6667620: (Escape Analysis) fix deoptimization for scalar replaced objects

Summary: Deoptimization code for reallocation and relocking scalar replaced objects has to be fixed.
Reviewed-by: rasbold, never

src/share/vm/ci/ciInstanceKlass.cpp

@@ -34,7 +34,9 @@
 // ciInstanceKlass::ciInstanceKlass
 //
 // Loaded instance klass.
-ciInstanceKlass::ciInstanceKlass(KlassHandle h_k) : ciKlass(h_k) {
+ciInstanceKlass::ciInstanceKlass(KlassHandle h_k) :
+  ciKlass(h_k), _non_static_fields(NULL)
+{
   assert(get_Klass()->oop_is_instance(), "wrong type");
   instanceKlass* ik = get_instanceKlass();
 
@@ -335,6 +337,37 @@ ciField* ciInstanceKlass::get_field_by_offset(int field_offset, bool is_static)
   return field;
 }
 
+// ------------------------------------------------------------------
+// ciInstanceKlass::non_static_fields.
+
+class NonStaticFieldFiller: public FieldClosure {
+  GrowableArray<ciField*>* _arr;
+  ciEnv* _curEnv;
+public:
+  NonStaticFieldFiller(ciEnv* curEnv, GrowableArray<ciField*>* arr) :
+    _curEnv(curEnv), _arr(arr)
+  {}
+  void do_field(fieldDescriptor* fd) {
+    ciField* field = new (_curEnv->arena()) ciField(fd);
+    _arr->append(field);
+  }
+};
+
+GrowableArray<ciField*>* ciInstanceKlass::non_static_fields() {
+  if (_non_static_fields == NULL) {
+    VM_ENTRY_MARK;
+    ciEnv* curEnv = ciEnv::current();
+    instanceKlass* ik = get_instanceKlass();
+    int max_n_fields = ik->fields()->length()/instanceKlass::next_offset;
+
+    _non_static_fields =
+      new (curEnv->arena()) GrowableArray<ciField*>(max_n_fields);
+    NonStaticFieldFiller filler(curEnv, _non_static_fields);
+    ik->do_nonstatic_fields(&filler);
+  }
+  return _non_static_fields;
+}
+
 static int sort_field_by_offset(ciField** a, ciField** b) {
   return (*a)->offset_in_bytes() - (*b)->offset_in_bytes();
   // (no worries about 32-bit overflow...)

src/share/vm/ci/ciInstanceKlass.hpp

@@ -46,6 +46,7 @@ private:
   bool                   _has_subklass;
   ciFlags                _flags;
   jint                   _nonstatic_field_size;
+  jint                   _nonstatic_oop_map_size;
 
   // Lazy fields get filled in only upon request.
   ciInstanceKlass*       _super;
@@ -58,6 +59,8 @@ private:
   ciInstanceKlass*       _implementors[implementors_limit];
   jint                   _nof_implementors;
 
+  GrowableArray<ciField*>* _non_static_fields;
+
 protected:
   ciInstanceKlass(KlassHandle h_k);
   ciInstanceKlass(ciSymbol* name, jobject loader, jobject protection_domain);
@@ -129,6 +132,9 @@ public:
   jint                   nonstatic_field_size()  {
     assert(is_loaded(), "must be loaded");
     return _nonstatic_field_size; }
+  jint                   nonstatic_oop_map_size()  {
+    assert(is_loaded(), "must be loaded");
+    return _nonstatic_oop_map_size; }
   ciInstanceKlass*       super();
   jint                   nof_implementors()  {
     assert(is_loaded(), "must be loaded");
@@ -138,6 +144,9 @@ public:
 
   ciInstanceKlass* get_canonical_holder(int offset);
   ciField* get_field_by_offset(int field_offset, bool is_static);
+
+  GrowableArray<ciField*>* non_static_fields();
+
   // total number of nonstatic fields (including inherited):
   int nof_nonstatic_fields() {
     if (_nonstatic_fields == NULL)

src/share/vm/code/debugInfo.cpp

@@ -47,7 +47,8 @@ ScopeValue* DebugInfoReadStream::read_object_value() {
   }
 #endif
   ObjectValue* result = new ObjectValue(id);
-  _obj_pool->append(result);
+  // Cache the object since an object field could reference it.
+  _obj_pool->push(result);
   result->read_object(this);
   return result;
 }
@@ -56,9 +57,9 @@ ScopeValue* DebugInfoReadStream::get_cached_object() {
   int id = read_int();
   assert(_obj_pool != NULL, "object pool does not exist");
   for (int i = _obj_pool->length() - 1; i >= 0; i--) {
-    ObjectValue* sv = (ObjectValue*) _obj_pool->at(i);
-    if (sv->id() == id) {
-      return sv;
+    ObjectValue* ov = (ObjectValue*) _obj_pool->at(i);
+    if (ov->id() == id) {
+      return ov;
     }
   }
   ShouldNotReachHere();

src/share/vm/code/scopeDesc.cpp

@@ -91,7 +91,9 @@ GrowableArray<ScopeValue*>* ScopeDesc::decode_object_values(int decode_offset) {
   DebugInfoReadStream* stream = new DebugInfoReadStream(_code, decode_offset, result);
   int length = stream->read_int();
   for (int index = 0; index < length; index++) {
-    result->push(ScopeValue::read_from(stream));
+    // Objects values are pushed to 'result' array during read so that
+    // object's fields could reference it (OBJECT_ID_CODE).
+    (void)ScopeValue::read_from(stream);
   }
   assert(result->length() == length, "inconsistent debug information");
   return result;

src/share/vm/oops/instanceKlass.cpp

@@ -791,17 +791,39 @@ void instanceKlass::do_local_static_fields_impl(instanceKlassHandle this_oop, vo
 }
 
 
+static int compare_fields_by_offset(int* a, int* b) {
+  return a[0] - b[0];
+}
+
 void instanceKlass::do_nonstatic_fields(FieldClosure* cl) {
-  fieldDescriptor fd;
   instanceKlass* super = superklass();
   if (super != NULL) {
     super->do_nonstatic_fields(cl);
   }
+  fieldDescriptor fd;
   int length = fields()->length();
+  // In DebugInfo nonstatic fields are sorted by offset.
+  int* fields_sorted = NEW_C_HEAP_ARRAY(int, 2*(length+1));
+  int j = 0;
   for (int i = 0; i < length; i += next_offset) {
     fd.initialize(as_klassOop(), i);
-    if (!(fd.is_static())) cl->do_field(&fd);
+    if (!fd.is_static()) {
+      fields_sorted[j + 0] = fd.offset();
+      fields_sorted[j + 1] = i;
+      j += 2;
+    }
+  }
+  if (j > 0) {
+    length = j;
+    // _sort_Fn is defined in growableArray.hpp.
+    qsort(fields_sorted, length/2, 2*sizeof(int), (_sort_Fn)compare_fields_by_offset);
+    for (int i = 0; i < length; i += 2) {
+      fd.initialize(as_klassOop(), fields_sorted[i + 1]);
+      assert(!fd.is_static() && fd.offset() == fields_sorted[i], "only nonstatic fields");
+      cl->do_field(&fd);
+    }
   }
+  FREE_C_HEAP_ARRAY(int, fields_sorted);
 }
 
 

src/share/vm/runtime/deoptimization.cpp

@@ -141,41 +141,45 @@ Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread
 #ifdef COMPILER2
   // Reallocate the non-escaping objects and restore their fields. Then
   // relock objects if synchronization on them was eliminated.
-  if (DoEscapeAnalysis && EliminateAllocations) {
-    GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
-    bool reallocated = false;
-    if (objects != NULL) {
-      JRT_BLOCK
-        reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
-      JRT_END
-    }
-    if (reallocated) {
-      reassign_fields(&deoptee, &map, objects);
+  if (DoEscapeAnalysis) {
+    if (EliminateAllocations) {
+      GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
+      bool reallocated = false;
+      if (objects != NULL) {
+        JRT_BLOCK
+          reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
+        JRT_END
+      }
+      if (reallocated) {
+        reassign_fields(&deoptee, &map, objects);
 #ifndef PRODUCT
-      if (TraceDeoptimization) {
-        ttyLocker ttyl;
-        tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
-        print_objects(objects);
+        if (TraceDeoptimization) {
+          ttyLocker ttyl;
+          tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
+          print_objects(objects);
       }
 #endif
+      }
     }
-    for (int i = 0; i < chunk->length(); i++) {
-      GrowableArray<MonitorValue*>* monitors = chunk->at(i)->scope()->monitors();
-      if (monitors != NULL) {
-        relock_objects(&deoptee, &map, monitors);
+    if (EliminateLocks) {
+      for (int i = 0; i < chunk->length(); i++) {
+        GrowableArray<MonitorValue*>* monitors = chunk->at(i)->scope()->monitors();
+        if (monitors != NULL) {
+          relock_objects(&deoptee, &map, monitors);
 #ifndef PRODUCT
-        if (TraceDeoptimization) {
-          ttyLocker ttyl;
-          tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
-          for (int j = 0; i < monitors->length(); i++) {
-            MonitorValue* mv = monitors->at(i);
-            if (mv->eliminated()) {
-              StackValue* owner = StackValue::create_stack_value(&deoptee, &map, mv->owner());
-              tty->print_cr("     object <" INTPTR_FORMAT "> locked", owner->get_obj()());
+          if (TraceDeoptimization) {
+            ttyLocker ttyl;
+            tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
+            for (int j = 0; j < monitors->length(); j++) {
+              MonitorValue* mv = monitors->at(j);
+              if (mv->eliminated()) {
+                StackValue* owner = StackValue::create_stack_value(&deoptee, &map, mv->owner());
+                tty->print_cr("     object <" INTPTR_FORMAT "> locked", owner->get_obj()());
+              }
             }
           }
-        }
 #endif
+        }
       }
     }
   }
@@ -656,6 +660,7 @@ public:
 
 
   void do_field(fieldDescriptor* fd) {
+    intptr_t val;
     StackValue* value =
       StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
     int offset = fd->offset();
@@ -669,24 +674,36 @@ public:
       assert(value->type() == T_INT, "Agreement.");
       StackValue* low =
         StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
+#ifdef _LP64
+      jlong res = (jlong)low->get_int();
+#else
+#ifdef SPARC
+      // For SPARC we have to swap high and low words.
+      jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
+#else
       jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
+#endif //SPARC
+#endif
       _obj->long_field_put(offset, res);
       break;
     }
-
+    // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
     case T_INT: case T_FLOAT: // 4 bytes.
       assert(value->type() == T_INT, "Agreement.");
-      _obj->int_field_put(offset, (jint)value->get_int());
+      val = value->get_int();
+      _obj->int_field_put(offset, (jint)*((jint*)&val));
       break;
 
     case T_SHORT: case T_CHAR: // 2 bytes
       assert(value->type() == T_INT, "Agreement.");
-      _obj->short_field_put(offset, (jshort)value->get_int());
+      val = value->get_int();
+      _obj->short_field_put(offset, (jshort)*((jint*)&val));
       break;
 
-    case T_BOOLEAN: // 1 byte
+    case T_BOOLEAN: case T_BYTE: // 1 byte
       assert(value->type() == T_INT, "Agreement.");
-      _obj->bool_field_put(offset, (jboolean)value->get_int());
+      val = value->get_int();
+      _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
       break;
 
     default:
@@ -698,25 +715,49 @@ public:
 
 // restore elements of an eliminated type array
 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
-  StackValue* low;
-  jlong lval;
   int index = 0;
+  intptr_t val;
 
   for (int i = 0; i < sv->field_size(); i++) {
     StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
     switch(type) {
-      case T_BOOLEAN: obj->bool_at_put (index, (jboolean) value->get_int()); break;
-      case T_BYTE:    obj->byte_at_put (index, (jbyte)    value->get_int()); break;
-      case T_CHAR:    obj->char_at_put (index, (jchar)    value->get_int()); break;
-      case T_SHORT:   obj->short_at_put(index, (jshort)   value->get_int()); break;
-      case T_INT:     obj->int_at_put  (index, (jint)     value->get_int()); break;
-      case T_FLOAT:   obj->float_at_put(index, (jfloat)   value->get_int()); break;
-      case T_LONG:
-      case T_DOUBLE:
-        low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
-        lval = jlong_from((jint)value->get_int(), (jint)low->get_int());
-        sv->value()->long_field_put(index, lval);
-        break;
+    case T_LONG: case T_DOUBLE: {
+      assert(value->type() == T_INT, "Agreement.");
+      StackValue* low =
+        StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
+#ifdef _LP64
+      jlong res = (jlong)low->get_int();
+#else
+#ifdef SPARC
+      // For SPARC we have to swap high and low words.
+      jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
+#else
+      jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
+#endif //SPARC
+#endif
+      obj->long_at_put(index, res);
+      break;
+    }
+
+    // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
+    case T_INT: case T_FLOAT: // 4 bytes.
+      assert(value->type() == T_INT, "Agreement.");
+      val = value->get_int();
+      obj->int_at_put(index, (jint)*((jint*)&val));
+      break;
+
+    case T_SHORT: case T_CHAR: // 2 bytes
+      assert(value->type() == T_INT, "Agreement.");
+      val = value->get_int();
+      obj->short_at_put(index, (jshort)*((jint*)&val));
+      break;
+
+    case T_BOOLEAN: case T_BYTE: // 1 byte
+      assert(value->type() == T_INT, "Agreement.");
+      val = value->get_int();
+      obj->bool_at_put(index, (jboolean)*((jint*)&val));
+      break;
+
       default:
         ShouldNotReachHere();
     }