< prev index next >

src/hotspot/share/opto/macro.cpp

Print this page

   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 

  25 #include "compiler/compileLog.hpp"
  26 #include "gc/shared/collectedHeap.inline.hpp"
  27 #include "gc/shared/tlab_globals.hpp"
  28 #include "libadt/vectset.hpp"
  29 #include "memory/universe.hpp"
  30 #include "opto/addnode.hpp"
  31 #include "opto/arraycopynode.hpp"
  32 #include "opto/callnode.hpp"
  33 #include "opto/castnode.hpp"
  34 #include "opto/cfgnode.hpp"
  35 #include "opto/compile.hpp"
  36 #include "opto/convertnode.hpp"
  37 #include "opto/graphKit.hpp"

  38 #include "opto/intrinsicnode.hpp"
  39 #include "opto/locknode.hpp"
  40 #include "opto/loopnode.hpp"
  41 #include "opto/macro.hpp"
  42 #include "opto/memnode.hpp"
  43 #include "opto/narrowptrnode.hpp"
  44 #include "opto/node.hpp"
  45 #include "opto/opaquenode.hpp"
  46 #include "opto/phaseX.hpp"
  47 #include "opto/rootnode.hpp"
  48 #include "opto/runtime.hpp"
  49 #include "opto/subnode.hpp"
  50 #include "opto/subtypenode.hpp"
  51 #include "opto/type.hpp"
  52 #include "prims/jvmtiExport.hpp"
  53 #include "runtime/continuation.hpp"
  54 #include "runtime/sharedRuntime.hpp"

  55 #include "utilities/macros.hpp"
  56 #include "utilities/powerOfTwo.hpp"
  57 #if INCLUDE_G1GC
  58 #include "gc/g1/g1ThreadLocalData.hpp"
  59 #endif // INCLUDE_G1GC
  60 
  61 
  62 //
  63 // Replace any references to "oldref" in inputs to "use" with "newref".
  64 // Returns the number of replacements made.
  65 //
  66 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
  67   int nreplacements = 0;
  68   uint req = use->req();
  69   for (uint j = 0; j < use->len(); j++) {
  70     Node *uin = use->in(j);
  71     if (uin == oldref) {
  72       if (j < req)
  73         use->set_req(j, newref);
  74       else
  75         use->set_prec(j, newref);
  76       nreplacements++;
  77     } else if (j >= req && uin == nullptr) {
  78       break;
  79     }
  80   }
  81   return nreplacements;
  82 }
  83 
  84 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
  85   assert(old != nullptr, "sanity");
  86   for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
  87     Node* use = old->fast_out(i);
  88     _igvn.rehash_node_delayed(use);
  89     imax -= replace_input(use, old, target);
  90     // back up iterator
  91     --i;
  92   }
  93   assert(old->outcnt() == 0, "all uses must be deleted");
  94 }
  95 
  96 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
  97   Node* cmp;
  98   if (mask != 0) {
  99     Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
 100     cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
 101   } else {
 102     cmp = word;
 103   }
 104   Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
 105   IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
 106   transform_later(iff);
 107 
 108   // Fast path taken.
 109   Node *fast_taken = transform_later(new IfFalseNode(iff));
 110 
 111   // Fast path not-taken, i.e. slow path
 112   Node *slow_taken = transform_later(new IfTrueNode(iff));
 113 
 114   if (return_fast_path) {
 115     region->init_req(edge, slow_taken); // Capture slow-control

 138   // Slow-path call
 139  CallNode *call = leaf_name
 140    ? (CallNode*)new CallLeafNode      ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
 141    : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
 142 
 143   // Slow path call has no side-effects, uses few values
 144   copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
 145   if (parm0 != nullptr)  call->init_req(TypeFunc::Parms+0, parm0);
 146   if (parm1 != nullptr)  call->init_req(TypeFunc::Parms+1, parm1);
 147   if (parm2 != nullptr)  call->init_req(TypeFunc::Parms+2, parm2);
 148   call->copy_call_debug_info(&_igvn, oldcall);
 149   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
 150   _igvn.replace_node(oldcall, call);
 151   transform_later(call);
 152 
 153   return call;
 154 }
 155 
 156 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
 157   BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
 158   bs->eliminate_gc_barrier(this, p2x);
 159 #ifndef PRODUCT
 160   if (PrintOptoStatistics) {
 161     Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
 162   }
 163 #endif
 164 }
 165 
 166 // Search for a memory operation for the specified memory slice.
 167 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
 168   Node *orig_mem = mem;
 169   Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 170   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 171   const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
 172   while (true) {
 173     if (mem == alloc_mem || mem == start_mem ) {
 174       return mem;  // hit one of our sentinels
 175     } else if (mem->is_MergeMem()) {
 176       mem = mem->as_MergeMem()->memory_at(alias_idx);
 177     } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
 178       Node *in = mem->in(0);

 193         ArrayCopyNode* ac = nullptr;
 194         if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
 195           if (ac != nullptr) {
 196             assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
 197             return ac;
 198           }
 199         }
 200         mem = in->in(TypeFunc::Memory);
 201       } else {
 202 #ifdef ASSERT
 203         in->dump();
 204         mem->dump();
 205         assert(false, "unexpected projection");
 206 #endif
 207       }
 208     } else if (mem->is_Store()) {
 209       const TypePtr* atype = mem->as_Store()->adr_type();
 210       int adr_idx = phase->C->get_alias_index(atype);
 211       if (adr_idx == alias_idx) {
 212         assert(atype->isa_oopptr(), "address type must be oopptr");
 213         int adr_offset = atype->offset();
 214         uint adr_iid = atype->is_oopptr()->instance_id();
 215         // Array elements references have the same alias_idx
 216         // but different offset and different instance_id.
 217         if (adr_offset == offset && adr_iid == alloc->_idx) {
 218           return mem;
 219         }
 220       } else {
 221         assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
 222       }
 223       mem = mem->in(MemNode::Memory);
 224     } else if (mem->is_ClearArray()) {
 225       if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
 226         // Can not bypass initialization of the instance
 227         // we are looking.
 228         debug_only(intptr_t offset;)
 229         assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
 230         InitializeNode* init = alloc->as_Allocate()->initialization();
 231         // We are looking for stored value, return Initialize node
 232         // or memory edge from Allocate node.
 233         if (init != nullptr) {

 238       }
 239       // Otherwise skip it (the call updated 'mem' value).
 240     } else if (mem->Opcode() == Op_SCMemProj) {
 241       mem = mem->in(0);
 242       Node* adr = nullptr;
 243       if (mem->is_LoadStore()) {
 244         adr = mem->in(MemNode::Address);
 245       } else {
 246         assert(mem->Opcode() == Op_EncodeISOArray ||
 247                mem->Opcode() == Op_StrCompressedCopy, "sanity");
 248         adr = mem->in(3); // Destination array
 249       }
 250       const TypePtr* atype = adr->bottom_type()->is_ptr();
 251       int adr_idx = phase->C->get_alias_index(atype);
 252       if (adr_idx == alias_idx) {
 253         DEBUG_ONLY(mem->dump();)
 254         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 255         return nullptr;
 256       }
 257       mem = mem->in(MemNode::Memory);
 258    } else if (mem->Opcode() == Op_StrInflatedCopy) {
 259       Node* adr = mem->in(3); // Destination array
 260       const TypePtr* atype = adr->bottom_type()->is_ptr();
 261       int adr_idx = phase->C->get_alias_index(atype);
 262       if (adr_idx == alias_idx) {
 263         DEBUG_ONLY(mem->dump();)
 264         assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
 265         return nullptr;
 266       }
 267       mem = mem->in(MemNode::Memory);
 268     } else {
 269       return mem;
 270     }
 271     assert(mem != orig_mem, "dead memory loop");
 272   }
 273 }
 274 
 275 // Generate loads from source of the arraycopy for fields of
 276 // destination needed at a deoptimization point
 277 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
 278   BasicType bt = ft;

 283   }
 284   Node* res = nullptr;
 285   if (ac->is_clonebasic()) {
 286     assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
 287     Node* base = ac->in(ArrayCopyNode::Src);
 288     Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
 289     const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
 290     MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
 291     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 292     res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
 293   } else {
 294     if (ac->modifies(offset, offset, &_igvn, true)) {
 295       assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
 296       uint shift = exact_log2(type2aelembytes(bt));
 297       Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
 298       Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
 299       const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
 300       const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
 301 
 302       Node* adr = nullptr;
 303       const TypePtr* adr_type = nullptr;




 304       if (src_pos_t->is_con() && dest_pos_t->is_con()) {
 305         intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
 306         Node* base = ac->in(ArrayCopyNode::Src);
 307         adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
 308         adr_type = _igvn.type(base)->is_ptr()->add_offset(off);

 309         if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
 310           // Don't emit a new load from src if src == dst but try to get the value from memory instead
 311           return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
 312         }
 313       } else {





 314         Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 315 #ifdef _LP64
 316         diff = _igvn.transform(new ConvI2LNode(diff));
 317 #endif
 318         diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
 319 
 320         Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
 321         Node* base = ac->in(ArrayCopyNode::Src);
 322         adr = _igvn.transform(new AddPNode(base, base, off));
 323         adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
 324         if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
 325           // Non constant offset in the array: we can't statically
 326           // determine the value
 327           return nullptr;
 328         }
 329       }
 330       MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
 331       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 332       res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
 333     }
 334   }
 335   if (res != nullptr) {
 336     if (ftype->isa_narrowoop()) {
 337       // PhaseMacroExpand::scalar_replacement adds DecodeN nodes

 338       res = _igvn.transform(new EncodePNode(res, ftype));
 339     }
 340     return res;
 341   }
 342   return nullptr;
 343 }
 344 
 345 //
 346 // Given a Memory Phi, compute a value Phi containing the values from stores
 347 // on the input paths.
 348 // Note: this function is recursive, its depth is limited by the "level" argument
 349 // Returns the computed Phi, or null if it cannot compute it.
 350 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
 351   assert(mem->is_Phi(), "sanity");
 352   int alias_idx = C->get_alias_index(adr_t);
 353   int offset = adr_t->offset();
 354   int instance_id = adr_t->instance_id();
 355 
 356   // Check if an appropriate value phi already exists.
 357   Node* region = mem->in(0);
 358   for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
 359     Node* phi = region->fast_out(k);
 360     if (phi->is_Phi() && phi != mem &&
 361         phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
 362       return phi;
 363     }
 364   }
 365   // Check if an appropriate new value phi already exists.
 366   Node* new_phi = value_phis->find(mem->_idx);
 367   if (new_phi != nullptr)
 368     return new_phi;
 369 
 370   if (level <= 0) {
 371     return nullptr; // Give up: phi tree too deep
 372   }
 373   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 374   Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 375   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 376 
 377   uint length = mem->req();
 378   GrowableArray <Node *> values(length, length, nullptr);
 379 
 380   // create a new Phi for the value
 381   PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
 382   transform_later(phi);
 383   value_phis->push(phi, mem->_idx);
 384 
 385   for (uint j = 1; j < length; j++) {
 386     Node *in = mem->in(j);
 387     if (in == nullptr || in->is_top()) {
 388       values.at_put(j, in);
 389     } else  {
 390       Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
 391       if (val == start_mem || val == alloc_mem) {
 392         // hit a sentinel, return appropriate 0 value
 393         values.at_put(j, _igvn.zerocon(ft));











 394         continue;
 395       }
 396       if (val->is_Initialize()) {
 397         val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 398       }
 399       if (val == nullptr) {
 400         return nullptr;  // can't find a value on this path
 401       }
 402       if (val == mem) {
 403         values.at_put(j, mem);
 404       } else if (val->is_Store()) {
 405         Node* n = val->in(MemNode::ValueIn);
 406         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 407         n = bs->step_over_gc_barrier(n);
 408         if (is_subword_type(ft)) {
 409           n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
 410         }
 411         values.at_put(j, n);
 412       } else if(val->is_Proj() && val->in(0) == alloc) {
 413         values.at_put(j, _igvn.zerocon(ft));







 414       } else if (val->is_Phi()) {
 415         val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
 416         if (val == nullptr) {
 417           return nullptr;
 418         }
 419         values.at_put(j, val);
 420       } else if (val->Opcode() == Op_SCMemProj) {
 421         assert(val->in(0)->is_LoadStore() ||
 422                val->in(0)->Opcode() == Op_EncodeISOArray ||
 423                val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
 424         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 425         return nullptr;
 426       } else if (val->is_ArrayCopy()) {
 427         Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
 428         if (res == nullptr) {
 429           return nullptr;
 430         }
 431         values.at_put(j, res);
 432       } else if (val->is_top()) {
 433         // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.

 441     }
 442   }
 443   // Set Phi's inputs
 444   for (uint j = 1; j < length; j++) {
 445     if (values.at(j) == mem) {
 446       phi->init_req(j, phi);
 447     } else {
 448       phi->init_req(j, values.at(j));
 449     }
 450   }
 451   return phi;
 452 }
 453 
 454 // Search the last value stored into the object's field.
 455 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
 456   assert(adr_t->is_known_instance_field(), "instance required");
 457   int instance_id = adr_t->instance_id();
 458   assert((uint)instance_id == alloc->_idx, "wrong allocation");
 459 
 460   int alias_idx = C->get_alias_index(adr_t);
 461   int offset = adr_t->offset();
 462   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 463   Node *alloc_ctrl = alloc->in(TypeFunc::Control);
 464   Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 465   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 466   VectorSet visited;
 467 
 468   bool done = sfpt_mem == alloc_mem;
 469   Node *mem = sfpt_mem;
 470   while (!done) {
 471     if (visited.test_set(mem->_idx)) {
 472       return nullptr;  // found a loop, give up
 473     }
 474     mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
 475     if (mem == start_mem || mem == alloc_mem) {
 476       done = true;  // hit a sentinel, return appropriate 0 value
 477     } else if (mem->is_Initialize()) {
 478       mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 479       if (mem == nullptr) {
 480         done = true; // Something go wrong.
 481       } else if (mem->is_Store()) {
 482         const TypePtr* atype = mem->as_Store()->adr_type();
 483         assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
 484         done = true;
 485       }
 486     } else if (mem->is_Store()) {
 487       const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
 488       assert(atype != nullptr, "address type must be oopptr");
 489       assert(C->get_alias_index(atype) == alias_idx &&
 490              atype->is_known_instance_field() && atype->offset() == offset &&
 491              atype->instance_id() == instance_id, "store is correct memory slice");
 492       done = true;
 493     } else if (mem->is_Phi()) {
 494       // try to find a phi's unique input
 495       Node *unique_input = nullptr;
 496       Node *top = C->top();
 497       for (uint i = 1; i < mem->req(); i++) {
 498         Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
 499         if (n == nullptr || n == top || n == mem) {
 500           continue;
 501         } else if (unique_input == nullptr) {
 502           unique_input = n;
 503         } else if (unique_input != n) {
 504           unique_input = top;
 505           break;
 506         }
 507       }
 508       if (unique_input != nullptr && unique_input != top) {
 509         mem = unique_input;
 510       } else {
 511         done = true;
 512       }
 513     } else if (mem->is_ArrayCopy()) {
 514       done = true;
 515     } else {
 516       DEBUG_ONLY( mem->dump(); )
 517       assert(false, "unexpected node");
 518     }
 519   }
 520   if (mem != nullptr) {
 521     if (mem == start_mem || mem == alloc_mem) {
 522       // hit a sentinel, return appropriate 0 value
















 523       return _igvn.zerocon(ft);
 524     } else if (mem->is_Store()) {
 525       Node* n = mem->in(MemNode::ValueIn);
 526       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 527       n = bs->step_over_gc_barrier(n);
 528       return n;
 529     } else if (mem->is_Phi()) {
 530       // attempt to produce a Phi reflecting the values on the input paths of the Phi
 531       Node_Stack value_phis(8);
 532       Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
 533       if (phi != nullptr) {
 534         return phi;
 535       } else {
 536         // Kill all new Phis
 537         while(value_phis.is_nonempty()) {
 538           Node* n = value_phis.node();
 539           _igvn.replace_node(n, C->top());
 540           value_phis.pop();
 541         }
 542       }
 543     } else if (mem->is_ArrayCopy()) {
 544       Node* ctl = mem->in(0);
 545       Node* m = mem->in(TypeFunc::Memory);
 546       if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
 547         // pin the loads in the uncommon trap path
 548         ctl = sfpt_ctl;
 549         m = sfpt_mem;
 550       }
 551       return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
 552     }
 553   }
 554   // Something go wrong.
 555   return nullptr;
 556 }
 557 















































 558 // Check the possibility of scalar replacement.
 559 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
 560   //  Scan the uses of the allocation to check for anything that would
 561   //  prevent us from eliminating it.
 562   NOT_PRODUCT( const char* fail_eliminate = nullptr; )
 563   DEBUG_ONLY( Node* disq_node = nullptr; )
 564   bool can_eliminate = true;
 565   bool reduce_merge_precheck = (safepoints == nullptr);
 566 

 567   Node* res = alloc->result_cast();
 568   const TypeOopPtr* res_type = nullptr;
 569   if (res == nullptr) {
 570     // All users were eliminated.
 571   } else if (!res->is_CheckCastPP()) {
 572     NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
 573     can_eliminate = false;
 574   } else {

 575     res_type = igvn->type(res)->isa_oopptr();
 576     if (res_type == nullptr) {
 577       NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
 578       can_eliminate = false;
 579     } else if (!res_type->klass_is_exact()) {
 580       NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
 581       can_eliminate = false;
 582     } else if (res_type->isa_aryptr()) {
 583       int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 584       if (length < 0) {
 585         NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
 586         can_eliminate = false;
 587       }
 588     }
 589   }
 590 
 591   if (can_eliminate && res != nullptr) {
 592     BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
 593     for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
 594                                j < jmax && can_eliminate; j++) {
 595       Node* use = res->fast_out(j);
 596 
 597       if (use->is_AddP()) {
 598         const TypePtr* addp_type = igvn->type(use)->is_ptr();
 599         int offset = addp_type->offset();
 600 
 601         if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
 602           NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
 603           can_eliminate = false;
 604           break;
 605         }
 606         for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
 607                                    k < kmax && can_eliminate; k++) {
 608           Node* n = use->fast_out(k);
 609           if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
 610             DEBUG_ONLY(disq_node = n;)
 611             if (n->is_Load() || n->is_LoadStore()) {
 612               NOT_PRODUCT(fail_eliminate = "Field load";)
 613             } else {
 614               NOT_PRODUCT(fail_eliminate = "Not store field reference";)

 620                  (use->as_ArrayCopy()->is_clonebasic() ||
 621                   use->as_ArrayCopy()->is_arraycopy_validated() ||
 622                   use->as_ArrayCopy()->is_copyof_validated() ||
 623                   use->as_ArrayCopy()->is_copyofrange_validated()) &&
 624                  use->in(ArrayCopyNode::Dest) == res) {
 625         // ok to eliminate
 626       } else if (use->is_SafePoint()) {
 627         SafePointNode* sfpt = use->as_SafePoint();
 628         if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
 629           // Object is passed as argument.
 630           DEBUG_ONLY(disq_node = use;)
 631           NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
 632           can_eliminate = false;
 633         }
 634         Node* sfptMem = sfpt->memory();
 635         if (sfptMem == nullptr || sfptMem->is_top()) {
 636           DEBUG_ONLY(disq_node = use;)
 637           NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
 638           can_eliminate = false;
 639         } else if (!reduce_merge_precheck) {

 640           safepoints->append_if_missing(sfpt);
 641         }























 642       } else if (reduce_merge_precheck &&
 643                  (use->is_Phi() || use->is_EncodeP() ||
 644                   use->Opcode() == Op_MemBarRelease ||
 645                   (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
 646         // Nothing to do
 647       } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
 648         if (use->is_Phi()) {
 649           if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
 650             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 651           } else {
 652             NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
 653           }
 654           DEBUG_ONLY(disq_node = use;)
 655         } else {
 656           if (use->Opcode() == Op_Return) {
 657             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 658           } else {
 659             NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
 660           }
 661           DEBUG_ONLY(disq_node = use;)
 662         }
 663         can_eliminate = false;



 664       }
 665     }
 666   }
 667 
 668 #ifndef PRODUCT
 669   if (PrintEliminateAllocations && safepoints != nullptr) {
 670     if (can_eliminate) {
 671       tty->print("Scalar ");
 672       if (res == nullptr)
 673         alloc->dump();
 674       else
 675         res->dump();
 676     } else if (alloc->_is_scalar_replaceable) {
 677       tty->print("NotScalar (%s)", fail_eliminate);
 678       if (res == nullptr)
 679         alloc->dump();
 680       else
 681         res->dump();
 682 #ifdef ASSERT
 683       if (disq_node != nullptr) {
 684           tty->print("  >>>> ");
 685           disq_node->dump();
 686       }
 687 #endif /*ASSERT*/
 688     }
 689   }
 690 
 691   if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
 692     tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
 693     DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
 694   }
 695 #endif
 696   return can_eliminate;

 726     JVMState *jvms = sfpt_done->jvms();
 727     jvms->set_endoff(sfpt_done->req());
 728     // Now make a pass over the debug information replacing any references
 729     // to SafePointScalarObjectNode with the allocated object.
 730     int start = jvms->debug_start();
 731     int end   = jvms->debug_end();
 732     for (int i = start; i < end; i++) {
 733       if (sfpt_done->in(i)->is_SafePointScalarObject()) {
 734         SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
 735         if (scobj->first_index(jvms) == sfpt_done->req() &&
 736             scobj->n_fields() == (uint)nfields) {
 737           assert(scobj->alloc() == alloc, "sanity");
 738           sfpt_done->set_req(i, res);
 739         }
 740       }
 741     }
 742     _igvn._worklist.push(sfpt_done);
 743   }
 744 }
 745 
 746 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) {

 747   // Fields of scalar objs are referenced only at the end
 748   // of regular debuginfo at the last (youngest) JVMS.
 749   // Record relative start index.
 750   ciInstanceKlass* iklass    = nullptr;
 751   BasicType basic_elem_type  = T_ILLEGAL;
 752   const Type* field_type     = nullptr;
 753   const TypeOopPtr* res_type = nullptr;
 754   int nfields                = 0;
 755   int array_base             = 0;
 756   int element_size           = 0;
 757   uint first_ind             = (sfpt->req() - sfpt->jvms()->scloff());
 758   Node* res                  = alloc->result_cast();
 759 
 760   assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
 761   assert(sfpt->jvms() != nullptr, "missed JVMS");
 762 
 763   if (res != nullptr) { // Could be null when there are no users
 764     res_type = _igvn.type(res)->isa_oopptr();
 765 
 766     if (res_type->isa_instptr()) {
 767       // find the fields of the class which will be needed for safepoint debug information
 768       iklass = res_type->is_instptr()->instance_klass();
 769       nfields = iklass->nof_nonstatic_fields();
 770     } else {
 771       // find the array's elements which will be needed for safepoint debug information
 772       nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 773       assert(nfields >= 0, "must be an array klass.");
 774       basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
 775       array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 776       element_size = type2aelembytes(basic_elem_type);
 777       field_type = res_type->is_aryptr()->elem();














 778     }
 779   }
 780 
 781   SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
 782   sobj->init_req(0, C->root());
 783   transform_later(sobj);
 784 
 785   // Scan object's fields adding an input to the safepoint for each field.
 786   for (int j = 0; j < nfields; j++) {
 787     intptr_t offset;
 788     ciField* field = nullptr;
 789     if (iklass != nullptr) {
 790       field = iklass->nonstatic_field_at(j);
 791       offset = field->offset_in_bytes();
 792       ciType* elem_type = field->type();
 793       basic_elem_type = field->layout_type();










 794 
 795       // The next code is taken from Parse::do_get_xxx().
 796       if (is_reference_type(basic_elem_type)) {
 797         if (!elem_type->is_loaded()) {
 798           field_type = TypeInstPtr::BOTTOM;
 799         } else if (field != nullptr && field->is_static_constant()) {
 800           ciObject* con = field->constant_value().as_object();
 801           // Do not "join" in the previous type; it doesn't add value,
 802           // and may yield a vacuous result if the field is of interface type.
 803           field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
 804           assert(field_type != nullptr, "field singleton type must be consistent");
 805         } else {
 806           field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
 807         }
 808         if (UseCompressedOops) {
 809           field_type = field_type->make_narrowoop();
 810           basic_elem_type = T_NARROWOOP;
 811         }
 812       } else {
 813         field_type = Type::get_const_basic_type(basic_elem_type);
 814       }
 815     } else {
 816       offset = array_base + j * (intptr_t)element_size;
 817     }
 818 
 819     const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
 820 
 821     Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);






 822 
 823     // We weren't able to find a value for this field,
 824     // give up on eliminating this allocation.
 825     if (field_val == nullptr) {
 826       uint last = sfpt->req() - 1;
 827       for (int k = 0;  k < j; k++) {
 828         sfpt->del_req(last--);
 829       }
 830       _igvn._worklist.push(sfpt);
 831 
 832 #ifndef PRODUCT
 833       if (PrintEliminateAllocations) {
 834         if (field != nullptr) {
 835           tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
 836           field->print();
 837           int field_idx = C->get_alias_index(field_addr_type);
 838           tty->print(" (alias_idx=%d)", field_idx);
 839         } else { // Array's element
 840           tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
 841         }
 842         tty->print(", which prevents elimination of: ");
 843         if (res == nullptr)
 844           alloc->dump();
 845         else
 846           res->dump();
 847       }
 848 #endif
 849 
 850       return nullptr;
 851     }
 852 
 853     if (UseCompressedOops && field_type->isa_narrowoop()) {
 854       // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
 855       // to be able scalar replace the allocation.
 856       if (field_val->is_EncodeP()) {
 857         field_val = field_val->in(1);
 858       } else {
 859         field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
 860       }
 861     }













 862     sfpt->add_req(field_val);
 863   }
 864 
 865   sfpt->jvms()->set_endoff(sfpt->req());
 866 
 867   return sobj;
 868 }
 869 
 870 // Do scalar replacement.
 871 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
 872   GrowableArray <SafePointNode *> safepoints_done;
 873   Node* res = alloc->result_cast();
 874   assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");




 875 
 876   // Process the safepoint uses



 877   while (safepoints.length() > 0) {
 878     SafePointNode* sfpt = safepoints.pop();
 879     SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt);
 880 
 881     if (sobj == nullptr) {
 882       undo_previous_scalarizations(safepoints_done, alloc);
 883       return false;
 884     }
 885 
 886     // Now make a pass over the debug information replacing any references
 887     // to the allocated object with "sobj"
 888     JVMState *jvms = sfpt->jvms();
 889     sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
 890     _igvn._worklist.push(sfpt);
 891 
 892     // keep it for rollback
 893     safepoints_done.append_if_missing(sfpt);
 894   }
 895 







 896   return true;
 897 }
 898 
 899 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
 900   Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
 901   Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
 902   if (ctl_proj != nullptr) {
 903     igvn.replace_node(ctl_proj, n->in(0));
 904   }
 905   if (mem_proj != nullptr) {
 906     igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
 907   }
 908 }
 909 
 910 // Process users of eliminated allocation.
 911 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {

 912   Node* res = alloc->result_cast();
 913   if (res != nullptr) {




 914     for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
 915       Node *use = res->last_out(j);
 916       uint oc1 = res->outcnt();
 917 
 918       if (use->is_AddP()) {
 919         for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
 920           Node *n = use->last_out(k);
 921           uint oc2 = use->outcnt();
 922           if (n->is_Store()) {
 923 #ifdef ASSERT
 924             // Verify that there is no dependent MemBarVolatile nodes,
 925             // they should be removed during IGVN, see MemBarNode::Ideal().
 926             for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
 927                                        p < pmax; p++) {
 928               Node* mb = n->fast_out(p);
 929               assert(mb->is_Initialize() || !mb->is_MemBar() ||
 930                      mb->req() <= MemBarNode::Precedent ||
 931                      mb->in(MemBarNode::Precedent) != n,
 932                      "MemBarVolatile should be eliminated for non-escaping object");
 933             }
 934 #endif
 935             _igvn.replace_node(n, n->in(MemNode::Memory));
 936           } else {
 937             eliminate_gc_barrier(n);
 938           }
 939           k -= (oc2 - use->outcnt());
 940         }
 941         _igvn.remove_dead_node(use);
 942       } else if (use->is_ArrayCopy()) {
 943         // Disconnect ArrayCopy node
 944         ArrayCopyNode* ac = use->as_ArrayCopy();
 945         if (ac->is_clonebasic()) {
 946           Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
 947           disconnect_projections(ac, _igvn);
 948           assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
 949           Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
 950           disconnect_projections(membar_before->as_MemBar(), _igvn);
 951           if (membar_after->is_MemBar()) {
 952             disconnect_projections(membar_after->as_MemBar(), _igvn);
 953           }
 954         } else {
 955           assert(ac->is_arraycopy_validated() ||
 956                  ac->is_copyof_validated() ||
 957                  ac->is_copyofrange_validated(), "unsupported");
 958           CallProjections callprojs;
 959           ac->extract_projections(&callprojs, true);
 960 
 961           _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
 962           _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
 963           _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
 964 
 965           // Set control to top. IGVN will remove the remaining projections
 966           ac->set_req(0, top());
 967           ac->replace_edge(res, top(), &_igvn);
 968 
 969           // Disconnect src right away: it can help find new
 970           // opportunities for allocation elimination
 971           Node* src = ac->in(ArrayCopyNode::Src);
 972           ac->replace_edge(src, top(), &_igvn);
 973           // src can be top at this point if src and dest of the
 974           // arraycopy were the same
 975           if (src->outcnt() == 0 && !src->is_top()) {
 976             _igvn.remove_dead_node(src);
 977           }
 978         }
 979         _igvn._worklist.push(ac);























 980       } else {
 981         eliminate_gc_barrier(use);
 982       }
 983       j -= (oc1 - res->outcnt());
 984     }
 985     assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
 986     _igvn.remove_dead_node(res);
 987   }
 988 
 989   //
 990   // Process other users of allocation's projections
 991   //
 992   if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) {
 993     // First disconnect stores captured by Initialize node.
 994     // If Initialize node is eliminated first in the following code,
 995     // it will kill such stores and DUIterator_Last will assert.
 996     for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax);  j < jmax; j++) {
 997       Node* use = _callprojs.resproj->fast_out(j);
 998       if (use->is_AddP()) {
 999         // raw memory addresses used only by the initialization
1000         _igvn.replace_node(use, C->top());
1001         --j; --jmax;
1002       }
1003     }
1004     for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
1005       Node* use = _callprojs.resproj->last_out(j);
1006       uint oc1 = _callprojs.resproj->outcnt();
1007       if (use->is_Initialize()) {
1008         // Eliminate Initialize node.
1009         InitializeNode *init = use->as_Initialize();
1010         assert(init->outcnt() <= 2, "only a control and memory projection expected");
1011         Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1012         if (ctrl_proj != nullptr) {
1013           _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1014 #ifdef ASSERT
1015           // If the InitializeNode has no memory out, it will die, and tmp will become null
1016           Node* tmp = init->in(TypeFunc::Control);
1017           assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
1018 #endif
1019         }
1020         Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1021         if (mem_proj != nullptr) {
1022           Node *mem = init->in(TypeFunc::Memory);
1023 #ifdef ASSERT
1024           if (mem->is_MergeMem()) {
1025             assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
1026           } else {
1027             assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
1028           }
1029 #endif
1030           _igvn.replace_node(mem_proj, mem);
1031         }




1032       } else  {
1033         assert(false, "only Initialize or AddP expected");
1034       }
1035       j -= (oc1 - _callprojs.resproj->outcnt());
1036     }
1037   }
1038   if (_callprojs.fallthrough_catchproj != nullptr) {
1039     _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
1040   }
1041   if (_callprojs.fallthrough_memproj != nullptr) {
1042     _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
1043   }
1044   if (_callprojs.catchall_memproj != nullptr) {
1045     _igvn.replace_node(_callprojs.catchall_memproj, C->top());
1046   }
1047   if (_callprojs.fallthrough_ioproj != nullptr) {
1048     _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1049   }
1050   if (_callprojs.catchall_ioproj != nullptr) {
1051     _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1052   }
1053   if (_callprojs.catchall_catchproj != nullptr) {
1054     _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1055   }
1056 }
1057 
1058 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1059   // If reallocation fails during deoptimization we'll pop all
1060   // interpreter frames for this compiled frame and that won't play
1061   // nice with JVMTI popframe.
1062   // We avoid this issue by eager reallocation when the popframe request
1063   // is received.
1064   if (!EliminateAllocations || !alloc->_is_non_escaping) {
1065     return false;
1066   }
1067   Node* klass = alloc->in(AllocateNode::KlassNode);
1068   const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1069   Node* res = alloc->result_cast();







1070   // Eliminate boxing allocations which are not used
1071   // regardless scalar replaceable status.
1072   bool boxing_alloc = C->eliminate_boxing() &&

1073                       tklass->isa_instklassptr() &&
1074                       tklass->is_instklassptr()->instance_klass()->is_box_klass();
1075   if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1076     return false;
1077   }
1078 
1079   alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1080 
1081   GrowableArray <SafePointNode *> safepoints;
1082   if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1083     return false;
1084   }
1085 
1086   if (!alloc->_is_scalar_replaceable) {
1087     assert(res == nullptr, "sanity");
1088     // We can only eliminate allocation if all debug info references
1089     // are already replaced with SafePointScalarObject because
1090     // we can't search for a fields value without instance_id.
1091     if (safepoints.length() > 0) {


1092       return false;
1093     }
1094   }
1095 
1096   if (!scalar_replacement(alloc, safepoints)) {
1097     return false;
1098   }
1099 
1100   CompileLog* log = C->log();
1101   if (log != nullptr) {
1102     log->head("eliminate_allocation type='%d'",
1103               log->identify(tklass->exact_klass()));
1104     JVMState* p = alloc->jvms();
1105     while (p != nullptr) {
1106       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1107       p = p->caller();
1108     }
1109     log->tail("eliminate_allocation");
1110   }
1111 
1112   process_users_of_allocation(alloc);
1113 
1114 #ifndef PRODUCT
1115   if (PrintEliminateAllocations) {
1116     if (alloc->is_AllocateArray())
1117       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1118     else
1119       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1120   }
1121 #endif
1122 
1123   return true;
1124 }
1125 
1126 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1127   // EA should remove all uses of non-escaping boxing node.
1128   if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1129     return false;
1130   }
1131 
1132   assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1133 
1134   boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1135 
1136   const TypeTuple* r = boxing->tf()->range();
1137   assert(r->cnt() > TypeFunc::Parms, "sanity");
1138   const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1139   assert(t != nullptr, "sanity");
1140 
1141   CompileLog* log = C->log();
1142   if (log != nullptr) {
1143     log->head("eliminate_boxing type='%d'",
1144               log->identify(t->instance_klass()));
1145     JVMState* p = boxing->jvms();
1146     while (p != nullptr) {
1147       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1148       p = p->caller();
1149     }
1150     log->tail("eliminate_boxing");
1151   }
1152 
1153   process_users_of_allocation(boxing);
1154 
1155 #ifndef PRODUCT
1156   if (PrintEliminateAllocations) {

1220 // oop flavor.
1221 //
1222 //=============================================================================
1223 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1224 // Allocations bigger than this always go the slow route.
1225 // This value must be small enough that allocation attempts that need to
1226 // trigger exceptions go the slow route.  Also, it must be small enough so
1227 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1228 //=============================================================================j//
1229 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1230 // The allocator will coalesce int->oop copies away.  See comment in
1231 // coalesce.cpp about how this works.  It depends critically on the exact
1232 // code shape produced here, so if you are changing this code shape
1233 // make sure the GC info for the heap-top is correct in and around the
1234 // slow-path call.
1235 //
1236 
1237 void PhaseMacroExpand::expand_allocate_common(
1238             AllocateNode* alloc, // allocation node to be expanded
1239             Node* length,  // array length for an array allocation

1240             const TypeFunc* slow_call_type, // Type of slow call
1241             address slow_call_address,  // Address of slow call
1242             Node* valid_length_test // whether length is valid or not
1243     )
1244 {
1245   Node* ctrl = alloc->in(TypeFunc::Control);
1246   Node* mem  = alloc->in(TypeFunc::Memory);
1247   Node* i_o  = alloc->in(TypeFunc::I_O);
1248   Node* size_in_bytes     = alloc->in(AllocateNode::AllocSize);
1249   Node* klass_node        = alloc->in(AllocateNode::KlassNode);
1250   Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1251   assert(ctrl != nullptr, "must have control");
1252 
1253   // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1254   // they will not be used if "always_slow" is set
1255   enum { slow_result_path = 1, fast_result_path = 2 };
1256   Node *result_region = nullptr;
1257   Node *result_phi_rawmem = nullptr;
1258   Node *result_phi_rawoop = nullptr;
1259   Node *result_phi_i_o = nullptr;

1302 #endif
1303       yank_alloc_node(alloc);
1304       return;
1305     }
1306   }
1307 
1308   enum { too_big_or_final_path = 1, need_gc_path = 2 };
1309   Node *slow_region = nullptr;
1310   Node *toobig_false = ctrl;
1311 
1312   // generate the initial test if necessary
1313   if (initial_slow_test != nullptr ) {
1314     assert (expand_fast_path, "Only need test if there is a fast path");
1315     slow_region = new RegionNode(3);
1316 
1317     // Now make the initial failure test.  Usually a too-big test but
1318     // might be a TRUE for finalizers.
1319     IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1320     transform_later(toobig_iff);
1321     // Plug the failing-too-big test into the slow-path region
1322     Node *toobig_true = new IfTrueNode( toobig_iff );
1323     transform_later(toobig_true);
1324     slow_region    ->init_req( too_big_or_final_path, toobig_true );
1325     toobig_false = new IfFalseNode( toobig_iff );
1326     transform_later(toobig_false);
1327   } else {
1328     // No initial test, just fall into next case
1329     assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1330     toobig_false = ctrl;
1331     debug_only(slow_region = NodeSentinel);
1332   }
1333 
1334   // If we are here there are several possibilities
1335   // - expand_fast_path is false - then only a slow path is expanded. That's it.
1336   // no_initial_check means a constant allocation.
1337   // - If check always evaluates to false -> expand_fast_path is false (see above)
1338   // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1339   // if !allocation_has_use the fast path is empty
1340   // if !allocation_has_use && no_initial_check
1341   // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1342   //   removed by yank_alloc_node above.
1343 
1344   Node *slow_mem = mem;  // save the current memory state for slow path
1345   // generate the fast allocation code unless we know that the initial test will always go slow
1346   if (expand_fast_path) {
1347     // Fast path modifies only raw memory.
1348     if (mem->is_MergeMem()) {
1349       mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1350     }
1351 
1352     // allocate the Region and Phi nodes for the result
1353     result_region = new RegionNode(3);
1354     result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1355     result_phi_i_o    = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1356 
1357     // Grab regular I/O before optional prefetch may change it.
1358     // Slow-path does no I/O so just set it to the original I/O.
1359     result_phi_i_o->init_req(slow_result_path, i_o);
1360 
1361     // Name successful fast-path variables
1362     Node* fast_oop_ctrl;
1363     Node* fast_oop_rawmem;

1364     if (allocation_has_use) {
1365       Node* needgc_ctrl = nullptr;
1366       result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1367 
1368       intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1369       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1370       Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1371                                         fast_oop_ctrl, fast_oop_rawmem,
1372                                         prefetch_lines);
1373 
1374       if (initial_slow_test != nullptr) {
1375         // This completes all paths into the slow merge point
1376         slow_region->init_req(need_gc_path, needgc_ctrl);
1377         transform_later(slow_region);
1378       } else {
1379         // No initial slow path needed!
1380         // Just fall from the need-GC path straight into the VM call.
1381         slow_region = needgc_ctrl;
1382       }
1383 

1401     result_phi_i_o   ->init_req(fast_result_path, i_o);
1402     result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1403   } else {
1404     slow_region = ctrl;
1405     result_phi_i_o = i_o; // Rename it to use in the following code.
1406   }
1407 
1408   // Generate slow-path call
1409   CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1410                                OptoRuntime::stub_name(slow_call_address),
1411                                TypePtr::BOTTOM);
1412   call->init_req(TypeFunc::Control,   slow_region);
1413   call->init_req(TypeFunc::I_O,       top());    // does no i/o
1414   call->init_req(TypeFunc::Memory,    slow_mem); // may gc ptrs
1415   call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1416   call->init_req(TypeFunc::FramePtr,  alloc->in(TypeFunc::FramePtr));
1417 
1418   call->init_req(TypeFunc::Parms+0, klass_node);
1419   if (length != nullptr) {
1420     call->init_req(TypeFunc::Parms+1, length);






1421   }
1422 
1423   // Copy debug information and adjust JVMState information, then replace
1424   // allocate node with the call
1425   call->copy_call_debug_info(&_igvn, alloc);
1426   // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1427   // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1428   // path dies).
1429   if (valid_length_test != nullptr) {
1430     call->add_req(valid_length_test);
1431   }
1432   if (expand_fast_path) {
1433     call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1434   } else {
1435     // Hook i_o projection to avoid its elimination during allocation
1436     // replacement (when only a slow call is generated).
1437     call->set_req(TypeFunc::I_O, result_phi_i_o);
1438   }
1439   _igvn.replace_node(alloc, call);
1440   transform_later(call);
1441 
1442   // Identify the output projections from the allocate node and
1443   // adjust any references to them.
1444   // The control and io projections look like:
1445   //
1446   //        v---Proj(ctrl) <-----+   v---CatchProj(ctrl)
1447   //  Allocate                   Catch
1448   //        ^---Proj(io) <-------+   ^---CatchProj(io)
1449   //
1450   //  We are interested in the CatchProj nodes.
1451   //
1452   call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1453 
1454   // An allocate node has separate memory projections for the uses on
1455   // the control and i_o paths. Replace the control memory projection with
1456   // result_phi_rawmem (unless we are only generating a slow call when
1457   // both memory projections are combined)
1458   if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) {
1459     migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1460   }
1461   // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1462   // catchall_memproj so we end up with a call that has only 1 memory projection.
1463   if (_callprojs.catchall_memproj != nullptr ) {
1464     if (_callprojs.fallthrough_memproj == nullptr) {
1465       _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1466       transform_later(_callprojs.fallthrough_memproj);
1467     }
1468     migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1469     _igvn.remove_dead_node(_callprojs.catchall_memproj);
1470   }
1471 
1472   // An allocate node has separate i_o projections for the uses on the control
1473   // and i_o paths. Always replace the control i_o projection with result i_o
1474   // otherwise incoming i_o become dead when only a slow call is generated
1475   // (it is different from memory projections where both projections are
1476   // combined in such case).
1477   if (_callprojs.fallthrough_ioproj != nullptr) {
1478     migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1479   }
1480   // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1481   // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1482   if (_callprojs.catchall_ioproj != nullptr ) {
1483     if (_callprojs.fallthrough_ioproj == nullptr) {
1484       _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1485       transform_later(_callprojs.fallthrough_ioproj);
1486     }
1487     migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1488     _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1489   }
1490 
1491   // if we generated only a slow call, we are done
1492   if (!expand_fast_path) {
1493     // Now we can unhook i_o.
1494     if (result_phi_i_o->outcnt() > 1) {
1495       call->set_req(TypeFunc::I_O, top());
1496     } else {
1497       assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1498       // Case of new array with negative size known during compilation.
1499       // AllocateArrayNode::Ideal() optimization disconnect unreachable
1500       // following code since call to runtime will throw exception.
1501       // As result there will be no users of i_o after the call.
1502       // Leave i_o attached to this call to avoid problems in preceding graph.
1503     }
1504     return;
1505   }
1506 
1507   if (_callprojs.fallthrough_catchproj != nullptr) {
1508     ctrl = _callprojs.fallthrough_catchproj->clone();
1509     transform_later(ctrl);
1510     _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1511   } else {
1512     ctrl = top();
1513   }
1514   Node *slow_result;
1515   if (_callprojs.resproj == nullptr) {
1516     // no uses of the allocation result
1517     slow_result = top();
1518   } else {
1519     slow_result = _callprojs.resproj->clone();
1520     transform_later(slow_result);
1521     _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1522   }
1523 
1524   // Plug slow-path into result merge point
1525   result_region->init_req( slow_result_path, ctrl);
1526   transform_later(result_region);
1527   if (allocation_has_use) {
1528     result_phi_rawoop->init_req(slow_result_path, slow_result);
1529     transform_later(result_phi_rawoop);
1530   }
1531   result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1532   transform_later(result_phi_rawmem);
1533   transform_later(result_phi_i_o);
1534   // This completes all paths into the result merge point
1535 }
1536 
1537 // Remove alloc node that has no uses.
1538 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1539   Node* ctrl = alloc->in(TypeFunc::Control);
1540   Node* mem  = alloc->in(TypeFunc::Memory);
1541   Node* i_o  = alloc->in(TypeFunc::I_O);
1542 
1543   alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1544   if (_callprojs.resproj != nullptr) {
1545     for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1546       Node* use = _callprojs.resproj->fast_out(i);
1547       use->isa_MemBar()->remove(&_igvn);
1548       --imax;
1549       --i; // back up iterator
1550     }
1551     assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1552     _igvn.remove_dead_node(_callprojs.resproj);
1553   }
1554   if (_callprojs.fallthrough_catchproj != nullptr) {
1555     migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1556     _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1557   }
1558   if (_callprojs.catchall_catchproj != nullptr) {
1559     _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1560     _callprojs.catchall_catchproj->set_req(0, top());
1561   }
1562   if (_callprojs.fallthrough_proj != nullptr) {
1563     Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1564     _igvn.remove_dead_node(catchnode);
1565     _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1566   }
1567   if (_callprojs.fallthrough_memproj != nullptr) {
1568     migrate_outs(_callprojs.fallthrough_memproj, mem);
1569     _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1570   }
1571   if (_callprojs.fallthrough_ioproj != nullptr) {
1572     migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1573     _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1574   }
1575   if (_callprojs.catchall_memproj != nullptr) {
1576     _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1577     _callprojs.catchall_memproj->set_req(0, top());
1578   }
1579   if (_callprojs.catchall_ioproj != nullptr) {
1580     _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1581     _callprojs.catchall_ioproj->set_req(0, top());
1582   }
1583 #ifndef PRODUCT
1584   if (PrintEliminateAllocations) {
1585     if (alloc->is_AllocateArray()) {
1586       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1587     } else {
1588       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1589     }
1590   }
1591 #endif
1592   _igvn.remove_dead_node(alloc);
1593 }
1594 
1595 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1596                                                 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1597   // If initialization is performed by an array copy, any required
1598   // MemBarStoreStore was already added. If the object does not
1599   // escape no need for a MemBarStoreStore. If the object does not
1600   // escape in its initializer and memory barrier (MemBarStoreStore or
1601   // stronger) is already added at exit of initializer, also no need

1679     Node* thread = new ThreadLocalNode();
1680     transform_later(thread);
1681 
1682     call->init_req(TypeFunc::Parms + 0, thread);
1683     call->init_req(TypeFunc::Parms + 1, oop);
1684     call->init_req(TypeFunc::Control, ctrl);
1685     call->init_req(TypeFunc::I_O    , top()); // does no i/o
1686     call->init_req(TypeFunc::Memory , rawmem);
1687     call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1688     call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1689     transform_later(call);
1690     ctrl = new ProjNode(call, TypeFunc::Control);
1691     transform_later(ctrl);
1692     rawmem = new ProjNode(call, TypeFunc::Memory);
1693     transform_later(rawmem);
1694   }
1695 }
1696 
1697 // Helper for PhaseMacroExpand::expand_allocate_common.
1698 // Initializes the newly-allocated storage.
1699 Node*
1700 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1701                                     Node* control, Node* rawmem, Node* object,
1702                                     Node* klass_node, Node* length,
1703                                     Node* size_in_bytes) {
1704   InitializeNode* init = alloc->initialization();
1705   // Store the klass & mark bits
1706   Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1707   if (!mark_node->is_Con()) {
1708     transform_later(mark_node);
1709   }
1710   rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1711 
1712   if (!UseCompactObjectHeaders) {
1713     rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1714   }
1715   int header_size = alloc->minimum_header_size();  // conservatively small
1716 
1717   // Array length
1718   if (length != nullptr) {         // Arrays need length field
1719     rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1720     // conservatively small header size:
1721     header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1722     if (_igvn.type(klass_node)->isa_aryklassptr()) {   // we know the exact header size in most cases:
1723       BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1724       if (is_reference_type(elem, true)) {
1725         elem = T_OBJECT;
1726       }
1727       header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1728     }
1729   }
1730 
1731   // Clear the object body, if necessary.
1732   if (init == nullptr) {
1733     // The init has somehow disappeared; be cautious and clear everything.
1734     //
1735     // This can happen if a node is allocated but an uncommon trap occurs
1736     // immediately.  In this case, the Initialize gets associated with the
1737     // trap, and may be placed in a different (outer) loop, if the Allocate
1738     // is in a loop.  If (this is rare) the inner loop gets unrolled, then
1739     // there can be two Allocates to one Initialize.  The answer in all these
1740     // edge cases is safety first.  It is always safe to clear immediately
1741     // within an Allocate, and then (maybe or maybe not) clear some more later.
1742     if (!(UseTLAB && ZeroTLAB)) {
1743       rawmem = ClearArrayNode::clear_memory(control, rawmem, object,


1744                                             header_size, size_in_bytes,
1745                                             &_igvn);
1746     }
1747   } else {
1748     if (!init->is_complete()) {
1749       // Try to win by zeroing only what the init does not store.
1750       // We can also try to do some peephole optimizations,
1751       // such as combining some adjacent subword stores.
1752       rawmem = init->complete_stores(control, rawmem, object,
1753                                      header_size, size_in_bytes, &_igvn);
1754     }
1755     // We have no more use for this link, since the AllocateNode goes away:
1756     init->set_req(InitializeNode::RawAddress, top());
1757     // (If we keep the link, it just confuses the register allocator,
1758     // who thinks he sees a real use of the address by the membar.)
1759   }
1760 
1761   return rawmem;
1762 }
1763 

1898       for ( intx i = 0; i < lines; i++ ) {
1899         prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1900                                             _igvn.MakeConX(distance) );
1901         transform_later(prefetch_adr);
1902         prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1903         // Do not let it float too high, since if eden_top == eden_end,
1904         // both might be null.
1905         if( i == 0 ) { // Set control for first prefetch, next follows it
1906           prefetch->init_req(0, needgc_false);
1907         }
1908         transform_later(prefetch);
1909         distance += step_size;
1910         i_o = prefetch;
1911       }
1912    }
1913    return i_o;
1914 }
1915 
1916 
1917 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1918   expand_allocate_common(alloc, nullptr,
1919                          OptoRuntime::new_instance_Type(),
1920                          OptoRuntime::new_instance_Java(), nullptr);
1921 }
1922 
1923 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1924   Node* length = alloc->in(AllocateNode::ALength);
1925   Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1926   InitializeNode* init = alloc->initialization();
1927   Node* klass_node = alloc->in(AllocateNode::KlassNode);

1928   const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();

1929   address slow_call_address;  // Address of slow call
1930   if (init != nullptr && init->is_complete_with_arraycopy() &&
1931       ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
1932     // Don't zero type array during slow allocation in VM since
1933     // it will be initialized later by arraycopy in compiled code.
1934     slow_call_address = OptoRuntime::new_array_nozero_Java();

1935   } else {
1936     slow_call_address = OptoRuntime::new_array_Java();







1937   }
1938   expand_allocate_common(alloc, length,
1939                          OptoRuntime::new_array_Type(),
1940                          slow_call_address, valid_length_test);
1941 }
1942 
1943 //-------------------mark_eliminated_box----------------------------------
1944 //
1945 // During EA obj may point to several objects but after few ideal graph
1946 // transformations (CCP) it may point to only one non escaping object
1947 // (but still using phi), corresponding locks and unlocks will be marked
1948 // for elimination. Later obj could be replaced with a new node (new phi)
1949 // and which does not have escape information. And later after some graph
1950 // reshape other locks and unlocks (which were not marked for elimination
1951 // before) are connected to this new obj (phi) but they still will not be
1952 // marked for elimination since new obj has no escape information.
1953 // Mark all associated (same box and obj) lock and unlock nodes for
1954 // elimination if some of them marked already.
1955 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
1956   BoxLockNode* oldbox = box->as_BoxLock();
1957   if (oldbox->is_eliminated()) {
1958     return; // This BoxLock node was processed already.
1959   }

2131 #ifdef ASSERT
2132   if (!alock->is_coarsened()) {
2133     // Check that new "eliminated" BoxLock node is created.
2134     BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2135     assert(oldbox->is_eliminated(), "should be done already");
2136   }
2137 #endif
2138 
2139   alock->log_lock_optimization(C, "eliminate_lock");
2140 
2141 #ifndef PRODUCT
2142   if (PrintEliminateLocks) {
2143     tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2144   }
2145 #endif
2146 
2147   Node* mem  = alock->in(TypeFunc::Memory);
2148   Node* ctrl = alock->in(TypeFunc::Control);
2149   guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2150 
2151   alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2152   // There are 2 projections from the lock.  The lock node will
2153   // be deleted when its last use is subsumed below.
2154   assert(alock->outcnt() == 2 &&
2155          _callprojs.fallthrough_proj != nullptr &&
2156          _callprojs.fallthrough_memproj != nullptr,
2157          "Unexpected projections from Lock/Unlock");
2158 
2159   Node* fallthroughproj = _callprojs.fallthrough_proj;
2160   Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2161 
2162   // The memory projection from a lock/unlock is RawMem
2163   // The input to a Lock is merged memory, so extract its RawMem input
2164   // (unless the MergeMem has been optimized away.)
2165   if (alock->is_Lock()) {
2166     // Search for MemBarAcquireLock node and delete it also.
2167     MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2168     assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2169     Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2170     Node* memproj = membar->proj_out(TypeFunc::Memory);
2171     _igvn.replace_node(ctrlproj, fallthroughproj);
2172     _igvn.replace_node(memproj, memproj_fallthrough);
2173 
2174     // Delete FastLock node also if this Lock node is unique user
2175     // (a loop peeling may clone a Lock node).
2176     Node* flock = alock->as_Lock()->fastlock_node();
2177     if (flock->outcnt() == 1) {
2178       assert(flock->unique_out() == alock, "sanity");
2179       _igvn.replace_node(flock, top());
2180     }

2211   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2212 
2213   // Make the merge point
2214   Node *region;
2215   Node *mem_phi;
2216   Node *slow_path;
2217 
2218   region  = new RegionNode(3);
2219   // create a Phi for the memory state
2220   mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2221 
2222   // Optimize test; set region slot 2
2223   slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2224   mem_phi->init_req(2, mem);
2225 
2226   // Make slow path call
2227   CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2228                                   OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2229                                   obj, box, nullptr);
2230 
2231   call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2232 
2233   // Slow path can only throw asynchronous exceptions, which are always
2234   // de-opted.  So the compiler thinks the slow-call can never throw an
2235   // exception.  If it DOES throw an exception we would need the debug
2236   // info removed first (since if it throws there is no monitor).
2237   assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2238          _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2239 
2240   // Capture slow path
2241   // disconnect fall-through projection from call and create a new one
2242   // hook up users of fall-through projection to region
2243   Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2244   transform_later(slow_ctrl);
2245   _igvn.hash_delete(_callprojs.fallthrough_proj);
2246   _callprojs.fallthrough_proj->disconnect_inputs(C);
2247   region->init_req(1, slow_ctrl);
2248   // region inputs are now complete
2249   transform_later(region);
2250   _igvn.replace_node(_callprojs.fallthrough_proj, region);
2251 
2252   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2253 
2254   mem_phi->init_req(1, memproj);
2255 
2256   transform_later(mem_phi);
2257 
2258   _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2259 }
2260 
2261 //------------------------------expand_unlock_node----------------------
2262 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2263 
2264   Node* ctrl = unlock->in(TypeFunc::Control);
2265   Node* mem = unlock->in(TypeFunc::Memory);
2266   Node* obj = unlock->obj_node();
2267   Node* box = unlock->box_node();
2268 
2269   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2270 
2271   // No need for a null check on unlock
2272 
2273   // Make the merge point
2274   Node *region;
2275   Node *mem_phi;
2276 
2277   region  = new RegionNode(3);
2278   // create a Phi for the memory state
2279   mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2280 
2281   FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2282   funlock = transform_later( funlock )->as_FastUnlock();
2283   // Optimize test; set region slot 2
2284   Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2285   Node *thread = transform_later(new ThreadLocalNode());
2286 
2287   CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2288                                   CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2289                                   "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2290 
2291   call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2292   assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2293          _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2294 
2295   // No exceptions for unlocking
2296   // Capture slow path
2297   // disconnect fall-through projection from call and create a new one
2298   // hook up users of fall-through projection to region
2299   Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2300   transform_later(slow_ctrl);
2301   _igvn.hash_delete(_callprojs.fallthrough_proj);
2302   _callprojs.fallthrough_proj->disconnect_inputs(C);
2303   region->init_req(1, slow_ctrl);
2304   // region inputs are now complete
2305   transform_later(region);
2306   _igvn.replace_node(_callprojs.fallthrough_proj, region);
2307 
2308   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2309   mem_phi->init_req(1, memproj );
2310   mem_phi->init_req(2, mem);
2311   transform_later(mem_phi);
2312 
2313   _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2314 }
2315 



















































































































































































































2316 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2317   assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2318   Node* bol = check->unique_out();
2319   Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2320   Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2321   assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2322 
2323   for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2324     Node* iff = bol->last_out(i);
2325     assert(iff->is_If(), "where's the if?");
2326 
2327     if (iff->in(0)->is_top()) {
2328       _igvn.replace_input_of(iff, 1, C->top());
2329       continue;
2330     }
2331 
2332     Node* iftrue = iff->as_If()->proj_out(1);
2333     Node* iffalse = iff->as_If()->proj_out(0);
2334     Node* ctrl = iff->in(0);
2335 
2336     Node* subklass = nullptr;
2337     if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2338       subklass = obj_or_subklass;
2339     } else {
2340       Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2341       subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2342     }
2343 
2344     Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2345 
2346     _igvn.replace_input_of(iff, 0, C->top());
2347     _igvn.replace_node(iftrue, not_subtype_ctrl);
2348     _igvn.replace_node(iffalse, ctrl);
2349   }
2350   _igvn.replace_node(check, C->top());
2351 }
2352 

















































































































2353 //---------------------------eliminate_macro_nodes----------------------
2354 // Eliminate scalar replaced allocations and associated locks.
2355 void PhaseMacroExpand::eliminate_macro_nodes() {
2356   if (C->macro_count() == 0)
2357     return;
2358   NOT_PRODUCT(int membar_before = count_MemBar(C);)
2359 
2360   // Before elimination may re-mark (change to Nested or NonEscObj)
2361   // all associated (same box and obj) lock and unlock nodes.
2362   int cnt = C->macro_count();
2363   for (int i=0; i < cnt; i++) {
2364     Node *n = C->macro_node(i);
2365     if (n->is_AbstractLock()) { // Lock and Unlock nodes
2366       mark_eliminated_locking_nodes(n->as_AbstractLock());
2367     }
2368   }
2369   // Re-marking may break consistency of Coarsened locks.
2370   if (!C->coarsened_locks_consistent()) {
2371     return; // recompile without Coarsened locks if broken
2372   } else {

2398   }
2399   // Next, attempt to eliminate allocations
2400   _has_locks = false;
2401   progress = true;
2402   while (progress) {
2403     progress = false;
2404     for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2405       Node* n = C->macro_node(i - 1);
2406       bool success = false;
2407       DEBUG_ONLY(int old_macro_count = C->macro_count();)
2408       switch (n->class_id()) {
2409       case Node::Class_Allocate:
2410       case Node::Class_AllocateArray:
2411         success = eliminate_allocate_node(n->as_Allocate());
2412 #ifndef PRODUCT
2413         if (success && PrintOptoStatistics) {
2414           Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2415         }
2416 #endif
2417         break;
2418       case Node::Class_CallStaticJava:
2419         success = eliminate_boxing_node(n->as_CallStaticJava());



2420         break;

2421       case Node::Class_Lock:
2422       case Node::Class_Unlock:
2423         assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2424         _has_locks = true;
2425         break;
2426       case Node::Class_ArrayCopy:
2427         break;
2428       case Node::Class_OuterStripMinedLoop:
2429         break;
2430       case Node::Class_SubTypeCheck:
2431         break;
2432       case Node::Class_Opaque1:
2433         break;


2434       default:
2435         assert(n->Opcode() == Op_LoopLimit ||
2436                n->Opcode() == Op_ModD ||
2437                n->Opcode() == Op_ModF ||
2438                n->is_OpaqueNotNull()       ||
2439                n->is_OpaqueInitializedAssertionPredicate() ||
2440                n->Opcode() == Op_MaxL      ||
2441                n->Opcode() == Op_MinL      ||
2442                BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2443                "unknown node type in macro list");
2444       }
2445       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2446       progress = progress || success;
2447     }
2448   }
2449 #ifndef PRODUCT
2450   if (PrintOptoStatistics) {
2451     int membar_after = count_MemBar(C);
2452     Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2453   }

2463     C->shuffle_macro_nodes();
2464   }
2465   // Last attempt to eliminate macro nodes.
2466   eliminate_macro_nodes();
2467   if (C->failing())  return true;
2468 
2469   // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2470   bool progress = true;
2471   while (progress) {
2472     progress = false;
2473     for (int i = C->macro_count(); i > 0; i--) {
2474       Node* n = C->macro_node(i-1);
2475       bool success = false;
2476       DEBUG_ONLY(int old_macro_count = C->macro_count();)
2477       if (n->Opcode() == Op_LoopLimit) {
2478         // Remove it from macro list and put on IGVN worklist to optimize.
2479         C->remove_macro_node(n);
2480         _igvn._worklist.push(n);
2481         success = true;
2482       } else if (n->Opcode() == Op_CallStaticJava) {
2483         // Remove it from macro list and put on IGVN worklist to optimize.
2484         C->remove_macro_node(n);
2485         _igvn._worklist.push(n);
2486         success = true;



2487       } else if (n->is_Opaque1()) {
2488         _igvn.replace_node(n, n->in(1));
2489         success = true;
2490       } else if (n->is_OpaqueNotNull()) {
2491         // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
2492         // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
2493         // we will execute a Halt node.
2494 #ifdef ASSERT
2495         _igvn.replace_node(n, n->in(1));
2496 #else
2497         _igvn.replace_node(n, _igvn.intcon(1));
2498 #endif
2499         success = true;
2500       } else if (n->is_OpaqueInitializedAssertionPredicate()) {
2501           // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
2502           // builds as they are useless. In debug builds we keep them as additional verification code. Even though
2503           // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
2504           // enable folding of dead control paths within which cast nodes become top after due to impossible types -
2505           // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
2506 #ifdef ASSERT

2568     // Worst case is a macro node gets expanded into about 200 nodes.
2569     // Allow 50% more for optimization.
2570     if (C->check_node_count(300, "out of nodes before macro expansion")) {
2571       return true;
2572     }
2573 
2574     DEBUG_ONLY(int old_macro_count = C->macro_count();)
2575     switch (n->class_id()) {
2576     case Node::Class_Lock:
2577       expand_lock_node(n->as_Lock());
2578       break;
2579     case Node::Class_Unlock:
2580       expand_unlock_node(n->as_Unlock());
2581       break;
2582     case Node::Class_ArrayCopy:
2583       expand_arraycopy_node(n->as_ArrayCopy());
2584       break;
2585     case Node::Class_SubTypeCheck:
2586       expand_subtypecheck_node(n->as_SubTypeCheck());
2587       break;







2588     default:
2589       switch (n->Opcode()) {
2590       case Op_ModD:
2591       case Op_ModF: {
2592         bool is_drem = n->Opcode() == Op_ModD;
2593         CallNode* mod_macro = n->as_Call();
2594         CallNode* call = new CallLeafNode(mod_macro->tf(),
2595                                           is_drem ? CAST_FROM_FN_PTR(address, SharedRuntime::drem)
2596                                                   : CAST_FROM_FN_PTR(address, SharedRuntime::frem),
2597                                           is_drem ? "drem" : "frem", TypeRawPtr::BOTTOM);
2598         call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
2599         call->init_req(TypeFunc::I_O, mod_macro->in(TypeFunc::I_O));
2600         call->init_req(TypeFunc::Memory, mod_macro->in(TypeFunc::Memory));
2601         call->init_req(TypeFunc::ReturnAdr, mod_macro->in(TypeFunc::ReturnAdr));
2602         call->init_req(TypeFunc::FramePtr, mod_macro->in(TypeFunc::FramePtr));
2603         for (unsigned int i = 0; i < mod_macro->tf()->domain()->cnt() - TypeFunc::Parms; i++) {
2604           call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
2605         }
2606         _igvn.replace_node(mod_macro, call);
2607         transform_later(call);
2608         break;
2609       }
2610       default:
2611         assert(false, "unknown node type in macro list");
2612       }
2613     }
2614     assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2615     if (C->failing())  return true;
2616     C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
2617 
2618     // Clean up the graph so we're less likely to hit the maximum node
2619     // limit
2620     _igvn.set_delay_transform(false);
2621     _igvn.optimize();
2622     if (C->failing())  return true;
2623     _igvn.set_delay_transform(true);

   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "ci/ciFlatArrayKlass.hpp"
  26 #include "compiler/compileLog.hpp"
  27 #include "gc/shared/collectedHeap.inline.hpp"
  28 #include "gc/shared/tlab_globals.hpp"
  29 #include "libadt/vectset.hpp"
  30 #include "memory/universe.hpp"
  31 #include "opto/addnode.hpp"
  32 #include "opto/arraycopynode.hpp"
  33 #include "opto/callnode.hpp"
  34 #include "opto/castnode.hpp"
  35 #include "opto/cfgnode.hpp"
  36 #include "opto/compile.hpp"
  37 #include "opto/convertnode.hpp"
  38 #include "opto/graphKit.hpp"
  39 #include "opto/inlinetypenode.hpp"
  40 #include "opto/intrinsicnode.hpp"
  41 #include "opto/locknode.hpp"
  42 #include "opto/loopnode.hpp"
  43 #include "opto/macro.hpp"
  44 #include "opto/memnode.hpp"
  45 #include "opto/narrowptrnode.hpp"
  46 #include "opto/node.hpp"
  47 #include "opto/opaquenode.hpp"
  48 #include "opto/phaseX.hpp"
  49 #include "opto/rootnode.hpp"
  50 #include "opto/runtime.hpp"
  51 #include "opto/subnode.hpp"
  52 #include "opto/subtypenode.hpp"
  53 #include "opto/type.hpp"
  54 #include "prims/jvmtiExport.hpp"
  55 #include "runtime/continuation.hpp"
  56 #include "runtime/sharedRuntime.hpp"
  57 #include "runtime/stubRoutines.hpp"
  58 #include "utilities/macros.hpp"
  59 #include "utilities/powerOfTwo.hpp"
  60 #if INCLUDE_G1GC
  61 #include "gc/g1/g1ThreadLocalData.hpp"
  62 #endif // INCLUDE_G1GC
  63 
  64 
  65 //
  66 // Replace any references to "oldref" in inputs to "use" with "newref".
  67 // Returns the number of replacements made.
  68 //
  69 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
  70   int nreplacements = 0;
  71   uint req = use->req();
  72   for (uint j = 0; j < use->len(); j++) {
  73     Node *uin = use->in(j);
  74     if (uin == oldref) {
  75       if (j < req)
  76         use->set_req(j, newref);
  77       else
  78         use->set_prec(j, newref);
  79       nreplacements++;
  80     } else if (j >= req && uin == nullptr) {
  81       break;
  82     }
  83   }
  84   return nreplacements;
  85 }
  86 












  87 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
  88   Node* cmp;
  89   if (mask != 0) {
  90     Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
  91     cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
  92   } else {
  93     cmp = word;
  94   }
  95   Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
  96   IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
  97   transform_later(iff);
  98 
  99   // Fast path taken.
 100   Node *fast_taken = transform_later(new IfFalseNode(iff));
 101 
 102   // Fast path not-taken, i.e. slow path
 103   Node *slow_taken = transform_later(new IfTrueNode(iff));
 104 
 105   if (return_fast_path) {
 106     region->init_req(edge, slow_taken); // Capture slow-control

 129   // Slow-path call
 130  CallNode *call = leaf_name
 131    ? (CallNode*)new CallLeafNode      ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
 132    : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
 133 
 134   // Slow path call has no side-effects, uses few values
 135   copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
 136   if (parm0 != nullptr)  call->init_req(TypeFunc::Parms+0, parm0);
 137   if (parm1 != nullptr)  call->init_req(TypeFunc::Parms+1, parm1);
 138   if (parm2 != nullptr)  call->init_req(TypeFunc::Parms+2, parm2);
 139   call->copy_call_debug_info(&_igvn, oldcall);
 140   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
 141   _igvn.replace_node(oldcall, call);
 142   transform_later(call);
 143 
 144   return call;
 145 }
 146 
 147 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
 148   BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
 149   bs->eliminate_gc_barrier(&_igvn, p2x);
 150 #ifndef PRODUCT
 151   if (PrintOptoStatistics) {
 152     Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
 153   }
 154 #endif
 155 }
 156 
 157 // Search for a memory operation for the specified memory slice.
 158 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
 159   Node *orig_mem = mem;
 160   Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 161   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 162   const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
 163   while (true) {
 164     if (mem == alloc_mem || mem == start_mem ) {
 165       return mem;  // hit one of our sentinels
 166     } else if (mem->is_MergeMem()) {
 167       mem = mem->as_MergeMem()->memory_at(alias_idx);
 168     } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
 169       Node *in = mem->in(0);

 184         ArrayCopyNode* ac = nullptr;
 185         if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
 186           if (ac != nullptr) {
 187             assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
 188             return ac;
 189           }
 190         }
 191         mem = in->in(TypeFunc::Memory);
 192       } else {
 193 #ifdef ASSERT
 194         in->dump();
 195         mem->dump();
 196         assert(false, "unexpected projection");
 197 #endif
 198       }
 199     } else if (mem->is_Store()) {
 200       const TypePtr* atype = mem->as_Store()->adr_type();
 201       int adr_idx = phase->C->get_alias_index(atype);
 202       if (adr_idx == alias_idx) {
 203         assert(atype->isa_oopptr(), "address type must be oopptr");
 204         int adr_offset = atype->flat_offset();
 205         uint adr_iid = atype->is_oopptr()->instance_id();
 206         // Array elements references have the same alias_idx
 207         // but different offset and different instance_id.
 208         if (adr_offset == offset && adr_iid == alloc->_idx) {
 209           return mem;
 210         }
 211       } else {
 212         assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
 213       }
 214       mem = mem->in(MemNode::Memory);
 215     } else if (mem->is_ClearArray()) {
 216       if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
 217         // Can not bypass initialization of the instance
 218         // we are looking.
 219         debug_only(intptr_t offset;)
 220         assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
 221         InitializeNode* init = alloc->as_Allocate()->initialization();
 222         // We are looking for stored value, return Initialize node
 223         // or memory edge from Allocate node.
 224         if (init != nullptr) {

 229       }
 230       // Otherwise skip it (the call updated 'mem' value).
 231     } else if (mem->Opcode() == Op_SCMemProj) {
 232       mem = mem->in(0);
 233       Node* adr = nullptr;
 234       if (mem->is_LoadStore()) {
 235         adr = mem->in(MemNode::Address);
 236       } else {
 237         assert(mem->Opcode() == Op_EncodeISOArray ||
 238                mem->Opcode() == Op_StrCompressedCopy, "sanity");
 239         adr = mem->in(3); // Destination array
 240       }
 241       const TypePtr* atype = adr->bottom_type()->is_ptr();
 242       int adr_idx = phase->C->get_alias_index(atype);
 243       if (adr_idx == alias_idx) {
 244         DEBUG_ONLY(mem->dump();)
 245         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 246         return nullptr;
 247       }
 248       mem = mem->in(MemNode::Memory);
 249     } else if (mem->Opcode() == Op_StrInflatedCopy) {
 250       Node* adr = mem->in(3); // Destination array
 251       const TypePtr* atype = adr->bottom_type()->is_ptr();
 252       int adr_idx = phase->C->get_alias_index(atype);
 253       if (adr_idx == alias_idx) {
 254         DEBUG_ONLY(mem->dump();)
 255         assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
 256         return nullptr;
 257       }
 258       mem = mem->in(MemNode::Memory);
 259     } else {
 260       return mem;
 261     }
 262     assert(mem != orig_mem, "dead memory loop");
 263   }
 264 }
 265 
 266 // Generate loads from source of the arraycopy for fields of
 267 // destination needed at a deoptimization point
 268 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
 269   BasicType bt = ft;

 274   }
 275   Node* res = nullptr;
 276   if (ac->is_clonebasic()) {
 277     assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
 278     Node* base = ac->in(ArrayCopyNode::Src);
 279     Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
 280     const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
 281     MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
 282     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 283     res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
 284   } else {
 285     if (ac->modifies(offset, offset, &_igvn, true)) {
 286       assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
 287       uint shift = exact_log2(type2aelembytes(bt));
 288       Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
 289       Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
 290       const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
 291       const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
 292 
 293       Node* adr = nullptr;
 294       Node* base = ac->in(ArrayCopyNode::Src);
 295       const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
 296       if (adr_type->is_flat()) {
 297         shift = adr_type->flat_log_elem_size();
 298       }
 299       if (src_pos_t->is_con() && dest_pos_t->is_con()) {
 300         intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;

 301         adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
 302         adr_type = _igvn.type(adr)->is_aryptr();
 303         assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
 304         if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
 305           // Don't emit a new load from src if src == dst but try to get the value from memory instead
 306           return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
 307         }
 308       } else {
 309         if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
 310           // Non constant offset in the array: we can't statically
 311           // determine the value
 312           return nullptr;
 313         }
 314         Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 315 #ifdef _LP64
 316         diff = _igvn.transform(new ConvI2LNode(diff));
 317 #endif
 318         diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
 319 
 320         Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));

 321         adr = _igvn.transform(new AddPNode(base, base, off));
 322         // In the case of a flat inline type array, each field has its
 323         // own slice so we need to extract the field being accessed from
 324         // the address computation
 325         adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
 326         adr = _igvn.transform(new CastPPNode(ctl, adr, adr_type));

 327       }
 328       MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
 329       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 330       res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
 331     }
 332   }
 333   if (res != nullptr) {
 334     if (ftype->isa_narrowoop()) {
 335       // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
 336       assert(res->isa_DecodeN(), "should be narrow oop");
 337       res = _igvn.transform(new EncodePNode(res, ftype));
 338     }
 339     return res;
 340   }
 341   return nullptr;
 342 }
 343 
 344 //
 345 // Given a Memory Phi, compute a value Phi containing the values from stores
 346 // on the input paths.
 347 // Note: this function is recursive, its depth is limited by the "level" argument
 348 // Returns the computed Phi, or null if it cannot compute it.
 349 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
 350   assert(mem->is_Phi(), "sanity");
 351   int alias_idx = C->get_alias_index(adr_t);
 352   int offset = adr_t->flat_offset();
 353   int instance_id = adr_t->instance_id();
 354 
 355   // Check if an appropriate value phi already exists.
 356   Node* region = mem->in(0);
 357   for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
 358     Node* phi = region->fast_out(k);
 359     if (phi->is_Phi() && phi != mem &&
 360         phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
 361       return phi;
 362     }
 363   }
 364   // Check if an appropriate new value phi already exists.
 365   Node* new_phi = value_phis->find(mem->_idx);
 366   if (new_phi != nullptr)
 367     return new_phi;
 368 
 369   if (level <= 0) {
 370     return nullptr; // Give up: phi tree too deep
 371   }
 372   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 373   Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 374   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 375 
 376   uint length = mem->req();
 377   GrowableArray <Node *> values(length, length, nullptr);
 378 
 379   // create a new Phi for the value
 380   PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
 381   transform_later(phi);
 382   value_phis->push(phi, mem->_idx);
 383 
 384   for (uint j = 1; j < length; j++) {
 385     Node *in = mem->in(j);
 386     if (in == nullptr || in->is_top()) {
 387       values.at_put(j, in);
 388     } else {
 389       Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
 390       if (val == start_mem || val == alloc_mem) {
 391         // hit a sentinel, return appropriate value
 392         Node* init_value = alloc->in(AllocateNode::InitValue);
 393         if (init_value != nullptr) {
 394           if (val == start_mem) {
 395             // TODO 8350865 Somehow we ended up with root mem and therefore walked past the alloc. Fix this. Triggered by TestGenerated::test15
 396             // Don't we need field_value_by_offset?
 397             return nullptr;
 398           }
 399           values.at_put(j, init_value);
 400         } else {
 401           assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
 402           values.at_put(j, _igvn.zerocon(ft));
 403         }
 404         continue;
 405       }
 406       if (val->is_Initialize()) {
 407         val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 408       }
 409       if (val == nullptr) {
 410         return nullptr;  // can't find a value on this path
 411       }
 412       if (val == mem) {
 413         values.at_put(j, mem);
 414       } else if (val->is_Store()) {
 415         Node* n = val->in(MemNode::ValueIn);
 416         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 417         n = bs->step_over_gc_barrier(n);
 418         if (is_subword_type(ft)) {
 419           n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
 420         }
 421         values.at_put(j, n);
 422       } else if (val->is_Proj() && val->in(0) == alloc) {
 423         Node* init_value = alloc->in(AllocateNode::InitValue);
 424         if (init_value != nullptr) {
 425           // TODO 8350865 Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
 426           values.at_put(j, init_value);
 427         } else {
 428           assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
 429           values.at_put(j, _igvn.zerocon(ft));
 430         }
 431       } else if (val->is_Phi()) {
 432         val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
 433         if (val == nullptr) {
 434           return nullptr;
 435         }
 436         values.at_put(j, val);
 437       } else if (val->Opcode() == Op_SCMemProj) {
 438         assert(val->in(0)->is_LoadStore() ||
 439                val->in(0)->Opcode() == Op_EncodeISOArray ||
 440                val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
 441         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 442         return nullptr;
 443       } else if (val->is_ArrayCopy()) {
 444         Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
 445         if (res == nullptr) {
 446           return nullptr;
 447         }
 448         values.at_put(j, res);
 449       } else if (val->is_top()) {
 450         // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region.

 458     }
 459   }
 460   // Set Phi's inputs
 461   for (uint j = 1; j < length; j++) {
 462     if (values.at(j) == mem) {
 463       phi->init_req(j, phi);
 464     } else {
 465       phi->init_req(j, values.at(j));
 466     }
 467   }
 468   return phi;
 469 }
 470 
 471 // Search the last value stored into the object's field.
 472 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
 473   assert(adr_t->is_known_instance_field(), "instance required");
 474   int instance_id = adr_t->instance_id();
 475   assert((uint)instance_id == alloc->_idx, "wrong allocation");
 476 
 477   int alias_idx = C->get_alias_index(adr_t);
 478   int offset = adr_t->flat_offset();
 479   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);

 480   Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false);
 481   assert(alloc_mem != nullptr, "Allocation without a memory projection.");
 482   VectorSet visited;
 483 
 484   bool done = sfpt_mem == alloc_mem;
 485   Node *mem = sfpt_mem;
 486   while (!done) {
 487     if (visited.test_set(mem->_idx)) {
 488       return nullptr;  // found a loop, give up
 489     }
 490     mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
 491     if (mem == start_mem || mem == alloc_mem) {
 492       done = true;  // hit a sentinel, return appropriate 0 value
 493     } else if (mem->is_Initialize()) {
 494       mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 495       if (mem == nullptr) {
 496         done = true; // Something went wrong.
 497       } else if (mem->is_Store()) {
 498         const TypePtr* atype = mem->as_Store()->adr_type();
 499         assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
 500         done = true;
 501       }
 502     } else if (mem->is_Store()) {
 503       const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
 504       assert(atype != nullptr, "address type must be oopptr");
 505       assert(C->get_alias_index(atype) == alias_idx &&
 506              atype->is_known_instance_field() && atype->flat_offset() == offset &&
 507              atype->instance_id() == instance_id, "store is correct memory slice");
 508       done = true;
 509     } else if (mem->is_Phi()) {
 510       // try to find a phi's unique input
 511       Node *unique_input = nullptr;
 512       Node *top = C->top();
 513       for (uint i = 1; i < mem->req(); i++) {
 514         Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
 515         if (n == nullptr || n == top || n == mem) {
 516           continue;
 517         } else if (unique_input == nullptr) {
 518           unique_input = n;
 519         } else if (unique_input != n) {
 520           unique_input = top;
 521           break;
 522         }
 523       }
 524       if (unique_input != nullptr && unique_input != top) {
 525         mem = unique_input;
 526       } else {
 527         done = true;
 528       }
 529     } else if (mem->is_ArrayCopy()) {
 530       done = true;
 531     } else {
 532       DEBUG_ONLY( mem->dump(); )
 533       assert(false, "unexpected node");
 534     }
 535   }
 536   if (mem != nullptr) {
 537     if (mem == start_mem || mem == alloc_mem) {
 538       // hit a sentinel, return appropriate value
 539       Node* init_value = alloc->in(AllocateNode::InitValue);
 540       if (init_value != nullptr) {
 541         if (adr_t->is_flat()) {
 542           if (init_value->is_EncodeP()) {
 543             init_value = init_value->in(1);
 544           }
 545           assert(adr_t->is_aryptr()->field_offset().get() != Type::OffsetBot, "Unknown offset");
 546           offset = adr_t->is_aryptr()->field_offset().get() + init_value->bottom_type()->inline_klass()->payload_offset();
 547           init_value = init_value->as_InlineType()->field_value_by_offset(offset, true);
 548           if (ft == T_NARROWOOP) {
 549             init_value = transform_later(new EncodePNode(init_value, init_value->bottom_type()->make_ptr()));
 550           }
 551         }
 552         return init_value;
 553       }
 554       assert(alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
 555       return _igvn.zerocon(ft);
 556     } else if (mem->is_Store()) {
 557       Node* n = mem->in(MemNode::ValueIn);
 558       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 559       n = bs->step_over_gc_barrier(n);
 560       return n;
 561     } else if (mem->is_Phi()) {
 562       // attempt to produce a Phi reflecting the values on the input paths of the Phi
 563       Node_Stack value_phis(8);
 564       Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
 565       if (phi != nullptr) {
 566         return phi;
 567       } else {
 568         // Kill all new Phis
 569         while(value_phis.is_nonempty()) {
 570           Node* n = value_phis.node();
 571           _igvn.replace_node(n, C->top());
 572           value_phis.pop();
 573         }
 574       }
 575     } else if (mem->is_ArrayCopy()) {
 576       Node* ctl = mem->in(0);
 577       Node* m = mem->in(TypeFunc::Memory);
 578       if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) {
 579         // pin the loads in the uncommon trap path
 580         ctl = sfpt_ctl;
 581         m = sfpt_mem;
 582       }
 583       return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
 584     }
 585   }
 586   // Something went wrong.
 587   return nullptr;
 588 }
 589 
 590 // Search the last value stored into the inline type's fields (for flat arrays).
 591 Node* PhaseMacroExpand::inline_type_from_mem(Node* mem, Node* ctl, ciInlineKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
 592   // Subtract the offset of the first field to account for the missing oop header
 593   offset -= vk->payload_offset();
 594   // Create a new InlineTypeNode and retrieve the field values from memory
 595   InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk);
 596   transform_later(vt);
 597   for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
 598     ciType* field_type = vt->field_type(i);
 599     int field_offset = offset + vt->field_offset(i);
 600     Node* value = nullptr;
 601     if (vt->field_is_flat(i)) {
 602       // TODO 8350865 Fix this
 603       // assert(vt->field_is_null_free(i), "Unexpected nullable flat field");
 604       if (!vt->field_is_null_free(i)) {
 605         return nullptr;
 606       }
 607       value = inline_type_from_mem(mem, ctl, field_type->as_inline_klass(), adr_type, field_offset, alloc);
 608     } else {
 609       const Type* ft = Type::get_const_type(field_type);
 610       BasicType bt = type2field[field_type->basic_type()];
 611       if (UseCompressedOops && !is_java_primitive(bt)) {
 612         ft = ft->make_narrowoop();
 613         bt = T_NARROWOOP;
 614       }
 615       // Each inline type field has its own memory slice
 616       adr_type = adr_type->with_field_offset(field_offset);
 617       value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
 618       if (value != nullptr && ft->isa_narrowoop()) {
 619         assert(UseCompressedOops, "unexpected narrow oop");
 620         if (value->is_EncodeP()) {
 621           value = value->in(1);
 622         } else if (!value->is_InlineType()) {
 623           value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
 624         }
 625       }
 626     }
 627     if (value != nullptr) {
 628       vt->set_field_value(i, value);
 629     } else {
 630       // We might have reached the TrackedInitializationLimit
 631       return nullptr;
 632     }
 633   }
 634   return vt;
 635 }
 636 
 637 // Check the possibility of scalar replacement.
 638 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
 639   //  Scan the uses of the allocation to check for anything that would
 640   //  prevent us from eliminating it.
 641   NOT_PRODUCT( const char* fail_eliminate = nullptr; )
 642   DEBUG_ONLY( Node* disq_node = nullptr; )
 643   bool can_eliminate = true;
 644   bool reduce_merge_precheck = (safepoints == nullptr);
 645 
 646   Unique_Node_List worklist;
 647   Node* res = alloc->result_cast();
 648   const TypeOopPtr* res_type = nullptr;
 649   if (res == nullptr) {
 650     // All users were eliminated.
 651   } else if (!res->is_CheckCastPP()) {
 652     NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
 653     can_eliminate = false;
 654   } else {
 655     worklist.push(res);
 656     res_type = igvn->type(res)->isa_oopptr();
 657     if (res_type == nullptr) {
 658       NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
 659       can_eliminate = false;
 660     } else if (!res_type->klass_is_exact()) {
 661       NOT_PRODUCT(fail_eliminate = "Not an exact type.";)
 662       can_eliminate = false;
 663     } else if (res_type->isa_aryptr()) {
 664       int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 665       if (length < 0) {
 666         NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
 667         can_eliminate = false;
 668       }
 669     }
 670   }
 671 
 672   while (can_eliminate && worklist.size() > 0) {
 673     BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
 674     res = worklist.pop();
 675     for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
 676       Node* use = res->fast_out(j);
 677 
 678       if (use->is_AddP()) {
 679         const TypePtr* addp_type = igvn->type(use)->is_ptr();
 680         int offset = addp_type->offset();
 681 
 682         if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
 683           NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
 684           can_eliminate = false;
 685           break;
 686         }
 687         for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
 688                                    k < kmax && can_eliminate; k++) {
 689           Node* n = use->fast_out(k);
 690           if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) {
 691             DEBUG_ONLY(disq_node = n;)
 692             if (n->is_Load() || n->is_LoadStore()) {
 693               NOT_PRODUCT(fail_eliminate = "Field load";)
 694             } else {
 695               NOT_PRODUCT(fail_eliminate = "Not store field reference";)

 701                  (use->as_ArrayCopy()->is_clonebasic() ||
 702                   use->as_ArrayCopy()->is_arraycopy_validated() ||
 703                   use->as_ArrayCopy()->is_copyof_validated() ||
 704                   use->as_ArrayCopy()->is_copyofrange_validated()) &&
 705                  use->in(ArrayCopyNode::Dest) == res) {
 706         // ok to eliminate
 707       } else if (use->is_SafePoint()) {
 708         SafePointNode* sfpt = use->as_SafePoint();
 709         if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
 710           // Object is passed as argument.
 711           DEBUG_ONLY(disq_node = use;)
 712           NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
 713           can_eliminate = false;
 714         }
 715         Node* sfptMem = sfpt->memory();
 716         if (sfptMem == nullptr || sfptMem->is_top()) {
 717           DEBUG_ONLY(disq_node = use;)
 718           NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
 719           can_eliminate = false;
 720         } else if (!reduce_merge_precheck) {
 721           assert(!res->is_Phi() || !res->as_Phi()->can_be_inline_type(), "Inline type allocations should not have safepoint uses");
 722           safepoints->append_if_missing(sfpt);
 723         }
 724       } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
 725         // Look at uses
 726         for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
 727           Node* u = use->fast_out(k);
 728           if (u->is_InlineType()) {
 729             // Use in flat field can be eliminated
 730             InlineTypeNode* vt = u->as_InlineType();
 731             for (uint i = 0; i < vt->field_count(); ++i) {
 732               if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
 733                 can_eliminate = false; // Use in non-flat field
 734                 break;
 735               }
 736             }
 737           } else {
 738             // Add other uses to the worklist to process individually
 739             worklist.push(use);
 740           }
 741         }
 742       } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
 743         // Store to mark word of inline type larval buffer
 744         assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
 745       } else if (res_type->is_inlinetypeptr() && (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore)) {
 746         // Inline type buffer allocations are followed by a membar
 747       } else if (reduce_merge_precheck &&
 748                  (use->is_Phi() || use->is_EncodeP() ||
 749                   use->Opcode() == Op_MemBarRelease ||
 750                   (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) {
 751         // Nothing to do
 752       } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
 753         if (use->is_Phi()) {
 754           if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
 755             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 756           } else {
 757             NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
 758           }
 759           DEBUG_ONLY(disq_node = use;)
 760         } else {
 761           if (use->Opcode() == Op_Return) {
 762             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 763           } else {
 764             NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
 765           }
 766           DEBUG_ONLY(disq_node = use;)
 767         }
 768         can_eliminate = false;
 769       } else {
 770         assert(use->Opcode() == Op_CastP2X, "should be");
 771         assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
 772       }
 773     }
 774   }
 775 
 776 #ifndef PRODUCT
 777   if (PrintEliminateAllocations && safepoints != nullptr) {
 778     if (can_eliminate) {
 779       tty->print("Scalar ");
 780       if (res == nullptr)
 781         alloc->dump();
 782       else
 783         res->dump();
 784     } else {
 785       tty->print("NotScalar (%s)", fail_eliminate);
 786       if (res == nullptr)
 787         alloc->dump();
 788       else
 789         res->dump();
 790 #ifdef ASSERT
 791       if (disq_node != nullptr) {
 792           tty->print("  >>>> ");
 793           disq_node->dump();
 794       }
 795 #endif /*ASSERT*/
 796     }
 797   }
 798 
 799   if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
 800     tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
 801     DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
 802   }
 803 #endif
 804   return can_eliminate;

 834     JVMState *jvms = sfpt_done->jvms();
 835     jvms->set_endoff(sfpt_done->req());
 836     // Now make a pass over the debug information replacing any references
 837     // to SafePointScalarObjectNode with the allocated object.
 838     int start = jvms->debug_start();
 839     int end   = jvms->debug_end();
 840     for (int i = start; i < end; i++) {
 841       if (sfpt_done->in(i)->is_SafePointScalarObject()) {
 842         SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
 843         if (scobj->first_index(jvms) == sfpt_done->req() &&
 844             scobj->n_fields() == (uint)nfields) {
 845           assert(scobj->alloc() == alloc, "sanity");
 846           sfpt_done->set_req(i, res);
 847         }
 848       }
 849     }
 850     _igvn._worklist.push(sfpt_done);
 851   }
 852 }
 853 
 854 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt,
 855                                                                                   Unique_Node_List* value_worklist) {
 856   // Fields of scalar objs are referenced only at the end
 857   // of regular debuginfo at the last (youngest) JVMS.
 858   // Record relative start index.
 859   ciInstanceKlass* iklass    = nullptr;
 860   BasicType basic_elem_type  = T_ILLEGAL;
 861   const Type* field_type     = nullptr;
 862   const TypeOopPtr* res_type = nullptr;
 863   int nfields                = 0;
 864   int array_base             = 0;
 865   int element_size           = 0;
 866   uint first_ind             = (sfpt->req() - sfpt->jvms()->scloff());
 867   Node* res                  = alloc->result_cast();
 868 
 869   assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
 870   assert(sfpt->jvms() != nullptr, "missed JVMS");
 871 
 872   if (res != nullptr) { // Could be null when there are no users
 873     res_type = _igvn.type(res)->isa_oopptr();
 874 
 875     if (res_type->isa_instptr()) {
 876       // find the fields of the class which will be needed for safepoint debug information
 877       iklass = res_type->is_instptr()->instance_klass();
 878       nfields = iklass->nof_nonstatic_fields();
 879     } else {
 880       // find the array's elements which will be needed for safepoint debug information
 881       nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 882       assert(nfields >= 0, "must be an array klass.");
 883       basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
 884       array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 885       element_size = type2aelembytes(basic_elem_type);
 886       field_type = res_type->is_aryptr()->elem();
 887       if (res_type->is_flat()) {
 888         // Flat inline type array
 889         element_size = res_type->is_aryptr()->flat_elem_size();
 890       }
 891     }
 892 
 893     if (res->bottom_type()->is_inlinetypeptr()) {
 894       // Nullable inline types have an IsInit field which is added to the safepoint when scalarizing them (see
 895       // InlineTypeNode::make_scalar_in_safepoint()). When having circular inline types, we stop scalarizing at depth 1
 896       // to avoid an endless recursion. Therefore, we do not have a SafePointScalarObjectNode node here, yet.
 897       // We are about to create a SafePointScalarObjectNode as if this is a normal object. Add an additional int input
 898       // with value 1 which sets IsInit to true to indicate that the object is always non-null. This input is checked
 899       // later in PhaseOutput::filLocArray() for inline types.
 900       sfpt->add_req(_igvn.intcon(1));
 901     }
 902   }
 903 
 904   SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields);
 905   sobj->init_req(0, C->root());
 906   transform_later(sobj);
 907 
 908   // Scan object's fields adding an input to the safepoint for each field.
 909   for (int j = 0; j < nfields; j++) {
 910     intptr_t offset;
 911     ciField* field = nullptr;
 912     if (iklass != nullptr) {
 913       field = iklass->nonstatic_field_at(j);
 914       offset = field->offset_in_bytes();
 915       ciType* elem_type = field->type();
 916       basic_elem_type = field->layout_type();
 917       assert(!field->is_flat(), "flat inline type fields should not have safepoint uses");
 918 
 919       ciField* flat_field = iklass->get_non_flat_field_by_offset(offset);
 920       if (flat_field != nullptr && flat_field->is_flat() && !flat_field->is_null_free()) {
 921         // TODO 8353432 Add support for nullable, flat fields in non-value class holders
 922         // Below code only iterates over the flat representation and therefore misses to
 923         // add null markers like we do in InlineTypeNode::add_fields_to_safepoint for value
 924         // class holders.
 925         return nullptr;
 926       }
 927 
 928       // The next code is taken from Parse::do_get_xxx().
 929       if (is_reference_type(basic_elem_type)) {
 930         if (!elem_type->is_loaded()) {
 931           field_type = TypeInstPtr::BOTTOM;
 932         } else if (field != nullptr && field->is_static_constant()) {
 933           ciObject* con = field->constant_value().as_object();
 934           // Do not "join" in the previous type; it doesn't add value,
 935           // and may yield a vacuous result if the field is of interface type.
 936           field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
 937           assert(field_type != nullptr, "field singleton type must be consistent");
 938         } else {
 939           field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
 940         }
 941         if (UseCompressedOops) {
 942           field_type = field_type->make_narrowoop();
 943           basic_elem_type = T_NARROWOOP;
 944         }
 945       } else {
 946         field_type = Type::get_const_basic_type(basic_elem_type);
 947       }
 948     } else {
 949       offset = array_base + j * (intptr_t)element_size;
 950     }
 951 
 952     Node* field_val = nullptr;
 953     const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
 954     if (res_type->is_flat()) {
 955       ciInlineKlass* inline_klass = res_type->is_aryptr()->elem()->inline_klass();
 956       assert(inline_klass->flat_in_array(), "must be flat in array");
 957       field_val = inline_type_from_mem(sfpt->memory(), sfpt->control(), inline_klass, field_addr_type->isa_aryptr(), 0, alloc);
 958     } else {
 959       field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
 960     }
 961 
 962     // We weren't able to find a value for this field,
 963     // give up on eliminating this allocation.
 964     if (field_val == nullptr) {
 965       uint last = sfpt->req() - 1;
 966       for (int k = 0;  k < j; k++) {
 967         sfpt->del_req(last--);
 968       }
 969       _igvn._worklist.push(sfpt);
 970 
 971 #ifndef PRODUCT
 972       if (PrintEliminateAllocations) {
 973         if (field != nullptr) {
 974           tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
 975           field->print();
 976           int field_idx = C->get_alias_index(field_addr_type);
 977           tty->print(" (alias_idx=%d)", field_idx);
 978         } else { // Array's element
 979           tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
 980         }
 981         tty->print(", which prevents elimination of: ");
 982         if (res == nullptr)
 983           alloc->dump();
 984         else
 985           res->dump();
 986       }
 987 #endif
 988 
 989       return nullptr;
 990     }
 991 
 992     if (UseCompressedOops && field_type->isa_narrowoop()) {
 993       // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
 994       // to be able scalar replace the allocation.
 995       if (field_val->is_EncodeP()) {
 996         field_val = field_val->in(1);
 997       } else if (!field_val->is_InlineType()) {
 998         field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
 999       }
1000     }
1001 
1002     // Keep track of inline types to scalarize them later
1003     if (field_val->is_InlineType()) {
1004       value_worklist->push(field_val);
1005     } else if (field_val->is_Phi()) {
1006       PhiNode* phi = field_val->as_Phi();
1007       // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
1008       // scalarize all its fields in safepoints.
1009       field_val = phi->try_push_inline_types_down(&_igvn, true);
1010       if (field_val->is_InlineType()) {
1011         value_worklist->push(field_val);
1012       }
1013     }
1014     sfpt->add_req(field_val);
1015   }
1016 
1017   sfpt->jvms()->set_endoff(sfpt->req());
1018 
1019   return sobj;
1020 }
1021 
1022 // Do scalar replacement.
1023 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
1024   GrowableArray <SafePointNode *> safepoints_done;
1025   Node* res = alloc->result_cast();
1026   assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
1027   const TypeOopPtr* res_type = nullptr;
1028   if (res != nullptr) { // Could be null when there are no users
1029     res_type = _igvn.type(res)->isa_oopptr();
1030   }
1031 
1032   // Process the safepoint uses
1033   assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr() || C->has_circular_inline_type(),
1034          "Inline type allocations should have been scalarized earlier");
1035   Unique_Node_List value_worklist;
1036   while (safepoints.length() > 0) {
1037     SafePointNode* sfpt = safepoints.pop();
1038     SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
1039 
1040     if (sobj == nullptr) {
1041       undo_previous_scalarizations(safepoints_done, alloc);
1042       return false;
1043     }
1044 
1045     // Now make a pass over the debug information replacing any references
1046     // to the allocated object with "sobj"
1047     JVMState *jvms = sfpt->jvms();
1048     sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1049     _igvn._worklist.push(sfpt);
1050 
1051     // keep it for rollback
1052     safepoints_done.append_if_missing(sfpt);
1053   }
1054   // Scalarize inline types that were added to the safepoint.
1055   // Don't allow linking a constant oop (if available) for flat array elements
1056   // because Deoptimization::reassign_flat_array_elements needs field values.
1057   bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1058   for (uint i = 0; i < value_worklist.size(); ++i) {
1059     InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1060     vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1061   }
1062   return true;
1063 }
1064 
1065 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1066   Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1067   Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1068   if (ctl_proj != nullptr) {
1069     igvn.replace_node(ctl_proj, n->in(0));
1070   }
1071   if (mem_proj != nullptr) {
1072     igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1073   }
1074 }
1075 
1076 // Process users of eliminated allocation.
1077 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1078   Unique_Node_List worklist;
1079   Node* res = alloc->result_cast();
1080   if (res != nullptr) {
1081     worklist.push(res);
1082   }
1083   while (worklist.size() > 0) {
1084     res = worklist.pop();
1085     for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1086       Node *use = res->last_out(j);
1087       uint oc1 = res->outcnt();
1088 
1089       if (use->is_AddP()) {
1090         for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1091           Node *n = use->last_out(k);
1092           uint oc2 = use->outcnt();
1093           if (n->is_Store()) {
1094             for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1095               MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1096               if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1097                 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1098                 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1099                 mb->remove(&_igvn);
1100               }



1101             }

1102             _igvn.replace_node(n, n->in(MemNode::Memory));
1103           } else {
1104             eliminate_gc_barrier(n);
1105           }
1106           k -= (oc2 - use->outcnt());
1107         }
1108         _igvn.remove_dead_node(use);
1109       } else if (use->is_ArrayCopy()) {
1110         // Disconnect ArrayCopy node
1111         ArrayCopyNode* ac = use->as_ArrayCopy();
1112         if (ac->is_clonebasic()) {
1113           Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1114           disconnect_projections(ac, _igvn);
1115           assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1116           Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1117           disconnect_projections(membar_before->as_MemBar(), _igvn);
1118           if (membar_after->is_MemBar()) {
1119             disconnect_projections(membar_after->as_MemBar(), _igvn);
1120           }
1121         } else {
1122           assert(ac->is_arraycopy_validated() ||
1123                  ac->is_copyof_validated() ||
1124                  ac->is_copyofrange_validated(), "unsupported");
1125           CallProjections* callprojs = ac->extract_projections(true);

1126 
1127           _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1128           _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1129           _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1130 
1131           // Set control to top. IGVN will remove the remaining projections
1132           ac->set_req(0, top());
1133           ac->replace_edge(res, top(), &_igvn);
1134 
1135           // Disconnect src right away: it can help find new
1136           // opportunities for allocation elimination
1137           Node* src = ac->in(ArrayCopyNode::Src);
1138           ac->replace_edge(src, top(), &_igvn);
1139           // src can be top at this point if src and dest of the
1140           // arraycopy were the same
1141           if (src->outcnt() == 0 && !src->is_top()) {
1142             _igvn.remove_dead_node(src);
1143           }
1144         }
1145         _igvn._worklist.push(ac);
1146       } else if (use->is_InlineType()) {
1147         assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1148         // Cut off oop input and remove known instance id from type
1149         _igvn.rehash_node_delayed(use);
1150         use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1151         const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1152         _igvn.set_type(use, toop);
1153         use->as_InlineType()->set_type(toop);
1154         // Process users
1155         for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1156           Node* u = use->fast_out(k);
1157           if (!u->is_InlineType()) {
1158             worklist.push(u);
1159           }
1160         }
1161       } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1162         // Store to mark word of inline type larval buffer
1163         assert(inline_alloc, "Unexpected store to mark word");
1164         _igvn.replace_node(use, use->in(MemNode::Memory));
1165       } else if (use->Opcode() == Op_MemBarRelease || use->Opcode() == Op_MemBarStoreStore) {
1166         // Inline type buffer allocations are followed by a membar
1167         assert(inline_alloc, "Unexpected MemBarRelease");
1168         use->as_MemBar()->remove(&_igvn);
1169       } else {
1170         eliminate_gc_barrier(use);
1171       }
1172       j -= (oc1 - res->outcnt());
1173     }
1174     assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1175     _igvn.remove_dead_node(res);
1176   }
1177 
1178   //
1179   // Process other users of allocation's projections
1180   //
1181   if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1182     // First disconnect stores captured by Initialize node.
1183     // If Initialize node is eliminated first in the following code,
1184     // it will kill such stores and DUIterator_Last will assert.
1185     for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax);  j < jmax; j++) {
1186       Node* use = _callprojs->resproj[0]->fast_out(j);
1187       if (use->is_AddP()) {
1188         // raw memory addresses used only by the initialization
1189         _igvn.replace_node(use, C->top());
1190         --j; --jmax;
1191       }
1192     }
1193     for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1194       Node* use = _callprojs->resproj[0]->last_out(j);
1195       uint oc1 = _callprojs->resproj[0]->outcnt();
1196       if (use->is_Initialize()) {
1197         // Eliminate Initialize node.
1198         InitializeNode *init = use->as_Initialize();
1199         assert(init->outcnt() <= 2, "only a control and memory projection expected");
1200         Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1201         if (ctrl_proj != nullptr) {
1202           _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1203 #ifdef ASSERT
1204           // If the InitializeNode has no memory out, it will die, and tmp will become null
1205           Node* tmp = init->in(TypeFunc::Control);
1206           assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1207 #endif
1208         }
1209         Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1210         if (mem_proj != nullptr) {
1211           Node *mem = init->in(TypeFunc::Memory);
1212 #ifdef ASSERT
1213           if (mem->is_MergeMem()) {
1214             assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1215           } else {
1216             assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1217           }
1218 #endif
1219           _igvn.replace_node(mem_proj, mem);
1220         }
1221       } else if (use->Opcode() == Op_MemBarStoreStore) {
1222         // Inline type buffer allocations are followed by a membar
1223         assert(inline_alloc, "Unexpected MemBarStoreStore");
1224         use->as_MemBar()->remove(&_igvn);
1225       } else  {
1226         assert(false, "only Initialize or AddP expected");
1227       }
1228       j -= (oc1 - _callprojs->resproj[0]->outcnt());
1229     }
1230   }
1231   if (_callprojs->fallthrough_catchproj != nullptr) {
1232     _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1233   }
1234   if (_callprojs->fallthrough_memproj != nullptr) {
1235     _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1236   }
1237   if (_callprojs->catchall_memproj != nullptr) {
1238     _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1239   }
1240   if (_callprojs->fallthrough_ioproj != nullptr) {
1241     _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1242   }
1243   if (_callprojs->catchall_ioproj != nullptr) {
1244     _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1245   }
1246   if (_callprojs->catchall_catchproj != nullptr) {
1247     _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1248   }
1249 }
1250 
1251 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1252   // If reallocation fails during deoptimization we'll pop all
1253   // interpreter frames for this compiled frame and that won't play
1254   // nice with JVMTI popframe.
1255   // We avoid this issue by eager reallocation when the popframe request
1256   // is received.
1257   if (!EliminateAllocations) {
1258     return false;
1259   }
1260   Node* klass = alloc->in(AllocateNode::KlassNode);
1261   const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1262 
1263   // Attempt to eliminate inline type buffer allocations
1264   // regardless of usage and escape/replaceable status.
1265   bool inline_alloc = tklass->isa_instklassptr() &&
1266                       tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1267   if (!alloc->_is_non_escaping && !inline_alloc) {
1268     return false;
1269   }
1270   // Eliminate boxing allocations which are not used
1271   // regardless scalar replaceable status.
1272   Node* res = alloc->result_cast();
1273   bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1274                       tklass->isa_instklassptr() &&
1275                       tklass->is_instklassptr()->instance_klass()->is_box_klass();
1276   if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1277     return false;
1278   }
1279 
1280   _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1281 
1282   GrowableArray <SafePointNode *> safepoints;
1283   if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1284     return false;
1285   }
1286 
1287   if (!alloc->_is_scalar_replaceable) {
1288     assert(res == nullptr || inline_alloc, "sanity");
1289     // We can only eliminate allocation if all debug info references
1290     // are already replaced with SafePointScalarObject because
1291     // we can't search for a fields value without instance_id.
1292     if (safepoints.length() > 0) {
1293       assert(!inline_alloc || C->has_circular_inline_type(),
1294              "Inline type allocations should have been scalarized earlier");
1295       return false;
1296     }
1297   }
1298 
1299   if (!scalar_replacement(alloc, safepoints)) {
1300     return false;
1301   }
1302 
1303   CompileLog* log = C->log();
1304   if (log != nullptr) {
1305     log->head("eliminate_allocation type='%d'",
1306               log->identify(tklass->exact_klass()));
1307     JVMState* p = alloc->jvms();
1308     while (p != nullptr) {
1309       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1310       p = p->caller();
1311     }
1312     log->tail("eliminate_allocation");
1313   }
1314 
1315   process_users_of_allocation(alloc, inline_alloc);
1316 
1317 #ifndef PRODUCT
1318   if (PrintEliminateAllocations) {
1319     if (alloc->is_AllocateArray())
1320       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1321     else
1322       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1323   }
1324 #endif
1325 
1326   return true;
1327 }
1328 
1329 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1330   // EA should remove all uses of non-escaping boxing node.
1331   if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1332     return false;
1333   }
1334 
1335   assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1336 
1337   _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1338 
1339   const TypeTuple* r = boxing->tf()->range_sig();
1340   assert(r->cnt() > TypeFunc::Parms, "sanity");
1341   const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1342   assert(t != nullptr, "sanity");
1343 
1344   CompileLog* log = C->log();
1345   if (log != nullptr) {
1346     log->head("eliminate_boxing type='%d'",
1347               log->identify(t->instance_klass()));
1348     JVMState* p = boxing->jvms();
1349     while (p != nullptr) {
1350       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1351       p = p->caller();
1352     }
1353     log->tail("eliminate_boxing");
1354   }
1355 
1356   process_users_of_allocation(boxing);
1357 
1358 #ifndef PRODUCT
1359   if (PrintEliminateAllocations) {

1423 // oop flavor.
1424 //
1425 //=============================================================================
1426 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1427 // Allocations bigger than this always go the slow route.
1428 // This value must be small enough that allocation attempts that need to
1429 // trigger exceptions go the slow route.  Also, it must be small enough so
1430 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1431 //=============================================================================j//
1432 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1433 // The allocator will coalesce int->oop copies away.  See comment in
1434 // coalesce.cpp about how this works.  It depends critically on the exact
1435 // code shape produced here, so if you are changing this code shape
1436 // make sure the GC info for the heap-top is correct in and around the
1437 // slow-path call.
1438 //
1439 
1440 void PhaseMacroExpand::expand_allocate_common(
1441             AllocateNode* alloc, // allocation node to be expanded
1442             Node* length,  // array length for an array allocation
1443             Node* init_val, // value to initialize the array with
1444             const TypeFunc* slow_call_type, // Type of slow call
1445             address slow_call_address,  // Address of slow call
1446             Node* valid_length_test // whether length is valid or not
1447     )
1448 {
1449   Node* ctrl = alloc->in(TypeFunc::Control);
1450   Node* mem  = alloc->in(TypeFunc::Memory);
1451   Node* i_o  = alloc->in(TypeFunc::I_O);
1452   Node* size_in_bytes     = alloc->in(AllocateNode::AllocSize);
1453   Node* klass_node        = alloc->in(AllocateNode::KlassNode);
1454   Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1455   assert(ctrl != nullptr, "must have control");
1456 
1457   // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1458   // they will not be used if "always_slow" is set
1459   enum { slow_result_path = 1, fast_result_path = 2 };
1460   Node *result_region = nullptr;
1461   Node *result_phi_rawmem = nullptr;
1462   Node *result_phi_rawoop = nullptr;
1463   Node *result_phi_i_o = nullptr;

1506 #endif
1507       yank_alloc_node(alloc);
1508       return;
1509     }
1510   }
1511 
1512   enum { too_big_or_final_path = 1, need_gc_path = 2 };
1513   Node *slow_region = nullptr;
1514   Node *toobig_false = ctrl;
1515 
1516   // generate the initial test if necessary
1517   if (initial_slow_test != nullptr ) {
1518     assert (expand_fast_path, "Only need test if there is a fast path");
1519     slow_region = new RegionNode(3);
1520 
1521     // Now make the initial failure test.  Usually a too-big test but
1522     // might be a TRUE for finalizers.
1523     IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1524     transform_later(toobig_iff);
1525     // Plug the failing-too-big test into the slow-path region
1526     Node* toobig_true = new IfTrueNode(toobig_iff);
1527     transform_later(toobig_true);
1528     slow_region    ->init_req( too_big_or_final_path, toobig_true );
1529     toobig_false = new IfFalseNode(toobig_iff);
1530     transform_later(toobig_false);
1531   } else {
1532     // No initial test, just fall into next case
1533     assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1534     toobig_false = ctrl;
1535     debug_only(slow_region = NodeSentinel);
1536   }
1537 
1538   // If we are here there are several possibilities
1539   // - expand_fast_path is false - then only a slow path is expanded. That's it.
1540   // no_initial_check means a constant allocation.
1541   // - If check always evaluates to false -> expand_fast_path is false (see above)
1542   // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1543   // if !allocation_has_use the fast path is empty
1544   // if !allocation_has_use && no_initial_check
1545   // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1546   //   removed by yank_alloc_node above.
1547 
1548   Node *slow_mem = mem;  // save the current memory state for slow path
1549   // generate the fast allocation code unless we know that the initial test will always go slow
1550   if (expand_fast_path) {
1551     // Fast path modifies only raw memory.
1552     if (mem->is_MergeMem()) {
1553       mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1554     }
1555 
1556     // allocate the Region and Phi nodes for the result
1557     result_region = new RegionNode(3);
1558     result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1559     result_phi_i_o    = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1560 
1561     // Grab regular I/O before optional prefetch may change it.
1562     // Slow-path does no I/O so just set it to the original I/O.
1563     result_phi_i_o->init_req(slow_result_path, i_o);
1564 
1565     // Name successful fast-path variables
1566     Node* fast_oop_ctrl;
1567     Node* fast_oop_rawmem;
1568 
1569     if (allocation_has_use) {
1570       Node* needgc_ctrl = nullptr;
1571       result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1572 
1573       intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1574       BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1575       Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1576                                         fast_oop_ctrl, fast_oop_rawmem,
1577                                         prefetch_lines);
1578 
1579       if (initial_slow_test != nullptr) {
1580         // This completes all paths into the slow merge point
1581         slow_region->init_req(need_gc_path, needgc_ctrl);
1582         transform_later(slow_region);
1583       } else {
1584         // No initial slow path needed!
1585         // Just fall from the need-GC path straight into the VM call.
1586         slow_region = needgc_ctrl;
1587       }
1588 

1606     result_phi_i_o   ->init_req(fast_result_path, i_o);
1607     result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1608   } else {
1609     slow_region = ctrl;
1610     result_phi_i_o = i_o; // Rename it to use in the following code.
1611   }
1612 
1613   // Generate slow-path call
1614   CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1615                                OptoRuntime::stub_name(slow_call_address),
1616                                TypePtr::BOTTOM);
1617   call->init_req(TypeFunc::Control,   slow_region);
1618   call->init_req(TypeFunc::I_O,       top());    // does no i/o
1619   call->init_req(TypeFunc::Memory,    slow_mem); // may gc ptrs
1620   call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1621   call->init_req(TypeFunc::FramePtr,  alloc->in(TypeFunc::FramePtr));
1622 
1623   call->init_req(TypeFunc::Parms+0, klass_node);
1624   if (length != nullptr) {
1625     call->init_req(TypeFunc::Parms+1, length);
1626     if (init_val != nullptr) {
1627       call->init_req(TypeFunc::Parms+2, init_val);
1628     }
1629   } else {
1630     // Let the runtime know if this is a larval allocation
1631     call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1632   }
1633 
1634   // Copy debug information and adjust JVMState information, then replace
1635   // allocate node with the call
1636   call->copy_call_debug_info(&_igvn, alloc);
1637   // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1638   // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1639   // path dies).
1640   if (valid_length_test != nullptr) {
1641     call->add_req(valid_length_test);
1642   }
1643   if (expand_fast_path) {
1644     call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1645   } else {
1646     // Hook i_o projection to avoid its elimination during allocation
1647     // replacement (when only a slow call is generated).
1648     call->set_req(TypeFunc::I_O, result_phi_i_o);
1649   }
1650   _igvn.replace_node(alloc, call);
1651   transform_later(call);
1652 
1653   // Identify the output projections from the allocate node and
1654   // adjust any references to them.
1655   // The control and io projections look like:
1656   //
1657   //        v---Proj(ctrl) <-----+   v---CatchProj(ctrl)
1658   //  Allocate                   Catch
1659   //        ^---Proj(io) <-------+   ^---CatchProj(io)
1660   //
1661   //  We are interested in the CatchProj nodes.
1662   //
1663   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1664 
1665   // An allocate node has separate memory projections for the uses on
1666   // the control and i_o paths. Replace the control memory projection with
1667   // result_phi_rawmem (unless we are only generating a slow call when
1668   // both memory projections are combined)
1669   if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1670     _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1671   }
1672   // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1673   // catchall_memproj so we end up with a call that has only 1 memory projection.
1674   if (_callprojs->catchall_memproj != nullptr) {
1675     if (_callprojs->fallthrough_memproj == nullptr) {
1676       _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1677       transform_later(_callprojs->fallthrough_memproj);
1678     }
1679     _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1680     _igvn.remove_dead_node(_callprojs->catchall_memproj);
1681   }
1682 
1683   // An allocate node has separate i_o projections for the uses on the control
1684   // and i_o paths. Always replace the control i_o projection with result i_o
1685   // otherwise incoming i_o become dead when only a slow call is generated
1686   // (it is different from memory projections where both projections are
1687   // combined in such case).
1688   if (_callprojs->fallthrough_ioproj != nullptr) {
1689     _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1690   }
1691   // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1692   // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1693   if (_callprojs->catchall_ioproj != nullptr) {
1694     if (_callprojs->fallthrough_ioproj == nullptr) {
1695       _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1696       transform_later(_callprojs->fallthrough_ioproj);
1697     }
1698     _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1699     _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1700   }
1701 
1702   // if we generated only a slow call, we are done
1703   if (!expand_fast_path) {
1704     // Now we can unhook i_o.
1705     if (result_phi_i_o->outcnt() > 1) {
1706       call->set_req(TypeFunc::I_O, top());
1707     } else {
1708       assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1709       // Case of new array with negative size known during compilation.
1710       // AllocateArrayNode::Ideal() optimization disconnect unreachable
1711       // following code since call to runtime will throw exception.
1712       // As result there will be no users of i_o after the call.
1713       // Leave i_o attached to this call to avoid problems in preceding graph.
1714     }
1715     return;
1716   }
1717 
1718   if (_callprojs->fallthrough_catchproj != nullptr) {
1719     ctrl = _callprojs->fallthrough_catchproj->clone();
1720     transform_later(ctrl);
1721     _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1722   } else {
1723     ctrl = top();
1724   }
1725   Node *slow_result;
1726   if (_callprojs->resproj[0] == nullptr) {
1727     // no uses of the allocation result
1728     slow_result = top();
1729   } else {
1730     slow_result = _callprojs->resproj[0]->clone();
1731     transform_later(slow_result);
1732     _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1733   }
1734 
1735   // Plug slow-path into result merge point
1736   result_region->init_req( slow_result_path, ctrl);
1737   transform_later(result_region);
1738   if (allocation_has_use) {
1739     result_phi_rawoop->init_req(slow_result_path, slow_result);
1740     transform_later(result_phi_rawoop);
1741   }
1742   result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1743   transform_later(result_phi_rawmem);
1744   transform_later(result_phi_i_o);
1745   // This completes all paths into the result merge point
1746 }
1747 
1748 // Remove alloc node that has no uses.
1749 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1750   Node* ctrl = alloc->in(TypeFunc::Control);
1751   Node* mem  = alloc->in(TypeFunc::Memory);
1752   Node* i_o  = alloc->in(TypeFunc::I_O);
1753 
1754   _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1755   if (_callprojs->resproj[0] != nullptr) {
1756     for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1757       Node* use = _callprojs->resproj[0]->fast_out(i);
1758       use->isa_MemBar()->remove(&_igvn);
1759       --imax;
1760       --i; // back up iterator
1761     }
1762     assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1763     _igvn.remove_dead_node(_callprojs->resproj[0]);
1764   }
1765   if (_callprojs->fallthrough_catchproj != nullptr) {
1766     _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1767     _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1768   }
1769   if (_callprojs->catchall_catchproj != nullptr) {
1770     _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1771     _callprojs->catchall_catchproj->set_req(0, top());
1772   }
1773   if (_callprojs->fallthrough_proj != nullptr) {
1774     Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1775     _igvn.remove_dead_node(catchnode);
1776     _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1777   }
1778   if (_callprojs->fallthrough_memproj != nullptr) {
1779     _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1780     _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1781   }
1782   if (_callprojs->fallthrough_ioproj != nullptr) {
1783     _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1784     _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1785   }
1786   if (_callprojs->catchall_memproj != nullptr) {
1787     _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1788     _callprojs->catchall_memproj->set_req(0, top());
1789   }
1790   if (_callprojs->catchall_ioproj != nullptr) {
1791     _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1792     _callprojs->catchall_ioproj->set_req(0, top());
1793   }
1794 #ifndef PRODUCT
1795   if (PrintEliminateAllocations) {
1796     if (alloc->is_AllocateArray()) {
1797       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1798     } else {
1799       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1800     }
1801   }
1802 #endif
1803   _igvn.remove_dead_node(alloc);
1804 }
1805 
1806 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1807                                                 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1808   // If initialization is performed by an array copy, any required
1809   // MemBarStoreStore was already added. If the object does not
1810   // escape no need for a MemBarStoreStore. If the object does not
1811   // escape in its initializer and memory barrier (MemBarStoreStore or
1812   // stronger) is already added at exit of initializer, also no need

1890     Node* thread = new ThreadLocalNode();
1891     transform_later(thread);
1892 
1893     call->init_req(TypeFunc::Parms + 0, thread);
1894     call->init_req(TypeFunc::Parms + 1, oop);
1895     call->init_req(TypeFunc::Control, ctrl);
1896     call->init_req(TypeFunc::I_O    , top()); // does no i/o
1897     call->init_req(TypeFunc::Memory , rawmem);
1898     call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1899     call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1900     transform_later(call);
1901     ctrl = new ProjNode(call, TypeFunc::Control);
1902     transform_later(ctrl);
1903     rawmem = new ProjNode(call, TypeFunc::Memory);
1904     transform_later(rawmem);
1905   }
1906 }
1907 
1908 // Helper for PhaseMacroExpand::expand_allocate_common.
1909 // Initializes the newly-allocated storage.
1910 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1911                                           Node* control, Node* rawmem, Node* object,
1912                                           Node* klass_node, Node* length,
1913                                           Node* size_in_bytes) {

1914   InitializeNode* init = alloc->initialization();
1915   // Store the klass & mark bits
1916   Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1917   if (!mark_node->is_Con()) {
1918     transform_later(mark_node);
1919   }
1920   rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1921 
1922   if (!UseCompactObjectHeaders) {
1923     rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1924   }
1925   int header_size = alloc->minimum_header_size();  // conservatively small
1926 
1927   // Array length
1928   if (length != nullptr) {         // Arrays need length field
1929     rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1930     // conservatively small header size:
1931     header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1932     if (_igvn.type(klass_node)->isa_aryklassptr()) {   // we know the exact header size in most cases:
1933       BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1934       if (is_reference_type(elem, true)) {
1935         elem = T_OBJECT;
1936       }
1937       header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1938     }
1939   }
1940 
1941   // Clear the object body, if necessary.
1942   if (init == nullptr) {
1943     // The init has somehow disappeared; be cautious and clear everything.
1944     //
1945     // This can happen if a node is allocated but an uncommon trap occurs
1946     // immediately.  In this case, the Initialize gets associated with the
1947     // trap, and may be placed in a different (outer) loop, if the Allocate
1948     // is in a loop.  If (this is rare) the inner loop gets unrolled, then
1949     // there can be two Allocates to one Initialize.  The answer in all these
1950     // edge cases is safety first.  It is always safe to clear immediately
1951     // within an Allocate, and then (maybe or maybe not) clear some more later.
1952     if (!(UseTLAB && ZeroTLAB)) {
1953       rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1954                                             alloc->in(AllocateNode::InitValue),
1955                                             alloc->in(AllocateNode::RawInitValue),
1956                                             header_size, size_in_bytes,
1957                                             &_igvn);
1958     }
1959   } else {
1960     if (!init->is_complete()) {
1961       // Try to win by zeroing only what the init does not store.
1962       // We can also try to do some peephole optimizations,
1963       // such as combining some adjacent subword stores.
1964       rawmem = init->complete_stores(control, rawmem, object,
1965                                      header_size, size_in_bytes, &_igvn);
1966     }
1967     // We have no more use for this link, since the AllocateNode goes away:
1968     init->set_req(InitializeNode::RawAddress, top());
1969     // (If we keep the link, it just confuses the register allocator,
1970     // who thinks he sees a real use of the address by the membar.)
1971   }
1972 
1973   return rawmem;
1974 }
1975 

2110       for ( intx i = 0; i < lines; i++ ) {
2111         prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2112                                             _igvn.MakeConX(distance) );
2113         transform_later(prefetch_adr);
2114         prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2115         // Do not let it float too high, since if eden_top == eden_end,
2116         // both might be null.
2117         if( i == 0 ) { // Set control for first prefetch, next follows it
2118           prefetch->init_req(0, needgc_false);
2119         }
2120         transform_later(prefetch);
2121         distance += step_size;
2122         i_o = prefetch;
2123       }
2124    }
2125    return i_o;
2126 }
2127 
2128 
2129 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2130   expand_allocate_common(alloc, nullptr, nullptr,
2131                          OptoRuntime::new_instance_Type(),
2132                          OptoRuntime::new_instance_Java(), nullptr);
2133 }
2134 
2135 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2136   Node* length = alloc->in(AllocateNode::ALength);
2137   Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
2138   InitializeNode* init = alloc->initialization();
2139   Node* klass_node = alloc->in(AllocateNode::KlassNode);
2140   Node* init_value = alloc->in(AllocateNode::InitValue);
2141   const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
2142   const TypeFunc* slow_call_type;
2143   address slow_call_address;  // Address of slow call
2144   if (init != nullptr && init->is_complete_with_arraycopy() &&
2145       ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) {
2146     // Don't zero type array during slow allocation in VM since
2147     // it will be initialized later by arraycopy in compiled code.
2148     slow_call_address = OptoRuntime::new_array_nozero_Java();
2149     slow_call_type = OptoRuntime::new_array_nozero_Type();
2150   } else {
2151     slow_call_address = OptoRuntime::new_array_Java();
2152     slow_call_type = OptoRuntime::new_array_Type();
2153 
2154     if (init_value == nullptr) {
2155       init_value = _igvn.zerocon(T_OBJECT);
2156     } else if (UseCompressedOops) {
2157       init_value = transform_later(new DecodeNNode(init_value, init_value->bottom_type()->make_ptr()));
2158     }
2159   }
2160   expand_allocate_common(alloc, length, init_value,
2161                          slow_call_type,
2162                          slow_call_address, valid_length_test);
2163 }
2164 
2165 //-------------------mark_eliminated_box----------------------------------
2166 //
2167 // During EA obj may point to several objects but after few ideal graph
2168 // transformations (CCP) it may point to only one non escaping object
2169 // (but still using phi), corresponding locks and unlocks will be marked
2170 // for elimination. Later obj could be replaced with a new node (new phi)
2171 // and which does not have escape information. And later after some graph
2172 // reshape other locks and unlocks (which were not marked for elimination
2173 // before) are connected to this new obj (phi) but they still will not be
2174 // marked for elimination since new obj has no escape information.
2175 // Mark all associated (same box and obj) lock and unlock nodes for
2176 // elimination if some of them marked already.
2177 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) {
2178   BoxLockNode* oldbox = box->as_BoxLock();
2179   if (oldbox->is_eliminated()) {
2180     return; // This BoxLock node was processed already.
2181   }

2353 #ifdef ASSERT
2354   if (!alock->is_coarsened()) {
2355     // Check that new "eliminated" BoxLock node is created.
2356     BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2357     assert(oldbox->is_eliminated(), "should be done already");
2358   }
2359 #endif
2360 
2361   alock->log_lock_optimization(C, "eliminate_lock");
2362 
2363 #ifndef PRODUCT
2364   if (PrintEliminateLocks) {
2365     tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2366   }
2367 #endif
2368 
2369   Node* mem  = alock->in(TypeFunc::Memory);
2370   Node* ctrl = alock->in(TypeFunc::Control);
2371   guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2372 
2373   _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2374   // There are 2 projections from the lock.  The lock node will
2375   // be deleted when its last use is subsumed below.
2376   assert(alock->outcnt() == 2 &&
2377          _callprojs->fallthrough_proj != nullptr &&
2378          _callprojs->fallthrough_memproj != nullptr,
2379          "Unexpected projections from Lock/Unlock");
2380 
2381   Node* fallthroughproj = _callprojs->fallthrough_proj;
2382   Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2383 
2384   // The memory projection from a lock/unlock is RawMem
2385   // The input to a Lock is merged memory, so extract its RawMem input
2386   // (unless the MergeMem has been optimized away.)
2387   if (alock->is_Lock()) {
2388     // Search for MemBarAcquireLock node and delete it also.
2389     MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2390     assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2391     Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2392     Node* memproj = membar->proj_out(TypeFunc::Memory);
2393     _igvn.replace_node(ctrlproj, fallthroughproj);
2394     _igvn.replace_node(memproj, memproj_fallthrough);
2395 
2396     // Delete FastLock node also if this Lock node is unique user
2397     // (a loop peeling may clone a Lock node).
2398     Node* flock = alock->as_Lock()->fastlock_node();
2399     if (flock->outcnt() == 1) {
2400       assert(flock->unique_out() == alock, "sanity");
2401       _igvn.replace_node(flock, top());
2402     }

2433   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2434 
2435   // Make the merge point
2436   Node *region;
2437   Node *mem_phi;
2438   Node *slow_path;
2439 
2440   region  = new RegionNode(3);
2441   // create a Phi for the memory state
2442   mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2443 
2444   // Optimize test; set region slot 2
2445   slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2446   mem_phi->init_req(2, mem);
2447 
2448   // Make slow path call
2449   CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(),
2450                                   OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2451                                   obj, box, nullptr);
2452 
2453   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2454 
2455   // Slow path can only throw asynchronous exceptions, which are always
2456   // de-opted.  So the compiler thinks the slow-call can never throw an
2457   // exception.  If it DOES throw an exception we would need the debug
2458   // info removed first (since if it throws there is no monitor).
2459   assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2460          _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2461 
2462   // Capture slow path
2463   // disconnect fall-through projection from call and create a new one
2464   // hook up users of fall-through projection to region
2465   Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2466   transform_later(slow_ctrl);
2467   _igvn.hash_delete(_callprojs->fallthrough_proj);
2468   _callprojs->fallthrough_proj->disconnect_inputs(C);
2469   region->init_req(1, slow_ctrl);
2470   // region inputs are now complete
2471   transform_later(region);
2472   _igvn.replace_node(_callprojs->fallthrough_proj, region);
2473 
2474   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2475 
2476   mem_phi->init_req(1, memproj);
2477 
2478   transform_later(mem_phi);
2479 
2480   _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2481 }
2482 
2483 //------------------------------expand_unlock_node----------------------
2484 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2485 
2486   Node* ctrl = unlock->in(TypeFunc::Control);
2487   Node* mem = unlock->in(TypeFunc::Memory);
2488   Node* obj = unlock->obj_node();
2489   Node* box = unlock->box_node();
2490 
2491   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2492 
2493   // No need for a null check on unlock
2494 
2495   // Make the merge point
2496   Node *region;
2497   Node *mem_phi;
2498 
2499   region  = new RegionNode(3);
2500   // create a Phi for the memory state
2501   mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2502 
2503   FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2504   funlock = transform_later( funlock )->as_FastUnlock();
2505   // Optimize test; set region slot 2
2506   Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2507   Node *thread = transform_later(new ThreadLocalNode());
2508 
2509   CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2510                                   CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2511                                   "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2512 
2513   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2514   assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2515          _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2516 
2517   // No exceptions for unlocking
2518   // Capture slow path
2519   // disconnect fall-through projection from call and create a new one
2520   // hook up users of fall-through projection to region
2521   Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2522   transform_later(slow_ctrl);
2523   _igvn.hash_delete(_callprojs->fallthrough_proj);
2524   _callprojs->fallthrough_proj->disconnect_inputs(C);
2525   region->init_req(1, slow_ctrl);
2526   // region inputs are now complete
2527   transform_later(region);
2528   _igvn.replace_node(_callprojs->fallthrough_proj, region);
2529 
2530   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2531   mem_phi->init_req(1, memproj );
2532   mem_phi->init_req(2, mem);
2533   transform_later(mem_phi);
2534 
2535   _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2536 }
2537 
2538 // An inline type might be returned from the call but we don't know its
2539 // type. Either we get a buffered inline type (and nothing needs to be done)
2540 // or one of the values being returned is the klass of the inline type
2541 // and we need to allocate an inline type instance of that type and
2542 // initialize it with other values being returned. In that case, we
2543 // first try a fast path allocation and initialize the value with the
2544 // inline klass's pack handler or we fall back to a runtime call.
2545 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2546   assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2547   Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2548   if (ret == nullptr) {
2549     return;
2550   }
2551   const TypeFunc* tf = call->_tf;
2552   const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2553   const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2554   call->_tf = new_tf;
2555   // Make sure the change of type is applied before projections are processed by igvn
2556   _igvn.set_type(call, call->Value(&_igvn));
2557   _igvn.set_type(ret, ret->Value(&_igvn));
2558 
2559   // Before any new projection is added:
2560   CallProjections* projs = call->extract_projections(true, true);
2561 
2562   // Create temporary hook nodes that will be replaced below.
2563   // Add an input to prevent hook nodes from being dead.
2564   Node* ctl = new Node(call);
2565   Node* mem = new Node(ctl);
2566   Node* io = new Node(ctl);
2567   Node* ex_ctl = new Node(ctl);
2568   Node* ex_mem = new Node(ctl);
2569   Node* ex_io = new Node(ctl);
2570   Node* res = new Node(ctl);
2571 
2572   // Allocate a new buffered inline type only if a new one is not returned
2573   Node* cast = transform_later(new CastP2XNode(ctl, res));
2574   Node* mask = MakeConX(0x1);
2575   Node* masked = transform_later(new AndXNode(cast, mask));
2576   Node* cmp = transform_later(new CmpXNode(masked, mask));
2577   Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2578   IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2579   transform_later(allocation_iff);
2580   Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2581   Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2582   Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2583 
2584   // Try to allocate a new buffered inline instance either from TLAB or eden space
2585   Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2586   CallLeafNoFPNode* handler_call;
2587   const bool alloc_in_place = UseTLAB;
2588   if (alloc_in_place) {
2589     Node* fast_oop_ctrl = nullptr;
2590     Node* fast_oop_rawmem = nullptr;
2591     Node* mask2 = MakeConX(-2);
2592     Node* masked2 = transform_later(new AndXNode(cast, mask2));
2593     Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2594     Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2595     Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2596     Node* size_in_bytes = ConvI2X(layout_val);
2597     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2598     Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2599                                       fast_oop_ctrl, fast_oop_rawmem,
2600                                       AllocateInstancePrefetchLines);
2601     // Allocation succeed, initialize buffered inline instance header firstly,
2602     // and then initialize its fields with an inline class specific handler
2603     Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2604     fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2605     fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2606     if (UseCompressedClassPointers) {
2607       fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2608     }
2609     Node* fixed_block  = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2610     Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2611     handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2612                                         nullptr,
2613                                         "pack handler",
2614                                         TypeRawPtr::BOTTOM);
2615     handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2616     handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2617     handler_call->init_req(TypeFunc::I_O, top());
2618     handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2619     handler_call->init_req(TypeFunc::ReturnAdr, top());
2620     handler_call->init_req(TypeFunc::Parms, pack_handler);
2621     handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2622   } else {
2623     needgc_ctrl = allocation_ctl;
2624   }
2625 
2626   // Allocation failed, fall back to a runtime call
2627   CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2628                                                          StubRoutines::store_inline_type_fields_to_buf(),
2629                                                          "store_inline_type_fields",
2630                                                          TypePtr::BOTTOM);
2631   slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2632   slow_call->init_req(TypeFunc::Memory, mem);
2633   slow_call->init_req(TypeFunc::I_O, io);
2634   slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2635   slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2636   slow_call->init_req(TypeFunc::Parms, res);
2637 
2638   Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2639   Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2640   Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2641   Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2642   Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2643   Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2644   Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci));
2645 
2646   Node* ex_r = new RegionNode(3);
2647   Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2648   Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2649   ex_r->init_req(1, slow_excp);
2650   ex_mem_phi->init_req(1, slow_mem);
2651   ex_io_phi->init_req(1, slow_io);
2652   ex_r->init_req(2, ex_ctl);
2653   ex_mem_phi->init_req(2, ex_mem);
2654   ex_io_phi->init_req(2, ex_io);
2655   transform_later(ex_r);
2656   transform_later(ex_mem_phi);
2657   transform_later(ex_io_phi);
2658 
2659   // We don't know how many values are returned. This assumes the
2660   // worst case, that all available registers are used.
2661   for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2662     if (domain->field_at(i) == Type::HALF) {
2663       slow_call->init_req(i, top());
2664       if (alloc_in_place) {
2665         handler_call->init_req(i+1, top());
2666       }
2667       continue;
2668     }
2669     Node* proj = transform_later(new ProjNode(call, i));
2670     slow_call->init_req(i, proj);
2671     if (alloc_in_place) {
2672       handler_call->init_req(i+1, proj);
2673     }
2674   }
2675   // We can safepoint at that new call
2676   slow_call->copy_call_debug_info(&_igvn, call);
2677   transform_later(slow_call);
2678   if (alloc_in_place) {
2679     transform_later(handler_call);
2680   }
2681 
2682   Node* fast_ctl = nullptr;
2683   Node* fast_res = nullptr;
2684   MergeMemNode* fast_mem = nullptr;
2685   if (alloc_in_place) {
2686     fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2687     Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2688     fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2689     fast_mem = MergeMemNode::make(mem);
2690     fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2691     transform_later(fast_mem);
2692   }
2693 
2694   Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2695   Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2696   Node* io_phi = new PhiNode(r, Type::ABIO);
2697   Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2698   r->init_req(1, no_allocation_ctl);
2699   mem_phi->init_req(1, mem);
2700   io_phi->init_req(1, io);
2701   res_phi->init_req(1, no_allocation_res);
2702   r->init_req(2, slow_norm);
2703   mem_phi->init_req(2, slow_mem);
2704   io_phi->init_req(2, slow_io);
2705   res_phi->init_req(2, slow_res);
2706   if (alloc_in_place) {
2707     r->init_req(3, fast_ctl);
2708     mem_phi->init_req(3, fast_mem);
2709     io_phi->init_req(3, io);
2710     res_phi->init_req(3, fast_res);
2711   }
2712   transform_later(r);
2713   transform_later(mem_phi);
2714   transform_later(io_phi);
2715   transform_later(res_phi);
2716 
2717   // Do not let stores that initialize this buffer be reordered with a subsequent
2718   // store that would make this buffer accessible by other threads.
2719   MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2720   transform_later(mb);
2721   mb->init_req(TypeFunc::Memory, mem_phi);
2722   mb->init_req(TypeFunc::Control, r);
2723   r = new ProjNode(mb, TypeFunc::Control);
2724   transform_later(r);
2725   mem_phi = new ProjNode(mb, TypeFunc::Memory);
2726   transform_later(mem_phi);
2727 
2728   assert(projs->nb_resproj == 1, "unexpected number of results");
2729   _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2730   _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2731   _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2732   _igvn.replace_in_uses(projs->resproj[0], res_phi);
2733   _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2734   _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2735   _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2736   // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2737   Node* cn = projs->fallthrough_catchproj->in(0);
2738   _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2739 
2740   _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2741   _igvn.replace_node(mem, projs->fallthrough_memproj);
2742   _igvn.replace_node(io, projs->fallthrough_ioproj);
2743   _igvn.replace_node(res, projs->resproj[0]);
2744   _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2745   _igvn.replace_node(ex_mem, projs->catchall_memproj);
2746   _igvn.replace_node(ex_io, projs->catchall_ioproj);
2747  }
2748 
2749 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2750   assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2751   Node* bol = check->unique_out();
2752   Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2753   Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2754   assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2755 
2756   for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2757     Node* iff = bol->last_out(i);
2758     assert(iff->is_If(), "where's the if?");
2759 
2760     if (iff->in(0)->is_top()) {
2761       _igvn.replace_input_of(iff, 1, C->top());
2762       continue;
2763     }
2764 
2765     Node* iftrue = iff->as_If()->proj_out(1);
2766     Node* iffalse = iff->as_If()->proj_out(0);
2767     Node* ctrl = iff->in(0);
2768 
2769     Node* subklass = nullptr;
2770     if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2771       subklass = obj_or_subklass;
2772     } else {
2773       Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2774       subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2775     }
2776 
2777     Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2778 
2779     _igvn.replace_input_of(iff, 0, C->top());
2780     _igvn.replace_node(iftrue, not_subtype_ctrl);
2781     _igvn.replace_node(iffalse, ctrl);
2782   }
2783   _igvn.replace_node(check, C->top());
2784 }
2785 
2786 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2787 //
2788 // long mark = array1.mark | array2.mark | ...;
2789 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2790 // if (locked_bit == 0) {
2791 //   // One array is locked, load prototype header from the klass
2792 //   mark = array1.klass.proto | array2.klass.proto | ...
2793 // }
2794 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2795 //    ...
2796 // }
2797 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2798   bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2799   if (array_inputs) {
2800     Node* mark = MakeConX(0);
2801     Node* locked_bit = MakeConX(markWord::unlocked_value);
2802     Node* mem = check->in(FlatArrayCheckNode::Memory);
2803     for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2804       Node* ary = check->in(i);
2805       const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2806       assert(t != nullptr, "Mixing array and klass inputs");
2807       assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2808       Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2809       Node* mark_load = _igvn.transform(LoadNode::make(_igvn, nullptr, mem, mark_adr, mark_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2810       mark = _igvn.transform(new OrXNode(mark, mark_load));
2811       locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2812     }
2813     assert(!mark->is_Con(), "Should have been optimized out");
2814     Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2815     Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2816 
2817     // BoolNode might be shared, replace each if user
2818     Node* old_bol = check->unique_out();
2819     assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2820     for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2821       IfNode* old_iff = old_bol->last_out(i)->as_If();
2822       Node* ctrl = old_iff->in(0);
2823       RegionNode* region = new RegionNode(3);
2824       Node* mark_phi = new PhiNode(region, TypeX_X);
2825 
2826       // Check if array is unlocked
2827       IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2828 
2829       // Unlocked: Use bits from mark word
2830       region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2831       mark_phi->init_req(1, mark);
2832 
2833       // Locked: Load prototype header from klass
2834       ctrl = _igvn.transform(new IfFalseNode(iff));
2835       Node* proto = MakeConX(0);
2836       for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2837         Node* ary = check->in(i);
2838         // Make loads control dependent to make sure they are only executed if array is locked
2839         Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2840         Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2841         Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2842         Node* proto_load = _igvn.transform(LoadNode::make(_igvn, ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2843         proto = _igvn.transform(new OrXNode(proto, proto_load));
2844       }
2845       region->init_req(2, ctrl);
2846       mark_phi->init_req(2, proto);
2847 
2848       // Check if flat array bits are set
2849       Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2850       Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2851       cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2852       Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2853 
2854       ctrl = _igvn.transform(region);
2855       iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2856       _igvn.replace_node(old_iff, iff);
2857     }
2858     _igvn.replace_node(check, C->top());
2859   } else {
2860     // Fall back to layout helper check
2861     Node* lhs = intcon(0);
2862     for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2863       Node* array_or_klass = check->in(i);
2864       Node* klass = nullptr;
2865       const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2866       assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2867       if (t->isa_oopptr() != nullptr) {
2868         Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2869         klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2870       } else {
2871         assert(t->isa_klassptr(), "Unexpected input type");
2872         klass = array_or_klass;
2873       }
2874       Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2875       Node* lh_val = _igvn.transform(LoadNode::make(_igvn, nullptr, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
2876       lhs = _igvn.transform(new OrINode(lhs, lh_val));
2877     }
2878     Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2879     Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2880     Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2881     Node* m2b = transform_later(new Conv2BNode(masked));
2882     // The matcher expects the input to If nodes to be produced by a Bool(CmpI..)
2883     // pattern, but the input to other potential users (e.g. Phi) to be some
2884     // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
2885     Node* old_bol = check->unique_out();
2886     for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2887       Node* user = old_bol->last_out(i);
2888       for (uint j = 0; j < user->req(); j++) {
2889         Node* n = user->in(j);
2890         if (n == old_bol) {
2891           _igvn.replace_input_of(user, j, user->is_If() ? bol : m2b);
2892         }
2893       }
2894     }
2895     _igvn.replace_node(check, C->top());
2896   }
2897 }
2898 
2899 //---------------------------eliminate_macro_nodes----------------------
2900 // Eliminate scalar replaced allocations and associated locks.
2901 void PhaseMacroExpand::eliminate_macro_nodes() {
2902   if (C->macro_count() == 0)
2903     return;
2904   NOT_PRODUCT(int membar_before = count_MemBar(C);)
2905 
2906   // Before elimination may re-mark (change to Nested or NonEscObj)
2907   // all associated (same box and obj) lock and unlock nodes.
2908   int cnt = C->macro_count();
2909   for (int i=0; i < cnt; i++) {
2910     Node *n = C->macro_node(i);
2911     if (n->is_AbstractLock()) { // Lock and Unlock nodes
2912       mark_eliminated_locking_nodes(n->as_AbstractLock());
2913     }
2914   }
2915   // Re-marking may break consistency of Coarsened locks.
2916   if (!C->coarsened_locks_consistent()) {
2917     return; // recompile without Coarsened locks if broken
2918   } else {

2944   }
2945   // Next, attempt to eliminate allocations
2946   _has_locks = false;
2947   progress = true;
2948   while (progress) {
2949     progress = false;
2950     for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2951       Node* n = C->macro_node(i - 1);
2952       bool success = false;
2953       DEBUG_ONLY(int old_macro_count = C->macro_count();)
2954       switch (n->class_id()) {
2955       case Node::Class_Allocate:
2956       case Node::Class_AllocateArray:
2957         success = eliminate_allocate_node(n->as_Allocate());
2958 #ifndef PRODUCT
2959         if (success && PrintOptoStatistics) {
2960           Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2961         }
2962 #endif
2963         break;
2964       case Node::Class_CallStaticJava: {
2965         CallStaticJavaNode* call = n->as_CallStaticJava();
2966         if (!call->method()->is_method_handle_intrinsic()) {
2967           success = eliminate_boxing_node(n->as_CallStaticJava());
2968         }
2969         break;
2970       }
2971       case Node::Class_Lock:
2972       case Node::Class_Unlock:
2973         assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2974         _has_locks = true;
2975         break;
2976       case Node::Class_ArrayCopy:
2977         break;
2978       case Node::Class_OuterStripMinedLoop:
2979         break;
2980       case Node::Class_SubTypeCheck:
2981         break;
2982       case Node::Class_Opaque1:
2983         break;
2984       case Node::Class_FlatArrayCheck:
2985         break;
2986       default:
2987         assert(n->Opcode() == Op_LoopLimit ||
2988                n->Opcode() == Op_ModD ||
2989                n->Opcode() == Op_ModF ||
2990                n->is_OpaqueNotNull()       ||
2991                n->is_OpaqueInitializedAssertionPredicate() ||
2992                n->Opcode() == Op_MaxL      ||
2993                n->Opcode() == Op_MinL      ||
2994                BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2995                "unknown node type in macro list");
2996       }
2997       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2998       progress = progress || success;
2999     }
3000   }
3001 #ifndef PRODUCT
3002   if (PrintOptoStatistics) {
3003     int membar_after = count_MemBar(C);
3004     Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
3005   }

3015     C->shuffle_macro_nodes();
3016   }
3017   // Last attempt to eliminate macro nodes.
3018   eliminate_macro_nodes();
3019   if (C->failing())  return true;
3020 
3021   // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
3022   bool progress = true;
3023   while (progress) {
3024     progress = false;
3025     for (int i = C->macro_count(); i > 0; i--) {
3026       Node* n = C->macro_node(i-1);
3027       bool success = false;
3028       DEBUG_ONLY(int old_macro_count = C->macro_count();)
3029       if (n->Opcode() == Op_LoopLimit) {
3030         // Remove it from macro list and put on IGVN worklist to optimize.
3031         C->remove_macro_node(n);
3032         _igvn._worklist.push(n);
3033         success = true;
3034       } else if (n->Opcode() == Op_CallStaticJava) {
3035         CallStaticJavaNode* call = n->as_CallStaticJava();
3036         if (!call->method()->is_method_handle_intrinsic()) {
3037           // Remove it from macro list and put on IGVN worklist to optimize.
3038           C->remove_macro_node(n);
3039           _igvn._worklist.push(n);
3040           success = true;
3041         }
3042       } else if (n->is_Opaque1()) {
3043         _igvn.replace_node(n, n->in(1));
3044         success = true;
3045       } else if (n->is_OpaqueNotNull()) {
3046         // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds,
3047         // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug),
3048         // we will execute a Halt node.
3049 #ifdef ASSERT
3050         _igvn.replace_node(n, n->in(1));
3051 #else
3052         _igvn.replace_node(n, _igvn.intcon(1));
3053 #endif
3054         success = true;
3055       } else if (n->is_OpaqueInitializedAssertionPredicate()) {
3056           // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product
3057           // builds as they are useless. In debug builds we keep them as additional verification code. Even though
3058           // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to
3059           // enable folding of dead control paths within which cast nodes become top after due to impossible types -
3060           // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now.
3061 #ifdef ASSERT

3123     // Worst case is a macro node gets expanded into about 200 nodes.
3124     // Allow 50% more for optimization.
3125     if (C->check_node_count(300, "out of nodes before macro expansion")) {
3126       return true;
3127     }
3128 
3129     DEBUG_ONLY(int old_macro_count = C->macro_count();)
3130     switch (n->class_id()) {
3131     case Node::Class_Lock:
3132       expand_lock_node(n->as_Lock());
3133       break;
3134     case Node::Class_Unlock:
3135       expand_unlock_node(n->as_Unlock());
3136       break;
3137     case Node::Class_ArrayCopy:
3138       expand_arraycopy_node(n->as_ArrayCopy());
3139       break;
3140     case Node::Class_SubTypeCheck:
3141       expand_subtypecheck_node(n->as_SubTypeCheck());
3142       break;
3143     case Node::Class_CallStaticJava:
3144       expand_mh_intrinsic_return(n->as_CallStaticJava());
3145       C->remove_macro_node(n);
3146       break;
3147     case Node::Class_FlatArrayCheck:
3148       expand_flatarraycheck_node(n->as_FlatArrayCheck());
3149       break;
3150     default:
3151       switch (n->Opcode()) {
3152       case Op_ModD:
3153       case Op_ModF: {
3154         bool is_drem = n->Opcode() == Op_ModD;
3155         CallNode* mod_macro = n->as_Call();
3156         CallNode* call = new CallLeafNode(mod_macro->tf(),
3157                                           is_drem ? CAST_FROM_FN_PTR(address, SharedRuntime::drem)
3158                                                   : CAST_FROM_FN_PTR(address, SharedRuntime::frem),
3159                                           is_drem ? "drem" : "frem", TypeRawPtr::BOTTOM);
3160         call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control));
3161         call->init_req(TypeFunc::I_O, mod_macro->in(TypeFunc::I_O));
3162         call->init_req(TypeFunc::Memory, mod_macro->in(TypeFunc::Memory));
3163         call->init_req(TypeFunc::ReturnAdr, mod_macro->in(TypeFunc::ReturnAdr));
3164         call->init_req(TypeFunc::FramePtr, mod_macro->in(TypeFunc::FramePtr));
3165         for (unsigned int i = 0; i < mod_macro->tf()->domain_cc()->cnt() - TypeFunc::Parms; i++) {
3166           call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i));
3167         }
3168         _igvn.replace_node(mod_macro, call);
3169         transform_later(call);
3170         break;
3171       }
3172       default:
3173         assert(false, "unknown node type in macro list");
3174       }
3175     }
3176     assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3177     if (C->failing())  return true;
3178     C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n);
3179 
3180     // Clean up the graph so we're less likely to hit the maximum node
3181     // limit
3182     _igvn.set_delay_transform(false);
3183     _igvn.optimize();
3184     if (C->failing())  return true;
3185     _igvn.set_delay_transform(true);
< prev index next >