1 /*
   2  * Copyright (c) 1997, 2012, 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/loopnode.hpp"
  35 #include "opto/machnode.hpp"
  36 #include "opto/matcher.hpp"
  37 #include "opto/memnode.hpp"
  38 #include "opto/mulnode.hpp"
  39 #include "opto/phaseX.hpp"
  40 #include "opto/regmask.hpp"
  41 
  42 // Portions of code courtesy of Clifford Click
  43 
  44 // Optimization - Graph Style
  45 
  46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
  47 
  48 //=============================================================================
  49 uint MemNode::size_of() const { return sizeof(*this); }
  50 
  51 const TypePtr *MemNode::adr_type() const {
  52   Node* adr = in(Address);
  53   const TypePtr* cross_check = NULL;
  54   DEBUG_ONLY(cross_check = _adr_type);
  55   return calculate_adr_type(adr->bottom_type(), cross_check);
  56 }
  57 
  58 #ifndef PRODUCT
  59 void MemNode::dump_spec(outputStream *st) const {
  60   if (in(Address) == NULL)  return; // node is dead
  61 #ifndef ASSERT
  62   // fake the missing field
  63   const TypePtr* _adr_type = NULL;
  64   if (in(Address) != NULL)
  65     _adr_type = in(Address)->bottom_type()->isa_ptr();
  66 #endif
  67   dump_adr_type(this, _adr_type, st);
  68 
  69   Compile* C = Compile::current();
  70   if( C->alias_type(_adr_type)->is_volatile() )
  71     st->print(" Volatile!");
  72 }
  73 
  74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
  75   st->print(" @");
  76   if (adr_type == NULL) {
  77     st->print("NULL");
  78   } else {
  79     adr_type->dump_on(st);
  80     Compile* C = Compile::current();
  81     Compile::AliasType* atp = NULL;
  82     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
  83     if (atp == NULL)
  84       st->print(", idx=?\?;");
  85     else if (atp->index() == Compile::AliasIdxBot)
  86       st->print(", idx=Bot;");
  87     else if (atp->index() == Compile::AliasIdxTop)
  88       st->print(", idx=Top;");
  89     else if (atp->index() == Compile::AliasIdxRaw)
  90       st->print(", idx=Raw;");
  91     else {
  92       ciField* field = atp->field();
  93       if (field) {
  94         st->print(", name=");
  95         field->print_name_on(st);
  96       }
  97       st->print(", idx=%d;", atp->index());
  98     }
  99   }
 100 }
 101 
 102 extern void print_alias_types();
 103 
 104 #endif
 105 
 106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
 107   assert((t_oop != NULL), "sanity");
 108   bool is_instance = t_oop->is_known_instance_field();
 109   bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
 110                              (load != NULL) && load->is_Load() &&
 111                              (phase->is_IterGVN() != NULL);
 112   if (!(is_instance || is_boxed_value_load))
 113     return mchain;  // don't try to optimize non-instance types
 114   uint instance_id = t_oop->instance_id();
 115   Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
 116   Node *prev = NULL;
 117   Node *result = mchain;
 118   while (prev != result) {
 119     prev = result;
 120     if (result == start_mem)
 121       break;  // hit one of our sentinels
 122     // skip over a call which does not affect this memory slice
 123     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
 124       Node *proj_in = result->in(0);
 125       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
 126         break;  // hit one of our sentinels
 127       } else if (proj_in->is_Call()) {
 128         CallNode *call = proj_in->as_Call();
 129         if (!call->may_modify(t_oop, phase)) { // returns false for instances
 130           result = call->in(TypeFunc::Memory);
 131         }
 132       } else if (proj_in->is_Initialize()) {
 133         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
 134         // Stop if this is the initialization for the object instance which
 135         // which contains this memory slice, otherwise skip over it.
 136         if ((alloc == NULL) || (alloc->_idx == instance_id)) {
 137           break;
 138         }
 139         if (is_instance) {
 140           result = proj_in->in(TypeFunc::Memory);
 141         } else if (is_boxed_value_load) {
 142           Node* klass = alloc->in(AllocateNode::KlassNode);
 143           const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
 144           if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
 145             result = proj_in->in(TypeFunc::Memory); // not related allocation
 146           }
 147         }
 148       } else if (proj_in->is_MemBar()) {
 149         result = proj_in->in(TypeFunc::Memory);
 150       } else {
 151         assert(false, "unexpected projection");
 152       }
 153     } else if (result->is_ClearArray()) {
 154       if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
 155         // Can not bypass initialization of the instance
 156         // we are looking for.
 157         break;
 158       }
 159       // Otherwise skip it (the call updated 'result' value).
 160     } else if (result->is_MergeMem()) {
 161       result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
 162     }
 163   }
 164   return result;
 165 }
 166 
 167 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
 168   const TypeOopPtr* t_oop = t_adr->isa_oopptr();
 169   if (t_oop == NULL)
 170     return mchain;  // don't try to optimize non-oop types
 171   Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
 172   bool is_instance = t_oop->is_known_instance_field();
 173   PhaseIterGVN *igvn = phase->is_IterGVN();
 174   if (is_instance && igvn != NULL  && result->is_Phi()) {
 175     PhiNode *mphi = result->as_Phi();
 176     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 177     const TypePtr *t = mphi->adr_type();
 178     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
 179         t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
 180         t->is_oopptr()->cast_to_exactness(true)
 181          ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 182          ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
 183       // clone the Phi with our address type
 184       result = mphi->split_out_instance(t_adr, igvn);
 185     } else {
 186       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 187     }
 188   }
 189   return result;
 190 }
 191 
 192 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 193   uint alias_idx = phase->C->get_alias_index(tp);
 194   Node *mem = mmem;
 195 #ifdef ASSERT
 196   {
 197     // Check that current type is consistent with the alias index used during graph construction
 198     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 199     bool consistent =  adr_check == NULL || adr_check->empty() ||
 200                        phase->C->must_alias(adr_check, alias_idx );
 201     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 202     if( !consistent && adr_check != NULL && !adr_check->empty() &&
 203                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 204         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 205         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 206           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 207           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 208       // don't assert if it is dead code.
 209       consistent = true;
 210     }
 211     if( !consistent ) {
 212       st->print("alias_idx==%d, adr_check==", alias_idx);
 213       if( adr_check == NULL ) {
 214         st->print("NULL");
 215       } else {
 216         adr_check->dump();
 217       }
 218       st->cr();
 219       print_alias_types();
 220       assert(consistent, "adr_check must match alias idx");
 221     }
 222   }
 223 #endif
 224   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
 225   // means an array I have not precisely typed yet.  Do not do any
 226   // alias stuff with it any time soon.
 227   const TypeOopPtr *toop = tp->isa_oopptr();
 228   if( tp->base() != Type::AnyPtr &&
 229       !(toop &&
 230         toop->klass() != NULL &&
 231         toop->klass()->is_java_lang_Object() &&
 232         toop->offset() == Type::OffsetBot) ) {
 233     // compress paths and change unreachable cycles to TOP
 234     // If not, we can update the input infinitely along a MergeMem cycle
 235     // Equivalent code in PhiNode::Ideal
 236     Node* m  = phase->transform(mmem);
 237     // If transformed to a MergeMem, get the desired slice
 238     // Otherwise the returned node represents memory for every slice
 239     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
 240     // Update input if it is progress over what we have now
 241   }
 242   return mem;
 243 }
 244 
 245 //--------------------------Ideal_common---------------------------------------
 246 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
 247 // Unhook non-raw memories from complete (macro-expanded) initializations.
 248 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
 249   // If our control input is a dead region, kill all below the region
 250   Node *ctl = in(MemNode::Control);
 251   if (ctl && remove_dead_region(phase, can_reshape))
 252     return this;
 253   ctl = in(MemNode::Control);
 254   // Don't bother trying to transform a dead node
 255   if (ctl && ctl->is_top())  return NodeSentinel;
 256 
 257   PhaseIterGVN *igvn = phase->is_IterGVN();
 258   // Wait if control on the worklist.
 259   if (ctl && can_reshape && igvn != NULL) {
 260     Node* bol = NULL;
 261     Node* cmp = NULL;
 262     if (ctl->in(0)->is_If()) {
 263       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
 264       bol = ctl->in(0)->in(1);
 265       if (bol->is_Bool())
 266         cmp = ctl->in(0)->in(1)->in(1);
 267     }
 268     if (igvn->_worklist.member(ctl) ||
 269         (bol != NULL && igvn->_worklist.member(bol)) ||
 270         (cmp != NULL && igvn->_worklist.member(cmp)) ) {
 271       // This control path may be dead.
 272       // Delay this memory node transformation until the control is processed.
 273       phase->is_IterGVN()->_worklist.push(this);
 274       return NodeSentinel; // caller will return NULL
 275     }
 276   }
 277   // Ignore if memory is dead, or self-loop
 278   Node *mem = in(MemNode::Memory);
 279   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
 280   assert(mem != this, "dead loop in MemNode::Ideal");
 281 
 282   if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
 283     // This memory slice may be dead.
 284     // Delay this mem node transformation until the memory is processed.
 285     phase->is_IterGVN()->_worklist.push(this);
 286     return NodeSentinel; // caller will return NULL
 287   }
 288 
 289   Node *address = in(MemNode::Address);
 290   const Type *t_adr = phase->type(address);
 291   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return NULL
 292 
 293   if (can_reshape && igvn != NULL &&
 294       (igvn->_worklist.member(address) ||
 295        igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
 296     // The address's base and type may change when the address is processed.
 297     // Delay this mem node transformation until the address is processed.
 298     phase->is_IterGVN()->_worklist.push(this);
 299     return NodeSentinel; // caller will return NULL
 300   }
 301 
 302   // Do NOT remove or optimize the next lines: ensure a new alias index
 303   // is allocated for an oop pointer type before Escape Analysis.
 304   // Note: C++ will not remove it since the call has side effect.
 305   if (t_adr->isa_oopptr()) {
 306     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 307   }
 308 
 309 #ifdef ASSERT
 310   Node* base = NULL;
 311   if (address->is_AddP())
 312     base = address->in(AddPNode::Base);
 313   if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
 314       !t_adr->isa_rawptr()) {
 315     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
 316     Compile* C = phase->C;
 317     tty->cr();
 318     tty->print_cr("===== NULL+offs not RAW address =====");
 319     if (C->is_dead_node(this->_idx))    tty->print_cr("'this' is dead");
 320     if ((ctl != NULL) && C->is_dead_node(ctl->_idx)) tty->print_cr("'ctl' is dead");
 321     if (C->is_dead_node(mem->_idx))     tty->print_cr("'mem' is dead");
 322     if (C->is_dead_node(address->_idx)) tty->print_cr("'address' is dead");
 323     if (C->is_dead_node(base->_idx))    tty->print_cr("'base' is dead");
 324     tty->cr();
 325     base->dump(1);
 326     tty->cr();
 327     this->dump(2);
 328     tty->print("this->adr_type():     "); adr_type()->dump(); tty->cr();
 329     tty->print("phase->type(address): "); t_adr->dump(); tty->cr();
 330     tty->print("phase->type(base):    "); phase->type(address)->dump(); tty->cr();
 331     tty->cr();
 332   }
 333   assert(base == NULL || t_adr->isa_rawptr() ||
 334         !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
 335 #endif
 336 
 337   // Avoid independent memory operations
 338   Node* old_mem = mem;
 339 
 340   // The code which unhooks non-raw memories from complete (macro-expanded)
 341   // initializations was removed. After macro-expansion all stores catched
 342   // by Initialize node became raw stores and there is no information
 343   // which memory slices they modify. So it is unsafe to move any memory
 344   // operation above these stores. Also in most cases hooked non-raw memories
 345   // were already unhooked by using information from detect_ptr_independence()
 346   // and find_previous_store().
 347 
 348   if (mem->is_MergeMem()) {
 349     MergeMemNode* mmem = mem->as_MergeMem();
 350     const TypePtr *tp = t_adr->is_ptr();
 351 
 352     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 353   }
 354 
 355   if (mem != old_mem) {
 356     set_req(MemNode::Memory, mem);
 357     if (can_reshape && old_mem->outcnt() == 0) {
 358         igvn->_worklist.push(old_mem);
 359     }
 360     if (phase->type( mem ) == Type::TOP) return NodeSentinel;
 361     return this;
 362   }
 363 
 364   // let the subclass continue analyzing...
 365   return NULL;
 366 }
 367 
 368 // Helper function for proving some simple control dominations.
 369 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 370 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 371 // is not a constant (dominated by the method's StartNode).
 372 // Used by MemNode::find_previous_store to prove that the
 373 // control input of a memory operation predates (dominates)
 374 // an allocation it wants to look past.
 375 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
 376   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
 377     return false; // Conservative answer for dead code
 378 
 379   // Check 'dom'. Skip Proj and CatchProj nodes.
 380   dom = dom->find_exact_control(dom);
 381   if (dom == NULL || dom->is_top())
 382     return false; // Conservative answer for dead code
 383 
 384   if (dom == sub) {
 385     // For the case when, for example, 'sub' is Initialize and the original
 386     // 'dom' is Proj node of the 'sub'.
 387     return false;
 388   }
 389 
 390   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
 391     return true;
 392 
 393   // 'dom' dominates 'sub' if its control edge and control edges
 394   // of all its inputs dominate or equal to sub's control edge.
 395 
 396   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 397   // Or Region for the check in LoadNode::Ideal();
 398   // 'sub' should have sub->in(0) != NULL.
 399   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 400          sub->is_Region() || sub->is_Call(), "expecting only these nodes");
 401 
 402   // Get control edge of 'sub'.
 403   Node* orig_sub = sub;
 404   sub = sub->find_exact_control(sub->in(0));
 405   if (sub == NULL || sub->is_top())
 406     return false; // Conservative answer for dead code
 407 
 408   assert(sub->is_CFG(), "expecting control");
 409 
 410   if (sub == dom)
 411     return true;
 412 
 413   if (sub->is_Start() || sub->is_Root())
 414     return false;
 415 
 416   {
 417     // Check all control edges of 'dom'.
 418 
 419     ResourceMark rm;
 420     Arena* arena = Thread::current()->resource_area();
 421     Node_List nlist(arena);
 422     Unique_Node_List dom_list(arena);
 423 
 424     dom_list.push(dom);
 425     bool only_dominating_controls = false;
 426 
 427     for (uint next = 0; next < dom_list.size(); next++) {
 428       Node* n = dom_list.at(next);
 429       if (n == orig_sub)
 430         return false; // One of dom's inputs dominated by sub.
 431       if (!n->is_CFG() && n->pinned()) {
 432         // Check only own control edge for pinned non-control nodes.
 433         n = n->find_exact_control(n->in(0));
 434         if (n == NULL || n->is_top())
 435           return false; // Conservative answer for dead code
 436         assert(n->is_CFG(), "expecting control");
 437         dom_list.push(n);
 438       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 439         only_dominating_controls = true;
 440       } else if (n->is_CFG()) {
 441         if (n->dominates(sub, nlist))
 442           only_dominating_controls = true;
 443         else
 444           return false;
 445       } else {
 446         // First, own control edge.
 447         Node* m = n->find_exact_control(n->in(0));
 448         if (m != NULL) {
 449           if (m->is_top())
 450             return false; // Conservative answer for dead code
 451           dom_list.push(m);
 452         }
 453         // Now, the rest of edges.
 454         uint cnt = n->req();
 455         for (uint i = 1; i < cnt; i++) {
 456           m = n->find_exact_control(n->in(i));
 457           if (m == NULL || m->is_top())
 458             continue;
 459           dom_list.push(m);
 460         }
 461       }
 462     }
 463     return only_dominating_controls;
 464   }
 465 }
 466 
 467 //---------------------detect_ptr_independence---------------------------------
 468 // Used by MemNode::find_previous_store to prove that two base
 469 // pointers are never equal.
 470 // The pointers are accompanied by their associated allocations,
 471 // if any, which have been previously discovered by the caller.
 472 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 473                                       Node* p2, AllocateNode* a2,
 474                                       PhaseTransform* phase) {
 475   // Attempt to prove that these two pointers cannot be aliased.
 476   // They may both manifestly be allocations, and they should differ.
 477   // Or, if they are not both allocations, they can be distinct constants.
 478   // Otherwise, one is an allocation and the other a pre-existing value.
 479   if (a1 == NULL && a2 == NULL) {           // neither an allocation
 480     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 481   } else if (a1 != NULL && a2 != NULL) {    // both allocations
 482     return (a1 != a2);
 483   } else if (a1 != NULL) {                  // one allocation a1
 484     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 485     return all_controls_dominate(p2, a1);
 486   } else { //(a2 != NULL)                   // one allocation a2
 487     return all_controls_dominate(p1, a2);
 488   }
 489   return false;
 490 }
 491 
 492 
 493 // The logic for reordering loads and stores uses four steps:
 494 // (a) Walk carefully past stores and initializations which we
 495 //     can prove are independent of this load.
 496 // (b) Observe that the next memory state makes an exact match
 497 //     with self (load or store), and locate the relevant store.
 498 // (c) Ensure that, if we were to wire self directly to the store,
 499 //     the optimizer would fold it up somehow.
 500 // (d) Do the rewiring, and return, depending on some other part of
 501 //     the optimizer to fold up the load.
 502 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 503 // specific to loads and stores, so they are handled by the callers.
 504 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 505 //
 506 Node* MemNode::find_previous_store(PhaseTransform* phase) {
 507   Node*         ctrl   = in(MemNode::Control);
 508   Node*         adr    = in(MemNode::Address);
 509   intptr_t      offset = 0;
 510   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 511   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
 512 
 513   if (offset == Type::OffsetBot)
 514     return NULL;            // cannot unalias unless there are precise offsets
 515 
 516   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 517 
 518   intptr_t size_in_bytes = memory_size();
 519 
 520   Node* mem = in(MemNode::Memory);   // start searching here...
 521 
 522   int cnt = 50;             // Cycle limiter
 523   for (;;) {                // While we can dance past unrelated stores...
 524     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 525 
 526     if (mem->is_Store()) {
 527       Node* st_adr = mem->in(MemNode::Address);
 528       intptr_t st_offset = 0;
 529       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 530       if (st_base == NULL)
 531         break;              // inscrutable pointer
 532       if (st_offset != offset && st_offset != Type::OffsetBot) {
 533         const int MAX_STORE = BytesPerLong;
 534         if (st_offset >= offset + size_in_bytes ||
 535             st_offset <= offset - MAX_STORE ||
 536             st_offset <= offset - mem->as_Store()->memory_size()) {
 537           // Success:  The offsets are provably independent.
 538           // (You may ask, why not just test st_offset != offset and be done?
 539           // The answer is that stores of different sizes can co-exist
 540           // in the same sequence of RawMem effects.  We sometimes initialize
 541           // a whole 'tile' of array elements with a single jint or jlong.)
 542           mem = mem->in(MemNode::Memory);
 543           continue;           // (a) advance through independent store memory
 544         }
 545       }
 546       if (st_base != base &&
 547           detect_ptr_independence(base, alloc,
 548                                   st_base,
 549                                   AllocateNode::Ideal_allocation(st_base, phase),
 550                                   phase)) {
 551         // Success:  The bases are provably independent.
 552         mem = mem->in(MemNode::Memory);
 553         continue;           // (a) advance through independent store memory
 554       }
 555 
 556       // (b) At this point, if the bases or offsets do not agree, we lose,
 557       // since we have not managed to prove 'this' and 'mem' independent.
 558       if (st_base == base && st_offset == offset) {
 559         return mem;         // let caller handle steps (c), (d)
 560       }
 561 
 562     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 563       InitializeNode* st_init = mem->in(0)->as_Initialize();
 564       AllocateNode*  st_alloc = st_init->allocation();
 565       if (st_alloc == NULL)
 566         break;              // something degenerated
 567       bool known_identical = false;
 568       bool known_independent = false;
 569       if (alloc == st_alloc)
 570         known_identical = true;
 571       else if (alloc != NULL)
 572         known_independent = true;
 573       else if (all_controls_dominate(this, st_alloc))
 574         known_independent = true;
 575 
 576       if (known_independent) {
 577         // The bases are provably independent: Either they are
 578         // manifestly distinct allocations, or else the control
 579         // of this load dominates the store's allocation.
 580         int alias_idx = phase->C->get_alias_index(adr_type());
 581         if (alias_idx == Compile::AliasIdxRaw) {
 582           mem = st_alloc->in(TypeFunc::Memory);
 583         } else {
 584           mem = st_init->memory(alias_idx);
 585         }
 586         continue;           // (a) advance through independent store memory
 587       }
 588 
 589       // (b) at this point, if we are not looking at a store initializing
 590       // the same allocation we are loading from, we lose.
 591       if (known_identical) {
 592         // From caller, can_see_stored_value will consult find_captured_store.
 593         return mem;         // let caller handle steps (c), (d)
 594       }
 595 
 596     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
 597       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 598       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 599         CallNode *call = mem->in(0)->as_Call();
 600         if (!call->may_modify(addr_t, phase)) {
 601           mem = call->in(TypeFunc::Memory);
 602           continue;         // (a) advance through independent call memory
 603         }
 604       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 605         mem = mem->in(0)->in(TypeFunc::Memory);
 606         continue;           // (a) advance through independent MemBar memory
 607       } else if (mem->is_ClearArray()) {
 608         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 609           // (the call updated 'mem' value)
 610           continue;         // (a) advance through independent allocation memory
 611         } else {
 612           // Can not bypass initialization of the instance
 613           // we are looking for.
 614           return mem;
 615         }
 616       } else if (mem->is_MergeMem()) {
 617         int alias_idx = phase->C->get_alias_index(adr_type());
 618         mem = mem->as_MergeMem()->memory_at(alias_idx);
 619         continue;           // (a) advance through independent MergeMem memory
 620       }
 621     }
 622 
 623     // Unless there is an explicit 'continue', we must bail out here,
 624     // because 'mem' is an inscrutable memory state (e.g., a call).
 625     break;
 626   }
 627 
 628   return NULL;              // bail out
 629 }
 630 
 631 //----------------------calculate_adr_type-------------------------------------
 632 // Helper function.  Notices when the given type of address hits top or bottom.
 633 // Also, asserts a cross-check of the type against the expected address type.
 634 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 635   if (t == Type::TOP)  return NULL; // does not touch memory any more?
 636   #ifdef PRODUCT
 637   cross_check = NULL;
 638   #else
 639   if (!VerifyAliases || is_error_reported() || Node::in_dump())  cross_check = NULL;
 640   #endif
 641   const TypePtr* tp = t->isa_ptr();
 642   if (tp == NULL) {
 643     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 644     return TypePtr::BOTTOM;           // touches lots of memory
 645   } else {
 646     #ifdef ASSERT
 647     // %%%% [phh] We don't check the alias index if cross_check is
 648     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 649     if (cross_check != NULL &&
 650         cross_check != TypePtr::BOTTOM &&
 651         cross_check != TypeRawPtr::BOTTOM) {
 652       // Recheck the alias index, to see if it has changed (due to a bug).
 653       Compile* C = Compile::current();
 654       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 655              "must stay in the original alias category");
 656       // The type of the address must be contained in the adr_type,
 657       // disregarding "null"-ness.
 658       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 659       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 660       assert(cross_check->meet(tp_notnull) == cross_check,
 661              "real address must not escape from expected memory type");
 662     }
 663     #endif
 664     return tp;
 665   }
 666 }
 667 
 668 //------------------------adr_phi_is_loop_invariant----------------------------
 669 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
 670 // loop is loop invariant. Make a quick traversal of Phi and associated
 671 // CastPP nodes, looking to see if they are a closed group within the loop.
 672 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
 673   // The idea is that the phi-nest must boil down to only CastPP nodes
 674   // with the same data. This implies that any path into the loop already
 675   // includes such a CastPP, and so the original cast, whatever its input,
 676   // must be covered by an equivalent cast, with an earlier control input.
 677   ResourceMark rm;
 678 
 679   // The loop entry input of the phi should be the unique dominating
 680   // node for every Phi/CastPP in the loop.
 681   Unique_Node_List closure;
 682   closure.push(adr_phi->in(LoopNode::EntryControl));
 683 
 684   // Add the phi node and the cast to the worklist.
 685   Unique_Node_List worklist;
 686   worklist.push(adr_phi);
 687   if( cast != NULL ){
 688     if( !cast->is_ConstraintCast() ) return false;
 689     worklist.push(cast);
 690   }
 691 
 692   // Begin recursive walk of phi nodes.
 693   while( worklist.size() ){
 694     // Take a node off the worklist
 695     Node *n = worklist.pop();
 696     if( !closure.member(n) ){
 697       // Add it to the closure.
 698       closure.push(n);
 699       // Make a sanity check to ensure we don't waste too much time here.
 700       if( closure.size() > 20) return false;
 701       // This node is OK if:
 702       //  - it is a cast of an identical value
 703       //  - or it is a phi node (then we add its inputs to the worklist)
 704       // Otherwise, the node is not OK, and we presume the cast is not invariant
 705       if( n->is_ConstraintCast() ){
 706         worklist.push(n->in(1));
 707       } else if( n->is_Phi() ) {
 708         for( uint i = 1; i < n->req(); i++ ) {
 709           worklist.push(n->in(i));
 710         }
 711       } else {
 712         return false;
 713       }
 714     }
 715   }
 716 
 717   // Quit when the worklist is empty, and we've found no offending nodes.
 718   return true;
 719 }
 720 
 721 //------------------------------Ideal_DU_postCCP-------------------------------
 722 // Find any cast-away of null-ness and keep its control.  Null cast-aways are
 723 // going away in this pass and we need to make this memory op depend on the
 724 // gating null check.
 725 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
 726   return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
 727 }
 728 
 729 // I tried to leave the CastPP's in.  This makes the graph more accurate in
 730 // some sense; we get to keep around the knowledge that an oop is not-null
 731 // after some test.  Alas, the CastPP's interfere with GVN (some values are
 732 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
 733 // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed
 734 // some of the more trivial cases in the optimizer.  Removing more useless
 735 // Phi's started allowing Loads to illegally float above null checks.  I gave
 736 // up on this approach.  CNC 10/20/2000
 737 // This static method may be called not from MemNode (EncodePNode calls it).
 738 // Only the control edge of the node 'n' might be updated.
 739 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
 740   Node *skipped_cast = NULL;
 741   // Need a null check?  Regular static accesses do not because they are
 742   // from constant addresses.  Array ops are gated by the range check (which
 743   // always includes a NULL check).  Just check field ops.
 744   if( n->in(MemNode::Control) == NULL ) {
 745     // Scan upwards for the highest location we can place this memory op.
 746     while( true ) {
 747       switch( adr->Opcode() ) {
 748 
 749       case Op_AddP:             // No change to NULL-ness, so peek thru AddP's
 750         adr = adr->in(AddPNode::Base);
 751         continue;
 752 
 753       case Op_DecodeN:         // No change to NULL-ness, so peek thru
 754       case Op_DecodeNKlass:
 755         adr = adr->in(1);
 756         continue;
 757 
 758       case Op_EncodeP:
 759       case Op_EncodePKlass:
 760         // EncodeP node's control edge could be set by this method
 761         // when EncodeP node depends on CastPP node.
 762         //
 763         // Use its control edge for memory op because EncodeP may go away
 764         // later when it is folded with following or preceding DecodeN node.
 765         if (adr->in(0) == NULL) {
 766           // Keep looking for cast nodes.
 767           adr = adr->in(1);
 768           continue;
 769         }
 770         ccp->hash_delete(n);
 771         n->set_req(MemNode::Control, adr->in(0));
 772         ccp->hash_insert(n);
 773         return n;
 774 
 775       case Op_CastPP:
 776         // If the CastPP is useless, just peek on through it.
 777         if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
 778           // Remember the cast that we've peeked though. If we peek
 779           // through more than one, then we end up remembering the highest
 780           // one, that is, if in a loop, the one closest to the top.
 781           skipped_cast = adr;
 782           adr = adr->in(1);
 783           continue;
 784         }
 785         // CastPP is going away in this pass!  We need this memory op to be
 786         // control-dependent on the test that is guarding the CastPP.
 787         ccp->hash_delete(n);
 788         n->set_req(MemNode::Control, adr->in(0));
 789         ccp->hash_insert(n);
 790         return n;
 791 
 792       case Op_Phi:
 793         // Attempt to float above a Phi to some dominating point.
 794         if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
 795           // If we've already peeked through a Cast (which could have set the
 796           // control), we can't float above a Phi, because the skipped Cast
 797           // may not be loop invariant.
 798           if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
 799             adr = adr->in(1);
 800             continue;
 801           }
 802         }
 803 
 804         // Intentional fallthrough!
 805 
 806         // No obvious dominating point.  The mem op is pinned below the Phi
 807         // by the Phi itself.  If the Phi goes away (no true value is merged)
 808         // then the mem op can float, but not indefinitely.  It must be pinned
 809         // behind the controls leading to the Phi.
 810       case Op_CheckCastPP:
 811         // These usually stick around to change address type, however a
 812         // useless one can be elided and we still need to pick up a control edge
 813         if (adr->in(0) == NULL) {
 814           // This CheckCastPP node has NO control and is likely useless. But we
 815           // need check further up the ancestor chain for a control input to keep
 816           // the node in place. 4959717.
 817           skipped_cast = adr;
 818           adr = adr->in(1);
 819           continue;
 820         }
 821         ccp->hash_delete(n);
 822         n->set_req(MemNode::Control, adr->in(0));
 823         ccp->hash_insert(n);
 824         return n;
 825 
 826         // List of "safe" opcodes; those that implicitly block the memory
 827         // op below any null check.
 828       case Op_CastX2P:          // no null checks on native pointers
 829       case Op_Parm:             // 'this' pointer is not null
 830       case Op_LoadP:            // Loading from within a klass
 831       case Op_LoadN:            // Loading from within a klass
 832       case Op_LoadKlass:        // Loading from within a klass
 833       case Op_LoadNKlass:       // Loading from within a klass
 834       case Op_ConP:             // Loading from a klass
 835       case Op_ConN:             // Loading from a klass
 836       case Op_ConNKlass:        // Loading from a klass
 837       case Op_CreateEx:         // Sucking up the guts of an exception oop
 838       case Op_Con:              // Reading from TLS
 839       case Op_CMoveP:           // CMoveP is pinned
 840       case Op_CMoveN:           // CMoveN is pinned
 841         break;                  // No progress
 842 
 843       case Op_Proj:             // Direct call to an allocation routine
 844       case Op_SCMemProj:        // Memory state from store conditional ops
 845 #ifdef ASSERT
 846         {
 847           assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
 848           const Node* call = adr->in(0);
 849           if (call->is_CallJava()) {
 850             const CallJavaNode* call_java = call->as_CallJava();
 851             const TypeTuple *r = call_java->tf()->range();
 852             assert(r->cnt() > TypeFunc::Parms, "must return value");
 853             const Type* ret_type = r->field_at(TypeFunc::Parms);
 854             assert(ret_type && ret_type->isa_ptr(), "must return pointer");
 855             // We further presume that this is one of
 856             // new_instance_Java, new_array_Java, or
 857             // the like, but do not assert for this.
 858           } else if (call->is_Allocate()) {
 859             // similar case to new_instance_Java, etc.
 860           } else if (!call->is_CallLeaf()) {
 861             // Projections from fetch_oop (OSR) are allowed as well.
 862             ShouldNotReachHere();
 863           }
 864         }
 865 #endif
 866         break;
 867       default:
 868         ShouldNotReachHere();
 869       }
 870       break;
 871     }
 872   }
 873 
 874   return  NULL;               // No progress
 875 }
 876 
 877 
 878 //=============================================================================
 879 uint LoadNode::size_of() const { return sizeof(*this); }
 880 uint LoadNode::cmp( const Node &n ) const
 881 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
 882 const Type *LoadNode::bottom_type() const { return _type; }
 883 uint LoadNode::ideal_reg() const {
 884   return _type->ideal_reg();
 885 }
 886 
 887 #ifndef PRODUCT
 888 void LoadNode::dump_spec(outputStream *st) const {
 889   MemNode::dump_spec(st);
 890   if( !Verbose && !WizardMode ) {
 891     // standard dump does this in Verbose and WizardMode
 892     st->print(" #"); _type->dump_on(st);
 893   }
 894 }
 895 #endif
 896 
 897 #ifdef ASSERT
 898 //----------------------------is_immutable_value-------------------------------
 899 // Helper function to allow a raw load without control edge for some cases
 900 bool LoadNode::is_immutable_value(Node* adr) {
 901   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 902           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 903           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 904            in_bytes(JavaThread::osthread_offset())));
 905 }
 906 #endif
 907 
 908 //----------------------------LoadNode::make-----------------------------------
 909 // Polymorphic factory method:
 910 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
 911   Compile* C = gvn.C;
 912 
 913   // sanity check the alias category against the created node type
 914   assert(!(adr_type->isa_oopptr() &&
 915            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
 916          "use LoadKlassNode instead");
 917   assert(!(adr_type->isa_aryptr() &&
 918            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
 919          "use LoadRangeNode instead");
 920   // Check control edge of raw loads
 921   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
 922           // oop will be recorded in oop map if load crosses safepoint
 923           rt->isa_oopptr() || is_immutable_value(adr),
 924           "raw memory operations should have control edge");
 925   switch (bt) {
 926   case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int()    );
 927   case T_BYTE:    return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int()    );
 928   case T_INT:     return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int()    );
 929   case T_CHAR:    return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int()    );
 930   case T_SHORT:   return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int()    );
 931   case T_LONG:    return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long()   );
 932   case T_FLOAT:   return new (C) LoadFNode (ctl, mem, adr, adr_type, rt              );
 933   case T_DOUBLE:  return new (C) LoadDNode (ctl, mem, adr, adr_type, rt              );
 934   case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr()    );
 935   case T_OBJECT:
 936 #ifdef _LP64
 937     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 938       Node* load  = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
 939       return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
 940     } else
 941 #endif
 942     {
 943       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
 944       return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
 945     }
 946   }
 947   ShouldNotReachHere();
 948   return (LoadNode*)NULL;
 949 }
 950 
 951 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
 952   bool require_atomic = true;
 953   return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
 954 }
 955 
 956 
 957 
 958 
 959 //------------------------------hash-------------------------------------------
 960 uint LoadNode::hash() const {
 961   // unroll addition of interesting fields
 962   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
 963 }
 964 
 965 //---------------------------can_see_stored_value------------------------------
 966 // This routine exists to make sure this set of tests is done the same
 967 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
 968 // will change the graph shape in a way which makes memory alive twice at the
 969 // same time (uses the Oracle model of aliasing), then some
 970 // LoadXNode::Identity will fold things back to the equivalence-class model
 971 // of aliasing.
 972 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
 973   Node* ld_adr = in(MemNode::Address);
 974   intptr_t ld_off = 0;
 975   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 976   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
 977   Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
 978   // This is more general than load from boxing objects.
 979   if (phase->C->eliminate_boxing() && (atp != NULL) &&
 980       (atp->index() >= Compile::AliasIdxRaw) &&
 981       (atp->field() != NULL) && !atp->field()->is_volatile()) {
 982     uint alias_idx = atp->index();
 983     bool final = atp->field()->is_final();
 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 || opc == Op_MemBarAcquireLock)) ||
 995           opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder ||
 996           opc == Op_MemBarReleaseLock) {
 997         Node* mem = current->in(0)->in(TypeFunc::Memory);
 998         if (mem->is_MergeMem()) {
 999           MergeMemNode* merge = mem->as_MergeMem();
1000           Node* new_st = merge->memory_at(alias_idx);
1001           if (new_st == merge->base_memory()) {
1002             // Keep searching
1003             current = new_st;
1004             continue;
1005           }
1006           // Save the new memory state for the slice and fall through
1007           // to exit.
1008           result = new_st;
1009         }
1010       }
1011       break;
1012     }
1013     if (result != NULL) {
1014       st = result;
1015     }
1016   }
1017 
1018 
1019   // Loop around twice in the case Load -> Initialize -> Store.
1020   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1021   for (int trip = 0; trip <= 1; trip++) {
1022 
1023     if (st->is_Store()) {
1024       Node* st_adr = st->in(MemNode::Address);
1025       if (!phase->eqv(st_adr, ld_adr)) {
1026         // Try harder before giving up...  Match raw and non-raw pointers.
1027         intptr_t st_off = 0;
1028         AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1029         if (alloc == NULL)       return NULL;
1030         if (alloc != ld_alloc)   return NULL;
1031         if (ld_off != st_off)    return NULL;
1032         // At this point we have proven something like this setup:
1033         //  A = Allocate(...)
1034         //  L = LoadQ(,  AddP(CastPP(, A.Parm),, #Off))
1035         //  S = StoreQ(, AddP(,        A.Parm  , #Off), V)
1036         // (Actually, we haven't yet proven the Q's are the same.)
1037         // In other words, we are loading from a casted version of
1038         // the same pointer-and-offset that we stored to.
1039         // Thus, we are able to replace L by V.
1040       }
1041       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1042       if (store_Opcode() != st->Opcode())
1043         return NULL;
1044       return st->in(MemNode::ValueIn);
1045     }
1046 
1047     // A load from a freshly-created object always returns zero.
1048     // (This can happen after LoadNode::Ideal resets the load's memory input
1049     // to find_captured_store, which returned InitializeNode::zero_memory.)
1050     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1051         (st->in(0) == ld_alloc) &&
1052         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1053       // return a zero value for the load's basic type
1054       // (This is one of the few places where a generic PhaseTransform
1055       // can create new nodes.  Think of it as lazily manifesting
1056       // virtually pre-existing constants.)
1057       return phase->zerocon(memory_type());
1058     }
1059 
1060     // A load from an initialization barrier can match a captured store.
1061     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1062       InitializeNode* init = st->in(0)->as_Initialize();
1063       AllocateNode* alloc = init->allocation();
1064       if ((alloc != NULL) && (alloc == ld_alloc)) {
1065         // examine a captured store value
1066         st = init->find_captured_store(ld_off, memory_size(), phase);
1067         if (st != NULL)
1068           continue;             // take one more trip around
1069       }
1070     }
1071 
1072     // Load boxed value from result of valueOf() call is input parameter.
1073     if (this->is_Load() && ld_adr->is_AddP() &&
1074         (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1075       intptr_t ignore = 0;
1076       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1077       if (base != NULL && base->is_Proj() &&
1078           base->as_Proj()->_con == TypeFunc::Parms &&
1079           base->in(0)->is_CallStaticJava() &&
1080           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1081         return base->in(0)->in(TypeFunc::Parms);
1082       }
1083     }
1084 
1085     break;
1086   }
1087 
1088   return NULL;
1089 }
1090 
1091 //----------------------is_instance_field_load_with_local_phi------------------
1092 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1093   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1094       in(Address)->is_AddP() ) {
1095     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1096     // Only instances and boxed values.
1097     if( t_oop != NULL &&
1098         (t_oop->is_ptr_to_boxed_value() ||
1099          t_oop->is_known_instance_field()) &&
1100         t_oop->offset() != Type::OffsetBot &&
1101         t_oop->offset() != Type::OffsetTop) {
1102       return true;
1103     }
1104   }
1105   return false;
1106 }
1107 
1108 //------------------------------Identity---------------------------------------
1109 // Loads are identity if previous store is to same address
1110 Node *LoadNode::Identity( PhaseTransform *phase ) {
1111   // If the previous store-maker is the right kind of Store, and the store is
1112   // to the same address, then we are equal to the value stored.
1113   Node* mem = in(Memory);
1114   Node* value = can_see_stored_value(mem, phase);
1115   if( value ) {
1116     // byte, short & char stores truncate naturally.
1117     // A load has to load the truncated value which requires
1118     // some sort of masking operation and that requires an
1119     // Ideal call instead of an Identity call.
1120     if (memory_size() < BytesPerInt) {
1121       // If the input to the store does not fit with the load's result type,
1122       // it must be truncated via an Ideal call.
1123       if (!phase->type(value)->higher_equal(phase->type(this)))
1124         return this;
1125     }
1126     // (This works even when value is a Con, but LoadNode::Value
1127     // usually runs first, producing the singleton type of the Con.)
1128     return value;
1129   }
1130 
1131   // Search for an existing data phi which was generated before for the same
1132   // instance's field to avoid infinite generation of phis in a loop.
1133   Node *region = mem->in(0);
1134   if (is_instance_field_load_with_local_phi(region)) {
1135     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1136     int this_index  = phase->C->get_alias_index(addr_t);
1137     int this_offset = addr_t->offset();
1138     int this_iid    = addr_t->instance_id();
1139     if (!addr_t->is_known_instance() &&
1140          addr_t->is_ptr_to_boxed_value()) {
1141       // Use _idx of address base (could be Phi node) for boxed values.
1142       intptr_t   ignore = 0;
1143       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1144       this_iid = base->_idx;
1145     }
1146     const Type* this_type = bottom_type();
1147     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1148       Node* phi = region->fast_out(i);
1149       if (phi->is_Phi() && phi != mem &&
1150           phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1151         return phi;
1152       }
1153     }
1154   }
1155 
1156   return this;
1157 }
1158 
1159 // We're loading from an object which has autobox behaviour.
1160 // If this object is result of a valueOf call we'll have a phi
1161 // merging a newly allocated object and a load from the cache.
1162 // We want to replace this load with the original incoming
1163 // argument to the valueOf call.
1164 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1165   assert(phase->C->eliminate_boxing(), "sanity");
1166   intptr_t ignore = 0;
1167   Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1168   if ((base == NULL) || base->is_Phi()) {
1169     // Push the loads from the phi that comes from valueOf up
1170     // through it to allow elimination of the loads and the recovery
1171     // of the original value. It is done in split_through_phi().
1172     return NULL;
1173   } else if (base->is_Load() ||
1174              base->is_DecodeN() && base->in(1)->is_Load()) {
1175     // Eliminate the load of boxed value for integer types from the cache
1176     // array by deriving the value from the index into the array.
1177     // Capture the offset of the load and then reverse the computation.
1178 
1179     // Get LoadN node which loads a boxing object from 'cache' array.
1180     if (base->is_DecodeN()) {
1181       base = base->in(1);
1182     }
1183     if (!base->in(Address)->is_AddP()) {
1184       return NULL; // Complex address
1185     }
1186     AddPNode* address = base->in(Address)->as_AddP();
1187     Node* cache_base = address->in(AddPNode::Base);
1188     if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1189       // Get ConP node which is static 'cache' field.
1190       cache_base = cache_base->in(1);
1191     }
1192     if ((cache_base != NULL) && cache_base->is_Con()) {
1193       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1194       if ((base_type != NULL) && base_type->is_autobox_cache()) {
1195         Node* elements[4];
1196         int shift = exact_log2(type2aelembytes(T_OBJECT));
1197         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1198         if ((count >  0) && elements[0]->is_Con() &&
1199             ((count == 1) ||
1200              (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1201                              elements[1]->in(2) == phase->intcon(shift))) {
1202           ciObjArray* array = base_type->const_oop()->as_obj_array();
1203           // Fetch the box object cache[0] at the base of the array and get its value
1204           ciInstance* box = array->obj_at(0)->as_instance();
1205           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1206           assert(ik->is_box_klass(), "sanity");
1207           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1208           if (ik->nof_nonstatic_fields() == 1) {
1209             // This should be true nonstatic_field_at requires calling
1210             // nof_nonstatic_fields so check it anyway
1211             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1212             BasicType bt = c.basic_type();
1213             // Only integer types have boxing cache.
1214             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1215                    bt == T_BYTE    || bt == T_SHORT ||
1216                    bt == T_INT     || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1217             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1218             if (cache_low != (int)cache_low) {
1219               return NULL; // should not happen since cache is array indexed by value
1220             }
1221             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1222             if (offset != (int)offset) {
1223               return NULL; // should not happen since cache is array indexed by value
1224             }
1225            // Add up all the offsets making of the address of the load
1226             Node* result = elements[0];
1227             for (int i = 1; i < count; i++) {
1228               result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1229             }
1230             // Remove the constant offset from the address and then
1231             result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));
1232             // remove the scaling of the offset to recover the original index.
1233             if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1234               // Peel the shift off directly but wrap it in a dummy node
1235               // since Ideal can't return existing nodes
1236               result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1237             } else if (result->is_Add() && result->in(2)->is_Con() &&
1238                        result->in(1)->Opcode() == Op_LShiftX &&
1239                        result->in(1)->in(2) == phase->intcon(shift)) {
1240               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1241               // but for boxing cache access we know that X<<Z will not overflow
1242               // (there is range check) so we do this optimizatrion by hand here.
1243               Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));
1244               result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));
1245             } else {
1246               result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1247             }
1248 #ifdef _LP64
1249             if (bt != T_LONG) {
1250               result = new (phase->C) ConvL2INode(phase->transform(result));
1251             }
1252 #else
1253             if (bt == T_LONG) {
1254               result = new (phase->C) ConvI2LNode(phase->transform(result));
1255             }
1256 #endif
1257             return result;
1258           }
1259         }
1260       }
1261     }
1262   }
1263   return NULL;
1264 }
1265 
1266 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1267   Node* region = phi->in(0);
1268   if (region == NULL) {
1269     return false; // Wait stable graph
1270   }
1271   uint cnt = phi->req();
1272   for (uint i = 1; i < cnt; i++) {
1273     Node* rc = region->in(i);
1274     if (rc == NULL || phase->type(rc) == Type::TOP)
1275       return false; // Wait stable graph
1276     Node* in = phi->in(i);
1277     if (in == NULL || phase->type(in) == Type::TOP)
1278       return false; // Wait stable graph
1279   }
1280   return true;
1281 }
1282 //------------------------------split_through_phi------------------------------
1283 // Split instance or boxed field load through Phi.
1284 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1285   Node* mem     = in(Memory);
1286   Node* address = in(Address);
1287   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1288 
1289   assert((t_oop != NULL) &&
1290          (t_oop->is_known_instance_field() ||
1291           t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1292 
1293   Compile* C = phase->C;
1294   intptr_t ignore = 0;
1295   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1296   bool base_is_phi = (base != NULL) && base->is_Phi();
1297   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1298                            (base != NULL) && (base == address->in(AddPNode::Base)) &&
1299                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1300 
1301   if (!((mem->is_Phi() || base_is_phi) &&
1302         (load_boxed_values || t_oop->is_known_instance_field()))) {
1303     return NULL; // memory is not Phi
1304   }
1305 
1306   if (mem->is_Phi()) {
1307     if (!stable_phi(mem->as_Phi(), phase)) {
1308       return NULL; // Wait stable graph
1309     }
1310     uint cnt = mem->req();
1311     // Check for loop invariant memory.
1312     if (cnt == 3) {
1313       for (uint i = 1; i < cnt; i++) {
1314         Node* in = mem->in(i);
1315         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1316         if (m == mem) {
1317           set_req(Memory, mem->in(cnt - i));
1318           return this; // made change
1319         }
1320       }
1321     }
1322   }
1323   if (base_is_phi) {
1324     if (!stable_phi(base->as_Phi(), phase)) {
1325       return NULL; // Wait stable graph
1326     }
1327     uint cnt = base->req();
1328     // Check for loop invariant memory.
1329     if (cnt == 3) {
1330       for (uint i = 1; i < cnt; i++) {
1331         if (base->in(i) == base) {
1332           return NULL; // Wait stable graph
1333         }
1334       }
1335     }
1336   }
1337 
1338   bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1339 
1340   // Split through Phi (see original code in loopopts.cpp).
1341   assert(C->have_alias_type(t_oop), "instance should have alias type");
1342 
1343   // Do nothing here if Identity will find a value
1344   // (to avoid infinite chain of value phis generation).
1345   if (!phase->eqv(this, this->Identity(phase)))
1346     return NULL;
1347 
1348   // Select Region to split through.
1349   Node* region;
1350   if (!base_is_phi) {
1351     assert(mem->is_Phi(), "sanity");
1352     region = mem->in(0);
1353     // Skip if the region dominates some control edge of the address.
1354     if (!MemNode::all_controls_dominate(address, region))
1355       return NULL;
1356   } else if (!mem->is_Phi()) {
1357     assert(base_is_phi, "sanity");
1358     region = base->in(0);
1359     // Skip if the region dominates some control edge of the memory.
1360     if (!MemNode::all_controls_dominate(mem, region))
1361       return NULL;
1362   } else if (base->in(0) != mem->in(0)) {
1363     assert(base_is_phi && mem->is_Phi(), "sanity");
1364     if (MemNode::all_controls_dominate(mem, base->in(0))) {
1365       region = base->in(0);
1366     } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1367       region = mem->in(0);
1368     } else {
1369       return NULL; // complex graph
1370     }
1371   } else {
1372     assert(base->in(0) == mem->in(0), "sanity");
1373     region = mem->in(0);
1374   }
1375 
1376   const Type* this_type = this->bottom_type();
1377   int this_index  = C->get_alias_index(t_oop);
1378   int this_offset = t_oop->offset();
1379   int this_iid    = t_oop->instance_id();
1380   if (!t_oop->is_known_instance() && load_boxed_values) {
1381     // Use _idx of address base for boxed values.
1382     this_iid = base->_idx;
1383   }
1384   PhaseIterGVN* igvn = phase->is_IterGVN();
1385   Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1386   for (uint i = 1; i < region->req(); i++) {
1387     Node* x;
1388     Node* the_clone = NULL;
1389     if (region->in(i) == C->top()) {
1390       x = C->top();      // Dead path?  Use a dead data op
1391     } else {
1392       x = this->clone();        // Else clone up the data op
1393       the_clone = x;            // Remember for possible deletion.
1394       // Alter data node to use pre-phi inputs
1395       if (this->in(0) == region) {
1396         x->set_req(0, region->in(i));
1397       } else {
1398         x->set_req(0, NULL);
1399       }
1400       if (mem->is_Phi() && (mem->in(0) == region)) {
1401         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1402       }
1403       if (address->is_Phi() && address->in(0) == region) {
1404         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1405       }
1406       if (base_is_phi && (base->in(0) == region)) {
1407         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1408         Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1409         x->set_req(Address, adr_x);
1410       }
1411     }
1412     // Check for a 'win' on some paths
1413     const Type *t = x->Value(igvn);
1414 
1415     bool singleton = t->singleton();
1416 
1417     // See comments in PhaseIdealLoop::split_thru_phi().
1418     if (singleton && t == Type::TOP) {
1419       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1420     }
1421 
1422     if (singleton) {
1423       x = igvn->makecon(t);
1424     } else {
1425       // We now call Identity to try to simplify the cloned node.
1426       // Note that some Identity methods call phase->type(this).
1427       // Make sure that the type array is big enough for
1428       // our new node, even though we may throw the node away.
1429       // (This tweaking with igvn only works because x is a new node.)
1430       igvn->set_type(x, t);
1431       // If x is a TypeNode, capture any more-precise type permanently into Node
1432       // otherwise it will be not updated during igvn->transform since
1433       // igvn->type(x) is set to x->Value() already.
1434       x->raise_bottom_type(t);
1435       Node *y = x->Identity(igvn);
1436       if (y != x) {
1437         x = y;
1438       } else {
1439         y = igvn->hash_find_insert(x);
1440         if (y) {
1441           x = y;
1442         } else {
1443           // Else x is a new node we are keeping
1444           // We do not need register_new_node_with_optimizer
1445           // because set_type has already been called.
1446           igvn->_worklist.push(x);
1447         }
1448       }
1449     }
1450     if (x != the_clone && the_clone != NULL) {
1451       igvn->remove_dead_node(the_clone);
1452     }
1453     phi->set_req(i, x);
1454   }
1455   // Record Phi
1456   igvn->register_new_node_with_optimizer(phi);
1457   return phi;
1458 }
1459 
1460 //------------------------------Ideal------------------------------------------
1461 // If the load is from Field memory and the pointer is non-null, we can
1462 // zero out the control input.
1463 // If the offset is constant and the base is an object allocation,
1464 // try to hook me up to the exact initializing store.
1465 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1466   Node* p = MemNode::Ideal_common(phase, can_reshape);
1467   if (p)  return (p == NodeSentinel) ? NULL : p;
1468 
1469   Node* ctrl    = in(MemNode::Control);
1470   Node* address = in(MemNode::Address);
1471 
1472   // Skip up past a SafePoint control.  Cannot do this for Stores because
1473   // pointer stores & cardmarks must stay on the same side of a SafePoint.
1474   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1475       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1476     ctrl = ctrl->in(0);
1477     set_req(MemNode::Control,ctrl);
1478   }
1479 
1480   intptr_t ignore = 0;
1481   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1482   if (base != NULL
1483       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1484     // Check for useless control edge in some common special cases
1485     if (in(MemNode::Control) != NULL
1486         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1487         && all_controls_dominate(base, phase->C->start())) {
1488       // A method-invariant, non-null address (constant or 'this' argument).
1489       set_req(MemNode::Control, NULL);
1490     }
1491   }
1492 
1493   Node* mem = in(MemNode::Memory);
1494   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1495 
1496   if (can_reshape && (addr_t != NULL)) {
1497     // try to optimize our memory input
1498     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1499     if (opt_mem != mem) {
1500       set_req(MemNode::Memory, opt_mem);
1501       if (phase->type( opt_mem ) == Type::TOP) return NULL;
1502       return this;
1503     }
1504     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1505     if ((t_oop != NULL) &&
1506         (t_oop->is_known_instance_field() ||
1507          t_oop->is_ptr_to_boxed_value())) {
1508       PhaseIterGVN *igvn = phase->is_IterGVN();
1509       if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1510         // Delay this transformation until memory Phi is processed.
1511         phase->is_IterGVN()->_worklist.push(this);
1512         return NULL;
1513       }
1514       // Split instance field load through Phi.
1515       Node* result = split_through_phi(phase);
1516       if (result != NULL) return result;
1517 
1518       if (t_oop->is_ptr_to_boxed_value()) {
1519         Node* result = eliminate_autobox(phase);
1520         if (result != NULL) return result;
1521       }
1522     }
1523   }
1524 
1525   // Check for prior store with a different base or offset; make Load
1526   // independent.  Skip through any number of them.  Bail out if the stores
1527   // are in an endless dead cycle and report no progress.  This is a key
1528   // transform for Reflection.  However, if after skipping through the Stores
1529   // we can't then fold up against a prior store do NOT do the transform as
1530   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
1531   // array memory alive twice: once for the hoisted Load and again after the
1532   // bypassed Store.  This situation only works if EVERYBODY who does
1533   // anti-dependence work knows how to bypass.  I.e. we need all
1534   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
1535   // the alias index stuff.  So instead, peek through Stores and IFF we can
1536   // fold up, do so.
1537   Node* prev_mem = find_previous_store(phase);
1538   // Steps (a), (b):  Walk past independent stores to find an exact match.
1539   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1540     // (c) See if we can fold up on the spot, but don't fold up here.
1541     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1542     // just return a prior value, which is done by Identity calls.
1543     if (can_see_stored_value(prev_mem, phase)) {
1544       // Make ready for step (d):
1545       set_req(MemNode::Memory, prev_mem);
1546       return this;
1547     }
1548   }
1549 
1550   return NULL;                  // No further progress
1551 }
1552 
1553 // Helper to recognize certain Klass fields which are invariant across
1554 // some group of array types (e.g., int[] or all T[] where T < Object).
1555 const Type*
1556 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1557                                  ciKlass* klass) const {
1558   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1559     // The field is Klass::_modifier_flags.  Return its (constant) value.
1560     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1561     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1562     return TypeInt::make(klass->modifier_flags());
1563   }
1564   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1565     // The field is Klass::_access_flags.  Return its (constant) value.
1566     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1567     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1568     return TypeInt::make(klass->access_flags());
1569   }
1570   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1571     // The field is Klass::_layout_helper.  Return its constant value if known.
1572     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1573     return TypeInt::make(klass->layout_helper());
1574   }
1575 
1576   // No match.
1577   return NULL;
1578 }
1579 
1580 //------------------------------Value-----------------------------------------
1581 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1582   // Either input is TOP ==> the result is TOP
1583   Node* mem = in(MemNode::Memory);
1584   const Type *t1 = phase->type(mem);
1585   if (t1 == Type::TOP)  return Type::TOP;
1586   Node* adr = in(MemNode::Address);
1587   const TypePtr* tp = phase->type(adr)->isa_ptr();
1588   if (tp == NULL || tp->empty())  return Type::TOP;
1589   int off = tp->offset();
1590   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1591   Compile* C = phase->C;
1592 
1593   // Try to guess loaded type from pointer type
1594   if (tp->base() == Type::AryPtr) {
1595     const Type *t = tp->is_aryptr()->elem();
1596     // Don't do this for integer types. There is only potential profit if
1597     // the element type t is lower than _type; that is, for int types, if _type is
1598     // more restrictive than t.  This only happens here if one is short and the other
1599     // char (both 16 bits), and in those cases we've made an intentional decision
1600     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1601     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1602     //
1603     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1604     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
1605     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1606     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
1607     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1608     // In fact, that could have been the original type of p1, and p1 could have
1609     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1610     // expression (LShiftL quux 3) independently optimized to the constant 8.
1611     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1612         && (_type->isa_vect() == NULL)
1613         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1614       // t might actually be lower than _type, if _type is a unique
1615       // concrete subclass of abstract class t.
1616       // Make sure the reference is not into the header, by comparing
1617       // the offset against the offset of the start of the array's data.
1618       // Different array types begin at slightly different offsets (12 vs. 16).
1619       // We choose T_BYTE as an example base type that is least restrictive
1620       // as to alignment, which will therefore produce the smallest
1621       // possible base offset.
1622       const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1623       if ((uint)off >= (uint)min_base_off) {  // is the offset beyond the header?
1624         const Type* jt = t->join(_type);
1625         // In any case, do not allow the join, per se, to empty out the type.
1626         if (jt->empty() && !t->empty()) {
1627           // This can happen if a interface-typed array narrows to a class type.
1628           jt = _type;
1629         }
1630 #ifdef ASSERT
1631         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1632           // The pointers in the autobox arrays are always non-null
1633           Node* base = adr->in(AddPNode::Base);
1634           if ((base != NULL) && base->is_DecodeN()) {
1635             // Get LoadN node which loads IntegerCache.cache field
1636             base = base->in(1);
1637           }
1638           if ((base != NULL) && base->is_Con()) {
1639             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1640             if ((base_type != NULL) && base_type->is_autobox_cache()) {
1641               // It could be narrow oop
1642               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1643             }
1644           }
1645         }
1646 #endif
1647         return jt;
1648       }
1649     }
1650   } else if (tp->base() == Type::InstPtr) {
1651     ciEnv* env = C->env();
1652     const TypeInstPtr* tinst = tp->is_instptr();
1653     ciKlass* klass = tinst->klass();
1654     assert( off != Type::OffsetBot ||
1655             // arrays can be cast to Objects
1656             tp->is_oopptr()->klass()->is_java_lang_Object() ||
1657             // unsafe field access may not have a constant offset
1658             C->has_unsafe_access(),
1659             "Field accesses must be precise" );
1660     // For oop loads, we expect the _type to be precise
1661     if (klass == env->String_klass() &&
1662         adr->is_AddP() && off != Type::OffsetBot) {
1663       // For constant Strings treat the final fields as compile time constants.
1664       Node* base = adr->in(AddPNode::Base);
1665       const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1666       if (t != NULL && t->singleton()) {
1667         ciField* field = env->String_klass()->get_field_by_offset(off, false);
1668         if (field != NULL && field->is_final()) {
1669           ciObject* string = t->const_oop();
1670           ciConstant constant = string->as_instance()->field_value(field);
1671           if (constant.basic_type() == T_INT) {
1672             return TypeInt::make(constant.as_int());
1673           } else if (constant.basic_type() == T_ARRAY) {
1674             if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1675               return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1676             } else {
1677               return TypeOopPtr::make_from_constant(constant.as_object(), true);
1678             }
1679           }
1680         }
1681       }
1682     }
1683     // Optimizations for constant objects
1684     ciObject* const_oop = tinst->const_oop();
1685     if (const_oop != NULL) {
1686       // For constant Boxed value treat the target field as a compile time constant.
1687       if (tinst->is_ptr_to_boxed_value()) {
1688         return tinst->get_const_boxed_value();
1689       } else
1690       // For constant CallSites treat the target field as a compile time constant.
1691       if (const_oop->is_call_site()) {
1692         ciCallSite* call_site = const_oop->as_call_site();
1693         ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1694         if (field != NULL && field->is_call_site_target()) {
1695           ciMethodHandle* target = call_site->get_target();
1696           if (target != NULL) {  // just in case
1697             ciConstant constant(T_OBJECT, target);
1698             const Type* t;
1699             if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1700               t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1701             } else {
1702               t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1703             }
1704             // Add a dependence for invalidation of the optimization.
1705             if (!call_site->is_constant_call_site()) {
1706               C->dependencies()->assert_call_site_target_value(call_site, target);
1707             }
1708             return t;
1709           }
1710         }
1711       }
1712     }
1713   } else if (tp->base() == Type::KlassPtr) {
1714     assert( off != Type::OffsetBot ||
1715             // arrays can be cast to Objects
1716             tp->is_klassptr()->klass()->is_java_lang_Object() ||
1717             // also allow array-loading from the primary supertype
1718             // array during subtype checks
1719             Opcode() == Op_LoadKlass,
1720             "Field accesses must be precise" );
1721     // For klass/static loads, we expect the _type to be precise
1722   }
1723 
1724   const TypeKlassPtr *tkls = tp->isa_klassptr();
1725   if (tkls != NULL && !StressReflectiveCode) {
1726     ciKlass* klass = tkls->klass();
1727     if (klass->is_loaded() && tkls->klass_is_exact()) {
1728       // We are loading a field from a Klass metaobject whose identity
1729       // is known at compile time (the type is "exact" or "precise").
1730       // Check for fields we know are maintained as constants by the VM.
1731       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1732         // The field is Klass::_super_check_offset.  Return its (constant) value.
1733         // (Folds up type checking code.)
1734         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1735         return TypeInt::make(klass->super_check_offset());
1736       }
1737       // Compute index into primary_supers array
1738       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1739       // Check for overflowing; use unsigned compare to handle the negative case.
1740       if( depth < ciKlass::primary_super_limit() ) {
1741         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1742         // (Folds up type checking code.)
1743         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1744         ciKlass *ss = klass->super_of_depth(depth);
1745         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1746       }
1747       const Type* aift = load_array_final_field(tkls, klass);
1748       if (aift != NULL)  return aift;
1749       if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1750           && klass->is_array_klass()) {
1751         // The field is ArrayKlass::_component_mirror.  Return its (constant) value.
1752         // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1753         assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1754         return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1755       }
1756       if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1757         // The field is Klass::_java_mirror.  Return its (constant) value.
1758         // (Folds up the 2nd indirection in anObjConstant.getClass().)
1759         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1760         return TypeInstPtr::make(klass->java_mirror());
1761       }
1762     }
1763 
1764     // We can still check if we are loading from the primary_supers array at a
1765     // shallow enough depth.  Even though the klass is not exact, entries less
1766     // than or equal to its super depth are correct.
1767     if (klass->is_loaded() ) {
1768       ciType *inner = klass;
1769       while( inner->is_obj_array_klass() )
1770         inner = inner->as_obj_array_klass()->base_element_type();
1771       if( inner->is_instance_klass() &&
1772           !inner->as_instance_klass()->flags().is_interface() ) {
1773         // Compute index into primary_supers array
1774         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1775         // Check for overflowing; use unsigned compare to handle the negative case.
1776         if( depth < ciKlass::primary_super_limit() &&
1777             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1778           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1779           // (Folds up type checking code.)
1780           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1781           ciKlass *ss = klass->super_of_depth(depth);
1782           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1783         }
1784       }
1785     }
1786 
1787     // If the type is enough to determine that the thing is not an array,
1788     // we can give the layout_helper a positive interval type.
1789     // This will help short-circuit some reflective code.
1790     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1791         && !klass->is_array_klass() // not directly typed as an array
1792         && !klass->is_interface()  // specifically not Serializable & Cloneable
1793         && !klass->is_java_lang_Object()   // not the supertype of all T[]
1794         ) {
1795       // Note:  When interfaces are reliable, we can narrow the interface
1796       // test to (klass != Serializable && klass != Cloneable).
1797       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1798       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1799       // The key property of this type is that it folds up tests
1800       // for array-ness, since it proves that the layout_helper is positive.
1801       // Thus, a generic value like the basic object layout helper works fine.
1802       return TypeInt::make(min_size, max_jint, Type::WidenMin);
1803     }
1804   }
1805 
1806   // If we are loading from a freshly-allocated object, produce a zero,
1807   // if the load is provably beyond the header of the object.
1808   // (Also allow a variable load from a fresh array to produce zero.)
1809   const TypeOopPtr *tinst = tp->isa_oopptr();
1810   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1811   bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1812   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1813     Node* value = can_see_stored_value(mem,phase);
1814     if (value != NULL && value->is_Con()) {
1815       assert(value->bottom_type()->higher_equal(_type),"sanity");
1816       return value->bottom_type();
1817     }
1818   }
1819 
1820   if (is_instance) {
1821     // If we have an instance type and our memory input is the
1822     // programs's initial memory state, there is no matching store,
1823     // so just return a zero of the appropriate type
1824     Node *mem = in(MemNode::Memory);
1825     if (mem->is_Parm() && mem->in(0)->is_Start()) {
1826       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1827       return Type::get_zero_type(_type->basic_type());
1828     }
1829   }
1830   return _type;
1831 }
1832 
1833 //------------------------------match_edge-------------------------------------
1834 // Do we Match on this edge index or not?  Match only the address.
1835 uint LoadNode::match_edge(uint idx) const {
1836   return idx == MemNode::Address;
1837 }
1838 
1839 //--------------------------LoadBNode::Ideal--------------------------------------
1840 //
1841 //  If the previous store is to the same address as this load,
1842 //  and the value stored was larger than a byte, replace this load
1843 //  with the value stored truncated to a byte.  If no truncation is
1844 //  needed, the replacement is done in LoadNode::Identity().
1845 //
1846 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1847   Node* mem = in(MemNode::Memory);
1848   Node* value = can_see_stored_value(mem,phase);
1849   if( value && !phase->type(value)->higher_equal( _type ) ) {
1850     Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1851     return new (phase->C) RShiftINode(result, phase->intcon(24));
1852   }
1853   // Identity call will handle the case where truncation is not needed.
1854   return LoadNode::Ideal(phase, can_reshape);
1855 }
1856 
1857 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1858   Node* mem = in(MemNode::Memory);
1859   Node* value = can_see_stored_value(mem,phase);
1860   if (value != NULL && value->is_Con() &&
1861       !value->bottom_type()->higher_equal(_type)) {
1862     // If the input to the store does not fit with the load's result type,
1863     // it must be truncated. We can't delay until Ideal call since
1864     // a singleton Value is needed for split_thru_phi optimization.
1865     int con = value->get_int();
1866     return TypeInt::make((con << 24) >> 24);
1867   }
1868   return LoadNode::Value(phase);
1869 }
1870 
1871 //--------------------------LoadUBNode::Ideal-------------------------------------
1872 //
1873 //  If the previous store is to the same address as this load,
1874 //  and the value stored was larger than a byte, replace this load
1875 //  with the value stored truncated to a byte.  If no truncation is
1876 //  needed, the replacement is done in LoadNode::Identity().
1877 //
1878 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1879   Node* mem = in(MemNode::Memory);
1880   Node* value = can_see_stored_value(mem, phase);
1881   if (value && !phase->type(value)->higher_equal(_type))
1882     return new (phase->C) AndINode(value, phase->intcon(0xFF));
1883   // Identity call will handle the case where truncation is not needed.
1884   return LoadNode::Ideal(phase, can_reshape);
1885 }
1886 
1887 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1888   Node* mem = in(MemNode::Memory);
1889   Node* value = can_see_stored_value(mem,phase);
1890   if (value != NULL && value->is_Con() &&
1891       !value->bottom_type()->higher_equal(_type)) {
1892     // If the input to the store does not fit with the load's result type,
1893     // it must be truncated. We can't delay until Ideal call since
1894     // a singleton Value is needed for split_thru_phi optimization.
1895     int con = value->get_int();
1896     return TypeInt::make(con & 0xFF);
1897   }
1898   return LoadNode::Value(phase);
1899 }
1900 
1901 //--------------------------LoadUSNode::Ideal-------------------------------------
1902 //
1903 //  If the previous store is to the same address as this load,
1904 //  and the value stored was larger than a char, replace this load
1905 //  with the value stored truncated to a char.  If no truncation is
1906 //  needed, the replacement is done in LoadNode::Identity().
1907 //
1908 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1909   Node* mem = in(MemNode::Memory);
1910   Node* value = can_see_stored_value(mem,phase);
1911   if( value && !phase->type(value)->higher_equal( _type ) )
1912     return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1913   // Identity call will handle the case where truncation is not needed.
1914   return LoadNode::Ideal(phase, can_reshape);
1915 }
1916 
1917 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1918   Node* mem = in(MemNode::Memory);
1919   Node* value = can_see_stored_value(mem,phase);
1920   if (value != NULL && value->is_Con() &&
1921       !value->bottom_type()->higher_equal(_type)) {
1922     // If the input to the store does not fit with the load's result type,
1923     // it must be truncated. We can't delay until Ideal call since
1924     // a singleton Value is needed for split_thru_phi optimization.
1925     int con = value->get_int();
1926     return TypeInt::make(con & 0xFFFF);
1927   }
1928   return LoadNode::Value(phase);
1929 }
1930 
1931 //--------------------------LoadSNode::Ideal--------------------------------------
1932 //
1933 //  If the previous store is to the same address as this load,
1934 //  and the value stored was larger than a short, replace this load
1935 //  with the value stored truncated to a short.  If no truncation is
1936 //  needed, the replacement is done in LoadNode::Identity().
1937 //
1938 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1939   Node* mem = in(MemNode::Memory);
1940   Node* value = can_see_stored_value(mem,phase);
1941   if( value && !phase->type(value)->higher_equal( _type ) ) {
1942     Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
1943     return new (phase->C) RShiftINode(result, phase->intcon(16));
1944   }
1945   // Identity call will handle the case where truncation is not needed.
1946   return LoadNode::Ideal(phase, can_reshape);
1947 }
1948 
1949 const Type* LoadSNode::Value(PhaseTransform *phase) const {
1950   Node* mem = in(MemNode::Memory);
1951   Node* value = can_see_stored_value(mem,phase);
1952   if (value != NULL && value->is_Con() &&
1953       !value->bottom_type()->higher_equal(_type)) {
1954     // If the input to the store does not fit with the load's result type,
1955     // it must be truncated. We can't delay until Ideal call since
1956     // a singleton Value is needed for split_thru_phi optimization.
1957     int con = value->get_int();
1958     return TypeInt::make((con << 16) >> 16);
1959   }
1960   return LoadNode::Value(phase);
1961 }
1962 
1963 //=============================================================================
1964 //----------------------------LoadKlassNode::make------------------------------
1965 // Polymorphic factory method:
1966 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1967   Compile* C = gvn.C;
1968   Node *ctl = NULL;
1969   // sanity check the alias category against the created node type
1970   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
1971   assert(adr_type != NULL, "expecting TypeKlassPtr");
1972 #ifdef _LP64
1973   if (adr_type->is_ptr_to_narrowklass()) {
1974     assert(UseCompressedKlassPointers, "no compressed klasses");
1975     Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass()));
1976     return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
1977   }
1978 #endif
1979   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1980   return new (C) LoadKlassNode(ctl, mem, adr, at, tk);
1981 }
1982 
1983 //------------------------------Value------------------------------------------
1984 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1985   return klass_value_common(phase);
1986 }
1987 
1988 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1989   // Either input is TOP ==> the result is TOP
1990   const Type *t1 = phase->type( in(MemNode::Memory) );
1991   if (t1 == Type::TOP)  return Type::TOP;
1992   Node *adr = in(MemNode::Address);
1993   const Type *t2 = phase->type( adr );
1994   if (t2 == Type::TOP)  return Type::TOP;
1995   const TypePtr *tp = t2->is_ptr();
1996   if (TypePtr::above_centerline(tp->ptr()) ||
1997       tp->ptr() == TypePtr::Null)  return Type::TOP;
1998 
1999   // Return a more precise klass, if possible
2000   const TypeInstPtr *tinst = tp->isa_instptr();
2001   if (tinst != NULL) {
2002     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2003     int offset = tinst->offset();
2004     if (ik == phase->C->env()->Class_klass()
2005         && (offset == java_lang_Class::klass_offset_in_bytes() ||
2006             offset == java_lang_Class::array_klass_offset_in_bytes())) {
2007       // We are loading a special hidden field from a Class mirror object,
2008       // the field which points to the VM's Klass metaobject.
2009       ciType* t = tinst->java_mirror_type();
2010       // java_mirror_type returns non-null for compile-time Class constants.
2011       if (t != NULL) {
2012         // constant oop => constant klass
2013         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2014           return TypeKlassPtr::make(ciArrayKlass::make(t));
2015         }
2016         if (!t->is_klass()) {
2017           // a primitive Class (e.g., int.class) has NULL for a klass field
2018           return TypePtr::NULL_PTR;
2019         }
2020         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2021         return TypeKlassPtr::make(t->as_klass());
2022       }
2023       // non-constant mirror, so we can't tell what's going on
2024     }
2025     if( !ik->is_loaded() )
2026       return _type;             // Bail out if not loaded
2027     if (offset == oopDesc::klass_offset_in_bytes()) {
2028       if (tinst->klass_is_exact()) {
2029         return TypeKlassPtr::make(ik);
2030       }
2031       // See if we can become precise: no subklasses and no interface
2032       // (Note:  We need to support verified interfaces.)
2033       if (!ik->is_interface() && !ik->has_subklass()) {
2034         //assert(!UseExactTypes, "this code should be useless with exact types");
2035         // Add a dependence; if any subclass added we need to recompile
2036         if (!ik->is_final()) {
2037           // %%% should use stronger assert_unique_concrete_subtype instead
2038           phase->C->dependencies()->assert_leaf_type(ik);
2039         }
2040         // Return precise klass
2041         return TypeKlassPtr::make(ik);
2042       }
2043 
2044       // Return root of possible klass
2045       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2046     }
2047   }
2048 
2049   // Check for loading klass from an array
2050   const TypeAryPtr *tary = tp->isa_aryptr();
2051   if( tary != NULL ) {
2052     ciKlass *tary_klass = tary->klass();
2053     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
2054         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2055       if (tary->klass_is_exact()) {
2056         return TypeKlassPtr::make(tary_klass);
2057       }
2058       ciArrayKlass *ak = tary->klass()->as_array_klass();
2059       // If the klass is an object array, we defer the question to the
2060       // array component klass.
2061       if( ak->is_obj_array_klass() ) {
2062         assert( ak->is_loaded(), "" );
2063         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2064         if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2065           ciInstanceKlass* ik = base_k->as_instance_klass();
2066           // See if we can become precise: no subklasses and no interface
2067           if (!ik->is_interface() && !ik->has_subklass()) {
2068             //assert(!UseExactTypes, "this code should be useless with exact types");
2069             // Add a dependence; if any subclass added we need to recompile
2070             if (!ik->is_final()) {
2071               phase->C->dependencies()->assert_leaf_type(ik);
2072             }
2073             // Return precise array klass
2074             return TypeKlassPtr::make(ak);
2075           }
2076         }
2077         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2078       } else {                  // Found a type-array?
2079         //assert(!UseExactTypes, "this code should be useless with exact types");
2080         assert( ak->is_type_array_klass(), "" );
2081         return TypeKlassPtr::make(ak); // These are always precise
2082       }
2083     }
2084   }
2085 
2086   // Check for loading klass from an array klass
2087   const TypeKlassPtr *tkls = tp->isa_klassptr();
2088   if (tkls != NULL && !StressReflectiveCode) {
2089     ciKlass* klass = tkls->klass();
2090     if( !klass->is_loaded() )
2091       return _type;             // Bail out if not loaded
2092     if( klass->is_obj_array_klass() &&
2093         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2094       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2095       // // Always returning precise element type is incorrect,
2096       // // e.g., element type could be object and array may contain strings
2097       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2098 
2099       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2100       // according to the element type's subclassing.
2101       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2102     }
2103     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2104         tkls->offset() == in_bytes(Klass::super_offset())) {
2105       ciKlass* sup = klass->as_instance_klass()->super();
2106       // The field is Klass::_super.  Return its (constant) value.
2107       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2108       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2109     }
2110   }
2111 
2112   // Bailout case
2113   return LoadNode::Value(phase);
2114 }
2115 
2116 //------------------------------Identity---------------------------------------
2117 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2118 // Also feed through the klass in Allocate(...klass...)._klass.
2119 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2120   return klass_identity_common(phase);
2121 }
2122 
2123 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2124   Node* x = LoadNode::Identity(phase);
2125   if (x != this)  return x;
2126 
2127   // Take apart the address into an oop and and offset.
2128   // Return 'this' if we cannot.
2129   Node*    adr    = in(MemNode::Address);
2130   intptr_t offset = 0;
2131   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2132   if (base == NULL)     return this;
2133   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2134   if (toop == NULL)     return this;
2135 
2136   // We can fetch the klass directly through an AllocateNode.
2137   // This works even if the klass is not constant (clone or newArray).
2138   if (offset == oopDesc::klass_offset_in_bytes()) {
2139     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2140     if (allocated_klass != NULL) {
2141       return allocated_klass;
2142     }
2143   }
2144 
2145   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2146   // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
2147   // See inline_native_Class_query for occurrences of these patterns.
2148   // Java Example:  x.getClass().isAssignableFrom(y)
2149   // Java Example:  Array.newInstance(x.getClass().getComponentType(), n)
2150   //
2151   // This improves reflective code, often making the Class
2152   // mirror go completely dead.  (Current exception:  Class
2153   // mirrors may appear in debug info, but we could clean them out by
2154   // introducing a new debug info operator for Klass*.java_mirror).
2155   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2156       && (offset == java_lang_Class::klass_offset_in_bytes() ||
2157           offset == java_lang_Class::array_klass_offset_in_bytes())) {
2158     // We are loading a special hidden field from a Class mirror,
2159     // the field which points to its Klass or ArrayKlass metaobject.
2160     if (base->is_Load()) {
2161       Node* adr2 = base->in(MemNode::Address);
2162       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2163       if (tkls != NULL && !tkls->empty()
2164           && (tkls->klass()->is_instance_klass() ||
2165               tkls->klass()->is_array_klass())
2166           && adr2->is_AddP()
2167           ) {
2168         int mirror_field = in_bytes(Klass::java_mirror_offset());
2169         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2170           mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
2171         }
2172         if (tkls->offset() == mirror_field) {
2173           return adr2->in(AddPNode::Base);
2174         }
2175       }
2176     }
2177   }
2178 
2179   return this;
2180 }
2181 
2182 
2183 //------------------------------Value------------------------------------------
2184 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2185   const Type *t = klass_value_common(phase);
2186   if (t == Type::TOP)
2187     return t;
2188 
2189   return t->make_narrowklass();
2190 }
2191 
2192 //------------------------------Identity---------------------------------------
2193 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2194 // Also feed through the klass in Allocate(...klass...)._klass.
2195 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2196   Node *x = klass_identity_common(phase);
2197 
2198   const Type *t = phase->type( x );
2199   if( t == Type::TOP ) return x;
2200   if( t->isa_narrowklass()) return x;
2201   assert (!t->isa_narrowoop(), "no narrow oop here");
2202 
2203   return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
2204 }
2205 
2206 //------------------------------Value-----------------------------------------
2207 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2208   // Either input is TOP ==> the result is TOP
2209   const Type *t1 = phase->type( in(MemNode::Memory) );
2210   if( t1 == Type::TOP ) return Type::TOP;
2211   Node *adr = in(MemNode::Address);
2212   const Type *t2 = phase->type( adr );
2213   if( t2 == Type::TOP ) return Type::TOP;
2214   const TypePtr *tp = t2->is_ptr();
2215   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2216   const TypeAryPtr *tap = tp->isa_aryptr();
2217   if( !tap ) return _type;
2218   return tap->size();
2219 }
2220 
2221 //-------------------------------Ideal---------------------------------------
2222 // Feed through the length in AllocateArray(...length...)._length.
2223 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2224   Node* p = MemNode::Ideal_common(phase, can_reshape);
2225   if (p)  return (p == NodeSentinel) ? NULL : p;
2226 
2227   // Take apart the address into an oop and and offset.
2228   // Return 'this' if we cannot.
2229   Node*    adr    = in(MemNode::Address);
2230   intptr_t offset = 0;
2231   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2232   if (base == NULL)     return NULL;
2233   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2234   if (tary == NULL)     return NULL;
2235 
2236   // We can fetch the length directly through an AllocateArrayNode.
2237   // This works even if the length is not constant (clone or newArray).
2238   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2239     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2240     if (alloc != NULL) {
2241       Node* allocated_length = alloc->Ideal_length();
2242       Node* len = alloc->make_ideal_length(tary, phase);
2243       if (allocated_length != len) {
2244         // New CastII improves on this.
2245         return len;
2246       }
2247     }
2248   }
2249 
2250   return NULL;
2251 }
2252 
2253 //------------------------------Identity---------------------------------------
2254 // Feed through the length in AllocateArray(...length...)._length.
2255 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2256   Node* x = LoadINode::Identity(phase);
2257   if (x != this)  return x;
2258 
2259   // Take apart the address into an oop and and offset.
2260   // Return 'this' if we cannot.
2261   Node*    adr    = in(MemNode::Address);
2262   intptr_t offset = 0;
2263   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2264   if (base == NULL)     return this;
2265   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2266   if (tary == NULL)     return this;
2267 
2268   // We can fetch the length directly through an AllocateArrayNode.
2269   // This works even if the length is not constant (clone or newArray).
2270   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2271     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2272     if (alloc != NULL) {
2273       Node* allocated_length = alloc->Ideal_length();
2274       // Do not allow make_ideal_length to allocate a CastII node.
2275       Node* len = alloc->make_ideal_length(tary, phase, false);
2276       if (allocated_length == len) {
2277         // Return allocated_length only if it would not be improved by a CastII.
2278         return allocated_length;
2279       }
2280     }
2281   }
2282 
2283   return this;
2284 
2285 }
2286 
2287 //=============================================================================
2288 //---------------------------StoreNode::make-----------------------------------
2289 // Polymorphic factory method:
2290 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
2291   Compile* C = gvn.C;
2292   assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2293           ctl != NULL, "raw memory operations should have control edge");
2294 
2295   switch (bt) {
2296   case T_BOOLEAN:
2297   case T_BYTE:    return new (C) StoreBNode(ctl, mem, adr, adr_type, val);
2298   case T_INT:     return new (C) StoreINode(ctl, mem, adr, adr_type, val);
2299   case T_CHAR:
2300   case T_SHORT:   return new (C) StoreCNode(ctl, mem, adr, adr_type, val);
2301   case T_LONG:    return new (C) StoreLNode(ctl, mem, adr, adr_type, val);
2302   case T_FLOAT:   return new (C) StoreFNode(ctl, mem, adr, adr_type, val);
2303   case T_DOUBLE:  return new (C) StoreDNode(ctl, mem, adr, adr_type, val);
2304   case T_METADATA:
2305   case T_ADDRESS:
2306   case T_OBJECT:
2307 #ifdef _LP64
2308     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2309       val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2310       return new (C) StoreNNode(ctl, mem, adr, adr_type, val);
2311     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2312                (UseCompressedKlassPointers && val->bottom_type()->isa_klassptr() &&
2313                 adr->bottom_type()->isa_rawptr())) {
2314       val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2315       return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val);
2316     }
2317 #endif
2318     {
2319       return new (C) StorePNode(ctl, mem, adr, adr_type, val);
2320     }
2321   }
2322   ShouldNotReachHere();
2323   return (StoreNode*)NULL;
2324 }
2325 
2326 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
2327   bool require_atomic = true;
2328   return new (C) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
2329 }
2330 
2331 
2332 //--------------------------bottom_type----------------------------------------
2333 const Type *StoreNode::bottom_type() const {
2334   return Type::MEMORY;
2335 }
2336 
2337 //------------------------------hash-------------------------------------------
2338 uint StoreNode::hash() const {
2339   // unroll addition of interesting fields
2340   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2341 
2342   // Since they are not commoned, do not hash them:
2343   return NO_HASH;
2344 }
2345 
2346 //------------------------------Ideal------------------------------------------
2347 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2348 // When a store immediately follows a relevant allocation/initialization,
2349 // try to capture it into the initialization, or hoist it above.
2350 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2351   Node* p = MemNode::Ideal_common(phase, can_reshape);
2352   if (p)  return (p == NodeSentinel) ? NULL : p;
2353 
2354   Node* mem     = in(MemNode::Memory);
2355   Node* address = in(MemNode::Address);
2356 
2357   // Back-to-back stores to same address?  Fold em up.  Generally
2358   // unsafe if I have intervening uses...  Also disallowed for StoreCM
2359   // since they must follow each StoreP operation.  Redundant StoreCMs
2360   // are eliminated just before matching in final_graph_reshape.
2361   if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2362       mem->Opcode() != Op_StoreCM) {
2363     // Looking at a dead closed cycle of memory?
2364     assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2365 
2366     assert(Opcode() == mem->Opcode() ||
2367            phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2368            "no mismatched stores, except on raw memory");
2369 
2370     if (mem->outcnt() == 1 &&           // check for intervening uses
2371         mem->as_Store()->memory_size() <= this->memory_size()) {
2372       // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2373       // For example, 'mem' might be the final state at a conditional return.
2374       // Or, 'mem' might be used by some node which is live at the same time
2375       // 'this' is live, which might be unschedulable.  So, require exactly
2376       // ONE user, the 'this' store, until such time as we clone 'mem' for
2377       // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2378       if (can_reshape) {  // (%%% is this an anachronism?)
2379         set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2380                   phase->is_IterGVN());
2381       } else {
2382         // It's OK to do this in the parser, since DU info is always accurate,
2383         // and the parser always refers to nodes via SafePointNode maps.
2384         set_req(MemNode::Memory, mem->in(MemNode::Memory));
2385       }
2386       return this;
2387     }
2388   }
2389 
2390   // Capture an unaliased, unconditional, simple store into an initializer.
2391   // Or, if it is independent of the allocation, hoist it above the allocation.
2392   if (ReduceFieldZeroing && /*can_reshape &&*/
2393       mem->is_Proj() && mem->in(0)->is_Initialize()) {
2394     InitializeNode* init = mem->in(0)->as_Initialize();
2395     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2396     if (offset > 0) {
2397       Node* moved = init->capture_store(this, offset, phase, can_reshape);
2398       // If the InitializeNode captured me, it made a raw copy of me,
2399       // and I need to disappear.
2400       if (moved != NULL) {
2401         // %%% hack to ensure that Ideal returns a new node:
2402         mem = MergeMemNode::make(phase->C, mem);
2403         return mem;             // fold me away
2404       }
2405     }
2406   }
2407 
2408   return NULL;                  // No further progress
2409 }
2410 
2411 //------------------------------Value-----------------------------------------
2412 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2413   // Either input is TOP ==> the result is TOP
2414   const Type *t1 = phase->type( in(MemNode::Memory) );
2415   if( t1 == Type::TOP ) return Type::TOP;
2416   const Type *t2 = phase->type( in(MemNode::Address) );
2417   if( t2 == Type::TOP ) return Type::TOP;
2418   const Type *t3 = phase->type( in(MemNode::ValueIn) );
2419   if( t3 == Type::TOP ) return Type::TOP;
2420   return Type::MEMORY;
2421 }
2422 
2423 //------------------------------Identity---------------------------------------
2424 // Remove redundant stores:
2425 //   Store(m, p, Load(m, p)) changes to m.
2426 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2427 Node *StoreNode::Identity( PhaseTransform *phase ) {
2428   Node* mem = in(MemNode::Memory);
2429   Node* adr = in(MemNode::Address);
2430   Node* val = in(MemNode::ValueIn);
2431 
2432   // Load then Store?  Then the Store is useless
2433   if (val->is_Load() &&
2434       val->in(MemNode::Address)->eqv_uncast(adr) &&
2435       val->in(MemNode::Memory )->eqv_uncast(mem) &&
2436       val->as_Load()->store_Opcode() == Opcode()) {
2437     return mem;
2438   }
2439 
2440   // Two stores in a row of the same value?
2441   if (mem->is_Store() &&
2442       mem->in(MemNode::Address)->eqv_uncast(adr) &&
2443       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2444       mem->Opcode() == Opcode()) {
2445     return mem;
2446   }
2447 
2448   // Store of zero anywhere into a freshly-allocated object?
2449   // Then the store is useless.
2450   // (It must already have been captured by the InitializeNode.)
2451   if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2452     // a newly allocated object is already all-zeroes everywhere
2453     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2454       return mem;
2455     }
2456 
2457     // the store may also apply to zero-bits in an earlier object
2458     Node* prev_mem = find_previous_store(phase);
2459     // Steps (a), (b):  Walk past independent stores to find an exact match.
2460     if (prev_mem != NULL) {
2461       Node* prev_val = can_see_stored_value(prev_mem, phase);
2462       if (prev_val != NULL && phase->eqv(prev_val, val)) {
2463         // prev_val and val might differ by a cast; it would be good
2464         // to keep the more informative of the two.
2465         return mem;
2466       }
2467     }
2468   }
2469 
2470   return this;
2471 }
2472 
2473 //------------------------------match_edge-------------------------------------
2474 // Do we Match on this edge index or not?  Match only memory & value
2475 uint StoreNode::match_edge(uint idx) const {
2476   return idx == MemNode::Address || idx == MemNode::ValueIn;
2477 }
2478 
2479 //------------------------------cmp--------------------------------------------
2480 // Do not common stores up together.  They generally have to be split
2481 // back up anyways, so do not bother.
2482 uint StoreNode::cmp( const Node &n ) const {
2483   return (&n == this);          // Always fail except on self
2484 }
2485 
2486 //------------------------------Ideal_masked_input-----------------------------
2487 // Check for a useless mask before a partial-word store
2488 // (StoreB ... (AndI valIn conIa) )
2489 // If (conIa & mask == mask) this simplifies to
2490 // (StoreB ... (valIn) )
2491 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2492   Node *val = in(MemNode::ValueIn);
2493   if( val->Opcode() == Op_AndI ) {
2494     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2495     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2496       set_req(MemNode::ValueIn, val->in(1));
2497       return this;
2498     }
2499   }
2500   return NULL;
2501 }
2502 
2503 
2504 //------------------------------Ideal_sign_extended_input----------------------
2505 // Check for useless sign-extension before a partial-word store
2506 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2507 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
2508 // (StoreB ... (valIn) )
2509 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2510   Node *val = in(MemNode::ValueIn);
2511   if( val->Opcode() == Op_RShiftI ) {
2512     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2513     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2514       Node *shl = val->in(1);
2515       if( shl->Opcode() == Op_LShiftI ) {
2516         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2517         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2518           set_req(MemNode::ValueIn, shl->in(1));
2519           return this;
2520         }
2521       }
2522     }
2523   }
2524   return NULL;
2525 }
2526 
2527 //------------------------------value_never_loaded-----------------------------------
2528 // Determine whether there are any possible loads of the value stored.
2529 // For simplicity, we actually check if there are any loads from the
2530 // address stored to, not just for loads of the value stored by this node.
2531 //
2532 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2533   Node *adr = in(Address);
2534   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2535   if (adr_oop == NULL)
2536     return false;
2537   if (!adr_oop->is_known_instance_field())
2538     return false; // if not a distinct instance, there may be aliases of the address
2539   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2540     Node *use = adr->fast_out(i);
2541     int opc = use->Opcode();
2542     if (use->is_Load() || use->is_LoadStore()) {
2543       return false;
2544     }
2545   }
2546   return true;
2547 }
2548 
2549 //=============================================================================
2550 //------------------------------Ideal------------------------------------------
2551 // If the store is from an AND mask that leaves the low bits untouched, then
2552 // we can skip the AND operation.  If the store is from a sign-extension
2553 // (a left shift, then right shift) we can skip both.
2554 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2555   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2556   if( progress != NULL ) return progress;
2557 
2558   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2559   if( progress != NULL ) return progress;
2560 
2561   // Finally check the default case
2562   return StoreNode::Ideal(phase, can_reshape);
2563 }
2564 
2565 //=============================================================================
2566 //------------------------------Ideal------------------------------------------
2567 // If the store is from an AND mask that leaves the low bits untouched, then
2568 // we can skip the AND operation
2569 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2570   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2571   if( progress != NULL ) return progress;
2572 
2573   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2574   if( progress != NULL ) return progress;
2575 
2576   // Finally check the default case
2577   return StoreNode::Ideal(phase, can_reshape);
2578 }
2579 
2580 //=============================================================================
2581 //------------------------------Identity---------------------------------------
2582 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2583   // No need to card mark when storing a null ptr
2584   Node* my_store = in(MemNode::OopStore);
2585   if (my_store->is_Store()) {
2586     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2587     if( t1 == TypePtr::NULL_PTR ) {
2588       return in(MemNode::Memory);
2589     }
2590   }
2591   return this;
2592 }
2593 
2594 //=============================================================================
2595 //------------------------------Ideal---------------------------------------
2596 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2597   Node* progress = StoreNode::Ideal(phase, can_reshape);
2598   if (progress != NULL) return progress;
2599 
2600   Node* my_store = in(MemNode::OopStore);
2601   if (my_store->is_MergeMem()) {
2602     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2603     set_req(MemNode::OopStore, mem);
2604     return this;
2605   }
2606 
2607   return NULL;
2608 }
2609 
2610 //------------------------------Value-----------------------------------------
2611 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2612   // Either input is TOP ==> the result is TOP
2613   const Type *t = phase->type( in(MemNode::Memory) );
2614   if( t == Type::TOP ) return Type::TOP;
2615   t = phase->type( in(MemNode::Address) );
2616   if( t == Type::TOP ) return Type::TOP;
2617   t = phase->type( in(MemNode::ValueIn) );
2618   if( t == Type::TOP ) return Type::TOP;
2619   // If extra input is TOP ==> the result is TOP
2620   t = phase->type( in(MemNode::OopStore) );
2621   if( t == Type::TOP ) return Type::TOP;
2622 
2623   return StoreNode::Value( phase );
2624 }
2625 
2626 
2627 //=============================================================================
2628 //----------------------------------SCMemProjNode------------------------------
2629 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2630 {
2631   return bottom_type();
2632 }
2633 
2634 //=============================================================================
2635 //----------------------------------LoadStoreNode------------------------------
2636 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2637   : Node(required),
2638     _type(rt),
2639     _adr_type(at)
2640 {
2641   init_req(MemNode::Control, c  );
2642   init_req(MemNode::Memory , mem);
2643   init_req(MemNode::Address, adr);
2644   init_req(MemNode::ValueIn, val);
2645   init_class_id(Class_LoadStore);
2646 }
2647 
2648 uint LoadStoreNode::ideal_reg() const {
2649   return _type->ideal_reg();
2650 }
2651 
2652 bool LoadStoreNode::result_not_used() const {
2653   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2654     Node *x = fast_out(i);
2655     if (x->Opcode() == Op_SCMemProj) continue;
2656     return false;
2657   }
2658   return true;
2659 }
2660 
2661 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2662 
2663 //=============================================================================
2664 //----------------------------------LoadStoreConditionalNode--------------------
2665 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2666   init_req(ExpectedIn, ex );
2667 }
2668 
2669 //=============================================================================
2670 //-------------------------------adr_type--------------------------------------
2671 // Do we Match on this edge index or not?  Do not match memory
2672 const TypePtr* ClearArrayNode::adr_type() const {
2673   Node *adr = in(3);
2674   return MemNode::calculate_adr_type(adr->bottom_type());
2675 }
2676 
2677 //------------------------------match_edge-------------------------------------
2678 // Do we Match on this edge index or not?  Do not match memory
2679 uint ClearArrayNode::match_edge(uint idx) const {
2680   return idx > 1;
2681 }
2682 
2683 //------------------------------Identity---------------------------------------
2684 // Clearing a zero length array does nothing
2685 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2686   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
2687 }
2688 
2689 //------------------------------Idealize---------------------------------------
2690 // Clearing a short array is faster with stores
2691 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2692   const int unit = BytesPerLong;
2693   const TypeX* t = phase->type(in(2))->isa_intptr_t();
2694   if (!t)  return NULL;
2695   if (!t->is_con())  return NULL;
2696   intptr_t raw_count = t->get_con();
2697   intptr_t size = raw_count;
2698   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2699   // Clearing nothing uses the Identity call.
2700   // Negative clears are possible on dead ClearArrays
2701   // (see jck test stmt114.stmt11402.val).
2702   if (size <= 0 || size % unit != 0)  return NULL;
2703   intptr_t count = size / unit;
2704   // Length too long; use fast hardware clear
2705   if (size > Matcher::init_array_short_size)  return NULL;
2706   Node *mem = in(1);
2707   if( phase->type(mem)==Type::TOP ) return NULL;
2708   Node *adr = in(3);
2709   const Type* at = phase->type(adr);
2710   if( at==Type::TOP ) return NULL;
2711   const TypePtr* atp = at->isa_ptr();
2712   // adjust atp to be the correct array element address type
2713   if (atp == NULL)  atp = TypePtr::BOTTOM;
2714   else              atp = atp->add_offset(Type::OffsetBot);
2715   // Get base for derived pointer purposes
2716   if( adr->Opcode() != Op_AddP ) Unimplemented();
2717   Node *base = adr->in(1);
2718 
2719   Node *zero = phase->makecon(TypeLong::ZERO);
2720   Node *off  = phase->MakeConX(BytesPerLong);
2721   mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero);
2722   count--;
2723   while( count-- ) {
2724     mem = phase->transform(mem);
2725     adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
2726     mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero);
2727   }
2728   return mem;
2729 }
2730 
2731 //----------------------------step_through----------------------------------
2732 // Return allocation input memory edge if it is different instance
2733 // or itself if it is the one we are looking for.
2734 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2735   Node* n = *np;
2736   assert(n->is_ClearArray(), "sanity");
2737   intptr_t offset;
2738   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2739   // This method is called only before Allocate nodes are expanded during
2740   // macro nodes expansion. Before that ClearArray nodes are only generated
2741   // in LibraryCallKit::generate_arraycopy() which follows allocations.
2742   assert(alloc != NULL, "should have allocation");
2743   if (alloc->_idx == instance_id) {
2744     // Can not bypass initialization of the instance we are looking for.
2745     return false;
2746   }
2747   // Otherwise skip it.
2748   InitializeNode* init = alloc->initialization();
2749   if (init != NULL)
2750     *np = init->in(TypeFunc::Memory);
2751   else
2752     *np = alloc->in(TypeFunc::Memory);
2753   return true;
2754 }
2755 
2756 //----------------------------clear_memory-------------------------------------
2757 // Generate code to initialize object storage to zero.
2758 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2759                                    intptr_t start_offset,
2760                                    Node* end_offset,
2761                                    PhaseGVN* phase) {
2762   Compile* C = phase->C;
2763   intptr_t offset = start_offset;
2764 
2765   int unit = BytesPerLong;
2766   if ((offset % unit) != 0) {
2767     Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
2768     adr = phase->transform(adr);
2769     const TypePtr* atp = TypeRawPtr::BOTTOM;
2770     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2771     mem = phase->transform(mem);
2772     offset += BytesPerInt;
2773   }
2774   assert((offset % unit) == 0, "");
2775 
2776   // Initialize the remaining stuff, if any, with a ClearArray.
2777   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2778 }
2779 
2780 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2781                                    Node* start_offset,
2782                                    Node* end_offset,
2783                                    PhaseGVN* phase) {
2784   if (start_offset == end_offset) {
2785     // nothing to do
2786     return mem;
2787   }
2788 
2789   Compile* C = phase->C;
2790   int unit = BytesPerLong;
2791   Node* zbase = start_offset;
2792   Node* zend  = end_offset;
2793 
2794   // Scale to the unit required by the CPU:
2795   if (!Matcher::init_array_count_is_in_bytes) {
2796     Node* shift = phase->intcon(exact_log2(unit));
2797     zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
2798     zend  = phase->transform( new(C) URShiftXNode(zend,  shift) );
2799   }
2800 
2801   // Bulk clear double-words
2802   Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
2803   Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
2804   mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
2805   return phase->transform(mem);
2806 }
2807 
2808 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2809                                    intptr_t start_offset,
2810                                    intptr_t end_offset,
2811                                    PhaseGVN* phase) {
2812   if (start_offset == end_offset) {
2813     // nothing to do
2814     return mem;
2815   }
2816 
2817   Compile* C = phase->C;
2818   assert((end_offset % BytesPerInt) == 0, "odd end offset");
2819   intptr_t done_offset = end_offset;
2820   if ((done_offset % BytesPerLong) != 0) {
2821     done_offset -= BytesPerInt;
2822   }
2823   if (done_offset > start_offset) {
2824     mem = clear_memory(ctl, mem, dest,
2825                        start_offset, phase->MakeConX(done_offset), phase);
2826   }
2827   if (done_offset < end_offset) { // emit the final 32-bit store
2828     Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
2829     adr = phase->transform(adr);
2830     const TypePtr* atp = TypeRawPtr::BOTTOM;
2831     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2832     mem = phase->transform(mem);
2833     done_offset += BytesPerInt;
2834   }
2835   assert(done_offset == end_offset, "");
2836   return mem;
2837 }
2838 
2839 //=============================================================================
2840 // Do not match memory edge.
2841 uint StrIntrinsicNode::match_edge(uint idx) const {
2842   return idx == 2 || idx == 3;
2843 }
2844 
2845 //------------------------------Ideal------------------------------------------
2846 // Return a node which is more "ideal" than the current node.  Strip out
2847 // control copies
2848 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2849   if (remove_dead_region(phase, can_reshape)) return this;
2850   // Don't bother trying to transform a dead node
2851   if (in(0) && in(0)->is_top())  return NULL;
2852 
2853   if (can_reshape) {
2854     Node* mem = phase->transform(in(MemNode::Memory));
2855     // If transformed to a MergeMem, get the desired slice
2856     uint alias_idx = phase->C->get_alias_index(adr_type());
2857     mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
2858     if (mem != in(MemNode::Memory)) {
2859       set_req(MemNode::Memory, mem);
2860       return this;
2861     }
2862   }
2863   return NULL;
2864 }
2865 
2866 //------------------------------Value------------------------------------------
2867 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
2868   if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2869   return bottom_type();
2870 }
2871 
2872 //=============================================================================
2873 //------------------------------match_edge-------------------------------------
2874 // Do not match memory edge
2875 uint EncodeISOArrayNode::match_edge(uint idx) const {
2876   return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len)
2877 }
2878 
2879 //------------------------------Ideal------------------------------------------
2880 // Return a node which is more "ideal" than the current node.  Strip out
2881 // control copies
2882 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2883   return remove_dead_region(phase, can_reshape) ? this : NULL;
2884 }
2885 
2886 //------------------------------Value------------------------------------------
2887 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const {
2888   if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2889   return bottom_type();
2890 }
2891 
2892 //=============================================================================
2893 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2894   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2895     _adr_type(C->get_adr_type(alias_idx))
2896 {
2897   init_class_id(Class_MemBar);
2898   Node* top = C->top();
2899   init_req(TypeFunc::I_O,top);
2900   init_req(TypeFunc::FramePtr,top);
2901   init_req(TypeFunc::ReturnAdr,top);
2902   if (precedent != NULL)
2903     init_req(TypeFunc::Parms, precedent);
2904 }
2905 
2906 //------------------------------cmp--------------------------------------------
2907 uint MemBarNode::hash() const { return NO_HASH; }
2908 uint MemBarNode::cmp( const Node &n ) const {
2909   return (&n == this);          // Always fail except on self
2910 }
2911 
2912 //------------------------------make-------------------------------------------
2913 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2914   switch (opcode) {
2915   case Op_MemBarAcquire:   return new(C) MemBarAcquireNode(C,  atp, pn);
2916   case Op_MemBarRelease:   return new(C) MemBarReleaseNode(C,  atp, pn);
2917   case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C,  atp, pn);
2918   case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C,  atp, pn);
2919   case Op_MemBarVolatile:  return new(C) MemBarVolatileNode(C, atp, pn);
2920   case Op_MemBarCPUOrder:  return new(C) MemBarCPUOrderNode(C, atp, pn);
2921   case Op_Initialize:      return new(C) InitializeNode(C,     atp, pn);
2922   case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C,  atp, pn);
2923   default:                 ShouldNotReachHere(); return NULL;
2924   }
2925 }
2926 
2927 //------------------------------Ideal------------------------------------------
2928 // Return a node which is more "ideal" than the current node.  Strip out
2929 // control copies
2930 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2931   if (remove_dead_region(phase, can_reshape)) return this;
2932   // Don't bother trying to transform a dead node
2933   if (in(0) && in(0)->is_top()) {
2934     return NULL;
2935   }
2936 
2937   // Eliminate volatile MemBars for scalar replaced objects.
2938   if (can_reshape && req() == (Precedent+1)) {
2939     bool eliminate = false;
2940     int opc = Opcode();
2941     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2942       // Volatile field loads and stores.
2943       Node* my_mem = in(MemBarNode::Precedent);
2944       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2945       if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2946         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2947         // replace this Precedent (decodeN) with the Load instead.
2948         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
2949           Node* load_node = my_mem->in(1);
2950           set_req(MemBarNode::Precedent, load_node);
2951           phase->is_IterGVN()->_worklist.push(my_mem);
2952           my_mem = load_node;
2953         } else {
2954           assert(my_mem->unique_out() == this, "sanity");
2955           del_req(Precedent);
2956           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2957           my_mem = NULL;
2958         }
2959       }
2960       if (my_mem != NULL && my_mem->is_Mem()) {
2961         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2962         // Check for scalar replaced object reference.
2963         if( t_oop != NULL && t_oop->is_known_instance_field() &&
2964             t_oop->offset() != Type::OffsetBot &&
2965             t_oop->offset() != Type::OffsetTop) {
2966           eliminate = true;
2967         }
2968       }
2969     } else if (opc == Op_MemBarRelease) {
2970       // Final field stores.
2971       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2972       if ((alloc != NULL) && alloc->is_Allocate() &&
2973           alloc->as_Allocate()->_is_non_escaping) {
2974         // The allocated object does not escape.
2975         eliminate = true;
2976       }
2977     }
2978     if (eliminate) {
2979       // Replace MemBar projections by its inputs.
2980       PhaseIterGVN* igvn = phase->is_IterGVN();
2981       igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2982       igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2983       // Must return either the original node (now dead) or a new node
2984       // (Do not return a top here, since that would break the uniqueness of top.)
2985       return new (phase->C) ConINode(TypeInt::ZERO);
2986     }
2987   }
2988   return NULL;
2989 }
2990 
2991 //------------------------------Value------------------------------------------
2992 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2993   if( !in(0) ) return Type::TOP;
2994   if( phase->type(in(0)) == Type::TOP )
2995     return Type::TOP;
2996   return TypeTuple::MEMBAR;
2997 }
2998 
2999 //------------------------------match------------------------------------------
3000 // Construct projections for memory.
3001 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3002   switch (proj->_con) {
3003   case TypeFunc::Control:
3004   case TypeFunc::Memory:
3005     return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3006   }
3007   ShouldNotReachHere();
3008   return NULL;
3009 }
3010 
3011 //===========================InitializeNode====================================
3012 // SUMMARY:
3013 // This node acts as a memory barrier on raw memory, after some raw stores.
3014 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3015 // The Initialize can 'capture' suitably constrained stores as raw inits.
3016 // It can coalesce related raw stores into larger units (called 'tiles').
3017 // It can avoid zeroing new storage for memory units which have raw inits.
3018 // At macro-expansion, it is marked 'complete', and does not optimize further.
3019 //
3020 // EXAMPLE:
3021 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3022 //   ctl = incoming control; mem* = incoming memory
3023 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
3024 // First allocate uninitialized memory and fill in the header:
3025 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
3026 //   ctl := alloc.Control; mem* := alloc.Memory*
3027 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
3028 // Then initialize to zero the non-header parts of the raw memory block:
3029 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3030 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3031 // After the initialize node executes, the object is ready for service:
3032 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3033 // Suppose its body is immediately initialized as {1,2}:
3034 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3035 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3036 //   mem.SLICE(#short[*]) := store2
3037 //
3038 // DETAILS:
3039 // An InitializeNode collects and isolates object initialization after
3040 // an AllocateNode and before the next possible safepoint.  As a
3041 // memory barrier (MemBarNode), it keeps critical stores from drifting
3042 // down past any safepoint or any publication of the allocation.
3043 // Before this barrier, a newly-allocated object may have uninitialized bits.
3044 // After this barrier, it may be treated as a real oop, and GC is allowed.
3045 //
3046 // The semantics of the InitializeNode include an implicit zeroing of
3047 // the new object from object header to the end of the object.
3048 // (The object header and end are determined by the AllocateNode.)
3049 //
3050 // Certain stores may be added as direct inputs to the InitializeNode.
3051 // These stores must update raw memory, and they must be to addresses
3052 // derived from the raw address produced by AllocateNode, and with
3053 // a constant offset.  They must be ordered by increasing offset.
3054 // The first one is at in(RawStores), the last at in(req()-1).
3055 // Unlike most memory operations, they are not linked in a chain,
3056 // but are displayed in parallel as users of the rawmem output of
3057 // the allocation.
3058 //
3059 // (See comments in InitializeNode::capture_store, which continue
3060 // the example given above.)
3061 //
3062 // When the associated Allocate is macro-expanded, the InitializeNode
3063 // may be rewritten to optimize collected stores.  A ClearArrayNode
3064 // may also be created at that point to represent any required zeroing.
3065 // The InitializeNode is then marked 'complete', prohibiting further
3066 // capturing of nearby memory operations.
3067 //
3068 // During macro-expansion, all captured initializations which store
3069 // constant values of 32 bits or smaller are coalesced (if advantageous)
3070 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
3071 // initialized in fewer memory operations.  Memory words which are
3072 // covered by neither tiles nor non-constant stores are pre-zeroed
3073 // by explicit stores of zero.  (The code shape happens to do all
3074 // zeroing first, then all other stores, with both sequences occurring
3075 // in order of ascending offsets.)
3076 //
3077 // Alternatively, code may be inserted between an AllocateNode and its
3078 // InitializeNode, to perform arbitrary initialization of the new object.
3079 // E.g., the object copying intrinsics insert complex data transfers here.
3080 // The initialization must then be marked as 'complete' disable the
3081 // built-in zeroing semantics and the collection of initializing stores.
3082 //
3083 // While an InitializeNode is incomplete, reads from the memory state
3084 // produced by it are optimizable if they match the control edge and
3085 // new oop address associated with the allocation/initialization.
3086 // They return a stored value (if the offset matches) or else zero.
3087 // A write to the memory state, if it matches control and address,
3088 // and if it is to a constant offset, may be 'captured' by the
3089 // InitializeNode.  It is cloned as a raw memory operation and rewired
3090 // inside the initialization, to the raw oop produced by the allocation.
3091 // Operations on addresses which are provably distinct (e.g., to
3092 // other AllocateNodes) are allowed to bypass the initialization.
3093 //
3094 // The effect of all this is to consolidate object initialization
3095 // (both arrays and non-arrays, both piecewise and bulk) into a
3096 // single location, where it can be optimized as a unit.
3097 //
3098 // Only stores with an offset less than TrackedInitializationLimit words
3099 // will be considered for capture by an InitializeNode.  This puts a
3100 // reasonable limit on the complexity of optimized initializations.
3101 
3102 //---------------------------InitializeNode------------------------------------
3103 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3104   : _is_complete(Incomplete), _does_not_escape(false),
3105     MemBarNode(C, adr_type, rawoop)
3106 {
3107   init_class_id(Class_Initialize);
3108 
3109   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3110   assert(in(RawAddress) == rawoop, "proper init");
3111   // Note:  allocation() can be NULL, for secondary initialization barriers
3112 }
3113 
3114 // Since this node is not matched, it will be processed by the
3115 // register allocator.  Declare that there are no constraints
3116 // on the allocation of the RawAddress edge.
3117 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3118   // This edge should be set to top, by the set_complete.  But be conservative.
3119   if (idx == InitializeNode::RawAddress)
3120     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3121   return RegMask::Empty;
3122 }
3123 
3124 Node* InitializeNode::memory(uint alias_idx) {
3125   Node* mem = in(Memory);
3126   if (mem->is_MergeMem()) {
3127     return mem->as_MergeMem()->memory_at(alias_idx);
3128   } else {
3129     // incoming raw memory is not split
3130     return mem;
3131   }
3132 }
3133 
3134 bool InitializeNode::is_non_zero() {
3135   if (is_complete())  return false;
3136   remove_extra_zeroes();
3137   return (req() > RawStores);
3138 }
3139 
3140 void InitializeNode::set_complete(PhaseGVN* phase) {
3141   assert(!is_complete(), "caller responsibility");
3142   _is_complete = Complete;
3143 
3144   // After this node is complete, it contains a bunch of
3145   // raw-memory initializations.  There is no need for
3146   // it to have anything to do with non-raw memory effects.
3147   // Therefore, tell all non-raw users to re-optimize themselves,
3148   // after skipping the memory effects of this initialization.
3149   PhaseIterGVN* igvn = phase->is_IterGVN();
3150   if (igvn)  igvn->add_users_to_worklist(this);
3151 }
3152 
3153 // convenience function
3154 // return false if the init contains any stores already
3155 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3156   InitializeNode* init = initialization();
3157   if (init == NULL || init->is_complete())  return false;
3158   init->remove_extra_zeroes();
3159   // for now, if this allocation has already collected any inits, bail:
3160   if (init->is_non_zero())  return false;
3161   init->set_complete(phase);
3162   return true;
3163 }
3164 
3165 void InitializeNode::remove_extra_zeroes() {
3166   if (req() == RawStores)  return;
3167   Node* zmem = zero_memory();
3168   uint fill = RawStores;
3169   for (uint i = fill; i < req(); i++) {
3170     Node* n = in(i);
3171     if (n->is_top() || n == zmem)  continue;  // skip
3172     if (fill < i)  set_req(fill, n);          // compact
3173     ++fill;
3174   }
3175   // delete any empty spaces created:
3176   while (fill < req()) {
3177     del_req(fill);
3178   }
3179 }
3180 
3181 // Helper for remembering which stores go with which offsets.
3182 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3183   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
3184   intptr_t offset = -1;
3185   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3186                                                phase, offset);
3187   if (base == NULL)     return -1;  // something is dead,
3188   if (offset < 0)       return -1;  //        dead, dead
3189   return offset;
3190 }
3191 
3192 // Helper for proving that an initialization expression is
3193 // "simple enough" to be folded into an object initialization.
3194 // Attempts to prove that a store's initial value 'n' can be captured
3195 // within the initialization without creating a vicious cycle, such as:
3196 //     { Foo p = new Foo(); p.next = p; }
3197 // True for constants and parameters and small combinations thereof.
3198 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3199   if (n == NULL)      return true;   // (can this really happen?)
3200   if (n->is_Proj())   n = n->in(0);
3201   if (n == this)      return false;  // found a cycle
3202   if (n->is_Con())    return true;
3203   if (n->is_Start())  return true;   // params, etc., are OK
3204   if (n->is_Root())   return true;   // even better
3205 
3206   Node* ctl = n->in(0);
3207   if (ctl != NULL && !ctl->is_top()) {
3208     if (ctl->is_Proj())  ctl = ctl->in(0);
3209     if (ctl == this)  return false;
3210 
3211     // If we already know that the enclosing memory op is pinned right after
3212     // the init, then any control flow that the store has picked up
3213     // must have preceded the init, or else be equal to the init.
3214     // Even after loop optimizations (which might change control edges)
3215     // a store is never pinned *before* the availability of its inputs.
3216     if (!MemNode::all_controls_dominate(n, this))
3217       return false;                  // failed to prove a good control
3218   }
3219 
3220   // Check data edges for possible dependencies on 'this'.
3221   if ((count += 1) > 20)  return false;  // complexity limit
3222   for (uint i = 1; i < n->req(); i++) {
3223     Node* m = n->in(i);
3224     if (m == NULL || m == n || m->is_top())  continue;
3225     uint first_i = n->find_edge(m);
3226     if (i != first_i)  continue;  // process duplicate edge just once
3227     if (!detect_init_independence(m, count)) {
3228       return false;
3229     }
3230   }
3231 
3232   return true;
3233 }
3234 
3235 // Here are all the checks a Store must pass before it can be moved into
3236 // an initialization.  Returns zero if a check fails.
3237 // On success, returns the (constant) offset to which the store applies,
3238 // within the initialized memory.
3239 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3240   const int FAIL = 0;
3241   if (st->req() != MemNode::ValueIn + 1)
3242     return FAIL;                // an inscrutable StoreNode (card mark?)
3243   Node* ctl = st->in(MemNode::Control);
3244   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3245     return FAIL;                // must be unconditional after the initialization
3246   Node* mem = st->in(MemNode::Memory);
3247   if (!(mem->is_Proj() && mem->in(0) == this))
3248     return FAIL;                // must not be preceded by other stores
3249   Node* adr = st->in(MemNode::Address);
3250   intptr_t offset;
3251   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3252   if (alloc == NULL)
3253     return FAIL;                // inscrutable address
3254   if (alloc != allocation())
3255     return FAIL;                // wrong allocation!  (store needs to float up)
3256   Node* val = st->in(MemNode::ValueIn);
3257   int complexity_count = 0;
3258   if (!detect_init_independence(val, complexity_count))
3259     return FAIL;                // stored value must be 'simple enough'
3260 
3261   // The Store can be captured only if nothing after the allocation
3262   // and before the Store is using the memory location that the store
3263   // overwrites.
3264   bool failed = false;
3265   // If is_complete_with_arraycopy() is true the shape of the graph is
3266   // well defined and is safe so no need for extra checks.
3267   if (!is_complete_with_arraycopy()) {
3268     // We are going to look at each use of the memory state following
3269     // the allocation to make sure nothing reads the memory that the
3270     // Store writes.
3271     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3272     int alias_idx = phase->C->get_alias_index(t_adr);
3273     ResourceMark rm;
3274     Unique_Node_List mems;
3275     mems.push(mem);
3276     Node* unique_merge = NULL;
3277     for (uint next = 0; next < mems.size(); ++next) {
3278       Node *m  = mems.at(next);
3279       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3280         Node *n = m->fast_out(j);
3281         if (n->outcnt() == 0) {
3282           continue;
3283         }
3284         if (n == st) {
3285           continue;
3286         } else if (n->in(0) != NULL && n->in(0) != ctl) {
3287           // If the control of this use is different from the control
3288           // of the Store which is right after the InitializeNode then
3289           // this node cannot be between the InitializeNode and the
3290           // Store.
3291           continue;
3292         } else if (n->is_MergeMem()) {
3293           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3294             // We can hit a MergeMemNode (that will likely go away
3295             // later) that is a direct use of the memory state
3296             // following the InitializeNode on the same slice as the
3297             // store node that we'd like to capture. We need to check
3298             // the uses of the MergeMemNode.
3299             mems.push(n);
3300           }
3301         } else if (n->is_Mem()) {
3302           Node* other_adr = n->in(MemNode::Address);
3303           if (other_adr == adr) {
3304             failed = true;
3305             break;
3306           } else {
3307             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3308             if (other_t_adr != NULL) {
3309               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3310               if (other_alias_idx == alias_idx) {
3311                 // A load from the same memory slice as the store right
3312                 // after the InitializeNode. We check the control of the
3313                 // object/array that is loaded from. If it's the same as
3314                 // the store control then we cannot capture the store.
3315                 assert(!n->is_Store(), "2 stores to same slice on same control?");
3316                 Node* base = other_adr;
3317                 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3318                 base = base->in(AddPNode::Base);
3319                 if (base != NULL) {
3320                   base = base->uncast();
3321                   if (base->is_Proj() && base->in(0) == alloc) {
3322                     failed = true;
3323                     break;
3324                   }
3325                 }
3326               }
3327             }
3328           }
3329         } else {
3330           failed = true;
3331           break;
3332         }
3333       }
3334     }
3335   }
3336   if (failed) {
3337     if (!can_reshape) {
3338       // We decided we couldn't capture the store during parsing. We
3339       // should try again during the next IGVN once the graph is
3340       // cleaner.
3341       phase->C->record_for_igvn(st);
3342     }
3343     return FAIL;
3344   }
3345 
3346   return offset;                // success
3347 }
3348 
3349 // Find the captured store in(i) which corresponds to the range
3350 // [start..start+size) in the initialized object.
3351 // If there is one, return its index i.  If there isn't, return the
3352 // negative of the index where it should be inserted.
3353 // Return 0 if the queried range overlaps an initialization boundary
3354 // or if dead code is encountered.
3355 // If size_in_bytes is zero, do not bother with overlap checks.
3356 int InitializeNode::captured_store_insertion_point(intptr_t start,
3357                                                    int size_in_bytes,
3358                                                    PhaseTransform* phase) {
3359   const int FAIL = 0, MAX_STORE = BytesPerLong;
3360 
3361   if (is_complete())
3362     return FAIL;                // arraycopy got here first; punt
3363 
3364   assert(allocation() != NULL, "must be present");
3365 
3366   // no negatives, no header fields:
3367   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
3368 
3369   // after a certain size, we bail out on tracking all the stores:
3370   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3371   if (start >= ti_limit)  return FAIL;
3372 
3373   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3374     if (i >= limit)  return -(int)i; // not found; here is where to put it
3375 
3376     Node*    st     = in(i);
3377     intptr_t st_off = get_store_offset(st, phase);
3378     if (st_off < 0) {
3379       if (st != zero_memory()) {
3380         return FAIL;            // bail out if there is dead garbage
3381       }
3382     } else if (st_off > start) {
3383       // ...we are done, since stores are ordered
3384       if (st_off < start + size_in_bytes) {
3385         return FAIL;            // the next store overlaps
3386       }
3387       return -(int)i;           // not found; here is where to put it
3388     } else if (st_off < start) {
3389       if (size_in_bytes != 0 &&
3390           start < st_off + MAX_STORE &&
3391           start < st_off + st->as_Store()->memory_size()) {
3392         return FAIL;            // the previous store overlaps
3393       }
3394     } else {
3395       if (size_in_bytes != 0 &&
3396           st->as_Store()->memory_size() != size_in_bytes) {
3397         return FAIL;            // mismatched store size
3398       }
3399       return i;
3400     }
3401 
3402     ++i;
3403   }
3404 }
3405 
3406 // Look for a captured store which initializes at the offset 'start'
3407 // with the given size.  If there is no such store, and no other
3408 // initialization interferes, then return zero_memory (the memory
3409 // projection of the AllocateNode).
3410 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3411                                           PhaseTransform* phase) {
3412   assert(stores_are_sane(phase), "");
3413   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3414   if (i == 0) {
3415     return NULL;                // something is dead
3416   } else if (i < 0) {
3417     return zero_memory();       // just primordial zero bits here
3418   } else {
3419     Node* st = in(i);           // here is the store at this position
3420     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3421     return st;
3422   }
3423 }
3424 
3425 // Create, as a raw pointer, an address within my new object at 'offset'.
3426 Node* InitializeNode::make_raw_address(intptr_t offset,
3427                                        PhaseTransform* phase) {
3428   Node* addr = in(RawAddress);
3429   if (offset != 0) {
3430     Compile* C = phase->C;
3431     addr = phase->transform( new (C) AddPNode(C->top(), addr,
3432                                                  phase->MakeConX(offset)) );
3433   }
3434   return addr;
3435 }
3436 
3437 // Clone the given store, converting it into a raw store
3438 // initializing a field or element of my new object.
3439 // Caller is responsible for retiring the original store,
3440 // with subsume_node or the like.
3441 //
3442 // From the example above InitializeNode::InitializeNode,
3443 // here are the old stores to be captured:
3444 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3445 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3446 //
3447 // Here is the changed code; note the extra edges on init:
3448 //   alloc = (Allocate ...)
3449 //   rawoop = alloc.RawAddress
3450 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3451 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3452 //   init = (Initialize alloc.Control alloc.Memory rawoop
3453 //                      rawstore1 rawstore2)
3454 //
3455 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3456                                     PhaseTransform* phase, bool can_reshape) {
3457   assert(stores_are_sane(phase), "");
3458 
3459   if (start < 0)  return NULL;
3460   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3461 
3462   Compile* C = phase->C;
3463   int size_in_bytes = st->memory_size();
3464   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3465   if (i == 0)  return NULL;     // bail out
3466   Node* prev_mem = NULL;        // raw memory for the captured store
3467   if (i > 0) {
3468     prev_mem = in(i);           // there is a pre-existing store under this one
3469     set_req(i, C->top());       // temporarily disconnect it
3470     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3471   } else {
3472     i = -i;                     // no pre-existing store
3473     prev_mem = zero_memory();   // a slice of the newly allocated object
3474     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3475       set_req(--i, C->top());   // reuse this edge; it has been folded away
3476     else
3477       ins_req(i, C->top());     // build a new edge
3478   }
3479   Node* new_st = st->clone();
3480   new_st->set_req(MemNode::Control, in(Control));
3481   new_st->set_req(MemNode::Memory,  prev_mem);
3482   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3483   new_st = phase->transform(new_st);
3484 
3485   // At this point, new_st might have swallowed a pre-existing store
3486   // at the same offset, or perhaps new_st might have disappeared,
3487   // if it redundantly stored the same value (or zero to fresh memory).
3488 
3489   // In any case, wire it in:
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);
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);
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);
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: %d > %d", 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(Compile* C, Node* mem) {
4120   return new(C) 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