1 /*
   2  * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "compiler/compileLog.hpp"
  27 #include "gc/shared/collectedHeap.inline.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/locknode.hpp"
  39 #include "opto/loopnode.hpp"
  40 #include "opto/macro.hpp"
  41 #include "opto/memnode.hpp"
  42 #include "opto/narrowptrnode.hpp"
  43 #include "opto/node.hpp"
  44 #include "opto/opaquenode.hpp"
  45 #include "opto/phaseX.hpp"
  46 #include "opto/rootnode.hpp"
  47 #include "opto/runtime.hpp"
  48 #include "opto/subnode.hpp"
  49 #include "opto/type.hpp"
  50 #include "runtime/sharedRuntime.hpp"
  51 #include "utilities/macros.hpp"
  52 #if INCLUDE_G1GC
  53 #include "gc/g1/g1ThreadLocalData.hpp"
  54 #endif // INCLUDE_G1GC
  55 #if INCLUDE_SHENANDOAHGC
  56 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  57 #endif
  58 
  59 
  60 //
  61 // Replace any references to "oldref" in inputs to "use" with "newref".
  62 // Returns the number of replacements made.
  63 //
  64 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
  65   int nreplacements = 0;
  66   uint req = use->req();
  67   for (uint j = 0; j < use->len(); j++) {
  68     Node *uin = use->in(j);
  69     if (uin == oldref) {
  70       if (j < req)
  71         use->set_req(j, newref);
  72       else
  73         use->set_prec(j, newref);
  74       nreplacements++;
  75     } else if (j >= req && uin == NULL) {
  76       break;
  77     }
  78   }
  79   return nreplacements;
  80 }
  81 
  82 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
  83   // Copy debug information and adjust JVMState information
  84   uint old_dbg_start = oldcall->tf()->domain()->cnt();
  85   uint new_dbg_start = newcall->tf()->domain()->cnt();
  86   int jvms_adj  = new_dbg_start - old_dbg_start;
  87   assert (new_dbg_start == newcall->req(), "argument count mismatch");
  88 
  89   // SafePointScalarObject node could be referenced several times in debug info.
  90   // Use Dict to record cloned nodes.
  91   Dict* sosn_map = new Dict(cmpkey,hashkey);
  92   for (uint i = old_dbg_start; i < oldcall->req(); i++) {
  93     Node* old_in = oldcall->in(i);
  94     // Clone old SafePointScalarObjectNodes, adjusting their field contents.
  95     if (old_in != NULL && old_in->is_SafePointScalarObject()) {
  96       SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
  97       uint old_unique = C->unique();
  98       Node* new_in = old_sosn->clone(sosn_map);
  99       if (old_unique != C->unique()) { // New node?
 100         new_in->set_req(0, C->root()); // reset control edge
 101         new_in = transform_later(new_in); // Register new node.
 102       }
 103       old_in = new_in;
 104     }
 105     newcall->add_req(old_in);
 106   }
 107 
 108   // JVMS may be shared so clone it before we modify it
 109   newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
 110   for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
 111     jvms->set_map(newcall);
 112     jvms->set_locoff(jvms->locoff()+jvms_adj);
 113     jvms->set_stkoff(jvms->stkoff()+jvms_adj);
 114     jvms->set_monoff(jvms->monoff()+jvms_adj);
 115     jvms->set_scloff(jvms->scloff()+jvms_adj);
 116     jvms->set_endoff(jvms->endoff()+jvms_adj);
 117   }
 118 }
 119 
 120 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
 121   Node* cmp;
 122   if (mask != 0) {
 123     Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
 124     cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
 125   } else {
 126     cmp = word;
 127   }
 128   Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
 129   IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
 130   transform_later(iff);
 131 
 132   // Fast path taken.
 133   Node *fast_taken = transform_later(new IfFalseNode(iff));
 134 
 135   // Fast path not-taken, i.e. slow path
 136   Node *slow_taken = transform_later(new IfTrueNode(iff));
 137 
 138   if (return_fast_path) {
 139     region->init_req(edge, slow_taken); // Capture slow-control
 140     return fast_taken;
 141   } else {
 142     region->init_req(edge, fast_taken); // Capture fast-control
 143     return slow_taken;
 144   }
 145 }
 146 
 147 //--------------------copy_predefined_input_for_runtime_call--------------------
 148 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
 149   // Set fixed predefined input arguments
 150   call->init_req( TypeFunc::Control, ctrl );
 151   call->init_req( TypeFunc::I_O    , oldcall->in( TypeFunc::I_O) );
 152   call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
 153   call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
 154   call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
 155 }
 156 
 157 //------------------------------make_slow_call---------------------------------
 158 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
 159                                            address slow_call, const char* leaf_name, Node* slow_path,
 160                                            Node* parm0, Node* parm1, Node* parm2) {
 161 
 162   // Slow-path call
 163  CallNode *call = leaf_name
 164    ? (CallNode*)new CallLeafNode      ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
 165    : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
 166 
 167   // Slow path call has no side-effects, uses few values
 168   copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
 169   if (parm0 != NULL)  call->init_req(TypeFunc::Parms+0, parm0);
 170   if (parm1 != NULL)  call->init_req(TypeFunc::Parms+1, parm1);
 171   if (parm2 != NULL)  call->init_req(TypeFunc::Parms+2, parm2);
 172   copy_call_debug_info(oldcall, call);
 173   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
 174   _igvn.replace_node(oldcall, call);
 175   transform_later(call);
 176 
 177   return call;
 178 }
 179 
 180 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
 181   _fallthroughproj = NULL;
 182   _fallthroughcatchproj = NULL;
 183   _ioproj_fallthrough = NULL;
 184   _ioproj_catchall = NULL;
 185   _catchallcatchproj = NULL;
 186   _memproj_fallthrough = NULL;
 187   _memproj_catchall = NULL;
 188   _resproj = NULL;
 189   for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
 190     ProjNode *pn = call->fast_out(i)->as_Proj();
 191     switch (pn->_con) {
 192       case TypeFunc::Control:
 193       {
 194         // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
 195         _fallthroughproj = pn;
 196         DUIterator_Fast jmax, j = pn->fast_outs(jmax);
 197         const Node *cn = pn->fast_out(j);
 198         if (cn->is_Catch()) {
 199           ProjNode *cpn = NULL;
 200           for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
 201             cpn = cn->fast_out(k)->as_Proj();
 202             assert(cpn->is_CatchProj(), "must be a CatchProjNode");
 203             if (cpn->_con == CatchProjNode::fall_through_index)
 204               _fallthroughcatchproj = cpn;
 205             else {
 206               assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
 207               _catchallcatchproj = cpn;
 208             }
 209           }
 210         }
 211         break;
 212       }
 213       case TypeFunc::I_O:
 214         if (pn->_is_io_use)
 215           _ioproj_catchall = pn;
 216         else
 217           _ioproj_fallthrough = pn;
 218         break;
 219       case TypeFunc::Memory:
 220         if (pn->_is_io_use)
 221           _memproj_catchall = pn;
 222         else
 223           _memproj_fallthrough = pn;
 224         break;
 225       case TypeFunc::Parms:
 226         _resproj = pn;
 227         break;
 228       default:
 229         assert(false, "unexpected projection from allocation node.");
 230     }
 231   }
 232 
 233 }
 234 
 235 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
 236   BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
 237   bs->eliminate_gc_barrier(this, p2x);
 238 }
 239 
 240 // Search for a memory operation for the specified memory slice.
 241 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
 242   Node *orig_mem = mem;
 243   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 244   const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
 245   while (true) {
 246     if (mem == alloc_mem || mem == start_mem ) {
 247       return mem;  // hit one of our sentinels
 248     } else if (mem->is_MergeMem()) {
 249       mem = mem->as_MergeMem()->memory_at(alias_idx);
 250     } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
 251       Node *in = mem->in(0);
 252       // we can safely skip over safepoints, calls, locks and membars because we
 253       // already know that the object is safe to eliminate.
 254       if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
 255         return in;
 256       } else if (in->is_Call()) {
 257         CallNode *call = in->as_Call();
 258         if (call->may_modify(tinst, phase)) {
 259           assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
 260           if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
 261             return in;
 262           }
 263         }
 264         mem = in->in(TypeFunc::Memory);
 265       } else if (in->is_MemBar()) {
 266         ArrayCopyNode* ac = NULL;
 267         if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
 268           assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
 269           return ac;
 270         }
 271         mem = in->in(TypeFunc::Memory);
 272       } else {
 273         assert(false, "unexpected projection");
 274       }
 275     } else if (mem->is_Store()) {
 276       const TypePtr* atype = mem->as_Store()->adr_type();
 277       int adr_idx = phase->C->get_alias_index(atype);
 278       if (adr_idx == alias_idx) {
 279         assert(atype->isa_oopptr(), "address type must be oopptr");
 280         int adr_offset = atype->offset();
 281         uint adr_iid = atype->is_oopptr()->instance_id();
 282         // Array elements references have the same alias_idx
 283         // but different offset and different instance_id.
 284         if (adr_offset == offset && adr_iid == alloc->_idx)
 285           return mem;
 286       } else {
 287         assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
 288       }
 289       mem = mem->in(MemNode::Memory);
 290     } else if (mem->is_ClearArray()) {
 291       if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
 292         // Can not bypass initialization of the instance
 293         // we are looking.
 294         debug_only(intptr_t offset;)
 295         assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
 296         InitializeNode* init = alloc->as_Allocate()->initialization();
 297         // We are looking for stored value, return Initialize node
 298         // or memory edge from Allocate node.
 299         if (init != NULL)
 300           return init;
 301         else
 302           return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
 303       }
 304       // Otherwise skip it (the call updated 'mem' value).
 305     } else if (mem->Opcode() == Op_SCMemProj) {
 306       mem = mem->in(0);
 307       Node* adr = NULL;
 308       if (mem->is_LoadStore()) {
 309         adr = mem->in(MemNode::Address);
 310       } else {
 311         assert(mem->Opcode() == Op_EncodeISOArray ||
 312                mem->Opcode() == Op_StrCompressedCopy, "sanity");
 313         adr = mem->in(3); // Destination array
 314       }
 315       const TypePtr* atype = adr->bottom_type()->is_ptr();
 316       int adr_idx = phase->C->get_alias_index(atype);
 317       if (adr_idx == alias_idx) {
 318         DEBUG_ONLY(mem->dump();)
 319         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 320         return NULL;
 321       }
 322       mem = mem->in(MemNode::Memory);
 323    } else if (mem->Opcode() == Op_StrInflatedCopy) {
 324       Node* adr = mem->in(3); // Destination array
 325       const TypePtr* atype = adr->bottom_type()->is_ptr();
 326       int adr_idx = phase->C->get_alias_index(atype);
 327       if (adr_idx == alias_idx) {
 328         DEBUG_ONLY(mem->dump();)
 329         assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
 330         return NULL;
 331       }
 332       mem = mem->in(MemNode::Memory);
 333     } else {
 334       return mem;
 335     }
 336     assert(mem != orig_mem, "dead memory loop");
 337   }
 338 }
 339 
 340 // Generate loads from source of the arraycopy for fields of
 341 // destination needed at a deoptimization point
 342 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
 343   BasicType bt = ft;
 344   const Type *type = ftype;
 345   if (ft == T_NARROWOOP) {
 346     bt = T_OBJECT;
 347     type = ftype->make_oopptr();
 348   }
 349   Node* res = NULL;
 350   if (ac->is_clonebasic()) {
 351     Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
 352     Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
 353     const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
 354     res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
 355   } else {
 356     if (ac->modifies(offset, offset, &_igvn, true)) {
 357       assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
 358       uint shift  = exact_log2(type2aelembytes(bt));
 359       Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 360 #ifdef _LP64
 361       diff = _igvn.transform(new ConvI2LNode(diff));
 362 #endif
 363       diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
 364 
 365       Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
 366       Node* base = ac->in(ArrayCopyNode::Src);
 367       Node* adr = _igvn.transform(new AddPNode(base, base, off));
 368       const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
 369       res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
 370     }
 371   }
 372   if (res != NULL) {
 373     res = _igvn.transform(res);
 374     if (ftype->isa_narrowoop()) {
 375       // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
 376       res = _igvn.transform(new EncodePNode(res, ftype));
 377     }
 378     return res;
 379   }
 380   return NULL;
 381 }
 382 
 383 //
 384 // Given a Memory Phi, compute a value Phi containing the values from stores
 385 // on the input paths.
 386 // Note: this function is recursive, its depth is limited by the "level" argument
 387 // Returns the computed Phi, or NULL if it cannot compute it.
 388 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) {
 389   assert(mem->is_Phi(), "sanity");
 390   int alias_idx = C->get_alias_index(adr_t);
 391   int offset = adr_t->offset();
 392   int instance_id = adr_t->instance_id();
 393 
 394   // Check if an appropriate value phi already exists.
 395   Node* region = mem->in(0);
 396   for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
 397     Node* phi = region->fast_out(k);
 398     if (phi->is_Phi() && phi != mem &&
 399         phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
 400       return phi;
 401     }
 402   }
 403   // Check if an appropriate new value phi already exists.
 404   Node* new_phi = value_phis->find(mem->_idx);
 405   if (new_phi != NULL)
 406     return new_phi;
 407 
 408   if (level <= 0) {
 409     return NULL; // Give up: phi tree too deep
 410   }
 411   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 412   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 413 
 414   uint length = mem->req();
 415   GrowableArray <Node *> values(length, length, NULL, false);
 416 
 417   // create a new Phi for the value
 418   PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
 419   transform_later(phi);
 420   value_phis->push(phi, mem->_idx);
 421 
 422   for (uint j = 1; j < length; j++) {
 423     Node *in = mem->in(j);
 424     if (in == NULL || in->is_top()) {
 425       values.at_put(j, in);
 426     } else  {
 427       Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
 428       if (val == start_mem || val == alloc_mem) {
 429         // hit a sentinel, return appropriate 0 value
 430         values.at_put(j, _igvn.zerocon(ft));
 431         continue;
 432       }
 433       if (val->is_Initialize()) {
 434         val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 435       }
 436       if (val == NULL) {
 437         return NULL;  // can't find a value on this path
 438       }
 439       if (val == mem) {
 440         values.at_put(j, mem);
 441       } else if (val->is_Store()) {
 442         Node* n = val->in(MemNode::ValueIn);
 443         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 444         n = bs->step_over_gc_barrier(n);
 445         values.at_put(j, n);
 446       } else if(val->is_Proj() && val->in(0) == alloc) {
 447         values.at_put(j, _igvn.zerocon(ft));
 448       } else if (val->is_Phi()) {
 449         val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
 450         if (val == NULL) {
 451           return NULL;
 452         }
 453         values.at_put(j, val);
 454       } else if (val->Opcode() == Op_SCMemProj) {
 455         assert(val->in(0)->is_LoadStore() ||
 456                val->in(0)->Opcode() == Op_EncodeISOArray ||
 457                val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
 458         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 459         return NULL;
 460       } else if (val->is_ArrayCopy()) {
 461         Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
 462         if (res == NULL) {
 463           return NULL;
 464         }
 465         values.at_put(j, res);
 466       } else {
 467 #ifdef ASSERT
 468         val->dump();
 469         assert(false, "unknown node on this path");
 470 #endif
 471         return NULL;  // unknown node on this path
 472       }
 473     }
 474   }
 475   // Set Phi's inputs
 476   for (uint j = 1; j < length; j++) {
 477     if (values.at(j) == mem) {
 478       phi->init_req(j, phi);
 479     } else {
 480       phi->init_req(j, values.at(j));
 481     }
 482   }
 483   return phi;
 484 }
 485 
 486 // Search the last value stored into the object's field.
 487 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
 488   assert(adr_t->is_known_instance_field(), "instance required");
 489   int instance_id = adr_t->instance_id();
 490   assert((uint)instance_id == alloc->_idx, "wrong allocation");
 491 
 492   int alias_idx = C->get_alias_index(adr_t);
 493   int offset = adr_t->offset();
 494   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 495   Node *alloc_ctrl = alloc->in(TypeFunc::Control);
 496   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 497   Arena *a = Thread::current()->resource_area();
 498   VectorSet visited(a);
 499 
 500 
 501   bool done = sfpt_mem == alloc_mem;
 502   Node *mem = sfpt_mem;
 503   while (!done) {
 504     if (visited.test_set(mem->_idx)) {
 505       return NULL;  // found a loop, give up
 506     }
 507     mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
 508     if (mem == start_mem || mem == alloc_mem) {
 509       done = true;  // hit a sentinel, return appropriate 0 value
 510     } else if (mem->is_Initialize()) {
 511       mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 512       if (mem == NULL) {
 513         done = true; // Something go wrong.
 514       } else if (mem->is_Store()) {
 515         const TypePtr* atype = mem->as_Store()->adr_type();
 516         assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
 517         done = true;
 518       }
 519     } else if (mem->is_Store()) {
 520       const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
 521       assert(atype != NULL, "address type must be oopptr");
 522       assert(C->get_alias_index(atype) == alias_idx &&
 523              atype->is_known_instance_field() && atype->offset() == offset &&
 524              atype->instance_id() == instance_id, "store is correct memory slice");
 525       done = true;
 526     } else if (mem->is_Phi()) {
 527       // try to find a phi's unique input
 528       Node *unique_input = NULL;
 529       Node *top = C->top();
 530       for (uint i = 1; i < mem->req(); i++) {
 531         Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
 532         if (n == NULL || n == top || n == mem) {
 533           continue;
 534         } else if (unique_input == NULL) {
 535           unique_input = n;
 536         } else if (unique_input != n) {
 537           unique_input = top;
 538           break;
 539         }
 540       }
 541       if (unique_input != NULL && unique_input != top) {
 542         mem = unique_input;
 543       } else {
 544         done = true;
 545       }
 546     } else if (mem->is_ArrayCopy()) {
 547       done = true;
 548     } else {
 549       assert(false, "unexpected node");
 550     }
 551   }
 552   if (mem != NULL) {
 553     if (mem == start_mem || mem == alloc_mem) {
 554       // hit a sentinel, return appropriate 0 value
 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(a, 8);
 564       Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
 565       if (phi != NULL) {
 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(Deoptimization::Reason_none)) {
 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 go wrong.
 587   return NULL;
 588 }
 589 
 590 // Check the possibility of scalar replacement.
 591 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
 592   //  Scan the uses of the allocation to check for anything that would
 593   //  prevent us from eliminating it.
 594   NOT_PRODUCT( const char* fail_eliminate = NULL; )
 595   DEBUG_ONLY( Node* disq_node = NULL; )
 596   bool  can_eliminate = true;
 597 
 598   Node* res = alloc->result_cast();
 599   const TypeOopPtr* res_type = NULL;
 600   if (res == NULL) {
 601     // All users were eliminated.
 602   } else if (!res->is_CheckCastPP()) {
 603     NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
 604     can_eliminate = false;
 605   } else {
 606     res_type = _igvn.type(res)->isa_oopptr();
 607     if (res_type == NULL) {
 608       NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
 609       can_eliminate = false;
 610     } else if (res_type->isa_aryptr()) {
 611       int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 612       if (length < 0) {
 613         NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
 614         can_eliminate = false;
 615       }
 616     }
 617   }
 618 
 619   if (can_eliminate && res != NULL) {
 620     for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
 621                                j < jmax && can_eliminate; j++) {
 622       Node* use = res->fast_out(j);
 623 
 624       if (use->is_AddP()) {
 625         const TypePtr* addp_type = _igvn.type(use)->is_ptr();
 626         int offset = addp_type->offset();
 627 
 628         if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
 629           NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
 630           can_eliminate = false;
 631           break;
 632         }
 633         for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
 634                                    k < kmax && can_eliminate; k++) {
 635           Node* n = use->fast_out(k);
 636           if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
 637               SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
 638               !(n->is_ArrayCopy() &&
 639                 n->as_ArrayCopy()->is_clonebasic() &&
 640                 n->in(ArrayCopyNode::Dest) == use)) {
 641             DEBUG_ONLY(disq_node = n;)
 642             if (n->is_Load() || n->is_LoadStore()) {
 643               NOT_PRODUCT(fail_eliminate = "Field load";)
 644             } else {
 645               NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
 646             }
 647             can_eliminate = false;
 648           }
 649         }
 650       } else if (use->is_ArrayCopy() &&
 651                  (use->as_ArrayCopy()->is_arraycopy_validated() ||
 652                   use->as_ArrayCopy()->is_copyof_validated() ||
 653                   use->as_ArrayCopy()->is_copyofrange_validated()) &&
 654                  use->in(ArrayCopyNode::Dest) == res) {
 655         // ok to eliminate
 656       } else if (use->is_SafePoint()) {
 657         SafePointNode* sfpt = use->as_SafePoint();
 658         if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
 659           // Object is passed as argument.
 660           DEBUG_ONLY(disq_node = use;)
 661           NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
 662           can_eliminate = false;
 663         }
 664         Node* sfptMem = sfpt->memory();
 665         if (sfptMem == NULL || sfptMem->is_top()) {
 666           DEBUG_ONLY(disq_node = use;)
 667           NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
 668           can_eliminate = false;
 669         } else {
 670           safepoints.append_if_missing(sfpt);
 671         }
 672       } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
 673         if (use->is_Phi()) {
 674           if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
 675             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 676           } else {
 677             NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
 678           }
 679           DEBUG_ONLY(disq_node = use;)
 680         } else {
 681           if (use->Opcode() == Op_Return) {
 682             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 683           }else {
 684             NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
 685           }
 686           DEBUG_ONLY(disq_node = use;)
 687         }
 688         can_eliminate = false;
 689       }
 690     }
 691   }
 692 
 693 #ifndef PRODUCT
 694   if (PrintEliminateAllocations) {
 695     if (can_eliminate) {
 696       tty->print("Scalar ");
 697       if (res == NULL)
 698         alloc->dump();
 699       else
 700         res->dump();
 701     } else if (alloc->_is_scalar_replaceable) {
 702       tty->print("NotScalar (%s)", fail_eliminate);
 703       if (res == NULL)
 704         alloc->dump();
 705       else
 706         res->dump();
 707 #ifdef ASSERT
 708       if (disq_node != NULL) {
 709           tty->print("  >>>> ");
 710           disq_node->dump();
 711       }
 712 #endif /*ASSERT*/
 713     }
 714   }
 715 #endif
 716   return can_eliminate;
 717 }
 718 
 719 // Do scalar replacement.
 720 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
 721   GrowableArray <SafePointNode *> safepoints_done;
 722 
 723   ciKlass* klass = NULL;
 724   ciInstanceKlass* iklass = NULL;
 725   int nfields = 0;
 726   int array_base = 0;
 727   int element_size = 0;
 728   BasicType basic_elem_type = T_ILLEGAL;
 729   ciType* elem_type = NULL;
 730 
 731   Node* res = alloc->result_cast();
 732   assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
 733   const TypeOopPtr* res_type = NULL;
 734   if (res != NULL) { // Could be NULL when there are no users
 735     res_type = _igvn.type(res)->isa_oopptr();
 736   }
 737 
 738   if (res != NULL) {
 739     klass = res_type->klass();
 740     if (res_type->isa_instptr()) {
 741       // find the fields of the class which will be needed for safepoint debug information
 742       assert(klass->is_instance_klass(), "must be an instance klass.");
 743       iklass = klass->as_instance_klass();
 744       nfields = iklass->nof_nonstatic_fields();
 745     } else {
 746       // find the array's elements which will be needed for safepoint debug information
 747       nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 748       assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
 749       elem_type = klass->as_array_klass()->element_type();
 750       basic_elem_type = elem_type->basic_type();
 751       array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 752       element_size = type2aelembytes(basic_elem_type);
 753     }
 754   }
 755   //
 756   // Process the safepoint uses
 757   //
 758   while (safepoints.length() > 0) {
 759     SafePointNode* sfpt = safepoints.pop();
 760     Node* mem = sfpt->memory();
 761     Node* ctl = sfpt->control();
 762     assert(sfpt->jvms() != NULL, "missed JVMS");
 763     // Fields of scalar objs are referenced only at the end
 764     // of regular debuginfo at the last (youngest) JVMS.
 765     // Record relative start index.
 766     uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
 767     SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
 768 #ifdef ASSERT
 769                                                  alloc,
 770 #endif
 771                                                  first_ind, nfields);
 772     sobj->init_req(0, C->root());
 773     transform_later(sobj);
 774 
 775     // Scan object's fields adding an input to the safepoint for each field.
 776     for (int j = 0; j < nfields; j++) {
 777       intptr_t offset;
 778       ciField* field = NULL;
 779       if (iklass != NULL) {
 780         field = iklass->nonstatic_field_at(j);
 781         offset = field->offset();
 782         elem_type = field->type();
 783         basic_elem_type = field->layout_type();
 784       } else {
 785         offset = array_base + j * (intptr_t)element_size;
 786       }
 787 
 788       const Type *field_type;
 789       // The next code is taken from Parse::do_get_xxx().
 790       if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
 791         if (!elem_type->is_loaded()) {
 792           field_type = TypeInstPtr::BOTTOM;
 793         } else if (field != NULL && field->is_static_constant()) {
 794           // This can happen if the constant oop is non-perm.
 795           ciObject* con = field->constant_value().as_object();
 796           // Do not "join" in the previous type; it doesn't add value,
 797           // and may yield a vacuous result if the field is of interface type.
 798           field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
 799           assert(field_type != NULL, "field singleton type must be consistent");
 800         } else {
 801           field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
 802         }
 803         if (UseCompressedOops) {
 804           field_type = field_type->make_narrowoop();
 805           basic_elem_type = T_NARROWOOP;
 806         }
 807       } else {
 808         field_type = Type::get_const_basic_type(basic_elem_type);
 809       }
 810 
 811       const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
 812 
 813       Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
 814       if (field_val == NULL) {
 815         // We weren't able to find a value for this field,
 816         // give up on eliminating this allocation.
 817 
 818         // Remove any extra entries we added to the safepoint.
 819         uint last = sfpt->req() - 1;
 820         for (int k = 0;  k < j; k++) {
 821           sfpt->del_req(last--);
 822         }
 823         _igvn._worklist.push(sfpt);
 824         // rollback processed safepoints
 825         while (safepoints_done.length() > 0) {
 826           SafePointNode* sfpt_done = safepoints_done.pop();
 827           // remove any extra entries we added to the safepoint
 828           last = sfpt_done->req() - 1;
 829           for (int k = 0;  k < nfields; k++) {
 830             sfpt_done->del_req(last--);
 831           }
 832           JVMState *jvms = sfpt_done->jvms();
 833           jvms->set_endoff(sfpt_done->req());
 834           // Now make a pass over the debug information replacing any references
 835           // to SafePointScalarObjectNode with the allocated object.
 836           int start = jvms->debug_start();
 837           int end   = jvms->debug_end();
 838           for (int i = start; i < end; i++) {
 839             if (sfpt_done->in(i)->is_SafePointScalarObject()) {
 840               SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
 841               if (scobj->first_index(jvms) == sfpt_done->req() &&
 842                   scobj->n_fields() == (uint)nfields) {
 843                 assert(scobj->alloc() == alloc, "sanity");
 844                 sfpt_done->set_req(i, res);
 845               }
 846             }
 847           }
 848           _igvn._worklist.push(sfpt_done);
 849         }
 850 #ifndef PRODUCT
 851         if (PrintEliminateAllocations) {
 852           if (field != NULL) {
 853             tty->print("=== At SafePoint node %d can't find value of Field: ",
 854                        sfpt->_idx);
 855             field->print();
 856             int field_idx = C->get_alias_index(field_addr_type);
 857             tty->print(" (alias_idx=%d)", field_idx);
 858           } else { // Array's element
 859             tty->print("=== At SafePoint node %d can't find value of array element [%d]",
 860                        sfpt->_idx, j);
 861           }
 862           tty->print(", which prevents elimination of: ");
 863           if (res == NULL)
 864             alloc->dump();
 865           else
 866             res->dump();
 867         }
 868 #endif
 869         return false;
 870       }
 871       if (UseCompressedOops && field_type->isa_narrowoop()) {
 872         // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
 873         // to be able scalar replace the allocation.
 874         if (field_val->is_EncodeP()) {
 875           field_val = field_val->in(1);
 876         } else {
 877           field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
 878         }
 879       }
 880       sfpt->add_req(field_val);
 881     }
 882     JVMState *jvms = sfpt->jvms();
 883     jvms->set_endoff(sfpt->req());
 884     // Now make a pass over the debug information replacing any references
 885     // to the allocated object with "sobj"
 886     int start = jvms->debug_start();
 887     int end   = jvms->debug_end();
 888     sfpt->replace_edges_in_range(res, sobj, start, end);
 889     _igvn._worklist.push(sfpt);
 890     safepoints_done.append_if_missing(sfpt); // keep it for rollback
 891   }
 892   return true;
 893 }
 894 
 895 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
 896   Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
 897   Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
 898   if (ctl_proj != NULL) {
 899     igvn.replace_node(ctl_proj, n->in(0));
 900   }
 901   if (mem_proj != NULL) {
 902     igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
 903   }
 904 }
 905 
 906 // Process users of eliminated allocation.
 907 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
 908   Node* res = alloc->result_cast();
 909   if (res != NULL) {
 910     for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
 911       Node *use = res->last_out(j);
 912       uint oc1 = res->outcnt();
 913 
 914       if (use->is_AddP()) {
 915         for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
 916           Node *n = use->last_out(k);
 917           uint oc2 = use->outcnt();
 918           if (n->is_Store()) {
 919 #ifdef ASSERT
 920             // Verify that there is no dependent MemBarVolatile nodes,
 921             // they should be removed during IGVN, see MemBarNode::Ideal().
 922             for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
 923                                        p < pmax; p++) {
 924               Node* mb = n->fast_out(p);
 925               assert(mb->is_Initialize() || !mb->is_MemBar() ||
 926                      mb->req() <= MemBarNode::Precedent ||
 927                      mb->in(MemBarNode::Precedent) != n,
 928                      "MemBarVolatile should be eliminated for non-escaping object");
 929             }
 930 #endif
 931             _igvn.replace_node(n, n->in(MemNode::Memory));
 932           } else if (n->is_ArrayCopy()) {
 933             // Disconnect ArrayCopy node
 934             ArrayCopyNode* ac = n->as_ArrayCopy();
 935             assert(ac->is_clonebasic(), "unexpected array copy kind");
 936             Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
 937             disconnect_projections(ac, _igvn);
 938             assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
 939             Node* membar_before = alloc->in(0)->in(0);
 940             disconnect_projections(membar_before->as_MemBar(), _igvn);
 941             if (membar_after->is_MemBar()) {
 942               disconnect_projections(membar_after->as_MemBar(), _igvn);
 943             }
 944           } else {
 945             eliminate_gc_barrier(n);
 946           }
 947           k -= (oc2 - use->outcnt());
 948         }
 949         _igvn.remove_dead_node(use);
 950       } else if (use->is_ArrayCopy()) {
 951         // Disconnect ArrayCopy node
 952         ArrayCopyNode* ac = use->as_ArrayCopy();
 953         assert(ac->is_arraycopy_validated() ||
 954                ac->is_copyof_validated() ||
 955                ac->is_copyofrange_validated(), "unsupported");
 956         CallProjections callprojs;
 957         ac->extract_projections(&callprojs, true);
 958 
 959         _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
 960         _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
 961         _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
 962 
 963         // Set control to top. IGVN will remove the remaining projections
 964         ac->set_req(0, top());
 965         ac->replace_edge(res, top());
 966 
 967         // Disconnect src right away: it can help find new
 968         // opportunities for allocation elimination
 969         Node* src = ac->in(ArrayCopyNode::Src);
 970         ac->replace_edge(src, top());
 971         // src can be top at this point if src and dest of the
 972         // arraycopy were the same
 973         if (src->outcnt() == 0 && !src->is_top()) {
 974           _igvn.remove_dead_node(src);
 975         }
 976 
 977         _igvn._worklist.push(ac);
 978       } else {
 979         eliminate_gc_barrier(use);
 980       }
 981       j -= (oc1 - res->outcnt());
 982     }
 983     assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
 984     _igvn.remove_dead_node(res);
 985   }
 986 
 987   //
 988   // Process other users of allocation's projections
 989   //
 990   if (_resproj != NULL && _resproj->outcnt() != 0) {
 991     // First disconnect stores captured by Initialize node.
 992     // If Initialize node is eliminated first in the following code,
 993     // it will kill such stores and DUIterator_Last will assert.
 994     for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax);  j < jmax; j++) {
 995       Node *use = _resproj->fast_out(j);
 996       if (use->is_AddP()) {
 997         // raw memory addresses used only by the initialization
 998         _igvn.replace_node(use, C->top());
 999         --j; --jmax;
1000       }
1001     }
1002     for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1003       Node *use = _resproj->last_out(j);
1004       uint oc1 = _resproj->outcnt();
1005       if (use->is_Initialize()) {
1006         // Eliminate Initialize node.
1007         InitializeNode *init = use->as_Initialize();
1008         assert(init->outcnt() <= 2, "only a control and memory projection expected");
1009         Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1010         if (ctrl_proj != NULL) {
1011            assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1012           _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1013         }
1014         Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1015         if (mem_proj != NULL) {
1016           Node *mem = init->in(TypeFunc::Memory);
1017 #ifdef ASSERT
1018           if (mem->is_MergeMem()) {
1019             assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1020           } else {
1021             assert(mem == _memproj_fallthrough, "allocation memory projection");
1022           }
1023 #endif
1024           _igvn.replace_node(mem_proj, mem);
1025         }
1026       } else  {
1027         assert(false, "only Initialize or AddP expected");
1028       }
1029       j -= (oc1 - _resproj->outcnt());
1030     }
1031   }
1032   if (_fallthroughcatchproj != NULL) {
1033     _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1034   }
1035   if (_memproj_fallthrough != NULL) {
1036     _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1037   }
1038   if (_memproj_catchall != NULL) {
1039     _igvn.replace_node(_memproj_catchall, C->top());
1040   }
1041   if (_ioproj_fallthrough != NULL) {
1042     _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1043   }
1044   if (_ioproj_catchall != NULL) {
1045     _igvn.replace_node(_ioproj_catchall, C->top());
1046   }
1047   if (_catchallcatchproj != NULL) {
1048     _igvn.replace_node(_catchallcatchproj, C->top());
1049   }
1050 }
1051 
1052 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1053   // Don't do scalar replacement if the frame can be popped by JVMTI:
1054   // if reallocation fails during deoptimization we'll pop all
1055   // interpreter frames for this compiled frame and that won't play
1056   // nice with JVMTI popframe.
1057   if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1058     return false;
1059   }
1060   Node* klass = alloc->in(AllocateNode::KlassNode);
1061   const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1062   Node* res = alloc->result_cast();
1063   // Eliminate boxing allocations which are not used
1064   // regardless scalar replacable status.
1065   bool boxing_alloc = C->eliminate_boxing() &&
1066                       tklass->klass()->is_instance_klass()  &&
1067                       tklass->klass()->as_instance_klass()->is_box_klass();
1068   if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1069     return false;
1070   }
1071 
1072   extract_call_projections(alloc);
1073 
1074   GrowableArray <SafePointNode *> safepoints;
1075   if (!can_eliminate_allocation(alloc, safepoints)) {
1076     return false;
1077   }
1078 
1079   if (!alloc->_is_scalar_replaceable) {
1080     assert(res == NULL, "sanity");
1081     // We can only eliminate allocation if all debug info references
1082     // are already replaced with SafePointScalarObject because
1083     // we can't search for a fields value without instance_id.
1084     if (safepoints.length() > 0) {
1085       return false;
1086     }
1087   }
1088 
1089   if (!scalar_replacement(alloc, safepoints)) {
1090     return false;
1091   }
1092 
1093   CompileLog* log = C->log();
1094   if (log != NULL) {
1095     log->head("eliminate_allocation type='%d'",
1096               log->identify(tklass->klass()));
1097     JVMState* p = alloc->jvms();
1098     while (p != NULL) {
1099       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1100       p = p->caller();
1101     }
1102     log->tail("eliminate_allocation");
1103   }
1104 
1105   process_users_of_allocation(alloc);
1106 
1107 #ifndef PRODUCT
1108   if (PrintEliminateAllocations) {
1109     if (alloc->is_AllocateArray())
1110       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1111     else
1112       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1113   }
1114 #endif
1115 
1116   return true;
1117 }
1118 
1119 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1120   // EA should remove all uses of non-escaping boxing node.
1121   if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1122     return false;
1123   }
1124 
1125   assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1126 
1127   extract_call_projections(boxing);
1128 
1129   const TypeTuple* r = boxing->tf()->range();
1130   assert(r->cnt() > TypeFunc::Parms, "sanity");
1131   const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1132   assert(t != NULL, "sanity");
1133 
1134   CompileLog* log = C->log();
1135   if (log != NULL) {
1136     log->head("eliminate_boxing type='%d'",
1137               log->identify(t->klass()));
1138     JVMState* p = boxing->jvms();
1139     while (p != NULL) {
1140       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1141       p = p->caller();
1142     }
1143     log->tail("eliminate_boxing");
1144   }
1145 
1146   process_users_of_allocation(boxing);
1147 
1148 #ifndef PRODUCT
1149   if (PrintEliminateAllocations) {
1150     tty->print("++++ Eliminated: %d ", boxing->_idx);
1151     boxing->method()->print_short_name(tty);
1152     tty->cr();
1153   }
1154 #endif
1155 
1156   return true;
1157 }
1158 
1159 //---------------------------set_eden_pointers-------------------------
1160 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1161   if (UseTLAB) {                // Private allocation: load from TLS
1162     Node* thread = transform_later(new ThreadLocalNode());
1163     int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1164     int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1165     eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1166     eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1167   } else {                      // Shared allocation: load from globals
1168     CollectedHeap* ch = Universe::heap();
1169     address top_adr = (address)ch->top_addr();
1170     address end_adr = (address)ch->end_addr();
1171     eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1172     eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1173   }
1174 }
1175 
1176 
1177 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1178   Node* adr = basic_plus_adr(base, offset);
1179   const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1180   Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1181   transform_later(value);
1182   return value;
1183 }
1184 
1185 
1186 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1187   Node* adr = basic_plus_adr(base, offset);
1188   mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1189   transform_later(mem);
1190   return mem;
1191 }
1192 
1193 //=============================================================================
1194 //
1195 //                              A L L O C A T I O N
1196 //
1197 // Allocation attempts to be fast in the case of frequent small objects.
1198 // It breaks down like this:
1199 //
1200 // 1) Size in doublewords is computed.  This is a constant for objects and
1201 // variable for most arrays.  Doubleword units are used to avoid size
1202 // overflow of huge doubleword arrays.  We need doublewords in the end for
1203 // rounding.
1204 //
1205 // 2) Size is checked for being 'too large'.  Too-large allocations will go
1206 // the slow path into the VM.  The slow path can throw any required
1207 // exceptions, and does all the special checks for very large arrays.  The
1208 // size test can constant-fold away for objects.  For objects with
1209 // finalizers it constant-folds the otherway: you always go slow with
1210 // finalizers.
1211 //
1212 // 3) If NOT using TLABs, this is the contended loop-back point.
1213 // Load-Locked the heap top.  If using TLABs normal-load the heap top.
1214 //
1215 // 4) Check that heap top + size*8 < max.  If we fail go the slow ` route.
1216 // NOTE: "top+size*8" cannot wrap the 4Gig line!  Here's why: for largish
1217 // "size*8" we always enter the VM, where "largish" is a constant picked small
1218 // enough that there's always space between the eden max and 4Gig (old space is
1219 // there so it's quite large) and large enough that the cost of entering the VM
1220 // is dwarfed by the cost to initialize the space.
1221 //
1222 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1223 // down.  If contended, repeat at step 3.  If using TLABs normal-store
1224 // adjusted heap top back down; there is no contention.
1225 //
1226 // 6) If !ZeroTLAB then Bulk-clear the object/array.  Fill in klass & mark
1227 // fields.
1228 //
1229 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1230 // oop flavor.
1231 //
1232 //=============================================================================
1233 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1234 // Allocations bigger than this always go the slow route.
1235 // This value must be small enough that allocation attempts that need to
1236 // trigger exceptions go the slow route.  Also, it must be small enough so
1237 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1238 //=============================================================================j//
1239 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1240 // The allocator will coalesce int->oop copies away.  See comment in
1241 // coalesce.cpp about how this works.  It depends critically on the exact
1242 // code shape produced here, so if you are changing this code shape
1243 // make sure the GC info for the heap-top is correct in and around the
1244 // slow-path call.
1245 //
1246 
1247 void PhaseMacroExpand::expand_allocate_common(
1248             AllocateNode* alloc, // allocation node to be expanded
1249             Node* length,  // array length for an array allocation
1250             const TypeFunc* slow_call_type, // Type of slow call
1251             address slow_call_address  // Address of slow call
1252     )
1253 {
1254 
1255   Node* ctrl = alloc->in(TypeFunc::Control);
1256   Node* mem  = alloc->in(TypeFunc::Memory);
1257   Node* i_o  = alloc->in(TypeFunc::I_O);
1258   Node* size_in_bytes     = alloc->in(AllocateNode::AllocSize);
1259   Node* klass_node        = alloc->in(AllocateNode::KlassNode);
1260   Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1261 
1262   assert(ctrl != NULL, "must have control");
1263   // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1264   // they will not be used if "always_slow" is set
1265   enum { slow_result_path = 1, fast_result_path = 2 };
1266   Node *result_region = NULL;
1267   Node *result_phi_rawmem = NULL;
1268   Node *result_phi_rawoop = NULL;
1269   Node *result_phi_i_o = NULL;
1270 
1271   // The initial slow comparison is a size check, the comparison
1272   // we want to do is a BoolTest::gt
1273   bool always_slow = false;
1274   int tv = _igvn.find_int_con(initial_slow_test, -1);
1275   if (tv >= 0) {
1276     always_slow = (tv == 1);
1277     initial_slow_test = NULL;
1278   } else {
1279     initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1280   }
1281 
1282   if (C->env()->dtrace_alloc_probes() ||
1283       (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1284     // Force slow-path allocation
1285     always_slow = true;
1286     initial_slow_test = NULL;
1287   }
1288 
1289 
1290   enum { too_big_or_final_path = 1, need_gc_path = 2 };
1291   Node *slow_region = NULL;
1292   Node *toobig_false = ctrl;
1293 
1294   assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1295   // generate the initial test if necessary
1296   if (initial_slow_test != NULL ) {
1297     slow_region = new RegionNode(3);
1298 
1299     // Now make the initial failure test.  Usually a too-big test but
1300     // might be a TRUE for finalizers or a fancy class check for
1301     // newInstance0.
1302     IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1303     transform_later(toobig_iff);
1304     // Plug the failing-too-big test into the slow-path region
1305     Node *toobig_true = new IfTrueNode( toobig_iff );
1306     transform_later(toobig_true);
1307     slow_region    ->init_req( too_big_or_final_path, toobig_true );
1308     toobig_false = new IfFalseNode( toobig_iff );
1309     transform_later(toobig_false);
1310   } else {         // No initial test, just fall into next case
1311     toobig_false = ctrl;
1312     debug_only(slow_region = NodeSentinel);
1313   }
1314 
1315   Node *slow_mem = mem;  // save the current memory state for slow path
1316   // generate the fast allocation code unless we know that the initial test will always go slow
1317   if (!always_slow) {
1318     // Fast path modifies only raw memory.
1319     if (mem->is_MergeMem()) {
1320       mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1321     }
1322 
1323     // allocate the Region and Phi nodes for the result
1324     result_region = new RegionNode(3);
1325     result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1326     result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1327     result_phi_i_o    = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1328 
1329     // Grab regular I/O before optional prefetch may change it.
1330     // Slow-path does no I/O so just set it to the original I/O.
1331     result_phi_i_o->init_req(slow_result_path, i_o);
1332 
1333     Node* needgc_ctrl = NULL;
1334     // Name successful fast-path variables
1335     Node* fast_oop_ctrl;
1336     Node* fast_oop_rawmem;
1337 
1338     intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1339 
1340     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1341     Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1342                                       fast_oop_ctrl, fast_oop_rawmem,
1343                                       prefetch_lines);
1344 
1345     if (initial_slow_test) {
1346       slow_region->init_req(need_gc_path, needgc_ctrl);
1347       // This completes all paths into the slow merge point
1348       transform_later(slow_region);
1349     } else {                      // No initial slow path needed!
1350       // Just fall from the need-GC path straight into the VM call.
1351       slow_region = needgc_ctrl;
1352     }
1353 
1354     InitializeNode* init = alloc->initialization();
1355     fast_oop_rawmem = initialize_object(alloc,
1356                                         fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1357                                         klass_node, length, size_in_bytes);
1358 
1359     // If initialization is performed by an array copy, any required
1360     // MemBarStoreStore was already added. If the object does not
1361     // escape no need for a MemBarStoreStore. If the object does not
1362     // escape in its initializer and memory barrier (MemBarStoreStore or
1363     // stronger) is already added at exit of initializer, also no need
1364     // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1365     // so that stores that initialize this object can't be reordered
1366     // with a subsequent store that makes this object accessible by
1367     // other threads.
1368     // Other threads include java threads and JVM internal threads
1369     // (for example concurrent GC threads). Current concurrent GC
1370     // implementation: CMS and G1 will not scan newly created object,
1371     // so it's safe to skip storestore barrier when allocation does
1372     // not escape.
1373     if (!alloc->does_not_escape_thread() &&
1374         !alloc->is_allocation_MemBar_redundant() &&
1375         (init == NULL || !init->is_complete_with_arraycopy())) {
1376       if (init == NULL || init->req() < InitializeNode::RawStores) {
1377         // No InitializeNode or no stores captured by zeroing
1378         // elimination. Simply add the MemBarStoreStore after object
1379         // initialization.
1380         MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1381         transform_later(mb);
1382 
1383         mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1384         mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1385         fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1386         transform_later(fast_oop_ctrl);
1387         fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1388         transform_later(fast_oop_rawmem);
1389       } else {
1390         // Add the MemBarStoreStore after the InitializeNode so that
1391         // all stores performing the initialization that were moved
1392         // before the InitializeNode happen before the storestore
1393         // barrier.
1394 
1395         Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1396         Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1397 
1398         MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1399         transform_later(mb);
1400 
1401         Node* ctrl = new ProjNode(init,TypeFunc::Control);
1402         transform_later(ctrl);
1403         Node* mem = new ProjNode(init,TypeFunc::Memory);
1404         transform_later(mem);
1405 
1406         // The MemBarStoreStore depends on control and memory coming
1407         // from the InitializeNode
1408         mb->init_req(TypeFunc::Memory, mem);
1409         mb->init_req(TypeFunc::Control, ctrl);
1410 
1411         ctrl = new ProjNode(mb,TypeFunc::Control);
1412         transform_later(ctrl);
1413         mem = new ProjNode(mb,TypeFunc::Memory);
1414         transform_later(mem);
1415 
1416         // All nodes that depended on the InitializeNode for control
1417         // and memory must now depend on the MemBarNode that itself
1418         // depends on the InitializeNode
1419         if (init_ctrl != NULL) {
1420           _igvn.replace_node(init_ctrl, ctrl);
1421         }
1422         if (init_mem != NULL) {
1423           _igvn.replace_node(init_mem, mem);
1424         }
1425       }
1426     }
1427 
1428     if (C->env()->dtrace_extended_probes()) {
1429       // Slow-path call
1430       int size = TypeFunc::Parms + 2;
1431       CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1432                                             CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1433                                             "dtrace_object_alloc",
1434                                             TypeRawPtr::BOTTOM);
1435 
1436       // Get base of thread-local storage area
1437       Node* thread = new ThreadLocalNode();
1438       transform_later(thread);
1439 
1440       call->init_req(TypeFunc::Parms+0, thread);
1441       call->init_req(TypeFunc::Parms+1, fast_oop);
1442       call->init_req(TypeFunc::Control, fast_oop_ctrl);
1443       call->init_req(TypeFunc::I_O    , top()); // does no i/o
1444       call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1445       call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1446       call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1447       transform_later(call);
1448       fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1449       transform_later(fast_oop_ctrl);
1450       fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1451       transform_later(fast_oop_rawmem);
1452     }
1453 
1454     // Plug in the successful fast-path into the result merge point
1455     result_region    ->init_req(fast_result_path, fast_oop_ctrl);
1456     result_phi_rawoop->init_req(fast_result_path, fast_oop);
1457     result_phi_i_o   ->init_req(fast_result_path, i_o);
1458     result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1459   } else {
1460     slow_region = ctrl;
1461     result_phi_i_o = i_o; // Rename it to use in the following code.
1462   }
1463 
1464   // Generate slow-path call
1465   CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1466                                OptoRuntime::stub_name(slow_call_address),
1467                                alloc->jvms()->bci(),
1468                                TypePtr::BOTTOM);
1469   call->init_req( TypeFunc::Control, slow_region );
1470   call->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
1471   call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1472   call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1473   call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1474 
1475   call->init_req(TypeFunc::Parms+0, klass_node);
1476   if (length != NULL) {
1477     call->init_req(TypeFunc::Parms+1, length);
1478   }
1479 
1480   // Copy debug information and adjust JVMState information, then replace
1481   // allocate node with the call
1482   copy_call_debug_info((CallNode *) alloc,  call);
1483   if (!always_slow) {
1484     call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1485   } else {
1486     // Hook i_o projection to avoid its elimination during allocation
1487     // replacement (when only a slow call is generated).
1488     call->set_req(TypeFunc::I_O, result_phi_i_o);
1489   }
1490   _igvn.replace_node(alloc, call);
1491   transform_later(call);
1492 
1493   // Identify the output projections from the allocate node and
1494   // adjust any references to them.
1495   // The control and io projections look like:
1496   //
1497   //        v---Proj(ctrl) <-----+   v---CatchProj(ctrl)
1498   //  Allocate                   Catch
1499   //        ^---Proj(io) <-------+   ^---CatchProj(io)
1500   //
1501   //  We are interested in the CatchProj nodes.
1502   //
1503   extract_call_projections(call);
1504 
1505   // An allocate node has separate memory projections for the uses on
1506   // the control and i_o paths. Replace the control memory projection with
1507   // result_phi_rawmem (unless we are only generating a slow call when
1508   // both memory projections are combined)
1509   if (!always_slow && _memproj_fallthrough != NULL) {
1510     for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1511       Node *use = _memproj_fallthrough->fast_out(i);
1512       _igvn.rehash_node_delayed(use);
1513       imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1514       // back up iterator
1515       --i;
1516     }
1517   }
1518   // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1519   // _memproj_catchall so we end up with a call that has only 1 memory projection.
1520   if (_memproj_catchall != NULL ) {
1521     if (_memproj_fallthrough == NULL) {
1522       _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1523       transform_later(_memproj_fallthrough);
1524     }
1525     for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1526       Node *use = _memproj_catchall->fast_out(i);
1527       _igvn.rehash_node_delayed(use);
1528       imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1529       // back up iterator
1530       --i;
1531     }
1532     assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1533     _igvn.remove_dead_node(_memproj_catchall);
1534   }
1535 
1536   // An allocate node has separate i_o projections for the uses on the control
1537   // and i_o paths. Always replace the control i_o projection with result i_o
1538   // otherwise incoming i_o become dead when only a slow call is generated
1539   // (it is different from memory projections where both projections are
1540   // combined in such case).
1541   if (_ioproj_fallthrough != NULL) {
1542     for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1543       Node *use = _ioproj_fallthrough->fast_out(i);
1544       _igvn.rehash_node_delayed(use);
1545       imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1546       // back up iterator
1547       --i;
1548     }
1549   }
1550   // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1551   // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1552   if (_ioproj_catchall != NULL ) {
1553     if (_ioproj_fallthrough == NULL) {
1554       _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1555       transform_later(_ioproj_fallthrough);
1556     }
1557     for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1558       Node *use = _ioproj_catchall->fast_out(i);
1559       _igvn.rehash_node_delayed(use);
1560       imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1561       // back up iterator
1562       --i;
1563     }
1564     assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1565     _igvn.remove_dead_node(_ioproj_catchall);
1566   }
1567 
1568   // if we generated only a slow call, we are done
1569   if (always_slow) {
1570     // Now we can unhook i_o.
1571     if (result_phi_i_o->outcnt() > 1) {
1572       call->set_req(TypeFunc::I_O, top());
1573     } else {
1574       assert(result_phi_i_o->unique_ctrl_out() == call, "");
1575       // Case of new array with negative size known during compilation.
1576       // AllocateArrayNode::Ideal() optimization disconnect unreachable
1577       // following code since call to runtime will throw exception.
1578       // As result there will be no users of i_o after the call.
1579       // Leave i_o attached to this call to avoid problems in preceding graph.
1580     }
1581     return;
1582   }
1583 
1584 
1585   if (_fallthroughcatchproj != NULL) {
1586     ctrl = _fallthroughcatchproj->clone();
1587     transform_later(ctrl);
1588     _igvn.replace_node(_fallthroughcatchproj, result_region);
1589   } else {
1590     ctrl = top();
1591   }
1592   Node *slow_result;
1593   if (_resproj == NULL) {
1594     // no uses of the allocation result
1595     slow_result = top();
1596   } else {
1597     slow_result = _resproj->clone();
1598     transform_later(slow_result);
1599     _igvn.replace_node(_resproj, result_phi_rawoop);
1600   }
1601 
1602   // Plug slow-path into result merge point
1603   result_region    ->init_req( slow_result_path, ctrl );
1604   result_phi_rawoop->init_req( slow_result_path, slow_result);
1605   result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1606   transform_later(result_region);
1607   transform_later(result_phi_rawoop);
1608   transform_later(result_phi_rawmem);
1609   transform_later(result_phi_i_o);
1610   // This completes all paths into the result merge point
1611 }
1612 
1613 
1614 // Helper for PhaseMacroExpand::expand_allocate_common.
1615 // Initializes the newly-allocated storage.
1616 Node*
1617 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1618                                     Node* control, Node* rawmem, Node* object,
1619                                     Node* klass_node, Node* length,
1620                                     Node* size_in_bytes) {
1621   InitializeNode* init = alloc->initialization();
1622   // Store the klass & mark bits
1623   Node* mark_node = NULL;
1624   // For now only enable fast locking for non-array types
1625   if (UseBiasedLocking && (length == NULL)) {
1626     mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1627   } else {
1628     mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1629   }
1630   rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1631 
1632   rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1633   int header_size = alloc->minimum_header_size();  // conservatively small
1634 
1635   // Array length
1636   if (length != NULL) {         // Arrays need length field
1637     rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1638     // conservatively small header size:
1639     header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1640     ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1641     if (k->is_array_klass())    // we know the exact header size in most cases:
1642       header_size = Klass::layout_helper_header_size(k->layout_helper());
1643   }
1644 
1645   // Clear the object body, if necessary.
1646   if (init == NULL) {
1647     // The init has somehow disappeared; be cautious and clear everything.
1648     //
1649     // This can happen if a node is allocated but an uncommon trap occurs
1650     // immediately.  In this case, the Initialize gets associated with the
1651     // trap, and may be placed in a different (outer) loop, if the Allocate
1652     // is in a loop.  If (this is rare) the inner loop gets unrolled, then
1653     // there can be two Allocates to one Initialize.  The answer in all these
1654     // edge cases is safety first.  It is always safe to clear immediately
1655     // within an Allocate, and then (maybe or maybe not) clear some more later.
1656     if (!(UseTLAB && ZeroTLAB)) {
1657       rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1658                                             header_size, size_in_bytes,
1659                                             &_igvn);
1660     }
1661   } else {
1662     if (!init->is_complete()) {
1663       // Try to win by zeroing only what the init does not store.
1664       // We can also try to do some peephole optimizations,
1665       // such as combining some adjacent subword stores.
1666       rawmem = init->complete_stores(control, rawmem, object,
1667                                      header_size, size_in_bytes, &_igvn);
1668     }
1669     // We have no more use for this link, since the AllocateNode goes away:
1670     init->set_req(InitializeNode::RawAddress, top());
1671     // (If we keep the link, it just confuses the register allocator,
1672     // who thinks he sees a real use of the address by the membar.)
1673   }
1674 
1675   return rawmem;
1676 }
1677 
1678 // Generate prefetch instructions for next allocations.
1679 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1680                                         Node*& contended_phi_rawmem,
1681                                         Node* old_eden_top, Node* new_eden_top,
1682                                         intx lines) {
1683    enum { fall_in_path = 1, pf_path = 2 };
1684    if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1685       // Generate prefetch allocation with watermark check.
1686       // As an allocation hits the watermark, we will prefetch starting
1687       // at a "distance" away from watermark.
1688 
1689       Node *pf_region = new RegionNode(3);
1690       Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1691                                                 TypeRawPtr::BOTTOM );
1692       // I/O is used for Prefetch
1693       Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1694 
1695       Node *thread = new ThreadLocalNode();
1696       transform_later(thread);
1697 
1698       Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1699                    _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1700       transform_later(eden_pf_adr);
1701 
1702       Node *old_pf_wm = new LoadPNode(needgc_false,
1703                                    contended_phi_rawmem, eden_pf_adr,
1704                                    TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1705                                    MemNode::unordered);
1706       transform_later(old_pf_wm);
1707 
1708       // check against new_eden_top
1709       Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1710       transform_later(need_pf_cmp);
1711       Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1712       transform_later(need_pf_bol);
1713       IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1714                                        PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1715       transform_later(need_pf_iff);
1716 
1717       // true node, add prefetchdistance
1718       Node *need_pf_true = new IfTrueNode( need_pf_iff );
1719       transform_later(need_pf_true);
1720 
1721       Node *need_pf_false = new IfFalseNode( need_pf_iff );
1722       transform_later(need_pf_false);
1723 
1724       Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1725                                     _igvn.MakeConX(AllocatePrefetchDistance) );
1726       transform_later(new_pf_wmt );
1727       new_pf_wmt->set_req(0, need_pf_true);
1728 
1729       Node *store_new_wmt = new StorePNode(need_pf_true,
1730                                        contended_phi_rawmem, eden_pf_adr,
1731                                        TypeRawPtr::BOTTOM, new_pf_wmt,
1732                                        MemNode::unordered);
1733       transform_later(store_new_wmt);
1734 
1735       // adding prefetches
1736       pf_phi_abio->init_req( fall_in_path, i_o );
1737 
1738       Node *prefetch_adr;
1739       Node *prefetch;
1740       uint step_size = AllocatePrefetchStepSize;
1741       uint distance = 0;
1742 
1743       for ( intx i = 0; i < lines; i++ ) {
1744         prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1745                                             _igvn.MakeConX(distance) );
1746         transform_later(prefetch_adr);
1747         prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1748         transform_later(prefetch);
1749         distance += step_size;
1750         i_o = prefetch;
1751       }
1752       pf_phi_abio->set_req( pf_path, i_o );
1753 
1754       pf_region->init_req( fall_in_path, need_pf_false );
1755       pf_region->init_req( pf_path, need_pf_true );
1756 
1757       pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1758       pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1759 
1760       transform_later(pf_region);
1761       transform_later(pf_phi_rawmem);
1762       transform_later(pf_phi_abio);
1763 
1764       needgc_false = pf_region;
1765       contended_phi_rawmem = pf_phi_rawmem;
1766       i_o = pf_phi_abio;
1767    } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1768       // Insert a prefetch instruction for each allocation.
1769       // This code is used to generate 1 prefetch instruction per cache line.
1770 
1771       // Generate several prefetch instructions.
1772       uint step_size = AllocatePrefetchStepSize;
1773       uint distance = AllocatePrefetchDistance;
1774 
1775       // Next cache address.
1776       Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1777                                      _igvn.MakeConX(step_size + distance));
1778       transform_later(cache_adr);
1779       cache_adr = new CastP2XNode(needgc_false, cache_adr);
1780       transform_later(cache_adr);
1781       // Address is aligned to execute prefetch to the beginning of cache line size
1782       // (it is important when BIS instruction is used on SPARC as prefetch).
1783       Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1784       cache_adr = new AndXNode(cache_adr, mask);
1785       transform_later(cache_adr);
1786       cache_adr = new CastX2PNode(cache_adr);
1787       transform_later(cache_adr);
1788 
1789       // Prefetch
1790       Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1791       prefetch->set_req(0, needgc_false);
1792       transform_later(prefetch);
1793       contended_phi_rawmem = prefetch;
1794       Node *prefetch_adr;
1795       distance = step_size;
1796       for ( intx i = 1; i < lines; i++ ) {
1797         prefetch_adr = new AddPNode( cache_adr, cache_adr,
1798                                             _igvn.MakeConX(distance) );
1799         transform_later(prefetch_adr);
1800         prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1801         transform_later(prefetch);
1802         distance += step_size;
1803         contended_phi_rawmem = prefetch;
1804       }
1805    } else if( AllocatePrefetchStyle > 0 ) {
1806       // Insert a prefetch for each allocation only on the fast-path
1807       Node *prefetch_adr;
1808       Node *prefetch;
1809       // Generate several prefetch instructions.
1810       uint step_size = AllocatePrefetchStepSize;
1811       uint distance = AllocatePrefetchDistance;
1812       for ( intx i = 0; i < lines; i++ ) {
1813         prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1814                                             _igvn.MakeConX(distance) );
1815         transform_later(prefetch_adr);
1816         prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1817         // Do not let it float too high, since if eden_top == eden_end,
1818         // both might be null.
1819         if( i == 0 ) { // Set control for first prefetch, next follows it
1820           prefetch->init_req(0, needgc_false);
1821         }
1822         transform_later(prefetch);
1823         distance += step_size;
1824         i_o = prefetch;
1825       }
1826    }
1827    return i_o;
1828 }
1829 
1830 
1831 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1832   expand_allocate_common(alloc, NULL,
1833                          OptoRuntime::new_instance_Type(),
1834                          OptoRuntime::new_instance_Java());
1835 }
1836 
1837 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1838   Node* length = alloc->in(AllocateNode::ALength);
1839   InitializeNode* init = alloc->initialization();
1840   Node* klass_node = alloc->in(AllocateNode::KlassNode);
1841   ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1842   address slow_call_address;  // Address of slow call
1843   if (init != NULL && init->is_complete_with_arraycopy() &&
1844       k->is_type_array_klass()) {
1845     // Don't zero type array during slow allocation in VM since
1846     // it will be initialized later by arraycopy in compiled code.
1847     slow_call_address = OptoRuntime::new_array_nozero_Java();
1848   } else {
1849     slow_call_address = OptoRuntime::new_array_Java();
1850   }
1851   expand_allocate_common(alloc, length,
1852                          OptoRuntime::new_array_Type(),
1853                          slow_call_address);
1854 }
1855 
1856 //-------------------mark_eliminated_box----------------------------------
1857 //
1858 // During EA obj may point to several objects but after few ideal graph
1859 // transformations (CCP) it may point to only one non escaping object
1860 // (but still using phi), corresponding locks and unlocks will be marked
1861 // for elimination. Later obj could be replaced with a new node (new phi)
1862 // and which does not have escape information. And later after some graph
1863 // reshape other locks and unlocks (which were not marked for elimination
1864 // before) are connected to this new obj (phi) but they still will not be
1865 // marked for elimination since new obj has no escape information.
1866 // Mark all associated (same box and obj) lock and unlock nodes for
1867 // elimination if some of them marked already.
1868 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1869   if (oldbox->as_BoxLock()->is_eliminated())
1870     return; // This BoxLock node was processed already.
1871 
1872   // New implementation (EliminateNestedLocks) has separate BoxLock
1873   // node for each locked region so mark all associated locks/unlocks as
1874   // eliminated even if different objects are referenced in one locked region
1875   // (for example, OSR compilation of nested loop inside locked scope).
1876   if (EliminateNestedLocks ||
1877       oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1878     // Box is used only in one lock region. Mark this box as eliminated.
1879     _igvn.hash_delete(oldbox);
1880     oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1881      _igvn.hash_insert(oldbox);
1882 
1883     for (uint i = 0; i < oldbox->outcnt(); i++) {
1884       Node* u = oldbox->raw_out(i);
1885       if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1886         AbstractLockNode* alock = u->as_AbstractLock();
1887         // Check lock's box since box could be referenced by Lock's debug info.
1888         if (alock->box_node() == oldbox) {
1889           // Mark eliminated all related locks and unlocks.
1890 #ifdef ASSERT
1891           alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1892 #endif
1893           alock->set_non_esc_obj();
1894         }
1895       }
1896     }
1897     return;
1898   }
1899 
1900   // Create new "eliminated" BoxLock node and use it in monitor debug info
1901   // instead of oldbox for the same object.
1902   BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1903 
1904   // Note: BoxLock node is marked eliminated only here and it is used
1905   // to indicate that all associated lock and unlock nodes are marked
1906   // for elimination.
1907   newbox->set_eliminated();
1908   transform_later(newbox);
1909 
1910   // Replace old box node with new box for all users of the same object.
1911   for (uint i = 0; i < oldbox->outcnt();) {
1912     bool next_edge = true;
1913 
1914     Node* u = oldbox->raw_out(i);
1915     if (u->is_AbstractLock()) {
1916       AbstractLockNode* alock = u->as_AbstractLock();
1917       if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1918         // Replace Box and mark eliminated all related locks and unlocks.
1919 #ifdef ASSERT
1920         alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1921 #endif
1922         alock->set_non_esc_obj();
1923         _igvn.rehash_node_delayed(alock);
1924         alock->set_box_node(newbox);
1925         next_edge = false;
1926       }
1927     }
1928     if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1929       FastLockNode* flock = u->as_FastLock();
1930       assert(flock->box_node() == oldbox, "sanity");
1931       _igvn.rehash_node_delayed(flock);
1932       flock->set_box_node(newbox);
1933       next_edge = false;
1934     }
1935 
1936     // Replace old box in monitor debug info.
1937     if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1938       SafePointNode* sfn = u->as_SafePoint();
1939       JVMState* youngest_jvms = sfn->jvms();
1940       int max_depth = youngest_jvms->depth();
1941       for (int depth = 1; depth <= max_depth; depth++) {
1942         JVMState* jvms = youngest_jvms->of_depth(depth);
1943         int num_mon  = jvms->nof_monitors();
1944         // Loop over monitors
1945         for (int idx = 0; idx < num_mon; idx++) {
1946           Node* obj_node = sfn->monitor_obj(jvms, idx);
1947           Node* box_node = sfn->monitor_box(jvms, idx);
1948           if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1949             int j = jvms->monitor_box_offset(idx);
1950             _igvn.replace_input_of(u, j, newbox);
1951             next_edge = false;
1952           }
1953         }
1954       }
1955     }
1956     if (next_edge) i++;
1957   }
1958 }
1959 
1960 //-----------------------mark_eliminated_locking_nodes-----------------------
1961 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1962   if (EliminateNestedLocks) {
1963     if (alock->is_nested()) {
1964        assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
1965        return;
1966     } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
1967       // Only Lock node has JVMState needed here.
1968       // Not that preceding claim is documented anywhere else.
1969       if (alock->jvms() != NULL) {
1970         if (alock->as_Lock()->is_nested_lock_region()) {
1971           // Mark eliminated related nested locks and unlocks.
1972           Node* obj = alock->obj_node();
1973           BoxLockNode* box_node = alock->box_node()->as_BoxLock();
1974           assert(!box_node->is_eliminated(), "should not be marked yet");
1975           // Note: BoxLock node is marked eliminated only here
1976           // and it is used to indicate that all associated lock
1977           // and unlock nodes are marked for elimination.
1978           box_node->set_eliminated(); // Box's hash is always NO_HASH here
1979           for (uint i = 0; i < box_node->outcnt(); i++) {
1980             Node* u = box_node->raw_out(i);
1981             if (u->is_AbstractLock()) {
1982               alock = u->as_AbstractLock();
1983               if (alock->box_node() == box_node) {
1984                 // Verify that this Box is referenced only by related locks.
1985                 assert(alock->obj_node()->eqv_uncast(obj), "");
1986                 // Mark all related locks and unlocks.
1987 #ifdef ASSERT
1988                 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
1989 #endif
1990                 alock->set_nested();
1991               }
1992             }
1993           }
1994         } else {
1995 #ifdef ASSERT
1996           alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
1997           if (C->log() != NULL)
1998             alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
1999 #endif
2000         }
2001       }
2002       return;
2003     }
2004     // Process locks for non escaping object
2005     assert(alock->is_non_esc_obj(), "");
2006   } // EliminateNestedLocks
2007 
2008   if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2009     // Look for all locks of this object and mark them and
2010     // corresponding BoxLock nodes as eliminated.
2011     Node* obj = alock->obj_node();
2012     for (uint j = 0; j < obj->outcnt(); j++) {
2013       Node* o = obj->raw_out(j);
2014       if (o->is_AbstractLock() &&
2015           o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2016         alock = o->as_AbstractLock();
2017         Node* box = alock->box_node();
2018         // Replace old box node with new eliminated box for all users
2019         // of the same object and mark related locks as eliminated.
2020         mark_eliminated_box(box, obj);
2021       }
2022     }
2023   }
2024 }
2025 
2026 // we have determined that this lock/unlock can be eliminated, we simply
2027 // eliminate the node without expanding it.
2028 //
2029 // Note:  The membar's associated with the lock/unlock are currently not
2030 //        eliminated.  This should be investigated as a future enhancement.
2031 //
2032 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2033 
2034   if (!alock->is_eliminated()) {
2035     return false;
2036   }
2037 #ifdef ASSERT
2038   if (!alock->is_coarsened()) {
2039     // Check that new "eliminated" BoxLock node is created.
2040     BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2041     assert(oldbox->is_eliminated(), "should be done already");
2042   }
2043 #endif
2044 
2045   alock->log_lock_optimization(C, "eliminate_lock");
2046 
2047 #ifndef PRODUCT
2048   if (PrintEliminateLocks) {
2049     if (alock->is_Lock()) {
2050       tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2051     } else {
2052       tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2053     }
2054   }
2055 #endif
2056 
2057   Node* mem  = alock->in(TypeFunc::Memory);
2058   Node* ctrl = alock->in(TypeFunc::Control);
2059   guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2060 
2061   extract_call_projections(alock);
2062   // There are 2 projections from the lock.  The lock node will
2063   // be deleted when its last use is subsumed below.
2064   assert(alock->outcnt() == 2 &&
2065          _fallthroughproj != NULL &&
2066          _memproj_fallthrough != NULL,
2067          "Unexpected projections from Lock/Unlock");
2068 
2069   Node* fallthroughproj = _fallthroughproj;
2070   Node* memproj_fallthrough = _memproj_fallthrough;
2071 
2072   // The memory projection from a lock/unlock is RawMem
2073   // The input to a Lock is merged memory, so extract its RawMem input
2074   // (unless the MergeMem has been optimized away.)
2075   if (alock->is_Lock()) {
2076     // Seach for MemBarAcquireLock node and delete it also.
2077     MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2078     assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2079     Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2080     Node* memproj = membar->proj_out(TypeFunc::Memory);
2081     _igvn.replace_node(ctrlproj, fallthroughproj);
2082     _igvn.replace_node(memproj, memproj_fallthrough);
2083 
2084     // Delete FastLock node also if this Lock node is unique user
2085     // (a loop peeling may clone a Lock node).
2086     Node* flock = alock->as_Lock()->fastlock_node();
2087     if (flock->outcnt() == 1) {
2088       assert(flock->unique_out() == alock, "sanity");
2089       _igvn.replace_node(flock, top());
2090     }
2091   }
2092 
2093   // Seach for MemBarReleaseLock node and delete it also.
2094   if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2095     MemBarNode* membar = ctrl->in(0)->as_MemBar();
2096     assert(membar->Opcode() == Op_MemBarReleaseLock &&
2097            mem->is_Proj() && membar == mem->in(0), "");
2098     _igvn.replace_node(fallthroughproj, ctrl);
2099     _igvn.replace_node(memproj_fallthrough, mem);
2100     fallthroughproj = ctrl;
2101     memproj_fallthrough = mem;
2102     ctrl = membar->in(TypeFunc::Control);
2103     mem  = membar->in(TypeFunc::Memory);
2104   }
2105 
2106   _igvn.replace_node(fallthroughproj, ctrl);
2107   _igvn.replace_node(memproj_fallthrough, mem);
2108   return true;
2109 }
2110 
2111 
2112 //------------------------------expand_lock_node----------------------
2113 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2114 
2115   Node* ctrl = lock->in(TypeFunc::Control);
2116   Node* mem = lock->in(TypeFunc::Memory);
2117   Node* obj = lock->obj_node();
2118   Node* box = lock->box_node();
2119   Node* flock = lock->fastlock_node();
2120 
2121   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2122 
2123   // Make the merge point
2124   Node *region;
2125   Node *mem_phi;
2126   Node *slow_path;
2127 
2128   if (UseOptoBiasInlining) {
2129     /*
2130      *  See the full description in MacroAssembler::biased_locking_enter().
2131      *
2132      *  if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2133      *    // The object is biased.
2134      *    proto_node = klass->prototype_header;
2135      *    o_node = thread | proto_node;
2136      *    x_node = o_node ^ mark_word;
2137      *    if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2138      *      // Done.
2139      *    } else {
2140      *      if( (x_node & biased_lock_mask) != 0 ) {
2141      *        // The klass's prototype header is no longer biased.
2142      *        cas(&mark_word, mark_word, proto_node)
2143      *        goto cas_lock;
2144      *      } else {
2145      *        // The klass's prototype header is still biased.
2146      *        if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2147      *          old = mark_word;
2148      *          new = o_node;
2149      *        } else {
2150      *          // Different thread or anonymous biased.
2151      *          old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2152      *          new = thread | old;
2153      *        }
2154      *        // Try to rebias.
2155      *        if( cas(&mark_word, old, new) == 0 ) {
2156      *          // Done.
2157      *        } else {
2158      *          goto slow_path; // Failed.
2159      *        }
2160      *      }
2161      *    }
2162      *  } else {
2163      *    // The object is not biased.
2164      *    cas_lock:
2165      *    if( FastLock(obj) == 0 ) {
2166      *      // Done.
2167      *    } else {
2168      *      slow_path:
2169      *      OptoRuntime::complete_monitor_locking_Java(obj);
2170      *    }
2171      *  }
2172      */
2173 
2174     region  = new RegionNode(5);
2175     // create a Phi for the memory state
2176     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2177 
2178     Node* fast_lock_region  = new RegionNode(3);
2179     Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2180 
2181     // First, check mark word for the biased lock pattern.
2182     Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2183 
2184     // Get fast path - mark word has the biased lock pattern.
2185     ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2186                          markOopDesc::biased_lock_mask_in_place,
2187                          markOopDesc::biased_lock_pattern, true);
2188     // fast_lock_region->in(1) is set to slow path.
2189     fast_lock_mem_phi->init_req(1, mem);
2190 
2191     // Now check that the lock is biased to the current thread and has
2192     // the same epoch and bias as Klass::_prototype_header.
2193 
2194     // Special-case a fresh allocation to avoid building nodes:
2195     Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2196     if (klass_node == NULL) {
2197       Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2198       klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2199 #ifdef _LP64
2200       if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2201         assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2202         klass_node->in(1)->init_req(0, ctrl);
2203       } else
2204 #endif
2205       klass_node->init_req(0, ctrl);
2206     }
2207     Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2208 
2209     Node* thread = transform_later(new ThreadLocalNode());
2210     Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2211     Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2212     Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2213 
2214     // Get slow path - mark word does NOT match the value.
2215     Node* not_biased_ctrl =  opt_bits_test(ctrl, region, 3, x_node,
2216                                       (~markOopDesc::age_mask_in_place), 0);
2217     // region->in(3) is set to fast path - the object is biased to the current thread.
2218     mem_phi->init_req(3, mem);
2219 
2220 
2221     // Mark word does NOT match the value (thread | Klass::_prototype_header).
2222 
2223 
2224     // First, check biased pattern.
2225     // Get fast path - _prototype_header has the same biased lock pattern.
2226     ctrl =  opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2227                           markOopDesc::biased_lock_mask_in_place, 0, true);
2228 
2229     not_biased_ctrl = fast_lock_region->in(2); // Slow path
2230     // fast_lock_region->in(2) - the prototype header is no longer biased
2231     // and we have to revoke the bias on this object.
2232     // We are going to try to reset the mark of this object to the prototype
2233     // value and fall through to the CAS-based locking scheme.
2234     Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2235     Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2236                                           proto_node, mark_node);
2237     transform_later(cas);
2238     Node* proj = transform_later(new SCMemProjNode(cas));
2239     fast_lock_mem_phi->init_req(2, proj);
2240 
2241 
2242     // Second, check epoch bits.
2243     Node* rebiased_region  = new RegionNode(3);
2244     Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2245     Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2246 
2247     // Get slow path - mark word does NOT match epoch bits.
2248     Node* epoch_ctrl =  opt_bits_test(ctrl, rebiased_region, 1, x_node,
2249                                       markOopDesc::epoch_mask_in_place, 0);
2250     // The epoch of the current bias is not valid, attempt to rebias the object
2251     // toward the current thread.
2252     rebiased_region->init_req(2, epoch_ctrl);
2253     old_phi->init_req(2, mark_node);
2254     new_phi->init_req(2, o_node);
2255 
2256     // rebiased_region->in(1) is set to fast path.
2257     // The epoch of the current bias is still valid but we know
2258     // nothing about the owner; it might be set or it might be clear.
2259     Node* cmask   = MakeConX(markOopDesc::biased_lock_mask_in_place |
2260                              markOopDesc::age_mask_in_place |
2261                              markOopDesc::epoch_mask_in_place);
2262     Node* old = transform_later(new AndXNode(mark_node, cmask));
2263     cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2264     Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2265     old_phi->init_req(1, old);
2266     new_phi->init_req(1, new_mark);
2267 
2268     transform_later(rebiased_region);
2269     transform_later(old_phi);
2270     transform_later(new_phi);
2271 
2272     // Try to acquire the bias of the object using an atomic operation.
2273     // If this fails we will go in to the runtime to revoke the object's bias.
2274     cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2275     transform_later(cas);
2276     proj = transform_later(new SCMemProjNode(cas));
2277 
2278     // Get slow path - Failed to CAS.
2279     not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2280     mem_phi->init_req(4, proj);
2281     // region->in(4) is set to fast path - the object is rebiased to the current thread.
2282 
2283     // Failed to CAS.
2284     slow_path  = new RegionNode(3);
2285     Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2286 
2287     slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2288     slow_mem->init_req(1, proj);
2289 
2290     // Call CAS-based locking scheme (FastLock node).
2291 
2292     transform_later(fast_lock_region);
2293     transform_later(fast_lock_mem_phi);
2294 
2295     // Get slow path - FastLock failed to lock the object.
2296     ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2297     mem_phi->init_req(2, fast_lock_mem_phi);
2298     // region->in(2) is set to fast path - the object is locked to the current thread.
2299 
2300     slow_path->init_req(2, ctrl); // Capture slow-control
2301     slow_mem->init_req(2, fast_lock_mem_phi);
2302 
2303     transform_later(slow_path);
2304     transform_later(slow_mem);
2305     // Reset lock's memory edge.
2306     lock->set_req(TypeFunc::Memory, slow_mem);
2307 
2308   } else {
2309     region  = new RegionNode(3);
2310     // create a Phi for the memory state
2311     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2312 
2313     // Optimize test; set region slot 2
2314     slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2315     mem_phi->init_req(2, mem);
2316   }
2317 
2318   // Make slow path call
2319   CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2320                                   OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2321                                   obj, box, NULL);
2322 
2323   extract_call_projections(call);
2324 
2325   // Slow path can only throw asynchronous exceptions, which are always
2326   // de-opted.  So the compiler thinks the slow-call can never throw an
2327   // exception.  If it DOES throw an exception we would need the debug
2328   // info removed first (since if it throws there is no monitor).
2329   assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2330            _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2331 
2332   // Capture slow path
2333   // disconnect fall-through projection from call and create a new one
2334   // hook up users of fall-through projection to region
2335   Node *slow_ctrl = _fallthroughproj->clone();
2336   transform_later(slow_ctrl);
2337   _igvn.hash_delete(_fallthroughproj);
2338   _fallthroughproj->disconnect_inputs(NULL, C);
2339   region->init_req(1, slow_ctrl);
2340   // region inputs are now complete
2341   transform_later(region);
2342   _igvn.replace_node(_fallthroughproj, region);
2343 
2344   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2345   mem_phi->init_req(1, memproj );
2346   transform_later(mem_phi);
2347   _igvn.replace_node(_memproj_fallthrough, mem_phi);
2348 }
2349 
2350 //------------------------------expand_unlock_node----------------------
2351 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2352 
2353   Node* ctrl = unlock->in(TypeFunc::Control);
2354   Node* mem = unlock->in(TypeFunc::Memory);
2355   Node* obj = unlock->obj_node();
2356   Node* box = unlock->box_node();
2357 
2358   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2359 
2360   // No need for a null check on unlock
2361 
2362   // Make the merge point
2363   Node *region;
2364   Node *mem_phi;
2365 
2366   if (UseOptoBiasInlining) {
2367     // Check for biased locking unlock case, which is a no-op.
2368     // See the full description in MacroAssembler::biased_locking_exit().
2369     region  = new RegionNode(4);
2370     // create a Phi for the memory state
2371     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2372     mem_phi->init_req(3, mem);
2373 
2374     Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2375     ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2376                          markOopDesc::biased_lock_mask_in_place,
2377                          markOopDesc::biased_lock_pattern);
2378   } else {
2379     region  = new RegionNode(3);
2380     // create a Phi for the memory state
2381     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2382   }
2383 
2384   FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2385   funlock = transform_later( funlock )->as_FastUnlock();
2386   // Optimize test; set region slot 2
2387   Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2388   Node *thread = transform_later(new ThreadLocalNode());
2389 
2390   CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2391                                   CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2392                                   "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2393 
2394   extract_call_projections(call);
2395 
2396   assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2397            _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2398 
2399   // No exceptions for unlocking
2400   // Capture slow path
2401   // disconnect fall-through projection from call and create a new one
2402   // hook up users of fall-through projection to region
2403   Node *slow_ctrl = _fallthroughproj->clone();
2404   transform_later(slow_ctrl);
2405   _igvn.hash_delete(_fallthroughproj);
2406   _fallthroughproj->disconnect_inputs(NULL, C);
2407   region->init_req(1, slow_ctrl);
2408   // region inputs are now complete
2409   transform_later(region);
2410   _igvn.replace_node(_fallthroughproj, region);
2411 
2412   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2413   mem_phi->init_req(1, memproj );
2414   mem_phi->init_req(2, mem);
2415   transform_later(mem_phi);
2416   _igvn.replace_node(_memproj_fallthrough, mem_phi);
2417 }
2418 
2419 //---------------------------eliminate_macro_nodes----------------------
2420 // Eliminate scalar replaced allocations and associated locks.
2421 void PhaseMacroExpand::eliminate_macro_nodes() {
2422   if (C->macro_count() == 0)
2423     return;
2424 
2425   // First, attempt to eliminate locks
2426   int cnt = C->macro_count();
2427   for (int i=0; i < cnt; i++) {
2428     Node *n = C->macro_node(i);
2429     if (n->is_AbstractLock()) { // Lock and Unlock nodes
2430       // Before elimination mark all associated (same box and obj)
2431       // lock and unlock nodes.
2432       mark_eliminated_locking_nodes(n->as_AbstractLock());
2433     }
2434   }
2435   bool progress = true;
2436   while (progress) {
2437     progress = false;
2438     for (int i = C->macro_count(); i > 0; i--) {
2439       Node * n = C->macro_node(i-1);
2440       bool success = false;
2441       debug_only(int old_macro_count = C->macro_count(););
2442       if (n->is_AbstractLock()) {
2443         success = eliminate_locking_node(n->as_AbstractLock());
2444       }
2445       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2446       progress = progress || success;
2447     }
2448   }
2449   // Next, attempt to eliminate allocations
2450   _has_locks = false;
2451   progress = true;
2452   while (progress) {
2453     progress = false;
2454     for (int i = C->macro_count(); i > 0; i--) {
2455       Node * n = C->macro_node(i-1);
2456       bool success = false;
2457       debug_only(int old_macro_count = C->macro_count(););
2458       switch (n->class_id()) {
2459       case Node::Class_Allocate:
2460       case Node::Class_AllocateArray:
2461         success = eliminate_allocate_node(n->as_Allocate());
2462         break;
2463       case Node::Class_CallStaticJava:
2464         success = eliminate_boxing_node(n->as_CallStaticJava());
2465         break;
2466       case Node::Class_Lock:
2467       case Node::Class_Unlock:
2468         assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2469         _has_locks = true;
2470         break;
2471       case Node::Class_ArrayCopy:
2472         break;
2473       case Node::Class_OuterStripMinedLoop:
2474         break;
2475       default:
2476         assert(n->Opcode() == Op_LoopLimit ||
2477                n->Opcode() == Op_Opaque1   ||
2478                n->Opcode() == Op_Opaque2   ||
2479                n->Opcode() == Op_Opaque3   ||
2480                BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2481                "unknown node type in macro list");
2482       }
2483       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2484       progress = progress || success;
2485     }
2486   }
2487 }
2488 
2489 //------------------------------expand_macro_nodes----------------------
2490 //  Returns true if a failure occurred.
2491 bool PhaseMacroExpand::expand_macro_nodes() {
2492   // Last attempt to eliminate macro nodes.
2493   eliminate_macro_nodes();
2494 
2495   // Make sure expansion will not cause node limit to be exceeded.
2496   // Worst case is a macro node gets expanded into about 200 nodes.
2497   // Allow 50% more for optimization.
2498   if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2499     return true;
2500 
2501   // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2502   bool progress = true;
2503   while (progress) {
2504     progress = false;
2505     for (int i = C->macro_count(); i > 0; i--) {
2506       Node * n = C->macro_node(i-1);
2507       bool success = false;
2508       debug_only(int old_macro_count = C->macro_count(););
2509       if (n->Opcode() == Op_LoopLimit) {
2510         // Remove it from macro list and put on IGVN worklist to optimize.
2511         C->remove_macro_node(n);
2512         _igvn._worklist.push(n);
2513         success = true;
2514       } else if (n->Opcode() == Op_CallStaticJava) {
2515         // Remove it from macro list and put on IGVN worklist to optimize.
2516         C->remove_macro_node(n);
2517         _igvn._worklist.push(n);
2518         success = true;
2519       } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2520         _igvn.replace_node(n, n->in(1));
2521         success = true;
2522 #if INCLUDE_RTM_OPT
2523       } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2524         assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2525         assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2526         Node* cmp = n->unique_out();
2527 #ifdef ASSERT
2528         // Validate graph.
2529         assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2530         BoolNode* bol = cmp->unique_out()->as_Bool();
2531         assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2532                (bol->_test._test == BoolTest::ne), "");
2533         IfNode* ifn = bol->unique_out()->as_If();
2534         assert((ifn->outcnt() == 2) &&
2535                ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2536 #endif
2537         Node* repl = n->in(1);
2538         if (!_has_locks) {
2539           // Remove RTM state check if there are no locks in the code.
2540           // Replace input to compare the same value.
2541           repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2542         }
2543         _igvn.replace_node(n, repl);
2544         success = true;
2545 #endif
2546       } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2547         n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2548         C->remove_macro_node(n);
2549         success = true;
2550       }
2551       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2552       progress = progress || success;
2553     }
2554   }
2555 
2556   // expand arraycopy "macro" nodes first
2557   // For ReduceBulkZeroing, we must first process all arraycopy nodes
2558   // before the allocate nodes are expanded.
2559   int macro_idx = C->macro_count() - 1;
2560   while (macro_idx >= 0) {
2561     Node * n = C->macro_node(macro_idx);
2562     assert(n->is_macro(), "only macro nodes expected here");
2563     if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2564       // node is unreachable, so don't try to expand it
2565       C->remove_macro_node(n);
2566     } else if (n->is_ArrayCopy()){
2567       int macro_count = C->macro_count();
2568       expand_arraycopy_node(n->as_ArrayCopy());
2569       assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2570     }
2571     if (C->failing())  return true;
2572     macro_idx --;
2573   }
2574 
2575   // expand "macro" nodes
2576   // nodes are removed from the macro list as they are processed
2577   while (C->macro_count() > 0) {
2578     int macro_count = C->macro_count();
2579     Node * n = C->macro_node(macro_count-1);
2580     assert(n->is_macro(), "only macro nodes expected here");
2581     if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2582       // node is unreachable, so don't try to expand it
2583       C->remove_macro_node(n);
2584       continue;
2585     }
2586     switch (n->class_id()) {
2587     case Node::Class_Allocate:
2588       expand_allocate(n->as_Allocate());
2589       break;
2590     case Node::Class_AllocateArray:
2591       expand_allocate_array(n->as_AllocateArray());
2592       break;
2593     case Node::Class_Lock:
2594       expand_lock_node(n->as_Lock());
2595       break;
2596     case Node::Class_Unlock:
2597       expand_unlock_node(n->as_Unlock());
2598       break;
2599     default:
2600       assert(false, "unknown node type in macro list");
2601     }
2602     assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2603     if (C->failing())  return true;
2604   }
2605 
2606   _igvn.set_delay_transform(false);
2607   _igvn.optimize();
2608   if (C->failing())  return true;
2609   return false;
2610 }