1 /*
   2  * Copyright (c) 1997, 2016, 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 "memory/resourceArea.hpp"
  30 #include "oops/objArrayKlass.hpp"
  31 #include "opto/addnode.hpp"
  32 #include "opto/arraycopynode.hpp"
  33 #include "opto/cfgnode.hpp"
  34 #include "opto/compile.hpp"
  35 #include "opto/connode.hpp"
  36 #include "opto/convertnode.hpp"
  37 #include "opto/loopnode.hpp"
  38 #include "opto/machnode.hpp"
  39 #include "opto/matcher.hpp"
  40 #include "opto/memnode.hpp"
  41 #include "opto/mulnode.hpp"
  42 #include "opto/narrowptrnode.hpp"
  43 #include "opto/phaseX.hpp"
  44 #include "opto/regmask.hpp"
  45 #include "utilities/copy.hpp"

  46 
  47 // Portions of code courtesy of Clifford Click
  48 
  49 // Optimization - Graph Style
  50 
  51 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
  52 
  53 //=============================================================================
  54 uint MemNode::size_of() const { return sizeof(*this); }
  55 
  56 const TypePtr *MemNode::adr_type() const {
  57   Node* adr = in(Address);
  58   if (adr == NULL)  return NULL; // node is dead
  59   const TypePtr* cross_check = NULL;
  60   DEBUG_ONLY(cross_check = _adr_type);
  61   return calculate_adr_type(adr->bottom_type(), cross_check);
  62 }
  63 
  64 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
  65   if (adr != NULL) {
  66     if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
  67       return true;
  68     }
  69   }
  70   return false;
  71 }
  72 
  73 #ifndef PRODUCT
  74 void MemNode::dump_spec(outputStream *st) const {
  75   if (in(Address) == NULL)  return; // node is dead
  76 #ifndef ASSERT
  77   // fake the missing field
  78   const TypePtr* _adr_type = NULL;
  79   if (in(Address) != NULL)
  80     _adr_type = in(Address)->bottom_type()->isa_ptr();
  81 #endif
  82   dump_adr_type(this, _adr_type, st);
  83 
  84   Compile* C = Compile::current();
  85   if (C->alias_type(_adr_type)->is_volatile()) {
  86     st->print(" Volatile!");
  87   }
  88   if (_unaligned_access) {
  89     st->print(" unaligned");
  90   }
  91   if (_mismatched_access) {
  92     st->print(" mismatched");
  93   }
  94 }
  95 
  96 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
  97   st->print(" @");
  98   if (adr_type == NULL) {
  99     st->print("NULL");
 100   } else {
 101     adr_type->dump_on(st);
 102     Compile* C = Compile::current();
 103     Compile::AliasType* atp = NULL;
 104     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
 105     if (atp == NULL)
 106       st->print(", idx=?\?;");
 107     else if (atp->index() == Compile::AliasIdxBot)
 108       st->print(", idx=Bot;");
 109     else if (atp->index() == Compile::AliasIdxTop)
 110       st->print(", idx=Top;");
 111     else if (atp->index() == Compile::AliasIdxRaw)
 112       st->print(", idx=Raw;");
 113     else {
 114       ciField* field = atp->field();
 115       if (field) {
 116         st->print(", name=");
 117         field->print_name_on(st);
 118       }
 119       st->print(", idx=%d;", atp->index());
 120     }
 121   }
 122 }
 123 
 124 extern void print_alias_types();
 125 
 126 #endif
 127 
 128 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
 129   assert((t_oop != NULL), "sanity");
 130   bool is_instance = t_oop->is_known_instance_field();
 131   bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
 132                              (load != NULL) && load->is_Load() &&
 133                              (phase->is_IterGVN() != NULL);
 134   if (!(is_instance || is_boxed_value_load))
 135     return mchain;  // don't try to optimize non-instance types
 136   uint instance_id = t_oop->instance_id();
 137   Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
 138   Node *prev = NULL;
 139   Node *result = mchain;
 140   while (prev != result) {
 141     prev = result;
 142     if (result == start_mem)
 143       break;  // hit one of our sentinels
 144     // skip over a call which does not affect this memory slice
 145     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
 146       Node *proj_in = result->in(0);
 147       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
 148         break;  // hit one of our sentinels
 149       } else if (proj_in->is_Call()) {
 150         // ArrayCopyNodes processed here as well
 151         CallNode *call = proj_in->as_Call();
 152         if (!call->may_modify(t_oop, phase)) { // returns false for instances
 153           result = call->in(TypeFunc::Memory);
 154         }
 155       } else if (proj_in->is_Initialize()) {
 156         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
 157         // Stop if this is the initialization for the object instance which
 158         // contains this memory slice, otherwise skip over it.
 159         if ((alloc == NULL) || (alloc->_idx == instance_id)) {
 160           break;
 161         }
 162         if (is_instance) {
 163           result = proj_in->in(TypeFunc::Memory);
 164         } else if (is_boxed_value_load) {
 165           Node* klass = alloc->in(AllocateNode::KlassNode);
 166           const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
 167           if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
 168             result = proj_in->in(TypeFunc::Memory); // not related allocation
 169           }
 170         }
 171       } else if (proj_in->is_MemBar()) {
 172         ArrayCopyNode* ac = NULL;
 173         if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
 174           break;
 175         }
 176         result = proj_in->in(TypeFunc::Memory);
 177       } else {
 178         assert(false, "unexpected projection");
 179       }
 180     } else if (result->is_ClearArray()) {
 181       if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
 182         // Can not bypass initialization of the instance
 183         // we are looking for.
 184         break;
 185       }
 186       // Otherwise skip it (the call updated 'result' value).
 187     } else if (result->is_MergeMem()) {
 188       result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
 189     }
 190   }
 191   return result;
 192 }
 193 
 194 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
 195   const TypeOopPtr* t_oop = t_adr->isa_oopptr();
 196   if (t_oop == NULL)
 197     return mchain;  // don't try to optimize non-oop types
 198   Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
 199   bool is_instance = t_oop->is_known_instance_field();
 200   PhaseIterGVN *igvn = phase->is_IterGVN();
 201   if (is_instance && igvn != NULL  && result->is_Phi()) {
 202     PhiNode *mphi = result->as_Phi();
 203     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 204     const TypePtr *t = mphi->adr_type();
 205     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
 206         t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
 207         t->is_oopptr()->cast_to_exactness(true)
 208          ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 209          ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
 210       // clone the Phi with our address type
 211       result = mphi->split_out_instance(t_adr, igvn);
 212     } else {
 213       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 214     }
 215   }
 216   return result;
 217 }
 218 
 219 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 220   uint alias_idx = phase->C->get_alias_index(tp);
 221   Node *mem = mmem;
 222 #ifdef ASSERT
 223   {
 224     // Check that current type is consistent with the alias index used during graph construction
 225     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 226     bool consistent =  adr_check == NULL || adr_check->empty() ||
 227                        phase->C->must_alias(adr_check, alias_idx );
 228     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 229     if( !consistent && adr_check != NULL && !adr_check->empty() &&
 230                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 231         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 232         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 233           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 234           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 235       // don't assert if it is dead code.
 236       consistent = true;
 237     }
 238     if( !consistent ) {
 239       st->print("alias_idx==%d, adr_check==", alias_idx);
 240       if( adr_check == NULL ) {
 241         st->print("NULL");
 242       } else {
 243         adr_check->dump();
 244       }
 245       st->cr();
 246       print_alias_types();
 247       assert(consistent, "adr_check must match alias idx");
 248     }
 249   }
 250 #endif
 251   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
 252   // means an array I have not precisely typed yet.  Do not do any
 253   // alias stuff with it any time soon.
 254   const TypeOopPtr *toop = tp->isa_oopptr();
 255   if( tp->base() != Type::AnyPtr &&
 256       !(toop &&
 257         toop->klass() != NULL &&
 258         toop->klass()->is_java_lang_Object() &&
 259         toop->offset() == Type::OffsetBot) ) {
 260     // compress paths and change unreachable cycles to TOP
 261     // If not, we can update the input infinitely along a MergeMem cycle
 262     // Equivalent code in PhiNode::Ideal
 263     Node* m  = phase->transform(mmem);
 264     // If transformed to a MergeMem, get the desired slice
 265     // Otherwise the returned node represents memory for every slice
 266     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
 267     // Update input if it is progress over what we have now
 268   }
 269   return mem;
 270 }
 271 
 272 //--------------------------Ideal_common---------------------------------------
 273 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
 274 // Unhook non-raw memories from complete (macro-expanded) initializations.
 275 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
 276   // If our control input is a dead region, kill all below the region
 277   Node *ctl = in(MemNode::Control);
 278   if (ctl && remove_dead_region(phase, can_reshape))
 279     return this;
 280   ctl = in(MemNode::Control);
 281   // Don't bother trying to transform a dead node
 282   if (ctl && ctl->is_top())  return NodeSentinel;
 283 
 284   PhaseIterGVN *igvn = phase->is_IterGVN();
 285   // Wait if control on the worklist.
 286   if (ctl && can_reshape && igvn != NULL) {
 287     Node* bol = NULL;
 288     Node* cmp = NULL;
 289     if (ctl->in(0)->is_If()) {
 290       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
 291       bol = ctl->in(0)->in(1);
 292       if (bol->is_Bool())
 293         cmp = ctl->in(0)->in(1)->in(1);
 294     }
 295     if (igvn->_worklist.member(ctl) ||
 296         (bol != NULL && igvn->_worklist.member(bol)) ||
 297         (cmp != NULL && igvn->_worklist.member(cmp)) ) {
 298       // This control path may be dead.
 299       // Delay this memory node transformation until the control is processed.
 300       phase->is_IterGVN()->_worklist.push(this);
 301       return NodeSentinel; // caller will return NULL
 302     }
 303   }
 304   // Ignore if memory is dead, or self-loop
 305   Node *mem = in(MemNode::Memory);
 306   if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
 307   assert(mem != this, "dead loop in MemNode::Ideal");
 308 
 309   if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
 310     // This memory slice may be dead.
 311     // Delay this mem node transformation until the memory is processed.
 312     phase->is_IterGVN()->_worklist.push(this);
 313     return NodeSentinel; // caller will return NULL
 314   }
 315 
 316   Node *address = in(MemNode::Address);
 317   const Type *t_adr = phase->type(address);
 318   if (t_adr == Type::TOP)              return NodeSentinel; // caller will return NULL
 319 
 320   if (can_reshape && igvn != NULL &&
 321       (igvn->_worklist.member(address) ||
 322        igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
 323     // The address's base and type may change when the address is processed.
 324     // Delay this mem node transformation until the address is processed.
 325     phase->is_IterGVN()->_worklist.push(this);
 326     return NodeSentinel; // caller will return NULL
 327   }
 328 
 329   // Do NOT remove or optimize the next lines: ensure a new alias index
 330   // is allocated for an oop pointer type before Escape Analysis.
 331   // Note: C++ will not remove it since the call has side effect.
 332   if (t_adr->isa_oopptr()) {
 333     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 334   }
 335 
 336   Node* base = NULL;
 337   if (address->is_AddP()) {
 338     base = address->in(AddPNode::Base);
 339   }
 340   if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
 341       !t_adr->isa_rawptr()) {
 342     // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
 343     // Skip this node optimization if its address has TOP base.
 344     return NodeSentinel; // caller will return NULL
 345   }
 346 
 347   // Avoid independent memory operations
 348   Node* old_mem = mem;
 349 
 350   // The code which unhooks non-raw memories from complete (macro-expanded)
 351   // initializations was removed. After macro-expansion all stores catched
 352   // by Initialize node became raw stores and there is no information
 353   // which memory slices they modify. So it is unsafe to move any memory
 354   // operation above these stores. Also in most cases hooked non-raw memories
 355   // were already unhooked by using information from detect_ptr_independence()
 356   // and find_previous_store().
 357 
 358   if (mem->is_MergeMem()) {
 359     MergeMemNode* mmem = mem->as_MergeMem();
 360     const TypePtr *tp = t_adr->is_ptr();
 361 
 362     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 363   }
 364 
 365   if (mem != old_mem) {
 366     set_req(MemNode::Memory, mem);
 367     if (can_reshape && old_mem->outcnt() == 0) {
 368         igvn->_worklist.push(old_mem);
 369     }
 370     if (phase->type( mem ) == Type::TOP) return NodeSentinel;
 371     return this;
 372   }
 373 
 374   // let the subclass continue analyzing...
 375   return NULL;
 376 }
 377 
 378 // Helper function for proving some simple control dominations.
 379 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 380 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 381 // is not a constant (dominated by the method's StartNode).
 382 // Used by MemNode::find_previous_store to prove that the
 383 // control input of a memory operation predates (dominates)
 384 // an allocation it wants to look past.
 385 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
 386   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
 387     return false; // Conservative answer for dead code
 388 
 389   // Check 'dom'. Skip Proj and CatchProj nodes.
 390   dom = dom->find_exact_control(dom);
 391   if (dom == NULL || dom->is_top())
 392     return false; // Conservative answer for dead code
 393 
 394   if (dom == sub) {
 395     // For the case when, for example, 'sub' is Initialize and the original
 396     // 'dom' is Proj node of the 'sub'.
 397     return false;
 398   }
 399 
 400   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
 401     return true;
 402 
 403   // 'dom' dominates 'sub' if its control edge and control edges
 404   // of all its inputs dominate or equal to sub's control edge.
 405 
 406   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 407   // Or Region for the check in LoadNode::Ideal();
 408   // 'sub' should have sub->in(0) != NULL.
 409   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 410          sub->is_Region() || sub->is_Call(), "expecting only these nodes");
 411 
 412   // Get control edge of 'sub'.
 413   Node* orig_sub = sub;
 414   sub = sub->find_exact_control(sub->in(0));
 415   if (sub == NULL || sub->is_top())
 416     return false; // Conservative answer for dead code
 417 
 418   assert(sub->is_CFG(), "expecting control");
 419 
 420   if (sub == dom)
 421     return true;
 422 
 423   if (sub->is_Start() || sub->is_Root())
 424     return false;
 425 
 426   {
 427     // Check all control edges of 'dom'.
 428 
 429     ResourceMark rm;
 430     Arena* arena = Thread::current()->resource_area();
 431     Node_List nlist(arena);
 432     Unique_Node_List dom_list(arena);
 433 
 434     dom_list.push(dom);
 435     bool only_dominating_controls = false;
 436 
 437     for (uint next = 0; next < dom_list.size(); next++) {
 438       Node* n = dom_list.at(next);
 439       if (n == orig_sub)
 440         return false; // One of dom's inputs dominated by sub.
 441       if (!n->is_CFG() && n->pinned()) {
 442         // Check only own control edge for pinned non-control nodes.
 443         n = n->find_exact_control(n->in(0));
 444         if (n == NULL || n->is_top())
 445           return false; // Conservative answer for dead code
 446         assert(n->is_CFG(), "expecting control");
 447         dom_list.push(n);
 448       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 449         only_dominating_controls = true;
 450       } else if (n->is_CFG()) {
 451         if (n->dominates(sub, nlist))
 452           only_dominating_controls = true;
 453         else
 454           return false;
 455       } else {
 456         // First, own control edge.
 457         Node* m = n->find_exact_control(n->in(0));
 458         if (m != NULL) {
 459           if (m->is_top())
 460             return false; // Conservative answer for dead code
 461           dom_list.push(m);
 462         }
 463         // Now, the rest of edges.
 464         uint cnt = n->req();
 465         for (uint i = 1; i < cnt; i++) {
 466           m = n->find_exact_control(n->in(i));
 467           if (m == NULL || m->is_top())
 468             continue;
 469           dom_list.push(m);
 470         }
 471       }
 472     }
 473     return only_dominating_controls;
 474   }
 475 }
 476 
 477 //---------------------detect_ptr_independence---------------------------------
 478 // Used by MemNode::find_previous_store to prove that two base
 479 // pointers are never equal.
 480 // The pointers are accompanied by their associated allocations,
 481 // if any, which have been previously discovered by the caller.
 482 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 483                                       Node* p2, AllocateNode* a2,
 484                                       PhaseTransform* phase) {
 485   // Attempt to prove that these two pointers cannot be aliased.
 486   // They may both manifestly be allocations, and they should differ.
 487   // Or, if they are not both allocations, they can be distinct constants.
 488   // Otherwise, one is an allocation and the other a pre-existing value.
 489   if (a1 == NULL && a2 == NULL) {           // neither an allocation
 490     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 491   } else if (a1 != NULL && a2 != NULL) {    // both allocations
 492     return (a1 != a2);
 493   } else if (a1 != NULL) {                  // one allocation a1
 494     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 495     return all_controls_dominate(p2, a1);
 496   } else { //(a2 != NULL)                   // one allocation a2
 497     return all_controls_dominate(p1, a2);
 498   }
 499   return false;
 500 }
 501 
 502 
 503 // Find an arraycopy that must have set (can_see_stored_value=true) or
 504 // could have set (can_see_stored_value=false) the value for this load
 505 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
 506   if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
 507                                                mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
 508     Node* mb = mem->in(0);
 509     if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
 510         mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
 511       ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
 512       if (ac->is_clonebasic()) {
 513         intptr_t offset;
 514         AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
 515         assert(alloc != NULL && (!ReduceBulkZeroing || alloc->initialization()->is_complete_with_arraycopy()), "broken allocation");
 516         if (alloc == ld_alloc) {
 517           return ac;
 518         }
 519       }
 520     }
 521   } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
 522     ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
 523 
 524     if (ac->is_arraycopy_validated() ||
 525         ac->is_copyof_validated() ||
 526         ac->is_copyofrange_validated()) {
 527       Node* ld_addp = in(MemNode::Address);
 528       if (ld_addp->is_AddP()) {
 529         Node* ld_base = ld_addp->in(AddPNode::Address);
 530         Node* ld_offs = ld_addp->in(AddPNode::Offset);
 531 
 532         Node* dest = ac->in(ArrayCopyNode::Dest);
 533 
 534         if (dest == ld_base) {
 535           const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
 536           if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
 537             return ac;
 538           }
 539           if (!can_see_stored_value) {
 540             mem = ac->in(TypeFunc::Memory);
 541           }
 542         }
 543       }
 544     }
 545   }
 546   return NULL;
 547 }
 548 
 549 // The logic for reordering loads and stores uses four steps:
 550 // (a) Walk carefully past stores and initializations which we
 551 //     can prove are independent of this load.
 552 // (b) Observe that the next memory state makes an exact match
 553 //     with self (load or store), and locate the relevant store.
 554 // (c) Ensure that, if we were to wire self directly to the store,
 555 //     the optimizer would fold it up somehow.
 556 // (d) Do the rewiring, and return, depending on some other part of
 557 //     the optimizer to fold up the load.
 558 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 559 // specific to loads and stores, so they are handled by the callers.
 560 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 561 //
 562 Node* MemNode::find_previous_store(PhaseTransform* phase) {
 563   Node*         ctrl   = in(MemNode::Control);
 564   Node*         adr    = in(MemNode::Address);
 565   intptr_t      offset = 0;
 566   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 567   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
 568 
 569   if (offset == Type::OffsetBot)
 570     return NULL;            // cannot unalias unless there are precise offsets
 571 
 572   const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
 573   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 574 
 575   intptr_t size_in_bytes = memory_size();
 576 
 577   Node* mem = in(MemNode::Memory);   // start searching here...
 578 
 579   int cnt = 50;             // Cycle limiter
 580   for (;;) {                // While we can dance past unrelated stores...
 581     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 582 
 583     Node* prev = mem;
 584     if (mem->is_Store()) {
 585       Node* st_adr = mem->in(MemNode::Address);
 586       intptr_t st_offset = 0;
 587       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 588       if (st_base == NULL)
 589         break;              // inscrutable pointer
 590 
 591       // For raw accesses it's not enough to prove that constant offsets don't intersect.
 592       // We need the bases to be the equal in order for the offset check to make sense.
 593       if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
 594         break;
 595       }
 596 
 597       if (st_offset != offset && st_offset != Type::OffsetBot) {
 598         const int MAX_STORE = BytesPerLong;
 599         if (st_offset >= offset + size_in_bytes ||
 600             st_offset <= offset - MAX_STORE ||
 601             st_offset <= offset - mem->as_Store()->memory_size()) {
 602           // Success:  The offsets are provably independent.
 603           // (You may ask, why not just test st_offset != offset and be done?
 604           // The answer is that stores of different sizes can co-exist
 605           // in the same sequence of RawMem effects.  We sometimes initialize
 606           // a whole 'tile' of array elements with a single jint or jlong.)
 607           mem = mem->in(MemNode::Memory);
 608           continue;           // (a) advance through independent store memory
 609         }
 610       }
 611       if (st_base != base &&
 612           detect_ptr_independence(base, alloc,
 613                                   st_base,
 614                                   AllocateNode::Ideal_allocation(st_base, phase),
 615                                   phase)) {
 616         // Success:  The bases are provably independent.
 617         mem = mem->in(MemNode::Memory);
 618         continue;           // (a) advance through independent store memory
 619       }
 620 
 621       // (b) At this point, if the bases or offsets do not agree, we lose,
 622       // since we have not managed to prove 'this' and 'mem' independent.
 623       if (st_base == base && st_offset == offset) {
 624         return mem;         // let caller handle steps (c), (d)
 625       }
 626 
 627     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 628       InitializeNode* st_init = mem->in(0)->as_Initialize();
 629       AllocateNode*  st_alloc = st_init->allocation();
 630       if (st_alloc == NULL)
 631         break;              // something degenerated
 632       bool known_identical = false;
 633       bool known_independent = false;
 634       if (alloc == st_alloc)
 635         known_identical = true;
 636       else if (alloc != NULL)
 637         known_independent = true;
 638       else if (all_controls_dominate(this, st_alloc))
 639         known_independent = true;
 640 
 641       if (known_independent) {
 642         // The bases are provably independent: Either they are
 643         // manifestly distinct allocations, or else the control
 644         // of this load dominates the store's allocation.
 645         int alias_idx = phase->C->get_alias_index(adr_type());
 646         if (alias_idx == Compile::AliasIdxRaw) {
 647           mem = st_alloc->in(TypeFunc::Memory);
 648         } else {
 649           mem = st_init->memory(alias_idx);
 650         }
 651         continue;           // (a) advance through independent store memory
 652       }
 653 
 654       // (b) at this point, if we are not looking at a store initializing
 655       // the same allocation we are loading from, we lose.
 656       if (known_identical) {
 657         // From caller, can_see_stored_value will consult find_captured_store.
 658         return mem;         // let caller handle steps (c), (d)
 659       }
 660 
 661     } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
 662       if (prev != mem) {
 663         // Found an arraycopy but it doesn't affect that load
 664         continue;
 665       }
 666       // Found an arraycopy that may affect that load
 667       return mem;
 668     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
 669       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 670       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 671         // ArrayCopyNodes processed here as well.
 672         CallNode *call = mem->in(0)->as_Call();
 673         if (!call->may_modify(addr_t, phase)) {
 674           mem = call->in(TypeFunc::Memory);
 675           continue;         // (a) advance through independent call memory
 676         }
 677       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 678         ArrayCopyNode* ac = NULL;
 679         if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
 680           break;
 681         }
 682         mem = mem->in(0)->in(TypeFunc::Memory);
 683         continue;           // (a) advance through independent MemBar memory
 684       } else if (mem->is_ClearArray()) {
 685         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 686           // (the call updated 'mem' value)
 687           continue;         // (a) advance through independent allocation memory
 688         } else {
 689           // Can not bypass initialization of the instance
 690           // we are looking for.
 691           return mem;
 692         }
 693       } else if (mem->is_MergeMem()) {
 694         int alias_idx = phase->C->get_alias_index(adr_type());
 695         mem = mem->as_MergeMem()->memory_at(alias_idx);
 696         continue;           // (a) advance through independent MergeMem memory
 697       }
 698     }
 699 
 700     // Unless there is an explicit 'continue', we must bail out here,
 701     // because 'mem' is an inscrutable memory state (e.g., a call).
 702     break;
 703   }
 704 
 705   return NULL;              // bail out
 706 }
 707 
 708 //----------------------calculate_adr_type-------------------------------------
 709 // Helper function.  Notices when the given type of address hits top or bottom.
 710 // Also, asserts a cross-check of the type against the expected address type.
 711 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 712   if (t == Type::TOP)  return NULL; // does not touch memory any more?
 713   #ifdef PRODUCT
 714   cross_check = NULL;
 715   #else
 716   if (!VerifyAliases || is_error_reported() || Node::in_dump())  cross_check = NULL;
 717   #endif
 718   const TypePtr* tp = t->isa_ptr();
 719   if (tp == NULL) {
 720     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 721     return TypePtr::BOTTOM;           // touches lots of memory
 722   } else {
 723     #ifdef ASSERT
 724     // %%%% [phh] We don't check the alias index if cross_check is
 725     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 726     if (cross_check != NULL &&
 727         cross_check != TypePtr::BOTTOM &&
 728         cross_check != TypeRawPtr::BOTTOM) {
 729       // Recheck the alias index, to see if it has changed (due to a bug).
 730       Compile* C = Compile::current();
 731       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 732              "must stay in the original alias category");
 733       // The type of the address must be contained in the adr_type,
 734       // disregarding "null"-ness.
 735       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 736       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 737       assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
 738              "real address must not escape from expected memory type");
 739     }
 740     #endif
 741     return tp;
 742   }
 743 }
 744 
 745 //=============================================================================
 746 // Should LoadNode::Ideal() attempt to remove control edges?
 747 bool LoadNode::can_remove_control() const {
 748   return true;
 749 }
 750 uint LoadNode::size_of() const { return sizeof(*this); }
 751 uint LoadNode::cmp( const Node &n ) const
 752 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
 753 const Type *LoadNode::bottom_type() const { return _type; }
 754 uint LoadNode::ideal_reg() const {
 755   return _type->ideal_reg();
 756 }
 757 
 758 #ifndef PRODUCT
 759 void LoadNode::dump_spec(outputStream *st) const {
 760   MemNode::dump_spec(st);
 761   if( !Verbose && !WizardMode ) {
 762     // standard dump does this in Verbose and WizardMode
 763     st->print(" #"); _type->dump_on(st);
 764   }
 765   if (!depends_only_on_test()) {
 766     st->print(" (does not depend only on test)");
 767   }
 768 }
 769 #endif
 770 
 771 #ifdef ASSERT
 772 //----------------------------is_immutable_value-------------------------------
 773 // Helper function to allow a raw load without control edge for some cases
 774 bool LoadNode::is_immutable_value(Node* adr) {
 775   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 776           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 777           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 778            in_bytes(JavaThread::osthread_offset())));
 779 }
 780 #endif
 781 
 782 //----------------------------LoadNode::make-----------------------------------
 783 // Polymorphic factory method:
 784 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
 785                      ControlDependency control_dependency, bool unaligned, bool mismatched) {
 786   Compile* C = gvn.C;
 787 
 788   // sanity check the alias category against the created node type
 789   assert(!(adr_type->isa_oopptr() &&
 790            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
 791          "use LoadKlassNode instead");
 792   assert(!(adr_type->isa_aryptr() &&
 793            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
 794          "use LoadRangeNode instead");
 795   // Check control edge of raw loads
 796   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
 797           // oop will be recorded in oop map if load crosses safepoint
 798           rt->isa_oopptr() || is_immutable_value(adr),
 799           "raw memory operations should have control edge");
 800   LoadNode* load = NULL;
 801   switch (bt) {
 802   case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 803   case T_BYTE:    load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 804   case T_INT:     load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 805   case T_CHAR:    load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 806   case T_SHORT:   load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(),  mo, control_dependency); break;
 807   case T_LONG:    load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
 808   case T_FLOAT:   load = new LoadFNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
 809   case T_DOUBLE:  load = new LoadDNode (ctl, mem, adr, adr_type, rt,            mo, control_dependency); break;
 810   case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(),  mo, control_dependency); break;
 811   case T_OBJECT:
 812 #ifdef _LP64
 813     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 814       load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
 815     } else
 816 #endif
 817     {
 818       assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
 819       load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
 820     }
 821     break;
 822   }
 823   assert(load != NULL, "LoadNode should have been created");
 824   if (unaligned) {
 825     load->set_unaligned_access();
 826   }
 827   if (mismatched) {
 828     load->set_mismatched_access();
 829   }
 830   if (load->Opcode() == Op_LoadN) {
 831     Node* ld = gvn.transform(load);
 832     return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
 833   }
 834 
 835   return load;
 836 }
 837 
 838 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
 839                                   ControlDependency control_dependency, bool unaligned, bool mismatched) {
 840   bool require_atomic = true;
 841   LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
 842   if (unaligned) {
 843     load->set_unaligned_access();
 844   }
 845   if (mismatched) {
 846     load->set_mismatched_access();
 847   }
 848   return load;
 849 }
 850 
 851 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
 852                                   ControlDependency control_dependency, bool unaligned, bool mismatched) {
 853   bool require_atomic = true;
 854   LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
 855   if (unaligned) {
 856     load->set_unaligned_access();
 857   }
 858   if (mismatched) {
 859     load->set_mismatched_access();
 860   }
 861   return load;
 862 }
 863 
 864 
 865 
 866 //------------------------------hash-------------------------------------------
 867 uint LoadNode::hash() const {
 868   // unroll addition of interesting fields
 869   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
 870 }
 871 
 872 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
 873   if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
 874     bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
 875     bool is_stable_ary = FoldStableValues &&
 876                          (tp != NULL) && (tp->isa_aryptr() != NULL) &&
 877                          tp->isa_aryptr()->is_stable();
 878 
 879     return (eliminate_boxing && non_volatile) || is_stable_ary;
 880   }
 881 
 882   return false;
 883 }
 884 
 885 // Is the value loaded previously stored by an arraycopy? If so return
 886 // a load node that reads from the source array so we may be able to
 887 // optimize out the ArrayCopy node later.
 888 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const {
 889   Node* ld_adr = in(MemNode::Address);
 890   intptr_t ld_off = 0;
 891   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 892   Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
 893   if (ac != NULL) {
 894     assert(ac->is_ArrayCopy(), "what kind of node can this be?");
 895 
 896     Node* ld = clone();
 897     if (ac->as_ArrayCopy()->is_clonebasic()) {
 898       assert(ld_alloc != NULL, "need an alloc");
 899       Node* addp = in(MemNode::Address)->clone();
 900       assert(addp->is_AddP(), "address must be addp");
 901       assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern");
 902       assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern");
 903       addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base));
 904       addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address));
 905       ld->set_req(MemNode::Address, phase->transform(addp));
 906       if (in(0) != NULL) {
 907         assert(ld_alloc->in(0) != NULL, "alloc must have control");
 908         ld->set_req(0, ld_alloc->in(0));
 909       }
 910     } else {
 911       Node* src = ac->in(ArrayCopyNode::Src);
 912       Node* addp = in(MemNode::Address)->clone();
 913       assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
 914       addp->set_req(AddPNode::Base, src);
 915       addp->set_req(AddPNode::Address, src);
 916 
 917       const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
 918       BasicType ary_elem  = ary_t->klass()->as_array_klass()->element_type()->basic_type();
 919       uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
 920       uint shift  = exact_log2(type2aelembytes(ary_elem));
 921 
 922       Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 923 #ifdef _LP64
 924       diff = phase->transform(new ConvI2LNode(diff));
 925 #endif
 926       diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
 927 
 928       Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
 929       addp->set_req(AddPNode::Offset, offset);
 930       ld->set_req(MemNode::Address, phase->transform(addp));
 931 
 932       const TypeX *ld_offs_t = phase->type(offset)->isa_intptr_t();
 933 
 934       if (!ac->as_ArrayCopy()->can_replace_dest_load_with_src_load(ld_offs_t->_lo, ld_offs_t->_hi, phase)) {
 935         return NULL;
 936       }
 937 
 938       if (in(0) != NULL) {
 939         assert(ac->in(0) != NULL, "alloc must have control");
 940         ld->set_req(0, ac->in(0));
 941       }
 942     }
 943     // load depends on the tests that validate the arraycopy
 944     ld->as_Load()->_control_dependency = Pinned;
 945     return ld;
 946   }
 947   return NULL;
 948 }
 949 
 950 
 951 //---------------------------can_see_stored_value------------------------------
 952 // This routine exists to make sure this set of tests is done the same
 953 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
 954 // will change the graph shape in a way which makes memory alive twice at the
 955 // same time (uses the Oracle model of aliasing), then some
 956 // LoadXNode::Identity will fold things back to the equivalence-class model
 957 // of aliasing.
 958 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
 959   Node* ld_adr = in(MemNode::Address);
 960   intptr_t ld_off = 0;
 961   AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 962   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
 963   Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
 964   // This is more general than load from boxing objects.
 965   if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
 966     uint alias_idx = atp->index();
 967     bool final = !atp->is_rewritable();
 968     Node* result = NULL;
 969     Node* current = st;
 970     // Skip through chains of MemBarNodes checking the MergeMems for
 971     // new states for the slice of this load.  Stop once any other
 972     // kind of node is encountered.  Loads from final memory can skip
 973     // through any kind of MemBar but normal loads shouldn't skip
 974     // through MemBarAcquire since the could allow them to move out of
 975     // a synchronized region.
 976     while (current->is_Proj()) {
 977       int opc = current->in(0)->Opcode();
 978       if ((final && (opc == Op_MemBarAcquire ||
 979                      opc == Op_MemBarAcquireLock ||
 980                      opc == Op_LoadFence)) ||
 981           opc == Op_MemBarRelease ||
 982           opc == Op_StoreFence ||
 983           opc == Op_MemBarReleaseLock ||
 984           opc == Op_MemBarStoreStore ||
 985           opc == Op_MemBarCPUOrder) {
 986         Node* mem = current->in(0)->in(TypeFunc::Memory);
 987         if (mem->is_MergeMem()) {
 988           MergeMemNode* merge = mem->as_MergeMem();
 989           Node* new_st = merge->memory_at(alias_idx);
 990           if (new_st == merge->base_memory()) {
 991             // Keep searching
 992             current = new_st;
 993             continue;
 994           }
 995           // Save the new memory state for the slice and fall through
 996           // to exit.
 997           result = new_st;
 998         }
 999       }
1000       break;
1001     }
1002     if (result != NULL) {
1003       st = result;
1004     }
1005   }
1006 
1007   // Loop around twice in the case Load -> Initialize -> Store.
1008   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1009   for (int trip = 0; trip <= 1; trip++) {
1010 
1011     if (st->is_Store()) {
1012       Node* st_adr = st->in(MemNode::Address);
1013       if (!phase->eqv(st_adr, ld_adr)) {
1014         // Try harder before giving up...  Match raw and non-raw pointers.
1015         intptr_t st_off = 0;
1016         AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1017         if (alloc == NULL)       return NULL;
1018         if (alloc != ld_alloc)   return NULL;
1019         if (ld_off != st_off)    return NULL;
1020         // At this point we have proven something like this setup:
1021         //  A = Allocate(...)
1022         //  L = LoadQ(,  AddP(CastPP(, A.Parm),, #Off))
1023         //  S = StoreQ(, AddP(,        A.Parm  , #Off), V)
1024         // (Actually, we haven't yet proven the Q's are the same.)
1025         // In other words, we are loading from a casted version of
1026         // the same pointer-and-offset that we stored to.
1027         // Thus, we are able to replace L by V.
1028       }
1029       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1030       if (store_Opcode() != st->Opcode())
1031         return NULL;
1032       return st->in(MemNode::ValueIn);
1033     }
1034 
1035     // A load from a freshly-created object always returns zero.
1036     // (This can happen after LoadNode::Ideal resets the load's memory input
1037     // to find_captured_store, which returned InitializeNode::zero_memory.)
1038     if (st->is_Proj() && st->in(0)->is_Allocate() &&
1039         (st->in(0) == ld_alloc) &&
1040         (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1041       // return a zero value for the load's basic type
1042       // (This is one of the few places where a generic PhaseTransform
1043       // can create new nodes.  Think of it as lazily manifesting
1044       // virtually pre-existing constants.)
1045       return phase->zerocon(memory_type());
1046     }
1047 
1048     // A load from an initialization barrier can match a captured store.
1049     if (st->is_Proj() && st->in(0)->is_Initialize()) {
1050       InitializeNode* init = st->in(0)->as_Initialize();
1051       AllocateNode* alloc = init->allocation();
1052       if ((alloc != NULL) && (alloc == ld_alloc)) {
1053         // examine a captured store value
1054         st = init->find_captured_store(ld_off, memory_size(), phase);
1055         if (st != NULL) {
1056           continue;             // take one more trip around
1057         }
1058       }
1059     }
1060 
1061     // Load boxed value from result of valueOf() call is input parameter.
1062     if (this->is_Load() && ld_adr->is_AddP() &&
1063         (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1064       intptr_t ignore = 0;
1065       Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1066       if (base != NULL && base->is_Proj() &&
1067           base->as_Proj()->_con == TypeFunc::Parms &&
1068           base->in(0)->is_CallStaticJava() &&
1069           base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1070         return base->in(0)->in(TypeFunc::Parms);
1071       }
1072     }
1073 
1074     break;
1075   }
1076 
1077   return NULL;
1078 }
1079 
1080 //----------------------is_instance_field_load_with_local_phi------------------
1081 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1082   if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1083       in(Address)->is_AddP() ) {
1084     const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1085     // Only instances and boxed values.
1086     if( t_oop != NULL &&
1087         (t_oop->is_ptr_to_boxed_value() ||
1088          t_oop->is_known_instance_field()) &&
1089         t_oop->offset() != Type::OffsetBot &&
1090         t_oop->offset() != Type::OffsetTop) {
1091       return true;
1092     }
1093   }
1094   return false;
1095 }
1096 
1097 //------------------------------Identity---------------------------------------
1098 // Loads are identity if previous store is to same address
1099 Node* LoadNode::Identity(PhaseGVN* phase) {
1100   // If the previous store-maker is the right kind of Store, and the store is
1101   // to the same address, then we are equal to the value stored.
1102   Node* mem = in(Memory);
1103   Node* value = can_see_stored_value(mem, phase);
1104   if( value ) {
1105     // byte, short & char stores truncate naturally.
1106     // A load has to load the truncated value which requires
1107     // some sort of masking operation and that requires an
1108     // Ideal call instead of an Identity call.
1109     if (memory_size() < BytesPerInt) {
1110       // If the input to the store does not fit with the load's result type,
1111       // it must be truncated via an Ideal call.
1112       if (!phase->type(value)->higher_equal(phase->type(this)))
1113         return this;
1114     }
1115     // (This works even when value is a Con, but LoadNode::Value
1116     // usually runs first, producing the singleton type of the Con.)
1117     return value;
1118   }
1119 
1120   // Search for an existing data phi which was generated before for the same
1121   // instance's field to avoid infinite generation of phis in a loop.
1122   Node *region = mem->in(0);
1123   if (is_instance_field_load_with_local_phi(region)) {
1124     const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1125     int this_index  = phase->C->get_alias_index(addr_t);
1126     int this_offset = addr_t->offset();
1127     int this_iid    = addr_t->instance_id();
1128     if (!addr_t->is_known_instance() &&
1129          addr_t->is_ptr_to_boxed_value()) {
1130       // Use _idx of address base (could be Phi node) for boxed values.
1131       intptr_t   ignore = 0;
1132       Node*      base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1133       if (base == NULL) {
1134         return this;
1135       }
1136       this_iid = base->_idx;
1137     }
1138     const Type* this_type = bottom_type();
1139     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1140       Node* phi = region->fast_out(i);
1141       if (phi->is_Phi() && phi != mem &&
1142           phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1143         return phi;
1144       }
1145     }
1146   }
1147 
1148   return this;
1149 }
1150 
1151 // Construct an equivalent unsigned load.
1152 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1153   BasicType bt = T_ILLEGAL;
1154   const Type* rt = NULL;
1155   switch (Opcode()) {
1156     case Op_LoadUB: return this;
1157     case Op_LoadUS: return this;
1158     case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1159     case Op_LoadS: bt = T_CHAR;    rt = TypeInt::CHAR;  break;
1160     default:
1161       assert(false, "no unsigned variant: %s", Name());
1162       return NULL;
1163   }
1164   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1165                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1166                         is_unaligned_access(), is_mismatched_access());
1167 }
1168 
1169 // Construct an equivalent signed load.
1170 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1171   BasicType bt = T_ILLEGAL;
1172   const Type* rt = NULL;
1173   switch (Opcode()) {
1174     case Op_LoadUB: bt = T_BYTE;  rt = TypeInt::BYTE;  break;
1175     case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1176     case Op_LoadB: // fall through
1177     case Op_LoadS: // fall through
1178     case Op_LoadI: // fall through
1179     case Op_LoadL: return this;
1180     default:
1181       assert(false, "no signed variant: %s", Name());
1182       return NULL;
1183   }
1184   return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1185                         raw_adr_type(), rt, bt, _mo, _control_dependency,
1186                         is_unaligned_access(), is_mismatched_access());
1187 }
1188 
1189 // We're loading from an object which has autobox behaviour.
1190 // If this object is result of a valueOf call we'll have a phi
1191 // merging a newly allocated object and a load from the cache.
1192 // We want to replace this load with the original incoming
1193 // argument to the valueOf call.
1194 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1195   assert(phase->C->eliminate_boxing(), "sanity");
1196   intptr_t ignore = 0;
1197   Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1198   if ((base == NULL) || base->is_Phi()) {
1199     // Push the loads from the phi that comes from valueOf up
1200     // through it to allow elimination of the loads and the recovery
1201     // of the original value. It is done in split_through_phi().
1202     return NULL;
1203   } else if (base->is_Load() ||
1204              base->is_DecodeN() && base->in(1)->is_Load()) {
1205     // Eliminate the load of boxed value for integer types from the cache
1206     // array by deriving the value from the index into the array.
1207     // Capture the offset of the load and then reverse the computation.
1208 
1209     // Get LoadN node which loads a boxing object from 'cache' array.
1210     if (base->is_DecodeN()) {
1211       base = base->in(1);
1212     }
1213     if (!base->in(Address)->is_AddP()) {
1214       return NULL; // Complex address
1215     }
1216     AddPNode* address = base->in(Address)->as_AddP();
1217     Node* cache_base = address->in(AddPNode::Base);
1218     if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1219       // Get ConP node which is static 'cache' field.
1220       cache_base = cache_base->in(1);
1221     }
1222     if ((cache_base != NULL) && cache_base->is_Con()) {
1223       const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1224       if ((base_type != NULL) && base_type->is_autobox_cache()) {
1225         Node* elements[4];
1226         int shift = exact_log2(type2aelembytes(T_OBJECT));
1227         int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1228         if ((count >  0) && elements[0]->is_Con() &&
1229             ((count == 1) ||
1230              (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1231                              elements[1]->in(2) == phase->intcon(shift))) {
1232           ciObjArray* array = base_type->const_oop()->as_obj_array();
1233           // Fetch the box object cache[0] at the base of the array and get its value
1234           ciInstance* box = array->obj_at(0)->as_instance();
1235           ciInstanceKlass* ik = box->klass()->as_instance_klass();
1236           assert(ik->is_box_klass(), "sanity");
1237           assert(ik->nof_nonstatic_fields() == 1, "change following code");
1238           if (ik->nof_nonstatic_fields() == 1) {
1239             // This should be true nonstatic_field_at requires calling
1240             // nof_nonstatic_fields so check it anyway
1241             ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1242             BasicType bt = c.basic_type();
1243             // Only integer types have boxing cache.
1244             assert(bt == T_BOOLEAN || bt == T_CHAR  ||
1245                    bt == T_BYTE    || bt == T_SHORT ||
1246                    bt == T_INT     || bt == T_LONG, "wrong type = %s", type2name(bt));
1247             jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1248             if (cache_low != (int)cache_low) {
1249               return NULL; // should not happen since cache is array indexed by value
1250             }
1251             jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1252             if (offset != (int)offset) {
1253               return NULL; // should not happen since cache is array indexed by value
1254             }
1255            // Add up all the offsets making of the address of the load
1256             Node* result = elements[0];
1257             for (int i = 1; i < count; i++) {
1258               result = phase->transform(new AddXNode(result, elements[i]));
1259             }
1260             // Remove the constant offset from the address and then
1261             result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1262             // remove the scaling of the offset to recover the original index.
1263             if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1264               // Peel the shift off directly but wrap it in a dummy node
1265               // since Ideal can't return existing nodes
1266               result = new RShiftXNode(result->in(1), phase->intcon(0));
1267             } else if (result->is_Add() && result->in(2)->is_Con() &&
1268                        result->in(1)->Opcode() == Op_LShiftX &&
1269                        result->in(1)->in(2) == phase->intcon(shift)) {
1270               // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1271               // but for boxing cache access we know that X<<Z will not overflow
1272               // (there is range check) so we do this optimizatrion by hand here.
1273               Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1274               result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1275             } else {
1276               result = new RShiftXNode(result, phase->intcon(shift));
1277             }
1278 #ifdef _LP64
1279             if (bt != T_LONG) {
1280               result = new ConvL2INode(phase->transform(result));
1281             }
1282 #else
1283             if (bt == T_LONG) {
1284               result = new ConvI2LNode(phase->transform(result));
1285             }
1286 #endif
1287             // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1288             // Need to preserve unboxing load type if it is unsigned.
1289             switch(this->Opcode()) {
1290               case Op_LoadUB:
1291                 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1292                 break;
1293               case Op_LoadUS:
1294                 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1295                 break;
1296             }
1297             return result;
1298           }
1299         }
1300       }
1301     }
1302   }
1303   return NULL;
1304 }
1305 
1306 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1307   Node* region = phi->in(0);
1308   if (region == NULL) {
1309     return false; // Wait stable graph
1310   }
1311   uint cnt = phi->req();
1312   for (uint i = 1; i < cnt; i++) {
1313     Node* rc = region->in(i);
1314     if (rc == NULL || phase->type(rc) == Type::TOP)
1315       return false; // Wait stable graph
1316     Node* in = phi->in(i);
1317     if (in == NULL || phase->type(in) == Type::TOP)
1318       return false; // Wait stable graph
1319   }
1320   return true;
1321 }
1322 //------------------------------split_through_phi------------------------------
1323 // Split instance or boxed field load through Phi.
1324 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1325   Node* mem     = in(Memory);
1326   Node* address = in(Address);
1327   const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1328 
1329   assert((t_oop != NULL) &&
1330          (t_oop->is_known_instance_field() ||
1331           t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1332 
1333   Compile* C = phase->C;
1334   intptr_t ignore = 0;
1335   Node*    base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1336   bool base_is_phi = (base != NULL) && base->is_Phi();
1337   bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1338                            (base != NULL) && (base == address->in(AddPNode::Base)) &&
1339                            phase->type(base)->higher_equal(TypePtr::NOTNULL);
1340 
1341   if (!((mem->is_Phi() || base_is_phi) &&
1342         (load_boxed_values || t_oop->is_known_instance_field()))) {
1343     return NULL; // memory is not Phi
1344   }
1345 
1346   if (mem->is_Phi()) {
1347     if (!stable_phi(mem->as_Phi(), phase)) {
1348       return NULL; // Wait stable graph
1349     }
1350     uint cnt = mem->req();
1351     // Check for loop invariant memory.
1352     if (cnt == 3) {
1353       for (uint i = 1; i < cnt; i++) {
1354         Node* in = mem->in(i);
1355         Node*  m = optimize_memory_chain(in, t_oop, this, phase);
1356         if (m == mem) {
1357           set_req(Memory, mem->in(cnt - i));
1358           return this; // made change
1359         }
1360       }
1361     }
1362   }
1363   if (base_is_phi) {
1364     if (!stable_phi(base->as_Phi(), phase)) {
1365       return NULL; // Wait stable graph
1366     }
1367     uint cnt = base->req();
1368     // Check for loop invariant memory.
1369     if (cnt == 3) {
1370       for (uint i = 1; i < cnt; i++) {
1371         if (base->in(i) == base) {
1372           return NULL; // Wait stable graph
1373         }
1374       }
1375     }
1376   }
1377 
1378   bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1379 
1380   // Split through Phi (see original code in loopopts.cpp).
1381   assert(C->have_alias_type(t_oop), "instance should have alias type");
1382 
1383   // Do nothing here if Identity will find a value
1384   // (to avoid infinite chain of value phis generation).
1385   if (!phase->eqv(this, this->Identity(phase)))
1386     return NULL;
1387 
1388   // Select Region to split through.
1389   Node* region;
1390   if (!base_is_phi) {
1391     assert(mem->is_Phi(), "sanity");
1392     region = mem->in(0);
1393     // Skip if the region dominates some control edge of the address.
1394     if (!MemNode::all_controls_dominate(address, region))
1395       return NULL;
1396   } else if (!mem->is_Phi()) {
1397     assert(base_is_phi, "sanity");
1398     region = base->in(0);
1399     // Skip if the region dominates some control edge of the memory.
1400     if (!MemNode::all_controls_dominate(mem, region))
1401       return NULL;
1402   } else if (base->in(0) != mem->in(0)) {
1403     assert(base_is_phi && mem->is_Phi(), "sanity");
1404     if (MemNode::all_controls_dominate(mem, base->in(0))) {
1405       region = base->in(0);
1406     } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1407       region = mem->in(0);
1408     } else {
1409       return NULL; // complex graph
1410     }
1411   } else {
1412     assert(base->in(0) == mem->in(0), "sanity");
1413     region = mem->in(0);
1414   }
1415 
1416   const Type* this_type = this->bottom_type();
1417   int this_index  = C->get_alias_index(t_oop);
1418   int this_offset = t_oop->offset();
1419   int this_iid    = t_oop->instance_id();
1420   if (!t_oop->is_known_instance() && load_boxed_values) {
1421     // Use _idx of address base for boxed values.
1422     this_iid = base->_idx;
1423   }
1424   PhaseIterGVN* igvn = phase->is_IterGVN();
1425   Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1426   for (uint i = 1; i < region->req(); i++) {
1427     Node* x;
1428     Node* the_clone = NULL;
1429     if (region->in(i) == C->top()) {
1430       x = C->top();      // Dead path?  Use a dead data op
1431     } else {
1432       x = this->clone();        // Else clone up the data op
1433       the_clone = x;            // Remember for possible deletion.
1434       // Alter data node to use pre-phi inputs
1435       if (this->in(0) == region) {
1436         x->set_req(0, region->in(i));
1437       } else {
1438         x->set_req(0, NULL);
1439       }
1440       if (mem->is_Phi() && (mem->in(0) == region)) {
1441         x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1442       }
1443       if (address->is_Phi() && address->in(0) == region) {
1444         x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1445       }
1446       if (base_is_phi && (base->in(0) == region)) {
1447         Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1448         Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1449         x->set_req(Address, adr_x);
1450       }
1451     }
1452     // Check for a 'win' on some paths
1453     const Type *t = x->Value(igvn);
1454 
1455     bool singleton = t->singleton();
1456 
1457     // See comments in PhaseIdealLoop::split_thru_phi().
1458     if (singleton && t == Type::TOP) {
1459       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1460     }
1461 
1462     if (singleton) {
1463       x = igvn->makecon(t);
1464     } else {
1465       // We now call Identity to try to simplify the cloned node.
1466       // Note that some Identity methods call phase->type(this).
1467       // Make sure that the type array is big enough for
1468       // our new node, even though we may throw the node away.
1469       // (This tweaking with igvn only works because x is a new node.)
1470       igvn->set_type(x, t);
1471       // If x is a TypeNode, capture any more-precise type permanently into Node
1472       // otherwise it will be not updated during igvn->transform since
1473       // igvn->type(x) is set to x->Value() already.
1474       x->raise_bottom_type(t);
1475       Node *y = x->Identity(igvn);
1476       if (y != x) {
1477         x = y;
1478       } else {
1479         y = igvn->hash_find_insert(x);
1480         if (y) {
1481           x = y;
1482         } else {
1483           // Else x is a new node we are keeping
1484           // We do not need register_new_node_with_optimizer
1485           // because set_type has already been called.
1486           igvn->_worklist.push(x);
1487         }
1488       }
1489     }
1490     if (x != the_clone && the_clone != NULL) {
1491       igvn->remove_dead_node(the_clone);
1492     }
1493     phi->set_req(i, x);
1494   }
1495   // Record Phi
1496   igvn->register_new_node_with_optimizer(phi);
1497   return phi;
1498 }
1499 
1500 //------------------------------Ideal------------------------------------------
1501 // If the load is from Field memory and the pointer is non-null, it might be possible to
1502 // zero out the control input.
1503 // If the offset is constant and the base is an object allocation,
1504 // try to hook me up to the exact initializing store.
1505 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1506   Node* p = MemNode::Ideal_common(phase, can_reshape);
1507   if (p)  return (p == NodeSentinel) ? NULL : p;
1508 
1509   Node* ctrl    = in(MemNode::Control);
1510   Node* address = in(MemNode::Address);
1511   bool progress = false;
1512 
1513   // Skip up past a SafePoint control.  Cannot do this for Stores because
1514   // pointer stores & cardmarks must stay on the same side of a SafePoint.
1515   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1516       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1517     ctrl = ctrl->in(0);
1518     set_req(MemNode::Control,ctrl);
1519     progress = true;
1520   }
1521 
1522   intptr_t ignore = 0;
1523   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1524   if (base != NULL
1525       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1526     // Check for useless control edge in some common special cases
1527     if (in(MemNode::Control) != NULL
1528         && can_remove_control()
1529         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1530         && all_controls_dominate(base, phase->C->start())) {
1531       // A method-invariant, non-null address (constant or 'this' argument).
1532       set_req(MemNode::Control, NULL);
1533       progress = true;
1534     }
1535   }
1536 
1537   Node* mem = in(MemNode::Memory);
1538   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1539 
1540   if (can_reshape && (addr_t != NULL)) {
1541     // try to optimize our memory input
1542     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1543     if (opt_mem != mem) {
1544       set_req(MemNode::Memory, opt_mem);
1545       if (phase->type( opt_mem ) == Type::TOP) return NULL;
1546       return this;
1547     }
1548     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1549     if ((t_oop != NULL) &&
1550         (t_oop->is_known_instance_field() ||
1551          t_oop->is_ptr_to_boxed_value())) {
1552       PhaseIterGVN *igvn = phase->is_IterGVN();
1553       if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1554         // Delay this transformation until memory Phi is processed.
1555         phase->is_IterGVN()->_worklist.push(this);
1556         return NULL;
1557       }
1558       // Split instance field load through Phi.
1559       Node* result = split_through_phi(phase);
1560       if (result != NULL) return result;
1561 
1562       if (t_oop->is_ptr_to_boxed_value()) {
1563         Node* result = eliminate_autobox(phase);
1564         if (result != NULL) return result;
1565       }
1566     }
1567   }
1568 
1569   // Is there a dominating load that loads the same value?  Leave
1570   // anything that is not a load of a field/array element (like
1571   // barriers etc.) alone
1572   if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) {
1573     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1574       Node *use = mem->fast_out(i);
1575       if (use != this &&
1576           use->Opcode() == Opcode() &&
1577           use->in(0) != NULL &&
1578           use->in(0) != in(0) &&
1579           use->in(Address) == in(Address)) {
1580         Node* ctl = in(0);
1581         for (int i = 0; i < 10 && ctl != NULL; i++) {
1582           ctl = IfNode::up_one_dom(ctl);
1583           if (ctl == use->in(0)) {
1584             set_req(0, use->in(0));
1585             return this;
1586           }
1587         }
1588       }
1589     }
1590   }
1591 
1592   // Check for prior store with a different base or offset; make Load
1593   // independent.  Skip through any number of them.  Bail out if the stores
1594   // are in an endless dead cycle and report no progress.  This is a key
1595   // transform for Reflection.  However, if after skipping through the Stores
1596   // we can't then fold up against a prior store do NOT do the transform as
1597   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
1598   // array memory alive twice: once for the hoisted Load and again after the
1599   // bypassed Store.  This situation only works if EVERYBODY who does
1600   // anti-dependence work knows how to bypass.  I.e. we need all
1601   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
1602   // the alias index stuff.  So instead, peek through Stores and IFF we can
1603   // fold up, do so.
1604   Node* prev_mem = find_previous_store(phase);
1605   if (prev_mem != NULL) {
1606     Node* value = can_see_arraycopy_value(prev_mem, phase);
1607     if (value != NULL) {
1608       return value;
1609     }
1610   }
1611   // Steps (a), (b):  Walk past independent stores to find an exact match.
1612   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1613     // (c) See if we can fold up on the spot, but don't fold up here.
1614     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1615     // just return a prior value, which is done by Identity calls.
1616     if (can_see_stored_value(prev_mem, phase)) {
1617       // Make ready for step (d):
1618       set_req(MemNode::Memory, prev_mem);
1619       return this;
1620     }
1621   }
1622 
1623   return progress ? this : NULL;
1624 }
1625 
1626 // Helper to recognize certain Klass fields which are invariant across
1627 // some group of array types (e.g., int[] or all T[] where T < Object).
1628 const Type*
1629 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1630                                  ciKlass* klass) const {
1631   if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1632     // The field is Klass::_modifier_flags.  Return its (constant) value.
1633     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1634     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1635     return TypeInt::make(klass->modifier_flags());
1636   }
1637   if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1638     // The field is Klass::_access_flags.  Return its (constant) value.
1639     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1640     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1641     return TypeInt::make(klass->access_flags());
1642   }
1643   if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1644     // The field is Klass::_layout_helper.  Return its constant value if known.
1645     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1646     return TypeInt::make(klass->layout_helper());
1647   }
1648 
1649   // No match.
1650   return NULL;
1651 }
1652 
1653 //------------------------------Value-----------------------------------------
1654 const Type* LoadNode::Value(PhaseGVN* phase) const {
1655   // Either input is TOP ==> the result is TOP
1656   Node* mem = in(MemNode::Memory);
1657   const Type *t1 = phase->type(mem);
1658   if (t1 == Type::TOP)  return Type::TOP;
1659   Node* adr = in(MemNode::Address);
1660   const TypePtr* tp = phase->type(adr)->isa_ptr();
1661   if (tp == NULL || tp->empty())  return Type::TOP;
1662   int off = tp->offset();
1663   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1664   Compile* C = phase->C;
1665 
1666   // Try to guess loaded type from pointer type
1667   if (tp->isa_aryptr()) {
1668     const TypeAryPtr* ary = tp->is_aryptr();
1669     const Type* t = ary->elem();
1670 
1671     // Determine whether the reference is beyond the header or not, by comparing
1672     // the offset against the offset of the start of the array's data.
1673     // Different array types begin at slightly different offsets (12 vs. 16).
1674     // We choose T_BYTE as an example base type that is least restrictive
1675     // as to alignment, which will therefore produce the smallest
1676     // possible base offset.
1677     const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1678     const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1679 
1680     // Try to constant-fold a stable array element.
1681     if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1682       // Make sure the reference is not into the header and the offset is constant
1683       ciObject* aobj = ary->const_oop();
1684       if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1685         int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1686         const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1687                                                                       stable_dimension,
1688                                                                       memory_type(), is_unsigned());
1689         if (con_type != NULL) {
1690           return con_type;
1691         }
1692       }
1693     }
1694 
1695     // Don't do this for integer types. There is only potential profit if
1696     // the element type t is lower than _type; that is, for int types, if _type is
1697     // more restrictive than t.  This only happens here if one is short and the other
1698     // char (both 16 bits), and in those cases we've made an intentional decision
1699     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1700     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1701     //
1702     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1703     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
1704     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1705     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
1706     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1707     // In fact, that could have been the original type of p1, and p1 could have
1708     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1709     // expression (LShiftL quux 3) independently optimized to the constant 8.
1710     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1711         && (_type->isa_vect() == NULL)
1712         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1713       // t might actually be lower than _type, if _type is a unique
1714       // concrete subclass of abstract class t.
1715       if (off_beyond_header) {  // is the offset beyond the header?
1716         const Type* jt = t->join_speculative(_type);
1717         // In any case, do not allow the join, per se, to empty out the type.
1718         if (jt->empty() && !t->empty()) {
1719           // This can happen if a interface-typed array narrows to a class type.
1720           jt = _type;
1721         }
1722 #ifdef ASSERT
1723         if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1724           // The pointers in the autobox arrays are always non-null
1725           Node* base = adr->in(AddPNode::Base);
1726           if ((base != NULL) && base->is_DecodeN()) {
1727             // Get LoadN node which loads IntegerCache.cache field
1728             base = base->in(1);
1729           }
1730           if ((base != NULL) && base->is_Con()) {
1731             const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1732             if ((base_type != NULL) && base_type->is_autobox_cache()) {
1733               // It could be narrow oop
1734               assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1735             }
1736           }
1737         }
1738 #endif
1739         return jt;
1740       }
1741     }
1742   } else if (tp->base() == Type::InstPtr) {
1743     assert( off != Type::OffsetBot ||
1744             // arrays can be cast to Objects
1745             tp->is_oopptr()->klass()->is_java_lang_Object() ||
1746             // unsafe field access may not have a constant offset
1747             C->has_unsafe_access(),
1748             "Field accesses must be precise" );
1749     // For oop loads, we expect the _type to be precise.
1750 
1751     // Optimize loads from constant fields.
1752     const TypeInstPtr* tinst = tp->is_instptr();
1753     ciObject* const_oop = tinst->const_oop();
1754     if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1755       const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1756       if (con_type != NULL) {
1757         return con_type;
1758       }
1759     }
1760   } else if (tp->base() == Type::KlassPtr) {
1761     assert( off != Type::OffsetBot ||
1762             // arrays can be cast to Objects
1763             tp->is_klassptr()->klass()->is_java_lang_Object() ||
1764             // also allow array-loading from the primary supertype
1765             // array during subtype checks
1766             Opcode() == Op_LoadKlass,
1767             "Field accesses must be precise" );
1768     // For klass/static loads, we expect the _type to be precise
1769   }
1770 
1771   const TypeKlassPtr *tkls = tp->isa_klassptr();
1772   if (tkls != NULL && !StressReflectiveCode) {
1773     ciKlass* klass = tkls->klass();
1774     if (klass->is_loaded() && tkls->klass_is_exact()) {
1775       // We are loading a field from a Klass metaobject whose identity
1776       // is known at compile time (the type is "exact" or "precise").
1777       // Check for fields we know are maintained as constants by the VM.
1778       if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1779         // The field is Klass::_super_check_offset.  Return its (constant) value.
1780         // (Folds up type checking code.)
1781         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1782         return TypeInt::make(klass->super_check_offset());
1783       }
1784       // Compute index into primary_supers array
1785       juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1786       // Check for overflowing; use unsigned compare to handle the negative case.
1787       if( depth < ciKlass::primary_super_limit() ) {
1788         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1789         // (Folds up type checking code.)
1790         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1791         ciKlass *ss = klass->super_of_depth(depth);
1792         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1793       }
1794       const Type* aift = load_array_final_field(tkls, klass);
1795       if (aift != NULL)  return aift;
1796       if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1797         // The field is Klass::_java_mirror.  Return its (constant) value.
1798         // (Folds up the 2nd indirection in anObjConstant.getClass().)
1799         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1800         return TypeInstPtr::make(klass->java_mirror());
1801       }
1802     }
1803 
1804     // We can still check if we are loading from the primary_supers array at a
1805     // shallow enough depth.  Even though the klass is not exact, entries less
1806     // than or equal to its super depth are correct.
1807     if (klass->is_loaded() ) {
1808       ciType *inner = klass;
1809       while( inner->is_obj_array_klass() )
1810         inner = inner->as_obj_array_klass()->base_element_type();
1811       if( inner->is_instance_klass() &&
1812           !inner->as_instance_klass()->flags().is_interface() ) {
1813         // Compute index into primary_supers array
1814         juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1815         // Check for overflowing; use unsigned compare to handle the negative case.
1816         if( depth < ciKlass::primary_super_limit() &&
1817             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1818           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1819           // (Folds up type checking code.)
1820           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1821           ciKlass *ss = klass->super_of_depth(depth);
1822           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1823         }
1824       }
1825     }
1826 
1827     // If the type is enough to determine that the thing is not an array,
1828     // we can give the layout_helper a positive interval type.
1829     // This will help short-circuit some reflective code.
1830     if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1831         && !klass->is_array_klass() // not directly typed as an array
1832         && !klass->is_interface()  // specifically not Serializable & Cloneable
1833         && !klass->is_java_lang_Object()   // not the supertype of all T[]
1834         ) {
1835       // Note:  When interfaces are reliable, we can narrow the interface
1836       // test to (klass != Serializable && klass != Cloneable).
1837       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1838       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1839       // The key property of this type is that it folds up tests
1840       // for array-ness, since it proves that the layout_helper is positive.
1841       // Thus, a generic value like the basic object layout helper works fine.
1842       return TypeInt::make(min_size, max_jint, Type::WidenMin);
1843     }
1844   }
1845 
1846   // If we are loading from a freshly-allocated object, produce a zero,
1847   // if the load is provably beyond the header of the object.
1848   // (Also allow a variable load from a fresh array to produce zero.)
1849   const TypeOopPtr *tinst = tp->isa_oopptr();
1850   bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1851   bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1852   if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1853     Node* value = can_see_stored_value(mem,phase);
1854     if (value != NULL && value->is_Con()) {
1855       assert(value->bottom_type()->higher_equal(_type),"sanity");
1856       return value->bottom_type();
1857     }
1858   }
1859 
1860   if (is_instance) {
1861     // If we have an instance type and our memory input is the
1862     // programs's initial memory state, there is no matching store,
1863     // so just return a zero of the appropriate type
1864     Node *mem = in(MemNode::Memory);
1865     if (mem->is_Parm() && mem->in(0)->is_Start()) {
1866       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1867       return Type::get_zero_type(_type->basic_type());
1868     }
1869   }
1870   return _type;
1871 }
1872 
1873 //------------------------------match_edge-------------------------------------
1874 // Do we Match on this edge index or not?  Match only the address.
1875 uint LoadNode::match_edge(uint idx) const {
1876   return idx == MemNode::Address;
1877 }
1878 
1879 //--------------------------LoadBNode::Ideal--------------------------------------
1880 //
1881 //  If the previous store is to the same address as this load,
1882 //  and the value stored was larger than a byte, replace this load
1883 //  with the value stored truncated to a byte.  If no truncation is
1884 //  needed, the replacement is done in LoadNode::Identity().
1885 //
1886 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1887   Node* mem = in(MemNode::Memory);
1888   Node* value = can_see_stored_value(mem,phase);
1889   if( value && !phase->type(value)->higher_equal( _type ) ) {
1890     Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1891     return new RShiftINode(result, phase->intcon(24));
1892   }
1893   // Identity call will handle the case where truncation is not needed.
1894   return LoadNode::Ideal(phase, can_reshape);
1895 }
1896 
1897 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1898   Node* mem = in(MemNode::Memory);
1899   Node* value = can_see_stored_value(mem,phase);
1900   if (value != NULL && value->is_Con() &&
1901       !value->bottom_type()->higher_equal(_type)) {
1902     // If the input to the store does not fit with the load's result type,
1903     // it must be truncated. We can't delay until Ideal call since
1904     // a singleton Value is needed for split_thru_phi optimization.
1905     int con = value->get_int();
1906     return TypeInt::make((con << 24) >> 24);
1907   }
1908   return LoadNode::Value(phase);
1909 }
1910 
1911 //--------------------------LoadUBNode::Ideal-------------------------------------
1912 //
1913 //  If the previous store is to the same address as this load,
1914 //  and the value stored was larger than a byte, replace this load
1915 //  with the value stored truncated to a byte.  If no truncation is
1916 //  needed, the replacement is done in LoadNode::Identity().
1917 //
1918 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1919   Node* mem = in(MemNode::Memory);
1920   Node* value = can_see_stored_value(mem, phase);
1921   if (value && !phase->type(value)->higher_equal(_type))
1922     return new AndINode(value, phase->intcon(0xFF));
1923   // Identity call will handle the case where truncation is not needed.
1924   return LoadNode::Ideal(phase, can_reshape);
1925 }
1926 
1927 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
1928   Node* mem = in(MemNode::Memory);
1929   Node* value = can_see_stored_value(mem,phase);
1930   if (value != NULL && value->is_Con() &&
1931       !value->bottom_type()->higher_equal(_type)) {
1932     // If the input to the store does not fit with the load's result type,
1933     // it must be truncated. We can't delay until Ideal call since
1934     // a singleton Value is needed for split_thru_phi optimization.
1935     int con = value->get_int();
1936     return TypeInt::make(con & 0xFF);
1937   }
1938   return LoadNode::Value(phase);
1939 }
1940 
1941 //--------------------------LoadUSNode::Ideal-------------------------------------
1942 //
1943 //  If the previous store is to the same address as this load,
1944 //  and the value stored was larger than a char, replace this load
1945 //  with the value stored truncated to a char.  If no truncation is
1946 //  needed, the replacement is done in LoadNode::Identity().
1947 //
1948 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1949   Node* mem = in(MemNode::Memory);
1950   Node* value = can_see_stored_value(mem,phase);
1951   if( value && !phase->type(value)->higher_equal( _type ) )
1952     return new AndINode(value,phase->intcon(0xFFFF));
1953   // Identity call will handle the case where truncation is not needed.
1954   return LoadNode::Ideal(phase, can_reshape);
1955 }
1956 
1957 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
1958   Node* mem = in(MemNode::Memory);
1959   Node* value = can_see_stored_value(mem,phase);
1960   if (value != NULL && value->is_Con() &&
1961       !value->bottom_type()->higher_equal(_type)) {
1962     // If the input to the store does not fit with the load's result type,
1963     // it must be truncated. We can't delay until Ideal call since
1964     // a singleton Value is needed for split_thru_phi optimization.
1965     int con = value->get_int();
1966     return TypeInt::make(con & 0xFFFF);
1967   }
1968   return LoadNode::Value(phase);
1969 }
1970 
1971 //--------------------------LoadSNode::Ideal--------------------------------------
1972 //
1973 //  If the previous store is to the same address as this load,
1974 //  and the value stored was larger than a short, replace this load
1975 //  with the value stored truncated to a short.  If no truncation is
1976 //  needed, the replacement is done in LoadNode::Identity().
1977 //
1978 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1979   Node* mem = in(MemNode::Memory);
1980   Node* value = can_see_stored_value(mem,phase);
1981   if( value && !phase->type(value)->higher_equal( _type ) ) {
1982     Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
1983     return new RShiftINode(result, phase->intcon(16));
1984   }
1985   // Identity call will handle the case where truncation is not needed.
1986   return LoadNode::Ideal(phase, can_reshape);
1987 }
1988 
1989 const Type* LoadSNode::Value(PhaseGVN* phase) const {
1990   Node* mem = in(MemNode::Memory);
1991   Node* value = can_see_stored_value(mem,phase);
1992   if (value != NULL && value->is_Con() &&
1993       !value->bottom_type()->higher_equal(_type)) {
1994     // If the input to the store does not fit with the load's result type,
1995     // it must be truncated. We can't delay until Ideal call since
1996     // a singleton Value is needed for split_thru_phi optimization.
1997     int con = value->get_int();
1998     return TypeInt::make((con << 16) >> 16);
1999   }
2000   return LoadNode::Value(phase);
2001 }
2002 
2003 //=============================================================================
2004 //----------------------------LoadKlassNode::make------------------------------
2005 // Polymorphic factory method:
2006 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2007   // sanity check the alias category against the created node type
2008   const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2009   assert(adr_type != NULL, "expecting TypeKlassPtr");
2010 #ifdef _LP64
2011   if (adr_type->is_ptr_to_narrowklass()) {
2012     assert(UseCompressedClassPointers, "no compressed klasses");
2013     Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2014     return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2015   }
2016 #endif
2017   assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2018   return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2019 }
2020 
2021 //------------------------------Value------------------------------------------
2022 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2023   return klass_value_common(phase);
2024 }
2025 
2026 // In most cases, LoadKlassNode does not have the control input set. If the control
2027 // input is set, it must not be removed (by LoadNode::Ideal()).
2028 bool LoadKlassNode::can_remove_control() const {
2029   return false;
2030 }
2031 
2032 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2033   // Either input is TOP ==> the result is TOP
2034   const Type *t1 = phase->type( in(MemNode::Memory) );
2035   if (t1 == Type::TOP)  return Type::TOP;
2036   Node *adr = in(MemNode::Address);
2037   const Type *t2 = phase->type( adr );
2038   if (t2 == Type::TOP)  return Type::TOP;
2039   const TypePtr *tp = t2->is_ptr();
2040   if (TypePtr::above_centerline(tp->ptr()) ||
2041       tp->ptr() == TypePtr::Null)  return Type::TOP;
2042 
2043   // Return a more precise klass, if possible
2044   const TypeInstPtr *tinst = tp->isa_instptr();
2045   if (tinst != NULL) {
2046     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2047     int offset = tinst->offset();
2048     if (ik == phase->C->env()->Class_klass()
2049         && (offset == java_lang_Class::klass_offset_in_bytes() ||
2050             offset == java_lang_Class::array_klass_offset_in_bytes())) {
2051       // We are loading a special hidden field from a Class mirror object,
2052       // the field which points to the VM's Klass metaobject.
2053       ciType* t = tinst->java_mirror_type();
2054       // java_mirror_type returns non-null for compile-time Class constants.
2055       if (t != NULL) {
2056         // constant oop => constant klass
2057         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2058           if (t->is_void()) {
2059             // We cannot create a void array.  Since void is a primitive type return null
2060             // klass.  Users of this result need to do a null check on the returned klass.
2061             return TypePtr::NULL_PTR;
2062           }
2063           return TypeKlassPtr::make(ciArrayKlass::make(t));
2064         }
2065         if (!t->is_klass()) {
2066           // a primitive Class (e.g., int.class) has NULL for a klass field
2067           return TypePtr::NULL_PTR;
2068         }
2069         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2070         return TypeKlassPtr::make(t->as_klass());
2071       }
2072       // non-constant mirror, so we can't tell what's going on
2073     }
2074     if( !ik->is_loaded() )
2075       return _type;             // Bail out if not loaded
2076     if (offset == oopDesc::klass_offset_in_bytes()) {
2077       if (tinst->klass_is_exact()) {
2078         return TypeKlassPtr::make(ik);
2079       }
2080       // See if we can become precise: no subklasses and no interface
2081       // (Note:  We need to support verified interfaces.)
2082       if (!ik->is_interface() && !ik->has_subklass()) {
2083         //assert(!UseExactTypes, "this code should be useless with exact types");
2084         // Add a dependence; if any subclass added we need to recompile
2085         if (!ik->is_final()) {
2086           // %%% should use stronger assert_unique_concrete_subtype instead
2087           phase->C->dependencies()->assert_leaf_type(ik);
2088         }
2089         // Return precise klass
2090         return TypeKlassPtr::make(ik);
2091       }
2092 
2093       // Return root of possible klass
2094       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2095     }
2096   }
2097 
2098   // Check for loading klass from an array
2099   const TypeAryPtr *tary = tp->isa_aryptr();
2100   if( tary != NULL ) {
2101     ciKlass *tary_klass = tary->klass();
2102     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
2103         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2104       if (tary->klass_is_exact()) {
2105         return TypeKlassPtr::make(tary_klass);
2106       }
2107       ciArrayKlass *ak = tary->klass()->as_array_klass();
2108       // If the klass is an object array, we defer the question to the
2109       // array component klass.
2110       if( ak->is_obj_array_klass() ) {
2111         assert( ak->is_loaded(), "" );
2112         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2113         if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2114           ciInstanceKlass* ik = base_k->as_instance_klass();
2115           // See if we can become precise: no subklasses and no interface
2116           if (!ik->is_interface() && !ik->has_subklass()) {
2117             //assert(!UseExactTypes, "this code should be useless with exact types");
2118             // Add a dependence; if any subclass added we need to recompile
2119             if (!ik->is_final()) {
2120               phase->C->dependencies()->assert_leaf_type(ik);
2121             }
2122             // Return precise array klass
2123             return TypeKlassPtr::make(ak);
2124           }
2125         }
2126         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2127       } else {                  // Found a type-array?
2128         //assert(!UseExactTypes, "this code should be useless with exact types");
2129         assert( ak->is_type_array_klass(), "" );
2130         return TypeKlassPtr::make(ak); // These are always precise
2131       }
2132     }
2133   }
2134 
2135   // Check for loading klass from an array klass
2136   const TypeKlassPtr *tkls = tp->isa_klassptr();
2137   if (tkls != NULL && !StressReflectiveCode) {
2138     ciKlass* klass = tkls->klass();
2139     if( !klass->is_loaded() )
2140       return _type;             // Bail out if not loaded
2141     if( klass->is_obj_array_klass() &&
2142         tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2143       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2144       // // Always returning precise element type is incorrect,
2145       // // e.g., element type could be object and array may contain strings
2146       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2147 
2148       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2149       // according to the element type's subclassing.
2150       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2151     }
2152     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2153         tkls->offset() == in_bytes(Klass::super_offset())) {
2154       ciKlass* sup = klass->as_instance_klass()->super();
2155       // The field is Klass::_super.  Return its (constant) value.
2156       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2157       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2158     }
2159   }
2160 
2161   // Bailout case
2162   return LoadNode::Value(phase);
2163 }
2164 
2165 //------------------------------Identity---------------------------------------
2166 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2167 // Also feed through the klass in Allocate(...klass...)._klass.
2168 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2169   return klass_identity_common(phase);
2170 }
2171 
2172 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2173   Node* x = LoadNode::Identity(phase);
2174   if (x != this)  return x;
2175 
2176   // Take apart the address into an oop and and offset.
2177   // Return 'this' if we cannot.
2178   Node*    adr    = in(MemNode::Address);
2179   intptr_t offset = 0;
2180   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2181   if (base == NULL)     return this;
2182   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2183   if (toop == NULL)     return this;
2184 
2185   // We can fetch the klass directly through an AllocateNode.
2186   // This works even if the klass is not constant (clone or newArray).
2187   if (offset == oopDesc::klass_offset_in_bytes()) {
2188     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2189     if (allocated_klass != NULL) {
2190       return allocated_klass;
2191     }
2192   }
2193 
2194   // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2195   // See inline_native_Class_query for occurrences of these patterns.
2196   // Java Example:  x.getClass().isAssignableFrom(y)
2197   //
2198   // This improves reflective code, often making the Class
2199   // mirror go completely dead.  (Current exception:  Class
2200   // mirrors may appear in debug info, but we could clean them out by
2201   // introducing a new debug info operator for Klass*.java_mirror).
2202   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2203       && offset == java_lang_Class::klass_offset_in_bytes()) {
2204     // We are loading a special hidden field from a Class mirror,
2205     // the field which points to its Klass or ArrayKlass metaobject.
2206     if (base->is_Load()) {
2207       Node* adr2 = base->in(MemNode::Address);
2208       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2209       if (tkls != NULL && !tkls->empty()
2210           && (tkls->klass()->is_instance_klass() ||
2211               tkls->klass()->is_array_klass())
2212           && adr2->is_AddP()
2213           ) {
2214         int mirror_field = in_bytes(Klass::java_mirror_offset());
2215         if (tkls->offset() == mirror_field) {
2216           return adr2->in(AddPNode::Base);
2217         }
2218       }
2219     }
2220   }
2221 
2222   return this;
2223 }
2224 
2225 
2226 //------------------------------Value------------------------------------------
2227 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2228   const Type *t = klass_value_common(phase);
2229   if (t == Type::TOP)
2230     return t;
2231 
2232   return t->make_narrowklass();
2233 }
2234 
2235 //------------------------------Identity---------------------------------------
2236 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2237 // Also feed through the klass in Allocate(...klass...)._klass.
2238 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2239   Node *x = klass_identity_common(phase);
2240 
2241   const Type *t = phase->type( x );
2242   if( t == Type::TOP ) return x;
2243   if( t->isa_narrowklass()) return x;
2244   assert (!t->isa_narrowoop(), "no narrow oop here");
2245 
2246   return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2247 }
2248 
2249 //------------------------------Value-----------------------------------------
2250 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2251   // Either input is TOP ==> the result is TOP
2252   const Type *t1 = phase->type( in(MemNode::Memory) );
2253   if( t1 == Type::TOP ) return Type::TOP;
2254   Node *adr = in(MemNode::Address);
2255   const Type *t2 = phase->type( adr );
2256   if( t2 == Type::TOP ) return Type::TOP;
2257   const TypePtr *tp = t2->is_ptr();
2258   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2259   const TypeAryPtr *tap = tp->isa_aryptr();
2260   if( !tap ) return _type;
2261   return tap->size();
2262 }
2263 
2264 //-------------------------------Ideal---------------------------------------
2265 // Feed through the length in AllocateArray(...length...)._length.
2266 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2267   Node* p = MemNode::Ideal_common(phase, can_reshape);
2268   if (p)  return (p == NodeSentinel) ? NULL : p;
2269 
2270   // Take apart the address into an oop and and offset.
2271   // Return 'this' if we cannot.
2272   Node*    adr    = in(MemNode::Address);
2273   intptr_t offset = 0;
2274   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2275   if (base == NULL)     return NULL;
2276   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2277   if (tary == NULL)     return NULL;
2278 
2279   // We can fetch the length directly through an AllocateArrayNode.
2280   // This works even if the length is not constant (clone or newArray).
2281   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2282     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2283     if (alloc != NULL) {
2284       Node* allocated_length = alloc->Ideal_length();
2285       Node* len = alloc->make_ideal_length(tary, phase);
2286       if (allocated_length != len) {
2287         // New CastII improves on this.
2288         return len;
2289       }
2290     }
2291   }
2292 
2293   return NULL;
2294 }
2295 
2296 //------------------------------Identity---------------------------------------
2297 // Feed through the length in AllocateArray(...length...)._length.
2298 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2299   Node* x = LoadINode::Identity(phase);
2300   if (x != this)  return x;
2301 
2302   // Take apart the address into an oop and and offset.
2303   // Return 'this' if we cannot.
2304   Node*    adr    = in(MemNode::Address);
2305   intptr_t offset = 0;
2306   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2307   if (base == NULL)     return this;
2308   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2309   if (tary == NULL)     return this;
2310 
2311   // We can fetch the length directly through an AllocateArrayNode.
2312   // This works even if the length is not constant (clone or newArray).
2313   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2314     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2315     if (alloc != NULL) {
2316       Node* allocated_length = alloc->Ideal_length();
2317       // Do not allow make_ideal_length to allocate a CastII node.
2318       Node* len = alloc->make_ideal_length(tary, phase, false);
2319       if (allocated_length == len) {
2320         // Return allocated_length only if it would not be improved by a CastII.
2321         return allocated_length;
2322       }
2323     }
2324   }
2325 
2326   return this;
2327 
2328 }
2329 
2330 //=============================================================================
2331 //---------------------------StoreNode::make-----------------------------------
2332 // Polymorphic factory method:
2333 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2334   assert((mo == unordered || mo == release), "unexpected");
2335   Compile* C = gvn.C;
2336   assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2337          ctl != NULL, "raw memory operations should have control edge");
2338 
2339   switch (bt) {
2340   case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2341   case T_BYTE:    return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2342   case T_INT:     return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2343   case T_CHAR:
2344   case T_SHORT:   return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2345   case T_LONG:    return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2346   case T_FLOAT:   return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2347   case T_DOUBLE:  return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2348   case T_METADATA:
2349   case T_ADDRESS:
2350   case T_OBJECT:
2351 #ifdef _LP64
2352     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2353       val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2354       return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2355     } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2356                (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2357                 adr->bottom_type()->isa_rawptr())) {
2358       val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2359       return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2360     }
2361 #endif
2362     {
2363       return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2364     }
2365   }
2366   ShouldNotReachHere();
2367   return (StoreNode*)NULL;
2368 }
2369 
2370 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2371   bool require_atomic = true;
2372   return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2373 }
2374 
2375 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2376   bool require_atomic = true;
2377   return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2378 }
2379 
2380 
2381 //--------------------------bottom_type----------------------------------------
2382 const Type *StoreNode::bottom_type() const {
2383   return Type::MEMORY;
2384 }
2385 
2386 //------------------------------hash-------------------------------------------
2387 uint StoreNode::hash() const {
2388   // unroll addition of interesting fields
2389   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2390 
2391   // Since they are not commoned, do not hash them:
2392   return NO_HASH;
2393 }
2394 
2395 //------------------------------Ideal------------------------------------------
2396 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2397 // When a store immediately follows a relevant allocation/initialization,
2398 // try to capture it into the initialization, or hoist it above.
2399 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2400   Node* p = MemNode::Ideal_common(phase, can_reshape);
2401   if (p)  return (p == NodeSentinel) ? NULL : p;
2402 
2403   Node* mem     = in(MemNode::Memory);
2404   Node* address = in(MemNode::Address);
2405   // Back-to-back stores to same address?  Fold em up.  Generally
2406   // unsafe if I have intervening uses...  Also disallowed for StoreCM
2407   // since they must follow each StoreP operation.  Redundant StoreCMs
2408   // are eliminated just before matching in final_graph_reshape.
2409   {
2410     Node* st = mem;
2411     // If Store 'st' has more than one use, we cannot fold 'st' away.
2412     // For example, 'st' might be the final state at a conditional
2413     // return.  Or, 'st' might be used by some node which is live at
2414     // the same time 'st' is live, which might be unschedulable.  So,
2415     // require exactly ONE user until such time as we clone 'mem' for
2416     // each of 'mem's uses (thus making the exactly-1-user-rule hold
2417     // true).
2418     while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2419       // Looking at a dead closed cycle of memory?
2420       assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2421       assert(Opcode() == st->Opcode() ||
2422              st->Opcode() == Op_StoreVector ||
2423              Opcode() == Op_StoreVector ||
2424              phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2425              (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2426              (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2427              "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2428 
2429       if (st->in(MemNode::Address)->eqv_uncast(address) &&
2430           st->as_Store()->memory_size() <= this->memory_size()) {
2431         Node* use = st->raw_out(0);
2432         phase->igvn_rehash_node_delayed(use);
2433         if (can_reshape) {
2434           use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2435         } else {
2436           // It's OK to do this in the parser, since DU info is always accurate,
2437           // and the parser always refers to nodes via SafePointNode maps.
2438           use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2439         }
2440         return this;
2441       }
2442       st = st->in(MemNode::Memory);
2443     }
2444   }
2445 
2446 
2447   // Capture an unaliased, unconditional, simple store into an initializer.
2448   // Or, if it is independent of the allocation, hoist it above the allocation.
2449   if (ReduceFieldZeroing && /*can_reshape &&*/
2450       mem->is_Proj() && mem->in(0)->is_Initialize()) {
2451     InitializeNode* init = mem->in(0)->as_Initialize();
2452     intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2453     if (offset > 0) {
2454       Node* moved = init->capture_store(this, offset, phase, can_reshape);
2455       // If the InitializeNode captured me, it made a raw copy of me,
2456       // and I need to disappear.
2457       if (moved != NULL) {
2458         // %%% hack to ensure that Ideal returns a new node:
2459         mem = MergeMemNode::make(mem);
2460         return mem;             // fold me away
2461       }
2462     }
2463   }
2464 
2465   return NULL;                  // No further progress
2466 }
2467 
2468 //------------------------------Value-----------------------------------------
2469 const Type* StoreNode::Value(PhaseGVN* phase) const {
2470   // Either input is TOP ==> the result is TOP
2471   const Type *t1 = phase->type( in(MemNode::Memory) );
2472   if( t1 == Type::TOP ) return Type::TOP;
2473   const Type *t2 = phase->type( in(MemNode::Address) );
2474   if( t2 == Type::TOP ) return Type::TOP;
2475   const Type *t3 = phase->type( in(MemNode::ValueIn) );
2476   if( t3 == Type::TOP ) return Type::TOP;
2477   return Type::MEMORY;
2478 }
2479 
2480 //------------------------------Identity---------------------------------------
2481 // Remove redundant stores:
2482 //   Store(m, p, Load(m, p)) changes to m.
2483 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2484 Node* StoreNode::Identity(PhaseGVN* phase) {
2485   Node* mem = in(MemNode::Memory);
2486   Node* adr = in(MemNode::Address);
2487   Node* val = in(MemNode::ValueIn);
2488 
2489   // Load then Store?  Then the Store is useless
2490   if (val->is_Load() &&
2491       val->in(MemNode::Address)->eqv_uncast(adr) &&
2492       val->in(MemNode::Memory )->eqv_uncast(mem) &&
2493       val->as_Load()->store_Opcode() == Opcode()) {
2494     return mem;
2495   }
2496 
2497   // Two stores in a row of the same value?
2498   if (mem->is_Store() &&
2499       mem->in(MemNode::Address)->eqv_uncast(adr) &&
2500       mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2501       mem->Opcode() == Opcode()) {
2502     return mem;
2503   }
2504 
2505   // Store of zero anywhere into a freshly-allocated object?
2506   // Then the store is useless.
2507   // (It must already have been captured by the InitializeNode.)
2508   if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2509     // a newly allocated object is already all-zeroes everywhere
2510     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2511       return mem;
2512     }
2513 
2514     // the store may also apply to zero-bits in an earlier object
2515     Node* prev_mem = find_previous_store(phase);
2516     // Steps (a), (b):  Walk past independent stores to find an exact match.
2517     if (prev_mem != NULL) {
2518       Node* prev_val = can_see_stored_value(prev_mem, phase);
2519       if (prev_val != NULL && phase->eqv(prev_val, val)) {
2520         // prev_val and val might differ by a cast; it would be good
2521         // to keep the more informative of the two.
2522         return mem;
2523       }
2524     }
2525   }
2526 
2527   return this;
2528 }
2529 
2530 //------------------------------match_edge-------------------------------------
2531 // Do we Match on this edge index or not?  Match only memory & value
2532 uint StoreNode::match_edge(uint idx) const {
2533   return idx == MemNode::Address || idx == MemNode::ValueIn;
2534 }
2535 
2536 //------------------------------cmp--------------------------------------------
2537 // Do not common stores up together.  They generally have to be split
2538 // back up anyways, so do not bother.
2539 uint StoreNode::cmp( const Node &n ) const {
2540   return (&n == this);          // Always fail except on self
2541 }
2542 
2543 //------------------------------Ideal_masked_input-----------------------------
2544 // Check for a useless mask before a partial-word store
2545 // (StoreB ... (AndI valIn conIa) )
2546 // If (conIa & mask == mask) this simplifies to
2547 // (StoreB ... (valIn) )
2548 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2549   Node *val = in(MemNode::ValueIn);
2550   if( val->Opcode() == Op_AndI ) {
2551     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2552     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2553       set_req(MemNode::ValueIn, val->in(1));
2554       return this;
2555     }
2556   }
2557   return NULL;
2558 }
2559 
2560 
2561 //------------------------------Ideal_sign_extended_input----------------------
2562 // Check for useless sign-extension before a partial-word store
2563 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2564 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
2565 // (StoreB ... (valIn) )
2566 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2567   Node *val = in(MemNode::ValueIn);
2568   if( val->Opcode() == Op_RShiftI ) {
2569     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2570     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2571       Node *shl = val->in(1);
2572       if( shl->Opcode() == Op_LShiftI ) {
2573         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2574         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2575           set_req(MemNode::ValueIn, shl->in(1));
2576           return this;
2577         }
2578       }
2579     }
2580   }
2581   return NULL;
2582 }
2583 
2584 //------------------------------value_never_loaded-----------------------------------
2585 // Determine whether there are any possible loads of the value stored.
2586 // For simplicity, we actually check if there are any loads from the
2587 // address stored to, not just for loads of the value stored by this node.
2588 //
2589 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2590   Node *adr = in(Address);
2591   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2592   if (adr_oop == NULL)
2593     return false;
2594   if (!adr_oop->is_known_instance_field())
2595     return false; // if not a distinct instance, there may be aliases of the address
2596   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2597     Node *use = adr->fast_out(i);
2598     if (use->is_Load() || use->is_LoadStore()) {
2599       return false;
2600     }
2601   }
2602   return true;
2603 }
2604 
2605 //=============================================================================
2606 //------------------------------Ideal------------------------------------------
2607 // If the store is from an AND mask that leaves the low bits untouched, then
2608 // we can skip the AND operation.  If the store is from a sign-extension
2609 // (a left shift, then right shift) we can skip both.
2610 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2611   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2612   if( progress != NULL ) return progress;
2613 
2614   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2615   if( progress != NULL ) return progress;
2616 
2617   // Finally check the default case
2618   return StoreNode::Ideal(phase, can_reshape);
2619 }
2620 
2621 //=============================================================================
2622 //------------------------------Ideal------------------------------------------
2623 // If the store is from an AND mask that leaves the low bits untouched, then
2624 // we can skip the AND operation
2625 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2626   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2627   if( progress != NULL ) return progress;
2628 
2629   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2630   if( progress != NULL ) return progress;
2631 
2632   // Finally check the default case
2633   return StoreNode::Ideal(phase, can_reshape);
2634 }
2635 
2636 //=============================================================================
2637 //------------------------------Identity---------------------------------------
2638 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2639   // No need to card mark when storing a null ptr
2640   Node* my_store = in(MemNode::OopStore);
2641   if (my_store->is_Store()) {
2642     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2643     if( t1 == TypePtr::NULL_PTR ) {
2644       return in(MemNode::Memory);
2645     }
2646   }
2647   return this;
2648 }
2649 
2650 //=============================================================================
2651 //------------------------------Ideal---------------------------------------
2652 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2653   Node* progress = StoreNode::Ideal(phase, can_reshape);
2654   if (progress != NULL) return progress;
2655 
2656   Node* my_store = in(MemNode::OopStore);
2657   if (my_store->is_MergeMem()) {
2658     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2659     set_req(MemNode::OopStore, mem);
2660     return this;
2661   }
2662 
2663   return NULL;
2664 }
2665 
2666 //------------------------------Value-----------------------------------------
2667 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2668   // Either input is TOP ==> the result is TOP
2669   const Type *t = phase->type( in(MemNode::Memory) );
2670   if( t == Type::TOP ) return Type::TOP;
2671   t = phase->type( in(MemNode::Address) );
2672   if( t == Type::TOP ) return Type::TOP;
2673   t = phase->type( in(MemNode::ValueIn) );
2674   if( t == Type::TOP ) return Type::TOP;
2675   // If extra input is TOP ==> the result is TOP
2676   t = phase->type( in(MemNode::OopStore) );
2677   if( t == Type::TOP ) return Type::TOP;
2678 
2679   return StoreNode::Value( phase );
2680 }
2681 
2682 
2683 //=============================================================================
2684 //----------------------------------SCMemProjNode------------------------------
2685 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2686 {
2687   return bottom_type();
2688 }
2689 
2690 //=============================================================================
2691 //----------------------------------LoadStoreNode------------------------------
2692 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2693   : Node(required),
2694     _type(rt),
2695     _adr_type(at)
2696 {
2697   init_req(MemNode::Control, c  );
2698   init_req(MemNode::Memory , mem);
2699   init_req(MemNode::Address, adr);
2700   init_req(MemNode::ValueIn, val);
2701   init_class_id(Class_LoadStore);
2702 }
2703 
2704 uint LoadStoreNode::ideal_reg() const {
2705   return _type->ideal_reg();
2706 }
2707 
2708 bool LoadStoreNode::result_not_used() const {
2709   for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2710     Node *x = fast_out(i);
2711     if (x->Opcode() == Op_SCMemProj) continue;
2712     return false;
2713   }
2714   return true;
2715 }
2716 
2717 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2718 
2719 //=============================================================================
2720 //----------------------------------LoadStoreConditionalNode--------------------
2721 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2722   init_req(ExpectedIn, ex );
2723 }
2724 
2725 //=============================================================================
2726 //-------------------------------adr_type--------------------------------------
2727 const TypePtr* ClearArrayNode::adr_type() const {
2728   Node *adr = in(3);
2729   if (adr == NULL)  return NULL; // node is dead
2730   return MemNode::calculate_adr_type(adr->bottom_type());
2731 }
2732 
2733 //------------------------------match_edge-------------------------------------
2734 // Do we Match on this edge index or not?  Do not match memory
2735 uint ClearArrayNode::match_edge(uint idx) const {
2736   return idx > 1;
2737 }
2738 
2739 //------------------------------Identity---------------------------------------
2740 // Clearing a zero length array does nothing
2741 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2742   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
2743 }
2744 
2745 //------------------------------Idealize---------------------------------------
2746 // Clearing a short array is faster with stores
2747 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2748   // Already know this is a large node, do not try to ideal it
2749   if (!IdealizeClearArrayNode || _is_large) return NULL;
2750 
2751   const int unit = BytesPerLong;
2752   const TypeX* t = phase->type(in(2))->isa_intptr_t();
2753   if (!t)  return NULL;
2754   if (!t->is_con())  return NULL;
2755   intptr_t raw_count = t->get_con();
2756   intptr_t size = raw_count;
2757   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2758   // Clearing nothing uses the Identity call.
2759   // Negative clears are possible on dead ClearArrays
2760   // (see jck test stmt114.stmt11402.val).
2761   if (size <= 0 || size % unit != 0)  return NULL;
2762   intptr_t count = size / unit;
2763   // Length too long; communicate this to matchers and assemblers.
2764   // Assemblers are responsible to produce fast hardware clears for it.
2765   if (size > InitArrayShortSize) {
2766     return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2767   }
2768   Node *mem = in(1);
2769   if( phase->type(mem)==Type::TOP ) return NULL;
2770   Node *adr = in(3);
2771   const Type* at = phase->type(adr);
2772   if( at==Type::TOP ) return NULL;
2773   const TypePtr* atp = at->isa_ptr();
2774   // adjust atp to be the correct array element address type
2775   if (atp == NULL)  atp = TypePtr::BOTTOM;
2776   else              atp = atp->add_offset(Type::OffsetBot);
2777   // Get base for derived pointer purposes
2778   if( adr->Opcode() != Op_AddP ) Unimplemented();
2779   Node *base = adr->in(1);
2780 
2781   Node *zero = phase->makecon(TypeLong::ZERO);
2782   Node *off  = phase->MakeConX(BytesPerLong);
2783   mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2784   count--;
2785   while( count-- ) {
2786     mem = phase->transform(mem);
2787     adr = phase->transform(new AddPNode(base,adr,off));
2788     mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2789   }
2790   return mem;
2791 }
2792 
2793 //----------------------------step_through----------------------------------
2794 // Return allocation input memory edge if it is different instance
2795 // or itself if it is the one we are looking for.
2796 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2797   Node* n = *np;
2798   assert(n->is_ClearArray(), "sanity");
2799   intptr_t offset;
2800   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2801   // This method is called only before Allocate nodes are expanded
2802   // during macro nodes expansion. Before that ClearArray nodes are
2803   // only generated in PhaseMacroExpand::generate_arraycopy() (before
2804   // Allocate nodes are expanded) which follows allocations.
2805   assert(alloc != NULL, "should have allocation");
2806   if (alloc->_idx == instance_id) {
2807     // Can not bypass initialization of the instance we are looking for.
2808     return false;
2809   }
2810   // Otherwise skip it.
2811   InitializeNode* init = alloc->initialization();
2812   if (init != NULL)
2813     *np = init->in(TypeFunc::Memory);
2814   else
2815     *np = alloc->in(TypeFunc::Memory);
2816   return true;
2817 }
2818 
2819 //----------------------------clear_memory-------------------------------------
2820 // Generate code to initialize object storage to zero.
2821 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2822                                    intptr_t start_offset,
2823                                    Node* end_offset,
2824                                    PhaseGVN* phase) {
2825   intptr_t offset = start_offset;
2826 
2827   int unit = BytesPerLong;
2828   if ((offset % unit) != 0) {
2829     Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2830     adr = phase->transform(adr);
2831     const TypePtr* atp = TypeRawPtr::BOTTOM;
2832     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2833     mem = phase->transform(mem);
2834     offset += BytesPerInt;
2835   }
2836   assert((offset % unit) == 0, "");
2837 
2838   // Initialize the remaining stuff, if any, with a ClearArray.
2839   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2840 }
2841 
2842 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2843                                    Node* start_offset,
2844                                    Node* end_offset,
2845                                    PhaseGVN* phase) {
2846   if (start_offset == end_offset) {
2847     // nothing to do
2848     return mem;
2849   }
2850 
2851   int unit = BytesPerLong;
2852   Node* zbase = start_offset;
2853   Node* zend  = end_offset;
2854 
2855   // Scale to the unit required by the CPU:
2856   if (!Matcher::init_array_count_is_in_bytes) {
2857     Node* shift = phase->intcon(exact_log2(unit));
2858     zbase = phase->transform(new URShiftXNode(zbase, shift) );
2859     zend  = phase->transform(new URShiftXNode(zend,  shift) );
2860   }
2861 
2862   // Bulk clear double-words
2863   Node* zsize = phase->transform(new SubXNode(zend, zbase) );
2864   Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
2865   mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
2866   return phase->transform(mem);
2867 }
2868 
2869 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2870                                    intptr_t start_offset,
2871                                    intptr_t end_offset,
2872                                    PhaseGVN* phase) {
2873   if (start_offset == end_offset) {
2874     // nothing to do
2875     return mem;
2876   }
2877 
2878   assert((end_offset % BytesPerInt) == 0, "odd end offset");
2879   intptr_t done_offset = end_offset;
2880   if ((done_offset % BytesPerLong) != 0) {
2881     done_offset -= BytesPerInt;
2882   }
2883   if (done_offset > start_offset) {
2884     mem = clear_memory(ctl, mem, dest,
2885                        start_offset, phase->MakeConX(done_offset), phase);
2886   }
2887   if (done_offset < end_offset) { // emit the final 32-bit store
2888     Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
2889     adr = phase->transform(adr);
2890     const TypePtr* atp = TypeRawPtr::BOTTOM;
2891     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2892     mem = phase->transform(mem);
2893     done_offset += BytesPerInt;
2894   }
2895   assert(done_offset == end_offset, "");
2896   return mem;
2897 }
2898 
2899 //=============================================================================
2900 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2901   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2902     _adr_type(C->get_adr_type(alias_idx))
2903 {
2904   init_class_id(Class_MemBar);
2905   Node* top = C->top();
2906   init_req(TypeFunc::I_O,top);
2907   init_req(TypeFunc::FramePtr,top);
2908   init_req(TypeFunc::ReturnAdr,top);
2909   if (precedent != NULL)
2910     init_req(TypeFunc::Parms, precedent);
2911 }
2912 
2913 //------------------------------cmp--------------------------------------------
2914 uint MemBarNode::hash() const { return NO_HASH; }
2915 uint MemBarNode::cmp( const Node &n ) const {
2916   return (&n == this);          // Always fail except on self
2917 }
2918 
2919 //------------------------------make-------------------------------------------
2920 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2921   switch (opcode) {
2922   case Op_MemBarAcquire:     return new MemBarAcquireNode(C, atp, pn);
2923   case Op_LoadFence:         return new LoadFenceNode(C, atp, pn);
2924   case Op_MemBarRelease:     return new MemBarReleaseNode(C, atp, pn);
2925   case Op_StoreFence:        return new StoreFenceNode(C, atp, pn);
2926   case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
2927   case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
2928   case Op_MemBarVolatile:    return new MemBarVolatileNode(C, atp, pn);
2929   case Op_MemBarCPUOrder:    return new MemBarCPUOrderNode(C, atp, pn);
2930   case Op_OnSpinWait:        return new OnSpinWaitNode(C, atp, pn);
2931   case Op_Initialize:        return new InitializeNode(C, atp, pn);
2932   case Op_MemBarStoreStore:  return new MemBarStoreStoreNode(C, atp, pn);
2933   default: ShouldNotReachHere(); return NULL;
2934   }
2935 }
2936 
2937 //------------------------------Ideal------------------------------------------
2938 // Return a node which is more "ideal" than the current node.  Strip out
2939 // control copies
2940 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2941   if (remove_dead_region(phase, can_reshape)) return this;
2942   // Don't bother trying to transform a dead node
2943   if (in(0) && in(0)->is_top()) {
2944     return NULL;
2945   }
2946 
2947   bool progress = false;
2948   // Eliminate volatile MemBars for scalar replaced objects.
2949   if (can_reshape && req() == (Precedent+1)) {
2950     bool eliminate = false;
2951     int opc = Opcode();
2952     if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2953       // Volatile field loads and stores.
2954       Node* my_mem = in(MemBarNode::Precedent);
2955       // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2956       if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2957         // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2958         // replace this Precedent (decodeN) with the Load instead.
2959         if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1))  {
2960           Node* load_node = my_mem->in(1);
2961           set_req(MemBarNode::Precedent, load_node);
2962           phase->is_IterGVN()->_worklist.push(my_mem);
2963           my_mem = load_node;
2964         } else {
2965           assert(my_mem->unique_out() == this, "sanity");
2966           del_req(Precedent);
2967           phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2968           my_mem = NULL;
2969         }
2970         progress = true;
2971       }
2972       if (my_mem != NULL && my_mem->is_Mem()) {
2973         const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2974         // Check for scalar replaced object reference.
2975         if( t_oop != NULL && t_oop->is_known_instance_field() &&
2976             t_oop->offset() != Type::OffsetBot &&
2977             t_oop->offset() != Type::OffsetTop) {
2978           eliminate = true;
2979         }
2980       }
2981     } else if (opc == Op_MemBarRelease) {
2982       // Final field stores.
2983       Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2984       if ((alloc != NULL) && alloc->is_Allocate() &&
2985           alloc->as_Allocate()->does_not_escape_thread()) {
2986         // The allocated object does not escape.
2987         eliminate = true;
2988       }
2989     }
2990     if (eliminate) {
2991       // Replace MemBar projections by its inputs.
2992       PhaseIterGVN* igvn = phase->is_IterGVN();
2993       igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2994       igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2995       // Must return either the original node (now dead) or a new node
2996       // (Do not return a top here, since that would break the uniqueness of top.)
2997       return new ConINode(TypeInt::ZERO);
2998     }
2999   }
3000   return progress ? this : NULL;
3001 }
3002 
3003 //------------------------------Value------------------------------------------
3004 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3005   if( !in(0) ) return Type::TOP;
3006   if( phase->type(in(0)) == Type::TOP )
3007     return Type::TOP;
3008   return TypeTuple::MEMBAR;
3009 }
3010 
3011 //------------------------------match------------------------------------------
3012 // Construct projections for memory.
3013 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3014   switch (proj->_con) {
3015   case TypeFunc::Control:
3016   case TypeFunc::Memory:
3017     return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3018   }
3019   ShouldNotReachHere();
3020   return NULL;
3021 }
3022 
3023 //===========================InitializeNode====================================
3024 // SUMMARY:
3025 // This node acts as a memory barrier on raw memory, after some raw stores.
3026 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3027 // The Initialize can 'capture' suitably constrained stores as raw inits.
3028 // It can coalesce related raw stores into larger units (called 'tiles').
3029 // It can avoid zeroing new storage for memory units which have raw inits.
3030 // At macro-expansion, it is marked 'complete', and does not optimize further.
3031 //
3032 // EXAMPLE:
3033 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3034 //   ctl = incoming control; mem* = incoming memory
3035 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
3036 // First allocate uninitialized memory and fill in the header:
3037 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
3038 //   ctl := alloc.Control; mem* := alloc.Memory*
3039 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
3040 // Then initialize to zero the non-header parts of the raw memory block:
3041 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3042 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3043 // After the initialize node executes, the object is ready for service:
3044 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3045 // Suppose its body is immediately initialized as {1,2}:
3046 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3047 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3048 //   mem.SLICE(#short[*]) := store2
3049 //
3050 // DETAILS:
3051 // An InitializeNode collects and isolates object initialization after
3052 // an AllocateNode and before the next possible safepoint.  As a
3053 // memory barrier (MemBarNode), it keeps critical stores from drifting
3054 // down past any safepoint or any publication of the allocation.
3055 // Before this barrier, a newly-allocated object may have uninitialized bits.
3056 // After this barrier, it may be treated as a real oop, and GC is allowed.
3057 //
3058 // The semantics of the InitializeNode include an implicit zeroing of
3059 // the new object from object header to the end of the object.
3060 // (The object header and end are determined by the AllocateNode.)
3061 //
3062 // Certain stores may be added as direct inputs to the InitializeNode.
3063 // These stores must update raw memory, and they must be to addresses
3064 // derived from the raw address produced by AllocateNode, and with
3065 // a constant offset.  They must be ordered by increasing offset.
3066 // The first one is at in(RawStores), the last at in(req()-1).
3067 // Unlike most memory operations, they are not linked in a chain,
3068 // but are displayed in parallel as users of the rawmem output of
3069 // the allocation.
3070 //
3071 // (See comments in InitializeNode::capture_store, which continue
3072 // the example given above.)
3073 //
3074 // When the associated Allocate is macro-expanded, the InitializeNode
3075 // may be rewritten to optimize collected stores.  A ClearArrayNode
3076 // may also be created at that point to represent any required zeroing.
3077 // The InitializeNode is then marked 'complete', prohibiting further
3078 // capturing of nearby memory operations.
3079 //
3080 // During macro-expansion, all captured initializations which store
3081 // constant values of 32 bits or smaller are coalesced (if advantageous)
3082 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
3083 // initialized in fewer memory operations.  Memory words which are
3084 // covered by neither tiles nor non-constant stores are pre-zeroed
3085 // by explicit stores of zero.  (The code shape happens to do all
3086 // zeroing first, then all other stores, with both sequences occurring
3087 // in order of ascending offsets.)
3088 //
3089 // Alternatively, code may be inserted between an AllocateNode and its
3090 // InitializeNode, to perform arbitrary initialization of the new object.
3091 // E.g., the object copying intrinsics insert complex data transfers here.
3092 // The initialization must then be marked as 'complete' disable the
3093 // built-in zeroing semantics and the collection of initializing stores.
3094 //
3095 // While an InitializeNode is incomplete, reads from the memory state
3096 // produced by it are optimizable if they match the control edge and
3097 // new oop address associated with the allocation/initialization.
3098 // They return a stored value (if the offset matches) or else zero.
3099 // A write to the memory state, if it matches control and address,
3100 // and if it is to a constant offset, may be 'captured' by the
3101 // InitializeNode.  It is cloned as a raw memory operation and rewired
3102 // inside the initialization, to the raw oop produced by the allocation.
3103 // Operations on addresses which are provably distinct (e.g., to
3104 // other AllocateNodes) are allowed to bypass the initialization.
3105 //
3106 // The effect of all this is to consolidate object initialization
3107 // (both arrays and non-arrays, both piecewise and bulk) into a
3108 // single location, where it can be optimized as a unit.
3109 //
3110 // Only stores with an offset less than TrackedInitializationLimit words
3111 // will be considered for capture by an InitializeNode.  This puts a
3112 // reasonable limit on the complexity of optimized initializations.
3113 
3114 //---------------------------InitializeNode------------------------------------
3115 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3116   : _is_complete(Incomplete), _does_not_escape(false),
3117     MemBarNode(C, adr_type, rawoop)
3118 {
3119   init_class_id(Class_Initialize);
3120 
3121   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3122   assert(in(RawAddress) == rawoop, "proper init");
3123   // Note:  allocation() can be NULL, for secondary initialization barriers
3124 }
3125 
3126 // Since this node is not matched, it will be processed by the
3127 // register allocator.  Declare that there are no constraints
3128 // on the allocation of the RawAddress edge.
3129 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3130   // This edge should be set to top, by the set_complete.  But be conservative.
3131   if (idx == InitializeNode::RawAddress)
3132     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3133   return RegMask::Empty;
3134 }
3135 
3136 Node* InitializeNode::memory(uint alias_idx) {
3137   Node* mem = in(Memory);
3138   if (mem->is_MergeMem()) {
3139     return mem->as_MergeMem()->memory_at(alias_idx);
3140   } else {
3141     // incoming raw memory is not split
3142     return mem;
3143   }
3144 }
3145 
3146 bool InitializeNode::is_non_zero() {
3147   if (is_complete())  return false;
3148   remove_extra_zeroes();
3149   return (req() > RawStores);
3150 }
3151 
3152 void InitializeNode::set_complete(PhaseGVN* phase) {
3153   assert(!is_complete(), "caller responsibility");
3154   _is_complete = Complete;
3155 
3156   // After this node is complete, it contains a bunch of
3157   // raw-memory initializations.  There is no need for
3158   // it to have anything to do with non-raw memory effects.
3159   // Therefore, tell all non-raw users to re-optimize themselves,
3160   // after skipping the memory effects of this initialization.
3161   PhaseIterGVN* igvn = phase->is_IterGVN();
3162   if (igvn)  igvn->add_users_to_worklist(this);
3163 }
3164 
3165 // convenience function
3166 // return false if the init contains any stores already
3167 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3168   InitializeNode* init = initialization();
3169   if (init == NULL || init->is_complete())  return false;
3170   init->remove_extra_zeroes();
3171   // for now, if this allocation has already collected any inits, bail:
3172   if (init->is_non_zero())  return false;
3173   init->set_complete(phase);
3174   return true;
3175 }
3176 
3177 void InitializeNode::remove_extra_zeroes() {
3178   if (req() == RawStores)  return;
3179   Node* zmem = zero_memory();
3180   uint fill = RawStores;
3181   for (uint i = fill; i < req(); i++) {
3182     Node* n = in(i);
3183     if (n->is_top() || n == zmem)  continue;  // skip
3184     if (fill < i)  set_req(fill, n);          // compact
3185     ++fill;
3186   }
3187   // delete any empty spaces created:
3188   while (fill < req()) {
3189     del_req(fill);
3190   }
3191 }
3192 
3193 // Helper for remembering which stores go with which offsets.
3194 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3195   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
3196   intptr_t offset = -1;
3197   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3198                                                phase, offset);
3199   if (base == NULL)     return -1;  // something is dead,
3200   if (offset < 0)       return -1;  //        dead, dead
3201   return offset;
3202 }
3203 
3204 // Helper for proving that an initialization expression is
3205 // "simple enough" to be folded into an object initialization.
3206 // Attempts to prove that a store's initial value 'n' can be captured
3207 // within the initialization without creating a vicious cycle, such as:
3208 //     { Foo p = new Foo(); p.next = p; }
3209 // True for constants and parameters and small combinations thereof.
3210 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3211   if (n == NULL)      return true;   // (can this really happen?)
3212   if (n->is_Proj())   n = n->in(0);
3213   if (n == this)      return false;  // found a cycle
3214   if (n->is_Con())    return true;
3215   if (n->is_Start())  return true;   // params, etc., are OK
3216   if (n->is_Root())   return true;   // even better
3217 
3218   Node* ctl = n->in(0);
3219   if (ctl != NULL && !ctl->is_top()) {
3220     if (ctl->is_Proj())  ctl = ctl->in(0);
3221     if (ctl == this)  return false;
3222 
3223     // If we already know that the enclosing memory op is pinned right after
3224     // the init, then any control flow that the store has picked up
3225     // must have preceded the init, or else be equal to the init.
3226     // Even after loop optimizations (which might change control edges)
3227     // a store is never pinned *before* the availability of its inputs.
3228     if (!MemNode::all_controls_dominate(n, this))
3229       return false;                  // failed to prove a good control
3230   }
3231 
3232   // Check data edges for possible dependencies on 'this'.
3233   if ((count += 1) > 20)  return false;  // complexity limit
3234   for (uint i = 1; i < n->req(); i++) {
3235     Node* m = n->in(i);
3236     if (m == NULL || m == n || m->is_top())  continue;
3237     uint first_i = n->find_edge(m);
3238     if (i != first_i)  continue;  // process duplicate edge just once
3239     if (!detect_init_independence(m, count)) {
3240       return false;
3241     }
3242   }
3243 
3244   return true;
3245 }
3246 
3247 // Here are all the checks a Store must pass before it can be moved into
3248 // an initialization.  Returns zero if a check fails.
3249 // On success, returns the (constant) offset to which the store applies,
3250 // within the initialized memory.
3251 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3252   const int FAIL = 0;
3253   if (st->is_unaligned_access()) {
3254     return FAIL;
3255   }
3256   if (st->req() != MemNode::ValueIn + 1)
3257     return FAIL;                // an inscrutable StoreNode (card mark?)
3258   Node* ctl = st->in(MemNode::Control);
3259   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3260     return FAIL;                // must be unconditional after the initialization
3261   Node* mem = st->in(MemNode::Memory);
3262   if (!(mem->is_Proj() && mem->in(0) == this))
3263     return FAIL;                // must not be preceded by other stores
3264   Node* adr = st->in(MemNode::Address);
3265   intptr_t offset;
3266   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3267   if (alloc == NULL)
3268     return FAIL;                // inscrutable address
3269   if (alloc != allocation())
3270     return FAIL;                // wrong allocation!  (store needs to float up)
3271   Node* val = st->in(MemNode::ValueIn);
3272   int complexity_count = 0;
3273   if (!detect_init_independence(val, complexity_count))
3274     return FAIL;                // stored value must be 'simple enough'
3275 
3276   // The Store can be captured only if nothing after the allocation
3277   // and before the Store is using the memory location that the store
3278   // overwrites.
3279   bool failed = false;
3280   // If is_complete_with_arraycopy() is true the shape of the graph is
3281   // well defined and is safe so no need for extra checks.
3282   if (!is_complete_with_arraycopy()) {
3283     // We are going to look at each use of the memory state following
3284     // the allocation to make sure nothing reads the memory that the
3285     // Store writes.
3286     const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3287     int alias_idx = phase->C->get_alias_index(t_adr);
3288     ResourceMark rm;
3289     Unique_Node_List mems;
3290     mems.push(mem);
3291     Node* unique_merge = NULL;
3292     for (uint next = 0; next < mems.size(); ++next) {
3293       Node *m  = mems.at(next);
3294       for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3295         Node *n = m->fast_out(j);
3296         if (n->outcnt() == 0) {
3297           continue;
3298         }
3299         if (n == st) {
3300           continue;
3301         } else if (n->in(0) != NULL && n->in(0) != ctl) {
3302           // If the control of this use is different from the control
3303           // of the Store which is right after the InitializeNode then
3304           // this node cannot be between the InitializeNode and the
3305           // Store.
3306           continue;
3307         } else if (n->is_MergeMem()) {
3308           if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3309             // We can hit a MergeMemNode (that will likely go away
3310             // later) that is a direct use of the memory state
3311             // following the InitializeNode on the same slice as the
3312             // store node that we'd like to capture. We need to check
3313             // the uses of the MergeMemNode.
3314             mems.push(n);
3315           }
3316         } else if (n->is_Mem()) {
3317           Node* other_adr = n->in(MemNode::Address);
3318           if (other_adr == adr) {
3319             failed = true;
3320             break;
3321           } else {
3322             const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3323             if (other_t_adr != NULL) {
3324               int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3325               if (other_alias_idx == alias_idx) {
3326                 // A load from the same memory slice as the store right
3327                 // after the InitializeNode. We check the control of the
3328                 // object/array that is loaded from. If it's the same as
3329                 // the store control then we cannot capture the store.
3330                 assert(!n->is_Store(), "2 stores to same slice on same control?");
3331                 Node* base = other_adr;
3332                 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3333                 base = base->in(AddPNode::Base);
3334                 if (base != NULL) {
3335                   base = base->uncast();
3336                   if (base->is_Proj() && base->in(0) == alloc) {
3337                     failed = true;
3338                     break;
3339                   }
3340                 }
3341               }
3342             }
3343           }
3344         } else {
3345           failed = true;
3346           break;
3347         }
3348       }
3349     }
3350   }
3351   if (failed) {
3352     if (!can_reshape) {
3353       // We decided we couldn't capture the store during parsing. We
3354       // should try again during the next IGVN once the graph is
3355       // cleaner.
3356       phase->C->record_for_igvn(st);
3357     }
3358     return FAIL;
3359   }
3360 
3361   return offset;                // success
3362 }
3363 
3364 // Find the captured store in(i) which corresponds to the range
3365 // [start..start+size) in the initialized object.
3366 // If there is one, return its index i.  If there isn't, return the
3367 // negative of the index where it should be inserted.
3368 // Return 0 if the queried range overlaps an initialization boundary
3369 // or if dead code is encountered.
3370 // If size_in_bytes is zero, do not bother with overlap checks.
3371 int InitializeNode::captured_store_insertion_point(intptr_t start,
3372                                                    int size_in_bytes,
3373                                                    PhaseTransform* phase) {
3374   const int FAIL = 0, MAX_STORE = BytesPerLong;
3375 
3376   if (is_complete())
3377     return FAIL;                // arraycopy got here first; punt
3378 
3379   assert(allocation() != NULL, "must be present");
3380 
3381   // no negatives, no header fields:
3382   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
3383 
3384   // after a certain size, we bail out on tracking all the stores:
3385   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3386   if (start >= ti_limit)  return FAIL;
3387 
3388   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3389     if (i >= limit)  return -(int)i; // not found; here is where to put it
3390 
3391     Node*    st     = in(i);
3392     intptr_t st_off = get_store_offset(st, phase);
3393     if (st_off < 0) {
3394       if (st != zero_memory()) {
3395         return FAIL;            // bail out if there is dead garbage
3396       }
3397     } else if (st_off > start) {
3398       // ...we are done, since stores are ordered
3399       if (st_off < start + size_in_bytes) {
3400         return FAIL;            // the next store overlaps
3401       }
3402       return -(int)i;           // not found; here is where to put it
3403     } else if (st_off < start) {
3404       if (size_in_bytes != 0 &&
3405           start < st_off + MAX_STORE &&
3406           start < st_off + st->as_Store()->memory_size()) {
3407         return FAIL;            // the previous store overlaps
3408       }
3409     } else {
3410       if (size_in_bytes != 0 &&
3411           st->as_Store()->memory_size() != size_in_bytes) {
3412         return FAIL;            // mismatched store size
3413       }
3414       return i;
3415     }
3416 
3417     ++i;
3418   }
3419 }
3420 
3421 // Look for a captured store which initializes at the offset 'start'
3422 // with the given size.  If there is no such store, and no other
3423 // initialization interferes, then return zero_memory (the memory
3424 // projection of the AllocateNode).
3425 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3426                                           PhaseTransform* phase) {
3427   assert(stores_are_sane(phase), "");
3428   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3429   if (i == 0) {
3430     return NULL;                // something is dead
3431   } else if (i < 0) {
3432     return zero_memory();       // just primordial zero bits here
3433   } else {
3434     Node* st = in(i);           // here is the store at this position
3435     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3436     return st;
3437   }
3438 }
3439 
3440 // Create, as a raw pointer, an address within my new object at 'offset'.
3441 Node* InitializeNode::make_raw_address(intptr_t offset,
3442                                        PhaseTransform* phase) {
3443   Node* addr = in(RawAddress);
3444   if (offset != 0) {
3445     Compile* C = phase->C;
3446     addr = phase->transform( new AddPNode(C->top(), addr,
3447                                                  phase->MakeConX(offset)) );
3448   }
3449   return addr;
3450 }
3451 
3452 // Clone the given store, converting it into a raw store
3453 // initializing a field or element of my new object.
3454 // Caller is responsible for retiring the original store,
3455 // with subsume_node or the like.
3456 //
3457 // From the example above InitializeNode::InitializeNode,
3458 // here are the old stores to be captured:
3459 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3460 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3461 //
3462 // Here is the changed code; note the extra edges on init:
3463 //   alloc = (Allocate ...)
3464 //   rawoop = alloc.RawAddress
3465 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3466 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3467 //   init = (Initialize alloc.Control alloc.Memory rawoop
3468 //                      rawstore1 rawstore2)
3469 //
3470 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3471                                     PhaseTransform* phase, bool can_reshape) {
3472   assert(stores_are_sane(phase), "");
3473 
3474   if (start < 0)  return NULL;
3475   assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3476 
3477   Compile* C = phase->C;
3478   int size_in_bytes = st->memory_size();
3479   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3480   if (i == 0)  return NULL;     // bail out
3481   Node* prev_mem = NULL;        // raw memory for the captured store
3482   if (i > 0) {
3483     prev_mem = in(i);           // there is a pre-existing store under this one
3484     set_req(i, C->top());       // temporarily disconnect it
3485     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3486   } else {
3487     i = -i;                     // no pre-existing store
3488     prev_mem = zero_memory();   // a slice of the newly allocated object
3489     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3490       set_req(--i, C->top());   // reuse this edge; it has been folded away
3491     else
3492       ins_req(i, C->top());     // build a new edge
3493   }
3494   Node* new_st = st->clone();
3495   new_st->set_req(MemNode::Control, in(Control));
3496   new_st->set_req(MemNode::Memory,  prev_mem);
3497   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3498   new_st = phase->transform(new_st);
3499 
3500   // At this point, new_st might have swallowed a pre-existing store
3501   // at the same offset, or perhaps new_st might have disappeared,
3502   // if it redundantly stored the same value (or zero to fresh memory).
3503 
3504   // In any case, wire it in:
3505   phase->igvn_rehash_node_delayed(this);
3506   set_req(i, new_st);
3507 
3508   // The caller may now kill the old guy.
3509   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3510   assert(check_st == new_st || check_st == NULL, "must be findable");
3511   assert(!is_complete(), "");
3512   return new_st;
3513 }
3514 
3515 static bool store_constant(jlong* tiles, int num_tiles,
3516                            intptr_t st_off, int st_size,
3517                            jlong con) {
3518   if ((st_off & (st_size-1)) != 0)
3519     return false;               // strange store offset (assume size==2**N)
3520   address addr = (address)tiles + st_off;
3521   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3522   switch (st_size) {
3523   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
3524   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
3525   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
3526   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
3527   default: return false;        // strange store size (detect size!=2**N here)
3528   }
3529   return true;                  // return success to caller
3530 }
3531 
3532 // Coalesce subword constants into int constants and possibly
3533 // into long constants.  The goal, if the CPU permits,
3534 // is to initialize the object with a small number of 64-bit tiles.
3535 // Also, convert floating-point constants to bit patterns.
3536 // Non-constants are not relevant to this pass.
3537 //
3538 // In terms of the running example on InitializeNode::InitializeNode
3539 // and InitializeNode::capture_store, here is the transformation
3540 // of rawstore1 and rawstore2 into rawstore12:
3541 //   alloc = (Allocate ...)
3542 //   rawoop = alloc.RawAddress
3543 //   tile12 = 0x00010002
3544 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3545 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3546 //
3547 void
3548 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3549                                         Node* size_in_bytes,
3550                                         PhaseGVN* phase) {
3551   Compile* C = phase->C;
3552 
3553   assert(stores_are_sane(phase), "");
3554   // Note:  After this pass, they are not completely sane,
3555   // since there may be some overlaps.
3556 
3557   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3558 
3559   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3560   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3561   size_limit = MIN2(size_limit, ti_limit);
3562   size_limit = align_size_up(size_limit, BytesPerLong);
3563   int num_tiles = size_limit / BytesPerLong;
3564 
3565   // allocate space for the tile map:
3566   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3567   jlong  tiles_buf[small_len];
3568   Node*  nodes_buf[small_len];
3569   jlong  inits_buf[small_len];
3570   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3571                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3572   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3573                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3574   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3575                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3576   // tiles: exact bitwise model of all primitive constants
3577   // nodes: last constant-storing node subsumed into the tiles model
3578   // inits: which bytes (in each tile) are touched by any initializations
3579 
3580   //// Pass A: Fill in the tile model with any relevant stores.
3581 
3582   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3583   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3584   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3585   Node* zmem = zero_memory(); // initially zero memory state
3586   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3587     Node* st = in(i);
3588     intptr_t st_off = get_store_offset(st, phase);
3589 
3590     // Figure out the store's offset and constant value:
3591     if (st_off < header_size)             continue; //skip (ignore header)
3592     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
3593     int st_size = st->as_Store()->memory_size();
3594     if (st_off + st_size > size_limit)    break;
3595 
3596     // Record which bytes are touched, whether by constant or not.
3597     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3598       continue;                 // skip (strange store size)
3599 
3600     const Type* val = phase->type(st->in(MemNode::ValueIn));
3601     if (!val->singleton())                continue; //skip (non-con store)
3602     BasicType type = val->basic_type();
3603 
3604     jlong con = 0;
3605     switch (type) {
3606     case T_INT:    con = val->is_int()->get_con();  break;
3607     case T_LONG:   con = val->is_long()->get_con(); break;
3608     case T_FLOAT:  con = jint_cast(val->getf());    break;
3609     case T_DOUBLE: con = jlong_cast(val->getd());   break;
3610     default:                              continue; //skip (odd store type)
3611     }
3612 
3613     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3614         st->Opcode() == Op_StoreL) {
3615       continue;                 // This StoreL is already optimal.
3616     }
3617 
3618     // Store down the constant.
3619     store_constant(tiles, num_tiles, st_off, st_size, con);
3620 
3621     intptr_t j = st_off >> LogBytesPerLong;
3622 
3623     if (type == T_INT && st_size == BytesPerInt
3624         && (st_off & BytesPerInt) == BytesPerInt) {
3625       jlong lcon = tiles[j];
3626       if (!Matcher::isSimpleConstant64(lcon) &&
3627           st->Opcode() == Op_StoreI) {
3628         // This StoreI is already optimal by itself.
3629         jint* intcon = (jint*) &tiles[j];
3630         intcon[1] = 0;  // undo the store_constant()
3631 
3632         // If the previous store is also optimal by itself, back up and
3633         // undo the action of the previous loop iteration... if we can.
3634         // But if we can't, just let the previous half take care of itself.
3635         st = nodes[j];
3636         st_off -= BytesPerInt;
3637         con = intcon[0];
3638         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3639           assert(st_off >= header_size, "still ignoring header");
3640           assert(get_store_offset(st, phase) == st_off, "must be");
3641           assert(in(i-1) == zmem, "must be");
3642           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3643           assert(con == tcon->is_int()->get_con(), "must be");
3644           // Undo the effects of the previous loop trip, which swallowed st:
3645           intcon[0] = 0;        // undo store_constant()
3646           set_req(i-1, st);     // undo set_req(i, zmem)
3647           nodes[j] = NULL;      // undo nodes[j] = st
3648           --old_subword;        // undo ++old_subword
3649         }
3650         continue;               // This StoreI is already optimal.
3651       }
3652     }
3653 
3654     // This store is not needed.
3655     set_req(i, zmem);
3656     nodes[j] = st;              // record for the moment
3657     if (st_size < BytesPerLong) // something has changed
3658           ++old_subword;        // includes int/float, but who's counting...
3659     else  ++old_long;
3660   }
3661 
3662   if ((old_subword + old_long) == 0)
3663     return;                     // nothing more to do
3664 
3665   //// Pass B: Convert any non-zero tiles into optimal constant stores.
3666   // Be sure to insert them before overlapping non-constant stores.
3667   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
3668   for (int j = 0; j < num_tiles; j++) {
3669     jlong con  = tiles[j];
3670     jlong init = inits[j];
3671     if (con == 0)  continue;
3672     jint con0,  con1;           // split the constant, address-wise
3673     jint init0, init1;          // split the init map, address-wise
3674     { union { jlong con; jint intcon[2]; } u;
3675       u.con = con;
3676       con0  = u.intcon[0];
3677       con1  = u.intcon[1];
3678       u.con = init;
3679       init0 = u.intcon[0];
3680       init1 = u.intcon[1];
3681     }
3682 
3683     Node* old = nodes[j];
3684     assert(old != NULL, "need the prior store");
3685     intptr_t offset = (j * BytesPerLong);
3686 
3687     bool split = !Matcher::isSimpleConstant64(con);
3688 
3689     if (offset < header_size) {
3690       assert(offset + BytesPerInt >= header_size, "second int counts");
3691       assert(*(jint*)&tiles[j] == 0, "junk in header");
3692       split = true;             // only the second word counts
3693       // Example:  int a[] = { 42 ... }
3694     } else if (con0 == 0 && init0 == -1) {
3695       split = true;             // first word is covered by full inits
3696       // Example:  int a[] = { ... foo(), 42 ... }
3697     } else if (con1 == 0 && init1 == -1) {
3698       split = true;             // second word is covered by full inits
3699       // Example:  int a[] = { ... 42, foo() ... }
3700     }
3701 
3702     // Here's a case where init0 is neither 0 nor -1:
3703     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
3704     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3705     // In this case the tile is not split; it is (jlong)42.
3706     // The big tile is stored down, and then the foo() value is inserted.
3707     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3708 
3709     Node* ctl = old->in(MemNode::Control);
3710     Node* adr = make_raw_address(offset, phase);
3711     const TypePtr* atp = TypeRawPtr::BOTTOM;
3712 
3713     // One or two coalesced stores to plop down.
3714     Node*    st[2];
3715     intptr_t off[2];
3716     int  nst = 0;
3717     if (!split) {
3718       ++new_long;
3719       off[nst] = offset;
3720       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3721                                   phase->longcon(con), T_LONG, MemNode::unordered);
3722     } else {
3723       // Omit either if it is a zero.
3724       if (con0 != 0) {
3725         ++new_int;
3726         off[nst]  = offset;
3727         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3728                                     phase->intcon(con0), T_INT, MemNode::unordered);
3729       }
3730       if (con1 != 0) {
3731         ++new_int;
3732         offset += BytesPerInt;
3733         adr = make_raw_address(offset, phase);
3734         off[nst]  = offset;
3735         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3736                                     phase->intcon(con1), T_INT, MemNode::unordered);
3737       }
3738     }
3739 
3740     // Insert second store first, then the first before the second.
3741     // Insert each one just before any overlapping non-constant stores.
3742     while (nst > 0) {
3743       Node* st1 = st[--nst];
3744       C->copy_node_notes_to(st1, old);
3745       st1 = phase->transform(st1);
3746       offset = off[nst];
3747       assert(offset >= header_size, "do not smash header");
3748       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3749       guarantee(ins_idx != 0, "must re-insert constant store");
3750       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
3751       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3752         set_req(--ins_idx, st1);
3753       else
3754         ins_req(ins_idx, st1);
3755     }
3756   }
3757 
3758   if (PrintCompilation && WizardMode)
3759     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3760                   old_subword, old_long, new_int, new_long);
3761   if (C->log() != NULL)
3762     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3763                    old_subword, old_long, new_int, new_long);
3764 
3765   // Clean up any remaining occurrences of zmem:
3766   remove_extra_zeroes();
3767 }
3768 
3769 // Explore forward from in(start) to find the first fully initialized
3770 // word, and return its offset.  Skip groups of subword stores which
3771 // together initialize full words.  If in(start) is itself part of a
3772 // fully initialized word, return the offset of in(start).  If there
3773 // are no following full-word stores, or if something is fishy, return
3774 // a negative value.
3775 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3776   int       int_map = 0;
3777   intptr_t  int_map_off = 0;
3778   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
3779 
3780   for (uint i = start, limit = req(); i < limit; i++) {
3781     Node* st = in(i);
3782 
3783     intptr_t st_off = get_store_offset(st, phase);
3784     if (st_off < 0)  break;  // return conservative answer
3785 
3786     int st_size = st->as_Store()->memory_size();
3787     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3788       return st_off;            // we found a complete word init
3789     }
3790 
3791     // update the map:
3792 
3793     intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3794     if (this_int_off != int_map_off) {
3795       // reset the map:
3796       int_map = 0;
3797       int_map_off = this_int_off;
3798     }
3799 
3800     int subword_off = st_off - this_int_off;
3801     int_map |= right_n_bits(st_size) << subword_off;
3802     if ((int_map & FULL_MAP) == FULL_MAP) {
3803       return this_int_off;      // we found a complete word init
3804     }
3805 
3806     // Did this store hit or cross the word boundary?
3807     intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3808     if (next_int_off == this_int_off + BytesPerInt) {
3809       // We passed the current int, without fully initializing it.
3810       int_map_off = next_int_off;
3811       int_map >>= BytesPerInt;
3812     } else if (next_int_off > this_int_off + BytesPerInt) {
3813       // We passed the current and next int.
3814       return this_int_off + BytesPerInt;
3815     }
3816   }
3817 
3818   return -1;
3819 }
3820 
3821 
3822 // Called when the associated AllocateNode is expanded into CFG.
3823 // At this point, we may perform additional optimizations.
3824 // Linearize the stores by ascending offset, to make memory
3825 // activity as coherent as possible.
3826 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3827                                       intptr_t header_size,
3828                                       Node* size_in_bytes,
3829                                       PhaseGVN* phase) {
3830   assert(!is_complete(), "not already complete");
3831   assert(stores_are_sane(phase), "");
3832   assert(allocation() != NULL, "must be present");
3833 
3834   remove_extra_zeroes();
3835 
3836   if (ReduceFieldZeroing || ReduceBulkZeroing)
3837     // reduce instruction count for common initialization patterns
3838     coalesce_subword_stores(header_size, size_in_bytes, phase);
3839 
3840   Node* zmem = zero_memory();   // initially zero memory state
3841   Node* inits = zmem;           // accumulating a linearized chain of inits
3842   #ifdef ASSERT
3843   intptr_t first_offset = allocation()->minimum_header_size();
3844   intptr_t last_init_off = first_offset;  // previous init offset
3845   intptr_t last_init_end = first_offset;  // previous init offset+size
3846   intptr_t last_tile_end = first_offset;  // previous tile offset+size
3847   #endif
3848   intptr_t zeroes_done = header_size;
3849 
3850   bool do_zeroing = true;       // we might give up if inits are very sparse
3851   int  big_init_gaps = 0;       // how many large gaps have we seen?
3852 
3853   if (UseTLAB && ZeroTLAB)  do_zeroing = false;
3854   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
3855 
3856   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3857     Node* st = in(i);
3858     intptr_t st_off = get_store_offset(st, phase);
3859     if (st_off < 0)
3860       break;                    // unknown junk in the inits
3861     if (st->in(MemNode::Memory) != zmem)
3862       break;                    // complicated store chains somehow in list
3863 
3864     int st_size = st->as_Store()->memory_size();
3865     intptr_t next_init_off = st_off + st_size;
3866 
3867     if (do_zeroing && zeroes_done < next_init_off) {
3868       // See if this store needs a zero before it or under it.
3869       intptr_t zeroes_needed = st_off;
3870 
3871       if (st_size < BytesPerInt) {
3872         // Look for subword stores which only partially initialize words.
3873         // If we find some, we must lay down some word-level zeroes first,
3874         // underneath the subword stores.
3875         //
3876         // Examples:
3877         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
3878         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
3879         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
3880         //
3881         // Note:  coalesce_subword_stores may have already done this,
3882         // if it was prompted by constant non-zero subword initializers.
3883         // But this case can still arise with non-constant stores.
3884 
3885         intptr_t next_full_store = find_next_fullword_store(i, phase);
3886 
3887         // In the examples above:
3888         //   in(i)          p   q   r   s     x   y     z
3889         //   st_off        12  13  14  15    12  13    14
3890         //   st_size        1   1   1   1     1   1     1
3891         //   next_full_s.  12  16  16  16    16  16    16
3892         //   z's_done      12  16  16  16    12  16    12
3893         //   z's_needed    12  16  16  16    16  16    16
3894         //   zsize          0   0   0   0     4   0     4
3895         if (next_full_store < 0) {
3896           // Conservative tack:  Zero to end of current word.
3897           zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3898         } else {
3899           // Zero to beginning of next fully initialized word.
3900           // Or, don't zero at all, if we are already in that word.
3901           assert(next_full_store >= zeroes_needed, "must go forward");
3902           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3903           zeroes_needed = next_full_store;
3904         }
3905       }
3906 
3907       if (zeroes_needed > zeroes_done) {
3908         intptr_t zsize = zeroes_needed - zeroes_done;
3909         // Do some incremental zeroing on rawmem, in parallel with inits.
3910         zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3911         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3912                                               zeroes_done, zeroes_needed,
3913                                               phase);
3914         zeroes_done = zeroes_needed;
3915         if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
3916           do_zeroing = false;   // leave the hole, next time
3917       }
3918     }
3919 
3920     // Collect the store and move on:
3921     st->set_req(MemNode::Memory, inits);
3922     inits = st;                 // put it on the linearized chain
3923     set_req(i, zmem);           // unhook from previous position
3924 
3925     if (zeroes_done == st_off)
3926       zeroes_done = next_init_off;
3927 
3928     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3929 
3930     #ifdef ASSERT
3931     // Various order invariants.  Weaker than stores_are_sane because
3932     // a large constant tile can be filled in by smaller non-constant stores.
3933     assert(st_off >= last_init_off, "inits do not reverse");
3934     last_init_off = st_off;
3935     const Type* val = NULL;
3936     if (st_size >= BytesPerInt &&
3937         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3938         (int)val->basic_type() < (int)T_OBJECT) {
3939       assert(st_off >= last_tile_end, "tiles do not overlap");
3940       assert(st_off >= last_init_end, "tiles do not overwrite inits");
3941       last_tile_end = MAX2(last_tile_end, next_init_off);
3942     } else {
3943       intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3944       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3945       assert(st_off      >= last_init_end, "inits do not overlap");
3946       last_init_end = next_init_off;  // it's a non-tile
3947     }
3948     #endif //ASSERT
3949   }
3950 
3951   remove_extra_zeroes();        // clear out all the zmems left over
3952   add_req(inits);
3953 
3954   if (!(UseTLAB && ZeroTLAB)) {
3955     // If anything remains to be zeroed, zero it all now.
3956     zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3957     // if it is the last unused 4 bytes of an instance, forget about it
3958     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3959     if (zeroes_done + BytesPerLong >= size_limit) {
3960       AllocateNode* alloc = allocation();
3961       assert(alloc != NULL, "must be present");
3962       if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
3963         Node* klass_node = alloc->in(AllocateNode::KlassNode);
3964         ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3965         if (zeroes_done == k->layout_helper())
3966           zeroes_done = size_limit;
3967       }
3968     }
3969     if (zeroes_done < size_limit) {
3970       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3971                                             zeroes_done, size_in_bytes, phase);
3972     }
3973   }
3974 
3975   set_complete(phase);
3976   return rawmem;
3977 }
3978 
3979 
3980 #ifdef ASSERT
3981 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3982   if (is_complete())
3983     return true;                // stores could be anything at this point
3984   assert(allocation() != NULL, "must be present");
3985   intptr_t last_off = allocation()->minimum_header_size();
3986   for (uint i = InitializeNode::RawStores; i < req(); i++) {
3987     Node* st = in(i);
3988     intptr_t st_off = get_store_offset(st, phase);
3989     if (st_off < 0)  continue;  // ignore dead garbage
3990     if (last_off > st_off) {
3991       tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
3992       this->dump(2);
3993       assert(false, "ascending store offsets");
3994       return false;
3995     }
3996     last_off = st_off + st->as_Store()->memory_size();
3997   }
3998   return true;
3999 }
4000 #endif //ASSERT
4001 
4002 
4003 
4004 
4005 //============================MergeMemNode=====================================
4006 //
4007 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
4008 // contributing store or call operations.  Each contributor provides the memory
4009 // state for a particular "alias type" (see Compile::alias_type).  For example,
4010 // if a MergeMem has an input X for alias category #6, then any memory reference
4011 // to alias category #6 may use X as its memory state input, as an exact equivalent
4012 // to using the MergeMem as a whole.
4013 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4014 //
4015 // (Here, the <N> notation gives the index of the relevant adr_type.)
4016 //
4017 // In one special case (and more cases in the future), alias categories overlap.
4018 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4019 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
4020 // it is exactly equivalent to that state W:
4021 //   MergeMem(<Bot>: W) <==> W
4022 //
4023 // Usually, the merge has more than one input.  In that case, where inputs
4024 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4025 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4026 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4027 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4028 //
4029 // A merge can take a "wide" memory state as one of its narrow inputs.
4030 // This simply means that the merge observes out only the relevant parts of
4031 // the wide input.  That is, wide memory states arriving at narrow merge inputs
4032 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
4033 //
4034 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4035 // and that memory slices "leak through":
4036 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4037 //
4038 // But, in such a cascade, repeated memory slices can "block the leak":
4039 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4040 //
4041 // In the last example, Y is not part of the combined memory state of the
4042 // outermost MergeMem.  The system must, of course, prevent unschedulable
4043 // memory states from arising, so you can be sure that the state Y is somehow
4044 // a precursor to state Y'.
4045 //
4046 //
4047 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4048 // of each MergeMemNode array are exactly the numerical alias indexes, including
4049 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
4050 // Compile::alias_type (and kin) produce and manage these indexes.
4051 //
4052 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4053 // (Note that this provides quick access to the top node inside MergeMem methods,
4054 // without the need to reach out via TLS to Compile::current.)
4055 //
4056 // As a consequence of what was just described, a MergeMem that represents a full
4057 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4058 // containing all alias categories.
4059 //
4060 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4061 //
4062 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4063 // a memory state for the alias type <N>, or else the top node, meaning that
4064 // there is no particular input for that alias type.  Note that the length of
4065 // a MergeMem is variable, and may be extended at any time to accommodate new
4066 // memory states at larger alias indexes.  When merges grow, they are of course
4067 // filled with "top" in the unused in() positions.
4068 //
4069 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4070 // (Top was chosen because it works smoothly with passes like GCM.)
4071 //
4072 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
4073 // the type of random VM bits like TLS references.)  Since it is always the
4074 // first non-Bot memory slice, some low-level loops use it to initialize an
4075 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
4076 //
4077 //
4078 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
4079 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
4080 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4081 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
4082 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
4083 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4084 //
4085 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4086 // really that different from the other memory inputs.  An abbreviation called
4087 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4088 //
4089 //
4090 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
4091 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
4092 // that "emerges though" the base memory will be marked as excluding the alias types
4093 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4094 //
4095 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4096 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4097 //
4098 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4099 // (It is currently unimplemented.)  As you can see, the resulting merge is
4100 // actually a disjoint union of memory states, rather than an overlay.
4101 //
4102 
4103 //------------------------------MergeMemNode-----------------------------------
4104 Node* MergeMemNode::make_empty_memory() {
4105   Node* empty_memory = (Node*) Compile::current()->top();
4106   assert(empty_memory->is_top(), "correct sentinel identity");
4107   return empty_memory;
4108 }
4109 
4110 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4111   init_class_id(Class_MergeMem);
4112   // all inputs are nullified in Node::Node(int)
4113   // set_input(0, NULL);  // no control input
4114 
4115   // Initialize the edges uniformly to top, for starters.
4116   Node* empty_mem = make_empty_memory();
4117   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4118     init_req(i,empty_mem);
4119   }
4120   assert(empty_memory() == empty_mem, "");
4121 
4122   if( new_base != NULL && new_base->is_MergeMem() ) {
4123     MergeMemNode* mdef = new_base->as_MergeMem();
4124     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4125     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4126       mms.set_memory(mms.memory2());
4127     }
4128     assert(base_memory() == mdef->base_memory(), "");
4129   } else {
4130     set_base_memory(new_base);
4131   }
4132 }
4133 
4134 // Make a new, untransformed MergeMem with the same base as 'mem'.
4135 // If mem is itself a MergeMem, populate the result with the same edges.
4136 MergeMemNode* MergeMemNode::make(Node* mem) {
4137   return new MergeMemNode(mem);
4138 }
4139 
4140 //------------------------------cmp--------------------------------------------
4141 uint MergeMemNode::hash() const { return NO_HASH; }
4142 uint MergeMemNode::cmp( const Node &n ) const {
4143   return (&n == this);          // Always fail except on self
4144 }
4145 
4146 //------------------------------Identity---------------------------------------
4147 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4148   // Identity if this merge point does not record any interesting memory
4149   // disambiguations.
4150   Node* base_mem = base_memory();
4151   Node* empty_mem = empty_memory();
4152   if (base_mem != empty_mem) {  // Memory path is not dead?
4153     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4154       Node* mem = in(i);
4155       if (mem != empty_mem && mem != base_mem) {
4156         return this;            // Many memory splits; no change
4157       }
4158     }
4159   }
4160   return base_mem;              // No memory splits; ID on the one true input
4161 }
4162 
4163 //------------------------------Ideal------------------------------------------
4164 // This method is invoked recursively on chains of MergeMem nodes
4165 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4166   // Remove chain'd MergeMems
4167   //
4168   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4169   // relative to the "in(Bot)".  Since we are patching both at the same time,
4170   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4171   // but rewrite each "in(i)" relative to the new "in(Bot)".
4172   Node *progress = NULL;
4173 
4174 
4175   Node* old_base = base_memory();
4176   Node* empty_mem = empty_memory();
4177   if (old_base == empty_mem)
4178     return NULL; // Dead memory path.
4179 
4180   MergeMemNode* old_mbase;
4181   if (old_base != NULL && old_base->is_MergeMem())
4182     old_mbase = old_base->as_MergeMem();
4183   else
4184     old_mbase = NULL;
4185   Node* new_base = old_base;
4186 
4187   // simplify stacked MergeMems in base memory
4188   if (old_mbase)  new_base = old_mbase->base_memory();
4189 
4190   // the base memory might contribute new slices beyond my req()
4191   if (old_mbase)  grow_to_match(old_mbase);
4192 
4193   // Look carefully at the base node if it is a phi.
4194   PhiNode* phi_base;
4195   if (new_base != NULL && new_base->is_Phi())
4196     phi_base = new_base->as_Phi();
4197   else
4198     phi_base = NULL;
4199 
4200   Node*    phi_reg = NULL;
4201   uint     phi_len = (uint)-1;
4202   if (phi_base != NULL && !phi_base->is_copy()) {
4203     // do not examine phi if degraded to a copy
4204     phi_reg = phi_base->region();
4205     phi_len = phi_base->req();
4206     // see if the phi is unfinished
4207     for (uint i = 1; i < phi_len; i++) {
4208       if (phi_base->in(i) == NULL) {
4209         // incomplete phi; do not look at it yet!
4210         phi_reg = NULL;
4211         phi_len = (uint)-1;
4212         break;
4213       }
4214     }
4215   }
4216 
4217   // Note:  We do not call verify_sparse on entry, because inputs
4218   // can normalize to the base_memory via subsume_node or similar
4219   // mechanisms.  This method repairs that damage.
4220 
4221   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4222 
4223   // Look at each slice.
4224   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4225     Node* old_in = in(i);
4226     // calculate the old memory value
4227     Node* old_mem = old_in;
4228     if (old_mem == empty_mem)  old_mem = old_base;
4229     assert(old_mem == memory_at(i), "");
4230 
4231     // maybe update (reslice) the old memory value
4232 
4233     // simplify stacked MergeMems
4234     Node* new_mem = old_mem;
4235     MergeMemNode* old_mmem;
4236     if (old_mem != NULL && old_mem->is_MergeMem())
4237       old_mmem = old_mem->as_MergeMem();
4238     else
4239       old_mmem = NULL;
4240     if (old_mmem == this) {
4241       // This can happen if loops break up and safepoints disappear.
4242       // A merge of BotPtr (default) with a RawPtr memory derived from a
4243       // safepoint can be rewritten to a merge of the same BotPtr with
4244       // the BotPtr phi coming into the loop.  If that phi disappears
4245       // also, we can end up with a self-loop of the mergemem.
4246       // In general, if loops degenerate and memory effects disappear,
4247       // a mergemem can be left looking at itself.  This simply means
4248       // that the mergemem's default should be used, since there is
4249       // no longer any apparent effect on this slice.
4250       // Note: If a memory slice is a MergeMem cycle, it is unreachable
4251       //       from start.  Update the input to TOP.
4252       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4253     }
4254     else if (old_mmem != NULL) {
4255       new_mem = old_mmem->memory_at(i);
4256     }
4257     // else preceding memory was not a MergeMem
4258 
4259     // replace equivalent phis (unfortunately, they do not GVN together)
4260     if (new_mem != NULL && new_mem != new_base &&
4261         new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4262       if (new_mem->is_Phi()) {
4263         PhiNode* phi_mem = new_mem->as_Phi();
4264         for (uint i = 1; i < phi_len; i++) {
4265           if (phi_base->in(i) != phi_mem->in(i)) {
4266             phi_mem = NULL;
4267             break;
4268           }
4269         }
4270         if (phi_mem != NULL) {
4271           // equivalent phi nodes; revert to the def
4272           new_mem = new_base;
4273         }
4274       }
4275     }
4276 
4277     // maybe store down a new value
4278     Node* new_in = new_mem;
4279     if (new_in == new_base)  new_in = empty_mem;
4280 
4281     if (new_in != old_in) {
4282       // Warning:  Do not combine this "if" with the previous "if"
4283       // A memory slice might have be be rewritten even if it is semantically
4284       // unchanged, if the base_memory value has changed.
4285       set_req(i, new_in);
4286       progress = this;          // Report progress
4287     }
4288   }
4289 
4290   if (new_base != old_base) {
4291     set_req(Compile::AliasIdxBot, new_base);
4292     // Don't use set_base_memory(new_base), because we need to update du.
4293     assert(base_memory() == new_base, "");
4294     progress = this;
4295   }
4296 
4297   if( base_memory() == this ) {
4298     // a self cycle indicates this memory path is dead
4299     set_req(Compile::AliasIdxBot, empty_mem);
4300   }
4301 
4302   // Resolve external cycles by calling Ideal on a MergeMem base_memory
4303   // Recursion must occur after the self cycle check above
4304   if( base_memory()->is_MergeMem() ) {
4305     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4306     Node *m = phase->transform(new_mbase);  // Rollup any cycles
4307     if( m != NULL && (m->is_top() ||
4308         m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4309       // propagate rollup of dead cycle to self
4310       set_req(Compile::AliasIdxBot, empty_mem);
4311     }
4312   }
4313 
4314   if( base_memory() == empty_mem ) {
4315     progress = this;
4316     // Cut inputs during Parse phase only.
4317     // During Optimize phase a dead MergeMem node will be subsumed by Top.
4318     if( !can_reshape ) {
4319       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4320         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4321       }
4322     }
4323   }
4324 
4325   if( !progress && base_memory()->is_Phi() && can_reshape ) {
4326     // Check if PhiNode::Ideal's "Split phis through memory merges"
4327     // transform should be attempted. Look for this->phi->this cycle.
4328     uint merge_width = req();
4329     if (merge_width > Compile::AliasIdxRaw) {
4330       PhiNode* phi = base_memory()->as_Phi();
4331       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4332         if (phi->in(i) == this) {
4333           phase->is_IterGVN()->_worklist.push(phi);
4334           break;
4335         }
4336       }
4337     }
4338   }
4339 
4340   assert(progress || verify_sparse(), "please, no dups of base");
4341   return progress;
4342 }
4343 
4344 //-------------------------set_base_memory-------------------------------------
4345 void MergeMemNode::set_base_memory(Node *new_base) {
4346   Node* empty_mem = empty_memory();
4347   set_req(Compile::AliasIdxBot, new_base);
4348   assert(memory_at(req()) == new_base, "must set default memory");
4349   // Clear out other occurrences of new_base:
4350   if (new_base != empty_mem) {
4351     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4352       if (in(i) == new_base)  set_req(i, empty_mem);
4353     }
4354   }
4355 }
4356 
4357 //------------------------------out_RegMask------------------------------------
4358 const RegMask &MergeMemNode::out_RegMask() const {
4359   return RegMask::Empty;
4360 }
4361 
4362 //------------------------------dump_spec--------------------------------------
4363 #ifndef PRODUCT
4364 void MergeMemNode::dump_spec(outputStream *st) const {
4365   st->print(" {");
4366   Node* base_mem = base_memory();
4367   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4368     Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4369     if (mem == base_mem) { st->print(" -"); continue; }
4370     st->print( " N%d:", mem->_idx );
4371     Compile::current()->get_adr_type(i)->dump_on(st);
4372   }
4373   st->print(" }");
4374 }
4375 #endif // !PRODUCT
4376 
4377 
4378 #ifdef ASSERT
4379 static bool might_be_same(Node* a, Node* b) {
4380   if (a == b)  return true;
4381   if (!(a->is_Phi() || b->is_Phi()))  return false;
4382   // phis shift around during optimization
4383   return true;  // pretty stupid...
4384 }
4385 
4386 // verify a narrow slice (either incoming or outgoing)
4387 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4388   if (!VerifyAliases)       return;  // don't bother to verify unless requested
4389   if (is_error_reported())  return;  // muzzle asserts when debugging an error
4390   if (Node::in_dump())      return;  // muzzle asserts when printing
4391   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4392   assert(n != NULL, "");
4393   // Elide intervening MergeMem's
4394   while (n->is_MergeMem()) {
4395     n = n->as_MergeMem()->memory_at(alias_idx);
4396   }
4397   Compile* C = Compile::current();
4398   const TypePtr* n_adr_type = n->adr_type();
4399   if (n == m->empty_memory()) {
4400     // Implicit copy of base_memory()
4401   } else if (n_adr_type != TypePtr::BOTTOM) {
4402     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4403     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4404   } else {
4405     // A few places like make_runtime_call "know" that VM calls are narrow,
4406     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4407     bool expected_wide_mem = false;
4408     if (n == m->base_memory()) {
4409       expected_wide_mem = true;
4410     } else if (alias_idx == Compile::AliasIdxRaw ||
4411                n == m->memory_at(Compile::AliasIdxRaw)) {
4412       expected_wide_mem = true;
4413     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4414       // memory can "leak through" calls on channels that
4415       // are write-once.  Allow this also.
4416       expected_wide_mem = true;
4417     }
4418     assert(expected_wide_mem, "expected narrow slice replacement");
4419   }
4420 }
4421 #else // !ASSERT
4422 #define verify_memory_slice(m,i,n) (void)(0)  // PRODUCT version is no-op
4423 #endif
4424 
4425 
4426 //-----------------------------memory_at---------------------------------------
4427 Node* MergeMemNode::memory_at(uint alias_idx) const {
4428   assert(alias_idx >= Compile::AliasIdxRaw ||
4429          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4430          "must avoid base_memory and AliasIdxTop");
4431 
4432   // Otherwise, it is a narrow slice.
4433   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4434   Compile *C = Compile::current();
4435   if (is_empty_memory(n)) {
4436     // the array is sparse; empty slots are the "top" node
4437     n = base_memory();
4438     assert(Node::in_dump()
4439            || n == NULL || n->bottom_type() == Type::TOP
4440            || n->adr_type() == NULL // address is TOP
4441            || n->adr_type() == TypePtr::BOTTOM
4442            || n->adr_type() == TypeRawPtr::BOTTOM
4443            || Compile::current()->AliasLevel() == 0,
4444            "must be a wide memory");
4445     // AliasLevel == 0 if we are organizing the memory states manually.
4446     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4447   } else {
4448     // make sure the stored slice is sane
4449     #ifdef ASSERT
4450     if (is_error_reported() || Node::in_dump()) {
4451     } else if (might_be_same(n, base_memory())) {
4452       // Give it a pass:  It is a mostly harmless repetition of the base.
4453       // This can arise normally from node subsumption during optimization.
4454     } else {
4455       verify_memory_slice(this, alias_idx, n);
4456     }
4457     #endif
4458   }
4459   return n;
4460 }
4461 
4462 //---------------------------set_memory_at-------------------------------------
4463 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4464   verify_memory_slice(this, alias_idx, n);
4465   Node* empty_mem = empty_memory();
4466   if (n == base_memory())  n = empty_mem;  // collapse default
4467   uint need_req = alias_idx+1;
4468   if (req() < need_req) {
4469     if (n == empty_mem)  return;  // already the default, so do not grow me
4470     // grow the sparse array
4471     do {
4472       add_req(empty_mem);
4473     } while (req() < need_req);
4474   }
4475   set_req( alias_idx, n );
4476 }
4477 
4478 
4479 
4480 //--------------------------iteration_setup------------------------------------
4481 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4482   if (other != NULL) {
4483     grow_to_match(other);
4484     // invariant:  the finite support of mm2 is within mm->req()
4485     #ifdef ASSERT
4486     for (uint i = req(); i < other->req(); i++) {
4487       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4488     }
4489     #endif
4490   }
4491   // Replace spurious copies of base_memory by top.
4492   Node* base_mem = base_memory();
4493   if (base_mem != NULL && !base_mem->is_top()) {
4494     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4495       if (in(i) == base_mem)
4496         set_req(i, empty_memory());
4497     }
4498   }
4499 }
4500 
4501 //---------------------------grow_to_match-------------------------------------
4502 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4503   Node* empty_mem = empty_memory();
4504   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4505   // look for the finite support of the other memory
4506   for (uint i = other->req(); --i >= req(); ) {
4507     if (other->in(i) != empty_mem) {
4508       uint new_len = i+1;
4509       while (req() < new_len)  add_req(empty_mem);
4510       break;
4511     }
4512   }
4513 }
4514 
4515 //---------------------------verify_sparse-------------------------------------
4516 #ifndef PRODUCT
4517 bool MergeMemNode::verify_sparse() const {
4518   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4519   Node* base_mem = base_memory();
4520   // The following can happen in degenerate cases, since empty==top.
4521   if (is_empty_memory(base_mem))  return true;
4522   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4523     assert(in(i) != NULL, "sane slice");
4524     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
4525   }
4526   return true;
4527 }
4528 
4529 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4530   Node* n;
4531   n = mm->in(idx);
4532   if (mem == n)  return true;  // might be empty_memory()
4533   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4534   if (mem == n)  return true;
4535   while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4536     if (mem == n)  return true;
4537     if (n == NULL)  break;
4538   }
4539   return false;
4540 }
4541 #endif // !PRODUCT
--- EOF ---