1 /*
   2  * Copyright (c) 1997, 2014, 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/cfgnode.hpp"
  32 #include "opto/compile.hpp"
  33 #include "opto/connode.hpp"
  34 #include "opto/convertnode.hpp"
  35 #include "opto/loopnode.hpp"
  36 #include "opto/machnode.hpp"
  37 #include "opto/matcher.hpp"
  38 #include "opto/memnode.hpp"
  39 #include "opto/mulnode.hpp"
  40 #include "opto/narrowptrnode.hpp"
  41 #include "opto/phaseX.hpp"
  42 #include "opto/regmask.hpp"
  43 
  44 // Portions of code courtesy of Clifford Click
  45 
  46 // Optimization - Graph Style
  47 
  48 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
  49 
  50 //=============================================================================
  51 uint MemNode::size_of() const { return sizeof(*this); }
  52 
  53 const TypePtr *MemNode::adr_type() const {
  54   Node* adr = in(Address);
  55   const TypePtr* cross_check = NULL;
  56   DEBUG_ONLY(cross_check = _adr_type);
  57   return calculate_adr_type(adr->bottom_type(), cross_check);
  58 }
  59 
  60 #ifndef PRODUCT
  61 void MemNode::dump_spec(outputStream *st) const {
  62   if (in(Address) == NULL)  return; // node is dead
  63 #ifndef ASSERT
  64   // fake the missing field
  65   const TypePtr* _adr_type = NULL;
  66   if (in(Address) != NULL)
  67     _adr_type = in(Address)->bottom_type()->isa_ptr();
  68 #endif
  69   dump_adr_type(this, _adr_type, st);
  70 
  71   Compile* C = Compile::current();
  72   if( C->alias_type(_adr_type)->is_volatile() )
  73     st->print(" Volatile!");
  74 }
  75 
  76 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
  77   st->print(" @");
  78   if (adr_type == NULL) {
  79     st->print("NULL");
  80   } else {
  81     adr_type->dump_on(st);
  82     Compile* C = Compile::current();
  83     Compile::AliasType* atp = NULL;
  84     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
  85     if (atp == NULL)
  86       st->print(", idx=?\?;");
  87     else if (atp->index() == Compile::AliasIdxBot)
  88       st->print(", idx=Bot;");
  89     else if (atp->index() == Compile::AliasIdxTop)
  90       st->print(", idx=Top;");
  91     else if (atp->index() == Compile::AliasIdxRaw)
  92       st->print(", idx=Raw;");
  93     else {
  94       ciField* field = atp->field();
  95       if (field) {
  96         st->print(", name=");
  97         field->print_name_on(st);
  98       }
  99       st->print(", idx=%d;", atp->index());
 100     }
 101   }
 102 }
 103 
 104 extern void print_alias_types();
 105 
 106 #endif
 107 
 108 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
 109   assert((t_oop != NULL), "sanity");
 110   bool is_instance = t_oop->is_known_instance_field();
 111   bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
 112                              (load != NULL) && load->is_Load() &&
 113                              (phase->is_IterGVN() != NULL);
 114   if (!(is_instance || is_boxed_value_load))
 115     return mchain;  // don't try to optimize non-instance types
 116   uint instance_id = t_oop->instance_id();
 117   Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
 118   Node *prev = NULL;
 119   Node *result = mchain;
 120   while (prev != result) {
 121     prev = result;
 122     if (result == start_mem)
 123       break;  // hit one of our sentinels
 124     // skip over a call which does not affect this memory slice
 125     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
 126       Node *proj_in = result->in(0);
 127       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
 128         break;  // hit one of our sentinels
 129       } else if (proj_in->is_Call()) {
 130         CallNode *call = proj_in->as_Call();
 131         if (!call->may_modify(t_oop, phase)) { // returns false for instances
 132           result = call->in(TypeFunc::Memory);
 133         }
 134       } else if (proj_in->is_Initialize()) {
 135         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
 136         // Stop if this is the initialization for the object instance which
 137         // which contains this memory slice, otherwise skip over it.
 138         if ((alloc == NULL) || (alloc->_idx == instance_id)) {
 139           break;
 140         }
 141         if (is_instance) {
 142           result = proj_in->in(TypeFunc::Memory);
 143         } else if (is_boxed_value_load) {
 144           Node* klass = alloc->in(AllocateNode::KlassNode);
 145           const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
 146           if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
 147             result = proj_in->in(TypeFunc::Memory); // not related allocation
 148           }
 149         }
 150       } else if (proj_in->is_MemBar()) {
 151         result = proj_in->in(TypeFunc::Memory);
 152       } else {
 153         assert(false, "unexpected projection");
 154       }
 155     } else if (result->is_ClearArray()) {
 156       if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
 157         // Can not bypass initialization of the instance
 158         // we are looking for.
 159         break;
 160       }
 161       // Otherwise skip it (the call updated 'result' value).
 162     } else if (result->is_MergeMem()) {
 163       result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
 164     }
 165   }
 166   return result;
 167 }
 168 
 169 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
 170   const TypeOopPtr* t_oop = t_adr->isa_oopptr();
 171   if (t_oop == NULL)
 172     return mchain;  // don't try to optimize non-oop types
 173   Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
 174   bool is_instance = t_oop->is_known_instance_field();
 175   PhaseIterGVN *igvn = phase->is_IterGVN();
 176   if (is_instance && igvn != NULL  && result->is_Phi()) {
 177     PhiNode *mphi = result->as_Phi();
 178     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 179     const TypePtr *t = mphi->adr_type();
 180     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
 181         t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
 182         t->is_oopptr()->cast_to_exactness(true)
 183          ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 184          ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
 185       // clone the Phi with our address type
 186       result = mphi->split_out_instance(t_adr, igvn);
 187     } else {
 188       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 189     }
 190   }
 191   return result;
 192 }
 193 
 194 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 195   uint alias_idx = phase->C->get_alias_index(tp);
 196   Node *mem = mmem;
 197 #ifdef ASSERT
 198   {
 199     // Check that current type is consistent with the alias index used during graph construction
 200     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 201     bool consistent =  adr_check == NULL || adr_check->empty() ||
 202                        phase->C->must_alias(adr_check, alias_idx );
 203     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 204     if( !consistent && adr_check != NULL && !adr_check->empty() &&
 205                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 206         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 207         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 208           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 209           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 210       // don't assert if it is dead code.
 211       consistent = true;
 212     }
 213     if( !consistent ) {
 214       st->print("alias_idx==%d, adr_check==", alias_idx);
 215       if( adr_check == NULL ) {
 216         st->print("NULL");
 217       } else {
 218         adr_check->dump();
 219       }
 220       st->cr();
 221       print_alias_types();
 222       assert(consistent, "adr_check must match alias idx");
 223     }
 224   }
 225 #endif
 226   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
 227   // means an array I have not precisely typed yet.  Do not do any
 228   // alias stuff with it any time soon.
 229   const TypeOopPtr *toop = tp->isa_oopptr();
 230   if( tp->base() != Type::AnyPtr &&
 231       !(toop &&
 232         toop->klass() != NULL &&
 233         toop->klass()->is_java_lang_Object() &&
 234         toop->offset() == Type::OffsetBot) ) {
 235     // compress paths and change unreachable cycles to TOP
 236     // If not, we can update the input infinitely along a MergeMem cycle
 237     // Equivalent code in PhiNode::Ideal
 238     Node* m  = phase->transform(mmem);
 239     // If transformed to a MergeMem, get the desired slice
 240     // Otherwise the returned node represents memory for every slice
 241     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
 242     // Update input if it is progress over what we have now
 243   }
 244   return mem;
 245 }
 246 
 247 //--------------------------Ideal_common---------------------------------------
 248 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
 249 // Unhook non-raw memories from complete (macro-expanded) initializations.
 250 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
 251   // If our control input is a dead region, kill all below the region
 252   Node *ctl = in(MemNode::Control);
 253   if (ctl && remove_dead_region(phase, can_reshape))
 254     return this;
 255   ctl = in(MemNode::Control);
 256   // Don't bother trying to transform a dead node
 257   if (ctl && ctl->is_top())  return NodeSentinel;
 258 
 259   PhaseIterGVN *igvn = phase->is_IterGVN();
 260   // Wait if control on the worklist.
 261   if (ctl && can_reshape && igvn != NULL) {
 262     Node* bol = NULL;
 263     Node* cmp = NULL;
 264     if (ctl->in(0)->is_If()) {
 265       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
 266       bol = ctl->in(0)->in(1);
 267       if (bol->is_Bool())
 268         cmp = ctl->in(0)->in(1)->in(1);
 269     }
 270     if (igvn->_worklist.member(ctl) ||
 271         (bol != NULL && igvn->_worklist.member(bol)) ||
 272         (cmp != NULL && igvn->_worklist.member(cmp)) ) {
 273       // This control path may be dead.
 274       // Delay this memory node transformation until the control is processed.
 275       phase->is_IterGVN()->_worklist.push(this);
 276       return NodeSentinel; // caller will return NULL
 277     }
 278   }
 279   // Ignore if memory is dead, or self-loop
 280   Node *mem = in(MemNode::Memory);
 281   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
 282   assert(mem != this, "dead loop in MemNode::Ideal");
 283 
 284   if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
 285     // This memory slice may be dead.
 286     // Delay this mem node transformation until the memory is processed.
 287     phase->is_IterGVN()->_worklist.push(this);
 288     return NodeSentinel; // caller will return NULL
 289   }
 290 
 291   Node *address = in(MemNode::Address);
 292   const Type *t_adr = phase->type(address);
 293   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return NULL
 294 
 295   if (can_reshape && igvn != NULL &&
 296       (igvn->_worklist.member(address) ||
 297        igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
 298     // The address's base and type may change when the address is processed.
 299     // Delay this mem node transformation until the address is processed.
 300     phase->is_IterGVN()->_worklist.push(this);
 301     return NodeSentinel; // caller will return NULL
 302   }
 303 
 304   // Do NOT remove or optimize the next lines: ensure a new alias index
 305   // is allocated for an oop pointer type before Escape Analysis.
 306   // Note: C++ will not remove it since the call has side effect.
 307   if (t_adr->isa_oopptr()) {
 308     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 309   }
 310 
 311   Node* base = NULL;
 312   if (address->is_AddP()) {
 313     base = address->in(AddPNode::Base);
 314   }
 315   if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
 316       !t_adr->isa_rawptr()) {
 317     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
 318     // Skip this node optimization if its address has TOP base.
 319     return NodeSentinel; // caller will return NULL
 320   }
 321 
 322   // Avoid independent memory operations
 323   Node* old_mem = mem;
 324 
 325   // The code which unhooks non-raw memories from complete (macro-expanded)
 326   // initializations was removed. After macro-expansion all stores catched
 327   // by Initialize node became raw stores and there is no information
 328   // which memory slices they modify. So it is unsafe to move any memory
 329   // operation above these stores. Also in most cases hooked non-raw memories
 330   // were already unhooked by using information from detect_ptr_independence()
 331   // and find_previous_store().
 332 
 333   if (mem->is_MergeMem()) {
 334     MergeMemNode* mmem = mem->as_MergeMem();
 335     const TypePtr *tp = t_adr->is_ptr();
 336 
 337     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 338   }
 339 
 340   if (mem != old_mem) {
 341     set_req(MemNode::Memory, mem);
 342     if (can_reshape && old_mem->outcnt() == 0) {
 343         igvn->_worklist.push(old_mem);
 344     }
 345     if (phase->type( mem ) == Type::TOP) return NodeSentinel;
 346     return this;
 347   }
 348 
 349   // let the subclass continue analyzing...
 350   return NULL;
 351 }
 352 
 353 // Helper function for proving some simple control dominations.
 354 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 355 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 356 // is not a constant (dominated by the method's StartNode).
 357 // Used by MemNode::find_previous_store to prove that the
 358 // control input of a memory operation predates (dominates)
 359 // an allocation it wants to look past.
 360 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
 361   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
 362     return false; // Conservative answer for dead code
 363 
 364   // Check 'dom'. Skip Proj and CatchProj nodes.
 365   dom = dom->find_exact_control(dom);
 366   if (dom == NULL || dom->is_top())
 367     return false; // Conservative answer for dead code
 368 
 369   if (dom == sub) {
 370     // For the case when, for example, 'sub' is Initialize and the original
 371     // 'dom' is Proj node of the 'sub'.
 372     return false;
 373   }
 374 
 375   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
 376     return true;
 377 
 378   // 'dom' dominates 'sub' if its control edge and control edges
 379   // of all its inputs dominate or equal to sub's control edge.
 380 
 381   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 382   // Or Region for the check in LoadNode::Ideal();
 383   // 'sub' should have sub->in(0) != NULL.
 384   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 385          sub->is_Region() || sub->is_Call(), "expecting only these nodes");
 386 
 387   // Get control edge of 'sub'.
 388   Node* orig_sub = sub;
 389   sub = sub->find_exact_control(sub->in(0));
 390   if (sub == NULL || sub->is_top())
 391     return false; // Conservative answer for dead code
 392 
 393   assert(sub->is_CFG(), "expecting control");
 394 
 395   if (sub == dom)
 396     return true;
 397 
 398   if (sub->is_Start() || sub->is_Root())
 399     return false;
 400 
 401   {
 402     // Check all control edges of 'dom'.
 403 
 404     ResourceMark rm;
 405     Arena* arena = Thread::current()->resource_area();
 406     Node_List nlist(arena);
 407     Unique_Node_List dom_list(arena);
 408 
 409     dom_list.push(dom);
 410     bool only_dominating_controls = false;
 411 
 412     for (uint next = 0; next < dom_list.size(); next++) {
 413       Node* n = dom_list.at(next);
 414       if (n == orig_sub)
 415         return false; // One of dom's inputs dominated by sub.
 416       if (!n->is_CFG() && n->pinned()) {
 417         // Check only own control edge for pinned non-control nodes.
 418         n = n->find_exact_control(n->in(0));
 419         if (n == NULL || n->is_top())
 420           return false; // Conservative answer for dead code
 421         assert(n->is_CFG(), "expecting control");
 422         dom_list.push(n);
 423       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 424         only_dominating_controls = true;
 425       } else if (n->is_CFG()) {
 426         if (n->dominates(sub, nlist))
 427           only_dominating_controls = true;
 428         else
 429           return false;
 430       } else {
 431         // First, own control edge.
 432         Node* m = n->find_exact_control(n->in(0));
 433         if (m != NULL) {
 434           if (m->is_top())
 435             return false; // Conservative answer for dead code
 436           dom_list.push(m);
 437         }
 438         // Now, the rest of edges.
 439         uint cnt = n->req();
 440         for (uint i = 1; i < cnt; i++) {
 441           m = n->find_exact_control(n->in(i));
 442           if (m == NULL || m->is_top())
 443             continue;
 444           dom_list.push(m);
 445         }
 446       }
 447     }
 448     return only_dominating_controls;
 449   }
 450 }
 451 
 452 //---------------------detect_ptr_independence---------------------------------
 453 // Used by MemNode::find_previous_store to prove that two base
 454 // pointers are never equal.
 455 // The pointers are accompanied by their associated allocations,
 456 // if any, which have been previously discovered by the caller.
 457 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 458                                       Node* p2, AllocateNode* a2,
 459                                       PhaseTransform* phase) {
 460   // Attempt to prove that these two pointers cannot be aliased.
 461   // They may both manifestly be allocations, and they should differ.
 462   // Or, if they are not both allocations, they can be distinct constants.
 463   // Otherwise, one is an allocation and the other a pre-existing value.
 464   if (a1 == NULL && a2 == NULL) {           // neither an allocation
 465     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 466   } else if (a1 != NULL && a2 != NULL) {    // both allocations
 467     return (a1 != a2);
 468   } else if (a1 != NULL) {                  // one allocation a1
 469     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 470     return all_controls_dominate(p2, a1);
 471   } else { //(a2 != NULL)                   // one allocation a2
 472     return all_controls_dominate(p1, a2);
 473   }
 474   return false;
 475 }
 476 
 477 
 478 // The logic for reordering loads and stores uses four steps:
 479 // (a) Walk carefully past stores and initializations which we
 480 //     can prove are independent of this load.
 481 // (b) Observe that the next memory state makes an exact match
 482 //     with self (load or store), and locate the relevant store.
 483 // (c) Ensure that, if we were to wire self directly to the store,
 484 //     the optimizer would fold it up somehow.
 485 // (d) Do the rewiring, and return, depending on some other part of
 486 //     the optimizer to fold up the load.
 487 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 488 // specific to loads and stores, so they are handled by the callers.
 489 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 490 //
 491 Node* MemNode::find_previous_store(PhaseTransform* phase) {
 492   Node*         ctrl   = in(MemNode::Control);
 493   Node*         adr    = in(MemNode::Address);
 494   intptr_t      offset = 0;
 495   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 496   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
 497 
 498   if (offset == Type::OffsetBot)
 499     return NULL;            // cannot unalias unless there are precise offsets
 500 
 501   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 502 
 503   intptr_t size_in_bytes = memory_size();
 504 
 505   Node* mem = in(MemNode::Memory);   // start searching here...
 506 
 507   int cnt = 50;             // Cycle limiter
 508   for (;;) {                // While we can dance past unrelated stores...
 509     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 510 
 511     if (mem->is_Store()) {
 512       Node* st_adr = mem->in(MemNode::Address);
 513       intptr_t st_offset = 0;
 514       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 515       if (st_base == NULL)
 516         break;              // inscrutable pointer
 517       if (st_offset != offset && st_offset != Type::OffsetBot) {
 518         const int MAX_STORE = BytesPerLong;
 519         if (st_offset >= offset + size_in_bytes ||
 520             st_offset <= offset - MAX_STORE ||
 521             st_offset <= offset - mem->as_Store()->memory_size()) {
 522           // Success:  The offsets are provably independent.
 523           // (You may ask, why not just test st_offset != offset and be done?
 524           // The answer is that stores of different sizes can co-exist
 525           // in the same sequence of RawMem effects.  We sometimes initialize
 526           // a whole 'tile' of array elements with a single jint or jlong.)
 527           mem = mem->in(MemNode::Memory);
 528           continue;           // (a) advance through independent store memory
 529         }
 530       }
 531       if (st_base != base &&
 532           detect_ptr_independence(base, alloc,
 533                                   st_base,
 534                                   AllocateNode::Ideal_allocation(st_base, phase),
 535                                   phase)) {
 536         // Success:  The bases are provably independent.
 537         mem = mem->in(MemNode::Memory);
 538         continue;           // (a) advance through independent store memory
 539       }
 540 
 541       // (b) At this point, if the bases or offsets do not agree, we lose,
 542       // since we have not managed to prove 'this' and 'mem' independent.
 543       if (st_base == base && st_offset == offset) {
 544         return mem;         // let caller handle steps (c), (d)
 545       }
 546 
 547     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 548       InitializeNode* st_init = mem->in(0)->as_Initialize();
 549       AllocateNode*  st_alloc = st_init->allocation();
 550       if (st_alloc == NULL)
 551         break;              // something degenerated
 552       bool known_identical = false;
 553       bool known_independent = false;
 554       if (alloc == st_alloc)
 555         known_identical = true;
 556       else if (alloc != NULL)
 557         known_independent = true;
 558       else if (all_controls_dominate(this, st_alloc))
 559         known_independent = true;
 560 
 561       if (known_independent) {
 562         // The bases are provably independent: Either they are
 563         // manifestly distinct allocations, or else the control
 564         // of this load dominates the store's allocation.
 565         int alias_idx = phase->C->get_alias_index(adr_type());
 566         if (alias_idx == Compile::AliasIdxRaw) {
 567           mem = st_alloc->in(TypeFunc::Memory);
 568         } else {
 569           mem = st_init->memory(alias_idx);
 570         }
 571         continue;           // (a) advance through independent store memory
 572       }
 573 
 574       // (b) at this point, if we are not looking at a store initializing
 575       // the same allocation we are loading from, we lose.
 576       if (known_identical) {
 577         // From caller, can_see_stored_value will consult find_captured_store.
 578         return mem;         // let caller handle steps (c), (d)
 579       }
 580 
 581     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
 582       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 583       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 584         CallNode *call = mem->in(0)->as_Call();
 585         if (!call->may_modify(addr_t, phase)) {
 586           mem = call->in(TypeFunc::Memory);
 587           continue;         // (a) advance through independent call memory
 588         }
 589       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 590         mem = mem->in(0)->in(TypeFunc::Memory);
 591         continue;           // (a) advance through independent MemBar memory
 592       } else if (mem->is_ClearArray()) {
 593         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 594           // (the call updated 'mem' value)
 595           continue;         // (a) advance through independent allocation memory
 596         } else {
 597           // Can not bypass initialization of the instance
 598           // we are looking for.
 599           return mem;
 600         }
 601       } else if (mem->is_MergeMem()) {
 602         int alias_idx = phase->C->get_alias_index(adr_type());
 603         mem = mem->as_MergeMem()->memory_at(alias_idx);
 604         continue;           // (a) advance through independent MergeMem memory
 605       }
 606     }
 607 
 608     // Unless there is an explicit 'continue', we must bail out here,
 609     // because 'mem' is an inscrutable memory state (e.g., a call).
 610     break;
 611   }
 612 
 613   return NULL;              // bail out
 614 }
 615 
 616 //----------------------calculate_adr_type-------------------------------------
 617 // Helper function.  Notices when the given type of address hits top or bottom.
 618 // Also, asserts a cross-check of the type against the expected address type.
 619 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 620   if (t == Type::TOP)  return NULL; // does not touch memory any more?
 621   #ifdef PRODUCT
 622   cross_check = NULL;
 623   #else
 624   if (!VerifyAliases || is_error_reported() || Node::in_dump())  cross_check = NULL;
 625   #endif
 626   const TypePtr* tp = t->isa_ptr();
 627   if (tp == NULL) {
 628     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 629     return TypePtr::BOTTOM;           // touches lots of memory
 630   } else {
 631     #ifdef ASSERT
 632     // %%%% [phh] We don't check the alias index if cross_check is
 633     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 634     if (cross_check != NULL &&
 635         cross_check != TypePtr::BOTTOM &&
 636         cross_check != TypeRawPtr::BOTTOM) {
 637       // Recheck the alias index, to see if it has changed (due to a bug).
 638       Compile* C = Compile::current();
 639       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 640              "must stay in the original alias category");
 641       // The type of the address must be contained in the adr_type,
 642       // disregarding "null"-ness.
 643       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 644       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 645       assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
 646              "real address must not escape from expected memory type");
 647     }
 648     #endif
 649     return tp;
 650   }
 651 }
 652 
 653 //------------------------adr_phi_is_loop_invariant----------------------------
 654 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
 655 // loop is loop invariant. Make a quick traversal of Phi and associated
 656 // CastPP nodes, looking to see if they are a closed group within the loop.
 657 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
 658   // The idea is that the phi-nest must boil down to only CastPP nodes
 659   // with the same data. This implies that any path into the loop already
 660   // includes such a CastPP, and so the original cast, whatever its input,
 661   // must be covered by an equivalent cast, with an earlier control input.
 662   ResourceMark rm;
 663 
 664   // The loop entry input of the phi should be the unique dominating
 665   // node for every Phi/CastPP in the loop.
 666   Unique_Node_List closure;
 667   closure.push(adr_phi->in(LoopNode::EntryControl));
 668 
 669   // Add the phi node and the cast to the worklist.
 670   Unique_Node_List worklist;
 671   worklist.push(adr_phi);
 672   if( cast != NULL ){
 673     if( !cast->is_ConstraintCast() ) return false;
 674     worklist.push(cast);
 675   }
 676 
 677   // Begin recursive walk of phi nodes.
 678   while( worklist.size() ){
 679     // Take a node off the worklist
 680     Node *n = worklist.pop();
 681     if( !closure.member(n) ){
 682       // Add it to the closure.
 683       closure.push(n);
 684       // Make a sanity check to ensure we don't waste too much time here.
 685       if( closure.size() > 20) return false;
 686       // This node is OK if:
 687       //  - it is a cast of an identical value
 688       //  - or it is a phi node (then we add its inputs to the worklist)
 689       // Otherwise, the node is not OK, and we presume the cast is not invariant
 690       if( n->is_ConstraintCast() ){
 691         worklist.push(n->in(1));
 692       } else if( n->is_Phi() ) {
 693         for( uint i = 1; i < n->req(); i++ ) {
 694           worklist.push(n->in(i));
 695         }
 696       } else {
 697         return false;
 698       }
 699     }
 700   }
 701 
 702   // Quit when the worklist is empty, and we've found no offending nodes.
 703   return true;
 704 }
 705 
 706 //------------------------------Ideal_DU_postCCP-------------------------------
 707 // Find any cast-away of null-ness and keep its control.  Null cast-aways are
 708 // going away in this pass and we need to make this memory op depend on the
 709 // gating null check.
 710 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
 711   return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
 712 }
 713 
 714 // I tried to leave the CastPP's in.  This makes the graph more accurate in
 715 // some sense; we get to keep around the knowledge that an oop is not-null
 716 // after some test.  Alas, the CastPP's interfere with GVN (some values are
 717 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
 718 // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed
 719 // some of the more trivial cases in the optimizer.  Removing more useless
 720 // Phi's started allowing Loads to illegally float above null checks.  I gave
 721 // up on this approach.  CNC 10/20/2000
 722 // This static method may be called not from MemNode (EncodePNode calls it).
 723 // Only the control edge of the node 'n' might be updated.
 724 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
 725   Node *skipped_cast = NULL;
 726   // Need a null check?  Regular static accesses do not because they are
 727   // from constant addresses.  Array ops are gated by the range check (which
 728   // always includes a NULL check).  Just check field ops.
 729   if( n->in(MemNode::Control) == NULL ) {
 730     // Scan upwards for the highest location we can place this memory op.
 731     while( true ) {
 732       switch( adr->Opcode() ) {
 733 
 734       case Op_AddP:             // No change to NULL-ness, so peek thru AddP's
 735         adr = adr->in(AddPNode::Base);
 736         continue;
 737 
 738       case Op_DecodeN:         // No change to NULL-ness, so peek thru
 739       case Op_DecodeNKlass:
 740         adr = adr->in(1);
 741         continue;
 742 
 743       case Op_EncodeP:
 744       case Op_EncodePKlass:
 745         // EncodeP node's control edge could be set by this method
 746         // when EncodeP node depends on CastPP node.
 747         //
 748         // Use its control edge for memory op because EncodeP may go away
 749         // later when it is folded with following or preceding DecodeN node.
 750         if (adr->in(0) == NULL) {
 751           // Keep looking for cast nodes.
 752           adr = adr->in(1);
 753           continue;
 754         }
 755         ccp->hash_delete(n);
 756         n->set_req(MemNode::Control, adr->in(0));
 757         ccp->hash_insert(n);
 758         return n;
 759 
 760       case Op_CastPP:
 761         // If the CastPP is useless, just peek on through it.
 762         if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
 763           // Remember the cast that we've peeked though. If we peek
 764           // through more than one, then we end up remembering the highest
 765           // one, that is, if in a loop, the one closest to the top.
 766           skipped_cast = adr;
 767           adr = adr->in(1);
 768           continue;
 769         }
 770         // CastPP is going away in this pass!  We need this memory op to be
 771         // control-dependent on the test that is guarding the CastPP.
 772         ccp->hash_delete(n);
 773         n->set_req(MemNode::Control, adr->in(0));
 774         ccp->hash_insert(n);
 775         return n;
 776 
 777       case Op_Phi:
 778         // Attempt to float above a Phi to some dominating point.
 779         if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
 780           // If we've already peeked through a Cast (which could have set the
 781           // control), we can't float above a Phi, because the skipped Cast
 782           // may not be loop invariant.
 783           if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
 784             adr = adr->in(1);
 785             continue;
 786           }
 787         }
 788 
 789         // Intentional fallthrough!
 790 
 791         // No obvious dominating point.  The mem op is pinned below the Phi
 792         // by the Phi itself.  If the Phi goes away (no true value is merged)
 793         // then the mem op can float, but not indefinitely.  It must be pinned
 794         // behind the controls leading to the Phi.
 795       case Op_CheckCastPP:
 796         // These usually stick around to change address type, however a
 797         // useless one can be elided and we still need to pick up a control edge
 798         if (adr->in(0) == NULL) {
 799           // This CheckCastPP node has NO control and is likely useless. But we
 800           // need check further up the ancestor chain for a control input to keep
 801           // the node in place. 4959717.
 802           skipped_cast = adr;
 803           adr = adr->in(1);
 804           continue;
 805         }
 806         ccp->hash_delete(n);
 807         n->set_req(MemNode::Control, adr->in(0));
 808         ccp->hash_insert(n);
 809         return n;
 810 
 811         // List of "safe" opcodes; those that implicitly block the memory
 812         // op below any null check.
 813       case Op_CastX2P:          // no null checks on native pointers
 814       case Op_Parm:             // 'this' pointer is not null
 815       case Op_LoadP:            // Loading from within a klass
 816       case Op_LoadN:            // Loading from within a klass
 817       case Op_LoadKlass:        // Loading from within a klass
 818       case Op_LoadNKlass:       // Loading from within a klass
 819       case Op_ConP:             // Loading from a klass
 820       case Op_ConN:             // Loading from a klass
 821       case Op_ConNKlass:        // Loading from a klass
 822       case Op_CreateEx:         // Sucking up the guts of an exception oop
 823       case Op_Con:              // Reading from TLS
 824       case Op_CMoveP:           // CMoveP is pinned
 825       case Op_CMoveN:           // CMoveN is pinned
 826         break;                  // No progress
 827 
 828       case Op_Proj:             // Direct call to an allocation routine
 829       case Op_SCMemProj:        // Memory state from store conditional ops
 830 #ifdef ASSERT
 831         {
 832           assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
 833           const Node* call = adr->in(0);
 834           if (call->is_CallJava()) {
 835             const CallJavaNode* call_java = call->as_CallJava();
 836             const TypeTuple *r = call_java->tf()->range();
 837             assert(r->cnt() > TypeFunc::Parms, "must return value");
 838             const Type* ret_type = r->field_at(TypeFunc::Parms);
 839             assert(ret_type && ret_type->isa_ptr(), "must return pointer");
 840             // We further presume that this is one of
 841             // new_instance_Java, new_array_Java, or
 842             // the like, but do not assert for this.
 843           } else if (call->is_Allocate()) {
 844             // similar case to new_instance_Java, etc.
 845           } else if (!call->is_CallLeaf()) {
 846             // Projections from fetch_oop (OSR) are allowed as well.
 847             ShouldNotReachHere();
 848           }
 849         }
 850 #endif
 851         break;
 852       default:
 853         ShouldNotReachHere();
 854       }
 855       break;
 856     }
 857   }
 858 
 859   return  NULL;               // No progress
 860 }
 861 
 862 
 863 //=============================================================================
 864 uint LoadNode::size_of() const { return sizeof(*this); }
 865 uint LoadNode::cmp( const Node &n ) const
 866 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
 867 const Type *LoadNode::bottom_type() const { return _type; }
 868 uint LoadNode::ideal_reg() const {
 869   return _type->ideal_reg();
 870 }
 871 
 872 #ifndef PRODUCT
 873 void LoadNode::dump_spec(outputStream *st) const {
 874   MemNode::dump_spec(st);
 875   if( !Verbose && !WizardMode ) {
 876     // standard dump does this in Verbose and WizardMode
 877     st->print(" #"); _type->dump_on(st);
 878   }
 879 }
 880 #endif
 881 
 882 #ifdef ASSERT
 883 //----------------------------is_immutable_value-------------------------------
 884 // Helper function to allow a raw load without control edge for some cases
 885 bool LoadNode::is_immutable_value(Node* adr) {
 886   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 887           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 888           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 889            in_bytes(JavaThread::osthread_offset())));
 890 }
 891 #endif
 892 
 893 //----------------------------LoadNode::make-----------------------------------
 894 // Polymorphic factory method:
 895 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) {
 896   Compile* C = gvn.C;
 897 
 898   // sanity check the alias category against the created node type
 899   assert(!(adr_type->isa_oopptr() &&
 900            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
 901          "use LoadKlassNode instead");
 902   assert(!(adr_type->isa_aryptr() &&
 903            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
 904          "use LoadRangeNode instead");
 905   // Check control edge of raw loads
 906   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
 907           // oop will be recorded in oop map if load crosses safepoint
 908           rt->isa_oopptr() || is_immutable_value(adr),
 909           "raw memory operations should have control edge");
 910   switch (bt) {
 911   case T_BOOLEAN: return new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(),  mo);
 912   case T_BYTE:    return new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(),  mo);
 913   case T_INT:     return new LoadINode (ctl, mem, adr, adr_type, rt->is_int(),  mo);
 914   case T_CHAR:    return new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(),  mo);
 915   case T_SHORT:   return new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(),  mo);
 916   case T_LONG:    return new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo);
 917   case T_FLOAT:   return new LoadFNode (ctl, mem, adr, adr_type, rt,            mo);
 918   case T_DOUBLE:  return new LoadDNode (ctl, mem, adr, adr_type, rt,            mo);
 919   case T_ADDRESS: return new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(),  mo);
 920   case T_OBJECT:
 921 #ifdef _LP64
 922     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 923       Node* load  = gvn.transform(new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo));
 924       return new DecodeNNode(load, load->bottom_type()->make_ptr());
 925     } else
 926 #endif
 927     {
 928       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
 929       return new LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo);
 930     }
 931   }
 932   ShouldNotReachHere();
 933   return (LoadNode*)NULL;
 934 }
 935 
 936 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
 937   bool require_atomic = true;
 938   return new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic);
 939 }
 940 
 941 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
 942   bool require_atomic = true;
 943   return new LoadDNode(ctl, mem, adr, adr_type, rt, mo, require_atomic);
 944 }
 945 
 946 
 947 
 948 //------------------------------hash-------------------------------------------
 949 uint LoadNode::hash() const {
 950   // unroll addition of interesting fields
 951   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
 952 }
 953 
 954 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
 955   if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
 956     bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
 957     bool is_stable_ary = FoldStableValues &&
 958                          (tp != NULL) && (tp->isa_aryptr() != NULL) &&
 959                          tp->isa_aryptr()->is_stable();
 960 
 961     return (eliminate_boxing && non_volatile) || is_stable_ary;
 962   }
 963 
 964   return false;
 965 }
 966 
 967 //---------------------------can_see_stored_value------------------------------
 968 // This routine exists to make sure this set of tests is done the same
 969 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
 970 // will change the graph shape in a way which makes memory alive twice at the
 971 // same time (uses the Oracle model of aliasing), then some
 972 // LoadXNode::Identity will fold things back to the equivalence-class model
 973 // of aliasing.
 974 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
 975   Node* ld_adr = in(MemNode::Address);
 976   intptr_t ld_off = 0;
 977   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 978   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
 979   Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
 980   // This is more general than load from boxing objects.
 981   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
 982     uint alias_idx = atp->index();
 983     bool final = !atp->is_rewritable();
 984     Node* result = NULL;
 985     Node* current = st;
 986     // Skip through chains of MemBarNodes checking the MergeMems for
 987     // new states for the slice of this load.  Stop once any other
 988     // kind of node is encountered.  Loads from final memory can skip
 989     // through any kind of MemBar but normal loads shouldn't skip
 990     // through MemBarAcquire since the could allow them to move out of
 991     // a synchronized region.
 992     while (current->is_Proj()) {
 993       int opc = current->in(0)->Opcode();
 994       if ((final && (opc == Op_MemBarAcquire ||
 995                      opc == Op_MemBarAcquireLock ||
 996                      opc == Op_LoadFence)) ||
 997           opc == Op_MemBarRelease ||
 998           opc == Op_StoreFence ||
 999           opc == Op_MemBarReleaseLock ||
1000           opc == Op_MemBarCPUOrder) {
1001         Node* mem = current->in(0)->in(TypeFunc::Memory);
1002         if (mem->is_MergeMem()) {
1003           MergeMemNode* merge = mem->as_MergeMem();
1004           Node* new_st = merge->memory_at(alias_idx);
1005           if (new_st == merge->base_memory()) {
1006             // Keep searching
1007             current = new_st;
1008             continue;
1009           }
1010           // Save the new memory state for the slice and fall through
1011           // to exit.
1012           result = new_st;
1013         }
1014       }
1015       break;
1016     }
1017     if (result != NULL) {
1018       st = result;
1019     }
1020   }
1021 
1022   // Loop around twice in the case Load -> Initialize -> Store.
1023   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1024   for (int trip = 0; trip <= 1; trip++) {
1025 
1026     if (st->is_Store()) {
1027       Node* st_adr = st->in(MemNode::Address);
1028       if (!phase->eqv(st_adr, ld_adr)) {
1029         // Try harder before giving up...  Match raw and non-raw pointers.
1030         intptr_t st_off = 0;
1031         AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1032         if (alloc == NULL)       return NULL;
1033         if (alloc != ld_alloc)   return NULL;
1034         if (ld_off != st_off)    return NULL;
1035         // At this point we have proven something like this setup:
1036         //  A = Allocate(...)
1037         //  L = LoadQ(,  AddP(CastPP(, A.Parm),, #Off))
1038         //  S = StoreQ(, AddP(,        A.Parm  , #Off), V)
1039         // (Actually, we haven't yet proven the Q's are the same.)
1040         // In other words, we are loading from a casted version of
1041         // the same pointer-and-offset that we stored to.
1042         // Thus, we are able to replace L by V.
1043       }
1044       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1045       if (store_Opcode() != st->Opcode())
1046         return NULL;
1047       return st->in(MemNode::ValueIn);
1048     }
1049 
1050     // A load from a freshly-created object always returns zero.
1051     // (This can happen after LoadNode::Ideal resets the load's memory input
1052     // to find_captured_store, which returned InitializeNode::zero_memory.)
1053     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1054         (st->in(0) == ld_alloc) &&
1055         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1056       // return a zero value for the load's basic type
1057       // (This is one of the few places where a generic PhaseTransform
1058       // can create new nodes.  Think of it as lazily manifesting
1059       // virtually pre-existing constants.)
1060       return phase->zerocon(memory_type());
1061     }
1062 
1063     // A load from an initialization barrier can match a captured store.
1064     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1065       InitializeNode* init = st->in(0)->as_Initialize();
1066       AllocateNode* alloc = init->allocation();
1067       if ((alloc != NULL) && (alloc == ld_alloc)) {
1068         // examine a captured store value
1069         st = init->find_captured_store(ld_off, memory_size(), phase);
1070         if (st != NULL)
1071           continue;             // take one more trip around
1072       }
1073     }
1074 
1075     // Load boxed value from result of valueOf() call is input parameter.
1076     if (this->is_Load() && ld_adr->is_AddP() &&
1077         (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1078       intptr_t ignore = 0;
1079       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1080       if (base != NULL && base->is_Proj() &&
1081           base->as_Proj()->_con == TypeFunc::Parms &&
1082           base->in(0)->is_CallStaticJava() &&
1083           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1084         return base->in(0)->in(TypeFunc::Parms);
1085       }
1086     }
1087 
1088     break;
1089   }
1090 
1091   return NULL;
1092 }
1093 
1094 //----------------------is_instance_field_load_with_local_phi------------------
1095 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1096   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1097       in(Address)->is_AddP() ) {
1098     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1099     // Only instances and boxed values.
1100     if( t_oop != NULL &&
1101         (t_oop->is_ptr_to_boxed_value() ||
1102          t_oop->is_known_instance_field()) &&
1103         t_oop->offset() != Type::OffsetBot &&
1104         t_oop->offset() != Type::OffsetTop) {
1105       return true;
1106     }
1107   }
1108   return false;
1109 }
1110 
1111 //------------------------------Identity---------------------------------------
1112 // Loads are identity if previous store is to same address
1113 Node *LoadNode::Identity( PhaseTransform *phase ) {
1114   // If the previous store-maker is the right kind of Store, and the store is
1115   // to the same address, then we are equal to the value stored.
1116   Node* mem = in(Memory);
1117   Node* value = can_see_stored_value(mem, phase);
1118   if( value ) {
1119     // byte, short & char stores truncate naturally.
1120     // A load has to load the truncated value which requires
1121     // some sort of masking operation and that requires an
1122     // Ideal call instead of an Identity call.
1123     if (memory_size() < BytesPerInt) {
1124       // If the input to the store does not fit with the load's result type,
1125       // it must be truncated via an Ideal call.
1126       if (!phase->type(value)->higher_equal(phase->type(this)))
1127         return this;
1128     }
1129     // (This works even when value is a Con, but LoadNode::Value
1130     // usually runs first, producing the singleton type of the Con.)
1131     return value;
1132   }
1133 
1134   // Search for an existing data phi which was generated before for the same
1135   // instance's field to avoid infinite generation of phis in a loop.
1136   Node *region = mem->in(0);
1137   if (is_instance_field_load_with_local_phi(region)) {
1138     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1139     int this_index  = phase->C->get_alias_index(addr_t);
1140     int this_offset = addr_t->offset();
1141     int this_iid    = addr_t->instance_id();
1142     if (!addr_t->is_known_instance() &&
1143          addr_t->is_ptr_to_boxed_value()) {
1144       // Use _idx of address base (could be Phi node) for boxed values.
1145       intptr_t   ignore = 0;
1146       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1147       this_iid = base->_idx;
1148     }
1149     const Type* this_type = bottom_type();
1150     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1151       Node* phi = region->fast_out(i);
1152       if (phi->is_Phi() && phi != mem &&
1153           phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1154         return phi;
1155       }
1156     }
1157   }
1158 
1159   return this;
1160 }
1161 
1162 // We're loading from an object which has autobox behaviour.
1163 // If this object is result of a valueOf call we'll have a phi
1164 // merging a newly allocated object and a load from the cache.
1165 // We want to replace this load with the original incoming
1166 // argument to the valueOf call.
1167 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1168   assert(phase->C->eliminate_boxing(), "sanity");
1169   intptr_t ignore = 0;
1170   Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1171   if ((base == NULL) || base->is_Phi()) {
1172     // Push the loads from the phi that comes from valueOf up
1173     // through it to allow elimination of the loads and the recovery
1174     // of the original value. It is done in split_through_phi().
1175     return NULL;
1176   } else if (base->is_Load() ||
1177              base->is_DecodeN() && base->in(1)->is_Load()) {
1178     // Eliminate the load of boxed value for integer types from the cache
1179     // array by deriving the value from the index into the array.
1180     // Capture the offset of the load and then reverse the computation.
1181 
1182     // Get LoadN node which loads a boxing object from 'cache' array.
1183     if (base->is_DecodeN()) {
1184       base = base->in(1);
1185     }
1186     if (!base->in(Address)->is_AddP()) {
1187       return NULL; // Complex address
1188     }
1189     AddPNode* address = base->in(Address)->as_AddP();
1190     Node* cache_base = address->in(AddPNode::Base);
1191     if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1192       // Get ConP node which is static 'cache' field.
1193       cache_base = cache_base->in(1);
1194     }
1195     if ((cache_base != NULL) && cache_base->is_Con()) {
1196       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1197       if ((base_type != NULL) && base_type->is_autobox_cache()) {
1198         Node* elements[4];
1199         int shift = exact_log2(type2aelembytes(T_OBJECT));
1200         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1201         if ((count >  0) && elements[0]->is_Con() &&
1202             ((count == 1) ||
1203              (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1204                              elements[1]->in(2) == phase->intcon(shift))) {
1205           ciObjArray* array = base_type->const_oop()->as_obj_array();
1206           // Fetch the box object cache[0] at the base of the array and get its value
1207           ciInstance* box = array->obj_at(0)->as_instance();
1208           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1209           assert(ik->is_box_klass(), "sanity");
1210           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1211           if (ik->nof_nonstatic_fields() == 1) {
1212             // This should be true nonstatic_field_at requires calling
1213             // nof_nonstatic_fields so check it anyway
1214             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1215             BasicType bt = c.basic_type();
1216             // Only integer types have boxing cache.
1217             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1218                    bt == T_BYTE    || bt == T_SHORT ||
1219                    bt == T_INT     || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1220             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1221             if (cache_low != (int)cache_low) {
1222               return NULL; // should not happen since cache is array indexed by value
1223             }
1224             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1225             if (offset != (int)offset) {
1226               return NULL; // should not happen since cache is array indexed by value
1227             }
1228            // Add up all the offsets making of the address of the load
1229             Node* result = elements[0];
1230             for (int i = 1; i < count; i++) {
1231               result = phase->transform(new AddXNode(result, elements[i]));
1232             }
1233             // Remove the constant offset from the address and then
1234             result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1235             // remove the scaling of the offset to recover the original index.
1236             if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1237               // Peel the shift off directly but wrap it in a dummy node
1238               // since Ideal can't return existing nodes
1239               result = new RShiftXNode(result->in(1), phase->intcon(0));
1240             } else if (result->is_Add() && result->in(2)->is_Con() &&
1241                        result->in(1)->Opcode() == Op_LShiftX &&
1242                        result->in(1)->in(2) == phase->intcon(shift)) {
1243               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1244               // but for boxing cache access we know that X<<Z will not overflow
1245               // (there is range check) so we do this optimizatrion by hand here.
1246               Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1247               result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1248             } else {
1249               result = new RShiftXNode(result, phase->intcon(shift));
1250             }
1251 #ifdef _LP64
1252             if (bt != T_LONG) {
1253               result = new ConvL2INode(phase->transform(result));
1254             }
1255 #else
1256             if (bt == T_LONG) {
1257               result = new ConvI2LNode(phase->transform(result));
1258             }
1259 #endif
1260             return result;
1261           }
1262         }
1263       }
1264     }
1265   }
1266   return NULL;
1267 }
1268 
1269 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1270   Node* region = phi->in(0);
1271   if (region == NULL) {
1272     return false; // Wait stable graph
1273   }
1274   uint cnt = phi->req();
1275   for (uint i = 1; i < cnt; i++) {
1276     Node* rc = region->in(i);
1277     if (rc == NULL || phase->type(rc) == Type::TOP)
1278       return false; // Wait stable graph
1279     Node* in = phi->in(i);
1280     if (in == NULL || phase->type(in) == Type::TOP)
1281       return false; // Wait stable graph
1282   }
1283   return true;
1284 }
1285 //------------------------------split_through_phi------------------------------
1286 // Split instance or boxed field load through Phi.
1287 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1288   Node* mem     = in(Memory);
1289   Node* address = in(Address);
1290   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1291 
1292   assert((t_oop != NULL) &&
1293          (t_oop->is_known_instance_field() ||
1294           t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1295 
1296   Compile* C = phase->C;
1297   intptr_t ignore = 0;
1298   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1299   bool base_is_phi = (base != NULL) && base->is_Phi();
1300   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1301                            (base != NULL) && (base == address->in(AddPNode::Base)) &&
1302                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1303 
1304   if (!((mem->is_Phi() || base_is_phi) &&
1305         (load_boxed_values || t_oop->is_known_instance_field()))) {
1306     return NULL; // memory is not Phi
1307   }
1308 
1309   if (mem->is_Phi()) {
1310     if (!stable_phi(mem->as_Phi(), phase)) {
1311       return NULL; // Wait stable graph
1312     }
1313     uint cnt = mem->req();
1314     // Check for loop invariant memory.
1315     if (cnt == 3) {
1316       for (uint i = 1; i < cnt; i++) {
1317         Node* in = mem->in(i);
1318         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1319         if (m == mem) {
1320           set_req(Memory, mem->in(cnt - i));
1321           return this; // made change
1322         }
1323       }
1324     }
1325   }
1326   if (base_is_phi) {
1327     if (!stable_phi(base->as_Phi(), phase)) {
1328       return NULL; // Wait stable graph
1329     }
1330     uint cnt = base->req();
1331     // Check for loop invariant memory.
1332     if (cnt == 3) {
1333       for (uint i = 1; i < cnt; i++) {
1334         if (base->in(i) == base) {
1335           return NULL; // Wait stable graph
1336         }
1337       }
1338     }
1339   }
1340 
1341   bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1342 
1343   // Split through Phi (see original code in loopopts.cpp).
1344   assert(C->have_alias_type(t_oop), "instance should have alias type");
1345 
1346   // Do nothing here if Identity will find a value
1347   // (to avoid infinite chain of value phis generation).
1348   if (!phase->eqv(this, this->Identity(phase)))
1349     return NULL;
1350 
1351   // Select Region to split through.
1352   Node* region;
1353   if (!base_is_phi) {
1354     assert(mem->is_Phi(), "sanity");
1355     region = mem->in(0);
1356     // Skip if the region dominates some control edge of the address.
1357     if (!MemNode::all_controls_dominate(address, region))
1358       return NULL;
1359   } else if (!mem->is_Phi()) {
1360     assert(base_is_phi, "sanity");
1361     region = base->in(0);
1362     // Skip if the region dominates some control edge of the memory.
1363     if (!MemNode::all_controls_dominate(mem, region))
1364       return NULL;
1365   } else if (base->in(0) != mem->in(0)) {
1366     assert(base_is_phi && mem->is_Phi(), "sanity");
1367     if (MemNode::all_controls_dominate(mem, base->in(0))) {
1368       region = base->in(0);
1369     } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1370       region = mem->in(0);
1371     } else {
1372       return NULL; // complex graph
1373     }
1374   } else {
1375     assert(base->in(0) == mem->in(0), "sanity");
1376     region = mem->in(0);
1377   }
1378 
1379   const Type* this_type = this->bottom_type();
1380   int this_index  = C->get_alias_index(t_oop);
1381   int this_offset = t_oop->offset();
1382   int this_iid    = t_oop->instance_id();
1383   if (!t_oop->is_known_instance() && load_boxed_values) {
1384     // Use _idx of address base for boxed values.
1385     this_iid = base->_idx;
1386   }
1387   PhaseIterGVN* igvn = phase->is_IterGVN();
1388   Node* phi = new PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1389   for (uint i = 1; i < region->req(); i++) {
1390     Node* x;
1391     Node* the_clone = NULL;
1392     if (region->in(i) == C->top()) {
1393       x = C->top();      // Dead path?  Use a dead data op
1394     } else {
1395       x = this->clone();        // Else clone up the data op
1396       the_clone = x;            // Remember for possible deletion.
1397       // Alter data node to use pre-phi inputs
1398       if (this->in(0) == region) {
1399         x->set_req(0, region->in(i));
1400       } else {
1401         x->set_req(0, NULL);
1402       }
1403       if (mem->is_Phi() && (mem->in(0) == region)) {
1404         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1405       }
1406       if (address->is_Phi() && address->in(0) == region) {
1407         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1408       }
1409       if (base_is_phi && (base->in(0) == region)) {
1410         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1411         Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1412         x->set_req(Address, adr_x);
1413       }
1414     }
1415     // Check for a 'win' on some paths
1416     const Type *t = x->Value(igvn);
1417 
1418     bool singleton = t->singleton();
1419 
1420     // See comments in PhaseIdealLoop::split_thru_phi().
1421     if (singleton && t == Type::TOP) {
1422       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1423     }
1424 
1425     if (singleton) {
1426       x = igvn->makecon(t);
1427     } else {
1428       // We now call Identity to try to simplify the cloned node.
1429       // Note that some Identity methods call phase->type(this).
1430       // Make sure that the type array is big enough for
1431       // our new node, even though we may throw the node away.
1432       // (This tweaking with igvn only works because x is a new node.)
1433       igvn->set_type(x, t);
1434       // If x is a TypeNode, capture any more-precise type permanently into Node
1435       // otherwise it will be not updated during igvn->transform since
1436       // igvn->type(x) is set to x->Value() already.
1437       x->raise_bottom_type(t);
1438       Node *y = x->Identity(igvn);
1439       if (y != x) {
1440         x = y;
1441       } else {
1442         y = igvn->hash_find_insert(x);
1443         if (y) {
1444           x = y;
1445         } else {
1446           // Else x is a new node we are keeping
1447           // We do not need register_new_node_with_optimizer
1448           // because set_type has already been called.
1449           igvn->_worklist.push(x);
1450         }
1451       }
1452     }
1453     if (x != the_clone && the_clone != NULL) {
1454       igvn->remove_dead_node(the_clone);
1455     }
1456     phi->set_req(i, x);
1457   }
1458   // Record Phi
1459   igvn->register_new_node_with_optimizer(phi);
1460   return phi;
1461 }
1462 
1463 //------------------------------Ideal------------------------------------------
1464 // If the load is from Field memory and the pointer is non-null, we can
1465 // zero out the control input.
1466 // If the offset is constant and the base is an object allocation,
1467 // try to hook me up to the exact initializing store.
1468 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1469   Node* p = MemNode::Ideal_common(phase, can_reshape);
1470   if (p)  return (p == NodeSentinel) ? NULL : p;
1471 
1472   Node* ctrl    = in(MemNode::Control);
1473   Node* address = in(MemNode::Address);
1474   bool progress = false;
1475 
1476   // Skip up past a SafePoint control.  Cannot do this for Stores because
1477   // pointer stores & cardmarks must stay on the same side of a SafePoint.
1478   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1479       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1480     ctrl = ctrl->in(0);
1481     set_req(MemNode::Control,ctrl);
1482     progress = true;
1483   }
1484 
1485   intptr_t ignore = 0;
1486   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1487   if (base != NULL
1488       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1489     // Check for useless control edge in some common special cases
1490     if (in(MemNode::Control) != NULL
1491         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1492         && all_controls_dominate(base, phase->C->start())) {
1493       // A method-invariant, non-null address (constant or 'this' argument).
1494       set_req(MemNode::Control, NULL);
1495       progress = true;
1496     }
1497   }
1498 
1499   Node* mem = in(MemNode::Memory);
1500   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1501 
1502   if (can_reshape && (addr_t != NULL)) {
1503     // try to optimize our memory input
1504     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1505     if (opt_mem != mem) {
1506       set_req(MemNode::Memory, opt_mem);
1507       if (phase->type( opt_mem ) == Type::TOP) return NULL;
1508       return this;
1509     }
1510     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1511     if ((t_oop != NULL) &&
1512         (t_oop->is_known_instance_field() ||
1513          t_oop->is_ptr_to_boxed_value())) {
1514       PhaseIterGVN *igvn = phase->is_IterGVN();
1515       if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1516         // Delay this transformation until memory Phi is processed.
1517         phase->is_IterGVN()->_worklist.push(this);
1518         return NULL;
1519       }
1520       // Split instance field load through Phi.
1521       Node* result = split_through_phi(phase);
1522       if (result != NULL) return result;
1523 
1524       if (t_oop->is_ptr_to_boxed_value()) {
1525         Node* result = eliminate_autobox(phase);
1526         if (result != NULL) return result;
1527       }
1528     }
1529   }
1530 
1531   // Check for prior store with a different base or offset; make Load
1532   // independent.  Skip through any number of them.  Bail out if the stores
1533   // are in an endless dead cycle and report no progress.  This is a key
1534   // transform for Reflection.  However, if after skipping through the Stores
1535   // we can't then fold up against a prior store do NOT do the transform as
1536   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
1537   // array memory alive twice: once for the hoisted Load and again after the
1538   // bypassed Store.  This situation only works if EVERYBODY who does
1539   // anti-dependence work knows how to bypass.  I.e. we need all
1540   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
1541   // the alias index stuff.  So instead, peek through Stores and IFF we can
1542   // fold up, do so.
1543   Node* prev_mem = find_previous_store(phase);
1544   // Steps (a), (b):  Walk past independent stores to find an exact match.
1545   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1546     // (c) See if we can fold up on the spot, but don't fold up here.
1547     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1548     // just return a prior value, which is done by Identity calls.
1549     if (can_see_stored_value(prev_mem, phase)) {
1550       // Make ready for step (d):
1551       set_req(MemNode::Memory, prev_mem);
1552       return this;
1553     }
1554   }
1555 
1556   return progress ? this : NULL;
1557 }
1558 
1559 // Helper to recognize certain Klass fields which are invariant across
1560 // some group of array types (e.g., int[] or all T[] where T < Object).
1561 const Type*
1562 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1563                                  ciKlass* klass) const {
1564   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1565     // The field is Klass::_modifier_flags.  Return its (constant) value.
1566     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1567     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1568     return TypeInt::make(klass->modifier_flags());
1569   }
1570   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1571     // The field is Klass::_access_flags.  Return its (constant) value.
1572     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1573     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1574     return TypeInt::make(klass->access_flags());
1575   }
1576   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1577     // The field is Klass::_layout_helper.  Return its constant value if known.
1578     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1579     return TypeInt::make(klass->layout_helper());
1580   }
1581 
1582   // No match.
1583   return NULL;
1584 }
1585 
1586 // Try to constant-fold a stable array element.
1587 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1588   assert(ary->const_oop(), "array should be constant");
1589   assert(ary->is_stable(), "array should be stable");
1590 
1591   // Decode the results of GraphKit::array_element_address.
1592   ciArray* aobj = ary->const_oop()->as_array();
1593   ciConstant con = aobj->element_value_by_offset(off);
1594 
1595   if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1596     const Type* con_type = Type::make_from_constant(con);
1597     if (con_type != NULL) {
1598       if (con_type->isa_aryptr()) {
1599         // Join with the array element type, in case it is also stable.
1600         int dim = ary->stable_dimension();
1601         con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1602       }
1603       if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1604         con_type = con_type->make_narrowoop();
1605       }
1606 #ifndef PRODUCT
1607       if (TraceIterativeGVN) {
1608         tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1609         con_type->dump(); tty->cr();
1610       }
1611 #endif //PRODUCT
1612       return con_type;
1613     }
1614   }
1615   return NULL;
1616 }
1617 
1618 //------------------------------Value-----------------------------------------
1619 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1620   // Either input is TOP ==> the result is TOP
1621   Node* mem = in(MemNode::Memory);
1622   const Type *t1 = phase->type(mem);
1623   if (t1 == Type::TOP)  return Type::TOP;
1624   Node* adr = in(MemNode::Address);
1625   const TypePtr* tp = phase->type(adr)->isa_ptr();
1626   if (tp == NULL || tp->empty())  return Type::TOP;
1627   int off = tp->offset();
1628   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1629   Compile* C = phase->C;
1630 
1631   // Try to guess loaded type from pointer type
1632   if (tp->isa_aryptr()) {
1633     const TypeAryPtr* ary = tp->is_aryptr();
1634     const Type* t = ary->elem();
1635 
1636     // Determine whether the reference is beyond the header or not, by comparing
1637     // the offset against the offset of the start of the array's data.
1638     // Different array types begin at slightly different offsets (12 vs. 16).
1639     // We choose T_BYTE as an example base type that is least restrictive
1640     // as to alignment, which will therefore produce the smallest
1641     // possible base offset.
1642     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1643     const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1644 
1645     // Try to constant-fold a stable array element.
1646     if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) {
1647       // Make sure the reference is not into the header and the offset is constant
1648       if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1649         const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1650         if (con_type != NULL) {
1651           return con_type;
1652         }
1653       }
1654     }
1655 
1656     // Don't do this for integer types. There is only potential profit if
1657     // the element type t is lower than _type; that is, for int types, if _type is
1658     // more restrictive than t.  This only happens here if one is short and the other
1659     // char (both 16 bits), and in those cases we've made an intentional decision
1660     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1661     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1662     //
1663     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1664     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
1665     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1666     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
1667     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1668     // In fact, that could have been the original type of p1, and p1 could have
1669     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1670     // expression (LShiftL quux 3) independently optimized to the constant 8.
1671     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1672         && (_type->isa_vect() == NULL)
1673         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1674       // t might actually be lower than _type, if _type is a unique
1675       // concrete subclass of abstract class t.
1676       if (off_beyond_header) {  // is the offset beyond the header?
1677         const Type* jt = t->join_speculative(_type);
1678         // In any case, do not allow the join, per se, to empty out the type.
1679         if (jt->empty() && !t->empty()) {
1680           // This can happen if a interface-typed array narrows to a class type.
1681           jt = _type;
1682         }
1683 #ifdef ASSERT
1684         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1685           // The pointers in the autobox arrays are always non-null
1686           Node* base = adr->in(AddPNode::Base);
1687           if ((base != NULL) && base->is_DecodeN()) {
1688             // Get LoadN node which loads IntegerCache.cache field
1689             base = base->in(1);
1690           }
1691           if ((base != NULL) && base->is_Con()) {
1692             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1693             if ((base_type != NULL) && base_type->is_autobox_cache()) {
1694               // It could be narrow oop
1695               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1696             }
1697           }
1698         }
1699 #endif
1700         return jt;
1701       }
1702     }
1703   } else if (tp->base() == Type::InstPtr) {
1704     ciEnv* env = C->env();
1705     const TypeInstPtr* tinst = tp->is_instptr();
1706     ciKlass* klass = tinst->klass();
1707     assert( off != Type::OffsetBot ||
1708             // arrays can be cast to Objects
1709             tp->is_oopptr()->klass()->is_java_lang_Object() ||
1710             // unsafe field access may not have a constant offset
1711             C->has_unsafe_access(),
1712             "Field accesses must be precise" );
1713     // For oop loads, we expect the _type to be precise
1714     if (klass == env->String_klass() &&
1715         adr->is_AddP() && off != Type::OffsetBot) {
1716       // For constant Strings treat the final fields as compile time constants.
1717       Node* base = adr->in(AddPNode::Base);
1718       const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1719       if (t != NULL && t->singleton()) {
1720         ciField* field = env->String_klass()->get_field_by_offset(off, false);
1721         if (field != NULL && field->is_final()) {
1722           ciObject* string = t->const_oop();
1723           ciConstant constant = string->as_instance()->field_value(field);
1724           if (constant.basic_type() == T_INT) {
1725             return TypeInt::make(constant.as_int());
1726           } else if (constant.basic_type() == T_ARRAY) {
1727             if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1728               return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1729             } else {
1730               return TypeOopPtr::make_from_constant(constant.as_object(), true);
1731             }
1732           }
1733         }
1734       }
1735     }
1736     // Optimizations for constant objects
1737     ciObject* const_oop = tinst->const_oop();
1738     if (const_oop != NULL) {
1739       // For constant Boxed value treat the target field as a compile time constant.
1740       if (tinst->is_ptr_to_boxed_value()) {
1741         return tinst->get_const_boxed_value();
1742       } else
1743       // For constant CallSites treat the target field as a compile time constant.
1744       if (const_oop->is_call_site()) {
1745         ciCallSite* call_site = const_oop->as_call_site();
1746         ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1747         if (field != NULL && field->is_call_site_target()) {
1748           ciMethodHandle* target = call_site->get_target();
1749           if (target != NULL) {  // just in case
1750             ciConstant constant(T_OBJECT, target);
1751             const Type* t;
1752             if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1753               t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1754             } else {
1755               t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1756             }
1757             // Add a dependence for invalidation of the optimization.
1758             if (!call_site->is_constant_call_site()) {
1759               C->dependencies()->assert_call_site_target_value(call_site, target);
1760             }
1761             return t;
1762           }
1763         }
1764       }
1765     }
1766   } else if (tp->base() == Type::KlassPtr) {
1767     assert( off != Type::OffsetBot ||
1768             // arrays can be cast to Objects
1769             tp->is_klassptr()->klass()->is_java_lang_Object() ||
1770             // also allow array-loading from the primary supertype
1771             // array during subtype checks
1772             Opcode() == Op_LoadKlass,
1773             "Field accesses must be precise" );
1774     // For klass/static loads, we expect the _type to be precise
1775   }
1776 
1777   const TypeKlassPtr *tkls = tp->isa_klassptr();
1778   if (tkls != NULL && !StressReflectiveCode) {
1779     ciKlass* klass = tkls->klass();
1780     if (klass->is_loaded() && tkls->klass_is_exact()) {
1781       // We are loading a field from a Klass metaobject whose identity
1782       // is known at compile time (the type is "exact" or "precise").
1783       // Check for fields we know are maintained as constants by the VM.
1784       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1785         // The field is Klass::_super_check_offset.  Return its (constant) value.
1786         // (Folds up type checking code.)
1787         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1788         return TypeInt::make(klass->super_check_offset());
1789       }
1790       // Compute index into primary_supers array
1791       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1792       // Check for overflowing; use unsigned compare to handle the negative case.
1793       if( depth < ciKlass::primary_super_limit() ) {
1794         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1795         // (Folds up type checking code.)
1796         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1797         ciKlass *ss = klass->super_of_depth(depth);
1798         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1799       }
1800       const Type* aift = load_array_final_field(tkls, klass);
1801       if (aift != NULL)  return aift;
1802       if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1803         // The field is Klass::_java_mirror.  Return its (constant) value.
1804         // (Folds up the 2nd indirection in anObjConstant.getClass().)
1805         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1806         return TypeInstPtr::make(klass->java_mirror());
1807       }
1808     }
1809 
1810     // We can still check if we are loading from the primary_supers array at a
1811     // shallow enough depth.  Even though the klass is not exact, entries less
1812     // than or equal to its super depth are correct.
1813     if (klass->is_loaded() ) {
1814       ciType *inner = klass;
1815       while( inner->is_obj_array_klass() )
1816         inner = inner->as_obj_array_klass()->base_element_type();
1817       if( inner->is_instance_klass() &&
1818           !inner->as_instance_klass()->flags().is_interface() ) {
1819         // Compute index into primary_supers array
1820         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1821         // Check for overflowing; use unsigned compare to handle the negative case.
1822         if( depth < ciKlass::primary_super_limit() &&
1823             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1824           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1825           // (Folds up type checking code.)
1826           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1827           ciKlass *ss = klass->super_of_depth(depth);
1828           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1829         }
1830       }
1831     }
1832 
1833     // If the type is enough to determine that the thing is not an array,
1834     // we can give the layout_helper a positive interval type.
1835     // This will help short-circuit some reflective code.
1836     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1837         && !klass->is_array_klass() // not directly typed as an array
1838         && !klass->is_interface()  // specifically not Serializable & Cloneable
1839         && !klass->is_java_lang_Object()   // not the supertype of all T[]
1840         ) {
1841       // Note:  When interfaces are reliable, we can narrow the interface
1842       // test to (klass != Serializable && klass != Cloneable).
1843       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1844       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1845       // The key property of this type is that it folds up tests
1846       // for array-ness, since it proves that the layout_helper is positive.
1847       // Thus, a generic value like the basic object layout helper works fine.
1848       return TypeInt::make(min_size, max_jint, Type::WidenMin);
1849     }
1850   }
1851 
1852   // If we are loading from a freshly-allocated object, produce a zero,
1853   // if the load is provably beyond the header of the object.
1854   // (Also allow a variable load from a fresh array to produce zero.)
1855   const TypeOopPtr *tinst = tp->isa_oopptr();
1856   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1857   bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1858   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1859     Node* value = can_see_stored_value(mem,phase);
1860     if (value != NULL && value->is_Con()) {
1861       assert(value->bottom_type()->higher_equal(_type),"sanity");
1862       return value->bottom_type();
1863     }
1864   }
1865 
1866   if (is_instance) {
1867     // If we have an instance type and our memory input is the
1868     // programs's initial memory state, there is no matching store,
1869     // so just return a zero of the appropriate type
1870     Node *mem = in(MemNode::Memory);
1871     if (mem->is_Parm() && mem->in(0)->is_Start()) {
1872       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1873       return Type::get_zero_type(_type->basic_type());
1874     }
1875   }
1876   return _type;
1877 }
1878 
1879 //------------------------------match_edge-------------------------------------
1880 // Do we Match on this edge index or not?  Match only the address.
1881 uint LoadNode::match_edge(uint idx) const {
1882   return idx == MemNode::Address;
1883 }
1884 
1885 //--------------------------LoadBNode::Ideal--------------------------------------
1886 //
1887 //  If the previous store is to the same address as this load,
1888 //  and the value stored was larger than a byte, replace this load
1889 //  with the value stored truncated to a byte.  If no truncation is
1890 //  needed, the replacement is done in LoadNode::Identity().
1891 //
1892 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1893   Node* mem = in(MemNode::Memory);
1894   Node* value = can_see_stored_value(mem,phase);
1895   if( value && !phase->type(value)->higher_equal( _type ) ) {
1896     Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1897     return new RShiftINode(result, phase->intcon(24));
1898   }
1899   // Identity call will handle the case where truncation is not needed.
1900   return LoadNode::Ideal(phase, can_reshape);
1901 }
1902 
1903 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1904   Node* mem = in(MemNode::Memory);
1905   Node* value = can_see_stored_value(mem,phase);
1906   if (value != NULL && value->is_Con() &&
1907       !value->bottom_type()->higher_equal(_type)) {
1908     // If the input to the store does not fit with the load's result type,
1909     // it must be truncated. We can't delay until Ideal call since
1910     // a singleton Value is needed for split_thru_phi optimization.
1911     int con = value->get_int();
1912     return TypeInt::make((con << 24) >> 24);
1913   }
1914   return LoadNode::Value(phase);
1915 }
1916 
1917 //--------------------------LoadUBNode::Ideal-------------------------------------
1918 //
1919 //  If the previous store is to the same address as this load,
1920 //  and the value stored was larger than a byte, replace this load
1921 //  with the value stored truncated to a byte.  If no truncation is
1922 //  needed, the replacement is done in LoadNode::Identity().
1923 //
1924 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1925   Node* mem = in(MemNode::Memory);
1926   Node* value = can_see_stored_value(mem, phase);
1927   if (value && !phase->type(value)->higher_equal(_type))
1928     return new AndINode(value, phase->intcon(0xFF));
1929   // Identity call will handle the case where truncation is not needed.
1930   return LoadNode::Ideal(phase, can_reshape);
1931 }
1932 
1933 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1934   Node* mem = in(MemNode::Memory);
1935   Node* value = can_see_stored_value(mem,phase);
1936   if (value != NULL && value->is_Con() &&
1937       !value->bottom_type()->higher_equal(_type)) {
1938     // If the input to the store does not fit with the load's result type,
1939     // it must be truncated. We can't delay until Ideal call since
1940     // a singleton Value is needed for split_thru_phi optimization.
1941     int con = value->get_int();
1942     return TypeInt::make(con & 0xFF);
1943   }
1944   return LoadNode::Value(phase);
1945 }
1946 
1947 //--------------------------LoadUSNode::Ideal-------------------------------------
1948 //
1949 //  If the previous store is to the same address as this load,
1950 //  and the value stored was larger than a char, replace this load
1951 //  with the value stored truncated to a char.  If no truncation is
1952 //  needed, the replacement is done in LoadNode::Identity().
1953 //
1954 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1955   Node* mem = in(MemNode::Memory);
1956   Node* value = can_see_stored_value(mem,phase);
1957   if( value && !phase->type(value)->higher_equal( _type ) )
1958     return new AndINode(value,phase->intcon(0xFFFF));
1959   // Identity call will handle the case where truncation is not needed.
1960   return LoadNode::Ideal(phase, can_reshape);
1961 }
1962 
1963 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1964   Node* mem = in(MemNode::Memory);
1965   Node* value = can_see_stored_value(mem,phase);
1966   if (value != NULL && value->is_Con() &&
1967       !value->bottom_type()->higher_equal(_type)) {
1968     // If the input to the store does not fit with the load's result type,
1969     // it must be truncated. We can't delay until Ideal call since
1970     // a singleton Value is needed for split_thru_phi optimization.
1971     int con = value->get_int();
1972     return TypeInt::make(con & 0xFFFF);
1973   }
1974   return LoadNode::Value(phase);
1975 }
1976 
1977 //--------------------------LoadSNode::Ideal--------------------------------------
1978 //
1979 //  If the previous store is to the same address as this load,
1980 //  and the value stored was larger than a short, replace this load
1981 //  with the value stored truncated to a short.  If no truncation is
1982 //  needed, the replacement is done in LoadNode::Identity().
1983 //
1984 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1985   Node* mem = in(MemNode::Memory);
1986   Node* value = can_see_stored_value(mem,phase);
1987   if( value && !phase->type(value)->higher_equal( _type ) ) {
1988     Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
1989     return new RShiftINode(result, phase->intcon(16));
1990   }
1991   // Identity call will handle the case where truncation is not needed.
1992   return LoadNode::Ideal(phase, can_reshape);
1993 }
1994 
1995 const Type* LoadSNode::Value(PhaseTransform *phase) const {
1996   Node* mem = in(MemNode::Memory);
1997   Node* value = can_see_stored_value(mem,phase);
1998   if (value != NULL && value->is_Con() &&
1999       !value->bottom_type()->higher_equal(_type)) {
2000     // If the input to the store does not fit with the load's result type,
2001     // it must be truncated. We can't delay until Ideal call since
2002     // a singleton Value is needed for split_thru_phi optimization.
2003     int con = value->get_int();
2004     return TypeInt::make((con << 16) >> 16);
2005   }
2006   return LoadNode::Value(phase);
2007 }
2008 
2009 //=============================================================================
2010 //----------------------------LoadKlassNode::make------------------------------
2011 // Polymorphic factory method:
2012 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
2013   Node *ctl = NULL;
2014   // sanity check the alias category against the created node type
2015   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2016   assert(adr_type != NULL, "expecting TypeKlassPtr");
2017 #ifdef _LP64
2018   if (adr_type->is_ptr_to_narrowklass()) {
2019     assert(UseCompressedClassPointers, "no compressed klasses");
2020     Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2021     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2022   }
2023 #endif
2024   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2025   return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2026 }
2027 
2028 //------------------------------Value------------------------------------------
2029 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
2030   return klass_value_common(phase);
2031 }
2032 
2033 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
2034   // Either input is TOP ==> the result is TOP
2035   const Type *t1 = phase->type( in(MemNode::Memory) );
2036   if (t1 == Type::TOP)  return Type::TOP;
2037   Node *adr = in(MemNode::Address);
2038   const Type *t2 = phase->type( adr );
2039   if (t2 == Type::TOP)  return Type::TOP;
2040   const TypePtr *tp = t2->is_ptr();
2041   if (TypePtr::above_centerline(tp->ptr()) ||
2042       tp->ptr() == TypePtr::Null)  return Type::TOP;
2043 
2044   // Return a more precise klass, if possible
2045   const TypeInstPtr *tinst = tp->isa_instptr();
2046   if (tinst != NULL) {
2047     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2048     int offset = tinst->offset();
2049     if (ik == phase->C->env()->Class_klass()
2050         && (offset == java_lang_Class::klass_offset_in_bytes() ||
2051             offset == java_lang_Class::array_klass_offset_in_bytes())) {
2052       // We are loading a special hidden field from a Class mirror object,
2053       // the field which points to the VM's Klass metaobject.
2054       ciType* t = tinst->java_mirror_type();
2055       // java_mirror_type returns non-null for compile-time Class constants.
2056       if (t != NULL) {
2057         // constant oop => constant klass
2058         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2059           if (t->is_void()) {
2060             // We cannot create a void array.  Since void is a primitive type return null
2061             // klass.  Users of this result need to do a null check on the returned klass.
2062             return TypePtr::NULL_PTR;
2063           }
2064           return TypeKlassPtr::make(ciArrayKlass::make(t));
2065         }
2066         if (!t->is_klass()) {
2067           // a primitive Class (e.g., int.class) has NULL for a klass field
2068           return TypePtr::NULL_PTR;
2069         }
2070         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2071         return TypeKlassPtr::make(t->as_klass());
2072       }
2073       // non-constant mirror, so we can't tell what's going on
2074     }
2075     if( !ik->is_loaded() )
2076       return _type;             // Bail out if not loaded
2077     if (offset == oopDesc::klass_offset_in_bytes()) {
2078       if (tinst->klass_is_exact()) {
2079         return TypeKlassPtr::make(ik);
2080       }
2081       // See if we can become precise: no subklasses and no interface
2082       // (Note:  We need to support verified interfaces.)
2083       if (!ik->is_interface() && !ik->has_subklass()) {
2084         //assert(!UseExactTypes, "this code should be useless with exact types");
2085         // Add a dependence; if any subclass added we need to recompile
2086         if (!ik->is_final()) {
2087           // %%% should use stronger assert_unique_concrete_subtype instead
2088           phase->C->dependencies()->assert_leaf_type(ik);
2089         }
2090         // Return precise klass
2091         return TypeKlassPtr::make(ik);
2092       }
2093 
2094       // Return root of possible klass
2095       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2096     }
2097   }
2098 
2099   // Check for loading klass from an array
2100   const TypeAryPtr *tary = tp->isa_aryptr();
2101   if( tary != NULL ) {
2102     ciKlass *tary_klass = tary->klass();
2103     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
2104         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2105       if (tary->klass_is_exact()) {
2106         return TypeKlassPtr::make(tary_klass);
2107       }
2108       ciArrayKlass *ak = tary->klass()->as_array_klass();
2109       // If the klass is an object array, we defer the question to the
2110       // array component klass.
2111       if( ak->is_obj_array_klass() ) {
2112         assert( ak->is_loaded(), "" );
2113         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2114         if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2115           ciInstanceKlass* ik = base_k->as_instance_klass();
2116           // See if we can become precise: no subklasses and no interface
2117           if (!ik->is_interface() && !ik->has_subklass()) {
2118             //assert(!UseExactTypes, "this code should be useless with exact types");
2119             // Add a dependence; if any subclass added we need to recompile
2120             if (!ik->is_final()) {
2121               phase->C->dependencies()->assert_leaf_type(ik);
2122             }
2123             // Return precise array klass
2124             return TypeKlassPtr::make(ak);
2125           }
2126         }
2127         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2128       } else {                  // Found a type-array?
2129         //assert(!UseExactTypes, "this code should be useless with exact types");
2130         assert( ak->is_type_array_klass(), "" );
2131         return TypeKlassPtr::make(ak); // These are always precise
2132       }
2133     }
2134   }
2135 
2136   // Check for loading klass from an array klass
2137   const TypeKlassPtr *tkls = tp->isa_klassptr();
2138   if (tkls != NULL && !StressReflectiveCode) {
2139     ciKlass* klass = tkls->klass();
2140     if( !klass->is_loaded() )
2141       return _type;             // Bail out if not loaded
2142     if( klass->is_obj_array_klass() &&
2143         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2144       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2145       // // Always returning precise element type is incorrect,
2146       // // e.g., element type could be object and array may contain strings
2147       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2148 
2149       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2150       // according to the element type's subclassing.
2151       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2152     }
2153     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2154         tkls->offset() == in_bytes(Klass::super_offset())) {
2155       ciKlass* sup = klass->as_instance_klass()->super();
2156       // The field is Klass::_super.  Return its (constant) value.
2157       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2158       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2159     }
2160   }
2161 
2162   // Bailout case
2163   return LoadNode::Value(phase);
2164 }
2165 
2166 //------------------------------Identity---------------------------------------
2167 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2168 // Also feed through the klass in Allocate(...klass...)._klass.
2169 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2170   return klass_identity_common(phase);
2171 }
2172 
2173 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2174   Node* x = LoadNode::Identity(phase);
2175   if (x != this)  return x;
2176 
2177   // Take apart the address into an oop and and offset.
2178   // Return 'this' if we cannot.
2179   Node*    adr    = in(MemNode::Address);
2180   intptr_t offset = 0;
2181   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2182   if (base == NULL)     return this;
2183   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2184   if (toop == NULL)     return this;
2185 
2186   // We can fetch the klass directly through an AllocateNode.
2187   // This works even if the klass is not constant (clone or newArray).
2188   if (offset == oopDesc::klass_offset_in_bytes()) {
2189     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2190     if (allocated_klass != NULL) {
2191       return allocated_klass;
2192     }
2193   }
2194 
2195   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2196   // See inline_native_Class_query for occurrences of these patterns.
2197   // Java Example:  x.getClass().isAssignableFrom(y)
2198   //
2199   // This improves reflective code, often making the Class
2200   // mirror go completely dead.  (Current exception:  Class
2201   // mirrors may appear in debug info, but we could clean them out by
2202   // introducing a new debug info operator for Klass*.java_mirror).
2203   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2204       && offset == java_lang_Class::klass_offset_in_bytes()) {
2205     // We are loading a special hidden field from a Class mirror,
2206     // the field which points to its Klass or ArrayKlass metaobject.
2207     if (base->is_Load()) {
2208       Node* adr2 = base->in(MemNode::Address);
2209       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2210       if (tkls != NULL && !tkls->empty()
2211           && (tkls->klass()->is_instance_klass() ||
2212               tkls->klass()->is_array_klass())
2213           && adr2->is_AddP()
2214           ) {
2215         int mirror_field = in_bytes(Klass::java_mirror_offset());
2216         if (tkls->offset() == mirror_field) {
2217           return adr2->in(AddPNode::Base);
2218         }
2219       }
2220     }
2221   }
2222 
2223   return this;
2224 }
2225 
2226 
2227 //------------------------------Value------------------------------------------
2228 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2229   const Type *t = klass_value_common(phase);
2230   if (t == Type::TOP)
2231     return t;
2232 
2233   return t->make_narrowklass();
2234 }
2235 
2236 //------------------------------Identity---------------------------------------
2237 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2238 // Also feed through the klass in Allocate(...klass...)._klass.
2239 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2240   Node *x = klass_identity_common(phase);
2241 
2242   const Type *t = phase->type( x );
2243   if( t == Type::TOP ) return x;
2244   if( t->isa_narrowklass()) return x;
2245   assert (!t->isa_narrowoop(), "no narrow oop here");
2246 
2247   return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2248 }
2249 
2250 //------------------------------Value-----------------------------------------
2251 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2252   // Either input is TOP ==> the result is TOP
2253   const Type *t1 = phase->type( in(MemNode::Memory) );
2254   if( t1 == Type::TOP ) return Type::TOP;
2255   Node *adr = in(MemNode::Address);
2256   const Type *t2 = phase->type( adr );
2257   if( t2 == Type::TOP ) return Type::TOP;
2258   const TypePtr *tp = t2->is_ptr();
2259   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2260   const TypeAryPtr *tap = tp->isa_aryptr();
2261   if( !tap ) return _type;
2262   return tap->size();
2263 }
2264 
2265 //-------------------------------Ideal---------------------------------------
2266 // Feed through the length in AllocateArray(...length...)._length.
2267 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2268   Node* p = MemNode::Ideal_common(phase, can_reshape);
2269   if (p)  return (p == NodeSentinel) ? NULL : p;
2270 
2271   // Take apart the address into an oop and and offset.
2272   // Return 'this' if we cannot.
2273   Node*    adr    = in(MemNode::Address);
2274   intptr_t offset = 0;
2275   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2276   if (base == NULL)     return NULL;
2277   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2278   if (tary == NULL)     return NULL;
2279 
2280   // We can fetch the length directly through an AllocateArrayNode.
2281   // This works even if the length is not constant (clone or newArray).
2282   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2283     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2284     if (alloc != NULL) {
2285       Node* allocated_length = alloc->Ideal_length();
2286       Node* len = alloc->make_ideal_length(tary, phase);
2287       if (allocated_length != len) {
2288         // New CastII improves on this.
2289         return len;
2290       }
2291     }
2292   }
2293 
2294   return NULL;
2295 }
2296 
2297 //------------------------------Identity---------------------------------------
2298 // Feed through the length in AllocateArray(...length...)._length.
2299 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2300   Node* x = LoadINode::Identity(phase);
2301   if (x != this)  return x;
2302 
2303   // Take apart the address into an oop and and offset.
2304   // Return 'this' if we cannot.
2305   Node*    adr    = in(MemNode::Address);
2306   intptr_t offset = 0;
2307   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2308   if (base == NULL)     return this;
2309   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2310   if (tary == NULL)     return this;
2311 
2312   // We can fetch the length directly through an AllocateArrayNode.
2313   // This works even if the length is not constant (clone or newArray).
2314   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2315     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2316     if (alloc != NULL) {
2317       Node* allocated_length = alloc->Ideal_length();
2318       // Do not allow make_ideal_length to allocate a CastII node.
2319       Node* len = alloc->make_ideal_length(tary, phase, false);
2320       if (allocated_length == len) {
2321         // Return allocated_length only if it would not be improved by a CastII.
2322         return allocated_length;
2323       }
2324     }
2325   }
2326 
2327   return this;
2328 
2329 }
2330 
2331 //=============================================================================
2332 //---------------------------StoreNode::make-----------------------------------
2333 // Polymorphic factory method:
2334 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2335   assert((mo == unordered || mo == release), "unexpected");
2336   Compile* C = gvn.C;
2337   assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2338          ctl != NULL, "raw memory operations should have control edge");
2339 
2340   switch (bt) {
2341   case T_BOOLEAN:
2342   case T_BYTE:    return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2343   case T_INT:     return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2344   case T_CHAR:
2345   case T_SHORT:   return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2346   case T_LONG:    return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2347   case T_FLOAT:   return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2348   case T_DOUBLE:  return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2349   case T_METADATA:
2350   case T_ADDRESS:
2351   case T_OBJECT:
2352 #ifdef _LP64
2353     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2354       val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2355       return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2356     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2357                (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2358                 adr->bottom_type()->isa_rawptr())) {
2359       val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2360       return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2361     }
2362 #endif
2363     {
2364       return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2365     }
2366   }
2367   ShouldNotReachHere();
2368   return (StoreNode*)NULL;
2369 }
2370 
2371 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2372   bool require_atomic = true;
2373   return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2374 }
2375 
2376 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2377   bool require_atomic = true;
2378   return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2379 }
2380 
2381 
2382 //--------------------------bottom_type----------------------------------------
2383 const Type *StoreNode::bottom_type() const {
2384   return Type::MEMORY;
2385 }
2386 
2387 //------------------------------hash-------------------------------------------
2388 uint StoreNode::hash() const {
2389   // unroll addition of interesting fields
2390   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2391 
2392   // Since they are not commoned, do not hash them:
2393   return NO_HASH;
2394 }
2395 
2396 //------------------------------Ideal------------------------------------------
2397 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2398 // When a store immediately follows a relevant allocation/initialization,
2399 // try to capture it into the initialization, or hoist it above.
2400 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2401   Node* p = MemNode::Ideal_common(phase, can_reshape);
2402   if (p)  return (p == NodeSentinel) ? NULL : p;
2403 
2404   Node* mem     = in(MemNode::Memory);
2405   Node* address = in(MemNode::Address);
2406 
2407   // Back-to-back stores to same address?  Fold em up.  Generally
2408   // unsafe if I have intervening uses...  Also disallowed for StoreCM
2409   // since they must follow each StoreP operation.  Redundant StoreCMs
2410   // are eliminated just before matching in final_graph_reshape.
2411   if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2412       mem->Opcode() != Op_StoreCM) {
2413     // Looking at a dead closed cycle of memory?
2414     assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2415 
2416     assert(Opcode() == mem->Opcode() ||
2417            phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2418            "no mismatched stores, except on raw memory");
2419 
2420     if (mem->outcnt() == 1 &&           // check for intervening uses
2421         mem->as_Store()->memory_size() <= this->memory_size()) {
2422       // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2423       // For example, 'mem' might be the final state at a conditional return.
2424       // Or, 'mem' might be used by some node which is live at the same time
2425       // 'this' is live, which might be unschedulable.  So, require exactly
2426       // ONE user, the 'this' store, until such time as we clone 'mem' for
2427       // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2428       if (can_reshape) {  // (%%% is this an anachronism?)
2429         set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2430                   phase->is_IterGVN());
2431       } else {
2432         // It's OK to do this in the parser, since DU info is always accurate,
2433         // and the parser always refers to nodes via SafePointNode maps.
2434         set_req(MemNode::Memory, mem->in(MemNode::Memory));
2435       }
2436       return this;
2437     }
2438   }
2439 
2440   // Capture an unaliased, unconditional, simple store into an initializer.
2441   // Or, if it is independent of the allocation, hoist it above the allocation.
2442   if (ReduceFieldZeroing && /*can_reshape &&*/
2443       mem->is_Proj() && mem->in(0)->is_Initialize()) {
2444     InitializeNode* init = mem->in(0)->as_Initialize();
2445     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2446     if (offset > 0) {
2447       Node* moved = init->capture_store(this, offset, phase, can_reshape);
2448       // If the InitializeNode captured me, it made a raw copy of me,
2449       // and I need to disappear.
2450       if (moved != NULL) {
2451         // %%% hack to ensure that Ideal returns a new node:
2452         mem = MergeMemNode::make(mem);
2453         return mem;             // fold me away
2454       }
2455     }
2456   }
2457 
2458   return NULL;                  // No further progress
2459 }
2460 
2461 //------------------------------Value-----------------------------------------
2462 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2463   // Either input is TOP ==> the result is TOP
2464   const Type *t1 = phase->type( in(MemNode::Memory) );
2465   if( t1 == Type::TOP ) return Type::TOP;
2466   const Type *t2 = phase->type( in(MemNode::Address) );
2467   if( t2 == Type::TOP ) return Type::TOP;
2468   const Type *t3 = phase->type( in(MemNode::ValueIn) );
2469   if( t3 == Type::TOP ) return Type::TOP;
2470   return Type::MEMORY;
2471 }
2472 
2473 //------------------------------Identity---------------------------------------
2474 // Remove redundant stores:
2475 //   Store(m, p, Load(m, p)) changes to m.
2476 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2477 Node *StoreNode::Identity( PhaseTransform *phase ) {
2478   Node* mem = in(MemNode::Memory);
2479   Node* adr = in(MemNode::Address);
2480   Node* val = in(MemNode::ValueIn);
2481 
2482   // Load then Store?  Then the Store is useless
2483   if (val->is_Load() &&
2484       val->in(MemNode::Address)->eqv_uncast(adr) &&
2485       val->in(MemNode::Memory )->eqv_uncast(mem) &&
2486       val->as_Load()->store_Opcode() == Opcode()) {
2487     return mem;
2488   }
2489 
2490   // Two stores in a row of the same value?
2491   if (mem->is_Store() &&
2492       mem->in(MemNode::Address)->eqv_uncast(adr) &&
2493       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2494       mem->Opcode() == Opcode()) {
2495     return mem;
2496   }
2497 
2498   // Store of zero anywhere into a freshly-allocated object?
2499   // Then the store is useless.
2500   // (It must already have been captured by the InitializeNode.)
2501   if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2502     // a newly allocated object is already all-zeroes everywhere
2503     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2504       return mem;
2505     }
2506 
2507     // the store may also apply to zero-bits in an earlier object
2508     Node* prev_mem = find_previous_store(phase);
2509     // Steps (a), (b):  Walk past independent stores to find an exact match.
2510     if (prev_mem != NULL) {
2511       Node* prev_val = can_see_stored_value(prev_mem, phase);
2512       if (prev_val != NULL && phase->eqv(prev_val, val)) {
2513         // prev_val and val might differ by a cast; it would be good
2514         // to keep the more informative of the two.
2515         return mem;
2516       }
2517     }
2518   }
2519 
2520   return this;
2521 }
2522 
2523 //------------------------------match_edge-------------------------------------
2524 // Do we Match on this edge index or not?  Match only memory & value
2525 uint StoreNode::match_edge(uint idx) const {
2526   return idx == MemNode::Address || idx == MemNode::ValueIn;
2527 }
2528 
2529 //------------------------------cmp--------------------------------------------
2530 // Do not common stores up together.  They generally have to be split
2531 // back up anyways, so do not bother.
2532 uint StoreNode::cmp( const Node &n ) const {
2533   return (&n == this);          // Always fail except on self
2534 }
2535 
2536 //------------------------------Ideal_masked_input-----------------------------
2537 // Check for a useless mask before a partial-word store
2538 // (StoreB ... (AndI valIn conIa) )
2539 // If (conIa & mask == mask) this simplifies to
2540 // (StoreB ... (valIn) )
2541 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2542   Node *val = in(MemNode::ValueIn);
2543   if( val->Opcode() == Op_AndI ) {
2544     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2545     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2546       set_req(MemNode::ValueIn, val->in(1));
2547       return this;
2548     }
2549   }
2550   return NULL;
2551 }
2552 
2553 
2554 //------------------------------Ideal_sign_extended_input----------------------
2555 // Check for useless sign-extension before a partial-word store
2556 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2557 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
2558 // (StoreB ... (valIn) )
2559 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2560   Node *val = in(MemNode::ValueIn);
2561   if( val->Opcode() == Op_RShiftI ) {
2562     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2563     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2564       Node *shl = val->in(1);
2565       if( shl->Opcode() == Op_LShiftI ) {
2566         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2567         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2568           set_req(MemNode::ValueIn, shl->in(1));
2569           return this;
2570         }
2571       }
2572     }
2573   }
2574   return NULL;
2575 }
2576 
2577 //------------------------------value_never_loaded-----------------------------------
2578 // Determine whether there are any possible loads of the value stored.
2579 // For simplicity, we actually check if there are any loads from the
2580 // address stored to, not just for loads of the value stored by this node.
2581 //
2582 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2583   Node *adr = in(Address);
2584   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2585   if (adr_oop == NULL)
2586     return false;
2587   if (!adr_oop->is_known_instance_field())
2588     return false; // if not a distinct instance, there may be aliases of the address
2589   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2590     Node *use = adr->fast_out(i);
2591     int opc = use->Opcode();
2592     if (use->is_Load() || use->is_LoadStore()) {
2593       return false;
2594     }
2595   }
2596   return true;
2597 }
2598 
2599 //=============================================================================
2600 //------------------------------Ideal------------------------------------------
2601 // If the store is from an AND mask that leaves the low bits untouched, then
2602 // we can skip the AND operation.  If the store is from a sign-extension
2603 // (a left shift, then right shift) we can skip both.
2604 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2605   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2606   if( progress != NULL ) return progress;
2607 
2608   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2609   if( progress != NULL ) return progress;
2610 
2611   // Finally check the default case
2612   return StoreNode::Ideal(phase, can_reshape);
2613 }
2614 
2615 //=============================================================================
2616 //------------------------------Ideal------------------------------------------
2617 // If the store is from an AND mask that leaves the low bits untouched, then
2618 // we can skip the AND operation
2619 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2620   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2621   if( progress != NULL ) return progress;
2622 
2623   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2624   if( progress != NULL ) return progress;
2625 
2626   // Finally check the default case
2627   return StoreNode::Ideal(phase, can_reshape);
2628 }
2629 
2630 //=============================================================================
2631 //------------------------------Identity---------------------------------------
2632 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2633   // No need to card mark when storing a null ptr
2634   Node* my_store = in(MemNode::OopStore);
2635   if (my_store->is_Store()) {
2636     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2637     if( t1 == TypePtr::NULL_PTR ) {
2638       return in(MemNode::Memory);
2639     }
2640   }
2641   return this;
2642 }
2643 
2644 //=============================================================================
2645 //------------------------------Ideal---------------------------------------
2646 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2647   Node* progress = StoreNode::Ideal(phase, can_reshape);
2648   if (progress != NULL) return progress;
2649 
2650   Node* my_store = in(MemNode::OopStore);
2651   if (my_store->is_MergeMem()) {
2652     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2653     set_req(MemNode::OopStore, mem);
2654     return this;
2655   }
2656 
2657   return NULL;
2658 }
2659 
2660 //------------------------------Value-----------------------------------------
2661 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2662   // Either input is TOP ==> the result is TOP
2663   const Type *t = phase->type( in(MemNode::Memory) );
2664   if( t == Type::TOP ) return Type::TOP;
2665   t = phase->type( in(MemNode::Address) );
2666   if( t == Type::TOP ) return Type::TOP;
2667   t = phase->type( in(MemNode::ValueIn) );
2668   if( t == Type::TOP ) return Type::TOP;
2669   // If extra input is TOP ==> the result is TOP
2670   t = phase->type( in(MemNode::OopStore) );
2671   if( t == Type::TOP ) return Type::TOP;
2672 
2673   return StoreNode::Value( phase );
2674 }
2675 
2676 
2677 //=============================================================================
2678 //----------------------------------SCMemProjNode------------------------------
2679 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2680 {
2681   return bottom_type();
2682 }
2683 
2684 //=============================================================================
2685 //----------------------------------LoadStoreNode------------------------------
2686 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2687   : Node(required),
2688     _type(rt),
2689     _adr_type(at)
2690 {
2691   init_req(MemNode::Control, c  );
2692   init_req(MemNode::Memory , mem);
2693   init_req(MemNode::Address, adr);
2694   init_req(MemNode::ValueIn, val);
2695   init_class_id(Class_LoadStore);
2696 }
2697 
2698 uint LoadStoreNode::ideal_reg() const {
2699   return _type->ideal_reg();
2700 }
2701 
2702 bool LoadStoreNode::result_not_used() const {
2703   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2704     Node *x = fast_out(i);
2705     if (x->Opcode() == Op_SCMemProj) continue;
2706     return false;
2707   }
2708   return true;
2709 }
2710 
2711 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2712 
2713 //=============================================================================
2714 //----------------------------------LoadStoreConditionalNode--------------------
2715 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2716   init_req(ExpectedIn, ex );
2717 }
2718 
2719 //=============================================================================
2720 //-------------------------------adr_type--------------------------------------
2721 // Do we Match on this edge index or not?  Do not match memory
2722 const TypePtr* ClearArrayNode::adr_type() const {
2723   Node *adr = in(3);
2724   return MemNode::calculate_adr_type(adr->bottom_type());
2725 }
2726 
2727 //------------------------------match_edge-------------------------------------
2728 // Do we Match on this edge index or not?  Do not match memory
2729 uint ClearArrayNode::match_edge(uint idx) const {
2730   return idx > 1;
2731 }
2732 
2733 //------------------------------Identity---------------------------------------
2734 // Clearing a zero length array does nothing
2735 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2736   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
2737 }
2738 
2739 //------------------------------Idealize---------------------------------------
2740 // Clearing a short array is faster with stores
2741 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2742   const int unit = BytesPerLong;
2743   const TypeX* t = phase->type(in(2))->isa_intptr_t();
2744   if (!t)  return NULL;
2745   if (!t->is_con())  return NULL;
2746   intptr_t raw_count = t->get_con();
2747   intptr_t size = raw_count;
2748   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2749   // Clearing nothing uses the Identity call.
2750   // Negative clears are possible on dead ClearArrays
2751   // (see jck test stmt114.stmt11402.val).
2752   if (size <= 0 || size % unit != 0)  return NULL;
2753   intptr_t count = size / unit;
2754   // Length too long; use fast hardware clear
2755   if (size > Matcher::init_array_short_size)  return NULL;
2756   Node *mem = in(1);
2757   if( phase->type(mem)==Type::TOP ) return NULL;
2758   Node *adr = in(3);
2759   const Type* at = phase->type(adr);
2760   if( at==Type::TOP ) return NULL;
2761   const TypePtr* atp = at->isa_ptr();
2762   // adjust atp to be the correct array element address type
2763   if (atp == NULL)  atp = TypePtr::BOTTOM;
2764   else              atp = atp->add_offset(Type::OffsetBot);
2765   // Get base for derived pointer purposes
2766   if( adr->Opcode() != Op_AddP ) Unimplemented();
2767   Node *base = adr->in(1);
2768 
2769   Node *zero = phase->makecon(TypeLong::ZERO);
2770   Node *off  = phase->MakeConX(BytesPerLong);
2771   mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2772   count--;
2773   while( count-- ) {
2774     mem = phase->transform(mem);
2775     adr = phase->transform(new AddPNode(base,adr,off));
2776     mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2777   }
2778   return mem;
2779 }
2780 
2781 //----------------------------step_through----------------------------------
2782 // Return allocation input memory edge if it is different instance
2783 // or itself if it is the one we are looking for.
2784 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2785   Node* n = *np;
2786   assert(n->is_ClearArray(), "sanity");
2787   intptr_t offset;
2788   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2789   // This method is called only before Allocate nodes are expanded
2790   // during macro nodes expansion. Before that ClearArray nodes are
2791   // only generated in PhaseMacroExpand::generate_arraycopy() (before
2792   // Allocate nodes are expanded) which follows allocations.
2793   assert(alloc != NULL, "should have allocation");
2794   if (alloc->_idx == instance_id) {
2795     // Can not bypass initialization of the instance we are looking for.
2796     return false;
2797   }
2798   // Otherwise skip it.
2799   InitializeNode* init = alloc->initialization();
2800   if (init != NULL)
2801     *np = init->in(TypeFunc::Memory);
2802   else
2803     *np = alloc->in(TypeFunc::Memory);
2804   return true;
2805 }
2806 
2807 //----------------------------clear_memory-------------------------------------
2808 // Generate code to initialize object storage to zero.
2809 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2810                                    intptr_t start_offset,
2811                                    Node* end_offset,
2812                                    PhaseGVN* phase) {
2813   intptr_t offset = start_offset;
2814 
2815   int unit = BytesPerLong;
2816   if ((offset % unit) != 0) {
2817     Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2818     adr = phase->transform(adr);
2819     const TypePtr* atp = TypeRawPtr::BOTTOM;
2820     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2821     mem = phase->transform(mem);
2822     offset += BytesPerInt;
2823   }
2824   assert((offset % unit) == 0, "");
2825 
2826   // Initialize the remaining stuff, if any, with a ClearArray.
2827   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2828 }
2829 
2830 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2831                                    Node* start_offset,
2832                                    Node* end_offset,
2833                                    PhaseGVN* phase) {
2834   if (start_offset == end_offset) {
2835     // nothing to do
2836     return mem;
2837   }
2838 
2839   int unit = BytesPerLong;
2840   Node* zbase = start_offset;
2841   Node* zend  = end_offset;
2842 
2843   // Scale to the unit required by the CPU:
2844   if (!Matcher::init_array_count_is_in_bytes) {
2845     Node* shift = phase->intcon(exact_log2(unit));
2846     zbase = phase->transform(new URShiftXNode(zbase, shift) );
2847     zend  = phase->transform(new URShiftXNode(zend,  shift) );
2848   }
2849 
2850   // Bulk clear double-words
2851   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
2852   Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
2853   mem = new ClearArrayNode(ctl, mem, zsize, adr);
2854   return phase->transform(mem);
2855 }
2856 
2857 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2858                                    intptr_t start_offset,
2859                                    intptr_t end_offset,
2860                                    PhaseGVN* phase) {
2861   if (start_offset == end_offset) {
2862     // nothing to do
2863     return mem;
2864   }
2865 
2866   assert((end_offset % BytesPerInt) == 0, "odd end offset");
2867   intptr_t done_offset = end_offset;
2868   if ((done_offset % BytesPerLong) != 0) {
2869     done_offset -= BytesPerInt;
2870   }
2871   if (done_offset > start_offset) {
2872     mem = clear_memory(ctl, mem, dest,
2873                        start_offset, phase->MakeConX(done_offset), phase);
2874   }
2875   if (done_offset < end_offset) { // emit the final 32-bit store
2876     Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
2877     adr = phase->transform(adr);
2878     const TypePtr* atp = TypeRawPtr::BOTTOM;
2879     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2880     mem = phase->transform(mem);
2881     done_offset += BytesPerInt;
2882   }
2883   assert(done_offset == end_offset, "");
2884   return mem;
2885 }
2886 
2887 //=============================================================================
2888 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2889   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2890     _adr_type(C->get_adr_type(alias_idx))
2891 {
2892   init_class_id(Class_MemBar);
2893   Node* top = C->top();
2894   init_req(TypeFunc::I_O,top);
2895   init_req(TypeFunc::FramePtr,top);
2896   init_req(TypeFunc::ReturnAdr,top);
2897   if (precedent != NULL)
2898     init_req(TypeFunc::Parms, precedent);
2899 }
2900 
2901 //------------------------------cmp--------------------------------------------
2902 uint MemBarNode::hash() const { return NO_HASH; }
2903 uint MemBarNode::cmp( const Node &n ) const {
2904   return (&n == this);          // Always fail except on self
2905 }
2906 
2907 //------------------------------make-------------------------------------------
2908 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2909   switch (opcode) {
2910   case Op_MemBarAcquire:     return new MemBarAcquireNode(C, atp, pn);
2911   case Op_LoadFence:         return new LoadFenceNode(C, atp, pn);
2912   case Op_MemBarRelease:     return new MemBarReleaseNode(C, atp, pn);
2913   case Op_StoreFence:        return new StoreFenceNode(C, atp, pn);
2914   case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
2915   case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
2916   case Op_MemBarVolatile:    return new MemBarVolatileNode(C, atp, pn);
2917   case Op_MemBarCPUOrder:    return new MemBarCPUOrderNode(C, atp, pn);
2918   case Op_Initialize:        return new InitializeNode(C, atp, pn);
2919   case Op_MemBarStoreStore:  return new MemBarStoreStoreNode(C, atp, pn);
2920   default: ShouldNotReachHere(); return NULL;
2921   }
2922 }
2923 
2924 //------------------------------Ideal------------------------------------------
2925 // Return a node which is more "ideal" than the current node.  Strip out
2926 // control copies
2927 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2928   if (remove_dead_region(phase, can_reshape)) return this;
2929   // Don't bother trying to transform a dead node
2930   if (in(0) && in(0)->is_top()) {
2931     return NULL;
2932   }
2933 
2934   bool progress = false;
2935   // Eliminate volatile MemBars for scalar replaced objects.
2936   if (can_reshape && req() == (Precedent+1)) {
2937     bool eliminate = false;
2938     int opc = Opcode();
2939     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2940       // Volatile field loads and stores.
2941       Node* my_mem = in(MemBarNode::Precedent);
2942       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2943       if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2944         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2945         // replace this Precedent (decodeN) with the Load instead.
2946         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
2947           Node* load_node = my_mem->in(1);
2948           set_req(MemBarNode::Precedent, load_node);
2949           phase->is_IterGVN()->_worklist.push(my_mem);
2950           my_mem = load_node;
2951         } else {
2952           assert(my_mem->unique_out() == this, "sanity");
2953           del_req(Precedent);
2954           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2955           my_mem = NULL;
2956         }
2957         progress = true;
2958       }
2959       if (my_mem != NULL && my_mem->is_Mem()) {
2960         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2961         // Check for scalar replaced object reference.
2962         if( t_oop != NULL && t_oop->is_known_instance_field() &&
2963             t_oop->offset() != Type::OffsetBot &&
2964             t_oop->offset() != Type::OffsetTop) {
2965           eliminate = true;
2966         }
2967       }
2968     } else if (opc == Op_MemBarRelease) {
2969       // Final field stores.
2970       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2971       if ((alloc != NULL) && alloc->is_Allocate() &&
2972           alloc->as_Allocate()->_is_non_escaping) {
2973         // The allocated object does not escape.
2974         eliminate = true;
2975       }
2976     }
2977     if (eliminate) {
2978       // Replace MemBar projections by its inputs.
2979       PhaseIterGVN* igvn = phase->is_IterGVN();
2980       igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2981       igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2982       // Must return either the original node (now dead) or a new node
2983       // (Do not return a top here, since that would break the uniqueness of top.)
2984       return new ConINode(TypeInt::ZERO);
2985     }
2986   }
2987   return progress ? this : NULL;
2988 }
2989 
2990 //------------------------------Value------------------------------------------
2991 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2992   if( !in(0) ) return Type::TOP;
2993   if( phase->type(in(0)) == Type::TOP )
2994     return Type::TOP;
2995   return TypeTuple::MEMBAR;
2996 }
2997 
2998 //------------------------------match------------------------------------------
2999 // Construct projections for memory.
3000 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3001   switch (proj->_con) {
3002   case TypeFunc::Control:
3003   case TypeFunc::Memory:
3004     return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3005   }
3006   ShouldNotReachHere();
3007   return NULL;
3008 }
3009 
3010 //===========================InitializeNode====================================
3011 // SUMMARY:
3012 // This node acts as a memory barrier on raw memory, after some raw stores.
3013 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3014 // The Initialize can 'capture' suitably constrained stores as raw inits.
3015 // It can coalesce related raw stores into larger units (called 'tiles').
3016 // It can avoid zeroing new storage for memory units which have raw inits.
3017 // At macro-expansion, it is marked 'complete', and does not optimize further.
3018 //
3019 // EXAMPLE:
3020 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3021 //   ctl = incoming control; mem* = incoming memory
3022 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
3023 // First allocate uninitialized memory and fill in the header:
3024 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
3025 //   ctl := alloc.Control; mem* := alloc.Memory*
3026 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
3027 // Then initialize to zero the non-header parts of the raw memory block:
3028 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3029 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3030 // After the initialize node executes, the object is ready for service:
3031 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3032 // Suppose its body is immediately initialized as {1,2}:
3033 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3034 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3035 //   mem.SLICE(#short[*]) := store2
3036 //
3037 // DETAILS:
3038 // An InitializeNode collects and isolates object initialization after
3039 // an AllocateNode and before the next possible safepoint.  As a
3040 // memory barrier (MemBarNode), it keeps critical stores from drifting
3041 // down past any safepoint or any publication of the allocation.
3042 // Before this barrier, a newly-allocated object may have uninitialized bits.
3043 // After this barrier, it may be treated as a real oop, and GC is allowed.
3044 //
3045 // The semantics of the InitializeNode include an implicit zeroing of
3046 // the new object from object header to the end of the object.
3047 // (The object header and end are determined by the AllocateNode.)
3048 //
3049 // Certain stores may be added as direct inputs to the InitializeNode.
3050 // These stores must update raw memory, and they must be to addresses
3051 // derived from the raw address produced by AllocateNode, and with
3052 // a constant offset.  They must be ordered by increasing offset.
3053 // The first one is at in(RawStores), the last at in(req()-1).
3054 // Unlike most memory operations, they are not linked in a chain,
3055 // but are displayed in parallel as users of the rawmem output of
3056 // the allocation.
3057 //
3058 // (See comments in InitializeNode::capture_store, which continue
3059 // the example given above.)
3060 //
3061 // When the associated Allocate is macro-expanded, the InitializeNode
3062 // may be rewritten to optimize collected stores.  A ClearArrayNode
3063 // may also be created at that point to represent any required zeroing.
3064 // The InitializeNode is then marked 'complete', prohibiting further
3065 // capturing of nearby memory operations.
3066 //
3067 // During macro-expansion, all captured initializations which store
3068 // constant values of 32 bits or smaller are coalesced (if advantageous)
3069 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
3070 // initialized in fewer memory operations.  Memory words which are
3071 // covered by neither tiles nor non-constant stores are pre-zeroed
3072 // by explicit stores of zero.  (The code shape happens to do all
3073 // zeroing first, then all other stores, with both sequences occurring
3074 // in order of ascending offsets.)
3075 //
3076 // Alternatively, code may be inserted between an AllocateNode and its
3077 // InitializeNode, to perform arbitrary initialization of the new object.
3078 // E.g., the object copying intrinsics insert complex data transfers here.
3079 // The initialization must then be marked as 'complete' disable the
3080 // built-in zeroing semantics and the collection of initializing stores.
3081 //
3082 // While an InitializeNode is incomplete, reads from the memory state
3083 // produced by it are optimizable if they match the control edge and
3084 // new oop address associated with the allocation/initialization.
3085 // They return a stored value (if the offset matches) or else zero.
3086 // A write to the memory state, if it matches control and address,
3087 // and if it is to a constant offset, may be 'captured' by the
3088 // InitializeNode.  It is cloned as a raw memory operation and rewired
3089 // inside the initialization, to the raw oop produced by the allocation.
3090 // Operations on addresses which are provably distinct (e.g., to
3091 // other AllocateNodes) are allowed to bypass the initialization.
3092 //
3093 // The effect of all this is to consolidate object initialization
3094 // (both arrays and non-arrays, both piecewise and bulk) into a
3095 // single location, where it can be optimized as a unit.
3096 //
3097 // Only stores with an offset less than TrackedInitializationLimit words
3098 // will be considered for capture by an InitializeNode.  This puts a
3099 // reasonable limit on the complexity of optimized initializations.
3100 
3101 //---------------------------InitializeNode------------------------------------
3102 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3103   : _is_complete(Incomplete), _does_not_escape(false),
3104     MemBarNode(C, adr_type, rawoop)
3105 {
3106   init_class_id(Class_Initialize);
3107 
3108   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3109   assert(in(RawAddress) == rawoop, "proper init");
3110   // Note:  allocation() can be NULL, for secondary initialization barriers
3111 }
3112 
3113 // Since this node is not matched, it will be processed by the
3114 // register allocator.  Declare that there are no constraints
3115 // on the allocation of the RawAddress edge.
3116 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3117   // This edge should be set to top, by the set_complete.  But be conservative.
3118   if (idx == InitializeNode::RawAddress)
3119     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3120   return RegMask::Empty;
3121 }
3122 
3123 Node* InitializeNode::memory(uint alias_idx) {
3124   Node* mem = in(Memory);
3125   if (mem->is_MergeMem()) {
3126     return mem->as_MergeMem()->memory_at(alias_idx);
3127   } else {
3128     // incoming raw memory is not split
3129     return mem;
3130   }
3131 }
3132 
3133 bool InitializeNode::is_non_zero() {
3134   if (is_complete())  return false;
3135   remove_extra_zeroes();
3136   return (req() > RawStores);
3137 }
3138 
3139 void InitializeNode::set_complete(PhaseGVN* phase) {
3140   assert(!is_complete(), "caller responsibility");
3141   _is_complete = Complete;
3142 
3143   // After this node is complete, it contains a bunch of
3144   // raw-memory initializations.  There is no need for
3145   // it to have anything to do with non-raw memory effects.
3146   // Therefore, tell all non-raw users to re-optimize themselves,
3147   // after skipping the memory effects of this initialization.
3148   PhaseIterGVN* igvn = phase->is_IterGVN();
3149   if (igvn)  igvn->add_users_to_worklist(this);
3150 }
3151 
3152 // convenience function
3153 // return false if the init contains any stores already
3154 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3155   InitializeNode* init = initialization();
3156   if (init == NULL || init->is_complete())  return false;
3157   init->remove_extra_zeroes();
3158   // for now, if this allocation has already collected any inits, bail:
3159   if (init->is_non_zero())  return false;
3160   init->set_complete(phase);
3161   return true;
3162 }
3163 
3164 void InitializeNode::remove_extra_zeroes() {
3165   if (req() == RawStores)  return;
3166   Node* zmem = zero_memory();
3167   uint fill = RawStores;
3168   for (uint i = fill; i < req(); i++) {
3169     Node* n = in(i);
3170     if (n->is_top() || n == zmem)  continue;  // skip
3171     if (fill < i)  set_req(fill, n);          // compact
3172     ++fill;
3173   }
3174   // delete any empty spaces created:
3175   while (fill < req()) {
3176     del_req(fill);
3177   }
3178 }
3179 
3180 // Helper for remembering which stores go with which offsets.
3181 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3182   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
3183   intptr_t offset = -1;
3184   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3185                                                phase, offset);
3186   if (base == NULL)     return -1;  // something is dead,
3187   if (offset < 0)       return -1;  //        dead, dead
3188   return offset;
3189 }
3190 
3191 // Helper for proving that an initialization expression is
3192 // "simple enough" to be folded into an object initialization.
3193 // Attempts to prove that a store's initial value 'n' can be captured
3194 // within the initialization without creating a vicious cycle, such as:
3195 //     { Foo p = new Foo(); p.next = p; }
3196 // True for constants and parameters and small combinations thereof.
3197 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3198   if (n == NULL)      return true;   // (can this really happen?)
3199   if (n->is_Proj())   n = n->in(0);
3200   if (n == this)      return false;  // found a cycle
3201   if (n->is_Con())    return true;
3202   if (n->is_Start())  return true;   // params, etc., are OK
3203   if (n->is_Root())   return true;   // even better
3204 
3205   Node* ctl = n->in(0);
3206   if (ctl != NULL && !ctl->is_top()) {
3207     if (ctl->is_Proj())  ctl = ctl->in(0);
3208     if (ctl == this)  return false;
3209 
3210     // If we already know that the enclosing memory op is pinned right after
3211     // the init, then any control flow that the store has picked up
3212     // must have preceded the init, or else be equal to the init.
3213     // Even after loop optimizations (which might change control edges)
3214     // a store is never pinned *before* the availability of its inputs.
3215     if (!MemNode::all_controls_dominate(n, this))
3216       return false;                  // failed to prove a good control
3217   }
3218 
3219   // Check data edges for possible dependencies on 'this'.
3220   if ((count += 1) > 20)  return false;  // complexity limit
3221   for (uint i = 1; i < n->req(); i++) {
3222     Node* m = n->in(i);
3223     if (m == NULL || m == n || m->is_top())  continue;
3224     uint first_i = n->find_edge(m);
3225     if (i != first_i)  continue;  // process duplicate edge just once
3226     if (!detect_init_independence(m, count)) {
3227       return false;
3228     }
3229   }
3230 
3231   return true;
3232 }
3233 
3234 // Here are all the checks a Store must pass before it can be moved into
3235 // an initialization.  Returns zero if a check fails.
3236 // On success, returns the (constant) offset to which the store applies,
3237 // within the initialized memory.
3238 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3239   const int FAIL = 0;
3240   if (st->req() != MemNode::ValueIn + 1)
3241     return FAIL;                // an inscrutable StoreNode (card mark?)
3242   Node* ctl = st->in(MemNode::Control);
3243   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3244     return FAIL;                // must be unconditional after the initialization
3245   Node* mem = st->in(MemNode::Memory);
3246   if (!(mem->is_Proj() && mem->in(0) == this))
3247     return FAIL;                // must not be preceded by other stores
3248   Node* adr = st->in(MemNode::Address);
3249   intptr_t offset;
3250   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3251   if (alloc == NULL)
3252     return FAIL;                // inscrutable address
3253   if (alloc != allocation())
3254     return FAIL;                // wrong allocation!  (store needs to float up)
3255   Node* val = st->in(MemNode::ValueIn);
3256   int complexity_count = 0;
3257   if (!detect_init_independence(val, complexity_count))
3258     return FAIL;                // stored value must be 'simple enough'
3259 
3260   // The Store can be captured only if nothing after the allocation
3261   // and before the Store is using the memory location that the store
3262   // overwrites.
3263   bool failed = false;
3264   // If is_complete_with_arraycopy() is true the shape of the graph is
3265   // well defined and is safe so no need for extra checks.
3266   if (!is_complete_with_arraycopy()) {
3267     // We are going to look at each use of the memory state following
3268     // the allocation to make sure nothing reads the memory that the
3269     // Store writes.
3270     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3271     int alias_idx = phase->C->get_alias_index(t_adr);
3272     ResourceMark rm;
3273     Unique_Node_List mems;
3274     mems.push(mem);
3275     Node* unique_merge = NULL;
3276     for (uint next = 0; next < mems.size(); ++next) {
3277       Node *m  = mems.at(next);
3278       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3279         Node *n = m->fast_out(j);
3280         if (n->outcnt() == 0) {
3281           continue;
3282         }
3283         if (n == st) {
3284           continue;
3285         } else if (n->in(0) != NULL && n->in(0) != ctl) {
3286           // If the control of this use is different from the control
3287           // of the Store which is right after the InitializeNode then
3288           // this node cannot be between the InitializeNode and the
3289           // Store.
3290           continue;
3291         } else if (n->is_MergeMem()) {
3292           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3293             // We can hit a MergeMemNode (that will likely go away
3294             // later) that is a direct use of the memory state
3295             // following the InitializeNode on the same slice as the
3296             // store node that we'd like to capture. We need to check
3297             // the uses of the MergeMemNode.
3298             mems.push(n);
3299           }
3300         } else if (n->is_Mem()) {
3301           Node* other_adr = n->in(MemNode::Address);
3302           if (other_adr == adr) {
3303             failed = true;
3304             break;
3305           } else {
3306             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3307             if (other_t_adr != NULL) {
3308               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3309               if (other_alias_idx == alias_idx) {
3310                 // A load from the same memory slice as the store right
3311                 // after the InitializeNode. We check the control of the
3312                 // object/array that is loaded from. If it's the same as
3313                 // the store control then we cannot capture the store.
3314                 assert(!n->is_Store(), "2 stores to same slice on same control?");
3315                 Node* base = other_adr;
3316                 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3317                 base = base->in(AddPNode::Base);
3318                 if (base != NULL) {
3319                   base = base->uncast();
3320                   if (base->is_Proj() && base->in(0) == alloc) {
3321                     failed = true;
3322                     break;
3323                   }
3324                 }
3325               }
3326             }
3327           }
3328         } else {
3329           failed = true;
3330           break;
3331         }
3332       }
3333     }
3334   }
3335   if (failed) {
3336     if (!can_reshape) {
3337       // We decided we couldn't capture the store during parsing. We
3338       // should try again during the next IGVN once the graph is
3339       // cleaner.
3340       phase->C->record_for_igvn(st);
3341     }
3342     return FAIL;
3343   }
3344 
3345   return offset;                // success
3346 }
3347 
3348 // Find the captured store in(i) which corresponds to the range
3349 // [start..start+size) in the initialized object.
3350 // If there is one, return its index i.  If there isn't, return the
3351 // negative of the index where it should be inserted.
3352 // Return 0 if the queried range overlaps an initialization boundary
3353 // or if dead code is encountered.
3354 // If size_in_bytes is zero, do not bother with overlap checks.
3355 int InitializeNode::captured_store_insertion_point(intptr_t start,
3356                                                    int size_in_bytes,
3357                                                    PhaseTransform* phase) {
3358   const int FAIL = 0, MAX_STORE = BytesPerLong;
3359 
3360   if (is_complete())
3361     return FAIL;                // arraycopy got here first; punt
3362 
3363   assert(allocation() != NULL, "must be present");
3364 
3365   // no negatives, no header fields:
3366   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
3367 
3368   // after a certain size, we bail out on tracking all the stores:
3369   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3370   if (start >= ti_limit)  return FAIL;
3371 
3372   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3373     if (i >= limit)  return -(int)i; // not found; here is where to put it
3374 
3375     Node*    st     = in(i);
3376     intptr_t st_off = get_store_offset(st, phase);
3377     if (st_off < 0) {
3378       if (st != zero_memory()) {
3379         return FAIL;            // bail out if there is dead garbage
3380       }
3381     } else if (st_off > start) {
3382       // ...we are done, since stores are ordered
3383       if (st_off < start + size_in_bytes) {
3384         return FAIL;            // the next store overlaps
3385       }
3386       return -(int)i;           // not found; here is where to put it
3387     } else if (st_off < start) {
3388       if (size_in_bytes != 0 &&
3389           start < st_off + MAX_STORE &&
3390           start < st_off + st->as_Store()->memory_size()) {
3391         return FAIL;            // the previous store overlaps
3392       }
3393     } else {
3394       if (size_in_bytes != 0 &&
3395           st->as_Store()->memory_size() != size_in_bytes) {
3396         return FAIL;            // mismatched store size
3397       }
3398       return i;
3399     }
3400 
3401     ++i;
3402   }
3403 }
3404 
3405 // Look for a captured store which initializes at the offset 'start'
3406 // with the given size.  If there is no such store, and no other
3407 // initialization interferes, then return zero_memory (the memory
3408 // projection of the AllocateNode).
3409 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3410                                           PhaseTransform* phase) {
3411   assert(stores_are_sane(phase), "");
3412   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3413   if (i == 0) {
3414     return NULL;                // something is dead
3415   } else if (i < 0) {
3416     return zero_memory();       // just primordial zero bits here
3417   } else {
3418     Node* st = in(i);           // here is the store at this position
3419     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3420     return st;
3421   }
3422 }
3423 
3424 // Create, as a raw pointer, an address within my new object at 'offset'.
3425 Node* InitializeNode::make_raw_address(intptr_t offset,
3426                                        PhaseTransform* phase) {
3427   Node* addr = in(RawAddress);
3428   if (offset != 0) {
3429     Compile* C = phase->C;
3430     addr = phase->transform( new AddPNode(C->top(), addr,
3431                                                  phase->MakeConX(offset)) );
3432   }
3433   return addr;
3434 }
3435 
3436 // Clone the given store, converting it into a raw store
3437 // initializing a field or element of my new object.
3438 // Caller is responsible for retiring the original store,
3439 // with subsume_node or the like.
3440 //
3441 // From the example above InitializeNode::InitializeNode,
3442 // here are the old stores to be captured:
3443 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3444 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3445 //
3446 // Here is the changed code; note the extra edges on init:
3447 //   alloc = (Allocate ...)
3448 //   rawoop = alloc.RawAddress
3449 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3450 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3451 //   init = (Initialize alloc.Control alloc.Memory rawoop
3452 //                      rawstore1 rawstore2)
3453 //
3454 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3455                                     PhaseTransform* phase, bool can_reshape) {
3456   assert(stores_are_sane(phase), "");
3457 
3458   if (start < 0)  return NULL;
3459   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3460 
3461   Compile* C = phase->C;
3462   int size_in_bytes = st->memory_size();
3463   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3464   if (i == 0)  return NULL;     // bail out
3465   Node* prev_mem = NULL;        // raw memory for the captured store
3466   if (i > 0) {
3467     prev_mem = in(i);           // there is a pre-existing store under this one
3468     set_req(i, C->top());       // temporarily disconnect it
3469     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3470   } else {
3471     i = -i;                     // no pre-existing store
3472     prev_mem = zero_memory();   // a slice of the newly allocated object
3473     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3474       set_req(--i, C->top());   // reuse this edge; it has been folded away
3475     else
3476       ins_req(i, C->top());     // build a new edge
3477   }
3478   Node* new_st = st->clone();
3479   new_st->set_req(MemNode::Control, in(Control));
3480   new_st->set_req(MemNode::Memory,  prev_mem);
3481   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3482   new_st = phase->transform(new_st);
3483 
3484   // At this point, new_st might have swallowed a pre-existing store
3485   // at the same offset, or perhaps new_st might have disappeared,
3486   // if it redundantly stored the same value (or zero to fresh memory).
3487 
3488   // In any case, wire it in:
3489   phase->igvn_rehash_node_delayed(this);
3490   set_req(i, new_st);
3491 
3492   // The caller may now kill the old guy.
3493   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3494   assert(check_st == new_st || check_st == NULL, "must be findable");
3495   assert(!is_complete(), "");
3496   return new_st;
3497 }
3498 
3499 static bool store_constant(jlong* tiles, int num_tiles,
3500                            intptr_t st_off, int st_size,
3501                            jlong con) {
3502   if ((st_off & (st_size-1)) != 0)
3503     return false;               // strange store offset (assume size==2**N)
3504   address addr = (address)tiles + st_off;
3505   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3506   switch (st_size) {
3507   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
3508   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
3509   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
3510   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
3511   default: return false;        // strange store size (detect size!=2**N here)
3512   }
3513   return true;                  // return success to caller
3514 }
3515 
3516 // Coalesce subword constants into int constants and possibly
3517 // into long constants.  The goal, if the CPU permits,
3518 // is to initialize the object with a small number of 64-bit tiles.
3519 // Also, convert floating-point constants to bit patterns.
3520 // Non-constants are not relevant to this pass.
3521 //
3522 // In terms of the running example on InitializeNode::InitializeNode
3523 // and InitializeNode::capture_store, here is the transformation
3524 // of rawstore1 and rawstore2 into rawstore12:
3525 //   alloc = (Allocate ...)
3526 //   rawoop = alloc.RawAddress
3527 //   tile12 = 0x00010002
3528 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3529 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3530 //
3531 void
3532 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3533                                         Node* size_in_bytes,
3534                                         PhaseGVN* phase) {
3535   Compile* C = phase->C;
3536 
3537   assert(stores_are_sane(phase), "");
3538   // Note:  After this pass, they are not completely sane,
3539   // since there may be some overlaps.
3540 
3541   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3542 
3543   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3544   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3545   size_limit = MIN2(size_limit, ti_limit);
3546   size_limit = align_size_up(size_limit, BytesPerLong);
3547   int num_tiles = size_limit / BytesPerLong;
3548 
3549   // allocate space for the tile map:
3550   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3551   jlong  tiles_buf[small_len];
3552   Node*  nodes_buf[small_len];
3553   jlong  inits_buf[small_len];
3554   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3555                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3556   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3557                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3558   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3559                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3560   // tiles: exact bitwise model of all primitive constants
3561   // nodes: last constant-storing node subsumed into the tiles model
3562   // inits: which bytes (in each tile) are touched by any initializations
3563 
3564   //// Pass A: Fill in the tile model with any relevant stores.
3565 
3566   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3567   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3568   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3569   Node* zmem = zero_memory(); // initially zero memory state
3570   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3571     Node* st = in(i);
3572     intptr_t st_off = get_store_offset(st, phase);
3573 
3574     // Figure out the store's offset and constant value:
3575     if (st_off < header_size)             continue; //skip (ignore header)
3576     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
3577     int st_size = st->as_Store()->memory_size();
3578     if (st_off + st_size > size_limit)    break;
3579 
3580     // Record which bytes are touched, whether by constant or not.
3581     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3582       continue;                 // skip (strange store size)
3583 
3584     const Type* val = phase->type(st->in(MemNode::ValueIn));
3585     if (!val->singleton())                continue; //skip (non-con store)
3586     BasicType type = val->basic_type();
3587 
3588     jlong con = 0;
3589     switch (type) {
3590     case T_INT:    con = val->is_int()->get_con();  break;
3591     case T_LONG:   con = val->is_long()->get_con(); break;
3592     case T_FLOAT:  con = jint_cast(val->getf());    break;
3593     case T_DOUBLE: con = jlong_cast(val->getd());   break;
3594     default:                              continue; //skip (odd store type)
3595     }
3596 
3597     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3598         st->Opcode() == Op_StoreL) {
3599       continue;                 // This StoreL is already optimal.
3600     }
3601 
3602     // Store down the constant.
3603     store_constant(tiles, num_tiles, st_off, st_size, con);
3604 
3605     intptr_t j = st_off >> LogBytesPerLong;
3606 
3607     if (type == T_INT && st_size == BytesPerInt
3608         && (st_off & BytesPerInt) == BytesPerInt) {
3609       jlong lcon = tiles[j];
3610       if (!Matcher::isSimpleConstant64(lcon) &&
3611           st->Opcode() == Op_StoreI) {
3612         // This StoreI is already optimal by itself.
3613         jint* intcon = (jint*) &tiles[j];
3614         intcon[1] = 0;  // undo the store_constant()
3615 
3616         // If the previous store is also optimal by itself, back up and
3617         // undo the action of the previous loop iteration... if we can.
3618         // But if we can't, just let the previous half take care of itself.
3619         st = nodes[j];
3620         st_off -= BytesPerInt;
3621         con = intcon[0];
3622         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3623           assert(st_off >= header_size, "still ignoring header");
3624           assert(get_store_offset(st, phase) == st_off, "must be");
3625           assert(in(i-1) == zmem, "must be");
3626           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3627           assert(con == tcon->is_int()->get_con(), "must be");
3628           // Undo the effects of the previous loop trip, which swallowed st:
3629           intcon[0] = 0;        // undo store_constant()
3630           set_req(i-1, st);     // undo set_req(i, zmem)
3631           nodes[j] = NULL;      // undo nodes[j] = st
3632           --old_subword;        // undo ++old_subword
3633         }
3634         continue;               // This StoreI is already optimal.
3635       }
3636     }
3637 
3638     // This store is not needed.
3639     set_req(i, zmem);
3640     nodes[j] = st;              // record for the moment
3641     if (st_size < BytesPerLong) // something has changed
3642           ++old_subword;        // includes int/float, but who's counting...
3643     else  ++old_long;
3644   }
3645 
3646   if ((old_subword + old_long) == 0)
3647     return;                     // nothing more to do
3648 
3649   //// Pass B: Convert any non-zero tiles into optimal constant stores.
3650   // Be sure to insert them before overlapping non-constant stores.
3651   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
3652   for (int j = 0; j < num_tiles; j++) {
3653     jlong con  = tiles[j];
3654     jlong init = inits[j];
3655     if (con == 0)  continue;
3656     jint con0,  con1;           // split the constant, address-wise
3657     jint init0, init1;          // split the init map, address-wise
3658     { union { jlong con; jint intcon[2]; } u;
3659       u.con = con;
3660       con0  = u.intcon[0];
3661       con1  = u.intcon[1];
3662       u.con = init;
3663       init0 = u.intcon[0];
3664       init1 = u.intcon[1];
3665     }
3666 
3667     Node* old = nodes[j];
3668     assert(old != NULL, "need the prior store");
3669     intptr_t offset = (j * BytesPerLong);
3670 
3671     bool split = !Matcher::isSimpleConstant64(con);
3672 
3673     if (offset < header_size) {
3674       assert(offset + BytesPerInt >= header_size, "second int counts");
3675       assert(*(jint*)&tiles[j] == 0, "junk in header");
3676       split = true;             // only the second word counts
3677       // Example:  int a[] = { 42 ... }
3678     } else if (con0 == 0 && init0 == -1) {
3679       split = true;             // first word is covered by full inits
3680       // Example:  int a[] = { ... foo(), 42 ... }
3681     } else if (con1 == 0 && init1 == -1) {
3682       split = true;             // second word is covered by full inits
3683       // Example:  int a[] = { ... 42, foo() ... }
3684     }
3685 
3686     // Here's a case where init0 is neither 0 nor -1:
3687     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
3688     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3689     // In this case the tile is not split; it is (jlong)42.
3690     // The big tile is stored down, and then the foo() value is inserted.
3691     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3692 
3693     Node* ctl = old->in(MemNode::Control);
3694     Node* adr = make_raw_address(offset, phase);
3695     const TypePtr* atp = TypeRawPtr::BOTTOM;
3696 
3697     // One or two coalesced stores to plop down.
3698     Node*    st[2];
3699     intptr_t off[2];
3700     int  nst = 0;
3701     if (!split) {
3702       ++new_long;
3703       off[nst] = offset;
3704       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3705                                   phase->longcon(con), T_LONG, MemNode::unordered);
3706     } else {
3707       // Omit either if it is a zero.
3708       if (con0 != 0) {
3709         ++new_int;
3710         off[nst]  = offset;
3711         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3712                                     phase->intcon(con0), T_INT, MemNode::unordered);
3713       }
3714       if (con1 != 0) {
3715         ++new_int;
3716         offset += BytesPerInt;
3717         adr = make_raw_address(offset, phase);
3718         off[nst]  = offset;
3719         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3720                                     phase->intcon(con1), T_INT, MemNode::unordered);
3721       }
3722     }
3723 
3724     // Insert second store first, then the first before the second.
3725     // Insert each one just before any overlapping non-constant stores.
3726     while (nst > 0) {
3727       Node* st1 = st[--nst];
3728       C->copy_node_notes_to(st1, old);
3729       st1 = phase->transform(st1);
3730       offset = off[nst];
3731       assert(offset >= header_size, "do not smash header");
3732       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3733       guarantee(ins_idx != 0, "must re-insert constant store");
3734       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
3735       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3736         set_req(--ins_idx, st1);
3737       else
3738         ins_req(ins_idx, st1);
3739     }
3740   }
3741 
3742   if (PrintCompilation && WizardMode)
3743     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3744                   old_subword, old_long, new_int, new_long);
3745   if (C->log() != NULL)
3746     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3747                    old_subword, old_long, new_int, new_long);
3748 
3749   // Clean up any remaining occurrences of zmem:
3750   remove_extra_zeroes();
3751 }
3752 
3753 // Explore forward from in(start) to find the first fully initialized
3754 // word, and return its offset.  Skip groups of subword stores which
3755 // together initialize full words.  If in(start) is itself part of a
3756 // fully initialized word, return the offset of in(start).  If there
3757 // are no following full-word stores, or if something is fishy, return
3758 // a negative value.
3759 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3760   int       int_map = 0;
3761   intptr_t  int_map_off = 0;
3762   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
3763 
3764   for (uint i = start, limit = req(); i < limit; i++) {
3765     Node* st = in(i);
3766 
3767     intptr_t st_off = get_store_offset(st, phase);
3768     if (st_off < 0)  break;  // return conservative answer
3769 
3770     int st_size = st->as_Store()->memory_size();
3771     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3772       return st_off;            // we found a complete word init
3773     }
3774 
3775     // update the map:
3776 
3777     intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3778     if (this_int_off != int_map_off) {
3779       // reset the map:
3780       int_map = 0;
3781       int_map_off = this_int_off;
3782     }
3783 
3784     int subword_off = st_off - this_int_off;
3785     int_map |= right_n_bits(st_size) << subword_off;
3786     if ((int_map & FULL_MAP) == FULL_MAP) {
3787       return this_int_off;      // we found a complete word init
3788     }
3789 
3790     // Did this store hit or cross the word boundary?
3791     intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3792     if (next_int_off == this_int_off + BytesPerInt) {
3793       // We passed the current int, without fully initializing it.
3794       int_map_off = next_int_off;
3795       int_map >>= BytesPerInt;
3796     } else if (next_int_off > this_int_off + BytesPerInt) {
3797       // We passed the current and next int.
3798       return this_int_off + BytesPerInt;
3799     }
3800   }
3801 
3802   return -1;
3803 }
3804 
3805 
3806 // Called when the associated AllocateNode is expanded into CFG.
3807 // At this point, we may perform additional optimizations.
3808 // Linearize the stores by ascending offset, to make memory
3809 // activity as coherent as possible.
3810 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3811                                       intptr_t header_size,
3812                                       Node* size_in_bytes,
3813                                       PhaseGVN* phase) {
3814   assert(!is_complete(), "not already complete");
3815   assert(stores_are_sane(phase), "");
3816   assert(allocation() != NULL, "must be present");
3817 
3818   remove_extra_zeroes();
3819 
3820   if (ReduceFieldZeroing || ReduceBulkZeroing)
3821     // reduce instruction count for common initialization patterns
3822     coalesce_subword_stores(header_size, size_in_bytes, phase);
3823 
3824   Node* zmem = zero_memory();   // initially zero memory state
3825   Node* inits = zmem;           // accumulating a linearized chain of inits
3826   #ifdef ASSERT
3827   intptr_t first_offset = allocation()->minimum_header_size();
3828   intptr_t last_init_off = first_offset;  // previous init offset
3829   intptr_t last_init_end = first_offset;  // previous init offset+size
3830   intptr_t last_tile_end = first_offset;  // previous tile offset+size
3831   #endif
3832   intptr_t zeroes_done = header_size;
3833 
3834   bool do_zeroing = true;       // we might give up if inits are very sparse
3835   int  big_init_gaps = 0;       // how many large gaps have we seen?
3836 
3837   if (ZeroTLAB)  do_zeroing = false;
3838   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
3839 
3840   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3841     Node* st = in(i);
3842     intptr_t st_off = get_store_offset(st, phase);
3843     if (st_off < 0)
3844       break;                    // unknown junk in the inits
3845     if (st->in(MemNode::Memory) != zmem)
3846       break;                    // complicated store chains somehow in list
3847 
3848     int st_size = st->as_Store()->memory_size();
3849     intptr_t next_init_off = st_off + st_size;
3850 
3851     if (do_zeroing && zeroes_done < next_init_off) {
3852       // See if this store needs a zero before it or under it.
3853       intptr_t zeroes_needed = st_off;
3854 
3855       if (st_size < BytesPerInt) {
3856         // Look for subword stores which only partially initialize words.
3857         // If we find some, we must lay down some word-level zeroes first,
3858         // underneath the subword stores.
3859         //
3860         // Examples:
3861         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
3862         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
3863         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
3864         //
3865         // Note:  coalesce_subword_stores may have already done this,
3866         // if it was prompted by constant non-zero subword initializers.
3867         // But this case can still arise with non-constant stores.
3868 
3869         intptr_t next_full_store = find_next_fullword_store(i, phase);
3870 
3871         // In the examples above:
3872         //   in(i)          p   q   r   s     x   y     z
3873         //   st_off        12  13  14  15    12  13    14
3874         //   st_size        1   1   1   1     1   1     1
3875         //   next_full_s.  12  16  16  16    16  16    16
3876         //   z's_done      12  16  16  16    12  16    12
3877         //   z's_needed    12  16  16  16    16  16    16
3878         //   zsize          0   0   0   0     4   0     4
3879         if (next_full_store < 0) {
3880           // Conservative tack:  Zero to end of current word.
3881           zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3882         } else {
3883           // Zero to beginning of next fully initialized word.
3884           // Or, don't zero at all, if we are already in that word.
3885           assert(next_full_store >= zeroes_needed, "must go forward");
3886           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3887           zeroes_needed = next_full_store;
3888         }
3889       }
3890 
3891       if (zeroes_needed > zeroes_done) {
3892         intptr_t zsize = zeroes_needed - zeroes_done;
3893         // Do some incremental zeroing on rawmem, in parallel with inits.
3894         zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3895         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3896                                               zeroes_done, zeroes_needed,
3897                                               phase);
3898         zeroes_done = zeroes_needed;
3899         if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3900           do_zeroing = false;   // leave the hole, next time
3901       }
3902     }
3903 
3904     // Collect the store and move on:
3905     st->set_req(MemNode::Memory, inits);
3906     inits = st;                 // put it on the linearized chain
3907     set_req(i, zmem);           // unhook from previous position
3908 
3909     if (zeroes_done == st_off)
3910       zeroes_done = next_init_off;
3911 
3912     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3913 
3914     #ifdef ASSERT
3915     // Various order invariants.  Weaker than stores_are_sane because
3916     // a large constant tile can be filled in by smaller non-constant stores.
3917     assert(st_off >= last_init_off, "inits do not reverse");
3918     last_init_off = st_off;
3919     const Type* val = NULL;
3920     if (st_size >= BytesPerInt &&
3921         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3922         (int)val->basic_type() < (int)T_OBJECT) {
3923       assert(st_off >= last_tile_end, "tiles do not overlap");
3924       assert(st_off >= last_init_end, "tiles do not overwrite inits");
3925       last_tile_end = MAX2(last_tile_end, next_init_off);
3926     } else {
3927       intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3928       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3929       assert(st_off      >= last_init_end, "inits do not overlap");
3930       last_init_end = next_init_off;  // it's a non-tile
3931     }
3932     #endif //ASSERT
3933   }
3934 
3935   remove_extra_zeroes();        // clear out all the zmems left over
3936   add_req(inits);
3937 
3938   if (!ZeroTLAB) {
3939     // If anything remains to be zeroed, zero it all now.
3940     zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3941     // if it is the last unused 4 bytes of an instance, forget about it
3942     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3943     if (zeroes_done + BytesPerLong >= size_limit) {
3944       assert(allocation() != NULL, "");
3945       if (allocation()->Opcode() == Op_Allocate) {
3946         Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3947         ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3948         if (zeroes_done == k->layout_helper())
3949           zeroes_done = size_limit;
3950       }
3951     }
3952     if (zeroes_done < size_limit) {
3953       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3954                                             zeroes_done, size_in_bytes, phase);
3955     }
3956   }
3957 
3958   set_complete(phase);
3959   return rawmem;
3960 }
3961 
3962 
3963 #ifdef ASSERT
3964 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3965   if (is_complete())
3966     return true;                // stores could be anything at this point
3967   assert(allocation() != NULL, "must be present");
3968   intptr_t last_off = allocation()->minimum_header_size();
3969   for (uint i = InitializeNode::RawStores; i < req(); i++) {
3970     Node* st = in(i);
3971     intptr_t st_off = get_store_offset(st, phase);
3972     if (st_off < 0)  continue;  // ignore dead garbage
3973     if (last_off > st_off) {
3974       tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
3975       this->dump(2);
3976       assert(false, "ascending store offsets");
3977       return false;
3978     }
3979     last_off = st_off + st->as_Store()->memory_size();
3980   }
3981   return true;
3982 }
3983 #endif //ASSERT
3984 
3985 
3986 
3987 
3988 //============================MergeMemNode=====================================
3989 //
3990 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
3991 // contributing store or call operations.  Each contributor provides the memory
3992 // state for a particular "alias type" (see Compile::alias_type).  For example,
3993 // if a MergeMem has an input X for alias category #6, then any memory reference
3994 // to alias category #6 may use X as its memory state input, as an exact equivalent
3995 // to using the MergeMem as a whole.
3996 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3997 //
3998 // (Here, the <N> notation gives the index of the relevant adr_type.)
3999 //
4000 // In one special case (and more cases in the future), alias categories overlap.
4001 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4002 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
4003 // it is exactly equivalent to that state W:
4004 //   MergeMem(<Bot>: W) <==> W
4005 //
4006 // Usually, the merge has more than one input.  In that case, where inputs
4007 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4008 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4009 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4010 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4011 //
4012 // A merge can take a "wide" memory state as one of its narrow inputs.
4013 // This simply means that the merge observes out only the relevant parts of
4014 // the wide input.  That is, wide memory states arriving at narrow merge inputs
4015 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
4016 //
4017 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4018 // and that memory slices "leak through":
4019 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4020 //
4021 // But, in such a cascade, repeated memory slices can "block the leak":
4022 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4023 //
4024 // In the last example, Y is not part of the combined memory state of the
4025 // outermost MergeMem.  The system must, of course, prevent unschedulable
4026 // memory states from arising, so you can be sure that the state Y is somehow
4027 // a precursor to state Y'.
4028 //
4029 //
4030 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4031 // of each MergeMemNode array are exactly the numerical alias indexes, including
4032 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
4033 // Compile::alias_type (and kin) produce and manage these indexes.
4034 //
4035 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4036 // (Note that this provides quick access to the top node inside MergeMem methods,
4037 // without the need to reach out via TLS to Compile::current.)
4038 //
4039 // As a consequence of what was just described, a MergeMem that represents a full
4040 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4041 // containing all alias categories.
4042 //
4043 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4044 //
4045 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4046 // a memory state for the alias type <N>, or else the top node, meaning that
4047 // there is no particular input for that alias type.  Note that the length of
4048 // a MergeMem is variable, and may be extended at any time to accommodate new
4049 // memory states at larger alias indexes.  When merges grow, they are of course
4050 // filled with "top" in the unused in() positions.
4051 //
4052 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4053 // (Top was chosen because it works smoothly with passes like GCM.)
4054 //
4055 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
4056 // the type of random VM bits like TLS references.)  Since it is always the
4057 // first non-Bot memory slice, some low-level loops use it to initialize an
4058 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
4059 //
4060 //
4061 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
4062 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
4063 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4064 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
4065 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
4066 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4067 //
4068 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4069 // really that different from the other memory inputs.  An abbreviation called
4070 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4071 //
4072 //
4073 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
4074 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
4075 // that "emerges though" the base memory will be marked as excluding the alias types
4076 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4077 //
4078 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4079 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4080 //
4081 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4082 // (It is currently unimplemented.)  As you can see, the resulting merge is
4083 // actually a disjoint union of memory states, rather than an overlay.
4084 //
4085 
4086 //------------------------------MergeMemNode-----------------------------------
4087 Node* MergeMemNode::make_empty_memory() {
4088   Node* empty_memory = (Node*) Compile::current()->top();
4089   assert(empty_memory->is_top(), "correct sentinel identity");
4090   return empty_memory;
4091 }
4092 
4093 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4094   init_class_id(Class_MergeMem);
4095   // all inputs are nullified in Node::Node(int)
4096   // set_input(0, NULL);  // no control input
4097 
4098   // Initialize the edges uniformly to top, for starters.
4099   Node* empty_mem = make_empty_memory();
4100   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4101     init_req(i,empty_mem);
4102   }
4103   assert(empty_memory() == empty_mem, "");
4104 
4105   if( new_base != NULL && new_base->is_MergeMem() ) {
4106     MergeMemNode* mdef = new_base->as_MergeMem();
4107     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4108     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4109       mms.set_memory(mms.memory2());
4110     }
4111     assert(base_memory() == mdef->base_memory(), "");
4112   } else {
4113     set_base_memory(new_base);
4114   }
4115 }
4116 
4117 // Make a new, untransformed MergeMem with the same base as 'mem'.
4118 // If mem is itself a MergeMem, populate the result with the same edges.
4119 MergeMemNode* MergeMemNode::make(Node* mem) {
4120   return new MergeMemNode(mem);
4121 }
4122 
4123 //------------------------------cmp--------------------------------------------
4124 uint MergeMemNode::hash() const { return NO_HASH; }
4125 uint MergeMemNode::cmp( const Node &n ) const {
4126   return (&n == this);          // Always fail except on self
4127 }
4128 
4129 //------------------------------Identity---------------------------------------
4130 Node* MergeMemNode::Identity(PhaseTransform *phase) {
4131   // Identity if this merge point does not record any interesting memory
4132   // disambiguations.
4133   Node* base_mem = base_memory();
4134   Node* empty_mem = empty_memory();
4135   if (base_mem != empty_mem) {  // Memory path is not dead?
4136     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4137       Node* mem = in(i);
4138       if (mem != empty_mem && mem != base_mem) {
4139         return this;            // Many memory splits; no change
4140       }
4141     }
4142   }
4143   return base_mem;              // No memory splits; ID on the one true input
4144 }
4145 
4146 //------------------------------Ideal------------------------------------------
4147 // This method is invoked recursively on chains of MergeMem nodes
4148 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4149   // Remove chain'd MergeMems
4150   //
4151   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4152   // relative to the "in(Bot)".  Since we are patching both at the same time,
4153   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4154   // but rewrite each "in(i)" relative to the new "in(Bot)".
4155   Node *progress = NULL;
4156 
4157 
4158   Node* old_base = base_memory();
4159   Node* empty_mem = empty_memory();
4160   if (old_base == empty_mem)
4161     return NULL; // Dead memory path.
4162 
4163   MergeMemNode* old_mbase;
4164   if (old_base != NULL && old_base->is_MergeMem())
4165     old_mbase = old_base->as_MergeMem();
4166   else
4167     old_mbase = NULL;
4168   Node* new_base = old_base;
4169 
4170   // simplify stacked MergeMems in base memory
4171   if (old_mbase)  new_base = old_mbase->base_memory();
4172 
4173   // the base memory might contribute new slices beyond my req()
4174   if (old_mbase)  grow_to_match(old_mbase);
4175 
4176   // Look carefully at the base node if it is a phi.
4177   PhiNode* phi_base;
4178   if (new_base != NULL && new_base->is_Phi())
4179     phi_base = new_base->as_Phi();
4180   else
4181     phi_base = NULL;
4182 
4183   Node*    phi_reg = NULL;
4184   uint     phi_len = (uint)-1;
4185   if (phi_base != NULL && !phi_base->is_copy()) {
4186     // do not examine phi if degraded to a copy
4187     phi_reg = phi_base->region();
4188     phi_len = phi_base->req();
4189     // see if the phi is unfinished
4190     for (uint i = 1; i < phi_len; i++) {
4191       if (phi_base->in(i) == NULL) {
4192         // incomplete phi; do not look at it yet!
4193         phi_reg = NULL;
4194         phi_len = (uint)-1;
4195         break;
4196       }
4197     }
4198   }
4199 
4200   // Note:  We do not call verify_sparse on entry, because inputs
4201   // can normalize to the base_memory via subsume_node or similar
4202   // mechanisms.  This method repairs that damage.
4203 
4204   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4205 
4206   // Look at each slice.
4207   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4208     Node* old_in = in(i);
4209     // calculate the old memory value
4210     Node* old_mem = old_in;
4211     if (old_mem == empty_mem)  old_mem = old_base;
4212     assert(old_mem == memory_at(i), "");
4213 
4214     // maybe update (reslice) the old memory value
4215 
4216     // simplify stacked MergeMems
4217     Node* new_mem = old_mem;
4218     MergeMemNode* old_mmem;
4219     if (old_mem != NULL && old_mem->is_MergeMem())
4220       old_mmem = old_mem->as_MergeMem();
4221     else
4222       old_mmem = NULL;
4223     if (old_mmem == this) {
4224       // This can happen if loops break up and safepoints disappear.
4225       // A merge of BotPtr (default) with a RawPtr memory derived from a
4226       // safepoint can be rewritten to a merge of the same BotPtr with
4227       // the BotPtr phi coming into the loop.  If that phi disappears
4228       // also, we can end up with a self-loop of the mergemem.
4229       // In general, if loops degenerate and memory effects disappear,
4230       // a mergemem can be left looking at itself.  This simply means
4231       // that the mergemem's default should be used, since there is
4232       // no longer any apparent effect on this slice.
4233       // Note: If a memory slice is a MergeMem cycle, it is unreachable
4234       //       from start.  Update the input to TOP.
4235       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4236     }
4237     else if (old_mmem != NULL) {
4238       new_mem = old_mmem->memory_at(i);
4239     }
4240     // else preceding memory was not a MergeMem
4241 
4242     // replace equivalent phis (unfortunately, they do not GVN together)
4243     if (new_mem != NULL && new_mem != new_base &&
4244         new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4245       if (new_mem->is_Phi()) {
4246         PhiNode* phi_mem = new_mem->as_Phi();
4247         for (uint i = 1; i < phi_len; i++) {
4248           if (phi_base->in(i) != phi_mem->in(i)) {
4249             phi_mem = NULL;
4250             break;
4251           }
4252         }
4253         if (phi_mem != NULL) {
4254           // equivalent phi nodes; revert to the def
4255           new_mem = new_base;
4256         }
4257       }
4258     }
4259 
4260     // maybe store down a new value
4261     Node* new_in = new_mem;
4262     if (new_in == new_base)  new_in = empty_mem;
4263 
4264     if (new_in != old_in) {
4265       // Warning:  Do not combine this "if" with the previous "if"
4266       // A memory slice might have be be rewritten even if it is semantically
4267       // unchanged, if the base_memory value has changed.
4268       set_req(i, new_in);
4269       progress = this;          // Report progress
4270     }
4271   }
4272 
4273   if (new_base != old_base) {
4274     set_req(Compile::AliasIdxBot, new_base);
4275     // Don't use set_base_memory(new_base), because we need to update du.
4276     assert(base_memory() == new_base, "");
4277     progress = this;
4278   }
4279 
4280   if( base_memory() == this ) {
4281     // a self cycle indicates this memory path is dead
4282     set_req(Compile::AliasIdxBot, empty_mem);
4283   }
4284 
4285   // Resolve external cycles by calling Ideal on a MergeMem base_memory
4286   // Recursion must occur after the self cycle check above
4287   if( base_memory()->is_MergeMem() ) {
4288     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4289     Node *m = phase->transform(new_mbase);  // Rollup any cycles
4290     if( m != NULL && (m->is_top() ||
4291         m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4292       // propagate rollup of dead cycle to self
4293       set_req(Compile::AliasIdxBot, empty_mem);
4294     }
4295   }
4296 
4297   if( base_memory() == empty_mem ) {
4298     progress = this;
4299     // Cut inputs during Parse phase only.
4300     // During Optimize phase a dead MergeMem node will be subsumed by Top.
4301     if( !can_reshape ) {
4302       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4303         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4304       }
4305     }
4306   }
4307 
4308   if( !progress && base_memory()->is_Phi() && can_reshape ) {
4309     // Check if PhiNode::Ideal's "Split phis through memory merges"
4310     // transform should be attempted. Look for this->phi->this cycle.
4311     uint merge_width = req();
4312     if (merge_width > Compile::AliasIdxRaw) {
4313       PhiNode* phi = base_memory()->as_Phi();
4314       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4315         if (phi->in(i) == this) {
4316           phase->is_IterGVN()->_worklist.push(phi);
4317           break;
4318         }
4319       }
4320     }
4321   }
4322 
4323   assert(progress || verify_sparse(), "please, no dups of base");
4324   return progress;
4325 }
4326 
4327 //-------------------------set_base_memory-------------------------------------
4328 void MergeMemNode::set_base_memory(Node *new_base) {
4329   Node* empty_mem = empty_memory();
4330   set_req(Compile::AliasIdxBot, new_base);
4331   assert(memory_at(req()) == new_base, "must set default memory");
4332   // Clear out other occurrences of new_base:
4333   if (new_base != empty_mem) {
4334     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4335       if (in(i) == new_base)  set_req(i, empty_mem);
4336     }
4337   }
4338 }
4339 
4340 //------------------------------out_RegMask------------------------------------
4341 const RegMask &MergeMemNode::out_RegMask() const {
4342   return RegMask::Empty;
4343 }
4344 
4345 //------------------------------dump_spec--------------------------------------
4346 #ifndef PRODUCT
4347 void MergeMemNode::dump_spec(outputStream *st) const {
4348   st->print(" {");
4349   Node* base_mem = base_memory();
4350   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4351     Node* mem = memory_at(i);
4352     if (mem == base_mem) { st->print(" -"); continue; }
4353     st->print( " N%d:", mem->_idx );
4354     Compile::current()->get_adr_type(i)->dump_on(st);
4355   }
4356   st->print(" }");
4357 }
4358 #endif // !PRODUCT
4359 
4360 
4361 #ifdef ASSERT
4362 static bool might_be_same(Node* a, Node* b) {
4363   if (a == b)  return true;
4364   if (!(a->is_Phi() || b->is_Phi()))  return false;
4365   // phis shift around during optimization
4366   return true;  // pretty stupid...
4367 }
4368 
4369 // verify a narrow slice (either incoming or outgoing)
4370 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4371   if (!VerifyAliases)       return;  // don't bother to verify unless requested
4372   if (is_error_reported())  return;  // muzzle asserts when debugging an error
4373   if (Node::in_dump())      return;  // muzzle asserts when printing
4374   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4375   assert(n != NULL, "");
4376   // Elide intervening MergeMem's
4377   while (n->is_MergeMem()) {
4378     n = n->as_MergeMem()->memory_at(alias_idx);
4379   }
4380   Compile* C = Compile::current();
4381   const TypePtr* n_adr_type = n->adr_type();
4382   if (n == m->empty_memory()) {
4383     // Implicit copy of base_memory()
4384   } else if (n_adr_type != TypePtr::BOTTOM) {
4385     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4386     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4387   } else {
4388     // A few places like make_runtime_call "know" that VM calls are narrow,
4389     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4390     bool expected_wide_mem = false;
4391     if (n == m->base_memory()) {
4392       expected_wide_mem = true;
4393     } else if (alias_idx == Compile::AliasIdxRaw ||
4394                n == m->memory_at(Compile::AliasIdxRaw)) {
4395       expected_wide_mem = true;
4396     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4397       // memory can "leak through" calls on channels that
4398       // are write-once.  Allow this also.
4399       expected_wide_mem = true;
4400     }
4401     assert(expected_wide_mem, "expected narrow slice replacement");
4402   }
4403 }
4404 #else // !ASSERT
4405 #define verify_memory_slice(m,i,n) (void)(0)  // PRODUCT version is no-op
4406 #endif
4407 
4408 
4409 //-----------------------------memory_at---------------------------------------
4410 Node* MergeMemNode::memory_at(uint alias_idx) const {
4411   assert(alias_idx >= Compile::AliasIdxRaw ||
4412          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4413          "must avoid base_memory and AliasIdxTop");
4414 
4415   // Otherwise, it is a narrow slice.
4416   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4417   Compile *C = Compile::current();
4418   if (is_empty_memory(n)) {
4419     // the array is sparse; empty slots are the "top" node
4420     n = base_memory();
4421     assert(Node::in_dump()
4422            || n == NULL || n->bottom_type() == Type::TOP
4423            || n->adr_type() == NULL // address is TOP
4424            || n->adr_type() == TypePtr::BOTTOM
4425            || n->adr_type() == TypeRawPtr::BOTTOM
4426            || Compile::current()->AliasLevel() == 0,
4427            "must be a wide memory");
4428     // AliasLevel == 0 if we are organizing the memory states manually.
4429     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4430   } else {
4431     // make sure the stored slice is sane
4432     #ifdef ASSERT
4433     if (is_error_reported() || Node::in_dump()) {
4434     } else if (might_be_same(n, base_memory())) {
4435       // Give it a pass:  It is a mostly harmless repetition of the base.
4436       // This can arise normally from node subsumption during optimization.
4437     } else {
4438       verify_memory_slice(this, alias_idx, n);
4439     }
4440     #endif
4441   }
4442   return n;
4443 }
4444 
4445 //---------------------------set_memory_at-------------------------------------
4446 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4447   verify_memory_slice(this, alias_idx, n);
4448   Node* empty_mem = empty_memory();
4449   if (n == base_memory())  n = empty_mem;  // collapse default
4450   uint need_req = alias_idx+1;
4451   if (req() < need_req) {
4452     if (n == empty_mem)  return;  // already the default, so do not grow me
4453     // grow the sparse array
4454     do {
4455       add_req(empty_mem);
4456     } while (req() < need_req);
4457   }
4458   set_req( alias_idx, n );
4459 }
4460 
4461 
4462 
4463 //--------------------------iteration_setup------------------------------------
4464 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4465   if (other != NULL) {
4466     grow_to_match(other);
4467     // invariant:  the finite support of mm2 is within mm->req()
4468     #ifdef ASSERT
4469     for (uint i = req(); i < other->req(); i++) {
4470       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4471     }
4472     #endif
4473   }
4474   // Replace spurious copies of base_memory by top.
4475   Node* base_mem = base_memory();
4476   if (base_mem != NULL && !base_mem->is_top()) {
4477     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4478       if (in(i) == base_mem)
4479         set_req(i, empty_memory());
4480     }
4481   }
4482 }
4483 
4484 //---------------------------grow_to_match-------------------------------------
4485 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4486   Node* empty_mem = empty_memory();
4487   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4488   // look for the finite support of the other memory
4489   for (uint i = other->req(); --i >= req(); ) {
4490     if (other->in(i) != empty_mem) {
4491       uint new_len = i+1;
4492       while (req() < new_len)  add_req(empty_mem);
4493       break;
4494     }
4495   }
4496 }
4497 
4498 //---------------------------verify_sparse-------------------------------------
4499 #ifndef PRODUCT
4500 bool MergeMemNode::verify_sparse() const {
4501   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4502   Node* base_mem = base_memory();
4503   // The following can happen in degenerate cases, since empty==top.
4504   if (is_empty_memory(base_mem))  return true;
4505   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4506     assert(in(i) != NULL, "sane slice");
4507     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
4508   }
4509   return true;
4510 }
4511 
4512 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4513   Node* n;
4514   n = mm->in(idx);
4515   if (mem == n)  return true;  // might be empty_memory()
4516   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4517   if (mem == n)  return true;
4518   while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4519     if (mem == n)  return true;
4520     if (n == NULL)  break;
4521   }
4522   return false;
4523 }
4524 #endif // !PRODUCT