1 /*
   2  * Copyright (c) 1997, 2009, 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 // Portions of code courtesy of Clifford Click
  26 
  27 // Optimization - Graph Style
  28 
  29 #include "incls/_precompiled.incl"
  30 #include "incls/_memnode.cpp.incl"
  31 
  32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
  33 
  34 //=============================================================================
  35 uint MemNode::size_of() const { return sizeof(*this); }
  36 
  37 const TypePtr *MemNode::adr_type() const {
  38   Node* adr = in(Address);
  39   const TypePtr* cross_check = NULL;
  40   DEBUG_ONLY(cross_check = _adr_type);
  41   return calculate_adr_type(adr->bottom_type(), cross_check);
  42 }
  43 
  44 #ifndef PRODUCT
  45 void MemNode::dump_spec(outputStream *st) const {
  46   if (in(Address) == NULL)  return; // node is dead
  47 #ifndef ASSERT
  48   // fake the missing field
  49   const TypePtr* _adr_type = NULL;
  50   if (in(Address) != NULL)
  51     _adr_type = in(Address)->bottom_type()->isa_ptr();
  52 #endif
  53   dump_adr_type(this, _adr_type, st);
  54 
  55   Compile* C = Compile::current();
  56   if( C->alias_type(_adr_type)->is_volatile() )
  57     st->print(" Volatile!");
  58 }
  59 
  60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
  61   st->print(" @");
  62   if (adr_type == NULL) {
  63     st->print("NULL");
  64   } else {
  65     adr_type->dump_on(st);
  66     Compile* C = Compile::current();
  67     Compile::AliasType* atp = NULL;
  68     if (C->have_alias_type(adr_type))  atp = C->alias_type(adr_type);
  69     if (atp == NULL)
  70       st->print(", idx=?\?;");
  71     else if (atp->index() == Compile::AliasIdxBot)
  72       st->print(", idx=Bot;");
  73     else if (atp->index() == Compile::AliasIdxTop)
  74       st->print(", idx=Top;");
  75     else if (atp->index() == Compile::AliasIdxRaw)
  76       st->print(", idx=Raw;");
  77     else {
  78       ciField* field = atp->field();
  79       if (field) {
  80         st->print(", name=");
  81         field->print_name_on(st);
  82       }
  83       st->print(", idx=%d;", atp->index());
  84     }
  85   }
  86 }
  87 
  88 extern void print_alias_types();
  89 
  90 #endif
  91 
  92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
  93   const TypeOopPtr *tinst = t_adr->isa_oopptr();
  94   if (tinst == NULL || !tinst->is_known_instance_field())
  95     return mchain;  // don't try to optimize non-instance types
  96   uint instance_id = tinst->instance_id();
  97   Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
  98   Node *prev = NULL;
  99   Node *result = mchain;
 100   while (prev != result) {
 101     prev = result;
 102     if (result == start_mem)
 103       break;  // hit one of our sentinels
 104     // skip over a call which does not affect this memory slice
 105     if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
 106       Node *proj_in = result->in(0);
 107       if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
 108         break;  // hit one of our sentinels
 109       } else if (proj_in->is_Call()) {
 110         CallNode *call = proj_in->as_Call();
 111         if (!call->may_modify(t_adr, phase)) {
 112           result = call->in(TypeFunc::Memory);
 113         }
 114       } else if (proj_in->is_Initialize()) {
 115         AllocateNode* alloc = proj_in->as_Initialize()->allocation();
 116         // Stop if this is the initialization for the object instance which
 117         // which contains this memory slice, otherwise skip over it.
 118         if (alloc != NULL && alloc->_idx != instance_id) {
 119           result = proj_in->in(TypeFunc::Memory);
 120         }
 121       } else if (proj_in->is_MemBar()) {
 122         result = proj_in->in(TypeFunc::Memory);
 123       } else {
 124         assert(false, "unexpected projection");
 125       }
 126     } else if (result->is_ClearArray()) {
 127       if (!ClearArrayNode::step_through(&result, instance_id, phase)) {
 128         // Can not bypass initialization of the instance
 129         // we are looking for.
 130         break;
 131       }
 132       // Otherwise skip it (the call updated 'result' value).
 133     } else if (result->is_MergeMem()) {
 134       result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
 135     }
 136   }
 137   return result;
 138 }
 139 
 140 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
 141   const TypeOopPtr *t_oop = t_adr->isa_oopptr();
 142   bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
 143   PhaseIterGVN *igvn = phase->is_IterGVN();
 144   Node *result = mchain;
 145   result = optimize_simple_memory_chain(result, t_adr, phase);
 146   if (is_instance && igvn != NULL  && result->is_Phi()) {
 147     PhiNode *mphi = result->as_Phi();
 148     assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
 149     const TypePtr *t = mphi->adr_type();
 150     if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
 151         t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
 152         t->is_oopptr()->cast_to_exactness(true)
 153          ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 154          ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
 155       // clone the Phi with our address type
 156       result = mphi->split_out_instance(t_adr, igvn);
 157     } else {
 158       assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
 159     }
 160   }
 161   return result;
 162 }
 163 
 164 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem,  const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
 165   uint alias_idx = phase->C->get_alias_index(tp);
 166   Node *mem = mmem;
 167 #ifdef ASSERT
 168   {
 169     // Check that current type is consistent with the alias index used during graph construction
 170     assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
 171     bool consistent =  adr_check == NULL || adr_check->empty() ||
 172                        phase->C->must_alias(adr_check, alias_idx );
 173     // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
 174     if( !consistent && adr_check != NULL && !adr_check->empty() &&
 175                tp->isa_aryptr() &&        tp->offset() == Type::OffsetBot &&
 176         adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
 177         ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
 178           adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
 179           adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
 180       // don't assert if it is dead code.
 181       consistent = true;
 182     }
 183     if( !consistent ) {
 184       st->print("alias_idx==%d, adr_check==", alias_idx);
 185       if( adr_check == NULL ) {
 186         st->print("NULL");
 187       } else {
 188         adr_check->dump();
 189       }
 190       st->cr();
 191       print_alias_types();
 192       assert(consistent, "adr_check must match alias idx");
 193     }
 194   }
 195 #endif
 196   // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
 197   // means an array I have not precisely typed yet.  Do not do any
 198   // alias stuff with it any time soon.
 199   const TypeOopPtr *toop = tp->isa_oopptr();
 200   if( tp->base() != Type::AnyPtr &&
 201       !(toop &&
 202         toop->klass() != NULL &&
 203         toop->klass()->is_java_lang_Object() &&
 204         toop->offset() == Type::OffsetBot) ) {
 205     // compress paths and change unreachable cycles to TOP
 206     // If not, we can update the input infinitely along a MergeMem cycle
 207     // Equivalent code in PhiNode::Ideal
 208     Node* m  = phase->transform(mmem);
 209     // If transformed to a MergeMem, get the desired slice
 210     // Otherwise the returned node represents memory for every slice
 211     mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
 212     // Update input if it is progress over what we have now
 213   }
 214   return mem;
 215 }
 216 
 217 //--------------------------Ideal_common---------------------------------------
 218 // Look for degenerate control and memory inputs.  Bypass MergeMem inputs.
 219 // Unhook non-raw memories from complete (macro-expanded) initializations.
 220 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
 221   // If our control input is a dead region, kill all below the region
 222   Node *ctl = in(MemNode::Control);
 223   if (ctl && remove_dead_region(phase, can_reshape))
 224     return this;
 225   ctl = in(MemNode::Control);
 226   // Don't bother trying to transform a dead node
 227   if( ctl && ctl->is_top() )  return NodeSentinel;
 228 
 229   PhaseIterGVN *igvn = phase->is_IterGVN();
 230   // Wait if control on the worklist.
 231   if (ctl && can_reshape && igvn != NULL) {
 232     Node* bol = NULL;
 233     Node* cmp = NULL;
 234     if (ctl->in(0)->is_If()) {
 235       assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
 236       bol = ctl->in(0)->in(1);
 237       if (bol->is_Bool())
 238         cmp = ctl->in(0)->in(1)->in(1);
 239     }
 240     if (igvn->_worklist.member(ctl) ||
 241         (bol != NULL && igvn->_worklist.member(bol)) ||
 242         (cmp != NULL && igvn->_worklist.member(cmp)) ) {
 243       // This control path may be dead.
 244       // Delay this memory node transformation until the control is processed.
 245       phase->is_IterGVN()->_worklist.push(this);
 246       return NodeSentinel; // caller will return NULL
 247     }
 248   }
 249   // Ignore if memory is dead, or self-loop
 250   Node *mem = in(MemNode::Memory);
 251   if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
 252   assert( mem != this, "dead loop in MemNode::Ideal" );
 253 
 254   Node *address = in(MemNode::Address);
 255   const Type *t_adr = phase->type( address );
 256   if( t_adr == Type::TOP )              return NodeSentinel; // caller will return NULL
 257 
 258   if( can_reshape && igvn != NULL &&
 259       (igvn->_worklist.member(address) || phase->type(address) != adr_type()) ) {
 260     // The address's base and type may change when the address is processed.
 261     // Delay this mem node transformation until the address is processed.
 262     phase->is_IterGVN()->_worklist.push(this);
 263     return NodeSentinel; // caller will return NULL
 264   }
 265 
 266   // Do NOT remove or optimize the next lines: ensure a new alias index
 267   // is allocated for an oop pointer type before Escape Analysis.
 268   // Note: C++ will not remove it since the call has side effect.
 269   if ( t_adr->isa_oopptr() ) {
 270     int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
 271   }
 272 
 273 #ifdef ASSERT
 274   Node* base = NULL;
 275   if (address->is_AddP())
 276     base = address->in(AddPNode::Base);
 277   assert(base == NULL || t_adr->isa_rawptr() ||
 278         !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
 279 #endif
 280 
 281   // Avoid independent memory operations
 282   Node* old_mem = mem;
 283 
 284   // The code which unhooks non-raw memories from complete (macro-expanded)
 285   // initializations was removed. After macro-expansion all stores catched
 286   // by Initialize node became raw stores and there is no information
 287   // which memory slices they modify. So it is unsafe to move any memory
 288   // operation above these stores. Also in most cases hooked non-raw memories
 289   // were already unhooked by using information from detect_ptr_independence()
 290   // and find_previous_store().
 291 
 292   if (mem->is_MergeMem()) {
 293     MergeMemNode* mmem = mem->as_MergeMem();
 294     const TypePtr *tp = t_adr->is_ptr();
 295 
 296     mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
 297   }
 298 
 299   if (mem != old_mem) {
 300     set_req(MemNode::Memory, mem);
 301     if (phase->type( mem ) == Type::TOP) return NodeSentinel;
 302     return this;
 303   }
 304 
 305   // let the subclass continue analyzing...
 306   return NULL;
 307 }
 308 
 309 // Helper function for proving some simple control dominations.
 310 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
 311 // Already assumes that 'dom' is available at 'sub', and that 'sub'
 312 // is not a constant (dominated by the method's StartNode).
 313 // Used by MemNode::find_previous_store to prove that the
 314 // control input of a memory operation predates (dominates)
 315 // an allocation it wants to look past.
 316 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
 317   if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
 318     return false; // Conservative answer for dead code
 319 
 320   // Check 'dom'. Skip Proj and CatchProj nodes.
 321   dom = dom->find_exact_control(dom);
 322   if (dom == NULL || dom->is_top())
 323     return false; // Conservative answer for dead code
 324 
 325   if (dom == sub) {
 326     // For the case when, for example, 'sub' is Initialize and the original
 327     // 'dom' is Proj node of the 'sub'.
 328     return false;
 329   }
 330 
 331   if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
 332     return true;
 333 
 334   // 'dom' dominates 'sub' if its control edge and control edges
 335   // of all its inputs dominate or equal to sub's control edge.
 336 
 337   // Currently 'sub' is either Allocate, Initialize or Start nodes.
 338   // Or Region for the check in LoadNode::Ideal();
 339   // 'sub' should have sub->in(0) != NULL.
 340   assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
 341          sub->is_Region(), "expecting only these nodes");
 342 
 343   // Get control edge of 'sub'.
 344   Node* orig_sub = sub;
 345   sub = sub->find_exact_control(sub->in(0));
 346   if (sub == NULL || sub->is_top())
 347     return false; // Conservative answer for dead code
 348 
 349   assert(sub->is_CFG(), "expecting control");
 350 
 351   if (sub == dom)
 352     return true;
 353 
 354   if (sub->is_Start() || sub->is_Root())
 355     return false;
 356 
 357   {
 358     // Check all control edges of 'dom'.
 359 
 360     ResourceMark rm;
 361     Arena* arena = Thread::current()->resource_area();
 362     Node_List nlist(arena);
 363     Unique_Node_List dom_list(arena);
 364 
 365     dom_list.push(dom);
 366     bool only_dominating_controls = false;
 367 
 368     for (uint next = 0; next < dom_list.size(); next++) {
 369       Node* n = dom_list.at(next);
 370       if (n == orig_sub)
 371         return false; // One of dom's inputs dominated by sub.
 372       if (!n->is_CFG() && n->pinned()) {
 373         // Check only own control edge for pinned non-control nodes.
 374         n = n->find_exact_control(n->in(0));
 375         if (n == NULL || n->is_top())
 376           return false; // Conservative answer for dead code
 377         assert(n->is_CFG(), "expecting control");
 378         dom_list.push(n);
 379       } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
 380         only_dominating_controls = true;
 381       } else if (n->is_CFG()) {
 382         if (n->dominates(sub, nlist))
 383           only_dominating_controls = true;
 384         else
 385           return false;
 386       } else {
 387         // First, own control edge.
 388         Node* m = n->find_exact_control(n->in(0));
 389         if (m != NULL) {
 390           if (m->is_top())
 391             return false; // Conservative answer for dead code
 392           dom_list.push(m);
 393         }
 394         // Now, the rest of edges.
 395         uint cnt = n->req();
 396         for (uint i = 1; i < cnt; i++) {
 397           m = n->find_exact_control(n->in(i));
 398           if (m == NULL || m->is_top())
 399             continue;
 400           dom_list.push(m);
 401         }
 402       }
 403     }
 404     return only_dominating_controls;
 405   }
 406 }
 407 
 408 //---------------------detect_ptr_independence---------------------------------
 409 // Used by MemNode::find_previous_store to prove that two base
 410 // pointers are never equal.
 411 // The pointers are accompanied by their associated allocations,
 412 // if any, which have been previously discovered by the caller.
 413 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
 414                                       Node* p2, AllocateNode* a2,
 415                                       PhaseTransform* phase) {
 416   // Attempt to prove that these two pointers cannot be aliased.
 417   // They may both manifestly be allocations, and they should differ.
 418   // Or, if they are not both allocations, they can be distinct constants.
 419   // Otherwise, one is an allocation and the other a pre-existing value.
 420   if (a1 == NULL && a2 == NULL) {           // neither an allocation
 421     return (p1 != p2) && p1->is_Con() && p2->is_Con();
 422   } else if (a1 != NULL && a2 != NULL) {    // both allocations
 423     return (a1 != a2);
 424   } else if (a1 != NULL) {                  // one allocation a1
 425     // (Note:  p2->is_Con implies p2->in(0)->is_Root, which dominates.)
 426     return all_controls_dominate(p2, a1);
 427   } else { //(a2 != NULL)                   // one allocation a2
 428     return all_controls_dominate(p1, a2);
 429   }
 430   return false;
 431 }
 432 
 433 
 434 // The logic for reordering loads and stores uses four steps:
 435 // (a) Walk carefully past stores and initializations which we
 436 //     can prove are independent of this load.
 437 // (b) Observe that the next memory state makes an exact match
 438 //     with self (load or store), and locate the relevant store.
 439 // (c) Ensure that, if we were to wire self directly to the store,
 440 //     the optimizer would fold it up somehow.
 441 // (d) Do the rewiring, and return, depending on some other part of
 442 //     the optimizer to fold up the load.
 443 // This routine handles steps (a) and (b).  Steps (c) and (d) are
 444 // specific to loads and stores, so they are handled by the callers.
 445 // (Currently, only LoadNode::Ideal has steps (c), (d).  More later.)
 446 //
 447 Node* MemNode::find_previous_store(PhaseTransform* phase) {
 448   Node*         ctrl   = in(MemNode::Control);
 449   Node*         adr    = in(MemNode::Address);
 450   intptr_t      offset = 0;
 451   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
 452   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
 453 
 454   if (offset == Type::OffsetBot)
 455     return NULL;            // cannot unalias unless there are precise offsets
 456 
 457   const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
 458 
 459   intptr_t size_in_bytes = memory_size();
 460 
 461   Node* mem = in(MemNode::Memory);   // start searching here...
 462 
 463   int cnt = 50;             // Cycle limiter
 464   for (;;) {                // While we can dance past unrelated stores...
 465     if (--cnt < 0)  break;  // Caught in cycle or a complicated dance?
 466 
 467     if (mem->is_Store()) {
 468       Node* st_adr = mem->in(MemNode::Address);
 469       intptr_t st_offset = 0;
 470       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
 471       if (st_base == NULL)
 472         break;              // inscrutable pointer
 473       if (st_offset != offset && st_offset != Type::OffsetBot) {
 474         const int MAX_STORE = BytesPerLong;
 475         if (st_offset >= offset + size_in_bytes ||
 476             st_offset <= offset - MAX_STORE ||
 477             st_offset <= offset - mem->as_Store()->memory_size()) {
 478           // Success:  The offsets are provably independent.
 479           // (You may ask, why not just test st_offset != offset and be done?
 480           // The answer is that stores of different sizes can co-exist
 481           // in the same sequence of RawMem effects.  We sometimes initialize
 482           // a whole 'tile' of array elements with a single jint or jlong.)
 483           mem = mem->in(MemNode::Memory);
 484           continue;           // (a) advance through independent store memory
 485         }
 486       }
 487       if (st_base != base &&
 488           detect_ptr_independence(base, alloc,
 489                                   st_base,
 490                                   AllocateNode::Ideal_allocation(st_base, phase),
 491                                   phase)) {
 492         // Success:  The bases are provably independent.
 493         mem = mem->in(MemNode::Memory);
 494         continue;           // (a) advance through independent store memory
 495       }
 496 
 497       // (b) At this point, if the bases or offsets do not agree, we lose,
 498       // since we have not managed to prove 'this' and 'mem' independent.
 499       if (st_base == base && st_offset == offset) {
 500         return mem;         // let caller handle steps (c), (d)
 501       }
 502 
 503     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
 504       InitializeNode* st_init = mem->in(0)->as_Initialize();
 505       AllocateNode*  st_alloc = st_init->allocation();
 506       if (st_alloc == NULL)
 507         break;              // something degenerated
 508       bool known_identical = false;
 509       bool known_independent = false;
 510       if (alloc == st_alloc)
 511         known_identical = true;
 512       else if (alloc != NULL)
 513         known_independent = true;
 514       else if (all_controls_dominate(this, st_alloc))
 515         known_independent = true;
 516 
 517       if (known_independent) {
 518         // The bases are provably independent: Either they are
 519         // manifestly distinct allocations, or else the control
 520         // of this load dominates the store's allocation.
 521         int alias_idx = phase->C->get_alias_index(adr_type());
 522         if (alias_idx == Compile::AliasIdxRaw) {
 523           mem = st_alloc->in(TypeFunc::Memory);
 524         } else {
 525           mem = st_init->memory(alias_idx);
 526         }
 527         continue;           // (a) advance through independent store memory
 528       }
 529 
 530       // (b) at this point, if we are not looking at a store initializing
 531       // the same allocation we are loading from, we lose.
 532       if (known_identical) {
 533         // From caller, can_see_stored_value will consult find_captured_store.
 534         return mem;         // let caller handle steps (c), (d)
 535       }
 536 
 537     } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
 538       // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
 539       if (mem->is_Proj() && mem->in(0)->is_Call()) {
 540         CallNode *call = mem->in(0)->as_Call();
 541         if (!call->may_modify(addr_t, phase)) {
 542           mem = call->in(TypeFunc::Memory);
 543           continue;         // (a) advance through independent call memory
 544         }
 545       } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
 546         mem = mem->in(0)->in(TypeFunc::Memory);
 547         continue;           // (a) advance through independent MemBar memory
 548       } else if (mem->is_ClearArray()) {
 549         if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
 550           // (the call updated 'mem' value)
 551           continue;         // (a) advance through independent allocation memory
 552         } else {
 553           // Can not bypass initialization of the instance
 554           // we are looking for.
 555           return mem;
 556         }
 557       } else if (mem->is_MergeMem()) {
 558         int alias_idx = phase->C->get_alias_index(adr_type());
 559         mem = mem->as_MergeMem()->memory_at(alias_idx);
 560         continue;           // (a) advance through independent MergeMem memory
 561       }
 562     }
 563 
 564     // Unless there is an explicit 'continue', we must bail out here,
 565     // because 'mem' is an inscrutable memory state (e.g., a call).
 566     break;
 567   }
 568 
 569   return NULL;              // bail out
 570 }
 571 
 572 //----------------------calculate_adr_type-------------------------------------
 573 // Helper function.  Notices when the given type of address hits top or bottom.
 574 // Also, asserts a cross-check of the type against the expected address type.
 575 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
 576   if (t == Type::TOP)  return NULL; // does not touch memory any more?
 577   #ifdef PRODUCT
 578   cross_check = NULL;
 579   #else
 580   if (!VerifyAliases || is_error_reported() || Node::in_dump())  cross_check = NULL;
 581   #endif
 582   const TypePtr* tp = t->isa_ptr();
 583   if (tp == NULL) {
 584     assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
 585     return TypePtr::BOTTOM;           // touches lots of memory
 586   } else {
 587     #ifdef ASSERT
 588     // %%%% [phh] We don't check the alias index if cross_check is
 589     //            TypeRawPtr::BOTTOM.  Needs to be investigated.
 590     if (cross_check != NULL &&
 591         cross_check != TypePtr::BOTTOM &&
 592         cross_check != TypeRawPtr::BOTTOM) {
 593       // Recheck the alias index, to see if it has changed (due to a bug).
 594       Compile* C = Compile::current();
 595       assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
 596              "must stay in the original alias category");
 597       // The type of the address must be contained in the adr_type,
 598       // disregarding "null"-ness.
 599       // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
 600       const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
 601       assert(cross_check->meet(tp_notnull) == cross_check,
 602              "real address must not escape from expected memory type");
 603     }
 604     #endif
 605     return tp;
 606   }
 607 }
 608 
 609 //------------------------adr_phi_is_loop_invariant----------------------------
 610 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
 611 // loop is loop invariant. Make a quick traversal of Phi and associated
 612 // CastPP nodes, looking to see if they are a closed group within the loop.
 613 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
 614   // The idea is that the phi-nest must boil down to only CastPP nodes
 615   // with the same data. This implies that any path into the loop already
 616   // includes such a CastPP, and so the original cast, whatever its input,
 617   // must be covered by an equivalent cast, with an earlier control input.
 618   ResourceMark rm;
 619 
 620   // The loop entry input of the phi should be the unique dominating
 621   // node for every Phi/CastPP in the loop.
 622   Unique_Node_List closure;
 623   closure.push(adr_phi->in(LoopNode::EntryControl));
 624 
 625   // Add the phi node and the cast to the worklist.
 626   Unique_Node_List worklist;
 627   worklist.push(adr_phi);
 628   if( cast != NULL ){
 629     if( !cast->is_ConstraintCast() ) return false;
 630     worklist.push(cast);
 631   }
 632 
 633   // Begin recursive walk of phi nodes.
 634   while( worklist.size() ){
 635     // Take a node off the worklist
 636     Node *n = worklist.pop();
 637     if( !closure.member(n) ){
 638       // Add it to the closure.
 639       closure.push(n);
 640       // Make a sanity check to ensure we don't waste too much time here.
 641       if( closure.size() > 20) return false;
 642       // This node is OK if:
 643       //  - it is a cast of an identical value
 644       //  - or it is a phi node (then we add its inputs to the worklist)
 645       // Otherwise, the node is not OK, and we presume the cast is not invariant
 646       if( n->is_ConstraintCast() ){
 647         worklist.push(n->in(1));
 648       } else if( n->is_Phi() ) {
 649         for( uint i = 1; i < n->req(); i++ ) {
 650           worklist.push(n->in(i));
 651         }
 652       } else {
 653         return false;
 654       }
 655     }
 656   }
 657 
 658   // Quit when the worklist is empty, and we've found no offending nodes.
 659   return true;
 660 }
 661 
 662 //------------------------------Ideal_DU_postCCP-------------------------------
 663 // Find any cast-away of null-ness and keep its control.  Null cast-aways are
 664 // going away in this pass and we need to make this memory op depend on the
 665 // gating null check.
 666 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
 667   return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
 668 }
 669 
 670 // I tried to leave the CastPP's in.  This makes the graph more accurate in
 671 // some sense; we get to keep around the knowledge that an oop is not-null
 672 // after some test.  Alas, the CastPP's interfere with GVN (some values are
 673 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
 674 // cannot collapse, etc).  This cost us 10% on SpecJVM, even when I removed
 675 // some of the more trivial cases in the optimizer.  Removing more useless
 676 // Phi's started allowing Loads to illegally float above null checks.  I gave
 677 // up on this approach.  CNC 10/20/2000
 678 // This static method may be called not from MemNode (EncodePNode calls it).
 679 // Only the control edge of the node 'n' might be updated.
 680 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
 681   Node *skipped_cast = NULL;
 682   // Need a null check?  Regular static accesses do not because they are
 683   // from constant addresses.  Array ops are gated by the range check (which
 684   // always includes a NULL check).  Just check field ops.
 685   if( n->in(MemNode::Control) == NULL ) {
 686     // Scan upwards for the highest location we can place this memory op.
 687     while( true ) {
 688       switch( adr->Opcode() ) {
 689 
 690       case Op_AddP:             // No change to NULL-ness, so peek thru AddP's
 691         adr = adr->in(AddPNode::Base);
 692         continue;
 693 
 694       case Op_DecodeN:         // No change to NULL-ness, so peek thru
 695         adr = adr->in(1);
 696         continue;
 697 
 698       case Op_CastPP:
 699         // If the CastPP is useless, just peek on through it.
 700         if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
 701           // Remember the cast that we've peeked though. If we peek
 702           // through more than one, then we end up remembering the highest
 703           // one, that is, if in a loop, the one closest to the top.
 704           skipped_cast = adr;
 705           adr = adr->in(1);
 706           continue;
 707         }
 708         // CastPP is going away in this pass!  We need this memory op to be
 709         // control-dependent on the test that is guarding the CastPP.
 710         ccp->hash_delete(n);
 711         n->set_req(MemNode::Control, adr->in(0));
 712         ccp->hash_insert(n);
 713         return n;
 714 
 715       case Op_Phi:
 716         // Attempt to float above a Phi to some dominating point.
 717         if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
 718           // If we've already peeked through a Cast (which could have set the
 719           // control), we can't float above a Phi, because the skipped Cast
 720           // may not be loop invariant.
 721           if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
 722             adr = adr->in(1);
 723             continue;
 724           }
 725         }
 726 
 727         // Intentional fallthrough!
 728 
 729         // No obvious dominating point.  The mem op is pinned below the Phi
 730         // by the Phi itself.  If the Phi goes away (no true value is merged)
 731         // then the mem op can float, but not indefinitely.  It must be pinned
 732         // behind the controls leading to the Phi.
 733       case Op_CheckCastPP:
 734         // These usually stick around to change address type, however a
 735         // useless one can be elided and we still need to pick up a control edge
 736         if (adr->in(0) == NULL) {
 737           // This CheckCastPP node has NO control and is likely useless. But we
 738           // need check further up the ancestor chain for a control input to keep
 739           // the node in place. 4959717.
 740           skipped_cast = adr;
 741           adr = adr->in(1);
 742           continue;
 743         }
 744         ccp->hash_delete(n);
 745         n->set_req(MemNode::Control, adr->in(0));
 746         ccp->hash_insert(n);
 747         return n;
 748 
 749         // List of "safe" opcodes; those that implicitly block the memory
 750         // op below any null check.
 751       case Op_CastX2P:          // no null checks on native pointers
 752       case Op_Parm:             // 'this' pointer is not null
 753       case Op_LoadP:            // Loading from within a klass
 754       case Op_LoadN:            // Loading from within a klass
 755       case Op_LoadKlass:        // Loading from within a klass
 756       case Op_LoadNKlass:       // Loading from within a klass
 757       case Op_ConP:             // Loading from a klass
 758       case Op_ConN:             // Loading from a klass
 759       case Op_CreateEx:         // Sucking up the guts of an exception oop
 760       case Op_Con:              // Reading from TLS
 761       case Op_CMoveP:           // CMoveP is pinned
 762       case Op_CMoveN:           // CMoveN is pinned
 763         break;                  // No progress
 764 
 765       case Op_Proj:             // Direct call to an allocation routine
 766       case Op_SCMemProj:        // Memory state from store conditional ops
 767 #ifdef ASSERT
 768         {
 769           assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
 770           const Node* call = adr->in(0);
 771           if (call->is_CallJava()) {
 772             const CallJavaNode* call_java = call->as_CallJava();
 773             const TypeTuple *r = call_java->tf()->range();
 774             assert(r->cnt() > TypeFunc::Parms, "must return value");
 775             const Type* ret_type = r->field_at(TypeFunc::Parms);
 776             assert(ret_type && ret_type->isa_ptr(), "must return pointer");
 777             // We further presume that this is one of
 778             // new_instance_Java, new_array_Java, or
 779             // the like, but do not assert for this.
 780           } else if (call->is_Allocate()) {
 781             // similar case to new_instance_Java, etc.
 782           } else if (!call->is_CallLeaf()) {
 783             // Projections from fetch_oop (OSR) are allowed as well.
 784             ShouldNotReachHere();
 785           }
 786         }
 787 #endif
 788         break;
 789       default:
 790         ShouldNotReachHere();
 791       }
 792       break;
 793     }
 794   }
 795 
 796   return  NULL;               // No progress
 797 }
 798 
 799 
 800 //=============================================================================
 801 uint LoadNode::size_of() const { return sizeof(*this); }
 802 uint LoadNode::cmp( const Node &n ) const
 803 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
 804 const Type *LoadNode::bottom_type() const { return _type; }
 805 uint LoadNode::ideal_reg() const {
 806   return Matcher::base2reg[_type->base()];
 807 }
 808 
 809 #ifndef PRODUCT
 810 void LoadNode::dump_spec(outputStream *st) const {
 811   MemNode::dump_spec(st);
 812   if( !Verbose && !WizardMode ) {
 813     // standard dump does this in Verbose and WizardMode
 814     st->print(" #"); _type->dump_on(st);
 815   }
 816 }
 817 #endif
 818 
 819 #ifdef ASSERT
 820 //----------------------------is_immutable_value-------------------------------
 821 // Helper function to allow a raw load without control edge for some cases
 822 bool LoadNode::is_immutable_value(Node* adr) {
 823   return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
 824           adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 825           (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
 826            in_bytes(JavaThread::osthread_offset())));
 827 }
 828 #endif
 829 
 830 //----------------------------LoadNode::make-----------------------------------
 831 // Polymorphic factory method:
 832 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
 833   Compile* C = gvn.C;
 834 
 835   // sanity check the alias category against the created node type
 836   assert(!(adr_type->isa_oopptr() &&
 837            adr_type->offset() == oopDesc::klass_offset_in_bytes()),
 838          "use LoadKlassNode instead");
 839   assert(!(adr_type->isa_aryptr() &&
 840            adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
 841          "use LoadRangeNode instead");
 842   // Check control edge of raw loads
 843   assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
 844           // oop will be recorded in oop map if load crosses safepoint
 845           rt->isa_oopptr() || is_immutable_value(adr),
 846           "raw memory operations should have control edge");
 847   switch (bt) {
 848   case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int()    );
 849   case T_BYTE:    return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int()    );
 850   case T_INT:     return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int()    );
 851   case T_CHAR:    return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int()    );
 852   case T_SHORT:   return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int()    );
 853   case T_LONG:    return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long()   );
 854   case T_FLOAT:   return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt              );
 855   case T_DOUBLE:  return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt              );
 856   case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr()    );
 857   case T_OBJECT:
 858 #ifdef _LP64
 859     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
 860       Node* load  = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
 861       return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
 862     } else
 863 #endif
 864     {
 865       assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
 866       return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
 867     }
 868   }
 869   ShouldNotReachHere();
 870   return (LoadNode*)NULL;
 871 }
 872 
 873 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
 874   bool require_atomic = true;
 875   return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
 876 }
 877 
 878 
 879 
 880 
 881 //------------------------------hash-------------------------------------------
 882 uint LoadNode::hash() const {
 883   // unroll addition of interesting fields
 884   return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
 885 }
 886 
 887 //---------------------------can_see_stored_value------------------------------
 888 // This routine exists to make sure this set of tests is done the same
 889 // everywhere.  We need to make a coordinated change: first LoadNode::Ideal
 890 // will change the graph shape in a way which makes memory alive twice at the
 891 // same time (uses the Oracle model of aliasing), then some
 892 // LoadXNode::Identity will fold things back to the equivalence-class model
 893 // of aliasing.
 894 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
 895   Node* ld_adr = in(MemNode::Address);
 896 
 897   const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
 898   Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
 899   if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
 900       atp->field() != NULL && !atp->field()->is_volatile()) {
 901     uint alias_idx = atp->index();
 902     bool final = atp->field()->is_final();
 903     Node* result = NULL;
 904     Node* current = st;
 905     // Skip through chains of MemBarNodes checking the MergeMems for
 906     // new states for the slice of this load.  Stop once any other
 907     // kind of node is encountered.  Loads from final memory can skip
 908     // through any kind of MemBar but normal loads shouldn't skip
 909     // through MemBarAcquire since the could allow them to move out of
 910     // a synchronized region.
 911     while (current->is_Proj()) {
 912       int opc = current->in(0)->Opcode();
 913       if ((final && opc == Op_MemBarAcquire) ||
 914           opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
 915         Node* mem = current->in(0)->in(TypeFunc::Memory);
 916         if (mem->is_MergeMem()) {
 917           MergeMemNode* merge = mem->as_MergeMem();
 918           Node* new_st = merge->memory_at(alias_idx);
 919           if (new_st == merge->base_memory()) {
 920             // Keep searching
 921             current = merge->base_memory();
 922             continue;
 923           }
 924           // Save the new memory state for the slice and fall through
 925           // to exit.
 926           result = new_st;
 927         }
 928       }
 929       break;
 930     }
 931     if (result != NULL) {
 932       st = result;
 933     }
 934   }
 935 
 936 
 937   // Loop around twice in the case Load -> Initialize -> Store.
 938   // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
 939   for (int trip = 0; trip <= 1; trip++) {
 940 
 941     if (st->is_Store()) {
 942       Node* st_adr = st->in(MemNode::Address);
 943       if (!phase->eqv(st_adr, ld_adr)) {
 944         // Try harder before giving up...  Match raw and non-raw pointers.
 945         intptr_t st_off = 0;
 946         AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
 947         if (alloc == NULL)       return NULL;
 948         intptr_t ld_off = 0;
 949         AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
 950         if (alloc != allo2)      return NULL;
 951         if (ld_off != st_off)    return NULL;
 952         // At this point we have proven something like this setup:
 953         //  A = Allocate(...)
 954         //  L = LoadQ(,  AddP(CastPP(, A.Parm),, #Off))
 955         //  S = StoreQ(, AddP(,        A.Parm  , #Off), V)
 956         // (Actually, we haven't yet proven the Q's are the same.)
 957         // In other words, we are loading from a casted version of
 958         // the same pointer-and-offset that we stored to.
 959         // Thus, we are able to replace L by V.
 960       }
 961       // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
 962       if (store_Opcode() != st->Opcode())
 963         return NULL;
 964       return st->in(MemNode::ValueIn);
 965     }
 966 
 967     intptr_t offset = 0;  // scratch
 968 
 969     // A load from a freshly-created object always returns zero.
 970     // (This can happen after LoadNode::Ideal resets the load's memory input
 971     // to find_captured_store, which returned InitializeNode::zero_memory.)
 972     if (st->is_Proj() && st->in(0)->is_Allocate() &&
 973         st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
 974         offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
 975       // return a zero value for the load's basic type
 976       // (This is one of the few places where a generic PhaseTransform
 977       // can create new nodes.  Think of it as lazily manifesting
 978       // virtually pre-existing constants.)
 979       return phase->zerocon(memory_type());
 980     }
 981 
 982     // A load from an initialization barrier can match a captured store.
 983     if (st->is_Proj() && st->in(0)->is_Initialize()) {
 984       InitializeNode* init = st->in(0)->as_Initialize();
 985       AllocateNode* alloc = init->allocation();
 986       if (alloc != NULL &&
 987           alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
 988         // examine a captured store value
 989         st = init->find_captured_store(offset, memory_size(), phase);
 990         if (st != NULL)
 991           continue;             // take one more trip around
 992       }
 993     }
 994 
 995     break;
 996   }
 997 
 998   return NULL;
 999 }
1000 
1001 //----------------------is_instance_field_load_with_local_phi------------------
1002 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1003   if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
1004       in(MemNode::Address)->is_AddP() ) {
1005     const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
1006     // Only instances.
1007     if( t_oop != NULL && t_oop->is_known_instance_field() &&
1008         t_oop->offset() != Type::OffsetBot &&
1009         t_oop->offset() != Type::OffsetTop) {
1010       return true;
1011     }
1012   }
1013   return false;
1014 }
1015 
1016 //------------------------------Identity---------------------------------------
1017 // Loads are identity if previous store is to same address
1018 Node *LoadNode::Identity( PhaseTransform *phase ) {
1019   // If the previous store-maker is the right kind of Store, and the store is
1020   // to the same address, then we are equal to the value stored.
1021   Node* mem = in(MemNode::Memory);
1022   Node* value = can_see_stored_value(mem, phase);
1023   if( value ) {
1024     // byte, short & char stores truncate naturally.
1025     // A load has to load the truncated value which requires
1026     // some sort of masking operation and that requires an
1027     // Ideal call instead of an Identity call.
1028     if (memory_size() < BytesPerInt) {
1029       // If the input to the store does not fit with the load's result type,
1030       // it must be truncated via an Ideal call.
1031       if (!phase->type(value)->higher_equal(phase->type(this)))
1032         return this;
1033     }
1034     // (This works even when value is a Con, but LoadNode::Value
1035     // usually runs first, producing the singleton type of the Con.)
1036     return value;
1037   }
1038 
1039   // Search for an existing data phi which was generated before for the same
1040   // instance's field to avoid infinite generation of phis in a loop.
1041   Node *region = mem->in(0);
1042   if (is_instance_field_load_with_local_phi(region)) {
1043     const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
1044     int this_index  = phase->C->get_alias_index(addr_t);
1045     int this_offset = addr_t->offset();
1046     int this_id    = addr_t->is_oopptr()->instance_id();
1047     const Type* this_type = bottom_type();
1048     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1049       Node* phi = region->fast_out(i);
1050       if (phi->is_Phi() && phi != mem &&
1051           phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
1052         return phi;
1053       }
1054     }
1055   }
1056 
1057   return this;
1058 }
1059 
1060 
1061 // Returns true if the AliasType refers to the field that holds the
1062 // cached box array.  Currently only handles the IntegerCache case.
1063 static bool is_autobox_cache(Compile::AliasType* atp) {
1064   if (atp != NULL && atp->field() != NULL) {
1065     ciField* field = atp->field();
1066     ciSymbol* klass = field->holder()->name();
1067     if (field->name() == ciSymbol::cache_field_name() &&
1068         field->holder()->uses_default_loader() &&
1069         klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1070       return true;
1071     }
1072   }
1073   return false;
1074 }
1075 
1076 // Fetch the base value in the autobox array
1077 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
1078   if (atp != NULL && atp->field() != NULL) {
1079     ciField* field = atp->field();
1080     ciSymbol* klass = field->holder()->name();
1081     if (field->name() == ciSymbol::cache_field_name() &&
1082         field->holder()->uses_default_loader() &&
1083         klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1084       assert(field->is_constant(), "what?");
1085       ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1086       // Fetch the box object at the base of the array and get its value
1087       ciInstance* box = array->obj_at(0)->as_instance();
1088       ciInstanceKlass* ik = box->klass()->as_instance_klass();
1089       if (ik->nof_nonstatic_fields() == 1) {
1090         // This should be true nonstatic_field_at requires calling
1091         // nof_nonstatic_fields so check it anyway
1092         ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1093         cache_offset = c.as_int();
1094       }
1095       return true;
1096     }
1097   }
1098   return false;
1099 }
1100 
1101 // Returns true if the AliasType refers to the value field of an
1102 // autobox object.  Currently only handles Integer.
1103 static bool is_autobox_object(Compile::AliasType* atp) {
1104   if (atp != NULL && atp->field() != NULL) {
1105     ciField* field = atp->field();
1106     ciSymbol* klass = field->holder()->name();
1107     if (field->name() == ciSymbol::value_name() &&
1108         field->holder()->uses_default_loader() &&
1109         klass == ciSymbol::java_lang_Integer()) {
1110       return true;
1111     }
1112   }
1113   return false;
1114 }
1115 
1116 
1117 // We're loading from an object which has autobox behaviour.
1118 // If this object is result of a valueOf call we'll have a phi
1119 // merging a newly allocated object and a load from the cache.
1120 // We want to replace this load with the original incoming
1121 // argument to the valueOf call.
1122 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1123   Node* base = in(Address)->in(AddPNode::Base);
1124   if (base->is_Phi() && base->req() == 3) {
1125     AllocateNode* allocation = NULL;
1126     int allocation_index = -1;
1127     int load_index = -1;
1128     for (uint i = 1; i < base->req(); i++) {
1129       allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1130       if (allocation != NULL) {
1131         allocation_index = i;
1132         load_index = 3 - allocation_index;
1133         break;
1134       }
1135     }
1136     bool has_load = ( allocation != NULL &&
1137                       (base->in(load_index)->is_Load() ||
1138                        base->in(load_index)->is_DecodeN() &&
1139                        base->in(load_index)->in(1)->is_Load()) );
1140     if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1141       // Push the loads from the phi that comes from valueOf up
1142       // through it to allow elimination of the loads and the recovery
1143       // of the original value.
1144       Node* mem_phi = in(Memory);
1145       Node* offset = in(Address)->in(AddPNode::Offset);
1146       Node* region = base->in(0);
1147 
1148       Node* in1 = clone();
1149       Node* in1_addr = in1->in(Address)->clone();
1150       in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1151       in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1152       in1_addr->set_req(AddPNode::Offset, offset);
1153       in1->set_req(0, region->in(allocation_index));
1154       in1->set_req(Address, in1_addr);
1155       in1->set_req(Memory, mem_phi->in(allocation_index));
1156 
1157       Node* in2 = clone();
1158       Node* in2_addr = in2->in(Address)->clone();
1159       in2_addr->set_req(AddPNode::Base, base->in(load_index));
1160       in2_addr->set_req(AddPNode::Address, base->in(load_index));
1161       in2_addr->set_req(AddPNode::Offset, offset);
1162       in2->set_req(0, region->in(load_index));
1163       in2->set_req(Address, in2_addr);
1164       in2->set_req(Memory, mem_phi->in(load_index));
1165 
1166       in1_addr = phase->transform(in1_addr);
1167       in1 =      phase->transform(in1);
1168       in2_addr = phase->transform(in2_addr);
1169       in2 =      phase->transform(in2);
1170 
1171       PhiNode* result = PhiNode::make_blank(region, this);
1172       result->set_req(allocation_index, in1);
1173       result->set_req(load_index, in2);
1174       return result;
1175     }
1176   } else if (base->is_Load() ||
1177              base->is_DecodeN() && base->in(1)->is_Load()) {
1178     if (base->is_DecodeN()) {
1179       // Get LoadN node which loads cached Integer object
1180       base = base->in(1);
1181     }
1182     // Eliminate the load of Integer.value for integers from the cache
1183     // array by deriving the value from the index into the array.
1184     // Capture the offset of the load and then reverse the computation.
1185     Node* load_base = base->in(Address)->in(AddPNode::Base);
1186     if (load_base->is_DecodeN()) {
1187       // Get LoadN node which loads IntegerCache.cache field
1188       load_base = load_base->in(1);
1189     }
1190     if (load_base != NULL) {
1191       Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1192       intptr_t cache_offset;
1193       int shift = -1;
1194       Node* cache = NULL;
1195       if (is_autobox_cache(atp)) {
1196         shift  = exact_log2(type2aelembytes(T_OBJECT));
1197         cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1198       }
1199       if (cache != NULL && base->in(Address)->is_AddP()) {
1200         Node* elements[4];
1201         int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1202         int cache_low;
1203         if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1204           int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1205           // Add up all the offsets making of the address of the load
1206           Node* result = elements[0];
1207           for (int i = 1; i < count; i++) {
1208             result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1209           }
1210           // Remove the constant offset from the address and then
1211           // remove the scaling of the offset to recover the original index.
1212           result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1213           if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1214             // Peel the shift off directly but wrap it in a dummy node
1215             // since Ideal can't return existing nodes
1216             result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1217           } else {
1218             result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1219           }
1220 #ifdef _LP64
1221           result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1222 #endif
1223           return result;
1224         }
1225       }
1226     }
1227   }
1228   return NULL;
1229 }
1230 
1231 //------------------------------split_through_phi------------------------------
1232 // Split instance field load through Phi.
1233 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1234   Node* mem     = in(MemNode::Memory);
1235   Node* address = in(MemNode::Address);
1236   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1237   const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1238 
1239   assert(mem->is_Phi() && (t_oop != NULL) &&
1240          t_oop->is_known_instance_field(), "invalide conditions");
1241 
1242   Node *region = mem->in(0);
1243   if (region == NULL) {
1244     return NULL; // Wait stable graph
1245   }
1246   uint cnt = mem->req();
1247   for( uint i = 1; i < cnt; i++ ) {
1248     Node *in = mem->in(i);
1249     if( in == NULL ) {
1250       return NULL; // Wait stable graph
1251     }
1252   }
1253   // Check for loop invariant.
1254   if (cnt == 3) {
1255     for( uint i = 1; i < cnt; i++ ) {
1256       Node *in = mem->in(i);
1257       Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1258       if (m == mem) {
1259         set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1260         return this;
1261       }
1262     }
1263   }
1264   // Split through Phi (see original code in loopopts.cpp).
1265   assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1266 
1267   // Do nothing here if Identity will find a value
1268   // (to avoid infinite chain of value phis generation).
1269   if ( !phase->eqv(this, this->Identity(phase)) )
1270     return NULL;
1271 
1272   // Skip the split if the region dominates some control edge of the address.
1273   if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
1274     return NULL;
1275 
1276   const Type* this_type = this->bottom_type();
1277   int this_index  = phase->C->get_alias_index(addr_t);
1278   int this_offset = addr_t->offset();
1279   int this_iid    = addr_t->is_oopptr()->instance_id();
1280   int wins = 0;
1281   PhaseIterGVN *igvn = phase->is_IterGVN();
1282   Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1283   for( uint i = 1; i < region->req(); i++ ) {
1284     Node *x;
1285     Node* the_clone = NULL;
1286     if( region->in(i) == phase->C->top() ) {
1287       x = phase->C->top();      // Dead path?  Use a dead data op
1288     } else {
1289       x = this->clone();        // Else clone up the data op
1290       the_clone = x;            // Remember for possible deletion.
1291       // Alter data node to use pre-phi inputs
1292       if( this->in(0) == region ) {
1293         x->set_req( 0, region->in(i) );
1294       } else {
1295         x->set_req( 0, NULL );
1296       }
1297       for( uint j = 1; j < this->req(); j++ ) {
1298         Node *in = this->in(j);
1299         if( in->is_Phi() && in->in(0) == region )
1300           x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1301       }
1302     }
1303     // Check for a 'win' on some paths
1304     const Type *t = x->Value(igvn);
1305 
1306     bool singleton = t->singleton();
1307 
1308     // See comments in PhaseIdealLoop::split_thru_phi().
1309     if( singleton && t == Type::TOP ) {
1310       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1311     }
1312 
1313     if( singleton ) {
1314       wins++;
1315       x = igvn->makecon(t);
1316     } else {
1317       // We now call Identity to try to simplify the cloned node.
1318       // Note that some Identity methods call phase->type(this).
1319       // Make sure that the type array is big enough for
1320       // our new node, even though we may throw the node away.
1321       // (This tweaking with igvn only works because x is a new node.)
1322       igvn->set_type(x, t);
1323       // If x is a TypeNode, capture any more-precise type permanently into Node
1324       // otherwise it will be not updated during igvn->transform since
1325       // igvn->type(x) is set to x->Value() already.
1326       x->raise_bottom_type(t);
1327       Node *y = x->Identity(igvn);
1328       if( y != x ) {
1329         wins++;
1330         x = y;
1331       } else {
1332         y = igvn->hash_find(x);
1333         if( y ) {
1334           wins++;
1335           x = y;
1336         } else {
1337           // Else x is a new node we are keeping
1338           // We do not need register_new_node_with_optimizer
1339           // because set_type has already been called.
1340           igvn->_worklist.push(x);
1341         }
1342       }
1343     }
1344     if (x != the_clone && the_clone != NULL)
1345       igvn->remove_dead_node(the_clone);
1346     phi->set_req(i, x);
1347   }
1348   if( wins > 0 ) {
1349     // Record Phi
1350     igvn->register_new_node_with_optimizer(phi);
1351     return phi;
1352   }
1353   igvn->remove_dead_node(phi);
1354   return NULL;
1355 }
1356 
1357 //------------------------------Ideal------------------------------------------
1358 // If the load is from Field memory and the pointer is non-null, we can
1359 // zero out the control input.
1360 // If the offset is constant and the base is an object allocation,
1361 // try to hook me up to the exact initializing store.
1362 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1363   Node* p = MemNode::Ideal_common(phase, can_reshape);
1364   if (p)  return (p == NodeSentinel) ? NULL : p;
1365 
1366   Node* ctrl    = in(MemNode::Control);
1367   Node* address = in(MemNode::Address);
1368 
1369   // Skip up past a SafePoint control.  Cannot do this for Stores because
1370   // pointer stores & cardmarks must stay on the same side of a SafePoint.
1371   if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1372       phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1373     ctrl = ctrl->in(0);
1374     set_req(MemNode::Control,ctrl);
1375   }
1376 
1377   intptr_t ignore = 0;
1378   Node*    base   = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1379   if (base != NULL
1380       && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1381     // Check for useless control edge in some common special cases
1382     if (in(MemNode::Control) != NULL
1383         && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1384         && all_controls_dominate(base, phase->C->start())) {
1385       // A method-invariant, non-null address (constant or 'this' argument).
1386       set_req(MemNode::Control, NULL);
1387     }
1388 
1389     if (EliminateAutoBox && can_reshape) {
1390       assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
1391       Compile::AliasType* atp = phase->C->alias_type(adr_type());
1392       if (is_autobox_object(atp)) {
1393         Node* result = eliminate_autobox(phase);
1394         if (result != NULL) return result;
1395       }
1396     }
1397   }
1398 
1399   Node* mem = in(MemNode::Memory);
1400   const TypePtr *addr_t = phase->type(address)->isa_ptr();
1401 
1402   if (addr_t != NULL) {
1403     // try to optimize our memory input
1404     Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1405     if (opt_mem != mem) {
1406       set_req(MemNode::Memory, opt_mem);
1407       if (phase->type( opt_mem ) == Type::TOP) return NULL;
1408       return this;
1409     }
1410     const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1411     if (can_reshape && opt_mem->is_Phi() &&
1412         (t_oop != NULL) && t_oop->is_known_instance_field()) {
1413       // Split instance field load through Phi.
1414       Node* result = split_through_phi(phase);
1415       if (result != NULL) return result;
1416     }
1417   }
1418 
1419   // Check for prior store with a different base or offset; make Load
1420   // independent.  Skip through any number of them.  Bail out if the stores
1421   // are in an endless dead cycle and report no progress.  This is a key
1422   // transform for Reflection.  However, if after skipping through the Stores
1423   // we can't then fold up against a prior store do NOT do the transform as
1424   // this amounts to using the 'Oracle' model of aliasing.  It leaves the same
1425   // array memory alive twice: once for the hoisted Load and again after the
1426   // bypassed Store.  This situation only works if EVERYBODY who does
1427   // anti-dependence work knows how to bypass.  I.e. we need all
1428   // anti-dependence checks to ask the same Oracle.  Right now, that Oracle is
1429   // the alias index stuff.  So instead, peek through Stores and IFF we can
1430   // fold up, do so.
1431   Node* prev_mem = find_previous_store(phase);
1432   // Steps (a), (b):  Walk past independent stores to find an exact match.
1433   if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1434     // (c) See if we can fold up on the spot, but don't fold up here.
1435     // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1436     // just return a prior value, which is done by Identity calls.
1437     if (can_see_stored_value(prev_mem, phase)) {
1438       // Make ready for step (d):
1439       set_req(MemNode::Memory, prev_mem);
1440       return this;
1441     }
1442   }
1443 
1444   return NULL;                  // No further progress
1445 }
1446 
1447 // Helper to recognize certain Klass fields which are invariant across
1448 // some group of array types (e.g., int[] or all T[] where T < Object).
1449 const Type*
1450 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1451                                  ciKlass* klass) const {
1452   if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1453     // The field is Klass::_modifier_flags.  Return its (constant) value.
1454     // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1455     assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1456     return TypeInt::make(klass->modifier_flags());
1457   }
1458   if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1459     // The field is Klass::_access_flags.  Return its (constant) value.
1460     // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1461     assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1462     return TypeInt::make(klass->access_flags());
1463   }
1464   if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1465     // The field is Klass::_layout_helper.  Return its constant value if known.
1466     assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1467     return TypeInt::make(klass->layout_helper());
1468   }
1469 
1470   // No match.
1471   return NULL;
1472 }
1473 
1474 //------------------------------Value-----------------------------------------
1475 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1476   // Either input is TOP ==> the result is TOP
1477   Node* mem = in(MemNode::Memory);
1478   const Type *t1 = phase->type(mem);
1479   if (t1 == Type::TOP)  return Type::TOP;
1480   Node* adr = in(MemNode::Address);
1481   const TypePtr* tp = phase->type(adr)->isa_ptr();
1482   if (tp == NULL || tp->empty())  return Type::TOP;
1483   int off = tp->offset();
1484   assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1485 
1486   // Try to guess loaded type from pointer type
1487   if (tp->base() == Type::AryPtr) {
1488     const Type *t = tp->is_aryptr()->elem();
1489     // Don't do this for integer types. There is only potential profit if
1490     // the element type t is lower than _type; that is, for int types, if _type is
1491     // more restrictive than t.  This only happens here if one is short and the other
1492     // char (both 16 bits), and in those cases we've made an intentional decision
1493     // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1494     // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1495     //
1496     // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1497     // where the _gvn.type of the AddP is wider than 8.  This occurs when an earlier
1498     // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1499     // subsumed by p1.  If p1 is on the worklist but has not yet been re-transformed,
1500     // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1501     // In fact, that could have been the original type of p1, and p1 could have
1502     // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1503     // expression (LShiftL quux 3) independently optimized to the constant 8.
1504     if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1505         && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1506       // t might actually be lower than _type, if _type is a unique
1507       // concrete subclass of abstract class t.
1508       // Make sure the reference is not into the header, by comparing
1509       // the offset against the offset of the start of the array's data.
1510       // Different array types begin at slightly different offsets (12 vs. 16).
1511       // We choose T_BYTE as an example base type that is least restrictive
1512       // as to alignment, which will therefore produce the smallest
1513       // possible base offset.
1514       const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1515       if ((uint)off >= (uint)min_base_off) {  // is the offset beyond the header?
1516         const Type* jt = t->join(_type);
1517         // In any case, do not allow the join, per se, to empty out the type.
1518         if (jt->empty() && !t->empty()) {
1519           // This can happen if a interface-typed array narrows to a class type.
1520           jt = _type;
1521         }
1522 
1523         if (EliminateAutoBox && adr->is_AddP()) {
1524           // The pointers in the autobox arrays are always non-null
1525           Node* base = adr->in(AddPNode::Base);
1526           if (base != NULL &&
1527               !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
1528             Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1529             if (is_autobox_cache(atp)) {
1530               return jt->join(TypePtr::NOTNULL)->is_ptr();
1531             }
1532           }
1533         }
1534         return jt;
1535       }
1536     }
1537   } else if (tp->base() == Type::InstPtr) {
1538     const TypeInstPtr* tinst = tp->is_instptr();
1539     ciKlass* klass = tinst->klass();
1540     assert( off != Type::OffsetBot ||
1541             // arrays can be cast to Objects
1542             tp->is_oopptr()->klass()->is_java_lang_Object() ||
1543             // unsafe field access may not have a constant offset
1544             phase->C->has_unsafe_access(),
1545             "Field accesses must be precise" );
1546     // For oop loads, we expect the _type to be precise
1547     if (OptimizeStringConcat && klass == phase->C->env()->String_klass() &&
1548         adr->is_AddP() && off != Type::OffsetBot) {
1549       // For constant Strings treat the fields as compile time constants.
1550       Node* base = adr->in(AddPNode::Base);
1551       const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1552       if (t != NULL && t->singleton()) {
1553         ciObject* string = t->const_oop();
1554         ciConstant constant = string->as_instance()->field_value_by_offset(off);
1555         if (constant.basic_type() == T_INT) {
1556           return TypeInt::make(constant.as_int());
1557         } else if (constant.basic_type() == T_ARRAY) {
1558           if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1559             return TypeNarrowOop::make_from_constant(constant.as_object());
1560           } else {
1561             return TypeOopPtr::make_from_constant(constant.as_object());
1562           }
1563         }
1564       }
1565     }
1566   } else if (tp->base() == Type::KlassPtr) {
1567     assert( off != Type::OffsetBot ||
1568             // arrays can be cast to Objects
1569             tp->is_klassptr()->klass()->is_java_lang_Object() ||
1570             // also allow array-loading from the primary supertype
1571             // array during subtype checks
1572             Opcode() == Op_LoadKlass,
1573             "Field accesses must be precise" );
1574     // For klass/static loads, we expect the _type to be precise
1575   }
1576 
1577   const TypeKlassPtr *tkls = tp->isa_klassptr();
1578   if (tkls != NULL && !StressReflectiveCode) {
1579     ciKlass* klass = tkls->klass();
1580     if (klass->is_loaded() && tkls->klass_is_exact()) {
1581       // We are loading a field from a Klass metaobject whose identity
1582       // is known at compile time (the type is "exact" or "precise").
1583       // Check for fields we know are maintained as constants by the VM.
1584       if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1585         // The field is Klass::_super_check_offset.  Return its (constant) value.
1586         // (Folds up type checking code.)
1587         assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1588         return TypeInt::make(klass->super_check_offset());
1589       }
1590       // Compute index into primary_supers array
1591       juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1592       // Check for overflowing; use unsigned compare to handle the negative case.
1593       if( depth < ciKlass::primary_super_limit() ) {
1594         // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1595         // (Folds up type checking code.)
1596         assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1597         ciKlass *ss = klass->super_of_depth(depth);
1598         return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1599       }
1600       const Type* aift = load_array_final_field(tkls, klass);
1601       if (aift != NULL)  return aift;
1602       if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1603           && klass->is_array_klass()) {
1604         // The field is arrayKlass::_component_mirror.  Return its (constant) value.
1605         // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1606         assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1607         return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1608       }
1609       if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1610         // The field is Klass::_java_mirror.  Return its (constant) value.
1611         // (Folds up the 2nd indirection in anObjConstant.getClass().)
1612         assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1613         return TypeInstPtr::make(klass->java_mirror());
1614       }
1615     }
1616 
1617     // We can still check if we are loading from the primary_supers array at a
1618     // shallow enough depth.  Even though the klass is not exact, entries less
1619     // than or equal to its super depth are correct.
1620     if (klass->is_loaded() ) {
1621       ciType *inner = klass->klass();
1622       while( inner->is_obj_array_klass() )
1623         inner = inner->as_obj_array_klass()->base_element_type();
1624       if( inner->is_instance_klass() &&
1625           !inner->as_instance_klass()->flags().is_interface() ) {
1626         // Compute index into primary_supers array
1627         juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1628         // Check for overflowing; use unsigned compare to handle the negative case.
1629         if( depth < ciKlass::primary_super_limit() &&
1630             depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1631           // The field is an element of Klass::_primary_supers.  Return its (constant) value.
1632           // (Folds up type checking code.)
1633           assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1634           ciKlass *ss = klass->super_of_depth(depth);
1635           return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1636         }
1637       }
1638     }
1639 
1640     // If the type is enough to determine that the thing is not an array,
1641     // we can give the layout_helper a positive interval type.
1642     // This will help short-circuit some reflective code.
1643     if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1644         && !klass->is_array_klass() // not directly typed as an array
1645         && !klass->is_interface()  // specifically not Serializable & Cloneable
1646         && !klass->is_java_lang_Object()   // not the supertype of all T[]
1647         ) {
1648       // Note:  When interfaces are reliable, we can narrow the interface
1649       // test to (klass != Serializable && klass != Cloneable).
1650       assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1651       jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1652       // The key property of this type is that it folds up tests
1653       // for array-ness, since it proves that the layout_helper is positive.
1654       // Thus, a generic value like the basic object layout helper works fine.
1655       return TypeInt::make(min_size, max_jint, Type::WidenMin);
1656     }
1657   }
1658 
1659   // If we are loading from a freshly-allocated object, produce a zero,
1660   // if the load is provably beyond the header of the object.
1661   // (Also allow a variable load from a fresh array to produce zero.)
1662   if (ReduceFieldZeroing) {
1663     Node* value = can_see_stored_value(mem,phase);
1664     if (value != NULL && value->is_Con())
1665       return value->bottom_type();
1666   }
1667 
1668   const TypeOopPtr *tinst = tp->isa_oopptr();
1669   if (tinst != NULL && tinst->is_known_instance_field()) {
1670     // If we have an instance type and our memory input is the
1671     // programs's initial memory state, there is no matching store,
1672     // so just return a zero of the appropriate type
1673     Node *mem = in(MemNode::Memory);
1674     if (mem->is_Parm() && mem->in(0)->is_Start()) {
1675       assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1676       return Type::get_zero_type(_type->basic_type());
1677     }
1678   }
1679   return _type;
1680 }
1681 
1682 //------------------------------match_edge-------------------------------------
1683 // Do we Match on this edge index or not?  Match only the address.
1684 uint LoadNode::match_edge(uint idx) const {
1685   return idx == MemNode::Address;
1686 }
1687 
1688 //--------------------------LoadBNode::Ideal--------------------------------------
1689 //
1690 //  If the previous store is to the same address as this load,
1691 //  and the value stored was larger than a byte, replace this load
1692 //  with the value stored truncated to a byte.  If no truncation is
1693 //  needed, the replacement is done in LoadNode::Identity().
1694 //
1695 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1696   Node* mem = in(MemNode::Memory);
1697   Node* value = can_see_stored_value(mem,phase);
1698   if( value && !phase->type(value)->higher_equal( _type ) ) {
1699     Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1700     return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1701   }
1702   // Identity call will handle the case where truncation is not needed.
1703   return LoadNode::Ideal(phase, can_reshape);
1704 }
1705 
1706 //--------------------------LoadUBNode::Ideal-------------------------------------
1707 //
1708 //  If the previous store is to the same address as this load,
1709 //  and the value stored was larger than a byte, replace this load
1710 //  with the value stored truncated to a byte.  If no truncation is
1711 //  needed, the replacement is done in LoadNode::Identity().
1712 //
1713 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1714   Node* mem = in(MemNode::Memory);
1715   Node* value = can_see_stored_value(mem, phase);
1716   if (value && !phase->type(value)->higher_equal(_type))
1717     return new (phase->C, 3) AndINode(value, phase->intcon(0xFF));
1718   // Identity call will handle the case where truncation is not needed.
1719   return LoadNode::Ideal(phase, can_reshape);
1720 }
1721 
1722 //--------------------------LoadUSNode::Ideal-------------------------------------
1723 //
1724 //  If the previous store is to the same address as this load,
1725 //  and the value stored was larger than a char, replace this load
1726 //  with the value stored truncated to a char.  If no truncation is
1727 //  needed, the replacement is done in LoadNode::Identity().
1728 //
1729 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1730   Node* mem = in(MemNode::Memory);
1731   Node* value = can_see_stored_value(mem,phase);
1732   if( value && !phase->type(value)->higher_equal( _type ) )
1733     return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1734   // Identity call will handle the case where truncation is not needed.
1735   return LoadNode::Ideal(phase, can_reshape);
1736 }
1737 
1738 //--------------------------LoadSNode::Ideal--------------------------------------
1739 //
1740 //  If the previous store is to the same address as this load,
1741 //  and the value stored was larger than a short, replace this load
1742 //  with the value stored truncated to a short.  If no truncation is
1743 //  needed, the replacement is done in LoadNode::Identity().
1744 //
1745 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1746   Node* mem = in(MemNode::Memory);
1747   Node* value = can_see_stored_value(mem,phase);
1748   if( value && !phase->type(value)->higher_equal( _type ) ) {
1749     Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1750     return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1751   }
1752   // Identity call will handle the case where truncation is not needed.
1753   return LoadNode::Ideal(phase, can_reshape);
1754 }
1755 
1756 //=============================================================================
1757 //----------------------------LoadKlassNode::make------------------------------
1758 // Polymorphic factory method:
1759 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1760   Compile* C = gvn.C;
1761   Node *ctl = NULL;
1762   // sanity check the alias category against the created node type
1763   const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
1764   assert(adr_type != NULL, "expecting TypeOopPtr");
1765 #ifdef _LP64
1766   if (adr_type->is_ptr_to_narrowoop()) {
1767     Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
1768     return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
1769   }
1770 #endif
1771   assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1772   return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
1773 }
1774 
1775 //------------------------------Value------------------------------------------
1776 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1777   return klass_value_common(phase);
1778 }
1779 
1780 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1781   // Either input is TOP ==> the result is TOP
1782   const Type *t1 = phase->type( in(MemNode::Memory) );
1783   if (t1 == Type::TOP)  return Type::TOP;
1784   Node *adr = in(MemNode::Address);
1785   const Type *t2 = phase->type( adr );
1786   if (t2 == Type::TOP)  return Type::TOP;
1787   const TypePtr *tp = t2->is_ptr();
1788   if (TypePtr::above_centerline(tp->ptr()) ||
1789       tp->ptr() == TypePtr::Null)  return Type::TOP;
1790 
1791   // Return a more precise klass, if possible
1792   const TypeInstPtr *tinst = tp->isa_instptr();
1793   if (tinst != NULL) {
1794     ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1795     int offset = tinst->offset();
1796     if (ik == phase->C->env()->Class_klass()
1797         && (offset == java_lang_Class::klass_offset_in_bytes() ||
1798             offset == java_lang_Class::array_klass_offset_in_bytes())) {
1799       // We are loading a special hidden field from a Class mirror object,
1800       // the field which points to the VM's Klass metaobject.
1801       ciType* t = tinst->java_mirror_type();
1802       // java_mirror_type returns non-null for compile-time Class constants.
1803       if (t != NULL) {
1804         // constant oop => constant klass
1805         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1806           return TypeKlassPtr::make(ciArrayKlass::make(t));
1807         }
1808         if (!t->is_klass()) {
1809           // a primitive Class (e.g., int.class) has NULL for a klass field
1810           return TypePtr::NULL_PTR;
1811         }
1812         // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1813         return TypeKlassPtr::make(t->as_klass());
1814       }
1815       // non-constant mirror, so we can't tell what's going on
1816     }
1817     if( !ik->is_loaded() )
1818       return _type;             // Bail out if not loaded
1819     if (offset == oopDesc::klass_offset_in_bytes()) {
1820       if (tinst->klass_is_exact()) {
1821         return TypeKlassPtr::make(ik);
1822       }
1823       // See if we can become precise: no subklasses and no interface
1824       // (Note:  We need to support verified interfaces.)
1825       if (!ik->is_interface() && !ik->has_subklass()) {
1826         //assert(!UseExactTypes, "this code should be useless with exact types");
1827         // Add a dependence; if any subclass added we need to recompile
1828         if (!ik->is_final()) {
1829           // %%% should use stronger assert_unique_concrete_subtype instead
1830           phase->C->dependencies()->assert_leaf_type(ik);
1831         }
1832         // Return precise klass
1833         return TypeKlassPtr::make(ik);
1834       }
1835 
1836       // Return root of possible klass
1837       return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1838     }
1839   }
1840 
1841   // Check for loading klass from an array
1842   const TypeAryPtr *tary = tp->isa_aryptr();
1843   if( tary != NULL ) {
1844     ciKlass *tary_klass = tary->klass();
1845     if (tary_klass != NULL   // can be NULL when at BOTTOM or TOP
1846         && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1847       if (tary->klass_is_exact()) {
1848         return TypeKlassPtr::make(tary_klass);
1849       }
1850       ciArrayKlass *ak = tary->klass()->as_array_klass();
1851       // If the klass is an object array, we defer the question to the
1852       // array component klass.
1853       if( ak->is_obj_array_klass() ) {
1854         assert( ak->is_loaded(), "" );
1855         ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1856         if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1857           ciInstanceKlass* ik = base_k->as_instance_klass();
1858           // See if we can become precise: no subklasses and no interface
1859           if (!ik->is_interface() && !ik->has_subklass()) {
1860             //assert(!UseExactTypes, "this code should be useless with exact types");
1861             // Add a dependence; if any subclass added we need to recompile
1862             if (!ik->is_final()) {
1863               phase->C->dependencies()->assert_leaf_type(ik);
1864             }
1865             // Return precise array klass
1866             return TypeKlassPtr::make(ak);
1867           }
1868         }
1869         return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1870       } else {                  // Found a type-array?
1871         //assert(!UseExactTypes, "this code should be useless with exact types");
1872         assert( ak->is_type_array_klass(), "" );
1873         return TypeKlassPtr::make(ak); // These are always precise
1874       }
1875     }
1876   }
1877 
1878   // Check for loading klass from an array klass
1879   const TypeKlassPtr *tkls = tp->isa_klassptr();
1880   if (tkls != NULL && !StressReflectiveCode) {
1881     ciKlass* klass = tkls->klass();
1882     if( !klass->is_loaded() )
1883       return _type;             // Bail out if not loaded
1884     if( klass->is_obj_array_klass() &&
1885         (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1886       ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1887       // // Always returning precise element type is incorrect,
1888       // // e.g., element type could be object and array may contain strings
1889       // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1890 
1891       // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1892       // according to the element type's subclassing.
1893       return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1894     }
1895     if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1896         (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1897       ciKlass* sup = klass->as_instance_klass()->super();
1898       // The field is Klass::_super.  Return its (constant) value.
1899       // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1900       return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1901     }
1902   }
1903 
1904   // Bailout case
1905   return LoadNode::Value(phase);
1906 }
1907 
1908 //------------------------------Identity---------------------------------------
1909 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1910 // Also feed through the klass in Allocate(...klass...)._klass.
1911 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1912   return klass_identity_common(phase);
1913 }
1914 
1915 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
1916   Node* x = LoadNode::Identity(phase);
1917   if (x != this)  return x;
1918 
1919   // Take apart the address into an oop and and offset.
1920   // Return 'this' if we cannot.
1921   Node*    adr    = in(MemNode::Address);
1922   intptr_t offset = 0;
1923   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1924   if (base == NULL)     return this;
1925   const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1926   if (toop == NULL)     return this;
1927 
1928   // We can fetch the klass directly through an AllocateNode.
1929   // This works even if the klass is not constant (clone or newArray).
1930   if (offset == oopDesc::klass_offset_in_bytes()) {
1931     Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1932     if (allocated_klass != NULL) {
1933       return allocated_klass;
1934     }
1935   }
1936 
1937   // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1938   // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1939   // See inline_native_Class_query for occurrences of these patterns.
1940   // Java Example:  x.getClass().isAssignableFrom(y)
1941   // Java Example:  Array.newInstance(x.getClass().getComponentType(), n)
1942   //
1943   // This improves reflective code, often making the Class
1944   // mirror go completely dead.  (Current exception:  Class
1945   // mirrors may appear in debug info, but we could clean them out by
1946   // introducing a new debug info operator for klassOop.java_mirror).
1947   if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1948       && (offset == java_lang_Class::klass_offset_in_bytes() ||
1949           offset == java_lang_Class::array_klass_offset_in_bytes())) {
1950     // We are loading a special hidden field from a Class mirror,
1951     // the field which points to its Klass or arrayKlass metaobject.
1952     if (base->is_Load()) {
1953       Node* adr2 = base->in(MemNode::Address);
1954       const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1955       if (tkls != NULL && !tkls->empty()
1956           && (tkls->klass()->is_instance_klass() ||
1957               tkls->klass()->is_array_klass())
1958           && adr2->is_AddP()
1959           ) {
1960         int mirror_field = Klass::java_mirror_offset_in_bytes();
1961         if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1962           mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1963         }
1964         if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1965           return adr2->in(AddPNode::Base);
1966         }
1967       }
1968     }
1969   }
1970 
1971   return this;
1972 }
1973 
1974 
1975 //------------------------------Value------------------------------------------
1976 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
1977   const Type *t = klass_value_common(phase);
1978   if (t == Type::TOP)
1979     return t;
1980 
1981   return t->make_narrowoop();
1982 }
1983 
1984 //------------------------------Identity---------------------------------------
1985 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
1986 // Also feed through the klass in Allocate(...klass...)._klass.
1987 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
1988   Node *x = klass_identity_common(phase);
1989 
1990   const Type *t = phase->type( x );
1991   if( t == Type::TOP ) return x;
1992   if( t->isa_narrowoop()) return x;
1993 
1994   return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
1995 }
1996 
1997 //------------------------------Value-----------------------------------------
1998 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1999   // Either input is TOP ==> the result is TOP
2000   const Type *t1 = phase->type( in(MemNode::Memory) );
2001   if( t1 == Type::TOP ) return Type::TOP;
2002   Node *adr = in(MemNode::Address);
2003   const Type *t2 = phase->type( adr );
2004   if( t2 == Type::TOP ) return Type::TOP;
2005   const TypePtr *tp = t2->is_ptr();
2006   if (TypePtr::above_centerline(tp->ptr()))  return Type::TOP;
2007   const TypeAryPtr *tap = tp->isa_aryptr();
2008   if( !tap ) return _type;
2009   return tap->size();
2010 }
2011 
2012 //-------------------------------Ideal---------------------------------------
2013 // Feed through the length in AllocateArray(...length...)._length.
2014 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2015   Node* p = MemNode::Ideal_common(phase, can_reshape);
2016   if (p)  return (p == NodeSentinel) ? NULL : p;
2017 
2018   // Take apart the address into an oop and and offset.
2019   // Return 'this' if we cannot.
2020   Node*    adr    = in(MemNode::Address);
2021   intptr_t offset = 0;
2022   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase,  offset);
2023   if (base == NULL)     return NULL;
2024   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2025   if (tary == NULL)     return NULL;
2026 
2027   // We can fetch the length directly through an AllocateArrayNode.
2028   // This works even if the length is not constant (clone or newArray).
2029   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2030     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2031     if (alloc != NULL) {
2032       Node* allocated_length = alloc->Ideal_length();
2033       Node* len = alloc->make_ideal_length(tary, phase);
2034       if (allocated_length != len) {
2035         // New CastII improves on this.
2036         return len;
2037       }
2038     }
2039   }
2040 
2041   return NULL;
2042 }
2043 
2044 //------------------------------Identity---------------------------------------
2045 // Feed through the length in AllocateArray(...length...)._length.
2046 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2047   Node* x = LoadINode::Identity(phase);
2048   if (x != this)  return x;
2049 
2050   // Take apart the address into an oop and and offset.
2051   // Return 'this' if we cannot.
2052   Node*    adr    = in(MemNode::Address);
2053   intptr_t offset = 0;
2054   Node*    base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2055   if (base == NULL)     return this;
2056   const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2057   if (tary == NULL)     return this;
2058 
2059   // We can fetch the length directly through an AllocateArrayNode.
2060   // This works even if the length is not constant (clone or newArray).
2061   if (offset == arrayOopDesc::length_offset_in_bytes()) {
2062     AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2063     if (alloc != NULL) {
2064       Node* allocated_length = alloc->Ideal_length();
2065       // Do not allow make_ideal_length to allocate a CastII node.
2066       Node* len = alloc->make_ideal_length(tary, phase, false);
2067       if (allocated_length == len) {
2068         // Return allocated_length only if it would not be improved by a CastII.
2069         return allocated_length;
2070       }
2071     }
2072   }
2073 
2074   return this;
2075 
2076 }
2077 
2078 //=============================================================================
2079 //---------------------------StoreNode::make-----------------------------------
2080 // Polymorphic factory method:
2081 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
2082   Compile* C = gvn.C;
2083   assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2084           ctl != NULL, "raw memory operations should have control edge");
2085 
2086   switch (bt) {
2087   case T_BOOLEAN:
2088   case T_BYTE:    return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
2089   case T_INT:     return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
2090   case T_CHAR:
2091   case T_SHORT:   return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
2092   case T_LONG:    return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
2093   case T_FLOAT:   return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
2094   case T_DOUBLE:  return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
2095   case T_ADDRESS:
2096   case T_OBJECT:
2097 #ifdef _LP64
2098     if (adr->bottom_type()->is_ptr_to_narrowoop() ||
2099         (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
2100          adr->bottom_type()->isa_rawptr())) {
2101       val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2102       return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
2103     } else
2104 #endif
2105     {
2106       return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
2107     }
2108   }
2109   ShouldNotReachHere();
2110   return (StoreNode*)NULL;
2111 }
2112 
2113 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
2114   bool require_atomic = true;
2115   return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
2116 }
2117 
2118 
2119 //--------------------------bottom_type----------------------------------------
2120 const Type *StoreNode::bottom_type() const {
2121   return Type::MEMORY;
2122 }
2123 
2124 //------------------------------hash-------------------------------------------
2125 uint StoreNode::hash() const {
2126   // unroll addition of interesting fields
2127   //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2128 
2129   // Since they are not commoned, do not hash them:
2130   return NO_HASH;
2131 }
2132 
2133 //------------------------------Ideal------------------------------------------
2134 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2135 // When a store immediately follows a relevant allocation/initialization,
2136 // try to capture it into the initialization, or hoist it above.
2137 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2138   Node* p = MemNode::Ideal_common(phase, can_reshape);
2139   if (p)  return (p == NodeSentinel) ? NULL : p;
2140 
2141   Node* mem     = in(MemNode::Memory);
2142   Node* address = in(MemNode::Address);
2143 
2144   // Back-to-back stores to same address?  Fold em up.
2145   // Generally unsafe if I have intervening uses...
2146   if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
2147     // Looking at a dead closed cycle of memory?
2148     assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2149 
2150     assert(Opcode() == mem->Opcode() ||
2151            phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2152            "no mismatched stores, except on raw memory");
2153 
2154     if (mem->outcnt() == 1 &&           // check for intervening uses
2155         mem->as_Store()->memory_size() <= this->memory_size()) {
2156       // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2157       // For example, 'mem' might be the final state at a conditional return.
2158       // Or, 'mem' might be used by some node which is live at the same time
2159       // 'this' is live, which might be unschedulable.  So, require exactly
2160       // ONE user, the 'this' store, until such time as we clone 'mem' for
2161       // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2162       if (can_reshape) {  // (%%% is this an anachronism?)
2163         set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2164                   phase->is_IterGVN());
2165       } else {
2166         // It's OK to do this in the parser, since DU info is always accurate,
2167         // and the parser always refers to nodes via SafePointNode maps.
2168         set_req(MemNode::Memory, mem->in(MemNode::Memory));
2169       }
2170       return this;
2171     }
2172   }
2173 
2174   // Capture an unaliased, unconditional, simple store into an initializer.
2175   // Or, if it is independent of the allocation, hoist it above the allocation.
2176   if (ReduceFieldZeroing && /*can_reshape &&*/
2177       mem->is_Proj() && mem->in(0)->is_Initialize()) {
2178     InitializeNode* init = mem->in(0)->as_Initialize();
2179     intptr_t offset = init->can_capture_store(this, phase);
2180     if (offset > 0) {
2181       Node* moved = init->capture_store(this, offset, phase);
2182       // If the InitializeNode captured me, it made a raw copy of me,
2183       // and I need to disappear.
2184       if (moved != NULL) {
2185         // %%% hack to ensure that Ideal returns a new node:
2186         mem = MergeMemNode::make(phase->C, mem);
2187         return mem;             // fold me away
2188       }
2189     }
2190   }
2191 
2192   return NULL;                  // No further progress
2193 }
2194 
2195 //------------------------------Value-----------------------------------------
2196 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2197   // Either input is TOP ==> the result is TOP
2198   const Type *t1 = phase->type( in(MemNode::Memory) );
2199   if( t1 == Type::TOP ) return Type::TOP;
2200   const Type *t2 = phase->type( in(MemNode::Address) );
2201   if( t2 == Type::TOP ) return Type::TOP;
2202   const Type *t3 = phase->type( in(MemNode::ValueIn) );
2203   if( t3 == Type::TOP ) return Type::TOP;
2204   return Type::MEMORY;
2205 }
2206 
2207 //------------------------------Identity---------------------------------------
2208 // Remove redundant stores:
2209 //   Store(m, p, Load(m, p)) changes to m.
2210 //   Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2211 Node *StoreNode::Identity( PhaseTransform *phase ) {
2212   Node* mem = in(MemNode::Memory);
2213   Node* adr = in(MemNode::Address);
2214   Node* val = in(MemNode::ValueIn);
2215 
2216   // Load then Store?  Then the Store is useless
2217   if (val->is_Load() &&
2218       phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
2219       phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
2220       val->as_Load()->store_Opcode() == Opcode()) {
2221     return mem;
2222   }
2223 
2224   // Two stores in a row of the same value?
2225   if (mem->is_Store() &&
2226       phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
2227       phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
2228       mem->Opcode() == Opcode()) {
2229     return mem;
2230   }
2231 
2232   // Store of zero anywhere into a freshly-allocated object?
2233   // Then the store is useless.
2234   // (It must already have been captured by the InitializeNode.)
2235   if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2236     // a newly allocated object is already all-zeroes everywhere
2237     if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2238       return mem;
2239     }
2240 
2241     // the store may also apply to zero-bits in an earlier object
2242     Node* prev_mem = find_previous_store(phase);
2243     // Steps (a), (b):  Walk past independent stores to find an exact match.
2244     if (prev_mem != NULL) {
2245       Node* prev_val = can_see_stored_value(prev_mem, phase);
2246       if (prev_val != NULL && phase->eqv(prev_val, val)) {
2247         // prev_val and val might differ by a cast; it would be good
2248         // to keep the more informative of the two.
2249         return mem;
2250       }
2251     }
2252   }
2253 
2254   return this;
2255 }
2256 
2257 //------------------------------match_edge-------------------------------------
2258 // Do we Match on this edge index or not?  Match only memory & value
2259 uint StoreNode::match_edge(uint idx) const {
2260   return idx == MemNode::Address || idx == MemNode::ValueIn;
2261 }
2262 
2263 //------------------------------cmp--------------------------------------------
2264 // Do not common stores up together.  They generally have to be split
2265 // back up anyways, so do not bother.
2266 uint StoreNode::cmp( const Node &n ) const {
2267   return (&n == this);          // Always fail except on self
2268 }
2269 
2270 //------------------------------Ideal_masked_input-----------------------------
2271 // Check for a useless mask before a partial-word store
2272 // (StoreB ... (AndI valIn conIa) )
2273 // If (conIa & mask == mask) this simplifies to
2274 // (StoreB ... (valIn) )
2275 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2276   Node *val = in(MemNode::ValueIn);
2277   if( val->Opcode() == Op_AndI ) {
2278     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2279     if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2280       set_req(MemNode::ValueIn, val->in(1));
2281       return this;
2282     }
2283   }
2284   return NULL;
2285 }
2286 
2287 
2288 //------------------------------Ideal_sign_extended_input----------------------
2289 // Check for useless sign-extension before a partial-word store
2290 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2291 // If (conIL == conIR && conIR <= num_bits)  this simplifies to
2292 // (StoreB ... (valIn) )
2293 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2294   Node *val = in(MemNode::ValueIn);
2295   if( val->Opcode() == Op_RShiftI ) {
2296     const TypeInt *t = phase->type( val->in(2) )->isa_int();
2297     if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2298       Node *shl = val->in(1);
2299       if( shl->Opcode() == Op_LShiftI ) {
2300         const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2301         if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2302           set_req(MemNode::ValueIn, shl->in(1));
2303           return this;
2304         }
2305       }
2306     }
2307   }
2308   return NULL;
2309 }
2310 
2311 //------------------------------value_never_loaded-----------------------------------
2312 // Determine whether there are any possible loads of the value stored.
2313 // For simplicity, we actually check if there are any loads from the
2314 // address stored to, not just for loads of the value stored by this node.
2315 //
2316 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2317   Node *adr = in(Address);
2318   const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2319   if (adr_oop == NULL)
2320     return false;
2321   if (!adr_oop->is_known_instance_field())
2322     return false; // if not a distinct instance, there may be aliases of the address
2323   for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2324     Node *use = adr->fast_out(i);
2325     int opc = use->Opcode();
2326     if (use->is_Load() || use->is_LoadStore()) {
2327       return false;
2328     }
2329   }
2330   return true;
2331 }
2332 
2333 //=============================================================================
2334 //------------------------------Ideal------------------------------------------
2335 // If the store is from an AND mask that leaves the low bits untouched, then
2336 // we can skip the AND operation.  If the store is from a sign-extension
2337 // (a left shift, then right shift) we can skip both.
2338 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2339   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2340   if( progress != NULL ) return progress;
2341 
2342   progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2343   if( progress != NULL ) return progress;
2344 
2345   // Finally check the default case
2346   return StoreNode::Ideal(phase, can_reshape);
2347 }
2348 
2349 //=============================================================================
2350 //------------------------------Ideal------------------------------------------
2351 // If the store is from an AND mask that leaves the low bits untouched, then
2352 // we can skip the AND operation
2353 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2354   Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2355   if( progress != NULL ) return progress;
2356 
2357   progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2358   if( progress != NULL ) return progress;
2359 
2360   // Finally check the default case
2361   return StoreNode::Ideal(phase, can_reshape);
2362 }
2363 
2364 //=============================================================================
2365 //------------------------------Identity---------------------------------------
2366 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2367   // No need to card mark when storing a null ptr
2368   Node* my_store = in(MemNode::OopStore);
2369   if (my_store->is_Store()) {
2370     const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2371     if( t1 == TypePtr::NULL_PTR ) {
2372       return in(MemNode::Memory);
2373     }
2374   }
2375   return this;
2376 }
2377 
2378 //=============================================================================
2379 //------------------------------Ideal---------------------------------------
2380 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2381   Node* progress = StoreNode::Ideal(phase, can_reshape);
2382   if (progress != NULL) return progress;
2383 
2384   Node* my_store = in(MemNode::OopStore);
2385   if (my_store->is_MergeMem()) {
2386     Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2387     set_req(MemNode::OopStore, mem);
2388     return this;
2389   }
2390 
2391   return NULL;
2392 }
2393 
2394 //------------------------------Value-----------------------------------------
2395 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2396   // Either input is TOP ==> the result is TOP
2397   const Type *t = phase->type( in(MemNode::Memory) );
2398   if( t == Type::TOP ) return Type::TOP;
2399   t = phase->type( in(MemNode::Address) );
2400   if( t == Type::TOP ) return Type::TOP;
2401   t = phase->type( in(MemNode::ValueIn) );
2402   if( t == Type::TOP ) return Type::TOP;
2403   // If extra input is TOP ==> the result is TOP
2404   t = phase->type( in(MemNode::OopStore) );
2405   if( t == Type::TOP ) return Type::TOP;
2406 
2407   return StoreNode::Value( phase );
2408 }
2409 
2410 
2411 //=============================================================================
2412 //----------------------------------SCMemProjNode------------------------------
2413 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2414 {
2415   return bottom_type();
2416 }
2417 
2418 //=============================================================================
2419 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2420   init_req(MemNode::Control, c  );
2421   init_req(MemNode::Memory , mem);
2422   init_req(MemNode::Address, adr);
2423   init_req(MemNode::ValueIn, val);
2424   init_req(         ExpectedIn, ex );
2425   init_class_id(Class_LoadStore);
2426 
2427 }
2428 
2429 //=============================================================================
2430 //-------------------------------adr_type--------------------------------------
2431 // Do we Match on this edge index or not?  Do not match memory
2432 const TypePtr* ClearArrayNode::adr_type() const {
2433   Node *adr = in(3);
2434   return MemNode::calculate_adr_type(adr->bottom_type());
2435 }
2436 
2437 //------------------------------match_edge-------------------------------------
2438 // Do we Match on this edge index or not?  Do not match memory
2439 uint ClearArrayNode::match_edge(uint idx) const {
2440   return idx > 1;
2441 }
2442 
2443 //------------------------------Identity---------------------------------------
2444 // Clearing a zero length array does nothing
2445 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2446   return phase->type(in(2))->higher_equal(TypeX::ZERO)  ? in(1) : this;
2447 }
2448 
2449 //------------------------------Idealize---------------------------------------
2450 // Clearing a short array is faster with stores
2451 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2452   const int unit = BytesPerLong;
2453   const TypeX* t = phase->type(in(2))->isa_intptr_t();
2454   if (!t)  return NULL;
2455   if (!t->is_con())  return NULL;
2456   intptr_t raw_count = t->get_con();
2457   intptr_t size = raw_count;
2458   if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2459   // Clearing nothing uses the Identity call.
2460   // Negative clears are possible on dead ClearArrays
2461   // (see jck test stmt114.stmt11402.val).
2462   if (size <= 0 || size % unit != 0)  return NULL;
2463   intptr_t count = size / unit;
2464   // Length too long; use fast hardware clear
2465   if (size > Matcher::init_array_short_size)  return NULL;
2466   Node *mem = in(1);
2467   if( phase->type(mem)==Type::TOP ) return NULL;
2468   Node *adr = in(3);
2469   const Type* at = phase->type(adr);
2470   if( at==Type::TOP ) return NULL;
2471   const TypePtr* atp = at->isa_ptr();
2472   // adjust atp to be the correct array element address type
2473   if (atp == NULL)  atp = TypePtr::BOTTOM;
2474   else              atp = atp->add_offset(Type::OffsetBot);
2475   // Get base for derived pointer purposes
2476   if( adr->Opcode() != Op_AddP ) Unimplemented();
2477   Node *base = adr->in(1);
2478 
2479   Node *zero = phase->makecon(TypeLong::ZERO);
2480   Node *off  = phase->MakeConX(BytesPerLong);
2481   mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2482   count--;
2483   while( count-- ) {
2484     mem = phase->transform(mem);
2485     adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2486     mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2487   }
2488   return mem;
2489 }
2490 
2491 //----------------------------step_through----------------------------------
2492 // Return allocation input memory edge if it is different instance
2493 // or itself if it is the one we are looking for.
2494 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2495   Node* n = *np;
2496   assert(n->is_ClearArray(), "sanity");
2497   intptr_t offset;
2498   AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2499   // This method is called only before Allocate nodes are expanded during
2500   // macro nodes expansion. Before that ClearArray nodes are only generated
2501   // in LibraryCallKit::generate_arraycopy() which follows allocations.
2502   assert(alloc != NULL, "should have allocation");
2503   if (alloc->_idx == instance_id) {
2504     // Can not bypass initialization of the instance we are looking for.
2505     return false;
2506   }
2507   // Otherwise skip it.
2508   InitializeNode* init = alloc->initialization();
2509   if (init != NULL)
2510     *np = init->in(TypeFunc::Memory);
2511   else
2512     *np = alloc->in(TypeFunc::Memory);
2513   return true;
2514 }
2515 
2516 //----------------------------clear_memory-------------------------------------
2517 // Generate code to initialize object storage to zero.
2518 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2519                                    intptr_t start_offset,
2520                                    Node* end_offset,
2521                                    PhaseGVN* phase) {
2522   Compile* C = phase->C;
2523   intptr_t offset = start_offset;
2524 
2525   int unit = BytesPerLong;
2526   if ((offset % unit) != 0) {
2527     Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2528     adr = phase->transform(adr);
2529     const TypePtr* atp = TypeRawPtr::BOTTOM;
2530     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2531     mem = phase->transform(mem);
2532     offset += BytesPerInt;
2533   }
2534   assert((offset % unit) == 0, "");
2535 
2536   // Initialize the remaining stuff, if any, with a ClearArray.
2537   return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2538 }
2539 
2540 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2541                                    Node* start_offset,
2542                                    Node* end_offset,
2543                                    PhaseGVN* phase) {
2544   if (start_offset == end_offset) {
2545     // nothing to do
2546     return mem;
2547   }
2548 
2549   Compile* C = phase->C;
2550   int unit = BytesPerLong;
2551   Node* zbase = start_offset;
2552   Node* zend  = end_offset;
2553 
2554   // Scale to the unit required by the CPU:
2555   if (!Matcher::init_array_count_is_in_bytes) {
2556     Node* shift = phase->intcon(exact_log2(unit));
2557     zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2558     zend  = phase->transform( new(C,3) URShiftXNode(zend,  shift) );
2559   }
2560 
2561   Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2562   Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2563 
2564   // Bulk clear double-words
2565   Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2566   mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2567   return phase->transform(mem);
2568 }
2569 
2570 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2571                                    intptr_t start_offset,
2572                                    intptr_t end_offset,
2573                                    PhaseGVN* phase) {
2574   if (start_offset == end_offset) {
2575     // nothing to do
2576     return mem;
2577   }
2578 
2579   Compile* C = phase->C;
2580   assert((end_offset % BytesPerInt) == 0, "odd end offset");
2581   intptr_t done_offset = end_offset;
2582   if ((done_offset % BytesPerLong) != 0) {
2583     done_offset -= BytesPerInt;
2584   }
2585   if (done_offset > start_offset) {
2586     mem = clear_memory(ctl, mem, dest,
2587                        start_offset, phase->MakeConX(done_offset), phase);
2588   }
2589   if (done_offset < end_offset) { // emit the final 32-bit store
2590     Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2591     adr = phase->transform(adr);
2592     const TypePtr* atp = TypeRawPtr::BOTTOM;
2593     mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2594     mem = phase->transform(mem);
2595     done_offset += BytesPerInt;
2596   }
2597   assert(done_offset == end_offset, "");
2598   return mem;
2599 }
2600 
2601 //=============================================================================
2602 // Do we match on this edge? No memory edges
2603 uint StrCompNode::match_edge(uint idx) const {
2604   return idx == 2 || idx == 3; // StrComp (Binary str1 cnt1) (Binary str2 cnt2)
2605 }
2606 
2607 //------------------------------Ideal------------------------------------------
2608 // Return a node which is more "ideal" than the current node.  Strip out
2609 // control copies
2610 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2611   return remove_dead_region(phase, can_reshape) ? this : NULL;
2612 }
2613 
2614 //=============================================================================
2615 // Do we match on this edge? No memory edges
2616 uint StrEqualsNode::match_edge(uint idx) const {
2617   return idx == 2 || idx == 3; // StrEquals (Binary str1 str2) cnt
2618 }
2619 
2620 //------------------------------Ideal------------------------------------------
2621 // Return a node which is more "ideal" than the current node.  Strip out
2622 // control copies
2623 Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){
2624   return remove_dead_region(phase, can_reshape) ? this : NULL;
2625 }
2626 
2627 //=============================================================================
2628 // Do we match on this edge? No memory edges
2629 uint StrIndexOfNode::match_edge(uint idx) const {
2630   return idx == 2 || idx == 3; // StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2)
2631 }
2632 
2633 //------------------------------Ideal------------------------------------------
2634 // Return a node which is more "ideal" than the current node.  Strip out
2635 // control copies
2636 Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){
2637   return remove_dead_region(phase, can_reshape) ? this : NULL;
2638 }
2639 
2640 //=============================================================================
2641 // Do we match on this edge? No memory edges
2642 uint AryEqNode::match_edge(uint idx) const {
2643   return idx == 2 || idx == 3; // StrEquals ary1 ary2
2644 }
2645 //------------------------------Ideal------------------------------------------
2646 // Return a node which is more "ideal" than the current node.  Strip out
2647 // control copies
2648 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
2649   return remove_dead_region(phase, can_reshape) ? this : NULL;
2650 }
2651 
2652 //=============================================================================
2653 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2654   : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2655     _adr_type(C->get_adr_type(alias_idx))
2656 {
2657   init_class_id(Class_MemBar);
2658   Node* top = C->top();
2659   init_req(TypeFunc::I_O,top);
2660   init_req(TypeFunc::FramePtr,top);
2661   init_req(TypeFunc::ReturnAdr,top);
2662   if (precedent != NULL)
2663     init_req(TypeFunc::Parms, precedent);
2664 }
2665 
2666 //------------------------------cmp--------------------------------------------
2667 uint MemBarNode::hash() const { return NO_HASH; }
2668 uint MemBarNode::cmp( const Node &n ) const {
2669   return (&n == this);          // Always fail except on self
2670 }
2671 
2672 //------------------------------make-------------------------------------------
2673 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2674   int len = Precedent + (pn == NULL? 0: 1);
2675   switch (opcode) {
2676   case Op_MemBarAcquire:   return new(C, len) MemBarAcquireNode(C,  atp, pn);
2677   case Op_MemBarRelease:   return new(C, len) MemBarReleaseNode(C,  atp, pn);
2678   case Op_MemBarVolatile:  return new(C, len) MemBarVolatileNode(C, atp, pn);
2679   case Op_MemBarCPUOrder:  return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2680   case Op_Initialize:      return new(C, len) InitializeNode(C,     atp, pn);
2681   default:                 ShouldNotReachHere(); return NULL;
2682   }
2683 }
2684 
2685 //------------------------------Ideal------------------------------------------
2686 // Return a node which is more "ideal" than the current node.  Strip out
2687 // control copies
2688 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2689   if (remove_dead_region(phase, can_reshape)) return this;
2690 
2691   // Eliminate volatile MemBars for scalar replaced objects.
2692   if (can_reshape && req() == (Precedent+1) &&
2693       (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) {
2694     // Volatile field loads and stores.
2695     Node* my_mem = in(MemBarNode::Precedent);
2696     if (my_mem != NULL && my_mem->is_Mem()) {
2697       const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2698       // Check for scalar replaced object reference.
2699       if( t_oop != NULL && t_oop->is_known_instance_field() &&
2700           t_oop->offset() != Type::OffsetBot &&
2701           t_oop->offset() != Type::OffsetTop) {
2702         // Replace MemBar projections by its inputs.
2703         PhaseIterGVN* igvn = phase->is_IterGVN();
2704         igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2705         igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2706         // Must return either the original node (now dead) or a new node
2707         // (Do not return a top here, since that would break the uniqueness of top.)
2708         return new (phase->C, 1) ConINode(TypeInt::ZERO);
2709       }
2710     }
2711   }
2712   return NULL;
2713 }
2714 
2715 //------------------------------Value------------------------------------------
2716 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2717   if( !in(0) ) return Type::TOP;
2718   if( phase->type(in(0)) == Type::TOP )
2719     return Type::TOP;
2720   return TypeTuple::MEMBAR;
2721 }
2722 
2723 //------------------------------match------------------------------------------
2724 // Construct projections for memory.
2725 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2726   switch (proj->_con) {
2727   case TypeFunc::Control:
2728   case TypeFunc::Memory:
2729     return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2730   }
2731   ShouldNotReachHere();
2732   return NULL;
2733 }
2734 
2735 //===========================InitializeNode====================================
2736 // SUMMARY:
2737 // This node acts as a memory barrier on raw memory, after some raw stores.
2738 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2739 // The Initialize can 'capture' suitably constrained stores as raw inits.
2740 // It can coalesce related raw stores into larger units (called 'tiles').
2741 // It can avoid zeroing new storage for memory units which have raw inits.
2742 // At macro-expansion, it is marked 'complete', and does not optimize further.
2743 //
2744 // EXAMPLE:
2745 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2746 //   ctl = incoming control; mem* = incoming memory
2747 // (Note:  A star * on a memory edge denotes I/O and other standard edges.)
2748 // First allocate uninitialized memory and fill in the header:
2749 //   alloc = (Allocate ctl mem* 16 #short[].klass ...)
2750 //   ctl := alloc.Control; mem* := alloc.Memory*
2751 //   rawmem = alloc.Memory; rawoop = alloc.RawAddress
2752 // Then initialize to zero the non-header parts of the raw memory block:
2753 //   init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2754 //   ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2755 // After the initialize node executes, the object is ready for service:
2756 //   oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2757 // Suppose its body is immediately initialized as {1,2}:
2758 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2759 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
2760 //   mem.SLICE(#short[*]) := store2
2761 //
2762 // DETAILS:
2763 // An InitializeNode collects and isolates object initialization after
2764 // an AllocateNode and before the next possible safepoint.  As a
2765 // memory barrier (MemBarNode), it keeps critical stores from drifting
2766 // down past any safepoint or any publication of the allocation.
2767 // Before this barrier, a newly-allocated object may have uninitialized bits.
2768 // After this barrier, it may be treated as a real oop, and GC is allowed.
2769 //
2770 // The semantics of the InitializeNode include an implicit zeroing of
2771 // the new object from object header to the end of the object.
2772 // (The object header and end are determined by the AllocateNode.)
2773 //
2774 // Certain stores may be added as direct inputs to the InitializeNode.
2775 // These stores must update raw memory, and they must be to addresses
2776 // derived from the raw address produced by AllocateNode, and with
2777 // a constant offset.  They must be ordered by increasing offset.
2778 // The first one is at in(RawStores), the last at in(req()-1).
2779 // Unlike most memory operations, they are not linked in a chain,
2780 // but are displayed in parallel as users of the rawmem output of
2781 // the allocation.
2782 //
2783 // (See comments in InitializeNode::capture_store, which continue
2784 // the example given above.)
2785 //
2786 // When the associated Allocate is macro-expanded, the InitializeNode
2787 // may be rewritten to optimize collected stores.  A ClearArrayNode
2788 // may also be created at that point to represent any required zeroing.
2789 // The InitializeNode is then marked 'complete', prohibiting further
2790 // capturing of nearby memory operations.
2791 //
2792 // During macro-expansion, all captured initializations which store
2793 // constant values of 32 bits or smaller are coalesced (if advantageous)
2794 // into larger 'tiles' 32 or 64 bits.  This allows an object to be
2795 // initialized in fewer memory operations.  Memory words which are
2796 // covered by neither tiles nor non-constant stores are pre-zeroed
2797 // by explicit stores of zero.  (The code shape happens to do all
2798 // zeroing first, then all other stores, with both sequences occurring
2799 // in order of ascending offsets.)
2800 //
2801 // Alternatively, code may be inserted between an AllocateNode and its
2802 // InitializeNode, to perform arbitrary initialization of the new object.
2803 // E.g., the object copying intrinsics insert complex data transfers here.
2804 // The initialization must then be marked as 'complete' disable the
2805 // built-in zeroing semantics and the collection of initializing stores.
2806 //
2807 // While an InitializeNode is incomplete, reads from the memory state
2808 // produced by it are optimizable if they match the control edge and
2809 // new oop address associated with the allocation/initialization.
2810 // They return a stored value (if the offset matches) or else zero.
2811 // A write to the memory state, if it matches control and address,
2812 // and if it is to a constant offset, may be 'captured' by the
2813 // InitializeNode.  It is cloned as a raw memory operation and rewired
2814 // inside the initialization, to the raw oop produced by the allocation.
2815 // Operations on addresses which are provably distinct (e.g., to
2816 // other AllocateNodes) are allowed to bypass the initialization.
2817 //
2818 // The effect of all this is to consolidate object initialization
2819 // (both arrays and non-arrays, both piecewise and bulk) into a
2820 // single location, where it can be optimized as a unit.
2821 //
2822 // Only stores with an offset less than TrackedInitializationLimit words
2823 // will be considered for capture by an InitializeNode.  This puts a
2824 // reasonable limit on the complexity of optimized initializations.
2825 
2826 //---------------------------InitializeNode------------------------------------
2827 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2828   : _is_complete(false),
2829     MemBarNode(C, adr_type, rawoop)
2830 {
2831   init_class_id(Class_Initialize);
2832 
2833   assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2834   assert(in(RawAddress) == rawoop, "proper init");
2835   // Note:  allocation() can be NULL, for secondary initialization barriers
2836 }
2837 
2838 // Since this node is not matched, it will be processed by the
2839 // register allocator.  Declare that there are no constraints
2840 // on the allocation of the RawAddress edge.
2841 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2842   // This edge should be set to top, by the set_complete.  But be conservative.
2843   if (idx == InitializeNode::RawAddress)
2844     return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2845   return RegMask::Empty;
2846 }
2847 
2848 Node* InitializeNode::memory(uint alias_idx) {
2849   Node* mem = in(Memory);
2850   if (mem->is_MergeMem()) {
2851     return mem->as_MergeMem()->memory_at(alias_idx);
2852   } else {
2853     // incoming raw memory is not split
2854     return mem;
2855   }
2856 }
2857 
2858 bool InitializeNode::is_non_zero() {
2859   if (is_complete())  return false;
2860   remove_extra_zeroes();
2861   return (req() > RawStores);
2862 }
2863 
2864 void InitializeNode::set_complete(PhaseGVN* phase) {
2865   assert(!is_complete(), "caller responsibility");
2866   _is_complete = true;
2867 
2868   // After this node is complete, it contains a bunch of
2869   // raw-memory initializations.  There is no need for
2870   // it to have anything to do with non-raw memory effects.
2871   // Therefore, tell all non-raw users to re-optimize themselves,
2872   // after skipping the memory effects of this initialization.
2873   PhaseIterGVN* igvn = phase->is_IterGVN();
2874   if (igvn)  igvn->add_users_to_worklist(this);
2875 }
2876 
2877 // convenience function
2878 // return false if the init contains any stores already
2879 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2880   InitializeNode* init = initialization();
2881   if (init == NULL || init->is_complete())  return false;
2882   init->remove_extra_zeroes();
2883   // for now, if this allocation has already collected any inits, bail:
2884   if (init->is_non_zero())  return false;
2885   init->set_complete(phase);
2886   return true;
2887 }
2888 
2889 void InitializeNode::remove_extra_zeroes() {
2890   if (req() == RawStores)  return;
2891   Node* zmem = zero_memory();
2892   uint fill = RawStores;
2893   for (uint i = fill; i < req(); i++) {
2894     Node* n = in(i);
2895     if (n->is_top() || n == zmem)  continue;  // skip
2896     if (fill < i)  set_req(fill, n);          // compact
2897     ++fill;
2898   }
2899   // delete any empty spaces created:
2900   while (fill < req()) {
2901     del_req(fill);
2902   }
2903 }
2904 
2905 // Helper for remembering which stores go with which offsets.
2906 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2907   if (!st->is_Store())  return -1;  // can happen to dead code via subsume_node
2908   intptr_t offset = -1;
2909   Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2910                                                phase, offset);
2911   if (base == NULL)     return -1;  // something is dead,
2912   if (offset < 0)       return -1;  //        dead, dead
2913   return offset;
2914 }
2915 
2916 // Helper for proving that an initialization expression is
2917 // "simple enough" to be folded into an object initialization.
2918 // Attempts to prove that a store's initial value 'n' can be captured
2919 // within the initialization without creating a vicious cycle, such as:
2920 //     { Foo p = new Foo(); p.next = p; }
2921 // True for constants and parameters and small combinations thereof.
2922 bool InitializeNode::detect_init_independence(Node* n,
2923                                               bool st_is_pinned,
2924                                               int& count) {
2925   if (n == NULL)      return true;   // (can this really happen?)
2926   if (n->is_Proj())   n = n->in(0);
2927   if (n == this)      return false;  // found a cycle
2928   if (n->is_Con())    return true;
2929   if (n->is_Start())  return true;   // params, etc., are OK
2930   if (n->is_Root())   return true;   // even better
2931 
2932   Node* ctl = n->in(0);
2933   if (ctl != NULL && !ctl->is_top()) {
2934     if (ctl->is_Proj())  ctl = ctl->in(0);
2935     if (ctl == this)  return false;
2936 
2937     // If we already know that the enclosing memory op is pinned right after
2938     // the init, then any control flow that the store has picked up
2939     // must have preceded the init, or else be equal to the init.
2940     // Even after loop optimizations (which might change control edges)
2941     // a store is never pinned *before* the availability of its inputs.
2942     if (!MemNode::all_controls_dominate(n, this))
2943       return false;                  // failed to prove a good control
2944 
2945   }
2946 
2947   // Check data edges for possible dependencies on 'this'.
2948   if ((count += 1) > 20)  return false;  // complexity limit
2949   for (uint i = 1; i < n->req(); i++) {
2950     Node* m = n->in(i);
2951     if (m == NULL || m == n || m->is_top())  continue;
2952     uint first_i = n->find_edge(m);
2953     if (i != first_i)  continue;  // process duplicate edge just once
2954     if (!detect_init_independence(m, st_is_pinned, count)) {
2955       return false;
2956     }
2957   }
2958 
2959   return true;
2960 }
2961 
2962 // Here are all the checks a Store must pass before it can be moved into
2963 // an initialization.  Returns zero if a check fails.
2964 // On success, returns the (constant) offset to which the store applies,
2965 // within the initialized memory.
2966 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2967   const int FAIL = 0;
2968   if (st->req() != MemNode::ValueIn + 1)
2969     return FAIL;                // an inscrutable StoreNode (card mark?)
2970   Node* ctl = st->in(MemNode::Control);
2971   if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2972     return FAIL;                // must be unconditional after the initialization
2973   Node* mem = st->in(MemNode::Memory);
2974   if (!(mem->is_Proj() && mem->in(0) == this))
2975     return FAIL;                // must not be preceded by other stores
2976   Node* adr = st->in(MemNode::Address);
2977   intptr_t offset;
2978   AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2979   if (alloc == NULL)
2980     return FAIL;                // inscrutable address
2981   if (alloc != allocation())
2982     return FAIL;                // wrong allocation!  (store needs to float up)
2983   Node* val = st->in(MemNode::ValueIn);
2984   int complexity_count = 0;
2985   if (!detect_init_independence(val, true, complexity_count))
2986     return FAIL;                // stored value must be 'simple enough'
2987 
2988   return offset;                // success
2989 }
2990 
2991 // Find the captured store in(i) which corresponds to the range
2992 // [start..start+size) in the initialized object.
2993 // If there is one, return its index i.  If there isn't, return the
2994 // negative of the index where it should be inserted.
2995 // Return 0 if the queried range overlaps an initialization boundary
2996 // or if dead code is encountered.
2997 // If size_in_bytes is zero, do not bother with overlap checks.
2998 int InitializeNode::captured_store_insertion_point(intptr_t start,
2999                                                    int size_in_bytes,
3000                                                    PhaseTransform* phase) {
3001   const int FAIL = 0, MAX_STORE = BytesPerLong;
3002 
3003   if (is_complete())
3004     return FAIL;                // arraycopy got here first; punt
3005 
3006   assert(allocation() != NULL, "must be present");
3007 
3008   // no negatives, no header fields:
3009   if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
3010 
3011   // after a certain size, we bail out on tracking all the stores:
3012   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3013   if (start >= ti_limit)  return FAIL;
3014 
3015   for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3016     if (i >= limit)  return -(int)i; // not found; here is where to put it
3017 
3018     Node*    st     = in(i);
3019     intptr_t st_off = get_store_offset(st, phase);
3020     if (st_off < 0) {
3021       if (st != zero_memory()) {
3022         return FAIL;            // bail out if there is dead garbage
3023       }
3024     } else if (st_off > start) {
3025       // ...we are done, since stores are ordered
3026       if (st_off < start + size_in_bytes) {
3027         return FAIL;            // the next store overlaps
3028       }
3029       return -(int)i;           // not found; here is where to put it
3030     } else if (st_off < start) {
3031       if (size_in_bytes != 0 &&
3032           start < st_off + MAX_STORE &&
3033           start < st_off + st->as_Store()->memory_size()) {
3034         return FAIL;            // the previous store overlaps
3035       }
3036     } else {
3037       if (size_in_bytes != 0 &&
3038           st->as_Store()->memory_size() != size_in_bytes) {
3039         return FAIL;            // mismatched store size
3040       }
3041       return i;
3042     }
3043 
3044     ++i;
3045   }
3046 }
3047 
3048 // Look for a captured store which initializes at the offset 'start'
3049 // with the given size.  If there is no such store, and no other
3050 // initialization interferes, then return zero_memory (the memory
3051 // projection of the AllocateNode).
3052 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3053                                           PhaseTransform* phase) {
3054   assert(stores_are_sane(phase), "");
3055   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3056   if (i == 0) {
3057     return NULL;                // something is dead
3058   } else if (i < 0) {
3059     return zero_memory();       // just primordial zero bits here
3060   } else {
3061     Node* st = in(i);           // here is the store at this position
3062     assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3063     return st;
3064   }
3065 }
3066 
3067 // Create, as a raw pointer, an address within my new object at 'offset'.
3068 Node* InitializeNode::make_raw_address(intptr_t offset,
3069                                        PhaseTransform* phase) {
3070   Node* addr = in(RawAddress);
3071   if (offset != 0) {
3072     Compile* C = phase->C;
3073     addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
3074                                                  phase->MakeConX(offset)) );
3075   }
3076   return addr;
3077 }
3078 
3079 // Clone the given store, converting it into a raw store
3080 // initializing a field or element of my new object.
3081 // Caller is responsible for retiring the original store,
3082 // with subsume_node or the like.
3083 //
3084 // From the example above InitializeNode::InitializeNode,
3085 // here are the old stores to be captured:
3086 //   store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3087 //   store2 = (StoreC init.Control store1      (+ oop 14) 2)
3088 //
3089 // Here is the changed code; note the extra edges on init:
3090 //   alloc = (Allocate ...)
3091 //   rawoop = alloc.RawAddress
3092 //   rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3093 //   rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3094 //   init = (Initialize alloc.Control alloc.Memory rawoop
3095 //                      rawstore1 rawstore2)
3096 //
3097 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3098                                     PhaseTransform* phase) {
3099   assert(stores_are_sane(phase), "");
3100 
3101   if (start < 0)  return NULL;
3102   assert(can_capture_store(st, phase) == start, "sanity");
3103 
3104   Compile* C = phase->C;
3105   int size_in_bytes = st->memory_size();
3106   int i = captured_store_insertion_point(start, size_in_bytes, phase);
3107   if (i == 0)  return NULL;     // bail out
3108   Node* prev_mem = NULL;        // raw memory for the captured store
3109   if (i > 0) {
3110     prev_mem = in(i);           // there is a pre-existing store under this one
3111     set_req(i, C->top());       // temporarily disconnect it
3112     // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3113   } else {
3114     i = -i;                     // no pre-existing store
3115     prev_mem = zero_memory();   // a slice of the newly allocated object
3116     if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3117       set_req(--i, C->top());   // reuse this edge; it has been folded away
3118     else
3119       ins_req(i, C->top());     // build a new edge
3120   }
3121   Node* new_st = st->clone();
3122   new_st->set_req(MemNode::Control, in(Control));
3123   new_st->set_req(MemNode::Memory,  prev_mem);
3124   new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3125   new_st = phase->transform(new_st);
3126 
3127   // At this point, new_st might have swallowed a pre-existing store
3128   // at the same offset, or perhaps new_st might have disappeared,
3129   // if it redundantly stored the same value (or zero to fresh memory).
3130 
3131   // In any case, wire it in:
3132   set_req(i, new_st);
3133 
3134   // The caller may now kill the old guy.
3135   DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3136   assert(check_st == new_st || check_st == NULL, "must be findable");
3137   assert(!is_complete(), "");
3138   return new_st;
3139 }
3140 
3141 static bool store_constant(jlong* tiles, int num_tiles,
3142                            intptr_t st_off, int st_size,
3143                            jlong con) {
3144   if ((st_off & (st_size-1)) != 0)
3145     return false;               // strange store offset (assume size==2**N)
3146   address addr = (address)tiles + st_off;
3147   assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3148   switch (st_size) {
3149   case sizeof(jbyte):  *(jbyte*) addr = (jbyte) con; break;
3150   case sizeof(jchar):  *(jchar*) addr = (jchar) con; break;
3151   case sizeof(jint):   *(jint*)  addr = (jint)  con; break;
3152   case sizeof(jlong):  *(jlong*) addr = (jlong) con; break;
3153   default: return false;        // strange store size (detect size!=2**N here)
3154   }
3155   return true;                  // return success to caller
3156 }
3157 
3158 // Coalesce subword constants into int constants and possibly
3159 // into long constants.  The goal, if the CPU permits,
3160 // is to initialize the object with a small number of 64-bit tiles.
3161 // Also, convert floating-point constants to bit patterns.
3162 // Non-constants are not relevant to this pass.
3163 //
3164 // In terms of the running example on InitializeNode::InitializeNode
3165 // and InitializeNode::capture_store, here is the transformation
3166 // of rawstore1 and rawstore2 into rawstore12:
3167 //   alloc = (Allocate ...)
3168 //   rawoop = alloc.RawAddress
3169 //   tile12 = 0x00010002
3170 //   rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3171 //   init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3172 //
3173 void
3174 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3175                                         Node* size_in_bytes,
3176                                         PhaseGVN* phase) {
3177   Compile* C = phase->C;
3178 
3179   assert(stores_are_sane(phase), "");
3180   // Note:  After this pass, they are not completely sane,
3181   // since there may be some overlaps.
3182 
3183   int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3184 
3185   intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3186   intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3187   size_limit = MIN2(size_limit, ti_limit);
3188   size_limit = align_size_up(size_limit, BytesPerLong);
3189   int num_tiles = size_limit / BytesPerLong;
3190 
3191   // allocate space for the tile map:
3192   const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3193   jlong  tiles_buf[small_len];
3194   Node*  nodes_buf[small_len];
3195   jlong  inits_buf[small_len];
3196   jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3197                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3198   Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3199                   : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3200   jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3201                   : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3202   // tiles: exact bitwise model of all primitive constants
3203   // nodes: last constant-storing node subsumed into the tiles model
3204   // inits: which bytes (in each tile) are touched by any initializations
3205 
3206   //// Pass A: Fill in the tile model with any relevant stores.
3207 
3208   Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3209   Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3210   Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3211   Node* zmem = zero_memory(); // initially zero memory state
3212   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3213     Node* st = in(i);
3214     intptr_t st_off = get_store_offset(st, phase);
3215 
3216     // Figure out the store's offset and constant value:
3217     if (st_off < header_size)             continue; //skip (ignore header)
3218     if (st->in(MemNode::Memory) != zmem)  continue; //skip (odd store chain)
3219     int st_size = st->as_Store()->memory_size();
3220     if (st_off + st_size > size_limit)    break;
3221 
3222     // Record which bytes are touched, whether by constant or not.
3223     if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3224       continue;                 // skip (strange store size)
3225 
3226     const Type* val = phase->type(st->in(MemNode::ValueIn));
3227     if (!val->singleton())                continue; //skip (non-con store)
3228     BasicType type = val->basic_type();
3229 
3230     jlong con = 0;
3231     switch (type) {
3232     case T_INT:    con = val->is_int()->get_con();  break;
3233     case T_LONG:   con = val->is_long()->get_con(); break;
3234     case T_FLOAT:  con = jint_cast(val->getf());    break;
3235     case T_DOUBLE: con = jlong_cast(val->getd());   break;
3236     default:                              continue; //skip (odd store type)
3237     }
3238 
3239     if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3240         st->Opcode() == Op_StoreL) {
3241       continue;                 // This StoreL is already optimal.
3242     }
3243 
3244     // Store down the constant.
3245     store_constant(tiles, num_tiles, st_off, st_size, con);
3246 
3247     intptr_t j = st_off >> LogBytesPerLong;
3248 
3249     if (type == T_INT && st_size == BytesPerInt
3250         && (st_off & BytesPerInt) == BytesPerInt) {
3251       jlong lcon = tiles[j];
3252       if (!Matcher::isSimpleConstant64(lcon) &&
3253           st->Opcode() == Op_StoreI) {
3254         // This StoreI is already optimal by itself.
3255         jint* intcon = (jint*) &tiles[j];
3256         intcon[1] = 0;  // undo the store_constant()
3257 
3258         // If the previous store is also optimal by itself, back up and
3259         // undo the action of the previous loop iteration... if we can.
3260         // But if we can't, just let the previous half take care of itself.
3261         st = nodes[j];
3262         st_off -= BytesPerInt;
3263         con = intcon[0];
3264         if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3265           assert(st_off >= header_size, "still ignoring header");
3266           assert(get_store_offset(st, phase) == st_off, "must be");
3267           assert(in(i-1) == zmem, "must be");
3268           DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3269           assert(con == tcon->is_int()->get_con(), "must be");
3270           // Undo the effects of the previous loop trip, which swallowed st:
3271           intcon[0] = 0;        // undo store_constant()
3272           set_req(i-1, st);     // undo set_req(i, zmem)
3273           nodes[j] = NULL;      // undo nodes[j] = st
3274           --old_subword;        // undo ++old_subword
3275         }
3276         continue;               // This StoreI is already optimal.
3277       }
3278     }
3279 
3280     // This store is not needed.
3281     set_req(i, zmem);
3282     nodes[j] = st;              // record for the moment
3283     if (st_size < BytesPerLong) // something has changed
3284           ++old_subword;        // includes int/float, but who's counting...
3285     else  ++old_long;
3286   }
3287 
3288   if ((old_subword + old_long) == 0)
3289     return;                     // nothing more to do
3290 
3291   //// Pass B: Convert any non-zero tiles into optimal constant stores.
3292   // Be sure to insert them before overlapping non-constant stores.
3293   // (E.g., byte[] x = { 1,2,y,4 }  =>  x[int 0] = 0x01020004, x[2]=y.)
3294   for (int j = 0; j < num_tiles; j++) {
3295     jlong con  = tiles[j];
3296     jlong init = inits[j];
3297     if (con == 0)  continue;
3298     jint con0,  con1;           // split the constant, address-wise
3299     jint init0, init1;          // split the init map, address-wise
3300     { union { jlong con; jint intcon[2]; } u;
3301       u.con = con;
3302       con0  = u.intcon[0];
3303       con1  = u.intcon[1];
3304       u.con = init;
3305       init0 = u.intcon[0];
3306       init1 = u.intcon[1];
3307     }
3308 
3309     Node* old = nodes[j];
3310     assert(old != NULL, "need the prior store");
3311     intptr_t offset = (j * BytesPerLong);
3312 
3313     bool split = !Matcher::isSimpleConstant64(con);
3314 
3315     if (offset < header_size) {
3316       assert(offset + BytesPerInt >= header_size, "second int counts");
3317       assert(*(jint*)&tiles[j] == 0, "junk in header");
3318       split = true;             // only the second word counts
3319       // Example:  int a[] = { 42 ... }
3320     } else if (con0 == 0 && init0 == -1) {
3321       split = true;             // first word is covered by full inits
3322       // Example:  int a[] = { ... foo(), 42 ... }
3323     } else if (con1 == 0 && init1 == -1) {
3324       split = true;             // second word is covered by full inits
3325       // Example:  int a[] = { ... 42, foo() ... }
3326     }
3327 
3328     // Here's a case where init0 is neither 0 nor -1:
3329     //   byte a[] = { ... 0,0,foo(),0,  0,0,0,42 ... }
3330     // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3331     // In this case the tile is not split; it is (jlong)42.
3332     // The big tile is stored down, and then the foo() value is inserted.
3333     // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3334 
3335     Node* ctl = old->in(MemNode::Control);
3336     Node* adr = make_raw_address(offset, phase);
3337     const TypePtr* atp = TypeRawPtr::BOTTOM;
3338 
3339     // One or two coalesced stores to plop down.
3340     Node*    st[2];
3341     intptr_t off[2];
3342     int  nst = 0;
3343     if (!split) {
3344       ++new_long;
3345       off[nst] = offset;
3346       st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3347                                   phase->longcon(con), T_LONG);
3348     } else {
3349       // Omit either if it is a zero.
3350       if (con0 != 0) {
3351         ++new_int;
3352         off[nst]  = offset;
3353         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3354                                     phase->intcon(con0), T_INT);
3355       }
3356       if (con1 != 0) {
3357         ++new_int;
3358         offset += BytesPerInt;
3359         adr = make_raw_address(offset, phase);
3360         off[nst]  = offset;
3361         st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3362                                     phase->intcon(con1), T_INT);
3363       }
3364     }
3365 
3366     // Insert second store first, then the first before the second.
3367     // Insert each one just before any overlapping non-constant stores.
3368     while (nst > 0) {
3369       Node* st1 = st[--nst];
3370       C->copy_node_notes_to(st1, old);
3371       st1 = phase->transform(st1);
3372       offset = off[nst];
3373       assert(offset >= header_size, "do not smash header");
3374       int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3375       guarantee(ins_idx != 0, "must re-insert constant store");
3376       if (ins_idx < 0)  ins_idx = -ins_idx;  // never overlap
3377       if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3378         set_req(--ins_idx, st1);
3379       else
3380         ins_req(ins_idx, st1);
3381     }
3382   }
3383 
3384   if (PrintCompilation && WizardMode)
3385     tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3386                   old_subword, old_long, new_int, new_long);
3387   if (C->log() != NULL)
3388     C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3389                    old_subword, old_long, new_int, new_long);
3390 
3391   // Clean up any remaining occurrences of zmem:
3392   remove_extra_zeroes();
3393 }
3394 
3395 // Explore forward from in(start) to find the first fully initialized
3396 // word, and return its offset.  Skip groups of subword stores which
3397 // together initialize full words.  If in(start) is itself part of a
3398 // fully initialized word, return the offset of in(start).  If there
3399 // are no following full-word stores, or if something is fishy, return
3400 // a negative value.
3401 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3402   int       int_map = 0;
3403   intptr_t  int_map_off = 0;
3404   const int FULL_MAP = right_n_bits(BytesPerInt);  // the int_map we hope for
3405 
3406   for (uint i = start, limit = req(); i < limit; i++) {
3407     Node* st = in(i);
3408 
3409     intptr_t st_off = get_store_offset(st, phase);
3410     if (st_off < 0)  break;  // return conservative answer
3411 
3412     int st_size = st->as_Store()->memory_size();
3413     if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3414       return st_off;            // we found a complete word init
3415     }
3416 
3417     // update the map:
3418 
3419     intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3420     if (this_int_off != int_map_off) {
3421       // reset the map:
3422       int_map = 0;
3423       int_map_off = this_int_off;
3424     }
3425 
3426     int subword_off = st_off - this_int_off;
3427     int_map |= right_n_bits(st_size) << subword_off;
3428     if ((int_map & FULL_MAP) == FULL_MAP) {
3429       return this_int_off;      // we found a complete word init
3430     }
3431 
3432     // Did this store hit or cross the word boundary?
3433     intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3434     if (next_int_off == this_int_off + BytesPerInt) {
3435       // We passed the current int, without fully initializing it.
3436       int_map_off = next_int_off;
3437       int_map >>= BytesPerInt;
3438     } else if (next_int_off > this_int_off + BytesPerInt) {
3439       // We passed the current and next int.
3440       return this_int_off + BytesPerInt;
3441     }
3442   }
3443 
3444   return -1;
3445 }
3446 
3447 
3448 // Called when the associated AllocateNode is expanded into CFG.
3449 // At this point, we may perform additional optimizations.
3450 // Linearize the stores by ascending offset, to make memory
3451 // activity as coherent as possible.
3452 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3453                                       intptr_t header_size,
3454                                       Node* size_in_bytes,
3455                                       PhaseGVN* phase) {
3456   assert(!is_complete(), "not already complete");
3457   assert(stores_are_sane(phase), "");
3458   assert(allocation() != NULL, "must be present");
3459 
3460   remove_extra_zeroes();
3461 
3462   if (ReduceFieldZeroing || ReduceBulkZeroing)
3463     // reduce instruction count for common initialization patterns
3464     coalesce_subword_stores(header_size, size_in_bytes, phase);
3465 
3466   Node* zmem = zero_memory();   // initially zero memory state
3467   Node* inits = zmem;           // accumulating a linearized chain of inits
3468   #ifdef ASSERT
3469   intptr_t first_offset = allocation()->minimum_header_size();
3470   intptr_t last_init_off = first_offset;  // previous init offset
3471   intptr_t last_init_end = first_offset;  // previous init offset+size
3472   intptr_t last_tile_end = first_offset;  // previous tile offset+size
3473   #endif
3474   intptr_t zeroes_done = header_size;
3475 
3476   bool do_zeroing = true;       // we might give up if inits are very sparse
3477   int  big_init_gaps = 0;       // how many large gaps have we seen?
3478 
3479   if (ZeroTLAB)  do_zeroing = false;
3480   if (!ReduceFieldZeroing && !ReduceBulkZeroing)  do_zeroing = false;
3481 
3482   for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3483     Node* st = in(i);
3484     intptr_t st_off = get_store_offset(st, phase);
3485     if (st_off < 0)
3486       break;                    // unknown junk in the inits
3487     if (st->in(MemNode::Memory) != zmem)
3488       break;                    // complicated store chains somehow in list
3489 
3490     int st_size = st->as_Store()->memory_size();
3491     intptr_t next_init_off = st_off + st_size;
3492 
3493     if (do_zeroing && zeroes_done < next_init_off) {
3494       // See if this store needs a zero before it or under it.
3495       intptr_t zeroes_needed = st_off;
3496 
3497       if (st_size < BytesPerInt) {
3498         // Look for subword stores which only partially initialize words.
3499         // If we find some, we must lay down some word-level zeroes first,
3500         // underneath the subword stores.
3501         //
3502         // Examples:
3503         //   byte[] a = { p,q,r,s }  =>  a[0]=p,a[1]=q,a[2]=r,a[3]=s
3504         //   byte[] a = { x,y,0,0 }  =>  a[0..3] = 0, a[0]=x,a[1]=y
3505         //   byte[] a = { 0,0,z,0 }  =>  a[0..3] = 0, a[2]=z
3506         //
3507         // Note:  coalesce_subword_stores may have already done this,
3508         // if it was prompted by constant non-zero subword initializers.
3509         // But this case can still arise with non-constant stores.
3510 
3511         intptr_t next_full_store = find_next_fullword_store(i, phase);
3512 
3513         // In the examples above:
3514         //   in(i)          p   q   r   s     x   y     z
3515         //   st_off        12  13  14  15    12  13    14
3516         //   st_size        1   1   1   1     1   1     1
3517         //   next_full_s.  12  16  16  16    16  16    16
3518         //   z's_done      12  16  16  16    12  16    12
3519         //   z's_needed    12  16  16  16    16  16    16
3520         //   zsize          0   0   0   0     4   0     4
3521         if (next_full_store < 0) {
3522           // Conservative tack:  Zero to end of current word.
3523           zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3524         } else {
3525           // Zero to beginning of next fully initialized word.
3526           // Or, don't zero at all, if we are already in that word.
3527           assert(next_full_store >= zeroes_needed, "must go forward");
3528           assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3529           zeroes_needed = next_full_store;
3530         }
3531       }
3532 
3533       if (zeroes_needed > zeroes_done) {
3534         intptr_t zsize = zeroes_needed - zeroes_done;
3535         // Do some incremental zeroing on rawmem, in parallel with inits.
3536         zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3537         rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3538                                               zeroes_done, zeroes_needed,
3539                                               phase);
3540         zeroes_done = zeroes_needed;
3541         if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3542           do_zeroing = false;   // leave the hole, next time
3543       }
3544     }
3545 
3546     // Collect the store and move on:
3547     st->set_req(MemNode::Memory, inits);
3548     inits = st;                 // put it on the linearized chain
3549     set_req(i, zmem);           // unhook from previous position
3550 
3551     if (zeroes_done == st_off)
3552       zeroes_done = next_init_off;
3553 
3554     assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3555 
3556     #ifdef ASSERT
3557     // Various order invariants.  Weaker than stores_are_sane because
3558     // a large constant tile can be filled in by smaller non-constant stores.
3559     assert(st_off >= last_init_off, "inits do not reverse");
3560     last_init_off = st_off;
3561     const Type* val = NULL;
3562     if (st_size >= BytesPerInt &&
3563         (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3564         (int)val->basic_type() < (int)T_OBJECT) {
3565       assert(st_off >= last_tile_end, "tiles do not overlap");
3566       assert(st_off >= last_init_end, "tiles do not overwrite inits");
3567       last_tile_end = MAX2(last_tile_end, next_init_off);
3568     } else {
3569       intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3570       assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3571       assert(st_off      >= last_init_end, "inits do not overlap");
3572       last_init_end = next_init_off;  // it's a non-tile
3573     }
3574     #endif //ASSERT
3575   }
3576 
3577   remove_extra_zeroes();        // clear out all the zmems left over
3578   add_req(inits);
3579 
3580   if (!ZeroTLAB) {
3581     // If anything remains to be zeroed, zero it all now.
3582     zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3583     // if it is the last unused 4 bytes of an instance, forget about it
3584     intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3585     if (zeroes_done + BytesPerLong >= size_limit) {
3586       assert(allocation() != NULL, "");
3587       Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3588       ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3589       if (zeroes_done == k->layout_helper())
3590         zeroes_done = size_limit;
3591     }
3592     if (zeroes_done < size_limit) {
3593       rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3594                                             zeroes_done, size_in_bytes, phase);
3595     }
3596   }
3597 
3598   set_complete(phase);
3599   return rawmem;
3600 }
3601 
3602 
3603 #ifdef ASSERT
3604 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3605   if (is_complete())
3606     return true;                // stores could be anything at this point
3607   assert(allocation() != NULL, "must be present");
3608   intptr_t last_off = allocation()->minimum_header_size();
3609   for (uint i = InitializeNode::RawStores; i < req(); i++) {
3610     Node* st = in(i);
3611     intptr_t st_off = get_store_offset(st, phase);
3612     if (st_off < 0)  continue;  // ignore dead garbage
3613     if (last_off > st_off) {
3614       tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3615       this->dump(2);
3616       assert(false, "ascending store offsets");
3617       return false;
3618     }
3619     last_off = st_off + st->as_Store()->memory_size();
3620   }
3621   return true;
3622 }
3623 #endif //ASSERT
3624 
3625 
3626 
3627 
3628 //============================MergeMemNode=====================================
3629 //
3630 // SEMANTICS OF MEMORY MERGES:  A MergeMem is a memory state assembled from several
3631 // contributing store or call operations.  Each contributor provides the memory
3632 // state for a particular "alias type" (see Compile::alias_type).  For example,
3633 // if a MergeMem has an input X for alias category #6, then any memory reference
3634 // to alias category #6 may use X as its memory state input, as an exact equivalent
3635 // to using the MergeMem as a whole.
3636 //   Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3637 //
3638 // (Here, the <N> notation gives the index of the relevant adr_type.)
3639 //
3640 // In one special case (and more cases in the future), alias categories overlap.
3641 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3642 // states.  Therefore, if a MergeMem has only one contributing input W for Bot,
3643 // it is exactly equivalent to that state W:
3644 //   MergeMem(<Bot>: W) <==> W
3645 //
3646 // Usually, the merge has more than one input.  In that case, where inputs
3647 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3648 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3649 //   Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3650 //   Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3651 //
3652 // A merge can take a "wide" memory state as one of its narrow inputs.
3653 // This simply means that the merge observes out only the relevant parts of
3654 // the wide input.  That is, wide memory states arriving at narrow merge inputs
3655 // are implicitly "filtered" or "sliced" as necessary.  (This is rare.)
3656 //
3657 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3658 // and that memory slices "leak through":
3659 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3660 //
3661 // But, in such a cascade, repeated memory slices can "block the leak":
3662 //   MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3663 //
3664 // In the last example, Y is not part of the combined memory state of the
3665 // outermost MergeMem.  The system must, of course, prevent unschedulable
3666 // memory states from arising, so you can be sure that the state Y is somehow
3667 // a precursor to state Y'.
3668 //
3669 //
3670 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3671 // of each MergeMemNode array are exactly the numerical alias indexes, including
3672 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw.  The functions
3673 // Compile::alias_type (and kin) produce and manage these indexes.
3674 //
3675 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3676 // (Note that this provides quick access to the top node inside MergeMem methods,
3677 // without the need to reach out via TLS to Compile::current.)
3678 //
3679 // As a consequence of what was just described, a MergeMem that represents a full
3680 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3681 // containing all alias categories.
3682 //
3683 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3684 //
3685 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3686 // a memory state for the alias type <N>, or else the top node, meaning that
3687 // there is no particular input for that alias type.  Note that the length of
3688 // a MergeMem is variable, and may be extended at any time to accommodate new
3689 // memory states at larger alias indexes.  When merges grow, they are of course
3690 // filled with "top" in the unused in() positions.
3691 //
3692 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3693 // (Top was chosen because it works smoothly with passes like GCM.)
3694 //
3695 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM.  (It is
3696 // the type of random VM bits like TLS references.)  Since it is always the
3697 // first non-Bot memory slice, some low-level loops use it to initialize an
3698 // index variable:  for (i = AliasIdxRaw; i < req(); i++).
3699 //
3700 //
3701 // ACCESSORS:  There is a special accessor MergeMemNode::base_memory which returns
3702 // the distinguished "wide" state.  The accessor MergeMemNode::memory_at(N) returns
3703 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3704 // it returns the base memory.  To prevent bugs, memory_at does not accept <Top>
3705 // or <Bot> indexes.  The iterator MergeMemStream provides robust iteration over
3706 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3707 //
3708 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3709 // really that different from the other memory inputs.  An abbreviation called
3710 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3711 //
3712 //
3713 // PARTIAL MEMORY STATES:  During optimization, MergeMem nodes may arise that represent
3714 // partial memory states.  When a Phi splits through a MergeMem, the copy of the Phi
3715 // that "emerges though" the base memory will be marked as excluding the alias types
3716 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3717 //
3718 //   Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3719 //     ==Ideal=>  MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3720 //
3721 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3722 // (It is currently unimplemented.)  As you can see, the resulting merge is
3723 // actually a disjoint union of memory states, rather than an overlay.
3724 //
3725 
3726 //------------------------------MergeMemNode-----------------------------------
3727 Node* MergeMemNode::make_empty_memory() {
3728   Node* empty_memory = (Node*) Compile::current()->top();
3729   assert(empty_memory->is_top(), "correct sentinel identity");
3730   return empty_memory;
3731 }
3732 
3733 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3734   init_class_id(Class_MergeMem);
3735   // all inputs are nullified in Node::Node(int)
3736   // set_input(0, NULL);  // no control input
3737 
3738   // Initialize the edges uniformly to top, for starters.
3739   Node* empty_mem = make_empty_memory();
3740   for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3741     init_req(i,empty_mem);
3742   }
3743   assert(empty_memory() == empty_mem, "");
3744 
3745   if( new_base != NULL && new_base->is_MergeMem() ) {
3746     MergeMemNode* mdef = new_base->as_MergeMem();
3747     assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3748     for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3749       mms.set_memory(mms.memory2());
3750     }
3751     assert(base_memory() == mdef->base_memory(), "");
3752   } else {
3753     set_base_memory(new_base);
3754   }
3755 }
3756 
3757 // Make a new, untransformed MergeMem with the same base as 'mem'.
3758 // If mem is itself a MergeMem, populate the result with the same edges.
3759 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3760   return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3761 }
3762 
3763 //------------------------------cmp--------------------------------------------
3764 uint MergeMemNode::hash() const { return NO_HASH; }
3765 uint MergeMemNode::cmp( const Node &n ) const {
3766   return (&n == this);          // Always fail except on self
3767 }
3768 
3769 //------------------------------Identity---------------------------------------
3770 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3771   // Identity if this merge point does not record any interesting memory
3772   // disambiguations.
3773   Node* base_mem = base_memory();
3774   Node* empty_mem = empty_memory();
3775   if (base_mem != empty_mem) {  // Memory path is not dead?
3776     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3777       Node* mem = in(i);
3778       if (mem != empty_mem && mem != base_mem) {
3779         return this;            // Many memory splits; no change
3780       }
3781     }
3782   }
3783   return base_mem;              // No memory splits; ID on the one true input
3784 }
3785 
3786 //------------------------------Ideal------------------------------------------
3787 // This method is invoked recursively on chains of MergeMem nodes
3788 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3789   // Remove chain'd MergeMems
3790   //
3791   // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3792   // relative to the "in(Bot)".  Since we are patching both at the same time,
3793   // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3794   // but rewrite each "in(i)" relative to the new "in(Bot)".
3795   Node *progress = NULL;
3796 
3797 
3798   Node* old_base = base_memory();
3799   Node* empty_mem = empty_memory();
3800   if (old_base == empty_mem)
3801     return NULL; // Dead memory path.
3802 
3803   MergeMemNode* old_mbase;
3804   if (old_base != NULL && old_base->is_MergeMem())
3805     old_mbase = old_base->as_MergeMem();
3806   else
3807     old_mbase = NULL;
3808   Node* new_base = old_base;
3809 
3810   // simplify stacked MergeMems in base memory
3811   if (old_mbase)  new_base = old_mbase->base_memory();
3812 
3813   // the base memory might contribute new slices beyond my req()
3814   if (old_mbase)  grow_to_match(old_mbase);
3815 
3816   // Look carefully at the base node if it is a phi.
3817   PhiNode* phi_base;
3818   if (new_base != NULL && new_base->is_Phi())
3819     phi_base = new_base->as_Phi();
3820   else
3821     phi_base = NULL;
3822 
3823   Node*    phi_reg = NULL;
3824   uint     phi_len = (uint)-1;
3825   if (phi_base != NULL && !phi_base->is_copy()) {
3826     // do not examine phi if degraded to a copy
3827     phi_reg = phi_base->region();
3828     phi_len = phi_base->req();
3829     // see if the phi is unfinished
3830     for (uint i = 1; i < phi_len; i++) {
3831       if (phi_base->in(i) == NULL) {
3832         // incomplete phi; do not look at it yet!
3833         phi_reg = NULL;
3834         phi_len = (uint)-1;
3835         break;
3836       }
3837     }
3838   }
3839 
3840   // Note:  We do not call verify_sparse on entry, because inputs
3841   // can normalize to the base_memory via subsume_node or similar
3842   // mechanisms.  This method repairs that damage.
3843 
3844   assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3845 
3846   // Look at each slice.
3847   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3848     Node* old_in = in(i);
3849     // calculate the old memory value
3850     Node* old_mem = old_in;
3851     if (old_mem == empty_mem)  old_mem = old_base;
3852     assert(old_mem == memory_at(i), "");
3853 
3854     // maybe update (reslice) the old memory value
3855 
3856     // simplify stacked MergeMems
3857     Node* new_mem = old_mem;
3858     MergeMemNode* old_mmem;
3859     if (old_mem != NULL && old_mem->is_MergeMem())
3860       old_mmem = old_mem->as_MergeMem();
3861     else
3862       old_mmem = NULL;
3863     if (old_mmem == this) {
3864       // This can happen if loops break up and safepoints disappear.
3865       // A merge of BotPtr (default) with a RawPtr memory derived from a
3866       // safepoint can be rewritten to a merge of the same BotPtr with
3867       // the BotPtr phi coming into the loop.  If that phi disappears
3868       // also, we can end up with a self-loop of the mergemem.
3869       // In general, if loops degenerate and memory effects disappear,
3870       // a mergemem can be left looking at itself.  This simply means
3871       // that the mergemem's default should be used, since there is
3872       // no longer any apparent effect on this slice.
3873       // Note: If a memory slice is a MergeMem cycle, it is unreachable
3874       //       from start.  Update the input to TOP.
3875       new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3876     }
3877     else if (old_mmem != NULL) {
3878       new_mem = old_mmem->memory_at(i);
3879     }
3880     // else preceding memory was not a MergeMem
3881 
3882     // replace equivalent phis (unfortunately, they do not GVN together)
3883     if (new_mem != NULL && new_mem != new_base &&
3884         new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3885       if (new_mem->is_Phi()) {
3886         PhiNode* phi_mem = new_mem->as_Phi();
3887         for (uint i = 1; i < phi_len; i++) {
3888           if (phi_base->in(i) != phi_mem->in(i)) {
3889             phi_mem = NULL;
3890             break;
3891           }
3892         }
3893         if (phi_mem != NULL) {
3894           // equivalent phi nodes; revert to the def
3895           new_mem = new_base;
3896         }
3897       }
3898     }
3899 
3900     // maybe store down a new value
3901     Node* new_in = new_mem;
3902     if (new_in == new_base)  new_in = empty_mem;
3903 
3904     if (new_in != old_in) {
3905       // Warning:  Do not combine this "if" with the previous "if"
3906       // A memory slice might have be be rewritten even if it is semantically
3907       // unchanged, if the base_memory value has changed.
3908       set_req(i, new_in);
3909       progress = this;          // Report progress
3910     }
3911   }
3912 
3913   if (new_base != old_base) {
3914     set_req(Compile::AliasIdxBot, new_base);
3915     // Don't use set_base_memory(new_base), because we need to update du.
3916     assert(base_memory() == new_base, "");
3917     progress = this;
3918   }
3919 
3920   if( base_memory() == this ) {
3921     // a self cycle indicates this memory path is dead
3922     set_req(Compile::AliasIdxBot, empty_mem);
3923   }
3924 
3925   // Resolve external cycles by calling Ideal on a MergeMem base_memory
3926   // Recursion must occur after the self cycle check above
3927   if( base_memory()->is_MergeMem() ) {
3928     MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3929     Node *m = phase->transform(new_mbase);  // Rollup any cycles
3930     if( m != NULL && (m->is_top() ||
3931         m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3932       // propagate rollup of dead cycle to self
3933       set_req(Compile::AliasIdxBot, empty_mem);
3934     }
3935   }
3936 
3937   if( base_memory() == empty_mem ) {
3938     progress = this;
3939     // Cut inputs during Parse phase only.
3940     // During Optimize phase a dead MergeMem node will be subsumed by Top.
3941     if( !can_reshape ) {
3942       for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3943         if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3944       }
3945     }
3946   }
3947 
3948   if( !progress && base_memory()->is_Phi() && can_reshape ) {
3949     // Check if PhiNode::Ideal's "Split phis through memory merges"
3950     // transform should be attempted. Look for this->phi->this cycle.
3951     uint merge_width = req();
3952     if (merge_width > Compile::AliasIdxRaw) {
3953       PhiNode* phi = base_memory()->as_Phi();
3954       for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3955         if (phi->in(i) == this) {
3956           phase->is_IterGVN()->_worklist.push(phi);
3957           break;
3958         }
3959       }
3960     }
3961   }
3962 
3963   assert(progress || verify_sparse(), "please, no dups of base");
3964   return progress;
3965 }
3966 
3967 //-------------------------set_base_memory-------------------------------------
3968 void MergeMemNode::set_base_memory(Node *new_base) {
3969   Node* empty_mem = empty_memory();
3970   set_req(Compile::AliasIdxBot, new_base);
3971   assert(memory_at(req()) == new_base, "must set default memory");
3972   // Clear out other occurrences of new_base:
3973   if (new_base != empty_mem) {
3974     for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3975       if (in(i) == new_base)  set_req(i, empty_mem);
3976     }
3977   }
3978 }
3979 
3980 //------------------------------out_RegMask------------------------------------
3981 const RegMask &MergeMemNode::out_RegMask() const {
3982   return RegMask::Empty;
3983 }
3984 
3985 //------------------------------dump_spec--------------------------------------
3986 #ifndef PRODUCT
3987 void MergeMemNode::dump_spec(outputStream *st) const {
3988   st->print(" {");
3989   Node* base_mem = base_memory();
3990   for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3991     Node* mem = memory_at(i);
3992     if (mem == base_mem) { st->print(" -"); continue; }
3993     st->print( " N%d:", mem->_idx );
3994     Compile::current()->get_adr_type(i)->dump_on(st);
3995   }
3996   st->print(" }");
3997 }
3998 #endif // !PRODUCT
3999 
4000 
4001 #ifdef ASSERT
4002 static bool might_be_same(Node* a, Node* b) {
4003   if (a == b)  return true;
4004   if (!(a->is_Phi() || b->is_Phi()))  return false;
4005   // phis shift around during optimization
4006   return true;  // pretty stupid...
4007 }
4008 
4009 // verify a narrow slice (either incoming or outgoing)
4010 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4011   if (!VerifyAliases)       return;  // don't bother to verify unless requested
4012   if (is_error_reported())  return;  // muzzle asserts when debugging an error
4013   if (Node::in_dump())      return;  // muzzle asserts when printing
4014   assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4015   assert(n != NULL, "");
4016   // Elide intervening MergeMem's
4017   while (n->is_MergeMem()) {
4018     n = n->as_MergeMem()->memory_at(alias_idx);
4019   }
4020   Compile* C = Compile::current();
4021   const TypePtr* n_adr_type = n->adr_type();
4022   if (n == m->empty_memory()) {
4023     // Implicit copy of base_memory()
4024   } else if (n_adr_type != TypePtr::BOTTOM) {
4025     assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4026     assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4027   } else {
4028     // A few places like make_runtime_call "know" that VM calls are narrow,
4029     // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4030     bool expected_wide_mem = false;
4031     if (n == m->base_memory()) {
4032       expected_wide_mem = true;
4033     } else if (alias_idx == Compile::AliasIdxRaw ||
4034                n == m->memory_at(Compile::AliasIdxRaw)) {
4035       expected_wide_mem = true;
4036     } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4037       // memory can "leak through" calls on channels that
4038       // are write-once.  Allow this also.
4039       expected_wide_mem = true;
4040     }
4041     assert(expected_wide_mem, "expected narrow slice replacement");
4042   }
4043 }
4044 #else // !ASSERT
4045 #define verify_memory_slice(m,i,n) (0)  // PRODUCT version is no-op
4046 #endif
4047 
4048 
4049 //-----------------------------memory_at---------------------------------------
4050 Node* MergeMemNode::memory_at(uint alias_idx) const {
4051   assert(alias_idx >= Compile::AliasIdxRaw ||
4052          alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4053          "must avoid base_memory and AliasIdxTop");
4054 
4055   // Otherwise, it is a narrow slice.
4056   Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4057   Compile *C = Compile::current();
4058   if (is_empty_memory(n)) {
4059     // the array is sparse; empty slots are the "top" node
4060     n = base_memory();
4061     assert(Node::in_dump()
4062            || n == NULL || n->bottom_type() == Type::TOP
4063            || n->adr_type() == TypePtr::BOTTOM
4064            || n->adr_type() == TypeRawPtr::BOTTOM
4065            || Compile::current()->AliasLevel() == 0,
4066            "must be a wide memory");
4067     // AliasLevel == 0 if we are organizing the memory states manually.
4068     // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4069   } else {
4070     // make sure the stored slice is sane
4071     #ifdef ASSERT
4072     if (is_error_reported() || Node::in_dump()) {
4073     } else if (might_be_same(n, base_memory())) {
4074       // Give it a pass:  It is a mostly harmless repetition of the base.
4075       // This can arise normally from node subsumption during optimization.
4076     } else {
4077       verify_memory_slice(this, alias_idx, n);
4078     }
4079     #endif
4080   }
4081   return n;
4082 }
4083 
4084 //---------------------------set_memory_at-------------------------------------
4085 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4086   verify_memory_slice(this, alias_idx, n);
4087   Node* empty_mem = empty_memory();
4088   if (n == base_memory())  n = empty_mem;  // collapse default
4089   uint need_req = alias_idx+1;
4090   if (req() < need_req) {
4091     if (n == empty_mem)  return;  // already the default, so do not grow me
4092     // grow the sparse array
4093     do {
4094       add_req(empty_mem);
4095     } while (req() < need_req);
4096   }
4097   set_req( alias_idx, n );
4098 }
4099 
4100 
4101 
4102 //--------------------------iteration_setup------------------------------------
4103 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4104   if (other != NULL) {
4105     grow_to_match(other);
4106     // invariant:  the finite support of mm2 is within mm->req()
4107     #ifdef ASSERT
4108     for (uint i = req(); i < other->req(); i++) {
4109       assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4110     }
4111     #endif
4112   }
4113   // Replace spurious copies of base_memory by top.
4114   Node* base_mem = base_memory();
4115   if (base_mem != NULL && !base_mem->is_top()) {
4116     for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4117       if (in(i) == base_mem)
4118         set_req(i, empty_memory());
4119     }
4120   }
4121 }
4122 
4123 //---------------------------grow_to_match-------------------------------------
4124 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4125   Node* empty_mem = empty_memory();
4126   assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4127   // look for the finite support of the other memory
4128   for (uint i = other->req(); --i >= req(); ) {
4129     if (other->in(i) != empty_mem) {
4130       uint new_len = i+1;
4131       while (req() < new_len)  add_req(empty_mem);
4132       break;
4133     }
4134   }
4135 }
4136 
4137 //---------------------------verify_sparse-------------------------------------
4138 #ifndef PRODUCT
4139 bool MergeMemNode::verify_sparse() const {
4140   assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4141   Node* base_mem = base_memory();
4142   // The following can happen in degenerate cases, since empty==top.
4143   if (is_empty_memory(base_mem))  return true;
4144   for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4145     assert(in(i) != NULL, "sane slice");
4146     if (in(i) == base_mem)  return false;  // should have been the sentinel value!
4147   }
4148   return true;
4149 }
4150 
4151 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4152   Node* n;
4153   n = mm->in(idx);
4154   if (mem == n)  return true;  // might be empty_memory()
4155   n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4156   if (mem == n)  return true;
4157   while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4158     if (mem == n)  return true;
4159     if (n == NULL)  break;
4160   }
4161   return false;
4162 }
4163 #endif // !PRODUCT