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