⚙️ V8 引擎优化深度解析

目录

V8 优化编译器在做什么？

本次解读的核心文件

src/compiler/pipeline.cc · graph-reducer.cc · simplified-lowering.cc · scheduler.cc

入口：Pipeline 类暴露的能力

class TFGraph;
class InstructionSequence;
class JSGraph;
class JSHeapBroker;
class MachineGraph;
class Schedule;
class SourcePositionTable;
struct WasmCompilationData;
class TFPipelineData;
class ZoneStats;

struct InstructionRangesAsJSON {
  const InstructionSequence* sequence;
  const ZoneVector<std::pair<int, int>>* instr_origins;
};

std::ostream& operator<<(std::ostream& out, const InstructionRangesAsJSON& s);

class Pipeline : public AllStatic {
 public:
  // Returns a new compilation job for the given JavaScript function.
  static V8_EXPORT_PRIVATE std::unique_ptr<TurbofanCompilationJob>
  NewCompilationJob(Isolate* isolate, Handle<JSFunction> function,
                    CodeKind code_kind, bool has_script,
                    BytecodeOffset osr_offset = BytecodeOffset::None());

  using CodeAssemblerGenerator =
      std::function<void(compiler::CodeAssemblerState*)>;
  using CodeAssemblerInstaller =
      std::function<void(Builtin builtin, Handle<Code> code)>;

  static std::unique_ptr<TurbofanCompilationJob>
  NewCSLinkageCodeStubBuiltinCompilationJob(
      Isolate* isolate, Builtin builtin, CodeAssemblerGenerator generator,
      CodeAssemblerInstaller installer,
      const AssemblerOptions& assembler_options,
      CallDescriptors::Key interface_descriptor, const char* name,
      const ProfileDataFromFile* profile_data, int finalize_order);

  static std::unique_ptr<TurbofanCompilationJob>
  NewJSLinkageCodeStubBuiltinCompilationJob(
      Isolate* isolate, Builtin builtin, CodeAssemblerGenerator generator,
      CodeAssemblerInstaller installer,
      const AssemblerOptions& assembler_options, int argc, const char* name,
      const ProfileDataFromFile* profile_data, int finalize_order);

  using TurboshaftAssemblerGenerator =
      std::function<void(compiler::turboshaft::PipelineData*, Isolate*,
                         compiler::turboshaft::Graph&, Zone*)>;
  using TurboshaftAssemblerInstaller = CodeAssemblerInstaller;

Pipeline 为什么是中心中枢？

它不直接完成所有优化，而是像 orchestrator 一样按阶段把数据交给专门组件。

NewCompilationJob：优化任务如何被创建

std::unique_ptr<TurbofanCompilationJob> Pipeline::NewCompilationJob(
    Isolate* isolate, Handle<JSFunction> function, CodeKind code_kind,
    bool has_script, BytecodeOffset osr_offset) {
  Handle<SharedFunctionInfo> shared(function->shared(), isolate);
  return std::make_unique<PipelineCompilationJob>(isolate, shared, function,
                                                  osr_offset, code_kind);
}

bool PipelineImpl::ComputeScheduledGraph() {
  TFPipelineData* data = this->data_;

  // We should only schedule the graph if it is not scheduled yet.
  DCHECK_NULL(data->schedule());

  RUN_MAYBE_ABORT(ComputeSchedulePhase);
  TraceScheduleAndVerify(data->info(), data, data->schedule(), "schedule");
  return true;
}

#if V8_ENABLE_WASM_SIMD256_REVEC
bool PipelineImpl::Revectorize() { return Run<RevectorizePhase>(); }
#endif  // V8_ENABLE_WASM_SIMD256_REVEC

OptimizedCompilationInfo* PipelineImpl::info() const { return data_->info(); }

Isolate* PipelineImpl::isolate() const { return data_->isolate(); }

CodeGenerator* PipelineImpl::code_generator() const {
  return data_->code_generator();
}

ObserveNodeManager* PipelineImpl::observe_node_manager() const {
  return data_->observe_node_manager();
}

std::ostream& operator<<(std::ostream& out, const InstructionRangesAsJSON& s) {
  const int max = static_cast<int>(s.sequence->LastInstructionIndex());

  out << ", \"nodeIdToInstructionRange\": {";
  bool need_comma = false;
  for (size_t i = 0; i < s.instr_origins->size(); ++i) {
    std::pair<int, int> offset = (*s.instr_origins)[i];
    if (offset.first == -1) continue;
    const int first = max - offset.first + 1;
    const int second = max - offset.second + 1;
    if (need_comma) out << ", ";
    out << "\"" << i << "\": [" << first << ", " << second << "]";
    need_comma = true;
  }
  out << "}";
  out << ", \"blockIdToInstructionRange\": {";
  need_comma = false;
  for (auto block : s.sequence->instruction_blocks()) {
    if (need_comma) out << ", ";
    out << "\"" << block->rpo_number() << "\": [" << block->code_start() << ", "
        << block->code_end() << "]";
    need_comma = true;
  }
  out << "}";
  return out;
}

}  // namespace compiler
}  // namespace internal
}  // namespace v8

优化编译全景图

Node：IR 图的最小计算单元

// Copyright 2013 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_COMPILER_NODE_H_
#define V8_COMPILER_NODE_H_

#include "src/common/globals.h"
#include "src/compiler/graph-zone-traits.h"
#include "src/compiler/opcodes.h"
#include "src/compiler/operator.h"
#include "src/compiler/turbofan-types.h"
#include "src/zone/zone-containers.h"

namespace v8 {
namespace internal {
namespace compiler {

// Forward declarations.
class Edge;
class TFGraph;

// Marks are used during traversal of the graph to distinguish states of nodes.
// Each node has a mark which is a monotonically increasing integer, and a
// {NodeMarker} has a range of values that indicate states of a node.
using Mark = uint32_t;

// NodeIds are identifying numbers for nodes that can be used to index auxiliary
// out-of-line data associated with each node.
using NodeId = uint32_t;

// A Node is the basic primitive of graphs. Nodes are chained together by
// input/use chains but by default otherwise contain only an identifying number
// which specific applications of graphs and nodes can use to index auxiliary
// out-of-line data, especially transient data.
//
// In addition Nodes only contain a mutable Operator that may change during
// compilation, e.g. during lowering passes. Other information that needs to be
// associated with Nodes during compilation must be stored out-of-line indexed
// by the Node's id.
class V8_EXPORT_PRIVATE Node final {
 public:
  static Node* New(Zone* zone, NodeId id, const Operator* op, int input_count,
                   Node* const* inputs, bool has_extensible_inputs);
  static Node* Clone(Zone* zone, NodeId id, const Node* node);

  inline bool IsDead() const;
  void Kill();

  const Operator* op() const { return op_; }

  constexpr IrOpcode::Value opcode() const {
    DCHECK_GE(IrOpcode::kLast, op_->opcode());
    return static_cast<IrOpcode::Value>(op_->opcode());
  }

  NodeId id() const { return IdField::decode(bit_field_); }

  int InputCount() const {
    return has_inline_inputs() ? InlineCountField::decode(bit_field_)
                               : outline_inputs()->count_;
  }

#ifdef DEBUG
  void Verify();
#else
  inline void Verify() {}
#endif

  Node* InputAt(int index) const {
    DCHECK_LE(0, index);
    DCHECK_LT(index, InputCount());
    return *GetInputPtrConst(index);
  }

  void ReplaceInput(int index, Node* new_to) {
    DCHECK_LE(0, index);
    DCHECK_LT(index, InputCount());
    ZoneNodePtr* input_ptr = GetInputPtr(index);
    Node* old_to = *input_ptr;
    if (old_to != new_to) {
      Use* use = GetUsePtr(index);
      if (old_to) old_to->RemoveUse(use);
      *input_ptr = new_to;
      if (new_to) new_to->AppendUse(use);
    }
  }

  void AppendInput(Zone* zone, Node* new_to);
  void InsertInput(Zone* zone, int index, Node* new_to);
  void InsertInputs(Zone* zone, int index, int count);
  // Returns the removed input.
  Node* RemoveInput(int index);
  void NullAllInputs();
  void TrimInputCount(int new_input_count);
  // Can trim, extend by appending new inputs, or do nothing.
  void EnsureInputCount(Zone* zone, int new_input_count);

  int UseCount() const;
  int BranchUseCount() const;
  void ReplaceUses(Node* replace_to);

  class InputEdges;
  inline InputEdges input_edges();

  class Inputs;
  inline Inputs inputs() const;
  inline base::Vector<Node*> inputs_vector() const;

  class UseEdges final {
   public:
    using value_type = Edge;

    class iterator;
    inline iterator begin() const;
    inline iterator end() const;

    bool empty() const;

    explicit UseEdges(Node* node) : node_(node) {}

Node 的输入边 / 使用边设计

class Node::InputEdges final {
 public:
  using value_type = Edge;

  class iterator;
  inline iterator begin() const;
  inline iterator end() const;

  bool empty() const { return count_ == 0; }
  int count() const { return count_; }

  inline value_type operator[](int index) const;

  InputEdges(ZoneNodePtr* input_root, Use* use_root, int count)
      : input_root_(input_root), use_root_(use_root), count_(count) {}

 private:
  ZoneNodePtr* input_root_;
  Use* use_root_;
  int count_;
};

class V8_EXPORT_PRIVATE Node::Inputs final {
 public:
  using value_type = Node*;

  class const_iterator;
  inline const_iterator begin() const;
  inline const_iterator end() const;

  bool empty() const { return count_ == 0; }
  int count() const { return count_; }

  inline value_type operator[](int index) const;

  explicit Inputs(ZoneNodePtr const* input_root, int count)
      : input_root_(input_root), count_(count) {}

 private:
  ZoneNodePtr const* input_root_;
  int count_;
};

// An encapsulation for information associated with a single use of a node as an
// input from another node, allowing access to both the defining node and
// the node having the input.
class Edge final {
 public:
  Node* from() const { return use_->from(); }
  Node* to() const { return *input_ptr_; }
  int index() const {
    int const index = use_->input_index();
    DCHECK_LT(index, use_->from()->InputCount());
    return index;
  }

  bool operator==(const Edge& other) { return input_ptr_ == other.input_ptr_; }
  bool operator!=(const Edge& other) { return !(*this == other); }

  void UpdateTo(Node* new_to) {
    Node* old_to = *input_ptr_;
    if (old_to != new_to) {
      if (old_to) old_to->RemoveUse(use_);
      *input_ptr_ = new_to;
      if (new_to) new_to->AppendUse(use_);
    }
  }

 private:
  friend class Node::UseEdges::iterator;
  friend class Node::InputEdges;
  friend class Node::InputEdges::iterator;

  Edge(Node::Use* use, ZoneNodePtr* input_ptr)
      : use_(use), input_ptr_(input_ptr) {
    DCHECK_NOT_NULL(use);
    DCHECK_NOT_NULL(input_ptr);
    DCHECK_EQ(input_ptr, use->input_ptr());
  }

  Node::Use* use_;
  ZoneNodePtr* input_ptr_;
};

bool Node::IsDead() const {
  Node::Inputs inputs = this->inputs();
  return inputs.count() > 0 && inputs[0] == nullptr;
}

Node::InputEdges Node::input_edges() {
  int inline_count = InlineCountField::decode(bit_field_);
  if (inline_count != kOutlineMarker) {
    return InputEdges(inline_inputs(), reinterpret_cast<Use*>(this) - 1,
                      inline_count);
  } else {
    return InputEdges(outline_inputs()->inputs(),
                      reinterpret_cast<Use*>(outline_inputs()) - 1,
                      outline_inputs()->count_);
  }
}

Node::Inputs Node::inputs() const {
  int inline_count = InlineCountField::decode(bit_field_);
  if (inline_count != kOutlineMarker) {
    return Inputs(inline_inputs(), inline_count);
  } else {
    return Inputs(outline_inputs()->inputs(), outline_inputs()->count_);
  }
}

base::Vector<Node*> Node::inputs_vector() const {
  int inline_count = InlineCountField::decode(bit_field_);
  if (inline_count != kOutlineMarker) {
    return base::VectorOf<Node*>(inline_inputs(), inline_count);
  } else {
    return base::VectorOf<Node*>(outline_inputs()->inputs(),
                                 outline_inputs()->count_);
  }
}

// A forward iterator to visit the edges for the input dependencies of a node.
class Node::InputEdges::iterator final {
 public:
  using iterator_category = std::forward_iterator_tag;
  using difference_type = std::ptrdiff_t;
  using value_type = Edge;
  using pointer = Edge*;
  using reference = Edge&;

  iterator() : use_(nullptr), input_ptr_(nullptr) {}
  iterator(const iterator& other) = default;

  Edge operator*() const { return Edge(use_, input_ptr_); }
  bool operator==(const iterator& other) const {
    return input_ptr_ == other.input_ptr_;
  }
  bool operator!=(const iterator& other) const { return !(*this == other); }
  iterator& operator++() {
    input_ptr_++;
    use_--;
    return *this;
  }
  iterator operator++(int);
  iterator& operator+=(difference_type offset) {
    input_ptr_ += offset;
    use_ -= offset;
    return *this;
  }
  iterator operator+(difference_type offset) const {
    return iterator(use_ - offset, input_ptr_ + offset);
  }
  difference_type operator-(const iterator& other) const {
    return input_ptr_ - other.input_ptr_;
  }

 private:
  friend class Node;

  explicit iterator(Use* use, ZoneNodePtr* input_ptr)
      : use_(use), input_ptr_(input_ptr) {}

  Use* use_;
  ZoneNodePtr* input_ptr_;
};


Node::InputEdges::iterator Node::InputEdges::begin() const {
  return Node::InputEdges::iterator(use_root_, input_root_);
}


Node::InputEdges::iterator Node::InputEdges::end() const {
  return Node::InputEdges::iterator(use_root_ - count_, input_root_ + count_);
}

Edge Node::InputEdges::operator[](int index) const {
  return Edge(use_root_ + index, input_root_ + index);
}

// A forward iterator to visit the inputs of a node.
class Node::Inputs::const_iterator final {
 public:
  using iterator_category = std::forward_iterator_tag;
  using difference_type = std::ptrdiff_t;
  using value_type = Node*;
  using pointer = const value_type*;
  using reference = value_type&;

  const_iterator(const const_iterator& other) = default;

  Node* operator*() const { return *input_ptr_; }
  bool operator==(const const_iterator& other) const {
    return input_ptr_ == other.input_ptr_;
  }
  bool operator!=(const const_iterator& other) const {
    return !(*this == other);
  }
  const_iterator& operator++() {
    ++input_ptr_;
    return *this;
  }
  const_iterator operator++(int);
  const_iterator& operator+=(difference_type offset) {
    input_ptr_ += offset;
    return *this;
  }
  const_iterator operator+(difference_type offset) const {
    return const_iterator(input_ptr_ + offset);
  }
  difference_type operator-(const const_iterator& other) const {
    return input_ptr_ - other.input_ptr_;
  }

 private:
  friend class Node::Inputs;

  explicit const_iterator(ZoneNodePtr const* input_ptr)
      : input_ptr_(input_ptr) {}

  ZoneNodePtr const* input_ptr_;
};


Node::Inputs::const_iterator Node::Inputs::begin() const {
  return const_iterator(input_root_);
}


Node::Inputs::const_iterator Node::Inputs::end() const {
  return const_iterator(input_root_ + count_);
}

Node* Node::Inputs::operator[](int index) const { return input_root_[index]; }

// A forward iterator to visit the uses edges of a node.
class Node::UseEdges::iterator final {
 public:
  iterator(const iterator& other) = default;

  Edge operator*() const { return Edge(current_, current_->input_ptr()); }
  bool operator==(const iterator& other) const {
    return current_ == other.current_;
  }
  bool operator!=(const iterator& other) const { return !(*this == other); }
  iterator& operator++() {
    DCHECK_NOT_NULL(current_);
    current_ = next_;
    next_ = current_ ? static_cast<Node::Use*>(current_->next) : nullptr;
    return *this;
  }
  iterator operator++(int);

 private:
  friend class Node::UseEdges;

  iterator() : current_(nullptr), next_(nullptr) {}
  explicit iterator(Node* node)
      : current_(node->first_use_),
        next_(current_ ? static_cast<Node::Use*>(current_->next) : nullptr) {}

  Node::Use* current_;
  Node::Use* next_;
};


Node::UseEdges::iterator Node::UseEdges::begin() const {
  return Node::UseEdges::iterator(this->node_);
}


Node::UseEdges::iterator Node::UseEdges::end() const {
  return Node::UseEdges::iterator();
}


// A forward iterator to visit the uses of a node.
class Node::Uses::const_iterator final {
 public:
  using iterator_category = std::forward_iterator_tag;
  using difference_type = int;
  using value_type = Node*;
  using pointer = Node**;
  using reference = Node*&;

  Node* operator*() const { return current_->from(); }
  bool operator==(const const_iterator& other) const {
    return other.current_ == current_;
  }
  bool operator!=(const const_iterator& other) const {
    return other.current_ != current_;
  }
  const_iterator& operator++() {
    DCHECK_NOT_NULL(current_);
    // Checking no use gets mutated while iterating through them, a potential
    // very tricky cause of bug.
    current_ = current_->next;
#ifdef DEBUG
    DCHECK_EQ(current_, next_);
    next_ = current_ ? current_->next : nullptr;
#endif
    return *this;
  }
  const_iterator operator++(int);

 private:
  friend class Node::Uses;

  const_iterator() : current_(nullptr) {}
  explicit const_iterator(Node* node)

为什么优化器喜欢图，不喜欢线性 IR？

GraphReducer：统一驱动所有 reducer

class TFGraph;
class JSHeapBroker;
class Node;
class ObserveNodeManager;

// NodeIds are identifying numbers for nodes that can be used to index auxiliary
// out-of-line data associated with each node.
using NodeId = uint32_t;

// Possible outcomes for decisions.
enum class Decision : uint8_t { kUnknown, kTrue, kFalse };

// Represents the result of trying to reduce a node in the graph.
class Reduction final {
 public:
  explicit Reduction(Node* replacement = nullptr) : replacement_(replacement) {}

  Node* replacement() const { return replacement_; }
  bool Changed() const { return replacement() != nullptr; }
  Reduction FollowedBy(Reduction next) const {
    if (next.Changed()) return next;
    return *this;
  }

 private:
  Node* replacement_;
};


// A reducer can reduce or simplify a given node based on its operator and
// inputs. This class functions as an extension point for the graph reducer for
// language-specific reductions (e.g. reduction based on types or constant
// folding of low-level operators) can be integrated into the graph reduction
// phase.
class V8_EXPORT_PRIVATE Reducer {
 public:
  virtual ~Reducer() = default;

  // Only used for tracing, when using the --trace_turbo_reduction flag.
  virtual const char* reducer_name() const = 0;

  // Try to reduce a node if possible.
  Reduction Reduce(Node* node, ObserveNodeManager* observe_node_manager);

  // Invoked by the {GraphReducer} when all nodes are done.  Can be used to
  // do additional reductions at the end, which in turn can cause a new round
  // of reductions.
  virtual void Finalize();

  // Helper functions for subclasses to produce reductions for a node.
  static Reduction NoChange() { return Reduction(); }
  static Reduction Replace(Node* node) { return Reduction(node); }
  static Reduction Changed(Node* node) { return Reduction(node); }

 private:
  virtual Reduction Reduce(Node* node) = 0;
};


// An advanced reducer can also edit the graphs by changing and replacing nodes
// other than the one currently being reduced.
class AdvancedReducer : public Reducer {
 public:
  // Observe the actions of this reducer.
  class Editor {
   public:
    virtual ~Editor() = default;

    // Replace {node} with {replacement}.
    virtual void Replace(Node* node, Node* replacement) = 0;
    virtual void Replace(Node* node, Node* replacement, NodeId max_id) = 0;
    // Revisit the {node} again later.
    virtual void Revisit(Node* node) = 0;
    // Replace value uses of {node} with {value} and effect uses of {node} with
    // {effect}. If {effect == nullptr}, then use the effect input to {node}.
    // All control uses will be relaxed assuming {node} cannot throw.
    virtual void ReplaceWithValue(Node* node, Node* value, Node* effect,
                                  Node* control) = 0;
  };

  explicit AdvancedReducer(Editor* editor) : editor_(editor) {}

 protected:
  // Helper functions for subclasses to produce reductions for a node.
  static Reduction Replace(Node* node) { return Reducer::Replace(node); }

  // Helper functions for subclasses to edit the graph.
  void Replace(Node* node, Node* replacement) {
    DCHECK_NOT_NULL(editor_);
    editor_->Replace(node, replacement);
  }
  void Replace(Node* node, Node* replacement, NodeId max_id) {
    return editor_->Replace(node, replacement, max_id);
  }
  void Revisit(Node* node) {
    DCHECK_NOT_NULL(editor_);
    editor_->Revisit(node);
  }
  void ReplaceWithValue(Node* node, Node* value, Node* effect = nullptr,
                        Node* control = nullptr) {
    DCHECK_NOT_NULL(editor_);
    editor_->ReplaceWithValue(node, value, effect, control);
  }

  // Relax the effects of {node} by immediately replacing effect and control
  // uses of {node} with the effect and control input to {node}.
  // TODO(turbofan): replace the effect input to {node} with {graph->start()}.
  void RelaxEffectsAndControls(Node* node) {
    ReplaceWithValue(node, node, nullptr, nullptr);
  }

  // Relax the control uses of {node} by immediately replacing them with the
  // either the given {control} node, or the control input to {node}.
  void RelaxControls(Node* node, Node* control = nullptr) {
    ReplaceWithValue(node, node, node, control);
  }

  void MergeControlToEnd(TFGraph* graph, CommonOperatorBuilder* common,
                         Node* node) {
    NodeProperties::MergeControlToEnd(graph, common, node);
    Revisit(graph->end());
  }

 private:
  Editor* const editor_;
};


// Performs an iterative reduction of a node graph.
class V8_EXPORT_PRIVATE GraphReducer
    : public NON_EXPORTED_BASE(AdvancedReducer::Editor) {
 public:
  GraphReducer(Zone* zone, TFGraph* graph, TickCounter* tick_counter,
               JSHeapBroker* broker, Node* dead = nullptr,
               ObserveNodeManager* observe_node_manager = nullptr);
  ~GraphReducer() override;

  GraphReducer(const GraphReducer&) = delete;
  GraphReducer& operator=(const GraphReducer&) = delete;

  TFGraph* graph() const { return graph_; }

  void AddReducer(Reducer* reducer);

  // Reduce a single node.
  void ReduceNode(Node* const);
  // Reduce the whole graph.
  void ReduceGraph();

 private:
  enum class State : uint8_t;
  struct NodeState {
    Node* node;
    int input_index;
  };

Reduction 对象：最简单但极关键的协议

这让每个 reducer 都只关心局部变换，复杂度被框架吸收。

ReduceGraph：从 graph end 开始反向吃全图

    }
  }
  DCHECK(revisit_.empty());
  DCHECK(stack_.empty());
}


void GraphReducer::ReduceGraph() { ReduceNode(graph()->end()); }


Reduction GraphReducer::Reduce(Node* const node) {
  auto skip = reducers_.end();
  for (auto i = reducers_.begin(); i != reducers_.end();) {
    if (i != skip) {
      tick_counter_->TickAndMaybeEnterSafepoint();
      Reduction reduction = (*i)->Reduce(node, observe_node_manager_);
      if (!reduction.Changed()) {
        // No change from this reducer.
      } else if (reduction.replacement() == node) {
        // {replacement} == {node} represents an in-place reduction. Rerun
        // all the other reducers for this node, as now there may be more
        // opportunities for reduction.
        if (v8_flags.trace_turbo_reduction) {
          UnparkedScopeIfNeeded unparked(broker_);
          // TODO(neis): Disallow racy handle dereference once we stop
          // supporting --no-local-heaps --no-concurrent-inlining.
          AllowHandleDereference allow_deref;
          StdoutStream{} << "- In-place update of #" << *node << " by reducer "
                         << (*i)->reducer_name() << std::endl;
        }
        skip = i;
        i = reducers_.begin();
        continue;
      } else {
        // {node} was replaced by another node.
        if (v8_flags.trace_turbo_reduction) {
          UnparkedScopeIfNeeded unparked(broker_);
          // TODO(neis): Disallow racy handle dereference once we stop
          // supporting --no-local-heaps --no-concurrent-inlining.
          AllowHandleDereference allow_deref;
          StdoutStream{} << "- Replacement of #" << *node << " with #"
                         << *(reduction.replacement()) << " by reducer "
                         << (*i)->reducer_name() << std::endl;
        }
        return reduction;
      }
    }
    ++i;
  }
  if (skip == reducers_.end()) {
    // No change from any reducer.
    return Reducer::NoChange();
  }
  // At least one reducer did some in-place reduction.
  return Reducer::Changed(node);
}


void GraphReducer::ReduceTop() {
  NodeState& entry = stack_.top();
  Node* node = entry.node;
  DCHECK_EQ(State::kOnStack, state_.Get(node));

  if (node->IsDead()) return Pop();  // Node was killed while on stack.

  Node::Inputs node_inputs = node->inputs();

  // Recurse on an input if necessary.
  int start = entry.input_index < node_inputs.count() ? entry.input_index : 0;
  for (int i = start; i < node_inputs.count(); ++i) {
    Node* input = node_inputs[i];

ReduceTop：单节点规约状态机

void GraphReducer::ReduceTop() {
  NodeState& entry = stack_.top();
  Node* node = entry.node;
  DCHECK_EQ(State::kOnStack, state_.Get(node));

  if (node->IsDead()) return Pop();  // Node was killed while on stack.

  Node::Inputs node_inputs = node->inputs();

  // Recurse on an input if necessary.
  int start = entry.input_index < node_inputs.count() ? entry.input_index : 0;
  for (int i = start; i < node_inputs.count(); ++i) {
    Node* input = node_inputs[i];
    if (input != node && Recurse(input)) {
      entry.input_index = i + 1;
      return;
    }
  }
  for (int i = 0; i < start; ++i) {
    Node* input = node_inputs[i];
    if (input != node && Recurse(input)) {
      entry.input_index = i + 1;
      return;
    }
  }

  // Remember the max node id before reduction.
  NodeId const max_id = static_cast<NodeId>(graph()->NodeCount() - 1);

  // All inputs should be visited or on stack. Apply reductions to node.
  Reduction reduction = Reduce(node);

  // If there was no reduction, pop {node} and continue.
  if (!reduction.Changed()) return Pop();

  // Check if the reduction is an in-place update of the {node}.
  Node* const replacement = reduction.replacement();
  if (replacement == node) {
    for (Node* const user : node->uses()) {
      DCHECK_IMPLIES(user == node, state_.Get(node) != State::kVisited);
      Revisit(user);
    }

    // In-place update of {node}, may need to recurse on an input.
    node_inputs = node->inputs();
    for (int i = 0; i < node_inputs.count(); ++i) {
      Node* input = node_inputs[i];
      if (input != node && Recurse(input)) {
        entry.input_index = i + 1;
        return;
      }
    }
  }

  // After reducing the node, pop it off the stack.
  Pop();

  // Check if we have a new replacement.
  if (replacement != node) {
    Replace(node, replacement, max_id);
  }
}


void GraphReducer::Replace(Node* node, Node* replacement) {
  Replace(node, replacement, std::numeric_limits<NodeId>::max());
}


void GraphReducer::Replace(Node* node, Node* replacement, NodeId max_id) {
  if (node == graph()->start()) graph()->SetStart(replacement);
  if (node == graph()->end()) graph()->SetEnd(replacement);
  if (replacement->id() <= max_id) {
    // {replacement} is an old node, so unlink {node} and assume that
    // {replacement} was already reduced and finish.
    for (Edge edge : node->use_edges()) {
      Node* const user = edge.from();
      Verifier::VerifyEdgeInputReplacement(edge, replacement);
      edge.UpdateTo(replacement);

ReplaceWithValue：值/副作用/控制三路分发

void GraphReducer::ReplaceWithValue(Node* node, Node* value, Node* effect,
                                    Node* control) {
  if (effect == nullptr && node->op()->EffectInputCount() > 0) {
    effect = NodeProperties::GetEffectInput(node);
  }
  if (control == nullptr && node->op()->ControlInputCount() > 0) {
    control = NodeProperties::GetControlInput(node);
  }

  // Requires distinguishing between value, effect and control edges.
  for (Edge edge : node->use_edges()) {
    Node* const user = edge.from();
    DCHECK(!user->IsDead());
    if (NodeProperties::IsControlEdge(edge)) {
      if (user->opcode() == IrOpcode::kIfSuccess) {
        Replace(user, control);
      } else if (user->opcode() == IrOpcode::kIfException) {
        DCHECK_NOT_NULL(dead_);
        edge.UpdateTo(dead_);
        Revisit(user);
      } else {
        DCHECK_NOT_NULL(control);
        edge.UpdateTo(control);
        Revisit(user);
      }
    } else if (NodeProperties::IsEffectEdge(edge)) {
      DCHECK_NOT_NULL(effect);
      edge.UpdateTo(effect);
      Revisit(user);
    } else {
      DCHECK_NOT_NULL(value);
      edge.UpdateTo(value);
      Revisit(user);
    }
  }
}


void GraphReducer::Pop() {
  Node* node = stack_.top().node;
  state_.Set(node, State::kVisited);
  stack_.pop();
}


void GraphReducer::Push(Node* const node) {
  DCHECK_NE(State::kOnStack, state_.Get(node));
  state_.Set(node, State::kOnStack);
  stack_.push({node, 0});
}


bool GraphReducer::Recurse(Node* node) {
  if (state_.Get(node) > State::kRevisit) return false;
  Push(node);
  return true;
}


void GraphReducer::Revisit(Node* node) {
  if (state_.Get(node) == State::kVisited) {
    state_.Set(node, State::kRevisit);
    revisit_.push(node);
  }
}

}  // namespace compiler
}  // namespace internal
}  // namespace v8

Reducer 链的设计思想

一串小而专的 reducer，比一个超级优化器更容易维护与组合。

案例：冗余检查如何被消掉

案例：控制流也能被优化

GraphReducer 的强项是让这些变化在 统一框架 中发生，而不是分散在几十处 ad-hoc 代码。

SimplifiedLowering：把“高级节点”压低到机器层

// Forward declarations.
class NodeOriginTable;
class ObserveNodeManager;
class RepresentationChanger;
class RepresentationSelector;
class SourcePositionTable;
class TypeCache;

class V8_EXPORT_PRIVATE SimplifiedLowering final {
 public:
  SimplifiedLowering(JSGraph* jsgraph, JSHeapBroker* broker, Zone* zone,
                     SourcePositionTable* source_position,
                     NodeOriginTable* node_origins, TickCounter* tick_counter,
                     Linkage* linkage, OptimizedCompilationInfo* info,
                     ObserveNodeManager* observe_node_manager = nullptr);
  ~SimplifiedLowering() = default;

  void LowerAllNodes();

  void DoMax(Node* node, Operator const* op, MachineRepresentation rep);
  void DoMin(Node* node, Operator const* op, MachineRepresentation rep);
  void DoJSToNumberOrNumericTruncatesToFloat64(
      Node* node, RepresentationSelector* selector);
  void DoJSToNumberOrNumericTruncatesToWord32(Node* node,
                                              RepresentationSelector* selector);
  void DoIntegral32ToBit(Node* node);
  void DoOrderedNumberToBit(Node* node);
  void DoNumberToBit(Node* node);
  void DoIntegerToUint8Clamped(Node* node);
  void DoNumberToUint8Clamped(Node* node);
  void DoSigned32ToUint8Clamped(Node* node);
  void DoUnsigned32ToUint8Clamped(Node* node);

 private:
  // The purpose of this nested class is to hide method
  // v8::internal::compiler::NodeProperties::ChangeOp which should not be
  // directly used by code in SimplifiedLowering.
  // SimplifiedLowering code should call SimplifiedLowering::ChangeOp instead,
  // in order to notify the changes to ObserveNodeManager and support the
  // %ObserveNode intrinsic.
  class NodeProperties : public compiler::NodeProperties {
    static void ChangeOp(Node* node, const Operator* new_op) { UNREACHABLE(); }
  };
  void ChangeOp(Node* node, const Operator* new_op);

  JSGraph* const jsgraph_;
  JSHeapBroker* broker_;
  Zone* const zone_;
  TypeCache const* type_cache_;
  SetOncePointer<Node> to_number_code_;
  SetOncePointer<Node> to_number_convert_big_int_code_;
  SetOncePointer<Node> to_numeric_code_;
  SetOncePointer<Operator const> to_number_operator_;
  SetOncePointer<Operator const> to_number_convert_big_int_operator_;
  SetOncePointer<Operator const> to_numeric_operator_;

  // TODO(danno): SimplifiedLowering shouldn't know anything about the source
  // positions table, but must for now since there currently is no other way to
  // pass down source position information to nodes created during
  // lowering. Once this phase becomes a vanilla reducer, it should get source
  // position information via the SourcePositionWrapper like all other reducers.
  SourcePositionTable* source_positions_;
  NodeOriginTable* node_origins_;

  TickCounter* const tick_counter_;
  Linkage* const linkage_;
  OptimizedCompilationInfo* info_;

  ObserveNodeManager* const observe_node_manager_;

  Node* Float64Round(Node* const node);
  Node* Float64Sign(Node* const node);
  Node* Int32Abs(Node* const node);
  Node* Int32Div(Node* const node);
  Node* Int32Mod(Node* const node);
  Node* Int32Sign(Node* const node);
  Node* Uint32Div(Node* const node);
  Node* Uint32Mod(Node* const node);

  Node* ToNumberCode();
  Node* ToNumberConvertBigIntCode();
  Node* ToNumericCode();
  Node* Ieee754Fp64ToFp16RawBitsCode();
  Operator const* ToNumberOperator();
  Operator const* ToNumberConvertBigIntOperator();
  Operator const* ToNumericOperator();
  Operator const* Ieee754Fp64ToFp16RawBitsOperator();

  friend class RepresentationSelector;

  Isolate* isolate() { return jsgraph_->isolate(); }
  Zone* zone() { return jsgraph_->zone(); }
  JSGraph* jsgraph() { return jsgraph_; }
  TFGraph* graph() { return jsgraph()->graph(); }
  CommonOperatorBuilder* common() { return jsgraph()->common(); }
  MachineOperatorBuilder* machine() { return jsgraph()->machine(); }
  SimplifiedOperatorBuilder* simplified() { return jsgraph()->simplified(); }
  Linkage* linkage() { return linkage_; }
};

}  // namespace compiler
}  // namespace internal
}  // namespace v8

#endif  // V8_COMPILER_SIMPLIFIED_LOWERING_H_

LowerAllNodes：Lowering 总入口

void SimplifiedLowering::LowerAllNodes() {
  SimplifiedLoweringVerifier* verifier = nullptr;
  if (v8_flags.verify_simplified_lowering) {
    verifier = zone_->New<SimplifiedLoweringVerifier>(zone_, graph());
  }
  RepresentationChanger changer(jsgraph(), broker_, verifier);
  RepresentationSelector selector(
      jsgraph(), broker_, zone_, &changer, source_positions_, node_origins_,
      tick_counter_, linkage_, observe_node_manager_, verifier);
  selector.Run(this);
}

void SimplifiedLowering::DoJSToNumberOrNumericTruncatesToFloat64(
    Node* node, RepresentationSelector* selector) {
  DCHECK(node->opcode() == IrOpcode::kJSToNumber ||
         node->opcode() == IrOpcode::kJSToNumberConvertBigInt ||
         node->opcode() == IrOpcode::kJSToNumeric);
  Node* value = node->InputAt(0);
  Node* context = node->InputAt(1);
  Node* frame_state = node->InputAt(2);
  Node* effect = node->InputAt(3);
  Node* control = node->InputAt(4);

  Node* check0 = graph()->NewNode(simplified()->ObjectIsSmi(), value);
  Node* branch0 = graph()->NewNode(
      common()->Branch(BranchHint::kTrue, BranchSemantics::kMachine), check0,
      control);

  Node* if_true0 = graph()->NewNode(common()->IfTrue(), branch0);
  Node* etrue0 = effect;
  Node* vtrue0;
  {
    vtrue0 = graph()->NewNode(simplified()->ChangeTaggedSignedToInt32(), value);
    vtrue0 = graph()->NewNode(machine()->ChangeInt32ToFloat64(), vtrue0);
  }

  Node* if_false0 = graph()->NewNode(common()->IfFalse(), branch0);
  Node* efalse0 = effect;
  Node* vfalse0;
  {
    Operator const* op =
        node->opcode() == IrOpcode::kJSToNumber
            ? (node->opcode() == IrOpcode::kJSToNumberConvertBigInt
                   ? ToNumberConvertBigIntOperator()
                   : ToNumberOperator())
            : ToNumericOperator();
    Node* code = node->opcode() == IrOpcode::kJSToNumber
                     ? ToNumberCode()
                     : (node->opcode() == IrOpcode::kJSToNumberConvertBigInt
                            ? ToNumberConvertBigIntCode()
                            : ToNumericCode());
    vfalse0 = efalse0 = if_false0 = graph()->NewNode(
        op, code, value, context, frame_state, efalse0, if_false0);

    // Update potential {IfException} uses of {node} to point to the above
    // stub call node instead.
    Node* on_exception = nullptr;
    if (NodeProperties::IsExceptionalCall(node, &on_exception)) {
      NodeProperties::ReplaceControlInput(on_exception, vfalse0);
      NodeProperties::ReplaceEffectInput(on_exception, efalse0);
      if_false0 = graph()->NewNode(common()->IfSuccess(), vfalse0);
    }

    Node* check1 = graph()->NewNode(simplified()->ObjectIsSmi(), vfalse0);
    Node* branch1 = graph()->NewNode(
        common()->Branch(BranchHint::kNone, BranchSemantics::kMachine), check1,
        if_false0);

    Node* if_true1 = graph()->NewNode(common()->IfTrue(), branch1);
    Node* etrue1 = efalse0;
    Node* vtrue1;
    {
      vtrue1 =
          graph()->NewNode(simplified()->ChangeTaggedSignedToInt32(), vfalse0);
      vtrue1 = graph()->NewNode(machine()->ChangeInt32ToFloat64(), vtrue1);
    }

    Node* if_false1 = graph()->NewNode(common()->IfFalse(), branch1);
    Node* efalse1 = efalse0;
    Node* vfalse1;
    {
      vfalse1 = efalse1 = graph()->NewNode(
          simplified()->LoadField(AccessBuilder::ForHeapNumberValue()), efalse0,
          efalse1, if_false1);
    }

    if_false0 = graph()->NewNode(common()->Merge(2), if_true1, if_false1);
    efalse0 =
        graph()->NewNode(common()->EffectPhi(2), etrue1, efalse1, if_false0);
    vfalse0 =

表示选择的核心矛盾

DoNumberToBit：数字到位表示的专门降级

}

void SimplifiedLowering::DoOrderedNumberToBit(Node* node) {
  Node* const input = node->InputAt(0);

  node->ReplaceInput(0, graph()->NewNode(machine()->Float64Equal(), input,
                                         jsgraph()->Float64Constant(0.0)));
  node->AppendInput(graph()->zone(), jsgraph()->Int32Constant(0));
  ChangeOp(node, machine()->Word32Equal());
}

void SimplifiedLowering::DoNumberToBit(Node* node) {
  Node* const input = node->InputAt(0);

  node->ReplaceInput(0, jsgraph()->Float64Constant(0.0));
  node->AppendInput(graph()->zone(),
                    graph()->NewNode(machine()->Float64Abs(), input));
  ChangeOp(node, machine()->Float64LessThan());
}

void SimplifiedLowering::DoIntegerToUint8Clamped(Node* node) {
  Node* const input = node->InputAt(0);
  Node* const min = jsgraph()->Float64Constant(0.0);
  Node* const max = jsgraph()->Float64Constant(255.0);

  node->ReplaceInput(
      0, graph()->NewNode(machine()->Float64LessThan(), min, input));
  node->AppendInput(
      graph()->zone(),
      graph()->NewNode(
          common()->Select(MachineRepresentation::kFloat64),
          graph()->NewNode(machine()->Float64LessThan(), input, max), input,
          max));
  node->AppendInput(graph()->zone(), min);
  ChangeOp(node, common()->Select(MachineRepresentation::kFloat64));
}

void SimplifiedLowering::DoNumberToUint8Clamped(Node* node) {
  Node* const input = node->InputAt(0);
  Node* const min = jsgraph()->Float64Constant(0.0);
  Node* const max = jsgraph()->Float64Constant(255.0);

  node->ReplaceInput(
      0, graph()->NewNode(
             common()->Select(MachineRepresentation::kFloat64),
             graph()->NewNode(machine()->Float64LessThan(), min, input),
             graph()->NewNode(
                 common()->Select(MachineRepresentation::kFloat64),
                 graph()->NewNode(machine()->Float64LessThan(), input, max),
                 input, max),
             min));
  ChangeOp(node, machine()->Float64RoundTiesEven().placeholder());
}

void SimplifiedLowering::DoSigned32ToUint8Clamped(Node* node) {
  Node* const input = node->InputAt(0);
  Node* const min = jsgraph()->Int32Constant(0);
  Node* const max = jsgraph()->Int32Constant(255);

  node->ReplaceInput(
      0, graph()->NewNode(machine()->Int32LessThanOrEqual(), input, max));
  node->AppendInput(
      graph()->zone(),
      graph()->NewNode(common()->Select(MachineRepresentation::kWord32),
                       graph()->NewNode(machine()->Int32LessThan(), input, min),
                       min, input));
  node->AppendInput(graph()->zone(), max);
  ChangeOp(node, common()->Select(MachineRepresentation::kWord32));
}

void SimplifiedLowering::DoUnsigned32ToUint8Clamped(Node* node) {
  Node* const input = node->InputAt(0);
  Node* const max = jsgraph()->Uint32Constant(255u);

几个典型 Lowering 模式

类型信息如何驱动优化？

为什么 Lowering 跟 Deopt 紧密相关

Scheduler：让图变成可执行顺序

Scheduler::Scheduler(Zone* zone, TFGraph* graph, Schedule* schedule,
                     Flags flags, size_t node_count_hint,
                     TickCounter* tick_counter,
                     const ProfileDataFromFile* profile_data)
    : zone_(zone),
      graph_(graph),
      schedule_(schedule),
      flags_(flags),
      scheduled_nodes_(zone),
      schedule_root_nodes_(zone),
      schedule_queue_(zone),
      node_data_(zone),
      tick_counter_(tick_counter),
      profile_data_(profile_data),
      common_dominator_cache_(zone) {
  node_data_.reserve(node_count_hint);
  node_data_.resize(graph->NodeCount(), DefaultSchedulerData());
}

Schedule* Scheduler::ComputeSchedule(Zone* zone, TFGraph* graph, Flags flags,
                                     TickCounter* tick_counter,
                                     const ProfileDataFromFile* profile_data) {
  Zone* schedule_zone =
      (flags & Scheduler::kTempSchedule) ? zone : graph->zone();

  // Reserve 10% more space for nodes if node splitting is enabled to try to
  // avoid resizing the vector since that would triple its zone memory usage.
  float node_hint_multiplier = (flags & Scheduler::kSplitNodes) ? 1.1 : 1;
  size_t node_count_hint = node_hint_multiplier * graph->NodeCount();

  Schedule* schedule =
      schedule_zone->New<Schedule>(schedule_zone, node_count_hint);
  Scheduler scheduler(zone, graph, schedule, flags, node_count_hint,
                      tick_counter, profile_data);

  scheduler.BuildCFG();
  scheduler.ComputeSpecialRPONumbering();
  scheduler.GenerateDominatorTree();

  scheduler.PrepareUses();
  scheduler.ScheduleEarly();
  scheduler.ScheduleLate();

  scheduler.SealFinalSchedule();

  return schedule;
}

Scheduler::SchedulerData Scheduler::DefaultSchedulerData() {
  SchedulerData def = {schedule_->start(), 0, kUnknown};
  return def;
}

为什么有了图还要 Schedule？

SchedulerData：每个节点的调度元信息

Scheduler::SchedulerData Scheduler::DefaultSchedulerData() {
  SchedulerData def = {schedule_->start(), 0, kUnknown};
  return def;
}


Scheduler::SchedulerData* Scheduler::GetData(Node* node) {
  return &node_data_[node->id()];
}

Scheduler::Placement Scheduler::InitializePlacement(Node* node) {
  SchedulerData* data = GetData(node);
  if (data->placement_ == kFixed) {
    // Nothing to do for control nodes that have been already fixed in
    // the schedule.
    return data->placement_;
  }
  DCHECK_EQ(kUnknown, data->placement_);
  switch (node->opcode()) {
    case IrOpcode::kParameter:
    case IrOpcode::kOsrValue:
      // Parameters and OSR values are always fixed to the start block.
      data->placement_ = kFixed;
      break;
    case IrOpcode::kPhi:
    case IrOpcode::kEffectPhi: {
      // Phis and effect phis are fixed if their control inputs are, whereas
      // otherwise they are coupled to a floating control node.
      Placement p = GetPlacement(NodeProperties::GetControlInput(node));
      data->placement_ = (p == kFixed ? kFixed : kCoupled);
      break;
    }
    default:
      // Control nodes that were not control-reachable from end may float.
      data->placement_ = kSchedulable;
      break;
  }
  return data->placement_;
}

Scheduler::Placement Scheduler::GetPlacement(Node* node) {
  return GetData(node)->placement_;
}

bool Scheduler::IsLive(Node* node) { return GetPlacement(node) != kUnknown; }

void Scheduler::UpdatePlacement(Node* node, Placement placement) {
  SchedulerData* data = GetData(node);
  if (data->placement_ == kUnknown) {
    // We only update control nodes from {kUnknown} to {kFixed}.  Ideally, we
    // should check that {node} is a control node (including exceptional calls),
    // but that is expensive.
    DCHECK_EQ(Scheduler::kFixed, placement);
    data->placement_ = placement;
    return;
  }

  switch (node->opcode()) {
    case IrOpcode::kParameter:
      // Parameters are fixed once and for all.
      UNREACHABLE();
    case IrOpcode::kPhi:
    case IrOpcode::kEffectPhi: {
      // Phis and effect phis are coupled to their respective blocks.
      DCHECK_EQ(Scheduler::kCoupled, data->placement_);
      DCHECK_EQ(Scheduler::kFixed, placement);
      Node* control = NodeProperties::GetControlInput(node);
      BasicBlock* block = schedule_->block(control);
      schedule_->AddNode(block, node);
      break;
    }
#define DEFINE_CONTROL_CASE(V) case IrOpcode::k##V:
      CONTROL_OP_LIST(DEFINE_CONTROL_CASE)
#undef DEFINE_CONTROL_CASE
      {
        // Control nodes force coupled uses to be placed.
        for (auto use : node->uses()) {
          if (GetPlacement(use) == Scheduler::kCoupled) {
            DCHECK_EQ(node, NodeProperties::GetControlInput(use));
            UpdatePlacement(use, placement);
          }
      }
      break;
    }
    default:
      DCHECK_EQ(Scheduler::kSchedulable, data->placement_);
      DCHECK_EQ(Scheduler::kScheduled, placement);
      break;
  }
  // Reduce the use count of the node's inputs to potentially make them
  // schedulable. If all the uses of a node have been scheduled, then the node
  // itself can be scheduled.

Schedule 最终产出什么？

Pipeline 中多次调用 Scheduler::ComputeSchedule

    reducer.ReduceGraph();
    // TODO(turbofan): Turn this into a debug mode check once we have
    // confidence.
    escape_reducer.VerifyReplacement();
  }
};

struct TypeAssertionsPhase {
  DECL_PIPELINE_PHASE_CONSTANTS(TypeAssertions)

  void Run(TFPipelineData* data, Zone* temp_zone) {
    Schedule* schedule = Scheduler::ComputeSchedule(
        temp_zone, data->graph(), Scheduler::kTempSchedule,
        &data->info()->tick_counter(), data->profile_data());

    AddTypeAssertions(data->jsgraph(), schedule, temp_zone);
  }
};

struct SimplifiedLoweringPhase {
  DECL_PIPELINE_PHASE_CONSTANTS(SimplifiedLowering)

  void Run(TFPipelineData* data, Zone* temp_zone, Linkage* linkage) {
    SimplifiedLowering lowering(data->jsgraph(), data->broker(), temp_zone,
                                data->source_positions(), data->node_origins(),
                                &data->info()->tick_counter(), linkage,
                                data->info(), data->observe_node_manager());

    }
  }
};
#endif  // V8_ENABLE_WEBASSEMBLY

struct ComputeSchedulePhase {
  DECL_PIPELINE_PHASE_CONSTANTS(Scheduling)

  void Run(TFPipelineData* data, Zone* temp_zone) {
    Schedule* schedule = Scheduler::ComputeSchedule(
        temp_zone, data->graph(),
        data->info()->splitting() ? Scheduler::kSplitNodes
                                  : Scheduler::kNoFlags,
        &data->info()->tick_counter(), data->profile_data());
    data->set_schedule(schedule);
  }
};

#if V8_ENABLE_WASM_SIMD256_REVEC
struct RevectorizePhase {
  DECL_PIPELINE_PHASE_CONSTANTS(Revectorizer)

  void Run(TFPipelineData* data, Zone* temp_zone) {
    Revectorizer revec(temp_zone, data->graph(), data->mcgraph(),
                       data->source_positions());
    revec.TryRevectorize(data->info()->GetDebugName().get());
  }
};
#endif  // V8_ENABLE_WASM_SIMD256_REVEC

struct PrintGraphPhase {
  DECL_PIPELINE_PHASE_CONSTANTS(PrintGraph)

  void Run(TFPipelineData* data, Zone* temp_zone, const char* phase) {
    OptimizedCompilationInfo* info = data->info();
    TFGraph* graph = data->graph();
    if (info->trace_turbo_json()) {  // Print JSON.
      UnparkedScopeIfNeeded scope(data->broker());
      AllowHandleDereference allow_deref;

      TurboJsonFile json_of(info, std::ios_base::app);
      json_of << "{\"name\":\"" << phase << "\",\"type\":\"graph\",\"data\":"
              << AsJSON(*graph, data->source_positions(), data->node_origins())
              << "},\n";
    }

    if (info->trace_turbo_scheduled()) {
      Schedule* schedule = data->schedule();
      if (schedule == nullptr) {
        schedule = Scheduler::ComputeSchedule(
            temp_zone, data->graph(), Scheduler::kNoFlags,
            &info->tick_counter(), data->profile_data());
      }

      UnparkedScopeIfNeeded scope(data->broker());
      AllowHandleDereference allow_deref;
      CodeTracer::StreamScope tracing_scope(data->GetCodeTracer());
      tracing_scope.stream()
          << "----- Graph after " << phase << " ----- " << std::endl
          << AsScheduledGraph(schedule);
    } else if (info->trace_turbo_graph()) {  // Simple textual RPO.

Schedule 之后：进入真正的后端世界

V8 优化器真正赚到的性能来自哪？

为什么 JS 这种动态语言也能跑得很快

设计模式：Pipeline + Pass Manager 思想

设计模式：Visitor / Reducer / Worklist

核心对象 UML

一次热点函数优化的时序图

细节：为什么 ReplaceWithValue 如此复杂

细节：为什么 Lowering 里有大量特化函数

细节：Scheduler 不是简单拓扑排序

阅读 V8 优化源码的常见误区

调优反模式：只盯单个函数

如何设计自己的 Benchmark 去观察这些优化

调试 V8 优化器常用方法

对业务工程的启示：把优化做成管线

对编译器学习者的启示

阅读这类源码的最佳实践

一句话总结 Turbofan 的优化哲学

它的本质不是“黑魔法优化”，而是大量工程化良好的中间表示转换。

本次源码解读带走什么

扩展阅读

完