1 /*
   2  * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/systemDictionary.hpp"
  27 #include "compiler/compileLog.hpp"
  28 #include "memory/allocation.inline.hpp"
  29 #include "oops/objArrayKlass.hpp"
  30 #include "opto/addnode.hpp"
  31 #include "opto/cfgnode.hpp"
  32 #include "opto/compile.hpp"
  33 #include "opto/connode.hpp"
  34 #include "opto/convertnode.hpp"
  35 #include "opto/loopnode.hpp"
  36 #include "opto/machnode.hpp"
  37 #include "opto/matcher.hpp"
  38 #include "opto/memnode.hpp"
  39 #include "opto/mulnode.hpp"
  40 #include "opto/narrowptrnode.hpp"
  41 #include "opto/phaseX.hpp"
  42 #include "opto/regmask.hpp"
  43 
  44 // Portions of code courtesy of Clifford Click
  45 
  46 // Optimization - Graph Style
  47 
  48 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
  49 
  50 //=============================================================================
  51 uint MemNode::size_of() const { return sizeof(*this); }
  52 
  53 const TypePtr *MemNode::adr_type() const {
  54   Node* adr = in(Address);
  55   const TypePtr* cross_check = NULL;
  56   DEBUG_ONLY(cross_check = _adr_type);
  57   return calculate_adr_type(adr->bottom_type(), cross_check);
  58 }
  59 
  60 #ifndef PRODUCT
  61 void MemNode::dump_spec(outputStream *st) const {
  62   if (in(Address) == NULL)  return; // node is dead
  63 #ifndef ASSERT
  64   // fake the missing field
  65   const TypePtr* _adr_type = NULL;
  66   if (in(Address) != NULL)
  67     _adr_type = in(Address)->bottom_type()->isa_ptr();
  68 #endif
  69   dump_adr_type(this, _adr_type, st);
  70 
  71   Compile* C = Compile::current();
  72   if( C->alias_type(_adr_type)->is_volatile() )
  73     st->print(" Volatile!");
  74 }
  75 
  76 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
  77   st->print(" @");
  78   if (adr_type == NULL) {
  79     st->print("NULL");
  80   } else {
  81     adr_type->dump_on(st);
  82     Compile* C = Compile::current();
  83     Compile::AliasType* atp = NULL;
  84     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
  85     if (atp == NULL)
  86       st->print(", idx=?\?;");
  87     else if (atp->index() == Compile::AliasIdxBot)
  88       st->print(", idx=Bot;");
  89     else if (atp->index() == Compile::AliasIdxTop)
  90       st->print(", idx=Top;");
  91     else if (atp->index() == Compile::AliasIdxRaw)
  92       st->print(", idx=Raw;");
  93     else {
  94       ciField* field = atp->field();
  95       if (field) {
  96         st->print(", name=");
  97         field->print_name_on(st);
  98       }
  99       st->print(", idx=%d;", atp->index());
 100     }
 101   }
 102 }
 103 
 104 extern void print_alias_types();
 105 
 106 #endif
 107 
 108 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
 109   assert((t_oop != NULL), "sanity");
 110   bool is_instance = t_oop->is_known_instance_field();
 111   bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
 112                              (load != NULL) && load->is_Load() &&
 113                              (phase->is_IterGVN() != NULL);
 114   if (!(is_instance || is_boxed_value_load))
 115     return mchain;  // don't try to optimize non-instance types
 116   uint instance_id = t_oop->instance_id();
 117   Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
 118   Node *prev = NULL;
 119   Node *result = mchain;
 120   while (prev != result) {
 121     prev = result;
 122     if (result == start_mem)
 123       break;  // hit one of our sentinels
 124     // skip over a call which does not affect this memory slice
 125     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
 126       Node *proj_in = result->in(0);
 127       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
 128         break;  // hit one of our sentinels
 129       } else if (proj_in->is_Call()) {
 130         CallNode *call = proj_in->as_Call();
 131         if (!call->may_modify(t_oop, phase)) { // returns false for instances
 132           result = call->in(TypeFunc::Memory);
 133         }
 134       } else if (proj_in->is_Initialize()) {
 135         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
 136         // Stop if this is the initialization for the object instance which
 137         // which contains this memory slice, otherwise skip over it.
 138         if ((alloc == NULL) || (alloc->_idx == instance_id)) {
 139           break;
 140         }
 141         if (is_instance) {
 142           result = proj_in->in(TypeFunc::Memory);
 143         } else if (is_boxed_value_load) {
 144           Node* klass = alloc->in(AllocateNode::KlassNode);
 145           const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
 146           if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
 147             result = proj_in->in(TypeFunc::Memory); // not related allocation
 148           }
 149         }
 150       } else if (proj_in->is_MemBar()) {
 151         result = proj_in->in(TypeFunc::Memory);
 152       } else {
 153         assert(false, "unexpected projection");
 154       }
 155     } else if (result->is_ClearArray()) {
 156       if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
 157         // Can not bypass initialization of the instance
 158         // we are looking for.
 159         break;
 160       }
 161       // Otherwise skip it (the call updated 'result' value).
 162     } else if (result->is_MergeMem()) {
 163       result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
 164     }
 165   }
 166   return result;
 167 }
 168 
 169 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
 170   const TypeOopPtr* t_oop = t_adr->isa_oopptr();
 171   if (t_oop == NULL)
 172     return mchain;  // don't try to optimize non-oop types
 173   Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
 174   bool is_instance = t_oop->is_known_instance_field();
 175   PhaseIterGVN *igvn = phase->is_IterGVN();
 176   if (is_instance && igvn != NULL  && result->is_Phi()) {
 177     PhiNode *mphi = result->as_Phi();
 178     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 179     const TypePtr *t = mphi->adr_type();
 180     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
 181         t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
 182         t->is_oopptr()->cast_to_exactness(true)
 183          ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 184          ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
 185       // clone the Phi with our address type
 186       result = mphi->split_out_instance(t_adr, igvn);
 187     } else {
 188       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 189     }
 190   }
 191   return result;
 192 }
 193 
 194 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 195   uint alias_idx = phase->C->get_alias_index(tp);
 196   Node *mem = mmem;
 197 #ifdef ASSERT
 198   {
 199     // Check that current type is consistent with the alias index used during graph construction
 200     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 201     bool consistent =  adr_check == NULL || adr_check->empty() ||
 202                        phase->C->must_alias(adr_check, alias_idx );
 203     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 204     if( !consistent && adr_check != NULL && !adr_check->empty() &&
 205                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 206         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 207         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 208           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 209           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 210       // don't assert if it is dead code.
 211       consistent = true;
 212     }
 213     if( !consistent ) {
 214       st->print("alias_idx==%d, adr_check==", alias_idx);
 215       if( adr_check == NULL ) {
 216         st->print("NULL");
 217       } else {
 218         adr_check->dump();
 219       }
 220       st->cr();
 221       print_alias_types();
 222       assert(consistent, "adr_check must match alias idx");
 223     }
 224   }
 225 #endif
 226   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
 227   // means an array I have not precisely typed yet.  Do not do any
 228   // alias stuff with it any time soon.
 229   const TypeOopPtr *toop = tp->isa_oopptr();
 230   if( tp->base() != Type::AnyPtr &&
 231       !(toop &&
 232         toop->klass() != NULL &&
 233         toop->klass()->is_java_lang_Object() &&
 234         toop->offset() == Type::OffsetBot) ) {
 235     // compress paths and change unreachable cycles to TOP
 236     // If not, we can update the input infinitely along a MergeMem cycle
 237     // Equivalent code in PhiNode::Ideal
 238     Node* m  = phase->transform(mmem);
 239     // If transformed to a MergeMem, get the desired slice
 240     // Otherwise the returned node represents memory for every slice
 241     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
 242     // Update input if it is progress over what we have now
 243   }
 244   return mem;
 245 }
 246 
 247 //--------------------------Ideal_common---------------------------------------
 248 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
 249 // Unhook non-raw memories from complete (macro-expanded) initializations.
 250 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
 251   // If our control input is a dead region, kill all below the region
 252   Node *ctl = in(MemNode::Control);
 253   if (ctl && remove_dead_region(phase, can_reshape))
 254     return this;
 255   ctl = in(MemNode::Control);
 256   // Don't bother trying to transform a dead node
 257   if (ctl && ctl->is_top())  return NodeSentinel;
 258 
 259   PhaseIterGVN *igvn = phase->is_IterGVN();
 260   // Wait if control on the worklist.
 261   if (ctl && can_reshape && igvn != NULL) {
 262     Node* bol = NULL;
 263     Node* cmp = NULL;
 264     if (ctl->in(0)->is_If()) {
 265       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
 266       bol = ctl->in(0)->in(1);
 267       if (bol->is_Bool())
 268         cmp = ctl->in(0)->in(1)->in(1);
 269     }
 270     if (igvn->_worklist.member(ctl) ||
 271         (bol != NULL && igvn->_worklist.member(bol)) ||
 272         (cmp != NULL && igvn->_worklist.member(cmp)) ) {
 273       // This control path may be dead.
 274       // Delay this memory node transformation until the control is processed.
 275       phase->is_IterGVN()->_worklist.push(this);
 276       return NodeSentinel; // caller will return NULL
 277     }
 278   }
 279   // Ignore if memory is dead, or self-loop
 280   Node *mem = in(MemNode::Memory);
 281   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
 282   assert(mem != this, "dead loop in MemNode::Ideal");
 283 
 284   if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
 285     // This memory slice may be dead.
 286     // Delay this mem node transformation until the memory is processed.
 287     phase->is_IterGVN()->_worklist.push(this);
 288     return NodeSentinel; // caller will return NULL
 289   }
 290 
 291   Node *address = in(MemNode::Address);
 292   const Type *t_adr = phase->type(address);
 293   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return NULL
 294 
 295   if (can_reshape && igvn != NULL &&
 296       (igvn->_worklist.member(address) ||
 297        igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
 298     // The address's base and type may change when the address is processed.
 299     // Delay this mem node transformation until the address is processed.
 300     phase->is_IterGVN()->_worklist.push(this);
 301     return NodeSentinel; // caller will return NULL
 302   }
 303 
 304   // Do NOT remove or optimize the next lines: ensure a new alias index
 305   // is allocated for an oop pointer type before Escape Analysis.
 306   // Note: C++ will not remove it since the call has side effect.
 307   if (t_adr->isa_oopptr()) {
 308     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 309   }
 310 
 311 #ifdef ASSERT
 312   Node* base = NULL;
 313   if (address->is_AddP())
 314     base = address->in(AddPNode::Base);
 315   if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
 316       !t_adr->isa_rawptr()) {
 317     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
 318     Compile* C = phase->C;
 319     tty->cr();
 320     tty->print_cr("===== NULL+offs not RAW address =====");
 321     if (C->is_dead_node(this->_idx))    tty->print_cr("'this' is dead");
 322     if ((ctl != NULL) && C->is_dead_node(ctl->_idx)) tty->print_cr("'ctl' is dead");
 323     if (C->is_dead_node(mem->_idx))     tty->print_cr("'mem' is dead");
 324     if (C->is_dead_node(address->_idx)) tty->print_cr("'address' is dead");
 325     if (C->is_dead_node(base->_idx))    tty->print_cr("'base' is dead");
 326     tty->cr();
 327     base->dump(1);
 328     tty->cr();
 329     this->dump(2);
 330     tty->print("this->adr_type():     "); adr_type()->dump(); tty->cr();
 331     tty->print("phase->type(address): "); t_adr->dump(); tty->cr();
 332     tty->print("phase->type(base):    "); phase->type(address)->dump(); tty->cr();
 333     tty->cr();
 334   }
 335   assert(base == NULL || t_adr->isa_rawptr() ||
 336         !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
 337 #endif
 338 
 339   // Avoid independent memory operations
 340   Node* old_mem = mem;
 341 
 342   // The code which unhooks non-raw memories from complete (macro-expanded)
 343   // initializations was removed. After macro-expansion all stores catched
 344   // by Initialize node became raw stores and there is no information
 345   // which memory slices they modify. So it is unsafe to move any memory
 346   // operation above these stores. Also in most cases hooked non-raw memories
 347   // were already unhooked by using information from detect_ptr_independence()
 348   // and find_previous_store().
 349 
 350   if (mem->is_MergeMem()) {
 351     MergeMemNode* mmem = mem->as_MergeMem();
 352     const TypePtr *tp = t_adr->is_ptr();
 353 
 354     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 355   }
 356 
 357   if (mem != old_mem) {
 358     set_req(MemNode::Memory, mem);
 359     if (can_reshape && old_mem->outcnt() == 0) {
 360         igvn->_worklist.push(old_mem);
 361     }
 362     if (phase->type( mem ) == Type::TOP) return NodeSentinel;
 363     return this;
 364   }
 365 
 366   // let the subclass continue analyzing...
 367   return NULL;
 368 }
 369 
 370 // Helper function for proving some simple control dominations.
 371 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 372 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 373 // is not a constant (dominated by the method's StartNode).
 374 // Used by MemNode::find_previous_store to prove that the
 375 // control input of a memory operation predates (dominates)
 376 // an allocation it wants to look past.
 377 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
 378   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
 379     return false; // Conservative answer for dead code
 380 
 381   // Check 'dom'. Skip Proj and CatchProj nodes.
 382   dom = dom->find_exact_control(dom);
 383   if (dom == NULL || dom->is_top())
 384     return false; // Conservative answer for dead code
 385 
 386   if (dom == sub) {
 387     // For the case when, for example, 'sub' is Initialize and the original
 388     // 'dom' is Proj node of the 'sub'.
 389     return false;
 390   }
 391 
 392   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
 393     return true;
 394 
 395   // 'dom' dominates 'sub' if its control edge and control edges
 396   // of all its inputs dominate or equal to sub's control edge.
 397 
 398   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 399   // Or Region for the check in LoadNode::Ideal();
 400   // 'sub' should have sub->in(0) != NULL.
 401   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 402          sub->is_Region() || sub->is_Call(), "expecting only these nodes");
 403 
 404   // Get control edge of 'sub'.
 405   Node* orig_sub = sub;
 406   sub = sub->find_exact_control(sub->in(0));
 407   if (sub == NULL || sub->is_top())
 408     return false; // Conservative answer for dead code
 409 
 410   assert(sub->is_CFG(), "expecting control");
 411 
 412   if (sub == dom)
 413     return true;
 414 
 415   if (sub->is_Start() || sub->is_Root())
 416     return false;
 417 
 418   {
 419     // Check all control edges of 'dom'.
 420 
 421     ResourceMark rm;
 422     Arena* arena = Thread::current()->resource_area();
 423     Node_List nlist(arena);
 424     Unique_Node_List dom_list(arena);
 425 
 426     dom_list.push(dom);
 427     bool only_dominating_controls = false;
 428 
 429     for (uint next = 0; next < dom_list.size(); next++) {
 430       Node* n = dom_list.at(next);
 431       if (n == orig_sub)
 432         return false; // One of dom's inputs dominated by sub.
 433       if (!n->is_CFG() && n->pinned()) {
 434         // Check only own control edge for pinned non-control nodes.
 435         n = n->find_exact_control(n->in(0));
 436         if (n == NULL || n->is_top())
 437           return false; // Conservative answer for dead code
 438         assert(n->is_CFG(), "expecting control");
 439         dom_list.push(n);
 440       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 441         only_dominating_controls = true;
 442       } else if (n->is_CFG()) {
 443         if (n->dominates(sub, nlist))
 444           only_dominating_controls = true;
 445         else
 446           return false;
 447       } else {
 448         // First, own control edge.
 449         Node* m = n->find_exact_control(n->in(0));
 450         if (m != NULL) {
 451           if (m->is_top())
 452             return false; // Conservative answer for dead code
 453           dom_list.push(m);
 454         }
 455         // Now, the rest of edges.
 456         uint cnt = n->req();
 457         for (uint i = 1; i < cnt; i++) {
 458           m = n->find_exact_control(n->in(i));
 459           if (m == NULL || m->is_top())
 460             continue;
 461           dom_list.push(m);
 462         }
 463       }
 464     }
 465     return only_dominating_controls;
 466   }
 467 }
 468 
 469 //---------------------detect_ptr_independence---------------------------------
 470 // Used by MemNode::find_previous_store to prove that two base
 471 // pointers are never equal.
 472 // The pointers are accompanied by their associated allocations,
 473 // if any, which have been previously discovered by the caller.
 474 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 475                                       Node* p2, AllocateNode* a2,
 476                                       PhaseTransform* phase) {
 477   // Attempt to prove that these two pointers cannot be aliased.
 478   // They may both manifestly be allocations, and they should differ.
 479   // Or, if they are not both allocations, they can be distinct constants.
 480   // Otherwise, one is an allocation and the other a pre-existing value.
 481   if (a1 == NULL && a2 == NULL) {           // neither an allocation
 482     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 483   } else if (a1 != NULL && a2 != NULL) {    // both allocations
 484     return (a1 != a2);
 485   } else if (a1 != NULL) {                  // one allocation a1
 486     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 487     return all_controls_dominate(p2, a1);
 488   } else { //(a2 != NULL)                   // one allocation a2
 489     return all_controls_dominate(p1, a2);
 490   }
 491   return false;
 492 }
 493 
 494 
 495 // The logic for reordering loads and stores uses four steps:
 496 // (a) Walk carefully past stores and initializations which we
 497 //     can prove are independent of this load.
 498 // (b) Observe that the next memory state makes an exact match
 499 //     with self (load or store), and locate the relevant store.
 500 // (c) Ensure that, if we were to wire self directly to the store,
 501 //     the optimizer would fold it up somehow.
 502 // (d) Do the rewiring, and return, depending on some other part of
 503 //     the optimizer to fold up the load.
 504 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 505 // specific to loads and stores, so they are handled by the callers.
 506 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 507 //
 508 Node* MemNode::find_previous_store(PhaseTransform* phase) {
 509   Node*         ctrl   = in(MemNode::Control);
 510   Node*         adr    = in(MemNode::Address);
 511   intptr_t      offset = 0;
 512   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 513   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
 514 
 515   if (offset == Type::OffsetBot)
 516     return NULL;            // cannot unalias unless there are precise offsets
 517 
 518   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 519 
 520   intptr_t size_in_bytes = memory_size();
 521 
 522   Node* mem = in(MemNode::Memory);   // start searching here...
 523 
 524   int cnt = 50;             // Cycle limiter
 525   for (;;) {                // While we can dance past unrelated stores...
 526     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 527 
 528     if (mem->is_Store()) {
 529       Node* st_adr = mem->in(MemNode::Address);
 530       intptr_t st_offset = 0;
 531       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 532       if (st_base == NULL)
 533         break;              // inscrutable pointer
 534       if (st_offset != offset && st_offset != Type::OffsetBot) {
 535         const int MAX_STORE = BytesPerLong;
 536         if (st_offset >= offset + size_in_bytes ||
 537             st_offset <= offset - MAX_STORE ||
 538             st_offset <= offset - mem->as_Store()->memory_size()) {
 539           // Success:  The offsets are provably independent.
 540           // (You may ask, why not just test st_offset != offset and be done?
 541           // The answer is that stores of different sizes can co-exist
 542           // in the same sequence of RawMem effects.  We sometimes initialize
 543           // a whole 'tile' of array elements with a single jint or jlong.)
 544           mem = mem->in(MemNode::Memory);
 545           continue;           // (a) advance through independent store memory
 546         }
 547       }
 548       if (st_base != base &&
 549           detect_ptr_independence(base, alloc,
 550                                   st_base,
 551                                   AllocateNode::Ideal_allocation(st_base, phase),
 552                                   phase)) {
 553         // Success:  The bases are provably independent.
 554         mem = mem->in(MemNode::Memory);
 555         continue;           // (a) advance through independent store memory
 556       }
 557 
 558       // (b) At this point, if the bases or offsets do not agree, we lose,
 559       // since we have not managed to prove 'this' and 'mem' independent.
 560       if (st_base == base && st_offset == offset) {
 561         return mem;         // let caller handle steps (c), (d)
 562       }
 563 
 564     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 565       InitializeNode* st_init = mem->in(0)->as_Initialize();
 566       AllocateNode*  st_alloc = st_init->allocation();
 567       if (st_alloc == NULL)
 568         break;              // something degenerated
 569       bool known_identical = false;
 570       bool known_independent = false;
 571       if (alloc == st_alloc)
 572         known_identical = true;
 573       else if (alloc != NULL)
 574         known_independent = true;
 575       else if (all_controls_dominate(this, st_alloc))
 576         known_independent = true;
 577 
 578       if (known_independent) {
 579         // The bases are provably independent: Either they are
 580         // manifestly distinct allocations, or else the control
 581         // of this load dominates the store's allocation.
 582         int alias_idx = phase->C->get_alias_index(adr_type());
 583         if (alias_idx == Compile::AliasIdxRaw) {
 584           mem = st_alloc->in(TypeFunc::Memory);
 585         } else {
 586           mem = st_init->memory(alias_idx);
 587         }
 588         continue;           // (a) advance through independent store memory
 589       }
 590 
 591       // (b) at this point, if we are not looking at a store initializing
 592       // the same allocation we are loading from, we lose.
 593       if (known_identical) {
 594         // From caller, can_see_stored_value will consult find_captured_store.
 595         return mem;         // let caller handle steps (c), (d)
 596       }
 597 
 598     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
 599       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 600       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 601         CallNode *call = mem->in(0)->as_Call();
 602         if (!call->may_modify(addr_t, phase)) {
 603           mem = call->in(TypeFunc::Memory);
 604           continue;         // (a) advance through independent call memory
 605         }
 606       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 607         mem = mem->in(0)->in(TypeFunc::Memory);
 608         continue;           // (a) advance through independent MemBar memory
 609       } else if (mem->is_ClearArray()) {
 610         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 611           // (the call updated 'mem' value)
 612           continue;         // (a) advance through independent allocation memory
 613         } else {
 614           // Can not bypass initialization of the instance
 615           // we are looking for.
 616           return mem;
 617         }
 618       } else if (mem->is_MergeMem()) {
 619         int alias_idx = phase->C->get_alias_index(adr_type());
 620         mem = mem->as_MergeMem()->memory_at(alias_idx);
 621         continue;           // (a) advance through independent MergeMem memory
 622       }
 623     }
 624 
 625     // Unless there is an explicit 'continue', we must bail out here,
 626     // because 'mem' is an inscrutable memory state (e.g., a call).
 627     break;
 628   }
 629 
 630   return NULL;              // bail out
 631 }
 632 
 633 //----------------------calculate_adr_type-------------------------------------
 634 // Helper function.  Notices when the given type of address hits top or bottom.
 635 // Also, asserts a cross-check of the type against the expected address type.
 636 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 637   if (t == Type::TOP)  return NULL; // does not touch memory any more?
 638   #ifdef PRODUCT
 639   cross_check = NULL;
 640   #else
 641   if (!VerifyAliases || is_error_reported() || Node::in_dump())  cross_check = NULL;
 642   #endif
 643   const TypePtr* tp = t->isa_ptr();
 644   if (tp == NULL) {
 645     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 646     return TypePtr::BOTTOM;           // touches lots of memory
 647   } else {
 648     #ifdef ASSERT
 649     // %%%% [phh] We don't check the alias index if cross_check is
 650     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 651     if (cross_check != NULL &&
 652         cross_check != TypePtr::BOTTOM &&
 653         cross_check != TypeRawPtr::BOTTOM) {
 654       // Recheck the alias index, to see if it has changed (due to a bug).
 655       Compile* C = Compile::current();
 656       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 657              "must stay in the original alias category");
 658       // The type of the address must be contained in the adr_type,
 659       // disregarding "null"-ness.
 660       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 661       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 662       assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
 663              "real address must not escape from expected memory type");
 664     }
 665     #endif
 666     return tp;
 667   }
 668 }
 669 
 670 //------------------------adr_phi_is_loop_invariant----------------------------
 671 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
 672 // loop is loop invariant. Make a quick traversal of Phi and associated
 673 // CastPP nodes, looking to see if they are a closed group within the loop.
 674 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
 675   // The idea is that the phi-nest must boil down to only CastPP nodes
 676   // with the same data. This implies that any path into the loop already
 677   // includes such a CastPP, and so the original cast, whatever its input,
 678   // must be covered by an equivalent cast, with an earlier control input.
 679   ResourceMark rm;
 680 
 681   // The loop entry input of the phi should be the unique dominating
 682   // node for every Phi/CastPP in the loop.
 683   Unique_Node_List closure;
 684   closure.push(adr_phi->in(LoopNode::EntryControl));
 685 
 686   // Add the phi node and the cast to the worklist.
 687   Unique_Node_List worklist;
 688   worklist.push(adr_phi);
 689   if( cast != NULL ){
 690     if( !cast->is_ConstraintCast() ) return false;
 691     worklist.push(cast);
 692   }
 693 
 694   // Begin recursive walk of phi nodes.
 695   while( worklist.size() ){
 696     // Take a node off the worklist
 697     Node *n = worklist.pop();
 698     if( !closure.member(n) ){
 699       // Add it to the closure.
 700       closure.push(n);
 701       // Make a sanity check to ensure we don't waste too much time here.
 702       if( closure.size() > 20) return false;
 703       // This node is OK if:
 704       //  - it is a cast of an identical value
 705       //  - or it is a phi node (then we add its inputs to the worklist)
 706       // Otherwise, the node is not OK, and we presume the cast is not invariant
 707       if( n->is_ConstraintCast() ){
 708         worklist.push(n->in(1));
 709       } else if( n->is_Phi() ) {
 710         for( uint i = 1; i < n->req(); i++ ) {
 711           worklist.push(n->in(i));
 712         }
 713       } else {
 714         return false;
 715       }
 716     }
 717   }
 718 
 719   // Quit when the worklist is empty, and we've found no offending nodes.
 720   return true;
 721 }
 722 
 723 //------------------------------Ideal_DU_postCCP-------------------------------
 724 // Find any cast-away of null-ness and keep its control.  Null cast-aways are
 725 // going away in this pass and we need to make this memory op depend on the
 726 // gating null check.
 727 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
 728   return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
 729 }
 730 
 731 // I tried to leave the CastPP's in.  This makes the graph more accurate in
 732 // some sense; we get to keep around the knowledge that an oop is not-null
 733 // after some test.  Alas, the CastPP's interfere with GVN (some values are
 734 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
 735 // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed
 736 // some of the more trivial cases in the optimizer.  Removing more useless
 737 // Phi's started allowing Loads to illegally float above null checks.  I gave
 738 // up on this approach.  CNC 10/20/2000
 739 // This static method may be called not from MemNode (EncodePNode calls it).
 740 // Only the control edge of the node 'n' might be updated.
 741 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
 742   Node *skipped_cast = NULL;
 743   // Need a null check?  Regular static accesses do not because they are
 744   // from constant addresses.  Array ops are gated by the range check (which
 745   // always includes a NULL check).  Just check field ops.
 746   if( n->in(MemNode::Control) == NULL ) {
 747     // Scan upwards for the highest location we can place this memory op.
 748     while( true ) {
 749       switch( adr->Opcode() ) {
 750 
 751       case Op_AddP:             // No change to NULL-ness, so peek thru AddP's
 752         adr = adr->in(AddPNode::Base);
 753         continue;
 754 
 755       case Op_DecodeN:         // No change to NULL-ness, so peek thru
 756       case Op_DecodeNKlass:
 757         adr = adr->in(1);
 758         continue;
 759 
 760       case Op_EncodeP:
 761       case Op_EncodePKlass:
 762         // EncodeP node's control edge could be set by this method
 763         // when EncodeP node depends on CastPP node.
 764         //
 765         // Use its control edge for memory op because EncodeP may go away
 766         // later when it is folded with following or preceding DecodeN node.
 767         if (adr->in(0) == NULL) {
 768           // Keep looking for cast nodes.
 769           adr = adr->in(1);
 770           continue;
 771         }
 772         ccp->hash_delete(n);
 773         n->set_req(MemNode::Control, adr->in(0));
 774         ccp->hash_insert(n);
 775         return n;
 776 
 777       case Op_CastPP:
 778         // If the CastPP is useless, just peek on through it.
 779         if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
 780           // Remember the cast that we've peeked though. If we peek
 781           // through more than one, then we end up remembering the highest
 782           // one, that is, if in a loop, the one closest to the top.
 783           skipped_cast = adr;
 784           adr = adr->in(1);
 785           continue;
 786         }
 787         // CastPP is going away in this pass!  We need this memory op to be
 788         // control-dependent on the test that is guarding the CastPP.
 789         ccp->hash_delete(n);
 790         n->set_req(MemNode::Control, adr->in(0));
 791         ccp->hash_insert(n);
 792         return n;
 793 
 794       case Op_Phi:
 795         // Attempt to float above a Phi to some dominating point.
 796         if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
 797           // If we've already peeked through a Cast (which could have set the
 798           // control), we can't float above a Phi, because the skipped Cast
 799           // may not be loop invariant.
 800           if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
 801             adr = adr->in(1);
 802             continue;
 803           }
 804         }
 805 
 806         // Intentional fallthrough!
 807 
 808         // No obvious dominating point.  The mem op is pinned below the Phi
 809         // by the Phi itself.  If the Phi goes away (no true value is merged)
 810         // then the mem op can float, but not indefinitely.  It must be pinned
 811         // behind the controls leading to the Phi.
 812       case Op_CheckCastPP:
 813         // These usually stick around to change address type, however a
 814         // useless one can be elided and we still need to pick up a control edge
 815         if (adr->in(0) == NULL) {
 816           // This CheckCastPP node has NO control and is likely useless. But we
 817           // need check further up the ancestor chain for a control input to keep
 818           // the node in place. 4959717.
 819           skipped_cast = adr;
 820           adr = adr->in(1);
 821           continue;
 822         }
 823         ccp->hash_delete(n);
 824         n->set_req(MemNode::Control, adr->in(0));
 825         ccp->hash_insert(n);
 826         return n;
 827 
 828         // List of "safe" opcodes; those that implicitly block the memory
 829         // op below any null check.
 830       case Op_CastX2P:          // no null checks on native pointers
 831       case Op_Parm:             // 'this' pointer is not null
 832       case Op_LoadP:            // Loading from within a klass
 833       case Op_LoadN:            // Loading from within a klass
 834       case Op_LoadKlass:        // Loading from within a klass
 835       case Op_LoadNKlass:       // Loading from within a klass
 836       case Op_ConP:             // Loading from a klass
 837       case Op_ConN:             // Loading from a klass
 838       case Op_ConNKlass:        // Loading from a klass
 839       case Op_CreateEx:         // Sucking up the guts of an exception oop
 840       case Op_Con:              // Reading from TLS
 841       case Op_CMoveP:           // CMoveP is pinned
 842       case Op_CMoveN:           // CMoveN is pinned
 843         break;                  // No progress
 844 
 845       case Op_Proj:             // Direct call to an allocation routine
 846       case Op_SCMemProj:        // Memory state from store conditional ops
 847 #ifdef ASSERT
 848         {
 849           assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
 850           const Node* call = adr->in(0);
 851           if (call->is_CallJava()) {
 852             const CallJavaNode* call_java = call->as_CallJava();
 853             const TypeTuple *r = call_java->tf()->range();
 854             assert(r->cnt() > TypeFunc::Parms, "must return value");
 855             const Type* ret_type = r->field_at(TypeFunc::Parms);
 856             assert(ret_type && ret_type->isa_ptr(), "must return pointer");
 857             // We further presume that this is one of
 858             // new_instance_Java, new_array_Java, or
 859             // the like, but do not assert for this.
 860           } else if (call->is_Allocate()) {
 861             // similar case to new_instance_Java, etc.
 862           } else if (!call->is_CallLeaf()) {
 863             // Projections from fetch_oop (OSR) are allowed as well.
 864             ShouldNotReachHere();
 865           }
 866         }
 867 #endif
 868         break;
 869       default:
 870         ShouldNotReachHere();
 871       }
 872       break;
 873     }
 874   }
 875 
 876   return  NULL;               // No progress
 877 }
 878 
 879 
 880 //=============================================================================
 881 uint LoadNode::size_of() const { return sizeof(*this); }
 882 uint LoadNode::cmp( const Node &n ) const
 883 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
 884 const Type *LoadNode::bottom_type() const { return _type; }
 885 uint LoadNode::ideal_reg() const {
 886   return _type->ideal_reg();
 887 }
 888 
 889 #ifndef PRODUCT
 890 void LoadNode::dump_spec(outputStream *st) const {
 891   MemNode::dump_spec(st);
 892   if( !Verbose && !WizardMode ) {
 893     // standard dump does this in Verbose and WizardMode
 894     st->print(" #"); _type->dump_on(st);
 895   }
 896 }
 897 #endif
 898 
 899 #ifdef ASSERT
 900 //----------------------------is_immutable_value-------------------------------
 901 // Helper function to allow a raw load without control edge for some cases
 902 bool LoadNode::is_immutable_value(Node* adr) {
 903   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 904           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 905           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 906            in_bytes(JavaThread::osthread_offset())));
 907 }
 908 #endif
 909 
 910 //----------------------------LoadNode::make-----------------------------------
 911 // Polymorphic factory method:
 912 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) {
 913   Compile* C = gvn.C;
 914 
 915   // sanity check the alias category against the created node type
 916   assert(!(adr_type->isa_oopptr() &&
 917            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
 918          "use LoadKlassNode instead");
 919   assert(!(adr_type->isa_aryptr() &&
 920            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
 921          "use LoadRangeNode instead");
 922   // Check control edge of raw loads
 923   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
 924           // oop will be recorded in oop map if load crosses safepoint
 925           rt->isa_oopptr() || is_immutable_value(adr),
 926           "raw memory operations should have control edge");
 927   switch (bt) {
 928   case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(),  mo);
 929   case T_BYTE:    return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(),  mo);
 930   case T_INT:     return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(),  mo);
 931   case T_CHAR:    return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(),  mo);
 932   case T_SHORT:   return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(),  mo);
 933   case T_LONG:    return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo);
 934   case T_FLOAT:   return new (C) LoadFNode (ctl, mem, adr, adr_type, rt,            mo);
 935   case T_DOUBLE:  return new (C) LoadDNode (ctl, mem, adr, adr_type, rt,            mo);
 936   case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(),  mo);
 937   case T_OBJECT:
 938 #ifdef _LP64
 939     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 940       Node* load  = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo));
 941       return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
 942     } else
 943 #endif
 944     {
 945       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
 946       return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo);
 947     }
 948   }
 949   ShouldNotReachHere();
 950   return (LoadNode*)NULL;
 951 }
 952 
 953 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
 954   bool require_atomic = true;
 955   return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic);
 956 }
 957 
 958 LoadDNode* LoadDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
 959   bool require_atomic = true;
 960   return new (C) LoadDNode(ctl, mem, adr, adr_type, rt, mo, require_atomic);
 961 }
 962 
 963 
 964 
 965 //------------------------------hash-------------------------------------------
 966 uint LoadNode::hash() const {
 967   // unroll addition of interesting fields
 968   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
 969 }
 970 
 971 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
 972   if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
 973     bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
 974     bool is_stable_ary = FoldStableValues &&
 975                          (tp != NULL) && (tp->isa_aryptr() != NULL) &&
 976                          tp->isa_aryptr()->is_stable();
 977 
 978     return (eliminate_boxing && non_volatile) || is_stable_ary;
 979   }
 980 
 981   return false;
 982 }
 983 
 984 //---------------------------can_see_stored_value------------------------------
 985 // This routine exists to make sure this set of tests is done the same
 986 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
 987 // will change the graph shape in a way which makes memory alive twice at the
 988 // same time (uses the Oracle model of aliasing), then some
 989 // LoadXNode::Identity will fold things back to the equivalence-class model
 990 // of aliasing.
 991 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
 992   Node* ld_adr = in(MemNode::Address);
 993   intptr_t ld_off = 0;
 994   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 995   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
 996   Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
 997   // This is more general than load from boxing objects.
 998   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
 999     uint alias_idx = atp->index();
1000     bool final = !atp->is_rewritable();
1001     Node* result = NULL;
1002     Node* current = st;
1003     // Skip through chains of MemBarNodes checking the MergeMems for
1004     // new states for the slice of this load.  Stop once any other
1005     // kind of node is encountered.  Loads from final memory can skip
1006     // through any kind of MemBar but normal loads shouldn't skip
1007     // through MemBarAcquire since the could allow them to move out of
1008     // a synchronized region.
1009     while (current->is_Proj()) {
1010       int opc = current->in(0)->Opcode();
1011       if ((final && (opc == Op_MemBarAcquire ||
1012                      opc == Op_MemBarAcquireLock ||
1013                      opc == Op_LoadFence)) ||
1014           opc == Op_MemBarRelease ||
1015           opc == Op_StoreFence ||
1016           opc == Op_MemBarReleaseLock ||
1017           opc == Op_MemBarCPUOrder) {
1018         Node* mem = current->in(0)->in(TypeFunc::Memory);
1019         if (mem->is_MergeMem()) {
1020           MergeMemNode* merge = mem->as_MergeMem();
1021           Node* new_st = merge->memory_at(alias_idx);
1022           if (new_st == merge->base_memory()) {
1023             // Keep searching
1024             current = new_st;
1025             continue;
1026           }
1027           // Save the new memory state for the slice and fall through
1028           // to exit.
1029           result = new_st;
1030         }
1031       }
1032       break;
1033     }
1034     if (result != NULL) {
1035       st = result;
1036     }
1037   }
1038 
1039   // Loop around twice in the case Load -> Initialize -> Store.
1040   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1041   for (int trip = 0; trip <= 1; trip++) {
1042 
1043     if (st->is_Store()) {
1044       Node* st_adr = st->in(MemNode::Address);
1045       if (!phase->eqv(st_adr, ld_adr)) {
1046         // Try harder before giving up...  Match raw and non-raw pointers.
1047         intptr_t st_off = 0;
1048         AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1049         if (alloc == NULL)       return NULL;
1050         if (alloc != ld_alloc)   return NULL;
1051         if (ld_off != st_off)    return NULL;
1052         // At this point we have proven something like this setup:
1053         //  A = Allocate(...)
1054         //  L = LoadQ(,  AddP(CastPP(, A.Parm),, #Off))
1055         //  S = StoreQ(, AddP(,        A.Parm  , #Off), V)
1056         // (Actually, we haven't yet proven the Q's are the same.)
1057         // In other words, we are loading from a casted version of
1058         // the same pointer-and-offset that we stored to.
1059         // Thus, we are able to replace L by V.
1060       }
1061       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1062       if (store_Opcode() != st->Opcode())
1063         return NULL;
1064       return st->in(MemNode::ValueIn);
1065     }
1066 
1067     // A load from a freshly-created object always returns zero.
1068     // (This can happen after LoadNode::Ideal resets the load's memory input
1069     // to find_captured_store, which returned InitializeNode::zero_memory.)
1070     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1071         (st->in(0) == ld_alloc) &&
1072         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1073       // return a zero value for the load's basic type
1074       // (This is one of the few places where a generic PhaseTransform
1075       // can create new nodes.  Think of it as lazily manifesting
1076       // virtually pre-existing constants.)
1077       return phase->zerocon(memory_type());
1078     }
1079 
1080     // A load from an initialization barrier can match a captured store.
1081     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1082       InitializeNode* init = st->in(0)->as_Initialize();
1083       AllocateNode* alloc = init->allocation();
1084       if ((alloc != NULL) && (alloc == ld_alloc)) {
1085         // examine a captured store value
1086         st = init->find_captured_store(ld_off, memory_size(), phase);
1087         if (st != NULL)
1088           continue;             // take one more trip around
1089       }
1090     }
1091 
1092     // Load boxed value from result of valueOf() call is input parameter.
1093     if (this->is_Load() && ld_adr->is_AddP() &&
1094         (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1095       intptr_t ignore = 0;
1096       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1097       if (base != NULL && base->is_Proj() &&
1098           base->as_Proj()->_con == TypeFunc::Parms &&
1099           base->in(0)->is_CallStaticJava() &&
1100           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1101         return base->in(0)->in(TypeFunc::Parms);
1102       }
1103     }
1104 
1105     break;
1106   }
1107 
1108   return NULL;
1109 }
1110 
1111 //----------------------is_instance_field_load_with_local_phi------------------
1112 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1113   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1114       in(Address)->is_AddP() ) {
1115     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1116     // Only instances and boxed values.
1117     if( t_oop != NULL &&
1118         (t_oop->is_ptr_to_boxed_value() ||
1119          t_oop->is_known_instance_field()) &&
1120         t_oop->offset() != Type::OffsetBot &&
1121         t_oop->offset() != Type::OffsetTop) {
1122       return true;
1123     }
1124   }
1125   return false;
1126 }
1127 
1128 //------------------------------Identity---------------------------------------
1129 // Loads are identity if previous store is to same address
1130 Node *LoadNode::Identity( PhaseTransform *phase ) {
1131   // If the previous store-maker is the right kind of Store, and the store is
1132   // to the same address, then we are equal to the value stored.
1133   Node* mem = in(Memory);
1134   Node* value = can_see_stored_value(mem, phase);
1135   if( value ) {
1136     // byte, short & char stores truncate naturally.
1137     // A load has to load the truncated value which requires
1138     // some sort of masking operation and that requires an
1139     // Ideal call instead of an Identity call.
1140     if (memory_size() < BytesPerInt) {
1141       // If the input to the store does not fit with the load's result type,
1142       // it must be truncated via an Ideal call.
1143       if (!phase->type(value)->higher_equal(phase->type(this)))
1144         return this;
1145     }
1146     // (This works even when value is a Con, but LoadNode::Value
1147     // usually runs first, producing the singleton type of the Con.)
1148     return value;
1149   }
1150 
1151   // Search for an existing data phi which was generated before for the same
1152   // instance's field to avoid infinite generation of phis in a loop.
1153   Node *region = mem->in(0);
1154   if (is_instance_field_load_with_local_phi(region)) {
1155     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1156     int this_index  = phase->C->get_alias_index(addr_t);
1157     int this_offset = addr_t->offset();
1158     int this_iid    = addr_t->instance_id();
1159     if (!addr_t->is_known_instance() &&
1160          addr_t->is_ptr_to_boxed_value()) {
1161       // Use _idx of address base (could be Phi node) for boxed values.
1162       intptr_t   ignore = 0;
1163       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1164       this_iid = base->_idx;
1165     }
1166     const Type* this_type = bottom_type();
1167     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1168       Node* phi = region->fast_out(i);
1169       if (phi->is_Phi() && phi != mem &&
1170           phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1171         return phi;
1172       }
1173     }
1174   }
1175 
1176   return this;
1177 }
1178 
1179 // We're loading from an object which has autobox behaviour.
1180 // If this object is result of a valueOf call we'll have a phi
1181 // merging a newly allocated object and a load from the cache.
1182 // We want to replace this load with the original incoming
1183 // argument to the valueOf call.
1184 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1185   assert(phase->C->eliminate_boxing(), "sanity");
1186   intptr_t ignore = 0;
1187   Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1188   if ((base == NULL) || base->is_Phi()) {
1189     // Push the loads from the phi that comes from valueOf up
1190     // through it to allow elimination of the loads and the recovery
1191     // of the original value. It is done in split_through_phi().
1192     return NULL;
1193   } else if (base->is_Load() ||
1194              base->is_DecodeN() && base->in(1)->is_Load()) {
1195     // Eliminate the load of boxed value for integer types from the cache
1196     // array by deriving the value from the index into the array.
1197     // Capture the offset of the load and then reverse the computation.
1198 
1199     // Get LoadN node which loads a boxing object from 'cache' array.
1200     if (base->is_DecodeN()) {
1201       base = base->in(1);
1202     }
1203     if (!base->in(Address)->is_AddP()) {
1204       return NULL; // Complex address
1205     }
1206     AddPNode* address = base->in(Address)->as_AddP();
1207     Node* cache_base = address->in(AddPNode::Base);
1208     if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1209       // Get ConP node which is static 'cache' field.
1210       cache_base = cache_base->in(1);
1211     }
1212     if ((cache_base != NULL) && cache_base->is_Con()) {
1213       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1214       if ((base_type != NULL) && base_type->is_autobox_cache()) {
1215         Node* elements[4];
1216         int shift = exact_log2(type2aelembytes(T_OBJECT));
1217         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1218         if ((count >  0) && elements[0]->is_Con() &&
1219             ((count == 1) ||
1220              (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1221                              elements[1]->in(2) == phase->intcon(shift))) {
1222           ciObjArray* array = base_type->const_oop()->as_obj_array();
1223           // Fetch the box object cache[0] at the base of the array and get its value
1224           ciInstance* box = array->obj_at(0)->as_instance();
1225           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1226           assert(ik->is_box_klass(), "sanity");
1227           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1228           if (ik->nof_nonstatic_fields() == 1) {
1229             // This should be true nonstatic_field_at requires calling
1230             // nof_nonstatic_fields so check it anyway
1231             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1232             BasicType bt = c.basic_type();
1233             // Only integer types have boxing cache.
1234             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1235                    bt == T_BYTE    || bt == T_SHORT ||
1236                    bt == T_INT     || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1237             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1238             if (cache_low != (int)cache_low) {
1239               return NULL; // should not happen since cache is array indexed by value
1240             }
1241             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1242             if (offset != (int)offset) {
1243               return NULL; // should not happen since cache is array indexed by value
1244             }
1245            // Add up all the offsets making of the address of the load
1246             Node* result = elements[0];
1247             for (int i = 1; i < count; i++) {
1248               result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1249             }
1250             // Remove the constant offset from the address and then
1251             result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));
1252             // remove the scaling of the offset to recover the original index.
1253             if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1254               // Peel the shift off directly but wrap it in a dummy node
1255               // since Ideal can't return existing nodes
1256               result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1257             } else if (result->is_Add() && result->in(2)->is_Con() &&
1258                        result->in(1)->Opcode() == Op_LShiftX &&
1259                        result->in(1)->in(2) == phase->intcon(shift)) {
1260               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1261               // but for boxing cache access we know that X<<Z will not overflow
1262               // (there is range check) so we do this optimizatrion by hand here.
1263               Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));
1264               result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));
1265             } else {
1266               result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1267             }
1268 #ifdef _LP64
1269             if (bt != T_LONG) {
1270               result = new (phase->C) ConvL2INode(phase->transform(result));
1271             }
1272 #else
1273             if (bt == T_LONG) {
1274               result = new (phase->C) ConvI2LNode(phase->transform(result));
1275             }
1276 #endif
1277             return result;
1278           }
1279         }
1280       }
1281     }
1282   }
1283   return NULL;
1284 }
1285 
1286 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1287   Node* region = phi->in(0);
1288   if (region == NULL) {
1289     return false; // Wait stable graph
1290   }
1291   uint cnt = phi->req();
1292   for (uint i = 1; i < cnt; i++) {
1293     Node* rc = region->in(i);
1294     if (rc == NULL || phase->type(rc) == Type::TOP)
1295       return false; // Wait stable graph
1296     Node* in = phi->in(i);
1297     if (in == NULL || phase->type(in) == Type::TOP)
1298       return false; // Wait stable graph
1299   }
1300   return true;
1301 }
1302 //------------------------------split_through_phi------------------------------
1303 // Split instance or boxed field load through Phi.
1304 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1305   Node* mem     = in(Memory);
1306   Node* address = in(Address);
1307   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1308 
1309   assert((t_oop != NULL) &&
1310          (t_oop->is_known_instance_field() ||
1311           t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1312 
1313   Compile* C = phase->C;
1314   intptr_t ignore = 0;
1315   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1316   bool base_is_phi = (base != NULL) && base->is_Phi();
1317   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1318                            (base != NULL) && (base == address->in(AddPNode::Base)) &&
1319                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1320 
1321   if (!((mem->is_Phi() || base_is_phi) &&
1322         (load_boxed_values || t_oop->is_known_instance_field()))) {
1323     return NULL; // memory is not Phi
1324   }
1325 
1326   if (mem->is_Phi()) {
1327     if (!stable_phi(mem->as_Phi(), phase)) {
1328       return NULL; // Wait stable graph
1329     }
1330     uint cnt = mem->req();
1331     // Check for loop invariant memory.
1332     if (cnt == 3) {
1333       for (uint i = 1; i < cnt; i++) {
1334         Node* in = mem->in(i);
1335         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1336         if (m == mem) {
1337           set_req(Memory, mem->in(cnt - i));
1338           return this; // made change
1339         }
1340       }
1341     }
1342   }
1343   if (base_is_phi) {
1344     if (!stable_phi(base->as_Phi(), phase)) {
1345       return NULL; // Wait stable graph
1346     }
1347     uint cnt = base->req();
1348     // Check for loop invariant memory.
1349     if (cnt == 3) {
1350       for (uint i = 1; i < cnt; i++) {
1351         if (base->in(i) == base) {
1352           return NULL; // Wait stable graph
1353         }
1354       }
1355     }
1356   }
1357 
1358   bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1359 
1360   // Split through Phi (see original code in loopopts.cpp).
1361   assert(C->have_alias_type(t_oop), "instance should have alias type");
1362 
1363   // Do nothing here if Identity will find a value
1364   // (to avoid infinite chain of value phis generation).
1365   if (!phase->eqv(this, this->Identity(phase)))
1366     return NULL;
1367 
1368   // Select Region to split through.
1369   Node* region;
1370   if (!base_is_phi) {
1371     assert(mem->is_Phi(), "sanity");
1372     region = mem->in(0);
1373     // Skip if the region dominates some control edge of the address.
1374     if (!MemNode::all_controls_dominate(address, region))
1375       return NULL;
1376   } else if (!mem->is_Phi()) {
1377     assert(base_is_phi, "sanity");
1378     region = base->in(0);
1379     // Skip if the region dominates some control edge of the memory.
1380     if (!MemNode::all_controls_dominate(mem, region))
1381       return NULL;
1382   } else if (base->in(0) != mem->in(0)) {
1383     assert(base_is_phi && mem->is_Phi(), "sanity");
1384     if (MemNode::all_controls_dominate(mem, base->in(0))) {
1385       region = base->in(0);
1386     } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1387       region = mem->in(0);
1388     } else {
1389       return NULL; // complex graph
1390     }
1391   } else {
1392     assert(base->in(0) == mem->in(0), "sanity");
1393     region = mem->in(0);
1394   }
1395 
1396   const Type* this_type = this->bottom_type();
1397   int this_index  = C->get_alias_index(t_oop);
1398   int this_offset = t_oop->offset();
1399   int this_iid    = t_oop->instance_id();
1400   if (!t_oop->is_known_instance() && load_boxed_values) {
1401     // Use _idx of address base for boxed values.
1402     this_iid = base->_idx;
1403   }
1404   PhaseIterGVN* igvn = phase->is_IterGVN();
1405   Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1406   for (uint i = 1; i < region->req(); i++) {
1407     Node* x;
1408     Node* the_clone = NULL;
1409     if (region->in(i) == C->top()) {
1410       x = C->top();      // Dead path?  Use a dead data op
1411     } else {
1412       x = this->clone();        // Else clone up the data op
1413       the_clone = x;            // Remember for possible deletion.
1414       // Alter data node to use pre-phi inputs
1415       if (this->in(0) == region) {
1416         x->set_req(0, region->in(i));
1417       } else {
1418         x->set_req(0, NULL);
1419       }
1420       if (mem->is_Phi() && (mem->in(0) == region)) {
1421         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1422       }
1423       if (address->is_Phi() && address->in(0) == region) {
1424         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1425       }
1426       if (base_is_phi && (base->in(0) == region)) {
1427         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1428         Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1429         x->set_req(Address, adr_x);
1430       }
1431     }
1432     // Check for a 'win' on some paths
1433     const Type *t = x->Value(igvn);
1434 
1435     bool singleton = t->singleton();
1436 
1437     // See comments in PhaseIdealLoop::split_thru_phi().
1438     if (singleton && t == Type::TOP) {
1439       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1440     }
1441 
1442     if (singleton) {
1443       x = igvn->makecon(t);
1444     } else {
1445       // We now call Identity to try to simplify the cloned node.
1446       // Note that some Identity methods call phase->type(this).
1447       // Make sure that the type array is big enough for
1448       // our new node, even though we may throw the node away.
1449       // (This tweaking with igvn only works because x is a new node.)
1450       igvn->set_type(x, t);
1451       // If x is a TypeNode, capture any more-precise type permanently into Node
1452       // otherwise it will be not updated during igvn->transform since
1453       // igvn->type(x) is set to x->Value() already.
1454       x->raise_bottom_type(t);
1455       Node *y = x->Identity(igvn);
1456       if (y != x) {
1457         x = y;
1458       } else {
1459         y = igvn->hash_find_insert(x);
1460         if (y) {
1461           x = y;
1462         } else {
1463           // Else x is a new node we are keeping
1464           // We do not need register_new_node_with_optimizer
1465           // because set_type has already been called.
1466           igvn->_worklist.push(x);
1467         }
1468       }
1469     }
1470     if (x != the_clone && the_clone != NULL) {
1471       igvn->remove_dead_node(the_clone);
1472     }
1473     phi->set_req(i, x);
1474   }
1475   // Record Phi
1476   igvn->register_new_node_with_optimizer(phi);
1477   return phi;
1478 }
1479 
1480 //------------------------------Ideal------------------------------------------
1481 // If the load is from Field memory and the pointer is non-null, we can
1482 // zero out the control input.
1483 // If the offset is constant and the base is an object allocation,
1484 // try to hook me up to the exact initializing store.
1485 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1486   Node* p = MemNode::Ideal_common(phase, can_reshape);
1487   if (p)  return (p == NodeSentinel) ? NULL : p;
1488 
1489   Node* ctrl    = in(MemNode::Control);
1490   Node* address = in(MemNode::Address);
1491 
1492   // Skip up past a SafePoint control.  Cannot do this for Stores because
1493   // pointer stores & cardmarks must stay on the same side of a SafePoint.
1494   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1495       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1496     ctrl = ctrl->in(0);
1497     set_req(MemNode::Control,ctrl);
1498   }
1499 
1500   intptr_t ignore = 0;
1501   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1502   if (base != NULL
1503       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1504     // Check for useless control edge in some common special cases
1505     if (in(MemNode::Control) != NULL
1506         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1507         && all_controls_dominate(base, phase->C->start())) {
1508       // A method-invariant, non-null address (constant or 'this' argument).
1509       set_req(MemNode::Control, NULL);
1510     }
1511   }
1512 
1513   Node* mem = in(MemNode::Memory);
1514   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1515 
1516   if (can_reshape && (addr_t != NULL)) {
1517     // try to optimize our memory input
1518     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1519     if (opt_mem != mem) {
1520       set_req(MemNode::Memory, opt_mem);
1521       if (phase->type( opt_mem ) == Type::TOP) return NULL;
1522       return this;
1523     }
1524     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1525     if ((t_oop != NULL) &&
1526         (t_oop->is_known_instance_field() ||
1527          t_oop->is_ptr_to_boxed_value())) {
1528       PhaseIterGVN *igvn = phase->is_IterGVN();
1529       if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1530         // Delay this transformation until memory Phi is processed.
1531         phase->is_IterGVN()->_worklist.push(this);
1532         return NULL;
1533       }
1534       // Split instance field load through Phi.
1535       Node* result = split_through_phi(phase);
1536       if (result != NULL) return result;
1537 
1538       if (t_oop->is_ptr_to_boxed_value()) {
1539         Node* result = eliminate_autobox(phase);
1540         if (result != NULL) return result;
1541       }
1542     }
1543   }
1544 
1545   // Check for prior store with a different base or offset; make Load
1546   // independent.  Skip through any number of them.  Bail out if the stores
1547   // are in an endless dead cycle and report no progress.  This is a key
1548   // transform for Reflection.  However, if after skipping through the Stores
1549   // we can't then fold up against a prior store do NOT do the transform as
1550   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
1551   // array memory alive twice: once for the hoisted Load and again after the
1552   // bypassed Store.  This situation only works if EVERYBODY who does
1553   // anti-dependence work knows how to bypass.  I.e. we need all
1554   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
1555   // the alias index stuff.  So instead, peek through Stores and IFF we can
1556   // fold up, do so.
1557   Node* prev_mem = find_previous_store(phase);
1558   // Steps (a), (b):  Walk past independent stores to find an exact match.
1559   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1560     // (c) See if we can fold up on the spot, but don't fold up here.
1561     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1562     // just return a prior value, which is done by Identity calls.
1563     if (can_see_stored_value(prev_mem, phase)) {
1564       // Make ready for step (d):
1565       set_req(MemNode::Memory, prev_mem);
1566       return this;
1567     }
1568   }
1569 
1570   return NULL;                  // No further progress
1571 }
1572 
1573 // Helper to recognize certain Klass fields which are invariant across
1574 // some group of array types (e.g., int[] or all T[] where T < Object).
1575 const Type*
1576 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1577                                  ciKlass* klass) const {
1578   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1579     // The field is Klass::_modifier_flags.  Return its (constant) value.
1580     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1581     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1582     return TypeInt::make(klass->modifier_flags());
1583   }
1584   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1585     // The field is Klass::_access_flags.  Return its (constant) value.
1586     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1587     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1588     return TypeInt::make(klass->access_flags());
1589   }
1590   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1591     // The field is Klass::_layout_helper.  Return its constant value if known.
1592     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1593     return TypeInt::make(klass->layout_helper());
1594   }
1595 
1596   // No match.
1597   return NULL;
1598 }
1599 
1600 // Try to constant-fold a stable array element.
1601 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1602   assert(ary->const_oop(), "array should be constant");
1603   assert(ary->is_stable(), "array should be stable");
1604 
1605   // Decode the results of GraphKit::array_element_address.
1606   ciArray* aobj = ary->const_oop()->as_array();
1607   ciConstant con = aobj->element_value_by_offset(off);
1608 
1609   if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1610     const Type* con_type = Type::make_from_constant(con);
1611     if (con_type != NULL) {
1612       if (con_type->isa_aryptr()) {
1613         // Join with the array element type, in case it is also stable.
1614         int dim = ary->stable_dimension();
1615         con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1616       }
1617       if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1618         con_type = con_type->make_narrowoop();
1619       }
1620 #ifndef PRODUCT
1621       if (TraceIterativeGVN) {
1622         tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1623         con_type->dump(); tty->cr();
1624       }
1625 #endif //PRODUCT
1626       return con_type;
1627     }
1628   }
1629   return NULL;
1630 }
1631 
1632 //------------------------------Value-----------------------------------------
1633 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1634   // Either input is TOP ==> the result is TOP
1635   Node* mem = in(MemNode::Memory);
1636   const Type *t1 = phase->type(mem);
1637   if (t1 == Type::TOP)  return Type::TOP;
1638   Node* adr = in(MemNode::Address);
1639   const TypePtr* tp = phase->type(adr)->isa_ptr();
1640   if (tp == NULL || tp->empty())  return Type::TOP;
1641   int off = tp->offset();
1642   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1643   Compile* C = phase->C;
1644 
1645   // Try to guess loaded type from pointer type
1646   if (tp->isa_aryptr()) {
1647     const TypeAryPtr* ary = tp->is_aryptr();
1648     const Type* t = ary->elem();
1649 
1650     // Determine whether the reference is beyond the header or not, by comparing
1651     // the offset against the offset of the start of the array's data.
1652     // Different array types begin at slightly different offsets (12 vs. 16).
1653     // We choose T_BYTE as an example base type that is least restrictive
1654     // as to alignment, which will therefore produce the smallest
1655     // possible base offset.
1656     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1657     const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1658 
1659     // Try to constant-fold a stable array element.
1660     if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) {
1661       // Make sure the reference is not into the header and the offset is constant
1662       if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1663         const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1664         if (con_type != NULL) {
1665           return con_type;
1666         }
1667       }
1668     }
1669 
1670     // Don't do this for integer types. There is only potential profit if
1671     // the element type t is lower than _type; that is, for int types, if _type is
1672     // more restrictive than t.  This only happens here if one is short and the other
1673     // char (both 16 bits), and in those cases we've made an intentional decision
1674     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1675     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1676     //
1677     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1678     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
1679     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1680     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
1681     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1682     // In fact, that could have been the original type of p1, and p1 could have
1683     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1684     // expression (LShiftL quux 3) independently optimized to the constant 8.
1685     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1686         && (_type->isa_vect() == NULL)
1687         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1688       // t might actually be lower than _type, if _type is a unique
1689       // concrete subclass of abstract class t.
1690       if (off_beyond_header) {  // is the offset beyond the header?
1691         const Type* jt = t->join_speculative(_type);
1692         // In any case, do not allow the join, per se, to empty out the type.
1693         if (jt->empty() && !t->empty()) {
1694           // This can happen if a interface-typed array narrows to a class type.
1695           jt = _type;
1696         }
1697 #ifdef ASSERT
1698         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1699           // The pointers in the autobox arrays are always non-null
1700           Node* base = adr->in(AddPNode::Base);
1701           if ((base != NULL) && base->is_DecodeN()) {
1702             // Get LoadN node which loads IntegerCache.cache field
1703             base = base->in(1);
1704           }
1705           if ((base != NULL) && base->is_Con()) {
1706             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1707             if ((base_type != NULL) && base_type->is_autobox_cache()) {
1708               // It could be narrow oop
1709               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1710             }
1711           }
1712         }
1713 #endif
1714         return jt;
1715       }
1716     }
1717   } else if (tp->base() == Type::InstPtr) {
1718     ciEnv* env = C->env();
1719     const TypeInstPtr* tinst = tp->is_instptr();
1720     ciKlass* klass = tinst->klass();
1721     assert( off != Type::OffsetBot ||
1722             // arrays can be cast to Objects
1723             tp->is_oopptr()->klass()->is_java_lang_Object() ||
1724             // unsafe field access may not have a constant offset
1725             C->has_unsafe_access(),
1726             "Field accesses must be precise" );
1727     // For oop loads, we expect the _type to be precise
1728     if (klass == env->String_klass() &&
1729         adr->is_AddP() && off != Type::OffsetBot) {
1730       // For constant Strings treat the final fields as compile time constants.
1731       Node* base = adr->in(AddPNode::Base);
1732       const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1733       if (t != NULL && t->singleton()) {
1734         ciField* field = env->String_klass()->get_field_by_offset(off, false);
1735         if (field != NULL && field->is_final()) {
1736           ciObject* string = t->const_oop();
1737           ciConstant constant = string->as_instance()->field_value(field);
1738           if (constant.basic_type() == T_INT) {
1739             return TypeInt::make(constant.as_int());
1740           } else if (constant.basic_type() == T_ARRAY) {
1741             if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1742               return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1743             } else {
1744               return TypeOopPtr::make_from_constant(constant.as_object(), true);
1745             }
1746           }
1747         }
1748       }
1749     }
1750     // Optimizations for constant objects
1751     ciObject* const_oop = tinst->const_oop();
1752     if (const_oop != NULL) {
1753       // For constant Boxed value treat the target field as a compile time constant.
1754       if (tinst->is_ptr_to_boxed_value()) {
1755         return tinst->get_const_boxed_value();
1756       } else
1757       // For constant CallSites treat the target field as a compile time constant.
1758       if (const_oop->is_call_site()) {
1759         ciCallSite* call_site = const_oop->as_call_site();
1760         ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1761         if (field != NULL && field->is_call_site_target()) {
1762           ciMethodHandle* target = call_site->get_target();
1763           if (target != NULL) {  // just in case
1764             ciConstant constant(T_OBJECT, target);
1765             const Type* t;
1766             if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1767               t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1768             } else {
1769               t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1770             }
1771             // Add a dependence for invalidation of the optimization.
1772             if (!call_site->is_constant_call_site()) {
1773               C->dependencies()->assert_call_site_target_value(call_site, target);
1774             }
1775             return t;
1776           }
1777         }
1778       }
1779     }
1780   } else if (tp->base() == Type::KlassPtr) {
1781     assert( off != Type::OffsetBot ||
1782             // arrays can be cast to Objects
1783             tp->is_klassptr()->klass()->is_java_lang_Object() ||
1784             // also allow array-loading from the primary supertype
1785             // array during subtype checks
1786             Opcode() == Op_LoadKlass,
1787             "Field accesses must be precise" );
1788     // For klass/static loads, we expect the _type to be precise
1789   }
1790 
1791   const TypeKlassPtr *tkls = tp->isa_klassptr();
1792   if (tkls != NULL && !StressReflectiveCode) {
1793     ciKlass* klass = tkls->klass();
1794     if (klass->is_loaded() && tkls->klass_is_exact()) {
1795       // We are loading a field from a Klass metaobject whose identity
1796       // is known at compile time (the type is "exact" or "precise").
1797       // Check for fields we know are maintained as constants by the VM.
1798       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1799         // The field is Klass::_super_check_offset.  Return its (constant) value.
1800         // (Folds up type checking code.)
1801         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1802         return TypeInt::make(klass->super_check_offset());
1803       }
1804       // Compute index into primary_supers array
1805       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1806       // Check for overflowing; use unsigned compare to handle the negative case.
1807       if( depth < ciKlass::primary_super_limit() ) {
1808         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1809         // (Folds up type checking code.)
1810         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1811         ciKlass *ss = klass->super_of_depth(depth);
1812         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1813       }
1814       const Type* aift = load_array_final_field(tkls, klass);
1815       if (aift != NULL)  return aift;
1816       if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1817           && klass->is_array_klass()) {
1818         // The field is ArrayKlass::_component_mirror.  Return its (constant) value.
1819         // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1820         assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1821         return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1822       }
1823       if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1824         // The field is Klass::_java_mirror.  Return its (constant) value.
1825         // (Folds up the 2nd indirection in anObjConstant.getClass().)
1826         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1827         return TypeInstPtr::make(klass->java_mirror());
1828       }
1829     }
1830 
1831     // We can still check if we are loading from the primary_supers array at a
1832     // shallow enough depth.  Even though the klass is not exact, entries less
1833     // than or equal to its super depth are correct.
1834     if (klass->is_loaded() ) {
1835       ciType *inner = klass;
1836       while( inner->is_obj_array_klass() )
1837         inner = inner->as_obj_array_klass()->base_element_type();
1838       if( inner->is_instance_klass() &&
1839           !inner->as_instance_klass()->flags().is_interface() ) {
1840         // Compute index into primary_supers array
1841         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1842         // Check for overflowing; use unsigned compare to handle the negative case.
1843         if( depth < ciKlass::primary_super_limit() &&
1844             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1845           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1846           // (Folds up type checking code.)
1847           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1848           ciKlass *ss = klass->super_of_depth(depth);
1849           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1850         }
1851       }
1852     }
1853 
1854     // If the type is enough to determine that the thing is not an array,
1855     // we can give the layout_helper a positive interval type.
1856     // This will help short-circuit some reflective code.
1857     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1858         && !klass->is_array_klass() // not directly typed as an array
1859         && !klass->is_interface()  // specifically not Serializable & Cloneable
1860         && !klass->is_java_lang_Object()   // not the supertype of all T[]
1861         ) {
1862       // Note:  When interfaces are reliable, we can narrow the interface
1863       // test to (klass != Serializable && klass != Cloneable).
1864       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1865       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1866       // The key property of this type is that it folds up tests
1867       // for array-ness, since it proves that the layout_helper is positive.
1868       // Thus, a generic value like the basic object layout helper works fine.
1869       return TypeInt::make(min_size, max_jint, Type::WidenMin);
1870     }
1871   }
1872 
1873   // If we are loading from a freshly-allocated object, produce a zero,
1874   // if the load is provably beyond the header of the object.
1875   // (Also allow a variable load from a fresh array to produce zero.)
1876   const TypeOopPtr *tinst = tp->isa_oopptr();
1877   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1878   bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1879   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1880     Node* value = can_see_stored_value(mem,phase);
1881     if (value != NULL && value->is_Con()) {
1882       assert(value->bottom_type()->higher_equal(_type),"sanity");
1883       return value->bottom_type();
1884     }
1885   }
1886 
1887   if (is_instance) {
1888     // If we have an instance type and our memory input is the
1889     // programs's initial memory state, there is no matching store,
1890     // so just return a zero of the appropriate type
1891     Node *mem = in(MemNode::Memory);
1892     if (mem->is_Parm() && mem->in(0)->is_Start()) {
1893       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1894       return Type::get_zero_type(_type->basic_type());
1895     }
1896   }
1897   return _type;
1898 }
1899 
1900 //------------------------------match_edge-------------------------------------
1901 // Do we Match on this edge index or not?  Match only the address.
1902 uint LoadNode::match_edge(uint idx) const {
1903   return idx == MemNode::Address;
1904 }
1905 
1906 //--------------------------LoadBNode::Ideal--------------------------------------
1907 //
1908 //  If the previous store is to the same address as this load,
1909 //  and the value stored was larger than a byte, replace this load
1910 //  with the value stored truncated to a byte.  If no truncation is
1911 //  needed, the replacement is done in LoadNode::Identity().
1912 //
1913 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1914   Node* mem = in(MemNode::Memory);
1915   Node* value = can_see_stored_value(mem,phase);
1916   if( value && !phase->type(value)->higher_equal( _type ) ) {
1917     Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1918     return new (phase->C) RShiftINode(result, phase->intcon(24));
1919   }
1920   // Identity call will handle the case where truncation is not needed.
1921   return LoadNode::Ideal(phase, can_reshape);
1922 }
1923 
1924 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1925   Node* mem = in(MemNode::Memory);
1926   Node* value = can_see_stored_value(mem,phase);
1927   if (value != NULL && value->is_Con() &&
1928       !value->bottom_type()->higher_equal(_type)) {
1929     // If the input to the store does not fit with the load's result type,
1930     // it must be truncated. We can't delay until Ideal call since
1931     // a singleton Value is needed for split_thru_phi optimization.
1932     int con = value->get_int();
1933     return TypeInt::make((con << 24) >> 24);
1934   }
1935   return LoadNode::Value(phase);
1936 }
1937 
1938 //--------------------------LoadUBNode::Ideal-------------------------------------
1939 //
1940 //  If the previous store is to the same address as this load,
1941 //  and the value stored was larger than a byte, replace this load
1942 //  with the value stored truncated to a byte.  If no truncation is
1943 //  needed, the replacement is done in LoadNode::Identity().
1944 //
1945 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1946   Node* mem = in(MemNode::Memory);
1947   Node* value = can_see_stored_value(mem, phase);
1948   if (value && !phase->type(value)->higher_equal(_type))
1949     return new (phase->C) AndINode(value, phase->intcon(0xFF));
1950   // Identity call will handle the case where truncation is not needed.
1951   return LoadNode::Ideal(phase, can_reshape);
1952 }
1953 
1954 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1955   Node* mem = in(MemNode::Memory);
1956   Node* value = can_see_stored_value(mem,phase);
1957   if (value != NULL && value->is_Con() &&
1958       !value->bottom_type()->higher_equal(_type)) {
1959     // If the input to the store does not fit with the load's result type,
1960     // it must be truncated. We can't delay until Ideal call since
1961     // a singleton Value is needed for split_thru_phi optimization.
1962     int con = value->get_int();
1963     return TypeInt::make(con & 0xFF);
1964   }
1965   return LoadNode::Value(phase);
1966 }
1967 
1968 //--------------------------LoadUSNode::Ideal-------------------------------------
1969 //
1970 //  If the previous store is to the same address as this load,
1971 //  and the value stored was larger than a char, replace this load
1972 //  with the value stored truncated to a char.  If no truncation is
1973 //  needed, the replacement is done in LoadNode::Identity().
1974 //
1975 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1976   Node* mem = in(MemNode::Memory);
1977   Node* value = can_see_stored_value(mem,phase);
1978   if( value && !phase->type(value)->higher_equal( _type ) )
1979     return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1980   // Identity call will handle the case where truncation is not needed.
1981   return LoadNode::Ideal(phase, can_reshape);
1982 }
1983 
1984 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1985   Node* mem = in(MemNode::Memory);
1986   Node* value = can_see_stored_value(mem,phase);
1987   if (value != NULL && value->is_Con() &&
1988       !value->bottom_type()->higher_equal(_type)) {
1989     // If the input to the store does not fit with the load's result type,
1990     // it must be truncated. We can't delay until Ideal call since
1991     // a singleton Value is needed for split_thru_phi optimization.
1992     int con = value->get_int();
1993     return TypeInt::make(con & 0xFFFF);
1994   }
1995   return LoadNode::Value(phase);
1996 }
1997 
1998 //--------------------------LoadSNode::Ideal--------------------------------------
1999 //
2000 //  If the previous store is to the same address as this load,
2001 //  and the value stored was larger than a short, replace this load
2002 //  with the value stored truncated to a short.  If no truncation is
2003 //  needed, the replacement is done in LoadNode::Identity().
2004 //
2005 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2006   Node* mem = in(MemNode::Memory);
2007   Node* value = can_see_stored_value(mem,phase);
2008   if( value && !phase->type(value)->higher_equal( _type ) ) {
2009     Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
2010     return new (phase->C) RShiftINode(result, phase->intcon(16));
2011   }
2012   // Identity call will handle the case where truncation is not needed.
2013   return LoadNode::Ideal(phase, can_reshape);
2014 }
2015 
2016 const Type* LoadSNode::Value(PhaseTransform *phase) const {
2017   Node* mem = in(MemNode::Memory);
2018   Node* value = can_see_stored_value(mem,phase);
2019   if (value != NULL && value->is_Con() &&
2020       !value->bottom_type()->higher_equal(_type)) {
2021     // If the input to the store does not fit with the load's result type,
2022     // it must be truncated. We can't delay until Ideal call since
2023     // a singleton Value is needed for split_thru_phi optimization.
2024     int con = value->get_int();
2025     return TypeInt::make((con << 16) >> 16);
2026   }
2027   return LoadNode::Value(phase);
2028 }
2029 
2030 //=============================================================================
2031 //----------------------------LoadKlassNode::make------------------------------
2032 // Polymorphic factory method:
2033 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
2034   Compile* C = gvn.C;
2035   Node *ctl = NULL;
2036   // sanity check the alias category against the created node type
2037   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2038   assert(adr_type != NULL, "expecting TypeKlassPtr");
2039 #ifdef _LP64
2040   if (adr_type->is_ptr_to_narrowklass()) {
2041     assert(UseCompressedClassPointers, "no compressed klasses");
2042     Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2043     return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2044   }
2045 #endif
2046   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2047   return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2048 }
2049 
2050 //------------------------------Value------------------------------------------
2051 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
2052   return klass_value_common(phase);
2053 }
2054 
2055 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
2056   // Either input is TOP ==> the result is TOP
2057   const Type *t1 = phase->type( in(MemNode::Memory) );
2058   if (t1 == Type::TOP)  return Type::TOP;
2059   Node *adr = in(MemNode::Address);
2060   const Type *t2 = phase->type( adr );
2061   if (t2 == Type::TOP)  return Type::TOP;
2062   const TypePtr *tp = t2->is_ptr();
2063   if (TypePtr::above_centerline(tp->ptr()) ||
2064       tp->ptr() == TypePtr::Null)  return Type::TOP;
2065 
2066   // Return a more precise klass, if possible
2067   const TypeInstPtr *tinst = tp->isa_instptr();
2068   if (tinst != NULL) {
2069     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2070     int offset = tinst->offset();
2071     if (ik == phase->C->env()->Class_klass()
2072         && (offset == java_lang_Class::klass_offset_in_bytes() ||
2073             offset == java_lang_Class::array_klass_offset_in_bytes())) {
2074       // We are loading a special hidden field from a Class mirror object,
2075       // the field which points to the VM's Klass metaobject.
2076       ciType* t = tinst->java_mirror_type();
2077       // java_mirror_type returns non-null for compile-time Class constants.
2078       if (t != NULL) {
2079         // constant oop => constant klass
2080         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2081           if (t->is_void()) {
2082             // We cannot create a void array.  Since void is a primitive type return null
2083             // klass.  Users of this result need to do a null check on the returned klass.
2084             return TypePtr::NULL_PTR;
2085           }
2086           return TypeKlassPtr::make(ciArrayKlass::make(t));
2087         }
2088         if (!t->is_klass()) {
2089           // a primitive Class (e.g., int.class) has NULL for a klass field
2090           return TypePtr::NULL_PTR;
2091         }
2092         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2093         return TypeKlassPtr::make(t->as_klass());
2094       }
2095       // non-constant mirror, so we can't tell what's going on
2096     }
2097     if( !ik->is_loaded() )
2098       return _type;             // Bail out if not loaded
2099     if (offset == oopDesc::klass_offset_in_bytes()) {
2100       if (tinst->klass_is_exact()) {
2101         return TypeKlassPtr::make(ik);
2102       }
2103       // See if we can become precise: no subklasses and no interface
2104       // (Note:  We need to support verified interfaces.)
2105       if (!ik->is_interface() && !ik->has_subklass()) {
2106         //assert(!UseExactTypes, "this code should be useless with exact types");
2107         // Add a dependence; if any subclass added we need to recompile
2108         if (!ik->is_final()) {
2109           // %%% should use stronger assert_unique_concrete_subtype instead
2110           phase->C->dependencies()->assert_leaf_type(ik);
2111         }
2112         // Return precise klass
2113         return TypeKlassPtr::make(ik);
2114       }
2115 
2116       // Return root of possible klass
2117       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2118     }
2119   }
2120 
2121   // Check for loading klass from an array
2122   const TypeAryPtr *tary = tp->isa_aryptr();
2123   if( tary != NULL ) {
2124     ciKlass *tary_klass = tary->klass();
2125     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
2126         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2127       if (tary->klass_is_exact()) {
2128         return TypeKlassPtr::make(tary_klass);
2129       }
2130       ciArrayKlass *ak = tary->klass()->as_array_klass();
2131       // If the klass is an object array, we defer the question to the
2132       // array component klass.
2133       if( ak->is_obj_array_klass() ) {
2134         assert( ak->is_loaded(), "" );
2135         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2136         if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2137           ciInstanceKlass* ik = base_k->as_instance_klass();
2138           // See if we can become precise: no subklasses and no interface
2139           if (!ik->is_interface() && !ik->has_subklass()) {
2140             //assert(!UseExactTypes, "this code should be useless with exact types");
2141             // Add a dependence; if any subclass added we need to recompile
2142             if (!ik->is_final()) {
2143               phase->C->dependencies()->assert_leaf_type(ik);
2144             }
2145             // Return precise array klass
2146             return TypeKlassPtr::make(ak);
2147           }
2148         }
2149         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2150       } else {                  // Found a type-array?
2151         //assert(!UseExactTypes, "this code should be useless with exact types");
2152         assert( ak->is_type_array_klass(), "" );
2153         return TypeKlassPtr::make(ak); // These are always precise
2154       }
2155     }
2156   }
2157 
2158   // Check for loading klass from an array klass
2159   const TypeKlassPtr *tkls = tp->isa_klassptr();
2160   if (tkls != NULL && !StressReflectiveCode) {
2161     ciKlass* klass = tkls->klass();
2162     if( !klass->is_loaded() )
2163       return _type;             // Bail out if not loaded
2164     if( klass->is_obj_array_klass() &&
2165         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2166       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2167       // // Always returning precise element type is incorrect,
2168       // // e.g., element type could be object and array may contain strings
2169       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2170 
2171       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2172       // according to the element type's subclassing.
2173       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2174     }
2175     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2176         tkls->offset() == in_bytes(Klass::super_offset())) {
2177       ciKlass* sup = klass->as_instance_klass()->super();
2178       // The field is Klass::_super.  Return its (constant) value.
2179       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2180       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2181     }
2182   }
2183 
2184   // Bailout case
2185   return LoadNode::Value(phase);
2186 }
2187 
2188 //------------------------------Identity---------------------------------------
2189 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2190 // Also feed through the klass in Allocate(...klass...)._klass.
2191 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2192   return klass_identity_common(phase);
2193 }
2194 
2195 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2196   Node* x = LoadNode::Identity(phase);
2197   if (x != this)  return x;
2198 
2199   // Take apart the address into an oop and and offset.
2200   // Return 'this' if we cannot.
2201   Node*    adr    = in(MemNode::Address);
2202   intptr_t offset = 0;
2203   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2204   if (base == NULL)     return this;
2205   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2206   if (toop == NULL)     return this;
2207 
2208   // We can fetch the klass directly through an AllocateNode.
2209   // This works even if the klass is not constant (clone or newArray).
2210   if (offset == oopDesc::klass_offset_in_bytes()) {
2211     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2212     if (allocated_klass != NULL) {
2213       return allocated_klass;
2214     }
2215   }
2216 
2217   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2218   // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
2219   // See inline_native_Class_query for occurrences of these patterns.
2220   // Java Example:  x.getClass().isAssignableFrom(y)
2221   // Java Example:  Array.newInstance(x.getClass().getComponentType(), n)
2222   //
2223   // This improves reflective code, often making the Class
2224   // mirror go completely dead.  (Current exception:  Class
2225   // mirrors may appear in debug info, but we could clean them out by
2226   // introducing a new debug info operator for Klass*.java_mirror).
2227   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2228       && (offset == java_lang_Class::klass_offset_in_bytes() ||
2229           offset == java_lang_Class::array_klass_offset_in_bytes())) {
2230     // We are loading a special hidden field from a Class mirror,
2231     // the field which points to its Klass or ArrayKlass metaobject.
2232     if (base->is_Load()) {
2233       Node* adr2 = base->in(MemNode::Address);
2234       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2235       if (tkls != NULL && !tkls->empty()
2236           && (tkls->klass()->is_instance_klass() ||
2237               tkls->klass()->is_array_klass())
2238           && adr2->is_AddP()
2239           ) {
2240         int mirror_field = in_bytes(Klass::java_mirror_offset());
2241         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2242           mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
2243         }
2244         if (tkls->offset() == mirror_field) {
2245           return adr2->in(AddPNode::Base);
2246         }
2247       }
2248     }
2249   }
2250 
2251   return this;
2252 }
2253 
2254 
2255 //------------------------------Value------------------------------------------
2256 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2257   const Type *t = klass_value_common(phase);
2258   if (t == Type::TOP)
2259     return t;
2260 
2261   return t->make_narrowklass();
2262 }
2263 
2264 //------------------------------Identity---------------------------------------
2265 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2266 // Also feed through the klass in Allocate(...klass...)._klass.
2267 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2268   Node *x = klass_identity_common(phase);
2269 
2270   const Type *t = phase->type( x );
2271   if( t == Type::TOP ) return x;
2272   if( t->isa_narrowklass()) return x;
2273   assert (!t->isa_narrowoop(), "no narrow oop here");
2274 
2275   return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
2276 }
2277 
2278 //------------------------------Value-----------------------------------------
2279 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2280   // Either input is TOP ==> the result is TOP
2281   const Type *t1 = phase->type( in(MemNode::Memory) );
2282   if( t1 == Type::TOP ) return Type::TOP;
2283   Node *adr = in(MemNode::Address);
2284   const Type *t2 = phase->type( adr );
2285   if( t2 == Type::TOP ) return Type::TOP;
2286   const TypePtr *tp = t2->is_ptr();
2287   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2288   const TypeAryPtr *tap = tp->isa_aryptr();
2289   if( !tap ) return _type;
2290   return tap->size();
2291 }
2292 
2293 //-------------------------------Ideal---------------------------------------
2294 // Feed through the length in AllocateArray(...length...)._length.
2295 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2296   Node* p = MemNode::Ideal_common(phase, can_reshape);
2297   if (p)  return (p == NodeSentinel) ? NULL : p;
2298 
2299   // Take apart the address into an oop and and offset.
2300   // Return 'this' if we cannot.
2301   Node*    adr    = in(MemNode::Address);
2302   intptr_t offset = 0;
2303   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2304   if (base == NULL)     return NULL;
2305   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2306   if (tary == NULL)     return NULL;
2307 
2308   // We can fetch the length directly through an AllocateArrayNode.
2309   // This works even if the length is not constant (clone or newArray).
2310   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2311     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2312     if (alloc != NULL) {
2313       Node* allocated_length = alloc->Ideal_length();
2314       Node* len = alloc->make_ideal_length(tary, phase);
2315       if (allocated_length != len) {
2316         // New CastII improves on this.
2317         return len;
2318       }
2319     }
2320   }
2321 
2322   return NULL;
2323 }
2324 
2325 //------------------------------Identity---------------------------------------
2326 // Feed through the length in AllocateArray(...length...)._length.
2327 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2328   Node* x = LoadINode::Identity(phase);
2329   if (x != this)  return x;
2330 
2331   // Take apart the address into an oop and and offset.
2332   // Return 'this' if we cannot.
2333   Node*    adr    = in(MemNode::Address);
2334   intptr_t offset = 0;
2335   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2336   if (base == NULL)     return this;
2337   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2338   if (tary == NULL)     return this;
2339 
2340   // We can fetch the length directly through an AllocateArrayNode.
2341   // This works even if the length is not constant (clone or newArray).
2342   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2343     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2344     if (alloc != NULL) {
2345       Node* allocated_length = alloc->Ideal_length();
2346       // Do not allow make_ideal_length to allocate a CastII node.
2347       Node* len = alloc->make_ideal_length(tary, phase, false);
2348       if (allocated_length == len) {
2349         // Return allocated_length only if it would not be improved by a CastII.
2350         return allocated_length;
2351       }
2352     }
2353   }
2354 
2355   return this;
2356 
2357 }
2358 
2359 //=============================================================================
2360 //---------------------------StoreNode::make-----------------------------------
2361 // Polymorphic factory method:
2362 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2363   assert((mo == unordered || mo == release), "unexpected");
2364   Compile* C = gvn.C;
2365   assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2366          ctl != NULL, "raw memory operations should have control edge");
2367 
2368   switch (bt) {
2369   case T_BOOLEAN:
2370   case T_BYTE:    return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo);
2371   case T_INT:     return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo);
2372   case T_CHAR:
2373   case T_SHORT:   return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo);
2374   case T_LONG:    return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo);
2375   case T_FLOAT:   return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo);
2376   case T_DOUBLE:  return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo);
2377   case T_METADATA:
2378   case T_ADDRESS:
2379   case T_OBJECT:
2380 #ifdef _LP64
2381     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2382       val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2383       return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo);
2384     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2385                (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2386                 adr->bottom_type()->isa_rawptr())) {
2387       val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2388       return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2389     }
2390 #endif
2391     {
2392       return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo);
2393     }
2394   }
2395   ShouldNotReachHere();
2396   return (StoreNode*)NULL;
2397 }
2398 
2399 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2400   bool require_atomic = true;
2401   return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2402 }
2403 
2404 StoreDNode* StoreDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2405   bool require_atomic = true;
2406   return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2407 }
2408 
2409 
2410 //--------------------------bottom_type----------------------------------------
2411 const Type *StoreNode::bottom_type() const {
2412   return Type::MEMORY;
2413 }
2414 
2415 //------------------------------hash-------------------------------------------
2416 uint StoreNode::hash() const {
2417   // unroll addition of interesting fields
2418   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2419 
2420   // Since they are not commoned, do not hash them:
2421   return NO_HASH;
2422 }
2423 
2424 //------------------------------Ideal------------------------------------------
2425 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2426 // When a store immediately follows a relevant allocation/initialization,
2427 // try to capture it into the initialization, or hoist it above.
2428 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2429   Node* p = MemNode::Ideal_common(phase, can_reshape);
2430   if (p)  return (p == NodeSentinel) ? NULL : p;
2431 
2432   Node* mem     = in(MemNode::Memory);
2433   Node* address = in(MemNode::Address);
2434 
2435   // Back-to-back stores to same address?  Fold em up.  Generally
2436   // unsafe if I have intervening uses...  Also disallowed for StoreCM
2437   // since they must follow each StoreP operation.  Redundant StoreCMs
2438   // are eliminated just before matching in final_graph_reshape.
2439   if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2440       mem->Opcode() != Op_StoreCM) {
2441     // Looking at a dead closed cycle of memory?
2442     assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2443 
2444     assert(Opcode() == mem->Opcode() ||
2445            phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2446            "no mismatched stores, except on raw memory");
2447 
2448     if (mem->outcnt() == 1 &&           // check for intervening uses
2449         mem->as_Store()->memory_size() <= this->memory_size()) {
2450       // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2451       // For example, 'mem' might be the final state at a conditional return.
2452       // Or, 'mem' might be used by some node which is live at the same time
2453       // 'this' is live, which might be unschedulable.  So, require exactly
2454       // ONE user, the 'this' store, until such time as we clone 'mem' for
2455       // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2456       if (can_reshape) {  // (%%% is this an anachronism?)
2457         set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2458                   phase->is_IterGVN());
2459       } else {
2460         // It's OK to do this in the parser, since DU info is always accurate,
2461         // and the parser always refers to nodes via SafePointNode maps.
2462         set_req(MemNode::Memory, mem->in(MemNode::Memory));
2463       }
2464       return this;
2465     }
2466   }
2467 
2468   // Capture an unaliased, unconditional, simple store into an initializer.
2469   // Or, if it is independent of the allocation, hoist it above the allocation.
2470   if (ReduceFieldZeroing && /*can_reshape &&*/
2471       mem->is_Proj() && mem->in(0)->is_Initialize()) {
2472     InitializeNode* init = mem->in(0)->as_Initialize();
2473     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2474     if (offset > 0) {
2475       Node* moved = init->capture_store(this, offset, phase, can_reshape);
2476       // If the InitializeNode captured me, it made a raw copy of me,
2477       // and I need to disappear.
2478       if (moved != NULL) {
2479         // %%% hack to ensure that Ideal returns a new node:
2480         mem = MergeMemNode::make(phase->C, mem);
2481         return mem;             // fold me away
2482       }
2483     }
2484   }
2485 
2486   return NULL;                  // No further progress
2487 }
2488 
2489 //------------------------------Value-----------------------------------------
2490 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2491   // Either input is TOP ==> the result is TOP
2492   const Type *t1 = phase->type( in(MemNode::Memory) );
2493   if( t1 == Type::TOP ) return Type::TOP;
2494   const Type *t2 = phase->type( in(MemNode::Address) );
2495   if( t2 == Type::TOP ) return Type::TOP;
2496   const Type *t3 = phase->type( in(MemNode::ValueIn) );
2497   if( t3 == Type::TOP ) return Type::TOP;
2498   return Type::MEMORY;
2499 }
2500 
2501 //------------------------------Identity---------------------------------------
2502 // Remove redundant stores:
2503 //   Store(m, p, Load(m, p)) changes to m.
2504 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2505 Node *StoreNode::Identity( PhaseTransform *phase ) {
2506   Node* mem = in(MemNode::Memory);
2507   Node* adr = in(MemNode::Address);
2508   Node* val = in(MemNode::ValueIn);
2509 
2510   // Load then Store?  Then the Store is useless
2511   if (val->is_Load() &&
2512       val->in(MemNode::Address)->eqv_uncast(adr) &&
2513       val->in(MemNode::Memory )->eqv_uncast(mem) &&
2514       val->as_Load()->store_Opcode() == Opcode()) {
2515     return mem;
2516   }
2517 
2518   // Two stores in a row of the same value?
2519   if (mem->is_Store() &&
2520       mem->in(MemNode::Address)->eqv_uncast(adr) &&
2521       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2522       mem->Opcode() == Opcode()) {
2523     return mem;
2524   }
2525 
2526   // Store of zero anywhere into a freshly-allocated object?
2527   // Then the store is useless.
2528   // (It must already have been captured by the InitializeNode.)
2529   if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2530     // a newly allocated object is already all-zeroes everywhere
2531     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2532       return mem;
2533     }
2534 
2535     // the store may also apply to zero-bits in an earlier object
2536     Node* prev_mem = find_previous_store(phase);
2537     // Steps (a), (b):  Walk past independent stores to find an exact match.
2538     if (prev_mem != NULL) {
2539       Node* prev_val = can_see_stored_value(prev_mem, phase);
2540       if (prev_val != NULL && phase->eqv(prev_val, val)) {
2541         // prev_val and val might differ by a cast; it would be good
2542         // to keep the more informative of the two.
2543         return mem;
2544       }
2545     }
2546   }
2547 
2548   return this;
2549 }
2550 
2551 //------------------------------match_edge-------------------------------------
2552 // Do we Match on this edge index or not?  Match only memory & value
2553 uint StoreNode::match_edge(uint idx) const {
2554   return idx == MemNode::Address || idx == MemNode::ValueIn;
2555 }
2556 
2557 //------------------------------cmp--------------------------------------------
2558 // Do not common stores up together.  They generally have to be split
2559 // back up anyways, so do not bother.
2560 uint StoreNode::cmp( const Node &n ) const {
2561   return (&n == this);          // Always fail except on self
2562 }
2563 
2564 //------------------------------Ideal_masked_input-----------------------------
2565 // Check for a useless mask before a partial-word store
2566 // (StoreB ... (AndI valIn conIa) )
2567 // If (conIa & mask == mask) this simplifies to
2568 // (StoreB ... (valIn) )
2569 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2570   Node *val = in(MemNode::ValueIn);
2571   if( val->Opcode() == Op_AndI ) {
2572     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2573     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2574       set_req(MemNode::ValueIn, val->in(1));
2575       return this;
2576     }
2577   }
2578   return NULL;
2579 }
2580 
2581 
2582 //------------------------------Ideal_sign_extended_input----------------------
2583 // Check for useless sign-extension before a partial-word store
2584 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2585 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
2586 // (StoreB ... (valIn) )
2587 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2588   Node *val = in(MemNode::ValueIn);
2589   if( val->Opcode() == Op_RShiftI ) {
2590     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2591     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2592       Node *shl = val->in(1);
2593       if( shl->Opcode() == Op_LShiftI ) {
2594         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2595         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2596           set_req(MemNode::ValueIn, shl->in(1));
2597           return this;
2598         }
2599       }
2600     }
2601   }
2602   return NULL;
2603 }
2604 
2605 //------------------------------value_never_loaded-----------------------------------
2606 // Determine whether there are any possible loads of the value stored.
2607 // For simplicity, we actually check if there are any loads from the
2608 // address stored to, not just for loads of the value stored by this node.
2609 //
2610 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2611   Node *adr = in(Address);
2612   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2613   if (adr_oop == NULL)
2614     return false;
2615   if (!adr_oop->is_known_instance_field())
2616     return false; // if not a distinct instance, there may be aliases of the address
2617   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2618     Node *use = adr->fast_out(i);
2619     int opc = use->Opcode();
2620     if (use->is_Load() || use->is_LoadStore()) {
2621       return false;
2622     }
2623   }
2624   return true;
2625 }
2626 
2627 //=============================================================================
2628 //------------------------------Ideal------------------------------------------
2629 // If the store is from an AND mask that leaves the low bits untouched, then
2630 // we can skip the AND operation.  If the store is from a sign-extension
2631 // (a left shift, then right shift) we can skip both.
2632 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2633   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2634   if( progress != NULL ) return progress;
2635 
2636   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2637   if( progress != NULL ) return progress;
2638 
2639   // Finally check the default case
2640   return StoreNode::Ideal(phase, can_reshape);
2641 }
2642 
2643 //=============================================================================
2644 //------------------------------Ideal------------------------------------------
2645 // If the store is from an AND mask that leaves the low bits untouched, then
2646 // we can skip the AND operation
2647 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2648   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2649   if( progress != NULL ) return progress;
2650 
2651   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2652   if( progress != NULL ) return progress;
2653 
2654   // Finally check the default case
2655   return StoreNode::Ideal(phase, can_reshape);
2656 }
2657 
2658 //=============================================================================
2659 //------------------------------Identity---------------------------------------
2660 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2661   // No need to card mark when storing a null ptr
2662   Node* my_store = in(MemNode::OopStore);
2663   if (my_store->is_Store()) {
2664     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2665     if( t1 == TypePtr::NULL_PTR ) {
2666       return in(MemNode::Memory);
2667     }
2668   }
2669   return this;
2670 }
2671 
2672 //=============================================================================
2673 //------------------------------Ideal---------------------------------------
2674 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2675   Node* progress = StoreNode::Ideal(phase, can_reshape);
2676   if (progress != NULL) return progress;
2677 
2678   Node* my_store = in(MemNode::OopStore);
2679   if (my_store->is_MergeMem()) {
2680     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2681     set_req(MemNode::OopStore, mem);
2682     return this;
2683   }
2684 
2685   return NULL;
2686 }
2687 
2688 //------------------------------Value-----------------------------------------
2689 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2690   // Either input is TOP ==> the result is TOP
2691   const Type *t = phase->type( in(MemNode::Memory) );
2692   if( t == Type::TOP ) return Type::TOP;
2693   t = phase->type( in(MemNode::Address) );
2694   if( t == Type::TOP ) return Type::TOP;
2695   t = phase->type( in(MemNode::ValueIn) );
2696   if( t == Type::TOP ) return Type::TOP;
2697   // If extra input is TOP ==> the result is TOP
2698   t = phase->type( in(MemNode::OopStore) );
2699   if( t == Type::TOP ) return Type::TOP;
2700 
2701   return StoreNode::Value( phase );
2702 }
2703 
2704 
2705 //=============================================================================
2706 //----------------------------------SCMemProjNode------------------------------
2707 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2708 {
2709   return bottom_type();
2710 }
2711 
2712 //=============================================================================
2713 //----------------------------------LoadStoreNode------------------------------
2714 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2715   : Node(required),
2716     _type(rt),
2717     _adr_type(at)
2718 {
2719   init_req(MemNode::Control, c  );
2720   init_req(MemNode::Memory , mem);
2721   init_req(MemNode::Address, adr);
2722   init_req(MemNode::ValueIn, val);
2723   init_class_id(Class_LoadStore);
2724 }
2725 
2726 uint LoadStoreNode::ideal_reg() const {
2727   return _type->ideal_reg();
2728 }
2729 
2730 bool LoadStoreNode::result_not_used() const {
2731   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2732     Node *x = fast_out(i);
2733     if (x->Opcode() == Op_SCMemProj) continue;
2734     return false;
2735   }
2736   return true;
2737 }
2738 
2739 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2740 
2741 //=============================================================================
2742 //----------------------------------LoadStoreConditionalNode--------------------
2743 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2744   init_req(ExpectedIn, ex );
2745 }
2746 
2747 //=============================================================================
2748 //-------------------------------adr_type--------------------------------------
2749 // Do we Match on this edge index or not?  Do not match memory
2750 const TypePtr* ClearArrayNode::adr_type() const {
2751   Node *adr = in(3);
2752   return MemNode::calculate_adr_type(adr->bottom_type());
2753 }
2754 
2755 //------------------------------match_edge-------------------------------------
2756 // Do we Match on this edge index or not?  Do not match memory
2757 uint ClearArrayNode::match_edge(uint idx) const {
2758   return idx > 1;
2759 }
2760 
2761 //------------------------------Identity---------------------------------------
2762 // Clearing a zero length array does nothing
2763 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2764   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
2765 }
2766 
2767 //------------------------------Idealize---------------------------------------
2768 // Clearing a short array is faster with stores
2769 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2770   const int unit = BytesPerLong;
2771   const TypeX* t = phase->type(in(2))->isa_intptr_t();
2772   if (!t)  return NULL;
2773   if (!t->is_con())  return NULL;
2774   intptr_t raw_count = t->get_con();
2775   intptr_t size = raw_count;
2776   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2777   // Clearing nothing uses the Identity call.
2778   // Negative clears are possible on dead ClearArrays
2779   // (see jck test stmt114.stmt11402.val).
2780   if (size <= 0 || size % unit != 0)  return NULL;
2781   intptr_t count = size / unit;
2782   // Length too long; use fast hardware clear
2783   if (size > Matcher::init_array_short_size)  return NULL;
2784   Node *mem = in(1);
2785   if( phase->type(mem)==Type::TOP ) return NULL;
2786   Node *adr = in(3);
2787   const Type* at = phase->type(adr);
2788   if( at==Type::TOP ) return NULL;
2789   const TypePtr* atp = at->isa_ptr();
2790   // adjust atp to be the correct array element address type
2791   if (atp == NULL)  atp = TypePtr::BOTTOM;
2792   else              atp = atp->add_offset(Type::OffsetBot);
2793   // Get base for derived pointer purposes
2794   if( adr->Opcode() != Op_AddP ) Unimplemented();
2795   Node *base = adr->in(1);
2796 
2797   Node *zero = phase->makecon(TypeLong::ZERO);
2798   Node *off  = phase->MakeConX(BytesPerLong);
2799   mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2800   count--;
2801   while( count-- ) {
2802     mem = phase->transform(mem);
2803     adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
2804     mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2805   }
2806   return mem;
2807 }
2808 
2809 //----------------------------step_through----------------------------------
2810 // Return allocation input memory edge if it is different instance
2811 // or itself if it is the one we are looking for.
2812 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2813   Node* n = *np;
2814   assert(n->is_ClearArray(), "sanity");
2815   intptr_t offset;
2816   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2817   // This method is called only before Allocate nodes are expanded during
2818   // macro nodes expansion. Before that ClearArray nodes are only generated
2819   // in LibraryCallKit::generate_arraycopy() which follows allocations.
2820   assert(alloc != NULL, "should have allocation");
2821   if (alloc->_idx == instance_id) {
2822     // Can not bypass initialization of the instance we are looking for.
2823     return false;
2824   }
2825   // Otherwise skip it.
2826   InitializeNode* init = alloc->initialization();
2827   if (init != NULL)
2828     *np = init->in(TypeFunc::Memory);
2829   else
2830     *np = alloc->in(TypeFunc::Memory);
2831   return true;
2832 }
2833 
2834 //----------------------------clear_memory-------------------------------------
2835 // Generate code to initialize object storage to zero.
2836 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2837                                    intptr_t start_offset,
2838                                    Node* end_offset,
2839                                    PhaseGVN* phase) {
2840   Compile* C = phase->C;
2841   intptr_t offset = start_offset;
2842 
2843   int unit = BytesPerLong;
2844   if ((offset % unit) != 0) {
2845     Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
2846     adr = phase->transform(adr);
2847     const TypePtr* atp = TypeRawPtr::BOTTOM;
2848     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2849     mem = phase->transform(mem);
2850     offset += BytesPerInt;
2851   }
2852   assert((offset % unit) == 0, "");
2853 
2854   // Initialize the remaining stuff, if any, with a ClearArray.
2855   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2856 }
2857 
2858 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2859                                    Node* start_offset,
2860                                    Node* end_offset,
2861                                    PhaseGVN* phase) {
2862   if (start_offset == end_offset) {
2863     // nothing to do
2864     return mem;
2865   }
2866 
2867   Compile* C = phase->C;
2868   int unit = BytesPerLong;
2869   Node* zbase = start_offset;
2870   Node* zend  = end_offset;
2871 
2872   // Scale to the unit required by the CPU:
2873   if (!Matcher::init_array_count_is_in_bytes) {
2874     Node* shift = phase->intcon(exact_log2(unit));
2875     zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
2876     zend  = phase->transform( new(C) URShiftXNode(zend,  shift) );
2877   }
2878 
2879   // Bulk clear double-words
2880   Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
2881   Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
2882   mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
2883   return phase->transform(mem);
2884 }
2885 
2886 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2887                                    intptr_t start_offset,
2888                                    intptr_t end_offset,
2889                                    PhaseGVN* phase) {
2890   if (start_offset == end_offset) {
2891     // nothing to do
2892     return mem;
2893   }
2894 
2895   Compile* C = phase->C;
2896   assert((end_offset % BytesPerInt) == 0, "odd end offset");
2897   intptr_t done_offset = end_offset;
2898   if ((done_offset % BytesPerLong) != 0) {
2899     done_offset -= BytesPerInt;
2900   }
2901   if (done_offset > start_offset) {
2902     mem = clear_memory(ctl, mem, dest,
2903                        start_offset, phase->MakeConX(done_offset), phase);
2904   }
2905   if (done_offset < end_offset) { // emit the final 32-bit store
2906     Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
2907     adr = phase->transform(adr);
2908     const TypePtr* atp = TypeRawPtr::BOTTOM;
2909     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2910     mem = phase->transform(mem);
2911     done_offset += BytesPerInt;
2912   }
2913   assert(done_offset == end_offset, "");
2914   return mem;
2915 }
2916 
2917 //=============================================================================
2918 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2919   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2920     _adr_type(C->get_adr_type(alias_idx))
2921 {
2922   init_class_id(Class_MemBar);
2923   Node* top = C->top();
2924   init_req(TypeFunc::I_O,top);
2925   init_req(TypeFunc::FramePtr,top);
2926   init_req(TypeFunc::ReturnAdr,top);
2927   if (precedent != NULL)
2928     init_req(TypeFunc::Parms, precedent);
2929 }
2930 
2931 //------------------------------cmp--------------------------------------------
2932 uint MemBarNode::hash() const { return NO_HASH; }
2933 uint MemBarNode::cmp( const Node &n ) const {
2934   return (&n == this);          // Always fail except on self
2935 }
2936 
2937 //------------------------------make-------------------------------------------
2938 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2939   switch (opcode) {
2940   case Op_MemBarAcquire:     return new(C) MemBarAcquireNode(C, atp, pn);
2941   case Op_LoadFence:         return new(C) LoadFenceNode(C, atp, pn);
2942   case Op_MemBarRelease:     return new(C) MemBarReleaseNode(C, atp, pn);
2943   case Op_StoreFence:        return new(C) StoreFenceNode(C, atp, pn);
2944   case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
2945   case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
2946   case Op_MemBarVolatile:    return new(C) MemBarVolatileNode(C, atp, pn);
2947   case Op_MemBarCPUOrder:    return new(C) MemBarCPUOrderNode(C, atp, pn);
2948   case Op_Initialize:        return new(C) InitializeNode(C, atp, pn);
2949   case Op_MemBarStoreStore:  return new(C) MemBarStoreStoreNode(C, atp, pn);
2950   default: ShouldNotReachHere(); return NULL;
2951   }
2952 }
2953 
2954 //------------------------------Ideal------------------------------------------
2955 // Return a node which is more "ideal" than the current node.  Strip out
2956 // control copies
2957 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2958   if (remove_dead_region(phase, can_reshape)) return this;
2959   // Don't bother trying to transform a dead node
2960   if (in(0) && in(0)->is_top()) {
2961     return NULL;
2962   }
2963 
2964   // Eliminate volatile MemBars for scalar replaced objects.
2965   if (can_reshape && req() == (Precedent+1)) {
2966     bool eliminate = false;
2967     int opc = Opcode();
2968     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2969       // Volatile field loads and stores.
2970       Node* my_mem = in(MemBarNode::Precedent);
2971       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2972       if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2973         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2974         // replace this Precedent (decodeN) with the Load instead.
2975         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
2976           Node* load_node = my_mem->in(1);
2977           set_req(MemBarNode::Precedent, load_node);
2978           phase->is_IterGVN()->_worklist.push(my_mem);
2979           my_mem = load_node;
2980         } else {
2981           assert(my_mem->unique_out() == this, "sanity");
2982           del_req(Precedent);
2983           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2984           my_mem = NULL;
2985         }
2986       }
2987       if (my_mem != NULL && my_mem->is_Mem()) {
2988         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2989         // Check for scalar replaced object reference.
2990         if( t_oop != NULL && t_oop->is_known_instance_field() &&
2991             t_oop->offset() != Type::OffsetBot &&
2992             t_oop->offset() != Type::OffsetTop) {
2993           eliminate = true;
2994         }
2995       }
2996     } else if (opc == Op_MemBarRelease) {
2997       // Final field stores.
2998       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2999       if ((alloc != NULL) && alloc->is_Allocate() &&
3000           alloc->as_Allocate()->_is_non_escaping) {
3001         // The allocated object does not escape.
3002         eliminate = true;
3003       }
3004     }
3005     if (eliminate) {
3006       // Replace MemBar projections by its inputs.
3007       PhaseIterGVN* igvn = phase->is_IterGVN();
3008       igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3009       igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3010       // Must return either the original node (now dead) or a new node
3011       // (Do not return a top here, since that would break the uniqueness of top.)
3012       return new (phase->C) ConINode(TypeInt::ZERO);
3013     }
3014   }
3015   return NULL;
3016 }
3017 
3018 //------------------------------Value------------------------------------------
3019 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
3020   if( !in(0) ) return Type::TOP;
3021   if( phase->type(in(0)) == Type::TOP )
3022     return Type::TOP;
3023   return TypeTuple::MEMBAR;
3024 }
3025 
3026 //------------------------------match------------------------------------------
3027 // Construct projections for memory.
3028 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3029   switch (proj->_con) {
3030   case TypeFunc::Control:
3031   case TypeFunc::Memory:
3032     return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3033   }
3034   ShouldNotReachHere();
3035   return NULL;
3036 }
3037 
3038 //===========================InitializeNode====================================
3039 // SUMMARY:
3040 // This node acts as a memory barrier on raw memory, after some raw stores.
3041 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3042 // The Initialize can 'capture' suitably constrained stores as raw inits.
3043 // It can coalesce related raw stores into larger units (called 'tiles').
3044 // It can avoid zeroing new storage for memory units which have raw inits.
3045 // At macro-expansion, it is marked 'complete', and does not optimize further.
3046 //
3047 // EXAMPLE:
3048 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3049 //   ctl = incoming control; mem* = incoming memory
3050 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
3051 // First allocate uninitialized memory and fill in the header:
3052 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
3053 //   ctl := alloc.Control; mem* := alloc.Memory*
3054 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
3055 // Then initialize to zero the non-header parts of the raw memory block:
3056 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3057 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3058 // After the initialize node executes, the object is ready for service:
3059 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3060 // Suppose its body is immediately initialized as {1,2}:
3061 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3062 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3063 //   mem.SLICE(#short[*]) := store2
3064 //
3065 // DETAILS:
3066 // An InitializeNode collects and isolates object initialization after
3067 // an AllocateNode and before the next possible safepoint.  As a
3068 // memory barrier (MemBarNode), it keeps critical stores from drifting
3069 // down past any safepoint or any publication of the allocation.
3070 // Before this barrier, a newly-allocated object may have uninitialized bits.
3071 // After this barrier, it may be treated as a real oop, and GC is allowed.
3072 //
3073 // The semantics of the InitializeNode include an implicit zeroing of
3074 // the new object from object header to the end of the object.
3075 // (The object header and end are determined by the AllocateNode.)
3076 //
3077 // Certain stores may be added as direct inputs to the InitializeNode.
3078 // These stores must update raw memory, and they must be to addresses
3079 // derived from the raw address produced by AllocateNode, and with
3080 // a constant offset.  They must be ordered by increasing offset.
3081 // The first one is at in(RawStores), the last at in(req()-1).
3082 // Unlike most memory operations, they are not linked in a chain,
3083 // but are displayed in parallel as users of the rawmem output of
3084 // the allocation.
3085 //
3086 // (See comments in InitializeNode::capture_store, which continue
3087 // the example given above.)
3088 //
3089 // When the associated Allocate is macro-expanded, the InitializeNode
3090 // may be rewritten to optimize collected stores.  A ClearArrayNode
3091 // may also be created at that point to represent any required zeroing.
3092 // The InitializeNode is then marked 'complete', prohibiting further
3093 // capturing of nearby memory operations.
3094 //
3095 // During macro-expansion, all captured initializations which store
3096 // constant values of 32 bits or smaller are coalesced (if advantageous)
3097 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
3098 // initialized in fewer memory operations.  Memory words which are
3099 // covered by neither tiles nor non-constant stores are pre-zeroed
3100 // by explicit stores of zero.  (The code shape happens to do all
3101 // zeroing first, then all other stores, with both sequences occurring
3102 // in order of ascending offsets.)
3103 //
3104 // Alternatively, code may be inserted between an AllocateNode and its
3105 // InitializeNode, to perform arbitrary initialization of the new object.
3106 // E.g., the object copying intrinsics insert complex data transfers here.
3107 // The initialization must then be marked as 'complete' disable the
3108 // built-in zeroing semantics and the collection of initializing stores.
3109 //
3110 // While an InitializeNode is incomplete, reads from the memory state
3111 // produced by it are optimizable if they match the control edge and
3112 // new oop address associated with the allocation/initialization.
3113 // They return a stored value (if the offset matches) or else zero.
3114 // A write to the memory state, if it matches control and address,
3115 // and if it is to a constant offset, may be 'captured' by the
3116 // InitializeNode.  It is cloned as a raw memory operation and rewired
3117 // inside the initialization, to the raw oop produced by the allocation.
3118 // Operations on addresses which are provably distinct (e.g., to
3119 // other AllocateNodes) are allowed to bypass the initialization.
3120 //
3121 // The effect of all this is to consolidate object initialization
3122 // (both arrays and non-arrays, both piecewise and bulk) into a
3123 // single location, where it can be optimized as a unit.
3124 //
3125 // Only stores with an offset less than TrackedInitializationLimit words
3126 // will be considered for capture by an InitializeNode.  This puts a
3127 // reasonable limit on the complexity of optimized initializations.
3128 
3129 //---------------------------InitializeNode------------------------------------
3130 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3131   : _is_complete(Incomplete), _does_not_escape(false),
3132     MemBarNode(C, adr_type, rawoop)
3133 {
3134   init_class_id(Class_Initialize);
3135 
3136   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3137   assert(in(RawAddress) == rawoop, "proper init");
3138   // Note:  allocation() can be NULL, for secondary initialization barriers
3139 }
3140 
3141 // Since this node is not matched, it will be processed by the
3142 // register allocator.  Declare that there are no constraints
3143 // on the allocation of the RawAddress edge.
3144 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3145   // This edge should be set to top, by the set_complete.  But be conservative.
3146   if (idx == InitializeNode::RawAddress)
3147     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3148   return RegMask::Empty;
3149 }
3150 
3151 Node* InitializeNode::memory(uint alias_idx) {
3152   Node* mem = in(Memory);
3153   if (mem->is_MergeMem()) {
3154     return mem->as_MergeMem()->memory_at(alias_idx);
3155   } else {
3156     // incoming raw memory is not split
3157     return mem;
3158   }
3159 }
3160 
3161 bool InitializeNode::is_non_zero() {
3162   if (is_complete())  return false;
3163   remove_extra_zeroes();
3164   return (req() > RawStores);
3165 }
3166 
3167 void InitializeNode::set_complete(PhaseGVN* phase) {
3168   assert(!is_complete(), "caller responsibility");
3169   _is_complete = Complete;
3170 
3171   // After this node is complete, it contains a bunch of
3172   // raw-memory initializations.  There is no need for
3173   // it to have anything to do with non-raw memory effects.
3174   // Therefore, tell all non-raw users to re-optimize themselves,
3175   // after skipping the memory effects of this initialization.
3176   PhaseIterGVN* igvn = phase->is_IterGVN();
3177   if (igvn)  igvn->add_users_to_worklist(this);
3178 }
3179 
3180 // convenience function
3181 // return false if the init contains any stores already
3182 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3183   InitializeNode* init = initialization();
3184   if (init == NULL || init->is_complete())  return false;
3185   init->remove_extra_zeroes();
3186   // for now, if this allocation has already collected any inits, bail:
3187   if (init->is_non_zero())  return false;
3188   init->set_complete(phase);
3189   return true;
3190 }
3191 
3192 void InitializeNode::remove_extra_zeroes() {
3193   if (req() == RawStores)  return;
3194   Node* zmem = zero_memory();
3195   uint fill = RawStores;
3196   for (uint i = fill; i < req(); i++) {
3197     Node* n = in(i);
3198     if (n->is_top() || n == zmem)  continue;  // skip
3199     if (fill < i)  set_req(fill, n);          // compact
3200     ++fill;
3201   }
3202   // delete any empty spaces created:
3203   while (fill < req()) {
3204     del_req(fill);
3205   }
3206 }
3207 
3208 // Helper for remembering which stores go with which offsets.
3209 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3210   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
3211   intptr_t offset = -1;
3212   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3213                                                phase, offset);
3214   if (base == NULL)     return -1;  // something is dead,
3215   if (offset < 0)       return -1;  //        dead, dead
3216   return offset;
3217 }
3218 
3219 // Helper for proving that an initialization expression is
3220 // "simple enough" to be folded into an object initialization.
3221 // Attempts to prove that a store's initial value 'n' can be captured
3222 // within the initialization without creating a vicious cycle, such as:
3223 //     { Foo p = new Foo(); p.next = p; }
3224 // True for constants and parameters and small combinations thereof.
3225 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3226   if (n == NULL)      return true;   // (can this really happen?)
3227   if (n->is_Proj())   n = n->in(0);
3228   if (n == this)      return false;  // found a cycle
3229   if (n->is_Con())    return true;
3230   if (n->is_Start())  return true;   // params, etc., are OK
3231   if (n->is_Root())   return true;   // even better
3232 
3233   Node* ctl = n->in(0);
3234   if (ctl != NULL && !ctl->is_top()) {
3235     if (ctl->is_Proj())  ctl = ctl->in(0);
3236     if (ctl == this)  return false;
3237 
3238     // If we already know that the enclosing memory op is pinned right after
3239     // the init, then any control flow that the store has picked up
3240     // must have preceded the init, or else be equal to the init.
3241     // Even after loop optimizations (which might change control edges)
3242     // a store is never pinned *before* the availability of its inputs.
3243     if (!MemNode::all_controls_dominate(n, this))
3244       return false;                  // failed to prove a good control
3245   }
3246 
3247   // Check data edges for possible dependencies on 'this'.
3248   if ((count += 1) > 20)  return false;  // complexity limit
3249   for (uint i = 1; i < n->req(); i++) {
3250     Node* m = n->in(i);
3251     if (m == NULL || m == n || m->is_top())  continue;
3252     uint first_i = n->find_edge(m);
3253     if (i != first_i)  continue;  // process duplicate edge just once
3254     if (!detect_init_independence(m, count)) {
3255       return false;
3256     }
3257   }
3258 
3259   return true;
3260 }
3261 
3262 // Here are all the checks a Store must pass before it can be moved into
3263 // an initialization.  Returns zero if a check fails.
3264 // On success, returns the (constant) offset to which the store applies,
3265 // within the initialized memory.
3266 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3267   const int FAIL = 0;
3268   if (st->req() != MemNode::ValueIn + 1)
3269     return FAIL;                // an inscrutable StoreNode (card mark?)
3270   Node* ctl = st->in(MemNode::Control);
3271   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3272     return FAIL;                // must be unconditional after the initialization
3273   Node* mem = st->in(MemNode::Memory);
3274   if (!(mem->is_Proj() && mem->in(0) == this))
3275     return FAIL;                // must not be preceded by other stores
3276   Node* adr = st->in(MemNode::Address);
3277   intptr_t offset;
3278   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3279   if (alloc == NULL)
3280     return FAIL;                // inscrutable address
3281   if (alloc != allocation())
3282     return FAIL;                // wrong allocation!  (store needs to float up)
3283   Node* val = st->in(MemNode::ValueIn);
3284   int complexity_count = 0;
3285   if (!detect_init_independence(val, complexity_count))
3286     return FAIL;                // stored value must be 'simple enough'
3287 
3288   // The Store can be captured only if nothing after the allocation
3289   // and before the Store is using the memory location that the store
3290   // overwrites.
3291   bool failed = false;
3292   // If is_complete_with_arraycopy() is true the shape of the graph is
3293   // well defined and is safe so no need for extra checks.
3294   if (!is_complete_with_arraycopy()) {
3295     // We are going to look at each use of the memory state following
3296     // the allocation to make sure nothing reads the memory that the
3297     // Store writes.
3298     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3299     int alias_idx = phase->C->get_alias_index(t_adr);
3300     ResourceMark rm;
3301     Unique_Node_List mems;
3302     mems.push(mem);
3303     Node* unique_merge = NULL;
3304     for (uint next = 0; next < mems.size(); ++next) {
3305       Node *m  = mems.at(next);
3306       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3307         Node *n = m->fast_out(j);
3308         if (n->outcnt() == 0) {
3309           continue;
3310         }
3311         if (n == st) {
3312           continue;
3313         } else if (n->in(0) != NULL && n->in(0) != ctl) {
3314           // If the control of this use is different from the control
3315           // of the Store which is right after the InitializeNode then
3316           // this node cannot be between the InitializeNode and the
3317           // Store.
3318           continue;
3319         } else if (n->is_MergeMem()) {
3320           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3321             // We can hit a MergeMemNode (that will likely go away
3322             // later) that is a direct use of the memory state
3323             // following the InitializeNode on the same slice as the
3324             // store node that we'd like to capture. We need to check
3325             // the uses of the MergeMemNode.
3326             mems.push(n);
3327           }
3328         } else if (n->is_Mem()) {
3329           Node* other_adr = n->in(MemNode::Address);
3330           if (other_adr == adr) {
3331             failed = true;
3332             break;
3333           } else {
3334             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3335             if (other_t_adr != NULL) {
3336               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3337               if (other_alias_idx == alias_idx) {
3338                 // A load from the same memory slice as the store right
3339                 // after the InitializeNode. We check the control of the
3340                 // object/array that is loaded from. If it's the same as
3341                 // the store control then we cannot capture the store.
3342                 assert(!n->is_Store(), "2 stores to same slice on same control?");
3343                 Node* base = other_adr;
3344                 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3345                 base = base->in(AddPNode::Base);
3346                 if (base != NULL) {
3347                   base = base->uncast();
3348                   if (base->is_Proj() && base->in(0) == alloc) {
3349                     failed = true;
3350                     break;
3351                   }
3352                 }
3353               }
3354             }
3355           }
3356         } else {
3357           failed = true;
3358           break;
3359         }
3360       }
3361     }
3362   }
3363   if (failed) {
3364     if (!can_reshape) {
3365       // We decided we couldn't capture the store during parsing. We
3366       // should try again during the next IGVN once the graph is
3367       // cleaner.
3368       phase->C->record_for_igvn(st);
3369     }
3370     return FAIL;
3371   }
3372 
3373   return offset;                // success
3374 }
3375 
3376 // Find the captured store in(i) which corresponds to the range
3377 // [start..start+size) in the initialized object.
3378 // If there is one, return its index i.  If there isn't, return the
3379 // negative of the index where it should be inserted.
3380 // Return 0 if the queried range overlaps an initialization boundary
3381 // or if dead code is encountered.
3382 // If size_in_bytes is zero, do not bother with overlap checks.
3383 int InitializeNode::captured_store_insertion_point(intptr_t start,
3384                                                    int size_in_bytes,
3385                                                    PhaseTransform* phase) {
3386   const int FAIL = 0, MAX_STORE = BytesPerLong;
3387 
3388   if (is_complete())
3389     return FAIL;                // arraycopy got here first; punt
3390 
3391   assert(allocation() != NULL, "must be present");
3392 
3393   // no negatives, no header fields:
3394   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
3395 
3396   // after a certain size, we bail out on tracking all the stores:
3397   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3398   if (start >= ti_limit)  return FAIL;
3399 
3400   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3401     if (i >= limit)  return -(int)i; // not found; here is where to put it
3402 
3403     Node*    st     = in(i);
3404     intptr_t st_off = get_store_offset(st, phase);
3405     if (st_off < 0) {
3406       if (st != zero_memory()) {
3407         return FAIL;            // bail out if there is dead garbage
3408       }
3409     } else if (st_off > start) {
3410       // ...we are done, since stores are ordered
3411       if (st_off < start + size_in_bytes) {
3412         return FAIL;            // the next store overlaps
3413       }
3414       return -(int)i;           // not found; here is where to put it
3415     } else if (st_off < start) {
3416       if (size_in_bytes != 0 &&
3417           start < st_off + MAX_STORE &&
3418           start < st_off + st->as_Store()->memory_size()) {
3419         return FAIL;            // the previous store overlaps
3420       }
3421     } else {
3422       if (size_in_bytes != 0 &&
3423           st->as_Store()->memory_size() != size_in_bytes) {
3424         return FAIL;            // mismatched store size
3425       }
3426       return i;
3427     }
3428 
3429     ++i;
3430   }
3431 }
3432 
3433 // Look for a captured store which initializes at the offset 'start'
3434 // with the given size.  If there is no such store, and no other
3435 // initialization interferes, then return zero_memory (the memory
3436 // projection of the AllocateNode).
3437 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3438                                           PhaseTransform* phase) {
3439   assert(stores_are_sane(phase), "");
3440   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3441   if (i == 0) {
3442     return NULL;                // something is dead
3443   } else if (i < 0) {
3444     return zero_memory();       // just primordial zero bits here
3445   } else {
3446     Node* st = in(i);           // here is the store at this position
3447     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3448     return st;
3449   }
3450 }
3451 
3452 // Create, as a raw pointer, an address within my new object at 'offset'.
3453 Node* InitializeNode::make_raw_address(intptr_t offset,
3454                                        PhaseTransform* phase) {
3455   Node* addr = in(RawAddress);
3456   if (offset != 0) {
3457     Compile* C = phase->C;
3458     addr = phase->transform( new (C) AddPNode(C->top(), addr,
3459                                                  phase->MakeConX(offset)) );
3460   }
3461   return addr;
3462 }
3463 
3464 // Clone the given store, converting it into a raw store
3465 // initializing a field or element of my new object.
3466 // Caller is responsible for retiring the original store,
3467 // with subsume_node or the like.
3468 //
3469 // From the example above InitializeNode::InitializeNode,
3470 // here are the old stores to be captured:
3471 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3472 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3473 //
3474 // Here is the changed code; note the extra edges on init:
3475 //   alloc = (Allocate ...)
3476 //   rawoop = alloc.RawAddress
3477 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3478 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3479 //   init = (Initialize alloc.Control alloc.Memory rawoop
3480 //                      rawstore1 rawstore2)
3481 //
3482 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3483                                     PhaseTransform* phase, bool can_reshape) {
3484   assert(stores_are_sane(phase), "");
3485 
3486   if (start < 0)  return NULL;
3487   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3488 
3489   Compile* C = phase->C;
3490   int size_in_bytes = st->memory_size();
3491   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3492   if (i == 0)  return NULL;     // bail out
3493   Node* prev_mem = NULL;        // raw memory for the captured store
3494   if (i > 0) {
3495     prev_mem = in(i);           // there is a pre-existing store under this one
3496     set_req(i, C->top());       // temporarily disconnect it
3497     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3498   } else {
3499     i = -i;                     // no pre-existing store
3500     prev_mem = zero_memory();   // a slice of the newly allocated object
3501     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3502       set_req(--i, C->top());   // reuse this edge; it has been folded away
3503     else
3504       ins_req(i, C->top());     // build a new edge
3505   }
3506   Node* new_st = st->clone();
3507   new_st->set_req(MemNode::Control, in(Control));
3508   new_st->set_req(MemNode::Memory,  prev_mem);
3509   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3510   new_st = phase->transform(new_st);
3511 
3512   // At this point, new_st might have swallowed a pre-existing store
3513   // at the same offset, or perhaps new_st might have disappeared,
3514   // if it redundantly stored the same value (or zero to fresh memory).
3515 
3516   // In any case, wire it in:
3517   set_req(i, new_st);
3518 
3519   // The caller may now kill the old guy.
3520   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3521   assert(check_st == new_st || check_st == NULL, "must be findable");
3522   assert(!is_complete(), "");
3523   return new_st;
3524 }
3525 
3526 static bool store_constant(jlong* tiles, int num_tiles,
3527                            intptr_t st_off, int st_size,
3528                            jlong con) {
3529   if ((st_off & (st_size-1)) != 0)
3530     return false;               // strange store offset (assume size==2**N)
3531   address addr = (address)tiles + st_off;
3532   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3533   switch (st_size) {
3534   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
3535   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
3536   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
3537   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
3538   default: return false;        // strange store size (detect size!=2**N here)
3539   }
3540   return true;                  // return success to caller
3541 }
3542 
3543 // Coalesce subword constants into int constants and possibly
3544 // into long constants.  The goal, if the CPU permits,
3545 // is to initialize the object with a small number of 64-bit tiles.
3546 // Also, convert floating-point constants to bit patterns.
3547 // Non-constants are not relevant to this pass.
3548 //
3549 // In terms of the running example on InitializeNode::InitializeNode
3550 // and InitializeNode::capture_store, here is the transformation
3551 // of rawstore1 and rawstore2 into rawstore12:
3552 //   alloc = (Allocate ...)
3553 //   rawoop = alloc.RawAddress
3554 //   tile12 = 0x00010002
3555 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3556 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3557 //
3558 void
3559 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3560                                         Node* size_in_bytes,
3561                                         PhaseGVN* phase) {
3562   Compile* C = phase->C;
3563 
3564   assert(stores_are_sane(phase), "");
3565   // Note:  After this pass, they are not completely sane,
3566   // since there may be some overlaps.
3567 
3568   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3569 
3570   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3571   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3572   size_limit = MIN2(size_limit, ti_limit);
3573   size_limit = align_size_up(size_limit, BytesPerLong);
3574   int num_tiles = size_limit / BytesPerLong;
3575 
3576   // allocate space for the tile map:
3577   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3578   jlong  tiles_buf[small_len];
3579   Node*  nodes_buf[small_len];
3580   jlong  inits_buf[small_len];
3581   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3582                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3583   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3584                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3585   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3586                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3587   // tiles: exact bitwise model of all primitive constants
3588   // nodes: last constant-storing node subsumed into the tiles model
3589   // inits: which bytes (in each tile) are touched by any initializations
3590 
3591   //// Pass A: Fill in the tile model with any relevant stores.
3592 
3593   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3594   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3595   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3596   Node* zmem = zero_memory(); // initially zero memory state
3597   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3598     Node* st = in(i);
3599     intptr_t st_off = get_store_offset(st, phase);
3600 
3601     // Figure out the store's offset and constant value:
3602     if (st_off < header_size)             continue; //skip (ignore header)
3603     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
3604     int st_size = st->as_Store()->memory_size();
3605     if (st_off + st_size > size_limit)    break;
3606 
3607     // Record which bytes are touched, whether by constant or not.
3608     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3609       continue;                 // skip (strange store size)
3610 
3611     const Type* val = phase->type(st->in(MemNode::ValueIn));
3612     if (!val->singleton())                continue; //skip (non-con store)
3613     BasicType type = val->basic_type();
3614 
3615     jlong con = 0;
3616     switch (type) {
3617     case T_INT:    con = val->is_int()->get_con();  break;
3618     case T_LONG:   con = val->is_long()->get_con(); break;
3619     case T_FLOAT:  con = jint_cast(val->getf());    break;
3620     case T_DOUBLE: con = jlong_cast(val->getd());   break;
3621     default:                              continue; //skip (odd store type)
3622     }
3623 
3624     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3625         st->Opcode() == Op_StoreL) {
3626       continue;                 // This StoreL is already optimal.
3627     }
3628 
3629     // Store down the constant.
3630     store_constant(tiles, num_tiles, st_off, st_size, con);
3631 
3632     intptr_t j = st_off >> LogBytesPerLong;
3633 
3634     if (type == T_INT && st_size == BytesPerInt
3635         && (st_off & BytesPerInt) == BytesPerInt) {
3636       jlong lcon = tiles[j];
3637       if (!Matcher::isSimpleConstant64(lcon) &&
3638           st->Opcode() == Op_StoreI) {
3639         // This StoreI is already optimal by itself.
3640         jint* intcon = (jint*) &tiles[j];
3641         intcon[1] = 0;  // undo the store_constant()
3642 
3643         // If the previous store is also optimal by itself, back up and
3644         // undo the action of the previous loop iteration... if we can.
3645         // But if we can't, just let the previous half take care of itself.
3646         st = nodes[j];
3647         st_off -= BytesPerInt;
3648         con = intcon[0];
3649         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3650           assert(st_off >= header_size, "still ignoring header");
3651           assert(get_store_offset(st, phase) == st_off, "must be");
3652           assert(in(i-1) == zmem, "must be");
3653           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3654           assert(con == tcon->is_int()->get_con(), "must be");
3655           // Undo the effects of the previous loop trip, which swallowed st:
3656           intcon[0] = 0;        // undo store_constant()
3657           set_req(i-1, st);     // undo set_req(i, zmem)
3658           nodes[j] = NULL;      // undo nodes[j] = st
3659           --old_subword;        // undo ++old_subword
3660         }
3661         continue;               // This StoreI is already optimal.
3662       }
3663     }
3664 
3665     // This store is not needed.
3666     set_req(i, zmem);
3667     nodes[j] = st;              // record for the moment
3668     if (st_size < BytesPerLong) // something has changed
3669           ++old_subword;        // includes int/float, but who's counting...
3670     else  ++old_long;
3671   }
3672 
3673   if ((old_subword + old_long) == 0)
3674     return;                     // nothing more to do
3675 
3676   //// Pass B: Convert any non-zero tiles into optimal constant stores.
3677   // Be sure to insert them before overlapping non-constant stores.
3678   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
3679   for (int j = 0; j < num_tiles; j++) {
3680     jlong con  = tiles[j];
3681     jlong init = inits[j];
3682     if (con == 0)  continue;
3683     jint con0,  con1;           // split the constant, address-wise
3684     jint init0, init1;          // split the init map, address-wise
3685     { union { jlong con; jint intcon[2]; } u;
3686       u.con = con;
3687       con0  = u.intcon[0];
3688       con1  = u.intcon[1];
3689       u.con = init;
3690       init0 = u.intcon[0];
3691       init1 = u.intcon[1];
3692     }
3693 
3694     Node* old = nodes[j];
3695     assert(old != NULL, "need the prior store");
3696     intptr_t offset = (j * BytesPerLong);
3697 
3698     bool split = !Matcher::isSimpleConstant64(con);
3699 
3700     if (offset < header_size) {
3701       assert(offset + BytesPerInt >= header_size, "second int counts");
3702       assert(*(jint*)&tiles[j] == 0, "junk in header");
3703       split = true;             // only the second word counts
3704       // Example:  int a[] = { 42 ... }
3705     } else if (con0 == 0 && init0 == -1) {
3706       split = true;             // first word is covered by full inits
3707       // Example:  int a[] = { ... foo(), 42 ... }
3708     } else if (con1 == 0 && init1 == -1) {
3709       split = true;             // second word is covered by full inits
3710       // Example:  int a[] = { ... 42, foo() ... }
3711     }
3712 
3713     // Here's a case where init0 is neither 0 nor -1:
3714     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
3715     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3716     // In this case the tile is not split; it is (jlong)42.
3717     // The big tile is stored down, and then the foo() value is inserted.
3718     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3719 
3720     Node* ctl = old->in(MemNode::Control);
3721     Node* adr = make_raw_address(offset, phase);
3722     const TypePtr* atp = TypeRawPtr::BOTTOM;
3723 
3724     // One or two coalesced stores to plop down.
3725     Node*    st[2];
3726     intptr_t off[2];
3727     int  nst = 0;
3728     if (!split) {
3729       ++new_long;
3730       off[nst] = offset;
3731       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3732                                   phase->longcon(con), T_LONG, MemNode::unordered);
3733     } else {
3734       // Omit either if it is a zero.
3735       if (con0 != 0) {
3736         ++new_int;
3737         off[nst]  = offset;
3738         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3739                                     phase->intcon(con0), T_INT, MemNode::unordered);
3740       }
3741       if (con1 != 0) {
3742         ++new_int;
3743         offset += BytesPerInt;
3744         adr = make_raw_address(offset, phase);
3745         off[nst]  = offset;
3746         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3747                                     phase->intcon(con1), T_INT, MemNode::unordered);
3748       }
3749     }
3750 
3751     // Insert second store first, then the first before the second.
3752     // Insert each one just before any overlapping non-constant stores.
3753     while (nst > 0) {
3754       Node* st1 = st[--nst];
3755       C->copy_node_notes_to(st1, old);
3756       st1 = phase->transform(st1);
3757       offset = off[nst];
3758       assert(offset >= header_size, "do not smash header");
3759       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3760       guarantee(ins_idx != 0, "must re-insert constant store");
3761       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
3762       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3763         set_req(--ins_idx, st1);
3764       else
3765         ins_req(ins_idx, st1);
3766     }
3767   }
3768 
3769   if (PrintCompilation && WizardMode)
3770     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3771                   old_subword, old_long, new_int, new_long);
3772   if (C->log() != NULL)
3773     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3774                    old_subword, old_long, new_int, new_long);
3775 
3776   // Clean up any remaining occurrences of zmem:
3777   remove_extra_zeroes();
3778 }
3779 
3780 // Explore forward from in(start) to find the first fully initialized
3781 // word, and return its offset.  Skip groups of subword stores which
3782 // together initialize full words.  If in(start) is itself part of a
3783 // fully initialized word, return the offset of in(start).  If there
3784 // are no following full-word stores, or if something is fishy, return
3785 // a negative value.
3786 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3787   int       int_map = 0;
3788   intptr_t  int_map_off = 0;
3789   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
3790 
3791   for (uint i = start, limit = req(); i < limit; i++) {
3792     Node* st = in(i);
3793 
3794     intptr_t st_off = get_store_offset(st, phase);
3795     if (st_off < 0)  break;  // return conservative answer
3796 
3797     int st_size = st->as_Store()->memory_size();
3798     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3799       return st_off;            // we found a complete word init
3800     }
3801 
3802     // update the map:
3803 
3804     intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3805     if (this_int_off != int_map_off) {
3806       // reset the map:
3807       int_map = 0;
3808       int_map_off = this_int_off;
3809     }
3810 
3811     int subword_off = st_off - this_int_off;
3812     int_map |= right_n_bits(st_size) << subword_off;
3813     if ((int_map & FULL_MAP) == FULL_MAP) {
3814       return this_int_off;      // we found a complete word init
3815     }
3816 
3817     // Did this store hit or cross the word boundary?
3818     intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3819     if (next_int_off == this_int_off + BytesPerInt) {
3820       // We passed the current int, without fully initializing it.
3821       int_map_off = next_int_off;
3822       int_map >>= BytesPerInt;
3823     } else if (next_int_off > this_int_off + BytesPerInt) {
3824       // We passed the current and next int.
3825       return this_int_off + BytesPerInt;
3826     }
3827   }
3828 
3829   return -1;
3830 }
3831 
3832 
3833 // Called when the associated AllocateNode is expanded into CFG.
3834 // At this point, we may perform additional optimizations.
3835 // Linearize the stores by ascending offset, to make memory
3836 // activity as coherent as possible.
3837 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3838                                       intptr_t header_size,
3839                                       Node* size_in_bytes,
3840                                       PhaseGVN* phase) {
3841   assert(!is_complete(), "not already complete");
3842   assert(stores_are_sane(phase), "");
3843   assert(allocation() != NULL, "must be present");
3844 
3845   remove_extra_zeroes();
3846 
3847   if (ReduceFieldZeroing || ReduceBulkZeroing)
3848     // reduce instruction count for common initialization patterns
3849     coalesce_subword_stores(header_size, size_in_bytes, phase);
3850 
3851   Node* zmem = zero_memory();   // initially zero memory state
3852   Node* inits = zmem;           // accumulating a linearized chain of inits
3853   #ifdef ASSERT
3854   intptr_t first_offset = allocation()->minimum_header_size();
3855   intptr_t last_init_off = first_offset;  // previous init offset
3856   intptr_t last_init_end = first_offset;  // previous init offset+size
3857   intptr_t last_tile_end = first_offset;  // previous tile offset+size
3858   #endif
3859   intptr_t zeroes_done = header_size;
3860 
3861   bool do_zeroing = true;       // we might give up if inits are very sparse
3862   int  big_init_gaps = 0;       // how many large gaps have we seen?
3863 
3864   if (ZeroTLAB)  do_zeroing = false;
3865   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
3866 
3867   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3868     Node* st = in(i);
3869     intptr_t st_off = get_store_offset(st, phase);
3870     if (st_off < 0)
3871       break;                    // unknown junk in the inits
3872     if (st->in(MemNode::Memory) != zmem)
3873       break;                    // complicated store chains somehow in list
3874 
3875     int st_size = st->as_Store()->memory_size();
3876     intptr_t next_init_off = st_off + st_size;
3877 
3878     if (do_zeroing && zeroes_done < next_init_off) {
3879       // See if this store needs a zero before it or under it.
3880       intptr_t zeroes_needed = st_off;
3881 
3882       if (st_size < BytesPerInt) {
3883         // Look for subword stores which only partially initialize words.
3884         // If we find some, we must lay down some word-level zeroes first,
3885         // underneath the subword stores.
3886         //
3887         // Examples:
3888         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
3889         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
3890         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
3891         //
3892         // Note:  coalesce_subword_stores may have already done this,
3893         // if it was prompted by constant non-zero subword initializers.
3894         // But this case can still arise with non-constant stores.
3895 
3896         intptr_t next_full_store = find_next_fullword_store(i, phase);
3897 
3898         // In the examples above:
3899         //   in(i)          p   q   r   s     x   y     z
3900         //   st_off        12  13  14  15    12  13    14
3901         //   st_size        1   1   1   1     1   1     1
3902         //   next_full_s.  12  16  16  16    16  16    16
3903         //   z's_done      12  16  16  16    12  16    12
3904         //   z's_needed    12  16  16  16    16  16    16
3905         //   zsize          0   0   0   0     4   0     4
3906         if (next_full_store < 0) {
3907           // Conservative tack:  Zero to end of current word.
3908           zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3909         } else {
3910           // Zero to beginning of next fully initialized word.
3911           // Or, don't zero at all, if we are already in that word.
3912           assert(next_full_store >= zeroes_needed, "must go forward");
3913           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3914           zeroes_needed = next_full_store;
3915         }
3916       }
3917 
3918       if (zeroes_needed > zeroes_done) {
3919         intptr_t zsize = zeroes_needed - zeroes_done;
3920         // Do some incremental zeroing on rawmem, in parallel with inits.
3921         zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3922         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3923                                               zeroes_done, zeroes_needed,
3924                                               phase);
3925         zeroes_done = zeroes_needed;
3926         if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3927           do_zeroing = false;   // leave the hole, next time
3928       }
3929     }
3930 
3931     // Collect the store and move on:
3932     st->set_req(MemNode::Memory, inits);
3933     inits = st;                 // put it on the linearized chain
3934     set_req(i, zmem);           // unhook from previous position
3935 
3936     if (zeroes_done == st_off)
3937       zeroes_done = next_init_off;
3938 
3939     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3940 
3941     #ifdef ASSERT
3942     // Various order invariants.  Weaker than stores_are_sane because
3943     // a large constant tile can be filled in by smaller non-constant stores.
3944     assert(st_off >= last_init_off, "inits do not reverse");
3945     last_init_off = st_off;
3946     const Type* val = NULL;
3947     if (st_size >= BytesPerInt &&
3948         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3949         (int)val->basic_type() < (int)T_OBJECT) {
3950       assert(st_off >= last_tile_end, "tiles do not overlap");
3951       assert(st_off >= last_init_end, "tiles do not overwrite inits");
3952       last_tile_end = MAX2(last_tile_end, next_init_off);
3953     } else {
3954       intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3955       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3956       assert(st_off      >= last_init_end, "inits do not overlap");
3957       last_init_end = next_init_off;  // it's a non-tile
3958     }
3959     #endif //ASSERT
3960   }
3961 
3962   remove_extra_zeroes();        // clear out all the zmems left over
3963   add_req(inits);
3964 
3965   if (!ZeroTLAB) {
3966     // If anything remains to be zeroed, zero it all now.
3967     zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3968     // if it is the last unused 4 bytes of an instance, forget about it
3969     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3970     if (zeroes_done + BytesPerLong >= size_limit) {
3971       assert(allocation() != NULL, "");
3972       if (allocation()->Opcode() == Op_Allocate) {
3973         Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3974         ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3975         if (zeroes_done == k->layout_helper())
3976           zeroes_done = size_limit;
3977       }
3978     }
3979     if (zeroes_done < size_limit) {
3980       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3981                                             zeroes_done, size_in_bytes, phase);
3982     }
3983   }
3984 
3985   set_complete(phase);
3986   return rawmem;
3987 }
3988 
3989 
3990 #ifdef ASSERT
3991 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3992   if (is_complete())
3993     return true;                // stores could be anything at this point
3994   assert(allocation() != NULL, "must be present");
3995   intptr_t last_off = allocation()->minimum_header_size();
3996   for (uint i = InitializeNode::RawStores; i < req(); i++) {
3997     Node* st = in(i);
3998     intptr_t st_off = get_store_offset(st, phase);
3999     if (st_off < 0)  continue;  // ignore dead garbage
4000     if (last_off > st_off) {
4001       tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
4002       this->dump(2);
4003       assert(false, "ascending store offsets");
4004       return false;
4005     }
4006     last_off = st_off + st->as_Store()->memory_size();
4007   }
4008   return true;
4009 }
4010 #endif //ASSERT
4011 
4012 
4013 
4014 
4015 //============================MergeMemNode=====================================
4016 //
4017 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
4018 // contributing store or call operations.  Each contributor provides the memory
4019 // state for a particular "alias type" (see Compile::alias_type).  For example,
4020 // if a MergeMem has an input X for alias category #6, then any memory reference
4021 // to alias category #6 may use X as its memory state input, as an exact equivalent
4022 // to using the MergeMem as a whole.
4023 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4024 //
4025 // (Here, the <N> notation gives the index of the relevant adr_type.)
4026 //
4027 // In one special case (and more cases in the future), alias categories overlap.
4028 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4029 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
4030 // it is exactly equivalent to that state W:
4031 //   MergeMem(<Bot>: W) <==> W
4032 //
4033 // Usually, the merge has more than one input.  In that case, where inputs
4034 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4035 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4036 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4037 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4038 //
4039 // A merge can take a "wide" memory state as one of its narrow inputs.
4040 // This simply means that the merge observes out only the relevant parts of
4041 // the wide input.  That is, wide memory states arriving at narrow merge inputs
4042 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
4043 //
4044 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4045 // and that memory slices "leak through":
4046 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4047 //
4048 // But, in such a cascade, repeated memory slices can "block the leak":
4049 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4050 //
4051 // In the last example, Y is not part of the combined memory state of the
4052 // outermost MergeMem.  The system must, of course, prevent unschedulable
4053 // memory states from arising, so you can be sure that the state Y is somehow
4054 // a precursor to state Y'.
4055 //
4056 //
4057 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4058 // of each MergeMemNode array are exactly the numerical alias indexes, including
4059 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
4060 // Compile::alias_type (and kin) produce and manage these indexes.
4061 //
4062 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4063 // (Note that this provides quick access to the top node inside MergeMem methods,
4064 // without the need to reach out via TLS to Compile::current.)
4065 //
4066 // As a consequence of what was just described, a MergeMem that represents a full
4067 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4068 // containing all alias categories.
4069 //
4070 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4071 //
4072 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4073 // a memory state for the alias type <N>, or else the top node, meaning that
4074 // there is no particular input for that alias type.  Note that the length of
4075 // a MergeMem is variable, and may be extended at any time to accommodate new
4076 // memory states at larger alias indexes.  When merges grow, they are of course
4077 // filled with "top" in the unused in() positions.
4078 //
4079 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4080 // (Top was chosen because it works smoothly with passes like GCM.)
4081 //
4082 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
4083 // the type of random VM bits like TLS references.)  Since it is always the
4084 // first non-Bot memory slice, some low-level loops use it to initialize an
4085 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
4086 //
4087 //
4088 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
4089 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
4090 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4091 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
4092 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
4093 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4094 //
4095 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4096 // really that different from the other memory inputs.  An abbreviation called
4097 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4098 //
4099 //
4100 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
4101 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
4102 // that "emerges though" the base memory will be marked as excluding the alias types
4103 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4104 //
4105 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4106 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4107 //
4108 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4109 // (It is currently unimplemented.)  As you can see, the resulting merge is
4110 // actually a disjoint union of memory states, rather than an overlay.
4111 //
4112 
4113 //------------------------------MergeMemNode-----------------------------------
4114 Node* MergeMemNode::make_empty_memory() {
4115   Node* empty_memory = (Node*) Compile::current()->top();
4116   assert(empty_memory->is_top(), "correct sentinel identity");
4117   return empty_memory;
4118 }
4119 
4120 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4121   init_class_id(Class_MergeMem);
4122   // all inputs are nullified in Node::Node(int)
4123   // set_input(0, NULL);  // no control input
4124 
4125   // Initialize the edges uniformly to top, for starters.
4126   Node* empty_mem = make_empty_memory();
4127   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4128     init_req(i,empty_mem);
4129   }
4130   assert(empty_memory() == empty_mem, "");
4131 
4132   if( new_base != NULL && new_base->is_MergeMem() ) {
4133     MergeMemNode* mdef = new_base->as_MergeMem();
4134     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4135     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4136       mms.set_memory(mms.memory2());
4137     }
4138     assert(base_memory() == mdef->base_memory(), "");
4139   } else {
4140     set_base_memory(new_base);
4141   }
4142 }
4143 
4144 // Make a new, untransformed MergeMem with the same base as 'mem'.
4145 // If mem is itself a MergeMem, populate the result with the same edges.
4146 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
4147   return new(C) MergeMemNode(mem);
4148 }
4149 
4150 //------------------------------cmp--------------------------------------------
4151 uint MergeMemNode::hash() const { return NO_HASH; }
4152 uint MergeMemNode::cmp( const Node &n ) const {
4153   return (&n == this);          // Always fail except on self
4154 }
4155 
4156 //------------------------------Identity---------------------------------------
4157 Node* MergeMemNode::Identity(PhaseTransform *phase) {
4158   // Identity if this merge point does not record any interesting memory
4159   // disambiguations.
4160   Node* base_mem = base_memory();
4161   Node* empty_mem = empty_memory();
4162   if (base_mem != empty_mem) {  // Memory path is not dead?
4163     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4164       Node* mem = in(i);
4165       if (mem != empty_mem && mem != base_mem) {
4166         return this;            // Many memory splits; no change
4167       }
4168     }
4169   }
4170   return base_mem;              // No memory splits; ID on the one true input
4171 }
4172 
4173 //------------------------------Ideal------------------------------------------
4174 // This method is invoked recursively on chains of MergeMem nodes
4175 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4176   // Remove chain'd MergeMems
4177   //
4178   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4179   // relative to the "in(Bot)".  Since we are patching both at the same time,
4180   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4181   // but rewrite each "in(i)" relative to the new "in(Bot)".
4182   Node *progress = NULL;
4183 
4184 
4185   Node* old_base = base_memory();
4186   Node* empty_mem = empty_memory();
4187   if (old_base == empty_mem)
4188     return NULL; // Dead memory path.
4189 
4190   MergeMemNode* old_mbase;
4191   if (old_base != NULL && old_base->is_MergeMem())
4192     old_mbase = old_base->as_MergeMem();
4193   else
4194     old_mbase = NULL;
4195   Node* new_base = old_base;
4196 
4197   // simplify stacked MergeMems in base memory
4198   if (old_mbase)  new_base = old_mbase->base_memory();
4199 
4200   // the base memory might contribute new slices beyond my req()
4201   if (old_mbase)  grow_to_match(old_mbase);
4202 
4203   // Look carefully at the base node if it is a phi.
4204   PhiNode* phi_base;
4205   if (new_base != NULL && new_base->is_Phi())
4206     phi_base = new_base->as_Phi();
4207   else
4208     phi_base = NULL;
4209 
4210   Node*    phi_reg = NULL;
4211   uint     phi_len = (uint)-1;
4212   if (phi_base != NULL && !phi_base->is_copy()) {
4213     // do not examine phi if degraded to a copy
4214     phi_reg = phi_base->region();
4215     phi_len = phi_base->req();
4216     // see if the phi is unfinished
4217     for (uint i = 1; i < phi_len; i++) {
4218       if (phi_base->in(i) == NULL) {
4219         // incomplete phi; do not look at it yet!
4220         phi_reg = NULL;
4221         phi_len = (uint)-1;
4222         break;
4223       }
4224     }
4225   }
4226 
4227   // Note:  We do not call verify_sparse on entry, because inputs
4228   // can normalize to the base_memory via subsume_node or similar
4229   // mechanisms.  This method repairs that damage.
4230 
4231   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4232 
4233   // Look at each slice.
4234   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4235     Node* old_in = in(i);
4236     // calculate the old memory value
4237     Node* old_mem = old_in;
4238     if (old_mem == empty_mem)  old_mem = old_base;
4239     assert(old_mem == memory_at(i), "");
4240 
4241     // maybe update (reslice) the old memory value
4242 
4243     // simplify stacked MergeMems
4244     Node* new_mem = old_mem;
4245     MergeMemNode* old_mmem;
4246     if (old_mem != NULL && old_mem->is_MergeMem())
4247       old_mmem = old_mem->as_MergeMem();
4248     else
4249       old_mmem = NULL;
4250     if (old_mmem == this) {
4251       // This can happen if loops break up and safepoints disappear.
4252       // A merge of BotPtr (default) with a RawPtr memory derived from a
4253       // safepoint can be rewritten to a merge of the same BotPtr with
4254       // the BotPtr phi coming into the loop.  If that phi disappears
4255       // also, we can end up with a self-loop of the mergemem.
4256       // In general, if loops degenerate and memory effects disappear,
4257       // a mergemem can be left looking at itself.  This simply means
4258       // that the mergemem's default should be used, since there is
4259       // no longer any apparent effect on this slice.
4260       // Note: If a memory slice is a MergeMem cycle, it is unreachable
4261       //       from start.  Update the input to TOP.
4262       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4263     }
4264     else if (old_mmem != NULL) {
4265       new_mem = old_mmem->memory_at(i);
4266     }
4267     // else preceding memory was not a MergeMem
4268 
4269     // replace equivalent phis (unfortunately, they do not GVN together)
4270     if (new_mem != NULL && new_mem != new_base &&
4271         new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4272       if (new_mem->is_Phi()) {
4273         PhiNode* phi_mem = new_mem->as_Phi();
4274         for (uint i = 1; i < phi_len; i++) {
4275           if (phi_base->in(i) != phi_mem->in(i)) {
4276             phi_mem = NULL;
4277             break;
4278           }
4279         }
4280         if (phi_mem != NULL) {
4281           // equivalent phi nodes; revert to the def
4282           new_mem = new_base;
4283         }
4284       }
4285     }
4286 
4287     // maybe store down a new value
4288     Node* new_in = new_mem;
4289     if (new_in == new_base)  new_in = empty_mem;
4290 
4291     if (new_in != old_in) {
4292       // Warning:  Do not combine this "if" with the previous "if"
4293       // A memory slice might have be be rewritten even if it is semantically
4294       // unchanged, if the base_memory value has changed.
4295       set_req(i, new_in);
4296       progress = this;          // Report progress
4297     }
4298   }
4299 
4300   if (new_base != old_base) {
4301     set_req(Compile::AliasIdxBot, new_base);
4302     // Don't use set_base_memory(new_base), because we need to update du.
4303     assert(base_memory() == new_base, "");
4304     progress = this;
4305   }
4306 
4307   if( base_memory() == this ) {
4308     // a self cycle indicates this memory path is dead
4309     set_req(Compile::AliasIdxBot, empty_mem);
4310   }
4311 
4312   // Resolve external cycles by calling Ideal on a MergeMem base_memory
4313   // Recursion must occur after the self cycle check above
4314   if( base_memory()->is_MergeMem() ) {
4315     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4316     Node *m = phase->transform(new_mbase);  // Rollup any cycles
4317     if( m != NULL && (m->is_top() ||
4318         m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4319       // propagate rollup of dead cycle to self
4320       set_req(Compile::AliasIdxBot, empty_mem);
4321     }
4322   }
4323 
4324   if( base_memory() == empty_mem ) {
4325     progress = this;
4326     // Cut inputs during Parse phase only.
4327     // During Optimize phase a dead MergeMem node will be subsumed by Top.
4328     if( !can_reshape ) {
4329       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4330         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4331       }
4332     }
4333   }
4334 
4335   if( !progress && base_memory()->is_Phi() && can_reshape ) {
4336     // Check if PhiNode::Ideal's "Split phis through memory merges"
4337     // transform should be attempted. Look for this->phi->this cycle.
4338     uint merge_width = req();
4339     if (merge_width > Compile::AliasIdxRaw) {
4340       PhiNode* phi = base_memory()->as_Phi();
4341       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4342         if (phi->in(i) == this) {
4343           phase->is_IterGVN()->_worklist.push(phi);
4344           break;
4345         }
4346       }
4347     }
4348   }
4349 
4350   assert(progress || verify_sparse(), "please, no dups of base");
4351   return progress;
4352 }
4353 
4354 //-------------------------set_base_memory-------------------------------------
4355 void MergeMemNode::set_base_memory(Node *new_base) {
4356   Node* empty_mem = empty_memory();
4357   set_req(Compile::AliasIdxBot, new_base);
4358   assert(memory_at(req()) == new_base, "must set default memory");
4359   // Clear out other occurrences of new_base:
4360   if (new_base != empty_mem) {
4361     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4362       if (in(i) == new_base)  set_req(i, empty_mem);
4363     }
4364   }
4365 }
4366 
4367 //------------------------------out_RegMask------------------------------------
4368 const RegMask &MergeMemNode::out_RegMask() const {
4369   return RegMask::Empty;
4370 }
4371 
4372 //------------------------------dump_spec--------------------------------------
4373 #ifndef PRODUCT
4374 void MergeMemNode::dump_spec(outputStream *st) const {
4375   st->print(" {");
4376   Node* base_mem = base_memory();
4377   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4378     Node* mem = memory_at(i);
4379     if (mem == base_mem) { st->print(" -"); continue; }
4380     st->print( " N%d:", mem->_idx );
4381     Compile::current()->get_adr_type(i)->dump_on(st);
4382   }
4383   st->print(" }");
4384 }
4385 #endif // !PRODUCT
4386 
4387 
4388 #ifdef ASSERT
4389 static bool might_be_same(Node* a, Node* b) {
4390   if (a == b)  return true;
4391   if (!(a->is_Phi() || b->is_Phi()))  return false;
4392   // phis shift around during optimization
4393   return true;  // pretty stupid...
4394 }
4395 
4396 // verify a narrow slice (either incoming or outgoing)
4397 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4398   if (!VerifyAliases)       return;  // don't bother to verify unless requested
4399   if (is_error_reported())  return;  // muzzle asserts when debugging an error
4400   if (Node::in_dump())      return;  // muzzle asserts when printing
4401   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4402   assert(n != NULL, "");
4403   // Elide intervening MergeMem's
4404   while (n->is_MergeMem()) {
4405     n = n->as_MergeMem()->memory_at(alias_idx);
4406   }
4407   Compile* C = Compile::current();
4408   const TypePtr* n_adr_type = n->adr_type();
4409   if (n == m->empty_memory()) {
4410     // Implicit copy of base_memory()
4411   } else if (n_adr_type != TypePtr::BOTTOM) {
4412     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4413     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4414   } else {
4415     // A few places like make_runtime_call "know" that VM calls are narrow,
4416     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4417     bool expected_wide_mem = false;
4418     if (n == m->base_memory()) {
4419       expected_wide_mem = true;
4420     } else if (alias_idx == Compile::AliasIdxRaw ||
4421                n == m->memory_at(Compile::AliasIdxRaw)) {
4422       expected_wide_mem = true;
4423     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4424       // memory can "leak through" calls on channels that
4425       // are write-once.  Allow this also.
4426       expected_wide_mem = true;
4427     }
4428     assert(expected_wide_mem, "expected narrow slice replacement");
4429   }
4430 }
4431 #else // !ASSERT
4432 #define verify_memory_slice(m,i,n) (void)(0)  // PRODUCT version is no-op
4433 #endif
4434 
4435 
4436 //-----------------------------memory_at---------------------------------------
4437 Node* MergeMemNode::memory_at(uint alias_idx) const {
4438   assert(alias_idx >= Compile::AliasIdxRaw ||
4439          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4440          "must avoid base_memory and AliasIdxTop");
4441 
4442   // Otherwise, it is a narrow slice.
4443   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4444   Compile *C = Compile::current();
4445   if (is_empty_memory(n)) {
4446     // the array is sparse; empty slots are the "top" node
4447     n = base_memory();
4448     assert(Node::in_dump()
4449            || n == NULL || n->bottom_type() == Type::TOP
4450            || n->adr_type() == NULL // address is TOP
4451            || n->adr_type() == TypePtr::BOTTOM
4452            || n->adr_type() == TypeRawPtr::BOTTOM
4453            || Compile::current()->AliasLevel() == 0,
4454            "must be a wide memory");
4455     // AliasLevel == 0 if we are organizing the memory states manually.
4456     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4457   } else {
4458     // make sure the stored slice is sane
4459     #ifdef ASSERT
4460     if (is_error_reported() || Node::in_dump()) {
4461     } else if (might_be_same(n, base_memory())) {
4462       // Give it a pass:  It is a mostly harmless repetition of the base.
4463       // This can arise normally from node subsumption during optimization.
4464     } else {
4465       verify_memory_slice(this, alias_idx, n);
4466     }
4467     #endif
4468   }
4469   return n;
4470 }
4471 
4472 //---------------------------set_memory_at-------------------------------------
4473 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4474   verify_memory_slice(this, alias_idx, n);
4475   Node* empty_mem = empty_memory();
4476   if (n == base_memory())  n = empty_mem;  // collapse default
4477   uint need_req = alias_idx+1;
4478   if (req() < need_req) {
4479     if (n == empty_mem)  return;  // already the default, so do not grow me
4480     // grow the sparse array
4481     do {
4482       add_req(empty_mem);
4483     } while (req() < need_req);
4484   }
4485   set_req( alias_idx, n );
4486 }
4487 
4488 
4489 
4490 //--------------------------iteration_setup------------------------------------
4491 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4492   if (other != NULL) {
4493     grow_to_match(other);
4494     // invariant:  the finite support of mm2 is within mm->req()
4495     #ifdef ASSERT
4496     for (uint i = req(); i < other->req(); i++) {
4497       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4498     }
4499     #endif
4500   }
4501   // Replace spurious copies of base_memory by top.
4502   Node* base_mem = base_memory();
4503   if (base_mem != NULL && !base_mem->is_top()) {
4504     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4505       if (in(i) == base_mem)
4506         set_req(i, empty_memory());
4507     }
4508   }
4509 }
4510 
4511 //---------------------------grow_to_match-------------------------------------
4512 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4513   Node* empty_mem = empty_memory();
4514   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4515   // look for the finite support of the other memory
4516   for (uint i = other->req(); --i >= req(); ) {
4517     if (other->in(i) != empty_mem) {
4518       uint new_len = i+1;
4519       while (req() < new_len)  add_req(empty_mem);
4520       break;
4521     }
4522   }
4523 }
4524 
4525 //---------------------------verify_sparse-------------------------------------
4526 #ifndef PRODUCT
4527 bool MergeMemNode::verify_sparse() const {
4528   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4529   Node* base_mem = base_memory();
4530   // The following can happen in degenerate cases, since empty==top.
4531   if (is_empty_memory(base_mem))  return true;
4532   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4533     assert(in(i) != NULL, "sane slice");
4534     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
4535   }
4536   return true;
4537 }
4538 
4539 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4540   Node* n;
4541   n = mm->in(idx);
4542   if (mem == n)  return true;  // might be empty_memory()
4543   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4544   if (mem == n)  return true;
4545   while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4546     if (mem == n)  return true;
4547     if (n == NULL)  break;
4548   }
4549   return false;
4550 }
4551 #endif // !PRODUCT
--- EOF ---