6672848: (Escape Analysis) improve lock elimination with EA

Summary: Remove lock/unlock MemBar nodes and specify locks in debug info for deoptimization.
Reviewed-by: never

src/share/vm/opto/callnode.cpp

@@ -1364,7 +1364,7 @@ void AbstractLockNode::set_eliminated() {
 //=============================================================================
 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 
-  // perform any generic optimizations first
+  // perform any generic optimizations first (returns 'this' or NULL)
   Node *result = SafePointNode::Ideal(phase, can_reshape);
 
   // Now see if we can optimize away this lock.  We don't actually
@@ -1372,7 +1372,20 @@ Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
   // prevents macro expansion from expanding the lock.  Since we don't
   // modify the graph, the value returned from this function is the
   // one computed above.
-  if (EliminateLocks && !is_eliminated()) {
+  if (result == NULL && can_reshape && EliminateLocks && !is_eliminated()) {
+    //
+    // If we are locking an unescaped object, the lock/unlock is unnecessary
+    //
+    ConnectionGraph *cgr = Compile::current()->congraph();
+    PointsToNode::EscapeState es = PointsToNode::GlobalEscape;
+    if (cgr != NULL)
+      es = cgr->escape_state(obj_node(), phase);
+    if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
+      // Mark it eliminated to update any counters
+      this->set_eliminated();
+      return result;
+    }
+
     //
     // Try lock coarsening
     //
@@ -1412,8 +1425,10 @@ Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
           int unlocks = 0;
           for (int i = 0; i < lock_ops.length(); i++) {
             AbstractLockNode* lock = lock_ops.at(i);
-            if (lock->Opcode() == Op_Lock) locks++;
-            else                               unlocks++;
+            if (lock->Opcode() == Op_Lock)
+              locks++;
+            else
+              unlocks++;
             if (Verbose) {
               lock->dump(1);
             }
@@ -1450,7 +1465,7 @@ uint UnlockNode::size_of() const { return sizeof(*this); }
 //=============================================================================
 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 
-  // perform any generic optimizations first
+  // perform any generic optimizations first (returns 'this' or NULL)
   Node * result = SafePointNode::Ideal(phase, can_reshape);
 
   // Now see if we can optimize away this unlock.  We don't actually
@@ -1458,66 +1473,18 @@ Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
   // prevents macro expansion from expanding the unlock.  Since we don't
   // modify the graph, the value returned from this function is the
   // one computed above.
-  if (EliminateLocks && !is_eliminated()) {
+  // Escape state is defined after Parse phase.
+  if (result == NULL && can_reshape && EliminateLocks && !is_eliminated()) {
     //
-    // If we are unlocking an unescaped object, the lock/unlock is unnecessary
-    // We can eliminate them if there are no safepoints in the locked region.
+    // If we are unlocking an unescaped object, the lock/unlock is unnecessary.
     //
     ConnectionGraph *cgr = Compile::current()->congraph();
-    if (cgr != NULL && cgr->escape_state(obj_node(), phase) == PointsToNode::NoEscape) {
-      GrowableArray<AbstractLockNode*>   lock_ops;
-      LockNode *lock = find_matching_lock(this);
-      if (lock != NULL) {
-        lock_ops.append(this);
-        lock_ops.append(lock);
-        // find other unlocks which pair with the lock we found and add them
-        // to the list
-        Node * box = box_node();
-
-        for (DUIterator_Fast imax, i = box->fast_outs(imax); i < imax; i++) {
-          Node *use = box->fast_out(i);
-          if (use->is_Unlock() && use != this) {
-            UnlockNode *unlock1 = use->as_Unlock();
-            if (!unlock1->is_eliminated()) {
-              LockNode *lock1 = find_matching_lock(unlock1);
-              if (lock == lock1)
-                lock_ops.append(unlock1);
-              else if (lock1 == NULL) {
-               // we can't find a matching lock, we must assume the worst
-                lock_ops.trunc_to(0);
-                break;
-              }
-            }
-          }
-        }
-        if (lock_ops.length() > 0) {
-
-  #ifndef PRODUCT
-          if (PrintEliminateLocks) {
-            int locks = 0;
-            int unlocks = 0;
-            for (int i = 0; i < lock_ops.length(); i++) {
-              AbstractLockNode* lock = lock_ops.at(i);
-              if (lock->Opcode() == Op_Lock) locks++;
-              else                               unlocks++;
-              if (Verbose) {
-                lock->dump(1);
-              }
-            }
-            tty->print_cr("***Eliminated %d unescaped unlocks and %d unescaped locks", unlocks, locks);
-          }
-  #endif
-
-          // for each of the identified locks, mark them
-          // as eliminatable
-          for (int i = 0; i < lock_ops.length(); i++) {
-            AbstractLockNode* lock = lock_ops.at(i);
-
-            // Mark it eliminated to update any counters
-            lock->set_eliminated();
-          }
-        }
-      }
+    PointsToNode::EscapeState es = PointsToNode::GlobalEscape;
+    if (cgr != NULL)
+      es = cgr->escape_state(obj_node(), phase);
+    if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
+      // Mark it eliminated to update any counters
+      this->set_eliminated();
     }
   }
   return result;

src/share/vm/opto/locknode.cpp

@@ -36,7 +36,8 @@ const RegMask &BoxLockNode::out_RegMask() const {
 
 uint BoxLockNode::size_of() const { return sizeof(*this); }
 
-BoxLockNode::BoxLockNode( int slot ) : Node( Compile::current()->root() ), _slot(slot) {
+BoxLockNode::BoxLockNode( int slot ) : Node( Compile::current()->root() ),
+                                       _slot(slot), _is_eliminated(false) {
   init_class_id(Class_BoxLock);
   init_flags(Flag_rematerialize);
   OptoReg::Name reg = OptoReg::stack2reg(_slot);

src/share/vm/opto/locknode.hpp

@@ -27,6 +27,7 @@ class BoxLockNode : public Node {
 public:
   const int _slot;
   RegMask   _inmask;
+  bool _is_eliminated;    // indicates this lock was safely eliminated
 
   BoxLockNode( int lock );
   virtual int Opcode() const;
@@ -42,6 +43,10 @@ public:
 
   static OptoReg::Name stack_slot(Node* box_node);
 
+  bool is_eliminated()  { return _is_eliminated; }
+  // mark lock as eliminated.
+  void set_eliminated() { _is_eliminated = true; }
+
 #ifndef PRODUCT
   virtual void format( PhaseRegAlloc *, outputStream *st ) const;
   virtual void dump_spec(outputStream *st) const { st->print("  Lock %d",_slot); }

src/share/vm/opto/macro.cpp

@@ -828,43 +828,102 @@ void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
 // Note:  The membar's associated with the lock/unlock are currently not
 //        eliminated.  This should be investigated as a future enhancement.
 //
-void PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
-  Node* mem = alock->in(TypeFunc::Memory);
+bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
+
+  if (!alock->is_eliminated()) {
+    return false;
+  }
+  // Mark the box lock as eliminated if all correspondent locks are eliminated
+  // to construct correct debug info.
+  BoxLockNode* box = alock->box_node()->as_BoxLock();
+  if (!box->is_eliminated()) {
+    bool eliminate = true;
+    for (DUIterator_Fast imax, i = box->fast_outs(imax); i < imax; i++) {
+      Node *lck = box->fast_out(i);
+      if (lck->is_Lock() && !lck->as_AbstractLock()->is_eliminated()) {
+        eliminate = false;
+        break;
+      }
+    }
+    if (eliminate)
+      box->set_eliminated();
+  }
+
+  #ifndef PRODUCT
+  if (PrintEliminateLocks) {
+    if (alock->is_Lock()) {
+      tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
+    } else {
+      tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
+    }
+  }
+  #endif
+
+  Node* mem  = alock->in(TypeFunc::Memory);
+  Node* ctrl = alock->in(TypeFunc::Control);
+
+  extract_call_projections(alock);
+  // There are 2 projections from the lock.  The lock node will
+  // be deleted when its last use is subsumed below.
+  assert(alock->outcnt() == 2 &&
+         _fallthroughproj != NULL &&
+         _memproj_fallthrough != NULL,
+         "Unexpected projections from Lock/Unlock");
+
+  Node* fallthroughproj = _fallthroughproj;
+  Node* memproj_fallthrough = _memproj_fallthrough;
 
   // The memory projection from a lock/unlock is RawMem
   // The input to a Lock is merged memory, so extract its RawMem input
   // (unless the MergeMem has been optimized away.)
   if (alock->is_Lock()) {
-    if (mem->is_MergeMem())
-      mem = mem->as_MergeMem()->in(Compile::AliasIdxRaw);
+    // Seach for MemBarAcquire node and delete it also.
+    MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
+    assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
+    Node* ctrlproj = membar->proj_out(TypeFunc::Control);
+    Node* memproj = membar->proj_out(TypeFunc::Memory);
+    _igvn.hash_delete(ctrlproj);
+    _igvn.subsume_node(ctrlproj, fallthroughproj);
+    _igvn.hash_delete(memproj);
+    _igvn.subsume_node(memproj, memproj_fallthrough);
   }
 
-  extract_call_projections(alock);
-  // There are 2 projections from the lock.  The lock node will
-  // be deleted when its last use is subsumed below.
-  assert(alock->outcnt() == 2 && _fallthroughproj != NULL &&
-          _memproj_fallthrough != NULL, "Unexpected projections from Lock/Unlock");
-  _igvn.hash_delete(_fallthroughproj);
-  _igvn.subsume_node(_fallthroughproj, alock->in(TypeFunc::Control));
-  _igvn.hash_delete(_memproj_fallthrough);
-  _igvn.subsume_node(_memproj_fallthrough, mem);
-  return;
+  // Seach for MemBarRelease node and delete it also.
+  if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
+      ctrl->in(0)->is_MemBar()) {
+    MemBarNode* membar = ctrl->in(0)->as_MemBar();
+    assert(membar->Opcode() == Op_MemBarRelease &&
+           mem->is_Proj() && membar == mem->in(0), "");
+    _igvn.hash_delete(fallthroughproj);
+    _igvn.subsume_node(fallthroughproj, ctrl);
+    _igvn.hash_delete(memproj_fallthrough);
+    _igvn.subsume_node(memproj_fallthrough, mem);
+    fallthroughproj = ctrl;
+    memproj_fallthrough = mem;
+    ctrl = membar->in(TypeFunc::Control);
+    mem  = membar->in(TypeFunc::Memory);
+  }
+
+  _igvn.hash_delete(fallthroughproj);
+  _igvn.subsume_node(fallthroughproj, ctrl);
+  _igvn.hash_delete(memproj_fallthrough);
+  _igvn.subsume_node(memproj_fallthrough, mem);
+  return true;
 }
 
 
 //------------------------------expand_lock_node----------------------
 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
 
+  if (eliminate_locking_node(lock)) {
+    return;
+  }
+
   Node* ctrl = lock->in(TypeFunc::Control);
   Node* mem = lock->in(TypeFunc::Memory);
   Node* obj = lock->obj_node();
   Node* box = lock->box_node();
-  Node *flock = lock->fastlock_node();
-
-  if (lock->is_eliminated()) {
-    eliminate_locking_node(lock);
-    return;
-  }
+  Node* flock = lock->fastlock_node();
 
   // Make the merge point
   Node *region = new (C, 3) RegionNode(3);
@@ -913,17 +972,15 @@ void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
 //------------------------------expand_unlock_node----------------------
 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
 
-  Node *ctrl = unlock->in(TypeFunc::Control);
+  if (eliminate_locking_node(unlock)) {
+    return;
+  }
+
+  Node* ctrl = unlock->in(TypeFunc::Control);
   Node* mem = unlock->in(TypeFunc::Memory);
   Node* obj = unlock->obj_node();
   Node* box = unlock->box_node();
 
-
-  if (unlock->is_eliminated()) {
-    eliminate_locking_node(unlock);
-    return;
-  }
-
   // No need for a null check on unlock
 
   // Make the merge point

src/share/vm/opto/macro.hpp

@@ -78,7 +78,7 @@ private:
                               Node* length,
                               const TypeFunc* slow_call_type,
                               address slow_call_address);
-  void eliminate_locking_node(AbstractLockNode *alock);
+  bool eliminate_locking_node(AbstractLockNode *alock);
   void expand_lock_node(LockNode *lock);
   void expand_unlock_node(UnlockNode *unlock);
 

src/share/vm/opto/output.cpp

@@ -882,7 +882,8 @@ void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
       }
 
       OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node);
-      monarray->append(new MonitorValue(scval, Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg))));
+      Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
+      monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated()));
     }
 
     // We dump the object pool first, since deoptimization reads it in first.