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