fewest + // characters, but on a negative lookahead the negative branch did not take + // part in that calculation (EatsAtLeast) so the assumptions don't hold. + bool try_to_emit_quick_check_for_alternative(bool is_first) override { + return !is_first; + } + RegExpNode* FilterOneByte(int depth) override; +}; + +class LoopChoiceNode : public ChoiceNode { + public: + LoopChoiceNode(bool body_can_be_zero_length, bool read_backward, Zone* zone) + : ChoiceNode(2, zone), + loop_node_(nullptr), + continue_node_(nullptr), + body_can_be_zero_length_(body_can_be_zero_length), + read_backward_(read_backward) {} + void AddLoopAlternative(GuardedAlternative alt); + void AddContinueAlternative(GuardedAlternative alt); + void Emit(RegExpCompiler* compiler, Trace* trace) override; + int EatsAtLeast(int still_to_find, int budget, bool not_at_start) override; + void GetQuickCheckDetails(QuickCheckDetails* details, + RegExpCompiler* compiler, int characters_filled_in, + bool not_at_start) override; + void FillInBMInfo(Isolate* isolate, int offset, int budget, + BoyerMooreLookahead* bm, bool not_at_start) override; + RegExpNode* loop_node() { return loop_node_; } + RegExpNode* continue_node() { return continue_node_; } + bool body_can_be_zero_length() { return body_can_be_zero_length_; } + bool read_backward() override { return read_backward_; } + void Accept(NodeVisitor* visitor) override; + RegExpNode* FilterOneByte(int depth) override; + + private: + // AddAlternative is made private for loop nodes because alternatives + // should not be added freely, we need to keep track of which node + // goes back to the node itself. + void AddAlternative(GuardedAlternative node) { + ChoiceNode::AddAlternative(node); + } + + RegExpNode* loop_node_; + RegExpNode* continue_node_; + bool body_can_be_zero_length_; + bool read_backward_; +}; + +class NodeVisitor { + public: + virtual ~NodeVisitor() = default; +#define DECLARE_VISIT(Type) virtual void Visit##Type(Type##Node* that) = 0; + FOR_EACH_NODE_TYPE(DECLARE_VISIT) +#undef DECLARE_VISIT + virtual void VisitLoopChoice(LoopChoiceNode* that) { VisitChoice(that); } +}; + +} // namespace internal +} // namespace v8 + +#endif // V8_REGEXP_REGEXP_NODES_H_ diff --git a/src/regexp/regexp-parser.cc b/src/regexp/regexp-parser.cc index 7cae456f56..df1f6de0e2 100644 --- a/src/regexp/regexp-parser.cc +++ b/src/regexp/regexp-parser.cc @@ -11,6 +11,7 @@ #include "src/objects/objects-inl.h" #include "src/regexp/jsregexp.h" #include "src/regexp/property-sequences.h" +#include "src/regexp/regexp-macro-assembler.h" #include "src/strings/char-predicates-inl.h" #include "src/utils/ostreams.h" #include "src/utils/utils.h"