Initial implementation of an edge-labeled instruction flow graph.

The flow graph is built by walking the AST.  Edges are labeled with
instructions (AST nodes).  Normal nodes have a single predecessor edge and a
single (labeled) successor edge.  Branch nodes are explicit, they have a
single predecessor edge and a pair of (unlabeled) successor edges.  Merge
nodes are explicit, they have a pair of predecessor edges and a single
(unlabeled) successor edge.

There is a distinguished (normal) entry node and a distinguished (special)
exit node with arbitrarily many predecessor edges and no successor edges.

The graph is intended to support graph-based analysis and transformation.

Review URL: http://codereview.chromium.org/660449

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@4051 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/ast.h

@@ -117,6 +117,9 @@ typedef ZoneList<Handle<Object> > ZoneObjectList;
 
 class AstNode: public ZoneObject {
  public:
+  static const int kNoNumber = -1;
+
+  AstNode() : num_(kNoNumber) {}
   virtual ~AstNode() { }
   virtual void Accept(AstVisitor* v) = 0;
 
@@ -141,6 +144,13 @@ class AstNode: public ZoneObject {
   virtual ObjectLiteral* AsObjectLiteral() { return NULL; }
   virtual ArrayLiteral* AsArrayLiteral() { return NULL; }
   virtual CompareOperation* AsCompareOperation() { return NULL; }
+
+  int num() { return num_; }
+  void set_num(int n) { num_ = n; }
+
+ private:
+  // Support for ast node numbering.
+  int num_;
 };
 
 
@@ -181,11 +191,8 @@ class Expression: public AstNode {
     kTestValue
   };
 
-  static const int kNoLabel = -1;
-
   Expression()
       : bitfields_(0),
-        num_(kNoLabel),
         def_(NULL),
         defined_vars_(NULL) {}
 
@@ -215,11 +222,6 @@ class Expression: public AstNode {
   // Static type information for this expression.
   StaticType* type() { return &type_; }
 
-  int num() { return num_; }
-
-  // AST node numbering ordered by evaluation order.
-  void set_num(int n) { num_ = n; }
-
   // Data flow information.
   DefinitionInfo* var_def() { return def_; }
   void set_var_def(DefinitionInfo* def) { def_ = def; }
@@ -261,7 +263,6 @@ class Expression: public AstNode {
   uint32_t bitfields_;
   StaticType type_;
 
-  int num_;
   DefinitionInfo* def_;
   ZoneList<DefinitionInfo*>* defined_vars_;
 

src/compiler.cc

@@ -31,6 +31,7 @@
 #include "codegen-inl.h"
 #include "compilation-cache.h"
 #include "compiler.h"
+#include "data-flow.h"
 #include "debug.h"
 #include "fast-codegen.h"
 #include "full-codegen.h"
@@ -79,6 +80,17 @@ static Handle<Code> MakeCode(Handle<Context> context, CompilationInfo* info) {
     return Handle<Code>::null();
   }
 
+  if (FLAG_use_flow_graph) {
+    FlowGraphBuilder builder;
+    builder.Build(function);
+
+#ifdef DEBUG
+    if (FLAG_print_graph_text) {
+      builder.graph()->PrintText(builder.postorder());
+    }
+#endif
+  }
+
   // Generate code and return it.  Code generator selection is governed by
   // which backends are enabled and whether the function is considered
   // run-once code or not:
@@ -452,6 +464,17 @@ Handle<JSFunction> Compiler::BuildBoilerplate(FunctionLiteral* literal,
       return Handle<JSFunction>::null();
     }
 
+    if (FLAG_use_flow_graph) {
+      FlowGraphBuilder builder;
+      builder.Build(literal);
+
+#ifdef DEBUG
+      if (FLAG_print_graph_text) {
+        builder.graph()->PrintText(builder.postorder());
+      }
+#endif
+    }
+
     // Generate code and return it.  The way that the compilation mode
     // is controlled by the command-line flags is described in
     // the static helper function MakeCode.

src/data-flow.h

@@ -95,6 +95,299 @@ class BitVector {
 };
 
 
+// Forward declarations of Node types.
+class Node;
+class BranchNode;
+class JoinNode;
+
+// Flow graphs have a single entry and single exit.  The empty flowgraph is
+// represented by both entry and exit being NULL.
+class FlowGraph BASE_EMBEDDED {
+ public:
+  FlowGraph() : entry_(NULL), exit_(NULL) {}
+
+  static FlowGraph Empty() { return FlowGraph(); }
+
+  bool is_empty() const { return entry_ == NULL; }
+  Node* entry() const { return entry_; }
+  Node* exit() const { return exit_; }
+
+  // Add a single instruction to the end of this flowgraph.
+  void AppendInstruction(AstNode* instruction);
+
+  // Add a single node to the end of this flow graph.
+  void AppendNode(Node* node);
+
+  // Add a flow graph fragment to the end of this one.
+  void AppendGraph(FlowGraph* graph);
+
+  // Concatenate an if-then-else flow-graph to this one.  Control is split
+  // and merged, so the graph remains single-entry, single-exit.
+  void Split(BranchNode* branch,
+             FlowGraph* left,
+             FlowGraph* right,
+             JoinNode* merge);
+
+  // Concatenate a forward loop (e.g., while or for loop) flow-graph to this
+  // one.  Control is split by the condition and merged back from the back
+  // edge at end of the body to the beginning of the condition.  The single
+  // (free) exit of the result graph is the right (false) arm of the branch
+  // node.
+  void Loop(JoinNode* merge,
+            FlowGraph* condition,
+            BranchNode* branch,
+            FlowGraph* body);
+
+#ifdef DEBUG
+  void PrintText(ZoneList<Node*>* postorder);
+#endif
+
+ private:
+  Node* entry_;
+  Node* exit_;
+};
+
+
+// Flow-graph nodes.
+class Node: public ZoneObject {
+ public:
+  Node() : number_(-1), mark_(false) {}
+
+  virtual ~Node() {}
+
+  virtual bool IsBlockNode() { return false; }
+  virtual bool IsJoinNode() { return false; }
+
+  virtual void AddPredecessor(Node* predecessor) = 0;
+  virtual void AddSuccessor(Node* successor) = 0;
+
+  bool IsMarkedWith(bool mark) { return mark_ == mark; }
+  void MarkWith(bool mark) { mark_ = mark; }
+
+  // Perform a depth first search and record preorder and postorder
+  // traversal orders.
+  virtual void Traverse(bool mark,
+                        ZoneList<Node*>* preorder,
+                        ZoneList<Node*>* postorder) = 0;
+
+  int number() { return number_; }
+  void set_number(int number) { number_ = number; }
+
+#ifdef DEBUG
+  virtual void AssignNumbers();
+  virtual void PrintText() = 0;
+#endif
+
+ private:
+  int number_;
+  bool mark_;
+
+  DISALLOW_COPY_AND_ASSIGN(Node);
+};
+
+
+// An entry node has no predecessors and a single successor.
+class EntryNode: public Node {
+ public:
+  EntryNode() : successor_(NULL) {}
+
+  void AddPredecessor(Node* predecessor) { UNREACHABLE(); }
+
+  void AddSuccessor(Node* successor) {
+    ASSERT(successor_ == NULL && successor != NULL);
+    successor_ = successor;
+  }
+
+  void Traverse(bool mark,
+                ZoneList<Node*>* preorder,
+                ZoneList<Node*>* postorder);
+
+#ifdef DEBUG
+  void PrintText();
+#endif
+
+ private:
+  Node* successor_;
+
+  DISALLOW_COPY_AND_ASSIGN(EntryNode);
+};
+
+
+// An exit node has a arbitrarily many predecessors and no successors.
+class ExitNode: public Node {
+ public:
+  ExitNode() : predecessors_(4) {}
+
+  void AddPredecessor(Node* predecessor) {
+    ASSERT(predecessor != NULL);
+    predecessors_.Add(predecessor);
+  }
+
+  void AddSuccessor(Node* successor) { /* Do nothing. */ }
+
+  void Traverse(bool mark,
+                ZoneList<Node*>* preorder,
+                ZoneList<Node*>* postorder);
+
+#ifdef DEBUG
+  void PrintText();
+#endif
+
+ private:
+  ZoneList<Node*> predecessors_;
+
+  DISALLOW_COPY_AND_ASSIGN(ExitNode);
+};
+
+
+// Block nodes have a single successor and predecessor and a list of
+// instructions.
+class BlockNode: public Node {
+ public:
+  BlockNode() : predecessor_(NULL), successor_(NULL), instructions_(4) {}
+
+  static BlockNode* cast(Node* node) {
+    ASSERT(node->IsBlockNode());
+    return reinterpret_cast<BlockNode*>(node);
+  }
+
+  bool IsBlockNode() { return true; }
+
+  void AddPredecessor(Node* predecessor) {
+    ASSERT(predecessor_ == NULL && predecessor != NULL);
+    predecessor_ = predecessor;
+  }
+
+  void AddSuccessor(Node* successor) {
+    ASSERT(successor_ == NULL && successor != NULL);
+    successor_ = successor;
+  }
+
+  void AddInstruction(AstNode* instruction) {
+    instructions_.Add(instruction);
+  }
+
+  void Traverse(bool mark,
+                ZoneList<Node*>* preorder,
+                ZoneList<Node*>* postorder);
+
+#ifdef DEBUG
+  void AssignNumbers();
+  void PrintText();
+#endif
+
+ private:
+  Node* predecessor_;
+  Node* successor_;
+  ZoneList<AstNode*> instructions_;
+
+  DISALLOW_COPY_AND_ASSIGN(BlockNode);
+};
+
+
+// Branch nodes have a single predecessor and a pair of successors.
+class BranchNode: public Node {
+ public:
+  BranchNode() : predecessor_(NULL), successor0_(NULL), successor1_(NULL) {}
+
+  void AddPredecessor(Node* predecessor) {
+    ASSERT(predecessor_ == NULL && predecessor != NULL);
+    predecessor_ = predecessor;
+  }
+
+  void AddSuccessor(Node* successor) {
+    ASSERT(successor1_ == NULL && successor != NULL);
+    if (successor0_ == NULL) {
+      successor0_ = successor;
+    } else {
+      successor1_ = successor;
+    }
+  }
+
+  void Traverse(bool mark,
+                ZoneList<Node*>* preorder,
+                ZoneList<Node*>* postorder);
+
+#ifdef DEBUG
+  void PrintText();
+#endif
+
+ private:
+  Node* predecessor_;
+  Node* successor0_;
+  Node* successor1_;
+
+  DISALLOW_COPY_AND_ASSIGN(BranchNode);
+};
+
+
+// Join nodes have arbitrarily many predecessors and a single successor.
+class JoinNode: public Node {
+ public:
+  JoinNode() : predecessors_(2), successor_(NULL) {}
+
+  static JoinNode* cast(Node* node) {
+    ASSERT(node->IsJoinNode());
+    return reinterpret_cast<JoinNode*>(node);
+  }
+
+  bool IsJoinNode() { return true; }
+
+  void AddPredecessor(Node* predecessor) {
+    ASSERT(predecessor != NULL);
+    predecessors_.Add(predecessor);
+  }
+
+  void AddSuccessor(Node* successor) {
+    ASSERT(successor_ == NULL && successor != NULL);
+    successor_ = successor;
+  }
+
+  void Traverse(bool mark,
+                ZoneList<Node*>* preorder,
+                ZoneList<Node*>* postorder);
+
+#ifdef DEBUG
+  void PrintText();
+#endif
+
+ private:
+  ZoneList<Node*> predecessors_;
+  Node* successor_;
+
+  DISALLOW_COPY_AND_ASSIGN(JoinNode);
+};
+
+
+// Construct a flow graph from a function literal.  Build pre- and postorder
+// traversal orders as a byproduct.
+class FlowGraphBuilder: public AstVisitor {
+ public:
+  FlowGraphBuilder() : global_exit_(NULL), preorder_(4), postorder_(4) {}
+
+  void Build(FunctionLiteral* lit);
+
+  FlowGraph* graph() { return &graph_; }
+
+  ZoneList<Node*>* postorder() { return &postorder_; }
+
+ private:
+  ExitNode* global_exit() { return global_exit_; }
+
+  // AST node visit functions.
+#define DECLARE_VISIT(type) virtual void Visit##type(type* node);
+  AST_NODE_LIST(DECLARE_VISIT)
+#undef DECLARE_VISIT
+
+  FlowGraph graph_;
+  ExitNode* global_exit_;
+  ZoneList<Node*> preorder_;
+  ZoneList<Node*> postorder_;
+
+  DISALLOW_COPY_AND_ASSIGN(FlowGraphBuilder);
+};
+
+
 // This class is used to number all expressions in the AST according to
 // their evaluation order (post-order left-to-right traversal).
 class AstLabeler: public AstVisitor {

src/flag-definitions.h

@@ -153,6 +153,7 @@ DEFINE_bool(always_fast_compiler, false,
             "try to use the speculative optimizing backend for all code")
 DEFINE_bool(trace_bailout, false,
             "print reasons for falling back to using the classic V8 backend")
+DEFINE_bool(use_flow_graph, false, "perform flow-graph based optimizations")
 
 // compilation-cache.cc
 DEFINE_bool(compilation_cache, true, "enable compilation cache")
@@ -305,6 +306,8 @@ DEFINE_string(stop_at, "", "function name where to insert a breakpoint")
 DEFINE_bool(print_builtin_scopes, false, "print scopes for builtins")
 DEFINE_bool(print_scopes, false, "print scopes")
 DEFINE_bool(print_ir, false, "print the AST as seen by the backend")
+DEFINE_bool(print_graph_text, false,
+            "print a text representation of the flow graph")
 
 // contexts.cc
 DEFINE_bool(trace_contexts, false, "trace contexts operations")

src/prettyprinter.cc

@@ -604,7 +604,7 @@ class IndentedScope BASE_EMBEDDED {
         ast_printer_->Print(StaticType::Type2String(expr->type()));
         printed_first = true;
       }
-      if (expr->num() != Expression::kNoLabel) {
+      if (expr->num() != AstNode::kNoNumber) {
         ast_printer_->Print(printed_first ? ", num = " : " (num = ");
         ast_printer_->Print("%d", expr->num());
         printed_first = true;
@@ -679,7 +679,7 @@ void AstPrinter::PrintLiteralWithModeIndented(const char* info,
       pos += OS::SNPrintF(buf + pos, ", type = %s",
                           StaticType::Type2String(type));
     }
-    if (num != Expression::kNoLabel) {
+    if (num != AstNode::kNoNumber) {
       pos += OS::SNPrintF(buf + pos, ", num = %d", num);
     }
     OS::SNPrintF(buf + pos, ")");
@@ -740,7 +740,7 @@ void AstPrinter::PrintParameters(Scope* scope) {
       PrintLiteralWithModeIndented("VAR", scope->parameter(i),
                                    scope->parameter(i)->name(),
                                    scope->parameter(i)->type(),
-                                   Expression::kNoLabel);
+                                   AstNode::kNoNumber);
     }
   }
 }
@@ -786,7 +786,7 @@ void AstPrinter::VisitDeclaration(Declaration* node) {
                                  node->proxy()->AsVariable(),
                                  node->proxy()->name(),
                                  node->proxy()->AsVariable()->type(),
-                                 Expression::kNoLabel);
+                                 AstNode::kNoNumber);
   } else {
     // function declarations
     PrintIndented("FUNCTION ");