6674588: (Escape Analysis) Improve Escape Analysis code

Summary: Current EA code has several problems which have to be fixed.
Reviewed-by: jrose, sgoldman

src/share/vm/includeDB_compiler2

@@ -164,6 +164,7 @@ callGenerator.hpp                       deoptimization.hpp
 callGenerator.hpp                       type.hpp
 
 callnode.cpp                            callnode.hpp
+callnode.cpp                            bcEscapeAnalyzer.hpp
 callnode.cpp                            escape.hpp
 callnode.cpp                            locknode.hpp
 callnode.cpp                            machnode.hpp
@@ -176,7 +177,6 @@ callnode.cpp                            rootnode.hpp
 callnode.cpp                            runtime.hpp
 
 callnode.hpp                            connode.hpp
-callnode.hpp                            escape.hpp
 callnode.hpp                            mulnode.hpp
 callnode.hpp                            multnode.hpp
 callnode.hpp                            opcodes.hpp
@@ -347,7 +347,6 @@ connode.cpp                             addnode.hpp
 connode.cpp                             allocation.inline.hpp
 connode.cpp                             compile.hpp
 connode.cpp                             connode.hpp
-connode.cpp                             escape.hpp
 connode.cpp                             machnode.hpp
 connode.cpp                             matcher.hpp
 connode.cpp                             memnode.hpp
@@ -844,7 +843,6 @@ phaseX.cpp                              block.hpp
 phaseX.cpp                              callnode.hpp
 phaseX.cpp                              cfgnode.hpp
 phaseX.cpp                              connode.hpp
-phaseX.cpp                              escape.hpp
 phaseX.cpp                              loopnode.hpp
 phaseX.cpp                              machnode.hpp
 phaseX.cpp                              opcodes.hpp

src/share/vm/opto/c2_globals.hpp

@@ -382,6 +382,12 @@
   product(bool, EliminateAllocations, true,                                 \
           "Use escape analysis to eliminate allocations")                   \
                                                                             \
+  notproduct(bool, PrintEliminateAllocations, false,                        \
+          "Print out when allocations are eliminated")                      \
+                                                                            \
+  product(intx, EliminateAllocationArraySizeLimit, 64,                      \
+          "Array size (number of elements) limit for scalar replacement")   \
+                                                                            \
   product(intx, MaxLabelRootDepth, 1100,                                    \
           "Maximum times call Label_Root to prevent stack overflow")        \
 

src/share/vm/opto/callnode.cpp

@@ -624,6 +624,87 @@ uint CallNode::match_edge(uint idx) const {
   return 0;
 }
 
+//
+// Determine whether the call could modify a memory value  of the
+// specified address type
+//
+bool CallNode::may_modify(const TypePtr *addr_t, PhaseTransform *phase) {
+  const TypeOopPtr *adrInst_t  = addr_t->isa_oopptr();
+
+  // if not an InstPtr or not an instance type, assume the worst
+  if (adrInst_t == NULL || !adrInst_t->is_instance_field()) {
+    return true;
+  }
+  Compile *C = phase->C;
+  int offset = adrInst_t->offset();
+  assert(offset >= 0, "should be valid offset");
+  assert(addr_t->isa_instptr() || addr_t->isa_aryptr(), "only instances or arrays are expected");
+
+  int base_idx = C->get_alias_index(adrInst_t);
+  ciMethod * meth = is_CallStaticJava() ?  as_CallStaticJava()->method() : NULL;
+  BCEscapeAnalyzer *bcea = (meth != NULL) ? meth->get_bcea() : NULL;
+
+  const TypeTuple * d = tf()->domain();
+  for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
+    const Type* t = d->field_at(i);
+    Node *arg = in(i);
+    const Type *at = phase->type(arg);
+    if (at == TypePtr::NULL_PTR || at == Type::TOP)
+      continue;  // null can't affect anything
+
+    const TypeOopPtr *at_ptr = at->isa_oopptr();
+    if (!arg->is_top() && (t->isa_oopptr() != NULL ||
+                           t->isa_ptr() && at_ptr != NULL)) {
+      assert(at_ptr != NULL, "expecting an OopPtr");
+      // If we have found an argument matching adr_base_t, check if the field
+      // at the specified offset is modified.  Since we don't know the size,
+      // assume 8.
+      int at_idx = C->get_alias_index(at_ptr->add_offset(offset)->isa_oopptr());
+      if (base_idx == at_idx &&
+          (bcea == NULL ||
+           bcea->is_arg_modified(i - TypeFunc::Parms, offset, 8))) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+// Does this call have a direct reference to n other than debug information?
+bool CallNode::has_non_debug_use(Node *n) {
+  const TypeTuple * d = tf()->domain();
+  for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
+    Node *arg = in(i);
+    if (arg == n) {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Returns the unique CheckCastPP of a call
+// or 'this' if there are several CheckCastPP
+// or returns NULL if there is no one.
+Node *CallNode::result_cast() {
+  Node *cast = NULL;
+
+  Node *p = proj_out(TypeFunc::Parms);
+  if (p == NULL)
+    return NULL;
+
+  for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
+    Node *use = p->fast_out(i);
+    if (use->is_CheckCastPP()) {
+      if (cast != NULL) {
+        return this;  // more than 1 CheckCastPP
+      }
+      cast = use;
+    }
+  }
+  return cast;
+}
+
+
 //=============================================================================
 uint CallJavaNode::size_of() const { return sizeof(*this); }
 uint CallJavaNode::cmp( const Node &n ) const {

src/share/vm/opto/callnode.hpp

@@ -388,6 +388,9 @@ public:
   void               set_next_exception(SafePointNode* n);
   bool                   has_exceptions() const { return next_exception() != NULL; }
 
+  // Does this node have a use of n other than in debug information?
+  virtual bool           has_non_debug_use(Node *n)  {return false; }
+
   // Standard Node stuff
   virtual int            Opcode() const;
   virtual bool           pinned() const { return true; }
@@ -457,7 +460,6 @@ public:
   const TypeFunc *_tf;        // Function type
   address      _entry_point;  // Address of method being called
   float        _cnt;          // Estimate of number of times called
-  PointsToNode::EscapeState _escape_state;
 
   CallNode(const TypeFunc* tf, address addr, const TypePtr* adr_type)
     : SafePointNode(tf->domain()->cnt(), NULL, adr_type),
@@ -467,7 +469,6 @@ public:
   {
     init_class_id(Class_Call);
     init_flags(Flag_is_Call);
-    _escape_state = PointsToNode::UnknownEscape;
   }
 
   const TypeFunc* tf()        const { return _tf; }
@@ -493,6 +494,15 @@ public:
   // the node the JVMState must be cloned.
   virtual void        clone_jvms() { }   // default is not to clone
 
+  // Returns true if the call may modify n
+  virtual bool        may_modify(const TypePtr *addr_t, PhaseTransform *phase);
+  // Does this node have a use of n other than in debug information?
+  virtual bool        has_non_debug_use(Node *n);
+  // Returns the unique CheckCastPP of a call
+  // or result projection is there are several CheckCastPP
+  // or returns NULL if there is no one.
+  Node *result_cast();
+
   virtual uint match_edge(uint idx) const;
 
 #ifndef PRODUCT
@@ -689,6 +699,9 @@ public:
   virtual uint ideal_reg() const { return Op_RegP; }
   virtual bool        guaranteed_safepoint()  { return false; }
 
+  // allocations do not modify their arguments
+  virtual bool        may_modify(const TypePtr *addr_t, PhaseTransform *phase) { return false;}
+
   // Pattern-match a possible usage of AllocateNode.
   // Return null if no allocation is recognized.
   // The operand is the pointer produced by the (possible) allocation.
@@ -801,6 +814,9 @@ public:
   // mark node as eliminated and update the counter if there is one
   void set_eliminated();
 
+  // locking does not modify its arguments
+  virtual bool        may_modify(const TypePtr *addr_t, PhaseTransform *phase){ return false;}
+
 #ifndef PRODUCT
   void create_lock_counter(JVMState* s);
   NamedCounter* counter() const { return _counter; }

src/share/vm/opto/compile.cpp

@@ -407,11 +407,6 @@ uint Compile::scratch_emit_size(const Node* n) {
   return buf.code_size();
 }
 
-void  Compile::record_for_escape_analysis(Node* n) {
-  if (_congraph != NULL)
-    _congraph->record_for_escape_analysis(n);
-}
-
 
 // ============================================================================
 //------------------------------Compile standard-------------------------------
@@ -494,9 +489,6 @@ Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr
   PhaseGVN gvn(node_arena(), estimated_size);
   set_initial_gvn(&gvn);
 
-  if (_do_escape_analysis)
-    _congraph = new ConnectionGraph(this);
-
   { // Scope for timing the parser
     TracePhase t3("parse", &_t_parser, true);
 
@@ -581,6 +573,8 @@ Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr
   NOT_PRODUCT( verify_graph_edges(); )
 
   // Perform escape analysis
+  if (_do_escape_analysis)
+    _congraph = new ConnectionGraph(this);
   if (_congraph != NULL) {
     NOT_PRODUCT( TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, TimeCompiler); )
     _congraph->compute_escape();

src/share/vm/opto/compile.hpp

@@ -485,7 +485,6 @@ class Compile : public Phase {
   PhaseGVN*         initial_gvn()               { return _initial_gvn; }
   Unique_Node_List* for_igvn()                  { return _for_igvn; }
   inline void       record_for_igvn(Node* n);   // Body is after class Unique_Node_List.
-  void              record_for_escape_analysis(Node* n);
   void          set_initial_gvn(PhaseGVN *gvn)           { _initial_gvn = gvn; }
   void          set_for_igvn(Unique_Node_List *for_igvn) { _for_igvn = for_igvn; }
 

src/share/vm/opto/escape.hpp

@@ -25,14 +25,15 @@
 //
 // Adaptation for C2 of the escape analysis algorithm described in:
 //
-//     [Choi99] Jong-Deok Shoi, Manish Gupta, Mauricio Seffano, Vugranam C. Sreedhar,
-//              Sam Midkiff,  "Escape Analysis for Java", Procedings of ACM SIGPLAN
+// [Choi99] Jong-Deok Shoi, Manish Gupta, Mauricio Seffano,
+//          Vugranam C. Sreedhar, Sam Midkiff,
+//          "Escape Analysis for Java", Procedings of ACM SIGPLAN
 //          OOPSLA  Conference, November 1, 1999
 //
 // The flow-insensitive analysis described in the paper has been implemented.
 //
-// The analysis requires construction of a "connection graph" (CG) for the method being
-// analyzed.  The nodes of the connection graph are:
+// The analysis requires construction of a "connection graph" (CG) for
+// the method being analyzed.  The nodes of the connection graph are:
 //
 //     -  Java objects (JO)
 //     -  Local variables (LV)
@@ -40,7 +41,7 @@
 //
 // The CG contains 3 types of edges:
 //
-//   -  PointsTo  (-P>)     {LV,OF}  to JO
+//   -  PointsTo  (-P>)    {LV, OF} to JO
 //   -  Deferred  (-D>)    from {LV, OF} to {LV, OF}
 //   -  Field     (-F>)    from JO to OF
 //
@@ -49,7 +50,8 @@
 //   PointsTo(n) - n is any CG node, it returns the set of JO that n could
 //                 point to.
 //
-// The algorithm describes how to construct the connection graph in the following 4 cases:
+// The algorithm describes how to construct the connection graph
+// in the following 4 cases:
 //
 //          Case                  Edges Created
 //
@@ -58,29 +60,32 @@
 // (3)   p.f = q                    JO -F> OF,  OF -D> LV
 // (4)   p   = q.f                  JO -F> OF,  LV -D> OF
 //
-// In all these cases, p and q are local variables.  For static field references, we can
-// construct a local variable containing a reference to the static memory.
+// In all these cases, p and q are local variables.  For static field
+// references, we can construct a local variable containing a reference
+// to the static memory.
 //
 // C2 does not have local variables.  However for the purposes of constructing
 // the connection graph, the following IR nodes are treated as local variables:
 //     Phi    (pointer values)
 //     LoadP
-//     Proj  (value returned from callnodes including allocations)
-//     CheckCastPP
+//     Proj#5 (value returned from callnodes including allocations)
+//     CheckCastPP, CastPP
 //
-// The LoadP, Proj and CheckCastPP behave like variables assigned to only once.  Only
-// a Phi can have multiple assignments.  Each input to a Phi is treated
+// The LoadP, Proj and CheckCastPP behave like variables assigned to only once.
+// Only a Phi can have multiple assignments.  Each input to a Phi is treated
 // as an assignment to it.
 //
-// The following note types are JavaObject:
+// The following node types are JavaObject:
 //
 //     top()
 //     Allocate
 //     AllocateArray
 //     Parm  (for incoming arguments)
+//     CastX2P ("unsafe" operations)
 //     CreateEx
 //     ConP
 //     LoadKlass
+//     ThreadLocal
 //
 // AddP nodes are fields.
 //
@@ -89,7 +94,7 @@
 // source.  This results in a graph with no deferred edges, only:
 //
 //    LV -P> JO
-//    OF -P> JO
+//    OF -P> JO (the object whose oop is stored in the field)
 //    JO -F> OF
 //
 // Then, for each node which is GlobalEscape, anything it could point to
@@ -118,9 +123,10 @@ public:
 
   typedef enum {
     UnknownEscape = 0,
-    NoEscape      = 1,
-    ArgEscape     = 2,
-    GlobalEscape  = 3
+    NoEscape      = 1, // A scalar replaceable object with unique type.
+    ArgEscape     = 2, // An object passed as argument or referenced by
+                       // argument (and not globally escape during call).
+    GlobalEscape  = 3  // An object escapes the method and thread.
   } EscapeState;
 
   typedef enum {
@@ -141,18 +147,24 @@ private:
   NodeType             _type;
   EscapeState          _escape;
   GrowableArray<uint>* _edges;   // outgoing edges
-  int                  _offset; // for fields
 
-  bool       _unique_type;       // For allocated objects, this node may be a unique type
 public:
-  Node*      _node;              // Ideal node corresponding to this PointsTo node
-  int        _inputs_processed;  // the number of Phi inputs that have been processed so far
-  bool       _hidden_alias;      // this node is an argument to a function which may return it
-                                 // creating a hidden alias
+  Node* _node;              // Ideal node corresponding to this PointsTo node.
+  int   _offset;            // Object fields offsets.
+  bool  _scalar_replaceable;// Not escaped object could be replaced with scalar
+  bool  _hidden_alias;      // This node is an argument to a function.
+                            // which may return it creating a hidden alias.
+
+  PointsToNode():
+    _type(UnknownType),
+    _escape(UnknownEscape),
+    _edges(NULL),
+    _node(NULL),
+    _offset(-1),
+    _scalar_replaceable(true),
+    _hidden_alias(false) {}
 
 
-  PointsToNode(): _offset(-1), _type(UnknownType), _escape(UnknownEscape), _edges(NULL), _node(NULL), _inputs_processed(0), _hidden_alias(false), _unique_type(true) {}
-
   EscapeState escape_state() const { return _escape; }
   NodeType node_type() const { return _type;}
   int offset() { return _offset;}
@@ -182,21 +194,27 @@ public:
 
 class ConnectionGraph: public ResourceObj {
 private:
-  enum {
-    INITIAL_NODE_COUNT = 100                    // initial size of _nodes array
-  };
+  GrowableArray<PointsToNode>* _nodes; // Connection graph nodes indexed
+                                       // by ideal node index.
+
+  Unique_Node_List  _delayed_worklist; // Nodes to be processed before
+                                       // the call build_connection_graph().
 
+  VectorSet                _processed; // Records which nodes have been
+                                       // processed.
+
+  bool                    _collecting; // Indicates whether escape information
+                                       // is still being collected. If false,
+                                       // no new nodes will be processed.
+
+  bool               _has_allocations; // Indicates whether method has any
+                                       // non-escaping allocations.
+
+  uint                _phantom_object; // Index of globally escaping object
+                                       // that pointer values loaded from
+                                       // a field which has not been set
+                                       // are assumed to point to.
 
-  GrowableArray<PointsToNode>* _nodes;          // connection graph nodes  Indexed by ideal
-                                                // node index
-  Unique_Node_List             _deferred;       // Phi's to be processed after parsing
-  VectorSet                    _processed;      // records which nodes have been processed
-  bool                         _collecting;     // indicates whether escape information is
-                                                // still being collected.  If false, no new
-                                                // nodes will be processed
-  uint                         _phantom_object; // index of globally escaping object that
-                                                // pointer values loaded from a field which
-                                                // has not been set are assumed to point to
   Compile *                  _compile; // Compile object for current compilation
 
   // address of an element in _nodes.  Used when the element is to be modified
@@ -208,8 +226,11 @@ private:
     return _nodes->adr_at(idx);
   }
 
+  // Add node to ConnectionGraph.
+  void add_node(Node *n, PointsToNode::NodeType nt, PointsToNode::EscapeState es, bool done);
+
   // offset of a field reference
-  int type_to_offset(const Type *t);
+  int address_offset(Node* adr, PhaseTransform *phase);
 
   // compute the escape state for arguments to a call
   void process_call_arguments(CallNode *call, PhaseTransform *phase);
@@ -217,12 +238,11 @@ private:
   // compute the escape state for the return value of a call
   void process_call_result(ProjNode *resproj, PhaseTransform *phase);
 
-  // compute the escape state of a Phi.  This may be called multiple
-  // times as new inputs are added to the Phi.
-  void process_phi_escape(PhiNode *phi, PhaseTransform *phase);
+  // Populate Connection Graph with Ideal nodes.
+  void record_for_escape_analysis(Node *n, PhaseTransform *phase);
 
-  // compute the escape state of an ideal node.
-  void record_escape_work(Node *n, PhaseTransform *phase);
+  // Build Connection Graph and set nodes escape state.
+  void build_connection_graph(Node *n, PhaseTransform *phase);
 
   // walk the connection graph starting at the node corresponding to "n" and
   // add the index of everything it could point to, to "ptset".  This may cause
@@ -241,8 +261,8 @@ private:
   // a pointsto edge is added if it is a JavaObject
   void add_edge_from_fields(uint adr, uint to_i, int offs);
 
-  // Add a deferred  edge from node given by "from_i" to any field of adr_i whose offset
-  // matches "offset"
+  // Add a deferred  edge from node given by "from_i" to any field
+  // of adr_i whose offset matches "offset"
   void add_deferred_edge_to_fields(uint from_i, uint adr, int offs);
 
 
@@ -262,6 +282,8 @@ private:
   PhiNode *create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *>  &orig_phi_worklist, PhaseGVN  *igvn, bool &new_created);
   PhiNode *split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *>  &orig_phi_worklist, PhaseGVN  *igvn);
   Node *find_mem(Node *mem, int alias_idx, PhaseGVN  *igvn);
+  Node *find_inst_mem(Node *mem, int alias_idx,GrowableArray<PhiNode *>  &orig_phi_worklist,  PhaseGVN  *igvn);
+
   // Propagate unique types created for unescaped allocated objects
   // through the graph
   void split_unique_types(GrowableArray<Node *>  &alloc_worklist);
@@ -285,26 +307,24 @@ private:
   // Set the escape state of a node
   void set_escape_state(uint ni, PointsToNode::EscapeState es);
 
-  // bypass any casts and return the node they refer to
-  Node * skip_casts(Node *n);
-
   // Get Compile object for current compilation.
   Compile *C() const        { return _compile; }
 
 public:
   ConnectionGraph(Compile *C);
 
-  // record a Phi for later processing.
-  void record_for_escape_analysis(Node *n);
-
-  // process a node and  fill in its connection graph node
-  void record_escape(Node *n, PhaseTransform *phase);
-
-  // All nodes have been recorded, compute the escape information
+  // Compute the escape information
   void compute_escape();
 
   // escape state of a node
   PointsToNode::EscapeState escape_state(Node *n, PhaseTransform *phase);
+  // other information we have collected
+  bool is_scalar_replaceable(Node *n) {
+    if (_collecting)
+      return false;
+    PointsToNode  ptn = _nodes->at_grow(n->_idx);
+    return ptn.escape_state() == PointsToNode::NoEscape && ptn._scalar_replaceable;
+  }
 
   bool hidden_alias(Node *n) {
     if (_collecting)

src/share/vm/opto/node.cpp

@@ -812,8 +812,7 @@ int Node::disconnect_inputs(Node *n) {
 Node* Node::uncast() const {
   // Should be inline:
   //return is_ConstraintCast() ? uncast_helper(this) : (Node*) this;
-  if (is_ConstraintCast() ||
-      (is_Type() && req() == 2 && Opcode() == Op_CheckCastPP))
+  if (is_ConstraintCast() || is_CheckCastPP())
     return uncast_helper(this);
   else
     return (Node*) this;
@@ -827,7 +826,7 @@ Node* Node::uncast_helper(const Node* p) {
       break;
     } else if (p->is_ConstraintCast()) {
       p = p->in(1);
-    } else if (p->is_Type() && p->Opcode() == Op_CheckCastPP) {
+    } else if (p->is_CheckCastPP()) {
       p = p->in(1);
     } else {
       break;

src/share/vm/opto/node.hpp

@@ -1328,7 +1328,6 @@ public:
 // Inline definition of Compile::record_for_igvn must be deferred to this point.
 inline void Compile::record_for_igvn(Node* n) {
   _for_igvn->push(n);
-  record_for_escape_analysis(n);
 }
 
 //------------------------------Node_Stack-------------------------------------

src/share/vm/opto/phaseX.cpp

@@ -587,11 +587,6 @@ ConNode* PhaseValues::uncached_makecon(const Type *t) {
       Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
       loc->clear(); // do not put debug info on constants
     }
-    // Collect points-to information for escape analysys
-    ConnectionGraph *cgr = C->congraph();
-    if (cgr != NULL) {
-      cgr->record_escape(x, this);
-    }
   } else {
     x->destruct();              // Hit, destroy duplicate constant
     x = k;                      // use existing constant
@@ -714,12 +709,6 @@ Node *PhaseGVN::transform_no_reclaim( Node *n ) {
     return i;
   }
 
-  // Collect points-to information for escape analysys
-  ConnectionGraph *cgr = C->congraph();
-  if (cgr != NULL) {
-    cgr->record_escape(k, this);
-  }
-
   // Return Idealized original
   return k;
 }
@@ -1245,7 +1234,7 @@ void PhaseIterGVN::add_users_to_worklist( Node *n ) {
 
     uint use_op = use->Opcode();
     // If changed Cast input, check Phi users for simple cycles
-    if( use->is_ConstraintCast() || use->Opcode() == Op_CheckCastPP ) {
+    if( use->is_ConstraintCast() || use->is_CheckCastPP() ) {
       for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
         Node* u = use->fast_out(i2);
         if (u->is_Phi())

src/share/vm/runtime/arguments.cpp

@@ -1276,6 +1276,9 @@ void Arguments::set_aggressive_opts_flags() {
     sprintf(buffer, "java.lang.Integer.IntegerCache.high=%d", AutoBoxCacheMax);
     add_property(buffer);
   }
+  if (AggressiveOpts && FLAG_IS_DEFAULT(DoEscapeAnalysis)) {
+    FLAG_SET_DEFAULT(DoEscapeAnalysis, true);
+  }
 #endif
 
   if (AggressiveOpts) {