1 /*
   2  * Copyright (c) 1997, 2020, 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 "ci/ciFlatArrayKlass.hpp"
  27 #include "classfile/javaClasses.hpp"
  28 #include "classfile/systemDictionary.hpp"
  29 #include "compiler/compileLog.hpp"
  30 #include "gc/shared/barrierSet.hpp"
  31 #include "gc/shared/c2/barrierSetC2.hpp"
  32 #include "memory/allocation.inline.hpp"
  33 #include "memory/resourceArea.hpp"
  34 #include "oops/objArrayKlass.hpp"
  35 #include "opto/addnode.hpp"
  36 #include "opto/arraycopynode.hpp"
  37 #include "opto/cfgnode.hpp"
  38 #include "opto/compile.hpp"
  39 #include "opto/connode.hpp"
  40 #include "opto/convertnode.hpp"
  41 #include "opto/inlinetypenode.hpp"
  42 #include "opto/loopnode.hpp"
  43 #include "opto/machnode.hpp"
  44 #include "opto/matcher.hpp"
  45 #include "opto/memnode.hpp"
  46 #include "opto/mulnode.hpp"
  47 #include "opto/narrowptrnode.hpp"
  48 #include "opto/phaseX.hpp"
  49 #include "opto/regmask.hpp"
  50 #include "opto/rootnode.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       phase->is_IterGVN()->_worklist.push(this);
 314       return NodeSentinel; // caller will return NULL
 315     }
 316   }
 317   // Ignore if memory is dead, or self-loop
 318   Node *mem = in(MemNode::Memory);
 319   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
 320   assert(mem != this, "dead loop in MemNode::Ideal");
 321 
 322   if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
 323     // This memory slice may be dead.
 324     // Delay this mem node transformation until the memory is processed.
 325     phase->is_IterGVN()->_worklist.push(this);
 326     return NodeSentinel; // caller will return NULL
 327   }
 328 
 329   Node *address = in(MemNode::Address);
 330   const Type *t_adr = phase->type(address);
 331   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return NULL
 332 
 333   if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
 334     // Unsafe off-heap access with zero address. Remove access and other control users
 335     // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
 336     assert(ctl != NULL, "unsafe accesses should be control dependent");
 337     for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
 338       Node* u = ctl->fast_out(i);
 339       if (u != ctl) {
 340         igvn->rehash_node_delayed(u);
 341         int nb = u->replace_edge(ctl, phase->C->top());
 342         --i, imax -= nb;
 343       }
 344     }
 345     Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
 346     Node* halt = igvn->transform(new HaltNode(ctl, frame, "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     phase->is_IterGVN()->_worklist.push(this);
 357     return NodeSentinel; // caller will return NULL
 358   }
 359 
 360   // Do NOT remove or optimize the next lines: ensure a new alias index
 361   // is allocated for an oop pointer type before Escape Analysis.
 362   // Note: C++ will not remove it since the call has side effect.
 363   if (t_adr->isa_oopptr()) {
 364     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 365   }
 366 
 367   Node* base = NULL;
 368   if (address->is_AddP()) {
 369     base = address->in(AddPNode::Base);
 370   }
 371   if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
 372       !t_adr->isa_rawptr()) {
 373     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
 374     // Skip this node optimization if its address has TOP base.
 375     return NodeSentinel; // caller will return NULL
 376   }
 377 
 378   // Avoid independent memory operations
 379   Node* old_mem = mem;
 380 
 381   // The code which unhooks non-raw memories from complete (macro-expanded)
 382   // initializations was removed. After macro-expansion all stores catched
 383   // by Initialize node became raw stores and there is no information
 384   // which memory slices they modify. So it is unsafe to move any memory
 385   // operation above these stores. Also in most cases hooked non-raw memories
 386   // were already unhooked by using information from detect_ptr_independence()
 387   // and find_previous_store().
 388 
 389   if (mem->is_MergeMem()) {
 390     MergeMemNode* mmem = mem->as_MergeMem();
 391     const TypePtr *tp = t_adr->is_ptr();
 392 
 393     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 394   }
 395 
 396   if (mem != old_mem) {
 397     set_req(MemNode::Memory, mem);
 398     if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
 399       igvn->_worklist.push(old_mem);
 400     }
 401     if (phase->type(mem) == Type::TOP) return NodeSentinel;
 402     return this;
 403   }
 404 
 405   // let the subclass continue analyzing...
 406   return NULL;
 407 }
 408 
 409 // Helper function for proving some simple control dominations.
 410 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 411 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 412 // is not a constant (dominated by the method's StartNode).
 413 // Used by MemNode::find_previous_store to prove that the
 414 // control input of a memory operation predates (dominates)
 415 // an allocation it wants to look past.
 416 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
 417   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
 418     return false; // Conservative answer for dead code
 419 
 420   // Check 'dom'. Skip Proj and CatchProj nodes.
 421   dom = dom->find_exact_control(dom);
 422   if (dom == NULL || dom->is_top())
 423     return false; // Conservative answer for dead code
 424 
 425   if (dom == sub) {
 426     // For the case when, for example, 'sub' is Initialize and the original
 427     // 'dom' is Proj node of the 'sub'.
 428     return false;
 429   }
 430 
 431   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
 432     return true;
 433 
 434   // 'dom' dominates 'sub' if its control edge and control edges
 435   // of all its inputs dominate or equal to sub's control edge.
 436 
 437   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 438   // Or Region for the check in LoadNode::Ideal();
 439   // 'sub' should have sub->in(0) != NULL.
 440   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 441          sub->is_Region() || sub->is_Call(), "expecting only these nodes");
 442 
 443   // Get control edge of 'sub'.
 444   Node* orig_sub = sub;
 445   sub = sub->find_exact_control(sub->in(0));
 446   if (sub == NULL || sub->is_top())
 447     return false; // Conservative answer for dead code
 448 
 449   assert(sub->is_CFG(), "expecting control");
 450 
 451   if (sub == dom)
 452     return true;
 453 
 454   if (sub->is_Start() || sub->is_Root())
 455     return false;
 456 
 457   {
 458     // Check all control edges of 'dom'.
 459 
 460     ResourceMark rm;
 461     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 that must have set (can_see_stored_value=true) or
 534 // could have set (can_see_stored_value=false) the value for this load
 535 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
 536   if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
 537                                                mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
 538     if (ld_alloc != NULL) {
 539       // Check if there is an array copy for a clone
 540       Node* mb = mem->in(0);
 541       ArrayCopyNode* ac = NULL;
 542       if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
 543           mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
 544         ac = mb->in(0)->in(0)->as_ArrayCopy();
 545       } else {
 546         // Step over GC barrier when ReduceInitialCardMarks is disabled
 547         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 548         Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
 549 
 550         if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
 551           ac = control_proj_ac->in(0)->as_ArrayCopy();
 552         }
 553       }
 554 
 555       if (ac != NULL && ac->is_clonebasic()) {
 556         AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase);
 557         if (alloc != NULL && alloc == ld_alloc) {
 558           return ac;
 559         }
 560       }
 561     }
 562   } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
 563     ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
 564 
 565     if (ac->is_arraycopy_validated() ||
 566         ac->is_copyof_validated() ||
 567         ac->is_copyofrange_validated()) {
 568       Node* ld_addp = in(MemNode::Address);
 569       if (ld_addp->is_AddP()) {
 570         Node* ld_base = ld_addp->in(AddPNode::Address);
 571         Node* ld_offs = ld_addp->in(AddPNode::Offset);
 572 
 573         Node* dest = ac->in(ArrayCopyNode::Dest);
 574 
 575         if (dest == ld_base) {
 576           const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 577           if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
 578             return ac;
 579           }
 580           if (!can_see_stored_value) {
 581             mem = ac->in(TypeFunc::Memory);
 582           }
 583         }
 584       }
 585     }
 586   }
 587   return NULL;
 588 }
 589 
 590 // The logic for reordering loads and stores uses four steps:
 591 // (a) Walk carefully past stores and initializations which we
 592 //     can prove are independent of this load.
 593 // (b) Observe that the next memory state makes an exact match
 594 //     with self (load or store), and locate the relevant store.
 595 // (c) Ensure that, if we were to wire self directly to the store,
 596 //     the optimizer would fold it up somehow.
 597 // (d) Do the rewiring, and return, depending on some other part of
 598 //     the optimizer to fold up the load.
 599 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 600 // specific to loads and stores, so they are handled by the callers.
 601 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 602 //
 603 Node* MemNode::find_previous_store(PhaseTransform* phase) {
 604   Node*         ctrl   = in(MemNode::Control);
 605   Node*         adr    = in(MemNode::Address);
 606   intptr_t      offset = 0;
 607   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 608   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
 609 
 610   if (offset == Type::OffsetBot)
 611     return NULL;            // cannot unalias unless there are precise offsets
 612 
 613   const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
 614   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 615 
 616   intptr_t size_in_bytes = memory_size();
 617 
 618   Node* mem = in(MemNode::Memory);   // start searching here...
 619 
 620   int cnt = 50;             // Cycle limiter
 621   for (;;) {                // While we can dance past unrelated stores...
 622     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 623 
 624     Node* prev = mem;
 625     if (mem->is_Store()) {
 626       Node* st_adr = mem->in(MemNode::Address);
 627       intptr_t st_offset = 0;
 628       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 629       if (st_base == NULL)
 630         break;              // inscrutable pointer
 631 
 632       // For raw accesses it's not enough to prove that constant offsets don't intersect.
 633       // We need the bases to be the equal in order for the offset check to make sense.
 634       if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
 635         break;
 636       }
 637 
 638       if (st_offset != offset && st_offset != Type::OffsetBot) {
 639         const int MAX_STORE = BytesPerLong;
 640         if (st_offset >= offset + size_in_bytes ||
 641             st_offset <= offset - MAX_STORE ||
 642             st_offset <= offset - mem->as_Store()->memory_size()) {
 643           // Success:  The offsets are provably independent.
 644           // (You may ask, why not just test st_offset != offset and be done?
 645           // The answer is that stores of different sizes can co-exist
 646           // in the same sequence of RawMem effects.  We sometimes initialize
 647           // a whole 'tile' of array elements with a single jint or jlong.)
 648           mem = mem->in(MemNode::Memory);
 649           continue;           // (a) advance through independent store memory
 650         }
 651       }
 652       if (st_base != base &&
 653           detect_ptr_independence(base, alloc,
 654                                   st_base,
 655                                   AllocateNode::Ideal_allocation(st_base, phase),
 656                                   phase)) {
 657         // Success:  The bases are provably independent.
 658         mem = mem->in(MemNode::Memory);
 659         continue;           // (a) advance through independent store memory
 660       }
 661 
 662       // (b) At this point, if the bases or offsets do not agree, we lose,
 663       // since we have not managed to prove 'this' and 'mem' independent.
 664       if (st_base == base && st_offset == offset) {
 665         return mem;         // let caller handle steps (c), (d)
 666       }
 667 
 668     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 669       InitializeNode* st_init = mem->in(0)->as_Initialize();
 670       AllocateNode*  st_alloc = st_init->allocation();
 671       if (st_alloc == NULL)
 672         break;              // something degenerated
 673       bool known_identical = false;
 674       bool known_independent = false;
 675       if (alloc == st_alloc)
 676         known_identical = true;
 677       else if (alloc != NULL)
 678         known_independent = true;
 679       else if (all_controls_dominate(this, st_alloc))
 680         known_independent = true;
 681 
 682       if (known_independent) {
 683         // The bases are provably independent: Either they are
 684         // manifestly distinct allocations, or else the control
 685         // of this load dominates the store's allocation.
 686         int alias_idx = phase->C->get_alias_index(adr_type());
 687         if (alias_idx == Compile::AliasIdxRaw) {
 688           mem = st_alloc->in(TypeFunc::Memory);
 689         } else {
 690           mem = st_init->memory(alias_idx);
 691         }
 692         continue;           // (a) advance through independent store memory
 693       }
 694 
 695       // (b) at this point, if we are not looking at a store initializing
 696       // the same allocation we are loading from, we lose.
 697       if (known_identical) {
 698         // From caller, can_see_stored_value will consult find_captured_store.
 699         return mem;         // let caller handle steps (c), (d)
 700       }
 701 
 702     } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
 703       if (prev != mem) {
 704         // Found an arraycopy but it doesn't affect that load
 705         continue;
 706       }
 707       // Found an arraycopy that may affect that load
 708       return mem;
 709     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
 710       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 711       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 712         // ArrayCopyNodes processed here as well.
 713         CallNode *call = mem->in(0)->as_Call();
 714         if (!call->may_modify(addr_t, phase)) {
 715           mem = call->in(TypeFunc::Memory);
 716           continue;         // (a) advance through independent call memory
 717         }
 718       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 719         ArrayCopyNode* ac = NULL;
 720         if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
 721           break;
 722         }
 723         mem = mem->in(0)->in(TypeFunc::Memory);
 724         continue;           // (a) advance through independent MemBar memory
 725       } else if (mem->is_ClearArray()) {
 726         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 727           // (the call updated 'mem' value)
 728           continue;         // (a) advance through independent allocation memory
 729         } else {
 730           // Can not bypass initialization of the instance
 731           // we are looking for.
 732           return mem;
 733         }
 734       } else if (mem->is_MergeMem()) {
 735         int alias_idx = phase->C->get_alias_index(adr_type());
 736         mem = mem->as_MergeMem()->memory_at(alias_idx);
 737         continue;           // (a) advance through independent MergeMem memory
 738       }
 739     }
 740 
 741     // Unless there is an explicit 'continue', we must bail out here,
 742     // because 'mem' is an inscrutable memory state (e.g., a call).
 743     break;
 744   }
 745 
 746   return NULL;              // bail out
 747 }
 748 
 749 //----------------------calculate_adr_type-------------------------------------
 750 // Helper function.  Notices when the given type of address hits top or bottom.
 751 // Also, asserts a cross-check of the type against the expected address type.
 752 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 753   if (t == Type::TOP)  return NULL; // does not touch memory any more?
 754   #ifdef ASSERT
 755   if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump())  cross_check = NULL;
 756   #endif
 757   const TypePtr* tp = t->isa_ptr();
 758   if (tp == NULL) {
 759     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 760     return TypePtr::BOTTOM;           // touches lots of memory
 761   } else {
 762     #ifdef ASSERT
 763     // %%%% [phh] We don't check the alias index if cross_check is
 764     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 765     if (cross_check != NULL &&
 766         cross_check != TypePtr::BOTTOM &&
 767         cross_check != TypeRawPtr::BOTTOM) {
 768       // Recheck the alias index, to see if it has changed (due to a bug).
 769       Compile* C = Compile::current();
 770       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 771              "must stay in the original alias category");
 772       // The type of the address must be contained in the adr_type,
 773       // disregarding "null"-ness.
 774       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 775       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 776       assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
 777              "real address must not escape from expected memory type");
 778     }
 779     #endif
 780     return tp;
 781   }
 782 }
 783 
 784 //=============================================================================
 785 // Should LoadNode::Ideal() attempt to remove control edges?
 786 bool LoadNode::can_remove_control() const {
 787   return true;
 788 }
 789 uint LoadNode::size_of() const { return sizeof(*this); }
 790 bool LoadNode::cmp( const Node &n ) const
 791 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
 792 const Type *LoadNode::bottom_type() const { return _type; }
 793 uint LoadNode::ideal_reg() const {
 794   return _type->ideal_reg();
 795 }
 796 
 797 #ifndef PRODUCT
 798 void LoadNode::dump_spec(outputStream *st) const {
 799   MemNode::dump_spec(st);
 800   if( !Verbose && !WizardMode ) {
 801     // standard dump does this in Verbose and WizardMode
 802     st->print(" #"); _type->dump_on(st);
 803   }
 804   if (!depends_only_on_test()) {
 805     st->print(" (does not depend only on test)");
 806   }
 807 }
 808 #endif
 809 
 810 #ifdef ASSERT
 811 //----------------------------is_immutable_value-------------------------------
 812 // Helper function to allow a raw load without control edge for some cases
 813 bool LoadNode::is_immutable_value(Node* adr) {
 814   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 815           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 816           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 817            in_bytes(JavaThread::osthread_offset())));
 818 }
 819 #endif
 820 
 821 //----------------------------LoadNode::make-----------------------------------
 822 // Polymorphic factory method:
 823 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
 824                      ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
 825   Compile* C = gvn.C;
 826 
 827   // sanity check the alias category against the created node type
 828   assert(!(adr_type->isa_oopptr() &&
 829            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
 830          "use LoadKlassNode instead");
 831   assert(!(adr_type->isa_aryptr() &&
 832            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
 833          "use LoadRangeNode instead");
 834   // Check control edge of raw loads
 835   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
 836           // oop will be recorded in oop map if load crosses safepoint
 837           rt->isa_oopptr() || is_immutable_value(adr),
 838           "raw memory operations should have control edge");
 839   LoadNode* load = NULL;
 840   switch (bt) {
 841   case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 842   case T_BYTE:    load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 843   case T_INT:     load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 844   case T_CHAR:    load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 845   case T_SHORT:   load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 846   case T_LONG:    load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
 847   case T_FLOAT:   load = new LoadFNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
 848   case T_DOUBLE:  load = new LoadDNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
 849   case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(),  mo, control_dependency); break;
 850   case T_INLINE_TYPE:
 851   case T_OBJECT:
 852 #ifdef _LP64
 853     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 854       load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
 855     } else
 856 #endif
 857     {
 858       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
 859       load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
 860     }
 861     break;
 862   default:
 863     ShouldNotReachHere();
 864     break;
 865   }
 866   assert(load != NULL, "LoadNode should have been created");
 867   if (unaligned) {
 868     load->set_unaligned_access();
 869   }
 870   if (mismatched) {
 871     load->set_mismatched_access();
 872   }
 873   if (unsafe) {
 874     load->set_unsafe_access();
 875   }
 876   load->set_barrier_data(barrier_data);
 877   if (load->Opcode() == Op_LoadN) {
 878     Node* ld = gvn.transform(load);
 879     return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
 880   }
 881 
 882   return load;
 883 }
 884 
 885 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
 886                                   ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
 887   bool require_atomic = true;
 888   LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
 889   if (unaligned) {
 890     load->set_unaligned_access();
 891   }
 892   if (mismatched) {
 893     load->set_mismatched_access();
 894   }
 895   if (unsafe) {
 896     load->set_unsafe_access();
 897   }
 898   load->set_barrier_data(barrier_data);
 899   return load;
 900 }
 901 
 902 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
 903                                   ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
 904   bool require_atomic = true;
 905   LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
 906   if (unaligned) {
 907     load->set_unaligned_access();
 908   }
 909   if (mismatched) {
 910     load->set_mismatched_access();
 911   }
 912   if (unsafe) {
 913     load->set_unsafe_access();
 914   }
 915   load->set_barrier_data(barrier_data);
 916   return load;
 917 }
 918 
 919 
 920 
 921 //------------------------------hash-------------------------------------------
 922 uint LoadNode::hash() const {
 923   // unroll addition of interesting fields
 924   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
 925 }
 926 
 927 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
 928   if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
 929     bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
 930     bool is_stable_ary = FoldStableValues &&
 931                          (tp != NULL) && (tp->isa_aryptr() != NULL) &&
 932                          tp->isa_aryptr()->is_stable();
 933 
 934     return (eliminate_boxing && non_volatile) || is_stable_ary;
 935   }
 936 
 937   return false;
 938 }
 939 
 940 // Is the value loaded previously stored by an arraycopy? If so return
 941 // a load node that reads from the source array so we may be able to
 942 // optimize out the ArrayCopy node later.
 943 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
 944   Node* ld_adr = in(MemNode::Address);
 945   intptr_t ld_off = 0;
 946   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 947   Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
 948   if (ac != NULL) {
 949     assert(ac->is_ArrayCopy(), "what kind of node can this be?");
 950 
 951     Node* mem = ac->in(TypeFunc::Memory);
 952     Node* ctl = ac->in(0);
 953     Node* src = ac->in(ArrayCopyNode::Src);
 954 
 955     if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
 956       return NULL;
 957     }
 958 
 959     LoadNode* ld = clone()->as_Load();
 960     Node* addp = in(MemNode::Address)->clone();
 961     if (ac->as_ArrayCopy()->is_clonebasic()) {
 962       assert(ld_alloc != NULL, "need an alloc");
 963       assert(addp->is_AddP(), "address must be addp");
 964       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 965       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
 966       assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
 967       addp->set_req(AddPNode::Base, src);
 968       addp->set_req(AddPNode::Address, src);
 969     } else {
 970       assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
 971              ac->as_ArrayCopy()->is_copyof_validated() ||
 972              ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
 973       assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
 974       addp->set_req(AddPNode::Base, src);
 975       addp->set_req(AddPNode::Address, src);
 976 
 977       const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
 978       BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
 979       uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
 980       uint shift  = exact_log2(type2aelembytes(ary_elem));
 981       if (ary_t->klass()->is_flat_array_klass()) {
 982         ciFlatArrayKlass* vak = ary_t->klass()->as_flat_array_klass();
 983         shift = vak->log2_element_size();
 984       }
 985 
 986       Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 987 #ifdef _LP64
 988       diff = phase->transform(new ConvI2LNode(diff));
 989 #endif
 990       diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
 991 
 992       Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
 993       addp->set_req(AddPNode::Offset, offset);
 994     }
 995     addp = phase->transform(addp);
 996 #ifdef ASSERT
 997     const TypePtr* adr_type = phase->type(addp)->is_ptr();
 998     ld->_adr_type = adr_type;
 999 #endif
1000     ld->set_req(MemNode::Address, addp);
1001     ld->set_req(0, ctl);
1002     ld->set_req(MemNode::Memory, mem);
1003     // load depends on the tests that validate the arraycopy
1004     ld->_control_dependency = UnknownControl;
1005     return ld;
1006   }
1007   return NULL;
1008 }
1009 
1010 
1011 //---------------------------can_see_stored_value------------------------------
1012 // This routine exists to make sure this set of tests is done the same
1013 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
1014 // will change the graph shape in a way which makes memory alive twice at the
1015 // same time (uses the Oracle model of aliasing), then some
1016 // LoadXNode::Identity will fold things back to the equivalence-class model
1017 // of aliasing.
1018 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1019   Node* ld_adr = in(MemNode::Address);
1020   intptr_t ld_off = 0;
1021   Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1022   Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1023   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1024   Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1025   // This is more general than load from boxing objects.
1026   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1027     uint alias_idx = atp->index();
1028     bool final = !atp->is_rewritable();
1029     Node* result = NULL;
1030     Node* current = st;
1031     // Skip through chains of MemBarNodes checking the MergeMems for
1032     // new states for the slice of this load.  Stop once any other
1033     // kind of node is encountered.  Loads from final memory can skip
1034     // through any kind of MemBar but normal loads shouldn't skip
1035     // through MemBarAcquire since the could allow them to move out of
1036     // a synchronized region.
1037     while (current->is_Proj()) {
1038       int opc = current->in(0)->Opcode();
1039       if ((final && (opc == Op_MemBarAcquire ||
1040                      opc == Op_MemBarAcquireLock ||
1041                      opc == Op_LoadFence)) ||
1042           opc == Op_MemBarRelease ||
1043           opc == Op_StoreFence ||
1044           opc == Op_MemBarReleaseLock ||
1045           opc == Op_MemBarStoreStore ||
1046           opc == Op_MemBarCPUOrder) {
1047         Node* mem = current->in(0)->in(TypeFunc::Memory);
1048         if (mem->is_MergeMem()) {
1049           MergeMemNode* merge = mem->as_MergeMem();
1050           Node* new_st = merge->memory_at(alias_idx);
1051           if (new_st == merge->base_memory()) {
1052             // Keep searching
1053             current = new_st;
1054             continue;
1055           }
1056           // Save the new memory state for the slice and fall through
1057           // to exit.
1058           result = new_st;
1059         }
1060       }
1061       break;
1062     }
1063     if (result != NULL) {
1064       st = result;
1065     }
1066   }
1067 
1068   // Loop around twice in the case Load -> Initialize -> Store.
1069   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1070   for (int trip = 0; trip <= 1; trip++) {
1071 
1072     if (st->is_Store()) {
1073       Node* st_adr = st->in(MemNode::Address);
1074       if (!phase->eqv(st_adr, ld_adr)) {
1075         // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1076         intptr_t st_off = 0;
1077         Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1078         if (ld_base == NULL)                                   return NULL;
1079         if (st_base == NULL)                                   return NULL;
1080         if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1081         if (ld_off != st_off)                                  return NULL;
1082         if (ld_off == Type::OffsetBot)                         return NULL;
1083         // Same base, same offset.
1084         // Possible improvement for arrays: check index value instead of absolute offset.
1085 
1086         // At this point we have proven something like this setup:
1087         //   B = << base >>
1088         //   L =  LoadQ(AddP(Check/CastPP(B), #Off))
1089         //   S = StoreQ(AddP(             B , #Off), V)
1090         // (Actually, we haven't yet proven the Q's are the same.)
1091         // In other words, we are loading from a casted version of
1092         // the same pointer-and-offset that we stored to.
1093         // Casted version may carry a dependency and it is respected.
1094         // Thus, we are able to replace L by V.
1095       }
1096       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1097       if (store_Opcode() != st->Opcode())
1098         return NULL;
1099       return st->in(MemNode::ValueIn);
1100     }
1101 
1102     // A load from a freshly-created object always returns zero.
1103     // (This can happen after LoadNode::Ideal resets the load's memory input
1104     // to find_captured_store, which returned InitializeNode::zero_memory.)
1105     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1106         (st->in(0) == ld_alloc) &&
1107         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1108       // return a zero value for the load's basic type
1109       // (This is one of the few places where a generic PhaseTransform
1110       // can create new nodes.  Think of it as lazily manifesting
1111       // virtually pre-existing constants.)
1112       assert(memory_type() != T_INLINE_TYPE, "should not be used for inline types");
1113       Node* default_value = ld_alloc->in(AllocateNode::DefaultValue);
1114       if (default_value != NULL) {
1115         return default_value;
1116       }
1117       assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
1118       return phase->zerocon(memory_type());
1119     }
1120 
1121     // A load from an initialization barrier can match a captured store.
1122     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1123       InitializeNode* init = st->in(0)->as_Initialize();
1124       AllocateNode* alloc = init->allocation();
1125       if ((alloc != NULL) && (alloc == ld_alloc)) {
1126         // examine a captured store value
1127         st = init->find_captured_store(ld_off, memory_size(), phase);
1128         if (st != NULL) {
1129           continue;             // take one more trip around
1130         }
1131       }
1132     }
1133 
1134     // Load boxed value from result of valueOf() call is input parameter.
1135     if (this->is_Load() && ld_adr->is_AddP() &&
1136         (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1137       intptr_t ignore = 0;
1138       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1139       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1140       base = bs->step_over_gc_barrier(base);
1141       if (base != NULL && base->is_Proj() &&
1142           base->as_Proj()->_con == TypeFunc::Parms &&
1143           base->in(0)->is_CallStaticJava() &&
1144           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1145         return base->in(0)->in(TypeFunc::Parms);
1146       }
1147     }
1148 
1149     break;
1150   }
1151 
1152   return NULL;
1153 }
1154 
1155 //----------------------is_instance_field_load_with_local_phi------------------
1156 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1157   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1158       in(Address)->is_AddP() ) {
1159     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1160     // Only instances and boxed values.
1161     if( t_oop != NULL &&
1162         (t_oop->is_ptr_to_boxed_value() ||
1163          t_oop->is_known_instance_field()) &&
1164         t_oop->offset() != Type::OffsetBot &&
1165         t_oop->offset() != Type::OffsetTop) {
1166       return true;
1167     }
1168   }
1169   return false;
1170 }
1171 
1172 //------------------------------Identity---------------------------------------
1173 // Loads are identity if previous store is to same address
1174 Node* LoadNode::Identity(PhaseGVN* phase) {
1175   // Loading from an InlineTypePtr? The InlineTypePtr has the values of
1176   // all fields as input. Look for the field with matching offset.
1177   Node* addr = in(Address);
1178   intptr_t offset;
1179   Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
1180   if (base != NULL && base->is_InlineTypePtr() && offset > oopDesc::klass_offset_in_bytes()) {
1181     Node* value = base->as_InlineTypePtr()->field_value_by_offset((int)offset, true);
1182     if (value->is_InlineType()) {
1183       // Non-flattened inline type field
1184       InlineTypeNode* vt = value->as_InlineType();
1185       if (vt->is_allocated(phase)) {
1186         value = vt->get_oop();
1187       } else {
1188         // Not yet allocated, bail out
1189         value = NULL;
1190       }
1191     }
1192     if (value != NULL) {
1193       if (Opcode() == Op_LoadN) {
1194         // Encode oop value if we are loading a narrow oop
1195         assert(!phase->type(value)->isa_narrowoop(), "should already be decoded");
1196         value = phase->transform(new EncodePNode(value, bottom_type()));
1197       }
1198       return value;
1199     }
1200   }
1201 
1202   // If the previous store-maker is the right kind of Store, and the store is
1203   // to the same address, then we are equal to the value stored.
1204   Node* mem = in(Memory);
1205   Node* value = can_see_stored_value(mem, phase);
1206   if( value ) {
1207     // byte, short & char stores truncate naturally.
1208     // A load has to load the truncated value which requires
1209     // some sort of masking operation and that requires an
1210     // Ideal call instead of an Identity call.
1211     if (memory_size() < BytesPerInt) {
1212       // If the input to the store does not fit with the load's result type,
1213       // it must be truncated via an Ideal call.
1214       if (!phase->type(value)->higher_equal(phase->type(this)))
1215         return this;
1216     }
1217     // (This works even when value is a Con, but LoadNode::Value
1218     // usually runs first, producing the singleton type of the Con.)
1219     return value;
1220   }
1221 
1222   // Search for an existing data phi which was generated before for the same
1223   // instance's field to avoid infinite generation of phis in a loop.
1224   Node *region = mem->in(0);
1225   if (is_instance_field_load_with_local_phi(region)) {
1226     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1227     int this_index  = phase->C->get_alias_index(addr_t);
1228     int this_offset = addr_t->offset();
1229     int this_iid    = addr_t->instance_id();
1230     if (!addr_t->is_known_instance() &&
1231          addr_t->is_ptr_to_boxed_value()) {
1232       // Use _idx of address base (could be Phi node) for boxed values.
1233       intptr_t   ignore = 0;
1234       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1235       if (base == NULL) {
1236         return this;
1237       }
1238       this_iid = base->_idx;
1239     }
1240     const Type* this_type = bottom_type();
1241     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1242       Node* phi = region->fast_out(i);
1243       if (phi->is_Phi() && phi != mem &&
1244           phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1245         return phi;
1246       }
1247     }
1248   }
1249 
1250   return this;
1251 }
1252 
1253 // Construct an equivalent unsigned load.
1254 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1255   BasicType bt = T_ILLEGAL;
1256   const Type* rt = NULL;
1257   switch (Opcode()) {
1258     case Op_LoadUB: return this;
1259     case Op_LoadUS: return this;
1260     case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1261     case Op_LoadS: bt = T_CHAR;    rt = TypeInt::CHAR;  break;
1262     default:
1263       assert(false, "no unsigned variant: %s", Name());
1264       return NULL;
1265   }
1266   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1267                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1268                         is_unaligned_access(), is_mismatched_access());
1269 }
1270 
1271 // Construct an equivalent signed load.
1272 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1273   BasicType bt = T_ILLEGAL;
1274   const Type* rt = NULL;
1275   switch (Opcode()) {
1276     case Op_LoadUB: bt = T_BYTE;  rt = TypeInt::BYTE;  break;
1277     case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1278     case Op_LoadB: // fall through
1279     case Op_LoadS: // fall through
1280     case Op_LoadI: // fall through
1281     case Op_LoadL: return this;
1282     default:
1283       assert(false, "no signed variant: %s", Name());
1284       return NULL;
1285   }
1286   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1287                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1288                         is_unaligned_access(), is_mismatched_access());
1289 }
1290 
1291 // We're loading from an object which has autobox behaviour.
1292 // If this object is result of a valueOf call we'll have a phi
1293 // merging a newly allocated object and a load from the cache.
1294 // We want to replace this load with the original incoming
1295 // argument to the valueOf call.
1296 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1297   assert(phase->C->eliminate_boxing(), "sanity");
1298   intptr_t ignore = 0;
1299   Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1300   if ((base == NULL) || base->is_Phi()) {
1301     // Push the loads from the phi that comes from valueOf up
1302     // through it to allow elimination of the loads and the recovery
1303     // of the original value. It is done in split_through_phi().
1304     return NULL;
1305   } else if (base->is_Load() ||
1306              (base->is_DecodeN() && base->in(1)->is_Load())) {
1307     // Eliminate the load of boxed value for integer types from the cache
1308     // array by deriving the value from the index into the array.
1309     // Capture the offset of the load and then reverse the computation.
1310 
1311     // Get LoadN node which loads a boxing object from 'cache' array.
1312     if (base->is_DecodeN()) {
1313       base = base->in(1);
1314     }
1315     if (!base->in(Address)->is_AddP()) {
1316       return NULL; // Complex address
1317     }
1318     AddPNode* address = base->in(Address)->as_AddP();
1319     Node* cache_base = address->in(AddPNode::Base);
1320     if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1321       // Get ConP node which is static 'cache' field.
1322       cache_base = cache_base->in(1);
1323     }
1324     if ((cache_base != NULL) && cache_base->is_Con()) {
1325       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1326       if ((base_type != NULL) && base_type->is_autobox_cache()) {
1327         Node* elements[4];
1328         int shift = exact_log2(type2aelembytes(T_OBJECT));
1329         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1330         if (count > 0 && elements[0]->is_Con() &&
1331             (count == 1 ||
1332              (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1333                             elements[1]->in(2) == phase->intcon(shift)))) {
1334           ciObjArray* array = base_type->const_oop()->as_obj_array();
1335           // Fetch the box object cache[0] at the base of the array and get its value
1336           ciInstance* box = array->obj_at(0)->as_instance();
1337           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1338           assert(ik->is_box_klass(), "sanity");
1339           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1340           if (ik->nof_nonstatic_fields() == 1) {
1341             // This should be true nonstatic_field_at requires calling
1342             // nof_nonstatic_fields so check it anyway
1343             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1344             BasicType bt = c.basic_type();
1345             // Only integer types have boxing cache.
1346             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1347                    bt == T_BYTE    || bt == T_SHORT ||
1348                    bt == T_INT     || bt == T_LONG, "wrong type = %s", type2name(bt));
1349             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1350             if (cache_low != (int)cache_low) {
1351               return NULL; // should not happen since cache is array indexed by value
1352             }
1353             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1354             if (offset != (int)offset) {
1355               return NULL; // should not happen since cache is array indexed by value
1356             }
1357            // Add up all the offsets making of the address of the load
1358             Node* result = elements[0];
1359             for (int i = 1; i < count; i++) {
1360               result = phase->transform(new AddXNode(result, elements[i]));
1361             }
1362             // Remove the constant offset from the address and then
1363             result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1364             // remove the scaling of the offset to recover the original index.
1365             if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1366               // Peel the shift off directly but wrap it in a dummy node
1367               // since Ideal can't return existing nodes
1368               result = new RShiftXNode(result->in(1), phase->intcon(0));
1369             } else if (result->is_Add() && result->in(2)->is_Con() &&
1370                        result->in(1)->Opcode() == Op_LShiftX &&
1371                        result->in(1)->in(2) == phase->intcon(shift)) {
1372               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1373               // but for boxing cache access we know that X<<Z will not overflow
1374               // (there is range check) so we do this optimizatrion by hand here.
1375               Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1376               result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1377             } else {
1378               result = new RShiftXNode(result, phase->intcon(shift));
1379             }
1380 #ifdef _LP64
1381             if (bt != T_LONG) {
1382               result = new ConvL2INode(phase->transform(result));
1383             }
1384 #else
1385             if (bt == T_LONG) {
1386               result = new ConvI2LNode(phase->transform(result));
1387             }
1388 #endif
1389             // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1390             // Need to preserve unboxing load type if it is unsigned.
1391             switch(this->Opcode()) {
1392               case Op_LoadUB:
1393                 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1394                 break;
1395               case Op_LoadUS:
1396                 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1397                 break;
1398             }
1399             return result;
1400           }
1401         }
1402       }
1403     }
1404   }
1405   return NULL;
1406 }
1407 
1408 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1409   Node* region = phi->in(0);
1410   if (region == NULL) {
1411     return false; // Wait stable graph
1412   }
1413   uint cnt = phi->req();
1414   for (uint i = 1; i < cnt; i++) {
1415     Node* rc = region->in(i);
1416     if (rc == NULL || phase->type(rc) == Type::TOP)
1417       return false; // Wait stable graph
1418     Node* in = phi->in(i);
1419     if (in == NULL || phase->type(in) == Type::TOP)
1420       return false; // Wait stable graph
1421   }
1422   return true;
1423 }
1424 //------------------------------split_through_phi------------------------------
1425 // Split instance or boxed field load through Phi.
1426 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1427   Node* mem     = in(Memory);
1428   Node* address = in(Address);
1429   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1430 
1431   assert((t_oop != NULL) &&
1432          (t_oop->is_known_instance_field() ||
1433           t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1434 
1435   Compile* C = phase->C;
1436   intptr_t ignore = 0;
1437   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1438   bool base_is_phi = (base != NULL) && base->is_Phi();
1439   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1440                            (base != NULL) && (base == address->in(AddPNode::Base)) &&
1441                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1442 
1443   if (!((mem->is_Phi() || base_is_phi) &&
1444         (load_boxed_values || t_oop->is_known_instance_field()))) {
1445     return NULL; // memory is not Phi
1446   }
1447 
1448   if (mem->is_Phi()) {
1449     if (!stable_phi(mem->as_Phi(), phase)) {
1450       return NULL; // Wait stable graph
1451     }
1452     uint cnt = mem->req();
1453     // Check for loop invariant memory.
1454     if (cnt == 3) {
1455       for (uint i = 1; i < cnt; i++) {
1456         Node* in = mem->in(i);
1457         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1458         if (m == mem) {
1459           if (i == 1) {
1460             // if the first edge was a loop, check second edge too.
1461             // If both are replaceable - we are in an infinite loop
1462             Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1463             if (n == mem) {
1464               break;
1465             }
1466           }
1467           set_req(Memory, mem->in(cnt - i));
1468           return this; // made change
1469         }
1470       }
1471     }
1472   }
1473   if (base_is_phi) {
1474     if (!stable_phi(base->as_Phi(), phase)) {
1475       return NULL; // Wait stable graph
1476     }
1477     uint cnt = base->req();
1478     // Check for loop invariant memory.
1479     if (cnt == 3) {
1480       for (uint i = 1; i < cnt; i++) {
1481         if (base->in(i) == base) {
1482           return NULL; // Wait stable graph
1483         }
1484       }
1485     }
1486   }
1487 
1488   // Split through Phi (see original code in loopopts.cpp).
1489   assert(C->have_alias_type(t_oop), "instance should have alias type");
1490 
1491   // Do nothing here if Identity will find a value
1492   // (to avoid infinite chain of value phis generation).
1493   if (!phase->eqv(this, this->Identity(phase))) {
1494     return NULL;
1495   }
1496 
1497   // Select Region to split through.
1498   Node* region;
1499   if (!base_is_phi) {
1500     assert(mem->is_Phi(), "sanity");
1501     region = mem->in(0);
1502     // Skip if the region dominates some control edge of the address.
1503     if (!MemNode::all_controls_dominate(address, region))
1504       return NULL;
1505   } else if (!mem->is_Phi()) {
1506     assert(base_is_phi, "sanity");
1507     region = base->in(0);
1508     // Skip if the region dominates some control edge of the memory.
1509     if (!MemNode::all_controls_dominate(mem, region))
1510       return NULL;
1511   } else if (base->in(0) != mem->in(0)) {
1512     assert(base_is_phi && mem->is_Phi(), "sanity");
1513     if (MemNode::all_controls_dominate(mem, base->in(0))) {
1514       region = base->in(0);
1515     } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1516       region = mem->in(0);
1517     } else {
1518       return NULL; // complex graph
1519     }
1520   } else {
1521     assert(base->in(0) == mem->in(0), "sanity");
1522     region = mem->in(0);
1523   }
1524 
1525   const Type* this_type = this->bottom_type();
1526   int this_index  = C->get_alias_index(t_oop);
1527   int this_offset = t_oop->offset();
1528   int this_iid    = t_oop->instance_id();
1529   if (!t_oop->is_known_instance() && load_boxed_values) {
1530     // Use _idx of address base for boxed values.
1531     this_iid = base->_idx;
1532   }
1533   PhaseIterGVN* igvn = phase->is_IterGVN();
1534   Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1535   for (uint i = 1; i < region->req(); i++) {
1536     Node* x;
1537     Node* the_clone = NULL;
1538     Node* in = region->in(i);
1539     if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1540         in != NULL && in->is_OuterStripMinedLoop()) {
1541       // No node should go in the outer strip mined loop
1542       in = in->in(LoopNode::EntryControl);
1543     }
1544     if (in == NULL || in == C->top()) {
1545       x = C->top();      // Dead path?  Use a dead data op
1546     } else {
1547       x = this->clone();        // Else clone up the data op
1548       the_clone = x;            // Remember for possible deletion.
1549       // Alter data node to use pre-phi inputs
1550       if (this->in(0) == region) {
1551         x->set_req(0, in);
1552       } else {
1553         x->set_req(0, NULL);
1554       }
1555       if (mem->is_Phi() && (mem->in(0) == region)) {
1556         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1557       }
1558       if (address->is_Phi() && address->in(0) == region) {
1559         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1560       }
1561       if (base_is_phi && (base->in(0) == region)) {
1562         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1563         Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1564         x->set_req(Address, adr_x);
1565       }
1566     }
1567     // Check for a 'win' on some paths
1568     const Type *t = x->Value(igvn);
1569 
1570     bool singleton = t->singleton();
1571 
1572     // See comments in PhaseIdealLoop::split_thru_phi().
1573     if (singleton && t == Type::TOP) {
1574       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1575     }
1576 
1577     if (singleton) {
1578       x = igvn->makecon(t);
1579     } else {
1580       // We now call Identity to try to simplify the cloned node.
1581       // Note that some Identity methods call phase->type(this).
1582       // Make sure that the type array is big enough for
1583       // our new node, even though we may throw the node away.
1584       // (This tweaking with igvn only works because x is a new node.)
1585       igvn->set_type(x, t);
1586       // If x is a TypeNode, capture any more-precise type permanently into Node
1587       // otherwise it will be not updated during igvn->transform since
1588       // igvn->type(x) is set to x->Value() already.
1589       x->raise_bottom_type(t);
1590       Node* y = x->Identity(igvn);
1591       if (y != x) {
1592         x = y;
1593       } else {
1594         y = igvn->hash_find_insert(x);
1595         if (y) {
1596           x = y;
1597         } else {
1598           // Else x is a new node we are keeping
1599           // We do not need register_new_node_with_optimizer
1600           // because set_type has already been called.
1601           igvn->_worklist.push(x);
1602         }
1603       }
1604     }
1605     if (x != the_clone && the_clone != NULL) {
1606       igvn->remove_dead_node(the_clone);
1607     }
1608     phi->set_req(i, x);
1609   }
1610   // Record Phi
1611   igvn->register_new_node_with_optimizer(phi);
1612   return phi;
1613 }
1614 
1615 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const {
1616   if (Opcode() == Op_LoadX) {
1617     Node* address = in(MemNode::Address);
1618     AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1619     Node* mem = in(MemNode::Memory);
1620     if (alloc != NULL && mem->is_Proj() &&
1621         mem->in(0) != NULL &&
1622         mem->in(0) == alloc->initialization() &&
1623         alloc->initialization()->proj_out_or_null(0) != NULL) {
1624       return alloc;
1625     }
1626   }
1627   return NULL;
1628 }
1629 
1630 
1631 //------------------------------Ideal------------------------------------------
1632 // If the load is from Field memory and the pointer is non-null, it might be possible to
1633 // zero out the control input.
1634 // If the offset is constant and the base is an object allocation,
1635 // try to hook me up to the exact initializing store.
1636 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1637   Node* p = MemNode::Ideal_common(phase, can_reshape);
1638   if (p)  return (p == NodeSentinel) ? NULL : p;
1639 
1640   Node* ctrl    = in(MemNode::Control);
1641   Node* address = in(MemNode::Address);
1642   bool progress = false;
1643 
1644   bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1645          phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1646 
1647   // Skip up past a SafePoint control.  Cannot do this for Stores because
1648   // pointer stores & cardmarks must stay on the same side of a SafePoint.
1649   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1650       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw  &&
1651       !addr_mark &&
1652       (depends_only_on_test() || has_unknown_control_dependency())) {
1653     ctrl = ctrl->in(0);
1654     set_req(MemNode::Control,ctrl);
1655     progress = true;
1656   }
1657 
1658   intptr_t ignore = 0;
1659   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1660   if (base != NULL
1661       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1662     // Check for useless control edge in some common special cases
1663     if (in(MemNode::Control) != NULL
1664         && can_remove_control()
1665         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1666         && all_controls_dominate(base, phase->C->start())) {
1667       // A method-invariant, non-null address (constant or 'this' argument).
1668       set_req(MemNode::Control, NULL);
1669       progress = true;
1670     }
1671   }
1672 
1673   Node* mem = in(MemNode::Memory);
1674   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1675 
1676   if (can_reshape && (addr_t != NULL)) {
1677     // try to optimize our memory input
1678     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1679     if (opt_mem != mem) {
1680       set_req(MemNode::Memory, opt_mem);
1681       if (phase->type( opt_mem ) == Type::TOP) return NULL;
1682       return this;
1683     }
1684     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1685     if ((t_oop != NULL) &&
1686         (t_oop->is_known_instance_field() ||
1687          t_oop->is_ptr_to_boxed_value())) {
1688       PhaseIterGVN *igvn = phase->is_IterGVN();
1689       if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1690         // Delay this transformation until memory Phi is processed.
1691         phase->is_IterGVN()->_worklist.push(this);
1692         return NULL;
1693       }
1694       // Split instance field load through Phi.
1695       Node* result = split_through_phi(phase);
1696       if (result != NULL) return result;
1697 
1698       if (t_oop->is_ptr_to_boxed_value()) {
1699         Node* result = eliminate_autobox(phase);
1700         if (result != NULL) return result;
1701       }
1702     }
1703   }
1704 
1705   // Is there a dominating load that loads the same value?  Leave
1706   // anything that is not a load of a field/array element (like
1707   // barriers etc.) alone
1708   if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1709     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1710       Node *use = mem->fast_out(i);
1711       if (use != this &&
1712           use->Opcode() == Opcode() &&
1713           use->in(0) != NULL &&
1714           use->in(0) != in(0) &&
1715           use->in(Address) == in(Address)) {
1716         Node* ctl = in(0);
1717         for (int i = 0; i < 10 && ctl != NULL; i++) {
1718           ctl = IfNode::up_one_dom(ctl);
1719           if (ctl == use->in(0)) {
1720             set_req(0, use->in(0));
1721             return this;
1722           }
1723         }
1724       }
1725     }
1726   }
1727 
1728   // Check for prior store with a different base or offset; make Load
1729   // independent.  Skip through any number of them.  Bail out if the stores
1730   // are in an endless dead cycle and report no progress.  This is a key
1731   // transform for Reflection.  However, if after skipping through the Stores
1732   // we can't then fold up against a prior store do NOT do the transform as
1733   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
1734   // array memory alive twice: once for the hoisted Load and again after the
1735   // bypassed Store.  This situation only works if EVERYBODY who does
1736   // anti-dependence work knows how to bypass.  I.e. we need all
1737   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
1738   // the alias index stuff.  So instead, peek through Stores and IFF we can
1739   // fold up, do so.
1740   Node* prev_mem = find_previous_store(phase);
1741   if (prev_mem != NULL) {
1742     Node* value = can_see_arraycopy_value(prev_mem, phase);
1743     if (value != NULL) {
1744       return value;
1745     }
1746   }
1747   // Steps (a), (b):  Walk past independent stores to find an exact match.
1748   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1749     // (c) See if we can fold up on the spot, but don't fold up here.
1750     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1751     // just return a prior value, which is done by Identity calls.
1752     if (can_see_stored_value(prev_mem, phase)) {
1753       // Make ready for step (d):
1754       set_req(MemNode::Memory, prev_mem);
1755       return this;
1756     }
1757   }
1758 
1759   AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1760   if (alloc != NULL && mem->is_Proj() &&
1761       mem->in(0) != NULL &&
1762       mem->in(0) == alloc->initialization() &&
1763       Opcode() == Op_LoadX &&
1764       alloc->initialization()->proj_out_or_null(0) != NULL) {
1765     InitializeNode* init = alloc->initialization();
1766     Node* control = init->proj_out(0);
1767     return alloc->make_ideal_mark(phase, control, mem);
1768   }
1769 
1770   return progress ? this : NULL;
1771 }
1772 
1773 // Helper to recognize certain Klass fields which are invariant across
1774 // some group of array types (e.g., int[] or all T[] where T < Object).
1775 const Type*
1776 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1777                                  ciKlass* klass) const {
1778   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1779     // The field is Klass::_modifier_flags.  Return its (constant) value.
1780     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1781     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1782     return TypeInt::make(klass->modifier_flags());
1783   }
1784   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1785     // The field is Klass::_access_flags.  Return its (constant) value.
1786     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1787     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1788     return TypeInt::make(klass->access_flags());
1789   }
1790   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1791     // The field is Klass::_layout_helper.  Return its constant value if known.
1792     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1793     return TypeInt::make(klass->layout_helper());
1794   }
1795 
1796   // No match.
1797   return NULL;
1798 }
1799 
1800 //------------------------------Value-----------------------------------------
1801 const Type* LoadNode::Value(PhaseGVN* phase) const {
1802   // Either input is TOP ==> the result is TOP
1803   Node* mem = in(MemNode::Memory);
1804   const Type *t1 = phase->type(mem);
1805   if (t1 == Type::TOP)  return Type::TOP;
1806   Node* adr = in(MemNode::Address);
1807   const TypePtr* tp = phase->type(adr)->isa_ptr();
1808   if (tp == NULL || tp->empty())  return Type::TOP;
1809   int off = tp->offset();
1810   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1811   Compile* C = phase->C;
1812 
1813   // Try to guess loaded type from pointer type
1814   if (tp->isa_aryptr()) {
1815     const TypeAryPtr* ary = tp->is_aryptr();
1816     const Type* t = ary->elem();
1817 
1818     // Determine whether the reference is beyond the header or not, by comparing
1819     // the offset against the offset of the start of the array's data.
1820     // Different array types begin at slightly different offsets (12 vs. 16).
1821     // We choose T_BYTE as an example base type that is least restrictive
1822     // as to alignment, which will therefore produce the smallest
1823     // possible base offset.
1824     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1825     const bool off_beyond_header = (off >= min_base_off);
1826 
1827     // Try to constant-fold a stable array element.
1828     if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1829       // Make sure the reference is not into the header and the offset is constant
1830       ciObject* aobj = ary->const_oop();
1831       if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1832         int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1833         const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1834                                                                       stable_dimension,
1835                                                                       memory_type(), is_unsigned());
1836         if (con_type != NULL) {
1837           return con_type;
1838         }
1839       }
1840     }
1841 
1842     // Don't do this for integer types. There is only potential profit if
1843     // the element type t is lower than _type; that is, for int types, if _type is
1844     // more restrictive than t.  This only happens here if one is short and the other
1845     // char (both 16 bits), and in those cases we've made an intentional decision
1846     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1847     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1848     //
1849     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1850     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
1851     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1852     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
1853     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1854     // In fact, that could have been the original type of p1, and p1 could have
1855     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1856     // expression (LShiftL quux 3) independently optimized to the constant 8.
1857     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1858         && (_type->isa_vect() == NULL)
1859         && t->isa_inlinetype() == NULL
1860         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1861       // t might actually be lower than _type, if _type is a unique
1862       // concrete subclass of abstract class t.
1863       if (off_beyond_header || off == Type::OffsetBot) {  // is the offset beyond the header?
1864         const Type* jt = t->join_speculative(_type);
1865         // In any case, do not allow the join, per se, to empty out the type.
1866         if (jt->empty() && !t->empty()) {
1867           // This can happen if a interface-typed array narrows to a class type.
1868           jt = _type;
1869         }
1870 #ifdef ASSERT
1871         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1872           // The pointers in the autobox arrays are always non-null
1873           Node* base = adr->in(AddPNode::Base);
1874           if ((base != NULL) && base->is_DecodeN()) {
1875             // Get LoadN node which loads IntegerCache.cache field
1876             base = base->in(1);
1877           }
1878           if ((base != NULL) && base->is_Con()) {
1879             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1880             if ((base_type != NULL) && base_type->is_autobox_cache()) {
1881               // It could be narrow oop
1882               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1883             }
1884           }
1885         }
1886 #endif
1887         return jt;
1888       }
1889     }
1890   } else if (tp->base() == Type::InstPtr) {
1891     assert( off != Type::OffsetBot ||
1892             // arrays can be cast to Objects
1893             tp->is_oopptr()->klass()->is_java_lang_Object() ||
1894             tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() ||
1895             // unsafe field access may not have a constant offset
1896             C->has_unsafe_access(),
1897             "Field accesses must be precise" );
1898     // For oop loads, we expect the _type to be precise.
1899 
1900     const TypeInstPtr* tinst = tp->is_instptr();
1901     BasicType bt = memory_type();
1902 
1903     // Optimize loads from constant fields.
1904     ciObject* const_oop = tinst->const_oop();
1905     if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1906       ciType* mirror_type = const_oop->as_instance()->java_mirror_type();
1907       if (mirror_type != NULL && mirror_type->is_inlinetype()) {
1908         ciInlineKlass* vk = mirror_type->as_inline_klass();
1909         if (off == vk->default_value_offset()) {
1910           // Loading a special hidden field that contains the oop of the default inline type
1911           const Type* const_oop = TypeInstPtr::make(vk->default_instance());
1912           return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
1913         }
1914       }
1915       const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
1916       if (con_type != NULL) {
1917         return con_type;
1918       }
1919     }
1920   } else if (tp->base() == Type::KlassPtr) {
1921     assert( off != Type::OffsetBot ||
1922             // arrays can be cast to Objects
1923             tp->is_klassptr()->klass() == NULL ||
1924             tp->is_klassptr()->klass()->is_java_lang_Object() ||
1925             // also allow array-loading from the primary supertype
1926             // array during subtype checks
1927             Opcode() == Op_LoadKlass,
1928             "Field accesses must be precise" );
1929     // For klass/static loads, we expect the _type to be precise
1930   } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) {
1931     if (adr->is_Load() && off == 0) {
1932       /* With mirrors being an indirect in the Klass*
1933        * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1934        * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1935        *
1936        * So check the type and klass of the node before the LoadP.
1937        */
1938       Node* adr2 = adr->in(MemNode::Address);
1939       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1940       if (tkls != NULL) {
1941         ciKlass* klass = tkls->klass();
1942         if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1943           assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1944           assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1945           return TypeInstPtr::make(klass->java_mirror());
1946         }
1947       }
1948     } else {
1949       // Check for a load of the default value offset from the InlineKlassFixedBlock:
1950       // LoadI(LoadP(inline_klass, adr_inlineklass_fixed_block_offset), default_value_offset_offset)
1951       intptr_t offset = 0;
1952       Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1953       if (base != NULL && base->is_Load() && offset == in_bytes(InlineKlass::default_value_offset_offset())) {
1954         const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr();
1955         if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_inlinetype() &&
1956             tkls->offset() == in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset())) {
1957           assert(base->Opcode() == Op_LoadP, "must load an oop from klass");
1958           assert(Opcode() == Op_LoadI, "must load an int from fixed block");
1959           return TypeInt::make(tkls->klass()->as_inline_klass()->default_value_offset());
1960         }
1961       }
1962     }
1963   }
1964 
1965   const TypeKlassPtr *tkls = tp->isa_klassptr();
1966   if (tkls != NULL && !StressReflectiveCode) {
1967     ciKlass* klass = tkls->klass();
1968     if (tkls->is_loaded() && tkls->klass_is_exact()) {
1969       // We are loading a field from a Klass metaobject whose identity
1970       // is known at compile time (the type is "exact" or "precise").
1971       // Check for fields we know are maintained as constants by the VM.
1972       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1973         // The field is Klass::_super_check_offset.  Return its (constant) value.
1974         // (Folds up type checking code.)
1975         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1976         return TypeInt::make(klass->super_check_offset());
1977       }
1978       // Compute index into primary_supers array
1979       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1980       // Check for overflowing; use unsigned compare to handle the negative case.
1981       if( depth < ciKlass::primary_super_limit() ) {
1982         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1983         // (Folds up type checking code.)
1984         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1985         ciKlass *ss = klass->super_of_depth(depth);
1986         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1987       }
1988       const Type* aift = load_array_final_field(tkls, klass);
1989       if (aift != NULL)  return aift;
1990     }
1991 
1992     // We can still check if we are loading from the primary_supers array at a
1993     // shallow enough depth.  Even though the klass is not exact, entries less
1994     // than or equal to its super depth are correct.
1995     if (tkls->is_loaded()) {
1996       ciType *inner = klass;
1997       while( inner->is_obj_array_klass() )
1998         inner = inner->as_obj_array_klass()->base_element_type();
1999       if( inner->is_instance_klass() &&
2000           !inner->as_instance_klass()->flags().is_interface() ) {
2001         // Compute index into primary_supers array
2002         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2003         // Check for overflowing; use unsigned compare to handle the negative case.
2004         if( depth < ciKlass::primary_super_limit() &&
2005             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
2006           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
2007           // (Folds up type checking code.)
2008           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2009           ciKlass *ss = klass->super_of_depth(depth);
2010           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2011         }
2012       }
2013     }
2014 
2015     // If the type is enough to determine that the thing is not an array,
2016     // we can give the layout_helper a positive interval type.
2017     // This will help short-circuit some reflective code.
2018     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
2019         && !klass->is_array_klass() // not directly typed as an array
2020         && !klass->is_interface()  // specifically not Serializable & Cloneable
2021         && !klass->is_java_lang_Object()   // not the supertype of all T[]
2022         ) {
2023       // Note:  When interfaces are reliable, we can narrow the interface
2024       // test to (klass != Serializable && klass != Cloneable).
2025       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2026       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2027       // The key property of this type is that it folds up tests
2028       // for array-ness, since it proves that the layout_helper is positive.
2029       // Thus, a generic value like the basic object layout helper works fine.
2030       return TypeInt::make(min_size, max_jint, Type::WidenMin);
2031     }
2032   }
2033 
2034   // If we are loading from a freshly-allocated object, produce a zero,
2035   // if the load is provably beyond the header of the object.
2036   // (Also allow a variable load from a fresh array to produce zero.)
2037   const TypeOopPtr *tinst = tp->isa_oopptr();
2038   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
2039   bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
2040   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2041     Node* value = can_see_stored_value(mem,phase);
2042     if (value != NULL && value->is_Con()) {
2043       assert(value->bottom_type()->higher_equal(_type),"sanity");
2044       return value->bottom_type();
2045     }
2046   }
2047 
2048   if (is_instance) {
2049     // If we have an instance type and our memory input is the
2050     // programs's initial memory state, there is no matching store,
2051     // so just return a zero of the appropriate type
2052     Node *mem = in(MemNode::Memory);
2053     if (mem->is_Parm() && mem->in(0)->is_Start()) {
2054       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2055       return Type::get_zero_type(_type->basic_type());
2056     }
2057   }
2058 
2059   Node* alloc = is_new_object_mark_load(phase);
2060   if (alloc != NULL && !(alloc->Opcode() == Op_Allocate && UseBiasedLocking)) {
2061     return TypeX::make(markWord::prototype().value());
2062   }
2063 
2064   return _type;
2065 }
2066 
2067 //------------------------------match_edge-------------------------------------
2068 // Do we Match on this edge index or not?  Match only the address.
2069 uint LoadNode::match_edge(uint idx) const {
2070   return idx == MemNode::Address;
2071 }
2072 
2073 //--------------------------LoadBNode::Ideal--------------------------------------
2074 //
2075 //  If the previous store is to the same address as this load,
2076 //  and the value stored was larger than a byte, replace this load
2077 //  with the value stored truncated to a byte.  If no truncation is
2078 //  needed, the replacement is done in LoadNode::Identity().
2079 //
2080 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2081   Node* mem = in(MemNode::Memory);
2082   Node* value = can_see_stored_value(mem,phase);
2083   if( value && !phase->type(value)->higher_equal( _type ) ) {
2084     Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
2085     return new RShiftINode(result, phase->intcon(24));
2086   }
2087   // Identity call will handle the case where truncation is not needed.
2088   return LoadNode::Ideal(phase, can_reshape);
2089 }
2090 
2091 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2092   Node* mem = in(MemNode::Memory);
2093   Node* value = can_see_stored_value(mem,phase);
2094   if (value != NULL && value->is_Con() &&
2095       !value->bottom_type()->higher_equal(_type)) {
2096     // If the input to the store does not fit with the load's result type,
2097     // it must be truncated. We can't delay until Ideal call since
2098     // a singleton Value is needed for split_thru_phi optimization.
2099     int con = value->get_int();
2100     return TypeInt::make((con << 24) >> 24);
2101   }
2102   return LoadNode::Value(phase);
2103 }
2104 
2105 //--------------------------LoadUBNode::Ideal-------------------------------------
2106 //
2107 //  If the previous store is to the same address as this load,
2108 //  and the value stored was larger than a byte, replace this load
2109 //  with the value stored truncated to a byte.  If no truncation is
2110 //  needed, the replacement is done in LoadNode::Identity().
2111 //
2112 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2113   Node* mem = in(MemNode::Memory);
2114   Node* value = can_see_stored_value(mem, phase);
2115   if (value && !phase->type(value)->higher_equal(_type))
2116     return new AndINode(value, phase->intcon(0xFF));
2117   // Identity call will handle the case where truncation is not needed.
2118   return LoadNode::Ideal(phase, can_reshape);
2119 }
2120 
2121 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2122   Node* mem = in(MemNode::Memory);
2123   Node* value = can_see_stored_value(mem,phase);
2124   if (value != NULL && value->is_Con() &&
2125       !value->bottom_type()->higher_equal(_type)) {
2126     // If the input to the store does not fit with the load's result type,
2127     // it must be truncated. We can't delay until Ideal call since
2128     // a singleton Value is needed for split_thru_phi optimization.
2129     int con = value->get_int();
2130     return TypeInt::make(con & 0xFF);
2131   }
2132   return LoadNode::Value(phase);
2133 }
2134 
2135 //--------------------------LoadUSNode::Ideal-------------------------------------
2136 //
2137 //  If the previous store is to the same address as this load,
2138 //  and the value stored was larger than a char, replace this load
2139 //  with the value stored truncated to a char.  If no truncation is
2140 //  needed, the replacement is done in LoadNode::Identity().
2141 //
2142 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2143   Node* mem = in(MemNode::Memory);
2144   Node* value = can_see_stored_value(mem,phase);
2145   if( value && !phase->type(value)->higher_equal( _type ) )
2146     return new AndINode(value,phase->intcon(0xFFFF));
2147   // Identity call will handle the case where truncation is not needed.
2148   return LoadNode::Ideal(phase, can_reshape);
2149 }
2150 
2151 const Type* LoadUSNode::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 & 0xFFFF);
2161   }
2162   return LoadNode::Value(phase);
2163 }
2164 
2165 //--------------------------LoadSNode::Ideal--------------------------------------
2166 //
2167 //  If the previous store is to the same address as this load,
2168 //  and the value stored was larger than a short, replace this load
2169 //  with the value stored truncated to a short.  If no truncation is
2170 //  needed, the replacement is done in LoadNode::Identity().
2171 //
2172 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2173   Node* mem = in(MemNode::Memory);
2174   Node* value = can_see_stored_value(mem,phase);
2175   if( value && !phase->type(value)->higher_equal( _type ) ) {
2176     Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2177     return new RShiftINode(result, phase->intcon(16));
2178   }
2179   // Identity call will handle the case where truncation is not needed.
2180   return LoadNode::Ideal(phase, can_reshape);
2181 }
2182 
2183 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2184   Node* mem = in(MemNode::Memory);
2185   Node* value = can_see_stored_value(mem,phase);
2186   if (value != NULL && value->is_Con() &&
2187       !value->bottom_type()->higher_equal(_type)) {
2188     // If the input to the store does not fit with the load's result type,
2189     // it must be truncated. We can't delay until Ideal call since
2190     // a singleton Value is needed for split_thru_phi optimization.
2191     int con = value->get_int();
2192     return TypeInt::make((con << 16) >> 16);
2193   }
2194   return LoadNode::Value(phase);
2195 }
2196 
2197 //=============================================================================
2198 //----------------------------LoadKlassNode::make------------------------------
2199 // Polymorphic factory method:
2200 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at,
2201                           const TypeKlassPtr* tk) {
2202   // sanity check the alias category against the created node type
2203   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2204   assert(adr_type != NULL, "expecting TypeKlassPtr");
2205 #ifdef _LP64
2206   if (adr_type->is_ptr_to_narrowklass()) {
2207     assert(UseCompressedClassPointers, "no compressed klasses");
2208     Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2209     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2210   }
2211 #endif
2212   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2213   return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2214 }
2215 
2216 //------------------------------Value------------------------------------------
2217 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2218   return klass_value_common(phase);
2219 }
2220 
2221 // In most cases, LoadKlassNode does not have the control input set. If the control
2222 // input is set, it must not be removed (by LoadNode::Ideal()).
2223 bool LoadKlassNode::can_remove_control() const {
2224   return false;
2225 }
2226 
2227 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2228   // Either input is TOP ==> the result is TOP
2229   const Type *t1 = phase->type( in(MemNode::Memory) );
2230   if (t1 == Type::TOP)  return Type::TOP;
2231   Node *adr = in(MemNode::Address);
2232   const Type *t2 = phase->type( adr );
2233   if (t2 == Type::TOP)  return Type::TOP;
2234   const TypePtr *tp = t2->is_ptr();
2235   if (TypePtr::above_centerline(tp->ptr()) ||
2236       tp->ptr() == TypePtr::Null)  return Type::TOP;
2237 
2238   // Return a more precise klass, if possible
2239   const TypeInstPtr *tinst = tp->isa_instptr();
2240   if (tinst != NULL) {
2241     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2242     int offset = tinst->offset();
2243     if (ik == phase->C->env()->Class_klass()
2244         && (offset == java_lang_Class::klass_offset() ||
2245             offset == java_lang_Class::array_klass_offset())) {
2246       // We are loading a special hidden field from a Class mirror object,
2247       // the field which points to the VM's Klass metaobject.
2248       ciType* t = tinst->java_mirror_type();
2249       // java_mirror_type returns non-null for compile-time Class constants.
2250       if (t != NULL) {
2251         // constant oop => constant klass
2252         if (offset == java_lang_Class::array_klass_offset()) {
2253           if (t->is_void()) {
2254             // We cannot create a void array.  Since void is a primitive type return null
2255             // klass.  Users of this result need to do a null check on the returned klass.
2256             return TypePtr::NULL_PTR;
2257           }
2258           return TypeKlassPtr::make(ciArrayKlass::make(t));
2259         }
2260         if (!t->is_klass()) {
2261           // a primitive Class (e.g., int.class) has NULL for a klass field
2262           return TypePtr::NULL_PTR;
2263         }
2264         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2265         return TypeKlassPtr::make(t->as_klass());
2266       }
2267       // non-constant mirror, so we can't tell what's going on
2268     }
2269     if( !ik->is_loaded() )
2270       return _type;             // Bail out if not loaded
2271     if (offset == oopDesc::klass_offset_in_bytes()) {
2272       if (tinst->klass_is_exact()) {
2273         return TypeKlassPtr::make(ik);
2274       }
2275       // See if we can become precise: no subklasses and no interface
2276       // (Note:  We need to support verified interfaces.)
2277       if (!ik->is_interface() && !ik->has_subklass()) {
2278         // Add a dependence; if any subclass added we need to recompile
2279         if (!ik->is_final()) {
2280           // %%% should use stronger assert_unique_concrete_subtype instead
2281           phase->C->dependencies()->assert_leaf_type(ik);
2282         }
2283         // Return precise klass
2284         return TypeKlassPtr::make(ik);
2285       }
2286 
2287       // Return root of possible klass
2288       return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0), tinst->flatten_array());
2289     }
2290   }
2291 
2292   // Check for loading klass from an array
2293   const TypeAryPtr *tary = tp->isa_aryptr();
2294   if (tary != NULL) {
2295     ciKlass *tary_klass = tary->klass();
2296     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
2297         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2298       if (tary->klass_is_exact()) {
2299         return TypeKlassPtr::make(tary_klass);
2300       }
2301       ciArrayKlass* ak = tary_klass->as_array_klass();
2302       // If the klass is an object array, we defer the question to the
2303       // array component klass.
2304       if (ak->is_obj_array_klass()) {
2305         assert(ak->is_loaded(), "");
2306         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2307         if (base_k->is_loaded() && base_k->is_instance_klass()) {
2308           ciInstanceKlass *ik = base_k->as_instance_klass();
2309           // See if we can become precise: no subklasses and no interface
2310           if (!ik->is_interface() && !ik->has_subklass()) {
2311             // Add a dependence; if any subclass added we need to recompile
2312             if (!ik->is_final()) {
2313               phase->C->dependencies()->assert_leaf_type(ik);
2314             }
2315             // Return precise array klass
2316             return TypeKlassPtr::make(ak);
2317           }
2318         }
2319         return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0));
2320       } else if (ak->is_type_array_klass()) {
2321         return TypeKlassPtr::make(ak); // These are always precise
2322       }
2323     }
2324   }
2325 
2326   // Check for loading klass from an array klass
2327   const TypeKlassPtr *tkls = tp->isa_klassptr();
2328   if (tkls != NULL && !StressReflectiveCode) {
2329     if (!tkls->is_loaded()) {
2330       return _type;             // Bail out if not loaded
2331     }
2332     ciKlass* klass = tkls->klass();
2333     if( klass->is_obj_array_klass() &&
2334         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2335       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2336       // // Always returning precise element type is incorrect,
2337       // // e.g., element type could be object and array may contain strings
2338       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2339 
2340       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2341       // according to the element type's subclassing.
2342       return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0));
2343     } else if (klass->is_flat_array_klass() &&
2344                tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2345       ciKlass* elem = klass->as_flat_array_klass()->element_klass();
2346       return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0), /* flatten_array= */ true);
2347     }
2348     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2349         tkls->offset() == in_bytes(Klass::super_offset())) {
2350       ciKlass* sup = klass->as_instance_klass()->super();
2351       // The field is Klass::_super.  Return its (constant) value.
2352       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2353       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2354     }
2355   }
2356 
2357   // Bailout case
2358   return LoadNode::Value(phase);
2359 }
2360 
2361 //------------------------------Identity---------------------------------------
2362 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2363 // Also feed through the klass in Allocate(...klass...)._klass.
2364 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2365   return klass_identity_common(phase);
2366 }
2367 
2368 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2369   Node* x = LoadNode::Identity(phase);
2370   if (x != this)  return x;
2371 
2372   // Take apart the address into an oop and and offset.
2373   // Return 'this' if we cannot.
2374   Node*    adr    = in(MemNode::Address);
2375   intptr_t offset = 0;
2376   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2377   if (base == NULL)     return this;
2378   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2379   if (toop == NULL)     return this;
2380 
2381   // Step over potential GC barrier for OopHandle resolve
2382   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2383   if (bs->is_gc_barrier_node(base)) {
2384     base = bs->step_over_gc_barrier(base);
2385   }
2386 
2387   // We can fetch the klass directly through an AllocateNode.
2388   // This works even if the klass is not constant (clone or newArray).
2389   if (offset == oopDesc::klass_offset_in_bytes()) {
2390     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2391     if (allocated_klass != NULL) {
2392       return allocated_klass;
2393     }
2394   }
2395 
2396   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2397   // See inline_native_Class_query for occurrences of these patterns.
2398   // Java Example:  x.getClass().isAssignableFrom(y)
2399   //
2400   // This improves reflective code, often making the Class
2401   // mirror go completely dead.  (Current exception:  Class
2402   // mirrors may appear in debug info, but we could clean them out by
2403   // introducing a new debug info operator for Klass.java_mirror).
2404 
2405   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2406       && offset == java_lang_Class::klass_offset()) {
2407     if (base->is_Load()) {
2408       Node* base2 = base->in(MemNode::Address);
2409       if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2410         Node* adr2 = base2->in(MemNode::Address);
2411         const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2412         if (tkls != NULL && !tkls->empty()
2413             && (tkls->klass()->is_instance_klass() ||
2414               tkls->klass()->is_array_klass())
2415             && adr2->is_AddP()
2416            ) {
2417           int mirror_field = in_bytes(Klass::java_mirror_offset());
2418           if (tkls->offset() == mirror_field) {
2419             return adr2->in(AddPNode::Base);
2420           }
2421         }
2422       }
2423     }
2424   }
2425 
2426   return this;
2427 }
2428 
2429 
2430 //------------------------------Value------------------------------------------
2431 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2432   const Type *t = klass_value_common(phase);
2433   if (t == Type::TOP)
2434     return t;
2435 
2436   return t->make_narrowklass();
2437 }
2438 
2439 //------------------------------Identity---------------------------------------
2440 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2441 // Also feed through the klass in Allocate(...klass...)._klass.
2442 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2443   Node *x = klass_identity_common(phase);
2444 
2445   const Type *t = phase->type( x );
2446   if( t == Type::TOP ) return x;
2447   if( t->isa_narrowklass()) return x;
2448   assert (!t->isa_narrowoop(), "no narrow oop here");
2449 
2450   return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2451 }
2452 
2453 //------------------------------Value-----------------------------------------
2454 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2455   // Either input is TOP ==> the result is TOP
2456   const Type *t1 = phase->type( in(MemNode::Memory) );
2457   if( t1 == Type::TOP ) return Type::TOP;
2458   Node *adr = in(MemNode::Address);
2459   const Type *t2 = phase->type( adr );
2460   if( t2 == Type::TOP ) return Type::TOP;
2461   const TypePtr *tp = t2->is_ptr();
2462   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2463   const TypeAryPtr *tap = tp->isa_aryptr();
2464   if( !tap ) return _type;
2465   return tap->size();
2466 }
2467 
2468 //-------------------------------Ideal---------------------------------------
2469 // Feed through the length in AllocateArray(...length...)._length.
2470 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2471   Node* p = MemNode::Ideal_common(phase, can_reshape);
2472   if (p)  return (p == NodeSentinel) ? NULL : p;
2473 
2474   // Take apart the address into an oop and and offset.
2475   // Return 'this' if we cannot.
2476   Node*    adr    = in(MemNode::Address);
2477   intptr_t offset = 0;
2478   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2479   if (base == NULL)     return NULL;
2480   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2481   if (tary == NULL)     return NULL;
2482 
2483   // We can fetch the length directly through an AllocateArrayNode.
2484   // This works even if the length is not constant (clone or newArray).
2485   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2486     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2487     if (alloc != NULL) {
2488       Node* allocated_length = alloc->Ideal_length();
2489       Node* len = alloc->make_ideal_length(tary, phase);
2490       if (allocated_length != len) {
2491         // New CastII improves on this.
2492         return len;
2493       }
2494     }
2495   }
2496 
2497   return NULL;
2498 }
2499 
2500 //------------------------------Identity---------------------------------------
2501 // Feed through the length in AllocateArray(...length...)._length.
2502 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2503   Node* x = LoadINode::Identity(phase);
2504   if (x != this)  return x;
2505 
2506   // Take apart the address into an oop and and offset.
2507   // Return 'this' if we cannot.
2508   Node*    adr    = in(MemNode::Address);
2509   intptr_t offset = 0;
2510   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2511   if (base == NULL)     return this;
2512   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2513   if (tary == NULL)     return this;
2514 
2515   // We can fetch the length directly through an AllocateArrayNode.
2516   // This works even if the length is not constant (clone or newArray).
2517   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2518     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2519     if (alloc != NULL) {
2520       Node* allocated_length = alloc->Ideal_length();
2521       // Do not allow make_ideal_length to allocate a CastII node.
2522       Node* len = alloc->make_ideal_length(tary, phase, false);
2523       if (allocated_length == len) {
2524         // Return allocated_length only if it would not be improved by a CastII.
2525         return allocated_length;
2526       }
2527     }
2528   }
2529 
2530   return this;
2531 
2532 }
2533 
2534 //=============================================================================
2535 //---------------------------StoreNode::make-----------------------------------
2536 // Polymorphic factory method:
2537 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2538   assert((mo == unordered || mo == release), "unexpected");
2539   Compile* C = gvn.C;
2540   assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2541          ctl != NULL, "raw memory operations should have control edge");
2542 
2543   switch (bt) {
2544   case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2545   case T_BYTE:    return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2546   case T_INT:     return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2547   case T_CHAR:
2548   case T_SHORT:   return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2549   case T_LONG:    return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2550   case T_FLOAT:   return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2551   case T_DOUBLE:  return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2552   case T_METADATA:
2553   case T_ADDRESS:
2554   case T_INLINE_TYPE:
2555   case T_OBJECT:
2556 #ifdef _LP64
2557     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2558       val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2559       return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2560     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2561                (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2562                 adr->bottom_type()->isa_rawptr())) {
2563       val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2564       return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2565     }
2566 #endif
2567     {
2568       return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2569     }
2570   default:
2571     ShouldNotReachHere();
2572     return (StoreNode*)NULL;
2573   }
2574 }
2575 
2576 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2577   bool require_atomic = true;
2578   return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2579 }
2580 
2581 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2582   bool require_atomic = true;
2583   return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2584 }
2585 
2586 
2587 //--------------------------bottom_type----------------------------------------
2588 const Type *StoreNode::bottom_type() const {
2589   return Type::MEMORY;
2590 }
2591 
2592 //------------------------------hash-------------------------------------------
2593 uint StoreNode::hash() const {
2594   // unroll addition of interesting fields
2595   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2596 
2597   // Since they are not commoned, do not hash them:
2598   return NO_HASH;
2599 }
2600 
2601 //------------------------------Ideal------------------------------------------
2602 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2603 // When a store immediately follows a relevant allocation/initialization,
2604 // try to capture it into the initialization, or hoist it above.
2605 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2606   Node* p = MemNode::Ideal_common(phase, can_reshape);
2607   if (p)  return (p == NodeSentinel) ? NULL : p;
2608 
2609   Node* mem     = in(MemNode::Memory);
2610   Node* address = in(MemNode::Address);
2611   // Back-to-back stores to same address?  Fold em up.  Generally
2612   // unsafe if I have intervening uses...  Also disallowed for StoreCM
2613   // since they must follow each StoreP operation.  Redundant StoreCMs
2614   // are eliminated just before matching in final_graph_reshape.
2615   if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
2616     Node* st = mem;
2617     // If Store 'st' has more than one use, we cannot fold 'st' away.
2618     // For example, 'st' might be the final state at a conditional
2619     // return.  Or, 'st' might be used by some node which is live at
2620     // the same time 'st' is live, which might be unschedulable.  So,
2621     // require exactly ONE user until such time as we clone 'mem' for
2622     // each of 'mem's uses (thus making the exactly-1-user-rule hold
2623     // true).
2624     while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2625       // Looking at a dead closed cycle of memory?
2626       assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2627       assert(Opcode() == st->Opcode() ||
2628              st->Opcode() == Op_StoreVector ||
2629              Opcode() == Op_StoreVector ||
2630              phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2631              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2632              (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2633              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
2634              (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2635              "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2636 
2637       if (st->in(MemNode::Address)->eqv_uncast(address) &&
2638           st->as_Store()->memory_size() <= this->memory_size()) {
2639         Node* use = st->raw_out(0);
2640         phase->igvn_rehash_node_delayed(use);
2641         if (can_reshape) {
2642           use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2643         } else {
2644           // It's OK to do this in the parser, since DU info is always accurate,
2645           // and the parser always refers to nodes via SafePointNode maps.
2646           use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2647         }
2648         return this;
2649       }
2650       st = st->in(MemNode::Memory);
2651     }
2652   }
2653 
2654 
2655   // Capture an unaliased, unconditional, simple store into an initializer.
2656   // Or, if it is independent of the allocation, hoist it above the allocation.
2657   if (ReduceFieldZeroing && /*can_reshape &&*/
2658       mem->is_Proj() && mem->in(0)->is_Initialize()) {
2659     InitializeNode* init = mem->in(0)->as_Initialize();
2660     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2661     if (offset > 0) {
2662       Node* moved = init->capture_store(this, offset, phase, can_reshape);
2663       // If the InitializeNode captured me, it made a raw copy of me,
2664       // and I need to disappear.
2665       if (moved != NULL) {
2666         // %%% hack to ensure that Ideal returns a new node:
2667         mem = MergeMemNode::make(mem);
2668         return mem;             // fold me away
2669       }
2670     }
2671   }
2672 
2673   return NULL;                  // No further progress
2674 }
2675 
2676 //------------------------------Value-----------------------------------------
2677 const Type* StoreNode::Value(PhaseGVN* phase) const {
2678   // Either input is TOP ==> the result is TOP
2679   const Type *t1 = phase->type( in(MemNode::Memory) );
2680   if( t1 == Type::TOP ) return Type::TOP;
2681   const Type *t2 = phase->type( in(MemNode::Address) );
2682   if( t2 == Type::TOP ) return Type::TOP;
2683   const Type *t3 = phase->type( in(MemNode::ValueIn) );
2684   if( t3 == Type::TOP ) return Type::TOP;
2685   return Type::MEMORY;
2686 }
2687 
2688 //------------------------------Identity---------------------------------------
2689 // Remove redundant stores:
2690 //   Store(m, p, Load(m, p)) changes to m.
2691 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2692 Node* StoreNode::Identity(PhaseGVN* phase) {
2693   Node* mem = in(MemNode::Memory);
2694   Node* adr = in(MemNode::Address);
2695   Node* val = in(MemNode::ValueIn);
2696 
2697   Node* result = this;
2698 
2699   // Load then Store?  Then the Store is useless
2700   if (val->is_Load() &&
2701       val->in(MemNode::Address)->eqv_uncast(adr) &&
2702       val->in(MemNode::Memory )->eqv_uncast(mem) &&
2703       val->as_Load()->store_Opcode() == Opcode()) {
2704     result = mem;
2705   }
2706 
2707   // Two stores in a row of the same value?
2708   if (result == this &&
2709       mem->is_Store() &&
2710       mem->in(MemNode::Address)->eqv_uncast(adr) &&
2711       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2712       mem->Opcode() == Opcode()) {
2713     result = mem;
2714   }
2715 
2716   // Store of zero anywhere into a freshly-allocated object?
2717   // Then the store is useless.
2718   // (It must already have been captured by the InitializeNode.)
2719   if (result == this && ReduceFieldZeroing) {
2720     // a newly allocated object is already all-zeroes everywhere
2721     if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
2722         (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) {
2723       assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to inline type array is forbidden");
2724       result = mem;
2725     }
2726 
2727     if (result == this) {
2728       // the store may also apply to zero-bits in an earlier object
2729       Node* prev_mem = find_previous_store(phase);
2730       // Steps (a), (b):  Walk past independent stores to find an exact match.
2731       if (prev_mem != NULL) {
2732         Node* prev_val = can_see_stored_value(prev_mem, phase);
2733         if (prev_val != NULL && phase->eqv(prev_val, val)) {
2734           // prev_val and val might differ by a cast; it would be good
2735           // to keep the more informative of the two.
2736           if (phase->type(val)->is_zero_type()) {
2737             result = mem;
2738           } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) {
2739             InitializeNode* init = prev_mem->in(0)->as_Initialize();
2740             AllocateNode* alloc = init->allocation();
2741             if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) {
2742               result = mem;
2743             }
2744           }
2745         }
2746       }
2747     }
2748   }
2749 
2750   if (result != this && phase->is_IterGVN() != NULL) {
2751     MemBarNode* trailing = trailing_membar();
2752     if (trailing != NULL) {
2753 #ifdef ASSERT
2754       const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2755       assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2756 #endif
2757       PhaseIterGVN* igvn = phase->is_IterGVN();
2758       trailing->remove(igvn);
2759     }
2760   }
2761 
2762   return result;
2763 }
2764 
2765 //------------------------------match_edge-------------------------------------
2766 // Do we Match on this edge index or not?  Match only memory & value
2767 uint StoreNode::match_edge(uint idx) const {
2768   return idx == MemNode::Address || idx == MemNode::ValueIn;
2769 }
2770 
2771 //------------------------------cmp--------------------------------------------
2772 // Do not common stores up together.  They generally have to be split
2773 // back up anyways, so do not bother.
2774 bool StoreNode::cmp( const Node &n ) const {
2775   return (&n == this);          // Always fail except on self
2776 }
2777 
2778 //------------------------------Ideal_masked_input-----------------------------
2779 // Check for a useless mask before a partial-word store
2780 // (StoreB ... (AndI valIn conIa) )
2781 // If (conIa & mask == mask) this simplifies to
2782 // (StoreB ... (valIn) )
2783 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2784   Node *val = in(MemNode::ValueIn);
2785   if( val->Opcode() == Op_AndI ) {
2786     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2787     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2788       set_req(MemNode::ValueIn, val->in(1));
2789       return this;
2790     }
2791   }
2792   return NULL;
2793 }
2794 
2795 
2796 //------------------------------Ideal_sign_extended_input----------------------
2797 // Check for useless sign-extension before a partial-word store
2798 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2799 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
2800 // (StoreB ... (valIn) )
2801 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2802   Node *val = in(MemNode::ValueIn);
2803   if( val->Opcode() == Op_RShiftI ) {
2804     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2805     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2806       Node *shl = val->in(1);
2807       if( shl->Opcode() == Op_LShiftI ) {
2808         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2809         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2810           set_req(MemNode::ValueIn, shl->in(1));
2811           return this;
2812         }
2813       }
2814     }
2815   }
2816   return NULL;
2817 }
2818 
2819 //------------------------------value_never_loaded-----------------------------------
2820 // Determine whether there are any possible loads of the value stored.
2821 // For simplicity, we actually check if there are any loads from the
2822 // address stored to, not just for loads of the value stored by this node.
2823 //
2824 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2825   Node *adr = in(Address);
2826   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2827   if (adr_oop == NULL)
2828     return false;
2829   if (!adr_oop->is_known_instance_field())
2830     return false; // if not a distinct instance, there may be aliases of the address
2831   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2832     Node *use = adr->fast_out(i);
2833     if (use->is_Load() || use->is_LoadStore()) {
2834       return false;
2835     }
2836   }
2837   return true;
2838 }
2839 
2840 MemBarNode* StoreNode::trailing_membar() const {
2841   if (is_release()) {
2842     MemBarNode* trailing_mb = NULL;
2843     for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2844       Node* u = fast_out(i);
2845       if (u->is_MemBar()) {
2846         if (u->as_MemBar()->trailing_store()) {
2847           assert(u->Opcode() == Op_MemBarVolatile, "");
2848           assert(trailing_mb == NULL, "only one");
2849           trailing_mb = u->as_MemBar();
2850 #ifdef ASSERT
2851           Node* leading = u->as_MemBar()->leading_membar();
2852           assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2853           assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2854           assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2855 #endif
2856         } else {
2857           assert(u->as_MemBar()->standalone(), "");
2858         }
2859       }
2860     }
2861     return trailing_mb;
2862   }
2863   return NULL;
2864 }
2865 
2866 
2867 //=============================================================================
2868 //------------------------------Ideal------------------------------------------
2869 // If the store is from an AND mask that leaves the low bits untouched, then
2870 // we can skip the AND operation.  If the store is from a sign-extension
2871 // (a left shift, then right shift) we can skip both.
2872 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2873   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2874   if( progress != NULL ) return progress;
2875 
2876   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2877   if( progress != NULL ) return progress;
2878 
2879   // Finally check the default case
2880   return StoreNode::Ideal(phase, can_reshape);
2881 }
2882 
2883 //=============================================================================
2884 //------------------------------Ideal------------------------------------------
2885 // If the store is from an AND mask that leaves the low bits untouched, then
2886 // we can skip the AND operation
2887 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2888   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2889   if( progress != NULL ) return progress;
2890 
2891   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2892   if( progress != NULL ) return progress;
2893 
2894   // Finally check the default case
2895   return StoreNode::Ideal(phase, can_reshape);
2896 }
2897 
2898 //=============================================================================
2899 //------------------------------Identity---------------------------------------
2900 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2901   // No need to card mark when storing a null ptr
2902   Node* my_store = in(MemNode::OopStore);
2903   if (my_store->is_Store()) {
2904     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2905     if( t1 == TypePtr::NULL_PTR ) {
2906       return in(MemNode::Memory);
2907     }
2908   }
2909   return this;
2910 }
2911 
2912 //=============================================================================
2913 //------------------------------Ideal---------------------------------------
2914 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2915   Node* progress = StoreNode::Ideal(phase, can_reshape);
2916   if (progress != NULL) return progress;
2917 
2918   Node* my_store = in(MemNode::OopStore);
2919   if (my_store->is_MergeMem()) {
2920     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2921     set_req(MemNode::OopStore, mem);
2922     return this;
2923   }
2924 
2925   return NULL;
2926 }
2927 
2928 //------------------------------Value-----------------------------------------
2929 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2930   // Either input is TOP ==> the result is TOP
2931   const Type *t = phase->type( in(MemNode::Memory) );
2932   if( t == Type::TOP ) return Type::TOP;
2933   t = phase->type( in(MemNode::Address) );
2934   if( t == Type::TOP ) return Type::TOP;
2935   t = phase->type( in(MemNode::ValueIn) );
2936   if( t == Type::TOP ) return Type::TOP;
2937   // If extra input is TOP ==> the result is TOP
2938   t = phase->type( in(MemNode::OopStore) );
2939   if( t == Type::TOP ) return Type::TOP;
2940 
2941   return StoreNode::Value( phase );
2942 }
2943 
2944 
2945 //=============================================================================
2946 //----------------------------------SCMemProjNode------------------------------
2947 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2948 {
2949   return bottom_type();
2950 }
2951 
2952 //=============================================================================
2953 //----------------------------------LoadStoreNode------------------------------
2954 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2955   : Node(required),
2956     _type(rt),
2957     _adr_type(at),
2958     _barrier(0)
2959 {
2960   init_req(MemNode::Control, c  );
2961   init_req(MemNode::Memory , mem);
2962   init_req(MemNode::Address, adr);
2963   init_req(MemNode::ValueIn, val);
2964   init_class_id(Class_LoadStore);
2965 }
2966 
2967 uint LoadStoreNode::ideal_reg() const {
2968   return _type->ideal_reg();
2969 }
2970 
2971 bool LoadStoreNode::result_not_used() const {
2972   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2973     Node *x = fast_out(i);
2974     if (x->Opcode() == Op_SCMemProj) continue;
2975     return false;
2976   }
2977   return true;
2978 }
2979 
2980 MemBarNode* LoadStoreNode::trailing_membar() const {
2981   MemBarNode* trailing = NULL;
2982   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2983     Node* u = fast_out(i);
2984     if (u->is_MemBar()) {
2985       if (u->as_MemBar()->trailing_load_store()) {
2986         assert(u->Opcode() == Op_MemBarAcquire, "");
2987         assert(trailing == NULL, "only one");
2988         trailing = u->as_MemBar();
2989 #ifdef ASSERT
2990         Node* leading = trailing->leading_membar();
2991         assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2992         assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2993         assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2994 #endif
2995       } else {
2996         assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2997       }
2998     }
2999   }
3000 
3001   return trailing;
3002 }
3003 
3004 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3005 
3006 //=============================================================================
3007 //----------------------------------LoadStoreConditionalNode--------------------
3008 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
3009   init_req(ExpectedIn, ex );
3010 }
3011 
3012 //=============================================================================
3013 //-------------------------------adr_type--------------------------------------
3014 const TypePtr* ClearArrayNode::adr_type() const {
3015   Node *adr = in(3);
3016   if (adr == NULL)  return NULL; // node is dead
3017   return MemNode::calculate_adr_type(adr->bottom_type());
3018 }
3019 
3020 //------------------------------match_edge-------------------------------------
3021 // Do we Match on this edge index or not?  Do not match memory
3022 uint ClearArrayNode::match_edge(uint idx) const {
3023   return idx > 1;
3024 }
3025 
3026 //------------------------------Identity---------------------------------------
3027 // Clearing a zero length array does nothing
3028 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3029   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
3030 }
3031 
3032 //------------------------------Idealize---------------------------------------
3033 // Clearing a short array is faster with stores
3034 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3035   // Already know this is a large node, do not try to ideal it
3036   if (!IdealizeClearArrayNode || _is_large) return NULL;
3037 
3038   const int unit = BytesPerLong;
3039   const TypeX* t = phase->type(in(2))->isa_intptr_t();
3040   if (!t)  return NULL;
3041   if (!t->is_con())  return NULL;
3042   intptr_t raw_count = t->get_con();
3043   intptr_t size = raw_count;
3044   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3045   // Clearing nothing uses the Identity call.
3046   // Negative clears are possible on dead ClearArrays
3047   // (see jck test stmt114.stmt11402.val).
3048   if (size <= 0 || size % unit != 0)  return NULL;
3049   intptr_t count = size / unit;
3050   // Length too long; communicate this to matchers and assemblers.
3051   // Assemblers are responsible to produce fast hardware clears for it.
3052   if (size > InitArrayShortSize) {
3053     return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
3054   }
3055   Node *mem = in(1);
3056   if( phase->type(mem)==Type::TOP ) return NULL;
3057   Node *adr = in(3);
3058   const Type* at = phase->type(adr);
3059   if( at==Type::TOP ) return NULL;
3060   const TypePtr* atp = at->isa_ptr();
3061   // adjust atp to be the correct array element address type
3062   if (atp == NULL)  atp = TypePtr::BOTTOM;
3063   else              atp = atp->add_offset(Type::OffsetBot);
3064   // Get base for derived pointer purposes
3065   if( adr->Opcode() != Op_AddP ) Unimplemented();
3066   Node *base = adr->in(1);
3067 
3068   Node *val = in(4);
3069   Node *off  = phase->MakeConX(BytesPerLong);
3070   mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3071   count--;
3072   while( count-- ) {
3073     mem = phase->transform(mem);
3074     adr = phase->transform(new AddPNode(base,adr,off));
3075     mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3076   }
3077   return mem;
3078 }
3079 
3080 //----------------------------step_through----------------------------------
3081 // Return allocation input memory edge if it is different instance
3082 // or itself if it is the one we are looking for.
3083 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3084   Node* n = *np;
3085   assert(n->is_ClearArray(), "sanity");
3086   intptr_t offset;
3087   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3088   // This method is called only before Allocate nodes are expanded
3089   // during macro nodes expansion. Before that ClearArray nodes are
3090   // only generated in PhaseMacroExpand::generate_arraycopy() (before
3091   // Allocate nodes are expanded) which follows allocations.
3092   assert(alloc != NULL, "should have allocation");
3093   if (alloc->_idx == instance_id) {
3094     // Can not bypass initialization of the instance we are looking for.
3095     return false;
3096   }
3097   // Otherwise skip it.
3098   InitializeNode* init = alloc->initialization();
3099   if (init != NULL)
3100     *np = init->in(TypeFunc::Memory);
3101   else
3102     *np = alloc->in(TypeFunc::Memory);
3103   return true;
3104 }
3105 
3106 //----------------------------clear_memory-------------------------------------
3107 // Generate code to initialize object storage to zero.
3108 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3109                                    Node* val,
3110                                    Node* raw_val,
3111                                    intptr_t start_offset,
3112                                    Node* end_offset,
3113                                    PhaseGVN* phase) {
3114   intptr_t offset = start_offset;
3115 
3116   int unit = BytesPerLong;
3117   if ((offset % unit) != 0) {
3118     Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3119     adr = phase->transform(adr);
3120     const TypePtr* atp = TypeRawPtr::BOTTOM;
3121     if (val != NULL) {
3122       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3123       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3124     } else {
3125       assert(raw_val == NULL, "val may not be null");
3126       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3127     }
3128     mem = phase->transform(mem);
3129     offset += BytesPerInt;
3130   }
3131   assert((offset % unit) == 0, "");
3132 
3133   // Initialize the remaining stuff, if any, with a ClearArray.
3134   return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
3135 }
3136 
3137 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3138                                    Node* raw_val,
3139                                    Node* start_offset,
3140                                    Node* end_offset,
3141                                    PhaseGVN* phase) {
3142   if (start_offset == end_offset) {
3143     // nothing to do
3144     return mem;
3145   }
3146 
3147   int unit = BytesPerLong;
3148   Node* zbase = start_offset;
3149   Node* zend  = end_offset;
3150 
3151   // Scale to the unit required by the CPU:
3152   if (!Matcher::init_array_count_is_in_bytes) {
3153     Node* shift = phase->intcon(exact_log2(unit));
3154     zbase = phase->transform(new URShiftXNode(zbase, shift) );
3155     zend  = phase->transform(new URShiftXNode(zend,  shift) );
3156   }
3157 
3158   // Bulk clear double-words
3159   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3160   Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3161   if (raw_val == NULL) {
3162     raw_val = phase->MakeConX(0);
3163   }
3164   mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
3165   return phase->transform(mem);
3166 }
3167 
3168 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3169                                    Node* val,
3170                                    Node* raw_val,
3171                                    intptr_t start_offset,
3172                                    intptr_t end_offset,
3173                                    PhaseGVN* phase) {
3174   if (start_offset == end_offset) {
3175     // nothing to do
3176     return mem;
3177   }
3178 
3179   assert((end_offset % BytesPerInt) == 0, "odd end offset");
3180   intptr_t done_offset = end_offset;
3181   if ((done_offset % BytesPerLong) != 0) {
3182     done_offset -= BytesPerInt;
3183   }
3184   if (done_offset > start_offset) {
3185     mem = clear_memory(ctl, mem, dest, val, raw_val,
3186                        start_offset, phase->MakeConX(done_offset), phase);
3187   }
3188   if (done_offset < end_offset) { // emit the final 32-bit store
3189     Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3190     adr = phase->transform(adr);
3191     const TypePtr* atp = TypeRawPtr::BOTTOM;
3192     if (val != NULL) {
3193       assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3194       mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3195     } else {
3196       assert(raw_val == NULL, "val may not be null");
3197       mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3198     }
3199     mem = phase->transform(mem);
3200     done_offset += BytesPerInt;
3201   }
3202   assert(done_offset == end_offset, "");
3203   return mem;
3204 }
3205 
3206 //=============================================================================
3207 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3208   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3209     _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3210 #ifdef ASSERT
3211   , _pair_idx(0)
3212 #endif
3213 {
3214   init_class_id(Class_MemBar);
3215   Node* top = C->top();
3216   init_req(TypeFunc::I_O,top);
3217   init_req(TypeFunc::FramePtr,top);
3218   init_req(TypeFunc::ReturnAdr,top);
3219   if (precedent != NULL)
3220     init_req(TypeFunc::Parms, precedent);
3221 }
3222 
3223 //------------------------------cmp--------------------------------------------
3224 uint MemBarNode::hash() const { return NO_HASH; }
3225 bool MemBarNode::cmp( const Node &n ) const {
3226   return (&n == this);          // Always fail except on self
3227 }
3228 
3229 //------------------------------make-------------------------------------------
3230 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3231   switch (opcode) {
3232   case Op_MemBarAcquire:     return new MemBarAcquireNode(C, atp, pn);
3233   case Op_LoadFence:         return new LoadFenceNode(C, atp, pn);
3234   case Op_MemBarRelease:     return new MemBarReleaseNode(C, atp, pn);
3235   case Op_StoreFence:        return new StoreFenceNode(C, atp, pn);
3236   case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3237   case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3238   case Op_MemBarVolatile:    return new MemBarVolatileNode(C, atp, pn);
3239   case Op_MemBarCPUOrder:    return new MemBarCPUOrderNode(C, atp, pn);
3240   case Op_OnSpinWait:        return new OnSpinWaitNode(C, atp, pn);
3241   case Op_Initialize:        return new InitializeNode(C, atp, pn);
3242   case Op_MemBarStoreStore:  return new MemBarStoreStoreNode(C, atp, pn);
3243   default: ShouldNotReachHere(); return NULL;
3244   }
3245 }
3246 
3247 void MemBarNode::remove(PhaseIterGVN *igvn) {
3248   if (outcnt() != 2) {
3249     return;
3250   }
3251   if (trailing_store() || trailing_load_store()) {
3252     MemBarNode* leading = leading_membar();
3253     if (leading != NULL) {
3254       assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3255       leading->remove(igvn);
3256     }
3257   }
3258   igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3259   igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3260 }
3261 
3262 //------------------------------Ideal------------------------------------------
3263 // Return a node which is more "ideal" than the current node.  Strip out
3264 // control copies
3265 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3266   if (remove_dead_region(phase, can_reshape)) return this;
3267   // Don't bother trying to transform a dead node
3268   if (in(0) && in(0)->is_top()) {
3269     return NULL;
3270   }
3271 
3272   bool progress = false;
3273   // Eliminate volatile MemBars for scalar replaced objects.
3274   if (can_reshape && req() == (Precedent+1)) {
3275     bool eliminate = false;
3276     int opc = Opcode();
3277     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3278       // Volatile field loads and stores.
3279       Node* my_mem = in(MemBarNode::Precedent);
3280       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3281       if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3282         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3283         // replace this Precedent (decodeN) with the Load instead.
3284         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
3285           Node* load_node = my_mem->in(1);
3286           set_req(MemBarNode::Precedent, load_node);
3287           phase->is_IterGVN()->_worklist.push(my_mem);
3288           my_mem = load_node;
3289         } else {
3290           assert(my_mem->unique_out() == this, "sanity");
3291           del_req(Precedent);
3292           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3293           my_mem = NULL;
3294         }
3295         progress = true;
3296       }
3297       if (my_mem != NULL && my_mem->is_Mem()) {
3298         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3299         // Check for scalar replaced object reference.
3300         if( t_oop != NULL && t_oop->is_known_instance_field() &&
3301             t_oop->offset() != Type::OffsetBot &&
3302             t_oop->offset() != Type::OffsetTop) {
3303           eliminate = true;
3304         }
3305       }
3306     } else if (opc == Op_MemBarRelease) {
3307       // Final field stores.
3308       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3309       if ((alloc != NULL) && alloc->is_Allocate() &&
3310           alloc->as_Allocate()->does_not_escape_thread()) {
3311         // The allocated object does not escape.
3312         eliminate = true;
3313       }
3314     }
3315     if (eliminate) {
3316       // Replace MemBar projections by its inputs.
3317       PhaseIterGVN* igvn = phase->is_IterGVN();
3318       remove(igvn);
3319       // Must return either the original node (now dead) or a new node
3320       // (Do not return a top here, since that would break the uniqueness of top.)
3321       return new ConINode(TypeInt::ZERO);
3322     }
3323   }
3324   return progress ? this : NULL;
3325 }
3326 
3327 //------------------------------Value------------------------------------------
3328 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3329   if( !in(0) ) return Type::TOP;
3330   if( phase->type(in(0)) == Type::TOP )
3331     return Type::TOP;
3332   return TypeTuple::MEMBAR;
3333 }
3334 
3335 //------------------------------match------------------------------------------
3336 // Construct projections for memory.
3337 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
3338   switch (proj->_con) {
3339   case TypeFunc::Control:
3340   case TypeFunc::Memory:
3341     return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3342   }
3343   ShouldNotReachHere();
3344   return NULL;
3345 }
3346 
3347 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3348   trailing->_kind = TrailingStore;
3349   leading->_kind = LeadingStore;
3350 #ifdef ASSERT
3351   trailing->_pair_idx = leading->_idx;
3352   leading->_pair_idx = leading->_idx;
3353 #endif
3354 }
3355 
3356 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3357   trailing->_kind = TrailingLoadStore;
3358   leading->_kind = LeadingLoadStore;
3359 #ifdef ASSERT
3360   trailing->_pair_idx = leading->_idx;
3361   leading->_pair_idx = leading->_idx;
3362 #endif
3363 }
3364 
3365 MemBarNode* MemBarNode::trailing_membar() const {
3366   ResourceMark rm;
3367   Node* trailing = (Node*)this;
3368   VectorSet seen;
3369   Node_Stack multis(0);
3370   do {
3371     Node* c = trailing;
3372     uint i = 0;
3373     do {
3374       trailing = NULL;
3375       for (; i < c->outcnt(); i++) {
3376         Node* next = c->raw_out(i);
3377         if (next != c && next->is_CFG()) {
3378           if (c->is_MultiBranch()) {
3379             if (multis.node() == c) {
3380               multis.set_index(i+1);
3381             } else {
3382               multis.push(c, i+1);
3383             }
3384           }
3385           trailing = next;
3386           break;
3387         }
3388       }
3389       if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3390         break;
3391       }
3392       while (multis.size() > 0) {
3393         c = multis.node();
3394         i = multis.index();
3395         if (i < c->req()) {
3396           break;
3397         }
3398         multis.pop();
3399       }
3400     } while (multis.size() > 0);
3401   } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3402 
3403   MemBarNode* mb = trailing->as_MemBar();
3404   assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3405          (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3406   assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3407   return mb;
3408 }
3409 
3410 MemBarNode* MemBarNode::leading_membar() const {
3411   ResourceMark rm;
3412   VectorSet seen;
3413   Node_Stack regions(0);
3414   Node* leading = in(0);
3415   while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3416     while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3417       leading = NULL;
3418       while (regions.size() > 0 && leading == NULL) {
3419         Node* r = regions.node();
3420         uint i = regions.index();
3421         if (i < r->req()) {
3422           leading = r->in(i);
3423           regions.set_index(i+1);
3424         } else {
3425           regions.pop();
3426         }
3427       }
3428       if (leading == NULL) {
3429         assert(regions.size() == 0, "all paths should have been tried");
3430         return NULL;
3431       }
3432     }
3433     if (leading->is_Region()) {
3434       regions.push(leading, 2);
3435       leading = leading->in(1);
3436     } else {
3437       leading = leading->in(0);
3438     }
3439   }
3440 #ifdef ASSERT
3441   Unique_Node_List wq;
3442   wq.push((Node*)this);
3443   uint found = 0;
3444   for (uint i = 0; i < wq.size(); i++) {
3445     Node* n = wq.at(i);
3446     if (n->is_Region()) {
3447       for (uint j = 1; j < n->req(); j++) {
3448         Node* in = n->in(j);
3449         if (in != NULL && !in->is_top()) {
3450           wq.push(in);
3451         }
3452       }
3453     } else {
3454       if (n->is_MemBar() && n->as_MemBar()->leading()) {
3455         assert(n == leading, "consistency check failed");
3456         found++;
3457       } else {
3458         Node* in = n->in(0);
3459         if (in != NULL && !in->is_top()) {
3460           wq.push(in);
3461         }
3462       }
3463     }
3464   }
3465   assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3466 #endif
3467   if (leading == NULL) {
3468     return NULL;
3469   }
3470   MemBarNode* mb = leading->as_MemBar();
3471   assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3472          (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3473   assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3474   return mb;
3475 }
3476 
3477 //===========================InitializeNode====================================
3478 // SUMMARY:
3479 // This node acts as a memory barrier on raw memory, after some raw stores.
3480 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3481 // The Initialize can 'capture' suitably constrained stores as raw inits.
3482 // It can coalesce related raw stores into larger units (called 'tiles').
3483 // It can avoid zeroing new storage for memory units which have raw inits.
3484 // At macro-expansion, it is marked 'complete', and does not optimize further.
3485 //
3486 // EXAMPLE:
3487 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3488 //   ctl = incoming control; mem* = incoming memory
3489 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
3490 // First allocate uninitialized memory and fill in the header:
3491 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
3492 //   ctl := alloc.Control; mem* := alloc.Memory*
3493 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
3494 // Then initialize to zero the non-header parts of the raw memory block:
3495 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3496 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3497 // After the initialize node executes, the object is ready for service:
3498 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3499 // Suppose its body is immediately initialized as {1,2}:
3500 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3501 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3502 //   mem.SLICE(#short[*]) := store2
3503 //
3504 // DETAILS:
3505 // An InitializeNode collects and isolates object initialization after
3506 // an AllocateNode and before the next possible safepoint.  As a
3507 // memory barrier (MemBarNode), it keeps critical stores from drifting
3508 // down past any safepoint or any publication of the allocation.
3509 // Before this barrier, a newly-allocated object may have uninitialized bits.
3510 // After this barrier, it may be treated as a real oop, and GC is allowed.
3511 //
3512 // The semantics of the InitializeNode include an implicit zeroing of
3513 // the new object from object header to the end of the object.
3514 // (The object header and end are determined by the AllocateNode.)
3515 //
3516 // Certain stores may be added as direct inputs to the InitializeNode.
3517 // These stores must update raw memory, and they must be to addresses
3518 // derived from the raw address produced by AllocateNode, and with
3519 // a constant offset.  They must be ordered by increasing offset.
3520 // The first one is at in(RawStores), the last at in(req()-1).
3521 // Unlike most memory operations, they are not linked in a chain,
3522 // but are displayed in parallel as users of the rawmem output of
3523 // the allocation.
3524 //
3525 // (See comments in InitializeNode::capture_store, which continue
3526 // the example given above.)
3527 //
3528 // When the associated Allocate is macro-expanded, the InitializeNode
3529 // may be rewritten to optimize collected stores.  A ClearArrayNode
3530 // may also be created at that point to represent any required zeroing.
3531 // The InitializeNode is then marked 'complete', prohibiting further
3532 // capturing of nearby memory operations.
3533 //
3534 // During macro-expansion, all captured initializations which store
3535 // constant values of 32 bits or smaller are coalesced (if advantageous)
3536 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
3537 // initialized in fewer memory operations.  Memory words which are
3538 // covered by neither tiles nor non-constant stores are pre-zeroed
3539 // by explicit stores of zero.  (The code shape happens to do all
3540 // zeroing first, then all other stores, with both sequences occurring
3541 // in order of ascending offsets.)
3542 //
3543 // Alternatively, code may be inserted between an AllocateNode and its
3544 // InitializeNode, to perform arbitrary initialization of the new object.
3545 // E.g., the object copying intrinsics insert complex data transfers here.
3546 // The initialization must then be marked as 'complete' disable the
3547 // built-in zeroing semantics and the collection of initializing stores.
3548 //
3549 // While an InitializeNode is incomplete, reads from the memory state
3550 // produced by it are optimizable if they match the control edge and
3551 // new oop address associated with the allocation/initialization.
3552 // They return a stored value (if the offset matches) or else zero.
3553 // A write to the memory state, if it matches control and address,
3554 // and if it is to a constant offset, may be 'captured' by the
3555 // InitializeNode.  It is cloned as a raw memory operation and rewired
3556 // inside the initialization, to the raw oop produced by the allocation.
3557 // Operations on addresses which are provably distinct (e.g., to
3558 // other AllocateNodes) are allowed to bypass the initialization.
3559 //
3560 // The effect of all this is to consolidate object initialization
3561 // (both arrays and non-arrays, both piecewise and bulk) into a
3562 // single location, where it can be optimized as a unit.
3563 //
3564 // Only stores with an offset less than TrackedInitializationLimit words
3565 // will be considered for capture by an InitializeNode.  This puts a
3566 // reasonable limit on the complexity of optimized initializations.
3567 
3568 //---------------------------InitializeNode------------------------------------
3569 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3570   : MemBarNode(C, adr_type, rawoop),
3571     _is_complete(Incomplete), _does_not_escape(false)
3572 {
3573   init_class_id(Class_Initialize);
3574 
3575   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3576   assert(in(RawAddress) == rawoop, "proper init");
3577   // Note:  allocation() can be NULL, for secondary initialization barriers
3578 }
3579 
3580 // Since this node is not matched, it will be processed by the
3581 // register allocator.  Declare that there are no constraints
3582 // on the allocation of the RawAddress edge.
3583 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3584   // This edge should be set to top, by the set_complete.  But be conservative.
3585   if (idx == InitializeNode::RawAddress)
3586     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3587   return RegMask::Empty;
3588 }
3589 
3590 Node* InitializeNode::memory(uint alias_idx) {
3591   Node* mem = in(Memory);
3592   if (mem->is_MergeMem()) {
3593     return mem->as_MergeMem()->memory_at(alias_idx);
3594   } else {
3595     // incoming raw memory is not split
3596     return mem;
3597   }
3598 }
3599 
3600 bool InitializeNode::is_non_zero() {
3601   if (is_complete())  return false;
3602   remove_extra_zeroes();
3603   return (req() > RawStores);
3604 }
3605 
3606 void InitializeNode::set_complete(PhaseGVN* phase) {
3607   assert(!is_complete(), "caller responsibility");
3608   _is_complete = Complete;
3609 
3610   // After this node is complete, it contains a bunch of
3611   // raw-memory initializations.  There is no need for
3612   // it to have anything to do with non-raw memory effects.
3613   // Therefore, tell all non-raw users to re-optimize themselves,
3614   // after skipping the memory effects of this initialization.
3615   PhaseIterGVN* igvn = phase->is_IterGVN();
3616   if (igvn)  igvn->add_users_to_worklist(this);
3617 }
3618 
3619 // convenience function
3620 // return false if the init contains any stores already
3621 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3622   InitializeNode* init = initialization();
3623   if (init == NULL || init->is_complete()) {
3624     return false;
3625   }
3626   init->remove_extra_zeroes();
3627   // for now, if this allocation has already collected any inits, bail:
3628   if (init->is_non_zero())  return false;
3629   init->set_complete(phase);
3630   return true;
3631 }
3632 
3633 void InitializeNode::remove_extra_zeroes() {
3634   if (req() == RawStores)  return;
3635   Node* zmem = zero_memory();
3636   uint fill = RawStores;
3637   for (uint i = fill; i < req(); i++) {
3638     Node* n = in(i);
3639     if (n->is_top() || n == zmem)  continue;  // skip
3640     if (fill < i)  set_req(fill, n);          // compact
3641     ++fill;
3642   }
3643   // delete any empty spaces created:
3644   while (fill < req()) {
3645     del_req(fill);
3646   }
3647 }
3648 
3649 // Helper for remembering which stores go with which offsets.
3650 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3651   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
3652   intptr_t offset = -1;
3653   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3654                                                phase, offset);
3655   if (base == NULL)     return -1;  // something is dead,
3656   if (offset < 0)       return -1;  //        dead, dead
3657   return offset;
3658 }
3659 
3660 // Helper for proving that an initialization expression is
3661 // "simple enough" to be folded into an object initialization.
3662 // Attempts to prove that a store's initial value 'n' can be captured
3663 // within the initialization without creating a vicious cycle, such as:
3664 //     { Foo p = new Foo(); p.next = p; }
3665 // True for constants and parameters and small combinations thereof.
3666 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3667   ResourceMark rm;
3668   Unique_Node_List worklist;
3669   worklist.push(value);
3670 
3671   uint complexity_limit = 20;
3672   for (uint j = 0; j < worklist.size(); j++) {
3673     if (j >= complexity_limit) {
3674       return false;  // Bail out if processed too many nodes
3675     }
3676 
3677     Node* n = worklist.at(j);
3678     if (n == NULL)      continue;   // (can this really happen?)
3679     if (n->is_Proj())   n = n->in(0);
3680     if (n == this)      return false;  // found a cycle
3681     if (n->is_Con())    continue;
3682     if (n->is_Start())  continue;   // params, etc., are OK
3683     if (n->is_Root())   continue;   // even better
3684 
3685     // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3686     if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3687       continue;
3688     }
3689 
3690     Node* ctl = n->in(0);
3691     if (ctl != NULL && !ctl->is_top()) {
3692       if (ctl->is_Proj())  ctl = ctl->in(0);
3693       if (ctl == this)  return false;
3694 
3695       // If we already know that the enclosing memory op is pinned right after
3696       // the init, then any control flow that the store has picked up
3697       // must have preceded the init, or else be equal to the init.
3698       // Even after loop optimizations (which might change control edges)
3699       // a store is never pinned *before* the availability of its inputs.
3700       if (!MemNode::all_controls_dominate(n, this))
3701         return false;                  // failed to prove a good control
3702     }
3703 
3704     // Check data edges for possible dependencies on 'this'.
3705     for (uint i = 1; i < n->req(); i++) {
3706       Node* m = n->in(i);
3707       if (m == NULL || m == n || m->is_top())  continue;
3708 
3709       // Only process data inputs once
3710       worklist.push(m);
3711     }
3712   }
3713 
3714   return true;
3715 }
3716 
3717 // Here are all the checks a Store must pass before it can be moved into
3718 // an initialization.  Returns zero if a check fails.
3719 // On success, returns the (constant) offset to which the store applies,
3720 // within the initialized memory.
3721 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3722   const int FAIL = 0;
3723   if (st->req() != MemNode::ValueIn + 1)
3724     return FAIL;                // an inscrutable StoreNode (card mark?)
3725   Node* ctl = st->in(MemNode::Control);
3726   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3727     return FAIL;                // must be unconditional after the initialization
3728   Node* mem = st->in(MemNode::Memory);
3729   if (!(mem->is_Proj() && mem->in(0) == this))
3730     return FAIL;                // must not be preceded by other stores
3731   Node* adr = st->in(MemNode::Address);
3732   intptr_t offset;
3733   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3734   if (alloc == NULL)
3735     return FAIL;                // inscrutable address
3736   if (alloc != allocation())
3737     return FAIL;                // wrong allocation!  (store needs to float up)
3738   int size_in_bytes = st->memory_size();
3739   if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3740     return FAIL;                // mismatched access
3741   }
3742   Node* val = st->in(MemNode::ValueIn);
3743 
3744   if (!detect_init_independence(val, phase))
3745     return FAIL;                // stored value must be 'simple enough'
3746 
3747   // The Store can be captured only if nothing after the allocation
3748   // and before the Store is using the memory location that the store
3749   // overwrites.
3750   bool failed = false;
3751   // If is_complete_with_arraycopy() is true the shape of the graph is
3752   // well defined and is safe so no need for extra checks.
3753   if (!is_complete_with_arraycopy()) {
3754     // We are going to look at each use of the memory state following
3755     // the allocation to make sure nothing reads the memory that the
3756     // Store writes.
3757     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3758     int alias_idx = phase->C->get_alias_index(t_adr);
3759     ResourceMark rm;
3760     Unique_Node_List mems;
3761     mems.push(mem);
3762     Node* unique_merge = NULL;
3763     for (uint next = 0; next < mems.size(); ++next) {
3764       Node *m  = mems.at(next);
3765       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3766         Node *n = m->fast_out(j);
3767         if (n->outcnt() == 0) {
3768           continue;
3769         }
3770         if (n == st) {
3771           continue;
3772         } else if (n->in(0) != NULL && n->in(0) != ctl) {
3773           // If the control of this use is different from the control
3774           // of the Store which is right after the InitializeNode then
3775           // this node cannot be between the InitializeNode and the
3776           // Store.
3777           continue;
3778         } else if (n->is_MergeMem()) {
3779           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3780             // We can hit a MergeMemNode (that will likely go away
3781             // later) that is a direct use of the memory state
3782             // following the InitializeNode on the same slice as the
3783             // store node that we'd like to capture. We need to check
3784             // the uses of the MergeMemNode.
3785             mems.push(n);
3786           }
3787         } else if (n->is_Mem()) {
3788           Node* other_adr = n->in(MemNode::Address);
3789           if (other_adr == adr) {
3790             failed = true;
3791             break;
3792           } else {
3793             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3794             if (other_t_adr != NULL) {
3795               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3796               if (other_alias_idx == alias_idx) {
3797                 // A load from the same memory slice as the store right
3798                 // after the InitializeNode. We check the control of the
3799                 // object/array that is loaded from. If it's the same as
3800                 // the store control then we cannot capture the store.
3801                 assert(!n->is_Store(), "2 stores to same slice on same control?");
3802                 Node* base = other_adr;
3803                 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3804                 base = base->in(AddPNode::Base);
3805                 if (base != NULL) {
3806                   base = base->uncast();
3807                   if (base->is_Proj() && base->in(0) == alloc) {
3808                     failed = true;
3809                     break;
3810                   }
3811                 }
3812               }
3813             }
3814           }
3815         } else {
3816           failed = true;
3817           break;
3818         }
3819       }
3820     }
3821   }
3822   if (failed) {
3823     if (!can_reshape) {
3824       // We decided we couldn't capture the store during parsing. We
3825       // should try again during the next IGVN once the graph is
3826       // cleaner.
3827       phase->C->record_for_igvn(st);
3828     }
3829     return FAIL;
3830   }
3831 
3832   return offset;                // success
3833 }
3834 
3835 // Find the captured store in(i) which corresponds to the range
3836 // [start..start+size) in the initialized object.
3837 // If there is one, return its index i.  If there isn't, return the
3838 // negative of the index where it should be inserted.
3839 // Return 0 if the queried range overlaps an initialization boundary
3840 // or if dead code is encountered.
3841 // If size_in_bytes is zero, do not bother with overlap checks.
3842 int InitializeNode::captured_store_insertion_point(intptr_t start,
3843                                                    int size_in_bytes,
3844                                                    PhaseTransform* phase) {
3845   const int FAIL = 0, MAX_STORE = BytesPerLong;
3846 
3847   if (is_complete())
3848     return FAIL;                // arraycopy got here first; punt
3849 
3850   assert(allocation() != NULL, "must be present");
3851 
3852   // no negatives, no header fields:
3853   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
3854 
3855   // after a certain size, we bail out on tracking all the stores:
3856   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3857   if (start >= ti_limit)  return FAIL;
3858 
3859   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3860     if (i >= limit)  return -(int)i; // not found; here is where to put it
3861 
3862     Node*    st     = in(i);
3863     intptr_t st_off = get_store_offset(st, phase);
3864     if (st_off < 0) {
3865       if (st != zero_memory()) {
3866         return FAIL;            // bail out if there is dead garbage
3867       }
3868     } else if (st_off > start) {
3869       // ...we are done, since stores are ordered
3870       if (st_off < start + size_in_bytes) {
3871         return FAIL;            // the next store overlaps
3872       }
3873       return -(int)i;           // not found; here is where to put it
3874     } else if (st_off < start) {
3875       if (size_in_bytes != 0 &&
3876           start < st_off + MAX_STORE &&
3877           start < st_off + st->as_Store()->memory_size()) {
3878         return FAIL;            // the previous store overlaps
3879       }
3880     } else {
3881       if (size_in_bytes != 0 &&
3882           st->as_Store()->memory_size() != size_in_bytes) {
3883         return FAIL;            // mismatched store size
3884       }
3885       return i;
3886     }
3887 
3888     ++i;
3889   }
3890 }
3891 
3892 // Look for a captured store which initializes at the offset 'start'
3893 // with the given size.  If there is no such store, and no other
3894 // initialization interferes, then return zero_memory (the memory
3895 // projection of the AllocateNode).
3896 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3897                                           PhaseTransform* phase) {
3898   assert(stores_are_sane(phase), "");
3899   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3900   if (i == 0) {
3901     return NULL;                // something is dead
3902   } else if (i < 0) {
3903     return zero_memory();       // just primordial zero bits here
3904   } else {
3905     Node* st = in(i);           // here is the store at this position
3906     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3907     return st;
3908   }
3909 }
3910 
3911 // Create, as a raw pointer, an address within my new object at 'offset'.
3912 Node* InitializeNode::make_raw_address(intptr_t offset,
3913                                        PhaseTransform* phase) {
3914   Node* addr = in(RawAddress);
3915   if (offset != 0) {
3916     Compile* C = phase->C;
3917     addr = phase->transform( new AddPNode(C->top(), addr,
3918                                                  phase->MakeConX(offset)) );
3919   }
3920   return addr;
3921 }
3922 
3923 // Clone the given store, converting it into a raw store
3924 // initializing a field or element of my new object.
3925 // Caller is responsible for retiring the original store,
3926 // with subsume_node or the like.
3927 //
3928 // From the example above InitializeNode::InitializeNode,
3929 // here are the old stores to be captured:
3930 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3931 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3932 //
3933 // Here is the changed code; note the extra edges on init:
3934 //   alloc = (Allocate ...)
3935 //   rawoop = alloc.RawAddress
3936 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3937 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3938 //   init = (Initialize alloc.Control alloc.Memory rawoop
3939 //                      rawstore1 rawstore2)
3940 //
3941 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3942                                     PhaseGVN* phase, bool can_reshape) {
3943   assert(stores_are_sane(phase), "");
3944 
3945   if (start < 0)  return NULL;
3946   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3947 
3948   Compile* C = phase->C;
3949   int size_in_bytes = st->memory_size();
3950   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3951   if (i == 0)  return NULL;     // bail out
3952   Node* prev_mem = NULL;        // raw memory for the captured store
3953   if (i > 0) {
3954     prev_mem = in(i);           // there is a pre-existing store under this one
3955     set_req(i, C->top());       // temporarily disconnect it
3956     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3957   } else {
3958     i = -i;                     // no pre-existing store
3959     prev_mem = zero_memory();   // a slice of the newly allocated object
3960     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3961       set_req(--i, C->top());   // reuse this edge; it has been folded away
3962     else
3963       ins_req(i, C->top());     // build a new edge
3964   }
3965   Node* new_st = st->clone();
3966   new_st->set_req(MemNode::Control, in(Control));
3967   new_st->set_req(MemNode::Memory,  prev_mem);
3968   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3969   new_st = phase->transform(new_st);
3970 
3971   // At this point, new_st might have swallowed a pre-existing store
3972   // at the same offset, or perhaps new_st might have disappeared,
3973   // if it redundantly stored the same value (or zero to fresh memory).
3974 
3975   // In any case, wire it in:
3976   phase->igvn_rehash_node_delayed(this);
3977   set_req(i, new_st);
3978 
3979   // The caller may now kill the old guy.
3980   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3981   assert(check_st == new_st || check_st == NULL, "must be findable");
3982   assert(!is_complete(), "");
3983   return new_st;
3984 }
3985 
3986 static bool store_constant(jlong* tiles, int num_tiles,
3987                            intptr_t st_off, int st_size,
3988                            jlong con) {
3989   if ((st_off & (st_size-1)) != 0)
3990     return false;               // strange store offset (assume size==2**N)
3991   address addr = (address)tiles + st_off;
3992   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3993   switch (st_size) {
3994   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
3995   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
3996   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
3997   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
3998   default: return false;        // strange store size (detect size!=2**N here)
3999   }
4000   return true;                  // return success to caller
4001 }
4002 
4003 // Coalesce subword constants into int constants and possibly
4004 // into long constants.  The goal, if the CPU permits,
4005 // is to initialize the object with a small number of 64-bit tiles.
4006 // Also, convert floating-point constants to bit patterns.
4007 // Non-constants are not relevant to this pass.
4008 //
4009 // In terms of the running example on InitializeNode::InitializeNode
4010 // and InitializeNode::capture_store, here is the transformation
4011 // of rawstore1 and rawstore2 into rawstore12:
4012 //   alloc = (Allocate ...)
4013 //   rawoop = alloc.RawAddress
4014 //   tile12 = 0x00010002
4015 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4016 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4017 //
4018 void
4019 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4020                                         Node* size_in_bytes,
4021                                         PhaseGVN* phase) {
4022   Compile* C = phase->C;
4023 
4024   assert(stores_are_sane(phase), "");
4025   // Note:  After this pass, they are not completely sane,
4026   // since there may be some overlaps.
4027 
4028   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4029 
4030   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4031   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4032   size_limit = MIN2(size_limit, ti_limit);
4033   size_limit = align_up(size_limit, BytesPerLong);
4034   int num_tiles = size_limit / BytesPerLong;
4035 
4036   // allocate space for the tile map:
4037   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4038   jlong  tiles_buf[small_len];
4039   Node*  nodes_buf[small_len];
4040   jlong  inits_buf[small_len];
4041   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4042                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4043   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4044                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4045   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4046                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4047   // tiles: exact bitwise model of all primitive constants
4048   // nodes: last constant-storing node subsumed into the tiles model
4049   // inits: which bytes (in each tile) are touched by any initializations
4050 
4051   //// Pass A: Fill in the tile model with any relevant stores.
4052 
4053   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4054   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4055   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4056   Node* zmem = zero_memory(); // initially zero memory state
4057   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4058     Node* st = in(i);
4059     intptr_t st_off = get_store_offset(st, phase);
4060 
4061     // Figure out the store's offset and constant value:
4062     if (st_off < header_size)             continue; //skip (ignore header)
4063     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
4064     int st_size = st->as_Store()->memory_size();
4065     if (st_off + st_size > size_limit)    break;
4066 
4067     // Record which bytes are touched, whether by constant or not.
4068     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4069       continue;                 // skip (strange store size)
4070 
4071     const Type* val = phase->type(st->in(MemNode::ValueIn));
4072     if (!val->singleton())                continue; //skip (non-con store)
4073     BasicType type = val->basic_type();
4074 
4075     jlong con = 0;
4076     switch (type) {
4077     case T_INT:    con = val->is_int()->get_con();  break;
4078     case T_LONG:   con = val->is_long()->get_con(); break;
4079     case T_FLOAT:  con = jint_cast(val->getf());    break;
4080     case T_DOUBLE: con = jlong_cast(val->getd());   break;
4081     default:                              continue; //skip (odd store type)
4082     }
4083 
4084     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4085         st->Opcode() == Op_StoreL) {
4086       continue;                 // This StoreL is already optimal.
4087     }
4088 
4089     // Store down the constant.
4090     store_constant(tiles, num_tiles, st_off, st_size, con);
4091 
4092     intptr_t j = st_off >> LogBytesPerLong;
4093 
4094     if (type == T_INT && st_size == BytesPerInt
4095         && (st_off & BytesPerInt) == BytesPerInt) {
4096       jlong lcon = tiles[j];
4097       if (!Matcher::isSimpleConstant64(lcon) &&
4098           st->Opcode() == Op_StoreI) {
4099         // This StoreI is already optimal by itself.
4100         jint* intcon = (jint*) &tiles[j];
4101         intcon[1] = 0;  // undo the store_constant()
4102 
4103         // If the previous store is also optimal by itself, back up and
4104         // undo the action of the previous loop iteration... if we can.
4105         // But if we can't, just let the previous half take care of itself.
4106         st = nodes[j];
4107         st_off -= BytesPerInt;
4108         con = intcon[0];
4109         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
4110           assert(st_off >= header_size, "still ignoring header");
4111           assert(get_store_offset(st, phase) == st_off, "must be");
4112           assert(in(i-1) == zmem, "must be");
4113           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4114           assert(con == tcon->is_int()->get_con(), "must be");
4115           // Undo the effects of the previous loop trip, which swallowed st:
4116           intcon[0] = 0;        // undo store_constant()
4117           set_req(i-1, st);     // undo set_req(i, zmem)
4118           nodes[j] = NULL;      // undo nodes[j] = st
4119           --old_subword;        // undo ++old_subword
4120         }
4121         continue;               // This StoreI is already optimal.
4122       }
4123     }
4124 
4125     // This store is not needed.
4126     set_req(i, zmem);
4127     nodes[j] = st;              // record for the moment
4128     if (st_size < BytesPerLong) // something has changed
4129           ++old_subword;        // includes int/float, but who's counting...
4130     else  ++old_long;
4131   }
4132 
4133   if ((old_subword + old_long) == 0)
4134     return;                     // nothing more to do
4135 
4136   //// Pass B: Convert any non-zero tiles into optimal constant stores.
4137   // Be sure to insert them before overlapping non-constant stores.
4138   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
4139   for (int j = 0; j < num_tiles; j++) {
4140     jlong con  = tiles[j];
4141     jlong init = inits[j];
4142     if (con == 0)  continue;
4143     jint con0,  con1;           // split the constant, address-wise
4144     jint init0, init1;          // split the init map, address-wise
4145     { union { jlong con; jint intcon[2]; } u;
4146       u.con = con;
4147       con0  = u.intcon[0];
4148       con1  = u.intcon[1];
4149       u.con = init;
4150       init0 = u.intcon[0];
4151       init1 = u.intcon[1];
4152     }
4153 
4154     Node* old = nodes[j];
4155     assert(old != NULL, "need the prior store");
4156     intptr_t offset = (j * BytesPerLong);
4157 
4158     bool split = !Matcher::isSimpleConstant64(con);
4159 
4160     if (offset < header_size) {
4161       assert(offset + BytesPerInt >= header_size, "second int counts");
4162       assert(*(jint*)&tiles[j] == 0, "junk in header");
4163       split = true;             // only the second word counts
4164       // Example:  int a[] = { 42 ... }
4165     } else if (con0 == 0 && init0 == -1) {
4166       split = true;             // first word is covered by full inits
4167       // Example:  int a[] = { ... foo(), 42 ... }
4168     } else if (con1 == 0 && init1 == -1) {
4169       split = true;             // second word is covered by full inits
4170       // Example:  int a[] = { ... 42, foo() ... }
4171     }
4172 
4173     // Here's a case where init0 is neither 0 nor -1:
4174     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
4175     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4176     // In this case the tile is not split; it is (jlong)42.
4177     // The big tile is stored down, and then the foo() value is inserted.
4178     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4179 
4180     Node* ctl = old->in(MemNode::Control);
4181     Node* adr = make_raw_address(offset, phase);
4182     const TypePtr* atp = TypeRawPtr::BOTTOM;
4183 
4184     // One or two coalesced stores to plop down.
4185     Node*    st[2];
4186     intptr_t off[2];
4187     int  nst = 0;
4188     if (!split) {
4189       ++new_long;
4190       off[nst] = offset;
4191       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4192                                   phase->longcon(con), T_LONG, MemNode::unordered);
4193     } else {
4194       // Omit either if it is a zero.
4195       if (con0 != 0) {
4196         ++new_int;
4197         off[nst]  = offset;
4198         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4199                                     phase->intcon(con0), T_INT, MemNode::unordered);
4200       }
4201       if (con1 != 0) {
4202         ++new_int;
4203         offset += BytesPerInt;
4204         adr = make_raw_address(offset, phase);
4205         off[nst]  = offset;
4206         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4207                                     phase->intcon(con1), T_INT, MemNode::unordered);
4208       }
4209     }
4210 
4211     // Insert second store first, then the first before the second.
4212     // Insert each one just before any overlapping non-constant stores.
4213     while (nst > 0) {
4214       Node* st1 = st[--nst];
4215       C->copy_node_notes_to(st1, old);
4216       st1 = phase->transform(st1);
4217       offset = off[nst];
4218       assert(offset >= header_size, "do not smash header");
4219       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4220       guarantee(ins_idx != 0, "must re-insert constant store");
4221       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
4222       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4223         set_req(--ins_idx, st1);
4224       else
4225         ins_req(ins_idx, st1);
4226     }
4227   }
4228 
4229   if (PrintCompilation && WizardMode)
4230     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4231                   old_subword, old_long, new_int, new_long);
4232   if (C->log() != NULL)
4233     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4234                    old_subword, old_long, new_int, new_long);
4235 
4236   // Clean up any remaining occurrences of zmem:
4237   remove_extra_zeroes();
4238 }
4239 
4240 // Explore forward from in(start) to find the first fully initialized
4241 // word, and return its offset.  Skip groups of subword stores which
4242 // together initialize full words.  If in(start) is itself part of a
4243 // fully initialized word, return the offset of in(start).  If there
4244 // are no following full-word stores, or if something is fishy, return
4245 // a negative value.
4246 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4247   int       int_map = 0;
4248   intptr_t  int_map_off = 0;
4249   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
4250 
4251   for (uint i = start, limit = req(); i < limit; i++) {
4252     Node* st = in(i);
4253 
4254     intptr_t st_off = get_store_offset(st, phase);
4255     if (st_off < 0)  break;  // return conservative answer
4256 
4257     int st_size = st->as_Store()->memory_size();
4258     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4259       return st_off;            // we found a complete word init
4260     }
4261 
4262     // update the map:
4263 
4264     intptr_t this_int_off = align_down(st_off, BytesPerInt);
4265     if (this_int_off != int_map_off) {
4266       // reset the map:
4267       int_map = 0;
4268       int_map_off = this_int_off;
4269     }
4270 
4271     int subword_off = st_off - this_int_off;
4272     int_map |= right_n_bits(st_size) << subword_off;
4273     if ((int_map & FULL_MAP) == FULL_MAP) {
4274       return this_int_off;      // we found a complete word init
4275     }
4276 
4277     // Did this store hit or cross the word boundary?
4278     intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4279     if (next_int_off == this_int_off + BytesPerInt) {
4280       // We passed the current int, without fully initializing it.
4281       int_map_off = next_int_off;
4282       int_map >>= BytesPerInt;
4283     } else if (next_int_off > this_int_off + BytesPerInt) {
4284       // We passed the current and next int.
4285       return this_int_off + BytesPerInt;
4286     }
4287   }
4288 
4289   return -1;
4290 }
4291 
4292 
4293 // Called when the associated AllocateNode is expanded into CFG.
4294 // At this point, we may perform additional optimizations.
4295 // Linearize the stores by ascending offset, to make memory
4296 // activity as coherent as possible.
4297 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4298                                       intptr_t header_size,
4299                                       Node* size_in_bytes,
4300                                       PhaseIterGVN* phase) {
4301   assert(!is_complete(), "not already complete");
4302   assert(stores_are_sane(phase), "");
4303   assert(allocation() != NULL, "must be present");
4304 
4305   remove_extra_zeroes();
4306 
4307   if (ReduceFieldZeroing || ReduceBulkZeroing)
4308     // reduce instruction count for common initialization patterns
4309     coalesce_subword_stores(header_size, size_in_bytes, phase);
4310 
4311   Node* zmem = zero_memory();   // initially zero memory state
4312   Node* inits = zmem;           // accumulating a linearized chain of inits
4313   #ifdef ASSERT
4314   intptr_t first_offset = allocation()->minimum_header_size();
4315   intptr_t last_init_off = first_offset;  // previous init offset
4316   intptr_t last_init_end = first_offset;  // previous init offset+size
4317   intptr_t last_tile_end = first_offset;  // previous tile offset+size
4318   #endif
4319   intptr_t zeroes_done = header_size;
4320 
4321   bool do_zeroing = true;       // we might give up if inits are very sparse
4322   int  big_init_gaps = 0;       // how many large gaps have we seen?
4323 
4324   if (UseTLAB && ZeroTLAB)  do_zeroing = false;
4325   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
4326 
4327   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4328     Node* st = in(i);
4329     intptr_t st_off = get_store_offset(st, phase);
4330     if (st_off < 0)
4331       break;                    // unknown junk in the inits
4332     if (st->in(MemNode::Memory) != zmem)
4333       break;                    // complicated store chains somehow in list
4334 
4335     int st_size = st->as_Store()->memory_size();
4336     intptr_t next_init_off = st_off + st_size;
4337 
4338     if (do_zeroing && zeroes_done < next_init_off) {
4339       // See if this store needs a zero before it or under it.
4340       intptr_t zeroes_needed = st_off;
4341 
4342       if (st_size < BytesPerInt) {
4343         // Look for subword stores which only partially initialize words.
4344         // If we find some, we must lay down some word-level zeroes first,
4345         // underneath the subword stores.
4346         //
4347         // Examples:
4348         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
4349         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
4350         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
4351         //
4352         // Note:  coalesce_subword_stores may have already done this,
4353         // if it was prompted by constant non-zero subword initializers.
4354         // But this case can still arise with non-constant stores.
4355 
4356         intptr_t next_full_store = find_next_fullword_store(i, phase);
4357 
4358         // In the examples above:
4359         //   in(i)          p   q   r   s     x   y     z
4360         //   st_off        12  13  14  15    12  13    14
4361         //   st_size        1   1   1   1     1   1     1
4362         //   next_full_s.  12  16  16  16    16  16    16
4363         //   z's_done      12  16  16  16    12  16    12
4364         //   z's_needed    12  16  16  16    16  16    16
4365         //   zsize          0   0   0   0     4   0     4
4366         if (next_full_store < 0) {
4367           // Conservative tack:  Zero to end of current word.
4368           zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4369         } else {
4370           // Zero to beginning of next fully initialized word.
4371           // Or, don't zero at all, if we are already in that word.
4372           assert(next_full_store >= zeroes_needed, "must go forward");
4373           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4374           zeroes_needed = next_full_store;
4375         }
4376       }
4377 
4378       if (zeroes_needed > zeroes_done) {
4379         intptr_t zsize = zeroes_needed - zeroes_done;
4380         // Do some incremental zeroing on rawmem, in parallel with inits.
4381         zeroes_done = align_down(zeroes_done, BytesPerInt);
4382         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4383                                               allocation()->in(AllocateNode::DefaultValue),
4384                                               allocation()->in(AllocateNode::RawDefaultValue),
4385                                               zeroes_done, zeroes_needed,
4386                                               phase);
4387         zeroes_done = zeroes_needed;
4388         if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4389           do_zeroing = false;   // leave the hole, next time
4390       }
4391     }
4392 
4393     // Collect the store and move on:
4394     phase->replace_input_of(st, MemNode::Memory, inits);
4395     inits = st;                 // put it on the linearized chain
4396     set_req(i, zmem);           // unhook from previous position
4397 
4398     if (zeroes_done == st_off)
4399       zeroes_done = next_init_off;
4400 
4401     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4402 
4403     #ifdef ASSERT
4404     // Various order invariants.  Weaker than stores_are_sane because
4405     // a large constant tile can be filled in by smaller non-constant stores.
4406     assert(st_off >= last_init_off, "inits do not reverse");
4407     last_init_off = st_off;
4408     const Type* val = NULL;
4409     if (st_size >= BytesPerInt &&
4410         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4411         (int)val->basic_type() < (int)T_OBJECT) {
4412       assert(st_off >= last_tile_end, "tiles do not overlap");
4413       assert(st_off >= last_init_end, "tiles do not overwrite inits");
4414       last_tile_end = MAX2(last_tile_end, next_init_off);
4415     } else {
4416       intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4417       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4418       assert(st_off      >= last_init_end, "inits do not overlap");
4419       last_init_end = next_init_off;  // it's a non-tile
4420     }
4421     #endif //ASSERT
4422   }
4423 
4424   remove_extra_zeroes();        // clear out all the zmems left over
4425   add_req(inits);
4426 
4427   if (!(UseTLAB && ZeroTLAB)) {
4428     // If anything remains to be zeroed, zero it all now.
4429     zeroes_done = align_down(zeroes_done, BytesPerInt);
4430     // if it is the last unused 4 bytes of an instance, forget about it
4431     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4432     if (zeroes_done + BytesPerLong >= size_limit) {
4433       AllocateNode* alloc = allocation();
4434       assert(alloc != NULL, "must be present");
4435       if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4436         Node* klass_node = alloc->in(AllocateNode::KlassNode);
4437         ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4438         if (zeroes_done == k->layout_helper())
4439           zeroes_done = size_limit;
4440       }
4441     }
4442     if (zeroes_done < size_limit) {
4443       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4444                                             allocation()->in(AllocateNode::DefaultValue),
4445                                             allocation()->in(AllocateNode::RawDefaultValue),
4446                                             zeroes_done, size_in_bytes, phase);
4447     }
4448   }
4449 
4450   set_complete(phase);
4451   return rawmem;
4452 }
4453 
4454 
4455 #ifdef ASSERT
4456 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4457   if (is_complete())
4458     return true;                // stores could be anything at this point
4459   assert(allocation() != NULL, "must be present");
4460   intptr_t last_off = allocation()->minimum_header_size();
4461   for (uint i = InitializeNode::RawStores; i < req(); i++) {
4462     Node* st = in(i);
4463     intptr_t st_off = get_store_offset(st, phase);
4464     if (st_off < 0)  continue;  // ignore dead garbage
4465     if (last_off > st_off) {
4466       tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4467       this->dump(2);
4468       assert(false, "ascending store offsets");
4469       return false;
4470     }
4471     last_off = st_off + st->as_Store()->memory_size();
4472   }
4473   return true;
4474 }
4475 #endif //ASSERT
4476 
4477 
4478 
4479 
4480 //============================MergeMemNode=====================================
4481 //
4482 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
4483 // contributing store or call operations.  Each contributor provides the memory
4484 // state for a particular "alias type" (see Compile::alias_type).  For example,
4485 // if a MergeMem has an input X for alias category #6, then any memory reference
4486 // to alias category #6 may use X as its memory state input, as an exact equivalent
4487 // to using the MergeMem as a whole.
4488 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4489 //
4490 // (Here, the <N> notation gives the index of the relevant adr_type.)
4491 //
4492 // In one special case (and more cases in the future), alias categories overlap.
4493 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4494 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
4495 // it is exactly equivalent to that state W:
4496 //   MergeMem(<Bot>: W) <==> W
4497 //
4498 // Usually, the merge has more than one input.  In that case, where inputs
4499 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4500 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4501 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4502 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4503 //
4504 // A merge can take a "wide" memory state as one of its narrow inputs.
4505 // This simply means that the merge observes out only the relevant parts of
4506 // the wide input.  That is, wide memory states arriving at narrow merge inputs
4507 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
4508 //
4509 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4510 // and that memory slices "leak through":
4511 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4512 //
4513 // But, in such a cascade, repeated memory slices can "block the leak":
4514 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4515 //
4516 // In the last example, Y is not part of the combined memory state of the
4517 // outermost MergeMem.  The system must, of course, prevent unschedulable
4518 // memory states from arising, so you can be sure that the state Y is somehow
4519 // a precursor to state Y'.
4520 //
4521 //
4522 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4523 // of each MergeMemNode array are exactly the numerical alias indexes, including
4524 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
4525 // Compile::alias_type (and kin) produce and manage these indexes.
4526 //
4527 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4528 // (Note that this provides quick access to the top node inside MergeMem methods,
4529 // without the need to reach out via TLS to Compile::current.)
4530 //
4531 // As a consequence of what was just described, a MergeMem that represents a full
4532 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4533 // containing all alias categories.
4534 //
4535 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4536 //
4537 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4538 // a memory state for the alias type <N>, or else the top node, meaning that
4539 // there is no particular input for that alias type.  Note that the length of
4540 // a MergeMem is variable, and may be extended at any time to accommodate new
4541 // memory states at larger alias indexes.  When merges grow, they are of course
4542 // filled with "top" in the unused in() positions.
4543 //
4544 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4545 // (Top was chosen because it works smoothly with passes like GCM.)
4546 //
4547 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
4548 // the type of random VM bits like TLS references.)  Since it is always the
4549 // first non-Bot memory slice, some low-level loops use it to initialize an
4550 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
4551 //
4552 //
4553 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
4554 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
4555 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4556 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
4557 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
4558 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4559 //
4560 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4561 // really that different from the other memory inputs.  An abbreviation called
4562 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4563 //
4564 //
4565 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
4566 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
4567 // that "emerges though" the base memory will be marked as excluding the alias types
4568 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4569 //
4570 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4571 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4572 //
4573 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4574 // (It is currently unimplemented.)  As you can see, the resulting merge is
4575 // actually a disjoint union of memory states, rather than an overlay.
4576 //
4577 
4578 //------------------------------MergeMemNode-----------------------------------
4579 Node* MergeMemNode::make_empty_memory() {
4580   Node* empty_memory = (Node*) Compile::current()->top();
4581   assert(empty_memory->is_top(), "correct sentinel identity");
4582   return empty_memory;
4583 }
4584 
4585 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4586   init_class_id(Class_MergeMem);
4587   // all inputs are nullified in Node::Node(int)
4588   // set_input(0, NULL);  // no control input
4589 
4590   // Initialize the edges uniformly to top, for starters.
4591   Node* empty_mem = make_empty_memory();
4592   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4593     init_req(i,empty_mem);
4594   }
4595   assert(empty_memory() == empty_mem, "");
4596 
4597   if( new_base != NULL && new_base->is_MergeMem() ) {
4598     MergeMemNode* mdef = new_base->as_MergeMem();
4599     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4600     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4601       mms.set_memory(mms.memory2());
4602     }
4603     assert(base_memory() == mdef->base_memory(), "");
4604   } else {
4605     set_base_memory(new_base);
4606   }
4607 }
4608 
4609 // Make a new, untransformed MergeMem with the same base as 'mem'.
4610 // If mem is itself a MergeMem, populate the result with the same edges.
4611 MergeMemNode* MergeMemNode::make(Node* mem) {
4612   return new MergeMemNode(mem);
4613 }
4614 
4615 //------------------------------cmp--------------------------------------------
4616 uint MergeMemNode::hash() const { return NO_HASH; }
4617 bool MergeMemNode::cmp( const Node &n ) const {
4618   return (&n == this);          // Always fail except on self
4619 }
4620 
4621 //------------------------------Identity---------------------------------------
4622 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4623   // Identity if this merge point does not record any interesting memory
4624   // disambiguations.
4625   Node* base_mem = base_memory();
4626   Node* empty_mem = empty_memory();
4627   if (base_mem != empty_mem) {  // Memory path is not dead?
4628     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4629       Node* mem = in(i);
4630       if (mem != empty_mem && mem != base_mem) {
4631         return this;            // Many memory splits; no change
4632       }
4633     }
4634   }
4635   return base_mem;              // No memory splits; ID on the one true input
4636 }
4637 
4638 //------------------------------Ideal------------------------------------------
4639 // This method is invoked recursively on chains of MergeMem nodes
4640 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4641   // Remove chain'd MergeMems
4642   //
4643   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4644   // relative to the "in(Bot)".  Since we are patching both at the same time,
4645   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4646   // but rewrite each "in(i)" relative to the new "in(Bot)".
4647   Node *progress = NULL;
4648 
4649 
4650   Node* old_base = base_memory();
4651   Node* empty_mem = empty_memory();
4652   if (old_base == empty_mem)
4653     return NULL; // Dead memory path.
4654 
4655   MergeMemNode* old_mbase;
4656   if (old_base != NULL && old_base->is_MergeMem())
4657     old_mbase = old_base->as_MergeMem();
4658   else
4659     old_mbase = NULL;
4660   Node* new_base = old_base;
4661 
4662   // simplify stacked MergeMems in base memory
4663   if (old_mbase)  new_base = old_mbase->base_memory();
4664 
4665   // the base memory might contribute new slices beyond my req()
4666   if (old_mbase)  grow_to_match(old_mbase);
4667 
4668   // Look carefully at the base node if it is a phi.
4669   PhiNode* phi_base;
4670   if (new_base != NULL && new_base->is_Phi())
4671     phi_base = new_base->as_Phi();
4672   else
4673     phi_base = NULL;
4674 
4675   Node*    phi_reg = NULL;
4676   uint     phi_len = (uint)-1;
4677   if (phi_base != NULL && !phi_base->is_copy()) {
4678     // do not examine phi if degraded to a copy
4679     phi_reg = phi_base->region();
4680     phi_len = phi_base->req();
4681     // see if the phi is unfinished
4682     for (uint i = 1; i < phi_len; i++) {
4683       if (phi_base->in(i) == NULL) {
4684         // incomplete phi; do not look at it yet!
4685         phi_reg = NULL;
4686         phi_len = (uint)-1;
4687         break;
4688       }
4689     }
4690   }
4691 
4692   // Note:  We do not call verify_sparse on entry, because inputs
4693   // can normalize to the base_memory via subsume_node or similar
4694   // mechanisms.  This method repairs that damage.
4695 
4696   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4697 
4698   // Look at each slice.
4699   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4700     Node* old_in = in(i);
4701     // calculate the old memory value
4702     Node* old_mem = old_in;
4703     if (old_mem == empty_mem)  old_mem = old_base;
4704     assert(old_mem == memory_at(i), "");
4705 
4706     // maybe update (reslice) the old memory value
4707 
4708     // simplify stacked MergeMems
4709     Node* new_mem = old_mem;
4710     MergeMemNode* old_mmem;
4711     if (old_mem != NULL && old_mem->is_MergeMem())
4712       old_mmem = old_mem->as_MergeMem();
4713     else
4714       old_mmem = NULL;
4715     if (old_mmem == this) {
4716       // This can happen if loops break up and safepoints disappear.
4717       // A merge of BotPtr (default) with a RawPtr memory derived from a
4718       // safepoint can be rewritten to a merge of the same BotPtr with
4719       // the BotPtr phi coming into the loop.  If that phi disappears
4720       // also, we can end up with a self-loop of the mergemem.
4721       // In general, if loops degenerate and memory effects disappear,
4722       // a mergemem can be left looking at itself.  This simply means
4723       // that the mergemem's default should be used, since there is
4724       // no longer any apparent effect on this slice.
4725       // Note: If a memory slice is a MergeMem cycle, it is unreachable
4726       //       from start.  Update the input to TOP.
4727       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4728     }
4729     else if (old_mmem != NULL) {
4730       new_mem = old_mmem->memory_at(i);
4731     }
4732     // else preceding memory was not a MergeMem
4733 
4734     // maybe store down a new value
4735     Node* new_in = new_mem;
4736     if (new_in == new_base)  new_in = empty_mem;
4737 
4738     if (new_in != old_in) {
4739       // Warning:  Do not combine this "if" with the previous "if"
4740       // A memory slice might have be be rewritten even if it is semantically
4741       // unchanged, if the base_memory value has changed.
4742       set_req(i, new_in);
4743       progress = this;          // Report progress
4744     }
4745   }
4746 
4747   if (new_base != old_base) {
4748     set_req(Compile::AliasIdxBot, new_base);
4749     // Don't use set_base_memory(new_base), because we need to update du.
4750     assert(base_memory() == new_base, "");
4751     progress = this;
4752   }
4753 
4754   if( base_memory() == this ) {
4755     // a self cycle indicates this memory path is dead
4756     set_req(Compile::AliasIdxBot, empty_mem);
4757   }
4758 
4759   // Resolve external cycles by calling Ideal on a MergeMem base_memory
4760   // Recursion must occur after the self cycle check above
4761   if( base_memory()->is_MergeMem() ) {
4762     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4763     Node *m = phase->transform(new_mbase);  // Rollup any cycles
4764     if( m != NULL &&
4765         (m->is_top() ||
4766          (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4767       // propagate rollup of dead cycle to self
4768       set_req(Compile::AliasIdxBot, empty_mem);
4769     }
4770   }
4771 
4772   if( base_memory() == empty_mem ) {
4773     progress = this;
4774     // Cut inputs during Parse phase only.
4775     // During Optimize phase a dead MergeMem node will be subsumed by Top.
4776     if( !can_reshape ) {
4777       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4778         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4779       }
4780     }
4781   }
4782 
4783   if( !progress && base_memory()->is_Phi() && can_reshape ) {
4784     // Check if PhiNode::Ideal's "Split phis through memory merges"
4785     // transform should be attempted. Look for this->phi->this cycle.
4786     uint merge_width = req();
4787     if (merge_width > Compile::AliasIdxRaw) {
4788       PhiNode* phi = base_memory()->as_Phi();
4789       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4790         if (phi->in(i) == this) {
4791           phase->is_IterGVN()->_worklist.push(phi);
4792           break;
4793         }
4794       }
4795     }
4796   }
4797 
4798   assert(progress || verify_sparse(), "please, no dups of base");
4799   return progress;
4800 }
4801 
4802 //-------------------------set_base_memory-------------------------------------
4803 void MergeMemNode::set_base_memory(Node *new_base) {
4804   Node* empty_mem = empty_memory();
4805   set_req(Compile::AliasIdxBot, new_base);
4806   assert(memory_at(req()) == new_base, "must set default memory");
4807   // Clear out other occurrences of new_base:
4808   if (new_base != empty_mem) {
4809     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4810       if (in(i) == new_base)  set_req(i, empty_mem);
4811     }
4812   }
4813 }
4814 
4815 //------------------------------out_RegMask------------------------------------
4816 const RegMask &MergeMemNode::out_RegMask() const {
4817   return RegMask::Empty;
4818 }
4819 
4820 //------------------------------dump_spec--------------------------------------
4821 #ifndef PRODUCT
4822 void MergeMemNode::dump_spec(outputStream *st) const {
4823   st->print(" {");
4824   Node* base_mem = base_memory();
4825   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4826     Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4827     if (mem == base_mem) { st->print(" -"); continue; }
4828     st->print( " N%d:", mem->_idx );
4829     Compile::current()->get_adr_type(i)->dump_on(st);
4830   }
4831   st->print(" }");
4832 }
4833 #endif // !PRODUCT
4834 
4835 
4836 #ifdef ASSERT
4837 static bool might_be_same(Node* a, Node* b) {
4838   if (a == b)  return true;
4839   if (!(a->is_Phi() || b->is_Phi()))  return false;
4840   // phis shift around during optimization
4841   return true;  // pretty stupid...
4842 }
4843 
4844 // verify a narrow slice (either incoming or outgoing)
4845 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4846   if (!VerifyAliases)                return;  // don't bother to verify unless requested
4847   if (VMError::is_error_reported())  return;  // muzzle asserts when debugging an error
4848   if (Node::in_dump())               return;  // muzzle asserts when printing
4849   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4850   assert(n != NULL, "");
4851   // Elide intervening MergeMem's
4852   while (n->is_MergeMem()) {
4853     n = n->as_MergeMem()->memory_at(alias_idx);
4854   }
4855   Compile* C = Compile::current();
4856   const TypePtr* n_adr_type = n->adr_type();
4857   if (n == m->empty_memory()) {
4858     // Implicit copy of base_memory()
4859   } else if (n_adr_type != TypePtr::BOTTOM) {
4860     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4861     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4862   } else {
4863     // A few places like make_runtime_call "know" that VM calls are narrow,
4864     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4865     bool expected_wide_mem = false;
4866     if (n == m->base_memory()) {
4867       expected_wide_mem = true;
4868     } else if (alias_idx == Compile::AliasIdxRaw ||
4869                n == m->memory_at(Compile::AliasIdxRaw)) {
4870       expected_wide_mem = true;
4871     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4872       // memory can "leak through" calls on channels that
4873       // are write-once.  Allow this also.
4874       expected_wide_mem = true;
4875     }
4876     assert(expected_wide_mem, "expected narrow slice replacement");
4877   }
4878 }
4879 #else // !ASSERT
4880 #define verify_memory_slice(m,i,n) (void)(0)  // PRODUCT version is no-op
4881 #endif
4882 
4883 
4884 //-----------------------------memory_at---------------------------------------
4885 Node* MergeMemNode::memory_at(uint alias_idx) const {
4886   assert(alias_idx >= Compile::AliasIdxRaw ||
4887          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4888          "must avoid base_memory and AliasIdxTop");
4889 
4890   // Otherwise, it is a narrow slice.
4891   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4892   Compile *C = Compile::current();
4893   if (is_empty_memory(n)) {
4894     // the array is sparse; empty slots are the "top" node
4895     n = base_memory();
4896     assert(Node::in_dump()
4897            || n == NULL || n->bottom_type() == Type::TOP
4898            || n->adr_type() == NULL // address is TOP
4899            || n->adr_type() == TypePtr::BOTTOM
4900            || n->adr_type() == TypeRawPtr::BOTTOM
4901            || Compile::current()->AliasLevel() == 0,
4902            "must be a wide memory");
4903     // AliasLevel == 0 if we are organizing the memory states manually.
4904     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4905   } else {
4906     // make sure the stored slice is sane
4907     #ifdef ASSERT
4908     if (VMError::is_error_reported() || Node::in_dump()) {
4909     } else if (might_be_same(n, base_memory())) {
4910       // Give it a pass:  It is a mostly harmless repetition of the base.
4911       // This can arise normally from node subsumption during optimization.
4912     } else {
4913       verify_memory_slice(this, alias_idx, n);
4914     }
4915     #endif
4916   }
4917   return n;
4918 }
4919 
4920 //---------------------------set_memory_at-------------------------------------
4921 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4922   verify_memory_slice(this, alias_idx, n);
4923   Node* empty_mem = empty_memory();
4924   if (n == base_memory())  n = empty_mem;  // collapse default
4925   uint need_req = alias_idx+1;
4926   if (req() < need_req) {
4927     if (n == empty_mem)  return;  // already the default, so do not grow me
4928     // grow the sparse array
4929     do {
4930       add_req(empty_mem);
4931     } while (req() < need_req);
4932   }
4933   set_req( alias_idx, n );
4934 }
4935 
4936 
4937 
4938 //--------------------------iteration_setup------------------------------------
4939 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4940   if (other != NULL) {
4941     grow_to_match(other);
4942     // invariant:  the finite support of mm2 is within mm->req()
4943     #ifdef ASSERT
4944     for (uint i = req(); i < other->req(); i++) {
4945       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4946     }
4947     #endif
4948   }
4949   // Replace spurious copies of base_memory by top.
4950   Node* base_mem = base_memory();
4951   if (base_mem != NULL && !base_mem->is_top()) {
4952     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4953       if (in(i) == base_mem)
4954         set_req(i, empty_memory());
4955     }
4956   }
4957 }
4958 
4959 //---------------------------grow_to_match-------------------------------------
4960 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4961   Node* empty_mem = empty_memory();
4962   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4963   // look for the finite support of the other memory
4964   for (uint i = other->req(); --i >= req(); ) {
4965     if (other->in(i) != empty_mem) {
4966       uint new_len = i+1;
4967       while (req() < new_len)  add_req(empty_mem);
4968       break;
4969     }
4970   }
4971 }
4972 
4973 //---------------------------verify_sparse-------------------------------------
4974 #ifndef PRODUCT
4975 bool MergeMemNode::verify_sparse() const {
4976   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4977   Node* base_mem = base_memory();
4978   // The following can happen in degenerate cases, since empty==top.
4979   if (is_empty_memory(base_mem))  return true;
4980   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4981     assert(in(i) != NULL, "sane slice");
4982     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
4983   }
4984   return true;
4985 }
4986 
4987 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4988   Node* n;
4989   n = mm->in(idx);
4990   if (mem == n)  return true;  // might be empty_memory()
4991   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4992   if (mem == n)  return true;
4993   while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4994     if (mem == n)  return true;
4995     if (n == NULL)  break;
4996   }
4997   return false;
4998 }
4999 #endif // !PRODUCT