6675699: need comprehensive fix for unconstrained ConvI2L with narrowed type

Summary: Emit CastII to make narrow ConvI2L dependent on the corresponding range check.
Reviewed-by: kvn, roland

src/share/vm/opto/compile.cpp

@@ -412,6 +412,13 @@ void Compile::remove_useless_nodes(Unique_Node_List &useful) {
       remove_macro_node(n);
     }
   }
+  // Remove useless CastII nodes with range check dependency
+  for (int i = range_check_cast_count() - 1; i >= 0; i--) {
+    Node* cast = range_check_cast_node(i);
+    if (!useful.member(cast)) {
+      remove_range_check_cast(cast);
+    }
+  }
   // Remove useless expensive node
   for (int i = C->expensive_count()-1; i >= 0; i--) {
     Node* n = C->expensive_node(i);
@@ -1148,6 +1155,7 @@ void Compile::Init(int aliaslevel) {
   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
+  _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
   register_library_intrinsics();
 }
 
@@ -1876,6 +1884,22 @@ void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
   assert(predicate_count()==0, "should be clean!");
 }
 
+void Compile::add_range_check_cast(Node* n) {
+  assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
+  assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
+  _range_check_casts->append(n);
+}
+
+// Remove all range check dependent CastIINodes.
+void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
+  for (int i = range_check_cast_count(); i > 0; i--) {
+    Node* cast = range_check_cast_node(i-1);
+    assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
+    igvn.replace_node(cast, cast->in(1));
+  }
+  assert(range_check_cast_count() == 0, "should be empty");
+}
+
 // StringOpts and late inlining of string methods
 void Compile::inline_string_calls(bool parse_time) {
   {
@@ -2218,6 +2242,12 @@ void Compile::Optimize() {
     PhaseIdealLoop::verify(igvn);
   }
 
+  if (range_check_cast_count() > 0) {
+    // No more loop optimizations. Remove all range check dependent CastIINodes.
+    C->remove_range_check_casts(igvn);
+    igvn.optimize();
+  }
+
   {
     NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
     PhaseMacroExpand  mex(igvn);
@@ -2987,6 +3017,16 @@ void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
 
 #endif
 
+#ifdef ASSERT
+  case Op_CastII:
+    // Verify that all range check dependent CastII nodes were removed.
+    if (n->isa_CastII()->has_range_check()) {
+      n->dump(3);
+      assert(false, "Range check dependent CastII node was not removed");
+    }
+    break;
+#endif
+
   case Op_ModI:
     if (UseDivMod) {
       // Check if a%b and a/b both exist
@@ -4024,6 +4064,24 @@ void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
   }
 }
 
+// Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
+Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
+  if (ctrl != NULL) {
+    // Express control dependency by a CastII node with a narrow type.
+    value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
+    // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
+    // node from floating above the range check during loop optimizations. Otherwise, the
+    // ConvI2L node may be eliminated independently of the range check, causing the data path
+    // to become TOP while the control path is still there (although it's unreachable).
+    value->set_req(0, ctrl);
+    // Save CastII node to remove it after loop optimizations.
+    phase->C->add_range_check_cast(value);
+    value = phase->transform(value);
+  }
+  const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
+  return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
+}
+
 // Auxiliary method to support randomized stressing/fuzzing.
 //
 // This method can be called the arbitrary number of times, with current count

src/share/vm/opto/compile.hpp

@@ -75,6 +75,7 @@ class SafePointNode;
 class JVMState;
 class Type;
 class TypeData;
+class TypeInt;
 class TypePtr;
 class TypeOopPtr;
 class TypeFunc;
@@ -334,6 +335,7 @@ class Compile : public Phase {
   GrowableArray<Node*>* _macro_nodes;           // List of nodes which need to be expanded before matching.
   GrowableArray<Node*>* _predicate_opaqs;       // List of Opaque1 nodes for the loop predicates.
   GrowableArray<Node*>* _expensive_nodes;       // List of nodes that are expensive to compute and that we'd better not let the GVN freely common
+  GrowableArray<Node*>* _range_check_casts;     // List of CastII nodes with a range check dependency
   ConnectionGraph*      _congraph;
 #ifndef PRODUCT
   IdealGraphPrinter*    _printer;
@@ -669,7 +671,7 @@ class Compile : public Phase {
   void set_congraph(ConnectionGraph* congraph)  { _congraph = congraph;}
   void add_macro_node(Node * n) {
     //assert(n->is_macro(), "must be a macro node");
-    assert(!_macro_nodes->contains(n), " duplicate entry in expand list");
+    assert(!_macro_nodes->contains(n), "duplicate entry in expand list");
     _macro_nodes->append(n);
   }
   void remove_macro_node(Node * n) {
@@ -689,10 +691,23 @@ class Compile : public Phase {
     }
   }
   void add_predicate_opaq(Node * n) {
-    assert(!_predicate_opaqs->contains(n), " duplicate entry in predicate opaque1");
+    assert(!_predicate_opaqs->contains(n), "duplicate entry in predicate opaque1");
     assert(_macro_nodes->contains(n), "should have already been in macro list");
     _predicate_opaqs->append(n);
   }
+
+  // Range check dependent CastII nodes that can be removed after loop optimizations
+  void add_range_check_cast(Node* n);
+  void remove_range_check_cast(Node* n) {
+    if (_range_check_casts->contains(n)) {
+      _range_check_casts->remove(n);
+    }
+  }
+  Node* range_check_cast_node(int idx) const { return _range_check_casts->at(idx);  }
+  int   range_check_cast_count()       const { return _range_check_casts->length(); }
+  // Remove all range check dependent CastIINodes.
+  void  remove_range_check_casts(PhaseIterGVN &igvn);
+
   // remove the opaque nodes that protect the predicates so that the unused checks and
   // uncommon traps will be eliminated from the graph.
   void cleanup_loop_predicates(PhaseIterGVN &igvn);
@@ -1201,6 +1216,9 @@ class Compile : public Phase {
   // Definitions of pd methods
   static void pd_compiler2_init();
 
+  // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
+  static Node* constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl);
+
   // Auxiliary method for randomized fuzzing/stressing
   static bool randomized_select(int count);
 };

src/share/vm/opto/connode.cpp

@@ -535,6 +535,9 @@ void CastIINode::dump_spec(outputStream *st) const {
   if (_carry_dependency) {
     st->print(" carry dependency");
   }
+  if (_range_check_dependency) {
+    st->print(" range check dependency");
+  }
 }
 #endif
 
@@ -994,7 +997,8 @@ Node *ConvI2LNode::Ideal(PhaseGVN *phase, bool can_reshape) {
   }
 
 #ifdef _LP64
-  // Convert ConvI2L(AddI(x, y)) to AddL(ConvI2L(x), ConvI2L(y)) ,
+  // Convert ConvI2L(AddI(x, y)) to AddL(ConvI2L(x), ConvI2L(y)) or
+  // ConvI2L(CastII(AddI(x, y))) to AddL(ConvI2L(CastII(x)), ConvI2L(CastII(y))),
   // but only if x and y have subranges that cannot cause 32-bit overflow,
   // under the assumption that x+y is in my own subrange this->type().
 
@@ -1018,6 +1022,13 @@ Node *ConvI2LNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 
   Node* z = in(1);
   int op = z->Opcode();
+  Node* ctrl = NULL;
+  if (op == Op_CastII && z->as_CastII()->has_range_check()) {
+    // Skip CastII node but save control dependency
+    ctrl = z->in(0);
+    z = z->in(1);
+    op = z->Opcode();
+  }
   if (op == Op_AddI || op == Op_SubI) {
     Node* x = z->in(1);
     Node* y = z->in(2);
@@ -1075,9 +1086,10 @@ Node *ConvI2LNode::Ideal(PhaseGVN *phase, bool can_reshape) {
       rylo = -ryhi;
       ryhi = -rylo0;
     }
-
-    Node* cx = phase->transform( new (phase->C) ConvI2LNode(x, TypeLong::make(rxlo, rxhi, widen)) );
-    Node* cy = phase->transform( new (phase->C) ConvI2LNode(y, TypeLong::make(rylo, ryhi, widen)) );
+    assert(rxlo == (int)rxlo && rxhi == (int)rxhi, "x should not overflow");
+    assert(rylo == (int)rylo && ryhi == (int)ryhi, "y should not overflow");
+    Node* cx = phase->C->constrained_convI2L(phase, x, TypeInt::make(rxlo, rxhi, widen), ctrl);
+    Node* cy = phase->C->constrained_convI2L(phase, y, TypeInt::make(rylo, ryhi, widen), ctrl);
     switch (op) {
     case Op_AddI:  return new (phase->C) AddLNode(cx, cy);
     case Op_SubI:  return new (phase->C) SubLNode(cx, cy);

src/share/vm/opto/connode.hpp

@@ -244,19 +244,31 @@ class CastIINode: public ConstraintCastNode {
   private:
   // Can this node be removed post CCP or does it carry a required dependency?
   const bool _carry_dependency;
+  // Is this node dependent on a range check?
+  const bool _range_check_dependency;
 
   protected:
   virtual uint cmp( const Node &n ) const;
   virtual uint size_of() const;
 
 public:
-  CastIINode(Node *n, const Type *t, bool carry_dependency = false)
-    : ConstraintCastNode(n,t), _carry_dependency(carry_dependency) {}
+  CastIINode(Node *n, const Type *t, bool carry_dependency = false, bool range_check_dependency = false)
+    : ConstraintCastNode(n,t), _carry_dependency(carry_dependency), _range_check_dependency(range_check_dependency) {
+    init_class_id(Class_CastII);
+  }
   virtual int Opcode() const;
   virtual uint ideal_reg() const { return Op_RegI; }
   virtual Node *Identity( PhaseTransform *phase );
   virtual const Type *Value( PhaseTransform *phase ) const;
   virtual Node *Ideal_DU_postCCP( PhaseCCP * );
+  const bool has_range_check() {
+ #ifdef _LP64
+     return _range_check_dependency;
+ #else
+     assert(!_range_check_dependency, "Should not have range check dependency");
+     return false;
+ #endif
+   }
 #ifndef PRODUCT
   virtual void dump_spec(outputStream *st) const;
 #endif

src/share/vm/opto/graphKit.cpp

@@ -1645,7 +1645,7 @@ Node* GraphKit::store_oop_to_unknown(Node* ctl,
 
 //-------------------------array_element_address-------------------------
 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
-                                      const TypeInt* sizetype) {
+                                      const TypeInt* sizetype, Node* ctrl) {
   uint shift  = exact_log2(type2aelembytes(elembt));
   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
 
@@ -1670,9 +1670,9 @@ Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
   // number.  (The prior range check has ensured this.)
   // This assertion is used by ConvI2LNode::Ideal.
   int index_max = max_jint - 1;  // array size is max_jint, index is one less
-  if (sizetype != NULL)  index_max = sizetype->_hi - 1;
-  const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
-  idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
+  if (sizetype != NULL) index_max = sizetype->_hi - 1;
+  const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
+  idx = C->constrained_convI2L(&_gvn, idx, iidxtype, ctrl);
 #endif
   Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
   return basic_plus_adr(ary, base, scale);
@@ -3491,10 +3491,6 @@ Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
 
   Node* initial_slow_cmp  = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
   Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
-  if (initial_slow_test->is_Bool()) {
-    // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
-    initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
-  }
 
   // --- Size Computation ---
   // array_size = round_to_heap(array_header + (length << elem_shift));
@@ -3540,13 +3536,35 @@ Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
   Node* lengthx = ConvI2X(length);
   Node* headerx = ConvI2X(header_size);
 #ifdef _LP64
-  { const TypeLong* tllen = _gvn.find_long_type(lengthx);
-    if (tllen != NULL && tllen->_lo < 0) {
+  { const TypeInt* tilen = _gvn.find_int_type(length);
+    if (tilen != NULL && tilen->_lo < 0) {
       // Add a manual constraint to a positive range.  Cf. array_element_address.
-      jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
-      if (size_max > tllen->_hi)  size_max = tllen->_hi;
-      const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
-      lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
+      jlong size_max = fast_size_limit;
+      if (size_max > tilen->_hi)  size_max = tilen->_hi;
+      const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
+
+      // Only do a narrow I2L conversion if the range check passed.
+      IfNode* iff = new (C) IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
+      _gvn.transform(iff);
+      RegionNode* region = new (C) RegionNode(3);
+      _gvn.set_type(region, Type::CONTROL);
+      lengthx = new (C) PhiNode(region, TypeLong::LONG);
+      _gvn.set_type(lengthx, TypeLong::LONG);
+
+      // Range check passed. Use ConvI2L node with narrow type.
+      Node* passed = IfFalse(iff);
+      region->init_req(1, passed);
+      // Make I2L conversion control dependent to prevent it from
+      // floating above the range check during loop optimizations.
+      lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
+
+      // Range check failed. Use ConvI2L with wide type because length may be invalid.
+      region->init_req(2, IfTrue(iff));
+      lengthx->init_req(2, ConvI2X(length));
+
+      set_control(region);
+      record_for_igvn(region);
+      record_for_igvn(lengthx);
     }
   }
 #endif
@@ -3577,6 +3595,11 @@ Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
   Node *mem = reset_memory();
   set_all_memory(mem); // Create new memory state
 
+  if (initial_slow_test->is_Bool()) {
+    // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
+    initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
+  }
+
   // Create the AllocateArrayNode and its result projections
   AllocateArrayNode* alloc
     = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),

src/share/vm/opto/graphKit.hpp

@@ -626,7 +626,9 @@ class GraphKit : public Phase {
   // Return addressing for an array element.
   Node* array_element_address(Node* ary, Node* idx, BasicType elembt,
                               // Optional constraint on the array size:
-                              const TypeInt* sizetype = NULL);
+                              const TypeInt* sizetype = NULL,
+                              // Optional control dependency (for example, on range check)
+                              Node* ctrl = NULL);
 
   // Return a load of array element at idx.
   Node* load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype);

src/share/vm/opto/loopTransform.cpp

@@ -2438,7 +2438,7 @@ bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new )
 
 //=============================================================================
 // Process all the loops in the loop tree and replace any fill
-// patterns with an intrisc version.
+// patterns with an intrinsic version.
 bool PhaseIdealLoop::do_intrinsify_fill() {
   bool changed = false;
   for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
@@ -2536,8 +2536,9 @@ bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& st
   }
 
   // Make sure the address expression can be handled.  It should be
-  // head->phi * elsize + con.  head->phi might have a ConvI2L.
+  // head->phi * elsize + con.  head->phi might have a ConvI2L(CastII()).
   Node* elements[4];
+  Node* cast = NULL;
   Node* conv = NULL;
   bool found_index = false;
   int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
@@ -2552,6 +2553,12 @@ bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& st
         conv = value;
         value = value->in(1);
       }
+      if (value->Opcode() == Op_CastII &&
+          value->as_CastII()->has_range_check()) {
+        // Skip range check dependent CastII nodes
+        cast = value;
+        value = value->in(1);
+      }
 #endif
       if (value != head->phi()) {
         msg = "unhandled shift in address";
@@ -2564,9 +2571,16 @@ bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& st
         }
       }
     } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
-      if (n->in(1) == head->phi()) {
+      conv = n;
+      n = n->in(1);
+      if (n->Opcode() == Op_CastII &&
+          n->as_CastII()->has_range_check()) {
+        // Skip range check dependent CastII nodes
+        cast = n;
+        n = n->in(1);
+      }
+      if (n == head->phi()) {
         found_index = true;
-        conv = n;
       } else {
         msg = "unhandled input to ConvI2L";
       }
@@ -2625,6 +2639,7 @@ bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& st
   // Address elements are ok
   if (con)   ok.set(con->_idx);
   if (shift) ok.set(shift->_idx);
+  if (cast)  ok.set(cast->_idx);
   if (conv)  ok.set(conv->_idx);
 
   for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {

src/share/vm/opto/loopopts.cpp

@@ -772,6 +772,9 @@ static bool merge_point_safe(Node* region) {
 #ifdef _LP64
         if (m->Opcode() == Op_ConvI2L)
           return false;
+        if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
+          return false;
+        }
 #endif
       }
     }

src/share/vm/opto/node.cpp

@@ -521,6 +521,11 @@ Node *Node::clone() const {
     C->add_macro_node(n);
   if (is_expensive())
     C->add_expensive_node(n);
+  // If the cloned node is a range check dependent CastII, add it to the list.
+  CastIINode* cast = n->isa_CastII();
+  if (cast != NULL && cast->has_range_check()) {
+    C->add_range_check_cast(cast);
+  }
 
   n->set_idx(C->next_unique()); // Get new unique index as well
   debug_only( n->verify_construction() );
@@ -649,6 +654,11 @@ void Node::destruct() {
   if (is_expensive()) {
     compile->remove_expensive_node(this);
   }
+  CastIINode* cast = isa_CastII();
+  if (cast != NULL && cast->has_range_check()) {
+    compile->remove_range_check_cast(cast);
+  }
+
   if (is_SafePoint()) {
     as_SafePoint()->delete_replaced_nodes();
   }
@@ -1344,6 +1354,10 @@ static void kill_dead_code( Node *dead, PhaseIterGVN *igvn ) {
       if (dead->is_expensive()) {
         igvn->C->remove_expensive_node(dead);
       }
+      CastIINode* cast = dead->isa_CastII();
+      if (cast != NULL && cast->has_range_check()) {
+        igvn->C->remove_range_check_cast(cast);
+      }
       igvn->C->record_dead_node(dead->_idx);
       // Kill all inputs to the dead guy
       for (uint i=0; i < dead->req(); i++) {

src/share/vm/opto/node.hpp

@@ -54,6 +54,7 @@ class CallStaticJavaNode;
 class CatchNode;
 class CatchProjNode;
 class CheckCastPPNode;
+class CastIINode;
 class ClearArrayNode;
 class CmpNode;
 class CodeBuffer;
@@ -603,6 +604,7 @@ public:
     DEFINE_CLASS_ID(Type,  Node, 2)
       DEFINE_CLASS_ID(Phi,   Type, 0)
       DEFINE_CLASS_ID(ConstraintCast, Type, 1)
+        DEFINE_CLASS_ID(CastII, ConstraintCast, 0)
       DEFINE_CLASS_ID(CheckCastPP, Type, 2)
       DEFINE_CLASS_ID(CMove, Type, 3)
       DEFINE_CLASS_ID(SafePointScalarObject, Type, 4)
@@ -727,6 +729,7 @@ public:
   DEFINE_CLASS_QUERY(Catch)
   DEFINE_CLASS_QUERY(CatchProj)
   DEFINE_CLASS_QUERY(CheckCastPP)
+  DEFINE_CLASS_QUERY(CastII)
   DEFINE_CLASS_QUERY(ConstraintCast)
   DEFINE_CLASS_QUERY(ClearArray)
   DEFINE_CLASS_QUERY(CMove)

src/share/vm/opto/parse2.cpp

@@ -154,7 +154,9 @@ Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) {
   // Check for always knowing you are throwing a range-check exception
   if (stopped())  return top();
 
-  Node* ptr = array_element_address(ary, idx, type, sizetype);
+  // Make array address computation control dependent to prevent it
+  // from floating above the range check during loop optimizations.
+  Node* ptr = array_element_address(ary, idx, type, sizetype, control());
 
   if (result2 != NULL)  *result2 = elemtype;
 
@@ -457,9 +459,12 @@ bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi)
 #ifdef _LP64
   // Clean the 32-bit int into a real 64-bit offset.
   // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
-  const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin);
-  key_val       = _gvn.transform( new (C) ConvI2LNode(key_val, lkeytype) );
+  const TypeInt* ikeytype = TypeInt::make(0, num_cases-1, Type::WidenMin);
+  // Make I2L conversion control dependent to prevent it from
+  // floating above the range check during loop optimizations.
+  key_val = C->constrained_convI2L(&_gvn, key_val, ikeytype, control());
 #endif
+
   // Shift the value by wordsize so we have an index into the table, rather
   // than a switch value
   Node *shiftWord = _gvn.MakeConX(wordSize);

src/share/vm/opto/phaseX.cpp

@@ -1339,6 +1339,10 @@ void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
       if (dead->is_expensive()) {
         C->remove_expensive_node(dead);
       }
+      CastIINode* cast = dead->isa_CastII();
+      if (cast != NULL && cast->has_range_check()) {
+        C->remove_range_check_cast(cast);
+      }
     }
   } // while (_stack.is_nonempty())
 }

src/share/vm/opto/superword.cpp

@@ -2388,6 +2388,11 @@ bool SWPointer::scaled_iv(Node* n) {
       return true;
     }
   } else if (opc == Op_ConvI2L) {
+    if (n->in(1)->Opcode() == Op_CastII &&
+        n->in(1)->as_CastII()->has_range_check()) {
+      // Skip range check dependent CastII nodes
+      n = n->in(1);
+    }
     if (scaled_iv_plus_offset(n->in(1))) {
       return true;
     }

test/compiler/loopopts/TestLoopPeeling.java

+/*
+ * Copyright (c) 2016, Oracle and/or its affiliates. All rights reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ */
+
+/*
+ * @test
+ * @bug 8078262
+ * @summary Tests correct dominator information after loop peeling.
+ * @run main/othervm -Xcomp -XX:CompileCommand=compileonly,TestLoopPeeling::test* TestLoopPeeling
+ */
+public class TestLoopPeeling {
+
+    public int[] array = new int[100];
+
+    public static void main(String args[]) {
+        TestLoopPeeling test = new TestLoopPeeling();
+        try {
+            test.testArrayAccess(0, 1);
+            test.testArrayAllocation(0, 1);
+        } catch (Exception e) {
+            // Ignore exceptions
+        }
+    }
+
+    public void testArrayAccess(int index, int inc) {
+        int storeIndex = -1;
+
+        for (; index < 10; index += inc) {
+            // This loop invariant check triggers loop peeling because it can
+            // be moved out of the loop (see 'IdealLoopTree::policy_peeling').
+            if (inc == 42) return;
+
+            // This loop variant usage of LShiftL( ConvI2L( Phi(storeIndex) ) )
+            // prevents the split if optimization that would otherwise clone the
+            // LShiftL and ConvI2L nodes and assign them to their corresponding array
+            // address computation (see 'PhaseIdealLoop::split_if_with_blocks_post').
+            if (storeIndex > 0 && array[storeIndex] == 42) return;
+
+            if (index == 42) {
+                // This store and the corresponding range check are moved out of the
+                // loop and both used after old loop and the peeled iteration exit.
+                // For the peeled iteration, storeIndex is always -1 and the ConvI2L
+                // is replaced by TOP. However, the range check is not folded because
+                // we don't do the split if optimization in PhaseIdealLoop2.
+                // As a result, we have a (dead) control path from the peeled iteration
+                // to the StoreI but the data path is removed.
+                array[storeIndex] = 1;
+                return;
+            }
+
+            storeIndex++;
+        }
+    }
+
+    public byte[] testArrayAllocation(int index, int inc) {
+        int allocationCount = -1;
+        byte[] result;
+
+        for (; index < 10; index += inc) {
+            // This loop invariant check triggers loop peeling because it can
+            // be moved out of the loop (see 'IdealLoopTree::policy_peeling').
+            if (inc == 42) return null;
+
+            if (index == 42) {
+                // This allocation and the corresponding size check are moved out of the
+                // loop and both used after old loop and the peeled iteration exit.
+                // For the peeled iteration, allocationCount is always -1 and the ConvI2L
+                // is replaced by TOP. However, the size check is not folded because
+                // we don't do the split if optimization in PhaseIdealLoop2.
+                // As a result, we have a (dead) control path from the peeled iteration
+                // to the allocation but the data path is removed.
+                result = new byte[allocationCount];
+                return result;
+            }
+
+            allocationCount++;
+        }
+        return null;
+    }
+}
+