1 /*
   2  * Copyright (c) 1997, 2021, 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/javaClasses.hpp"
  27 #include "compiler/compileLog.hpp"
  28 #include "gc/shared/barrierSet.hpp"
  29 #include "gc/shared/c2/barrierSetC2.hpp"
  30 #include "gc/shared/tlab_globals.hpp"
  31 #include "memory/allocation.inline.hpp"
  32 #include "memory/resourceArea.hpp"
  33 #include "oops/objArrayKlass.hpp"
  34 #include "opto/addnode.hpp"
  35 #include "opto/arraycopynode.hpp"
  36 #include "opto/cfgnode.hpp"
  37 #include "opto/regalloc.hpp"
  38 #include "opto/compile.hpp"
  39 #include "opto/connode.hpp"
  40 #include "opto/convertnode.hpp"
  41 #include "opto/loopnode.hpp"
  42 #include "opto/machnode.hpp"
  43 #include "opto/matcher.hpp"
  44 #include "opto/memnode.hpp"
  45 #include "opto/mulnode.hpp"
  46 #include "opto/narrowptrnode.hpp"
  47 #include "opto/phaseX.hpp"
  48 #include "opto/regmask.hpp"
  49 #include "opto/rootnode.hpp"
  50 #include "opto/vectornode.hpp"
  51 #include "utilities/align.hpp"
  52 #include "utilities/copy.hpp"
  53 #include "utilities/macros.hpp"
  54 #include "utilities/powerOfTwo.hpp"
  55 #include "utilities/vmError.hpp"
  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       igvn->_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     igvn->_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(), igvn);
 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, "unsafe off-heap access with zero address"));
 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     igvn->_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     Node_List nlist;
 462     Unique_Node_List dom_list;
 463 
 464     dom_list.push(dom);
 465     bool only_dominating_controls = false;
 466 
 467     for (uint next = 0; next < dom_list.size(); next++) {
 468       Node* n = dom_list.at(next);
 469       if (n == orig_sub)
 470         return false; // One of dom's inputs dominated by sub.
 471       if (!n->is_CFG() && n->pinned()) {
 472         // Check only own control edge for pinned non-control nodes.
 473         n = n->find_exact_control(n->in(0));
 474         if (n == NULL || n->is_top())
 475           return false; // Conservative answer for dead code
 476         assert(n->is_CFG(), "expecting control");
 477         dom_list.push(n);
 478       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 479         only_dominating_controls = true;
 480       } else if (n->is_CFG()) {
 481         if (n->dominates(sub, nlist))
 482           only_dominating_controls = true;
 483         else
 484           return false;
 485       } else {
 486         // First, own control edge.
 487         Node* m = n->find_exact_control(n->in(0));
 488         if (m != NULL) {
 489           if (m->is_top())
 490             return false; // Conservative answer for dead code
 491           dom_list.push(m);
 492         }
 493         // Now, the rest of edges.
 494         uint cnt = n->req();
 495         for (uint i = 1; i < cnt; i++) {
 496           m = n->find_exact_control(n->in(i));
 497           if (m == NULL || m->is_top())
 498             continue;
 499           dom_list.push(m);
 500         }
 501       }
 502     }
 503     return only_dominating_controls;
 504   }
 505 }
 506 
 507 //---------------------detect_ptr_independence---------------------------------
 508 // Used by MemNode::find_previous_store to prove that two base
 509 // pointers are never equal.
 510 // The pointers are accompanied by their associated allocations,
 511 // if any, which have been previously discovered by the caller.
 512 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 513                                       Node* p2, AllocateNode* a2,
 514                                       PhaseTransform* phase) {
 515   // Attempt to prove that these two pointers cannot be aliased.
 516   // They may both manifestly be allocations, and they should differ.
 517   // Or, if they are not both allocations, they can be distinct constants.
 518   // Otherwise, one is an allocation and the other a pre-existing value.
 519   if (a1 == NULL && a2 == NULL) {           // neither an allocation
 520     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 521   } else if (a1 != NULL && a2 != NULL) {    // both allocations
 522     return (a1 != a2);
 523   } else if (a1 != NULL) {                  // one allocation a1
 524     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 525     return all_controls_dominate(p2, a1);
 526   } else { //(a2 != NULL)                   // one allocation a2
 527     return all_controls_dominate(p1, a2);
 528   }
 529   return false;
 530 }
 531 
 532 
 533 // Find an arraycopy ac that produces the memory state represented by parameter mem.
 534 // Return ac if
 535 // (a) can_see_stored_value=true  and ac must have set the value for this load or if
 536 // (b) can_see_stored_value=false and ac could have set the value for this load or if
 537 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
 538 // In case (c) change the parameter mem to the memory input of ac to skip it
 539 // when searching stored value.
 540 // Otherwise return NULL.
 541 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
 542   ArrayCopyNode* ac = find_array_copy_clone(phase, ld_alloc, mem);
 543   if (ac != NULL) {
 544     Node* ld_addp = in(MemNode::Address);
 545     Node* src = ac->in(ArrayCopyNode::Src);
 546     const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
 547 
 548     // This is a load from a cloned array. The corresponding arraycopy ac must
 549     // have set the value for the load and we can return ac but only if the load
 550     // is known to be within bounds. This is checked below.
 551     if (ary_t != NULL && ld_addp->is_AddP()) {
 552       Node* ld_offs = ld_addp->in(AddPNode::Offset);
 553       BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
 554       jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
 555       jlong elemsize = type2aelembytes(ary_elem);
 556 
 557       const TypeX*   ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 558       const TypeInt* sizetype  = ary_t->size();
 559 
 560       if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
 561         // The load is known to be within bounds. It receives its value from ac.
 562         return ac;
 563       }
 564       // The load is known to be out-of-bounds.
 565     }
 566     // The load could be out-of-bounds. It must not be hoisted but must remain
 567     // dependent on the runtime range check. This is achieved by returning NULL.
 568   } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
 569     ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
 570 
 571     if (ac->is_arraycopy_validated() ||
 572         ac->is_copyof_validated() ||
 573         ac->is_copyofrange_validated()) {
 574       Node* ld_addp = in(MemNode::Address);
 575       if (ld_addp->is_AddP()) {
 576         Node* ld_base = ld_addp->in(AddPNode::Address);
 577         Node* ld_offs = ld_addp->in(AddPNode::Offset);
 578 
 579         Node* dest = ac->in(ArrayCopyNode::Dest);
 580 
 581         if (dest == ld_base) {
 582           const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 583           if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
 584             return ac;
 585           }
 586           if (!can_see_stored_value) {
 587             mem = ac->in(TypeFunc::Memory);
 588             return ac;
 589           }
 590         }
 591       }
 592     }
 593   }
 594   return NULL;
 595 }
 596 
 597 ArrayCopyNode* MemNode::find_array_copy_clone(PhaseTransform* phase, Node* ld_alloc, Node* mem) const {
 598   if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
 599                                                mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
 600     if (ld_alloc != NULL) {
 601       // Check if there is an array copy for a clone
 602       Node* mb = mem->in(0);
 603       ArrayCopyNode* ac = NULL;
 604       if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
 605           mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
 606         ac = mb->in(0)->in(0)->as_ArrayCopy();
 607       } else {
 608         // Step over GC barrier when ReduceInitialCardMarks is disabled
 609         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 610         Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
 611 
 612         if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
 613           ac = control_proj_ac->in(0)->as_ArrayCopy();
 614         }
 615       }
 616 
 617       if (ac != NULL && ac->is_clonebasic()) {
 618         AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase);
 619         if (alloc != NULL && alloc == ld_alloc) {
 620           return ac;
 621         }
 622       }
 623     }
 624   }
 625   return NULL;
 626 }
 627 
 628 // The logic for reordering loads and stores uses four steps:
 629 // (a) Walk carefully past stores and initializations which we
 630 //     can prove are independent of this load.
 631 // (b) Observe that the next memory state makes an exact match
 632 //     with self (load or store), and locate the relevant store.
 633 // (c) Ensure that, if we were to wire self directly to the store,
 634 //     the optimizer would fold it up somehow.
 635 // (d) Do the rewiring, and return, depending on some other part of
 636 //     the optimizer to fold up the load.
 637 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 638 // specific to loads and stores, so they are handled by the callers.
 639 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 640 //
 641 Node* MemNode::find_previous_store(PhaseTransform* phase) {
 642   Node*         ctrl   = in(MemNode::Control);
 643   Node*         adr    = in(MemNode::Address);
 644   intptr_t      offset = 0;
 645   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 646   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
 647 
 648   if (offset == Type::OffsetBot)
 649     return NULL;            // cannot unalias unless there are precise offsets
 650 
 651   const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
 652   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 653 
 654   intptr_t size_in_bytes = memory_size();
 655 
 656   Node* mem = in(MemNode::Memory);   // start searching here...
 657 
 658   int cnt = 50;             // Cycle limiter
 659   for (;;) {                // While we can dance past unrelated stores...
 660     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 661 
 662     Node* prev = mem;
 663     if (mem->is_Store()) {
 664       Node* st_adr = mem->in(MemNode::Address);
 665       intptr_t st_offset = 0;
 666       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 667       if (st_base == NULL)
 668         break;              // inscrutable pointer
 669 
 670       // For raw accesses it's not enough to prove that constant offsets don't intersect.
 671       // We need the bases to be the equal in order for the offset check to make sense.
 672       if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
 673         break;
 674       }
 675 
 676       if (st_offset != offset && st_offset != Type::OffsetBot) {
 677         const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
 678         assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
 679         if (st_offset >= offset + size_in_bytes ||
 680             st_offset <= offset - MAX_STORE ||
 681             st_offset <= offset - mem->as_Store()->memory_size()) {
 682           // Success:  The offsets are provably independent.
 683           // (You may ask, why not just test st_offset != offset and be done?
 684           // The answer is that stores of different sizes can co-exist
 685           // in the same sequence of RawMem effects.  We sometimes initialize
 686           // a whole 'tile' of array elements with a single jint or jlong.)
 687           mem = mem->in(MemNode::Memory);
 688           continue;           // (a) advance through independent store memory
 689         }
 690       }
 691       if (st_base != base &&
 692           detect_ptr_independence(base, alloc,
 693                                   st_base,
 694                                   AllocateNode::Ideal_allocation(st_base, phase),
 695                                   phase)) {
 696         // Success:  The bases are provably independent.
 697         mem = mem->in(MemNode::Memory);
 698         continue;           // (a) advance through independent store memory
 699       }
 700 
 701       // (b) At this point, if the bases or offsets do not agree, we lose,
 702       // since we have not managed to prove 'this' and 'mem' independent.
 703       if (st_base == base && st_offset == offset) {
 704         return mem;         // let caller handle steps (c), (d)
 705       }
 706 
 707     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 708       InitializeNode* st_init = mem->in(0)->as_Initialize();
 709       AllocateNode*  st_alloc = st_init->allocation();
 710       if (st_alloc == NULL)
 711         break;              // something degenerated
 712       bool known_identical = false;
 713       bool known_independent = false;
 714       if (alloc == st_alloc)
 715         known_identical = true;
 716       else if (alloc != NULL)
 717         known_independent = true;
 718       else if (all_controls_dominate(this, st_alloc))
 719         known_independent = true;
 720 
 721       if (known_independent) {
 722         // The bases are provably independent: Either they are
 723         // manifestly distinct allocations, or else the control
 724         // of this load dominates the store's allocation.
 725         int alias_idx = phase->C->get_alias_index(adr_type());
 726         if (alias_idx == Compile::AliasIdxRaw) {
 727           mem = st_alloc->in(TypeFunc::Memory);
 728         } else {
 729           mem = st_init->memory(alias_idx);
 730         }
 731         continue;           // (a) advance through independent store memory
 732       }
 733 
 734       // (b) at this point, if we are not looking at a store initializing
 735       // the same allocation we are loading from, we lose.
 736       if (known_identical) {
 737         // From caller, can_see_stored_value will consult find_captured_store.
 738         return mem;         // let caller handle steps (c), (d)
 739       }
 740 
 741     } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
 742       if (prev != mem) {
 743         // Found an arraycopy but it doesn't affect that load
 744         continue;
 745       }
 746       // Found an arraycopy that may affect that load
 747       return mem;
 748     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
 749       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 750       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 751         // ArrayCopyNodes processed here as well.
 752         CallNode *call = mem->in(0)->as_Call();
 753         if (!call->may_modify(addr_t, phase)) {
 754           mem = call->in(TypeFunc::Memory);
 755           continue;         // (a) advance through independent call memory
 756         }
 757       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 758         ArrayCopyNode* ac = NULL;
 759         if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
 760           break;
 761         }
 762         mem = mem->in(0)->in(TypeFunc::Memory);
 763         continue;           // (a) advance through independent MemBar memory
 764       } else if (mem->is_ClearArray()) {
 765         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 766           // (the call updated 'mem' value)
 767           continue;         // (a) advance through independent allocation memory
 768         } else {
 769           // Can not bypass initialization of the instance
 770           // we are looking for.
 771           return mem;
 772         }
 773       } else if (mem->is_MergeMem()) {
 774         int alias_idx = phase->C->get_alias_index(adr_type());
 775         mem = mem->as_MergeMem()->memory_at(alias_idx);
 776         continue;           // (a) advance through independent MergeMem memory
 777       }
 778     }
 779 
 780     // Unless there is an explicit 'continue', we must bail out here,
 781     // because 'mem' is an inscrutable memory state (e.g., a call).
 782     break;
 783   }
 784 
 785   return NULL;              // bail out
 786 }
 787 
 788 //----------------------calculate_adr_type-------------------------------------
 789 // Helper function.  Notices when the given type of address hits top or bottom.
 790 // Also, asserts a cross-check of the type against the expected address type.
 791 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 792   if (t == Type::TOP)  return NULL; // does not touch memory any more?
 793   #ifdef ASSERT
 794   if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump())  cross_check = NULL;
 795   #endif
 796   const TypePtr* tp = t->isa_ptr();
 797   if (tp == NULL) {
 798     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 799     return TypePtr::BOTTOM;           // touches lots of memory
 800   } else {
 801     #ifdef ASSERT
 802     // %%%% [phh] We don't check the alias index if cross_check is
 803     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 804     if (cross_check != NULL &&
 805         cross_check != TypePtr::BOTTOM &&
 806         cross_check != TypeRawPtr::BOTTOM) {
 807       // Recheck the alias index, to see if it has changed (due to a bug).
 808       Compile* C = Compile::current();
 809       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 810              "must stay in the original alias category");
 811       // The type of the address must be contained in the adr_type,
 812       // disregarding "null"-ness.
 813       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 814       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 815       assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
 816              "real address must not escape from expected memory type");
 817     }
 818     #endif
 819     return tp;
 820   }
 821 }
 822 
 823 //=============================================================================
 824 // Should LoadNode::Ideal() attempt to remove control edges?
 825 bool LoadNode::can_remove_control() const {
 826   return true;
 827 }
 828 uint LoadNode::size_of() const { return sizeof(*this); }
 829 bool LoadNode::cmp( const Node &n ) const
 830 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
 831 const Type *LoadNode::bottom_type() const { return _type; }
 832 uint LoadNode::ideal_reg() const {
 833   return _type->ideal_reg();
 834 }
 835 
 836 #ifndef PRODUCT
 837 void LoadNode::dump_spec(outputStream *st) const {
 838   MemNode::dump_spec(st);
 839   if( !Verbose && !WizardMode ) {
 840     // standard dump does this in Verbose and WizardMode
 841     st->print(" #"); _type->dump_on(st);
 842   }
 843   if (!depends_only_on_test()) {
 844     st->print(" (does not depend only on test)");
 845   }
 846 }
 847 #endif
 848 
 849 #ifdef ASSERT
 850 //----------------------------is_immutable_value-------------------------------
 851 // Helper function to allow a raw load without control edge for some cases
 852 bool LoadNode::is_immutable_value(Node* adr) {
 853   if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 854       adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
 855 
 856     jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
 857     int offsets[] = {
 858       in_bytes(JavaThread::osthread_offset()),
 859       in_bytes(JavaThread::threadObj_offset()),
 860       in_bytes(JavaThread::vthread_offset()),
 861       in_bytes(JavaThread::scopeLocalCache_offset()),
 862     };
 863 
 864     for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
 865       if (offset == offsets[i])
 866         return true;
 867     }
 868   }
 869 
 870   return false;
 871 }
 872 #endif
 873 
 874 //----------------------------LoadNode::make-----------------------------------
 875 // Polymorphic factory method:
 876 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
 877                      ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
 878   Compile* C = gvn.C;
 879 
 880   // sanity check the alias category against the created node type
 881   assert(!(adr_type->isa_oopptr() &&
 882            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
 883          "use LoadKlassNode instead");
 884   assert(!(adr_type->isa_aryptr() &&
 885            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
 886          "use LoadRangeNode instead");
 887   // Check control edge of raw loads
 888   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
 889           // oop will be recorded in oop map if load crosses safepoint
 890           rt->isa_oopptr() || is_immutable_value(adr),
 891           "raw memory operations should have control edge");
 892   LoadNode* load = NULL;
 893   switch (bt) {
 894   case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 895   case T_BYTE:    load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 896   case T_INT:     load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 897   case T_CHAR:    load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 898   case T_SHORT:   load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 899   case T_LONG:    load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
 900   case T_FLOAT:   load = new LoadFNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
 901   case T_DOUBLE:  load = new LoadDNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
 902   case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(),  mo, control_dependency); break;
 903   case T_OBJECT:
 904 #ifdef _LP64
 905     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 906       load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
 907     } else
 908 #endif
 909     {
 910       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
 911       load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
 912     }
 913     break;
 914   default:
 915     ShouldNotReachHere();
 916     break;
 917   }
 918   assert(load != NULL, "LoadNode should have been created");
 919   if (unaligned) {
 920     load->set_unaligned_access();
 921   }
 922   if (mismatched) {
 923     load->set_mismatched_access();
 924   }
 925   if (unsafe) {
 926     load->set_unsafe_access();
 927   }
 928   load->set_barrier_data(barrier_data);
 929   if (load->Opcode() == Op_LoadN) {
 930     Node* ld = gvn.transform(load);
 931     return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
 932   }
 933 
 934   return load;
 935 }
 936 
 937 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
 938                                   ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
 939   bool require_atomic = true;
 940   LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
 941   if (unaligned) {
 942     load->set_unaligned_access();
 943   }
 944   if (mismatched) {
 945     load->set_mismatched_access();
 946   }
 947   if (unsafe) {
 948     load->set_unsafe_access();
 949   }
 950   load->set_barrier_data(barrier_data);
 951   return load;
 952 }
 953 
 954 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
 955                                   ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
 956   bool require_atomic = true;
 957   LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
 958   if (unaligned) {
 959     load->set_unaligned_access();
 960   }
 961   if (mismatched) {
 962     load->set_mismatched_access();
 963   }
 964   if (unsafe) {
 965     load->set_unsafe_access();
 966   }
 967   load->set_barrier_data(barrier_data);
 968   return load;
 969 }
 970 
 971 
 972 
 973 //------------------------------hash-------------------------------------------
 974 uint LoadNode::hash() const {
 975   // unroll addition of interesting fields
 976   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
 977 }
 978 
 979 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
 980   if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
 981     bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
 982     bool is_stable_ary = FoldStableValues &&
 983                          (tp != NULL) && (tp->isa_aryptr() != NULL) &&
 984                          tp->isa_aryptr()->is_stable();
 985 
 986     return (eliminate_boxing && non_volatile) || is_stable_ary;
 987   }
 988 
 989   return false;
 990 }
 991 
 992 // Is the value loaded previously stored by an arraycopy? If so return
 993 // a load node that reads from the source array so we may be able to
 994 // optimize out the ArrayCopy node later.
 995 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
 996   Node* ld_adr = in(MemNode::Address);
 997   intptr_t ld_off = 0;
 998   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 999   Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1000   if (ac != NULL) {
1001     assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1002 
1003     Node* mem = ac->in(TypeFunc::Memory);
1004     Node* ctl = ac->in(0);
1005     Node* src = ac->in(ArrayCopyNode::Src);
1006 
1007     if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1008       return NULL;
1009     }
1010 
1011     LoadNode* ld = clone()->as_Load();
1012     Node* addp = in(MemNode::Address)->clone();
1013     if (ac->as_ArrayCopy()->is_clonebasic()) {
1014       assert(ld_alloc != NULL, "need an alloc");
1015       assert(addp->is_AddP(), "address must be addp");
1016       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1017       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1018       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1019       addp->set_req(AddPNode::Base, src);
1020       addp->set_req(AddPNode::Address, src);
1021     } else {
1022       assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1023              ac->as_ArrayCopy()->is_copyof_validated() ||
1024              ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1025       assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1026       addp->set_req(AddPNode::Base, src);
1027       addp->set_req(AddPNode::Address, src);
1028 
1029       const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1030       BasicType ary_elem  = ary_t->klass()->as_array_klass()->element_type()->basic_type();
1031       uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1032       uint shift  = exact_log2(type2aelembytes(ary_elem));
1033 
1034       Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1035 #ifdef _LP64
1036       diff = phase->transform(new ConvI2LNode(diff));
1037 #endif
1038       diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1039 
1040       Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1041       addp->set_req(AddPNode::Offset, offset);
1042     }
1043     addp = phase->transform(addp);
1044 #ifdef ASSERT
1045     const TypePtr* adr_type = phase->type(addp)->is_ptr();
1046     ld->_adr_type = adr_type;
1047 #endif
1048     ld->set_req(MemNode::Address, addp);
1049     ld->set_req(0, ctl);
1050     ld->set_req(MemNode::Memory, mem);
1051     // load depends on the tests that validate the arraycopy
1052     ld->_control_dependency = UnknownControl;
1053     return ld;
1054   }
1055   return NULL;
1056 }
1057 
1058 
1059 //---------------------------can_see_stored_value------------------------------
1060 // This routine exists to make sure this set of tests is done the same
1061 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
1062 // will change the graph shape in a way which makes memory alive twice at the
1063 // same time (uses the Oracle model of aliasing), then some
1064 // LoadXNode::Identity will fold things back to the equivalence-class model
1065 // of aliasing.
1066 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1067   Node* ld_adr = in(MemNode::Address);
1068   intptr_t ld_off = 0;
1069   Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1070   Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1071   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1072   Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1073   // This is more general than load from boxing objects.
1074   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1075     uint alias_idx = atp->index();
1076     Node* result = NULL;
1077     Node* current = st;
1078     // Skip through chains of MemBarNodes checking the MergeMems for
1079     // new states for the slice of this load.  Stop once any other
1080     // kind of node is encountered.  Loads from final memory can skip
1081     // through any kind of MemBar but normal loads shouldn't skip
1082     // through MemBarAcquire since the could allow them to move out of
1083     // a synchronized region. It is not safe to step over MemBarCPUOrder,
1084     // because alias info above them may be inaccurate (e.g., due to
1085     // mixed/mismatched unsafe accesses).
1086     bool is_final_mem = !atp->is_rewritable();
1087     while (current->is_Proj()) {
1088       int opc = current->in(0)->Opcode();
1089       if ((is_final_mem && (opc == Op_MemBarAcquire ||
1090                             opc == Op_MemBarAcquireLock ||
1091                             opc == Op_LoadFence)) ||
1092           opc == Op_MemBarRelease ||
1093           opc == Op_StoreFence ||
1094           opc == Op_MemBarReleaseLock ||
1095           opc == Op_MemBarStoreStore) {
1096         Node* mem = current->in(0)->in(TypeFunc::Memory);
1097         if (mem->is_MergeMem()) {
1098           MergeMemNode* merge = mem->as_MergeMem();
1099           Node* new_st = merge->memory_at(alias_idx);
1100           if (new_st == merge->base_memory()) {
1101             // Keep searching
1102             current = new_st;
1103             continue;
1104           }
1105           // Save the new memory state for the slice and fall through
1106           // to exit.
1107           result = new_st;
1108         }
1109       }
1110       break;
1111     }
1112     if (result != NULL) {
1113       st = result;
1114     }
1115   }
1116 
1117   // Loop around twice in the case Load -> Initialize -> Store.
1118   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1119   for (int trip = 0; trip <= 1; trip++) {
1120 
1121     if (st->is_Store()) {
1122       Node* st_adr = st->in(MemNode::Address);
1123       if (st_adr != ld_adr) {
1124         // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1125         intptr_t st_off = 0;
1126         Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1127         if (ld_base == NULL)                                   return NULL;
1128         if (st_base == NULL)                                   return NULL;
1129         if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1130         if (ld_off != st_off)                                  return NULL;
1131         if (ld_off == Type::OffsetBot)                         return NULL;
1132         // Same base, same offset.
1133         // Possible improvement for arrays: check index value instead of absolute offset.
1134 
1135         // At this point we have proven something like this setup:
1136         //   B = << base >>
1137         //   L =  LoadQ(AddP(Check/CastPP(B), #Off))
1138         //   S = StoreQ(AddP(             B , #Off), V)
1139         // (Actually, we haven't yet proven the Q's are the same.)
1140         // In other words, we are loading from a casted version of
1141         // the same pointer-and-offset that we stored to.
1142         // Casted version may carry a dependency and it is respected.
1143         // Thus, we are able to replace L by V.
1144       }
1145       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1146       if (store_Opcode() != st->Opcode()) {
1147         return NULL;
1148       }
1149       // LoadVector/StoreVector needs additional check to ensure the types match.
1150       if (store_Opcode() == Op_StoreVector) {
1151         const TypeVect*  in_vt = st->as_StoreVector()->vect_type();
1152         const TypeVect* out_vt = as_LoadVector()->vect_type();
1153         if (in_vt != out_vt) {
1154           return NULL;
1155         }
1156       }
1157       return st->in(MemNode::ValueIn);
1158     }
1159 
1160     // A load from a freshly-created object always returns zero.
1161     // (This can happen after LoadNode::Ideal resets the load's memory input
1162     // to find_captured_store, which returned InitializeNode::zero_memory.)
1163     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1164         (st->in(0) == ld_alloc) &&
1165         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1166       // return a zero value for the load's basic type
1167       // (This is one of the few places where a generic PhaseTransform
1168       // can create new nodes.  Think of it as lazily manifesting
1169       // virtually pre-existing constants.)
1170       if (memory_type() != T_VOID) {
1171         if (ReduceBulkZeroing || find_array_copy_clone(phase, ld_alloc, in(MemNode::Memory)) == NULL) {
1172           // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1173           // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1174           // by the ArrayCopyNode.
1175           return phase->zerocon(memory_type());
1176         }
1177       } else {
1178         // TODO: materialize all-zero vector constant
1179         assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1180       }
1181     }
1182 
1183     // A load from an initialization barrier can match a captured store.
1184     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1185       InitializeNode* init = st->in(0)->as_Initialize();
1186       AllocateNode* alloc = init->allocation();
1187       if ((alloc != NULL) && (alloc == ld_alloc)) {
1188         // examine a captured store value
1189         st = init->find_captured_store(ld_off, memory_size(), phase);
1190         if (st != NULL) {
1191           continue;             // take one more trip around
1192         }
1193       }
1194     }
1195 
1196     // Load boxed value from result of valueOf() call is input parameter.
1197     if (this->is_Load() && ld_adr->is_AddP() &&
1198         (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1199       intptr_t ignore = 0;
1200       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1201       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1202       base = bs->step_over_gc_barrier(base);
1203       if (base != NULL && base->is_Proj() &&
1204           base->as_Proj()->_con == TypeFunc::Parms &&
1205           base->in(0)->is_CallStaticJava() &&
1206           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1207         return base->in(0)->in(TypeFunc::Parms);
1208       }
1209     }
1210 
1211     break;
1212   }
1213 
1214   return NULL;
1215 }
1216 
1217 //----------------------is_instance_field_load_with_local_phi------------------
1218 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1219   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1220       in(Address)->is_AddP() ) {
1221     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1222     // Only instances and boxed values.
1223     if( t_oop != NULL &&
1224         (t_oop->is_ptr_to_boxed_value() ||
1225          t_oop->is_known_instance_field()) &&
1226         t_oop->offset() != Type::OffsetBot &&
1227         t_oop->offset() != Type::OffsetTop) {
1228       return true;
1229     }
1230   }
1231   return false;
1232 }
1233 
1234 //------------------------------Identity---------------------------------------
1235 // Loads are identity if previous store is to same address
1236 Node* LoadNode::Identity(PhaseGVN* phase) {
1237   // If the previous store-maker is the right kind of Store, and the store is
1238   // to the same address, then we are equal to the value stored.
1239   Node* mem = in(Memory);
1240   Node* value = can_see_stored_value(mem, phase);
1241   if( value ) {
1242     // byte, short & char stores truncate naturally.
1243     // A load has to load the truncated value which requires
1244     // some sort of masking operation and that requires an
1245     // Ideal call instead of an Identity call.
1246     if (memory_size() < BytesPerInt) {
1247       // If the input to the store does not fit with the load's result type,
1248       // it must be truncated via an Ideal call.
1249       if (!phase->type(value)->higher_equal(phase->type(this)))
1250         return this;
1251     }
1252     // (This works even when value is a Con, but LoadNode::Value
1253     // usually runs first, producing the singleton type of the Con.)
1254     return value;
1255   }
1256 
1257   // Search for an existing data phi which was generated before for the same
1258   // instance's field to avoid infinite generation of phis in a loop.
1259   Node *region = mem->in(0);
1260   if (is_instance_field_load_with_local_phi(region)) {
1261     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1262     int this_index  = phase->C->get_alias_index(addr_t);
1263     int this_offset = addr_t->offset();
1264     int this_iid    = addr_t->instance_id();
1265     if (!addr_t->is_known_instance() &&
1266          addr_t->is_ptr_to_boxed_value()) {
1267       // Use _idx of address base (could be Phi node) for boxed values.
1268       intptr_t   ignore = 0;
1269       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1270       if (base == NULL) {
1271         return this;
1272       }
1273       this_iid = base->_idx;
1274     }
1275     const Type* this_type = bottom_type();
1276     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1277       Node* phi = region->fast_out(i);
1278       if (phi->is_Phi() && phi != mem &&
1279           phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1280         return phi;
1281       }
1282     }
1283   }
1284 
1285   return this;
1286 }
1287 
1288 // Construct an equivalent unsigned load.
1289 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1290   BasicType bt = T_ILLEGAL;
1291   const Type* rt = NULL;
1292   switch (Opcode()) {
1293     case Op_LoadUB: return this;
1294     case Op_LoadUS: return this;
1295     case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1296     case Op_LoadS: bt = T_CHAR;    rt = TypeInt::CHAR;  break;
1297     default:
1298       assert(false, "no unsigned variant: %s", Name());
1299       return NULL;
1300   }
1301   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1302                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1303                         is_unaligned_access(), is_mismatched_access());
1304 }
1305 
1306 // Construct an equivalent signed load.
1307 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1308   BasicType bt = T_ILLEGAL;
1309   const Type* rt = NULL;
1310   switch (Opcode()) {
1311     case Op_LoadUB: bt = T_BYTE;  rt = TypeInt::BYTE;  break;
1312     case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1313     case Op_LoadB: // fall through
1314     case Op_LoadS: // fall through
1315     case Op_LoadI: // fall through
1316     case Op_LoadL: return this;
1317     default:
1318       assert(false, "no signed variant: %s", Name());
1319       return NULL;
1320   }
1321   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1322                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1323                         is_unaligned_access(), is_mismatched_access());
1324 }
1325 
1326 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1327   BasicType bt = rt->basic_type();
1328   switch (Opcode()) {
1329     case Op_LoadI: return (bt == T_FLOAT);
1330     case Op_LoadL: return (bt == T_DOUBLE);
1331     case Op_LoadF: return (bt == T_INT);
1332     case Op_LoadD: return (bt == T_LONG);
1333 
1334     default: return false;
1335   }
1336 }
1337 
1338 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1339   BasicType bt = rt->basic_type();
1340   assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1341   bool is_mismatched = is_mismatched_access();
1342   const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1343   if (raw_type == NULL) {
1344     is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1345   }
1346   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1347                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1348                         is_unaligned_access(), is_mismatched);
1349 }
1350 
1351 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1352   BasicType bt = vt->basic_type();
1353   switch (Opcode()) {
1354     case Op_StoreI: return (bt == T_FLOAT);
1355     case Op_StoreL: return (bt == T_DOUBLE);
1356     case Op_StoreF: return (bt == T_INT);
1357     case Op_StoreD: return (bt == T_LONG);
1358 
1359     default: return false;
1360   }
1361 }
1362 
1363 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1364   BasicType bt = vt->basic_type();
1365   assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1366   StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), raw_adr_type(), val, bt, _mo);
1367 
1368   bool is_mismatched = is_mismatched_access();
1369   const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1370   if (raw_type == NULL) {
1371     is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1372   }
1373   if (is_mismatched) {
1374     st->set_mismatched_access();
1375   }
1376   return st;
1377 }
1378 
1379 // We're loading from an object which has autobox behaviour.
1380 // If this object is result of a valueOf call we'll have a phi
1381 // merging a newly allocated object and a load from the cache.
1382 // We want to replace this load with the original incoming
1383 // argument to the valueOf call.
1384 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1385   assert(igvn->C->eliminate_boxing(), "sanity");
1386   intptr_t ignore = 0;
1387   Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1388   if ((base == NULL) || base->is_Phi()) {
1389     // Push the loads from the phi that comes from valueOf up
1390     // through it to allow elimination of the loads and the recovery
1391     // of the original value. It is done in split_through_phi().
1392     return NULL;
1393   } else if (base->is_Load() ||
1394              (base->is_DecodeN() && base->in(1)->is_Load())) {
1395     // Eliminate the load of boxed value for integer types from the cache
1396     // array by deriving the value from the index into the array.
1397     // Capture the offset of the load and then reverse the computation.
1398 
1399     // Get LoadN node which loads a boxing object from 'cache' array.
1400     if (base->is_DecodeN()) {
1401       base = base->in(1);
1402     }
1403     if (!base->in(Address)->is_AddP()) {
1404       return NULL; // Complex address
1405     }
1406     AddPNode* address = base->in(Address)->as_AddP();
1407     Node* cache_base = address->in(AddPNode::Base);
1408     if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1409       // Get ConP node which is static 'cache' field.
1410       cache_base = cache_base->in(1);
1411     }
1412     if ((cache_base != NULL) && cache_base->is_Con()) {
1413       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1414       if ((base_type != NULL) && base_type->is_autobox_cache()) {
1415         Node* elements[4];
1416         int shift = exact_log2(type2aelembytes(T_OBJECT));
1417         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1418         if (count > 0 && elements[0]->is_Con() &&
1419             (count == 1 ||
1420              (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1421                             elements[1]->in(2) == igvn->intcon(shift)))) {
1422           ciObjArray* array = base_type->const_oop()->as_obj_array();
1423           // Fetch the box object cache[0] at the base of the array and get its value
1424           ciInstance* box = array->obj_at(0)->as_instance();
1425           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1426           assert(ik->is_box_klass(), "sanity");
1427           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1428           if (ik->nof_nonstatic_fields() == 1) {
1429             // This should be true nonstatic_field_at requires calling
1430             // nof_nonstatic_fields so check it anyway
1431             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1432             BasicType bt = c.basic_type();
1433             // Only integer types have boxing cache.
1434             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1435                    bt == T_BYTE    || bt == T_SHORT ||
1436                    bt == T_INT     || bt == T_LONG, "wrong type = %s", type2name(bt));
1437             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1438             if (cache_low != (int)cache_low) {
1439               return NULL; // should not happen since cache is array indexed by value
1440             }
1441             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1442             if (offset != (int)offset) {
1443               return NULL; // should not happen since cache is array indexed by value
1444             }
1445            // Add up all the offsets making of the address of the load
1446             Node* result = elements[0];
1447             for (int i = 1; i < count; i++) {
1448               result = igvn->transform(new AddXNode(result, elements[i]));
1449             }
1450             // Remove the constant offset from the address and then
1451             result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1452             // remove the scaling of the offset to recover the original index.
1453             if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1454               // Peel the shift off directly but wrap it in a dummy node
1455               // since Ideal can't return existing nodes
1456               igvn->_worklist.push(result); // remove dead node later
1457               result = new RShiftXNode(result->in(1), igvn->intcon(0));
1458             } else if (result->is_Add() && result->in(2)->is_Con() &&
1459                        result->in(1)->Opcode() == Op_LShiftX &&
1460                        result->in(1)->in(2) == igvn->intcon(shift)) {
1461               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1462               // but for boxing cache access we know that X<<Z will not overflow
1463               // (there is range check) so we do this optimizatrion by hand here.
1464               igvn->_worklist.push(result); // remove dead node later
1465               Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1466               result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1467             } else {
1468               result = new RShiftXNode(result, igvn->intcon(shift));
1469             }
1470 #ifdef _LP64
1471             if (bt != T_LONG) {
1472               result = new ConvL2INode(igvn->transform(result));
1473             }
1474 #else
1475             if (bt == T_LONG) {
1476               result = new ConvI2LNode(igvn->transform(result));
1477             }
1478 #endif
1479             // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1480             // Need to preserve unboxing load type if it is unsigned.
1481             switch(this->Opcode()) {
1482               case Op_LoadUB:
1483                 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1484                 break;
1485               case Op_LoadUS:
1486                 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1487                 break;
1488             }
1489             return result;
1490           }
1491         }
1492       }
1493     }
1494   }
1495   return NULL;
1496 }
1497 
1498 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1499   Node* region = phi->in(0);
1500   if (region == NULL) {
1501     return false; // Wait stable graph
1502   }
1503   uint cnt = phi->req();
1504   for (uint i = 1; i < cnt; i++) {
1505     Node* rc = region->in(i);
1506     if (rc == NULL || phase->type(rc) == Type::TOP)
1507       return false; // Wait stable graph
1508     Node* in = phi->in(i);
1509     if (in == NULL || phase->type(in) == Type::TOP)
1510       return false; // Wait stable graph
1511   }
1512   return true;
1513 }
1514 //------------------------------split_through_phi------------------------------
1515 // Split instance or boxed field load through Phi.
1516 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1517   Node* mem     = in(Memory);
1518   Node* address = in(Address);
1519   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1520 
1521   assert((t_oop != NULL) &&
1522          (t_oop->is_known_instance_field() ||
1523           t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1524 
1525   Compile* C = phase->C;
1526   intptr_t ignore = 0;
1527   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1528   bool base_is_phi = (base != NULL) && base->is_Phi();
1529   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1530                            (base != NULL) && (base == address->in(AddPNode::Base)) &&
1531                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1532 
1533   if (!((mem->is_Phi() || base_is_phi) &&
1534         (load_boxed_values || t_oop->is_known_instance_field()))) {
1535     return NULL; // memory is not Phi
1536   }
1537 
1538   if (mem->is_Phi()) {
1539     if (!stable_phi(mem->as_Phi(), phase)) {
1540       return NULL; // Wait stable graph
1541     }
1542     uint cnt = mem->req();
1543     // Check for loop invariant memory.
1544     if (cnt == 3) {
1545       for (uint i = 1; i < cnt; i++) {
1546         Node* in = mem->in(i);
1547         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1548         if (m == mem) {
1549           if (i == 1) {
1550             // if the first edge was a loop, check second edge too.
1551             // If both are replaceable - we are in an infinite loop
1552             Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1553             if (n == mem) {
1554               break;
1555             }
1556           }
1557           set_req(Memory, mem->in(cnt - i));
1558           return this; // made change
1559         }
1560       }
1561     }
1562   }
1563   if (base_is_phi) {
1564     if (!stable_phi(base->as_Phi(), phase)) {
1565       return NULL; // Wait stable graph
1566     }
1567     uint cnt = base->req();
1568     // Check for loop invariant memory.
1569     if (cnt == 3) {
1570       for (uint i = 1; i < cnt; i++) {
1571         if (base->in(i) == base) {
1572           return NULL; // Wait stable graph
1573         }
1574       }
1575     }
1576   }
1577 
1578   // Split through Phi (see original code in loopopts.cpp).
1579   assert(C->have_alias_type(t_oop), "instance should have alias type");
1580 
1581   // Do nothing here if Identity will find a value
1582   // (to avoid infinite chain of value phis generation).
1583   if (this != Identity(phase)) {
1584     return NULL;
1585   }
1586 
1587   // Select Region to split through.
1588   Node* region;
1589   if (!base_is_phi) {
1590     assert(mem->is_Phi(), "sanity");
1591     region = mem->in(0);
1592     // Skip if the region dominates some control edge of the address.
1593     if (!MemNode::all_controls_dominate(address, region))
1594       return NULL;
1595   } else if (!mem->is_Phi()) {
1596     assert(base_is_phi, "sanity");
1597     region = base->in(0);
1598     // Skip if the region dominates some control edge of the memory.
1599     if (!MemNode::all_controls_dominate(mem, region))
1600       return NULL;
1601   } else if (base->in(0) != mem->in(0)) {
1602     assert(base_is_phi && mem->is_Phi(), "sanity");
1603     if (MemNode::all_controls_dominate(mem, base->in(0))) {
1604       region = base->in(0);
1605     } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1606       region = mem->in(0);
1607     } else {
1608       return NULL; // complex graph
1609     }
1610   } else {
1611     assert(base->in(0) == mem->in(0), "sanity");
1612     region = mem->in(0);
1613   }
1614 
1615   const Type* this_type = this->bottom_type();
1616   int this_index  = C->get_alias_index(t_oop);
1617   int this_offset = t_oop->offset();
1618   int this_iid    = t_oop->instance_id();
1619   if (!t_oop->is_known_instance() && load_boxed_values) {
1620     // Use _idx of address base for boxed values.
1621     this_iid = base->_idx;
1622   }
1623   PhaseIterGVN* igvn = phase->is_IterGVN();
1624   Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1625   for (uint i = 1; i < region->req(); i++) {
1626     Node* x;
1627     Node* the_clone = NULL;
1628     Node* in = region->in(i);
1629     if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1630         in != NULL && in->is_OuterStripMinedLoop()) {
1631       // No node should go in the outer strip mined loop
1632       in = in->in(LoopNode::EntryControl);
1633     }
1634     if (in == NULL || in == C->top()) {
1635       x = C->top();      // Dead path?  Use a dead data op
1636     } else {
1637       x = this->clone();        // Else clone up the data op
1638       the_clone = x;            // Remember for possible deletion.
1639       // Alter data node to use pre-phi inputs
1640       if (this->in(0) == region) {
1641         if (mem->is_Phi() && (mem->in(0) == region) && mem->in(i)->in(0) != NULL &&
1642             MemNode::all_controls_dominate(address, region)) {
1643           // Enable other optimizations such as loop predication which does not work
1644           // if we directly pin the node to node `in`
1645           x->set_req(0, mem->in(i)->in(0)); // Use same control as memory
1646         } else {
1647           x->set_req(0, in);
1648         }
1649       } else {
1650         x->set_req(0, NULL);
1651       }
1652       if (mem->is_Phi() && (mem->in(0) == region)) {
1653         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1654       }
1655       if (address->is_Phi() && address->in(0) == region) {
1656         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1657       }
1658       if (base_is_phi && (base->in(0) == region)) {
1659         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1660         Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1661         x->set_req(Address, adr_x);
1662       }
1663     }
1664     // Check for a 'win' on some paths
1665     const Type *t = x->Value(igvn);
1666 
1667     bool singleton = t->singleton();
1668 
1669     // See comments in PhaseIdealLoop::split_thru_phi().
1670     if (singleton && t == Type::TOP) {
1671       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1672     }
1673 
1674     if (singleton) {
1675       x = igvn->makecon(t);
1676     } else {
1677       // We now call Identity to try to simplify the cloned node.
1678       // Note that some Identity methods call phase->type(this).
1679       // Make sure that the type array is big enough for
1680       // our new node, even though we may throw the node away.
1681       // (This tweaking with igvn only works because x is a new node.)
1682       igvn->set_type(x, t);
1683       // If x is a TypeNode, capture any more-precise type permanently into Node
1684       // otherwise it will be not updated during igvn->transform since
1685       // igvn->type(x) is set to x->Value() already.
1686       x->raise_bottom_type(t);
1687       Node* y = x->Identity(igvn);
1688       if (y != x) {
1689         x = y;
1690       } else {
1691         y = igvn->hash_find_insert(x);
1692         if (y) {
1693           x = y;
1694         } else {
1695           // Else x is a new node we are keeping
1696           // We do not need register_new_node_with_optimizer
1697           // because set_type has already been called.
1698           igvn->_worklist.push(x);
1699         }
1700       }
1701     }
1702     if (x != the_clone && the_clone != NULL) {
1703       igvn->remove_dead_node(the_clone);
1704     }
1705     phi->set_req(i, x);
1706   }
1707   // Record Phi
1708   igvn->register_new_node_with_optimizer(phi);
1709   return phi;
1710 }
1711 
1712 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const {
1713   if (Opcode() == Op_LoadX) {
1714     Node* address = in(MemNode::Address);
1715     AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1716     Node* mem = in(MemNode::Memory);
1717     if (alloc != NULL && mem->is_Proj() &&
1718         mem->in(0) != NULL &&
1719         mem->in(0) == alloc->initialization() &&
1720         alloc->initialization()->proj_out_or_null(0) != NULL) {
1721       return alloc;
1722     }
1723   }
1724   return NULL;
1725 }
1726 
1727 
1728 //------------------------------Ideal------------------------------------------
1729 // If the load is from Field memory and the pointer is non-null, it might be possible to
1730 // zero out the control input.
1731 // If the offset is constant and the base is an object allocation,
1732 // try to hook me up to the exact initializing store.
1733 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1734   Node* p = MemNode::Ideal_common(phase, can_reshape);
1735   if (p)  return (p == NodeSentinel) ? NULL : p;
1736 
1737   Node* ctrl    = in(MemNode::Control);
1738   Node* address = in(MemNode::Address);
1739   bool progress = false;
1740 
1741   bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1742          phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1743 
1744   // Skip up past a SafePoint control.  Cannot do this for Stores because
1745   // pointer stores & cardmarks must stay on the same side of a SafePoint.
1746   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1747       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw  &&
1748       !addr_mark &&
1749       (depends_only_on_test() || has_unknown_control_dependency())) {
1750     ctrl = ctrl->in(0);
1751     set_req(MemNode::Control,ctrl);
1752     progress = true;
1753   }
1754 
1755   intptr_t ignore = 0;
1756   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1757   if (base != NULL
1758       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1759     // Check for useless control edge in some common special cases
1760     if (in(MemNode::Control) != NULL
1761         && can_remove_control()
1762         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1763         && all_controls_dominate(base, phase->C->start())) {
1764       // A method-invariant, non-null address (constant or 'this' argument).
1765       set_req(MemNode::Control, NULL);
1766       progress = true;
1767     }
1768   }
1769 
1770   Node* mem = in(MemNode::Memory);
1771   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1772 
1773   if (can_reshape && (addr_t != NULL)) {
1774     // try to optimize our memory input
1775     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1776     if (opt_mem != mem) {
1777       set_req_X(MemNode::Memory, opt_mem, phase);
1778       if (phase->type( opt_mem ) == Type::TOP) return NULL;
1779       return this;
1780     }
1781     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1782     if ((t_oop != NULL) &&
1783         (t_oop->is_known_instance_field() ||
1784          t_oop->is_ptr_to_boxed_value())) {
1785       PhaseIterGVN *igvn = phase->is_IterGVN();
1786       assert(igvn != NULL, "must be PhaseIterGVN when can_reshape is true");
1787       if (igvn->_worklist.member(opt_mem)) {
1788         // Delay this transformation until memory Phi is processed.
1789         igvn->_worklist.push(this);
1790         return NULL;
1791       }
1792       // Split instance field load through Phi.
1793       Node* result = split_through_phi(phase);
1794       if (result != NULL) return result;
1795 
1796       if (t_oop->is_ptr_to_boxed_value()) {
1797         Node* result = eliminate_autobox(igvn);
1798         if (result != NULL) return result;
1799       }
1800     }
1801   }
1802 
1803   // Is there a dominating load that loads the same value?  Leave
1804   // anything that is not a load of a field/array element (like
1805   // barriers etc.) alone
1806   if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1807     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1808       Node *use = mem->fast_out(i);
1809       if (use != this &&
1810           use->Opcode() == Opcode() &&
1811           use->in(0) != NULL &&
1812           use->in(0) != in(0) &&
1813           use->in(Address) == in(Address)) {
1814         Node* ctl = in(0);
1815         for (int i = 0; i < 10 && ctl != NULL; i++) {
1816           ctl = IfNode::up_one_dom(ctl);
1817           if (ctl == use->in(0)) {
1818             set_req(0, use->in(0));
1819             return this;
1820           }
1821         }
1822       }
1823     }
1824   }
1825 
1826   // Check for prior store with a different base or offset; make Load
1827   // independent.  Skip through any number of them.  Bail out if the stores
1828   // are in an endless dead cycle and report no progress.  This is a key
1829   // transform for Reflection.  However, if after skipping through the Stores
1830   // we can't then fold up against a prior store do NOT do the transform as
1831   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
1832   // array memory alive twice: once for the hoisted Load and again after the
1833   // bypassed Store.  This situation only works if EVERYBODY who does
1834   // anti-dependence work knows how to bypass.  I.e. we need all
1835   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
1836   // the alias index stuff.  So instead, peek through Stores and IFF we can
1837   // fold up, do so.
1838   Node* prev_mem = find_previous_store(phase);
1839   if (prev_mem != NULL) {
1840     Node* value = can_see_arraycopy_value(prev_mem, phase);
1841     if (value != NULL) {
1842       return value;
1843     }
1844   }
1845   // Steps (a), (b):  Walk past independent stores to find an exact match.
1846   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1847     // (c) See if we can fold up on the spot, but don't fold up here.
1848     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1849     // just return a prior value, which is done by Identity calls.
1850     if (can_see_stored_value(prev_mem, phase)) {
1851       // Make ready for step (d):
1852       set_req_X(MemNode::Memory, prev_mem, phase);
1853       return this;
1854     }
1855   }
1856 
1857   return progress ? this : NULL;
1858 }
1859 
1860 // Helper to recognize certain Klass fields which are invariant across
1861 // some group of array types (e.g., int[] or all T[] where T < Object).
1862 const Type*
1863 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1864                                  ciKlass* klass) const {
1865   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1866     // The field is Klass::_modifier_flags.  Return its (constant) value.
1867     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1868     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1869     return TypeInt::make(klass->modifier_flags());
1870   }
1871   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1872     // The field is Klass::_access_flags.  Return its (constant) value.
1873     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1874     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1875     return TypeInt::make(klass->access_flags());
1876   }
1877   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1878     // The field is Klass::_layout_helper.  Return its constant value if known.
1879     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1880     return TypeInt::make(klass->layout_helper());
1881   }
1882 
1883   // No match.
1884   return NULL;
1885 }
1886 
1887 //------------------------------Value-----------------------------------------
1888 const Type* LoadNode::Value(PhaseGVN* phase) const {
1889   // Either input is TOP ==> the result is TOP
1890   Node* mem = in(MemNode::Memory);
1891   const Type *t1 = phase->type(mem);
1892   if (t1 == Type::TOP)  return Type::TOP;
1893   Node* adr = in(MemNode::Address);
1894   const TypePtr* tp = phase->type(adr)->isa_ptr();
1895   if (tp == NULL || tp->empty())  return Type::TOP;
1896   int off = tp->offset();
1897   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1898   Compile* C = phase->C;
1899 
1900   // Try to guess loaded type from pointer type
1901   if (tp->isa_aryptr()) {
1902     const TypeAryPtr* ary = tp->is_aryptr();
1903     const Type* t = ary->elem();
1904 
1905     // Determine whether the reference is beyond the header or not, by comparing
1906     // the offset against the offset of the start of the array's data.
1907     // Different array types begin at slightly different offsets (12 vs. 16).
1908     // We choose T_BYTE as an example base type that is least restrictive
1909     // as to alignment, which will therefore produce the smallest
1910     // possible base offset.
1911     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1912     const bool off_beyond_header = (off >= min_base_off);
1913 
1914     // Try to constant-fold a stable array element.
1915     if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1916       // Make sure the reference is not into the header and the offset is constant
1917       ciObject* aobj = ary->const_oop();
1918       if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1919         int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1920         const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1921                                                                       stable_dimension,
1922                                                                       memory_type(), is_unsigned());
1923         if (con_type != NULL) {
1924           return con_type;
1925         }
1926       }
1927     }
1928 
1929     // Don't do this for integer types. There is only potential profit if
1930     // the element type t is lower than _type; that is, for int types, if _type is
1931     // more restrictive than t.  This only happens here if one is short and the other
1932     // char (both 16 bits), and in those cases we've made an intentional decision
1933     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1934     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1935     //
1936     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1937     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
1938     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1939     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
1940     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1941     // In fact, that could have been the original type of p1, and p1 could have
1942     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1943     // expression (LShiftL quux 3) independently optimized to the constant 8.
1944     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1945         && (_type->isa_vect() == NULL)
1946         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1947       // t might actually be lower than _type, if _type is a unique
1948       // concrete subclass of abstract class t.
1949       if (off_beyond_header || off == Type::OffsetBot) {  // is the offset beyond the header?
1950         const Type* jt = t->join_speculative(_type);
1951         // In any case, do not allow the join, per se, to empty out the type.
1952         if (jt->empty() && !t->empty()) {
1953           // This can happen if a interface-typed array narrows to a class type.
1954           jt = _type;
1955         }
1956 #ifdef ASSERT
1957         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1958           // The pointers in the autobox arrays are always non-null
1959           Node* base = adr->in(AddPNode::Base);
1960           if ((base != NULL) && base->is_DecodeN()) {
1961             // Get LoadN node which loads IntegerCache.cache field
1962             base = base->in(1);
1963           }
1964           if ((base != NULL) && base->is_Con()) {
1965             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1966             if ((base_type != NULL) && base_type->is_autobox_cache()) {
1967               // It could be narrow oop
1968               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1969             }
1970           }
1971         }
1972 #endif
1973         return jt;
1974       }
1975     }
1976   } else if (tp->base() == Type::InstPtr) {
1977     assert( off != Type::OffsetBot ||
1978             // arrays can be cast to Objects
1979             tp->is_oopptr()->klass()->is_java_lang_Object() ||
1980             // unsafe field access may not have a constant offset
1981             C->has_unsafe_access(),
1982             "Field accesses must be precise" );
1983     // For oop loads, we expect the _type to be precise.
1984 
1985     // Optimize loads from constant fields.
1986     const TypeInstPtr* tinst = tp->is_instptr();
1987     ciObject* const_oop = tinst->const_oop();
1988     if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1989       const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1990       if (con_type != NULL) {
1991         return con_type;
1992       }
1993     }
1994   } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
1995     assert( off != Type::OffsetBot ||
1996             // arrays can be cast to Objects
1997             tp->is_klassptr()->klass()->is_java_lang_Object() ||
1998             // also allow array-loading from the primary supertype
1999             // array during subtype checks
2000             Opcode() == Op_LoadKlass,
2001             "Field accesses must be precise" );
2002     // For klass/static loads, we expect the _type to be precise
2003   } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2004     /* With mirrors being an indirect in the Klass*
2005      * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2006      * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2007      *
2008      * So check the type and klass of the node before the LoadP.
2009      */
2010     Node* adr2 = adr->in(MemNode::Address);
2011     const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2012     if (tkls != NULL && !StressReflectiveCode) {
2013       ciKlass* klass = tkls->klass();
2014       if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2015         assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2016         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2017         return TypeInstPtr::make(klass->java_mirror());
2018       }
2019     }
2020   }
2021 
2022   const TypeKlassPtr *tkls = tp->isa_klassptr();
2023   if (tkls != NULL && !StressReflectiveCode) {
2024     ciKlass* klass = tkls->klass();
2025     if (klass->is_loaded() && tkls->klass_is_exact()) {
2026       // We are loading a field from a Klass metaobject whose identity
2027       // is known at compile time (the type is "exact" or "precise").
2028       // Check for fields we know are maintained as constants by the VM.
2029       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2030         // The field is Klass::_super_check_offset.  Return its (constant) value.
2031         // (Folds up type checking code.)
2032         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2033         return TypeInt::make(klass->super_check_offset());
2034       }
2035       // Compute index into primary_supers array
2036       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2037       // Check for overflowing; use unsigned compare to handle the negative case.
2038       if( depth < ciKlass::primary_super_limit() ) {
2039         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
2040         // (Folds up type checking code.)
2041         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2042         ciKlass *ss = klass->super_of_depth(depth);
2043         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2044       }
2045       const Type* aift = load_array_final_field(tkls, klass);
2046       if (aift != NULL)  return aift;
2047     }
2048 
2049     // We can still check if we are loading from the primary_supers array at a
2050     // shallow enough depth.  Even though the klass is not exact, entries less
2051     // than or equal to its super depth are correct.
2052     if (klass->is_loaded() ) {
2053       ciType *inner = klass;
2054       while( inner->is_obj_array_klass() )
2055         inner = inner->as_obj_array_klass()->base_element_type();
2056       if( inner->is_instance_klass() &&
2057           !inner->as_instance_klass()->flags().is_interface() ) {
2058         // Compute index into primary_supers array
2059         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2060         // Check for overflowing; use unsigned compare to handle the negative case.
2061         if( depth < ciKlass::primary_super_limit() &&
2062             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
2063           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
2064           // (Folds up type checking code.)
2065           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2066           ciKlass *ss = klass->super_of_depth(depth);
2067           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2068         }
2069       }
2070     }
2071 
2072     // If the type is enough to determine that the thing is not an array,
2073     // we can give the layout_helper a positive interval type.
2074     // This will help short-circuit some reflective code.
2075     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
2076         && !klass->is_array_klass() // not directly typed as an array
2077         && !klass->is_interface()  // specifically not Serializable & Cloneable
2078         && !klass->is_java_lang_Object()   // not the supertype of all T[]
2079         ) {
2080       // Note:  When interfaces are reliable, we can narrow the interface
2081       // test to (klass != Serializable && klass != Cloneable).
2082       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2083       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2084       // The key property of this type is that it folds up tests
2085       // for array-ness, since it proves that the layout_helper is positive.
2086       // Thus, a generic value like the basic object layout helper works fine.
2087       return TypeInt::make(min_size, max_jint, Type::WidenMin);
2088     }
2089   }
2090 
2091   // If we are loading from a freshly-allocated object, produce a zero,
2092   // if the load is provably beyond the header of the object.
2093   // (Also allow a variable load from a fresh array to produce zero.)
2094   const TypeOopPtr *tinst = tp->isa_oopptr();
2095   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
2096   bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
2097   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2098     Node* value = can_see_stored_value(mem,phase);
2099     if (value != NULL && value->is_Con()) {
2100       assert(value->bottom_type()->higher_equal(_type),"sanity");
2101       return value->bottom_type();
2102     }
2103   }
2104 
2105   bool is_vect = (_type->isa_vect() != NULL);
2106   if (is_instance && !is_vect) {
2107     // If we have an instance type and our memory input is the
2108     // programs's initial memory state, there is no matching store,
2109     // so just return a zero of the appropriate type -
2110     // except if it is vectorized - then we have no zero constant.
2111     Node *mem = in(MemNode::Memory);
2112     if (mem->is_Parm() && mem->in(0)->is_Start()) {
2113       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2114       return Type::get_zero_type(_type->basic_type());
2115     }
2116   }
2117 
2118   Node* alloc = is_new_object_mark_load(phase);
2119   if (alloc != NULL) {
2120     return TypeX::make(markWord::prototype().value());
2121   }
2122 
2123   return _type;
2124 }
2125 
2126 //------------------------------match_edge-------------------------------------
2127 // Do we Match on this edge index or not?  Match only the address.
2128 uint LoadNode::match_edge(uint idx) const {
2129   return idx == MemNode::Address;
2130 }
2131 
2132 //--------------------------LoadBNode::Ideal--------------------------------------
2133 //
2134 //  If the previous store is to the same address as this load,
2135 //  and the value stored was larger than a byte, replace this load
2136 //  with the value stored truncated to a byte.  If no truncation is
2137 //  needed, the replacement is done in LoadNode::Identity().
2138 //
2139 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2140   Node* mem = in(MemNode::Memory);
2141   Node* value = can_see_stored_value(mem,phase);
2142   if (value != NULL) {
2143     Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2144     if (narrow != value) {
2145       return narrow;
2146     }
2147   }
2148   // Identity call will handle the case where truncation is not needed.
2149   return LoadNode::Ideal(phase, can_reshape);
2150 }
2151 
2152 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2153   Node* mem = in(MemNode::Memory);
2154   Node* value = can_see_stored_value(mem,phase);
2155   if (value != NULL && value->is_Con() &&
2156       !value->bottom_type()->higher_equal(_type)) {
2157     // If the input to the store does not fit with the load's result type,
2158     // it must be truncated. We can't delay until Ideal call since
2159     // a singleton Value is needed for split_thru_phi optimization.
2160     int con = value->get_int();
2161     return TypeInt::make((con << 24) >> 24);
2162   }
2163   return LoadNode::Value(phase);
2164 }
2165 
2166 //--------------------------LoadUBNode::Ideal-------------------------------------
2167 //
2168 //  If the previous store is to the same address as this load,
2169 //  and the value stored was larger than a byte, replace this load
2170 //  with the value stored truncated to a byte.  If no truncation is
2171 //  needed, the replacement is done in LoadNode::Identity().
2172 //
2173 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2174   Node* mem = in(MemNode::Memory);
2175   Node* value = can_see_stored_value(mem, phase);
2176   if (value != NULL) {
2177     Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2178     if (narrow != value) {
2179       return narrow;
2180     }
2181   }
2182   // Identity call will handle the case where truncation is not needed.
2183   return LoadNode::Ideal(phase, can_reshape);
2184 }
2185 
2186 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2187   Node* mem = in(MemNode::Memory);
2188   Node* value = can_see_stored_value(mem,phase);
2189   if (value != NULL && value->is_Con() &&
2190       !value->bottom_type()->higher_equal(_type)) {
2191     // If the input to the store does not fit with the load's result type,
2192     // it must be truncated. We can't delay until Ideal call since
2193     // a singleton Value is needed for split_thru_phi optimization.
2194     int con = value->get_int();
2195     return TypeInt::make(con & 0xFF);
2196   }
2197   return LoadNode::Value(phase);
2198 }
2199 
2200 //--------------------------LoadUSNode::Ideal-------------------------------------
2201 //
2202 //  If the previous store is to the same address as this load,
2203 //  and the value stored was larger than a char, replace this load
2204 //  with the value stored truncated to a char.  If no truncation is
2205 //  needed, the replacement is done in LoadNode::Identity().
2206 //
2207 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2208   Node* mem = in(MemNode::Memory);
2209   Node* value = can_see_stored_value(mem,phase);
2210   if (value != NULL) {
2211     Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2212     if (narrow != value) {
2213       return narrow;
2214     }
2215   }
2216   // Identity call will handle the case where truncation is not needed.
2217   return LoadNode::Ideal(phase, can_reshape);
2218 }
2219 
2220 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2221   Node* mem = in(MemNode::Memory);
2222   Node* value = can_see_stored_value(mem,phase);
2223   if (value != NULL && value->is_Con() &&
2224       !value->bottom_type()->higher_equal(_type)) {
2225     // If the input to the store does not fit with the load's result type,
2226     // it must be truncated. We can't delay until Ideal call since
2227     // a singleton Value is needed for split_thru_phi optimization.
2228     int con = value->get_int();
2229     return TypeInt::make(con & 0xFFFF);
2230   }
2231   return LoadNode::Value(phase);
2232 }
2233 
2234 //--------------------------LoadSNode::Ideal--------------------------------------
2235 //
2236 //  If the previous store is to the same address as this load,
2237 //  and the value stored was larger than a short, replace this load
2238 //  with the value stored truncated to a short.  If no truncation is
2239 //  needed, the replacement is done in LoadNode::Identity().
2240 //
2241 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2242   Node* mem = in(MemNode::Memory);
2243   Node* value = can_see_stored_value(mem,phase);
2244   if (value != NULL) {
2245     Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2246     if (narrow != value) {
2247       return narrow;
2248     }
2249   }
2250   // Identity call will handle the case where truncation is not needed.
2251   return LoadNode::Ideal(phase, can_reshape);
2252 }
2253 
2254 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2255   Node* mem = in(MemNode::Memory);
2256   Node* value = can_see_stored_value(mem,phase);
2257   if (value != NULL && value->is_Con() &&
2258       !value->bottom_type()->higher_equal(_type)) {
2259     // If the input to the store does not fit with the load's result type,
2260     // it must be truncated. We can't delay until Ideal call since
2261     // a singleton Value is needed for split_thru_phi optimization.
2262     int con = value->get_int();
2263     return TypeInt::make((con << 16) >> 16);
2264   }
2265   return LoadNode::Value(phase);
2266 }
2267 
2268 //=============================================================================
2269 //----------------------------LoadKlassNode::make------------------------------
2270 // Polymorphic factory method:
2271 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2272   // sanity check the alias category against the created node type
2273   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2274   assert(adr_type != NULL, "expecting TypeKlassPtr");
2275 #ifdef _LP64
2276   if (adr_type->is_ptr_to_narrowklass()) {
2277     assert(UseCompressedClassPointers, "no compressed klasses");
2278     Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2279     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2280   }
2281 #endif
2282   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2283   return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2284 }
2285 
2286 //------------------------------Value------------------------------------------
2287 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2288   return klass_value_common(phase);
2289 }
2290 
2291 // In most cases, LoadKlassNode does not have the control input set. If the control
2292 // input is set, it must not be removed (by LoadNode::Ideal()).
2293 bool LoadKlassNode::can_remove_control() const {
2294   return false;
2295 }
2296 
2297 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2298   // Either input is TOP ==> the result is TOP
2299   const Type *t1 = phase->type( in(MemNode::Memory) );
2300   if (t1 == Type::TOP)  return Type::TOP;
2301   Node *adr = in(MemNode::Address);
2302   const Type *t2 = phase->type( adr );
2303   if (t2 == Type::TOP)  return Type::TOP;
2304   const TypePtr *tp = t2->is_ptr();
2305   if (TypePtr::above_centerline(tp->ptr()) ||
2306       tp->ptr() == TypePtr::Null)  return Type::TOP;
2307 
2308   // Return a more precise klass, if possible
2309   const TypeInstPtr *tinst = tp->isa_instptr();
2310   if (tinst != NULL) {
2311     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2312     int offset = tinst->offset();
2313     if (ik == phase->C->env()->Class_klass()
2314         && (offset == java_lang_Class::klass_offset() ||
2315             offset == java_lang_Class::array_klass_offset())) {
2316       // We are loading a special hidden field from a Class mirror object,
2317       // the field which points to the VM's Klass metaobject.
2318       ciType* t = tinst->java_mirror_type();
2319       // java_mirror_type returns non-null for compile-time Class constants.
2320       if (t != NULL) {
2321         // constant oop => constant klass
2322         if (offset == java_lang_Class::array_klass_offset()) {
2323           if (t->is_void()) {
2324             // We cannot create a void array.  Since void is a primitive type return null
2325             // klass.  Users of this result need to do a null check on the returned klass.
2326             return TypePtr::NULL_PTR;
2327           }
2328           return TypeKlassPtr::make(ciArrayKlass::make(t));
2329         }
2330         if (!t->is_klass()) {
2331           // a primitive Class (e.g., int.class) has NULL for a klass field
2332           return TypePtr::NULL_PTR;
2333         }
2334         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2335         return TypeKlassPtr::make(t->as_klass());
2336       }
2337       // non-constant mirror, so we can't tell what's going on
2338     }
2339     if( !ik->is_loaded() )
2340       return _type;             // Bail out if not loaded
2341     if (offset == oopDesc::klass_offset_in_bytes()) {
2342       if (tinst->klass_is_exact()) {
2343         return TypeKlassPtr::make(ik);
2344       }
2345       // See if we can become precise: no subklasses and no interface
2346       // (Note:  We need to support verified interfaces.)
2347       if (!ik->is_interface() && !ik->has_subklass()) {
2348         // Add a dependence; if any subclass added we need to recompile
2349         if (!ik->is_final()) {
2350           // %%% should use stronger assert_unique_concrete_subtype instead
2351           phase->C->dependencies()->assert_leaf_type(ik);
2352         }
2353         // Return precise klass
2354         return TypeKlassPtr::make(ik);
2355       }
2356 
2357       // Return root of possible klass
2358       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2359     }
2360   }
2361 
2362   // Check for loading klass from an array
2363   const TypeAryPtr *tary = tp->isa_aryptr();
2364   if( tary != NULL ) {
2365     ciKlass *tary_klass = tary->klass();
2366     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
2367         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2368       if (tary->klass_is_exact()) {
2369         return TypeKlassPtr::make(tary_klass);
2370       }
2371       ciArrayKlass *ak = tary->klass()->as_array_klass();
2372       // If the klass is an object array, we defer the question to the
2373       // array component klass.
2374       if( ak->is_obj_array_klass() ) {
2375         assert( ak->is_loaded(), "" );
2376         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2377         if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2378           ciInstanceKlass* ik = base_k->as_instance_klass();
2379           // See if we can become precise: no subklasses and no interface
2380           if (!ik->is_interface() && !ik->has_subklass()) {
2381             // Add a dependence; if any subclass added we need to recompile
2382             if (!ik->is_final()) {
2383               phase->C->dependencies()->assert_leaf_type(ik);
2384             }
2385             // Return precise array klass
2386             return TypeKlassPtr::make(ak);
2387           }
2388         }
2389         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2390       } else {                  // Found a type-array?
2391         assert( ak->is_type_array_klass(), "" );
2392         return TypeKlassPtr::make(ak); // These are always precise
2393       }
2394     }
2395   }
2396 
2397   // Check for loading klass from an array klass
2398   const TypeKlassPtr *tkls = tp->isa_klassptr();
2399   if (tkls != NULL && !StressReflectiveCode) {
2400     ciKlass* klass = tkls->klass();
2401     if( !klass->is_loaded() )
2402       return _type;             // Bail out if not loaded
2403     if( klass->is_obj_array_klass() &&
2404         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2405       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2406       // // Always returning precise element type is incorrect,
2407       // // e.g., element type could be object and array may contain strings
2408       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2409 
2410       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2411       // according to the element type's subclassing.
2412       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2413     }
2414     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2415         tkls->offset() == in_bytes(Klass::super_offset())) {
2416       ciKlass* sup = klass->as_instance_klass()->super();
2417       // The field is Klass::_super.  Return its (constant) value.
2418       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2419       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2420     }
2421   }
2422 
2423   // Bailout case
2424   return LoadNode::Value(phase);
2425 }
2426 
2427 //------------------------------Identity---------------------------------------
2428 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2429 // Also feed through the klass in Allocate(...klass...)._klass.
2430 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2431   return klass_identity_common(phase);
2432 }
2433 
2434 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2435   Node* x = LoadNode::Identity(phase);
2436   if (x != this)  return x;
2437 
2438   // Take apart the address into an oop and and offset.
2439   // Return 'this' if we cannot.
2440   Node*    adr    = in(MemNode::Address);
2441   intptr_t offset = 0;
2442   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2443   if (base == NULL)     return this;
2444   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2445   if (toop == NULL)     return this;
2446 
2447   // Step over potential GC barrier for OopHandle resolve
2448   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2449   if (bs->is_gc_barrier_node(base)) {
2450     base = bs->step_over_gc_barrier(base);
2451   }
2452 
2453   // We can fetch the klass directly through an AllocateNode.
2454   // This works even if the klass is not constant (clone or newArray).
2455   if (offset == oopDesc::klass_offset_in_bytes()) {
2456     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2457     if (allocated_klass != NULL) {
2458       return allocated_klass;
2459     }
2460   }
2461 
2462   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2463   // See inline_native_Class_query for occurrences of these patterns.
2464   // Java Example:  x.getClass().isAssignableFrom(y)
2465   //
2466   // This improves reflective code, often making the Class
2467   // mirror go completely dead.  (Current exception:  Class
2468   // mirrors may appear in debug info, but we could clean them out by
2469   // introducing a new debug info operator for Klass.java_mirror).
2470 
2471   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2472       && offset == java_lang_Class::klass_offset()) {
2473     if (base->is_Load()) {
2474       Node* base2 = base->in(MemNode::Address);
2475       if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2476         Node* adr2 = base2->in(MemNode::Address);
2477         const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2478         if (tkls != NULL && !tkls->empty()
2479             && (tkls->klass()->is_instance_klass() ||
2480               tkls->klass()->is_array_klass())
2481             && adr2->is_AddP()
2482            ) {
2483           int mirror_field = in_bytes(Klass::java_mirror_offset());
2484           if (tkls->offset() == mirror_field) {
2485             return adr2->in(AddPNode::Base);
2486           }
2487         }
2488       }
2489     }
2490   }
2491 
2492   return this;
2493 }
2494 
2495 
2496 //------------------------------Value------------------------------------------
2497 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2498   const Type *t = klass_value_common(phase);
2499   if (t == Type::TOP)
2500     return t;
2501 
2502   return t->make_narrowklass();
2503 }
2504 
2505 //------------------------------Identity---------------------------------------
2506 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2507 // Also feed through the klass in Allocate(...klass...)._klass.
2508 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2509   Node *x = klass_identity_common(phase);
2510 
2511   const Type *t = phase->type( x );
2512   if( t == Type::TOP ) return x;
2513   if( t->isa_narrowklass()) return x;
2514   assert (!t->isa_narrowoop(), "no narrow oop here");
2515 
2516   return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2517 }
2518 
2519 //------------------------------Value-----------------------------------------
2520 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2521   // Either input is TOP ==> the result is TOP
2522   const Type *t1 = phase->type( in(MemNode::Memory) );
2523   if( t1 == Type::TOP ) return Type::TOP;
2524   Node *adr = in(MemNode::Address);
2525   const Type *t2 = phase->type( adr );
2526   if( t2 == Type::TOP ) return Type::TOP;
2527   const TypePtr *tp = t2->is_ptr();
2528   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2529   const TypeAryPtr *tap = tp->isa_aryptr();
2530   if( !tap ) return _type;
2531   return tap->size();
2532 }
2533 
2534 //-------------------------------Ideal---------------------------------------
2535 // Feed through the length in AllocateArray(...length...)._length.
2536 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2537   Node* p = MemNode::Ideal_common(phase, can_reshape);
2538   if (p)  return (p == NodeSentinel) ? NULL : p;
2539 
2540   // Take apart the address into an oop and and offset.
2541   // Return 'this' if we cannot.
2542   Node*    adr    = in(MemNode::Address);
2543   intptr_t offset = 0;
2544   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2545   if (base == NULL)     return NULL;
2546   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2547   if (tary == NULL)     return NULL;
2548 
2549   // We can fetch the length directly through an AllocateArrayNode.
2550   // This works even if the length is not constant (clone or newArray).
2551   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2552     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2553     if (alloc != NULL) {
2554       Node* allocated_length = alloc->Ideal_length();
2555       Node* len = alloc->make_ideal_length(tary, phase);
2556       if (allocated_length != len) {
2557         // New CastII improves on this.
2558         return len;
2559       }
2560     }
2561   }
2562 
2563   return NULL;
2564 }
2565 
2566 //------------------------------Identity---------------------------------------
2567 // Feed through the length in AllocateArray(...length...)._length.
2568 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2569   Node* x = LoadINode::Identity(phase);
2570   if (x != this)  return x;
2571 
2572   // Take apart the address into an oop and and offset.
2573   // Return 'this' if we cannot.
2574   Node*    adr    = in(MemNode::Address);
2575   intptr_t offset = 0;
2576   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2577   if (base == NULL)     return this;
2578   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2579   if (tary == NULL)     return this;
2580 
2581   // We can fetch the length directly through an AllocateArrayNode.
2582   // This works even if the length is not constant (clone or newArray).
2583   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2584     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2585     if (alloc != NULL) {
2586       Node* allocated_length = alloc->Ideal_length();
2587       // Do not allow make_ideal_length to allocate a CastII node.
2588       Node* len = alloc->make_ideal_length(tary, phase, false);
2589       if (allocated_length == len) {
2590         // Return allocated_length only if it would not be improved by a CastII.
2591         return allocated_length;
2592       }
2593     }
2594   }
2595 
2596   return this;
2597 
2598 }
2599 
2600 //=============================================================================
2601 //---------------------------StoreNode::make-----------------------------------
2602 // Polymorphic factory method:
2603 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2604   assert((mo == unordered || mo == release), "unexpected");
2605   Compile* C = gvn.C;
2606   assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2607          ctl != NULL, "raw memory operations should have control edge");
2608 
2609   switch (bt) {
2610   case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2611   case T_BYTE:    return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2612   case T_INT:     return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2613   case T_CHAR:
2614   case T_SHORT:   return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2615   case T_LONG:    return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2616   case T_FLOAT:   return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2617   case T_DOUBLE:  return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2618   case T_METADATA:
2619   case T_ADDRESS:
2620   case T_OBJECT:
2621 #ifdef _LP64
2622     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2623       val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2624       return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2625     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2626                (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2627                 adr->bottom_type()->isa_rawptr())) {
2628       val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2629       return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2630     }
2631 #endif
2632     {
2633       return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2634     }
2635   default:
2636     ShouldNotReachHere();
2637     return (StoreNode*)NULL;
2638   }
2639 }
2640 
2641 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2642   bool require_atomic = true;
2643   return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2644 }
2645 
2646 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2647   bool require_atomic = true;
2648   return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2649 }
2650 
2651 
2652 //--------------------------bottom_type----------------------------------------
2653 const Type *StoreNode::bottom_type() const {
2654   return Type::MEMORY;
2655 }
2656 
2657 //------------------------------hash-------------------------------------------
2658 uint StoreNode::hash() const {
2659   // unroll addition of interesting fields
2660   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2661 
2662   // Since they are not commoned, do not hash them:
2663   return NO_HASH;
2664 }
2665 
2666 //------------------------------Ideal------------------------------------------
2667 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2668 // When a store immediately follows a relevant allocation/initialization,
2669 // try to capture it into the initialization, or hoist it above.
2670 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2671   Node* p = MemNode::Ideal_common(phase, can_reshape);
2672   if (p)  return (p == NodeSentinel) ? NULL : p;
2673 
2674   Node* mem     = in(MemNode::Memory);
2675   Node* address = in(MemNode::Address);
2676   Node* value   = in(MemNode::ValueIn);
2677   // Back-to-back stores to same address?  Fold em up.  Generally
2678   // unsafe if I have intervening uses...  Also disallowed for StoreCM
2679   // since they must follow each StoreP operation.  Redundant StoreCMs
2680   // are eliminated just before matching in final_graph_reshape.
2681   {
2682     Node* st = mem;
2683     // If Store 'st' has more than one use, we cannot fold 'st' away.
2684     // For example, 'st' might be the final state at a conditional
2685     // return.  Or, 'st' might be used by some node which is live at
2686     // the same time 'st' is live, which might be unschedulable.  So,
2687     // require exactly ONE user until such time as we clone 'mem' for
2688     // each of 'mem's uses (thus making the exactly-1-user-rule hold
2689     // true).
2690     while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2691       // Looking at a dead closed cycle of memory?
2692       assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2693       assert(Opcode() == st->Opcode() ||
2694              st->Opcode() == Op_StoreVector ||
2695              Opcode() == Op_StoreVector ||
2696              st->Opcode() == Op_StoreVectorScatter ||
2697              Opcode() == Op_StoreVectorScatter ||
2698              phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2699              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2700              (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2701              (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2702              "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2703 
2704       if (st->in(MemNode::Address)->eqv_uncast(address) &&
2705           st->as_Store()->memory_size() <= this->memory_size()) {
2706         Node* use = st->raw_out(0);
2707         if (phase->is_IterGVN()) {
2708           phase->is_IterGVN()->rehash_node_delayed(use);
2709         }
2710         // It's OK to do this in the parser, since DU info is always accurate,
2711         // and the parser always refers to nodes via SafePointNode maps.
2712         use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
2713         return this;
2714       }
2715       st = st->in(MemNode::Memory);
2716     }
2717   }
2718 
2719 
2720   // Capture an unaliased, unconditional, simple store into an initializer.
2721   // Or, if it is independent of the allocation, hoist it above the allocation.
2722   if (ReduceFieldZeroing && /*can_reshape &&*/
2723       mem->is_Proj() && mem->in(0)->is_Initialize()) {
2724     InitializeNode* init = mem->in(0)->as_Initialize();
2725     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2726     if (offset > 0) {
2727       Node* moved = init->capture_store(this, offset, phase, can_reshape);
2728       // If the InitializeNode captured me, it made a raw copy of me,
2729       // and I need to disappear.
2730       if (moved != NULL) {
2731         // %%% hack to ensure that Ideal returns a new node:
2732         mem = MergeMemNode::make(mem);
2733         return mem;             // fold me away
2734       }
2735     }
2736   }
2737 
2738   // Fold reinterpret cast into memory operation:
2739   //    StoreX mem (MoveY2X v) => StoreY mem v
2740   if (value->is_Move()) {
2741     const Type* vt = value->in(1)->bottom_type();
2742     if (has_reinterpret_variant(vt)) {
2743       if (phase->C->post_loop_opts_phase()) {
2744         return convert_to_reinterpret_store(*phase, value->in(1), vt);
2745       } else {
2746         phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
2747       }
2748     }
2749   }
2750 
2751   return NULL;                  // No further progress
2752 }
2753 
2754 //------------------------------Value-----------------------------------------
2755 const Type* StoreNode::Value(PhaseGVN* phase) const {
2756   // Either input is TOP ==> the result is TOP
2757   const Type *t1 = phase->type( in(MemNode::Memory) );
2758   if( t1 == Type::TOP ) return Type::TOP;
2759   const Type *t2 = phase->type( in(MemNode::Address) );
2760   if( t2 == Type::TOP ) return Type::TOP;
2761   const Type *t3 = phase->type( in(MemNode::ValueIn) );
2762   if( t3 == Type::TOP ) return Type::TOP;
2763   return Type::MEMORY;
2764 }
2765 
2766 //------------------------------Identity---------------------------------------
2767 // Remove redundant stores:
2768 //   Store(m, p, Load(m, p)) changes to m.
2769 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2770 Node* StoreNode::Identity(PhaseGVN* phase) {
2771   Node* mem = in(MemNode::Memory);
2772   Node* adr = in(MemNode::Address);
2773   Node* val = in(MemNode::ValueIn);
2774 
2775   Node* result = this;
2776 
2777   // Load then Store?  Then the Store is useless
2778   if (val->is_Load() &&
2779       val->in(MemNode::Address)->eqv_uncast(adr) &&
2780       val->in(MemNode::Memory )->eqv_uncast(mem) &&
2781       val->as_Load()->store_Opcode() == Opcode()) {
2782     result = mem;
2783   }
2784 
2785   // Two stores in a row of the same value?
2786   if (result == this &&
2787       mem->is_Store() &&
2788       mem->in(MemNode::Address)->eqv_uncast(adr) &&
2789       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2790       mem->Opcode() == Opcode()) {
2791     result = mem;
2792   }
2793 
2794   // Store of zero anywhere into a freshly-allocated object?
2795   // Then the store is useless.
2796   // (It must already have been captured by the InitializeNode.)
2797   if (result == this &&
2798       ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2799     // a newly allocated object is already all-zeroes everywhere
2800     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2801       result = mem;
2802     }
2803 
2804     if (result == this) {
2805       // the store may also apply to zero-bits in an earlier object
2806       Node* prev_mem = find_previous_store(phase);
2807       // Steps (a), (b):  Walk past independent stores to find an exact match.
2808       if (prev_mem != NULL) {
2809         Node* prev_val = can_see_stored_value(prev_mem, phase);
2810         if (prev_val != NULL && prev_val == val) {
2811           // prev_val and val might differ by a cast; it would be good
2812           // to keep the more informative of the two.
2813           result = mem;
2814         }
2815       }
2816     }
2817   }
2818 
2819   PhaseIterGVN* igvn = phase->is_IterGVN();
2820   if (result != this && igvn != NULL) {
2821     MemBarNode* trailing = trailing_membar();
2822     if (trailing != NULL) {
2823 #ifdef ASSERT
2824       const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2825       assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2826 #endif
2827       trailing->remove(igvn);
2828     }
2829   }
2830 
2831   return result;
2832 }
2833 
2834 //------------------------------match_edge-------------------------------------
2835 // Do we Match on this edge index or not?  Match only memory & value
2836 uint StoreNode::match_edge(uint idx) const {
2837   return idx == MemNode::Address || idx == MemNode::ValueIn;
2838 }
2839 
2840 //------------------------------cmp--------------------------------------------
2841 // Do not common stores up together.  They generally have to be split
2842 // back up anyways, so do not bother.
2843 bool StoreNode::cmp( const Node &n ) const {
2844   return (&n == this);          // Always fail except on self
2845 }
2846 
2847 //------------------------------Ideal_masked_input-----------------------------
2848 // Check for a useless mask before a partial-word store
2849 // (StoreB ... (AndI valIn conIa) )
2850 // If (conIa & mask == mask) this simplifies to
2851 // (StoreB ... (valIn) )
2852 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2853   Node *val = in(MemNode::ValueIn);
2854   if( val->Opcode() == Op_AndI ) {
2855     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2856     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2857       set_req_X(MemNode::ValueIn, val->in(1), phase);
2858       return this;
2859     }
2860   }
2861   return NULL;
2862 }
2863 
2864 
2865 //------------------------------Ideal_sign_extended_input----------------------
2866 // Check for useless sign-extension before a partial-word store
2867 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2868 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
2869 // (StoreB ... (valIn) )
2870 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2871   Node *val = in(MemNode::ValueIn);
2872   if( val->Opcode() == Op_RShiftI ) {
2873     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2874     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2875       Node *shl = val->in(1);
2876       if( shl->Opcode() == Op_LShiftI ) {
2877         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2878         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2879           set_req_X(MemNode::ValueIn, shl->in(1), phase);
2880           return this;
2881         }
2882       }
2883     }
2884   }
2885   return NULL;
2886 }
2887 
2888 //------------------------------value_never_loaded-----------------------------------
2889 // Determine whether there are any possible loads of the value stored.
2890 // For simplicity, we actually check if there are any loads from the
2891 // address stored to, not just for loads of the value stored by this node.
2892 //
2893 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2894   Node *adr = in(Address);
2895   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2896   if (adr_oop == NULL)
2897     return false;
2898   if (!adr_oop->is_known_instance_field())
2899     return false; // if not a distinct instance, there may be aliases of the address
2900   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2901     Node *use = adr->fast_out(i);
2902     if (use->is_Load() || use->is_LoadStore()) {
2903       return false;
2904     }
2905   }
2906   return true;
2907 }
2908 
2909 MemBarNode* StoreNode::trailing_membar() const {
2910   if (is_release()) {
2911     MemBarNode* trailing_mb = NULL;
2912     for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2913       Node* u = fast_out(i);
2914       if (u->is_MemBar()) {
2915         if (u->as_MemBar()->trailing_store()) {
2916           assert(u->Opcode() == Op_MemBarVolatile, "");
2917           assert(trailing_mb == NULL, "only one");
2918           trailing_mb = u->as_MemBar();
2919 #ifdef ASSERT
2920           Node* leading = u->as_MemBar()->leading_membar();
2921           assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2922           assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2923           assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2924 #endif
2925         } else {
2926           assert(u->as_MemBar()->standalone(), "");
2927         }
2928       }
2929     }
2930     return trailing_mb;
2931   }
2932   return NULL;
2933 }
2934 
2935 
2936 //=============================================================================
2937 //------------------------------Ideal------------------------------------------
2938 // If the store is from an AND mask that leaves the low bits untouched, then
2939 // we can skip the AND operation.  If the store is from a sign-extension
2940 // (a left shift, then right shift) we can skip both.
2941 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2942   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2943   if( progress != NULL ) return progress;
2944 
2945   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2946   if( progress != NULL ) return progress;
2947 
2948   // Finally check the default case
2949   return StoreNode::Ideal(phase, can_reshape);
2950 }
2951 
2952 //=============================================================================
2953 //------------------------------Ideal------------------------------------------
2954 // If the store is from an AND mask that leaves the low bits untouched, then
2955 // we can skip the AND operation
2956 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2957   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2958   if( progress != NULL ) return progress;
2959 
2960   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2961   if( progress != NULL ) return progress;
2962 
2963   // Finally check the default case
2964   return StoreNode::Ideal(phase, can_reshape);
2965 }
2966 
2967 //=============================================================================
2968 //------------------------------Identity---------------------------------------
2969 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2970   // No need to card mark when storing a null ptr
2971   Node* my_store = in(MemNode::OopStore);
2972   if (my_store->is_Store()) {
2973     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2974     if( t1 == TypePtr::NULL_PTR ) {
2975       return in(MemNode::Memory);
2976     }
2977   }
2978   return this;
2979 }
2980 
2981 //=============================================================================
2982 //------------------------------Ideal---------------------------------------
2983 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2984   Node* progress = StoreNode::Ideal(phase, can_reshape);
2985   if (progress != NULL) return progress;
2986 
2987   Node* my_store = in(MemNode::OopStore);
2988   if (my_store->is_MergeMem()) {
2989     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2990     set_req_X(MemNode::OopStore, mem, phase);
2991     return this;
2992   }
2993 
2994   return NULL;
2995 }
2996 
2997 //------------------------------Value-----------------------------------------
2998 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2999   // Either input is TOP ==> the result is TOP (checked in StoreNode::Value).
3000   // If extra input is TOP ==> the result is TOP
3001   const Type* t = phase->type(in(MemNode::OopStore));
3002   if (t == Type::TOP) {
3003     return Type::TOP;
3004   }
3005   return StoreNode::Value(phase);
3006 }
3007 
3008 
3009 //=============================================================================
3010 //----------------------------------SCMemProjNode------------------------------
3011 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3012 {
3013   if (in(0) == NULL || phase->type(in(0)) == Type::TOP) {
3014     return Type::TOP;
3015   }
3016   return bottom_type();
3017 }
3018 
3019 //=============================================================================
3020 //----------------------------------LoadStoreNode------------------------------
3021 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3022   : Node(required),
3023     _type(rt),
3024     _adr_type(at),
3025     _barrier_data(0)
3026 {
3027   init_req(MemNode::Control, c  );
3028   init_req(MemNode::Memory , mem);
3029   init_req(MemNode::Address, adr);
3030   init_req(MemNode::ValueIn, val);
3031   init_class_id(Class_LoadStore);
3032 }
3033 
3034 //------------------------------Value-----------------------------------------
3035 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3036   // Either input is TOP ==> the result is TOP
3037   if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3038     return Type::TOP;
3039   }
3040   const Type* t = phase->type(in(MemNode::Memory));
3041   if (t == Type::TOP) {
3042     return Type::TOP;
3043   }
3044   t = phase->type(in(MemNode::Address));
3045   if (t == Type::TOP) {
3046     return Type::TOP;
3047   }
3048   t = phase->type(in(MemNode::ValueIn));
3049   if (t == Type::TOP) {
3050     return Type::TOP;
3051   }
3052   return bottom_type();
3053 }
3054 
3055 uint LoadStoreNode::ideal_reg() const {
3056   return _type->ideal_reg();
3057 }
3058 
3059 bool LoadStoreNode::result_not_used() const {
3060   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
3061     Node *x = fast_out(i);
3062     if (x->Opcode() == Op_SCMemProj) continue;
3063     return false;
3064   }
3065   return true;
3066 }
3067 
3068 MemBarNode* LoadStoreNode::trailing_membar() const {
3069   MemBarNode* trailing = NULL;
3070   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3071     Node* u = fast_out(i);
3072     if (u->is_MemBar()) {
3073       if (u->as_MemBar()->trailing_load_store()) {
3074         assert(u->Opcode() == Op_MemBarAcquire, "");
3075         assert(trailing == NULL, "only one");
3076         trailing = u->as_MemBar();
3077 #ifdef ASSERT
3078         Node* leading = trailing->leading_membar();
3079         assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3080         assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3081         assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3082 #endif
3083       } else {
3084         assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3085       }
3086     }
3087   }
3088 
3089   return trailing;
3090 }
3091 
3092 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3093 
3094 //=============================================================================
3095 //----------------------------------LoadStoreConditionalNode--------------------
3096 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
3097   init_req(ExpectedIn, ex );
3098 }
3099 
3100 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3101   // Either input is TOP ==> the result is TOP
3102   const Type* t = phase->type(in(ExpectedIn));
3103   if (t == Type::TOP) {
3104     return Type::TOP;
3105   }
3106   return LoadStoreNode::Value(phase);
3107 }
3108 
3109 //=============================================================================
3110 //-------------------------------adr_type--------------------------------------
3111 const TypePtr* ClearArrayNode::adr_type() const {
3112   Node *adr = in(3);
3113   if (adr == NULL)  return NULL; // node is dead
3114   return MemNode::calculate_adr_type(adr->bottom_type());
3115 }
3116 
3117 //------------------------------match_edge-------------------------------------
3118 // Do we Match on this edge index or not?  Do not match memory
3119 uint ClearArrayNode::match_edge(uint idx) const {
3120   return idx > 1;
3121 }
3122 
3123 //------------------------------Identity---------------------------------------
3124 // Clearing a zero length array does nothing
3125 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3126   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
3127 }
3128 
3129 //------------------------------Idealize---------------------------------------
3130 // Clearing a short array is faster with stores
3131 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3132   // Already know this is a large node, do not try to ideal it
3133   if (!IdealizeClearArrayNode || _is_large) return NULL;
3134 
3135   const int unit = BytesPerLong;
3136   const TypeX* t = phase->type(in(2))->isa_intptr_t();
3137   if (!t)  return NULL;
3138   if (!t->is_con())  return NULL;
3139   intptr_t raw_count = t->get_con();
3140   intptr_t size = raw_count;
3141   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3142   // Clearing nothing uses the Identity call.
3143   // Negative clears are possible on dead ClearArrays
3144   // (see jck test stmt114.stmt11402.val).
3145   if (size <= 0 || size % unit != 0)  return NULL;
3146   intptr_t count = size / unit;
3147   // Length too long; communicate this to matchers and assemblers.
3148   // Assemblers are responsible to produce fast hardware clears for it.
3149   if (size > InitArrayShortSize) {
3150     return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3151   } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3152     return NULL;
3153   }
3154   Node *mem = in(1);
3155   if( phase->type(mem)==Type::TOP ) return NULL;
3156   Node *adr = in(3);
3157   const Type* at = phase->type(adr);
3158   if( at==Type::TOP ) return NULL;
3159   const TypePtr* atp = at->isa_ptr();
3160   // adjust atp to be the correct array element address type
3161   if (atp == NULL)  atp = TypePtr::BOTTOM;
3162   else              atp = atp->add_offset(Type::OffsetBot);
3163   // Get base for derived pointer purposes
3164   if( adr->Opcode() != Op_AddP ) Unimplemented();
3165   Node *base = adr->in(1);
3166 
3167   Node *zero = phase->makecon(TypeLong::ZERO);
3168   Node *off  = phase->MakeConX(BytesPerLong);
3169   mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3170   count--;
3171   while( count-- ) {
3172     mem = phase->transform(mem);
3173     adr = phase->transform(new AddPNode(base,adr,off));
3174     mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3175   }
3176   return mem;
3177 }
3178 
3179 //----------------------------step_through----------------------------------
3180 // Return allocation input memory edge if it is different instance
3181 // or itself if it is the one we are looking for.
3182 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3183   Node* n = *np;
3184   assert(n->is_ClearArray(), "sanity");
3185   intptr_t offset;
3186   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3187   // This method is called only before Allocate nodes are expanded
3188   // during macro nodes expansion. Before that ClearArray nodes are
3189   // only generated in PhaseMacroExpand::generate_arraycopy() (before
3190   // Allocate nodes are expanded) which follows allocations.
3191   assert(alloc != NULL, "should have allocation");
3192   if (alloc->_idx == instance_id) {
3193     // Can not bypass initialization of the instance we are looking for.
3194     return false;
3195   }
3196   // Otherwise skip it.
3197   InitializeNode* init = alloc->initialization();
3198   if (init != NULL)
3199     *np = init->in(TypeFunc::Memory);
3200   else
3201     *np = alloc->in(TypeFunc::Memory);
3202   return true;
3203 }
3204 
3205 //----------------------------clear_memory-------------------------------------
3206 // Generate code to initialize object storage to zero.
3207 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3208                                    intptr_t start_offset,
3209                                    Node* end_offset,
3210                                    PhaseGVN* phase) {
3211   intptr_t offset = start_offset;
3212 
3213   int unit = BytesPerLong;
3214   if ((offset % unit) != 0) {
3215     Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3216     adr = phase->transform(adr);
3217     const TypePtr* atp = TypeRawPtr::BOTTOM;
3218     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3219     mem = phase->transform(mem);
3220     offset += BytesPerInt;
3221   }
3222   assert((offset % unit) == 0, "");
3223 
3224   // Initialize the remaining stuff, if any, with a ClearArray.
3225   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
3226 }
3227 
3228 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3229                                    Node* start_offset,
3230                                    Node* end_offset,
3231                                    PhaseGVN* phase) {
3232   if (start_offset == end_offset) {
3233     // nothing to do
3234     return mem;
3235   }
3236 
3237   int unit = BytesPerLong;
3238   Node* zbase = start_offset;
3239   Node* zend  = end_offset;
3240 
3241   // Scale to the unit required by the CPU:
3242   if (!Matcher::init_array_count_is_in_bytes) {
3243     Node* shift = phase->intcon(exact_log2(unit));
3244     zbase = phase->transform(new URShiftXNode(zbase, shift) );
3245     zend  = phase->transform(new URShiftXNode(zend,  shift) );
3246   }
3247 
3248   // Bulk clear double-words
3249   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3250   Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3251   mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3252   return phase->transform(mem);
3253 }
3254 
3255 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3256                                    intptr_t start_offset,
3257                                    intptr_t end_offset,
3258                                    PhaseGVN* phase) {
3259   if (start_offset == end_offset) {
3260     // nothing to do
3261     return mem;
3262   }
3263 
3264   assert((end_offset % BytesPerInt) == 0, "odd end offset");
3265   intptr_t done_offset = end_offset;
3266   if ((done_offset % BytesPerLong) != 0) {
3267     done_offset -= BytesPerInt;
3268   }
3269   if (done_offset > start_offset) {
3270     mem = clear_memory(ctl, mem, dest,
3271                        start_offset, phase->MakeConX(done_offset), phase);
3272   }
3273   if (done_offset < end_offset) { // emit the final 32-bit store
3274     Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3275     adr = phase->transform(adr);
3276     const TypePtr* atp = TypeRawPtr::BOTTOM;
3277     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3278     mem = phase->transform(mem);
3279     done_offset += BytesPerInt;
3280   }
3281   assert(done_offset == end_offset, "");
3282   return mem;
3283 }
3284 
3285 //=============================================================================
3286 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3287   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3288     _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3289 #ifdef ASSERT
3290   , _pair_idx(0)
3291 #endif
3292 {
3293   init_class_id(Class_MemBar);
3294   Node* top = C->top();
3295   init_req(TypeFunc::I_O,top);
3296   init_req(TypeFunc::FramePtr,top);
3297   init_req(TypeFunc::ReturnAdr,top);
3298   if (precedent != NULL)
3299     init_req(TypeFunc::Parms, precedent);
3300 }
3301 
3302 //------------------------------cmp--------------------------------------------
3303 uint MemBarNode::hash() const { return NO_HASH; }
3304 bool MemBarNode::cmp( const Node &n ) const {
3305   return (&n == this);          // Always fail except on self
3306 }
3307 
3308 //------------------------------make-------------------------------------------
3309 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3310   switch (opcode) {
3311   case Op_MemBarAcquire:     return new MemBarAcquireNode(C, atp, pn);
3312   case Op_LoadFence:         return new LoadFenceNode(C, atp, pn);
3313   case Op_MemBarRelease:     return new MemBarReleaseNode(C, atp, pn);
3314   case Op_StoreFence:        return new StoreFenceNode(C, atp, pn);
3315   case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3316   case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3317   case Op_MemBarVolatile:    return new MemBarVolatileNode(C, atp, pn);
3318   case Op_MemBarCPUOrder:    return new MemBarCPUOrderNode(C, atp, pn);
3319   case Op_OnSpinWait:        return new OnSpinWaitNode(C, atp, pn);
3320   case Op_Initialize:        return new InitializeNode(C, atp, pn);
3321   case Op_MemBarStoreStore:  return new MemBarStoreStoreNode(C, atp, pn);
3322   case Op_Blackhole:         return new BlackholeNode(C, atp, pn);
3323   default: ShouldNotReachHere(); return NULL;
3324   }
3325 }
3326 
3327 void MemBarNode::remove(PhaseIterGVN *igvn) {
3328   if (outcnt() != 2) {
3329     assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
3330     assert(outcnt() == 1, "Only control then");
3331   }
3332   if (trailing_store() || trailing_load_store()) {
3333     MemBarNode* leading = leading_membar();
3334     if (leading != NULL) {
3335       assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3336       leading->remove(igvn);
3337     }
3338   }
3339   if (proj_out_or_null(TypeFunc::Memory) != NULL) {
3340     igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3341   }
3342   if (proj_out_or_null(TypeFunc::Control) != NULL) {
3343     igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3344   }
3345 }
3346 
3347 //------------------------------Ideal------------------------------------------
3348 // Return a node which is more "ideal" than the current node.  Strip out
3349 // control copies
3350 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3351   if (remove_dead_region(phase, can_reshape)) return this;
3352   // Don't bother trying to transform a dead node
3353   if (in(0) && in(0)->is_top()) {
3354     return NULL;
3355   }
3356 
3357   bool progress = false;
3358   // Eliminate volatile MemBars for scalar replaced objects.
3359   if (can_reshape && req() == (Precedent+1)) {
3360     bool eliminate = false;
3361     int opc = Opcode();
3362     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3363       // Volatile field loads and stores.
3364       Node* my_mem = in(MemBarNode::Precedent);
3365       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3366       if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3367         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3368         // replace this Precedent (decodeN) with the Load instead.
3369         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
3370           Node* load_node = my_mem->in(1);
3371           set_req(MemBarNode::Precedent, load_node);
3372           phase->is_IterGVN()->_worklist.push(my_mem);
3373           my_mem = load_node;
3374         } else {
3375           assert(my_mem->unique_out() == this, "sanity");
3376           del_req(Precedent);
3377           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3378           my_mem = NULL;
3379         }
3380         progress = true;
3381       }
3382       if (my_mem != NULL && my_mem->is_Mem()) {
3383         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3384         // Check for scalar replaced object reference.
3385         if( t_oop != NULL && t_oop->is_known_instance_field() &&
3386             t_oop->offset() != Type::OffsetBot &&
3387             t_oop->offset() != Type::OffsetTop) {
3388           eliminate = true;
3389         }
3390       }
3391     } else if (opc == Op_MemBarRelease) {
3392       // Final field stores.
3393       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3394       if ((alloc != NULL) && alloc->is_Allocate() &&
3395           alloc->as_Allocate()->does_not_escape_thread()) {
3396         // The allocated object does not escape.
3397         eliminate = true;
3398       }
3399     }
3400     if (eliminate) {
3401       // Replace MemBar projections by its inputs.
3402       PhaseIterGVN* igvn = phase->is_IterGVN();
3403       remove(igvn);
3404       // Must return either the original node (now dead) or a new node
3405       // (Do not return a top here, since that would break the uniqueness of top.)
3406       return new ConINode(TypeInt::ZERO);
3407     }
3408   }
3409   return progress ? this : NULL;
3410 }
3411 
3412 //------------------------------Value------------------------------------------
3413 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3414   if( !in(0) ) return Type::TOP;
3415   if( phase->type(in(0)) == Type::TOP )
3416     return Type::TOP;
3417   return TypeTuple::MEMBAR;
3418 }
3419 
3420 //------------------------------match------------------------------------------
3421 // Construct projections for memory.
3422 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3423   switch (proj->_con) {
3424   case TypeFunc::Control:
3425   case TypeFunc::Memory:
3426     return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3427   }
3428   ShouldNotReachHere();
3429   return NULL;
3430 }
3431 
3432 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3433   trailing->_kind = TrailingStore;
3434   leading->_kind = LeadingStore;
3435 #ifdef ASSERT
3436   trailing->_pair_idx = leading->_idx;
3437   leading->_pair_idx = leading->_idx;
3438 #endif
3439 }
3440 
3441 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3442   trailing->_kind = TrailingLoadStore;
3443   leading->_kind = LeadingLoadStore;
3444 #ifdef ASSERT
3445   trailing->_pair_idx = leading->_idx;
3446   leading->_pair_idx = leading->_idx;
3447 #endif
3448 }
3449 
3450 MemBarNode* MemBarNode::trailing_membar() const {
3451   ResourceMark rm;
3452   Node* trailing = (Node*)this;
3453   VectorSet seen;
3454   Node_Stack multis(0);
3455   do {
3456     Node* c = trailing;
3457     uint i = 0;
3458     do {
3459       trailing = NULL;
3460       for (; i < c->outcnt(); i++) {
3461         Node* next = c->raw_out(i);
3462         if (next != c && next->is_CFG()) {
3463           if (c->is_MultiBranch()) {
3464             if (multis.node() == c) {
3465               multis.set_index(i+1);
3466             } else {
3467               multis.push(c, i+1);
3468             }
3469           }
3470           trailing = next;
3471           break;
3472         }
3473       }
3474       if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3475         break;
3476       }
3477       while (multis.size() > 0) {
3478         c = multis.node();
3479         i = multis.index();
3480         if (i < c->req()) {
3481           break;
3482         }
3483         multis.pop();
3484       }
3485     } while (multis.size() > 0);
3486   } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3487 
3488   MemBarNode* mb = trailing->as_MemBar();
3489   assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3490          (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3491   assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3492   return mb;
3493 }
3494 
3495 MemBarNode* MemBarNode::leading_membar() const {
3496   ResourceMark rm;
3497   VectorSet seen;
3498   Node_Stack regions(0);
3499   Node* leading = in(0);
3500   while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3501     while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3502       leading = NULL;
3503       while (regions.size() > 0 && leading == NULL) {
3504         Node* r = regions.node();
3505         uint i = regions.index();
3506         if (i < r->req()) {
3507           leading = r->in(i);
3508           regions.set_index(i+1);
3509         } else {
3510           regions.pop();
3511         }
3512       }
3513       if (leading == NULL) {
3514         assert(regions.size() == 0, "all paths should have been tried");
3515         return NULL;
3516       }
3517     }
3518     if (leading->is_Region()) {
3519       regions.push(leading, 2);
3520       leading = leading->in(1);
3521     } else {
3522       leading = leading->in(0);
3523     }
3524   }
3525 #ifdef ASSERT
3526   Unique_Node_List wq;
3527   wq.push((Node*)this);
3528   uint found = 0;
3529   for (uint i = 0; i < wq.size(); i++) {
3530     Node* n = wq.at(i);
3531     if (n->is_Region()) {
3532       for (uint j = 1; j < n->req(); j++) {
3533         Node* in = n->in(j);
3534         if (in != NULL && !in->is_top()) {
3535           wq.push(in);
3536         }
3537       }
3538     } else {
3539       if (n->is_MemBar() && n->as_MemBar()->leading()) {
3540         assert(n == leading, "consistency check failed");
3541         found++;
3542       } else {
3543         Node* in = n->in(0);
3544         if (in != NULL && !in->is_top()) {
3545           wq.push(in);
3546         }
3547       }
3548     }
3549   }
3550   assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3551 #endif
3552   if (leading == NULL) {
3553     return NULL;
3554   }
3555   MemBarNode* mb = leading->as_MemBar();
3556   assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3557          (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3558   assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3559   return mb;
3560 }
3561 
3562 #ifndef PRODUCT
3563 void BlackholeNode::format(PhaseRegAlloc* ra, outputStream* st) const {
3564   st->print("blackhole ");
3565   bool first = true;
3566   for (uint i = 0; i < req(); i++) {
3567     Node* n = in(i);
3568     if (n != NULL && OptoReg::is_valid(ra->get_reg_first(n))) {
3569       if (first) {
3570         first = false;
3571       } else {
3572         st->print(", ");
3573       }
3574       char buf[128];
3575       ra->dump_register(n, buf);
3576       st->print("%s", buf);
3577     }
3578   }
3579   st->cr();
3580 }
3581 #endif
3582 
3583 //===========================InitializeNode====================================
3584 // SUMMARY:
3585 // This node acts as a memory barrier on raw memory, after some raw stores.
3586 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3587 // The Initialize can 'capture' suitably constrained stores as raw inits.
3588 // It can coalesce related raw stores into larger units (called 'tiles').
3589 // It can avoid zeroing new storage for memory units which have raw inits.
3590 // At macro-expansion, it is marked 'complete', and does not optimize further.
3591 //
3592 // EXAMPLE:
3593 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3594 //   ctl = incoming control; mem* = incoming memory
3595 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
3596 // First allocate uninitialized memory and fill in the header:
3597 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
3598 //   ctl := alloc.Control; mem* := alloc.Memory*
3599 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
3600 // Then initialize to zero the non-header parts of the raw memory block:
3601 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3602 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3603 // After the initialize node executes, the object is ready for service:
3604 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3605 // Suppose its body is immediately initialized as {1,2}:
3606 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3607 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3608 //   mem.SLICE(#short[*]) := store2
3609 //
3610 // DETAILS:
3611 // An InitializeNode collects and isolates object initialization after
3612 // an AllocateNode and before the next possible safepoint.  As a
3613 // memory barrier (MemBarNode), it keeps critical stores from drifting
3614 // down past any safepoint or any publication of the allocation.
3615 // Before this barrier, a newly-allocated object may have uninitialized bits.
3616 // After this barrier, it may be treated as a real oop, and GC is allowed.
3617 //
3618 // The semantics of the InitializeNode include an implicit zeroing of
3619 // the new object from object header to the end of the object.
3620 // (The object header and end are determined by the AllocateNode.)
3621 //
3622 // Certain stores may be added as direct inputs to the InitializeNode.
3623 // These stores must update raw memory, and they must be to addresses
3624 // derived from the raw address produced by AllocateNode, and with
3625 // a constant offset.  They must be ordered by increasing offset.
3626 // The first one is at in(RawStores), the last at in(req()-1).
3627 // Unlike most memory operations, they are not linked in a chain,
3628 // but are displayed in parallel as users of the rawmem output of
3629 // the allocation.
3630 //
3631 // (See comments in InitializeNode::capture_store, which continue
3632 // the example given above.)
3633 //
3634 // When the associated Allocate is macro-expanded, the InitializeNode
3635 // may be rewritten to optimize collected stores.  A ClearArrayNode
3636 // may also be created at that point to represent any required zeroing.
3637 // The InitializeNode is then marked 'complete', prohibiting further
3638 // capturing of nearby memory operations.
3639 //
3640 // During macro-expansion, all captured initializations which store
3641 // constant values of 32 bits or smaller are coalesced (if advantageous)
3642 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
3643 // initialized in fewer memory operations.  Memory words which are
3644 // covered by neither tiles nor non-constant stores are pre-zeroed
3645 // by explicit stores of zero.  (The code shape happens to do all
3646 // zeroing first, then all other stores, with both sequences occurring
3647 // in order of ascending offsets.)
3648 //
3649 // Alternatively, code may be inserted between an AllocateNode and its
3650 // InitializeNode, to perform arbitrary initialization of the new object.
3651 // E.g., the object copying intrinsics insert complex data transfers here.
3652 // The initialization must then be marked as 'complete' disable the
3653 // built-in zeroing semantics and the collection of initializing stores.
3654 //
3655 // While an InitializeNode is incomplete, reads from the memory state
3656 // produced by it are optimizable if they match the control edge and
3657 // new oop address associated with the allocation/initialization.
3658 // They return a stored value (if the offset matches) or else zero.
3659 // A write to the memory state, if it matches control and address,
3660 // and if it is to a constant offset, may be 'captured' by the
3661 // InitializeNode.  It is cloned as a raw memory operation and rewired
3662 // inside the initialization, to the raw oop produced by the allocation.
3663 // Operations on addresses which are provably distinct (e.g., to
3664 // other AllocateNodes) are allowed to bypass the initialization.
3665 //
3666 // The effect of all this is to consolidate object initialization
3667 // (both arrays and non-arrays, both piecewise and bulk) into a
3668 // single location, where it can be optimized as a unit.
3669 //
3670 // Only stores with an offset less than TrackedInitializationLimit words
3671 // will be considered for capture by an InitializeNode.  This puts a
3672 // reasonable limit on the complexity of optimized initializations.
3673 
3674 //---------------------------InitializeNode------------------------------------
3675 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3676   : MemBarNode(C, adr_type, rawoop),
3677     _is_complete(Incomplete), _does_not_escape(false)
3678 {
3679   init_class_id(Class_Initialize);
3680 
3681   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3682   assert(in(RawAddress) == rawoop, "proper init");
3683   // Note:  allocation() can be NULL, for secondary initialization barriers
3684 }
3685 
3686 // Since this node is not matched, it will be processed by the
3687 // register allocator.  Declare that there are no constraints
3688 // on the allocation of the RawAddress edge.
3689 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3690   // This edge should be set to top, by the set_complete.  But be conservative.
3691   if (idx == InitializeNode::RawAddress)
3692     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3693   return RegMask::Empty;
3694 }
3695 
3696 Node* InitializeNode::memory(uint alias_idx) {
3697   Node* mem = in(Memory);
3698   if (mem->is_MergeMem()) {
3699     return mem->as_MergeMem()->memory_at(alias_idx);
3700   } else {
3701     // incoming raw memory is not split
3702     return mem;
3703   }
3704 }
3705 
3706 bool InitializeNode::is_non_zero() {
3707   if (is_complete())  return false;
3708   remove_extra_zeroes();
3709   return (req() > RawStores);
3710 }
3711 
3712 void InitializeNode::set_complete(PhaseGVN* phase) {
3713   assert(!is_complete(), "caller responsibility");
3714   _is_complete = Complete;
3715 
3716   // After this node is complete, it contains a bunch of
3717   // raw-memory initializations.  There is no need for
3718   // it to have anything to do with non-raw memory effects.
3719   // Therefore, tell all non-raw users to re-optimize themselves,
3720   // after skipping the memory effects of this initialization.
3721   PhaseIterGVN* igvn = phase->is_IterGVN();
3722   if (igvn)  igvn->add_users_to_worklist(this);
3723 }
3724 
3725 // convenience function
3726 // return false if the init contains any stores already
3727 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3728   InitializeNode* init = initialization();
3729   if (init == NULL || init->is_complete())  return false;
3730   init->remove_extra_zeroes();
3731   // for now, if this allocation has already collected any inits, bail:
3732   if (init->is_non_zero())  return false;
3733   init->set_complete(phase);
3734   return true;
3735 }
3736 
3737 void InitializeNode::remove_extra_zeroes() {
3738   if (req() == RawStores)  return;
3739   Node* zmem = zero_memory();
3740   uint fill = RawStores;
3741   for (uint i = fill; i < req(); i++) {
3742     Node* n = in(i);
3743     if (n->is_top() || n == zmem)  continue;  // skip
3744     if (fill < i)  set_req(fill, n);          // compact
3745     ++fill;
3746   }
3747   // delete any empty spaces created:
3748   while (fill < req()) {
3749     del_req(fill);
3750   }
3751 }
3752 
3753 // Helper for remembering which stores go with which offsets.
3754 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3755   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
3756   intptr_t offset = -1;
3757   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3758                                                phase, offset);
3759   if (base == NULL)     return -1;  // something is dead,
3760   if (offset < 0)       return -1;  //        dead, dead
3761   return offset;
3762 }
3763 
3764 // Helper for proving that an initialization expression is
3765 // "simple enough" to be folded into an object initialization.
3766 // Attempts to prove that a store's initial value 'n' can be captured
3767 // within the initialization without creating a vicious cycle, such as:
3768 //     { Foo p = new Foo(); p.next = p; }
3769 // True for constants and parameters and small combinations thereof.
3770 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3771   ResourceMark rm;
3772   Unique_Node_List worklist;
3773   worklist.push(value);
3774 
3775   uint complexity_limit = 20;
3776   for (uint j = 0; j < worklist.size(); j++) {
3777     if (j >= complexity_limit) {
3778       return false;  // Bail out if processed too many nodes
3779     }
3780 
3781     Node* n = worklist.at(j);
3782     if (n == NULL)      continue;   // (can this really happen?)
3783     if (n->is_Proj())   n = n->in(0);
3784     if (n == this)      return false;  // found a cycle
3785     if (n->is_Con())    continue;
3786     if (n->is_Start())  continue;   // params, etc., are OK
3787     if (n->is_Root())   continue;   // even better
3788 
3789     // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3790     if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3791       continue;
3792     }
3793 
3794     Node* ctl = n->in(0);
3795     if (ctl != NULL && !ctl->is_top()) {
3796       if (ctl->is_Proj())  ctl = ctl->in(0);
3797       if (ctl == this)  return false;
3798 
3799       // If we already know that the enclosing memory op is pinned right after
3800       // the init, then any control flow that the store has picked up
3801       // must have preceded the init, or else be equal to the init.
3802       // Even after loop optimizations (which might change control edges)
3803       // a store is never pinned *before* the availability of its inputs.
3804       if (!MemNode::all_controls_dominate(n, this))
3805         return false;                  // failed to prove a good control
3806     }
3807 
3808     // Check data edges for possible dependencies on 'this'.
3809     for (uint i = 1; i < n->req(); i++) {
3810       Node* m = n->in(i);
3811       if (m == NULL || m == n || m->is_top())  continue;
3812 
3813       // Only process data inputs once
3814       worklist.push(m);
3815     }
3816   }
3817 
3818   return true;
3819 }
3820 
3821 // Here are all the checks a Store must pass before it can be moved into
3822 // an initialization.  Returns zero if a check fails.
3823 // On success, returns the (constant) offset to which the store applies,
3824 // within the initialized memory.
3825 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3826   const int FAIL = 0;
3827   if (st->req() != MemNode::ValueIn + 1)
3828     return FAIL;                // an inscrutable StoreNode (card mark?)
3829   Node* ctl = st->in(MemNode::Control);
3830   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3831     return FAIL;                // must be unconditional after the initialization
3832   Node* mem = st->in(MemNode::Memory);
3833   if (!(mem->is_Proj() && mem->in(0) == this))
3834     return FAIL;                // must not be preceded by other stores
3835   Node* adr = st->in(MemNode::Address);
3836   intptr_t offset;
3837   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3838   if (alloc == NULL)
3839     return FAIL;                // inscrutable address
3840   if (alloc != allocation())
3841     return FAIL;                // wrong allocation!  (store needs to float up)
3842   int size_in_bytes = st->memory_size();
3843   if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3844     return FAIL;                // mismatched access
3845   }
3846   Node* val = st->in(MemNode::ValueIn);
3847 
3848   if (!detect_init_independence(val, phase))
3849     return FAIL;                // stored value must be 'simple enough'
3850 
3851   // The Store can be captured only if nothing after the allocation
3852   // and before the Store is using the memory location that the store
3853   // overwrites.
3854   bool failed = false;
3855   // If is_complete_with_arraycopy() is true the shape of the graph is
3856   // well defined and is safe so no need for extra checks.
3857   if (!is_complete_with_arraycopy()) {
3858     // We are going to look at each use of the memory state following
3859     // the allocation to make sure nothing reads the memory that the
3860     // Store writes.
3861     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3862     int alias_idx = phase->C->get_alias_index(t_adr);
3863     ResourceMark rm;
3864     Unique_Node_List mems;
3865     mems.push(mem);
3866     Node* unique_merge = NULL;
3867     for (uint next = 0; next < mems.size(); ++next) {
3868       Node *m  = mems.at(next);
3869       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3870         Node *n = m->fast_out(j);
3871         if (n->outcnt() == 0) {
3872           continue;
3873         }
3874         if (n == st) {
3875           continue;
3876         } else if (n->in(0) != NULL && n->in(0) != ctl) {
3877           // If the control of this use is different from the control
3878           // of the Store which is right after the InitializeNode then
3879           // this node cannot be between the InitializeNode and the
3880           // Store.
3881           continue;
3882         } else if (n->is_MergeMem()) {
3883           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3884             // We can hit a MergeMemNode (that will likely go away
3885             // later) that is a direct use of the memory state
3886             // following the InitializeNode on the same slice as the
3887             // store node that we'd like to capture. We need to check
3888             // the uses of the MergeMemNode.
3889             mems.push(n);
3890           }
3891         } else if (n->is_Mem()) {
3892           Node* other_adr = n->in(MemNode::Address);
3893           if (other_adr == adr) {
3894             failed = true;
3895             break;
3896           } else {
3897             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3898             if (other_t_adr != NULL) {
3899               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3900               if (other_alias_idx == alias_idx) {
3901                 // A load from the same memory slice as the store right
3902                 // after the InitializeNode. We check the control of the
3903                 // object/array that is loaded from. If it's the same as
3904                 // the store control then we cannot capture the store.
3905                 assert(!n->is_Store(), "2 stores to same slice on same control?");
3906                 Node* base = other_adr;
3907                 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3908                 base = base->in(AddPNode::Base);
3909                 if (base != NULL) {
3910                   base = base->uncast();
3911                   if (base->is_Proj() && base->in(0) == alloc) {
3912                     failed = true;
3913                     break;
3914                   }
3915                 }
3916               }
3917             }
3918           }
3919         } else {
3920           failed = true;
3921           break;
3922         }
3923       }
3924     }
3925   }
3926   if (failed) {
3927     if (!can_reshape) {
3928       // We decided we couldn't capture the store during parsing. We
3929       // should try again during the next IGVN once the graph is
3930       // cleaner.
3931       phase->C->record_for_igvn(st);
3932     }
3933     return FAIL;
3934   }
3935 
3936   return offset;                // success
3937 }
3938 
3939 // Find the captured store in(i) which corresponds to the range
3940 // [start..start+size) in the initialized object.
3941 // If there is one, return its index i.  If there isn't, return the
3942 // negative of the index where it should be inserted.
3943 // Return 0 if the queried range overlaps an initialization boundary
3944 // or if dead code is encountered.
3945 // If size_in_bytes is zero, do not bother with overlap checks.
3946 int InitializeNode::captured_store_insertion_point(intptr_t start,
3947                                                    int size_in_bytes,
3948                                                    PhaseTransform* phase) {
3949   const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
3950 
3951   if (is_complete())
3952     return FAIL;                // arraycopy got here first; punt
3953 
3954   assert(allocation() != NULL, "must be present");
3955 
3956   // no negatives, no header fields:
3957   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
3958 
3959   // after a certain size, we bail out on tracking all the stores:
3960   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3961   if (start >= ti_limit)  return FAIL;
3962 
3963   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3964     if (i >= limit)  return -(int)i; // not found; here is where to put it
3965 
3966     Node*    st     = in(i);
3967     intptr_t st_off = get_store_offset(st, phase);
3968     if (st_off < 0) {
3969       if (st != zero_memory()) {
3970         return FAIL;            // bail out if there is dead garbage
3971       }
3972     } else if (st_off > start) {
3973       // ...we are done, since stores are ordered
3974       if (st_off < start + size_in_bytes) {
3975         return FAIL;            // the next store overlaps
3976       }
3977       return -(int)i;           // not found; here is where to put it
3978     } else if (st_off < start) {
3979       assert(st->as_Store()->memory_size() <= MAX_STORE, "");
3980       if (size_in_bytes != 0 &&
3981           start < st_off + MAX_STORE &&
3982           start < st_off + st->as_Store()->memory_size()) {
3983         return FAIL;            // the previous store overlaps
3984       }
3985     } else {
3986       if (size_in_bytes != 0 &&
3987           st->as_Store()->memory_size() != size_in_bytes) {
3988         return FAIL;            // mismatched store size
3989       }
3990       return i;
3991     }
3992 
3993     ++i;
3994   }
3995 }
3996 
3997 // Look for a captured store which initializes at the offset 'start'
3998 // with the given size.  If there is no such store, and no other
3999 // initialization interferes, then return zero_memory (the memory
4000 // projection of the AllocateNode).
4001 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4002                                           PhaseTransform* phase) {
4003   assert(stores_are_sane(phase), "");
4004   int i = captured_store_insertion_point(start, size_in_bytes, phase);
4005   if (i == 0) {
4006     return NULL;                // something is dead
4007   } else if (i < 0) {
4008     return zero_memory();       // just primordial zero bits here
4009   } else {
4010     Node* st = in(i);           // here is the store at this position
4011     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4012     return st;
4013   }
4014 }
4015 
4016 // Create, as a raw pointer, an address within my new object at 'offset'.
4017 Node* InitializeNode::make_raw_address(intptr_t offset,
4018                                        PhaseTransform* phase) {
4019   Node* addr = in(RawAddress);
4020   if (offset != 0) {
4021     Compile* C = phase->C;
4022     addr = phase->transform( new AddPNode(C->top(), addr,
4023                                                  phase->MakeConX(offset)) );
4024   }
4025   return addr;
4026 }
4027 
4028 // Clone the given store, converting it into a raw store
4029 // initializing a field or element of my new object.
4030 // Caller is responsible for retiring the original store,
4031 // with subsume_node or the like.
4032 //
4033 // From the example above InitializeNode::InitializeNode,
4034 // here are the old stores to be captured:
4035 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4036 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
4037 //
4038 // Here is the changed code; note the extra edges on init:
4039 //   alloc = (Allocate ...)
4040 //   rawoop = alloc.RawAddress
4041 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4042 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4043 //   init = (Initialize alloc.Control alloc.Memory rawoop
4044 //                      rawstore1 rawstore2)
4045 //
4046 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4047                                     PhaseGVN* phase, bool can_reshape) {
4048   assert(stores_are_sane(phase), "");
4049 
4050   if (start < 0)  return NULL;
4051   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4052 
4053   Compile* C = phase->C;
4054   int size_in_bytes = st->memory_size();
4055   int i = captured_store_insertion_point(start, size_in_bytes, phase);
4056   if (i == 0)  return NULL;     // bail out
4057   Node* prev_mem = NULL;        // raw memory for the captured store
4058   if (i > 0) {
4059     prev_mem = in(i);           // there is a pre-existing store under this one
4060     set_req(i, C->top());       // temporarily disconnect it
4061     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4062   } else {
4063     i = -i;                     // no pre-existing store
4064     prev_mem = zero_memory();   // a slice of the newly allocated object
4065     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4066       set_req(--i, C->top());   // reuse this edge; it has been folded away
4067     else
4068       ins_req(i, C->top());     // build a new edge
4069   }
4070   Node* new_st = st->clone();
4071   new_st->set_req(MemNode::Control, in(Control));
4072   new_st->set_req(MemNode::Memory,  prev_mem);
4073   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4074   new_st = phase->transform(new_st);
4075 
4076   // At this point, new_st might have swallowed a pre-existing store
4077   // at the same offset, or perhaps new_st might have disappeared,
4078   // if it redundantly stored the same value (or zero to fresh memory).
4079 
4080   // In any case, wire it in:
4081   PhaseIterGVN* igvn = phase->is_IterGVN();
4082   if (igvn) {
4083     igvn->rehash_node_delayed(this);
4084   }
4085   set_req(i, new_st);
4086 
4087   // The caller may now kill the old guy.
4088   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4089   assert(check_st == new_st || check_st == NULL, "must be findable");
4090   assert(!is_complete(), "");
4091   return new_st;
4092 }
4093 
4094 static bool store_constant(jlong* tiles, int num_tiles,
4095                            intptr_t st_off, int st_size,
4096                            jlong con) {
4097   if ((st_off & (st_size-1)) != 0)
4098     return false;               // strange store offset (assume size==2**N)
4099   address addr = (address)tiles + st_off;
4100   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4101   switch (st_size) {
4102   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
4103   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
4104   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
4105   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
4106   default: return false;        // strange store size (detect size!=2**N here)
4107   }
4108   return true;                  // return success to caller
4109 }
4110 
4111 // Coalesce subword constants into int constants and possibly
4112 // into long constants.  The goal, if the CPU permits,
4113 // is to initialize the object with a small number of 64-bit tiles.
4114 // Also, convert floating-point constants to bit patterns.
4115 // Non-constants are not relevant to this pass.
4116 //
4117 // In terms of the running example on InitializeNode::InitializeNode
4118 // and InitializeNode::capture_store, here is the transformation
4119 // of rawstore1 and rawstore2 into rawstore12:
4120 //   alloc = (Allocate ...)
4121 //   rawoop = alloc.RawAddress
4122 //   tile12 = 0x00010002
4123 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4124 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4125 //
4126 void
4127 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4128                                         Node* size_in_bytes,
4129                                         PhaseGVN* phase) {
4130   Compile* C = phase->C;
4131 
4132   assert(stores_are_sane(phase), "");
4133   // Note:  After this pass, they are not completely sane,
4134   // since there may be some overlaps.
4135 
4136   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4137 
4138   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4139   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4140   size_limit = MIN2(size_limit, ti_limit);
4141   size_limit = align_up(size_limit, BytesPerLong);
4142   int num_tiles = size_limit / BytesPerLong;
4143 
4144   // allocate space for the tile map:
4145   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4146   jlong  tiles_buf[small_len];
4147   Node*  nodes_buf[small_len];
4148   jlong  inits_buf[small_len];
4149   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4150                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4151   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4152                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4153   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4154                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4155   // tiles: exact bitwise model of all primitive constants
4156   // nodes: last constant-storing node subsumed into the tiles model
4157   // inits: which bytes (in each tile) are touched by any initializations
4158 
4159   //// Pass A: Fill in the tile model with any relevant stores.
4160 
4161   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4162   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4163   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4164   Node* zmem = zero_memory(); // initially zero memory state
4165   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4166     Node* st = in(i);
4167     intptr_t st_off = get_store_offset(st, phase);
4168 
4169     // Figure out the store's offset and constant value:
4170     if (st_off < header_size)             continue; //skip (ignore header)
4171     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
4172     int st_size = st->as_Store()->memory_size();
4173     if (st_off + st_size > size_limit)    break;
4174 
4175     // Record which bytes are touched, whether by constant or not.
4176     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4177       continue;                 // skip (strange store size)
4178 
4179     const Type* val = phase->type(st->in(MemNode::ValueIn));
4180     if (!val->singleton())                continue; //skip (non-con store)
4181     BasicType type = val->basic_type();
4182 
4183     jlong con = 0;
4184     switch (type) {
4185     case T_INT:    con = val->is_int()->get_con();  break;
4186     case T_LONG:   con = val->is_long()->get_con(); break;
4187     case T_FLOAT:  con = jint_cast(val->getf());    break;
4188     case T_DOUBLE: con = jlong_cast(val->getd());   break;
4189     default:                              continue; //skip (odd store type)
4190     }
4191 
4192     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4193         st->Opcode() == Op_StoreL) {
4194       continue;                 // This StoreL is already optimal.
4195     }
4196 
4197     // Store down the constant.
4198     store_constant(tiles, num_tiles, st_off, st_size, con);
4199 
4200     intptr_t j = st_off >> LogBytesPerLong;
4201 
4202     if (type == T_INT && st_size == BytesPerInt
4203         && (st_off & BytesPerInt) == BytesPerInt) {
4204       jlong lcon = tiles[j];
4205       if (!Matcher::isSimpleConstant64(lcon) &&
4206           st->Opcode() == Op_StoreI) {
4207         // This StoreI is already optimal by itself.
4208         jint* intcon = (jint*) &tiles[j];
4209         intcon[1] = 0;  // undo the store_constant()
4210 
4211         // If the previous store is also optimal by itself, back up and
4212         // undo the action of the previous loop iteration... if we can.
4213         // But if we can't, just let the previous half take care of itself.
4214         st = nodes[j];
4215         st_off -= BytesPerInt;
4216         con = intcon[0];
4217         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
4218           assert(st_off >= header_size, "still ignoring header");
4219           assert(get_store_offset(st, phase) == st_off, "must be");
4220           assert(in(i-1) == zmem, "must be");
4221           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4222           assert(con == tcon->is_int()->get_con(), "must be");
4223           // Undo the effects of the previous loop trip, which swallowed st:
4224           intcon[0] = 0;        // undo store_constant()
4225           set_req(i-1, st);     // undo set_req(i, zmem)
4226           nodes[j] = NULL;      // undo nodes[j] = st
4227           --old_subword;        // undo ++old_subword
4228         }
4229         continue;               // This StoreI is already optimal.
4230       }
4231     }
4232 
4233     // This store is not needed.
4234     set_req(i, zmem);
4235     nodes[j] = st;              // record for the moment
4236     if (st_size < BytesPerLong) // something has changed
4237           ++old_subword;        // includes int/float, but who's counting...
4238     else  ++old_long;
4239   }
4240 
4241   if ((old_subword + old_long) == 0)
4242     return;                     // nothing more to do
4243 
4244   //// Pass B: Convert any non-zero tiles into optimal constant stores.
4245   // Be sure to insert them before overlapping non-constant stores.
4246   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
4247   for (int j = 0; j < num_tiles; j++) {
4248     jlong con  = tiles[j];
4249     jlong init = inits[j];
4250     if (con == 0)  continue;
4251     jint con0,  con1;           // split the constant, address-wise
4252     jint init0, init1;          // split the init map, address-wise
4253     { union { jlong con; jint intcon[2]; } u;
4254       u.con = con;
4255       con0  = u.intcon[0];
4256       con1  = u.intcon[1];
4257       u.con = init;
4258       init0 = u.intcon[0];
4259       init1 = u.intcon[1];
4260     }
4261 
4262     Node* old = nodes[j];
4263     assert(old != NULL, "need the prior store");
4264     intptr_t offset = (j * BytesPerLong);
4265 
4266     bool split = !Matcher::isSimpleConstant64(con);
4267 
4268     if (offset < header_size) {
4269       assert(offset + BytesPerInt >= header_size, "second int counts");
4270       assert(*(jint*)&tiles[j] == 0, "junk in header");
4271       split = true;             // only the second word counts
4272       // Example:  int a[] = { 42 ... }
4273     } else if (con0 == 0 && init0 == -1) {
4274       split = true;             // first word is covered by full inits
4275       // Example:  int a[] = { ... foo(), 42 ... }
4276     } else if (con1 == 0 && init1 == -1) {
4277       split = true;             // second word is covered by full inits
4278       // Example:  int a[] = { ... 42, foo() ... }
4279     }
4280 
4281     // Here's a case where init0 is neither 0 nor -1:
4282     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
4283     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4284     // In this case the tile is not split; it is (jlong)42.
4285     // The big tile is stored down, and then the foo() value is inserted.
4286     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4287 
4288     Node* ctl = old->in(MemNode::Control);
4289     Node* adr = make_raw_address(offset, phase);
4290     const TypePtr* atp = TypeRawPtr::BOTTOM;
4291 
4292     // One or two coalesced stores to plop down.
4293     Node*    st[2];
4294     intptr_t off[2];
4295     int  nst = 0;
4296     if (!split) {
4297       ++new_long;
4298       off[nst] = offset;
4299       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4300                                   phase->longcon(con), T_LONG, MemNode::unordered);
4301     } else {
4302       // Omit either if it is a zero.
4303       if (con0 != 0) {
4304         ++new_int;
4305         off[nst]  = offset;
4306         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4307                                     phase->intcon(con0), T_INT, MemNode::unordered);
4308       }
4309       if (con1 != 0) {
4310         ++new_int;
4311         offset += BytesPerInt;
4312         adr = make_raw_address(offset, phase);
4313         off[nst]  = offset;
4314         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4315                                     phase->intcon(con1), T_INT, MemNode::unordered);
4316       }
4317     }
4318 
4319     // Insert second store first, then the first before the second.
4320     // Insert each one just before any overlapping non-constant stores.
4321     while (nst > 0) {
4322       Node* st1 = st[--nst];
4323       C->copy_node_notes_to(st1, old);
4324       st1 = phase->transform(st1);
4325       offset = off[nst];
4326       assert(offset >= header_size, "do not smash header");
4327       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4328       guarantee(ins_idx != 0, "must re-insert constant store");
4329       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
4330       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4331         set_req(--ins_idx, st1);
4332       else
4333         ins_req(ins_idx, st1);
4334     }
4335   }
4336 
4337   if (PrintCompilation && WizardMode)
4338     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4339                   old_subword, old_long, new_int, new_long);
4340   if (C->log() != NULL)
4341     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4342                    old_subword, old_long, new_int, new_long);
4343 
4344   // Clean up any remaining occurrences of zmem:
4345   remove_extra_zeroes();
4346 }
4347 
4348 // Explore forward from in(start) to find the first fully initialized
4349 // word, and return its offset.  Skip groups of subword stores which
4350 // together initialize full words.  If in(start) is itself part of a
4351 // fully initialized word, return the offset of in(start).  If there
4352 // are no following full-word stores, or if something is fishy, return
4353 // a negative value.
4354 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4355   int       int_map = 0;
4356   intptr_t  int_map_off = 0;
4357   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
4358 
4359   for (uint i = start, limit = req(); i < limit; i++) {
4360     Node* st = in(i);
4361 
4362     intptr_t st_off = get_store_offset(st, phase);
4363     if (st_off < 0)  break;  // return conservative answer
4364 
4365     int st_size = st->as_Store()->memory_size();
4366     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4367       return st_off;            // we found a complete word init
4368     }
4369 
4370     // update the map:
4371 
4372     intptr_t this_int_off = align_down(st_off, BytesPerInt);
4373     if (this_int_off != int_map_off) {
4374       // reset the map:
4375       int_map = 0;
4376       int_map_off = this_int_off;
4377     }
4378 
4379     int subword_off = st_off - this_int_off;
4380     int_map |= right_n_bits(st_size) << subword_off;
4381     if ((int_map & FULL_MAP) == FULL_MAP) {
4382       return this_int_off;      // we found a complete word init
4383     }
4384 
4385     // Did this store hit or cross the word boundary?
4386     intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4387     if (next_int_off == this_int_off + BytesPerInt) {
4388       // We passed the current int, without fully initializing it.
4389       int_map_off = next_int_off;
4390       int_map >>= BytesPerInt;
4391     } else if (next_int_off > this_int_off + BytesPerInt) {
4392       // We passed the current and next int.
4393       return this_int_off + BytesPerInt;
4394     }
4395   }
4396 
4397   return -1;
4398 }
4399 
4400 
4401 // Called when the associated AllocateNode is expanded into CFG.
4402 // At this point, we may perform additional optimizations.
4403 // Linearize the stores by ascending offset, to make memory
4404 // activity as coherent as possible.
4405 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4406                                       intptr_t header_size,
4407                                       Node* size_in_bytes,
4408                                       PhaseIterGVN* phase) {
4409   assert(!is_complete(), "not already complete");
4410   assert(stores_are_sane(phase), "");
4411   assert(allocation() != NULL, "must be present");
4412 
4413   remove_extra_zeroes();
4414 
4415   if (ReduceFieldZeroing || ReduceBulkZeroing)
4416     // reduce instruction count for common initialization patterns
4417     coalesce_subword_stores(header_size, size_in_bytes, phase);
4418 
4419   Node* zmem = zero_memory();   // initially zero memory state
4420   Node* inits = zmem;           // accumulating a linearized chain of inits
4421   #ifdef ASSERT
4422   intptr_t first_offset = allocation()->minimum_header_size();
4423   intptr_t last_init_off = first_offset;  // previous init offset
4424   intptr_t last_init_end = first_offset;  // previous init offset+size
4425   intptr_t last_tile_end = first_offset;  // previous tile offset+size
4426   #endif
4427   intptr_t zeroes_done = header_size;
4428 
4429   bool do_zeroing = true;       // we might give up if inits are very sparse
4430   int  big_init_gaps = 0;       // how many large gaps have we seen?
4431 
4432   if (UseTLAB && ZeroTLAB)  do_zeroing = false;
4433   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
4434 
4435   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4436     Node* st = in(i);
4437     intptr_t st_off = get_store_offset(st, phase);
4438     if (st_off < 0)
4439       break;                    // unknown junk in the inits
4440     if (st->in(MemNode::Memory) != zmem)
4441       break;                    // complicated store chains somehow in list
4442 
4443     int st_size = st->as_Store()->memory_size();
4444     intptr_t next_init_off = st_off + st_size;
4445 
4446     if (do_zeroing && zeroes_done < next_init_off) {
4447       // See if this store needs a zero before it or under it.
4448       intptr_t zeroes_needed = st_off;
4449 
4450       if (st_size < BytesPerInt) {
4451         // Look for subword stores which only partially initialize words.
4452         // If we find some, we must lay down some word-level zeroes first,
4453         // underneath the subword stores.
4454         //
4455         // Examples:
4456         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
4457         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
4458         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
4459         //
4460         // Note:  coalesce_subword_stores may have already done this,
4461         // if it was prompted by constant non-zero subword initializers.
4462         // But this case can still arise with non-constant stores.
4463 
4464         intptr_t next_full_store = find_next_fullword_store(i, phase);
4465 
4466         // In the examples above:
4467         //   in(i)          p   q   r   s     x   y     z
4468         //   st_off        12  13  14  15    12  13    14
4469         //   st_size        1   1   1   1     1   1     1
4470         //   next_full_s.  12  16  16  16    16  16    16
4471         //   z's_done      12  16  16  16    12  16    12
4472         //   z's_needed    12  16  16  16    16  16    16
4473         //   zsize          0   0   0   0     4   0     4
4474         if (next_full_store < 0) {
4475           // Conservative tack:  Zero to end of current word.
4476           zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4477         } else {
4478           // Zero to beginning of next fully initialized word.
4479           // Or, don't zero at all, if we are already in that word.
4480           assert(next_full_store >= zeroes_needed, "must go forward");
4481           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4482           zeroes_needed = next_full_store;
4483         }
4484       }
4485 
4486       if (zeroes_needed > zeroes_done) {
4487         intptr_t zsize = zeroes_needed - zeroes_done;
4488         // Do some incremental zeroing on rawmem, in parallel with inits.
4489         zeroes_done = align_down(zeroes_done, BytesPerInt);
4490         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4491                                               zeroes_done, zeroes_needed,
4492                                               phase);
4493         zeroes_done = zeroes_needed;
4494         if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4495           do_zeroing = false;   // leave the hole, next time
4496       }
4497     }
4498 
4499     // Collect the store and move on:
4500     phase->replace_input_of(st, MemNode::Memory, inits);
4501     inits = st;                 // put it on the linearized chain
4502     set_req(i, zmem);           // unhook from previous position
4503 
4504     if (zeroes_done == st_off)
4505       zeroes_done = next_init_off;
4506 
4507     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4508 
4509     #ifdef ASSERT
4510     // Various order invariants.  Weaker than stores_are_sane because
4511     // a large constant tile can be filled in by smaller non-constant stores.
4512     assert(st_off >= last_init_off, "inits do not reverse");
4513     last_init_off = st_off;
4514     const Type* val = NULL;
4515     if (st_size >= BytesPerInt &&
4516         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4517         (int)val->basic_type() < (int)T_OBJECT) {
4518       assert(st_off >= last_tile_end, "tiles do not overlap");
4519       assert(st_off >= last_init_end, "tiles do not overwrite inits");
4520       last_tile_end = MAX2(last_tile_end, next_init_off);
4521     } else {
4522       intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4523       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4524       assert(st_off      >= last_init_end, "inits do not overlap");
4525       last_init_end = next_init_off;  // it's a non-tile
4526     }
4527     #endif //ASSERT
4528   }
4529 
4530   remove_extra_zeroes();        // clear out all the zmems left over
4531   add_req(inits);
4532 
4533   if (!(UseTLAB && ZeroTLAB)) {
4534     // If anything remains to be zeroed, zero it all now.
4535     zeroes_done = align_down(zeroes_done, BytesPerInt);
4536     // if it is the last unused 4 bytes of an instance, forget about it
4537     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4538     if (zeroes_done + BytesPerLong >= size_limit) {
4539       AllocateNode* alloc = allocation();
4540       assert(alloc != NULL, "must be present");
4541       if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4542         Node* klass_node = alloc->in(AllocateNode::KlassNode);
4543         ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4544         if (zeroes_done == k->layout_helper())
4545           zeroes_done = size_limit;
4546       }
4547     }
4548     if (zeroes_done < size_limit) {
4549       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4550                                             zeroes_done, size_in_bytes, phase);
4551     }
4552   }
4553 
4554   set_complete(phase);
4555   return rawmem;
4556 }
4557 
4558 
4559 #ifdef ASSERT
4560 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4561   if (is_complete())
4562     return true;                // stores could be anything at this point
4563   assert(allocation() != NULL, "must be present");
4564   intptr_t last_off = allocation()->minimum_header_size();
4565   for (uint i = InitializeNode::RawStores; i < req(); i++) {
4566     Node* st = in(i);
4567     intptr_t st_off = get_store_offset(st, phase);
4568     if (st_off < 0)  continue;  // ignore dead garbage
4569     if (last_off > st_off) {
4570       tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4571       this->dump(2);
4572       assert(false, "ascending store offsets");
4573       return false;
4574     }
4575     last_off = st_off + st->as_Store()->memory_size();
4576   }
4577   return true;
4578 }
4579 #endif //ASSERT
4580 
4581 
4582 
4583 
4584 //============================MergeMemNode=====================================
4585 //
4586 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
4587 // contributing store or call operations.  Each contributor provides the memory
4588 // state for a particular "alias type" (see Compile::alias_type).  For example,
4589 // if a MergeMem has an input X for alias category #6, then any memory reference
4590 // to alias category #6 may use X as its memory state input, as an exact equivalent
4591 // to using the MergeMem as a whole.
4592 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4593 //
4594 // (Here, the <N> notation gives the index of the relevant adr_type.)
4595 //
4596 // In one special case (and more cases in the future), alias categories overlap.
4597 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4598 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
4599 // it is exactly equivalent to that state W:
4600 //   MergeMem(<Bot>: W) <==> W
4601 //
4602 // Usually, the merge has more than one input.  In that case, where inputs
4603 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4604 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4605 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4606 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4607 //
4608 // A merge can take a "wide" memory state as one of its narrow inputs.
4609 // This simply means that the merge observes out only the relevant parts of
4610 // the wide input.  That is, wide memory states arriving at narrow merge inputs
4611 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
4612 //
4613 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4614 // and that memory slices "leak through":
4615 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4616 //
4617 // But, in such a cascade, repeated memory slices can "block the leak":
4618 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4619 //
4620 // In the last example, Y is not part of the combined memory state of the
4621 // outermost MergeMem.  The system must, of course, prevent unschedulable
4622 // memory states from arising, so you can be sure that the state Y is somehow
4623 // a precursor to state Y'.
4624 //
4625 //
4626 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4627 // of each MergeMemNode array are exactly the numerical alias indexes, including
4628 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
4629 // Compile::alias_type (and kin) produce and manage these indexes.
4630 //
4631 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4632 // (Note that this provides quick access to the top node inside MergeMem methods,
4633 // without the need to reach out via TLS to Compile::current.)
4634 //
4635 // As a consequence of what was just described, a MergeMem that represents a full
4636 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4637 // containing all alias categories.
4638 //
4639 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4640 //
4641 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4642 // a memory state for the alias type <N>, or else the top node, meaning that
4643 // there is no particular input for that alias type.  Note that the length of
4644 // a MergeMem is variable, and may be extended at any time to accommodate new
4645 // memory states at larger alias indexes.  When merges grow, they are of course
4646 // filled with "top" in the unused in() positions.
4647 //
4648 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4649 // (Top was chosen because it works smoothly with passes like GCM.)
4650 //
4651 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
4652 // the type of random VM bits like TLS references.)  Since it is always the
4653 // first non-Bot memory slice, some low-level loops use it to initialize an
4654 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
4655 //
4656 //
4657 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
4658 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
4659 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4660 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
4661 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
4662 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4663 //
4664 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4665 // really that different from the other memory inputs.  An abbreviation called
4666 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4667 //
4668 //
4669 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
4670 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
4671 // that "emerges though" the base memory will be marked as excluding the alias types
4672 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4673 //
4674 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4675 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4676 //
4677 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4678 // (It is currently unimplemented.)  As you can see, the resulting merge is
4679 // actually a disjoint union of memory states, rather than an overlay.
4680 //
4681 
4682 //------------------------------MergeMemNode-----------------------------------
4683 Node* MergeMemNode::make_empty_memory() {
4684   Node* empty_memory = (Node*) Compile::current()->top();
4685   assert(empty_memory->is_top(), "correct sentinel identity");
4686   return empty_memory;
4687 }
4688 
4689 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4690   init_class_id(Class_MergeMem);
4691   // all inputs are nullified in Node::Node(int)
4692   // set_input(0, NULL);  // no control input
4693 
4694   // Initialize the edges uniformly to top, for starters.
4695   Node* empty_mem = make_empty_memory();
4696   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4697     init_req(i,empty_mem);
4698   }
4699   assert(empty_memory() == empty_mem, "");
4700 
4701   if( new_base != NULL && new_base->is_MergeMem() ) {
4702     MergeMemNode* mdef = new_base->as_MergeMem();
4703     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4704     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4705       mms.set_memory(mms.memory2());
4706     }
4707     assert(base_memory() == mdef->base_memory(), "");
4708   } else {
4709     set_base_memory(new_base);
4710   }
4711 }
4712 
4713 // Make a new, untransformed MergeMem with the same base as 'mem'.
4714 // If mem is itself a MergeMem, populate the result with the same edges.
4715 MergeMemNode* MergeMemNode::make(Node* mem) {
4716   return new MergeMemNode(mem);
4717 }
4718 
4719 //------------------------------cmp--------------------------------------------
4720 uint MergeMemNode::hash() const { return NO_HASH; }
4721 bool MergeMemNode::cmp( const Node &n ) const {
4722   return (&n == this);          // Always fail except on self
4723 }
4724 
4725 //------------------------------Identity---------------------------------------
4726 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4727   // Identity if this merge point does not record any interesting memory
4728   // disambiguations.
4729   Node* base_mem = base_memory();
4730   Node* empty_mem = empty_memory();
4731   if (base_mem != empty_mem) {  // Memory path is not dead?
4732     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4733       Node* mem = in(i);
4734       if (mem != empty_mem && mem != base_mem) {
4735         return this;            // Many memory splits; no change
4736       }
4737     }
4738   }
4739   return base_mem;              // No memory splits; ID on the one true input
4740 }
4741 
4742 //------------------------------Ideal------------------------------------------
4743 // This method is invoked recursively on chains of MergeMem nodes
4744 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4745   // Remove chain'd MergeMems
4746   //
4747   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4748   // relative to the "in(Bot)".  Since we are patching both at the same time,
4749   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4750   // but rewrite each "in(i)" relative to the new "in(Bot)".
4751   Node *progress = NULL;
4752 
4753 
4754   Node* old_base = base_memory();
4755   Node* empty_mem = empty_memory();
4756   if (old_base == empty_mem)
4757     return NULL; // Dead memory path.
4758 
4759   MergeMemNode* old_mbase;
4760   if (old_base != NULL && old_base->is_MergeMem())
4761     old_mbase = old_base->as_MergeMem();
4762   else
4763     old_mbase = NULL;
4764   Node* new_base = old_base;
4765 
4766   // simplify stacked MergeMems in base memory
4767   if (old_mbase)  new_base = old_mbase->base_memory();
4768 
4769   // the base memory might contribute new slices beyond my req()
4770   if (old_mbase)  grow_to_match(old_mbase);
4771 
4772   // Look carefully at the base node if it is a phi.
4773   PhiNode* phi_base;
4774   if (new_base != NULL && new_base->is_Phi())
4775     phi_base = new_base->as_Phi();
4776   else
4777     phi_base = NULL;
4778 
4779   Node*    phi_reg = NULL;
4780   uint     phi_len = (uint)-1;
4781   if (phi_base != NULL) {
4782     phi_reg = phi_base->region();
4783     phi_len = phi_base->req();
4784     // see if the phi is unfinished
4785     for (uint i = 1; i < phi_len; i++) {
4786       if (phi_base->in(i) == NULL) {
4787         // incomplete phi; do not look at it yet!
4788         phi_reg = NULL;
4789         phi_len = (uint)-1;
4790         break;
4791       }
4792     }
4793   }
4794 
4795   // Note:  We do not call verify_sparse on entry, because inputs
4796   // can normalize to the base_memory via subsume_node or similar
4797   // mechanisms.  This method repairs that damage.
4798 
4799   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4800 
4801   // Look at each slice.
4802   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4803     Node* old_in = in(i);
4804     // calculate the old memory value
4805     Node* old_mem = old_in;
4806     if (old_mem == empty_mem)  old_mem = old_base;
4807     assert(old_mem == memory_at(i), "");
4808 
4809     // maybe update (reslice) the old memory value
4810 
4811     // simplify stacked MergeMems
4812     Node* new_mem = old_mem;
4813     MergeMemNode* old_mmem;
4814     if (old_mem != NULL && old_mem->is_MergeMem())
4815       old_mmem = old_mem->as_MergeMem();
4816     else
4817       old_mmem = NULL;
4818     if (old_mmem == this) {
4819       // This can happen if loops break up and safepoints disappear.
4820       // A merge of BotPtr (default) with a RawPtr memory derived from a
4821       // safepoint can be rewritten to a merge of the same BotPtr with
4822       // the BotPtr phi coming into the loop.  If that phi disappears
4823       // also, we can end up with a self-loop of the mergemem.
4824       // In general, if loops degenerate and memory effects disappear,
4825       // a mergemem can be left looking at itself.  This simply means
4826       // that the mergemem's default should be used, since there is
4827       // no longer any apparent effect on this slice.
4828       // Note: If a memory slice is a MergeMem cycle, it is unreachable
4829       //       from start.  Update the input to TOP.
4830       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4831     }
4832     else if (old_mmem != NULL) {
4833       new_mem = old_mmem->memory_at(i);
4834     }
4835     // else preceding memory was not a MergeMem
4836 
4837     // maybe store down a new value
4838     Node* new_in = new_mem;
4839     if (new_in == new_base)  new_in = empty_mem;
4840 
4841     if (new_in != old_in) {
4842       // Warning:  Do not combine this "if" with the previous "if"
4843       // A memory slice might have be be rewritten even if it is semantically
4844       // unchanged, if the base_memory value has changed.
4845       set_req_X(i, new_in, phase);
4846       progress = this;          // Report progress
4847     }
4848   }
4849 
4850   if (new_base != old_base) {
4851     set_req_X(Compile::AliasIdxBot, new_base, phase);
4852     // Don't use set_base_memory(new_base), because we need to update du.
4853     assert(base_memory() == new_base, "");
4854     progress = this;
4855   }
4856 
4857   if( base_memory() == this ) {
4858     // a self cycle indicates this memory path is dead
4859     set_req(Compile::AliasIdxBot, empty_mem);
4860   }
4861 
4862   // Resolve external cycles by calling Ideal on a MergeMem base_memory
4863   // Recursion must occur after the self cycle check above
4864   if( base_memory()->is_MergeMem() ) {
4865     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4866     Node *m = phase->transform(new_mbase);  // Rollup any cycles
4867     if( m != NULL &&
4868         (m->is_top() ||
4869          (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4870       // propagate rollup of dead cycle to self
4871       set_req(Compile::AliasIdxBot, empty_mem);
4872     }
4873   }
4874 
4875   if( base_memory() == empty_mem ) {
4876     progress = this;
4877     // Cut inputs during Parse phase only.
4878     // During Optimize phase a dead MergeMem node will be subsumed by Top.
4879     if( !can_reshape ) {
4880       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4881         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4882       }
4883     }
4884   }
4885 
4886   if( !progress && base_memory()->is_Phi() && can_reshape ) {
4887     // Check if PhiNode::Ideal's "Split phis through memory merges"
4888     // transform should be attempted. Look for this->phi->this cycle.
4889     uint merge_width = req();
4890     if (merge_width > Compile::AliasIdxRaw) {
4891       PhiNode* phi = base_memory()->as_Phi();
4892       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4893         if (phi->in(i) == this) {
4894           phase->is_IterGVN()->_worklist.push(phi);
4895           break;
4896         }
4897       }
4898     }
4899   }
4900 
4901   assert(progress || verify_sparse(), "please, no dups of base");
4902   return progress;
4903 }
4904 
4905 //-------------------------set_base_memory-------------------------------------
4906 void MergeMemNode::set_base_memory(Node *new_base) {
4907   Node* empty_mem = empty_memory();
4908   set_req(Compile::AliasIdxBot, new_base);
4909   assert(memory_at(req()) == new_base, "must set default memory");
4910   // Clear out other occurrences of new_base:
4911   if (new_base != empty_mem) {
4912     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4913       if (in(i) == new_base)  set_req(i, empty_mem);
4914     }
4915   }
4916 }
4917 
4918 //------------------------------out_RegMask------------------------------------
4919 const RegMask &MergeMemNode::out_RegMask() const {
4920   return RegMask::Empty;
4921 }
4922 
4923 //------------------------------dump_spec--------------------------------------
4924 #ifndef PRODUCT
4925 void MergeMemNode::dump_spec(outputStream *st) const {
4926   st->print(" {");
4927   Node* base_mem = base_memory();
4928   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4929     Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4930     if (mem == base_mem) { st->print(" -"); continue; }
4931     st->print( " N%d:", mem->_idx );
4932     Compile::current()->get_adr_type(i)->dump_on(st);
4933   }
4934   st->print(" }");
4935 }
4936 #endif // !PRODUCT
4937 
4938 
4939 #ifdef ASSERT
4940 static bool might_be_same(Node* a, Node* b) {
4941   if (a == b)  return true;
4942   if (!(a->is_Phi() || b->is_Phi()))  return false;
4943   // phis shift around during optimization
4944   return true;  // pretty stupid...
4945 }
4946 
4947 // verify a narrow slice (either incoming or outgoing)
4948 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4949   if (!VerifyAliases)                return;  // don't bother to verify unless requested
4950   if (VMError::is_error_reported())  return;  // muzzle asserts when debugging an error
4951   if (Node::in_dump())               return;  // muzzle asserts when printing
4952   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4953   assert(n != NULL, "");
4954   // Elide intervening MergeMem's
4955   while (n->is_MergeMem()) {
4956     n = n->as_MergeMem()->memory_at(alias_idx);
4957   }
4958   Compile* C = Compile::current();
4959   const TypePtr* n_adr_type = n->adr_type();
4960   if (n == m->empty_memory()) {
4961     // Implicit copy of base_memory()
4962   } else if (n_adr_type != TypePtr::BOTTOM) {
4963     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4964     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4965   } else {
4966     // A few places like make_runtime_call "know" that VM calls are narrow,
4967     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4968     bool expected_wide_mem = false;
4969     if (n == m->base_memory()) {
4970       expected_wide_mem = true;
4971     } else if (alias_idx == Compile::AliasIdxRaw ||
4972                n == m->memory_at(Compile::AliasIdxRaw)) {
4973       expected_wide_mem = true;
4974     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4975       // memory can "leak through" calls on channels that
4976       // are write-once.  Allow this also.
4977       expected_wide_mem = true;
4978     }
4979     assert(expected_wide_mem, "expected narrow slice replacement");
4980   }
4981 }
4982 #else // !ASSERT
4983 #define verify_memory_slice(m,i,n) (void)(0)  // PRODUCT version is no-op
4984 #endif
4985 
4986 
4987 //-----------------------------memory_at---------------------------------------
4988 Node* MergeMemNode::memory_at(uint alias_idx) const {
4989   assert(alias_idx >= Compile::AliasIdxRaw ||
4990          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4991          "must avoid base_memory and AliasIdxTop");
4992 
4993   // Otherwise, it is a narrow slice.
4994   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4995   Compile *C = Compile::current();
4996   if (is_empty_memory(n)) {
4997     // the array is sparse; empty slots are the "top" node
4998     n = base_memory();
4999     assert(Node::in_dump()
5000            || n == NULL || n->bottom_type() == Type::TOP
5001            || n->adr_type() == NULL // address is TOP
5002            || n->adr_type() == TypePtr::BOTTOM
5003            || n->adr_type() == TypeRawPtr::BOTTOM
5004            || Compile::current()->AliasLevel() == 0,
5005            "must be a wide memory");
5006     // AliasLevel == 0 if we are organizing the memory states manually.
5007     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5008   } else {
5009     // make sure the stored slice is sane
5010     #ifdef ASSERT
5011     if (VMError::is_error_reported() || Node::in_dump()) {
5012     } else if (might_be_same(n, base_memory())) {
5013       // Give it a pass:  It is a mostly harmless repetition of the base.
5014       // This can arise normally from node subsumption during optimization.
5015     } else {
5016       verify_memory_slice(this, alias_idx, n);
5017     }
5018     #endif
5019   }
5020   return n;
5021 }
5022 
5023 //---------------------------set_memory_at-------------------------------------
5024 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5025   verify_memory_slice(this, alias_idx, n);
5026   Node* empty_mem = empty_memory();
5027   if (n == base_memory())  n = empty_mem;  // collapse default
5028   uint need_req = alias_idx+1;
5029   if (req() < need_req) {
5030     if (n == empty_mem)  return;  // already the default, so do not grow me
5031     // grow the sparse array
5032     do {
5033       add_req(empty_mem);
5034     } while (req() < need_req);
5035   }
5036   set_req( alias_idx, n );
5037 }
5038 
5039 
5040 
5041 //--------------------------iteration_setup------------------------------------
5042 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5043   if (other != NULL) {
5044     grow_to_match(other);
5045     // invariant:  the finite support of mm2 is within mm->req()
5046     #ifdef ASSERT
5047     for (uint i = req(); i < other->req(); i++) {
5048       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5049     }
5050     #endif
5051   }
5052   // Replace spurious copies of base_memory by top.
5053   Node* base_mem = base_memory();
5054   if (base_mem != NULL && !base_mem->is_top()) {
5055     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5056       if (in(i) == base_mem)
5057         set_req(i, empty_memory());
5058     }
5059   }
5060 }
5061 
5062 //---------------------------grow_to_match-------------------------------------
5063 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5064   Node* empty_mem = empty_memory();
5065   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5066   // look for the finite support of the other memory
5067   for (uint i = other->req(); --i >= req(); ) {
5068     if (other->in(i) != empty_mem) {
5069       uint new_len = i+1;
5070       while (req() < new_len)  add_req(empty_mem);
5071       break;
5072     }
5073   }
5074 }
5075 
5076 //---------------------------verify_sparse-------------------------------------
5077 #ifndef PRODUCT
5078 bool MergeMemNode::verify_sparse() const {
5079   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5080   Node* base_mem = base_memory();
5081   // The following can happen in degenerate cases, since empty==top.
5082   if (is_empty_memory(base_mem))  return true;
5083   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5084     assert(in(i) != NULL, "sane slice");
5085     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
5086   }
5087   return true;
5088 }
5089 
5090 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5091   Node* n;
5092   n = mm->in(idx);
5093   if (mem == n)  return true;  // might be empty_memory()
5094   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5095   if (mem == n)  return true;
5096   return false;
5097 }
5098 #endif // !PRODUCT