Mercurial > jdk8-mips64-public > hotspot / file revision

 /*

  * Copyright (c) 1997, 2012, 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.

  *

  */

 #include "precompiled.hpp"

 #include "classfile/systemDictionary.hpp"

 #include "compiler/compileLog.hpp"

 #include "memory/allocation.inline.hpp"

 #include "oops/objArrayKlass.hpp"

 #include "opto/addnode.hpp"

 #include "opto/cfgnode.hpp"

 #include "opto/compile.hpp"

 #include "opto/connode.hpp"

 #include "opto/loopnode.hpp"

 #include "opto/machnode.hpp"

 #include "opto/matcher.hpp"

 #include "opto/memnode.hpp"

 #include "opto/mulnode.hpp"

 #include "opto/phaseX.hpp"

 #include "opto/regmask.hpp"

 // Portions of code courtesy of Clifford Click

 // Optimization - Graph Style

 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);

 //=============================================================================

 uint MemNode::size_of() const { return sizeof(*this); }

 const TypePtr *MemNode::adr_type() const {

   Node* adr = in(Address);

   const TypePtr* cross_check = NULL;

   DEBUG_ONLY(cross_check = _adr_type);

   return calculate_adr_type(adr->bottom_type(), cross_check);

 }

 #ifndef PRODUCT

 void MemNode::dump_spec(outputStream *st) const {

   if (in(Address) == NULL)  return; // node is dead

 #ifndef ASSERT

   // fake the missing field

   const TypePtr* _adr_type = NULL;

   if (in(Address) != NULL)

     _adr_type = in(Address)->bottom_type()->isa_ptr();

 #endif

   dump_adr_type(this, _adr_type, st);

   Compile* C = Compile::current();

   if( C->alias_type(_adr_type)->is_volatile() )

     st->print(" Volatile!");

 }

 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {

   st->print(" @");

   if (adr_type == NULL) {

     st->print("NULL");

   } else {

     adr_type->dump_on(st);

     Compile* C = Compile::current();

     Compile::AliasType* atp = NULL;

     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);

     if (atp == NULL)

       st->print(", idx=?\?;");

     else if (atp->index() == Compile::AliasIdxBot)

       st->print(", idx=Bot;");

     else if (atp->index() == Compile::AliasIdxTop)

       st->print(", idx=Top;");

     else if (atp->index() == Compile::AliasIdxRaw)

       st->print(", idx=Raw;");

     else {

       ciField* field = atp->field();

       if (field) {

         st->print(", name=");

         field->print_name_on(st);

       }

       st->print(", idx=%d;", atp->index());

     }

   }

 }

 extern void print_alias_types();

 #endif

 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {

   assert((t_oop != NULL), "sanity");

   bool is_instance = t_oop->is_known_instance_field();

   bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&

                              (load != NULL) && load->is_Load() &&

                              (phase->is_IterGVN() != NULL);

   if (!(is_instance || is_boxed_value_load))

     return mchain;  // don't try to optimize non-instance types

   uint instance_id = t_oop->instance_id();

   Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);

   Node *prev = NULL;

   Node *result = mchain;

   while (prev != result) {

     prev = result;

     if (result == start_mem)

       break;  // hit one of our sentinels

     // skip over a call which does not affect this memory slice

     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {

       Node *proj_in = result->in(0);

       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {

         break;  // hit one of our sentinels

       } else if (proj_in->is_Call()) {

         CallNode *call = proj_in->as_Call();

         if (!call->may_modify(t_oop, phase)) { // returns false for instances

           result = call->in(TypeFunc::Memory);

         }

       } else if (proj_in->is_Initialize()) {

         AllocateNode* alloc = proj_in->as_Initialize()->allocation();

         // Stop if this is the initialization for the object instance which

         // which contains this memory slice, otherwise skip over it.

         if ((alloc == NULL) || (alloc->_idx == instance_id)) {

           break;

         }

         if (is_instance) {

           result = proj_in->in(TypeFunc::Memory);

         } else if (is_boxed_value_load) {

           Node* klass = alloc->in(AllocateNode::KlassNode);

           const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();

           if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {

             result = proj_in->in(TypeFunc::Memory); // not related allocation

           }

         }

       } else if (proj_in->is_MemBar()) {

         result = proj_in->in(TypeFunc::Memory);

       } else {

         assert(false, "unexpected projection");

       }

     } else if (result->is_ClearArray()) {

       if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {

         // Can not bypass initialization of the instance

         // we are looking for.

         break;

       }

       // Otherwise skip it (the call updated 'result' value).

     } else if (result->is_MergeMem()) {

       result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);

     }

   }

   return result;

 }

 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {

   const TypeOopPtr* t_oop = t_adr->isa_oopptr();

   if (t_oop == NULL)

     return mchain;  // don't try to optimize non-oop types

   Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);

   bool is_instance = t_oop->is_known_instance_field();

   PhaseIterGVN *igvn = phase->is_IterGVN();

   if (is_instance && igvn != NULL  && result->is_Phi()) {

     PhiNode *mphi = result->as_Phi();

     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");

     const TypePtr *t = mphi->adr_type();

     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||

         t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&

         t->is_oopptr()->cast_to_exactness(true)

          ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())

          ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {

       // clone the Phi with our address type

       result = mphi->split_out_instance(t_adr, igvn);

     } else {

       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");

     }

   }

   return result;

 }

 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {

   uint alias_idx = phase->C->get_alias_index(tp);

   Node *mem = mmem;

 #ifdef ASSERT

   {

     // Check that current type is consistent with the alias index used during graph construction

     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");

     bool consistent =  adr_check == NULL || adr_check->empty() ||

                        phase->C->must_alias(adr_check, alias_idx );

     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]

     if( !consistent && adr_check != NULL && !adr_check->empty() &&

                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&

         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&

         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||

           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||

           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {

       // don't assert if it is dead code.

       consistent = true;

     }

     if( !consistent ) {

       st->print("alias_idx==%d, adr_check==", alias_idx);

       if( adr_check == NULL ) {

         st->print("NULL");

       } else {

         adr_check->dump();

       }

       st->cr();

       print_alias_types();

       assert(consistent, "adr_check must match alias idx");

     }

   }

 #endif

   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally

   // means an array I have not precisely typed yet.  Do not do any

   // alias stuff with it any time soon.

   const TypeOopPtr *toop = tp->isa_oopptr();

   if( tp->base() != Type::AnyPtr &&

       !(toop &&

         toop->klass() != NULL &&

         toop->klass()->is_java_lang_Object() &&

         toop->offset() == Type::OffsetBot) ) {

     // compress paths and change unreachable cycles to TOP

     // If not, we can update the input infinitely along a MergeMem cycle

     // Equivalent code in PhiNode::Ideal

     Node* m  = phase->transform(mmem);

     // If transformed to a MergeMem, get the desired slice

     // Otherwise the returned node represents memory for every slice

     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;

     // Update input if it is progress over what we have now

   }

   return mem;

 }

 //--------------------------Ideal_common---------------------------------------

 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.

 // Unhook non-raw memories from complete (macro-expanded) initializations.

 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {

   // If our control input is a dead region, kill all below the region

   Node *ctl = in(MemNode::Control);

   if (ctl && remove_dead_region(phase, can_reshape))

     return this;

   ctl = in(MemNode::Control);

   // Don't bother trying to transform a dead node

   if (ctl && ctl->is_top())  return NodeSentinel;

   PhaseIterGVN *igvn = phase->is_IterGVN();

   // Wait if control on the worklist.

   if (ctl && can_reshape && igvn != NULL) {

     Node* bol = NULL;

     Node* cmp = NULL;

     if (ctl->in(0)->is_If()) {

       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");

       bol = ctl->in(0)->in(1);

       if (bol->is_Bool())

         cmp = ctl->in(0)->in(1)->in(1);

     }

     if (igvn->_worklist.member(ctl) ||

         (bol != NULL && igvn->_worklist.member(bol)) ||

         (cmp != NULL && igvn->_worklist.member(cmp)) ) {

       // This control path may be dead.

       // Delay this memory node transformation until the control is processed.

       phase->is_IterGVN()->_worklist.push(this);

       return NodeSentinel; // caller will return NULL

     }

   }

   // Ignore if memory is dead, or self-loop

   Node *mem = in(MemNode::Memory);

   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL

   assert(mem != this, "dead loop in MemNode::Ideal");

   if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {

     // This memory slice may be dead.

     // Delay this mem node transformation until the memory is processed.

     phase->is_IterGVN()->_worklist.push(this);

     return NodeSentinel; // caller will return NULL

   }

   Node *address = in(MemNode::Address);

   const Type *t_adr = phase->type(address);

   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return NULL

   if (can_reshape && igvn != NULL &&

       (igvn->_worklist.member(address) ||

        igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {

     // The address's base and type may change when the address is processed.

     // Delay this mem node transformation until the address is processed.

     phase->is_IterGVN()->_worklist.push(this);

     return NodeSentinel; // caller will return NULL

   }

   // Do NOT remove or optimize the next lines: ensure a new alias index

   // is allocated for an oop pointer type before Escape Analysis.

   // Note: C++ will not remove it since the call has side effect.

   if (t_adr->isa_oopptr()) {

     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());

   }

 #ifdef ASSERT

   Node* base = NULL;

   if (address->is_AddP())

     base = address->in(AddPNode::Base);

   if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&

       !t_adr->isa_rawptr()) {

     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.

     Compile* C = phase->C;

     tty->cr();

     tty->print_cr("===== NULL+offs not RAW address =====");

     if (C->is_dead_node(this->_idx))    tty->print_cr("'this' is dead");

     if ((ctl != NULL) && C->is_dead_node(ctl->_idx)) tty->print_cr("'ctl' is dead");

     if (C->is_dead_node(mem->_idx))     tty->print_cr("'mem' is dead");

     if (C->is_dead_node(address->_idx)) tty->print_cr("'address' is dead");

     if (C->is_dead_node(base->_idx))    tty->print_cr("'base' is dead");

     tty->cr();

     base->dump(1);

     tty->cr();

     this->dump(2);

     tty->print("this->adr_type():     "); adr_type()->dump(); tty->cr();

     tty->print("phase->type(address): "); t_adr->dump(); tty->cr();

     tty->print("phase->type(base):    "); phase->type(address)->dump(); tty->cr();

     tty->cr();

   }

   assert(base == NULL || t_adr->isa_rawptr() ||

         !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");

 #endif

   // Avoid independent memory operations

   Node* old_mem = mem;

   // The code which unhooks non-raw memories from complete (macro-expanded)

   // initializations was removed. After macro-expansion all stores catched

   // by Initialize node became raw stores and there is no information

   // which memory slices they modify. So it is unsafe to move any memory

   // operation above these stores. Also in most cases hooked non-raw memories

   // were already unhooked by using information from detect_ptr_independence()

   // and find_previous_store().

   if (mem->is_MergeMem()) {

     MergeMemNode* mmem = mem->as_MergeMem();

     const TypePtr *tp = t_adr->is_ptr();

     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);

   }

   if (mem != old_mem) {

     set_req(MemNode::Memory, mem);

     if (can_reshape && old_mem->outcnt() == 0) {

         igvn->_worklist.push(old_mem);

     }

     if (phase->type( mem ) == Type::TOP) return NodeSentinel;

     return this;

   }

   // let the subclass continue analyzing...

   return NULL;

 }

 // Helper function for proving some simple control dominations.

 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.

 // Already assumes that 'dom' is available at 'sub', and that 'sub'

 // is not a constant (dominated by the method's StartNode).

 // Used by MemNode::find_previous_store to prove that the

 // control input of a memory operation predates (dominates)

 // an allocation it wants to look past.

 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {

   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())

     return false; // Conservative answer for dead code

   // Check 'dom'. Skip Proj and CatchProj nodes.

   dom = dom->find_exact_control(dom);

   if (dom == NULL || dom->is_top())

     return false; // Conservative answer for dead code

   if (dom == sub) {

     // For the case when, for example, 'sub' is Initialize and the original

     // 'dom' is Proj node of the 'sub'.

     return false;

   }

   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)

     return true;

   // 'dom' dominates 'sub' if its control edge and control edges

   // of all its inputs dominate or equal to sub's control edge.

   // Currently 'sub' is either Allocate, Initialize or Start nodes.

   // Or Region for the check in LoadNode::Ideal();

   // 'sub' should have sub->in(0) != NULL.

   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||

          sub->is_Region() || sub->is_Call(), "expecting only these nodes");

   // Get control edge of 'sub'.

   Node* orig_sub = sub;

   sub = sub->find_exact_control(sub->in(0));

   if (sub == NULL || sub->is_top())

     return false; // Conservative answer for dead code

   assert(sub->is_CFG(), "expecting control");

   if (sub == dom)

     return true;

   if (sub->is_Start() || sub->is_Root())

     return false;

   {

     // Check all control edges of 'dom'.

     ResourceMark rm;

     Arena* arena = Thread::current()->resource_area();

     Node_List nlist(arena);

     Unique_Node_List dom_list(arena);

     dom_list.push(dom);

     bool only_dominating_controls = false;

     for (uint next = 0; next < dom_list.size(); next++) {

       Node* n = dom_list.at(next);

       if (n == orig_sub)

         return false; // One of dom's inputs dominated by sub.

       if (!n->is_CFG() && n->pinned()) {

         // Check only own control edge for pinned non-control nodes.

         n = n->find_exact_control(n->in(0));

         if (n == NULL || n->is_top())

           return false; // Conservative answer for dead code

         assert(n->is_CFG(), "expecting control");

         dom_list.push(n);

       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {

         only_dominating_controls = true;

       } else if (n->is_CFG()) {

         if (n->dominates(sub, nlist))

           only_dominating_controls = true;

         else

           return false;

       } else {

         // First, own control edge.

         Node* m = n->find_exact_control(n->in(0));

         if (m != NULL) {

           if (m->is_top())

             return false; // Conservative answer for dead code

           dom_list.push(m);

         }

         // Now, the rest of edges.

         uint cnt = n->req();

         for (uint i = 1; i < cnt; i++) {

           m = n->find_exact_control(n->in(i));

           if (m == NULL || m->is_top())

             continue;

           dom_list.push(m);

         }

       }

     }

     return only_dominating_controls;

   }

 }

 //---------------------detect_ptr_independence---------------------------------

 // Used by MemNode::find_previous_store to prove that two base

 // pointers are never equal.

 // The pointers are accompanied by their associated allocations,

 // if any, which have been previously discovered by the caller.

 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,

                                       Node* p2, AllocateNode* a2,

                                       PhaseTransform* phase) {

   // Attempt to prove that these two pointers cannot be aliased.

   // They may both manifestly be allocations, and they should differ.

   // Or, if they are not both allocations, they can be distinct constants.

   // Otherwise, one is an allocation and the other a pre-existing value.

   if (a1 == NULL && a2 == NULL) {           // neither an allocation

     return (p1 != p2) && p1->is_Con() && p2->is_Con();

   } else if (a1 != NULL && a2 != NULL) {    // both allocations

     return (a1 != a2);

   } else if (a1 != NULL) {                  // one allocation a1

     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)

     return all_controls_dominate(p2, a1);

   } else { //(a2 != NULL)                   // one allocation a2

     return all_controls_dominate(p1, a2);

   }

   return false;

 }

 // The logic for reordering loads and stores uses four steps:

 // (a) Walk carefully past stores and initializations which we

 //     can prove are independent of this load.

 // (b) Observe that the next memory state makes an exact match

 //     with self (load or store), and locate the relevant store.

 // (c) Ensure that, if we were to wire self directly to the store,

 //     the optimizer would fold it up somehow.

 // (d) Do the rewiring, and return, depending on some other part of

 //     the optimizer to fold up the load.

 // This routine handles steps (a) and (b).  Steps (c) and (d) are

 // specific to loads and stores, so they are handled by the callers.

 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)

 //

 Node* MemNode::find_previous_store(PhaseTransform* phase) {

   Node*         ctrl   = in(MemNode::Control);

   Node*         adr    = in(MemNode::Address);

   intptr_t      offset = 0;

   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);

   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);

   if (offset == Type::OffsetBot)

     return NULL;            // cannot unalias unless there are precise offsets

   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();

   intptr_t size_in_bytes = memory_size();

   Node* mem = in(MemNode::Memory);   // start searching here...

   int cnt = 50;             // Cycle limiter

   for (;;) {                // While we can dance past unrelated stores...

     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?

     if (mem->is_Store()) {

       Node* st_adr = mem->in(MemNode::Address);

       intptr_t st_offset = 0;

       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);

       if (st_base == NULL)

         break;              // inscrutable pointer

       if (st_offset != offset && st_offset != Type::OffsetBot) {

         const int MAX_STORE = BytesPerLong;

         if (st_offset >= offset + size_in_bytes ||

             st_offset <= offset - MAX_STORE ||

             st_offset <= offset - mem->as_Store()->memory_size()) {

           // Success:  The offsets are provably independent.

           // (You may ask, why not just test st_offset != offset and be done?

           // The answer is that stores of different sizes can co-exist

           // in the same sequence of RawMem effects.  We sometimes initialize

           // a whole 'tile' of array elements with a single jint or jlong.)

           mem = mem->in(MemNode::Memory);

           continue;           // (a) advance through independent store memory

         }

       }

       if (st_base != base &&

           detect_ptr_independence(base, alloc,

                                   st_base,

                                   AllocateNode::Ideal_allocation(st_base, phase),

                                   phase)) {

         // Success:  The bases are provably independent.

         mem = mem->in(MemNode::Memory);

         continue;           // (a) advance through independent store memory

       }

       // (b) At this point, if the bases or offsets do not agree, we lose,

       // since we have not managed to prove 'this' and 'mem' independent.

       if (st_base == base && st_offset == offset) {

         return mem;         // let caller handle steps (c), (d)

       }

     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {

       InitializeNode* st_init = mem->in(0)->as_Initialize();

       AllocateNode*  st_alloc = st_init->allocation();

       if (st_alloc == NULL)

         break;              // something degenerated

       bool known_identical = false;

       bool known_independent = false;

       if (alloc == st_alloc)

         known_identical = true;

       else if (alloc != NULL)

         known_independent = true;

       else if (all_controls_dominate(this, st_alloc))

         known_independent = true;

       if (known_independent) {

         // The bases are provably independent: Either they are

         // manifestly distinct allocations, or else the control

         // of this load dominates the store's allocation.

         int alias_idx = phase->C->get_alias_index(adr_type());

         if (alias_idx == Compile::AliasIdxRaw) {

           mem = st_alloc->in(TypeFunc::Memory);

         } else {

           mem = st_init->memory(alias_idx);

         }

         continue;           // (a) advance through independent store memory

       }

       // (b) at this point, if we are not looking at a store initializing

       // the same allocation we are loading from, we lose.

       if (known_identical) {

         // From caller, can_see_stored_value will consult find_captured_store.

         return mem;         // let caller handle steps (c), (d)

       }

     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {

       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.

       if (mem->is_Proj() && mem->in(0)->is_Call()) {

         CallNode *call = mem->in(0)->as_Call();

         if (!call->may_modify(addr_t, phase)) {

           mem = call->in(TypeFunc::Memory);

           continue;         // (a) advance through independent call memory

         }

       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {

         mem = mem->in(0)->in(TypeFunc::Memory);

         continue;           // (a) advance through independent MemBar memory

       } else if (mem->is_ClearArray()) {

         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {

           // (the call updated 'mem' value)

           continue;         // (a) advance through independent allocation memory

         } else {

           // Can not bypass initialization of the instance

           // we are looking for.

           return mem;

         }

       } else if (mem->is_MergeMem()) {

         int alias_idx = phase->C->get_alias_index(adr_type());

         mem = mem->as_MergeMem()->memory_at(alias_idx);

         continue;           // (a) advance through independent MergeMem memory

       }

     }

     // Unless there is an explicit 'continue', we must bail out here,

     // because 'mem' is an inscrutable memory state (e.g., a call).

     break;

   }

   return NULL;              // bail out

 }

 //----------------------calculate_adr_type-------------------------------------

 // Helper function.  Notices when the given type of address hits top or bottom.

 // Also, asserts a cross-check of the type against the expected address type.

 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {

   if (t == Type::TOP)  return NULL; // does not touch memory any more?

   #ifdef PRODUCT

   cross_check = NULL;

   #else

   if (!VerifyAliases || is_error_reported() || Node::in_dump())  cross_check = NULL;

   #endif

   const TypePtr* tp = t->isa_ptr();

   if (tp == NULL) {

     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");

     return TypePtr::BOTTOM;           // touches lots of memory

   } else {

     #ifdef ASSERT

     // %%%% [phh] We don't check the alias index if cross_check is

     //            TypeRawPtr::BOTTOM.  Needs to be investigated.

     if (cross_check != NULL &&

         cross_check != TypePtr::BOTTOM &&

         cross_check != TypeRawPtr::BOTTOM) {

       // Recheck the alias index, to see if it has changed (due to a bug).

       Compile* C = Compile::current();

       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),

              "must stay in the original alias category");

       // The type of the address must be contained in the adr_type,

       // disregarding "null"-ness.

       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)

       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();

       assert(cross_check->meet(tp_notnull) == cross_check,

              "real address must not escape from expected memory type");

     }

     #endif

     return tp;

   }

 }

 //------------------------adr_phi_is_loop_invariant----------------------------

 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted

 // loop is loop invariant. Make a quick traversal of Phi and associated

 // CastPP nodes, looking to see if they are a closed group within the loop.

 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {

   // The idea is that the phi-nest must boil down to only CastPP nodes

   // with the same data. This implies that any path into the loop already

   // includes such a CastPP, and so the original cast, whatever its input,

   // must be covered by an equivalent cast, with an earlier control input.

   ResourceMark rm;

   // The loop entry input of the phi should be the unique dominating

   // node for every Phi/CastPP in the loop.

   Unique_Node_List closure;

   closure.push(adr_phi->in(LoopNode::EntryControl));

   // Add the phi node and the cast to the worklist.

   Unique_Node_List worklist;

   worklist.push(adr_phi);

   if( cast != NULL ){

     if( !cast->is_ConstraintCast() ) return false;

     worklist.push(cast);

   }

   // Begin recursive walk of phi nodes.

   while( worklist.size() ){

     // Take a node off the worklist

     Node *n = worklist.pop();

     if( !closure.member(n) ){

       // Add it to the closure.

       closure.push(n);

       // Make a sanity check to ensure we don't waste too much time here.

       if( closure.size() > 20) return false;

       // This node is OK if:

       //  - it is a cast of an identical value

       //  - or it is a phi node (then we add its inputs to the worklist)

       // Otherwise, the node is not OK, and we presume the cast is not invariant

       if( n->is_ConstraintCast() ){

         worklist.push(n->in(1));

       } else if( n->is_Phi() ) {

         for( uint i = 1; i < n->req(); i++ ) {

           worklist.push(n->in(i));

         }

       } else {

         return false;

       }

     }

   }

   // Quit when the worklist is empty, and we've found no offending nodes.

   return true;

 }

 //------------------------------Ideal_DU_postCCP-------------------------------

 // Find any cast-away of null-ness and keep its control.  Null cast-aways are

 // going away in this pass and we need to make this memory op depend on the

 // gating null check.

 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {

   return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));

 }

 // I tried to leave the CastPP's in.  This makes the graph more accurate in

 // some sense; we get to keep around the knowledge that an oop is not-null

 // after some test.  Alas, the CastPP's interfere with GVN (some values are

 // the regular oop, some are the CastPP of the oop, all merge at Phi's which

 // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed

 // some of the more trivial cases in the optimizer.  Removing more useless

 // Phi's started allowing Loads to illegally float above null checks.  I gave

 // up on this approach.  CNC 10/20/2000

 // This static method may be called not from MemNode (EncodePNode calls it).

 // Only the control edge of the node 'n' might be updated.

 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {

   Node *skipped_cast = NULL;

   // Need a null check?  Regular static accesses do not because they are

   // from constant addresses.  Array ops are gated by the range check (which

   // always includes a NULL check).  Just check field ops.

   if( n->in(MemNode::Control) == NULL ) {

     // Scan upwards for the highest location we can place this memory op.

     while( true ) {

       switch( adr->Opcode() ) {

       case Op_AddP:             // No change to NULL-ness, so peek thru AddP's

         adr = adr->in(AddPNode::Base);

         continue;

       case Op_DecodeN:         // No change to NULL-ness, so peek thru

       case Op_DecodeNKlass:

         adr = adr->in(1);

         continue;

       case Op_EncodeP:

       case Op_EncodePKlass:

         // EncodeP node's control edge could be set by this method

         // when EncodeP node depends on CastPP node.

         //

         // Use its control edge for memory op because EncodeP may go away

         // later when it is folded with following or preceding DecodeN node.

         if (adr->in(0) == NULL) {

           // Keep looking for cast nodes.

           adr = adr->in(1);

           continue;

         }

         ccp->hash_delete(n);

         n->set_req(MemNode::Control, adr->in(0));

         ccp->hash_insert(n);

         return n;

       case Op_CastPP:

         // If the CastPP is useless, just peek on through it.

         if( ccp->type(adr) == ccp->type(adr->in(1)) ) {

           // Remember the cast that we've peeked though. If we peek

           // through more than one, then we end up remembering the highest

           // one, that is, if in a loop, the one closest to the top.

           skipped_cast = adr;

           adr = adr->in(1);

           continue;

         }

         // CastPP is going away in this pass!  We need this memory op to be

         // control-dependent on the test that is guarding the CastPP.

         ccp->hash_delete(n);

         n->set_req(MemNode::Control, adr->in(0));

         ccp->hash_insert(n);

         return n;

       case Op_Phi:

         // Attempt to float above a Phi to some dominating point.

         if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {

           // If we've already peeked through a Cast (which could have set the

           // control), we can't float above a Phi, because the skipped Cast

           // may not be loop invariant.

           if (adr_phi_is_loop_invariant(adr, skipped_cast)) {

             adr = adr->in(1);

             continue;

           }

         }

         // Intentional fallthrough!

         // No obvious dominating point.  The mem op is pinned below the Phi

         // by the Phi itself.  If the Phi goes away (no true value is merged)

         // then the mem op can float, but not indefinitely.  It must be pinned

         // behind the controls leading to the Phi.

       case Op_CheckCastPP:

         // These usually stick around to change address type, however a

         // useless one can be elided and we still need to pick up a control edge

         if (adr->in(0) == NULL) {

           // This CheckCastPP node has NO control and is likely useless. But we

           // need check further up the ancestor chain for a control input to keep

           // the node in place. 4959717.

           skipped_cast = adr;

           adr = adr->in(1);

           continue;

         }

         ccp->hash_delete(n);

         n->set_req(MemNode::Control, adr->in(0));

         ccp->hash_insert(n);

         return n;

         // List of "safe" opcodes; those that implicitly block the memory

         // op below any null check.

       case Op_CastX2P:          // no null checks on native pointers

       case Op_Parm:             // 'this' pointer is not null

       case Op_LoadP:            // Loading from within a klass

       case Op_LoadN:            // Loading from within a klass

       case Op_LoadKlass:        // Loading from within a klass

       case Op_LoadNKlass:       // Loading from within a klass

       case Op_ConP:             // Loading from a klass

       case Op_ConN:             // Loading from a klass

       case Op_ConNKlass:        // Loading from a klass

       case Op_CreateEx:         // Sucking up the guts of an exception oop

       case Op_Con:              // Reading from TLS

       case Op_CMoveP:           // CMoveP is pinned

       case Op_CMoveN:           // CMoveN is pinned

         break;                  // No progress

       case Op_Proj:             // Direct call to an allocation routine

       case Op_SCMemProj:        // Memory state from store conditional ops

 #ifdef ASSERT

         {

           assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");

           const Node* call = adr->in(0);

           if (call->is_CallJava()) {

             const CallJavaNode* call_java = call->as_CallJava();

             const TypeTuple *r = call_java->tf()->range();

             assert(r->cnt() > TypeFunc::Parms, "must return value");

             const Type* ret_type = r->field_at(TypeFunc::Parms);

             assert(ret_type && ret_type->isa_ptr(), "must return pointer");

             // We further presume that this is one of

             // new_instance_Java, new_array_Java, or

             // the like, but do not assert for this.

           } else if (call->is_Allocate()) {

             // similar case to new_instance_Java, etc.

           } else if (!call->is_CallLeaf()) {

             // Projections from fetch_oop (OSR) are allowed as well.

             ShouldNotReachHere();

           }

         }

 #endif

         break;

       default:

         ShouldNotReachHere();

       }

       break;

     }

   }

   return  NULL;               // No progress

 }

 //=============================================================================

 uint LoadNode::size_of() const { return sizeof(*this); }

 uint LoadNode::cmp( const Node &n ) const

 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }

 const Type *LoadNode::bottom_type() const { return _type; }

 uint LoadNode::ideal_reg() const {

   return _type->ideal_reg();

 }

 #ifndef PRODUCT

 void LoadNode::dump_spec(outputStream *st) const {

   MemNode::dump_spec(st);

   if( !Verbose && !WizardMode ) {

     // standard dump does this in Verbose and WizardMode

     st->print(" #"); _type->dump_on(st);

   }

 }

 #endif

 #ifdef ASSERT

 //----------------------------is_immutable_value-------------------------------

 // Helper function to allow a raw load without control edge for some cases

 bool LoadNode::is_immutable_value(Node* adr) {

   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&

           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&

           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==

            in_bytes(JavaThread::osthread_offset())));

 }

 #endif

 //----------------------------LoadNode::make-----------------------------------

 // Polymorphic factory method:

 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {

   Compile* C = gvn.C;

   // sanity check the alias category against the created node type

   assert(!(adr_type->isa_oopptr() &&

            adr_type->offset() == oopDesc::klass_offset_in_bytes()),

          "use LoadKlassNode instead");

   assert(!(adr_type->isa_aryptr() &&

            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),

          "use LoadRangeNode instead");

   // Check control edge of raw loads

   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||

           // oop will be recorded in oop map if load crosses safepoint

           rt->isa_oopptr() || is_immutable_value(adr),

           "raw memory operations should have control edge");

   switch (bt) {

   case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int()    );

   case T_BYTE:    return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int()    );

   case T_INT:     return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int()    );

   case T_CHAR:    return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int()    );

   case T_SHORT:   return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int()    );

   case T_LONG:    return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long()   );

   case T_FLOAT:   return new (C) LoadFNode (ctl, mem, adr, adr_type, rt              );

   case T_DOUBLE:  return new (C) LoadDNode (ctl, mem, adr, adr_type, rt              );

   case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr()    );

   case T_OBJECT:

 #ifdef _LP64

     if (adr->bottom_type()->is_ptr_to_narrowoop()) {

       Node* load  = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));

       return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());

     } else

 #endif

     {

       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");

       return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());

     }

   }

   ShouldNotReachHere();

   return (LoadNode*)NULL;

 }

 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {

   bool require_atomic = true;

   return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);

 }

 //------------------------------hash-------------------------------------------

 uint LoadNode::hash() const {

   // unroll addition of interesting fields

   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);

 }

 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {

   if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {

     bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();

     bool is_stable_ary = FoldStableValues &&

                          (tp != NULL) && (tp->isa_aryptr() != NULL) &&

                          tp->isa_aryptr()->is_stable();

     return (eliminate_boxing && non_volatile) || is_stable_ary;

   }

   return false;

 }

 //---------------------------can_see_stored_value------------------------------

 // This routine exists to make sure this set of tests is done the same

 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal

 // will change the graph shape in a way which makes memory alive twice at the

 // same time (uses the Oracle model of aliasing), then some

 // LoadXNode::Identity will fold things back to the equivalence-class model

 // of aliasing.

 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {

   Node* ld_adr = in(MemNode::Address);

   intptr_t ld_off = 0;

   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);

   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();

   Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;

   // This is more general than load from boxing objects.

   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {

     uint alias_idx = atp->index();

     bool final = !atp->is_rewritable();

     Node* result = NULL;

     Node* current = st;

     // Skip through chains of MemBarNodes checking the MergeMems for

     // new states for the slice of this load.  Stop once any other

     // kind of node is encountered.  Loads from final memory can skip

     // through any kind of MemBar but normal loads shouldn't skip

     // through MemBarAcquire since the could allow them to move out of

     // a synchronized region.

     while (current->is_Proj()) {

       int opc = current->in(0)->Opcode();

       if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) ||

           opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder ||

           opc == Op_MemBarReleaseLock) {

         Node* mem = current->in(0)->in(TypeFunc::Memory);

         if (mem->is_MergeMem()) {

           MergeMemNode* merge = mem->as_MergeMem();

           Node* new_st = merge->memory_at(alias_idx);

           if (new_st == merge->base_memory()) {

             // Keep searching

             current = new_st;

             continue;

           }

           // Save the new memory state for the slice and fall through

           // to exit.

           result = new_st;

         }

       }

       break;

     }

     if (result != NULL) {

       st = result;

     }

   }

   // Loop around twice in the case Load -> Initialize -> Store.

   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)

   for (int trip = 0; trip <= 1; trip++) {

     if (st->is_Store()) {

       Node* st_adr = st->in(MemNode::Address);

       if (!phase->eqv(st_adr, ld_adr)) {

         // Try harder before giving up...  Match raw and non-raw pointers.

         intptr_t st_off = 0;

         AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);

         if (alloc == NULL)       return NULL;

         if (alloc != ld_alloc)   return NULL;

         if (ld_off != st_off)    return NULL;

         // At this point we have proven something like this setup:

         //  A = Allocate(...)

         //  L = LoadQ(,  AddP(CastPP(, A.Parm),, #Off))

         //  S = StoreQ(, AddP(,        A.Parm  , #Off), V)

         // (Actually, we haven't yet proven the Q's are the same.)

         // In other words, we are loading from a casted version of

         // the same pointer-and-offset that we stored to.

         // Thus, we are able to replace L by V.

       }

       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.

       if (store_Opcode() != st->Opcode())

         return NULL;

       return st->in(MemNode::ValueIn);

     }

     // A load from a freshly-created object always returns zero.

     // (This can happen after LoadNode::Ideal resets the load's memory input

     // to find_captured_store, which returned InitializeNode::zero_memory.)

     if (st->is_Proj() && st->in(0)->is_Allocate() &&

         (st->in(0) == ld_alloc) &&

         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {

       // return a zero value for the load's basic type

       // (This is one of the few places where a generic PhaseTransform

       // can create new nodes.  Think of it as lazily manifesting

       // virtually pre-existing constants.)

       return phase->zerocon(memory_type());

     }

     // A load from an initialization barrier can match a captured store.

     if (st->is_Proj() && st->in(0)->is_Initialize()) {

       InitializeNode* init = st->in(0)->as_Initialize();

       AllocateNode* alloc = init->allocation();

       if ((alloc != NULL) && (alloc == ld_alloc)) {

         // examine a captured store value

         st = init->find_captured_store(ld_off, memory_size(), phase);

         if (st != NULL)

           continue;             // take one more trip around

       }

     }

     // Load boxed value from result of valueOf() call is input parameter.

     if (this->is_Load() && ld_adr->is_AddP() &&

         (tp != NULL) && tp->is_ptr_to_boxed_value()) {

       intptr_t ignore = 0;

       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);

       if (base != NULL && base->is_Proj() &&

           base->as_Proj()->_con == TypeFunc::Parms &&

           base->in(0)->is_CallStaticJava() &&

           base->in(0)->as_CallStaticJava()->is_boxing_method()) {

         return base->in(0)->in(TypeFunc::Parms);

       }

     }

     break;

   }

   return NULL;

 }

 //----------------------is_instance_field_load_with_local_phi------------------

 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {

   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&

       in(Address)->is_AddP() ) {

     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();

     // Only instances and boxed values.

     if( t_oop != NULL &&

         (t_oop->is_ptr_to_boxed_value() ||

          t_oop->is_known_instance_field()) &&

         t_oop->offset() != Type::OffsetBot &&

         t_oop->offset() != Type::OffsetTop) {

       return true;

     }

   }

   return false;

 }

 //------------------------------Identity---------------------------------------

 // Loads are identity if previous store is to same address

 Node *LoadNode::Identity( PhaseTransform *phase ) {

   // If the previous store-maker is the right kind of Store, and the store is

   // to the same address, then we are equal to the value stored.

   Node* mem = in(Memory);

   Node* value = can_see_stored_value(mem, phase);

   if( value ) {

     // byte, short & char stores truncate naturally.

     // A load has to load the truncated value which requires

     // some sort of masking operation and that requires an

     // Ideal call instead of an Identity call.

     if (memory_size() < BytesPerInt) {

       // If the input to the store does not fit with the load's result type,

       // it must be truncated via an Ideal call.

       if (!phase->type(value)->higher_equal(phase->type(this)))

         return this;

     }

     // (This works even when value is a Con, but LoadNode::Value

     // usually runs first, producing the singleton type of the Con.)

     return value;

   }

   // Search for an existing data phi which was generated before for the same

   // instance's field to avoid infinite generation of phis in a loop.

   Node *region = mem->in(0);

   if (is_instance_field_load_with_local_phi(region)) {

     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();

     int this_index  = phase->C->get_alias_index(addr_t);

     int this_offset = addr_t->offset();

     int this_iid    = addr_t->instance_id();

     if (!addr_t->is_known_instance() &&

          addr_t->is_ptr_to_boxed_value()) {

       // Use _idx of address base (could be Phi node) for boxed values.

       intptr_t   ignore = 0;

       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);

       this_iid = base->_idx;

     }

     const Type* this_type = bottom_type();

     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {

       Node* phi = region->fast_out(i);

       if (phi->is_Phi() && phi != mem &&

           phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {

         return phi;

       }

     }

   }

   return this;

 }

 // We're loading from an object which has autobox behaviour.

 // If this object is result of a valueOf call we'll have a phi

 // merging a newly allocated object and a load from the cache.

 // We want to replace this load with the original incoming

 // argument to the valueOf call.

 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {

   assert(phase->C->eliminate_boxing(), "sanity");

   intptr_t ignore = 0;

   Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);

   if ((base == NULL) || base->is_Phi()) {

     // Push the loads from the phi that comes from valueOf up

     // through it to allow elimination of the loads and the recovery

     // of the original value. It is done in split_through_phi().

     return NULL;

   } else if (base->is_Load() ||

              base->is_DecodeN() && base->in(1)->is_Load()) {

     // Eliminate the load of boxed value for integer types from the cache

     // array by deriving the value from the index into the array.

     // Capture the offset of the load and then reverse the computation.

     // Get LoadN node which loads a boxing object from 'cache' array.

     if (base->is_DecodeN()) {

       base = base->in(1);

     }

     if (!base->in(Address)->is_AddP()) {

       return NULL; // Complex address

     }

     AddPNode* address = base->in(Address)->as_AddP();

     Node* cache_base = address->in(AddPNode::Base);

     if ((cache_base != NULL) && cache_base->is_DecodeN()) {

       // Get ConP node which is static 'cache' field.

       cache_base = cache_base->in(1);

     }

     if ((cache_base != NULL) && cache_base->is_Con()) {

       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();

       if ((base_type != NULL) && base_type->is_autobox_cache()) {

         Node* elements[4];

         int shift = exact_log2(type2aelembytes(T_OBJECT));

         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));

         if ((count >  0) && elements[0]->is_Con() &&

             ((count == 1) ||

              (count == 2) && elements[1]->Opcode() == Op_LShiftX &&

                              elements[1]->in(2) == phase->intcon(shift))) {

           ciObjArray* array = base_type->const_oop()->as_obj_array();

           // Fetch the box object cache[0] at the base of the array and get its value

           ciInstance* box = array->obj_at(0)->as_instance();

           ciInstanceKlass* ik = box->klass()->as_instance_klass();

           assert(ik->is_box_klass(), "sanity");

           assert(ik->nof_nonstatic_fields() == 1, "change following code");

           if (ik->nof_nonstatic_fields() == 1) {

             // This should be true nonstatic_field_at requires calling

             // nof_nonstatic_fields so check it anyway

             ciConstant c = box->field_value(ik->nonstatic_field_at(0));

             BasicType bt = c.basic_type();

             // Only integer types have boxing cache.

             assert(bt == T_BOOLEAN || bt == T_CHAR  ||

                    bt == T_BYTE    || bt == T_SHORT ||

                    bt == T_INT     || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));

             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();

             if (cache_low != (int)cache_low) {

               return NULL; // should not happen since cache is array indexed by value

             }

             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);

             if (offset != (int)offset) {

               return NULL; // should not happen since cache is array indexed by value

             }

            // Add up all the offsets making of the address of the load

             Node* result = elements[0];

             for (int i = 1; i < count; i++) {

               result = phase->transform(new (phase->C) AddXNode(result, elements[i]));

             }

             // Remove the constant offset from the address and then

             result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));

             // remove the scaling of the offset to recover the original index.

             if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {

               // Peel the shift off directly but wrap it in a dummy node

               // since Ideal can't return existing nodes

               result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));

             } else if (result->is_Add() && result->in(2)->is_Con() &&

                        result->in(1)->Opcode() == Op_LShiftX &&

                        result->in(1)->in(2) == phase->intcon(shift)) {

               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)

               // but for boxing cache access we know that X<<Z will not overflow

               // (there is range check) so we do this optimizatrion by hand here.

               Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));

               result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));

             } else {

               result = new (phase->C) RShiftXNode(result, phase->intcon(shift));

             }

 #ifdef _LP64

             if (bt != T_LONG) {

               result = new (phase->C) ConvL2INode(phase->transform(result));

             }

 #else

             if (bt == T_LONG) {

               result = new (phase->C) ConvI2LNode(phase->transform(result));

             }

 #endif

             return result;

           }

         }

       }

     }

   }

   return NULL;

 }

 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {

   Node* region = phi->in(0);

   if (region == NULL) {

     return false; // Wait stable graph

   }

   uint cnt = phi->req();

   for (uint i = 1; i < cnt; i++) {

     Node* rc = region->in(i);

     if (rc == NULL || phase->type(rc) == Type::TOP)

       return false; // Wait stable graph

     Node* in = phi->in(i);

     if (in == NULL || phase->type(in) == Type::TOP)

       return false; // Wait stable graph

   }

   return true;

 }

 //------------------------------split_through_phi------------------------------

 // Split instance or boxed field load through Phi.

 Node *LoadNode::split_through_phi(PhaseGVN *phase) {

   Node* mem     = in(Memory);

   Node* address = in(Address);

   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();

   assert((t_oop != NULL) &&

          (t_oop->is_known_instance_field() ||

           t_oop->is_ptr_to_boxed_value()), "invalide conditions");

   Compile* C = phase->C;

   intptr_t ignore = 0;

   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);

   bool base_is_phi = (base != NULL) && base->is_Phi();

   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&

                            (base != NULL) && (base == address->in(AddPNode::Base)) &&

                            phase->type(base)->higher_equal(TypePtr::NOTNULL);

   if (!((mem->is_Phi() || base_is_phi) &&

         (load_boxed_values || t_oop->is_known_instance_field()))) {

     return NULL; // memory is not Phi

   }

   if (mem->is_Phi()) {

     if (!stable_phi(mem->as_Phi(), phase)) {

       return NULL; // Wait stable graph

     }

     uint cnt = mem->req();

     // Check for loop invariant memory.

     if (cnt == 3) {

       for (uint i = 1; i < cnt; i++) {

         Node* in = mem->in(i);

         Node*  m = optimize_memory_chain(in, t_oop, this, phase);

         if (m == mem) {

           set_req(Memory, mem->in(cnt - i));

           return this; // made change

         }

       }

     }

   }

   if (base_is_phi) {

     if (!stable_phi(base->as_Phi(), phase)) {

       return NULL; // Wait stable graph

     }

     uint cnt = base->req();

     // Check for loop invariant memory.

     if (cnt == 3) {

       for (uint i = 1; i < cnt; i++) {

         if (base->in(i) == base) {

           return NULL; // Wait stable graph

         }

       }

     }

   }

   bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));

   // Split through Phi (see original code in loopopts.cpp).

   assert(C->have_alias_type(t_oop), "instance should have alias type");

   // Do nothing here if Identity will find a value

   // (to avoid infinite chain of value phis generation).

   if (!phase->eqv(this, this->Identity(phase)))

     return NULL;

   // Select Region to split through.

   Node* region;

   if (!base_is_phi) {

     assert(mem->is_Phi(), "sanity");

     region = mem->in(0);

     // Skip if the region dominates some control edge of the address.

     if (!MemNode::all_controls_dominate(address, region))

       return NULL;

   } else if (!mem->is_Phi()) {

     assert(base_is_phi, "sanity");

     region = base->in(0);

     // Skip if the region dominates some control edge of the memory.

     if (!MemNode::all_controls_dominate(mem, region))

       return NULL;

   } else if (base->in(0) != mem->in(0)) {

     assert(base_is_phi && mem->is_Phi(), "sanity");

     if (MemNode::all_controls_dominate(mem, base->in(0))) {

       region = base->in(0);

     } else if (MemNode::all_controls_dominate(address, mem->in(0))) {

       region = mem->in(0);

     } else {

       return NULL; // complex graph

     }

   } else {

     assert(base->in(0) == mem->in(0), "sanity");

     region = mem->in(0);

   }

   const Type* this_type = this->bottom_type();

   int this_index  = C->get_alias_index(t_oop);

   int this_offset = t_oop->offset();

   int this_iid    = t_oop->instance_id();

   if (!t_oop->is_known_instance() && load_boxed_values) {

     // Use _idx of address base for boxed values.

     this_iid = base->_idx;

   }

   PhaseIterGVN* igvn = phase->is_IterGVN();

   Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);

   for (uint i = 1; i < region->req(); i++) {

     Node* x;

     Node* the_clone = NULL;

     if (region->in(i) == C->top()) {

       x = C->top();      // Dead path?  Use a dead data op

     } else {

       x = this->clone();        // Else clone up the data op

       the_clone = x;            // Remember for possible deletion.

       // Alter data node to use pre-phi inputs

       if (this->in(0) == region) {

         x->set_req(0, region->in(i));

       } else {

         x->set_req(0, NULL);

       }

       if (mem->is_Phi() && (mem->in(0) == region)) {

         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.

       }

       if (address->is_Phi() && address->in(0) == region) {

         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone

       }

       if (base_is_phi && (base->in(0) == region)) {

         Node* base_x = base->in(i); // Clone address for loads from boxed objects.

         Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));

         x->set_req(Address, adr_x);

       }

     }

     // Check for a 'win' on some paths

     const Type *t = x->Value(igvn);

     bool singleton = t->singleton();

     // See comments in PhaseIdealLoop::split_thru_phi().

     if (singleton && t == Type::TOP) {

       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);

     }

     if (singleton) {

       x = igvn->makecon(t);

     } else {

       // We now call Identity to try to simplify the cloned node.

       // Note that some Identity methods call phase->type(this).

       // Make sure that the type array is big enough for

       // our new node, even though we may throw the node away.

       // (This tweaking with igvn only works because x is a new node.)

       igvn->set_type(x, t);

       // If x is a TypeNode, capture any more-precise type permanently into Node

       // otherwise it will be not updated during igvn->transform since

       // igvn->type(x) is set to x->Value() already.

       x->raise_bottom_type(t);

       Node *y = x->Identity(igvn);

       if (y != x) {

         x = y;

       } else {

         y = igvn->hash_find_insert(x);

         if (y) {

           x = y;

         } else {

           // Else x is a new node we are keeping

           // We do not need register_new_node_with_optimizer

           // because set_type has already been called.

           igvn->_worklist.push(x);

         }

       }

     }

     if (x != the_clone && the_clone != NULL) {

       igvn->remove_dead_node(the_clone);

     }

     phi->set_req(i, x);

   }

   // Record Phi

   igvn->register_new_node_with_optimizer(phi);

   return phi;

 }

 //------------------------------Ideal------------------------------------------

 // If the load is from Field memory and the pointer is non-null, we can

 // zero out the control input.

 // If the offset is constant and the base is an object allocation,

 // try to hook me up to the exact initializing store.

 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   Node* p = MemNode::Ideal_common(phase, can_reshape);

   if (p)  return (p == NodeSentinel) ? NULL : p;

   Node* ctrl    = in(MemNode::Control);

   Node* address = in(MemNode::Address);

   // Skip up past a SafePoint control.  Cannot do this for Stores because

   // pointer stores & cardmarks must stay on the same side of a SafePoint.

   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&

       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {

     ctrl = ctrl->in(0);

     set_req(MemNode::Control,ctrl);

   }

   intptr_t ignore = 0;

   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);

   if (base != NULL

       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {

     // Check for useless control edge in some common special cases

     if (in(MemNode::Control) != NULL

         && phase->type(base)->higher_equal(TypePtr::NOTNULL)

         && all_controls_dominate(base, phase->C->start())) {

       // A method-invariant, non-null address (constant or 'this' argument).

       set_req(MemNode::Control, NULL);

     }

   }

   Node* mem = in(MemNode::Memory);

   const TypePtr *addr_t = phase->type(address)->isa_ptr();

   if (can_reshape && (addr_t != NULL)) {

     // try to optimize our memory input

     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);

     if (opt_mem != mem) {

       set_req(MemNode::Memory, opt_mem);

       if (phase->type( opt_mem ) == Type::TOP) return NULL;

       return this;

     }

     const TypeOopPtr *t_oop = addr_t->isa_oopptr();

     if ((t_oop != NULL) &&

         (t_oop->is_known_instance_field() ||

          t_oop->is_ptr_to_boxed_value())) {

       PhaseIterGVN *igvn = phase->is_IterGVN();

       if (igvn != NULL && igvn->_worklist.member(opt_mem)) {

         // Delay this transformation until memory Phi is processed.

         phase->is_IterGVN()->_worklist.push(this);

         return NULL;

       }

       // Split instance field load through Phi.

       Node* result = split_through_phi(phase);

       if (result != NULL) return result;

       if (t_oop->is_ptr_to_boxed_value()) {

         Node* result = eliminate_autobox(phase);

         if (result != NULL) return result;

       }

     }

   }

   // Check for prior store with a different base or offset; make Load

   // independent.  Skip through any number of them.  Bail out if the stores

   // are in an endless dead cycle and report no progress.  This is a key

   // transform for Reflection.  However, if after skipping through the Stores

   // we can't then fold up against a prior store do NOT do the transform as

   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same

   // array memory alive twice: once for the hoisted Load and again after the

   // bypassed Store.  This situation only works if EVERYBODY who does

   // anti-dependence work knows how to bypass.  I.e. we need all

   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is

   // the alias index stuff.  So instead, peek through Stores and IFF we can

   // fold up, do so.

   Node* prev_mem = find_previous_store(phase);

   // Steps (a), (b):  Walk past independent stores to find an exact match.

   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {

     // (c) See if we can fold up on the spot, but don't fold up here.

     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or

     // just return a prior value, which is done by Identity calls.

     if (can_see_stored_value(prev_mem, phase)) {

       // Make ready for step (d):

       set_req(MemNode::Memory, prev_mem);

       return this;

     }

   }

   return NULL;                  // No further progress

 }

 // Helper to recognize certain Klass fields which are invariant across

 // some group of array types (e.g., int[] or all T[] where T < Object).

 const Type*

 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,

                                  ciKlass* klass) const {

   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {

     // The field is Klass::_modifier_flags.  Return its (constant) value.

     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)

     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");

     return TypeInt::make(klass->modifier_flags());

   }

   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {

     // The field is Klass::_access_flags.  Return its (constant) value.

     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)

     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");

     return TypeInt::make(klass->access_flags());

   }

   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {

     // The field is Klass::_layout_helper.  Return its constant value if known.

     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");

     return TypeInt::make(klass->layout_helper());

   }

   // No match.

   return NULL;

 }

 // Try to constant-fold a stable array element.

 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {

   assert(ary->is_stable(), "array should be stable");

   if (ary->const_oop() != NULL) {

     // Decode the results of GraphKit::array_element_address.

     ciArray* aobj = ary->const_oop()->as_array();

     ciConstant con = aobj->element_value_by_offset(off);

     if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {

       const Type* con_type = Type::make_from_constant(con);

       if (con_type != NULL) {

         if (con_type->isa_aryptr()) {

           // Join with the array element type, in case it is also stable.

           int dim = ary->stable_dimension();

           con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);

         }

         if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {

           con_type = con_type->make_narrowoop();

         }

 #ifndef PRODUCT

         if (TraceIterativeGVN) {

           tty->print("FoldStableValues: array element [off=%d]: con_type=", off);

           con_type->dump(); tty->cr();

         }

 #endif //PRODUCT

         return con_type;

       }

     }

   }

   return NULL;

 }

 //------------------------------Value-----------------------------------------

 const Type *LoadNode::Value( PhaseTransform *phase ) const {

   // Either input is TOP ==> the result is TOP

   Node* mem = in(MemNode::Memory);

   const Type *t1 = phase->type(mem);

   if (t1 == Type::TOP)  return Type::TOP;

   Node* adr = in(MemNode::Address);

   const TypePtr* tp = phase->type(adr)->isa_ptr();

   if (tp == NULL || tp->empty())  return Type::TOP;

   int off = tp->offset();

   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");

   Compile* C = phase->C;

   // Try to guess loaded type from pointer type

   if (tp->isa_aryptr()) {

     const TypeAryPtr* ary = tp->is_aryptr();

     const Type *t = ary->elem();

     // Determine whether the reference is beyond the header or not, by comparing

     // the offset against the offset of the start of the array's data.

     // Different array types begin at slightly different offsets (12 vs. 16).

     // We choose T_BYTE as an example base type that is least restrictive

     // as to alignment, which will therefore produce the smallest

     // possible base offset.

     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);

     const bool off_beyond_header = ((uint)off >= (uint)min_base_off);

     // Try to constant-fold a stable array element.

     if (FoldStableValues && ary->is_stable()) {

       // Make sure the reference is not into the header

       if (off_beyond_header && off != Type::OffsetBot) {

         assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con(), "offset is a constant");

         const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());

         if (con_type != NULL) {

           return con_type;

         }

       }

     }

     // Don't do this for integer types. There is only potential profit if

     // the element type t is lower than _type; that is, for int types, if _type is

     // more restrictive than t.  This only happens here if one is short and the other

     // char (both 16 bits), and in those cases we've made an intentional decision

     // to use one kind of load over the other. See AndINode::Ideal and 4965907.

     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.

     //

     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))

     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier

     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been

     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,

     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.

     // In fact, that could have been the original type of p1, and p1 could have

     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the

     // expression (LShiftL quux 3) independently optimized to the constant 8.

     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)

         && (_type->isa_vect() == NULL)

         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {

       // t might actually be lower than _type, if _type is a unique

       // concrete subclass of abstract class t.

       if (off_beyond_header) {  // is the offset beyond the header?

         const Type* jt = t->join(_type);

         // In any case, do not allow the join, per se, to empty out the type.

         if (jt->empty() && !t->empty()) {

           // This can happen if a interface-typed array narrows to a class type.

           jt = _type;

         }

 #ifdef ASSERT

         if (phase->C->eliminate_boxing() && adr->is_AddP()) {

           // The pointers in the autobox arrays are always non-null

           Node* base = adr->in(AddPNode::Base);

           if ((base != NULL) && base->is_DecodeN()) {

             // Get LoadN node which loads IntegerCache.cache field

             base = base->in(1);

           }

           if ((base != NULL) && base->is_Con()) {

             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();

             if ((base_type != NULL) && base_type->is_autobox_cache()) {

               // It could be narrow oop

               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");

             }

           }

         }

 #endif

         return jt;

       }

     }

   } else if (tp->base() == Type::InstPtr) {

     ciEnv* env = C->env();

     const TypeInstPtr* tinst = tp->is_instptr();

     ciKlass* klass = tinst->klass();

     assert( off != Type::OffsetBot ||

             // arrays can be cast to Objects

             tp->is_oopptr()->klass()->is_java_lang_Object() ||

             // unsafe field access may not have a constant offset

             C->has_unsafe_access(),

             "Field accesses must be precise" );

     // For oop loads, we expect the _type to be precise

     if (klass == env->String_klass() &&

         adr->is_AddP() && off != Type::OffsetBot) {

       // For constant Strings treat the final fields as compile time constants.

       Node* base = adr->in(AddPNode::Base);

       const TypeOopPtr* t = phase->type(base)->isa_oopptr();

       if (t != NULL && t->singleton()) {

         ciField* field = env->String_klass()->get_field_by_offset(off, false);

         if (field != NULL && field->is_final()) {

           ciObject* string = t->const_oop();

           ciConstant constant = string->as_instance()->field_value(field);

           if (constant.basic_type() == T_INT) {

             return TypeInt::make(constant.as_int());

           } else if (constant.basic_type() == T_ARRAY) {

             if (adr->bottom_type()->is_ptr_to_narrowoop()) {

               return TypeNarrowOop::make_from_constant(constant.as_object(), true);

             } else {

               return TypeOopPtr::make_from_constant(constant.as_object(), true);

             }

           }

         }

       }

     }

     // Optimizations for constant objects

     ciObject* const_oop = tinst->const_oop();

     if (const_oop != NULL) {

       // For constant Boxed value treat the target field as a compile time constant.

       if (tinst->is_ptr_to_boxed_value()) {

         return tinst->get_const_boxed_value();

       } else

       // For constant CallSites treat the target field as a compile time constant.

       if (const_oop->is_call_site()) {

         ciCallSite* call_site = const_oop->as_call_site();

         ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);

         if (field != NULL && field->is_call_site_target()) {

           ciMethodHandle* target = call_site->get_target();

           if (target != NULL) {  // just in case

             ciConstant constant(T_OBJECT, target);

             const Type* t;

             if (adr->bottom_type()->is_ptr_to_narrowoop()) {

               t = TypeNarrowOop::make_from_constant(constant.as_object(), true);

             } else {

               t = TypeOopPtr::make_from_constant(constant.as_object(), true);

             }

             // Add a dependence for invalidation of the optimization.

             if (!call_site->is_constant_call_site()) {

               C->dependencies()->assert_call_site_target_value(call_site, target);

             }

             return t;

           }

         }

       }

     }

   } else if (tp->base() == Type::KlassPtr) {

     assert( off != Type::OffsetBot ||

             // arrays can be cast to Objects

             tp->is_klassptr()->klass()->is_java_lang_Object() ||

             // also allow array-loading from the primary supertype

             // array during subtype checks

             Opcode() == Op_LoadKlass,

             "Field accesses must be precise" );

     // For klass/static loads, we expect the _type to be precise

   }

   const TypeKlassPtr *tkls = tp->isa_klassptr();

   if (tkls != NULL && !StressReflectiveCode) {

     ciKlass* klass = tkls->klass();

     if (klass->is_loaded() && tkls->klass_is_exact()) {

       // We are loading a field from a Klass metaobject whose identity

       // is known at compile time (the type is "exact" or "precise").

       // Check for fields we know are maintained as constants by the VM.

       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {

         // The field is Klass::_super_check_offset.  Return its (constant) value.

         // (Folds up type checking code.)

         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");

         return TypeInt::make(klass->super_check_offset());

       }

       // Compute index into primary_supers array

       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);

       // Check for overflowing; use unsigned compare to handle the negative case.

       if( depth < ciKlass::primary_super_limit() ) {

         // The field is an element of Klass::_primary_supers.  Return its (constant) value.

         // (Folds up type checking code.)

         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");

         ciKlass *ss = klass->super_of_depth(depth);

         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;

       }

       const Type* aift = load_array_final_field(tkls, klass);

       if (aift != NULL)  return aift;

       if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())

           && klass->is_array_klass()) {

         // The field is ArrayKlass::_component_mirror.  Return its (constant) value.

         // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)

         assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");

         return TypeInstPtr::make(klass->as_array_klass()->component_mirror());

       }

       if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {

         // The field is Klass::_java_mirror.  Return its (constant) value.

         // (Folds up the 2nd indirection in anObjConstant.getClass().)

         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");

         return TypeInstPtr::make(klass->java_mirror());

       }

     }

     // We can still check if we are loading from the primary_supers array at a

     // shallow enough depth.  Even though the klass is not exact, entries less

     // than or equal to its super depth are correct.

     if (klass->is_loaded() ) {

       ciType *inner = klass;

       while( inner->is_obj_array_klass() )

         inner = inner->as_obj_array_klass()->base_element_type();

       if( inner->is_instance_klass() &&

           !inner->as_instance_klass()->flags().is_interface() ) {

         // Compute index into primary_supers array

         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);

         // Check for overflowing; use unsigned compare to handle the negative case.

         if( depth < ciKlass::primary_super_limit() &&

             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case

           // The field is an element of Klass::_primary_supers.  Return its (constant) value.

           // (Folds up type checking code.)

           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");

           ciKlass *ss = klass->super_of_depth(depth);

           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;

         }

       }

     }

     // If the type is enough to determine that the thing is not an array,

     // we can give the layout_helper a positive interval type.

     // This will help short-circuit some reflective code.

     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())

         && !klass->is_array_klass() // not directly typed as an array

         && !klass->is_interface()  // specifically not Serializable & Cloneable

         && !klass->is_java_lang_Object()   // not the supertype of all T[]

         ) {

       // Note:  When interfaces are reliable, we can narrow the interface

       // test to (klass != Serializable && klass != Cloneable).

       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");

       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);

       // The key property of this type is that it folds up tests

       // for array-ness, since it proves that the layout_helper is positive.

       // Thus, a generic value like the basic object layout helper works fine.

       return TypeInt::make(min_size, max_jint, Type::WidenMin);

     }

   }

   // If we are loading from a freshly-allocated object, produce a zero,

   // if the load is provably beyond the header of the object.

   // (Also allow a variable load from a fresh array to produce zero.)

   const TypeOopPtr *tinst = tp->isa_oopptr();

   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();

   bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();

   if (ReduceFieldZeroing || is_instance || is_boxed_value) {

     Node* value = can_see_stored_value(mem,phase);

     if (value != NULL && value->is_Con()) {

       assert(value->bottom_type()->higher_equal(_type),"sanity");

       return value->bottom_type();

     }

   }

   if (is_instance) {

     // If we have an instance type and our memory input is the

     // programs's initial memory state, there is no matching store,

     // so just return a zero of the appropriate type

     Node *mem = in(MemNode::Memory);

     if (mem->is_Parm() && mem->in(0)->is_Start()) {

       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");

       return Type::get_zero_type(_type->basic_type());

     }

   }

   return _type;

 }

 //------------------------------match_edge-------------------------------------

 // Do we Match on this edge index or not?  Match only the address.

 uint LoadNode::match_edge(uint idx) const {

   return idx == MemNode::Address;

 }

 //--------------------------LoadBNode::Ideal--------------------------------------

 //

 //  If the previous store is to the same address as this load,

 //  and the value stored was larger than a byte, replace this load

 //  with the value stored truncated to a byte.  If no truncation is

 //  needed, the replacement is done in LoadNode::Identity().

 //

 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   Node* mem = in(MemNode::Memory);

   Node* value = can_see_stored_value(mem,phase);

   if( value && !phase->type(value)->higher_equal( _type ) ) {

     Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );

     return new (phase->C) RShiftINode(result, phase->intcon(24));

   }

   // Identity call will handle the case where truncation is not needed.

   return LoadNode::Ideal(phase, can_reshape);

 }

 const Type* LoadBNode::Value(PhaseTransform *phase) const {

   Node* mem = in(MemNode::Memory);

   Node* value = can_see_stored_value(mem,phase);

   if (value != NULL && value->is_Con() &&

       !value->bottom_type()->higher_equal(_type)) {

     // If the input to the store does not fit with the load's result type,

     // it must be truncated. We can't delay until Ideal call since

     // a singleton Value is needed for split_thru_phi optimization.

     int con = value->get_int();

     return TypeInt::make((con << 24) >> 24);

   }

   return LoadNode::Value(phase);

 }

 //--------------------------LoadUBNode::Ideal-------------------------------------

 //

 //  If the previous store is to the same address as this load,

 //  and the value stored was larger than a byte, replace this load

 //  with the value stored truncated to a byte.  If no truncation is

 //  needed, the replacement is done in LoadNode::Identity().

 //

 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {

   Node* mem = in(MemNode::Memory);

   Node* value = can_see_stored_value(mem, phase);

   if (value && !phase->type(value)->higher_equal(_type))

     return new (phase->C) AndINode(value, phase->intcon(0xFF));

   // Identity call will handle the case where truncation is not needed.

   return LoadNode::Ideal(phase, can_reshape);

 }

 const Type* LoadUBNode::Value(PhaseTransform *phase) const {

   Node* mem = in(MemNode::Memory);

   Node* value = can_see_stored_value(mem,phase);

   if (value != NULL && value->is_Con() &&

       !value->bottom_type()->higher_equal(_type)) {

     // If the input to the store does not fit with the load's result type,

     // it must be truncated. We can't delay until Ideal call since

     // a singleton Value is needed for split_thru_phi optimization.

     int con = value->get_int();

     return TypeInt::make(con & 0xFF);

   }

   return LoadNode::Value(phase);

 }

 //--------------------------LoadUSNode::Ideal-------------------------------------

 //

 //  If the previous store is to the same address as this load,

 //  and the value stored was larger than a char, replace this load

 //  with the value stored truncated to a char.  If no truncation is

 //  needed, the replacement is done in LoadNode::Identity().

 //

 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   Node* mem = in(MemNode::Memory);

   Node* value = can_see_stored_value(mem,phase);

   if( value && !phase->type(value)->higher_equal( _type ) )

     return new (phase->C) AndINode(value,phase->intcon(0xFFFF));

   // Identity call will handle the case where truncation is not needed.

   return LoadNode::Ideal(phase, can_reshape);

 }

 const Type* LoadUSNode::Value(PhaseTransform *phase) const {

   Node* mem = in(MemNode::Memory);

   Node* value = can_see_stored_value(mem,phase);

   if (value != NULL && value->is_Con() &&

       !value->bottom_type()->higher_equal(_type)) {

     // If the input to the store does not fit with the load's result type,

     // it must be truncated. We can't delay until Ideal call since

     // a singleton Value is needed for split_thru_phi optimization.

     int con = value->get_int();

     return TypeInt::make(con & 0xFFFF);

   }

   return LoadNode::Value(phase);

 }

 //--------------------------LoadSNode::Ideal--------------------------------------

 //

 //  If the previous store is to the same address as this load,

 //  and the value stored was larger than a short, replace this load

 //  with the value stored truncated to a short.  If no truncation is

 //  needed, the replacement is done in LoadNode::Identity().

 //

 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   Node* mem = in(MemNode::Memory);

   Node* value = can_see_stored_value(mem,phase);

   if( value && !phase->type(value)->higher_equal( _type ) ) {

     Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );

     return new (phase->C) RShiftINode(result, phase->intcon(16));

   }

   // Identity call will handle the case where truncation is not needed.

   return LoadNode::Ideal(phase, can_reshape);

 }

 const Type* LoadSNode::Value(PhaseTransform *phase) const {

   Node* mem = in(MemNode::Memory);

   Node* value = can_see_stored_value(mem,phase);

   if (value != NULL && value->is_Con() &&

       !value->bottom_type()->higher_equal(_type)) {

     // If the input to the store does not fit with the load's result type,

     // it must be truncated. We can't delay until Ideal call since

     // a singleton Value is needed for split_thru_phi optimization.

     int con = value->get_int();

     return TypeInt::make((con << 16) >> 16);

   }

   return LoadNode::Value(phase);

 }

 //=============================================================================

 //----------------------------LoadKlassNode::make------------------------------

 // Polymorphic factory method:

 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {

   Compile* C = gvn.C;

   Node *ctl = NULL;

   // sanity check the alias category against the created node type

   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();

   assert(adr_type != NULL, "expecting TypeKlassPtr");

 #ifdef _LP64

   if (adr_type->is_ptr_to_narrowklass()) {

     assert(UseCompressedClassPointers, "no compressed klasses");

     Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass()));

     return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());

   }

 #endif

   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");

   return new (C) LoadKlassNode(ctl, mem, adr, at, tk);

 }

 //------------------------------Value------------------------------------------

 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {

   return klass_value_common(phase);

 }

 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {

   // Either input is TOP ==> the result is TOP

   const Type *t1 = phase->type( in(MemNode::Memory) );

   if (t1 == Type::TOP)  return Type::TOP;

   Node *adr = in(MemNode::Address);

   const Type *t2 = phase->type( adr );

   if (t2 == Type::TOP)  return Type::TOP;

   const TypePtr *tp = t2->is_ptr();

   if (TypePtr::above_centerline(tp->ptr()) ||

       tp->ptr() == TypePtr::Null)  return Type::TOP;

   // Return a more precise klass, if possible

   const TypeInstPtr *tinst = tp->isa_instptr();

   if (tinst != NULL) {

     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();

     int offset = tinst->offset();

     if (ik == phase->C->env()->Class_klass()

         && (offset == java_lang_Class::klass_offset_in_bytes() ||

             offset == java_lang_Class::array_klass_offset_in_bytes())) {

       // We are loading a special hidden field from a Class mirror object,

       // the field which points to the VM's Klass metaobject.

       ciType* t = tinst->java_mirror_type();

       // java_mirror_type returns non-null for compile-time Class constants.

       if (t != NULL) {

         // constant oop => constant klass

         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {

           return TypeKlassPtr::make(ciArrayKlass::make(t));

         }

         if (!t->is_klass()) {

           // a primitive Class (e.g., int.class) has NULL for a klass field

           return TypePtr::NULL_PTR;

         }

         // (Folds up the 1st indirection in aClassConstant.getModifiers().)

         return TypeKlassPtr::make(t->as_klass());

       }

       // non-constant mirror, so we can't tell what's going on

     }

     if( !ik->is_loaded() )

       return _type;             // Bail out if not loaded

     if (offset == oopDesc::klass_offset_in_bytes()) {

       if (tinst->klass_is_exact()) {

         return TypeKlassPtr::make(ik);

       }

       // See if we can become precise: no subklasses and no interface

       // (Note:  We need to support verified interfaces.)

       if (!ik->is_interface() && !ik->has_subklass()) {

         //assert(!UseExactTypes, "this code should be useless with exact types");

         // Add a dependence; if any subclass added we need to recompile

         if (!ik->is_final()) {

           // %%% should use stronger assert_unique_concrete_subtype instead

           phase->C->dependencies()->assert_leaf_type(ik);

         }

         // Return precise klass

         return TypeKlassPtr::make(ik);

       }

       // Return root of possible klass

       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);

     }

   }

   // Check for loading klass from an array

   const TypeAryPtr *tary = tp->isa_aryptr();

   if( tary != NULL ) {

     ciKlass *tary_klass = tary->klass();

     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP

         && tary->offset() == oopDesc::klass_offset_in_bytes()) {

       if (tary->klass_is_exact()) {

         return TypeKlassPtr::make(tary_klass);

       }

       ciArrayKlass *ak = tary->klass()->as_array_klass();

       // If the klass is an object array, we defer the question to the

       // array component klass.

       if( ak->is_obj_array_klass() ) {

         assert( ak->is_loaded(), "" );

         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();

         if( base_k->is_loaded() && base_k->is_instance_klass() ) {

           ciInstanceKlass* ik = base_k->as_instance_klass();

           // See if we can become precise: no subklasses and no interface

           if (!ik->is_interface() && !ik->has_subklass()) {

             //assert(!UseExactTypes, "this code should be useless with exact types");

             // Add a dependence; if any subclass added we need to recompile

             if (!ik->is_final()) {

               phase->C->dependencies()->assert_leaf_type(ik);

             }

             // Return precise array klass

             return TypeKlassPtr::make(ak);

           }

         }

         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);

       } else {                  // Found a type-array?

         //assert(!UseExactTypes, "this code should be useless with exact types");

         assert( ak->is_type_array_klass(), "" );

         return TypeKlassPtr::make(ak); // These are always precise

       }

     }

   }

   // Check for loading klass from an array klass

   const TypeKlassPtr *tkls = tp->isa_klassptr();

   if (tkls != NULL && !StressReflectiveCode) {

     ciKlass* klass = tkls->klass();

     if( !klass->is_loaded() )

       return _type;             // Bail out if not loaded

     if( klass->is_obj_array_klass() &&

         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {

       ciKlass* elem = klass->as_obj_array_klass()->element_klass();

       // // Always returning precise element type is incorrect,

       // // e.g., element type could be object and array may contain strings

       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);

       // The array's TypeKlassPtr was declared 'precise' or 'not precise'

       // according to the element type's subclassing.

       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);

     }

     if( klass->is_instance_klass() && tkls->klass_is_exact() &&

         tkls->offset() == in_bytes(Klass::super_offset())) {

       ciKlass* sup = klass->as_instance_klass()->super();

       // The field is Klass::_super.  Return its (constant) value.

       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)

       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;

     }

   }

   // Bailout case

   return LoadNode::Value(phase);

 }

 //------------------------------Identity---------------------------------------

 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.

 // Also feed through the klass in Allocate(...klass...)._klass.

 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {

   return klass_identity_common(phase);

 }

 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {

   Node* x = LoadNode::Identity(phase);

   if (x != this)  return x;

   // Take apart the address into an oop and and offset.

   // Return 'this' if we cannot.

   Node*    adr    = in(MemNode::Address);

   intptr_t offset = 0;

   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);

   if (base == NULL)     return this;

   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();

   if (toop == NULL)     return this;

   // We can fetch the klass directly through an AllocateNode.

   // This works even if the klass is not constant (clone or newArray).

   if (offset == oopDesc::klass_offset_in_bytes()) {

     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);

     if (allocated_klass != NULL) {

       return allocated_klass;

     }

   }

   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.

   // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.

   // See inline_native_Class_query for occurrences of these patterns.

   // Java Example:  x.getClass().isAssignableFrom(y)

   // Java Example:  Array.newInstance(x.getClass().getComponentType(), n)

   //

   // This improves reflective code, often making the Class

   // mirror go completely dead.  (Current exception:  Class

   // mirrors may appear in debug info, but we could clean them out by

   // introducing a new debug info operator for Klass*.java_mirror).

   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()

       && (offset == java_lang_Class::klass_offset_in_bytes() ||

           offset == java_lang_Class::array_klass_offset_in_bytes())) {

     // We are loading a special hidden field from a Class mirror,

     // the field which points to its Klass or ArrayKlass metaobject.

     if (base->is_Load()) {

       Node* adr2 = base->in(MemNode::Address);

       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();

       if (tkls != NULL && !tkls->empty()

           && (tkls->klass()->is_instance_klass() ||

               tkls->klass()->is_array_klass())

           && adr2->is_AddP()

           ) {

         int mirror_field = in_bytes(Klass::java_mirror_offset());

         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {

           mirror_field = in_bytes(ArrayKlass::component_mirror_offset());

         }

         if (tkls->offset() == mirror_field) {

           return adr2->in(AddPNode::Base);

         }

       }

     }

   }

   return this;

 }

 //------------------------------Value------------------------------------------

 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {

   const Type *t = klass_value_common(phase);

   if (t == Type::TOP)

     return t;

   return t->make_narrowklass();

 }

 //------------------------------Identity---------------------------------------

 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.

 // Also feed through the klass in Allocate(...klass...)._klass.

 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {

   Node *x = klass_identity_common(phase);

   const Type *t = phase->type( x );

   if( t == Type::TOP ) return x;

   if( t->isa_narrowklass()) return x;

   assert (!t->isa_narrowoop(), "no narrow oop here");

   return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));

 }

 //------------------------------Value-----------------------------------------

 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {

   // Either input is TOP ==> the result is TOP

   const Type *t1 = phase->type( in(MemNode::Memory) );

   if( t1 == Type::TOP ) return Type::TOP;

   Node *adr = in(MemNode::Address);

   const Type *t2 = phase->type( adr );

   if( t2 == Type::TOP ) return Type::TOP;

   const TypePtr *tp = t2->is_ptr();

   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;

   const TypeAryPtr *tap = tp->isa_aryptr();

   if( !tap ) return _type;

   return tap->size();

 }

 //-------------------------------Ideal---------------------------------------

 // Feed through the length in AllocateArray(...length...)._length.

 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   Node* p = MemNode::Ideal_common(phase, can_reshape);

   if (p)  return (p == NodeSentinel) ? NULL : p;

   // Take apart the address into an oop and and offset.

   // Return 'this' if we cannot.

   Node*    adr    = in(MemNode::Address);

   intptr_t offset = 0;

   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);

   if (base == NULL)     return NULL;

   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();

   if (tary == NULL)     return NULL;

   // We can fetch the length directly through an AllocateArrayNode.

   // This works even if the length is not constant (clone or newArray).

   if (offset == arrayOopDesc::length_offset_in_bytes()) {

     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);

     if (alloc != NULL) {

       Node* allocated_length = alloc->Ideal_length();

       Node* len = alloc->make_ideal_length(tary, phase);

       if (allocated_length != len) {

         // New CastII improves on this.

         return len;

       }

     }

   }

   return NULL;

 }

 //------------------------------Identity---------------------------------------

 // Feed through the length in AllocateArray(...length...)._length.

 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {

   Node* x = LoadINode::Identity(phase);

   if (x != this)  return x;

   // Take apart the address into an oop and and offset.

   // Return 'this' if we cannot.

   Node*    adr    = in(MemNode::Address);

   intptr_t offset = 0;

   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);

   if (base == NULL)     return this;

   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();

   if (tary == NULL)     return this;

   // We can fetch the length directly through an AllocateArrayNode.

   // This works even if the length is not constant (clone or newArray).

   if (offset == arrayOopDesc::length_offset_in_bytes()) {

     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);

     if (alloc != NULL) {

       Node* allocated_length = alloc->Ideal_length();

       // Do not allow make_ideal_length to allocate a CastII node.

       Node* len = alloc->make_ideal_length(tary, phase, false);

       if (allocated_length == len) {

         // Return allocated_length only if it would not be improved by a CastII.

         return allocated_length;

       }

     }

   }

   return this;

 }

 //=============================================================================

 //---------------------------StoreNode::make-----------------------------------

 // Polymorphic factory method:

 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {

   Compile* C = gvn.C;

   assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||

           ctl != NULL, "raw memory operations should have control edge");

   switch (bt) {

   case T_BOOLEAN:

   case T_BYTE:    return new (C) StoreBNode(ctl, mem, adr, adr_type, val);

   case T_INT:     return new (C) StoreINode(ctl, mem, adr, adr_type, val);

   case T_CHAR:

   case T_SHORT:   return new (C) StoreCNode(ctl, mem, adr, adr_type, val);

   case T_LONG:    return new (C) StoreLNode(ctl, mem, adr, adr_type, val);

   case T_FLOAT:   return new (C) StoreFNode(ctl, mem, adr, adr_type, val);

   case T_DOUBLE:  return new (C) StoreDNode(ctl, mem, adr, adr_type, val);

   case T_METADATA:

   case T_ADDRESS:

   case T_OBJECT:

 #ifdef _LP64

     if (adr->bottom_type()->is_ptr_to_narrowoop()) {

       val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));

       return new (C) StoreNNode(ctl, mem, adr, adr_type, val);

     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||

                (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&

                 adr->bottom_type()->isa_rawptr())) {

       val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));

       return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val);

     }

 #endif

     {

       return new (C) StorePNode(ctl, mem, adr, adr_type, val);

     }

   }

   ShouldNotReachHere();

   return (StoreNode*)NULL;

 }

 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {

   bool require_atomic = true;

   return new (C) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);

 }

 //--------------------------bottom_type----------------------------------------

 const Type *StoreNode::bottom_type() const {

   return Type::MEMORY;

 }

 //------------------------------hash-------------------------------------------

 uint StoreNode::hash() const {

   // unroll addition of interesting fields

   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);

   // Since they are not commoned, do not hash them:

   return NO_HASH;

 }

 //------------------------------Ideal------------------------------------------

 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).

 // When a store immediately follows a relevant allocation/initialization,

 // try to capture it into the initialization, or hoist it above.

 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   Node* p = MemNode::Ideal_common(phase, can_reshape);

   if (p)  return (p == NodeSentinel) ? NULL : p;

   Node* mem     = in(MemNode::Memory);

   Node* address = in(MemNode::Address);

   // Back-to-back stores to same address?  Fold em up.  Generally

   // unsafe if I have intervening uses...  Also disallowed for StoreCM

   // since they must follow each StoreP operation.  Redundant StoreCMs

   // are eliminated just before matching in final_graph_reshape.

   if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&

       mem->Opcode() != Op_StoreCM) {

     // Looking at a dead closed cycle of memory?

     assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");

     assert(Opcode() == mem->Opcode() ||

            phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,

            "no mismatched stores, except on raw memory");

     if (mem->outcnt() == 1 &&           // check for intervening uses

         mem->as_Store()->memory_size() <= this->memory_size()) {

       // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.

       // For example, 'mem' might be the final state at a conditional return.

       // Or, 'mem' might be used by some node which is live at the same time

       // 'this' is live, which might be unschedulable.  So, require exactly

       // ONE user, the 'this' store, until such time as we clone 'mem' for

       // each of 'mem's uses (thus making the exactly-1-user-rule hold true).

       if (can_reshape) {  // (%%% is this an anachronism?)

         set_req_X(MemNode::Memory, mem->in(MemNode::Memory),

                   phase->is_IterGVN());

       } else {

         // It's OK to do this in the parser, since DU info is always accurate,

         // and the parser always refers to nodes via SafePointNode maps.

         set_req(MemNode::Memory, mem->in(MemNode::Memory));

       }

       return this;

     }

   }

   // Capture an unaliased, unconditional, simple store into an initializer.

   // Or, if it is independent of the allocation, hoist it above the allocation.

   if (ReduceFieldZeroing && /*can_reshape &&*/

       mem->is_Proj() && mem->in(0)->is_Initialize()) {

     InitializeNode* init = mem->in(0)->as_Initialize();

     intptr_t offset = init->can_capture_store(this, phase, can_reshape);

     if (offset > 0) {

       Node* moved = init->capture_store(this, offset, phase, can_reshape);

       // If the InitializeNode captured me, it made a raw copy of me,

       // and I need to disappear.

       if (moved != NULL) {

         // %%% hack to ensure that Ideal returns a new node:

         mem = MergeMemNode::make(phase->C, mem);

         return mem;             // fold me away

       }

     }

   }

   return NULL;                  // No further progress

 }

 //------------------------------Value-----------------------------------------

 const Type *StoreNode::Value( PhaseTransform *phase ) const {

   // Either input is TOP ==> the result is TOP

   const Type *t1 = phase->type( in(MemNode::Memory) );

   if( t1 == Type::TOP ) return Type::TOP;

   const Type *t2 = phase->type( in(MemNode::Address) );

   if( t2 == Type::TOP ) return Type::TOP;

   const Type *t3 = phase->type( in(MemNode::ValueIn) );

   if( t3 == Type::TOP ) return Type::TOP;

   return Type::MEMORY;

 }

 //------------------------------Identity---------------------------------------

 // Remove redundant stores:

 //   Store(m, p, Load(m, p)) changes to m.

 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).

 Node *StoreNode::Identity( PhaseTransform *phase ) {

   Node* mem = in(MemNode::Memory);

   Node* adr = in(MemNode::Address);

   Node* val = in(MemNode::ValueIn);

   // Load then Store?  Then the Store is useless

   if (val->is_Load() &&

       val->in(MemNode::Address)->eqv_uncast(adr) &&

       val->in(MemNode::Memory )->eqv_uncast(mem) &&

       val->as_Load()->store_Opcode() == Opcode()) {

     return mem;

   }

   // Two stores in a row of the same value?

   if (mem->is_Store() &&

       mem->in(MemNode::Address)->eqv_uncast(adr) &&

       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&

       mem->Opcode() == Opcode()) {

     return mem;

   }

   // Store of zero anywhere into a freshly-allocated object?

   // Then the store is useless.

   // (It must already have been captured by the InitializeNode.)

   if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {

     // a newly allocated object is already all-zeroes everywhere

     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {

       return mem;

     }

     // the store may also apply to zero-bits in an earlier object

     Node* prev_mem = find_previous_store(phase);

     // Steps (a), (b):  Walk past independent stores to find an exact match.

     if (prev_mem != NULL) {

       Node* prev_val = can_see_stored_value(prev_mem, phase);

       if (prev_val != NULL && phase->eqv(prev_val, val)) {

         // prev_val and val might differ by a cast; it would be good

         // to keep the more informative of the two.

         return mem;

       }

     }

   }

   return this;

 }

 //------------------------------match_edge-------------------------------------

 // Do we Match on this edge index or not?  Match only memory & value

 uint StoreNode::match_edge(uint idx) const {

   return idx == MemNode::Address || idx == MemNode::ValueIn;

 }

 //------------------------------cmp--------------------------------------------

 // Do not common stores up together.  They generally have to be split

 // back up anyways, so do not bother.

 uint StoreNode::cmp( const Node &n ) const {

   return (&n == this);          // Always fail except on self

 }

 //------------------------------Ideal_masked_input-----------------------------

 // Check for a useless mask before a partial-word store

 // (StoreB ... (AndI valIn conIa) )

 // If (conIa & mask == mask) this simplifies to

 // (StoreB ... (valIn) )

 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {

   Node *val = in(MemNode::ValueIn);

   if( val->Opcode() == Op_AndI ) {

     const TypeInt *t = phase->type( val->in(2) )->isa_int();

     if( t && t->is_con() && (t->get_con() & mask) == mask ) {

       set_req(MemNode::ValueIn, val->in(1));

       return this;

     }

   }

   return NULL;

 }

 //------------------------------Ideal_sign_extended_input----------------------

 // Check for useless sign-extension before a partial-word store

 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )

 // If (conIL == conIR && conIR <= num_bits)  this simplifies to

 // (StoreB ... (valIn) )

 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {

   Node *val = in(MemNode::ValueIn);

   if( val->Opcode() == Op_RShiftI ) {

     const TypeInt *t = phase->type( val->in(2) )->isa_int();

     if( t && t->is_con() && (t->get_con() <= num_bits) ) {

       Node *shl = val->in(1);

       if( shl->Opcode() == Op_LShiftI ) {

         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();

         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {

           set_req(MemNode::ValueIn, shl->in(1));

           return this;

         }

       }

     }

   }

   return NULL;

 }

 //------------------------------value_never_loaded-----------------------------------

 // Determine whether there are any possible loads of the value stored.

 // For simplicity, we actually check if there are any loads from the

 // address stored to, not just for loads of the value stored by this node.

 //

 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {

   Node *adr = in(Address);

   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();

   if (adr_oop == NULL)

     return false;

   if (!adr_oop->is_known_instance_field())

     return false; // if not a distinct instance, there may be aliases of the address

   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {

     Node *use = adr->fast_out(i);

     int opc = use->Opcode();

     if (use->is_Load() || use->is_LoadStore()) {

       return false;

     }

   }

   return true;

 }

 //=============================================================================

 //------------------------------Ideal------------------------------------------

 // If the store is from an AND mask that leaves the low bits untouched, then

 // we can skip the AND operation.  If the store is from a sign-extension

 // (a left shift, then right shift) we can skip both.

 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){

   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);

   if( progress != NULL ) return progress;

   progress = StoreNode::Ideal_sign_extended_input(phase, 24);

   if( progress != NULL ) return progress;

   // Finally check the default case

   return StoreNode::Ideal(phase, can_reshape);

 }

 //=============================================================================

 //------------------------------Ideal------------------------------------------

 // If the store is from an AND mask that leaves the low bits untouched, then

 // we can skip the AND operation

 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){

   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);

   if( progress != NULL ) return progress;

   progress = StoreNode::Ideal_sign_extended_input(phase, 16);

   if( progress != NULL ) return progress;

   // Finally check the default case

   return StoreNode::Ideal(phase, can_reshape);

 }

 //=============================================================================

 //------------------------------Identity---------------------------------------

 Node *StoreCMNode::Identity( PhaseTransform *phase ) {

   // No need to card mark when storing a null ptr