1 /*
   2  * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/systemDictionary.hpp"
  27 #include "compiler/compileLog.hpp"
  28 #include "gc/shared/barrierSet.hpp"
  29 #include "gc/shared/c2/barrierSetC2.hpp"
  30 #include "memory/allocation.inline.hpp"
  31 #include "memory/resourceArea.hpp"
  32 #include "oops/objArrayKlass.hpp"
  33 #include "opto/addnode.hpp"
  34 #include "opto/arraycopynode.hpp"
  35 #include "opto/cfgnode.hpp"
  36 #include "opto/compile.hpp"
  37 #include "opto/connode.hpp"
  38 #include "opto/convertnode.hpp"
  39 #include "opto/loopnode.hpp"
  40 #include "opto/machnode.hpp"
  41 #include "opto/matcher.hpp"
  42 #include "opto/memnode.hpp"
  43 #include "opto/mulnode.hpp"
  44 #include "opto/narrowptrnode.hpp"
  45 #include "opto/phaseX.hpp"
  46 #include "opto/regmask.hpp"
  47 #include "opto/rootnode.hpp"
  48 #include "opto/valuetypenode.hpp"
  49 #include "utilities/align.hpp"
  50 #include "utilities/copy.hpp"
  51 #include "utilities/macros.hpp"
  52 #include "utilities/vmError.hpp"
  53 #if INCLUDE_ZGC
  54 #include "gc/z/c2/zBarrierSetC2.hpp"
  55 #endif
  56 
  57 // Portions of code courtesy of Clifford Click
  58 
  59 // Optimization - Graph Style
  60 
  61 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
  62 
  63 //=============================================================================
  64 uint MemNode::size_of() const { return sizeof(*this); }
  65 
  66 const TypePtr *MemNode::adr_type() const {
  67   Node* adr = in(Address);
  68   if (adr == NULL)  return NULL; // node is dead
  69   const TypePtr* cross_check = NULL;
  70   DEBUG_ONLY(cross_check = _adr_type);
  71   return calculate_adr_type(adr->bottom_type(), cross_check);
  72 }
  73 
  74 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
  75   if (adr != NULL) {
  76     if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
  77       return true;
  78     }
  79   }
  80   return false;
  81 }
  82 
  83 #ifndef PRODUCT
  84 void MemNode::dump_spec(outputStream *st) const {
  85   if (in(Address) == NULL)  return; // node is dead
  86 #ifndef ASSERT
  87   // fake the missing field
  88   const TypePtr* _adr_type = NULL;
  89   if (in(Address) != NULL)
  90     _adr_type = in(Address)->bottom_type()->isa_ptr();
  91 #endif
  92   dump_adr_type(this, _adr_type, st);
  93 
  94   Compile* C = Compile::current();
  95   if (C->alias_type(_adr_type)->is_volatile()) {
  96     st->print(" Volatile!");
  97   }
  98   if (_unaligned_access) {
  99     st->print(" unaligned");
 100   }
 101   if (_mismatched_access) {
 102     st->print(" mismatched");
 103   }
 104   if (_unsafe_access) {
 105     st->print(" unsafe");
 106   }
 107 }
 108 
 109 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
 110   st->print(" @");
 111   if (adr_type == NULL) {
 112     st->print("NULL");
 113   } else {
 114     adr_type->dump_on(st);
 115     Compile* C = Compile::current();
 116     Compile::AliasType* atp = NULL;
 117     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
 118     if (atp == NULL)
 119       st->print(", idx=?\?;");
 120     else if (atp->index() == Compile::AliasIdxBot)
 121       st->print(", idx=Bot;");
 122     else if (atp->index() == Compile::AliasIdxTop)
 123       st->print(", idx=Top;");
 124     else if (atp->index() == Compile::AliasIdxRaw)
 125       st->print(", idx=Raw;");
 126     else {
 127       ciField* field = atp->field();
 128       if (field) {
 129         st->print(", name=");
 130         field->print_name_on(st);
 131       }
 132       st->print(", idx=%d;", atp->index());
 133     }
 134   }
 135 }
 136 
 137 extern void print_alias_types();
 138 
 139 #endif
 140 
 141 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
 142   assert((t_oop != NULL), "sanity");
 143   bool is_instance = t_oop->is_known_instance_field();
 144   bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
 145                              (load != NULL) && load->is_Load() &&
 146                              (phase->is_IterGVN() != NULL);
 147   if (!(is_instance || is_boxed_value_load))
 148     return mchain;  // don't try to optimize non-instance types
 149   uint instance_id = t_oop->instance_id();
 150   Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
 151   Node *prev = NULL;
 152   Node *result = mchain;
 153   while (prev != result) {
 154     prev = result;
 155     if (result == start_mem)
 156       break;  // hit one of our sentinels
 157     // skip over a call which does not affect this memory slice
 158     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
 159       Node *proj_in = result->in(0);
 160       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
 161         break;  // hit one of our sentinels
 162       } else if (proj_in->is_Call()) {
 163         // ArrayCopyNodes processed here as well
 164         CallNode *call = proj_in->as_Call();
 165         if (!call->may_modify(t_oop, phase)) { // returns false for instances
 166           result = call->in(TypeFunc::Memory);
 167         }
 168       } else if (proj_in->is_Initialize()) {
 169         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
 170         // Stop if this is the initialization for the object instance which
 171         // contains this memory slice, otherwise skip over it.
 172         if ((alloc == NULL) || (alloc->_idx == instance_id)) {
 173           break;
 174         }
 175         if (is_instance) {
 176           result = proj_in->in(TypeFunc::Memory);
 177         } else if (is_boxed_value_load) {
 178           Node* klass = alloc->in(AllocateNode::KlassNode);
 179           const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
 180           if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
 181             result = proj_in->in(TypeFunc::Memory); // not related allocation
 182           }
 183         }
 184       } else if (proj_in->is_MemBar()) {
 185         ArrayCopyNode* ac = NULL;
 186         if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
 187           break;
 188         }
 189         result = proj_in->in(TypeFunc::Memory);
 190       } else {
 191         assert(false, "unexpected projection");
 192       }
 193     } else if (result->is_ClearArray()) {
 194       if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
 195         // Can not bypass initialization of the instance
 196         // we are looking for.
 197         break;
 198       }
 199       // Otherwise skip it (the call updated 'result' value).
 200     } else if (result->is_MergeMem()) {
 201       result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
 202     }
 203   }
 204   return result;
 205 }
 206 
 207 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
 208   const TypeOopPtr* t_oop = t_adr->isa_oopptr();
 209   if (t_oop == NULL)
 210     return mchain;  // don't try to optimize non-oop types
 211   Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
 212   bool is_instance = t_oop->is_known_instance_field();
 213   PhaseIterGVN *igvn = phase->is_IterGVN();
 214   if (is_instance && igvn != NULL  && result->is_Phi()) {
 215     PhiNode *mphi = result->as_Phi();
 216     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 217     const TypePtr *t = mphi->adr_type();
 218     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
 219         (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
 220          t->is_oopptr()->cast_to_exactness(true)
 221            ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 222             ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
 223       // clone the Phi with our address type
 224       result = mphi->split_out_instance(t_adr, igvn);
 225     } else {
 226       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 227     }
 228   }
 229   return result;
 230 }
 231 
 232 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 233   uint alias_idx = phase->C->get_alias_index(tp);
 234   Node *mem = mmem;
 235 #ifdef ASSERT
 236   {
 237     // Check that current type is consistent with the alias index used during graph construction
 238     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 239     bool consistent =  adr_check == NULL || adr_check->empty() ||
 240                        phase->C->must_alias(adr_check, alias_idx );
 241     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 242     if( !consistent && adr_check != NULL && !adr_check->empty() &&
 243         tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 244         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 245         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 246           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 247           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 248       // don't assert if it is dead code.
 249       consistent = true;
 250     }
 251     if( !consistent ) {
 252       st->print("alias_idx==%d, adr_check==", alias_idx);
 253       if( adr_check == NULL ) {
 254         st->print("NULL");
 255       } else {
 256         adr_check->dump();
 257       }
 258       st->cr();
 259       print_alias_types();
 260       assert(consistent, "adr_check must match alias idx");
 261     }
 262   }
 263 #endif
 264   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
 265   // means an array I have not precisely typed yet.  Do not do any
 266   // alias stuff with it any time soon.
 267   const TypeOopPtr *toop = tp->isa_oopptr();
 268   if( tp->base() != Type::AnyPtr &&
 269       !(toop &&
 270         toop->klass() != NULL &&
 271         toop->klass()->is_java_lang_Object() &&
 272         toop->offset() == Type::OffsetBot) ) {
 273     // compress paths and change unreachable cycles to TOP
 274     // If not, we can update the input infinitely along a MergeMem cycle
 275     // Equivalent code in PhiNode::Ideal
 276     Node* m  = phase->transform(mmem);
 277     // If transformed to a MergeMem, get the desired slice
 278     // Otherwise the returned node represents memory for every slice
 279     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
 280     // Update input if it is progress over what we have now
 281   }
 282   return mem;
 283 }
 284 
 285 //--------------------------Ideal_common---------------------------------------
 286 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
 287 // Unhook non-raw memories from complete (macro-expanded) initializations.
 288 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
 289   // If our control input is a dead region, kill all below the region
 290   Node *ctl = in(MemNode::Control);
 291   if (ctl && remove_dead_region(phase, can_reshape))
 292     return this;
 293   ctl = in(MemNode::Control);
 294   // Don't bother trying to transform a dead node
 295   if (ctl && ctl->is_top())  return NodeSentinel;
 296 
 297   PhaseIterGVN *igvn = phase->is_IterGVN();
 298   // Wait if control on the worklist.
 299   if (ctl && can_reshape && igvn != NULL) {
 300     Node* bol = NULL;
 301     Node* cmp = NULL;
 302     if (ctl->in(0)->is_If()) {
 303       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
 304       bol = ctl->in(0)->in(1);
 305       if (bol->is_Bool())
 306         cmp = ctl->in(0)->in(1)->in(1);
 307     }
 308     if (igvn->_worklist.member(ctl) ||
 309         (bol != NULL && igvn->_worklist.member(bol)) ||
 310         (cmp != NULL && igvn->_worklist.member(cmp)) ) {
 311       // This control path may be dead.
 312       // Delay this memory node transformation until the control is processed.
 313       phase->is_IterGVN()->_worklist.push(this);
 314       return NodeSentinel; // caller will return NULL
 315     }
 316   }
 317   // Ignore if memory is dead, or self-loop
 318   Node *mem = in(MemNode::Memory);
 319   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
 320   assert(mem != this, "dead loop in MemNode::Ideal");
 321 
 322   if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
 323     // This memory slice may be dead.
 324     // Delay this mem node transformation until the memory is processed.
 325     phase->is_IterGVN()->_worklist.push(this);
 326     return NodeSentinel; // caller will return NULL
 327   }
 328 
 329   Node *address = in(MemNode::Address);
 330   const Type *t_adr = phase->type(address);
 331   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return NULL
 332 
 333   if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
 334     // Unsafe off-heap access with zero address. Remove access and other control users
 335     // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
 336     assert(ctl != NULL, "unsafe accesses should be control dependent");
 337     for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
 338       Node* u = ctl->fast_out(i);
 339       if (u != ctl) {
 340         igvn->rehash_node_delayed(u);
 341         int nb = u->replace_edge(ctl, phase->C->top());
 342         --i, imax -= nb;
 343       }
 344     }
 345     Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
 346     Node* halt = igvn->transform(new HaltNode(ctl, frame));
 347     phase->C->root()->add_req(halt);
 348     return this;
 349   }
 350 
 351   if (can_reshape && igvn != NULL &&
 352       (igvn->_worklist.member(address) ||
 353        (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
 354     // The address's base and type may change when the address is processed.
 355     // Delay this mem node transformation until the address is processed.
 356     phase->is_IterGVN()->_worklist.push(this);
 357     return NodeSentinel; // caller will return NULL
 358   }
 359 
 360   // Do NOT remove or optimize the next lines: ensure a new alias index
 361   // is allocated for an oop pointer type before Escape Analysis.
 362   // Note: C++ will not remove it since the call has side effect.
 363   if (t_adr->isa_oopptr()) {
 364     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 365   }
 366 
 367   Node* base = NULL;
 368   if (address->is_AddP()) {
 369     base = address->in(AddPNode::Base);
 370   }
 371   if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
 372       !t_adr->isa_rawptr()) {
 373     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
 374     // Skip this node optimization if its address has TOP base.
 375     return NodeSentinel; // caller will return NULL
 376   }
 377 
 378   // Avoid independent memory operations
 379   Node* old_mem = mem;
 380 
 381   // The code which unhooks non-raw memories from complete (macro-expanded)
 382   // initializations was removed. After macro-expansion all stores catched
 383   // by Initialize node became raw stores and there is no information
 384   // which memory slices they modify. So it is unsafe to move any memory
 385   // operation above these stores. Also in most cases hooked non-raw memories
 386   // were already unhooked by using information from detect_ptr_independence()
 387   // and find_previous_store().
 388 
 389   if (mem->is_MergeMem()) {
 390     MergeMemNode* mmem = mem->as_MergeMem();
 391     const TypePtr *tp = t_adr->is_ptr();
 392 
 393     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 394   }
 395 
 396   if (mem != old_mem) {
 397     set_req(MemNode::Memory, mem);
 398     if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
 399       igvn->_worklist.push(old_mem);
 400     }
 401     if (phase->type(mem) == Type::TOP) return NodeSentinel;
 402     return this;
 403   }
 404 
 405   // let the subclass continue analyzing...
 406   return NULL;
 407 }
 408 
 409 // Helper function for proving some simple control dominations.
 410 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 411 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 412 // is not a constant (dominated by the method's StartNode).
 413 // Used by MemNode::find_previous_store to prove that the
 414 // control input of a memory operation predates (dominates)
 415 // an allocation it wants to look past.
 416 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
 417   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
 418     return false; // Conservative answer for dead code
 419 
 420   // Check 'dom'. Skip Proj and CatchProj nodes.
 421   dom = dom->find_exact_control(dom);
 422   if (dom == NULL || dom->is_top())
 423     return false; // Conservative answer for dead code
 424 
 425   if (dom == sub) {
 426     // For the case when, for example, 'sub' is Initialize and the original
 427     // 'dom' is Proj node of the 'sub'.
 428     return false;
 429   }
 430 
 431   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
 432     return true;
 433 
 434   // 'dom' dominates 'sub' if its control edge and control edges
 435   // of all its inputs dominate or equal to sub's control edge.
 436 
 437   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 438   // Or Region for the check in LoadNode::Ideal();
 439   // 'sub' should have sub->in(0) != NULL.
 440   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 441          sub->is_Region() || sub->is_Call(), "expecting only these nodes");
 442 
 443   // Get control edge of 'sub'.
 444   Node* orig_sub = sub;
 445   sub = sub->find_exact_control(sub->in(0));
 446   if (sub == NULL || sub->is_top())
 447     return false; // Conservative answer for dead code
 448 
 449   assert(sub->is_CFG(), "expecting control");
 450 
 451   if (sub == dom)
 452     return true;
 453 
 454   if (sub->is_Start() || sub->is_Root())
 455     return false;
 456 
 457   {
 458     // Check all control edges of 'dom'.
 459 
 460     ResourceMark rm;
 461     Arena* arena = Thread::current()->resource_area();
 462     Node_List nlist(arena);
 463     Unique_Node_List dom_list(arena);
 464 
 465     dom_list.push(dom);
 466     bool only_dominating_controls = false;
 467 
 468     for (uint next = 0; next < dom_list.size(); next++) {
 469       Node* n = dom_list.at(next);
 470       if (n == orig_sub)
 471         return false; // One of dom's inputs dominated by sub.
 472       if (!n->is_CFG() && n->pinned()) {
 473         // Check only own control edge for pinned non-control nodes.
 474         n = n->find_exact_control(n->in(0));
 475         if (n == NULL || n->is_top())
 476           return false; // Conservative answer for dead code
 477         assert(n->is_CFG(), "expecting control");
 478         dom_list.push(n);
 479       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 480         only_dominating_controls = true;
 481       } else if (n->is_CFG()) {
 482         if (n->dominates(sub, nlist))
 483           only_dominating_controls = true;
 484         else
 485           return false;
 486       } else {
 487         // First, own control edge.
 488         Node* m = n->find_exact_control(n->in(0));
 489         if (m != NULL) {
 490           if (m->is_top())
 491             return false; // Conservative answer for dead code
 492           dom_list.push(m);
 493         }
 494         // Now, the rest of edges.
 495         uint cnt = n->req();
 496         for (uint i = 1; i < cnt; i++) {
 497           m = n->find_exact_control(n->in(i));
 498           if (m == NULL || m->is_top())
 499             continue;
 500           dom_list.push(m);
 501         }
 502       }
 503     }
 504     return only_dominating_controls;
 505   }
 506 }
 507 
 508 //---------------------detect_ptr_independence---------------------------------
 509 // Used by MemNode::find_previous_store to prove that two base
 510 // pointers are never equal.
 511 // The pointers are accompanied by their associated allocations,
 512 // if any, which have been previously discovered by the caller.
 513 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 514                                       Node* p2, AllocateNode* a2,
 515                                       PhaseTransform* phase) {
 516   // Attempt to prove that these two pointers cannot be aliased.
 517   // They may both manifestly be allocations, and they should differ.
 518   // Or, if they are not both allocations, they can be distinct constants.
 519   // Otherwise, one is an allocation and the other a pre-existing value.
 520   if (a1 == NULL && a2 == NULL) {           // neither an allocation
 521     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 522   } else if (a1 != NULL && a2 != NULL) {    // both allocations
 523     return (a1 != a2);
 524   } else if (a1 != NULL) {                  // one allocation a1
 525     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 526     return all_controls_dominate(p2, a1);
 527   } else { //(a2 != NULL)                   // one allocation a2
 528     return all_controls_dominate(p1, a2);
 529   }
 530   return false;
 531 }
 532 
 533 
 534 // Find an arraycopy that must have set (can_see_stored_value=true) or
 535 // could have set (can_see_stored_value=false) the value for this load
 536 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
 537   if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
 538                                                mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
 539     Node* mb = mem->in(0);
 540     if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
 541         mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
 542       ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
 543       if (ac->is_clonebasic()) {
 544         intptr_t offset;
 545         AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
 546         if (alloc != NULL && alloc == ld_alloc) {
 547           return ac;
 548         }
 549       }
 550     }
 551   } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
 552     ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
 553 
 554     if (ac->is_arraycopy_validated() ||
 555         ac->is_copyof_validated() ||
 556         ac->is_copyofrange_validated()) {
 557       Node* ld_addp = in(MemNode::Address);
 558       if (ld_addp->is_AddP()) {
 559         Node* ld_base = ld_addp->in(AddPNode::Address);
 560         Node* ld_offs = ld_addp->in(AddPNode::Offset);
 561 
 562         Node* dest = ac->in(ArrayCopyNode::Dest);
 563 
 564         if (dest == ld_base) {
 565           const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 566           if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
 567             return ac;
 568           }
 569           if (!can_see_stored_value) {
 570             mem = ac->in(TypeFunc::Memory);
 571           }
 572         }
 573       }
 574     }
 575   }
 576   return NULL;
 577 }
 578 
 579 // The logic for reordering loads and stores uses four steps:
 580 // (a) Walk carefully past stores and initializations which we
 581 //     can prove are independent of this load.
 582 // (b) Observe that the next memory state makes an exact match
 583 //     with self (load or store), and locate the relevant store.
 584 // (c) Ensure that, if we were to wire self directly to the store,
 585 //     the optimizer would fold it up somehow.
 586 // (d) Do the rewiring, and return, depending on some other part of
 587 //     the optimizer to fold up the load.
 588 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 589 // specific to loads and stores, so they are handled by the callers.
 590 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 591 //
 592 Node* MemNode::find_previous_store(PhaseTransform* phase) {
 593   Node*         ctrl   = in(MemNode::Control);
 594   Node*         adr    = in(MemNode::Address);
 595   intptr_t      offset = 0;
 596   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 597   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
 598 
 599   if (offset == Type::OffsetBot)
 600     return NULL;            // cannot unalias unless there are precise offsets
 601 
 602   const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
 603   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 604 
 605   intptr_t size_in_bytes = memory_size();
 606 
 607   Node* mem = in(MemNode::Memory);   // start searching here...
 608 
 609   int cnt = 50;             // Cycle limiter
 610   for (;;) {                // While we can dance past unrelated stores...
 611     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 612 
 613     Node* prev = mem;
 614     if (mem->is_Store()) {
 615       Node* st_adr = mem->in(MemNode::Address);
 616       intptr_t st_offset = 0;
 617       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 618       if (st_base == NULL)
 619         break;              // inscrutable pointer
 620 
 621       // For raw accesses it's not enough to prove that constant offsets don't intersect.
 622       // We need the bases to be the equal in order for the offset check to make sense.
 623       if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
 624         break;
 625       }
 626 
 627       if (st_offset != offset && st_offset != Type::OffsetBot) {
 628         const int MAX_STORE = BytesPerLong;
 629         if (st_offset >= offset + size_in_bytes ||
 630             st_offset <= offset - MAX_STORE ||
 631             st_offset <= offset - mem->as_Store()->memory_size()) {
 632           // Success:  The offsets are provably independent.
 633           // (You may ask, why not just test st_offset != offset and be done?
 634           // The answer is that stores of different sizes can co-exist
 635           // in the same sequence of RawMem effects.  We sometimes initialize
 636           // a whole 'tile' of array elements with a single jint or jlong.)
 637           mem = mem->in(MemNode::Memory);
 638           continue;           // (a) advance through independent store memory
 639         }
 640       }
 641       if (st_base != base &&
 642           detect_ptr_independence(base, alloc,
 643                                   st_base,
 644                                   AllocateNode::Ideal_allocation(st_base, phase),
 645                                   phase)) {
 646         // Success:  The bases are provably independent.
 647         mem = mem->in(MemNode::Memory);
 648         continue;           // (a) advance through independent store memory
 649       }
 650 
 651       // (b) At this point, if the bases or offsets do not agree, we lose,
 652       // since we have not managed to prove 'this' and 'mem' independent.
 653       if (st_base == base && st_offset == offset) {
 654         return mem;         // let caller handle steps (c), (d)
 655       }
 656 
 657     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 658       InitializeNode* st_init = mem->in(0)->as_Initialize();
 659       AllocateNode*  st_alloc = st_init->allocation();
 660       if (st_alloc == NULL)
 661         break;              // something degenerated
 662       bool known_identical = false;
 663       bool known_independent = false;
 664       if (alloc == st_alloc)
 665         known_identical = true;
 666       else if (alloc != NULL)
 667         known_independent = true;
 668       else if (all_controls_dominate(this, st_alloc))
 669         known_independent = true;
 670 
 671       if (known_independent) {
 672         // The bases are provably independent: Either they are
 673         // manifestly distinct allocations, or else the control
 674         // of this load dominates the store's allocation.
 675         int alias_idx = phase->C->get_alias_index(adr_type());
 676         if (alias_idx == Compile::AliasIdxRaw) {
 677           mem = st_alloc->in(TypeFunc::Memory);
 678         } else {
 679           mem = st_init->memory(alias_idx);
 680         }
 681         continue;           // (a) advance through independent store memory
 682       }
 683 
 684       // (b) at this point, if we are not looking at a store initializing
 685       // the same allocation we are loading from, we lose.
 686       if (known_identical) {
 687         // From caller, can_see_stored_value will consult find_captured_store.
 688         return mem;         // let caller handle steps (c), (d)
 689       }
 690 
 691     } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
 692       if (prev != mem) {
 693         // Found an arraycopy but it doesn't affect that load
 694         continue;
 695       }
 696       // Found an arraycopy that may affect that load
 697       return mem;
 698     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
 699       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 700       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 701         // ArrayCopyNodes processed here as well.
 702         CallNode *call = mem->in(0)->as_Call();
 703         if (!call->may_modify(addr_t, phase)) {
 704           mem = call->in(TypeFunc::Memory);
 705           continue;         // (a) advance through independent call memory
 706         }
 707       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 708         ArrayCopyNode* ac = NULL;
 709         if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
 710           break;
 711         }
 712         mem = mem->in(0)->in(TypeFunc::Memory);
 713         continue;           // (a) advance through independent MemBar memory
 714       } else if (mem->is_ClearArray()) {
 715         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 716           // (the call updated 'mem' value)
 717           continue;         // (a) advance through independent allocation memory
 718         } else {
 719           // Can not bypass initialization of the instance
 720           // we are looking for.
 721           return mem;
 722         }
 723       } else if (mem->is_MergeMem()) {
 724         int alias_idx = phase->C->get_alias_index(adr_type());
 725         mem = mem->as_MergeMem()->memory_at(alias_idx);
 726         continue;           // (a) advance through independent MergeMem memory
 727       }
 728     }
 729 
 730     // Unless there is an explicit 'continue', we must bail out here,
 731     // because 'mem' is an inscrutable memory state (e.g., a call).
 732     break;
 733   }
 734 
 735   return NULL;              // bail out
 736 }
 737 
 738 //----------------------calculate_adr_type-------------------------------------
 739 // Helper function.  Notices when the given type of address hits top or bottom.
 740 // Also, asserts a cross-check of the type against the expected address type.
 741 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 742   if (t == Type::TOP)  return NULL; // does not touch memory any more?
 743   #ifdef PRODUCT
 744   cross_check = NULL;
 745   #else
 746   if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump())  cross_check = NULL;
 747   #endif
 748   const TypePtr* tp = t->isa_ptr();
 749   if (tp == NULL) {
 750     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 751     return TypePtr::BOTTOM;           // touches lots of memory
 752   } else {
 753     #ifdef ASSERT
 754     // %%%% [phh] We don't check the alias index if cross_check is
 755     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 756     if (cross_check != NULL &&
 757         cross_check != TypePtr::BOTTOM &&
 758         cross_check != TypeRawPtr::BOTTOM) {
 759       // Recheck the alias index, to see if it has changed (due to a bug).
 760       Compile* C = Compile::current();
 761       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 762              "must stay in the original alias category");
 763       // The type of the address must be contained in the adr_type,
 764       // disregarding "null"-ness.
 765       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 766       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 767       assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
 768              "real address must not escape from expected memory type");
 769     }
 770     #endif
 771     return tp;
 772   }
 773 }
 774 
 775 //=============================================================================
 776 // Should LoadNode::Ideal() attempt to remove control edges?
 777 bool LoadNode::can_remove_control() const {
 778   return true;
 779 }
 780 uint LoadNode::size_of() const { return sizeof(*this); }
 781 bool LoadNode::cmp( const Node &n ) const
 782 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
 783 const Type *LoadNode::bottom_type() const { return _type; }
 784 uint LoadNode::ideal_reg() const {
 785   return _type->ideal_reg();
 786 }
 787 
 788 #ifndef PRODUCT
 789 void LoadNode::dump_spec(outputStream *st) const {
 790   MemNode::dump_spec(st);
 791   if( !Verbose && !WizardMode ) {
 792     // standard dump does this in Verbose and WizardMode
 793     st->print(" #"); _type->dump_on(st);
 794   }
 795   if (!depends_only_on_test()) {
 796     st->print(" (does not depend only on test)");
 797   }
 798 }
 799 #endif
 800 
 801 #ifdef ASSERT
 802 //----------------------------is_immutable_value-------------------------------
 803 // Helper function to allow a raw load without control edge for some cases
 804 bool LoadNode::is_immutable_value(Node* adr) {
 805   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 806           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 807           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 808            in_bytes(JavaThread::osthread_offset())));
 809 }
 810 #endif
 811 
 812 //----------------------------LoadNode::make-----------------------------------
 813 // Polymorphic factory method:
 814 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
 815                      ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
 816   Compile* C = gvn.C;
 817 
 818   // sanity check the alias category against the created node type
 819   assert(!(adr_type->isa_oopptr() &&
 820            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
 821          "use LoadKlassNode instead");
 822   assert(!(adr_type->isa_aryptr() &&
 823            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
 824          "use LoadRangeNode instead");
 825   // Check control edge of raw loads
 826   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
 827           // oop will be recorded in oop map if load crosses safepoint
 828           rt->isa_oopptr() || is_immutable_value(adr),
 829           "raw memory operations should have control edge");
 830   LoadNode* load = NULL;
 831   switch (bt) {
 832   case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 833   case T_BYTE:    load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 834   case T_INT:     load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 835   case T_CHAR:    load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 836   case T_SHORT:   load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 837   case T_LONG:    load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
 838   case T_FLOAT:   load = new LoadFNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
 839   case T_DOUBLE:  load = new LoadDNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
 840   case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(),  mo, control_dependency); break;
 841   case T_VALUETYPE:
 842   case T_OBJECT:
 843 #ifdef _LP64
 844     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 845       load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
 846     } else
 847 #endif
 848     {
 849       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
 850       load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
 851     }
 852     break;
 853   default:
 854     ShouldNotReachHere();
 855     break;
 856   }
 857   assert(load != NULL, "LoadNode should have been created");
 858   if (unaligned) {
 859     load->set_unaligned_access();
 860   }
 861   if (mismatched) {
 862     load->set_mismatched_access();
 863   }
 864   if (unsafe) {
 865     load->set_unsafe_access();
 866   }
 867   if (load->Opcode() == Op_LoadN) {
 868     Node* ld = gvn.transform(load);
 869     return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
 870   }
 871 
 872   return load;
 873 }
 874 
 875 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
 876                                   ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
 877   bool require_atomic = true;
 878   LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
 879   if (unaligned) {
 880     load->set_unaligned_access();
 881   }
 882   if (mismatched) {
 883     load->set_mismatched_access();
 884   }
 885   if (unsafe) {
 886     load->set_unsafe_access();
 887   }
 888   return load;
 889 }
 890 
 891 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
 892                                   ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
 893   bool require_atomic = true;
 894   LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
 895   if (unaligned) {
 896     load->set_unaligned_access();
 897   }
 898   if (mismatched) {
 899     load->set_mismatched_access();
 900   }
 901   if (unsafe) {
 902     load->set_unsafe_access();
 903   }
 904   return load;
 905 }
 906 
 907 
 908 
 909 //------------------------------hash-------------------------------------------
 910 uint LoadNode::hash() const {
 911   // unroll addition of interesting fields
 912   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
 913 }
 914 
 915 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
 916   if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
 917     bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
 918     bool is_stable_ary = FoldStableValues &&
 919                          (tp != NULL) && (tp->isa_aryptr() != NULL) &&
 920                          tp->isa_aryptr()->is_stable();
 921 
 922     return (eliminate_boxing && non_volatile) || is_stable_ary;
 923   }
 924 
 925   return false;
 926 }
 927 
 928 // Is the value loaded previously stored by an arraycopy? If so return
 929 // a load node that reads from the source array so we may be able to
 930 // optimize out the ArrayCopy node later.
 931 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
 932   Node* ld_adr = in(MemNode::Address);
 933   intptr_t ld_off = 0;
 934   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 935   Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
 936   if (ac != NULL) {
 937     assert(ac->is_ArrayCopy(), "what kind of node can this be?");
 938 
 939     Node* mem = ac->in(TypeFunc::Memory);
 940     Node* ctl = ac->in(0);
 941     Node* src = ac->in(ArrayCopyNode::Src);
 942 
 943     if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
 944       return NULL;
 945     }
 946 
 947     LoadNode* ld = clone()->as_Load();
 948     Node* addp = in(MemNode::Address)->clone();
 949     if (ac->as_ArrayCopy()->is_clonebasic()) {
 950       assert(ld_alloc != NULL, "need an alloc");
 951       assert(addp->is_AddP(), "address must be addp");
 952       assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address");
 953       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 954       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern");
 955       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern");
 956       addp->set_req(AddPNode::Base, src->in(AddPNode::Base));
 957       addp->set_req(AddPNode::Address, src->in(AddPNode::Address));
 958     } else {
 959       assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
 960              ac->as_ArrayCopy()->is_copyof_validated() ||
 961              ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
 962       assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
 963       addp->set_req(AddPNode::Base, src);
 964       addp->set_req(AddPNode::Address, src);
 965 
 966       const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
 967       BasicType ary_elem  = ary_t->klass()->as_array_klass()->element_type()->basic_type();
 968       uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
 969       uint shift  = exact_log2(type2aelembytes(ary_elem));
 970 
 971       Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 972 #ifdef _LP64
 973       diff = phase->transform(new ConvI2LNode(diff));
 974 #endif
 975       diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
 976 
 977       Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
 978       addp->set_req(AddPNode::Offset, offset);
 979     }
 980     addp = phase->transform(addp);
 981 #ifdef ASSERT
 982     const TypePtr* adr_type = phase->type(addp)->is_ptr();
 983     ld->_adr_type = adr_type;
 984 #endif
 985     ld->set_req(MemNode::Address, addp);
 986     ld->set_req(0, ctl);
 987     ld->set_req(MemNode::Memory, mem);
 988     // load depends on the tests that validate the arraycopy
 989     ld->_control_dependency = UnknownControl;
 990     return ld;
 991   }
 992   return NULL;
 993 }
 994 
 995 
 996 //---------------------------can_see_stored_value------------------------------
 997 // This routine exists to make sure this set of tests is done the same
 998 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
 999 // will change the graph shape in a way which makes memory alive twice at the
1000 // same time (uses the Oracle model of aliasing), then some
1001 // LoadXNode::Identity will fold things back to the equivalence-class model
1002 // of aliasing.
1003 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1004   Node* ld_adr = in(MemNode::Address);
1005   intptr_t ld_off = 0;
1006   Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1007   Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1008   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1009   Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1010   // This is more general than load from boxing objects.
1011   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1012     uint alias_idx = atp->index();
1013     bool final = !atp->is_rewritable();
1014     Node* result = NULL;
1015     Node* current = st;
1016     // Skip through chains of MemBarNodes checking the MergeMems for
1017     // new states for the slice of this load.  Stop once any other
1018     // kind of node is encountered.  Loads from final memory can skip
1019     // through any kind of MemBar but normal loads shouldn't skip
1020     // through MemBarAcquire since the could allow them to move out of
1021     // a synchronized region.
1022     while (current->is_Proj()) {
1023       int opc = current->in(0)->Opcode();
1024       if ((final && (opc == Op_MemBarAcquire ||
1025                      opc == Op_MemBarAcquireLock ||
1026                      opc == Op_LoadFence)) ||
1027           opc == Op_MemBarRelease ||
1028           opc == Op_StoreFence ||
1029           opc == Op_MemBarReleaseLock ||
1030           opc == Op_MemBarStoreStore ||
1031           opc == Op_MemBarCPUOrder) {
1032         Node* mem = current->in(0)->in(TypeFunc::Memory);
1033         if (mem->is_MergeMem()) {
1034           MergeMemNode* merge = mem->as_MergeMem();
1035           Node* new_st = merge->memory_at(alias_idx);
1036           if (new_st == merge->base_memory()) {
1037             // Keep searching
1038             current = new_st;
1039             continue;
1040           }
1041           // Save the new memory state for the slice and fall through
1042           // to exit.
1043           result = new_st;
1044         }
1045       }
1046       break;
1047     }
1048     if (result != NULL) {
1049       st = result;
1050     }
1051   }
1052 
1053   // Loop around twice in the case Load -> Initialize -> Store.
1054   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1055   for (int trip = 0; trip <= 1; trip++) {
1056 
1057     if (st->is_Store()) {
1058       Node* st_adr = st->in(MemNode::Address);
1059       if (!phase->eqv(st_adr, ld_adr)) {
1060         // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1061         intptr_t st_off = 0;
1062         Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1063         if (ld_base == NULL)                                   return NULL;
1064         if (st_base == NULL)                                   return NULL;
1065         if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1066         if (ld_off != st_off)                                  return NULL;
1067         if (ld_off == Type::OffsetBot)                         return NULL;
1068         // Same base, same offset.
1069         // Possible improvement for arrays: check index value instead of absolute offset.
1070 
1071         // At this point we have proven something like this setup:
1072         //   B = << base >>
1073         //   L =  LoadQ(AddP(Check/CastPP(B), #Off))
1074         //   S = StoreQ(AddP(             B , #Off), V)
1075         // (Actually, we haven't yet proven the Q's are the same.)
1076         // In other words, we are loading from a casted version of
1077         // the same pointer-and-offset that we stored to.
1078         // Casted version may carry a dependency and it is respected.
1079         // Thus, we are able to replace L by V.
1080       }
1081       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1082       if (store_Opcode() != st->Opcode())
1083         return NULL;
1084       return st->in(MemNode::ValueIn);
1085     }
1086 
1087     // A load from a freshly-created object always returns zero.
1088     // (This can happen after LoadNode::Ideal resets the load's memory input
1089     // to find_captured_store, which returned InitializeNode::zero_memory.)
1090     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1091         (st->in(0) == ld_alloc) &&
1092         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1093       // return a zero value for the load's basic type
1094       // (This is one of the few places where a generic PhaseTransform
1095       // can create new nodes.  Think of it as lazily manifesting
1096       // virtually pre-existing constants.)
1097       assert(memory_type() != T_VALUETYPE, "should not be used for value types");
1098       Node* default_value = ld_alloc->in(AllocateNode::DefaultValue);
1099       if (default_value != NULL) {
1100         return default_value;
1101       }
1102       assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
1103       return phase->zerocon(memory_type());
1104     }
1105 
1106     // A load from an initialization barrier can match a captured store.
1107     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1108       InitializeNode* init = st->in(0)->as_Initialize();
1109       AllocateNode* alloc = init->allocation();
1110       if ((alloc != NULL) && (alloc == ld_alloc)) {
1111         // examine a captured store value
1112         st = init->find_captured_store(ld_off, memory_size(), phase);
1113         if (st != NULL) {
1114           continue;             // take one more trip around
1115         }
1116       }
1117     }
1118 
1119     // Load boxed value from result of valueOf() call is input parameter.
1120     if (this->is_Load() && ld_adr->is_AddP() &&
1121         (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1122       intptr_t ignore = 0;
1123       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1124       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1125       base = bs->step_over_gc_barrier(base);
1126       if (base != NULL && base->is_Proj() &&
1127           base->as_Proj()->_con == TypeFunc::Parms &&
1128           base->in(0)->is_CallStaticJava() &&
1129           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1130         return base->in(0)->in(TypeFunc::Parms);
1131       }
1132     }
1133 
1134     break;
1135   }
1136 
1137   return NULL;
1138 }
1139 
1140 //----------------------is_instance_field_load_with_local_phi------------------
1141 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1142   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1143       in(Address)->is_AddP() ) {
1144     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1145     // Only instances and boxed values.
1146     if( t_oop != NULL &&
1147         (t_oop->is_ptr_to_boxed_value() ||
1148          t_oop->is_known_instance_field()) &&
1149         t_oop->offset() != Type::OffsetBot &&
1150         t_oop->offset() != Type::OffsetTop) {
1151       return true;
1152     }
1153   }
1154   return false;
1155 }
1156 
1157 //------------------------------Identity---------------------------------------
1158 // Loads are identity if previous store is to same address
1159 Node* LoadNode::Identity(PhaseGVN* phase) {
1160   // Loading from a ValueTypePtr? The ValueTypePtr has the values of
1161   // all fields as input. Look for the field with matching offset.
1162   Node* addr = in(Address);
1163   intptr_t offset;
1164   Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
1165   if (base != NULL && base->is_ValueTypePtr() && offset > oopDesc::klass_offset_in_bytes()) {
1166     Node* value = base->as_ValueTypePtr()->field_value_by_offset((int)offset, true);
1167     if (value->is_ValueType()) {
1168       // Non-flattened value type field
1169       ValueTypeNode* vt = value->as_ValueType();
1170       if (vt->is_allocated(phase)) {
1171         value = vt->get_oop();
1172       } else {
1173         // Not yet allocated, bail out
1174         value = NULL;
1175       }
1176     }
1177     if (value != NULL) {
1178       if (Opcode() == Op_LoadN) {
1179         // Encode oop value if we are loading a narrow oop
1180         assert(!phase->type(value)->isa_narrowoop(), "should already be decoded");
1181         value = phase->transform(new EncodePNode(value, bottom_type()));
1182       }
1183       return value;
1184     }
1185   }
1186 
1187   // If the previous store-maker is the right kind of Store, and the store is
1188   // to the same address, then we are equal to the value stored.
1189   Node* mem = in(Memory);
1190   Node* value = can_see_stored_value(mem, phase);
1191   if( value ) {
1192     // byte, short & char stores truncate naturally.
1193     // A load has to load the truncated value which requires
1194     // some sort of masking operation and that requires an
1195     // Ideal call instead of an Identity call.
1196     if (memory_size() < BytesPerInt) {
1197       // If the input to the store does not fit with the load's result type,
1198       // it must be truncated via an Ideal call.
1199       if (!phase->type(value)->higher_equal(phase->type(this)))
1200         return this;
1201     }
1202     // (This works even when value is a Con, but LoadNode::Value
1203     // usually runs first, producing the singleton type of the Con.)
1204     return value;
1205   }
1206 
1207   // Search for an existing data phi which was generated before for the same
1208   // instance's field to avoid infinite generation of phis in a loop.
1209   Node *region = mem->in(0);
1210   if (is_instance_field_load_with_local_phi(region)) {
1211     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1212     int this_index  = phase->C->get_alias_index(addr_t);
1213     int this_offset = addr_t->offset();
1214     int this_iid    = addr_t->instance_id();
1215     if (!addr_t->is_known_instance() &&
1216          addr_t->is_ptr_to_boxed_value()) {
1217       // Use _idx of address base (could be Phi node) for boxed values.
1218       intptr_t   ignore = 0;
1219       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1220       if (base == NULL) {
1221         return this;
1222       }
1223       this_iid = base->_idx;
1224     }
1225     const Type* this_type = bottom_type();
1226     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1227       Node* phi = region->fast_out(i);
1228       if (phi->is_Phi() && phi != mem &&
1229           phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1230         return phi;
1231       }
1232     }
1233   }
1234 
1235   return this;
1236 }
1237 
1238 // Construct an equivalent unsigned load.
1239 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1240   BasicType bt = T_ILLEGAL;
1241   const Type* rt = NULL;
1242   switch (Opcode()) {
1243     case Op_LoadUB: return this;
1244     case Op_LoadUS: return this;
1245     case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1246     case Op_LoadS: bt = T_CHAR;    rt = TypeInt::CHAR;  break;
1247     default:
1248       assert(false, "no unsigned variant: %s", Name());
1249       return NULL;
1250   }
1251   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1252                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1253                         is_unaligned_access(), is_mismatched_access());
1254 }
1255 
1256 // Construct an equivalent signed load.
1257 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1258   BasicType bt = T_ILLEGAL;
1259   const Type* rt = NULL;
1260   switch (Opcode()) {
1261     case Op_LoadUB: bt = T_BYTE;  rt = TypeInt::BYTE;  break;
1262     case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1263     case Op_LoadB: // fall through
1264     case Op_LoadS: // fall through
1265     case Op_LoadI: // fall through
1266     case Op_LoadL: return this;
1267     default:
1268       assert(false, "no signed variant: %s", Name());
1269       return NULL;
1270   }
1271   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1272                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1273                         is_unaligned_access(), is_mismatched_access());
1274 }
1275 
1276 // We're loading from an object which has autobox behaviour.
1277 // If this object is result of a valueOf call we'll have a phi
1278 // merging a newly allocated object and a load from the cache.
1279 // We want to replace this load with the original incoming
1280 // argument to the valueOf call.
1281 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1282   assert(phase->C->eliminate_boxing(), "sanity");
1283   intptr_t ignore = 0;
1284   Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1285   if ((base == NULL) || base->is_Phi()) {
1286     // Push the loads from the phi that comes from valueOf up
1287     // through it to allow elimination of the loads and the recovery
1288     // of the original value. It is done in split_through_phi().
1289     return NULL;
1290   } else if (base->is_Load() ||
1291              (base->is_DecodeN() && base->in(1)->is_Load())) {
1292     // Eliminate the load of boxed value for integer types from the cache
1293     // array by deriving the value from the index into the array.
1294     // Capture the offset of the load and then reverse the computation.
1295 
1296     // Get LoadN node which loads a boxing object from 'cache' array.
1297     if (base->is_DecodeN()) {
1298       base = base->in(1);
1299     }
1300     if (!base->in(Address)->is_AddP()) {
1301       return NULL; // Complex address
1302     }
1303     AddPNode* address = base->in(Address)->as_AddP();
1304     Node* cache_base = address->in(AddPNode::Base);
1305     if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1306       // Get ConP node which is static 'cache' field.
1307       cache_base = cache_base->in(1);
1308     }
1309     if ((cache_base != NULL) && cache_base->is_Con()) {
1310       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1311       if ((base_type != NULL) && base_type->is_autobox_cache()) {
1312         Node* elements[4];
1313         int shift = exact_log2(type2aelembytes(T_OBJECT));
1314         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1315         if (count > 0 && elements[0]->is_Con() &&
1316             (count == 1 ||
1317              (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1318                             elements[1]->in(2) == phase->intcon(shift)))) {
1319           ciObjArray* array = base_type->const_oop()->as_obj_array();
1320           // Fetch the box object cache[0] at the base of the array and get its value
1321           ciInstance* box = array->obj_at(0)->as_instance();
1322           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1323           assert(ik->is_box_klass(), "sanity");
1324           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1325           if (ik->nof_nonstatic_fields() == 1) {
1326             // This should be true nonstatic_field_at requires calling
1327             // nof_nonstatic_fields so check it anyway
1328             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1329             BasicType bt = c.basic_type();
1330             // Only integer types have boxing cache.
1331             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1332                    bt == T_BYTE    || bt == T_SHORT ||
1333                    bt == T_INT     || bt == T_LONG, "wrong type = %s", type2name(bt));
1334             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1335             if (cache_low != (int)cache_low) {
1336               return NULL; // should not happen since cache is array indexed by value
1337             }
1338             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1339             if (offset != (int)offset) {
1340               return NULL; // should not happen since cache is array indexed by value
1341             }
1342            // Add up all the offsets making of the address of the load
1343             Node* result = elements[0];
1344             for (int i = 1; i < count; i++) {
1345               result = phase->transform(new AddXNode(result, elements[i]));
1346             }
1347             // Remove the constant offset from the address and then
1348             result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1349             // remove the scaling of the offset to recover the original index.
1350             if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1351               // Peel the shift off directly but wrap it in a dummy node
1352               // since Ideal can't return existing nodes
1353               result = new RShiftXNode(result->in(1), phase->intcon(0));
1354             } else if (result->is_Add() && result->in(2)->is_Con() &&
1355                        result->in(1)->Opcode() == Op_LShiftX &&
1356                        result->in(1)->in(2) == phase->intcon(shift)) {
1357               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1358               // but for boxing cache access we know that X<<Z will not overflow
1359               // (there is range check) so we do this optimizatrion by hand here.
1360               Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1361               result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1362             } else {
1363               result = new RShiftXNode(result, phase->intcon(shift));
1364             }
1365 #ifdef _LP64
1366             if (bt != T_LONG) {
1367               result = new ConvL2INode(phase->transform(result));
1368             }
1369 #else
1370             if (bt == T_LONG) {
1371               result = new ConvI2LNode(phase->transform(result));
1372             }
1373 #endif
1374             // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1375             // Need to preserve unboxing load type if it is unsigned.
1376             switch(this->Opcode()) {
1377               case Op_LoadUB:
1378                 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1379                 break;
1380               case Op_LoadUS:
1381                 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1382                 break;
1383             }
1384             return result;
1385           }
1386         }
1387       }
1388     }
1389   }
1390   return NULL;
1391 }
1392 
1393 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1394   Node* region = phi->in(0);
1395   if (region == NULL) {
1396     return false; // Wait stable graph
1397   }
1398   uint cnt = phi->req();
1399   for (uint i = 1; i < cnt; i++) {
1400     Node* rc = region->in(i);
1401     if (rc == NULL || phase->type(rc) == Type::TOP)
1402       return false; // Wait stable graph
1403     Node* in = phi->in(i);
1404     if (in == NULL || phase->type(in) == Type::TOP)
1405       return false; // Wait stable graph
1406   }
1407   return true;
1408 }
1409 //------------------------------split_through_phi------------------------------
1410 // Split instance or boxed field load through Phi.
1411 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1412   Node* mem     = in(Memory);
1413   Node* address = in(Address);
1414   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1415 
1416   assert((t_oop != NULL) &&
1417          (t_oop->is_known_instance_field() ||
1418           t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1419 
1420   Compile* C = phase->C;
1421   intptr_t ignore = 0;
1422   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1423   bool base_is_phi = (base != NULL) && base->is_Phi();
1424   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1425                            (base != NULL) && (base == address->in(AddPNode::Base)) &&
1426                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1427 
1428   if (!((mem->is_Phi() || base_is_phi) &&
1429         (load_boxed_values || t_oop->is_known_instance_field()))) {
1430     return NULL; // memory is not Phi
1431   }
1432 
1433   if (mem->is_Phi()) {
1434     if (!stable_phi(mem->as_Phi(), phase)) {
1435       return NULL; // Wait stable graph
1436     }
1437     uint cnt = mem->req();
1438     // Check for loop invariant memory.
1439     if (cnt == 3) {
1440       for (uint i = 1; i < cnt; i++) {
1441         Node* in = mem->in(i);
1442         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1443         if (m == mem) {
1444           if (i == 1) {
1445             // if the first edge was a loop, check second edge too.
1446             // If both are replaceable - we are in an infinite loop
1447             Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1448             if (n == mem) {
1449               break;
1450             }
1451           }
1452           set_req(Memory, mem->in(cnt - i));
1453           return this; // made change
1454         }
1455       }
1456     }
1457   }
1458   if (base_is_phi) {
1459     if (!stable_phi(base->as_Phi(), phase)) {
1460       return NULL; // Wait stable graph
1461     }
1462     uint cnt = base->req();
1463     // Check for loop invariant memory.
1464     if (cnt == 3) {
1465       for (uint i = 1; i < cnt; i++) {
1466         if (base->in(i) == base) {
1467           return NULL; // Wait stable graph
1468         }
1469       }
1470     }
1471   }
1472 
1473   // Split through Phi (see original code in loopopts.cpp).
1474   assert(C->have_alias_type(t_oop), "instance should have alias type");
1475 
1476   // Do nothing here if Identity will find a value
1477   // (to avoid infinite chain of value phis generation).
1478   if (!phase->eqv(this, phase->apply_identity(this)))
1479     return NULL;
1480 
1481   // Select Region to split through.
1482   Node* region;
1483   if (!base_is_phi) {
1484     assert(mem->is_Phi(), "sanity");
1485     region = mem->in(0);
1486     // Skip if the region dominates some control edge of the address.
1487     if (!MemNode::all_controls_dominate(address, region))
1488       return NULL;
1489   } else if (!mem->is_Phi()) {
1490     assert(base_is_phi, "sanity");
1491     region = base->in(0);
1492     // Skip if the region dominates some control edge of the memory.
1493     if (!MemNode::all_controls_dominate(mem, region))
1494       return NULL;
1495   } else if (base->in(0) != mem->in(0)) {
1496     assert(base_is_phi && mem->is_Phi(), "sanity");
1497     if (MemNode::all_controls_dominate(mem, base->in(0))) {
1498       region = base->in(0);
1499     } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1500       region = mem->in(0);
1501     } else {
1502       return NULL; // complex graph
1503     }
1504   } else {
1505     assert(base->in(0) == mem->in(0), "sanity");
1506     region = mem->in(0);
1507   }
1508 
1509   const Type* this_type = this->bottom_type();
1510   int this_index  = C->get_alias_index(t_oop);
1511   int this_offset = t_oop->offset();
1512   int this_iid    = t_oop->instance_id();
1513   if (!t_oop->is_known_instance() && load_boxed_values) {
1514     // Use _idx of address base for boxed values.
1515     this_iid = base->_idx;
1516   }
1517   PhaseIterGVN* igvn = phase->is_IterGVN();
1518   Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1519   for (uint i = 1; i < region->req(); i++) {
1520     Node* x;
1521     Node* the_clone = NULL;
1522     if (region->in(i) == C->top()) {
1523       x = C->top();      // Dead path?  Use a dead data op
1524     } else {
1525       x = this->clone();        // Else clone up the data op
1526       the_clone = x;            // Remember for possible deletion.
1527       // Alter data node to use pre-phi inputs
1528       if (this->in(0) == region) {
1529         x->set_req(0, region->in(i));
1530       } else {
1531         x->set_req(0, NULL);
1532       }
1533       if (mem->is_Phi() && (mem->in(0) == region)) {
1534         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1535       }
1536       if (address->is_Phi() && address->in(0) == region) {
1537         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1538       }
1539       if (base_is_phi && (base->in(0) == region)) {
1540         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1541         Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1542         x->set_req(Address, adr_x);
1543       }
1544     }
1545     // Check for a 'win' on some paths
1546     const Type *t = x->Value(igvn);
1547 
1548     bool singleton = t->singleton();
1549 
1550     // See comments in PhaseIdealLoop::split_thru_phi().
1551     if (singleton && t == Type::TOP) {
1552       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1553     }
1554 
1555     if (singleton) {
1556       x = igvn->makecon(t);
1557     } else {
1558       // We now call Identity to try to simplify the cloned node.
1559       // Note that some Identity methods call phase->type(this).
1560       // Make sure that the type array is big enough for
1561       // our new node, even though we may throw the node away.
1562       // (This tweaking with igvn only works because x is a new node.)
1563       igvn->set_type(x, t);
1564       // If x is a TypeNode, capture any more-precise type permanently into Node
1565       // otherwise it will be not updated during igvn->transform since
1566       // igvn->type(x) is set to x->Value() already.
1567       x->raise_bottom_type(t);
1568       Node *y = igvn->apply_identity(x);
1569       if (y != x) {
1570         x = y;
1571       } else {
1572         y = igvn->hash_find_insert(x);
1573         if (y) {
1574           x = y;
1575         } else {
1576           // Else x is a new node we are keeping
1577           // We do not need register_new_node_with_optimizer
1578           // because set_type has already been called.
1579           igvn->_worklist.push(x);
1580         }
1581       }
1582     }
1583     if (x != the_clone && the_clone != NULL) {
1584       igvn->remove_dead_node(the_clone);
1585     }
1586     phi->set_req(i, x);
1587   }
1588   // Record Phi
1589   igvn->register_new_node_with_optimizer(phi);
1590   return phi;
1591 }
1592 
1593 //------------------------------Ideal------------------------------------------
1594 // If the load is from Field memory and the pointer is non-null, it might be possible to
1595 // zero out the control input.
1596 // If the offset is constant and the base is an object allocation,
1597 // try to hook me up to the exact initializing store.
1598 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1599   Node* p = MemNode::Ideal_common(phase, can_reshape);
1600   if (p)  return (p == NodeSentinel) ? NULL : p;
1601 
1602   Node* ctrl    = in(MemNode::Control);
1603   Node* address = in(MemNode::Address);
1604   bool progress = false;
1605 
1606   bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1607          phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1608 
1609   // Skip up past a SafePoint control.  Cannot do this for Stores because
1610   // pointer stores & cardmarks must stay on the same side of a SafePoint.
1611   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1612       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw  &&
1613       !addr_mark &&
1614       (depends_only_on_test() || has_unknown_control_dependency())) {
1615     ctrl = ctrl->in(0);
1616     set_req(MemNode::Control,ctrl);
1617     progress = true;
1618   }
1619 
1620   intptr_t ignore = 0;
1621   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1622   if (base != NULL
1623       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1624     // Check for useless control edge in some common special cases
1625     if (in(MemNode::Control) != NULL
1626         && can_remove_control()
1627         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1628         && all_controls_dominate(base, phase->C->start())) {
1629       // A method-invariant, non-null address (constant or 'this' argument).
1630       set_req(MemNode::Control, NULL);
1631       progress = true;
1632     }
1633   }
1634 
1635   Node* mem = in(MemNode::Memory);
1636   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1637 
1638   if (can_reshape && (addr_t != NULL)) {
1639     // try to optimize our memory input
1640     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1641     if (opt_mem != mem) {
1642       set_req(MemNode::Memory, opt_mem);
1643       if (phase->type( opt_mem ) == Type::TOP) return NULL;
1644       return this;
1645     }
1646     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1647     if ((t_oop != NULL) &&
1648         (t_oop->is_known_instance_field() ||
1649          t_oop->is_ptr_to_boxed_value())) {
1650       PhaseIterGVN *igvn = phase->is_IterGVN();
1651       if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1652         // Delay this transformation until memory Phi is processed.
1653         phase->is_IterGVN()->_worklist.push(this);
1654         return NULL;
1655       }
1656       // Split instance field load through Phi.
1657       Node* result = split_through_phi(phase);
1658       if (result != NULL) return result;
1659 
1660       if (t_oop->is_ptr_to_boxed_value()) {
1661         Node* result = eliminate_autobox(phase);
1662         if (result != NULL) return result;
1663       }
1664     }
1665   }
1666 
1667   // Is there a dominating load that loads the same value?  Leave
1668   // anything that is not a load of a field/array element (like
1669   // barriers etc.) alone
1670   if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1671     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1672       Node *use = mem->fast_out(i);
1673       if (use != this &&
1674           use->Opcode() == Opcode() &&
1675           use->in(0) != NULL &&
1676           use->in(0) != in(0) &&
1677           use->in(Address) == in(Address)) {
1678         Node* ctl = in(0);
1679         for (int i = 0; i < 10 && ctl != NULL; i++) {
1680           ctl = IfNode::up_one_dom(ctl);
1681           if (ctl == use->in(0)) {
1682             set_req(0, use->in(0));
1683             return this;
1684           }
1685         }
1686       }
1687     }
1688   }
1689 
1690   // Check for prior store with a different base or offset; make Load
1691   // independent.  Skip through any number of them.  Bail out if the stores
1692   // are in an endless dead cycle and report no progress.  This is a key
1693   // transform for Reflection.  However, if after skipping through the Stores
1694   // we can't then fold up against a prior store do NOT do the transform as
1695   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
1696   // array memory alive twice: once for the hoisted Load and again after the
1697   // bypassed Store.  This situation only works if EVERYBODY who does
1698   // anti-dependence work knows how to bypass.  I.e. we need all
1699   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
1700   // the alias index stuff.  So instead, peek through Stores and IFF we can
1701   // fold up, do so.
1702   Node* prev_mem = find_previous_store(phase);
1703   if (prev_mem != NULL) {
1704     Node* value = can_see_arraycopy_value(prev_mem, phase);
1705     if (value != NULL) {
1706       return value;
1707     }
1708   }
1709   // Steps (a), (b):  Walk past independent stores to find an exact match.
1710   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1711     // (c) See if we can fold up on the spot, but don't fold up here.
1712     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1713     // just return a prior value, which is done by Identity calls.
1714     if (can_see_stored_value(prev_mem, phase)) {
1715       // Make ready for step (d):
1716       set_req(MemNode::Memory, prev_mem);
1717       return this;
1718     }
1719   }
1720 
1721   AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1722   if (alloc != NULL && mem->is_Proj() &&
1723       mem->in(0) != NULL &&
1724       mem->in(0) == alloc->initialization() &&
1725       Opcode() == Op_LoadX &&
1726       alloc->initialization()->proj_out_or_null(0) != NULL) {
1727     InitializeNode* init = alloc->initialization();
1728     Node* control = init->proj_out(0);
1729     return alloc->make_ideal_mark(phase, address, control, mem, NULL);
1730   }
1731 
1732   return progress ? this : NULL;
1733 }
1734 
1735 // Helper to recognize certain Klass fields which are invariant across
1736 // some group of array types (e.g., int[] or all T[] where T < Object).
1737 const Type*
1738 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1739                                  ciKlass* klass) const {
1740   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1741     // The field is Klass::_modifier_flags.  Return its (constant) value.
1742     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1743     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1744     return TypeInt::make(klass->modifier_flags());
1745   }
1746   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1747     // The field is Klass::_access_flags.  Return its (constant) value.
1748     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1749     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1750     return TypeInt::make(klass->access_flags());
1751   }
1752   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1753     // The field is Klass::_layout_helper.  Return its constant value if known.
1754     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1755     return TypeInt::make(klass->layout_helper());
1756   }
1757 
1758   // No match.
1759   return NULL;
1760 }
1761 
1762 //------------------------------Value-----------------------------------------
1763 const Type* LoadNode::Value(PhaseGVN* phase) const {
1764   // Either input is TOP ==> the result is TOP
1765   Node* mem = in(MemNode::Memory);
1766   const Type *t1 = phase->type(mem);
1767   if (t1 == Type::TOP)  return Type::TOP;
1768   Node* adr = in(MemNode::Address);
1769   const TypePtr* tp = phase->type(adr)->isa_ptr();
1770   if (tp == NULL || tp->empty())  return Type::TOP;
1771   int off = tp->offset();
1772   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1773   Compile* C = phase->C;
1774 
1775   // Try to guess loaded type from pointer type
1776   if (tp->isa_aryptr()) {
1777     const TypeAryPtr* ary = tp->is_aryptr();
1778     const Type* t = ary->elem();
1779 
1780     // Determine whether the reference is beyond the header or not, by comparing
1781     // the offset against the offset of the start of the array's data.
1782     // Different array types begin at slightly different offsets (12 vs. 16).
1783     // We choose T_BYTE as an example base type that is least restrictive
1784     // as to alignment, which will therefore produce the smallest
1785     // possible base offset.
1786     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1787     const bool off_beyond_header = (off >= min_base_off);
1788 
1789     // Try to constant-fold a stable array element.
1790     if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1791       // Make sure the reference is not into the header and the offset is constant
1792       ciObject* aobj = ary->const_oop();
1793       if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1794         int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1795         const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1796                                                                       stable_dimension,
1797                                                                       memory_type(), is_unsigned());
1798         if (con_type != NULL) {
1799           return con_type;
1800         }
1801       }
1802     }
1803 
1804     // Don't do this for integer types. There is only potential profit if
1805     // the element type t is lower than _type; that is, for int types, if _type is
1806     // more restrictive than t.  This only happens here if one is short and the other
1807     // char (both 16 bits), and in those cases we've made an intentional decision
1808     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1809     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1810     //
1811     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1812     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
1813     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1814     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
1815     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1816     // In fact, that could have been the original type of p1, and p1 could have
1817     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1818     // expression (LShiftL quux 3) independently optimized to the constant 8.
1819     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1820         && (_type->isa_vect() == NULL)
1821         && t->isa_valuetype() == NULL
1822         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1823       // t might actually be lower than _type, if _type is a unique
1824       // concrete subclass of abstract class t.
1825       if (off_beyond_header || off == Type::OffsetBot) {  // is the offset beyond the header?
1826         const Type* jt = t->join_speculative(_type);
1827         // In any case, do not allow the join, per se, to empty out the type.
1828         if (jt->empty() && !t->empty()) {
1829           // This can happen if a interface-typed array narrows to a class type.
1830           jt = _type;
1831         }
1832 #ifdef ASSERT
1833         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1834           // The pointers in the autobox arrays are always non-null
1835           Node* base = adr->in(AddPNode::Base);
1836           if ((base != NULL) && base->is_DecodeN()) {
1837             // Get LoadN node which loads IntegerCache.cache field
1838             base = base->in(1);
1839           }
1840           if ((base != NULL) && base->is_Con()) {
1841             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1842             if ((base_type != NULL) && base_type->is_autobox_cache()) {
1843               // It could be narrow oop
1844               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1845             }
1846           }
1847         }
1848 #endif
1849         return jt;
1850       }
1851     }
1852   } else if (tp->base() == Type::InstPtr) {
1853     assert( off != Type::OffsetBot ||
1854             // arrays can be cast to Objects
1855             tp->is_oopptr()->klass()->is_java_lang_Object() ||
1856             tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() ||
1857             // unsafe field access may not have a constant offset
1858             C->has_unsafe_access(),
1859             "Field accesses must be precise" );
1860     // For oop loads, we expect the _type to be precise.
1861 
1862     const TypeInstPtr* tinst = tp->is_instptr();
1863     BasicType bt = memory_type();
1864 
1865     // Fold component and value mirror loads
1866     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1867     if (ik == phase->C->env()->Class_klass() && (off == java_lang_Class::component_mirror_offset_in_bytes() ||
1868                                                  off == java_lang_Class::inline_mirror_offset_in_bytes())) {
1869       ciType* mirror_type = tinst->java_mirror_type();
1870       if (mirror_type != NULL) {
1871         const Type* const_oop = TypePtr::NULL_PTR;
1872         if (mirror_type->is_array_klass()) {
1873           const_oop = TypeInstPtr::make(mirror_type->as_array_klass()->component_mirror_instance());
1874         } else if (mirror_type->is_valuetype()) {
1875           const_oop = TypeInstPtr::make(mirror_type->as_value_klass()->inline_mirror_instance());
1876         }
1877         return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
1878       }
1879     }
1880 
1881     // Optimize loads from constant fields.
1882     ciObject* const_oop = tinst->const_oop();
1883     if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1884       ciType* mirror_type = const_oop->as_instance()->java_mirror_type();
1885       if (mirror_type != NULL && mirror_type->is_valuetype()) {
1886         ciValueKlass* vk = mirror_type->as_value_klass();
1887         if (off == vk->default_value_offset()) {
1888           // Loading a special hidden field that contains the oop of the default value type
1889           const Type* const_oop = TypeInstPtr::make(vk->default_value_instance());
1890           return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
1891         }
1892       }
1893       const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
1894       if (con_type != NULL) {
1895         return con_type;
1896       }
1897     }
1898   } else if (tp->base() == Type::KlassPtr) {
1899     assert( off != Type::OffsetBot ||
1900             // arrays can be cast to Objects
1901             tp->is_klassptr()->klass() == NULL ||
1902             tp->is_klassptr()->klass()->is_java_lang_Object() ||
1903             // also allow array-loading from the primary supertype
1904             // array during subtype checks
1905             Opcode() == Op_LoadKlass,
1906             "Field accesses must be precise" );
1907     // For klass/static loads, we expect the _type to be precise
1908   } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) {
1909     if (adr->is_Load() && off == 0) {
1910       /* With mirrors being an indirect in the Klass*
1911        * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1912        * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1913        *
1914        * So check the type and klass of the node before the LoadP.
1915        */
1916       Node* adr2 = adr->in(MemNode::Address);
1917       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1918       if (tkls != NULL) {
1919         ciKlass* klass = tkls->klass();
1920         if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1921           assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1922           assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1923           return TypeInstPtr::make(klass->java_mirror());
1924         }
1925       }
1926     } else {
1927       // Check for a load of the default value offset from the ValueKlassFixedBlock:
1928       // LoadI(LoadP(value_klass, adr_valueklass_fixed_block_offset), default_value_offset_offset)
1929       intptr_t offset = 0;
1930       Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1931       if (base != NULL && base->is_Load() && offset == in_bytes(ValueKlass::default_value_offset_offset())) {
1932         const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr();
1933         if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_valuetype() &&
1934             tkls->offset() == in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())) {
1935           assert(base->Opcode() == Op_LoadP, "must load an oop from klass");
1936           assert(Opcode() == Op_LoadI, "must load an int from fixed block");
1937           return TypeInt::make(tkls->klass()->as_value_klass()->default_value_offset());
1938         }
1939       }
1940     }
1941   }
1942 
1943   const TypeKlassPtr *tkls = tp->isa_klassptr();
1944   if (tkls != NULL && !StressReflectiveCode) {
1945     ciKlass* klass = tkls->klass();
1946     if (tkls->is_loaded() && tkls->klass_is_exact()) {
1947       // We are loading a field from a Klass metaobject whose identity
1948       // is known at compile time (the type is "exact" or "precise").
1949       // Check for fields we know are maintained as constants by the VM.
1950       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1951         // The field is Klass::_super_check_offset.  Return its (constant) value.
1952         // (Folds up type checking code.)
1953         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1954         return TypeInt::make(klass->super_check_offset());
1955       }
1956       // Compute index into primary_supers array
1957       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1958       // Check for overflowing; use unsigned compare to handle the negative case.
1959       if( depth < ciKlass::primary_super_limit() ) {
1960         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1961         // (Folds up type checking code.)
1962         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1963         ciKlass *ss = klass->super_of_depth(depth);
1964         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1965       }
1966       const Type* aift = load_array_final_field(tkls, klass);
1967       if (aift != NULL)  return aift;
1968     }
1969 
1970     // We can still check if we are loading from the primary_supers array at a
1971     // shallow enough depth.  Even though the klass is not exact, entries less
1972     // than or equal to its super depth are correct.
1973     if (tkls->is_loaded()) {
1974       ciType *inner = klass;
1975       while( inner->is_obj_array_klass() )
1976         inner = inner->as_obj_array_klass()->base_element_type();
1977       if( inner->is_instance_klass() &&
1978           !inner->as_instance_klass()->flags().is_interface() ) {
1979         // Compute index into primary_supers array
1980         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1981         // Check for overflowing; use unsigned compare to handle the negative case.
1982         if( depth < ciKlass::primary_super_limit() &&
1983             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1984           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1985           // (Folds up type checking code.)
1986           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1987           ciKlass *ss = klass->super_of_depth(depth);
1988           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1989         }
1990       }
1991     }
1992 
1993     // If the type is enough to determine that the thing is not an array,
1994     // we can give the layout_helper a positive interval type.
1995     // This will help short-circuit some reflective code.
1996     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1997         && !klass->is_array_klass() // not directly typed as an array
1998         && !klass->is_interface()  // specifically not Serializable & Cloneable
1999         && !klass->is_java_lang_Object()   // not the supertype of all T[]
2000         ) {
2001       // Note:  When interfaces are reliable, we can narrow the interface
2002       // test to (klass != Serializable && klass != Cloneable).
2003       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2004       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2005       // The key property of this type is that it folds up tests
2006       // for array-ness, since it proves that the layout_helper is positive.
2007       // Thus, a generic value like the basic object layout helper works fine.
2008       return TypeInt::make(min_size, max_jint, Type::WidenMin);
2009     }
2010   }
2011 
2012   // If we are loading from a freshly-allocated object, produce a zero,
2013   // if the load is provably beyond the header of the object.
2014   // (Also allow a variable load from a fresh array to produce zero.)
2015   const TypeOopPtr *tinst = tp->isa_oopptr();
2016   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
2017   bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
2018   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2019     Node* value = can_see_stored_value(mem,phase);
2020     if (value != NULL && value->is_Con()) {
2021       assert(value->bottom_type()->higher_equal(_type),"sanity");
2022       return value->bottom_type();
2023     }
2024   }
2025 
2026   if (is_instance) {
2027     // If we have an instance type and our memory input is the
2028     // programs's initial memory state, there is no matching store,
2029     // so just return a zero of the appropriate type
2030     Node *mem = in(MemNode::Memory);
2031     if (mem->is_Parm() && mem->in(0)->is_Start()) {
2032       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2033       return Type::get_zero_type(_type->basic_type());
2034     }
2035   }
2036   return _type;
2037 }
2038 
2039 //------------------------------match_edge-------------------------------------
2040 // Do we Match on this edge index or not?  Match only the address.
2041 uint LoadNode::match_edge(uint idx) const {
2042   return idx == MemNode::Address;
2043 }
2044 
2045 //--------------------------LoadBNode::Ideal--------------------------------------
2046 //
2047 //  If the previous store is to the same address as this load,
2048 //  and the value stored was larger than a byte, replace this load
2049 //  with the value stored truncated to a byte.  If no truncation is
2050 //  needed, the replacement is done in LoadNode::Identity().
2051 //
2052 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2053   Node* mem = in(MemNode::Memory);
2054   Node* value = can_see_stored_value(mem,phase);
2055   if( value && !phase->type(value)->higher_equal( _type ) ) {
2056     Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
2057     return new RShiftINode(result, phase->intcon(24));
2058   }
2059   // Identity call will handle the case where truncation is not needed.
2060   return LoadNode::Ideal(phase, can_reshape);
2061 }
2062 
2063 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2064   Node* mem = in(MemNode::Memory);
2065   Node* value = can_see_stored_value(mem,phase);
2066   if (value != NULL && value->is_Con() &&
2067       !value->bottom_type()->higher_equal(_type)) {
2068     // If the input to the store does not fit with the load's result type,
2069     // it must be truncated. We can't delay until Ideal call since
2070     // a singleton Value is needed for split_thru_phi optimization.
2071     int con = value->get_int();
2072     return TypeInt::make((con << 24) >> 24);
2073   }
2074   return LoadNode::Value(phase);
2075 }
2076 
2077 //--------------------------LoadUBNode::Ideal-------------------------------------
2078 //
2079 //  If the previous store is to the same address as this load,
2080 //  and the value stored was larger than a byte, replace this load
2081 //  with the value stored truncated to a byte.  If no truncation is
2082 //  needed, the replacement is done in LoadNode::Identity().
2083 //
2084 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2085   Node* mem = in(MemNode::Memory);
2086   Node* value = can_see_stored_value(mem, phase);
2087   if (value && !phase->type(value)->higher_equal(_type))
2088     return new AndINode(value, phase->intcon(0xFF));
2089   // Identity call will handle the case where truncation is not needed.
2090   return LoadNode::Ideal(phase, can_reshape);
2091 }
2092 
2093 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2094   Node* mem = in(MemNode::Memory);
2095   Node* value = can_see_stored_value(mem,phase);
2096   if (value != NULL && value->is_Con() &&
2097       !value->bottom_type()->higher_equal(_type)) {
2098     // If the input to the store does not fit with the load's result type,
2099     // it must be truncated. We can't delay until Ideal call since
2100     // a singleton Value is needed for split_thru_phi optimization.
2101     int con = value->get_int();
2102     return TypeInt::make(con & 0xFF);
2103   }
2104   return LoadNode::Value(phase);
2105 }
2106 
2107 //--------------------------LoadUSNode::Ideal-------------------------------------
2108 //
2109 //  If the previous store is to the same address as this load,
2110 //  and the value stored was larger than a char, replace this load
2111 //  with the value stored truncated to a char.  If no truncation is
2112 //  needed, the replacement is done in LoadNode::Identity().
2113 //
2114 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2115   Node* mem = in(MemNode::Memory);
2116   Node* value = can_see_stored_value(mem,phase);
2117   if( value && !phase->type(value)->higher_equal( _type ) )
2118     return new AndINode(value,phase->intcon(0xFFFF));
2119   // Identity call will handle the case where truncation is not needed.
2120   return LoadNode::Ideal(phase, can_reshape);
2121 }
2122 
2123 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2124   Node* mem = in(MemNode::Memory);
2125   Node* value = can_see_stored_value(mem,phase);
2126   if (value != NULL && value->is_Con() &&
2127       !value->bottom_type()->higher_equal(_type)) {
2128     // If the input to the store does not fit with the load's result type,
2129     // it must be truncated. We can't delay until Ideal call since
2130     // a singleton Value is needed for split_thru_phi optimization.
2131     int con = value->get_int();
2132     return TypeInt::make(con & 0xFFFF);
2133   }
2134   return LoadNode::Value(phase);
2135 }
2136 
2137 //--------------------------LoadSNode::Ideal--------------------------------------
2138 //
2139 //  If the previous store is to the same address as this load,
2140 //  and the value stored was larger than a short, replace this load
2141 //  with the value stored truncated to a short.  If no truncation is
2142 //  needed, the replacement is done in LoadNode::Identity().
2143 //
2144 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2145   Node* mem = in(MemNode::Memory);
2146   Node* value = can_see_stored_value(mem,phase);
2147   if( value && !phase->type(value)->higher_equal( _type ) ) {
2148     Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2149     return new RShiftINode(result, phase->intcon(16));
2150   }
2151   // Identity call will handle the case where truncation is not needed.
2152   return LoadNode::Ideal(phase, can_reshape);
2153 }
2154 
2155 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2156   Node* mem = in(MemNode::Memory);
2157   Node* value = can_see_stored_value(mem,phase);
2158   if (value != NULL && value->is_Con() &&
2159       !value->bottom_type()->higher_equal(_type)) {
2160     // If the input to the store does not fit with the load's result type,
2161     // it must be truncated. We can't delay until Ideal call since
2162     // a singleton Value is needed for split_thru_phi optimization.
2163     int con = value->get_int();
2164     return TypeInt::make((con << 16) >> 16);
2165   }
2166   return LoadNode::Value(phase);
2167 }
2168 
2169 //=============================================================================
2170 //----------------------------LoadKlassNode::make------------------------------
2171 // Polymorphic factory method:
2172 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2173   // sanity check the alias category against the created node type
2174   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2175   assert(adr_type != NULL, "expecting TypeKlassPtr");
2176 #ifdef _LP64
2177   if (adr_type->is_ptr_to_narrowklass()) {
2178     assert(UseCompressedClassPointers, "no compressed klasses");
2179     Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2180     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2181   }
2182 #endif
2183   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2184   return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2185 }
2186 
2187 //------------------------------Value------------------------------------------
2188 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2189   return klass_value_common(phase);
2190 }
2191 
2192 // In most cases, LoadKlassNode does not have the control input set. If the control
2193 // input is set, it must not be removed (by LoadNode::Ideal()).
2194 bool LoadKlassNode::can_remove_control() const {
2195   return false;
2196 }
2197 
2198 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2199   // Either input is TOP ==> the result is TOP
2200   const Type *t1 = phase->type( in(MemNode::Memory) );
2201   if (t1 == Type::TOP)  return Type::TOP;
2202   Node *adr = in(MemNode::Address);
2203   const Type *t2 = phase->type( adr );
2204   if (t2 == Type::TOP)  return Type::TOP;
2205   const TypePtr *tp = t2->is_ptr();
2206   if (TypePtr::above_centerline(tp->ptr()) ||
2207       tp->ptr() == TypePtr::Null)  return Type::TOP;
2208 
2209   // Return a more precise klass, if possible
2210   const TypeInstPtr *tinst = tp->isa_instptr();
2211   if (tinst != NULL) {
2212     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2213     int offset = tinst->offset();
2214     if (ik == phase->C->env()->Class_klass()
2215         && (offset == java_lang_Class::klass_offset_in_bytes() ||
2216             offset == java_lang_Class::array_klass_offset_in_bytes())) {
2217       // We are loading a special hidden field from a Class mirror object,
2218       // the field which points to the VM's Klass metaobject.
2219       bool is_indirect_type = true;
2220       ciType* t = tinst->java_mirror_type(&is_indirect_type);
2221       // java_mirror_type returns non-null for compile-time Class constants.
2222       if (t != NULL) {
2223         // constant oop => constant klass
2224         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2225           if (t->is_void()) {
2226             // We cannot create a void array.  Since void is a primitive type return null
2227             // klass.  Users of this result need to do a null check on the returned klass.
2228             return TypePtr::NULL_PTR;
2229           }
2230           return TypeKlassPtr::make(ciArrayKlass::make(t, /* never_null */ !is_indirect_type));
2231         }
2232         if (!t->is_klass()) {
2233           // a primitive Class (e.g., int.class) has NULL for a klass field
2234           return TypePtr::NULL_PTR;
2235         }
2236         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2237         return TypeKlassPtr::make(t->as_klass());
2238       }
2239       // non-constant mirror, so we can't tell what's going on
2240     }
2241     if( !ik->is_loaded() )
2242       return _type;             // Bail out if not loaded
2243     if (offset == oopDesc::klass_offset_in_bytes()) {
2244       if (tinst->klass_is_exact()) {
2245         return TypeKlassPtr::make(ik);
2246       }
2247       // See if we can become precise: no subklasses and no interface
2248       // (Note:  We need to support verified interfaces.)
2249       if (!ik->is_interface() && !ik->has_subklass()) {
2250         //assert(!UseExactTypes, "this code should be useless with exact types");
2251         // Add a dependence; if any subclass added we need to recompile
2252         if (!ik->is_final()) {
2253           // %%% should use stronger assert_unique_concrete_subtype instead
2254           phase->C->dependencies()->assert_leaf_type(ik);
2255         }
2256         // Return precise klass
2257         return TypeKlassPtr::make(ik);
2258       }
2259 
2260       // Return root of possible klass
2261       return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0));
2262     }
2263   }
2264 
2265   // Check for loading klass from an array
2266   const TypeAryPtr *tary = tp->isa_aryptr();
2267   if (tary != NULL) {
2268     ciKlass *tary_klass = tary->klass();
2269     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
2270         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2271       ciArrayKlass* ak = tary_klass->as_array_klass();
2272       // Do not fold klass loads from [V?. The runtime type might be [V due to [V <: [V?
2273       // and the klass for [V is not equal to the klass for [V?.
2274       if (tary->klass_is_exact()) {
2275         return TypeKlassPtr::make(tary_klass);
2276       }
2277 
2278       // If the klass is an object array, we defer the question to the
2279       // array component klass.
2280       if (ak->is_obj_array_klass()) {
2281         assert(ak->is_loaded(), "");
2282         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2283         if (base_k->is_loaded() && base_k->is_instance_klass()) {
2284           ciInstanceKlass *ik = base_k->as_instance_klass();
2285           // See if we can become precise: no subklasses and no interface
2286           if (!ik->is_interface() && !ik->has_subklass() && (!ik->is_valuetype() || ak->storage_properties().is_null_free())) {
2287             //assert(!UseExactTypes, "this code should be useless with exact types");
2288             // Add a dependence; if any subclass added we need to recompile
2289             if (!ik->is_final()) {
2290               phase->C->dependencies()->assert_leaf_type(ik);
2291             }
2292             // Return precise array klass
2293             return TypeKlassPtr::make(ak);
2294           }
2295         }
2296         return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0));
2297       } else if (ak->is_type_array_klass()) {
2298         //assert(!UseExactTypes, "this code should be useless with exact types");
2299         return TypeKlassPtr::make(ak); // These are always precise
2300       }
2301     }
2302   }
2303 
2304   // Check for loading klass from an array klass
2305   const TypeKlassPtr *tkls = tp->isa_klassptr();
2306   if (tkls != NULL && !StressReflectiveCode) {
2307     if (!tkls->is_loaded()) {
2308       return _type;             // Bail out if not loaded
2309     }
2310     ciKlass* klass = tkls->klass();
2311     if( klass->is_obj_array_klass() &&
2312         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2313       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2314       // // Always returning precise element type is incorrect,
2315       // // e.g., element type could be object and array may contain strings
2316       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2317 
2318       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2319       // according to the element type's subclassing.
2320       return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0));
2321     } else if (klass->is_value_array_klass() &&
2322                tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2323       ciKlass* elem = klass->as_value_array_klass()->element_klass();
2324       return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0));
2325     }
2326     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2327         tkls->offset() == in_bytes(Klass::super_offset())) {
2328       ciKlass* sup = klass->as_instance_klass()->super();
2329       // The field is Klass::_super.  Return its (constant) value.
2330       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2331       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2332     }
2333   }
2334 
2335   // Bailout case
2336   return LoadNode::Value(phase);
2337 }
2338 
2339 //------------------------------Identity---------------------------------------
2340 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2341 // Also feed through the klass in Allocate(...klass...)._klass.
2342 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2343   return klass_identity_common(phase);
2344 }
2345 
2346 const Type* GetNullFreePropertyNode::Value(PhaseGVN* phase) const {
2347   if (in(1) != NULL) {
2348     const Type* in1_t = phase->type(in(1));
2349     if (in1_t == Type::TOP) {
2350       return Type::TOP;
2351     }
2352     const TypeKlassPtr* tk = in1_t->make_ptr()->is_klassptr();
2353     ciArrayKlass* ak = tk->klass()->as_array_klass();
2354     ciKlass* elem = ak->element_klass();
2355     if (tk->klass_is_exact() || (!elem->is_java_lang_Object() && !elem->is_interface() && !elem->is_valuetype())) {
2356       int props_shift = in1_t->isa_narrowklass() ? oopDesc::narrow_storage_props_shift : oopDesc::wide_storage_props_shift;
2357       ArrayStorageProperties props = ak->storage_properties();
2358       intptr_t storage_properties = props.encode<intptr_t>(props_shift);
2359       if (in1_t->isa_narrowklass()) {
2360         return TypeInt::make((int)storage_properties);
2361       }
2362       return TypeX::make(storage_properties);
2363     }
2364   }
2365   return bottom_type();
2366 }
2367 
2368 Node* GetNullFreePropertyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2369   if (!can_reshape) {
2370     return NULL;
2371   }
2372   if (in(1) != NULL && in(1)->is_Phi()) {
2373     Node* phi = in(1);
2374     Node* r = phi->in(0);
2375     Node* new_phi = new PhiNode(r, bottom_type());
2376     for (uint i = 1; i < r->req(); i++) {
2377       Node* in = phi->in(i);
2378       if (in == NULL) continue;
2379       new_phi->init_req(i, phase->transform(new GetNullFreePropertyNode(in)));
2380     }
2381     return new_phi;
2382   }
2383   return NULL;
2384 }
2385 
2386 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2387   Node* x = LoadNode::Identity(phase);
2388   if (x != this)  return x;
2389 
2390   // Take apart the address into an oop and and offset.
2391   // Return 'this' if we cannot.
2392   Node*    adr    = in(MemNode::Address);
2393   intptr_t offset = 0;
2394   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2395   if (base == NULL)     return this;
2396   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2397   if (toop == NULL)     return this;
2398 
2399   // Step over potential GC barrier for OopHandle resolve
2400   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2401   if (bs->is_gc_barrier_node(base)) {
2402     base = bs->step_over_gc_barrier(base);
2403   }
2404 
2405   // We can fetch the klass directly through an AllocateNode.
2406   // This works even if the klass is not constant (clone or newArray).
2407   if (offset == oopDesc::klass_offset_in_bytes()) {
2408     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2409     if (allocated_klass != NULL) {
2410       return allocated_klass;
2411     }
2412   }
2413 
2414   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2415   // See inline_native_Class_query for occurrences of these patterns.
2416   // Java Example:  x.getClass().isAssignableFrom(y)
2417   //
2418   // This improves reflective code, often making the Class
2419   // mirror go completely dead.  (Current exception:  Class
2420   // mirrors may appear in debug info, but we could clean them out by
2421   // introducing a new debug info operator for Klass.java_mirror).
2422 
2423   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2424       && offset == java_lang_Class::klass_offset_in_bytes()) {
2425     if (base->is_Load()) {
2426       Node* base2 = base->in(MemNode::Address);
2427       if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2428         Node* adr2 = base2->in(MemNode::Address);
2429         const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2430         if (tkls != NULL && !tkls->empty()
2431             && (tkls->klass()->is_instance_klass() ||
2432               tkls->klass()->is_array_klass())
2433             && adr2->is_AddP()
2434            ) {
2435           int mirror_field = in_bytes(Klass::java_mirror_offset());
2436           if (tkls->offset() == mirror_field) {
2437             return adr2->in(AddPNode::Base);
2438           }
2439         }
2440       }
2441     }
2442   }
2443 
2444   return this;
2445 }
2446 
2447 
2448 //------------------------------Value------------------------------------------
2449 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2450   const Type *t = klass_value_common(phase);
2451   if (t == Type::TOP)
2452     return t;
2453 
2454   return t->make_narrowklass();
2455 }
2456 
2457 //------------------------------Identity---------------------------------------
2458 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2459 // Also feed through the klass in Allocate(...klass...)._klass.
2460 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2461   Node *x = klass_identity_common(phase);
2462 
2463   const Type *t = phase->type( x );
2464   if( t == Type::TOP ) return x;
2465   if( t->isa_narrowklass()) return x;
2466   assert (!t->isa_narrowoop(), "no narrow oop here");
2467 
2468   return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2469 }
2470 
2471 //------------------------------Value-----------------------------------------
2472 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2473   // Either input is TOP ==> the result is TOP
2474   const Type *t1 = phase->type( in(MemNode::Memory) );
2475   if( t1 == Type::TOP ) return Type::TOP;
2476   Node *adr = in(MemNode::Address);
2477   const Type *t2 = phase->type( adr );
2478   if( t2 == Type::TOP ) return Type::TOP;
2479   const TypePtr *tp = t2->is_ptr();
2480   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2481   const TypeAryPtr *tap = tp->isa_aryptr();
2482   if( !tap ) return _type;
2483   return tap->size();
2484 }
2485 
2486 //-------------------------------Ideal---------------------------------------
2487 // Feed through the length in AllocateArray(...length...)._length.
2488 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2489   Node* p = MemNode::Ideal_common(phase, can_reshape);
2490   if (p)  return (p == NodeSentinel) ? NULL : p;
2491 
2492   // Take apart the address into an oop and and offset.
2493   // Return 'this' if we cannot.
2494   Node*    adr    = in(MemNode::Address);
2495   intptr_t offset = 0;
2496   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2497   if (base == NULL)     return NULL;
2498   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2499   if (tary == NULL)     return NULL;
2500 
2501   // We can fetch the length directly through an AllocateArrayNode.
2502   // This works even if the length is not constant (clone or newArray).
2503   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2504     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2505     if (alloc != NULL) {
2506       Node* allocated_length = alloc->Ideal_length();
2507       Node* len = alloc->make_ideal_length(tary, phase);
2508       if (allocated_length != len) {
2509         // New CastII improves on this.
2510         return len;
2511       }
2512     }
2513   }
2514 
2515   return NULL;
2516 }
2517 
2518 //------------------------------Identity---------------------------------------
2519 // Feed through the length in AllocateArray(...length...)._length.
2520 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2521   Node* x = LoadINode::Identity(phase);
2522   if (x != this)  return x;
2523 
2524   // Take apart the address into an oop and and offset.
2525   // Return 'this' if we cannot.
2526   Node*    adr    = in(MemNode::Address);
2527   intptr_t offset = 0;
2528   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2529   if (base == NULL)     return this;
2530   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2531   if (tary == NULL)     return this;
2532 
2533   // We can fetch the length directly through an AllocateArrayNode.
2534   // This works even if the length is not constant (clone or newArray).
2535   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2536     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2537     if (alloc != NULL) {
2538       Node* allocated_length = alloc->Ideal_length();
2539       // Do not allow make_ideal_length to allocate a CastII node.
2540       Node* len = alloc->make_ideal_length(tary, phase, false);
2541       if (allocated_length == len) {
2542         // Return allocated_length only if it would not be improved by a CastII.
2543         return allocated_length;
2544       }
2545     }
2546   }
2547 
2548   return this;
2549 
2550 }
2551 
2552 //=============================================================================
2553 //---------------------------StoreNode::make-----------------------------------
2554 // Polymorphic factory method:
2555 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2556   assert((mo == unordered || mo == release), "unexpected");
2557   Compile* C = gvn.C;
2558   assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2559          ctl != NULL, "raw memory operations should have control edge");
2560 
2561   switch (bt) {
2562   case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2563   case T_BYTE:    return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2564   case T_INT:     return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2565   case T_CHAR:
2566   case T_SHORT:   return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2567   case T_LONG:    return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2568   case T_FLOAT:   return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2569   case T_DOUBLE:  return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2570   case T_METADATA:
2571   case T_ADDRESS:
2572   case T_VALUETYPE:
2573   case T_OBJECT:
2574 #ifdef _LP64
2575     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2576       val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2577       return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2578     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2579                (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2580                 adr->bottom_type()->isa_rawptr())) {
2581       val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2582       return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2583     }
2584 #endif
2585     {
2586       return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2587     }
2588   default:
2589     ShouldNotReachHere();
2590     return (StoreNode*)NULL;
2591   }
2592 }
2593 
2594 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2595   bool require_atomic = true;
2596   return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2597 }
2598 
2599 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2600   bool require_atomic = true;
2601   return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2602 }
2603 
2604 
2605 //--------------------------bottom_type----------------------------------------
2606 const Type *StoreNode::bottom_type() const {
2607   return Type::MEMORY;
2608 }
2609 
2610 //------------------------------hash-------------------------------------------
2611 uint StoreNode::hash() const {
2612   // unroll addition of interesting fields
2613   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2614 
2615   // Since they are not commoned, do not hash them:
2616   return NO_HASH;
2617 }
2618 
2619 //------------------------------Ideal------------------------------------------
2620 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2621 // When a store immediately follows a relevant allocation/initialization,
2622 // try to capture it into the initialization, or hoist it above.
2623 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2624   Node* p = MemNode::Ideal_common(phase, can_reshape);
2625   if (p)  return (p == NodeSentinel) ? NULL : p;
2626 
2627   Node* mem     = in(MemNode::Memory);
2628   Node* address = in(MemNode::Address);
2629   // Back-to-back stores to same address?  Fold em up.  Generally
2630   // unsafe if I have intervening uses...  Also disallowed for StoreCM
2631   // since they must follow each StoreP operation.  Redundant StoreCMs
2632   // are eliminated just before matching in final_graph_reshape.
2633   if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::VALUES) {
2634     Node* st = mem;
2635     // If Store 'st' has more than one use, we cannot fold 'st' away.
2636     // For example, 'st' might be the final state at a conditional
2637     // return.  Or, 'st' might be used by some node which is live at
2638     // the same time 'st' is live, which might be unschedulable.  So,
2639     // require exactly ONE user until such time as we clone 'mem' for
2640     // each of 'mem's uses (thus making the exactly-1-user-rule hold
2641     // true).
2642     while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2643       // Looking at a dead closed cycle of memory?
2644       assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2645       assert(Opcode() == st->Opcode() ||
2646              st->Opcode() == Op_StoreVector ||
2647              Opcode() == Op_StoreVector ||
2648              phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2649              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2650              (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2651              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
2652              (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2653              "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2654 
2655       if (st->in(MemNode::Address)->eqv_uncast(address) &&
2656           st->as_Store()->memory_size() <= this->memory_size()) {
2657         Node* use = st->raw_out(0);
2658         phase->igvn_rehash_node_delayed(use);
2659         if (can_reshape) {
2660           use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2661         } else {
2662           // It's OK to do this in the parser, since DU info is always accurate,
2663           // and the parser always refers to nodes via SafePointNode maps.
2664           use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2665         }
2666         return this;
2667       }
2668       st = st->in(MemNode::Memory);
2669     }
2670   }
2671 
2672 
2673   // Capture an unaliased, unconditional, simple store into an initializer.
2674   // Or, if it is independent of the allocation, hoist it above the allocation.
2675   if (ReduceFieldZeroing && /*can_reshape &&*/
2676       mem->is_Proj() && mem->in(0)->is_Initialize()) {
2677     InitializeNode* init = mem->in(0)->as_Initialize();
2678     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2679     if (offset > 0) {
2680       Node* moved = init->capture_store(this, offset, phase, can_reshape);
2681       // If the InitializeNode captured me, it made a raw copy of me,
2682       // and I need to disappear.
2683       if (moved != NULL) {
2684         // %%% hack to ensure that Ideal returns a new node:
2685         mem = MergeMemNode::make(mem);
2686         return mem;             // fold me away
2687       }
2688     }
2689   }
2690 
2691   return NULL;                  // No further progress
2692 }
2693 
2694 //------------------------------Value-----------------------------------------
2695 const Type* StoreNode::Value(PhaseGVN* phase) const {
2696   // Either input is TOP ==> the result is TOP
2697   const Type *t1 = phase->type( in(MemNode::Memory) );
2698   if( t1 == Type::TOP ) return Type::TOP;
2699   const Type *t2 = phase->type( in(MemNode::Address) );
2700   if( t2 == Type::TOP ) return Type::TOP;
2701   const Type *t3 = phase->type( in(MemNode::ValueIn) );
2702   if( t3 == Type::TOP ) return Type::TOP;
2703   return Type::MEMORY;
2704 }
2705 
2706 //------------------------------Identity---------------------------------------
2707 // Remove redundant stores:
2708 //   Store(m, p, Load(m, p)) changes to m.
2709 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2710 Node* StoreNode::Identity(PhaseGVN* phase) {
2711   Node* mem = in(MemNode::Memory);
2712   Node* adr = in(MemNode::Address);
2713   Node* val = in(MemNode::ValueIn);
2714 
2715   Node* result = this;
2716 
2717   // Load then Store?  Then the Store is useless
2718   if (val->is_Load() &&
2719       val->in(MemNode::Address)->eqv_uncast(adr) &&
2720       val->in(MemNode::Memory )->eqv_uncast(mem) &&
2721       val->as_Load()->store_Opcode() == Opcode()) {
2722     result = mem;
2723   }
2724 
2725   // Two stores in a row of the same value?
2726   if (result == this &&
2727       mem->is_Store() &&
2728       mem->in(MemNode::Address)->eqv_uncast(adr) &&
2729       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2730       mem->Opcode() == Opcode()) {
2731     result = mem;
2732   }
2733 
2734   // Store of zero anywhere into a freshly-allocated object?
2735   // Then the store is useless.
2736   // (It must already have been captured by the InitializeNode.)
2737   if (result == this && ReduceFieldZeroing) {
2738     // a newly allocated object is already all-zeroes everywhere
2739     if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
2740         (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) {
2741       assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to value array is forbidden");
2742       result = mem;
2743     }
2744 
2745     if (result == this) {
2746       // the store may also apply to zero-bits in an earlier object
2747       Node* prev_mem = find_previous_store(phase);
2748       // Steps (a), (b):  Walk past independent stores to find an exact match.
2749       if (prev_mem != NULL) {
2750         Node* prev_val = can_see_stored_value(prev_mem, phase);
2751         if (prev_val != NULL && phase->eqv(prev_val, val)) {
2752           // prev_val and val might differ by a cast; it would be good
2753           // to keep the more informative of the two.
2754           if (phase->type(val)->is_zero_type()) {
2755             result = mem;
2756           } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) {
2757             InitializeNode* init = prev_mem->in(0)->as_Initialize();
2758             AllocateNode* alloc = init->allocation();
2759             if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) {
2760               result = mem;
2761             }
2762           }
2763         }
2764       }
2765     }
2766   }
2767 
2768   if (result != this && phase->is_IterGVN() != NULL) {
2769     MemBarNode* trailing = trailing_membar();
2770     if (trailing != NULL) {
2771 #ifdef ASSERT
2772       const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2773       assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2774 #endif
2775       PhaseIterGVN* igvn = phase->is_IterGVN();
2776       trailing->remove(igvn);
2777     }
2778   }
2779 
2780   return result;
2781 }
2782 
2783 //------------------------------match_edge-------------------------------------
2784 // Do we Match on this edge index or not?  Match only memory & value
2785 uint StoreNode::match_edge(uint idx) const {
2786   return idx == MemNode::Address || idx == MemNode::ValueIn;
2787 }
2788 
2789 //------------------------------cmp--------------------------------------------
2790 // Do not common stores up together.  They generally have to be split
2791 // back up anyways, so do not bother.
2792 bool StoreNode::cmp( const Node &n ) const {
2793   return (&n == this);          // Always fail except on self
2794 }
2795 
2796 //------------------------------Ideal_masked_input-----------------------------
2797 // Check for a useless mask before a partial-word store
2798 // (StoreB ... (AndI valIn conIa) )
2799 // If (conIa & mask == mask) this simplifies to
2800 // (StoreB ... (valIn) )
2801 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2802   Node *val = in(MemNode::ValueIn);
2803   if( val->Opcode() == Op_AndI ) {
2804     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2805     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2806       set_req(MemNode::ValueIn, val->in(1));
2807       return this;
2808     }
2809   }
2810   return NULL;
2811 }
2812 
2813 
2814 //------------------------------Ideal_sign_extended_input----------------------
2815 // Check for useless sign-extension before a partial-word store
2816 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2817 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
2818 // (StoreB ... (valIn) )
2819 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2820   Node *val = in(MemNode::ValueIn);
2821   if( val->Opcode() == Op_RShiftI ) {
2822     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2823     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2824       Node *shl = val->in(1);
2825       if( shl->Opcode() == Op_LShiftI ) {
2826         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2827         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2828           set_req(MemNode::ValueIn, shl->in(1));
2829           return this;
2830         }
2831       }
2832     }
2833   }
2834   return NULL;
2835 }
2836 
2837 //------------------------------value_never_loaded-----------------------------------
2838 // Determine whether there are any possible loads of the value stored.
2839 // For simplicity, we actually check if there are any loads from the
2840 // address stored to, not just for loads of the value stored by this node.
2841 //
2842 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2843   Node *adr = in(Address);
2844   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2845   if (adr_oop == NULL)
2846     return false;
2847   if (!adr_oop->is_known_instance_field())
2848     return false; // if not a distinct instance, there may be aliases of the address
2849   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2850     Node *use = adr->fast_out(i);
2851     if (use->is_Load() || use->is_LoadStore()) {
2852       return false;
2853     }
2854   }
2855   return true;
2856 }
2857 
2858 MemBarNode* StoreNode::trailing_membar() const {
2859   if (is_release()) {
2860     MemBarNode* trailing_mb = NULL;
2861     for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2862       Node* u = fast_out(i);
2863       if (u->is_MemBar()) {
2864         if (u->as_MemBar()->trailing_store()) {
2865           assert(u->Opcode() == Op_MemBarVolatile, "");
2866           assert(trailing_mb == NULL, "only one");
2867           trailing_mb = u->as_MemBar();
2868 #ifdef ASSERT
2869           Node* leading = u->as_MemBar()->leading_membar();
2870           assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2871           assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2872           assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2873 #endif
2874         } else {
2875           assert(u->as_MemBar()->standalone(), "");
2876         }
2877       }
2878     }
2879     return trailing_mb;
2880   }
2881   return NULL;
2882 }
2883 
2884 
2885 //=============================================================================
2886 //------------------------------Ideal------------------------------------------
2887 // If the store is from an AND mask that leaves the low bits untouched, then
2888 // we can skip the AND operation.  If the store is from a sign-extension
2889 // (a left shift, then right shift) we can skip both.
2890 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2891   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2892   if( progress != NULL ) return progress;
2893 
2894   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2895   if( progress != NULL ) return progress;
2896 
2897   // Finally check the default case
2898   return StoreNode::Ideal(phase, can_reshape);
2899 }
2900 
2901 //=============================================================================
2902 //------------------------------Ideal------------------------------------------
2903 // If the store is from an AND mask that leaves the low bits untouched, then
2904 // we can skip the AND operation
2905 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2906   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2907   if( progress != NULL ) return progress;
2908 
2909   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2910   if( progress != NULL ) return progress;
2911 
2912   // Finally check the default case
2913   return StoreNode::Ideal(phase, can_reshape);
2914 }
2915 
2916 //=============================================================================
2917 //------------------------------Identity---------------------------------------
2918 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2919   // No need to card mark when storing a null ptr
2920   Node* my_store = in(MemNode::OopStore);
2921   if (my_store->is_Store()) {
2922     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2923     if( t1 == TypePtr::NULL_PTR ) {
2924       return in(MemNode::Memory);
2925     }
2926   }
2927   return this;
2928 }
2929 
2930 //=============================================================================
2931 //------------------------------Ideal---------------------------------------
2932 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2933   Node* progress = StoreNode::Ideal(phase, can_reshape);
2934   if (progress != NULL) return progress;
2935 
2936   Node* my_store = in(MemNode::OopStore);
2937   if (my_store->is_MergeMem()) {
2938     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2939     set_req(MemNode::OopStore, mem);
2940     return this;
2941   }
2942 
2943   return NULL;
2944 }
2945 
2946 //------------------------------Value-----------------------------------------
2947 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2948   // Either input is TOP ==> the result is TOP
2949   const Type *t = phase->type( in(MemNode::Memory) );
2950   if( t == Type::TOP ) return Type::TOP;
2951   t = phase->type( in(MemNode::Address) );
2952   if( t == Type::TOP ) return Type::TOP;
2953   t = phase->type( in(MemNode::ValueIn) );
2954   if( t == Type::TOP ) return Type::TOP;
2955   // If extra input is TOP ==> the result is TOP
2956   t = phase->type( in(MemNode::OopStore) );
2957   if( t == Type::TOP ) return Type::TOP;
2958 
2959   return StoreNode::Value( phase );
2960 }
2961 
2962 
2963 //=============================================================================
2964 //----------------------------------SCMemProjNode------------------------------
2965 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2966 {
2967   return bottom_type();
2968 }
2969 
2970 //=============================================================================
2971 //----------------------------------LoadStoreNode------------------------------
2972 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2973   : Node(required),
2974     _type(rt),
2975     _adr_type(at),
2976     _has_barrier(false)
2977 {
2978   init_req(MemNode::Control, c  );
2979   init_req(MemNode::Memory , mem);
2980   init_req(MemNode::Address, adr);
2981   init_req(MemNode::ValueIn, val);
2982   init_class_id(Class_LoadStore);
2983 }
2984 
2985 uint LoadStoreNode::ideal_reg() const {
2986   return _type->ideal_reg();
2987 }
2988 
2989 bool LoadStoreNode::result_not_used() const {
2990   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2991     Node *x = fast_out(i);
2992     if (x->Opcode() == Op_SCMemProj) continue;
2993     return false;
2994   }
2995   return true;
2996 }
2997 
2998 MemBarNode* LoadStoreNode::trailing_membar() const {
2999   MemBarNode* trailing = NULL;
3000   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3001     Node* u = fast_out(i);
3002     if (u->is_MemBar()) {
3003       if (u->as_MemBar()->trailing_load_store()) {
3004         assert(u->Opcode() == Op_MemBarAcquire, "");
3005         assert(trailing == NULL, "only one");
3006         trailing = u->as_MemBar();
3007 #ifdef ASSERT
3008         Node* leading = trailing->leading_membar();
3009         assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3010         assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3011         assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3012 #endif
3013       } else {
3014         assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3015       }
3016     }
3017   }
3018 
3019   return trailing;
3020 }
3021 
3022 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3023 
3024 //=============================================================================
3025 //----------------------------------LoadStoreConditionalNode--------------------
3026 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
3027   init_req(ExpectedIn, ex );
3028 }
3029 
3030 //=============================================================================
3031 //-------------------------------adr_type--------------------------------------
3032 const TypePtr* ClearArrayNode::adr_type() const {
3033   Node *adr = in(3);
3034   if (adr == NULL)  return NULL; // node is dead
3035   return MemNode::calculate_adr_type(adr->bottom_type());
3036 }
3037 
3038 //------------------------------match_edge-------------------------------------
3039 // Do we Match on this edge index or not?  Do not match memory
3040 uint ClearArrayNode::match_edge(uint idx) const {
3041   return idx > 1;
3042 }
3043 
3044 //------------------------------Identity---------------------------------------
3045 // Clearing a zero length array does nothing
3046 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3047   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
3048 }
3049 
3050 //------------------------------Idealize---------------------------------------
3051 // Clearing a short array is faster with stores
3052 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3053   // Already know this is a large node, do not try to ideal it
3054   if (!IdealizeClearArrayNode || _is_large) return NULL;
3055 
3056   const int unit = BytesPerLong;
3057   const TypeX* t = phase->type(in(2))->isa_intptr_t();
3058   if (!t)  return NULL;
3059   if (!t->is_con())  return NULL;
3060   intptr_t raw_count = t->get_con();
3061   intptr_t size = raw_count;
3062   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3063   // Clearing nothing uses the Identity call.
3064   // Negative clears are possible on dead ClearArrays
3065   // (see jck test stmt114.stmt11402.val).
3066   if (size <= 0 || size % unit != 0)  return NULL;
3067   intptr_t count = size / unit;
3068   // Length too long; communicate this to matchers and assemblers.
3069   // Assemblers are responsible to produce fast hardware clears for it.
3070   if (size > InitArrayShortSize) {
3071     return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
3072   }
3073   Node *mem = in(1);
3074   if( phase->type(mem)==Type::TOP ) return NULL;
3075   Node *adr = in(3);
3076   const Type* at = phase->type(adr);
3077   if( at==Type::TOP ) return NULL;
3078   const TypePtr* atp = at->isa_ptr();
3079   // adjust atp to be the correct array element address type
3080   if (atp == NULL)  atp = TypePtr::BOTTOM;
3081   else              atp = atp->add_offset(Type::OffsetBot);
3082   // Get base for derived pointer purposes
3083   if( adr->Opcode() != Op_AddP ) Unimplemented();
3084   Node *base = adr->in(1);
3085 
3086   Node *val = in(4);
3087   Node *off  = phase->MakeConX(BytesPerLong);
3088   mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3089   count--;
3090   while( count-- ) {
3091     mem = phase->transform(mem);
3092     adr = phase->transform(new AddPNode(base,adr,off));
3093     mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3094   }
3095   return mem;
3096 }
3097 
3098 //----------------------------step_through----------------------------------
3099 // Return allocation input memory edge if it is different instance
3100 // or itself if it is the one we are looking for.
3101 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3102   Node* n = *np;
3103   assert(n->is_ClearArray(), "sanity");
3104   intptr_t offset;
3105   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3106   // This method is called only before Allocate nodes are expanded
3107   // during macro nodes expansion. Before that ClearArray nodes are
3108   // only generated in PhaseMacroExpand::generate_arraycopy() (before
3109   // Allocate nodes are expanded) which follows allocations.
3110   assert(alloc != NULL, "should have allocation");
3111   if (alloc->_idx == instance_id) {
3112     // Can not bypass initialization of the instance we are looking for.
3113     return false;
3114   }
3115   // Otherwise skip it.
3116   InitializeNode* init = alloc->initialization();
3117   if (init != NULL)
3118     *np = init->in(TypeFunc::Memory);
3119   else
3120     *np = alloc->in(TypeFunc::Memory);
3121   return true;
3122 }
3123 
3124 //----------------------------clear_memory-------------------------------------
3125 // Generate code to initialize object storage to zero.
3126 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3127                                    Node* val,
3128                                    Node* raw_val,
3129                                    intptr_t start_offset,
3130                                    Node* end_offset,
3131                                    PhaseGVN* phase) {
3132   intptr_t offset = start_offset;
3133 
3134   int unit = BytesPerLong;
3135   if ((offset % unit) != 0) {
3136     Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3137     adr = phase->transform(adr);
3138     const TypePtr* atp = TypeRawPtr::BOTTOM;
3139     if (val != NULL) {
3140       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3141       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3142     } else {
3143       assert(raw_val == NULL, "val may not be null");
3144       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3145     }
3146     mem = phase->transform(mem);
3147     offset += BytesPerInt;
3148   }
3149   assert((offset % unit) == 0, "");
3150 
3151   // Initialize the remaining stuff, if any, with a ClearArray.
3152   return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
3153 }
3154 
3155 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3156                                    Node* raw_val,
3157                                    Node* start_offset,
3158                                    Node* end_offset,
3159                                    PhaseGVN* phase) {
3160   if (start_offset == end_offset) {
3161     // nothing to do
3162     return mem;
3163   }
3164 
3165   int unit = BytesPerLong;
3166   Node* zbase = start_offset;
3167   Node* zend  = end_offset;
3168 
3169   // Scale to the unit required by the CPU:
3170   if (!Matcher::init_array_count_is_in_bytes) {
3171     Node* shift = phase->intcon(exact_log2(unit));
3172     zbase = phase->transform(new URShiftXNode(zbase, shift) );
3173     zend  = phase->transform(new URShiftXNode(zend,  shift) );
3174   }
3175 
3176   // Bulk clear double-words
3177   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3178   Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3179   if (raw_val == NULL) {
3180     raw_val = phase->MakeConX(0);
3181   }
3182   mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
3183   return phase->transform(mem);
3184 }
3185 
3186 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3187                                    Node* val,
3188                                    Node* raw_val,
3189                                    intptr_t start_offset,
3190                                    intptr_t end_offset,
3191                                    PhaseGVN* phase) {
3192   if (start_offset == end_offset) {
3193     // nothing to do
3194     return mem;
3195   }
3196 
3197   assert((end_offset % BytesPerInt) == 0, "odd end offset");
3198   intptr_t done_offset = end_offset;
3199   if ((done_offset % BytesPerLong) != 0) {
3200     done_offset -= BytesPerInt;
3201   }
3202   if (done_offset > start_offset) {
3203     mem = clear_memory(ctl, mem, dest, val, raw_val,
3204                        start_offset, phase->MakeConX(done_offset), phase);
3205   }
3206   if (done_offset < end_offset) { // emit the final 32-bit store
3207     Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3208     adr = phase->transform(adr);
3209     const TypePtr* atp = TypeRawPtr::BOTTOM;
3210     if (val != NULL) {
3211       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3212       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3213     } else {
3214       assert(raw_val == NULL, "val may not be null");
3215       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3216     }
3217     mem = phase->transform(mem);
3218     done_offset += BytesPerInt;
3219   }
3220   assert(done_offset == end_offset, "");
3221   return mem;
3222 }
3223 
3224 //=============================================================================
3225 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3226   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3227     _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3228 #ifdef ASSERT
3229   , _pair_idx(0)
3230 #endif
3231 {
3232   init_class_id(Class_MemBar);
3233   Node* top = C->top();
3234   init_req(TypeFunc::I_O,top);
3235   init_req(TypeFunc::FramePtr,top);
3236   init_req(TypeFunc::ReturnAdr,top);
3237   if (precedent != NULL)
3238     init_req(TypeFunc::Parms, precedent);
3239 }
3240 
3241 //------------------------------cmp--------------------------------------------
3242 uint MemBarNode::hash() const { return NO_HASH; }
3243 bool MemBarNode::cmp( const Node &n ) const {
3244   return (&n == this);          // Always fail except on self
3245 }
3246 
3247 //------------------------------make-------------------------------------------
3248 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3249   switch (opcode) {
3250   case Op_MemBarAcquire:     return new MemBarAcquireNode(C, atp, pn);
3251   case Op_LoadFence:         return new LoadFenceNode(C, atp, pn);
3252   case Op_MemBarRelease:     return new MemBarReleaseNode(C, atp, pn);
3253   case Op_StoreFence:        return new StoreFenceNode(C, atp, pn);
3254   case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3255   case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3256   case Op_MemBarVolatile:    return new MemBarVolatileNode(C, atp, pn);
3257   case Op_MemBarCPUOrder:    return new MemBarCPUOrderNode(C, atp, pn);
3258   case Op_OnSpinWait:        return new OnSpinWaitNode(C, atp, pn);
3259   case Op_Initialize:        return new InitializeNode(C, atp, pn);
3260   case Op_MemBarStoreStore:  return new MemBarStoreStoreNode(C, atp, pn);
3261   default: ShouldNotReachHere(); return NULL;
3262   }
3263 }
3264 
3265 void MemBarNode::remove(PhaseIterGVN *igvn) {
3266   if (outcnt() != 2) {
3267     return;
3268   }
3269   if (trailing_store() || trailing_load_store()) {
3270     MemBarNode* leading = leading_membar();
3271     if (leading != NULL) {
3272       assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3273       leading->remove(igvn);
3274     }
3275   }
3276   igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3277   igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3278 }
3279 
3280 //------------------------------Ideal------------------------------------------
3281 // Return a node which is more "ideal" than the current node.  Strip out
3282 // control copies
3283 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3284   if (remove_dead_region(phase, can_reshape)) return this;
3285   // Don't bother trying to transform a dead node
3286   if (in(0) && in(0)->is_top()) {
3287     return NULL;
3288   }
3289 
3290   bool progress = false;
3291   // Eliminate volatile MemBars for scalar replaced objects.
3292   if (can_reshape && req() == (Precedent+1)) {
3293     bool eliminate = false;
3294     int opc = Opcode();
3295     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3296       // Volatile field loads and stores.
3297       Node* my_mem = in(MemBarNode::Precedent);
3298       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3299       if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3300         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3301         // replace this Precedent (decodeN) with the Load instead.
3302         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
3303           Node* load_node = my_mem->in(1);
3304           set_req(MemBarNode::Precedent, load_node);
3305           phase->is_IterGVN()->_worklist.push(my_mem);
3306           my_mem = load_node;
3307         } else {
3308           assert(my_mem->unique_out() == this, "sanity");
3309           del_req(Precedent);
3310           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3311           my_mem = NULL;
3312         }
3313         progress = true;
3314       }
3315       if (my_mem != NULL && my_mem->is_Mem()) {
3316         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3317         // Check for scalar replaced object reference.
3318         if( t_oop != NULL && t_oop->is_known_instance_field() &&
3319             t_oop->offset() != Type::OffsetBot &&
3320             t_oop->offset() != Type::OffsetTop) {
3321           eliminate = true;
3322         }
3323       }
3324     } else if (opc == Op_MemBarRelease) {
3325       // Final field stores.
3326       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3327       if ((alloc != NULL) && alloc->is_Allocate() &&
3328           alloc->as_Allocate()->does_not_escape_thread()) {
3329         // The allocated object does not escape.
3330         eliminate = true;
3331       }
3332     }
3333     if (eliminate) {
3334       // Replace MemBar projections by its inputs.
3335       PhaseIterGVN* igvn = phase->is_IterGVN();
3336       remove(igvn);
3337       // Must return either the original node (now dead) or a new node
3338       // (Do not return a top here, since that would break the uniqueness of top.)
3339       return new ConINode(TypeInt::ZERO);
3340     }
3341   }
3342   return progress ? this : NULL;
3343 }
3344 
3345 //------------------------------Value------------------------------------------
3346 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3347   if( !in(0) ) return Type::TOP;
3348   if( phase->type(in(0)) == Type::TOP )
3349     return Type::TOP;
3350   return TypeTuple::MEMBAR;
3351 }
3352 
3353 //------------------------------match------------------------------------------
3354 // Construct projections for memory.
3355 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
3356   switch (proj->_con) {
3357   case TypeFunc::Control:
3358   case TypeFunc::Memory:
3359     return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3360   }
3361   ShouldNotReachHere();
3362   return NULL;
3363 }
3364 
3365 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3366   trailing->_kind = TrailingStore;
3367   leading->_kind = LeadingStore;
3368 #ifdef ASSERT
3369   trailing->_pair_idx = leading->_idx;
3370   leading->_pair_idx = leading->_idx;
3371 #endif
3372 }
3373 
3374 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3375   trailing->_kind = TrailingLoadStore;
3376   leading->_kind = LeadingLoadStore;
3377 #ifdef ASSERT
3378   trailing->_pair_idx = leading->_idx;
3379   leading->_pair_idx = leading->_idx;
3380 #endif
3381 }
3382 
3383 MemBarNode* MemBarNode::trailing_membar() const {
3384   ResourceMark rm;
3385   Node* trailing = (Node*)this;
3386   VectorSet seen(Thread::current()->resource_area());
3387   Node_Stack multis(0);
3388   do {
3389     Node* c = trailing;
3390     uint i = 0;
3391     do {
3392       trailing = NULL;
3393       for (; i < c->outcnt(); i++) {
3394         Node* next = c->raw_out(i);
3395         if (next != c && next->is_CFG()) {
3396           if (c->is_MultiBranch()) {
3397             if (multis.node() == c) {
3398               multis.set_index(i+1);
3399             } else {
3400               multis.push(c, i+1);
3401             }
3402           }
3403           trailing = next;
3404           break;
3405         }
3406       }
3407       if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3408         break;
3409       }
3410       while (multis.size() > 0) {
3411         c = multis.node();
3412         i = multis.index();
3413         if (i < c->req()) {
3414           break;
3415         }
3416         multis.pop();
3417       }
3418     } while (multis.size() > 0);
3419   } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3420 
3421   MemBarNode* mb = trailing->as_MemBar();
3422   assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3423          (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3424   assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3425   return mb;
3426 }
3427 
3428 MemBarNode* MemBarNode::leading_membar() const {
3429   ResourceMark rm;
3430   VectorSet seen(Thread::current()->resource_area());
3431   Node_Stack regions(0);
3432   Node* leading = in(0);
3433   while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3434     while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3435       leading = NULL;
3436       while (regions.size() > 0 && leading == NULL) {
3437         Node* r = regions.node();
3438         uint i = regions.index();
3439         if (i < r->req()) {
3440           leading = r->in(i);
3441           regions.set_index(i+1);
3442         } else {
3443           regions.pop();
3444         }
3445       }
3446       if (leading == NULL) {
3447         assert(regions.size() == 0, "all paths should have been tried");
3448         return NULL;
3449       }
3450     }
3451     if (leading->is_Region()) {
3452       regions.push(leading, 2);
3453       leading = leading->in(1);
3454     } else {
3455       leading = leading->in(0);
3456     }
3457   }
3458 #ifdef ASSERT
3459   Unique_Node_List wq;
3460   wq.push((Node*)this);
3461   uint found = 0;
3462   for (uint i = 0; i < wq.size(); i++) {
3463     Node* n = wq.at(i);
3464     if (n->is_Region()) {
3465       for (uint j = 1; j < n->req(); j++) {
3466         Node* in = n->in(j);
3467         if (in != NULL && !in->is_top()) {
3468           wq.push(in);
3469         }
3470       }
3471     } else {
3472       if (n->is_MemBar() && n->as_MemBar()->leading()) {
3473         assert(n == leading, "consistency check failed");
3474         found++;
3475       } else {
3476         Node* in = n->in(0);
3477         if (in != NULL && !in->is_top()) {
3478           wq.push(in);
3479         }
3480       }
3481     }
3482   }
3483   assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3484 #endif
3485   if (leading == NULL) {
3486     return NULL;
3487   }
3488   MemBarNode* mb = leading->as_MemBar();
3489   assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3490          (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3491   assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3492   return mb;
3493 }
3494 
3495 //===========================InitializeNode====================================
3496 // SUMMARY:
3497 // This node acts as a memory barrier on raw memory, after some raw stores.
3498 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3499 // The Initialize can 'capture' suitably constrained stores as raw inits.
3500 // It can coalesce related raw stores into larger units (called 'tiles').
3501 // It can avoid zeroing new storage for memory units which have raw inits.
3502 // At macro-expansion, it is marked 'complete', and does not optimize further.
3503 //
3504 // EXAMPLE:
3505 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3506 //   ctl = incoming control; mem* = incoming memory
3507 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
3508 // First allocate uninitialized memory and fill in the header:
3509 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
3510 //   ctl := alloc.Control; mem* := alloc.Memory*
3511 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
3512 // Then initialize to zero the non-header parts of the raw memory block:
3513 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3514 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3515 // After the initialize node executes, the object is ready for service:
3516 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3517 // Suppose its body is immediately initialized as {1,2}:
3518 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3519 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3520 //   mem.SLICE(#short[*]) := store2
3521 //
3522 // DETAILS:
3523 // An InitializeNode collects and isolates object initialization after
3524 // an AllocateNode and before the next possible safepoint.  As a
3525 // memory barrier (MemBarNode), it keeps critical stores from drifting
3526 // down past any safepoint or any publication of the allocation.
3527 // Before this barrier, a newly-allocated object may have uninitialized bits.
3528 // After this barrier, it may be treated as a real oop, and GC is allowed.
3529 //
3530 // The semantics of the InitializeNode include an implicit zeroing of
3531 // the new object from object header to the end of the object.
3532 // (The object header and end are determined by the AllocateNode.)
3533 //
3534 // Certain stores may be added as direct inputs to the InitializeNode.
3535 // These stores must update raw memory, and they must be to addresses
3536 // derived from the raw address produced by AllocateNode, and with
3537 // a constant offset.  They must be ordered by increasing offset.
3538 // The first one is at in(RawStores), the last at in(req()-1).
3539 // Unlike most memory operations, they are not linked in a chain,
3540 // but are displayed in parallel as users of the rawmem output of
3541 // the allocation.
3542 //
3543 // (See comments in InitializeNode::capture_store, which continue
3544 // the example given above.)
3545 //
3546 // When the associated Allocate is macro-expanded, the InitializeNode
3547 // may be rewritten to optimize collected stores.  A ClearArrayNode
3548 // may also be created at that point to represent any required zeroing.
3549 // The InitializeNode is then marked 'complete', prohibiting further
3550 // capturing of nearby memory operations.
3551 //
3552 // During macro-expansion, all captured initializations which store
3553 // constant values of 32 bits or smaller are coalesced (if advantageous)
3554 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
3555 // initialized in fewer memory operations.  Memory words which are
3556 // covered by neither tiles nor non-constant stores are pre-zeroed
3557 // by explicit stores of zero.  (The code shape happens to do all
3558 // zeroing first, then all other stores, with both sequences occurring
3559 // in order of ascending offsets.)
3560 //
3561 // Alternatively, code may be inserted between an AllocateNode and its
3562 // InitializeNode, to perform arbitrary initialization of the new object.
3563 // E.g., the object copying intrinsics insert complex data transfers here.
3564 // The initialization must then be marked as 'complete' disable the
3565 // built-in zeroing semantics and the collection of initializing stores.
3566 //
3567 // While an InitializeNode is incomplete, reads from the memory state
3568 // produced by it are optimizable if they match the control edge and
3569 // new oop address associated with the allocation/initialization.
3570 // They return a stored value (if the offset matches) or else zero.
3571 // A write to the memory state, if it matches control and address,
3572 // and if it is to a constant offset, may be 'captured' by the
3573 // InitializeNode.  It is cloned as a raw memory operation and rewired
3574 // inside the initialization, to the raw oop produced by the allocation.
3575 // Operations on addresses which are provably distinct (e.g., to
3576 // other AllocateNodes) are allowed to bypass the initialization.
3577 //
3578 // The effect of all this is to consolidate object initialization
3579 // (both arrays and non-arrays, both piecewise and bulk) into a
3580 // single location, where it can be optimized as a unit.
3581 //
3582 // Only stores with an offset less than TrackedInitializationLimit words
3583 // will be considered for capture by an InitializeNode.  This puts a
3584 // reasonable limit on the complexity of optimized initializations.
3585 
3586 //---------------------------InitializeNode------------------------------------
3587 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3588   : MemBarNode(C, adr_type, rawoop),
3589     _is_complete(Incomplete), _does_not_escape(false)
3590 {
3591   init_class_id(Class_Initialize);
3592 
3593   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3594   assert(in(RawAddress) == rawoop, "proper init");
3595   // Note:  allocation() can be NULL, for secondary initialization barriers
3596 }
3597 
3598 // Since this node is not matched, it will be processed by the
3599 // register allocator.  Declare that there are no constraints
3600 // on the allocation of the RawAddress edge.
3601 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3602   // This edge should be set to top, by the set_complete.  But be conservative.
3603   if (idx == InitializeNode::RawAddress)
3604     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3605   return RegMask::Empty;
3606 }
3607 
3608 Node* InitializeNode::memory(uint alias_idx) {
3609   Node* mem = in(Memory);
3610   if (mem->is_MergeMem()) {
3611     return mem->as_MergeMem()->memory_at(alias_idx);
3612   } else {
3613     // incoming raw memory is not split
3614     return mem;
3615   }
3616 }
3617 
3618 bool InitializeNode::is_non_zero() {
3619   if (is_complete())  return false;
3620   remove_extra_zeroes();
3621   return (req() > RawStores);
3622 }
3623 
3624 void InitializeNode::set_complete(PhaseGVN* phase) {
3625   assert(!is_complete(), "caller responsibility");
3626   _is_complete = Complete;
3627 
3628   // After this node is complete, it contains a bunch of
3629   // raw-memory initializations.  There is no need for
3630   // it to have anything to do with non-raw memory effects.
3631   // Therefore, tell all non-raw users to re-optimize themselves,
3632   // after skipping the memory effects of this initialization.
3633   PhaseIterGVN* igvn = phase->is_IterGVN();
3634   if (igvn)  igvn->add_users_to_worklist(this);
3635 }
3636 
3637 // convenience function
3638 // return false if the init contains any stores already
3639 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3640   InitializeNode* init = initialization();
3641   if (init == NULL || init->is_complete()) {
3642     return false;
3643   }
3644   init->remove_extra_zeroes();
3645   // for now, if this allocation has already collected any inits, bail:
3646   if (init->is_non_zero())  return false;
3647   init->set_complete(phase);
3648   return true;
3649 }
3650 
3651 void InitializeNode::remove_extra_zeroes() {
3652   if (req() == RawStores)  return;
3653   Node* zmem = zero_memory();
3654   uint fill = RawStores;
3655   for (uint i = fill; i < req(); i++) {
3656     Node* n = in(i);
3657     if (n->is_top() || n == zmem)  continue;  // skip
3658     if (fill < i)  set_req(fill, n);          // compact
3659     ++fill;
3660   }
3661   // delete any empty spaces created:
3662   while (fill < req()) {
3663     del_req(fill);
3664   }
3665 }
3666 
3667 // Helper for remembering which stores go with which offsets.
3668 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3669   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
3670   intptr_t offset = -1;
3671   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3672                                                phase, offset);
3673   if (base == NULL)     return -1;  // something is dead,
3674   if (offset < 0)       return -1;  //        dead, dead
3675   return offset;
3676 }
3677 
3678 // Helper for proving that an initialization expression is
3679 // "simple enough" to be folded into an object initialization.
3680 // Attempts to prove that a store's initial value 'n' can be captured
3681 // within the initialization without creating a vicious cycle, such as:
3682 //     { Foo p = new Foo(); p.next = p; }
3683 // True for constants and parameters and small combinations thereof.
3684 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3685   if (n == NULL)      return true;   // (can this really happen?)
3686   if (n->is_Proj())   n = n->in(0);
3687   if (n == this)      return false;  // found a cycle
3688   if (n->is_Con())    return true;
3689   if (n->is_Start())  return true;   // params, etc., are OK
3690   if (n->is_Root())   return true;   // even better
3691 
3692   Node* ctl = n->in(0);
3693   if (ctl != NULL && !ctl->is_top()) {
3694     if (ctl->is_Proj())  ctl = ctl->in(0);
3695     if (ctl == this)  return false;
3696 
3697     // If we already know that the enclosing memory op is pinned right after
3698     // the init, then any control flow that the store has picked up
3699     // must have preceded the init, or else be equal to the init.
3700     // Even after loop optimizations (which might change control edges)
3701     // a store is never pinned *before* the availability of its inputs.
3702     if (!MemNode::all_controls_dominate(n, this))
3703       return false;                  // failed to prove a good control
3704   }
3705 
3706   // Check data edges for possible dependencies on 'this'.
3707   if ((count += 1) > 20)  return false;  // complexity limit
3708   for (uint i = 1; i < n->req(); i++) {
3709     Node* m = n->in(i);
3710     if (m == NULL || m == n || m->is_top())  continue;
3711     uint first_i = n->find_edge(m);
3712     if (i != first_i)  continue;  // process duplicate edge just once
3713     if (!detect_init_independence(m, count)) {
3714       return false;
3715     }
3716   }
3717 
3718   return true;
3719 }
3720 
3721 // Here are all the checks a Store must pass before it can be moved into
3722 // an initialization.  Returns zero if a check fails.
3723 // On success, returns the (constant) offset to which the store applies,
3724 // within the initialized memory.
3725 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3726   const int FAIL = 0;
3727   if (st->req() != MemNode::ValueIn + 1)
3728     return FAIL;                // an inscrutable StoreNode (card mark?)
3729   Node* ctl = st->in(MemNode::Control);
3730   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3731     return FAIL;                // must be unconditional after the initialization
3732   Node* mem = st->in(MemNode::Memory);
3733   if (!(mem->is_Proj() && mem->in(0) == this))
3734     return FAIL;                // must not be preceded by other stores
3735   Node* adr = st->in(MemNode::Address);
3736   intptr_t offset;
3737   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3738   if (alloc == NULL)
3739     return FAIL;                // inscrutable address
3740   if (alloc != allocation())
3741     return FAIL;                // wrong allocation!  (store needs to float up)
3742   int size_in_bytes = st->memory_size();
3743   if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3744     return FAIL;                // mismatched access
3745   }
3746   Node* val = st->in(MemNode::ValueIn);
3747   int complexity_count = 0;
3748   if (!detect_init_independence(val, complexity_count))
3749     return FAIL;                // stored value must be 'simple enough'
3750 
3751   // The Store can be captured only if nothing after the allocation
3752   // and before the Store is using the memory location that the store
3753   // overwrites.
3754   bool failed = false;
3755   // If is_complete_with_arraycopy() is true the shape of the graph is
3756   // well defined and is safe so no need for extra checks.
3757   if (!is_complete_with_arraycopy()) {
3758     // We are going to look at each use of the memory state following
3759     // the allocation to make sure nothing reads the memory that the
3760     // Store writes.
3761     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3762     int alias_idx = phase->C->get_alias_index(t_adr);
3763     ResourceMark rm;
3764     Unique_Node_List mems;
3765     mems.push(mem);
3766     Node* unique_merge = NULL;
3767     for (uint next = 0; next < mems.size(); ++next) {
3768       Node *m  = mems.at(next);
3769       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3770         Node *n = m->fast_out(j);
3771         if (n->outcnt() == 0) {
3772           continue;
3773         }
3774         if (n == st) {
3775           continue;
3776         } else if (n->in(0) != NULL && n->in(0) != ctl) {
3777           // If the control of this use is different from the control
3778           // of the Store which is right after the InitializeNode then
3779           // this node cannot be between the InitializeNode and the
3780           // Store.
3781           continue;
3782         } else if (n->is_MergeMem()) {
3783           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3784             // We can hit a MergeMemNode (that will likely go away
3785             // later) that is a direct use of the memory state
3786             // following the InitializeNode on the same slice as the
3787             // store node that we'd like to capture. We need to check
3788             // the uses of the MergeMemNode.
3789             mems.push(n);
3790           }
3791         } else if (n->is_Mem()) {
3792           Node* other_adr = n->in(MemNode::Address);
3793           if (other_adr == adr) {
3794             failed = true;
3795             break;
3796           } else {
3797             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3798             if (other_t_adr != NULL) {
3799               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3800               if (other_alias_idx == alias_idx) {
3801                 // A load from the same memory slice as the store right
3802                 // after the InitializeNode. We check the control of the
3803                 // object/array that is loaded from. If it's the same as
3804                 // the store control then we cannot capture the store.
3805                 assert(!n->is_Store(), "2 stores to same slice on same control?");
3806                 Node* base = other_adr;
3807                 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3808                 base = base->in(AddPNode::Base);
3809                 if (base != NULL) {
3810                   base = base->uncast();
3811                   if (base->is_Proj() && base->in(0) == alloc) {
3812                     failed = true;
3813                     break;
3814                   }
3815                 }
3816               }
3817             }
3818           }
3819         } else {
3820           failed = true;
3821           break;
3822         }
3823       }
3824     }
3825   }
3826   if (failed) {
3827     if (!can_reshape) {
3828       // We decided we couldn't capture the store during parsing. We
3829       // should try again during the next IGVN once the graph is
3830       // cleaner.
3831       phase->C->record_for_igvn(st);
3832     }
3833     return FAIL;
3834   }
3835 
3836   return offset;                // success
3837 }
3838 
3839 // Find the captured store in(i) which corresponds to the range
3840 // [start..start+size) in the initialized object.
3841 // If there is one, return its index i.  If there isn't, return the
3842 // negative of the index where it should be inserted.
3843 // Return 0 if the queried range overlaps an initialization boundary
3844 // or if dead code is encountered.
3845 // If size_in_bytes is zero, do not bother with overlap checks.
3846 int InitializeNode::captured_store_insertion_point(intptr_t start,
3847                                                    int size_in_bytes,
3848                                                    PhaseTransform* phase) {
3849   const int FAIL = 0, MAX_STORE = BytesPerLong;
3850 
3851   if (is_complete())
3852     return FAIL;                // arraycopy got here first; punt
3853 
3854   assert(allocation() != NULL, "must be present");
3855 
3856   // no negatives, no header fields:
3857   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
3858 
3859   // after a certain size, we bail out on tracking all the stores:
3860   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3861   if (start >= ti_limit)  return FAIL;
3862 
3863   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3864     if (i >= limit)  return -(int)i; // not found; here is where to put it
3865 
3866     Node*    st     = in(i);
3867     intptr_t st_off = get_store_offset(st, phase);
3868     if (st_off < 0) {
3869       if (st != zero_memory()) {
3870         return FAIL;            // bail out if there is dead garbage
3871       }
3872     } else if (st_off > start) {
3873       // ...we are done, since stores are ordered
3874       if (st_off < start + size_in_bytes) {
3875         return FAIL;            // the next store overlaps
3876       }
3877       return -(int)i;           // not found; here is where to put it
3878     } else if (st_off < start) {
3879       if (size_in_bytes != 0 &&
3880           start < st_off + MAX_STORE &&
3881           start < st_off + st->as_Store()->memory_size()) {
3882         return FAIL;            // the previous store overlaps
3883       }
3884     } else {
3885       if (size_in_bytes != 0 &&
3886           st->as_Store()->memory_size() != size_in_bytes) {
3887         return FAIL;            // mismatched store size
3888       }
3889       return i;
3890     }
3891 
3892     ++i;
3893   }
3894 }
3895 
3896 // Look for a captured store which initializes at the offset 'start'
3897 // with the given size.  If there is no such store, and no other
3898 // initialization interferes, then return zero_memory (the memory
3899 // projection of the AllocateNode).
3900 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3901                                           PhaseTransform* phase) {
3902   assert(stores_are_sane(phase), "");
3903   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3904   if (i == 0) {
3905     return NULL;                // something is dead
3906   } else if (i < 0) {
3907     return zero_memory();       // just primordial zero bits here
3908   } else {
3909     Node* st = in(i);           // here is the store at this position
3910     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3911     return st;
3912   }
3913 }
3914 
3915 // Create, as a raw pointer, an address within my new object at 'offset'.
3916 Node* InitializeNode::make_raw_address(intptr_t offset,
3917                                        PhaseTransform* phase) {
3918   Node* addr = in(RawAddress);
3919   if (offset != 0) {
3920     Compile* C = phase->C;
3921     addr = phase->transform( new AddPNode(C->top(), addr,
3922                                                  phase->MakeConX(offset)) );
3923   }
3924   return addr;
3925 }
3926 
3927 // Clone the given store, converting it into a raw store
3928 // initializing a field or element of my new object.
3929 // Caller is responsible for retiring the original store,
3930 // with subsume_node or the like.
3931 //
3932 // From the example above InitializeNode::InitializeNode,
3933 // here are the old stores to be captured:
3934 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3935 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3936 //
3937 // Here is the changed code; note the extra edges on init:
3938 //   alloc = (Allocate ...)
3939 //   rawoop = alloc.RawAddress
3940 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3941 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3942 //   init = (Initialize alloc.Control alloc.Memory rawoop
3943 //                      rawstore1 rawstore2)
3944 //
3945 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3946                                     PhaseTransform* phase, bool can_reshape) {
3947   assert(stores_are_sane(phase), "");
3948 
3949   if (start < 0)  return NULL;
3950   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3951 
3952   Compile* C = phase->C;
3953   int size_in_bytes = st->memory_size();
3954   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3955   if (i == 0)  return NULL;     // bail out
3956   Node* prev_mem = NULL;        // raw memory for the captured store
3957   if (i > 0) {
3958     prev_mem = in(i);           // there is a pre-existing store under this one
3959     set_req(i, C->top());       // temporarily disconnect it
3960     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3961   } else {
3962     i = -i;                     // no pre-existing store
3963     prev_mem = zero_memory();   // a slice of the newly allocated object
3964     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3965       set_req(--i, C->top());   // reuse this edge; it has been folded away
3966     else
3967       ins_req(i, C->top());     // build a new edge
3968   }
3969   Node* new_st = st->clone();
3970   new_st->set_req(MemNode::Control, in(Control));
3971   new_st->set_req(MemNode::Memory,  prev_mem);
3972   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3973   new_st = phase->transform(new_st);
3974 
3975   // At this point, new_st might have swallowed a pre-existing store
3976   // at the same offset, or perhaps new_st might have disappeared,
3977   // if it redundantly stored the same value (or zero to fresh memory).
3978 
3979   // In any case, wire it in:
3980   phase->igvn_rehash_node_delayed(this);
3981   set_req(i, new_st);
3982 
3983   // The caller may now kill the old guy.
3984   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3985   assert(check_st == new_st || check_st == NULL, "must be findable");
3986   assert(!is_complete(), "");
3987   return new_st;
3988 }
3989 
3990 static bool store_constant(jlong* tiles, int num_tiles,
3991                            intptr_t st_off, int st_size,
3992                            jlong con) {
3993   if ((st_off & (st_size-1)) != 0)
3994     return false;               // strange store offset (assume size==2**N)
3995   address addr = (address)tiles + st_off;
3996   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3997   switch (st_size) {
3998   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
3999   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
4000   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
4001   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
4002   default: return false;        // strange store size (detect size!=2**N here)
4003   }
4004   return true;                  // return success to caller
4005 }
4006 
4007 // Coalesce subword constants into int constants and possibly
4008 // into long constants.  The goal, if the CPU permits,
4009 // is to initialize the object with a small number of 64-bit tiles.
4010 // Also, convert floating-point constants to bit patterns.
4011 // Non-constants are not relevant to this pass.
4012 //
4013 // In terms of the running example on InitializeNode::InitializeNode
4014 // and InitializeNode::capture_store, here is the transformation
4015 // of rawstore1 and rawstore2 into rawstore12:
4016 //   alloc = (Allocate ...)
4017 //   rawoop = alloc.RawAddress
4018 //   tile12 = 0x00010002
4019 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4020 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4021 //
4022 void
4023 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4024                                         Node* size_in_bytes,
4025                                         PhaseGVN* phase) {
4026   Compile* C = phase->C;
4027 
4028   assert(stores_are_sane(phase), "");
4029   // Note:  After this pass, they are not completely sane,
4030   // since there may be some overlaps.
4031 
4032   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4033 
4034   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4035   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4036   size_limit = MIN2(size_limit, ti_limit);
4037   size_limit = align_up(size_limit, BytesPerLong);
4038   int num_tiles = size_limit / BytesPerLong;
4039 
4040   // allocate space for the tile map:
4041   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4042   jlong  tiles_buf[small_len];
4043   Node*  nodes_buf[small_len];
4044   jlong  inits_buf[small_len];
4045   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4046                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4047   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4048                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4049   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4050                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4051   // tiles: exact bitwise model of all primitive constants
4052   // nodes: last constant-storing node subsumed into the tiles model
4053   // inits: which bytes (in each tile) are touched by any initializations
4054 
4055   //// Pass A: Fill in the tile model with any relevant stores.
4056 
4057   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4058   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4059   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4060   Node* zmem = zero_memory(); // initially zero memory state
4061   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4062     Node* st = in(i);
4063     intptr_t st_off = get_store_offset(st, phase);
4064 
4065     // Figure out the store's offset and constant value:
4066     if (st_off < header_size)             continue; //skip (ignore header)
4067     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
4068     int st_size = st->as_Store()->memory_size();
4069     if (st_off + st_size > size_limit)    break;
4070 
4071     // Record which bytes are touched, whether by constant or not.
4072     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4073       continue;                 // skip (strange store size)
4074 
4075     const Type* val = phase->type(st->in(MemNode::ValueIn));
4076     if (!val->singleton())                continue; //skip (non-con store)
4077     BasicType type = val->basic_type();
4078 
4079     jlong con = 0;
4080     switch (type) {
4081     case T_INT:    con = val->is_int()->get_con();  break;
4082     case T_LONG:   con = val->is_long()->get_con(); break;
4083     case T_FLOAT:  con = jint_cast(val->getf());    break;
4084     case T_DOUBLE: con = jlong_cast(val->getd());   break;
4085     default:                              continue; //skip (odd store type)
4086     }
4087 
4088     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4089         st->Opcode() == Op_StoreL) {
4090       continue;                 // This StoreL is already optimal.
4091     }
4092 
4093     // Store down the constant.
4094     store_constant(tiles, num_tiles, st_off, st_size, con);
4095 
4096     intptr_t j = st_off >> LogBytesPerLong;
4097 
4098     if (type == T_INT && st_size == BytesPerInt
4099         && (st_off & BytesPerInt) == BytesPerInt) {
4100       jlong lcon = tiles[j];
4101       if (!Matcher::isSimpleConstant64(lcon) &&
4102           st->Opcode() == Op_StoreI) {
4103         // This StoreI is already optimal by itself.
4104         jint* intcon = (jint*) &tiles[j];
4105         intcon[1] = 0;  // undo the store_constant()
4106 
4107         // If the previous store is also optimal by itself, back up and
4108         // undo the action of the previous loop iteration... if we can.
4109         // But if we can't, just let the previous half take care of itself.
4110         st = nodes[j];
4111         st_off -= BytesPerInt;
4112         con = intcon[0];
4113         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
4114           assert(st_off >= header_size, "still ignoring header");
4115           assert(get_store_offset(st, phase) == st_off, "must be");
4116           assert(in(i-1) == zmem, "must be");
4117           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4118           assert(con == tcon->is_int()->get_con(), "must be");
4119           // Undo the effects of the previous loop trip, which swallowed st:
4120           intcon[0] = 0;        // undo store_constant()
4121           set_req(i-1, st);     // undo set_req(i, zmem)
4122           nodes[j] = NULL;      // undo nodes[j] = st
4123           --old_subword;        // undo ++old_subword
4124         }
4125         continue;               // This StoreI is already optimal.
4126       }
4127     }
4128 
4129     // This store is not needed.
4130     set_req(i, zmem);
4131     nodes[j] = st;              // record for the moment
4132     if (st_size < BytesPerLong) // something has changed
4133           ++old_subword;        // includes int/float, but who's counting...
4134     else  ++old_long;
4135   }
4136 
4137   if ((old_subword + old_long) == 0)
4138     return;                     // nothing more to do
4139 
4140   //// Pass B: Convert any non-zero tiles into optimal constant stores.
4141   // Be sure to insert them before overlapping non-constant stores.
4142   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
4143   for (int j = 0; j < num_tiles; j++) {
4144     jlong con  = tiles[j];
4145     jlong init = inits[j];
4146     if (con == 0)  continue;
4147     jint con0,  con1;           // split the constant, address-wise
4148     jint init0, init1;          // split the init map, address-wise
4149     { union { jlong con; jint intcon[2]; } u;
4150       u.con = con;
4151       con0  = u.intcon[0];
4152       con1  = u.intcon[1];
4153       u.con = init;
4154       init0 = u.intcon[0];
4155       init1 = u.intcon[1];
4156     }
4157 
4158     Node* old = nodes[j];
4159     assert(old != NULL, "need the prior store");
4160     intptr_t offset = (j * BytesPerLong);
4161 
4162     bool split = !Matcher::isSimpleConstant64(con);
4163 
4164     if (offset < header_size) {
4165       assert(offset + BytesPerInt >= header_size, "second int counts");
4166       assert(*(jint*)&tiles[j] == 0, "junk in header");
4167       split = true;             // only the second word counts
4168       // Example:  int a[] = { 42 ... }
4169     } else if (con0 == 0 && init0 == -1) {
4170       split = true;             // first word is covered by full inits
4171       // Example:  int a[] = { ... foo(), 42 ... }
4172     } else if (con1 == 0 && init1 == -1) {
4173       split = true;             // second word is covered by full inits
4174       // Example:  int a[] = { ... 42, foo() ... }
4175     }
4176 
4177     // Here's a case where init0 is neither 0 nor -1:
4178     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
4179     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4180     // In this case the tile is not split; it is (jlong)42.
4181     // The big tile is stored down, and then the foo() value is inserted.
4182     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4183 
4184     Node* ctl = old->in(MemNode::Control);
4185     Node* adr = make_raw_address(offset, phase);
4186     const TypePtr* atp = TypeRawPtr::BOTTOM;
4187 
4188     // One or two coalesced stores to plop down.
4189     Node*    st[2];
4190     intptr_t off[2];
4191     int  nst = 0;
4192     if (!split) {
4193       ++new_long;
4194       off[nst] = offset;
4195       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4196                                   phase->longcon(con), T_LONG, MemNode::unordered);
4197     } else {
4198       // Omit either if it is a zero.
4199       if (con0 != 0) {
4200         ++new_int;
4201         off[nst]  = offset;
4202         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4203                                     phase->intcon(con0), T_INT, MemNode::unordered);
4204       }
4205       if (con1 != 0) {
4206         ++new_int;
4207         offset += BytesPerInt;
4208         adr = make_raw_address(offset, phase);
4209         off[nst]  = offset;
4210         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4211                                     phase->intcon(con1), T_INT, MemNode::unordered);
4212       }
4213     }
4214 
4215     // Insert second store first, then the first before the second.
4216     // Insert each one just before any overlapping non-constant stores.
4217     while (nst > 0) {
4218       Node* st1 = st[--nst];
4219       C->copy_node_notes_to(st1, old);
4220       st1 = phase->transform(st1);
4221       offset = off[nst];
4222       assert(offset >= header_size, "do not smash header");
4223       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4224       guarantee(ins_idx != 0, "must re-insert constant store");
4225       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
4226       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4227         set_req(--ins_idx, st1);
4228       else
4229         ins_req(ins_idx, st1);
4230     }
4231   }
4232 
4233   if (PrintCompilation && WizardMode)
4234     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4235                   old_subword, old_long, new_int, new_long);
4236   if (C->log() != NULL)
4237     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4238                    old_subword, old_long, new_int, new_long);
4239 
4240   // Clean up any remaining occurrences of zmem:
4241   remove_extra_zeroes();
4242 }
4243 
4244 // Explore forward from in(start) to find the first fully initialized
4245 // word, and return its offset.  Skip groups of subword stores which
4246 // together initialize full words.  If in(start) is itself part of a
4247 // fully initialized word, return the offset of in(start).  If there
4248 // are no following full-word stores, or if something is fishy, return
4249 // a negative value.
4250 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4251   int       int_map = 0;
4252   intptr_t  int_map_off = 0;
4253   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
4254 
4255   for (uint i = start, limit = req(); i < limit; i++) {
4256     Node* st = in(i);
4257 
4258     intptr_t st_off = get_store_offset(st, phase);
4259     if (st_off < 0)  break;  // return conservative answer
4260 
4261     int st_size = st->as_Store()->memory_size();
4262     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4263       return st_off;            // we found a complete word init
4264     }
4265 
4266     // update the map:
4267 
4268     intptr_t this_int_off = align_down(st_off, BytesPerInt);
4269     if (this_int_off != int_map_off) {
4270       // reset the map:
4271       int_map = 0;
4272       int_map_off = this_int_off;
4273     }
4274 
4275     int subword_off = st_off - this_int_off;
4276     int_map |= right_n_bits(st_size) << subword_off;
4277     if ((int_map & FULL_MAP) == FULL_MAP) {
4278       return this_int_off;      // we found a complete word init
4279     }
4280 
4281     // Did this store hit or cross the word boundary?
4282     intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4283     if (next_int_off == this_int_off + BytesPerInt) {
4284       // We passed the current int, without fully initializing it.
4285       int_map_off = next_int_off;
4286       int_map >>= BytesPerInt;
4287     } else if (next_int_off > this_int_off + BytesPerInt) {
4288       // We passed the current and next int.
4289       return this_int_off + BytesPerInt;
4290     }
4291   }
4292 
4293   return -1;
4294 }
4295 
4296 
4297 // Called when the associated AllocateNode is expanded into CFG.
4298 // At this point, we may perform additional optimizations.
4299 // Linearize the stores by ascending offset, to make memory
4300 // activity as coherent as possible.
4301 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4302                                       intptr_t header_size,
4303                                       Node* size_in_bytes,
4304                                       PhaseGVN* phase) {
4305   assert(!is_complete(), "not already complete");
4306   assert(stores_are_sane(phase), "");
4307   assert(allocation() != NULL, "must be present");
4308 
4309   remove_extra_zeroes();
4310 
4311   if (ReduceFieldZeroing || ReduceBulkZeroing)
4312     // reduce instruction count for common initialization patterns
4313     coalesce_subword_stores(header_size, size_in_bytes, phase);
4314 
4315   Node* zmem = zero_memory();   // initially zero memory state
4316   Node* inits = zmem;           // accumulating a linearized chain of inits
4317   #ifdef ASSERT
4318   intptr_t first_offset = allocation()->minimum_header_size();
4319   intptr_t last_init_off = first_offset;  // previous init offset
4320   intptr_t last_init_end = first_offset;  // previous init offset+size
4321   intptr_t last_tile_end = first_offset;  // previous tile offset+size
4322   #endif
4323   intptr_t zeroes_done = header_size;
4324 
4325   bool do_zeroing = true;       // we might give up if inits are very sparse
4326   int  big_init_gaps = 0;       // how many large gaps have we seen?
4327 
4328   if (UseTLAB && ZeroTLAB)  do_zeroing = false;
4329   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
4330 
4331   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4332     Node* st = in(i);
4333     intptr_t st_off = get_store_offset(st, phase);
4334     if (st_off < 0)
4335       break;                    // unknown junk in the inits
4336     if (st->in(MemNode::Memory) != zmem)
4337       break;                    // complicated store chains somehow in list
4338 
4339     int st_size = st->as_Store()->memory_size();
4340     intptr_t next_init_off = st_off + st_size;
4341 
4342     if (do_zeroing && zeroes_done < next_init_off) {
4343       // See if this store needs a zero before it or under it.
4344       intptr_t zeroes_needed = st_off;
4345 
4346       if (st_size < BytesPerInt) {
4347         // Look for subword stores which only partially initialize words.
4348         // If we find some, we must lay down some word-level zeroes first,
4349         // underneath the subword stores.
4350         //
4351         // Examples:
4352         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
4353         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
4354         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
4355         //
4356         // Note:  coalesce_subword_stores may have already done this,
4357         // if it was prompted by constant non-zero subword initializers.
4358         // But this case can still arise with non-constant stores.
4359 
4360         intptr_t next_full_store = find_next_fullword_store(i, phase);
4361 
4362         // In the examples above:
4363         //   in(i)          p   q   r   s     x   y     z
4364         //   st_off        12  13  14  15    12  13    14
4365         //   st_size        1   1   1   1     1   1     1
4366         //   next_full_s.  12  16  16  16    16  16    16
4367         //   z's_done      12  16  16  16    12  16    12
4368         //   z's_needed    12  16  16  16    16  16    16
4369         //   zsize          0   0   0   0     4   0     4
4370         if (next_full_store < 0) {
4371           // Conservative tack:  Zero to end of current word.
4372           zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4373         } else {
4374           // Zero to beginning of next fully initialized word.
4375           // Or, don't zero at all, if we are already in that word.
4376           assert(next_full_store >= zeroes_needed, "must go forward");
4377           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4378           zeroes_needed = next_full_store;
4379         }
4380       }
4381 
4382       if (zeroes_needed > zeroes_done) {
4383         intptr_t zsize = zeroes_needed - zeroes_done;
4384         // Do some incremental zeroing on rawmem, in parallel with inits.
4385         zeroes_done = align_down(zeroes_done, BytesPerInt);
4386         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4387                                               allocation()->in(AllocateNode::DefaultValue),
4388                                               allocation()->in(AllocateNode::RawDefaultValue),
4389                                               zeroes_done, zeroes_needed,
4390                                               phase);
4391         zeroes_done = zeroes_needed;
4392         if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4393           do_zeroing = false;   // leave the hole, next time
4394       }
4395     }
4396 
4397     // Collect the store and move on:
4398     st->set_req(MemNode::Memory, inits);
4399     inits = st;                 // put it on the linearized chain
4400     set_req(i, zmem);           // unhook from previous position
4401 
4402     if (zeroes_done == st_off)
4403       zeroes_done = next_init_off;
4404 
4405     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4406 
4407     #ifdef ASSERT
4408     // Various order invariants.  Weaker than stores_are_sane because
4409     // a large constant tile can be filled in by smaller non-constant stores.
4410     assert(st_off >= last_init_off, "inits do not reverse");
4411     last_init_off = st_off;
4412     const Type* val = NULL;
4413     if (st_size >= BytesPerInt &&
4414         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4415         (int)val->basic_type() < (int)T_OBJECT) {
4416       assert(st_off >= last_tile_end, "tiles do not overlap");
4417       assert(st_off >= last_init_end, "tiles do not overwrite inits");
4418       last_tile_end = MAX2(last_tile_end, next_init_off);
4419     } else {
4420       intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4421       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4422       assert(st_off      >= last_init_end, "inits do not overlap");
4423       last_init_end = next_init_off;  // it's a non-tile
4424     }
4425     #endif //ASSERT
4426   }
4427 
4428   remove_extra_zeroes();        // clear out all the zmems left over
4429   add_req(inits);
4430 
4431   if (!(UseTLAB && ZeroTLAB)) {
4432     // If anything remains to be zeroed, zero it all now.
4433     zeroes_done = align_down(zeroes_done, BytesPerInt);
4434     // if it is the last unused 4 bytes of an instance, forget about it
4435     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4436     if (zeroes_done + BytesPerLong >= size_limit) {
4437       AllocateNode* alloc = allocation();
4438       assert(alloc != NULL, "must be present");
4439       if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4440         Node* klass_node = alloc->in(AllocateNode::KlassNode);
4441         ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4442         if (zeroes_done == k->layout_helper())
4443           zeroes_done = size_limit;
4444       }
4445     }
4446     if (zeroes_done < size_limit) {
4447       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4448                                             allocation()->in(AllocateNode::DefaultValue),
4449                                             allocation()->in(AllocateNode::RawDefaultValue),
4450                                             zeroes_done, size_in_bytes, phase);
4451     }
4452   }
4453 
4454   set_complete(phase);
4455   return rawmem;
4456 }
4457 
4458 
4459 #ifdef ASSERT
4460 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4461   if (is_complete())
4462     return true;                // stores could be anything at this point
4463   assert(allocation() != NULL, "must be present");
4464   intptr_t last_off = allocation()->minimum_header_size();
4465   for (uint i = InitializeNode::RawStores; i < req(); i++) {
4466     Node* st = in(i);
4467     intptr_t st_off = get_store_offset(st, phase);
4468     if (st_off < 0)  continue;  // ignore dead garbage
4469     if (last_off > st_off) {
4470       tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4471       this->dump(2);
4472       assert(false, "ascending store offsets");
4473       return false;
4474     }
4475     last_off = st_off + st->as_Store()->memory_size();
4476   }
4477   return true;
4478 }
4479 #endif //ASSERT
4480 
4481 
4482 
4483 
4484 //============================MergeMemNode=====================================
4485 //
4486 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
4487 // contributing store or call operations.  Each contributor provides the memory
4488 // state for a particular "alias type" (see Compile::alias_type).  For example,
4489 // if a MergeMem has an input X for alias category #6, then any memory reference
4490 // to alias category #6 may use X as its memory state input, as an exact equivalent
4491 // to using the MergeMem as a whole.
4492 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4493 //
4494 // (Here, the <N> notation gives the index of the relevant adr_type.)
4495 //
4496 // In one special case (and more cases in the future), alias categories overlap.
4497 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4498 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
4499 // it is exactly equivalent to that state W:
4500 //   MergeMem(<Bot>: W) <==> W
4501 //
4502 // Usually, the merge has more than one input.  In that case, where inputs
4503 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4504 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4505 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4506 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4507 //
4508 // A merge can take a "wide" memory state as one of its narrow inputs.
4509 // This simply means that the merge observes out only the relevant parts of
4510 // the wide input.  That is, wide memory states arriving at narrow merge inputs
4511 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
4512 //
4513 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4514 // and that memory slices "leak through":
4515 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4516 //
4517 // But, in such a cascade, repeated memory slices can "block the leak":
4518 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4519 //
4520 // In the last example, Y is not part of the combined memory state of the
4521 // outermost MergeMem.  The system must, of course, prevent unschedulable
4522 // memory states from arising, so you can be sure that the state Y is somehow
4523 // a precursor to state Y'.
4524 //
4525 //
4526 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4527 // of each MergeMemNode array are exactly the numerical alias indexes, including
4528 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
4529 // Compile::alias_type (and kin) produce and manage these indexes.
4530 //
4531 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4532 // (Note that this provides quick access to the top node inside MergeMem methods,
4533 // without the need to reach out via TLS to Compile::current.)
4534 //
4535 // As a consequence of what was just described, a MergeMem that represents a full
4536 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4537 // containing all alias categories.
4538 //
4539 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4540 //
4541 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4542 // a memory state for the alias type <N>, or else the top node, meaning that
4543 // there is no particular input for that alias type.  Note that the length of
4544 // a MergeMem is variable, and may be extended at any time to accommodate new
4545 // memory states at larger alias indexes.  When merges grow, they are of course
4546 // filled with "top" in the unused in() positions.
4547 //
4548 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4549 // (Top was chosen because it works smoothly with passes like GCM.)
4550 //
4551 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
4552 // the type of random VM bits like TLS references.)  Since it is always the
4553 // first non-Bot memory slice, some low-level loops use it to initialize an
4554 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
4555 //
4556 //
4557 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
4558 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
4559 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4560 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
4561 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
4562 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4563 //
4564 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4565 // really that different from the other memory inputs.  An abbreviation called
4566 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4567 //
4568 //
4569 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
4570 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
4571 // that "emerges though" the base memory will be marked as excluding the alias types
4572 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4573 //
4574 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4575 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4576 //
4577 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4578 // (It is currently unimplemented.)  As you can see, the resulting merge is
4579 // actually a disjoint union of memory states, rather than an overlay.
4580 //
4581 
4582 //------------------------------MergeMemNode-----------------------------------
4583 Node* MergeMemNode::make_empty_memory() {
4584   Node* empty_memory = (Node*) Compile::current()->top();
4585   assert(empty_memory->is_top(), "correct sentinel identity");
4586   return empty_memory;
4587 }
4588 
4589 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4590   init_class_id(Class_MergeMem);
4591   // all inputs are nullified in Node::Node(int)
4592   // set_input(0, NULL);  // no control input
4593 
4594   // Initialize the edges uniformly to top, for starters.
4595   Node* empty_mem = make_empty_memory();
4596   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4597     init_req(i,empty_mem);
4598   }
4599   assert(empty_memory() == empty_mem, "");
4600 
4601   if( new_base != NULL && new_base->is_MergeMem() ) {
4602     MergeMemNode* mdef = new_base->as_MergeMem();
4603     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4604     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4605       mms.set_memory(mms.memory2());
4606     }
4607     assert(base_memory() == mdef->base_memory(), "");
4608   } else {
4609     set_base_memory(new_base);
4610   }
4611 }
4612 
4613 // Make a new, untransformed MergeMem with the same base as 'mem'.
4614 // If mem is itself a MergeMem, populate the result with the same edges.
4615 MergeMemNode* MergeMemNode::make(Node* mem) {
4616   return new MergeMemNode(mem);
4617 }
4618 
4619 //------------------------------cmp--------------------------------------------
4620 uint MergeMemNode::hash() const { return NO_HASH; }
4621 bool MergeMemNode::cmp( const Node &n ) const {
4622   return (&n == this);          // Always fail except on self
4623 }
4624 
4625 //------------------------------Identity---------------------------------------
4626 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4627   // Identity if this merge point does not record any interesting memory
4628   // disambiguations.
4629   Node* base_mem = base_memory();
4630   Node* empty_mem = empty_memory();
4631   if (base_mem != empty_mem) {  // Memory path is not dead?
4632     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4633       Node* mem = in(i);
4634       if (mem != empty_mem && mem != base_mem) {
4635         return this;            // Many memory splits; no change
4636       }
4637     }
4638   }
4639   return base_mem;              // No memory splits; ID on the one true input
4640 }
4641 
4642 //------------------------------Ideal------------------------------------------
4643 // This method is invoked recursively on chains of MergeMem nodes
4644 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4645   // Remove chain'd MergeMems
4646   //
4647   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4648   // relative to the "in(Bot)".  Since we are patching both at the same time,
4649   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4650   // but rewrite each "in(i)" relative to the new "in(Bot)".
4651   Node *progress = NULL;
4652 
4653 
4654   Node* old_base = base_memory();
4655   Node* empty_mem = empty_memory();
4656   if (old_base == empty_mem)
4657     return NULL; // Dead memory path.
4658 
4659   MergeMemNode* old_mbase;
4660   if (old_base != NULL && old_base->is_MergeMem())
4661     old_mbase = old_base->as_MergeMem();
4662   else
4663     old_mbase = NULL;
4664   Node* new_base = old_base;
4665 
4666   // simplify stacked MergeMems in base memory
4667   if (old_mbase)  new_base = old_mbase->base_memory();
4668 
4669   // the base memory might contribute new slices beyond my req()
4670   if (old_mbase)  grow_to_match(old_mbase);
4671 
4672   // Look carefully at the base node if it is a phi.
4673   PhiNode* phi_base;
4674   if (new_base != NULL && new_base->is_Phi())
4675     phi_base = new_base->as_Phi();
4676   else
4677     phi_base = NULL;
4678 
4679   Node*    phi_reg = NULL;
4680   uint     phi_len = (uint)-1;
4681   if (phi_base != NULL && !phi_base->is_copy()) {
4682     // do not examine phi if degraded to a copy
4683     phi_reg = phi_base->region();
4684     phi_len = phi_base->req();
4685     // see if the phi is unfinished
4686     for (uint i = 1; i < phi_len; i++) {
4687       if (phi_base->in(i) == NULL) {
4688         // incomplete phi; do not look at it yet!
4689         phi_reg = NULL;
4690         phi_len = (uint)-1;
4691         break;
4692       }
4693     }
4694   }
4695 
4696   // Note:  We do not call verify_sparse on entry, because inputs
4697   // can normalize to the base_memory via subsume_node or similar
4698   // mechanisms.  This method repairs that damage.
4699 
4700   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4701 
4702   // Look at each slice.
4703   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4704     Node* old_in = in(i);
4705     // calculate the old memory value
4706     Node* old_mem = old_in;
4707     if (old_mem == empty_mem)  old_mem = old_base;
4708     assert(old_mem == memory_at(i), "");
4709 
4710     // maybe update (reslice) the old memory value
4711 
4712     // simplify stacked MergeMems
4713     Node* new_mem = old_mem;
4714     MergeMemNode* old_mmem;
4715     if (old_mem != NULL && old_mem->is_MergeMem())
4716       old_mmem = old_mem->as_MergeMem();
4717     else
4718       old_mmem = NULL;
4719     if (old_mmem == this) {
4720       // This can happen if loops break up and safepoints disappear.
4721       // A merge of BotPtr (default) with a RawPtr memory derived from a
4722       // safepoint can be rewritten to a merge of the same BotPtr with
4723       // the BotPtr phi coming into the loop.  If that phi disappears
4724       // also, we can end up with a self-loop of the mergemem.
4725       // In general, if loops degenerate and memory effects disappear,
4726       // a mergemem can be left looking at itself.  This simply means
4727       // that the mergemem's default should be used, since there is
4728       // no longer any apparent effect on this slice.
4729       // Note: If a memory slice is a MergeMem cycle, it is unreachable
4730       //       from start.  Update the input to TOP.
4731       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4732     }
4733     else if (old_mmem != NULL) {
4734       new_mem = old_mmem->memory_at(i);
4735     }
4736     // else preceding memory was not a MergeMem
4737 
4738     // replace equivalent phis (unfortunately, they do not GVN together)
4739     if (new_mem != NULL && new_mem != new_base &&
4740         new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4741       if (new_mem->is_Phi()) {
4742         PhiNode* phi_mem = new_mem->as_Phi();
4743         for (uint i = 1; i < phi_len; i++) {
4744           if (phi_base->in(i) != phi_mem->in(i)) {
4745             phi_mem = NULL;
4746             break;
4747           }
4748         }
4749         if (phi_mem != NULL) {
4750           // equivalent phi nodes; revert to the def
4751           new_mem = new_base;
4752         }
4753       }
4754     }
4755 
4756     // maybe store down a new value
4757     Node* new_in = new_mem;
4758     if (new_in == new_base)  new_in = empty_mem;
4759 
4760     if (new_in != old_in) {
4761       // Warning:  Do not combine this "if" with the previous "if"
4762       // A memory slice might have be be rewritten even if it is semantically
4763       // unchanged, if the base_memory value has changed.
4764       set_req(i, new_in);
4765       progress = this;          // Report progress
4766     }
4767   }
4768 
4769   if (new_base != old_base) {
4770     set_req(Compile::AliasIdxBot, new_base);
4771     // Don't use set_base_memory(new_base), because we need to update du.
4772     assert(base_memory() == new_base, "");
4773     progress = this;
4774   }
4775 
4776   if( base_memory() == this ) {
4777     // a self cycle indicates this memory path is dead
4778     set_req(Compile::AliasIdxBot, empty_mem);
4779   }
4780 
4781   // Resolve external cycles by calling Ideal on a MergeMem base_memory
4782   // Recursion must occur after the self cycle check above
4783   if( base_memory()->is_MergeMem() ) {
4784     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4785     Node *m = phase->transform(new_mbase);  // Rollup any cycles
4786     if( m != NULL &&
4787         (m->is_top() ||
4788          (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4789       // propagate rollup of dead cycle to self
4790       set_req(Compile::AliasIdxBot, empty_mem);
4791     }
4792   }
4793 
4794   if( base_memory() == empty_mem ) {
4795     progress = this;
4796     // Cut inputs during Parse phase only.
4797     // During Optimize phase a dead MergeMem node will be subsumed by Top.
4798     if( !can_reshape ) {
4799       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4800         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4801       }
4802     }
4803   }
4804 
4805   if( !progress && base_memory()->is_Phi() && can_reshape ) {
4806     // Check if PhiNode::Ideal's "Split phis through memory merges"
4807     // transform should be attempted. Look for this->phi->this cycle.
4808     uint merge_width = req();
4809     if (merge_width > Compile::AliasIdxRaw) {
4810       PhiNode* phi = base_memory()->as_Phi();
4811       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4812         if (phi->in(i) == this) {
4813           phase->is_IterGVN()->_worklist.push(phi);
4814           break;
4815         }
4816       }
4817     }
4818   }
4819 
4820   assert(progress || verify_sparse(), "please, no dups of base");
4821   return progress;
4822 }
4823 
4824 //-------------------------set_base_memory-------------------------------------
4825 void MergeMemNode::set_base_memory(Node *new_base) {
4826   Node* empty_mem = empty_memory();
4827   set_req(Compile::AliasIdxBot, new_base);
4828   assert(memory_at(req()) == new_base, "must set default memory");
4829   // Clear out other occurrences of new_base:
4830   if (new_base != empty_mem) {
4831     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4832       if (in(i) == new_base)  set_req(i, empty_mem);
4833     }
4834   }
4835 }
4836 
4837 //------------------------------out_RegMask------------------------------------
4838 const RegMask &MergeMemNode::out_RegMask() const {
4839   return RegMask::Empty;
4840 }
4841 
4842 //------------------------------dump_spec--------------------------------------
4843 #ifndef PRODUCT
4844 void MergeMemNode::dump_spec(outputStream *st) const {
4845   st->print(" {");
4846   Node* base_mem = base_memory();
4847   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4848     Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4849     if (mem == base_mem) { st->print(" -"); continue; }
4850     st->print( " N%d:", mem->_idx );
4851     Compile::current()->get_adr_type(i)->dump_on(st);
4852   }
4853   st->print(" }");
4854 }
4855 #endif // !PRODUCT
4856 
4857 
4858 #ifdef ASSERT
4859 static bool might_be_same(Node* a, Node* b) {
4860   if (a == b)  return true;
4861   if (!(a->is_Phi() || b->is_Phi()))  return false;
4862   // phis shift around during optimization
4863   return true;  // pretty stupid...
4864 }
4865 
4866 // verify a narrow slice (either incoming or outgoing)
4867 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4868   if (!VerifyAliases)                return;  // don't bother to verify unless requested
4869   if (VMError::is_error_reported())  return;  // muzzle asserts when debugging an error
4870   if (Node::in_dump())               return;  // muzzle asserts when printing
4871   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4872   assert(n != NULL, "");
4873   // Elide intervening MergeMem's
4874   while (n->is_MergeMem()) {
4875     n = n->as_MergeMem()->memory_at(alias_idx);
4876   }
4877   Compile* C = Compile::current();
4878   const TypePtr* n_adr_type = n->adr_type();
4879   if (n == m->empty_memory()) {
4880     // Implicit copy of base_memory()
4881   } else if (n_adr_type != TypePtr::BOTTOM) {
4882     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4883     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4884   } else {
4885     // A few places like make_runtime_call "know" that VM calls are narrow,
4886     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4887     bool expected_wide_mem = false;
4888     if (n == m->base_memory()) {
4889       expected_wide_mem = true;
4890     } else if (alias_idx == Compile::AliasIdxRaw ||
4891                n == m->memory_at(Compile::AliasIdxRaw)) {
4892       expected_wide_mem = true;
4893     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4894       // memory can "leak through" calls on channels that
4895       // are write-once.  Allow this also.
4896       expected_wide_mem = true;
4897     }
4898     assert(expected_wide_mem, "expected narrow slice replacement");
4899   }
4900 }
4901 #else // !ASSERT
4902 #define verify_memory_slice(m,i,n) (void)(0)  // PRODUCT version is no-op
4903 #endif
4904 
4905 
4906 //-----------------------------memory_at---------------------------------------
4907 Node* MergeMemNode::memory_at(uint alias_idx) const {
4908   assert(alias_idx >= Compile::AliasIdxRaw ||
4909          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4910          "must avoid base_memory and AliasIdxTop");
4911 
4912   // Otherwise, it is a narrow slice.
4913   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4914   Compile *C = Compile::current();
4915   if (is_empty_memory(n)) {
4916     // the array is sparse; empty slots are the "top" node
4917     n = base_memory();
4918     assert(Node::in_dump()
4919            || n == NULL || n->bottom_type() == Type::TOP
4920            || n->adr_type() == NULL // address is TOP
4921            || n->adr_type() == TypePtr::BOTTOM
4922            || n->adr_type() == TypeRawPtr::BOTTOM
4923            || Compile::current()->AliasLevel() == 0,
4924            "must be a wide memory");
4925     // AliasLevel == 0 if we are organizing the memory states manually.
4926     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4927   } else {
4928     // make sure the stored slice is sane
4929     #ifdef ASSERT
4930     if (VMError::is_error_reported() || Node::in_dump()) {
4931     } else if (might_be_same(n, base_memory())) {
4932       // Give it a pass:  It is a mostly harmless repetition of the base.
4933       // This can arise normally from node subsumption during optimization.
4934     } else {
4935       verify_memory_slice(this, alias_idx, n);
4936     }
4937     #endif
4938   }
4939   return n;
4940 }
4941 
4942 //---------------------------set_memory_at-------------------------------------
4943 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4944   verify_memory_slice(this, alias_idx, n);
4945   Node* empty_mem = empty_memory();
4946   if (n == base_memory())  n = empty_mem;  // collapse default
4947   uint need_req = alias_idx+1;
4948   if (req() < need_req) {
4949     if (n == empty_mem)  return;  // already the default, so do not grow me
4950     // grow the sparse array
4951     do {
4952       add_req(empty_mem);
4953     } while (req() < need_req);
4954   }
4955   set_req( alias_idx, n );
4956 }
4957 
4958 
4959 
4960 //--------------------------iteration_setup------------------------------------
4961 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4962   if (other != NULL) {
4963     grow_to_match(other);
4964     // invariant:  the finite support of mm2 is within mm->req()
4965     #ifdef ASSERT
4966     for (uint i = req(); i < other->req(); i++) {
4967       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4968     }
4969     #endif
4970   }
4971   // Replace spurious copies of base_memory by top.
4972   Node* base_mem = base_memory();
4973   if (base_mem != NULL && !base_mem->is_top()) {
4974     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4975       if (in(i) == base_mem)
4976         set_req(i, empty_memory());
4977     }
4978   }
4979 }
4980 
4981 //---------------------------grow_to_match-------------------------------------
4982 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4983   Node* empty_mem = empty_memory();
4984   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4985   // look for the finite support of the other memory
4986   for (uint i = other->req(); --i >= req(); ) {
4987     if (other->in(i) != empty_mem) {
4988       uint new_len = i+1;
4989       while (req() < new_len)  add_req(empty_mem);
4990       break;
4991     }
4992   }
4993 }
4994 
4995 //---------------------------verify_sparse-------------------------------------
4996 #ifndef PRODUCT
4997 bool MergeMemNode::verify_sparse() const {
4998   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4999   Node* base_mem = base_memory();
5000   // The following can happen in degenerate cases, since empty==top.
5001   if (is_empty_memory(base_mem))  return true;
5002   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5003     assert(in(i) != NULL, "sane slice");
5004     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
5005   }
5006   return true;
5007 }
5008 
5009 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5010   Node* n;
5011   n = mm->in(idx);
5012   if (mem == n)  return true;  // might be empty_memory()
5013   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5014   if (mem == n)  return true;
5015   while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
5016     if (mem == n)  return true;
5017     if (n == NULL)  break;
5018   }
5019   return false;
5020 }
5021 #endif // !PRODUCT