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 "opto/addnode.hpp"
  30 #include "opto/arraycopynode.hpp"
  31 #include "opto/callnode.hpp"
  32 #include "opto/castnode.hpp"
  33 #include "opto/cfgnode.hpp"
  34 #include "opto/compile.hpp"
  35 #include "opto/convertnode.hpp"
  36 #include "opto/graphKit.hpp"
  37 #include "opto/locknode.hpp"
  38 #include "opto/loopnode.hpp"
  39 #include "opto/macro.hpp"
  40 #include "opto/memnode.hpp"
  41 #include "opto/narrowptrnode.hpp"
  42 #include "opto/node.hpp"
  43 #include "opto/opaquenode.hpp"
  44 #include "opto/phaseX.hpp"
  45 #include "opto/rootnode.hpp"
  46 #include "opto/runtime.hpp"
  47 #include "opto/subnode.hpp"
  48 #include "opto/type.hpp"
  49 #include "runtime/sharedRuntime.hpp"
  50 #include "utilities/macros.hpp"
  51 #if INCLUDE_G1GC
  52 #include "gc/g1/g1ThreadLocalData.hpp"
  53 #endif // INCLUDE_G1GC
  54 #if INCLUDE_SHENANDOAHGC
  55 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  56 #endif
  57 
  58 
  59 //
  60 // Replace any references to "oldref" in inputs to "use" with "newref".
  61 // Returns the number of replacements made.
  62 //
  63 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
  64   int nreplacements = 0;
  65   uint req = use->req();
  66   for (uint j = 0; j < use->len(); j++) {
  67     Node *uin = use->in(j);
  68     if (uin == oldref) {
  69       if (j < req)
  70         use->set_req(j, newref);
  71       else
  72         use->set_prec(j, newref);
  73       nreplacements++;
  74     } else if (j >= req && uin == NULL) {
  75       break;
  76     }
  77   }
  78   return nreplacements;
  79 }
  80 
  81 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
  82   // Copy debug information and adjust JVMState information
  83   uint old_dbg_start = oldcall->tf()->domain()->cnt();
  84   uint new_dbg_start = newcall->tf()->domain()->cnt();
  85   int jvms_adj  = new_dbg_start - old_dbg_start;
  86   assert (new_dbg_start == newcall->req(), "argument count mismatch");
  87 
  88   // SafePointScalarObject node could be referenced several times in debug info.
  89   // Use Dict to record cloned nodes.
  90   Dict* sosn_map = new Dict(cmpkey,hashkey);
  91   for (uint i = old_dbg_start; i < oldcall->req(); i++) {
  92     Node* old_in = oldcall->in(i);
  93     // Clone old SafePointScalarObjectNodes, adjusting their field contents.
  94     if (old_in != NULL && old_in->is_SafePointScalarObject()) {
  95       SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
  96       uint old_unique = C->unique();
  97       Node* new_in = old_sosn->clone(sosn_map);
  98       if (old_unique != C->unique()) { // New node?
  99         new_in->set_req(0, C->root()); // reset control edge
 100         new_in = transform_later(new_in); // Register new node.
 101       }
 102       old_in = new_in;
 103     }
 104     newcall->add_req(old_in);
 105   }
 106 
 107   // JVMS may be shared so clone it before we modify it
 108   newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
 109   for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
 110     jvms->set_map(newcall);
 111     jvms->set_locoff(jvms->locoff()+jvms_adj);
 112     jvms->set_stkoff(jvms->stkoff()+jvms_adj);
 113     jvms->set_monoff(jvms->monoff()+jvms_adj);
 114     jvms->set_scloff(jvms->scloff()+jvms_adj);
 115     jvms->set_endoff(jvms->endoff()+jvms_adj);
 116   }
 117 }
 118 
 119 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
 120   Node* cmp;
 121   if (mask != 0) {
 122     Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
 123     cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
 124   } else {
 125     cmp = word;
 126   }
 127   Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
 128   IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
 129   transform_later(iff);
 130 
 131   // Fast path taken.
 132   Node *fast_taken = transform_later(new IfFalseNode(iff));
 133 
 134   // Fast path not-taken, i.e. slow path
 135   Node *slow_taken = transform_later(new IfTrueNode(iff));
 136 
 137   if (return_fast_path) {
 138     region->init_req(edge, slow_taken); // Capture slow-control
 139     return fast_taken;
 140   } else {
 141     region->init_req(edge, fast_taken); // Capture fast-control
 142     return slow_taken;
 143   }
 144 }
 145 
 146 //--------------------copy_predefined_input_for_runtime_call--------------------
 147 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
 148   // Set fixed predefined input arguments
 149   call->init_req( TypeFunc::Control, ctrl );
 150   call->init_req( TypeFunc::I_O    , oldcall->in( TypeFunc::I_O) );
 151   call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
 152   call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
 153   call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
 154 }
 155 
 156 //------------------------------make_slow_call---------------------------------
 157 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
 158                                            address slow_call, const char* leaf_name, Node* slow_path,
 159                                            Node* parm0, Node* parm1, Node* parm2) {
 160 
 161   // Slow-path call
 162  CallNode *call = leaf_name
 163    ? (CallNode*)new CallLeafNode      ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
 164    : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
 165 
 166   // Slow path call has no side-effects, uses few values
 167   copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
 168   if (parm0 != NULL)  call->init_req(TypeFunc::Parms+0, parm0);
 169   if (parm1 != NULL)  call->init_req(TypeFunc::Parms+1, parm1);
 170   if (parm2 != NULL)  call->init_req(TypeFunc::Parms+2, parm2);
 171   copy_call_debug_info(oldcall, call);
 172   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
 173   _igvn.replace_node(oldcall, call);
 174   transform_later(call);
 175 
 176   return call;
 177 }
 178 
 179 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
 180   _fallthroughproj = NULL;
 181   _fallthroughcatchproj = NULL;
 182   _ioproj_fallthrough = NULL;
 183   _ioproj_catchall = NULL;
 184   _catchallcatchproj = NULL;
 185   _memproj_fallthrough = NULL;
 186   _memproj_catchall = NULL;
 187   _resproj = NULL;
 188   for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
 189     ProjNode *pn = call->fast_out(i)->as_Proj();
 190     switch (pn->_con) {
 191       case TypeFunc::Control:
 192       {
 193         // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
 194         _fallthroughproj = pn;
 195         DUIterator_Fast jmax, j = pn->fast_outs(jmax);
 196         const Node *cn = pn->fast_out(j);
 197         if (cn->is_Catch()) {
 198           ProjNode *cpn = NULL;
 199           for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
 200             cpn = cn->fast_out(k)->as_Proj();
 201             assert(cpn->is_CatchProj(), "must be a CatchProjNode");
 202             if (cpn->_con == CatchProjNode::fall_through_index)
 203               _fallthroughcatchproj = cpn;
 204             else {
 205               assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
 206               _catchallcatchproj = cpn;
 207             }
 208           }
 209         }
 210         break;
 211       }
 212       case TypeFunc::I_O:
 213         if (pn->_is_io_use)
 214           _ioproj_catchall = pn;
 215         else
 216           _ioproj_fallthrough = pn;
 217         break;
 218       case TypeFunc::Memory:
 219         if (pn->_is_io_use)
 220           _memproj_catchall = pn;
 221         else
 222           _memproj_fallthrough = pn;
 223         break;
 224       case TypeFunc::Parms:
 225         _resproj = pn;
 226         break;
 227       default:
 228         assert(false, "unexpected projection from allocation node.");
 229     }
 230   }
 231 
 232 }
 233 
 234 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
 235   BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
 236   bs->eliminate_gc_barrier(this, p2x);
 237 }
 238 
 239 // Search for a memory operation for the specified memory slice.
 240 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
 241   Node *orig_mem = mem;
 242   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 243   const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
 244   while (true) {
 245     if (mem == alloc_mem || mem == start_mem ) {
 246       return mem;  // hit one of our sentinels
 247     } else if (mem->is_MergeMem()) {
 248       mem = mem->as_MergeMem()->memory_at(alias_idx);
 249     } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
 250       Node *in = mem->in(0);
 251       // we can safely skip over safepoints, calls, locks and membars because we
 252       // already know that the object is safe to eliminate.
 253       if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
 254         return in;
 255       } else if (in->is_Call()) {
 256         CallNode *call = in->as_Call();
 257         if (call->may_modify(tinst, phase)) {
 258           assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
 259           if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
 260             return in;
 261           }
 262         }
 263         mem = in->in(TypeFunc::Memory);
 264       } else if (in->is_MemBar()) {
 265         ArrayCopyNode* ac = NULL;
 266         if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
 267           assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
 268           return ac;
 269         }
 270         mem = in->in(TypeFunc::Memory);
 271       } else {
 272         assert(false, "unexpected projection");
 273       }
 274     } else if (mem->is_Store()) {
 275       const TypePtr* atype = mem->as_Store()->adr_type();
 276       int adr_idx = phase->C->get_alias_index(atype);
 277       if (adr_idx == alias_idx) {
 278         assert(atype->isa_oopptr(), "address type must be oopptr");
 279         int adr_offset = atype->offset();
 280         uint adr_iid = atype->is_oopptr()->instance_id();
 281         // Array elements references have the same alias_idx
 282         // but different offset and different instance_id.
 283         if (adr_offset == offset && adr_iid == alloc->_idx)
 284           return mem;
 285       } else {
 286         assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
 287       }
 288       mem = mem->in(MemNode::Memory);
 289     } else if (mem->is_ClearArray()) {
 290       if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
 291         // Can not bypass initialization of the instance
 292         // we are looking.
 293         debug_only(intptr_t offset;)
 294         assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
 295         InitializeNode* init = alloc->as_Allocate()->initialization();
 296         // We are looking for stored value, return Initialize node
 297         // or memory edge from Allocate node.
 298         if (init != NULL)
 299           return init;
 300         else
 301           return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
 302       }
 303       // Otherwise skip it (the call updated 'mem' value).
 304     } else if (mem->Opcode() == Op_SCMemProj) {
 305       mem = mem->in(0);
 306       Node* adr = NULL;
 307       if (mem->is_LoadStore()) {
 308         adr = mem->in(MemNode::Address);
 309       } else {
 310         assert(mem->Opcode() == Op_EncodeISOArray ||
 311                mem->Opcode() == Op_StrCompressedCopy, "sanity");
 312         adr = mem->in(3); // Destination array
 313       }
 314       const TypePtr* atype = adr->bottom_type()->is_ptr();
 315       int adr_idx = phase->C->get_alias_index(atype);
 316       if (adr_idx == alias_idx) {
 317         DEBUG_ONLY(mem->dump();)
 318         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 319         return NULL;
 320       }
 321       mem = mem->in(MemNode::Memory);
 322    } else if (mem->Opcode() == Op_StrInflatedCopy) {
 323       Node* adr = mem->in(3); // Destination array
 324       const TypePtr* atype = adr->bottom_type()->is_ptr();
 325       int adr_idx = phase->C->get_alias_index(atype);
 326       if (adr_idx == alias_idx) {
 327         DEBUG_ONLY(mem->dump();)
 328         assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
 329         return NULL;
 330       }
 331       mem = mem->in(MemNode::Memory);
 332     } else {
 333       return mem;
 334     }
 335     assert(mem != orig_mem, "dead memory loop");
 336   }
 337 }
 338 
 339 // Generate loads from source of the arraycopy for fields of
 340 // destination needed at a deoptimization point
 341 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
 342   BasicType bt = ft;
 343   const Type *type = ftype;
 344   if (ft == T_NARROWOOP) {
 345     bt = T_OBJECT;
 346     type = ftype->make_oopptr();
 347   }
 348   Node* res = NULL;
 349   if (ac->is_clonebasic()) {
 350     Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
 351     Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
 352     const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
 353     res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
 354   } else {
 355     if (ac->modifies(offset, offset, &_igvn, true)) {
 356       assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
 357       uint shift  = exact_log2(type2aelembytes(bt));
 358       Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 359 #ifdef _LP64
 360       diff = _igvn.transform(new ConvI2LNode(diff));
 361 #endif
 362       diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
 363 
 364       Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
 365       Node* base = ac->in(ArrayCopyNode::Src);
 366       Node* adr = _igvn.transform(new AddPNode(base, base, off));
 367       const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
 368       res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
 369     }
 370   }
 371   if (res != NULL) {
 372     res = _igvn.transform(res);
 373     if (ftype->isa_narrowoop()) {
 374       // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
 375       res = _igvn.transform(new EncodePNode(res, ftype));
 376     }
 377     return res;
 378   }
 379   return NULL;
 380 }
 381 
 382 //
 383 // Given a Memory Phi, compute a value Phi containing the values from stores
 384 // on the input paths.
 385 // Note: this function is recursive, its depth is limited by the "level" argument
 386 // Returns the computed Phi, or NULL if it cannot compute it.
 387 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) {
 388   assert(mem->is_Phi(), "sanity");
 389   int alias_idx = C->get_alias_index(adr_t);
 390   int offset = adr_t->offset();
 391   int instance_id = adr_t->instance_id();
 392 
 393   // Check if an appropriate value phi already exists.
 394   Node* region = mem->in(0);
 395   for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
 396     Node* phi = region->fast_out(k);
 397     if (phi->is_Phi() && phi != mem &&
 398         phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
 399       return phi;
 400     }
 401   }
 402   // Check if an appropriate new value phi already exists.
 403   Node* new_phi = value_phis->find(mem->_idx);
 404   if (new_phi != NULL)
 405     return new_phi;
 406 
 407   if (level <= 0) {
 408     return NULL; // Give up: phi tree too deep
 409   }
 410   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 411   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 412 
 413   uint length = mem->req();
 414   GrowableArray <Node *> values(length, length, NULL, false);
 415 
 416   // create a new Phi for the value
 417   PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
 418   transform_later(phi);
 419   value_phis->push(phi, mem->_idx);
 420 
 421   for (uint j = 1; j < length; j++) {
 422     Node *in = mem->in(j);
 423     if (in == NULL || in->is_top()) {
 424       values.at_put(j, in);
 425     } else  {
 426       Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
 427       if (val == start_mem || val == alloc_mem) {
 428         // hit a sentinel, return appropriate 0 value
 429         values.at_put(j, _igvn.zerocon(ft));
 430         continue;
 431       }
 432       if (val->is_Initialize()) {
 433         val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 434       }
 435       if (val == NULL) {
 436         return NULL;  // can't find a value on this path
 437       }
 438       if (val == mem) {
 439         values.at_put(j, mem);
 440       } else if (val->is_Store()) {
 441         Node* n = val->in(MemNode::ValueIn);
 442 #if INCLUDE_SHENANDOAHGC
 443         if (UseShenandoahGC) {
 444           BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 445           n = bs->step_over_gc_barrier(n);
 446         }
 447 #endif
 448         values.at_put(j, n);
 449       } else if(val->is_Proj() && val->in(0) == alloc) {
 450         values.at_put(j, _igvn.zerocon(ft));
 451       } else if (val->is_Phi()) {
 452         val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
 453         if (val == NULL) {
 454           return NULL;
 455         }
 456         values.at_put(j, val);
 457       } else if (val->Opcode() == Op_SCMemProj) {
 458         assert(val->in(0)->is_LoadStore() ||
 459                val->in(0)->Opcode() == Op_EncodeISOArray ||
 460                val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
 461         assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
 462         return NULL;
 463       } else if (val->is_ArrayCopy()) {
 464         Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
 465         if (res == NULL) {
 466           return NULL;
 467         }
 468         values.at_put(j, res);
 469       } else {
 470 #ifdef ASSERT
 471         val->dump();
 472         assert(false, "unknown node on this path");
 473 #endif
 474         return NULL;  // unknown node on this path
 475       }
 476     }
 477   }
 478   // Set Phi's inputs
 479   for (uint j = 1; j < length; j++) {
 480     if (values.at(j) == mem) {
 481       phi->init_req(j, phi);
 482     } else {
 483       phi->init_req(j, values.at(j));
 484     }
 485   }
 486   return phi;
 487 }
 488 
 489 // Search the last value stored into the object's field.
 490 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
 491   assert(adr_t->is_known_instance_field(), "instance required");
 492   int instance_id = adr_t->instance_id();
 493   assert((uint)instance_id == alloc->_idx, "wrong allocation");
 494 
 495   int alias_idx = C->get_alias_index(adr_t);
 496   int offset = adr_t->offset();
 497   Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
 498   Node *alloc_ctrl = alloc->in(TypeFunc::Control);
 499   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 500   Arena *a = Thread::current()->resource_area();
 501   VectorSet visited(a);
 502 
 503 
 504   bool done = sfpt_mem == alloc_mem;
 505   Node *mem = sfpt_mem;
 506   while (!done) {
 507     if (visited.test_set(mem->_idx)) {
 508       return NULL;  // found a loop, give up
 509     }
 510     mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
 511     if (mem == start_mem || mem == alloc_mem) {
 512       done = true;  // hit a sentinel, return appropriate 0 value
 513     } else if (mem->is_Initialize()) {
 514       mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 515       if (mem == NULL) {
 516         done = true; // Something go wrong.
 517       } else if (mem->is_Store()) {
 518         const TypePtr* atype = mem->as_Store()->adr_type();
 519         assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
 520         done = true;
 521       }
 522     } else if (mem->is_Store()) {
 523       const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
 524       assert(atype != NULL, "address type must be oopptr");
 525       assert(C->get_alias_index(atype) == alias_idx &&
 526              atype->is_known_instance_field() && atype->offset() == offset &&
 527              atype->instance_id() == instance_id, "store is correct memory slice");
 528       done = true;
 529     } else if (mem->is_Phi()) {
 530       // try to find a phi's unique input
 531       Node *unique_input = NULL;
 532       Node *top = C->top();
 533       for (uint i = 1; i < mem->req(); i++) {
 534         Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
 535         if (n == NULL || n == top || n == mem) {
 536           continue;
 537         } else if (unique_input == NULL) {
 538           unique_input = n;
 539         } else if (unique_input != n) {
 540           unique_input = top;
 541           break;
 542         }
 543       }
 544       if (unique_input != NULL && unique_input != top) {
 545         mem = unique_input;
 546       } else {
 547         done = true;
 548       }
 549     } else if (mem->is_ArrayCopy()) {
 550       done = true;
 551     } else {
 552       assert(false, "unexpected node");
 553     }
 554   }
 555   if (mem != NULL) {
 556     if (mem == start_mem || mem == alloc_mem) {
 557       // hit a sentinel, return appropriate 0 value
 558       return _igvn.zerocon(ft);
 559     } else if (mem->is_Store()) {
 560       Node* n = mem->in(MemNode::ValueIn);
 561 #if INCLUDE_SHENANDOAHGC
 562       if (UseShenandoahGC) {
 563         BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 564         n = bs->step_over_gc_barrier(n);
 565       }
 566 #endif
 567       return n;
 568     } else if (mem->is_Phi()) {
 569       // attempt to produce a Phi reflecting the values on the input paths of the Phi
 570       Node_Stack value_phis(a, 8);
 571       Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
 572       if (phi != NULL) {
 573         return phi;
 574       } else {
 575         // Kill all new Phis
 576         while(value_phis.is_nonempty()) {
 577           Node* n = value_phis.node();
 578           _igvn.replace_node(n, C->top());
 579           value_phis.pop();
 580         }
 581       }
 582     } else if (mem->is_ArrayCopy()) {
 583       Node* ctl = mem->in(0);
 584       Node* m = mem->in(TypeFunc::Memory);
 585       if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
 586         // pin the loads in the uncommon trap path
 587         ctl = sfpt_ctl;
 588         m = sfpt_mem;
 589       }
 590       return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
 591     }
 592   }
 593   // Something go wrong.
 594   return NULL;
 595 }
 596 
 597 // Check the possibility of scalar replacement.
 598 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
 599   //  Scan the uses of the allocation to check for anything that would
 600   //  prevent us from eliminating it.
 601   NOT_PRODUCT( const char* fail_eliminate = NULL; )
 602   DEBUG_ONLY( Node* disq_node = NULL; )
 603   bool  can_eliminate = true;
 604 
 605   Node* res = alloc->result_cast();
 606   const TypeOopPtr* res_type = NULL;
 607   if (res == NULL) {
 608     // All users were eliminated.
 609   } else if (!res->is_CheckCastPP()) {
 610     NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
 611     can_eliminate = false;
 612   } else {
 613     res_type = _igvn.type(res)->isa_oopptr();
 614     if (res_type == NULL) {
 615       NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
 616       can_eliminate = false;
 617     } else if (res_type->isa_aryptr()) {
 618       int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 619       if (length < 0) {
 620         NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
 621         can_eliminate = false;
 622       }
 623     }
 624   }
 625 
 626   if (can_eliminate && res != NULL) {
 627     for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
 628                                j < jmax && can_eliminate; j++) {
 629       Node* use = res->fast_out(j);
 630 
 631       if (use->is_AddP()) {
 632         const TypePtr* addp_type = _igvn.type(use)->is_ptr();
 633         int offset = addp_type->offset();
 634 
 635         if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
 636           NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
 637           can_eliminate = false;
 638           break;
 639         }
 640         for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
 641                                    k < kmax && can_eliminate; k++) {
 642           Node* n = use->fast_out(k);
 643           if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
 644               SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
 645               !(n->is_ArrayCopy() &&
 646                 n->as_ArrayCopy()->is_clonebasic() &&
 647                 n->in(ArrayCopyNode::Dest) == use)) {
 648             DEBUG_ONLY(disq_node = n;)
 649             if (n->is_Load() || n->is_LoadStore()) {
 650               NOT_PRODUCT(fail_eliminate = "Field load";)
 651             } else {
 652               NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
 653             }
 654             can_eliminate = false;
 655           }
 656         }
 657       } else if (use->is_ArrayCopy() &&
 658                  (use->as_ArrayCopy()->is_arraycopy_validated() ||
 659                   use->as_ArrayCopy()->is_copyof_validated() ||
 660                   use->as_ArrayCopy()->is_copyofrange_validated()) &&
 661                  use->in(ArrayCopyNode::Dest) == res) {
 662         // ok to eliminate
 663       } else if (use->is_SafePoint()) {
 664         SafePointNode* sfpt = use->as_SafePoint();
 665         if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
 666           // Object is passed as argument.
 667           DEBUG_ONLY(disq_node = use;)
 668           NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
 669           can_eliminate = false;
 670         }
 671         Node* sfptMem = sfpt->memory();
 672         if (sfptMem == NULL || sfptMem->is_top()) {
 673           DEBUG_ONLY(disq_node = use;)
 674           NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
 675           can_eliminate = false;
 676         } else {
 677           safepoints.append_if_missing(sfpt);
 678         }
 679       } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
 680         if (use->is_Phi()) {
 681           if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
 682             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 683           } else {
 684             NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
 685           }
 686           DEBUG_ONLY(disq_node = use;)
 687         } else {
 688           if (use->Opcode() == Op_Return) {
 689             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 690           }else {
 691             NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
 692           }
 693           DEBUG_ONLY(disq_node = use;)
 694         }
 695         can_eliminate = false;
 696       }
 697     }
 698   }
 699 
 700 #ifndef PRODUCT
 701   if (PrintEliminateAllocations) {
 702     if (can_eliminate) {
 703       tty->print("Scalar ");
 704       if (res == NULL)
 705         alloc->dump();
 706       else
 707         res->dump();
 708     } else if (alloc->_is_scalar_replaceable) {
 709       tty->print("NotScalar (%s)", fail_eliminate);
 710       if (res == NULL)
 711         alloc->dump();
 712       else
 713         res->dump();
 714 #ifdef ASSERT
 715       if (disq_node != NULL) {
 716           tty->print("  >>>> ");
 717           disq_node->dump();
 718       }
 719 #endif /*ASSERT*/
 720     }
 721   }
 722 #endif
 723   return can_eliminate;
 724 }
 725 
 726 // Do scalar replacement.
 727 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
 728   GrowableArray <SafePointNode *> safepoints_done;
 729 
 730   ciKlass* klass = NULL;
 731   ciInstanceKlass* iklass = NULL;
 732   int nfields = 0;
 733   int array_base = 0;
 734   int element_size = 0;
 735   BasicType basic_elem_type = T_ILLEGAL;
 736   ciType* elem_type = NULL;
 737 
 738   Node* res = alloc->result_cast();
 739   assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
 740   const TypeOopPtr* res_type = NULL;
 741   if (res != NULL) { // Could be NULL when there are no users
 742     res_type = _igvn.type(res)->isa_oopptr();
 743   }
 744 
 745   if (res != NULL) {
 746     klass = res_type->klass();
 747     if (res_type->isa_instptr()) {
 748       // find the fields of the class which will be needed for safepoint debug information
 749       assert(klass->is_instance_klass(), "must be an instance klass.");
 750       iklass = klass->as_instance_klass();
 751       nfields = iklass->nof_nonstatic_fields();
 752     } else {
 753       // find the array's elements which will be needed for safepoint debug information
 754       nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 755       assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
 756       elem_type = klass->as_array_klass()->element_type();
 757       basic_elem_type = elem_type->basic_type();
 758       array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 759       element_size = type2aelembytes(basic_elem_type);
 760     }
 761   }
 762   //
 763   // Process the safepoint uses
 764   //
 765   while (safepoints.length() > 0) {
 766     SafePointNode* sfpt = safepoints.pop();
 767     Node* mem = sfpt->memory();
 768     Node* ctl = sfpt->control();
 769     assert(sfpt->jvms() != NULL, "missed JVMS");
 770     // Fields of scalar objs are referenced only at the end
 771     // of regular debuginfo at the last (youngest) JVMS.
 772     // Record relative start index.
 773     uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
 774     SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
 775 #ifdef ASSERT
 776                                                  alloc,
 777 #endif
 778                                                  first_ind, nfields);
 779     sobj->init_req(0, C->root());
 780     transform_later(sobj);
 781 
 782     // Scan object's fields adding an input to the safepoint for each field.
 783     for (int j = 0; j < nfields; j++) {
 784       intptr_t offset;
 785       ciField* field = NULL;
 786       if (iklass != NULL) {
 787         field = iklass->nonstatic_field_at(j);
 788         offset = field->offset();
 789         elem_type = field->type();
 790         basic_elem_type = field->layout_type();
 791       } else {
 792         offset = array_base + j * (intptr_t)element_size;
 793       }
 794 
 795       const Type *field_type;
 796       // The next code is taken from Parse::do_get_xxx().
 797       if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
 798         if (!elem_type->is_loaded()) {
 799           field_type = TypeInstPtr::BOTTOM;
 800         } else if (field != NULL && field->is_static_constant()) {
 801           // This can happen if the constant oop is non-perm.
 802           ciObject* con = field->constant_value().as_object();
 803           // Do not "join" in the previous type; it doesn't add value,
 804           // and may yield a vacuous result if the field is of interface type.
 805           field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
 806           assert(field_type != NULL, "field singleton type must be consistent");
 807         } else {
 808           field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
 809         }
 810         if (UseCompressedOops) {
 811           field_type = field_type->make_narrowoop();
 812           basic_elem_type = T_NARROWOOP;
 813         }
 814       } else {
 815         field_type = Type::get_const_basic_type(basic_elem_type);
 816       }
 817 
 818       const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
 819 
 820       Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
 821       if (field_val == NULL) {
 822         // We weren't able to find a value for this field,
 823         // give up on eliminating this allocation.
 824 
 825         // Remove any extra entries we added to the safepoint.
 826         uint last = sfpt->req() - 1;
 827         for (int k = 0;  k < j; k++) {
 828           sfpt->del_req(last--);
 829         }
 830         _igvn._worklist.push(sfpt);
 831         // rollback processed safepoints
 832         while (safepoints_done.length() > 0) {
 833           SafePointNode* sfpt_done = safepoints_done.pop();
 834           // remove any extra entries we added to the safepoint
 835           last = sfpt_done->req() - 1;
 836           for (int k = 0;  k < nfields; k++) {
 837             sfpt_done->del_req(last--);
 838           }
 839           JVMState *jvms = sfpt_done->jvms();
 840           jvms->set_endoff(sfpt_done->req());
 841           // Now make a pass over the debug information replacing any references
 842           // to SafePointScalarObjectNode with the allocated object.
 843           int start = jvms->debug_start();
 844           int end   = jvms->debug_end();
 845           for (int i = start; i < end; i++) {
 846             if (sfpt_done->in(i)->is_SafePointScalarObject()) {
 847               SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
 848               if (scobj->first_index(jvms) == sfpt_done->req() &&
 849                   scobj->n_fields() == (uint)nfields) {
 850                 assert(scobj->alloc() == alloc, "sanity");
 851                 sfpt_done->set_req(i, res);
 852               }
 853             }
 854           }
 855           _igvn._worklist.push(sfpt_done);
 856         }
 857 #ifndef PRODUCT
 858         if (PrintEliminateAllocations) {
 859           if (field != NULL) {
 860             tty->print("=== At SafePoint node %d can't find value of Field: ",
 861                        sfpt->_idx);
 862             field->print();
 863             int field_idx = C->get_alias_index(field_addr_type);
 864             tty->print(" (alias_idx=%d)", field_idx);
 865           } else { // Array's element
 866             tty->print("=== At SafePoint node %d can't find value of array element [%d]",
 867                        sfpt->_idx, j);
 868           }
 869           tty->print(", which prevents elimination of: ");
 870           if (res == NULL)
 871             alloc->dump();
 872           else
 873             res->dump();
 874         }
 875 #endif
 876         return false;
 877       }
 878       if (UseCompressedOops && field_type->isa_narrowoop()) {
 879         // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
 880         // to be able scalar replace the allocation.
 881         if (field_val->is_EncodeP()) {
 882           field_val = field_val->in(1);
 883         } else {
 884           field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
 885         }
 886       }
 887       sfpt->add_req(field_val);
 888     }
 889     JVMState *jvms = sfpt->jvms();
 890     jvms->set_endoff(sfpt->req());
 891     // Now make a pass over the debug information replacing any references
 892     // to the allocated object with "sobj"
 893     int start = jvms->debug_start();
 894     int end   = jvms->debug_end();
 895     sfpt->replace_edges_in_range(res, sobj, start, end);
 896     _igvn._worklist.push(sfpt);
 897     safepoints_done.append_if_missing(sfpt); // keep it for rollback
 898   }
 899   return true;
 900 }
 901 
 902 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
 903   Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
 904   Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
 905   if (ctl_proj != NULL) {
 906     igvn.replace_node(ctl_proj, n->in(0));
 907   }
 908   if (mem_proj != NULL) {
 909     igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
 910   }
 911 }
 912 
 913 // Process users of eliminated allocation.
 914 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
 915   Node* res = alloc->result_cast();
 916   if (res != NULL) {
 917     for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
 918       Node *use = res->last_out(j);
 919       uint oc1 = res->outcnt();
 920 
 921       if (use->is_AddP()) {
 922         for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
 923           Node *n = use->last_out(k);
 924           uint oc2 = use->outcnt();
 925           if (n->is_Store()) {
 926 #ifdef ASSERT
 927             // Verify that there is no dependent MemBarVolatile nodes,
 928             // they should be removed during IGVN, see MemBarNode::Ideal().
 929             for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
 930                                        p < pmax; p++) {
 931               Node* mb = n->fast_out(p);
 932               assert(mb->is_Initialize() || !mb->is_MemBar() ||
 933                      mb->req() <= MemBarNode::Precedent ||
 934                      mb->in(MemBarNode::Precedent) != n,
 935                      "MemBarVolatile should be eliminated for non-escaping object");
 936             }
 937 #endif
 938             _igvn.replace_node(n, n->in(MemNode::Memory));
 939           } else if (n->is_ArrayCopy()) {
 940             // Disconnect ArrayCopy node
 941             ArrayCopyNode* ac = n->as_ArrayCopy();
 942             assert(ac->is_clonebasic(), "unexpected array copy kind");
 943             Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
 944             disconnect_projections(ac, _igvn);
 945             assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
 946             Node* membar_before = alloc->in(0)->in(0);
 947             disconnect_projections(membar_before->as_MemBar(), _igvn);
 948             if (membar_after->is_MemBar()) {
 949               disconnect_projections(membar_after->as_MemBar(), _igvn);
 950             }
 951           } else {
 952             eliminate_gc_barrier(n);
 953           }
 954           k -= (oc2 - use->outcnt());
 955         }
 956         if (UseShenandoahGC) {
 957           _igvn.remove_dead_node(use);
 958         }
 959       } else if (use->is_ArrayCopy()) {
 960         // Disconnect ArrayCopy node
 961         ArrayCopyNode* ac = use->as_ArrayCopy();
 962         assert(ac->is_arraycopy_validated() ||
 963                ac->is_copyof_validated() ||
 964                ac->is_copyofrange_validated(), "unsupported");
 965         CallProjections callprojs;
 966         ac->extract_projections(&callprojs, true);
 967 
 968         _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
 969         _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
 970         _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
 971 
 972         // Set control to top. IGVN will remove the remaining projections
 973         ac->set_req(0, top());
 974         ac->replace_edge(res, top());
 975 
 976         // Disconnect src right away: it can help find new
 977         // opportunities for allocation elimination
 978         Node* src = ac->in(ArrayCopyNode::Src);
 979         ac->replace_edge(src, top());
 980         // src can be top at this point if src and dest of the
 981         // arraycopy were the same
 982         if (src->outcnt() == 0 && !src->is_top()) {
 983           _igvn.remove_dead_node(src);
 984         }
 985 
 986         _igvn._worklist.push(ac);
 987       } else {
 988         eliminate_gc_barrier(use);
 989       }
 990       j -= (oc1 - res->outcnt());
 991     }
 992     assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
 993     _igvn.remove_dead_node(res);
 994   }
 995 
 996   //
 997   // Process other users of allocation's projections
 998   //
 999   if (_resproj != NULL && _resproj->outcnt() != 0) {
1000     // First disconnect stores captured by Initialize node.
1001     // If Initialize node is eliminated first in the following code,
1002     // it will kill such stores and DUIterator_Last will assert.
1003     for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax);  j < jmax; j++) {
1004       Node *use = _resproj->fast_out(j);
1005       if (use->is_AddP()) {
1006         // raw memory addresses used only by the initialization
1007         _igvn.replace_node(use, C->top());
1008         --j; --jmax;
1009       }
1010     }
1011     for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1012       Node *use = _resproj->last_out(j);
1013       uint oc1 = _resproj->outcnt();
1014       if (use->is_Initialize()) {
1015         // Eliminate Initialize node.
1016         InitializeNode *init = use->as_Initialize();
1017         assert(init->outcnt() <= 2, "only a control and memory projection expected");
1018         Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1019         if (ctrl_proj != NULL) {
1020            assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1021           _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1022         }
1023         Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1024         if (mem_proj != NULL) {
1025           Node *mem = init->in(TypeFunc::Memory);
1026 #ifdef ASSERT
1027           if (mem->is_MergeMem()) {
1028             assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1029           } else {
1030             assert(mem == _memproj_fallthrough, "allocation memory projection");
1031           }
1032 #endif
1033           _igvn.replace_node(mem_proj, mem);
1034         }
1035       } else  {
1036         assert(false, "only Initialize or AddP expected");
1037       }
1038       j -= (oc1 - _resproj->outcnt());
1039     }
1040   }
1041   if (_fallthroughcatchproj != NULL) {
1042     _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1043   }
1044   if (_memproj_fallthrough != NULL) {
1045     _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1046   }
1047   if (_memproj_catchall != NULL) {
1048     _igvn.replace_node(_memproj_catchall, C->top());
1049   }
1050   if (_ioproj_fallthrough != NULL) {
1051     _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1052   }
1053   if (_ioproj_catchall != NULL) {
1054     _igvn.replace_node(_ioproj_catchall, C->top());
1055   }
1056   if (_catchallcatchproj != NULL) {
1057     _igvn.replace_node(_catchallcatchproj, C->top());
1058   }
1059 }
1060 
1061 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1062   // Don't do scalar replacement if the frame can be popped by JVMTI:
1063   // if reallocation fails during deoptimization we'll pop all
1064   // interpreter frames for this compiled frame and that won't play
1065   // nice with JVMTI popframe.
1066   if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1067     return false;
1068   }
1069   Node* klass = alloc->in(AllocateNode::KlassNode);
1070   const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1071   Node* res = alloc->result_cast();
1072   // Eliminate boxing allocations which are not used
1073   // regardless scalar replacable status.
1074   bool boxing_alloc = C->eliminate_boxing() &&
1075                       tklass->klass()->is_instance_klass()  &&
1076                       tklass->klass()->as_instance_klass()->is_box_klass();
1077   if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1078     return false;
1079   }
1080 
1081   extract_call_projections(alloc);
1082 
1083   GrowableArray <SafePointNode *> safepoints;
1084   if (!can_eliminate_allocation(alloc, safepoints)) {
1085     return false;
1086   }
1087 
1088   if (!alloc->_is_scalar_replaceable) {
1089     assert(res == NULL, "sanity");
1090     // We can only eliminate allocation if all debug info references
1091     // are already replaced with SafePointScalarObject because
1092     // we can't search for a fields value without instance_id.
1093     if (safepoints.length() > 0) {
1094       return false;
1095     }
1096   }
1097 
1098   if (!scalar_replacement(alloc, safepoints)) {
1099     return false;
1100   }
1101 
1102   CompileLog* log = C->log();
1103   if (log != NULL) {
1104     log->head("eliminate_allocation type='%d'",
1105               log->identify(tklass->klass()));
1106     JVMState* p = alloc->jvms();
1107     while (p != NULL) {
1108       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1109       p = p->caller();
1110     }
1111     log->tail("eliminate_allocation");
1112   }
1113 
1114   process_users_of_allocation(alloc);
1115 
1116 #ifndef PRODUCT
1117   if (PrintEliminateAllocations) {
1118     if (alloc->is_AllocateArray())
1119       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1120     else
1121       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1122   }
1123 #endif
1124 
1125   return true;
1126 }
1127 
1128 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1129   // EA should remove all uses of non-escaping boxing node.
1130   if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1131     return false;
1132   }
1133 
1134   assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1135 
1136   extract_call_projections(boxing);
1137 
1138   const TypeTuple* r = boxing->tf()->range();
1139   assert(r->cnt() > TypeFunc::Parms, "sanity");
1140   const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1141   assert(t != NULL, "sanity");
1142 
1143   CompileLog* log = C->log();
1144   if (log != NULL) {
1145     log->head("eliminate_boxing type='%d'",
1146               log->identify(t->klass()));
1147     JVMState* p = boxing->jvms();
1148     while (p != NULL) {
1149       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1150       p = p->caller();
1151     }
1152     log->tail("eliminate_boxing");
1153   }
1154 
1155   process_users_of_allocation(boxing);
1156 
1157 #ifndef PRODUCT
1158   if (PrintEliminateAllocations) {
1159     tty->print("++++ Eliminated: %d ", boxing->_idx);
1160     boxing->method()->print_short_name(tty);
1161     tty->cr();
1162   }
1163 #endif
1164 
1165   return true;
1166 }
1167 
1168 //---------------------------set_eden_pointers-------------------------
1169 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1170   if (UseTLAB) {                // Private allocation: load from TLS
1171     Node* thread = transform_later(new ThreadLocalNode());
1172     int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1173     int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1174     eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1175     eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1176   } else {                      // Shared allocation: load from globals
1177     CollectedHeap* ch = Universe::heap();
1178     address top_adr = (address)ch->top_addr();
1179     address end_adr = (address)ch->end_addr();
1180     eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1181     eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1182   }
1183 }
1184 
1185 
1186 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1187   Node* adr = basic_plus_adr(base, offset);
1188   const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1189   Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1190   transform_later(value);
1191   return value;
1192 }
1193 
1194 
1195 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1196   Node* adr = basic_plus_adr(base, offset);
1197   mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1198   transform_later(mem);
1199   return mem;
1200 }
1201 
1202 //=============================================================================
1203 //
1204 //                              A L L O C A T I O N
1205 //
1206 // Allocation attempts to be fast in the case of frequent small objects.
1207 // It breaks down like this:
1208 //
1209 // 1) Size in doublewords is computed.  This is a constant for objects and
1210 // variable for most arrays.  Doubleword units are used to avoid size
1211 // overflow of huge doubleword arrays.  We need doublewords in the end for
1212 // rounding.
1213 //
1214 // 2) Size is checked for being 'too large'.  Too-large allocations will go
1215 // the slow path into the VM.  The slow path can throw any required
1216 // exceptions, and does all the special checks for very large arrays.  The
1217 // size test can constant-fold away for objects.  For objects with
1218 // finalizers it constant-folds the otherway: you always go slow with
1219 // finalizers.
1220 //
1221 // 3) If NOT using TLABs, this is the contended loop-back point.
1222 // Load-Locked the heap top.  If using TLABs normal-load the heap top.
1223 //
1224 // 4) Check that heap top + size*8 < max.  If we fail go the slow ` route.
1225 // NOTE: "top+size*8" cannot wrap the 4Gig line!  Here's why: for largish
1226 // "size*8" we always enter the VM, where "largish" is a constant picked small
1227 // enough that there's always space between the eden max and 4Gig (old space is
1228 // there so it's quite large) and large enough that the cost of entering the VM
1229 // is dwarfed by the cost to initialize the space.
1230 //
1231 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1232 // down.  If contended, repeat at step 3.  If using TLABs normal-store
1233 // adjusted heap top back down; there is no contention.
1234 //
1235 // 6) If !ZeroTLAB then Bulk-clear the object/array.  Fill in klass & mark
1236 // fields.
1237 //
1238 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1239 // oop flavor.
1240 //
1241 //=============================================================================
1242 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1243 // Allocations bigger than this always go the slow route.
1244 // This value must be small enough that allocation attempts that need to
1245 // trigger exceptions go the slow route.  Also, it must be small enough so
1246 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1247 //=============================================================================j//
1248 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1249 // The allocator will coalesce int->oop copies away.  See comment in
1250 // coalesce.cpp about how this works.  It depends critically on the exact
1251 // code shape produced here, so if you are changing this code shape
1252 // make sure the GC info for the heap-top is correct in and around the
1253 // slow-path call.
1254 //
1255 
1256 void PhaseMacroExpand::expand_allocate_common(
1257             AllocateNode* alloc, // allocation node to be expanded
1258             Node* length,  // array length for an array allocation
1259             const TypeFunc* slow_call_type, // Type of slow call
1260             address slow_call_address  // Address of slow call
1261     )
1262 {
1263 
1264   Node* ctrl = alloc->in(TypeFunc::Control);
1265   Node* mem  = alloc->in(TypeFunc::Memory);
1266   Node* i_o  = alloc->in(TypeFunc::I_O);
1267   Node* size_in_bytes     = alloc->in(AllocateNode::AllocSize);
1268   Node* klass_node        = alloc->in(AllocateNode::KlassNode);
1269   Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1270 
1271   assert(ctrl != NULL, "must have control");
1272   // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1273   // they will not be used if "always_slow" is set
1274   enum { slow_result_path = 1, fast_result_path = 2 };
1275   Node *result_region = NULL;
1276   Node *result_phi_rawmem = NULL;
1277   Node *result_phi_rawoop = NULL;
1278   Node *result_phi_i_o = NULL;
1279 
1280   // The initial slow comparison is a size check, the comparison
1281   // we want to do is a BoolTest::gt
1282   bool always_slow = false;
1283   int tv = _igvn.find_int_con(initial_slow_test, -1);
1284   if (tv >= 0) {
1285     always_slow = (tv == 1);
1286     initial_slow_test = NULL;
1287   } else {
1288     initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1289   }
1290 
1291   if (C->env()->dtrace_alloc_probes() ||
1292       (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1293     // Force slow-path allocation
1294     always_slow = true;
1295     initial_slow_test = NULL;
1296   }
1297 
1298 
1299   enum { too_big_or_final_path = 1, need_gc_path = 2 };
1300   Node *slow_region = NULL;
1301   Node *toobig_false = ctrl;
1302 
1303   assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1304   // generate the initial test if necessary
1305   if (initial_slow_test != NULL ) {
1306     slow_region = new RegionNode(3);
1307 
1308     // Now make the initial failure test.  Usually a too-big test but
1309     // might be a TRUE for finalizers or a fancy class check for
1310     // newInstance0.
1311     IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1312     transform_later(toobig_iff);
1313     // Plug the failing-too-big test into the slow-path region
1314     Node *toobig_true = new IfTrueNode( toobig_iff );
1315     transform_later(toobig_true);
1316     slow_region    ->init_req( too_big_or_final_path, toobig_true );
1317     toobig_false = new IfFalseNode( toobig_iff );
1318     transform_later(toobig_false);
1319   } else {         // No initial test, just fall into next case
1320     toobig_false = ctrl;
1321     debug_only(slow_region = NodeSentinel);
1322   }
1323 
1324   Node *slow_mem = mem;  // save the current memory state for slow path
1325   // generate the fast allocation code unless we know that the initial test will always go slow
1326   if (!always_slow) {
1327     // Fast path modifies only raw memory.
1328     if (mem->is_MergeMem()) {
1329       mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1330     }
1331 
1332     Node* eden_top_adr;
1333     Node* eden_end_adr;
1334 
1335     set_eden_pointers(eden_top_adr, eden_end_adr);
1336 
1337     // Load Eden::end.  Loop invariant and hoisted.
1338     //
1339     // Note: We set the control input on "eden_end" and "old_eden_top" when using
1340     //       a TLAB to work around a bug where these values were being moved across
1341     //       a safepoint.  These are not oops, so they cannot be include in the oop
1342     //       map, but they can be changed by a GC.   The proper way to fix this would
1343     //       be to set the raw memory state when generating a  SafepointNode.  However
1344     //       this will require extensive changes to the loop optimization in order to
1345     //       prevent a degradation of the optimization.
1346     //       See comment in memnode.hpp, around line 227 in class LoadPNode.
1347     Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1348 
1349     // allocate the Region and Phi nodes for the result
1350     result_region = new RegionNode(3);
1351     result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1352     result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1353     result_phi_i_o    = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1354 
1355     // We need a Region for the loop-back contended case.
1356     enum { fall_in_path = 1, contended_loopback_path = 2 };
1357     Node *contended_region;
1358     Node *contended_phi_rawmem;
1359     if (UseTLAB) {
1360       contended_region = toobig_false;
1361       contended_phi_rawmem = mem;
1362     } else {
1363       contended_region = new RegionNode(3);
1364       contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1365       // Now handle the passing-too-big test.  We fall into the contended
1366       // loop-back merge point.
1367       contended_region    ->init_req(fall_in_path, toobig_false);
1368       contended_phi_rawmem->init_req(fall_in_path, mem);
1369       transform_later(contended_region);
1370       transform_later(contended_phi_rawmem);
1371     }
1372 
1373     // Load(-locked) the heap top.
1374     // See note above concerning the control input when using a TLAB
1375     Node *old_eden_top = UseTLAB
1376       ? new LoadPNode      (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1377       : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1378 
1379     transform_later(old_eden_top);
1380     // Add to heap top to get a new heap top
1381     Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1382     transform_later(new_eden_top);
1383     // Check for needing a GC; compare against heap end
1384     Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1385     transform_later(needgc_cmp);
1386     Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1387     transform_later(needgc_bol);
1388     IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1389     transform_later(needgc_iff);
1390 
1391     // Plug the failing-heap-space-need-gc test into the slow-path region
1392     Node *needgc_true = new IfTrueNode(needgc_iff);
1393     transform_later(needgc_true);
1394     if (initial_slow_test) {
1395       slow_region->init_req(need_gc_path, needgc_true);
1396       // This completes all paths into the slow merge point
1397       transform_later(slow_region);
1398     } else {                      // No initial slow path needed!
1399       // Just fall from the need-GC path straight into the VM call.
1400       slow_region = needgc_true;
1401     }
1402     // No need for a GC.  Setup for the Store-Conditional
1403     Node *needgc_false = new IfFalseNode(needgc_iff);
1404     transform_later(needgc_false);
1405 
1406     // Grab regular I/O before optional prefetch may change it.
1407     // Slow-path does no I/O so just set it to the original I/O.
1408     result_phi_i_o->init_req(slow_result_path, i_o);
1409 
1410     i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1411                               old_eden_top, new_eden_top, length);
1412 
1413     // Name successful fast-path variables
1414     Node* fast_oop = old_eden_top;
1415     Node* fast_oop_ctrl;
1416     Node* fast_oop_rawmem;
1417 
1418     // Store (-conditional) the modified eden top back down.
1419     // StorePConditional produces flags for a test PLUS a modified raw
1420     // memory state.
1421     if (UseTLAB) {
1422       Node* store_eden_top =
1423         new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1424                               TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1425       transform_later(store_eden_top);
1426       fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1427       fast_oop_rawmem = store_eden_top;
1428     } else {
1429       Node* store_eden_top =
1430         new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1431                                          new_eden_top, fast_oop/*old_eden_top*/);
1432       transform_later(store_eden_top);
1433       Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1434       transform_later(contention_check);
1435       store_eden_top = new SCMemProjNode(store_eden_top);
1436       transform_later(store_eden_top);
1437 
1438       // If not using TLABs, check to see if there was contention.
1439       IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1440       transform_later(contention_iff);
1441       Node *contention_true = new IfTrueNode(contention_iff);
1442       transform_later(contention_true);
1443       // If contention, loopback and try again.
1444       contended_region->init_req(contended_loopback_path, contention_true);
1445       contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1446 
1447       // Fast-path succeeded with no contention!
1448       Node *contention_false = new IfFalseNode(contention_iff);
1449       transform_later(contention_false);
1450       fast_oop_ctrl = contention_false;
1451 
1452       // Bump total allocated bytes for this thread
1453       Node* thread = new ThreadLocalNode();
1454       transform_later(thread);
1455       Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1456                                              in_bytes(JavaThread::allocated_bytes_offset()));
1457       Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1458                                     0, TypeLong::LONG, T_LONG);
1459 #ifdef _LP64
1460       Node* alloc_size = size_in_bytes;
1461 #else
1462       Node* alloc_size = new ConvI2LNode(size_in_bytes);
1463       transform_later(alloc_size);
1464 #endif
1465       Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1466       transform_later(new_alloc_bytes);
1467       fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1468                                    0, new_alloc_bytes, T_LONG);
1469     }
1470 
1471     InitializeNode* init = alloc->initialization();
1472     fast_oop_rawmem = initialize_object(alloc,
1473                                         fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1474                                         klass_node, length, size_in_bytes);
1475 
1476     // If initialization is performed by an array copy, any required
1477     // MemBarStoreStore was already added. If the object does not
1478     // escape no need for a MemBarStoreStore. If the object does not
1479     // escape in its initializer and memory barrier (MemBarStoreStore or
1480     // stronger) is already added at exit of initializer, also no need
1481     // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1482     // so that stores that initialize this object can't be reordered
1483     // with a subsequent store that makes this object accessible by
1484     // other threads.
1485     // Other threads include java threads and JVM internal threads
1486     // (for example concurrent GC threads). Current concurrent GC
1487     // implementation: CMS and G1 will not scan newly created object,
1488     // so it's safe to skip storestore barrier when allocation does
1489     // not escape.
1490     if (!alloc->does_not_escape_thread() &&
1491         !alloc->is_allocation_MemBar_redundant() &&
1492         (init == NULL || !init->is_complete_with_arraycopy())) {
1493       if (init == NULL || init->req() < InitializeNode::RawStores) {
1494         // No InitializeNode or no stores captured by zeroing
1495         // elimination. Simply add the MemBarStoreStore after object
1496         // initialization.
1497         MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1498         transform_later(mb);
1499 
1500         mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1501         mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1502         fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1503         transform_later(fast_oop_ctrl);
1504         fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1505         transform_later(fast_oop_rawmem);
1506       } else {
1507         // Add the MemBarStoreStore after the InitializeNode so that
1508         // all stores performing the initialization that were moved
1509         // before the InitializeNode happen before the storestore
1510         // barrier.
1511 
1512         Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1513         Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1514 
1515         MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1516         transform_later(mb);
1517 
1518         Node* ctrl = new ProjNode(init,TypeFunc::Control);
1519         transform_later(ctrl);
1520         Node* mem = new ProjNode(init,TypeFunc::Memory);
1521         transform_later(mem);
1522 
1523         // The MemBarStoreStore depends on control and memory coming
1524         // from the InitializeNode
1525         mb->init_req(TypeFunc::Memory, mem);
1526         mb->init_req(TypeFunc::Control, ctrl);
1527 
1528         ctrl = new ProjNode(mb,TypeFunc::Control);
1529         transform_later(ctrl);
1530         mem = new ProjNode(mb,TypeFunc::Memory);
1531         transform_later(mem);
1532 
1533         // All nodes that depended on the InitializeNode for control
1534         // and memory must now depend on the MemBarNode that itself
1535         // depends on the InitializeNode
1536         if (init_ctrl != NULL) {
1537           _igvn.replace_node(init_ctrl, ctrl);
1538         }
1539         if (init_mem != NULL) {
1540           _igvn.replace_node(init_mem, mem);
1541         }
1542       }
1543     }
1544 
1545     if (C->env()->dtrace_extended_probes()) {
1546       // Slow-path call
1547       int size = TypeFunc::Parms + 2;
1548       CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1549                                             CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1550                                             "dtrace_object_alloc",
1551                                             TypeRawPtr::BOTTOM);
1552 
1553       // Get base of thread-local storage area
1554       Node* thread = new ThreadLocalNode();
1555       transform_later(thread);
1556 
1557       call->init_req(TypeFunc::Parms+0, thread);
1558       call->init_req(TypeFunc::Parms+1, fast_oop);
1559       call->init_req(TypeFunc::Control, fast_oop_ctrl);
1560       call->init_req(TypeFunc::I_O    , top()); // does no i/o
1561       call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1562       call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1563       call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1564       transform_later(call);
1565       fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1566       transform_later(fast_oop_ctrl);
1567       fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1568       transform_later(fast_oop_rawmem);
1569     }
1570 
1571     // Plug in the successful fast-path into the result merge point
1572     result_region    ->init_req(fast_result_path, fast_oop_ctrl);
1573     result_phi_rawoop->init_req(fast_result_path, fast_oop);
1574     result_phi_i_o   ->init_req(fast_result_path, i_o);
1575     result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1576   } else {
1577     slow_region = ctrl;
1578     result_phi_i_o = i_o; // Rename it to use in the following code.
1579   }
1580 
1581   // Generate slow-path call
1582   CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1583                                OptoRuntime::stub_name(slow_call_address),
1584                                alloc->jvms()->bci(),
1585                                TypePtr::BOTTOM);
1586   call->init_req( TypeFunc::Control, slow_region );
1587   call->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
1588   call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1589   call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1590   call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1591 
1592   call->init_req(TypeFunc::Parms+0, klass_node);
1593   if (length != NULL) {
1594     call->init_req(TypeFunc::Parms+1, length);
1595   }
1596 
1597   // Copy debug information and adjust JVMState information, then replace
1598   // allocate node with the call
1599   copy_call_debug_info((CallNode *) alloc,  call);
1600   if (!always_slow) {
1601     call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1602   } else {
1603     // Hook i_o projection to avoid its elimination during allocation
1604     // replacement (when only a slow call is generated).
1605     call->set_req(TypeFunc::I_O, result_phi_i_o);
1606   }
1607   _igvn.replace_node(alloc, call);
1608   transform_later(call);
1609 
1610   // Identify the output projections from the allocate node and
1611   // adjust any references to them.
1612   // The control and io projections look like:
1613   //
1614   //        v---Proj(ctrl) <-----+   v---CatchProj(ctrl)
1615   //  Allocate                   Catch
1616   //        ^---Proj(io) <-------+   ^---CatchProj(io)
1617   //
1618   //  We are interested in the CatchProj nodes.
1619   //
1620   extract_call_projections(call);
1621 
1622   // An allocate node has separate memory projections for the uses on
1623   // the control and i_o paths. Replace the control memory projection with
1624   // result_phi_rawmem (unless we are only generating a slow call when
1625   // both memory projections are combined)
1626   if (!always_slow && _memproj_fallthrough != NULL) {
1627     for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1628       Node *use = _memproj_fallthrough->fast_out(i);
1629       _igvn.rehash_node_delayed(use);
1630       imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1631       // back up iterator
1632       --i;
1633     }
1634   }
1635   // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1636   // _memproj_catchall so we end up with a call that has only 1 memory projection.
1637   if (_memproj_catchall != NULL ) {
1638     if (_memproj_fallthrough == NULL) {
1639       _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1640       transform_later(_memproj_fallthrough);
1641     }
1642     for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1643       Node *use = _memproj_catchall->fast_out(i);
1644       _igvn.rehash_node_delayed(use);
1645       imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1646       // back up iterator
1647       --i;
1648     }
1649     assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1650     _igvn.remove_dead_node(_memproj_catchall);
1651   }
1652 
1653   // An allocate node has separate i_o projections for the uses on the control
1654   // and i_o paths. Always replace the control i_o projection with result i_o
1655   // otherwise incoming i_o become dead when only a slow call is generated
1656   // (it is different from memory projections where both projections are
1657   // combined in such case).
1658   if (_ioproj_fallthrough != NULL) {
1659     for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1660       Node *use = _ioproj_fallthrough->fast_out(i);
1661       _igvn.rehash_node_delayed(use);
1662       imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1663       // back up iterator
1664       --i;
1665     }
1666   }
1667   // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1668   // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1669   if (_ioproj_catchall != NULL ) {
1670     if (_ioproj_fallthrough == NULL) {
1671       _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1672       transform_later(_ioproj_fallthrough);
1673     }
1674     for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1675       Node *use = _ioproj_catchall->fast_out(i);
1676       _igvn.rehash_node_delayed(use);
1677       imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1678       // back up iterator
1679       --i;
1680     }
1681     assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1682     _igvn.remove_dead_node(_ioproj_catchall);
1683   }
1684 
1685   // if we generated only a slow call, we are done
1686   if (always_slow) {
1687     // Now we can unhook i_o.
1688     if (result_phi_i_o->outcnt() > 1) {
1689       call->set_req(TypeFunc::I_O, top());
1690     } else {
1691       assert(result_phi_i_o->unique_ctrl_out() == call, "");
1692       // Case of new array with negative size known during compilation.
1693       // AllocateArrayNode::Ideal() optimization disconnect unreachable
1694       // following code since call to runtime will throw exception.
1695       // As result there will be no users of i_o after the call.
1696       // Leave i_o attached to this call to avoid problems in preceding graph.
1697     }
1698     return;
1699   }
1700 
1701 
1702   if (_fallthroughcatchproj != NULL) {
1703     ctrl = _fallthroughcatchproj->clone();
1704     transform_later(ctrl);
1705     _igvn.replace_node(_fallthroughcatchproj, result_region);
1706   } else {
1707     ctrl = top();
1708   }
1709   Node *slow_result;
1710   if (_resproj == NULL) {
1711     // no uses of the allocation result
1712     slow_result = top();
1713   } else {
1714     slow_result = _resproj->clone();
1715     transform_later(slow_result);
1716     _igvn.replace_node(_resproj, result_phi_rawoop);
1717   }
1718 
1719   // Plug slow-path into result merge point
1720   result_region    ->init_req( slow_result_path, ctrl );
1721   result_phi_rawoop->init_req( slow_result_path, slow_result);
1722   result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1723   transform_later(result_region);
1724   transform_later(result_phi_rawoop);
1725   transform_later(result_phi_rawmem);
1726   transform_later(result_phi_i_o);
1727   // This completes all paths into the result merge point
1728 }
1729 
1730 
1731 // Helper for PhaseMacroExpand::expand_allocate_common.
1732 // Initializes the newly-allocated storage.
1733 Node*
1734 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1735                                     Node* control, Node* rawmem, Node* object,
1736                                     Node* klass_node, Node* length,
1737                                     Node* size_in_bytes) {
1738   InitializeNode* init = alloc->initialization();
1739   // Store the klass & mark bits
1740   Node* mark_node = NULL;
1741   // For now only enable fast locking for non-array types
1742   if (UseBiasedLocking && (length == NULL)) {
1743     mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1744   } else {
1745     mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1746   }
1747   rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1748 
1749   rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1750   int header_size = alloc->minimum_header_size();  // conservatively small
1751 
1752   // Array length
1753   if (length != NULL) {         // Arrays need length field
1754     rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1755     // conservatively small header size:
1756     header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1757     ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1758     if (k->is_array_klass())    // we know the exact header size in most cases:
1759       header_size = Klass::layout_helper_header_size(k->layout_helper());
1760   }
1761 
1762   // Clear the object body, if necessary.
1763   if (init == NULL) {
1764     // The init has somehow disappeared; be cautious and clear everything.
1765     //
1766     // This can happen if a node is allocated but an uncommon trap occurs
1767     // immediately.  In this case, the Initialize gets associated with the
1768     // trap, and may be placed in a different (outer) loop, if the Allocate
1769     // is in a loop.  If (this is rare) the inner loop gets unrolled, then
1770     // there can be two Allocates to one Initialize.  The answer in all these
1771     // edge cases is safety first.  It is always safe to clear immediately
1772     // within an Allocate, and then (maybe or maybe not) clear some more later.
1773     if (!(UseTLAB && ZeroTLAB)) {
1774       rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1775                                             header_size, size_in_bytes,
1776                                             &_igvn);
1777     }
1778   } else {
1779     if (!init->is_complete()) {
1780       // Try to win by zeroing only what the init does not store.
1781       // We can also try to do some peephole optimizations,
1782       // such as combining some adjacent subword stores.
1783       rawmem = init->complete_stores(control, rawmem, object,
1784                                      header_size, size_in_bytes, &_igvn);
1785     }
1786     // We have no more use for this link, since the AllocateNode goes away:
1787     init->set_req(InitializeNode::RawAddress, top());
1788     // (If we keep the link, it just confuses the register allocator,
1789     // who thinks he sees a real use of the address by the membar.)
1790   }
1791 
1792   return rawmem;
1793 }
1794 
1795 // Generate prefetch instructions for next allocations.
1796 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1797                                         Node*& contended_phi_rawmem,
1798                                         Node* old_eden_top, Node* new_eden_top,
1799                                         Node* length) {
1800    enum { fall_in_path = 1, pf_path = 2 };
1801    if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1802       // Generate prefetch allocation with watermark check.
1803       // As an allocation hits the watermark, we will prefetch starting
1804       // at a "distance" away from watermark.
1805 
1806       Node *pf_region = new RegionNode(3);
1807       Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1808                                                 TypeRawPtr::BOTTOM );
1809       // I/O is used for Prefetch
1810       Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1811 
1812       Node *thread = new ThreadLocalNode();
1813       transform_later(thread);
1814 
1815       Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1816                    _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1817       transform_later(eden_pf_adr);
1818 
1819       Node *old_pf_wm = new LoadPNode(needgc_false,
1820                                    contended_phi_rawmem, eden_pf_adr,
1821                                    TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1822                                    MemNode::unordered);
1823       transform_later(old_pf_wm);
1824 
1825       // check against new_eden_top
1826       Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1827       transform_later(need_pf_cmp);
1828       Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1829       transform_later(need_pf_bol);
1830       IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1831                                        PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1832       transform_later(need_pf_iff);
1833 
1834       // true node, add prefetchdistance
1835       Node *need_pf_true = new IfTrueNode( need_pf_iff );
1836       transform_later(need_pf_true);
1837 
1838       Node *need_pf_false = new IfFalseNode( need_pf_iff );
1839       transform_later(need_pf_false);
1840 
1841       Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1842                                     _igvn.MakeConX(AllocatePrefetchDistance) );
1843       transform_later(new_pf_wmt );
1844       new_pf_wmt->set_req(0, need_pf_true);
1845 
1846       Node *store_new_wmt = new StorePNode(need_pf_true,
1847                                        contended_phi_rawmem, eden_pf_adr,
1848                                        TypeRawPtr::BOTTOM, new_pf_wmt,
1849                                        MemNode::unordered);
1850       transform_later(store_new_wmt);
1851 
1852       // adding prefetches
1853       pf_phi_abio->init_req( fall_in_path, i_o );
1854 
1855       Node *prefetch_adr;
1856       Node *prefetch;
1857       uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1858       uint step_size = AllocatePrefetchStepSize;
1859       uint distance = 0;
1860 
1861       for ( uint i = 0; i < lines; i++ ) {
1862         prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1863                                             _igvn.MakeConX(distance) );
1864         transform_later(prefetch_adr);
1865         prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1866         transform_later(prefetch);
1867         distance += step_size;
1868         i_o = prefetch;
1869       }
1870       pf_phi_abio->set_req( pf_path, i_o );
1871 
1872       pf_region->init_req( fall_in_path, need_pf_false );
1873       pf_region->init_req( pf_path, need_pf_true );
1874 
1875       pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1876       pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1877 
1878       transform_later(pf_region);
1879       transform_later(pf_phi_rawmem);
1880       transform_later(pf_phi_abio);
1881 
1882       needgc_false = pf_region;
1883       contended_phi_rawmem = pf_phi_rawmem;
1884       i_o = pf_phi_abio;
1885    } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1886       // Insert a prefetch instruction for each allocation.
1887       // This code is used to generate 1 prefetch instruction per cache line.
1888 
1889       // Generate several prefetch instructions.
1890       uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1891       uint step_size = AllocatePrefetchStepSize;
1892       uint distance = AllocatePrefetchDistance;
1893 
1894       // Next cache address.
1895       Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1896                                      _igvn.MakeConX(step_size + distance));
1897       transform_later(cache_adr);
1898       cache_adr = new CastP2XNode(needgc_false, cache_adr);
1899       transform_later(cache_adr);
1900       // Address is aligned to execute prefetch to the beginning of cache line size
1901       // (it is important when BIS instruction is used on SPARC as prefetch).
1902       Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1903       cache_adr = new AndXNode(cache_adr, mask);
1904       transform_later(cache_adr);
1905       cache_adr = new CastX2PNode(cache_adr);
1906       transform_later(cache_adr);
1907 
1908       // Prefetch
1909       Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1910       prefetch->set_req(0, needgc_false);
1911       transform_later(prefetch);
1912       contended_phi_rawmem = prefetch;
1913       Node *prefetch_adr;
1914       distance = step_size;
1915       for ( uint i = 1; i < lines; i++ ) {
1916         prefetch_adr = new AddPNode( cache_adr, cache_adr,
1917                                             _igvn.MakeConX(distance) );
1918         transform_later(prefetch_adr);
1919         prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1920         transform_later(prefetch);
1921         distance += step_size;
1922         contended_phi_rawmem = prefetch;
1923       }
1924    } else if( AllocatePrefetchStyle > 0 ) {
1925       // Insert a prefetch for each allocation only on the fast-path
1926       Node *prefetch_adr;
1927       Node *prefetch;
1928       // Generate several prefetch instructions.
1929       uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1930       uint step_size = AllocatePrefetchStepSize;
1931       uint distance = AllocatePrefetchDistance;
1932       for ( uint i = 0; i < lines; i++ ) {
1933         prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1934                                             _igvn.MakeConX(distance) );
1935         transform_later(prefetch_adr);
1936         prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1937         // Do not let it float too high, since if eden_top == eden_end,
1938         // both might be null.
1939         if( i == 0 ) { // Set control for first prefetch, next follows it
1940           prefetch->init_req(0, needgc_false);
1941         }
1942         transform_later(prefetch);
1943         distance += step_size;
1944         i_o = prefetch;
1945       }
1946    }
1947    return i_o;
1948 }
1949 
1950 
1951 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1952   expand_allocate_common(alloc, NULL,
1953                          OptoRuntime::new_instance_Type(),
1954                          OptoRuntime::new_instance_Java());
1955 }
1956 
1957 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1958   Node* length = alloc->in(AllocateNode::ALength);
1959   InitializeNode* init = alloc->initialization();
1960   Node* klass_node = alloc->in(AllocateNode::KlassNode);
1961   ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1962   address slow_call_address;  // Address of slow call
1963   if (init != NULL && init->is_complete_with_arraycopy() &&
1964       k->is_type_array_klass()) {
1965     // Don't zero type array during slow allocation in VM since
1966     // it will be initialized later by arraycopy in compiled code.
1967     slow_call_address = OptoRuntime::new_array_nozero_Java();
1968   } else {
1969     slow_call_address = OptoRuntime::new_array_Java();
1970   }
1971   expand_allocate_common(alloc, length,
1972                          OptoRuntime::new_array_Type(),
1973                          slow_call_address);
1974 }
1975 
1976 //-------------------mark_eliminated_box----------------------------------
1977 //
1978 // During EA obj may point to several objects but after few ideal graph
1979 // transformations (CCP) it may point to only one non escaping object
1980 // (but still using phi), corresponding locks and unlocks will be marked
1981 // for elimination. Later obj could be replaced with a new node (new phi)
1982 // and which does not have escape information. And later after some graph
1983 // reshape other locks and unlocks (which were not marked for elimination
1984 // before) are connected to this new obj (phi) but they still will not be
1985 // marked for elimination since new obj has no escape information.
1986 // Mark all associated (same box and obj) lock and unlock nodes for
1987 // elimination if some of them marked already.
1988 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1989   if (oldbox->as_BoxLock()->is_eliminated())
1990     return; // This BoxLock node was processed already.
1991 
1992   // New implementation (EliminateNestedLocks) has separate BoxLock
1993   // node for each locked region so mark all associated locks/unlocks as
1994   // eliminated even if different objects are referenced in one locked region
1995   // (for example, OSR compilation of nested loop inside locked scope).
1996   if (EliminateNestedLocks ||
1997       oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1998     // Box is used only in one lock region. Mark this box as eliminated.
1999     _igvn.hash_delete(oldbox);
2000     oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2001      _igvn.hash_insert(oldbox);
2002 
2003     for (uint i = 0; i < oldbox->outcnt(); i++) {
2004       Node* u = oldbox->raw_out(i);
2005       if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2006         AbstractLockNode* alock = u->as_AbstractLock();
2007         // Check lock's box since box could be referenced by Lock's debug info.
2008         if (alock->box_node() == oldbox) {
2009           // Mark eliminated all related locks and unlocks.
2010 #ifdef ASSERT
2011           alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2012 #endif
2013           alock->set_non_esc_obj();
2014         }
2015       }
2016     }
2017     return;
2018   }
2019 
2020   // Create new "eliminated" BoxLock node and use it in monitor debug info
2021   // instead of oldbox for the same object.
2022   BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2023 
2024   // Note: BoxLock node is marked eliminated only here and it is used
2025   // to indicate that all associated lock and unlock nodes are marked
2026   // for elimination.
2027   newbox->set_eliminated();
2028   transform_later(newbox);
2029 
2030   // Replace old box node with new box for all users of the same object.
2031   for (uint i = 0; i < oldbox->outcnt();) {
2032     bool next_edge = true;
2033 
2034     Node* u = oldbox->raw_out(i);
2035     if (u->is_AbstractLock()) {
2036       AbstractLockNode* alock = u->as_AbstractLock();
2037       if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2038         // Replace Box and mark eliminated all related locks and unlocks.
2039 #ifdef ASSERT
2040         alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2041 #endif
2042         alock->set_non_esc_obj();
2043         _igvn.rehash_node_delayed(alock);
2044         alock->set_box_node(newbox);
2045         next_edge = false;
2046       }
2047     }
2048     if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2049       FastLockNode* flock = u->as_FastLock();
2050       assert(flock->box_node() == oldbox, "sanity");
2051       _igvn.rehash_node_delayed(flock);
2052       flock->set_box_node(newbox);
2053       next_edge = false;
2054     }
2055 
2056     // Replace old box in monitor debug info.
2057     if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2058       SafePointNode* sfn = u->as_SafePoint();
2059       JVMState* youngest_jvms = sfn->jvms();
2060       int max_depth = youngest_jvms->depth();
2061       for (int depth = 1; depth <= max_depth; depth++) {
2062         JVMState* jvms = youngest_jvms->of_depth(depth);
2063         int num_mon  = jvms->nof_monitors();
2064         // Loop over monitors
2065         for (int idx = 0; idx < num_mon; idx++) {
2066           Node* obj_node = sfn->monitor_obj(jvms, idx);
2067           Node* box_node = sfn->monitor_box(jvms, idx);
2068           if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2069             int j = jvms->monitor_box_offset(idx);
2070             _igvn.replace_input_of(u, j, newbox);
2071             next_edge = false;
2072           }
2073         }
2074       }
2075     }
2076     if (next_edge) i++;
2077   }
2078 }
2079 
2080 //-----------------------mark_eliminated_locking_nodes-----------------------
2081 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2082   if (EliminateNestedLocks) {
2083     if (alock->is_nested()) {
2084        assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2085        return;
2086     } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2087       // Only Lock node has JVMState needed here.
2088       // Not that preceding claim is documented anywhere else.
2089       if (alock->jvms() != NULL) {
2090         if (alock->as_Lock()->is_nested_lock_region()) {
2091           // Mark eliminated related nested locks and unlocks.
2092           Node* obj = alock->obj_node();
2093           BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2094           assert(!box_node->is_eliminated(), "should not be marked yet");
2095           // Note: BoxLock node is marked eliminated only here
2096           // and it is used to indicate that all associated lock
2097           // and unlock nodes are marked for elimination.
2098           box_node->set_eliminated(); // Box's hash is always NO_HASH here
2099           for (uint i = 0; i < box_node->outcnt(); i++) {
2100             Node* u = box_node->raw_out(i);
2101             if (u->is_AbstractLock()) {
2102               alock = u->as_AbstractLock();
2103               if (alock->box_node() == box_node) {
2104                 // Verify that this Box is referenced only by related locks.
2105                 assert(alock->obj_node()->eqv_uncast(obj), "");
2106                 // Mark all related locks and unlocks.
2107 #ifdef ASSERT
2108                 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2109 #endif
2110                 alock->set_nested();
2111               }
2112             }
2113           }
2114         } else {
2115 #ifdef ASSERT
2116           alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2117           if (C->log() != NULL)
2118             alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2119 #endif
2120         }
2121       }
2122       return;
2123     }
2124     // Process locks for non escaping object
2125     assert(alock->is_non_esc_obj(), "");
2126   } // EliminateNestedLocks
2127 
2128   if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2129     // Look for all locks of this object and mark them and
2130     // corresponding BoxLock nodes as eliminated.
2131     Node* obj = alock->obj_node();
2132     for (uint j = 0; j < obj->outcnt(); j++) {
2133       Node* o = obj->raw_out(j);
2134       if (o->is_AbstractLock() &&
2135           o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2136         alock = o->as_AbstractLock();
2137         Node* box = alock->box_node();
2138         // Replace old box node with new eliminated box for all users
2139         // of the same object and mark related locks as eliminated.
2140         mark_eliminated_box(box, obj);
2141       }
2142     }
2143   }
2144 }
2145 
2146 // we have determined that this lock/unlock can be eliminated, we simply
2147 // eliminate the node without expanding it.
2148 //
2149 // Note:  The membar's associated with the lock/unlock are currently not
2150 //        eliminated.  This should be investigated as a future enhancement.
2151 //
2152 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2153 
2154   if (!alock->is_eliminated()) {
2155     return false;
2156   }
2157 #ifdef ASSERT
2158   if (!alock->is_coarsened()) {
2159     // Check that new "eliminated" BoxLock node is created.
2160     BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2161     assert(oldbox->is_eliminated(), "should be done already");
2162   }
2163 #endif
2164 
2165   alock->log_lock_optimization(C, "eliminate_lock");
2166 
2167 #ifndef PRODUCT
2168   if (PrintEliminateLocks) {
2169     if (alock->is_Lock()) {
2170       tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2171     } else {
2172       tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2173     }
2174   }
2175 #endif
2176 
2177   Node* mem  = alock->in(TypeFunc::Memory);
2178   Node* ctrl = alock->in(TypeFunc::Control);
2179   guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2180 
2181   extract_call_projections(alock);
2182   // There are 2 projections from the lock.  The lock node will
2183   // be deleted when its last use is subsumed below.
2184   assert(alock->outcnt() == 2 &&
2185          _fallthroughproj != NULL &&
2186          _memproj_fallthrough != NULL,
2187          "Unexpected projections from Lock/Unlock");
2188 
2189   Node* fallthroughproj = _fallthroughproj;
2190   Node* memproj_fallthrough = _memproj_fallthrough;
2191 
2192   // The memory projection from a lock/unlock is RawMem
2193   // The input to a Lock is merged memory, so extract its RawMem input
2194   // (unless the MergeMem has been optimized away.)
2195   if (alock->is_Lock()) {
2196     // Seach for MemBarAcquireLock node and delete it also.
2197     MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2198     assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2199     Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2200     Node* memproj = membar->proj_out(TypeFunc::Memory);
2201     _igvn.replace_node(ctrlproj, fallthroughproj);
2202     _igvn.replace_node(memproj, memproj_fallthrough);
2203 
2204     // Delete FastLock node also if this Lock node is unique user
2205     // (a loop peeling may clone a Lock node).
2206     Node* flock = alock->as_Lock()->fastlock_node();
2207     if (flock->outcnt() == 1) {
2208       assert(flock->unique_out() == alock, "sanity");
2209       _igvn.replace_node(flock, top());
2210     }
2211   }
2212 
2213   // Seach for MemBarReleaseLock node and delete it also.
2214   if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2215     MemBarNode* membar = ctrl->in(0)->as_MemBar();
2216     assert(membar->Opcode() == Op_MemBarReleaseLock &&
2217            mem->is_Proj() && membar == mem->in(0), "");
2218     _igvn.replace_node(fallthroughproj, ctrl);
2219     _igvn.replace_node(memproj_fallthrough, mem);
2220     fallthroughproj = ctrl;
2221     memproj_fallthrough = mem;
2222     ctrl = membar->in(TypeFunc::Control);
2223     mem  = membar->in(TypeFunc::Memory);
2224   }
2225 
2226   _igvn.replace_node(fallthroughproj, ctrl);
2227   _igvn.replace_node(memproj_fallthrough, mem);
2228   return true;
2229 }
2230 
2231 
2232 //------------------------------expand_lock_node----------------------
2233 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2234 
2235   Node* ctrl = lock->in(TypeFunc::Control);
2236   Node* mem = lock->in(TypeFunc::Memory);
2237   Node* obj = lock->obj_node();
2238   Node* box = lock->box_node();
2239   Node* flock = lock->fastlock_node();
2240 
2241   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2242 
2243   // Make the merge point
2244   Node *region;
2245   Node *mem_phi;
2246   Node *slow_path;
2247 
2248   if (UseOptoBiasInlining) {
2249     /*
2250      *  See the full description in MacroAssembler::biased_locking_enter().
2251      *
2252      *  if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2253      *    // The object is biased.
2254      *    proto_node = klass->prototype_header;
2255      *    o_node = thread | proto_node;
2256      *    x_node = o_node ^ mark_word;
2257      *    if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2258      *      // Done.
2259      *    } else {
2260      *      if( (x_node & biased_lock_mask) != 0 ) {
2261      *        // The klass's prototype header is no longer biased.
2262      *        cas(&mark_word, mark_word, proto_node)
2263      *        goto cas_lock;
2264      *      } else {
2265      *        // The klass's prototype header is still biased.
2266      *        if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2267      *          old = mark_word;
2268      *          new = o_node;
2269      *        } else {
2270      *          // Different thread or anonymous biased.
2271      *          old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2272      *          new = thread | old;
2273      *        }
2274      *        // Try to rebias.
2275      *        if( cas(&mark_word, old, new) == 0 ) {
2276      *          // Done.
2277      *        } else {
2278      *          goto slow_path; // Failed.
2279      *        }
2280      *      }
2281      *    }
2282      *  } else {
2283      *    // The object is not biased.
2284      *    cas_lock:
2285      *    if( FastLock(obj) == 0 ) {
2286      *      // Done.
2287      *    } else {
2288      *      slow_path:
2289      *      OptoRuntime::complete_monitor_locking_Java(obj);
2290      *    }
2291      *  }
2292      */
2293 
2294     region  = new RegionNode(5);
2295     // create a Phi for the memory state
2296     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2297 
2298     Node* fast_lock_region  = new RegionNode(3);
2299     Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2300 
2301     // First, check mark word for the biased lock pattern.
2302     Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2303 
2304     // Get fast path - mark word has the biased lock pattern.
2305     ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2306                          markOopDesc::biased_lock_mask_in_place,
2307                          markOopDesc::biased_lock_pattern, true);
2308     // fast_lock_region->in(1) is set to slow path.
2309     fast_lock_mem_phi->init_req(1, mem);
2310 
2311     // Now check that the lock is biased to the current thread and has
2312     // the same epoch and bias as Klass::_prototype_header.
2313 
2314     // Special-case a fresh allocation to avoid building nodes:
2315     Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2316     if (klass_node == NULL) {
2317       Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2318       klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2319 #ifdef _LP64
2320       if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2321         assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2322         klass_node->in(1)->init_req(0, ctrl);
2323       } else
2324 #endif
2325       klass_node->init_req(0, ctrl);
2326     }
2327     Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2328 
2329     Node* thread = transform_later(new ThreadLocalNode());
2330     Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2331     Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2332     Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2333 
2334     // Get slow path - mark word does NOT match the value.
2335     Node* not_biased_ctrl =  opt_bits_test(ctrl, region, 3, x_node,
2336                                       (~markOopDesc::age_mask_in_place), 0);
2337     // region->in(3) is set to fast path - the object is biased to the current thread.
2338     mem_phi->init_req(3, mem);
2339 
2340 
2341     // Mark word does NOT match the value (thread | Klass::_prototype_header).
2342 
2343 
2344     // First, check biased pattern.
2345     // Get fast path - _prototype_header has the same biased lock pattern.
2346     ctrl =  opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2347                           markOopDesc::biased_lock_mask_in_place, 0, true);
2348 
2349     not_biased_ctrl = fast_lock_region->in(2); // Slow path
2350     // fast_lock_region->in(2) - the prototype header is no longer biased
2351     // and we have to revoke the bias on this object.
2352     // We are going to try to reset the mark of this object to the prototype
2353     // value and fall through to the CAS-based locking scheme.
2354     Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2355     Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2356                                           proto_node, mark_node);
2357     transform_later(cas);
2358     Node* proj = transform_later(new SCMemProjNode(cas));
2359     fast_lock_mem_phi->init_req(2, proj);
2360 
2361 
2362     // Second, check epoch bits.
2363     Node* rebiased_region  = new RegionNode(3);
2364     Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2365     Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2366 
2367     // Get slow path - mark word does NOT match epoch bits.
2368     Node* epoch_ctrl =  opt_bits_test(ctrl, rebiased_region, 1, x_node,
2369                                       markOopDesc::epoch_mask_in_place, 0);
2370     // The epoch of the current bias is not valid, attempt to rebias the object
2371     // toward the current thread.
2372     rebiased_region->init_req(2, epoch_ctrl);
2373     old_phi->init_req(2, mark_node);
2374     new_phi->init_req(2, o_node);
2375 
2376     // rebiased_region->in(1) is set to fast path.
2377     // The epoch of the current bias is still valid but we know
2378     // nothing about the owner; it might be set or it might be clear.
2379     Node* cmask   = MakeConX(markOopDesc::biased_lock_mask_in_place |
2380                              markOopDesc::age_mask_in_place |
2381                              markOopDesc::epoch_mask_in_place);
2382     Node* old = transform_later(new AndXNode(mark_node, cmask));
2383     cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2384     Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2385     old_phi->init_req(1, old);
2386     new_phi->init_req(1, new_mark);
2387 
2388     transform_later(rebiased_region);
2389     transform_later(old_phi);
2390     transform_later(new_phi);
2391 
2392     // Try to acquire the bias of the object using an atomic operation.
2393     // If this fails we will go in to the runtime to revoke the object's bias.
2394     cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2395     transform_later(cas);
2396     proj = transform_later(new SCMemProjNode(cas));
2397 
2398     // Get slow path - Failed to CAS.
2399     not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2400     mem_phi->init_req(4, proj);
2401     // region->in(4) is set to fast path - the object is rebiased to the current thread.
2402 
2403     // Failed to CAS.
2404     slow_path  = new RegionNode(3);
2405     Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2406 
2407     slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2408     slow_mem->init_req(1, proj);
2409 
2410     // Call CAS-based locking scheme (FastLock node).
2411 
2412     transform_later(fast_lock_region);
2413     transform_later(fast_lock_mem_phi);
2414 
2415     // Get slow path - FastLock failed to lock the object.
2416     ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2417     mem_phi->init_req(2, fast_lock_mem_phi);
2418     // region->in(2) is set to fast path - the object is locked to the current thread.
2419 
2420     slow_path->init_req(2, ctrl); // Capture slow-control
2421     slow_mem->init_req(2, fast_lock_mem_phi);
2422 
2423     transform_later(slow_path);
2424     transform_later(slow_mem);
2425     // Reset lock's memory edge.
2426     lock->set_req(TypeFunc::Memory, slow_mem);
2427 
2428   } else {
2429     region  = new RegionNode(3);
2430     // create a Phi for the memory state
2431     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2432 
2433     // Optimize test; set region slot 2
2434     slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2435     mem_phi->init_req(2, mem);
2436   }
2437 
2438   // Make slow path call
2439   CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2440                                   OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2441                                   obj, box, NULL);
2442 
2443   extract_call_projections(call);
2444 
2445   // Slow path can only throw asynchronous exceptions, which are always
2446   // de-opted.  So the compiler thinks the slow-call can never throw an
2447   // exception.  If it DOES throw an exception we would need the debug
2448   // info removed first (since if it throws there is no monitor).
2449   assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2450            _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2451 
2452   // Capture slow path
2453   // disconnect fall-through projection from call and create a new one
2454   // hook up users of fall-through projection to region
2455   Node *slow_ctrl = _fallthroughproj->clone();
2456   transform_later(slow_ctrl);
2457   _igvn.hash_delete(_fallthroughproj);
2458   _fallthroughproj->disconnect_inputs(NULL, C);
2459   region->init_req(1, slow_ctrl);
2460   // region inputs are now complete
2461   transform_later(region);
2462   _igvn.replace_node(_fallthroughproj, region);
2463 
2464   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2465   mem_phi->init_req(1, memproj );
2466   transform_later(mem_phi);
2467   _igvn.replace_node(_memproj_fallthrough, mem_phi);
2468 }
2469 
2470 //------------------------------expand_unlock_node----------------------
2471 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2472 
2473   Node* ctrl = unlock->in(TypeFunc::Control);
2474   Node* mem = unlock->in(TypeFunc::Memory);
2475   Node* obj = unlock->obj_node();
2476   Node* box = unlock->box_node();
2477 
2478   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2479 
2480   // No need for a null check on unlock
2481 
2482   // Make the merge point
2483   Node *region;
2484   Node *mem_phi;
2485 
2486   if (UseOptoBiasInlining) {
2487     // Check for biased locking unlock case, which is a no-op.
2488     // See the full description in MacroAssembler::biased_locking_exit().
2489     region  = new RegionNode(4);
2490     // create a Phi for the memory state
2491     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2492     mem_phi->init_req(3, mem);
2493 
2494     Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2495     ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2496                          markOopDesc::biased_lock_mask_in_place,
2497                          markOopDesc::biased_lock_pattern);
2498   } else {
2499     region  = new RegionNode(3);
2500     // create a Phi for the memory state
2501     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2502   }
2503 
2504   FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2505   funlock = transform_later( funlock )->as_FastUnlock();
2506   // Optimize test; set region slot 2
2507   Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2508   Node *thread = transform_later(new ThreadLocalNode());
2509 
2510   CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2511                                   CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2512                                   "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2513 
2514   extract_call_projections(call);
2515 
2516   assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2517            _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2518 
2519   // No exceptions for unlocking
2520   // Capture slow path
2521   // disconnect fall-through projection from call and create a new one
2522   // hook up users of fall-through projection to region
2523   Node *slow_ctrl = _fallthroughproj->clone();
2524   transform_later(slow_ctrl);
2525   _igvn.hash_delete(_fallthroughproj);
2526   _fallthroughproj->disconnect_inputs(NULL, C);
2527   region->init_req(1, slow_ctrl);
2528   // region inputs are now complete
2529   transform_later(region);
2530   _igvn.replace_node(_fallthroughproj, region);
2531 
2532   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2533   mem_phi->init_req(1, memproj );
2534   mem_phi->init_req(2, mem);
2535   transform_later(mem_phi);
2536   _igvn.replace_node(_memproj_fallthrough, mem_phi);
2537 }
2538 
2539 //---------------------------eliminate_macro_nodes----------------------
2540 // Eliminate scalar replaced allocations and associated locks.
2541 void PhaseMacroExpand::eliminate_macro_nodes() {
2542   if (C->macro_count() == 0)
2543     return;
2544 
2545   // First, attempt to eliminate locks
2546   int cnt = C->macro_count();
2547   for (int i=0; i < cnt; i++) {
2548     Node *n = C->macro_node(i);
2549     if (n->is_AbstractLock()) { // Lock and Unlock nodes
2550       // Before elimination mark all associated (same box and obj)
2551       // lock and unlock nodes.
2552       mark_eliminated_locking_nodes(n->as_AbstractLock());
2553     }
2554   }
2555   bool progress = true;
2556   while (progress) {
2557     progress = false;
2558     for (int i = C->macro_count(); i > 0; i--) {
2559       Node * n = C->macro_node(i-1);
2560       bool success = false;
2561       debug_only(int old_macro_count = C->macro_count(););
2562       if (n->is_AbstractLock()) {
2563         success = eliminate_locking_node(n->as_AbstractLock());
2564       }
2565       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2566       progress = progress || success;
2567     }
2568   }
2569   // Next, attempt to eliminate allocations
2570   _has_locks = false;
2571   progress = true;
2572   while (progress) {
2573     progress = false;
2574     for (int i = C->macro_count(); i > 0; i--) {
2575       Node * n = C->macro_node(i-1);
2576       bool success = false;
2577       debug_only(int old_macro_count = C->macro_count(););
2578       switch (n->class_id()) {
2579       case Node::Class_Allocate:
2580       case Node::Class_AllocateArray:
2581         success = eliminate_allocate_node(n->as_Allocate());
2582         break;
2583       case Node::Class_CallStaticJava:
2584         success = eliminate_boxing_node(n->as_CallStaticJava());
2585         break;
2586       case Node::Class_Lock:
2587       case Node::Class_Unlock:
2588         assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2589         _has_locks = true;
2590         break;
2591       case Node::Class_ArrayCopy:
2592         break;
2593       case Node::Class_OuterStripMinedLoop:
2594         break;
2595       default:
2596         assert(n->Opcode() == Op_LoopLimit ||
2597                n->Opcode() == Op_Opaque1   ||
2598                n->Opcode() == Op_Opaque2   ||
2599                n->Opcode() == Op_Opaque3   ||
2600                BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2601                "unknown node type in macro list");
2602       }
2603       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2604       progress = progress || success;
2605     }
2606   }
2607 }
2608 
2609 //------------------------------expand_macro_nodes----------------------
2610 //  Returns true if a failure occurred.
2611 bool PhaseMacroExpand::expand_macro_nodes() {
2612   // Last attempt to eliminate macro nodes.
2613   eliminate_macro_nodes();
2614 
2615   // Make sure expansion will not cause node limit to be exceeded.
2616   // Worst case is a macro node gets expanded into about 200 nodes.
2617   // Allow 50% more for optimization.
2618   if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2619     return true;
2620 
2621   // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2622   bool progress = true;
2623   while (progress) {
2624     progress = false;
2625     for (int i = C->macro_count(); i > 0; i--) {
2626       Node * n = C->macro_node(i-1);
2627       bool success = false;
2628       debug_only(int old_macro_count = C->macro_count(););
2629       if (n->Opcode() == Op_LoopLimit) {
2630         // Remove it from macro list and put on IGVN worklist to optimize.
2631         C->remove_macro_node(n);
2632         _igvn._worklist.push(n);
2633         success = true;
2634       } else if (n->Opcode() == Op_CallStaticJava) {
2635         // Remove it from macro list and put on IGVN worklist to optimize.
2636         C->remove_macro_node(n);
2637         _igvn._worklist.push(n);
2638         success = true;
2639       } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2640         _igvn.replace_node(n, n->in(1));
2641         success = true;
2642 #if INCLUDE_RTM_OPT
2643       } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2644         assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2645         assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2646         Node* cmp = n->unique_out();
2647 #ifdef ASSERT
2648         // Validate graph.
2649         assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2650         BoolNode* bol = cmp->unique_out()->as_Bool();
2651         assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2652                (bol->_test._test == BoolTest::ne), "");
2653         IfNode* ifn = bol->unique_out()->as_If();
2654         assert((ifn->outcnt() == 2) &&
2655                ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2656 #endif
2657         Node* repl = n->in(1);
2658         if (!_has_locks) {
2659           // Remove RTM state check if there are no locks in the code.
2660           // Replace input to compare the same value.
2661           repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2662         }
2663         _igvn.replace_node(n, repl);
2664         success = true;
2665 #endif
2666       } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2667         n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2668         C->remove_macro_node(n);
2669         success = true;
2670       }
2671       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2672       progress = progress || success;
2673     }
2674   }
2675 
2676   // expand arraycopy "macro" nodes first
2677   // For ReduceBulkZeroing, we must first process all arraycopy nodes
2678   // before the allocate nodes are expanded.
2679   int macro_idx = C->macro_count() - 1;
2680   while (macro_idx >= 0) {
2681     Node * n = C->macro_node(macro_idx);
2682     assert(n->is_macro(), "only macro nodes expected here");
2683     if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2684       // node is unreachable, so don't try to expand it
2685       C->remove_macro_node(n);
2686     } else if (n->is_ArrayCopy()){
2687       int macro_count = C->macro_count();
2688       expand_arraycopy_node(n->as_ArrayCopy());
2689       assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2690     }
2691     if (C->failing())  return true;
2692     macro_idx --;
2693   }
2694 
2695   // expand "macro" nodes
2696   // nodes are removed from the macro list as they are processed
2697   while (C->macro_count() > 0) {
2698     int macro_count = C->macro_count();
2699     Node * n = C->macro_node(macro_count-1);
2700     assert(n->is_macro(), "only macro nodes expected here");
2701     if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2702       // node is unreachable, so don't try to expand it
2703       C->remove_macro_node(n);
2704       continue;
2705     }
2706     switch (n->class_id()) {
2707     case Node::Class_Allocate:
2708       expand_allocate(n->as_Allocate());
2709       break;
2710     case Node::Class_AllocateArray:
2711       expand_allocate_array(n->as_AllocateArray());
2712       break;
2713     case Node::Class_Lock:
2714       expand_lock_node(n->as_Lock());
2715       break;
2716     case Node::Class_Unlock:
2717       expand_unlock_node(n->as_Unlock());
2718       break;
2719     default:
2720       assert(false, "unknown node type in macro list");
2721     }
2722     assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2723     if (C->failing())  return true;
2724   }
2725 
2726   _igvn.set_delay_transform(false);
2727   _igvn.optimize();
2728   if (C->failing())  return true;
2729   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2730   return bs->expand_macro_nodes(this);
2731 }