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   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 854           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 855           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 856            in_bytes(JavaThread::osthread_offset()) ||
 857            adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 858            in_bytes(JavaThread::threadObj_offset())));












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