1 /*
   2  * Copyright (c) 2005, 2015, 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 "libadt/vectset.hpp"
  28 #include "opto/addnode.hpp"
  29 #include "opto/arraycopynode.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/castnode.hpp"
  32 #include "opto/cfgnode.hpp"
  33 #include "opto/compile.hpp"
  34 #include "opto/convertnode.hpp"
  35 #include "opto/locknode.hpp"
  36 #include "opto/loopnode.hpp"
  37 #include "opto/macro.hpp"
  38 #include "opto/memnode.hpp"
  39 #include "opto/narrowptrnode.hpp"
  40 #include "opto/node.hpp"
  41 #include "opto/opaquenode.hpp"
  42 #include "opto/phaseX.hpp"
  43 #include "opto/rootnode.hpp"
  44 #include "opto/runtime.hpp"
  45 #include "opto/subnode.hpp"
  46 #include "opto/type.hpp"
  47 #include "runtime/sharedRuntime.hpp"
  48 
  49 
  50 //
  51 // Replace any references to "oldref" in inputs to "use" with "newref".
  52 // Returns the number of replacements made.
  53 //
  54 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
  55   int nreplacements = 0;
  56   uint req = use->req();
  57   for (uint j = 0; j < use->len(); j++) {
  58     Node *uin = use->in(j);
  59     if (uin == oldref) {
  60       if (j < req)
  61         use->set_req(j, newref);
  62       else
  63         use->set_prec(j, newref);
  64       nreplacements++;
  65     } else if (j >= req && uin == NULL) {
  66       break;
  67     }
  68   }
  69   return nreplacements;
  70 }
  71 
  72 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
  73   // Copy debug information and adjust JVMState information
  74   uint old_dbg_start = oldcall->tf()->domain()->cnt();
  75   uint new_dbg_start = newcall->tf()->domain()->cnt();
  76   int jvms_adj  = new_dbg_start - old_dbg_start;
  77   assert (new_dbg_start == newcall->req(), "argument count mismatch");
  78 
  79   // SafePointScalarObject node could be referenced several times in debug info.
  80   // Use Dict to record cloned nodes.
  81   Dict* sosn_map = new Dict(cmpkey,hashkey);
  82   for (uint i = old_dbg_start; i < oldcall->req(); i++) {
  83     Node* old_in = oldcall->in(i);
  84     // Clone old SafePointScalarObjectNodes, adjusting their field contents.
  85     if (old_in != NULL && old_in->is_SafePointScalarObject()) {
  86       SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
  87       uint old_unique = C->unique();
  88       Node* new_in = old_sosn->clone(sosn_map);
  89       if (old_unique != C->unique()) { // New node?
  90         new_in->set_req(0, C->root()); // reset control edge
  91         new_in = transform_later(new_in); // Register new node.
  92       }
  93       old_in = new_in;
  94     }
  95     newcall->add_req(old_in);
  96   }
  97 
  98   // JVMS may be shared so clone it before we modify it
  99   newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
 100   for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
 101     jvms->set_map(newcall);
 102     jvms->set_locoff(jvms->locoff()+jvms_adj);
 103     jvms->set_stkoff(jvms->stkoff()+jvms_adj);
 104     jvms->set_monoff(jvms->monoff()+jvms_adj);
 105     jvms->set_scloff(jvms->scloff()+jvms_adj);
 106     jvms->set_endoff(jvms->endoff()+jvms_adj);
 107   }
 108 }
 109 
 110 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
 111   Node* cmp;
 112   if (mask != 0) {
 113     Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
 114     cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
 115   } else {
 116     cmp = word;
 117   }
 118   Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
 119   IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
 120   transform_later(iff);
 121 
 122   // Fast path taken.
 123   Node *fast_taken = transform_later(new IfFalseNode(iff));
 124 
 125   // Fast path not-taken, i.e. slow path
 126   Node *slow_taken = transform_later(new IfTrueNode(iff));
 127 
 128   if (return_fast_path) {
 129     region->init_req(edge, slow_taken); // Capture slow-control
 130     return fast_taken;
 131   } else {
 132     region->init_req(edge, fast_taken); // Capture fast-control
 133     return slow_taken;
 134   }
 135 }
 136 
 137 //--------------------copy_predefined_input_for_runtime_call--------------------
 138 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
 139   // Set fixed predefined input arguments
 140   call->init_req( TypeFunc::Control, ctrl );
 141   call->init_req( TypeFunc::I_O    , oldcall->in( TypeFunc::I_O) );
 142   call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
 143   call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
 144   call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
 145 }
 146 
 147 //------------------------------make_slow_call---------------------------------
 148 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
 149                                            address slow_call, const char* leaf_name, Node* slow_path,
 150                                            Node* parm0, Node* parm1, Node* parm2) {
 151 
 152   // Slow-path call
 153  CallNode *call = leaf_name
 154    ? (CallNode*)new CallLeafNode      ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
 155    : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
 156 
 157   // Slow path call has no side-effects, uses few values
 158   copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
 159   if (parm0 != NULL)  call->init_req(TypeFunc::Parms+0, parm0);
 160   if (parm1 != NULL)  call->init_req(TypeFunc::Parms+1, parm1);
 161   if (parm2 != NULL)  call->init_req(TypeFunc::Parms+2, parm2);
 162   copy_call_debug_info(oldcall, call);
 163   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
 164   _igvn.replace_node(oldcall, call);
 165   transform_later(call);
 166 
 167   return call;
 168 }
 169 
 170 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
 171   _fallthroughproj = NULL;
 172   _fallthroughcatchproj = NULL;
 173   _ioproj_fallthrough = NULL;
 174   _ioproj_catchall = NULL;
 175   _catchallcatchproj = NULL;
 176   _memproj_fallthrough = NULL;
 177   _memproj_catchall = NULL;
 178   _resproj = NULL;
 179   for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
 180     ProjNode *pn = call->fast_out(i)->as_Proj();
 181     switch (pn->_con) {
 182       case TypeFunc::Control:
 183       {
 184         // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
 185         _fallthroughproj = pn;
 186         DUIterator_Fast jmax, j = pn->fast_outs(jmax);
 187         const Node *cn = pn->fast_out(j);
 188         if (cn->is_Catch()) {
 189           ProjNode *cpn = NULL;
 190           for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
 191             cpn = cn->fast_out(k)->as_Proj();
 192             assert(cpn->is_CatchProj(), "must be a CatchProjNode");
 193             if (cpn->_con == CatchProjNode::fall_through_index)
 194               _fallthroughcatchproj = cpn;
 195             else {
 196               assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
 197               _catchallcatchproj = cpn;
 198             }
 199           }
 200         }
 201         break;
 202       }
 203       case TypeFunc::I_O:
 204         if (pn->_is_io_use)
 205           _ioproj_catchall = pn;
 206         else
 207           _ioproj_fallthrough = pn;
 208         break;
 209       case TypeFunc::Memory:
 210         if (pn->_is_io_use)
 211           _memproj_catchall = pn;
 212         else
 213           _memproj_fallthrough = pn;
 214         break;
 215       case TypeFunc::Parms:
 216         _resproj = pn;
 217         break;
 218       default:
 219         assert(false, "unexpected projection from allocation node.");
 220     }
 221   }
 222 
 223 }
 224 
 225 // Eliminate a card mark sequence.  p2x is a ConvP2XNode
 226 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
 227   assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
 228   if (!UseG1GC) {
 229     // vanilla/CMS post barrier
 230     Node *shift = p2x->unique_out();
 231     Node *addp = shift->unique_out();
 232     for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
 233       Node *mem = addp->last_out(j);
 234       if (UseCondCardMark && mem->is_Load()) {
 235         assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
 236         // The load is checking if the card has been written so
 237         // replace it with zero to fold the test.
 238         _igvn.replace_node(mem, intcon(0));
 239         continue;
 240       }
 241       assert(mem->is_Store(), "store required");
 242       _igvn.replace_node(mem, mem->in(MemNode::Memory));
 243     }
 244   } else {
 245     // G1 pre/post barriers
 246     assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
 247     // It could be only one user, URShift node, in Object.clone() instrinsic
 248     // but the new allocation is passed to arraycopy stub and it could not
 249     // be scalar replaced. So we don't check the case.
 250 
 251     // An other case of only one user (Xor) is when the value check for NULL
 252     // in G1 post barrier is folded after CCP so the code which used URShift
 253     // is removed.
 254 
 255     // Take Region node before eliminating post barrier since it also
 256     // eliminates CastP2X node when it has only one user.
 257     Node* this_region = p2x->in(0);
 258     assert(this_region != NULL, "");
 259 
 260     // Remove G1 post barrier.
 261 
 262     // Search for CastP2X->Xor->URShift->Cmp path which
 263     // checks if the store done to a different from the value's region.
 264     // And replace Cmp with #0 (false) to collapse G1 post barrier.
 265     Node* xorx = p2x->find_out_with(Op_XorX);
 266     assert(xorx != NULL, "missing G1 post barrier");
 267     Node* shift = xorx->unique_out();
 268     Node* cmpx = shift->unique_out();
 269     assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
 270     cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
 271     "missing region check in G1 post barrier");
 272     _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
 273 
 274     // Remove G1 pre barrier.
 275 
 276     // Search "if (marking != 0)" check and set it to "false".
 277     // There is no G1 pre barrier if previous stored value is NULL
 278     // (for example, after initialization).
 279     if (this_region->is_Region() && this_region->req() == 3) {
 280       int ind = 1;
 281       if (!this_region->in(ind)->is_IfFalse()) {
 282         ind = 2;
 283       }
 284       if (this_region->in(ind)->is_IfFalse()) {
 285         Node* bol = this_region->in(ind)->in(0)->in(1);
 286         assert(bol->is_Bool(), "");
 287         cmpx = bol->in(1);
 288         if (bol->as_Bool()->_test._test == BoolTest::ne &&
 289             cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
 290             cmpx->in(1)->is_Load()) {
 291           Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
 292           const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
 293                                               PtrQueue::byte_offset_of_active());
 294           if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
 295               adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 296               adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
 297             _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
 298           }
 299         }
 300       }
 301     }
 302     // Now CastP2X can be removed since it is used only on dead path
 303     // which currently still alive until igvn optimize it.
 304     assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
 305     _igvn.replace_node(p2x, top());
 306   }
 307 }
 308 
 309 // Search for a memory operation for the specified memory slice.
 310 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
 311   Node *orig_mem = mem;
 312   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 313   const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
 314   while (true) {
 315     if (mem == alloc_mem || mem == start_mem ) {
 316       return mem;  // hit one of our sentinels
 317     } else if (mem->is_MergeMem()) {
 318       mem = mem->as_MergeMem()->memory_at(alias_idx);
 319     } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
 320       Node *in = mem->in(0);
 321       // we can safely skip over safepoints, calls, locks and membars because we
 322       // already know that the object is safe to eliminate.
 323       if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
 324         return in;
 325       } else if (in->is_Call()) {
 326         CallNode *call = in->as_Call();
 327         if (call->may_modify(tinst, phase)) {
 328           assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
 329 
 330           if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
 331             return in;
 332           }
 333         }
 334         mem = in->in(TypeFunc::Memory);
 335       } else if (in->is_MemBar()) {
 336         if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase)) {
 337           assert(in->in(0)->is_Proj() && in->in(0)->in(0)->is_ArrayCopy(), "should be arraycopy");
 338           ArrayCopyNode* ac = in->in(0)->in(0)->as_ArrayCopy();
 339           assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
 340           return ac;
 341         }
 342         mem = in->in(TypeFunc::Memory);
 343       } else {
 344         assert(false, "unexpected projection");
 345       }
 346     } else if (mem->is_Store()) {
 347       const TypePtr* atype = mem->as_Store()->adr_type();
 348       int adr_idx = phase->C->get_alias_index(atype);
 349       if (adr_idx == alias_idx) {
 350         assert(atype->isa_oopptr(), "address type must be oopptr");
 351         int adr_offset = atype->offset();
 352         uint adr_iid = atype->is_oopptr()->instance_id();
 353         // Array elements references have the same alias_idx
 354         // but different offset and different instance_id.
 355         if (adr_offset == offset && adr_iid == alloc->_idx)
 356           return mem;
 357       } else {
 358         assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
 359       }
 360       mem = mem->in(MemNode::Memory);
 361     } else if (mem->is_ClearArray()) {
 362       if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
 363         // Can not bypass initialization of the instance
 364         // we are looking.
 365         debug_only(intptr_t offset;)
 366         assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
 367         InitializeNode* init = alloc->as_Allocate()->initialization();
 368         // We are looking for stored value, return Initialize node
 369         // or memory edge from Allocate node.
 370         if (init != NULL)
 371           return init;
 372         else
 373           return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
 374       }
 375       // Otherwise skip it (the call updated 'mem' value).
 376     } else if (mem->Opcode() == Op_SCMemProj) {
 377       mem = mem->in(0);
 378       Node* adr = NULL;
 379       if (mem->is_LoadStore()) {
 380         adr = mem->in(MemNode::Address);
 381       } else {
 382         assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
 383         adr = mem->in(3); // Destination array
 384       }
 385       const TypePtr* atype = adr->bottom_type()->is_ptr();
 386       int adr_idx = phase->C->get_alias_index(atype);
 387       if (adr_idx == alias_idx) {
 388         assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
 389         return NULL;
 390       }
 391       mem = mem->in(MemNode::Memory);
 392     } else {
 393       return mem;
 394     }
 395     assert(mem != orig_mem, "dead memory loop");
 396   }
 397 }
 398 
 399 // Generate loads from source of the arraycopy for fields of
 400 // destination needed at a deoptimization point
 401 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, BasicType ft, const Type *ftype, AllocateNode *alloc) {
 402   BasicType bt = ft;
 403   const Type *type = ftype;
 404   if (ft == T_NARROWOOP) {
 405     bt = T_OBJECT;
 406     type = ftype->make_oopptr();
 407   }
 408   Node* res = NULL;
 409   if (ac->is_clonebasic()) {
 410     Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
 411     Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
 412     const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
 413     Node* m = ac->in(TypeFunc::Memory);
 414     while (m->is_MergeMem()) {
 415       m = m->as_MergeMem()->memory_at(C->get_alias_index(adr_type));
 416       if (m->is_Proj() && m->in(0)->is_MemBar()) {
 417         m = m->in(0)->in(TypeFunc::Memory);
 418       }
 419     }
 420     res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
 421   } else {
 422     if (ac->modifies(offset, offset, &_igvn, true)) {
 423       assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
 424       uint shift  = exact_log2(type2aelembytes(bt));
 425       Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
 426 #ifdef _LP64
 427       diff = _igvn.transform(new ConvI2LNode(diff));
 428 #endif
 429       diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
 430 
 431       Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
 432       Node* base = ac->in(ArrayCopyNode::Src);
 433       Node* adr = _igvn.transform(new AddPNode(base, base, off));
 434       const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
 435       Node* m = ac->in(TypeFunc::Memory);
 436       res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
 437     }
 438   }
 439   if (res != NULL) {
 440     res = _igvn.transform(res);
 441     if (ftype->isa_narrowoop()) {
 442       // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
 443       res = _igvn.transform(new EncodePNode(res, ftype));
 444     }
 445     return res;
 446   }
 447   return NULL;
 448 }
 449 
 450 //
 451 // Given a Memory Phi, compute a value Phi containing the values from stores
 452 // on the input paths.
 453 // Note: this function is recursive, its depth is limited by the "level" argument
 454 // Returns the computed Phi, or NULL if it cannot compute it.
 455 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) {
 456   assert(mem->is_Phi(), "sanity");
 457   int alias_idx = C->get_alias_index(adr_t);
 458   int offset = adr_t->offset();
 459   int instance_id = adr_t->instance_id();
 460 
 461   // Check if an appropriate value phi already exists.
 462   Node* region = mem->in(0);
 463   for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
 464     Node* phi = region->fast_out(k);
 465     if (phi->is_Phi() && phi != mem &&
 466         phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
 467       return phi;
 468     }
 469   }
 470   // Check if an appropriate new value phi already exists.
 471   Node* new_phi = value_phis->find(mem->_idx);
 472   if (new_phi != NULL)
 473     return new_phi;
 474 
 475   if (level <= 0) {
 476     return NULL; // Give up: phi tree too deep
 477   }
 478   Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
 479   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 480 
 481   uint length = mem->req();
 482   GrowableArray <Node *> values(length, length, NULL, false);
 483 
 484   // create a new Phi for the value
 485   PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
 486   transform_later(phi);
 487   value_phis->push(phi, mem->_idx);
 488 
 489   for (uint j = 1; j < length; j++) {
 490     Node *in = mem->in(j);
 491     if (in == NULL || in->is_top()) {
 492       values.at_put(j, in);
 493     } else  {
 494       Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
 495       if (val == start_mem || val == alloc_mem) {
 496         // hit a sentinel, return appropriate 0 value
 497         values.at_put(j, _igvn.zerocon(ft));
 498         continue;
 499       }
 500       if (val->is_Initialize()) {
 501         val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 502       }
 503       if (val == NULL) {
 504         return NULL;  // can't find a value on this path
 505       }
 506       if (val == mem) {
 507         values.at_put(j, mem);
 508       } else if (val->is_Store()) {
 509         values.at_put(j, val->in(MemNode::ValueIn));
 510       } else if(val->is_Proj() && val->in(0) == alloc) {
 511         values.at_put(j, _igvn.zerocon(ft));
 512       } else if (val->is_Phi()) {
 513         val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
 514         if (val == NULL) {
 515           return NULL;
 516         }
 517         values.at_put(j, val);
 518       } else if (val->Opcode() == Op_SCMemProj) {
 519         assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity");
 520         assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
 521         return NULL;
 522       } else if (val->is_ArrayCopy()) {
 523         Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), ft, phi_type, alloc);
 524         if (res == NULL) {
 525           return NULL;
 526         }
 527         values.at_put(j, res);
 528       } else {
 529 #ifdef ASSERT
 530         val->dump();
 531         assert(false, "unknown node on this path");
 532 #endif
 533         return NULL;  // unknown node on this path
 534       }
 535     }
 536   }
 537   // Set Phi's inputs
 538   for (uint j = 1; j < length; j++) {
 539     if (values.at(j) == mem) {
 540       phi->init_req(j, phi);
 541     } else {
 542       phi->init_req(j, values.at(j));
 543     }
 544   }
 545   return phi;
 546 }
 547 
 548 // Search the last value stored into the object's field.
 549 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
 550   assert(adr_t->is_known_instance_field(), "instance required");
 551   int instance_id = adr_t->instance_id();
 552   assert((uint)instance_id == alloc->_idx, "wrong allocation");
 553 
 554   int alias_idx = C->get_alias_index(adr_t);
 555   int offset = adr_t->offset();
 556   Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
 557   Node *alloc_ctrl = alloc->in(TypeFunc::Control);
 558   Node *alloc_mem = alloc->in(TypeFunc::Memory);
 559   Arena *a = Thread::current()->resource_area();
 560   VectorSet visited(a);
 561 
 562 
 563   bool done = sfpt_mem == alloc_mem;
 564   Node *mem = sfpt_mem;
 565   while (!done) {
 566     if (visited.test_set(mem->_idx)) {
 567       return NULL;  // found a loop, give up
 568     }
 569     mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
 570     if (mem == start_mem || mem == alloc_mem) {
 571       done = true;  // hit a sentinel, return appropriate 0 value
 572     } else if (mem->is_Initialize()) {
 573       mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
 574       if (mem == NULL) {
 575         done = true; // Something go wrong.
 576       } else if (mem->is_Store()) {
 577         const TypePtr* atype = mem->as_Store()->adr_type();
 578         assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
 579         done = true;
 580       }
 581     } else if (mem->is_Store()) {
 582       const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
 583       assert(atype != NULL, "address type must be oopptr");
 584       assert(C->get_alias_index(atype) == alias_idx &&
 585              atype->is_known_instance_field() && atype->offset() == offset &&
 586              atype->instance_id() == instance_id, "store is correct memory slice");
 587       done = true;
 588     } else if (mem->is_Phi()) {
 589       // try to find a phi's unique input
 590       Node *unique_input = NULL;
 591       Node *top = C->top();
 592       for (uint i = 1; i < mem->req(); i++) {
 593         Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
 594         if (n == NULL || n == top || n == mem) {
 595           continue;
 596         } else if (unique_input == NULL) {
 597           unique_input = n;
 598         } else if (unique_input != n) {
 599           unique_input = top;
 600           break;
 601         }
 602       }
 603       if (unique_input != NULL && unique_input != top) {
 604         mem = unique_input;
 605       } else {
 606         done = true;
 607       }
 608     } else if (mem->is_ArrayCopy()) {
 609       done = true;
 610     } else {
 611       assert(false, "unexpected node");
 612     }
 613   }
 614   if (mem != NULL) {
 615     if (mem == start_mem || mem == alloc_mem) {
 616       // hit a sentinel, return appropriate 0 value
 617       return _igvn.zerocon(ft);
 618     } else if (mem->is_Store()) {
 619       return mem->in(MemNode::ValueIn);
 620     } else if (mem->is_Phi()) {
 621       // attempt to produce a Phi reflecting the values on the input paths of the Phi
 622       Node_Stack value_phis(a, 8);
 623       Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
 624       if (phi != NULL) {
 625         return phi;
 626       } else {
 627         // Kill all new Phis
 628         while(value_phis.is_nonempty()) {
 629           Node* n = value_phis.node();
 630           _igvn.replace_node(n, C->top());
 631           value_phis.pop();
 632         }
 633       }
 634     } else if (mem->is_ArrayCopy()) {
 635       Node* ctl = mem->in(0);
 636       if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
 637         // pin the loads in the uncommon trap path
 638         ctl = sfpt_ctl;
 639       }
 640       return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, ft, ftype, alloc);
 641     }
 642   }
 643   // Something go wrong.
 644   return NULL;
 645 }
 646 
 647 // Check the possibility of scalar replacement.
 648 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
 649   //  Scan the uses of the allocation to check for anything that would
 650   //  prevent us from eliminating it.
 651   NOT_PRODUCT( const char* fail_eliminate = NULL; )
 652   DEBUG_ONLY( Node* disq_node = NULL; )
 653   bool  can_eliminate = true;
 654 
 655   Node* res = alloc->result_cast();
 656   const TypeOopPtr* res_type = NULL;
 657   if (res == NULL) {
 658     // All users were eliminated.
 659   } else if (!res->is_CheckCastPP()) {
 660     NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
 661     can_eliminate = false;
 662   } else {
 663     res_type = _igvn.type(res)->isa_oopptr();
 664     if (res_type == NULL) {
 665       NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
 666       can_eliminate = false;
 667     } else if (res_type->isa_aryptr()) {
 668       int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 669       if (length < 0) {
 670         NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
 671         can_eliminate = false;
 672       }
 673     }
 674   }
 675 
 676   if (can_eliminate && res != NULL) {
 677     for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
 678                                j < jmax && can_eliminate; j++) {
 679       Node* use = res->fast_out(j);
 680 
 681       if (use->is_AddP()) {
 682         const TypePtr* addp_type = _igvn.type(use)->is_ptr();
 683         int offset = addp_type->offset();
 684 
 685         if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
 686           NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
 687           can_eliminate = false;
 688           break;
 689         }
 690         for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
 691                                    k < kmax && can_eliminate; k++) {
 692           Node* n = use->fast_out(k);
 693           if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
 694               !(n->is_ArrayCopy() &&
 695                 n->as_ArrayCopy()->is_clonebasic() &&
 696                 n->in(ArrayCopyNode::Dest) == use)) {
 697             DEBUG_ONLY(disq_node = n;)
 698             if (n->is_Load() || n->is_LoadStore()) {
 699               NOT_PRODUCT(fail_eliminate = "Field load";)
 700             } else {
 701               NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
 702             }
 703             can_eliminate = false;
 704           }
 705         }
 706       } else if (use->is_ArrayCopy() &&
 707                  (use->as_ArrayCopy()->is_arraycopy_validated() ||
 708                   use->as_ArrayCopy()->is_copyof_validated() ||
 709                   use->as_ArrayCopy()->is_copyofrange_validated()) &&
 710                  use->in(ArrayCopyNode::Dest) == res) {
 711         // ok to eliminate
 712       } else if (use->is_SafePoint()) {
 713         SafePointNode* sfpt = use->as_SafePoint();
 714         if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
 715           // Object is passed as argument.
 716           DEBUG_ONLY(disq_node = use;)
 717           NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
 718           can_eliminate = false;
 719         }
 720         Node* sfptMem = sfpt->memory();
 721         if (sfptMem == NULL || sfptMem->is_top()) {
 722           DEBUG_ONLY(disq_node = use;)
 723           NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
 724           can_eliminate = false;
 725         } else {
 726           safepoints.append_if_missing(sfpt);
 727         }
 728       } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
 729         if (use->is_Phi()) {
 730           if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
 731             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 732           } else {
 733             NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
 734           }
 735           DEBUG_ONLY(disq_node = use;)
 736         } else {
 737           if (use->Opcode() == Op_Return) {
 738             NOT_PRODUCT(fail_eliminate = "Object is return value";)
 739           }else {
 740             NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
 741           }
 742           DEBUG_ONLY(disq_node = use;)
 743         }
 744         can_eliminate = false;
 745       }
 746     }
 747   }
 748 
 749 #ifndef PRODUCT
 750   if (PrintEliminateAllocations) {
 751     if (can_eliminate) {
 752       tty->print("Scalar ");
 753       if (res == NULL)
 754         alloc->dump();
 755       else
 756         res->dump();
 757     } else if (alloc->_is_scalar_replaceable) {
 758       tty->print("NotScalar (%s)", fail_eliminate);
 759       if (res == NULL)
 760         alloc->dump();
 761       else
 762         res->dump();
 763 #ifdef ASSERT
 764       if (disq_node != NULL) {
 765           tty->print("  >>>> ");
 766           disq_node->dump();
 767       }
 768 #endif /*ASSERT*/
 769     }
 770   }
 771 #endif
 772   return can_eliminate;
 773 }
 774 
 775 // Do scalar replacement.
 776 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
 777   GrowableArray <SafePointNode *> safepoints_done;
 778 
 779   ciKlass* klass = NULL;
 780   ciInstanceKlass* iklass = NULL;
 781   int nfields = 0;
 782   int array_base;
 783   int element_size;
 784   BasicType basic_elem_type;
 785   ciType* elem_type;
 786 
 787   Node* res = alloc->result_cast();
 788   assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
 789   const TypeOopPtr* res_type = NULL;
 790   if (res != NULL) { // Could be NULL when there are no users
 791     res_type = _igvn.type(res)->isa_oopptr();
 792   }
 793 
 794   if (res != NULL) {
 795     klass = res_type->klass();
 796     if (res_type->isa_instptr()) {
 797       // find the fields of the class which will be needed for safepoint debug information
 798       assert(klass->is_instance_klass(), "must be an instance klass.");
 799       iklass = klass->as_instance_klass();
 800       nfields = iklass->nof_nonstatic_fields();
 801     } else {
 802       // find the array's elements which will be needed for safepoint debug information
 803       nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
 804       assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
 805       elem_type = klass->as_array_klass()->element_type();
 806       basic_elem_type = elem_type->basic_type();
 807       array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 808       element_size = type2aelembytes(basic_elem_type);
 809     }
 810   }
 811   //
 812   // Process the safepoint uses
 813   //
 814   while (safepoints.length() > 0) {
 815     SafePointNode* sfpt = safepoints.pop();
 816     Node* mem = sfpt->memory();
 817     Node* ctl = sfpt->control();
 818     assert(sfpt->jvms() != NULL, "missed JVMS");
 819     // Fields of scalar objs are referenced only at the end
 820     // of regular debuginfo at the last (youngest) JVMS.
 821     // Record relative start index.
 822     uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
 823     SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
 824 #ifdef ASSERT
 825                                                  alloc,
 826 #endif
 827                                                  first_ind, nfields);
 828     sobj->init_req(0, C->root());
 829     transform_later(sobj);
 830 
 831     // Scan object's fields adding an input to the safepoint for each field.
 832     for (int j = 0; j < nfields; j++) {
 833       intptr_t offset;
 834       ciField* field = NULL;
 835       if (iklass != NULL) {
 836         field = iklass->nonstatic_field_at(j);
 837         offset = field->offset();
 838         elem_type = field->type();
 839         basic_elem_type = field->layout_type();
 840       } else {
 841         offset = array_base + j * (intptr_t)element_size;
 842       }
 843 
 844       const Type *field_type;
 845       // The next code is taken from Parse::do_get_xxx().
 846       if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
 847         if (!elem_type->is_loaded()) {
 848           field_type = TypeInstPtr::BOTTOM;
 849         } else if (field != NULL && field->is_constant() && field->is_static()) {
 850           // This can happen if the constant oop is non-perm.
 851           ciObject* con = field->constant_value().as_object();
 852           // Do not "join" in the previous type; it doesn't add value,
 853           // and may yield a vacuous result if the field is of interface type.
 854           field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
 855           assert(field_type != NULL, "field singleton type must be consistent");
 856         } else {
 857           field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
 858         }
 859         if (UseCompressedOops) {
 860           field_type = field_type->make_narrowoop();
 861           basic_elem_type = T_NARROWOOP;
 862         }
 863       } else {
 864         field_type = Type::get_const_basic_type(basic_elem_type);
 865       }
 866 
 867       const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
 868 
 869       Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
 870       if (field_val == NULL) {
 871         // We weren't able to find a value for this field,
 872         // give up on eliminating this allocation.
 873 
 874         // Remove any extra entries we added to the safepoint.
 875         uint last = sfpt->req() - 1;
 876         for (int k = 0;  k < j; k++) {
 877           sfpt->del_req(last--);
 878         }
 879         _igvn._worklist.push(sfpt);
 880         // rollback processed safepoints
 881         while (safepoints_done.length() > 0) {
 882           SafePointNode* sfpt_done = safepoints_done.pop();
 883           // remove any extra entries we added to the safepoint
 884           last = sfpt_done->req() - 1;
 885           for (int k = 0;  k < nfields; k++) {
 886             sfpt_done->del_req(last--);
 887           }
 888           JVMState *jvms = sfpt_done->jvms();
 889           jvms->set_endoff(sfpt_done->req());
 890           // Now make a pass over the debug information replacing any references
 891           // to SafePointScalarObjectNode with the allocated object.
 892           int start = jvms->debug_start();
 893           int end   = jvms->debug_end();
 894           for (int i = start; i < end; i++) {
 895             if (sfpt_done->in(i)->is_SafePointScalarObject()) {
 896               SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
 897               if (scobj->first_index(jvms) == sfpt_done->req() &&
 898                   scobj->n_fields() == (uint)nfields) {
 899                 assert(scobj->alloc() == alloc, "sanity");
 900                 sfpt_done->set_req(i, res);
 901               }
 902             }
 903           }
 904           _igvn._worklist.push(sfpt_done);
 905         }
 906 #ifndef PRODUCT
 907         if (PrintEliminateAllocations) {
 908           if (field != NULL) {
 909             tty->print("=== At SafePoint node %d can't find value of Field: ",
 910                        sfpt->_idx);
 911             field->print();
 912             int field_idx = C->get_alias_index(field_addr_type);
 913             tty->print(" (alias_idx=%d)", field_idx);
 914           } else { // Array's element
 915             tty->print("=== At SafePoint node %d can't find value of array element [%d]",
 916                        sfpt->_idx, j);
 917           }
 918           tty->print(", which prevents elimination of: ");
 919           if (res == NULL)
 920             alloc->dump();
 921           else
 922             res->dump();
 923         }
 924 #endif
 925         return false;
 926       }
 927       if (UseCompressedOops && field_type->isa_narrowoop()) {
 928         // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
 929         // to be able scalar replace the allocation.
 930         if (field_val->is_EncodeP()) {
 931           field_val = field_val->in(1);
 932         } else {
 933           field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
 934         }
 935       }
 936       sfpt->add_req(field_val);
 937     }
 938     JVMState *jvms = sfpt->jvms();
 939     jvms->set_endoff(sfpt->req());
 940     // Now make a pass over the debug information replacing any references
 941     // to the allocated object with "sobj"
 942     int start = jvms->debug_start();
 943     int end   = jvms->debug_end();
 944     sfpt->replace_edges_in_range(res, sobj, start, end);
 945     _igvn._worklist.push(sfpt);
 946     safepoints_done.append_if_missing(sfpt); // keep it for rollback
 947   }
 948   return true;
 949 }
 950 
 951 // Process users of eliminated allocation.
 952 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
 953   Node* res = alloc->result_cast();
 954   if (res != NULL) {
 955     for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
 956       Node *use = res->last_out(j);
 957       uint oc1 = res->outcnt();
 958 
 959       if (use->is_AddP()) {
 960         for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
 961           Node *n = use->last_out(k);
 962           uint oc2 = use->outcnt();
 963           if (n->is_Store()) {
 964 #ifdef ASSERT
 965             // Verify that there is no dependent MemBarVolatile nodes,
 966             // they should be removed during IGVN, see MemBarNode::Ideal().
 967             for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
 968                                        p < pmax; p++) {
 969               Node* mb = n->fast_out(p);
 970               assert(mb->is_Initialize() || !mb->is_MemBar() ||
 971                      mb->req() <= MemBarNode::Precedent ||
 972                      mb->in(MemBarNode::Precedent) != n,
 973                      "MemBarVolatile should be eliminated for non-escaping object");
 974             }
 975 #endif
 976             _igvn.replace_node(n, n->in(MemNode::Memory));
 977           } else if (n->is_ArrayCopy()) {
 978             // Disconnect ArrayCopy node
 979             ArrayCopyNode* ac = n->as_ArrayCopy();
 980             assert(ac->is_clonebasic(), "unexpected array copy kind");
 981             Node* ctl_proj = ac->proj_out(TypeFunc::Control);
 982             Node* mem_proj = ac->proj_out(TypeFunc::Memory);
 983             if (ctl_proj != NULL) {
 984               _igvn.replace_node(ctl_proj, n->in(0));
 985             }
 986             if (mem_proj != NULL) {
 987               _igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
 988             }
 989           } else {
 990             eliminate_card_mark(n);
 991           }
 992           k -= (oc2 - use->outcnt());
 993         }
 994       } else if (use->is_ArrayCopy()) {
 995         // Disconnect ArrayCopy node
 996         ArrayCopyNode* ac = use->as_ArrayCopy();
 997         assert(ac->is_arraycopy_validated() ||
 998                ac->is_copyof_validated() ||
 999                ac->is_copyofrange_validated(), "unsupported");
1000         CallProjections callprojs;
1001         ac->extract_projections(&callprojs, true);
1002 
1003         _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1004         _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1005         _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1006 
1007         // Set control to top. IGVN will remove the remaining projections
1008         ac->set_req(0, top());
1009         ac->replace_edge(res, top());
1010 
1011         // Disconnect src right away: it can help find new
1012         // opportunities for allocation elimination
1013         Node* src = ac->in(ArrayCopyNode::Src);
1014         ac->replace_edge(src, top());
1015         if (src->outcnt() == 0) {
1016           _igvn.remove_dead_node(src);
1017         }
1018 
1019         _igvn._worklist.push(ac);
1020       } else {
1021         eliminate_card_mark(use);
1022       }
1023       j -= (oc1 - res->outcnt());
1024     }
1025     assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1026     _igvn.remove_dead_node(res);
1027   }
1028 
1029   //
1030   // Process other users of allocation's projections
1031   //
1032   if (_resproj != NULL && _resproj->outcnt() != 0) {
1033     // First disconnect stores captured by Initialize node.
1034     // If Initialize node is eliminated first in the following code,
1035     // it will kill such stores and DUIterator_Last will assert.
1036     for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax);  j < jmax; j++) {
1037       Node *use = _resproj->fast_out(j);
1038       if (use->is_AddP()) {
1039         // raw memory addresses used only by the initialization
1040         _igvn.replace_node(use, C->top());
1041         --j; --jmax;
1042       }
1043     }
1044     for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1045       Node *use = _resproj->last_out(j);
1046       uint oc1 = _resproj->outcnt();
1047       if (use->is_Initialize()) {
1048         // Eliminate Initialize node.
1049         InitializeNode *init = use->as_Initialize();
1050         assert(init->outcnt() <= 2, "only a control and memory projection expected");
1051         Node *ctrl_proj = init->proj_out(TypeFunc::Control);
1052         if (ctrl_proj != NULL) {
1053            assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1054           _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1055         }
1056         Node *mem_proj = init->proj_out(TypeFunc::Memory);
1057         if (mem_proj != NULL) {
1058           Node *mem = init->in(TypeFunc::Memory);
1059 #ifdef ASSERT
1060           if (mem->is_MergeMem()) {
1061             assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1062           } else {
1063             assert(mem == _memproj_fallthrough, "allocation memory projection");
1064           }
1065 #endif
1066           _igvn.replace_node(mem_proj, mem);
1067         }
1068       } else  {
1069         assert(false, "only Initialize or AddP expected");
1070       }
1071       j -= (oc1 - _resproj->outcnt());
1072     }
1073   }
1074   if (_fallthroughcatchproj != NULL) {
1075     _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1076   }
1077   if (_memproj_fallthrough != NULL) {
1078     _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1079   }
1080   if (_memproj_catchall != NULL) {
1081     _igvn.replace_node(_memproj_catchall, C->top());
1082   }
1083   if (_ioproj_fallthrough != NULL) {
1084     _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1085   }
1086   if (_ioproj_catchall != NULL) {
1087     _igvn.replace_node(_ioproj_catchall, C->top());
1088   }
1089   if (_catchallcatchproj != NULL) {
1090     _igvn.replace_node(_catchallcatchproj, C->top());
1091   }
1092 }
1093 
1094 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1095   // Don't do scalar replacement if the frame can be popped by JVMTI:
1096   // if reallocation fails during deoptimization we'll pop all
1097   // interpreter frames for this compiled frame and that won't play
1098   // nice with JVMTI popframe.
1099   if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1100     return false;
1101   }
1102   Node* klass = alloc->in(AllocateNode::KlassNode);
1103   const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1104   Node* res = alloc->result_cast();
1105   // Eliminate boxing allocations which are not used
1106   // regardless scalar replacable status.
1107   bool boxing_alloc = C->eliminate_boxing() &&
1108                       tklass->klass()->is_instance_klass()  &&
1109                       tklass->klass()->as_instance_klass()->is_box_klass();
1110   if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1111     return false;
1112   }
1113 
1114   extract_call_projections(alloc);
1115 
1116   GrowableArray <SafePointNode *> safepoints;
1117   if (!can_eliminate_allocation(alloc, safepoints)) {
1118     return false;
1119   }
1120 
1121   if (!alloc->_is_scalar_replaceable) {
1122     assert(res == NULL, "sanity");
1123     // We can only eliminate allocation if all debug info references
1124     // are already replaced with SafePointScalarObject because
1125     // we can't search for a fields value without instance_id.
1126     if (safepoints.length() > 0) {
1127       return false;
1128     }
1129   }
1130 
1131   if (!scalar_replacement(alloc, safepoints)) {
1132     return false;
1133   }
1134 
1135   CompileLog* log = C->log();
1136   if (log != NULL) {
1137     log->head("eliminate_allocation type='%d'",
1138               log->identify(tklass->klass()));
1139     JVMState* p = alloc->jvms();
1140     while (p != NULL) {
1141       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1142       p = p->caller();
1143     }
1144     log->tail("eliminate_allocation");
1145   }
1146 
1147   process_users_of_allocation(alloc);
1148 
1149 #ifndef PRODUCT
1150   if (PrintEliminateAllocations) {
1151     if (alloc->is_AllocateArray())
1152       tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1153     else
1154       tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1155   }
1156 #endif
1157 
1158   return true;
1159 }
1160 
1161 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1162   // EA should remove all uses of non-escaping boxing node.
1163   if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1164     return false;
1165   }
1166 
1167   assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1168 
1169   extract_call_projections(boxing);
1170 
1171   const TypeTuple* r = boxing->tf()->range();
1172   assert(r->cnt() > TypeFunc::Parms, "sanity");
1173   const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1174   assert(t != NULL, "sanity");
1175 
1176   CompileLog* log = C->log();
1177   if (log != NULL) {
1178     log->head("eliminate_boxing type='%d'",
1179               log->identify(t->klass()));
1180     JVMState* p = boxing->jvms();
1181     while (p != NULL) {
1182       log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1183       p = p->caller();
1184     }
1185     log->tail("eliminate_boxing");
1186   }
1187 
1188   process_users_of_allocation(boxing);
1189 
1190 #ifndef PRODUCT
1191   if (PrintEliminateAllocations) {
1192     tty->print("++++ Eliminated: %d ", boxing->_idx);
1193     boxing->method()->print_short_name(tty);
1194     tty->cr();
1195   }
1196 #endif
1197 
1198   return true;
1199 }
1200 
1201 //---------------------------set_eden_pointers-------------------------
1202 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1203   if (UseTLAB) {                // Private allocation: load from TLS
1204     Node* thread = transform_later(new ThreadLocalNode());
1205     int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1206     int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1207     eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1208     eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1209   } else {                      // Shared allocation: load from globals
1210     CollectedHeap* ch = Universe::heap();
1211     address top_adr = (address)ch->top_addr();
1212     address end_adr = (address)ch->end_addr();
1213     eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1214     eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1215   }
1216 }
1217 
1218 
1219 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1220   Node* adr = basic_plus_adr(base, offset);
1221   const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1222   Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1223   transform_later(value);
1224   return value;
1225 }
1226 
1227 
1228 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1229   Node* adr = basic_plus_adr(base, offset);
1230   mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1231   transform_later(mem);
1232   return mem;
1233 }
1234 
1235 //=============================================================================
1236 //
1237 //                              A L L O C A T I O N
1238 //
1239 // Allocation attempts to be fast in the case of frequent small objects.
1240 // It breaks down like this:
1241 //
1242 // 1) Size in doublewords is computed.  This is a constant for objects and
1243 // variable for most arrays.  Doubleword units are used to avoid size
1244 // overflow of huge doubleword arrays.  We need doublewords in the end for
1245 // rounding.
1246 //
1247 // 2) Size is checked for being 'too large'.  Too-large allocations will go
1248 // the slow path into the VM.  The slow path can throw any required
1249 // exceptions, and does all the special checks for very large arrays.  The
1250 // size test can constant-fold away for objects.  For objects with
1251 // finalizers it constant-folds the otherway: you always go slow with
1252 // finalizers.
1253 //
1254 // 3) If NOT using TLABs, this is the contended loop-back point.
1255 // Load-Locked the heap top.  If using TLABs normal-load the heap top.
1256 //
1257 // 4) Check that heap top + size*8 < max.  If we fail go the slow ` route.
1258 // NOTE: "top+size*8" cannot wrap the 4Gig line!  Here's why: for largish
1259 // "size*8" we always enter the VM, where "largish" is a constant picked small
1260 // enough that there's always space between the eden max and 4Gig (old space is
1261 // there so it's quite large) and large enough that the cost of entering the VM
1262 // is dwarfed by the cost to initialize the space.
1263 //
1264 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1265 // down.  If contended, repeat at step 3.  If using TLABs normal-store
1266 // adjusted heap top back down; there is no contention.
1267 //
1268 // 6) If !ZeroTLAB then Bulk-clear the object/array.  Fill in klass & mark
1269 // fields.
1270 //
1271 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1272 // oop flavor.
1273 //
1274 //=============================================================================
1275 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1276 // Allocations bigger than this always go the slow route.
1277 // This value must be small enough that allocation attempts that need to
1278 // trigger exceptions go the slow route.  Also, it must be small enough so
1279 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1280 //=============================================================================j//
1281 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1282 // The allocator will coalesce int->oop copies away.  See comment in
1283 // coalesce.cpp about how this works.  It depends critically on the exact
1284 // code shape produced here, so if you are changing this code shape
1285 // make sure the GC info for the heap-top is correct in and around the
1286 // slow-path call.
1287 //
1288 
1289 void PhaseMacroExpand::expand_allocate_common(
1290             AllocateNode* alloc, // allocation node to be expanded
1291             Node* length,  // array length for an array allocation
1292             const TypeFunc* slow_call_type, // Type of slow call
1293             address slow_call_address  // Address of slow call
1294     )
1295 {
1296 
1297   Node* ctrl = alloc->in(TypeFunc::Control);
1298   Node* mem  = alloc->in(TypeFunc::Memory);
1299   Node* i_o  = alloc->in(TypeFunc::I_O);
1300   Node* size_in_bytes     = alloc->in(AllocateNode::AllocSize);
1301   Node* klass_node        = alloc->in(AllocateNode::KlassNode);
1302   Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1303 
1304   assert(ctrl != NULL, "must have control");
1305   // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1306   // they will not be used if "always_slow" is set
1307   enum { slow_result_path = 1, fast_result_path = 2 };
1308   Node *result_region;
1309   Node *result_phi_rawmem;
1310   Node *result_phi_rawoop;
1311   Node *result_phi_i_o;
1312 
1313   // The initial slow comparison is a size check, the comparison
1314   // we want to do is a BoolTest::gt
1315   bool always_slow = false;
1316   int tv = _igvn.find_int_con(initial_slow_test, -1);
1317   if (tv >= 0) {
1318     always_slow = (tv == 1);
1319     initial_slow_test = NULL;
1320   } else {
1321     initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1322   }
1323 
1324   if (C->env()->dtrace_alloc_probes() ||
1325       !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) {
1326     // Force slow-path allocation
1327     always_slow = true;
1328     initial_slow_test = NULL;
1329   }
1330 
1331 
1332   enum { too_big_or_final_path = 1, need_gc_path = 2 };
1333   Node *slow_region = NULL;
1334   Node *toobig_false = ctrl;
1335 
1336   assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1337   // generate the initial test if necessary
1338   if (initial_slow_test != NULL ) {
1339     slow_region = new RegionNode(3);
1340 
1341     // Now make the initial failure test.  Usually a too-big test but
1342     // might be a TRUE for finalizers or a fancy class check for
1343     // newInstance0.
1344     IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1345     transform_later(toobig_iff);
1346     // Plug the failing-too-big test into the slow-path region
1347     Node *toobig_true = new IfTrueNode( toobig_iff );
1348     transform_later(toobig_true);
1349     slow_region    ->init_req( too_big_or_final_path, toobig_true );
1350     toobig_false = new IfFalseNode( toobig_iff );
1351     transform_later(toobig_false);
1352   } else {         // No initial test, just fall into next case
1353     toobig_false = ctrl;
1354     debug_only(slow_region = NodeSentinel);
1355   }
1356 
1357   Node *slow_mem = mem;  // save the current memory state for slow path
1358   // generate the fast allocation code unless we know that the initial test will always go slow
1359   if (!always_slow) {
1360     // Fast path modifies only raw memory.
1361     if (mem->is_MergeMem()) {
1362       mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1363     }
1364 
1365     Node* eden_top_adr;
1366     Node* eden_end_adr;
1367 
1368     set_eden_pointers(eden_top_adr, eden_end_adr);
1369 
1370     // Load Eden::end.  Loop invariant and hoisted.
1371     //
1372     // Note: We set the control input on "eden_end" and "old_eden_top" when using
1373     //       a TLAB to work around a bug where these values were being moved across
1374     //       a safepoint.  These are not oops, so they cannot be include in the oop
1375     //       map, but they can be changed by a GC.   The proper way to fix this would
1376     //       be to set the raw memory state when generating a  SafepointNode.  However
1377     //       this will require extensive changes to the loop optimization in order to
1378     //       prevent a degradation of the optimization.
1379     //       See comment in memnode.hpp, around line 227 in class LoadPNode.
1380     Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1381 
1382     // allocate the Region and Phi nodes for the result
1383     result_region = new RegionNode(3);
1384     result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1385     result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1386     result_phi_i_o    = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1387 
1388     // We need a Region for the loop-back contended case.
1389     enum { fall_in_path = 1, contended_loopback_path = 2 };
1390     Node *contended_region;
1391     Node *contended_phi_rawmem;
1392     if (UseTLAB) {
1393       contended_region = toobig_false;
1394       contended_phi_rawmem = mem;
1395     } else {
1396       contended_region = new RegionNode(3);
1397       contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1398       // Now handle the passing-too-big test.  We fall into the contended
1399       // loop-back merge point.
1400       contended_region    ->init_req(fall_in_path, toobig_false);
1401       contended_phi_rawmem->init_req(fall_in_path, mem);
1402       transform_later(contended_region);
1403       transform_later(contended_phi_rawmem);
1404     }
1405 
1406     // Load(-locked) the heap top.
1407     // See note above concerning the control input when using a TLAB
1408     Node *old_eden_top = UseTLAB
1409       ? new LoadPNode      (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1410       : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1411 
1412     transform_later(old_eden_top);
1413     // Add to heap top to get a new heap top
1414     Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1415     transform_later(new_eden_top);
1416     // Check for needing a GC; compare against heap end
1417     Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1418     transform_later(needgc_cmp);
1419     Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1420     transform_later(needgc_bol);
1421     IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1422     transform_later(needgc_iff);
1423 
1424     // Plug the failing-heap-space-need-gc test into the slow-path region
1425     Node *needgc_true = new IfTrueNode(needgc_iff);
1426     transform_later(needgc_true);
1427     if (initial_slow_test) {
1428       slow_region->init_req(need_gc_path, needgc_true);
1429       // This completes all paths into the slow merge point
1430       transform_later(slow_region);
1431     } else {                      // No initial slow path needed!
1432       // Just fall from the need-GC path straight into the VM call.
1433       slow_region = needgc_true;
1434     }
1435     // No need for a GC.  Setup for the Store-Conditional
1436     Node *needgc_false = new IfFalseNode(needgc_iff);
1437     transform_later(needgc_false);
1438 
1439     // Grab regular I/O before optional prefetch may change it.
1440     // Slow-path does no I/O so just set it to the original I/O.
1441     result_phi_i_o->init_req(slow_result_path, i_o);
1442 
1443     i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1444                               old_eden_top, new_eden_top, length);
1445 
1446     // Name successful fast-path variables
1447     Node* fast_oop = old_eden_top;
1448     Node* fast_oop_ctrl;
1449     Node* fast_oop_rawmem;
1450 
1451     // Store (-conditional) the modified eden top back down.
1452     // StorePConditional produces flags for a test PLUS a modified raw
1453     // memory state.
1454     if (UseTLAB) {
1455       Node* store_eden_top =
1456         new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1457                               TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1458       transform_later(store_eden_top);
1459       fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1460       fast_oop_rawmem = store_eden_top;
1461     } else {
1462       Node* store_eden_top =
1463         new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1464                                          new_eden_top, fast_oop/*old_eden_top*/);
1465       transform_later(store_eden_top);
1466       Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1467       transform_later(contention_check);
1468       store_eden_top = new SCMemProjNode(store_eden_top);
1469       transform_later(store_eden_top);
1470 
1471       // If not using TLABs, check to see if there was contention.
1472       IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1473       transform_later(contention_iff);
1474       Node *contention_true = new IfTrueNode(contention_iff);
1475       transform_later(contention_true);
1476       // If contention, loopback and try again.
1477       contended_region->init_req(contended_loopback_path, contention_true);
1478       contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1479 
1480       // Fast-path succeeded with no contention!
1481       Node *contention_false = new IfFalseNode(contention_iff);
1482       transform_later(contention_false);
1483       fast_oop_ctrl = contention_false;
1484 
1485       // Bump total allocated bytes for this thread
1486       Node* thread = new ThreadLocalNode();
1487       transform_later(thread);
1488       Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1489                                              in_bytes(JavaThread::allocated_bytes_offset()));
1490       Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1491                                     0, TypeLong::LONG, T_LONG);
1492 #ifdef _LP64
1493       Node* alloc_size = size_in_bytes;
1494 #else
1495       Node* alloc_size = new ConvI2LNode(size_in_bytes);
1496       transform_later(alloc_size);
1497 #endif
1498       Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1499       transform_later(new_alloc_bytes);
1500       fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1501                                    0, new_alloc_bytes, T_LONG);
1502     }
1503 
1504     InitializeNode* init = alloc->initialization();
1505     fast_oop_rawmem = initialize_object(alloc,
1506                                         fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1507                                         klass_node, length, size_in_bytes);
1508 
1509     // If initialization is performed by an array copy, any required
1510     // MemBarStoreStore was already added. If the object does not
1511     // escape no need for a MemBarStoreStore. Otherwise we need a
1512     // MemBarStoreStore so that stores that initialize this object
1513     // can't be reordered with a subsequent store that makes this
1514     // object accessible by other threads.
1515     if (!alloc->does_not_escape_thread() &&
1516         (init == NULL || !init->is_complete_with_arraycopy())) {
1517       if (init == NULL || init->req() < InitializeNode::RawStores) {
1518         // No InitializeNode or no stores captured by zeroing
1519         // elimination. Simply add the MemBarStoreStore after object
1520         // initialization.
1521         MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1522         transform_later(mb);
1523 
1524         mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1525         mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1526         fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1527         transform_later(fast_oop_ctrl);
1528         fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1529         transform_later(fast_oop_rawmem);
1530       } else {
1531         // Add the MemBarStoreStore after the InitializeNode so that
1532         // all stores performing the initialization that were moved
1533         // before the InitializeNode happen before the storestore
1534         // barrier.
1535 
1536         Node* init_ctrl = init->proj_out(TypeFunc::Control);
1537         Node* init_mem = init->proj_out(TypeFunc::Memory);
1538 
1539         MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1540         transform_later(mb);
1541 
1542         Node* ctrl = new ProjNode(init,TypeFunc::Control);
1543         transform_later(ctrl);
1544         Node* mem = new ProjNode(init,TypeFunc::Memory);
1545         transform_later(mem);
1546 
1547         // The MemBarStoreStore depends on control and memory coming
1548         // from the InitializeNode
1549         mb->init_req(TypeFunc::Memory, mem);
1550         mb->init_req(TypeFunc::Control, ctrl);
1551 
1552         ctrl = new ProjNode(mb,TypeFunc::Control);
1553         transform_later(ctrl);
1554         mem = new ProjNode(mb,TypeFunc::Memory);
1555         transform_later(mem);
1556 
1557         // All nodes that depended on the InitializeNode for control
1558         // and memory must now depend on the MemBarNode that itself
1559         // depends on the InitializeNode
1560         _igvn.replace_node(init_ctrl, ctrl);
1561         _igvn.replace_node(init_mem, mem);
1562       }
1563     }
1564 
1565     if (C->env()->dtrace_extended_probes()) {
1566       // Slow-path call
1567       int size = TypeFunc::Parms + 2;
1568       CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1569                                             CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1570                                             "dtrace_object_alloc",
1571                                             TypeRawPtr::BOTTOM);
1572 
1573       // Get base of thread-local storage area
1574       Node* thread = new ThreadLocalNode();
1575       transform_later(thread);
1576 
1577       call->init_req(TypeFunc::Parms+0, thread);
1578       call->init_req(TypeFunc::Parms+1, fast_oop);
1579       call->init_req(TypeFunc::Control, fast_oop_ctrl);
1580       call->init_req(TypeFunc::I_O    , top()); // does no i/o
1581       call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1582       call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1583       call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1584       transform_later(call);
1585       fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1586       transform_later(fast_oop_ctrl);
1587       fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1588       transform_later(fast_oop_rawmem);
1589     }
1590 
1591     // Plug in the successful fast-path into the result merge point
1592     result_region    ->init_req(fast_result_path, fast_oop_ctrl);
1593     result_phi_rawoop->init_req(fast_result_path, fast_oop);
1594     result_phi_i_o   ->init_req(fast_result_path, i_o);
1595     result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1596   } else {
1597     slow_region = ctrl;
1598     result_phi_i_o = i_o; // Rename it to use in the following code.
1599   }
1600 
1601   // Generate slow-path call
1602   CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1603                                OptoRuntime::stub_name(slow_call_address),
1604                                alloc->jvms()->bci(),
1605                                TypePtr::BOTTOM);
1606   call->init_req( TypeFunc::Control, slow_region );
1607   call->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
1608   call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1609   call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1610   call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1611 
1612   call->init_req(TypeFunc::Parms+0, klass_node);
1613   if (length != NULL) {
1614     call->init_req(TypeFunc::Parms+1, length);
1615   }
1616 
1617   // Copy debug information and adjust JVMState information, then replace
1618   // allocate node with the call
1619   copy_call_debug_info((CallNode *) alloc,  call);
1620   if (!always_slow) {
1621     call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1622   } else {
1623     // Hook i_o projection to avoid its elimination during allocation
1624     // replacement (when only a slow call is generated).
1625     call->set_req(TypeFunc::I_O, result_phi_i_o);
1626   }
1627   _igvn.replace_node(alloc, call);
1628   transform_later(call);
1629 
1630   // Identify the output projections from the allocate node and
1631   // adjust any references to them.
1632   // The control and io projections look like:
1633   //
1634   //        v---Proj(ctrl) <-----+   v---CatchProj(ctrl)
1635   //  Allocate                   Catch
1636   //        ^---Proj(io) <-------+   ^---CatchProj(io)
1637   //
1638   //  We are interested in the CatchProj nodes.
1639   //
1640   extract_call_projections(call);
1641 
1642   // An allocate node has separate memory projections for the uses on
1643   // the control and i_o paths. Replace the control memory projection with
1644   // result_phi_rawmem (unless we are only generating a slow call when
1645   // both memory projections are combined)
1646   if (!always_slow && _memproj_fallthrough != NULL) {
1647     for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1648       Node *use = _memproj_fallthrough->fast_out(i);
1649       _igvn.rehash_node_delayed(use);
1650       imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1651       // back up iterator
1652       --i;
1653     }
1654   }
1655   // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1656   // _memproj_catchall so we end up with a call that has only 1 memory projection.
1657   if (_memproj_catchall != NULL ) {
1658     if (_memproj_fallthrough == NULL) {
1659       _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1660       transform_later(_memproj_fallthrough);
1661     }
1662     for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1663       Node *use = _memproj_catchall->fast_out(i);
1664       _igvn.rehash_node_delayed(use);
1665       imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1666       // back up iterator
1667       --i;
1668     }
1669     assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1670     _igvn.remove_dead_node(_memproj_catchall);
1671   }
1672 
1673   // An allocate node has separate i_o projections for the uses on the control
1674   // and i_o paths. Always replace the control i_o projection with result i_o
1675   // otherwise incoming i_o become dead when only a slow call is generated
1676   // (it is different from memory projections where both projections are
1677   // combined in such case).
1678   if (_ioproj_fallthrough != NULL) {
1679     for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1680       Node *use = _ioproj_fallthrough->fast_out(i);
1681       _igvn.rehash_node_delayed(use);
1682       imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1683       // back up iterator
1684       --i;
1685     }
1686   }
1687   // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1688   // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1689   if (_ioproj_catchall != NULL ) {
1690     if (_ioproj_fallthrough == NULL) {
1691       _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1692       transform_later(_ioproj_fallthrough);
1693     }
1694     for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1695       Node *use = _ioproj_catchall->fast_out(i);
1696       _igvn.rehash_node_delayed(use);
1697       imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1698       // back up iterator
1699       --i;
1700     }
1701     assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1702     _igvn.remove_dead_node(_ioproj_catchall);
1703   }
1704 
1705   // if we generated only a slow call, we are done
1706   if (always_slow) {
1707     // Now we can unhook i_o.
1708     if (result_phi_i_o->outcnt() > 1) {
1709       call->set_req(TypeFunc::I_O, top());
1710     } else {
1711       assert(result_phi_i_o->unique_ctrl_out() == call, "");
1712       // Case of new array with negative size known during compilation.
1713       // AllocateArrayNode::Ideal() optimization disconnect unreachable
1714       // following code since call to runtime will throw exception.
1715       // As result there will be no users of i_o after the call.
1716       // Leave i_o attached to this call to avoid problems in preceding graph.
1717     }
1718     return;
1719   }
1720 
1721 
1722   if (_fallthroughcatchproj != NULL) {
1723     ctrl = _fallthroughcatchproj->clone();
1724     transform_later(ctrl);
1725     _igvn.replace_node(_fallthroughcatchproj, result_region);
1726   } else {
1727     ctrl = top();
1728   }
1729   Node *slow_result;
1730   if (_resproj == NULL) {
1731     // no uses of the allocation result
1732     slow_result = top();
1733   } else {
1734     slow_result = _resproj->clone();
1735     transform_later(slow_result);
1736     _igvn.replace_node(_resproj, result_phi_rawoop);
1737   }
1738 
1739   // Plug slow-path into result merge point
1740   result_region    ->init_req( slow_result_path, ctrl );
1741   result_phi_rawoop->init_req( slow_result_path, slow_result);
1742   result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1743   transform_later(result_region);
1744   transform_later(result_phi_rawoop);
1745   transform_later(result_phi_rawmem);
1746   transform_later(result_phi_i_o);
1747   // This completes all paths into the result merge point
1748 }
1749 
1750 
1751 // Helper for PhaseMacroExpand::expand_allocate_common.
1752 // Initializes the newly-allocated storage.
1753 Node*
1754 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1755                                     Node* control, Node* rawmem, Node* object,
1756                                     Node* klass_node, Node* length,
1757                                     Node* size_in_bytes) {
1758   InitializeNode* init = alloc->initialization();
1759   // Store the klass & mark bits
1760   Node* mark_node = NULL;
1761   // For now only enable fast locking for non-array types
1762   if (UseBiasedLocking && (length == NULL)) {
1763     mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1764   } else {
1765     mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1766   }
1767   rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1768 
1769   rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1770   int header_size = alloc->minimum_header_size();  // conservatively small
1771 
1772   // Array length
1773   if (length != NULL) {         // Arrays need length field
1774     rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1775     // conservatively small header size:
1776     header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1777     ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1778     if (k->is_array_klass())    // we know the exact header size in most cases:
1779       header_size = Klass::layout_helper_header_size(k->layout_helper());
1780   }
1781 
1782   // Clear the object body, if necessary.
1783   if (init == NULL) {
1784     // The init has somehow disappeared; be cautious and clear everything.
1785     //
1786     // This can happen if a node is allocated but an uncommon trap occurs
1787     // immediately.  In this case, the Initialize gets associated with the
1788     // trap, and may be placed in a different (outer) loop, if the Allocate
1789     // is in a loop.  If (this is rare) the inner loop gets unrolled, then
1790     // there can be two Allocates to one Initialize.  The answer in all these
1791     // edge cases is safety first.  It is always safe to clear immediately
1792     // within an Allocate, and then (maybe or maybe not) clear some more later.
1793     if (!ZeroTLAB)
1794       rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1795                                             header_size, size_in_bytes,
1796                                             &_igvn);
1797   } else {
1798     if (!init->is_complete()) {
1799       // Try to win by zeroing only what the init does not store.
1800       // We can also try to do some peephole optimizations,
1801       // such as combining some adjacent subword stores.
1802       rawmem = init->complete_stores(control, rawmem, object,
1803                                      header_size, size_in_bytes, &_igvn);
1804     }
1805     // We have no more use for this link, since the AllocateNode goes away:
1806     init->set_req(InitializeNode::RawAddress, top());
1807     // (If we keep the link, it just confuses the register allocator,
1808     // who thinks he sees a real use of the address by the membar.)
1809   }
1810 
1811   return rawmem;
1812 }
1813 
1814 // Generate prefetch instructions for next allocations.
1815 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1816                                         Node*& contended_phi_rawmem,
1817                                         Node* old_eden_top, Node* new_eden_top,
1818                                         Node* length) {
1819    enum { fall_in_path = 1, pf_path = 2 };
1820    if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1821       // Generate prefetch allocation with watermark check.
1822       // As an allocation hits the watermark, we will prefetch starting
1823       // at a "distance" away from watermark.
1824 
1825       Node *pf_region = new RegionNode(3);
1826       Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1827                                                 TypeRawPtr::BOTTOM );
1828       // I/O is used for Prefetch
1829       Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1830 
1831       Node *thread = new ThreadLocalNode();
1832       transform_later(thread);
1833 
1834       Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1835                    _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1836       transform_later(eden_pf_adr);
1837 
1838       Node *old_pf_wm = new LoadPNode(needgc_false,
1839                                    contended_phi_rawmem, eden_pf_adr,
1840                                    TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1841                                    MemNode::unordered);
1842       transform_later(old_pf_wm);
1843 
1844       // check against new_eden_top
1845       Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1846       transform_later(need_pf_cmp);
1847       Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1848       transform_later(need_pf_bol);
1849       IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1850                                        PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1851       transform_later(need_pf_iff);
1852 
1853       // true node, add prefetchdistance
1854       Node *need_pf_true = new IfTrueNode( need_pf_iff );
1855       transform_later(need_pf_true);
1856 
1857       Node *need_pf_false = new IfFalseNode( need_pf_iff );
1858       transform_later(need_pf_false);
1859 
1860       Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1861                                     _igvn.MakeConX(AllocatePrefetchDistance) );
1862       transform_later(new_pf_wmt );
1863       new_pf_wmt->set_req(0, need_pf_true);
1864 
1865       Node *store_new_wmt = new StorePNode(need_pf_true,
1866                                        contended_phi_rawmem, eden_pf_adr,
1867                                        TypeRawPtr::BOTTOM, new_pf_wmt,
1868                                        MemNode::unordered);
1869       transform_later(store_new_wmt);
1870 
1871       // adding prefetches
1872       pf_phi_abio->init_req( fall_in_path, i_o );
1873 
1874       Node *prefetch_adr;
1875       Node *prefetch;
1876       uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1877       uint step_size = AllocatePrefetchStepSize;
1878       uint distance = 0;
1879 
1880       for ( uint i = 0; i < lines; i++ ) {
1881         prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1882                                             _igvn.MakeConX(distance) );
1883         transform_later(prefetch_adr);
1884         prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1885         transform_later(prefetch);
1886         distance += step_size;
1887         i_o = prefetch;
1888       }
1889       pf_phi_abio->set_req( pf_path, i_o );
1890 
1891       pf_region->init_req( fall_in_path, need_pf_false );
1892       pf_region->init_req( pf_path, need_pf_true );
1893 
1894       pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1895       pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1896 
1897       transform_later(pf_region);
1898       transform_later(pf_phi_rawmem);
1899       transform_later(pf_phi_abio);
1900 
1901       needgc_false = pf_region;
1902       contended_phi_rawmem = pf_phi_rawmem;
1903       i_o = pf_phi_abio;
1904    } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1905       // Insert a prefetch for each allocation.
1906       // This code is used for Sparc with BIS.
1907       Node *pf_region = new RegionNode(3);
1908       Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1909                                              TypeRawPtr::BOTTOM );
1910       transform_later(pf_region);
1911 
1912       // Generate several prefetch instructions.
1913       uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1914       uint step_size = AllocatePrefetchStepSize;
1915       uint distance = AllocatePrefetchDistance;
1916 
1917       // Next cache address.
1918       Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1919                                             _igvn.MakeConX(distance));
1920       transform_later(cache_adr);
1921       cache_adr = new CastP2XNode(needgc_false, cache_adr);
1922       transform_later(cache_adr);
1923       Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1924       cache_adr = new AndXNode(cache_adr, mask);
1925       transform_later(cache_adr);
1926       cache_adr = new CastX2PNode(cache_adr);
1927       transform_later(cache_adr);
1928 
1929       // Prefetch
1930       Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1931       prefetch->set_req(0, needgc_false);
1932       transform_later(prefetch);
1933       contended_phi_rawmem = prefetch;
1934       Node *prefetch_adr;
1935       distance = step_size;
1936       for ( uint i = 1; i < lines; i++ ) {
1937         prefetch_adr = new AddPNode( cache_adr, cache_adr,
1938                                             _igvn.MakeConX(distance) );
1939         transform_later(prefetch_adr);
1940         prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1941         transform_later(prefetch);
1942         distance += step_size;
1943         contended_phi_rawmem = prefetch;
1944       }
1945    } else if( AllocatePrefetchStyle > 0 ) {
1946       // Insert a prefetch for each allocation only on the fast-path
1947       Node *prefetch_adr;
1948       Node *prefetch;
1949       // Generate several prefetch instructions.
1950       uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1951       uint step_size = AllocatePrefetchStepSize;
1952       uint distance = AllocatePrefetchDistance;
1953       for ( uint i = 0; i < lines; i++ ) {
1954         prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1955                                             _igvn.MakeConX(distance) );
1956         transform_later(prefetch_adr);
1957         prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1958         // Do not let it float too high, since if eden_top == eden_end,
1959         // both might be null.
1960         if( i == 0 ) { // Set control for first prefetch, next follows it
1961           prefetch->init_req(0, needgc_false);
1962         }
1963         transform_later(prefetch);
1964         distance += step_size;
1965         i_o = prefetch;
1966       }
1967    }
1968    return i_o;
1969 }
1970 
1971 
1972 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1973   expand_allocate_common(alloc, NULL,
1974                          OptoRuntime::new_instance_Type(),
1975                          OptoRuntime::new_instance_Java());
1976 }
1977 
1978 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1979   Node* length = alloc->in(AllocateNode::ALength);
1980   InitializeNode* init = alloc->initialization();
1981   Node* klass_node = alloc->in(AllocateNode::KlassNode);
1982   ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1983   address slow_call_address;  // Address of slow call
1984   if (init != NULL && init->is_complete_with_arraycopy() &&
1985       k->is_type_array_klass()) {
1986     // Don't zero type array during slow allocation in VM since
1987     // it will be initialized later by arraycopy in compiled code.
1988     slow_call_address = OptoRuntime::new_array_nozero_Java();
1989   } else {
1990     slow_call_address = OptoRuntime::new_array_Java();
1991   }
1992   expand_allocate_common(alloc, length,
1993                          OptoRuntime::new_array_Type(),
1994                          slow_call_address);
1995 }
1996 
1997 //-------------------mark_eliminated_box----------------------------------
1998 //
1999 // During EA obj may point to several objects but after few ideal graph
2000 // transformations (CCP) it may point to only one non escaping object
2001 // (but still using phi), corresponding locks and unlocks will be marked
2002 // for elimination. Later obj could be replaced with a new node (new phi)
2003 // and which does not have escape information. And later after some graph
2004 // reshape other locks and unlocks (which were not marked for elimination
2005 // before) are connected to this new obj (phi) but they still will not be
2006 // marked for elimination since new obj has no escape information.
2007 // Mark all associated (same box and obj) lock and unlock nodes for
2008 // elimination if some of them marked already.
2009 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2010   if (oldbox->as_BoxLock()->is_eliminated())
2011     return; // This BoxLock node was processed already.
2012 
2013   // New implementation (EliminateNestedLocks) has separate BoxLock
2014   // node for each locked region so mark all associated locks/unlocks as
2015   // eliminated even if different objects are referenced in one locked region
2016   // (for example, OSR compilation of nested loop inside locked scope).
2017   if (EliminateNestedLocks ||
2018       oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2019     // Box is used only in one lock region. Mark this box as eliminated.
2020     _igvn.hash_delete(oldbox);
2021     oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2022      _igvn.hash_insert(oldbox);
2023 
2024     for (uint i = 0; i < oldbox->outcnt(); i++) {
2025       Node* u = oldbox->raw_out(i);
2026       if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2027         AbstractLockNode* alock = u->as_AbstractLock();
2028         // Check lock's box since box could be referenced by Lock's debug info.
2029         if (alock->box_node() == oldbox) {
2030           // Mark eliminated all related locks and unlocks.
2031 #ifdef ASSERT
2032           alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2033 #endif
2034           alock->set_non_esc_obj();
2035         }
2036       }
2037     }
2038     return;
2039   }
2040 
2041   // Create new "eliminated" BoxLock node and use it in monitor debug info
2042   // instead of oldbox for the same object.
2043   BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2044 
2045   // Note: BoxLock node is marked eliminated only here and it is used
2046   // to indicate that all associated lock and unlock nodes are marked
2047   // for elimination.
2048   newbox->set_eliminated();
2049   transform_later(newbox);
2050 
2051   // Replace old box node with new box for all users of the same object.
2052   for (uint i = 0; i < oldbox->outcnt();) {
2053     bool next_edge = true;
2054 
2055     Node* u = oldbox->raw_out(i);
2056     if (u->is_AbstractLock()) {
2057       AbstractLockNode* alock = u->as_AbstractLock();
2058       if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2059         // Replace Box and mark eliminated all related locks and unlocks.
2060 #ifdef ASSERT
2061         alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2062 #endif
2063         alock->set_non_esc_obj();
2064         _igvn.rehash_node_delayed(alock);
2065         alock->set_box_node(newbox);
2066         next_edge = false;
2067       }
2068     }
2069     if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2070       FastLockNode* flock = u->as_FastLock();
2071       assert(flock->box_node() == oldbox, "sanity");
2072       _igvn.rehash_node_delayed(flock);
2073       flock->set_box_node(newbox);
2074       next_edge = false;
2075     }
2076 
2077     // Replace old box in monitor debug info.
2078     if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2079       SafePointNode* sfn = u->as_SafePoint();
2080       JVMState* youngest_jvms = sfn->jvms();
2081       int max_depth = youngest_jvms->depth();
2082       for (int depth = 1; depth <= max_depth; depth++) {
2083         JVMState* jvms = youngest_jvms->of_depth(depth);
2084         int num_mon  = jvms->nof_monitors();
2085         // Loop over monitors
2086         for (int idx = 0; idx < num_mon; idx++) {
2087           Node* obj_node = sfn->monitor_obj(jvms, idx);
2088           Node* box_node = sfn->monitor_box(jvms, idx);
2089           if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2090             int j = jvms->monitor_box_offset(idx);
2091             _igvn.replace_input_of(u, j, newbox);
2092             next_edge = false;
2093           }
2094         }
2095       }
2096     }
2097     if (next_edge) i++;
2098   }
2099 }
2100 
2101 //-----------------------mark_eliminated_locking_nodes-----------------------
2102 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2103   if (EliminateNestedLocks) {
2104     if (alock->is_nested()) {
2105        assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2106        return;
2107     } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2108       // Only Lock node has JVMState needed here.
2109       // Not that preceding claim is documented anywhere else.
2110       if (alock->jvms() != NULL) {
2111         if (alock->as_Lock()->is_nested_lock_region()) {
2112           // Mark eliminated related nested locks and unlocks.
2113           Node* obj = alock->obj_node();
2114           BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2115           assert(!box_node->is_eliminated(), "should not be marked yet");
2116           // Note: BoxLock node is marked eliminated only here
2117           // and it is used to indicate that all associated lock
2118           // and unlock nodes are marked for elimination.
2119           box_node->set_eliminated(); // Box's hash is always NO_HASH here
2120           for (uint i = 0; i < box_node->outcnt(); i++) {
2121             Node* u = box_node->raw_out(i);
2122             if (u->is_AbstractLock()) {
2123               alock = u->as_AbstractLock();
2124               if (alock->box_node() == box_node) {
2125                 // Verify that this Box is referenced only by related locks.
2126                 assert(alock->obj_node()->eqv_uncast(obj), "");
2127                 // Mark all related locks and unlocks.
2128 #ifdef ASSERT
2129                 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2130 #endif
2131                 alock->set_nested();
2132               }
2133             }
2134           }
2135         } else {
2136 #ifdef ASSERT
2137           alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2138           if (C->log() != NULL)
2139             alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2140 #endif
2141         }
2142       }
2143       return;
2144     }
2145     // Process locks for non escaping object
2146     assert(alock->is_non_esc_obj(), "");
2147   } // EliminateNestedLocks
2148 
2149   if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2150     // Look for all locks of this object and mark them and
2151     // corresponding BoxLock nodes as eliminated.
2152     Node* obj = alock->obj_node();
2153     for (uint j = 0; j < obj->outcnt(); j++) {
2154       Node* o = obj->raw_out(j);
2155       if (o->is_AbstractLock() &&
2156           o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2157         alock = o->as_AbstractLock();
2158         Node* box = alock->box_node();
2159         // Replace old box node with new eliminated box for all users
2160         // of the same object and mark related locks as eliminated.
2161         mark_eliminated_box(box, obj);
2162       }
2163     }
2164   }
2165 }
2166 
2167 // we have determined that this lock/unlock can be eliminated, we simply
2168 // eliminate the node without expanding it.
2169 //
2170 // Note:  The membar's associated with the lock/unlock are currently not
2171 //        eliminated.  This should be investigated as a future enhancement.
2172 //
2173 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2174 
2175   if (!alock->is_eliminated()) {
2176     return false;
2177   }
2178 #ifdef ASSERT
2179   if (!alock->is_coarsened()) {
2180     // Check that new "eliminated" BoxLock node is created.
2181     BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2182     assert(oldbox->is_eliminated(), "should be done already");
2183   }
2184 #endif
2185 
2186   alock->log_lock_optimization(C, "eliminate_lock");
2187 
2188 #ifndef PRODUCT
2189   if (PrintEliminateLocks) {
2190     if (alock->is_Lock()) {
2191       tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2192     } else {
2193       tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2194     }
2195   }
2196 #endif
2197 
2198   Node* mem  = alock->in(TypeFunc::Memory);
2199   Node* ctrl = alock->in(TypeFunc::Control);
2200 
2201   extract_call_projections(alock);
2202   // There are 2 projections from the lock.  The lock node will
2203   // be deleted when its last use is subsumed below.
2204   assert(alock->outcnt() == 2 &&
2205          _fallthroughproj != NULL &&
2206          _memproj_fallthrough != NULL,
2207          "Unexpected projections from Lock/Unlock");
2208 
2209   Node* fallthroughproj = _fallthroughproj;
2210   Node* memproj_fallthrough = _memproj_fallthrough;
2211 
2212   // The memory projection from a lock/unlock is RawMem
2213   // The input to a Lock is merged memory, so extract its RawMem input
2214   // (unless the MergeMem has been optimized away.)
2215   if (alock->is_Lock()) {
2216     // Seach for MemBarAcquireLock node and delete it also.
2217     MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2218     assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2219     Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2220     Node* memproj = membar->proj_out(TypeFunc::Memory);
2221     _igvn.replace_node(ctrlproj, fallthroughproj);
2222     _igvn.replace_node(memproj, memproj_fallthrough);
2223 
2224     // Delete FastLock node also if this Lock node is unique user
2225     // (a loop peeling may clone a Lock node).
2226     Node* flock = alock->as_Lock()->fastlock_node();
2227     if (flock->outcnt() == 1) {
2228       assert(flock->unique_out() == alock, "sanity");
2229       _igvn.replace_node(flock, top());
2230     }
2231   }
2232 
2233   // Seach for MemBarReleaseLock node and delete it also.
2234   if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2235       ctrl->in(0)->is_MemBar()) {
2236     MemBarNode* membar = ctrl->in(0)->as_MemBar();
2237     assert(membar->Opcode() == Op_MemBarReleaseLock &&
2238            mem->is_Proj() && membar == mem->in(0), "");
2239     _igvn.replace_node(fallthroughproj, ctrl);
2240     _igvn.replace_node(memproj_fallthrough, mem);
2241     fallthroughproj = ctrl;
2242     memproj_fallthrough = mem;
2243     ctrl = membar->in(TypeFunc::Control);
2244     mem  = membar->in(TypeFunc::Memory);
2245   }
2246 
2247   _igvn.replace_node(fallthroughproj, ctrl);
2248   _igvn.replace_node(memproj_fallthrough, mem);
2249   return true;
2250 }
2251 
2252 
2253 //------------------------------expand_lock_node----------------------
2254 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2255 
2256   Node* ctrl = lock->in(TypeFunc::Control);
2257   Node* mem = lock->in(TypeFunc::Memory);
2258   Node* obj = lock->obj_node();
2259   Node* box = lock->box_node();
2260   Node* flock = lock->fastlock_node();
2261 
2262   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2263 
2264   // Make the merge point
2265   Node *region;
2266   Node *mem_phi;
2267   Node *slow_path;
2268 
2269   if (UseOptoBiasInlining) {
2270     /*
2271      *  See the full description in MacroAssembler::biased_locking_enter().
2272      *
2273      *  if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2274      *    // The object is biased.
2275      *    proto_node = klass->prototype_header;
2276      *    o_node = thread | proto_node;
2277      *    x_node = o_node ^ mark_word;
2278      *    if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2279      *      // Done.
2280      *    } else {
2281      *      if( (x_node & biased_lock_mask) != 0 ) {
2282      *        // The klass's prototype header is no longer biased.
2283      *        cas(&mark_word, mark_word, proto_node)
2284      *        goto cas_lock;
2285      *      } else {
2286      *        // The klass's prototype header is still biased.
2287      *        if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2288      *          old = mark_word;
2289      *          new = o_node;
2290      *        } else {
2291      *          // Different thread or anonymous biased.
2292      *          old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2293      *          new = thread | old;
2294      *        }
2295      *        // Try to rebias.
2296      *        if( cas(&mark_word, old, new) == 0 ) {
2297      *          // Done.
2298      *        } else {
2299      *          goto slow_path; // Failed.
2300      *        }
2301      *      }
2302      *    }
2303      *  } else {
2304      *    // The object is not biased.
2305      *    cas_lock:
2306      *    if( FastLock(obj) == 0 ) {
2307      *      // Done.
2308      *    } else {
2309      *      slow_path:
2310      *      OptoRuntime::complete_monitor_locking_Java(obj);
2311      *    }
2312      *  }
2313      */
2314 
2315     region  = new RegionNode(5);
2316     // create a Phi for the memory state
2317     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2318 
2319     Node* fast_lock_region  = new RegionNode(3);
2320     Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2321 
2322     // First, check mark word for the biased lock pattern.
2323     Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2324 
2325     // Get fast path - mark word has the biased lock pattern.
2326     ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2327                          markOopDesc::biased_lock_mask_in_place,
2328                          markOopDesc::biased_lock_pattern, true);
2329     // fast_lock_region->in(1) is set to slow path.
2330     fast_lock_mem_phi->init_req(1, mem);
2331 
2332     // Now check that the lock is biased to the current thread and has
2333     // the same epoch and bias as Klass::_prototype_header.
2334 
2335     // Special-case a fresh allocation to avoid building nodes:
2336     Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2337     if (klass_node == NULL) {
2338       Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2339       klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2340 #ifdef _LP64
2341       if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2342         assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2343         klass_node->in(1)->init_req(0, ctrl);
2344       } else
2345 #endif
2346       klass_node->init_req(0, ctrl);
2347     }
2348     Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2349 
2350     Node* thread = transform_later(new ThreadLocalNode());
2351     Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2352     Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2353     Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2354 
2355     // Get slow path - mark word does NOT match the value.
2356     Node* not_biased_ctrl =  opt_bits_test(ctrl, region, 3, x_node,
2357                                       (~markOopDesc::age_mask_in_place), 0);
2358     // region->in(3) is set to fast path - the object is biased to the current thread.
2359     mem_phi->init_req(3, mem);
2360 
2361 
2362     // Mark word does NOT match the value (thread | Klass::_prototype_header).
2363 
2364 
2365     // First, check biased pattern.
2366     // Get fast path - _prototype_header has the same biased lock pattern.
2367     ctrl =  opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2368                           markOopDesc::biased_lock_mask_in_place, 0, true);
2369 
2370     not_biased_ctrl = fast_lock_region->in(2); // Slow path
2371     // fast_lock_region->in(2) - the prototype header is no longer biased
2372     // and we have to revoke the bias on this object.
2373     // We are going to try to reset the mark of this object to the prototype
2374     // value and fall through to the CAS-based locking scheme.
2375     Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2376     Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2377                                           proto_node, mark_node);
2378     transform_later(cas);
2379     Node* proj = transform_later(new SCMemProjNode(cas));
2380     fast_lock_mem_phi->init_req(2, proj);
2381 
2382 
2383     // Second, check epoch bits.
2384     Node* rebiased_region  = new RegionNode(3);
2385     Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2386     Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2387 
2388     // Get slow path - mark word does NOT match epoch bits.
2389     Node* epoch_ctrl =  opt_bits_test(ctrl, rebiased_region, 1, x_node,
2390                                       markOopDesc::epoch_mask_in_place, 0);
2391     // The epoch of the current bias is not valid, attempt to rebias the object
2392     // toward the current thread.
2393     rebiased_region->init_req(2, epoch_ctrl);
2394     old_phi->init_req(2, mark_node);
2395     new_phi->init_req(2, o_node);
2396 
2397     // rebiased_region->in(1) is set to fast path.
2398     // The epoch of the current bias is still valid but we know
2399     // nothing about the owner; it might be set or it might be clear.
2400     Node* cmask   = MakeConX(markOopDesc::biased_lock_mask_in_place |
2401                              markOopDesc::age_mask_in_place |
2402                              markOopDesc::epoch_mask_in_place);
2403     Node* old = transform_later(new AndXNode(mark_node, cmask));
2404     cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2405     Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2406     old_phi->init_req(1, old);
2407     new_phi->init_req(1, new_mark);
2408 
2409     transform_later(rebiased_region);
2410     transform_later(old_phi);
2411     transform_later(new_phi);
2412 
2413     // Try to acquire the bias of the object using an atomic operation.
2414     // If this fails we will go in to the runtime to revoke the object's bias.
2415     cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2416     transform_later(cas);
2417     proj = transform_later(new SCMemProjNode(cas));
2418 
2419     // Get slow path - Failed to CAS.
2420     not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2421     mem_phi->init_req(4, proj);
2422     // region->in(4) is set to fast path - the object is rebiased to the current thread.
2423 
2424     // Failed to CAS.
2425     slow_path  = new RegionNode(3);
2426     Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2427 
2428     slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2429     slow_mem->init_req(1, proj);
2430 
2431     // Call CAS-based locking scheme (FastLock node).
2432 
2433     transform_later(fast_lock_region);
2434     transform_later(fast_lock_mem_phi);
2435 
2436     // Get slow path - FastLock failed to lock the object.
2437     ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2438     mem_phi->init_req(2, fast_lock_mem_phi);
2439     // region->in(2) is set to fast path - the object is locked to the current thread.
2440 
2441     slow_path->init_req(2, ctrl); // Capture slow-control
2442     slow_mem->init_req(2, fast_lock_mem_phi);
2443 
2444     transform_later(slow_path);
2445     transform_later(slow_mem);
2446     // Reset lock's memory edge.
2447     lock->set_req(TypeFunc::Memory, slow_mem);
2448 
2449   } else {
2450     region  = new RegionNode(3);
2451     // create a Phi for the memory state
2452     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2453 
2454     // Optimize test; set region slot 2
2455     slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2456     mem_phi->init_req(2, mem);
2457   }
2458 
2459   // Make slow path call
2460   CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2461                                   OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2462                                   obj, box, NULL);
2463 
2464   extract_call_projections(call);
2465 
2466   // Slow path can only throw asynchronous exceptions, which are always
2467   // de-opted.  So the compiler thinks the slow-call can never throw an
2468   // exception.  If it DOES throw an exception we would need the debug
2469   // info removed first (since if it throws there is no monitor).
2470   assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2471            _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2472 
2473   // Capture slow path
2474   // disconnect fall-through projection from call and create a new one
2475   // hook up users of fall-through projection to region
2476   Node *slow_ctrl = _fallthroughproj->clone();
2477   transform_later(slow_ctrl);
2478   _igvn.hash_delete(_fallthroughproj);
2479   _fallthroughproj->disconnect_inputs(NULL, C);
2480   region->init_req(1, slow_ctrl);
2481   // region inputs are now complete
2482   transform_later(region);
2483   _igvn.replace_node(_fallthroughproj, region);
2484 
2485   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2486   mem_phi->init_req(1, memproj );
2487   transform_later(mem_phi);
2488   _igvn.replace_node(_memproj_fallthrough, mem_phi);
2489 }
2490 
2491 //------------------------------expand_unlock_node----------------------
2492 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2493 
2494   Node* ctrl = unlock->in(TypeFunc::Control);
2495   Node* mem = unlock->in(TypeFunc::Memory);
2496   Node* obj = unlock->obj_node();
2497   Node* box = unlock->box_node();
2498 
2499   assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2500 
2501   // No need for a null check on unlock
2502 
2503   // Make the merge point
2504   Node *region;
2505   Node *mem_phi;
2506 
2507   if (UseOptoBiasInlining) {
2508     // Check for biased locking unlock case, which is a no-op.
2509     // See the full description in MacroAssembler::biased_locking_exit().
2510     region  = new RegionNode(4);
2511     // create a Phi for the memory state
2512     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2513     mem_phi->init_req(3, mem);
2514 
2515     Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2516     ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2517                          markOopDesc::biased_lock_mask_in_place,
2518                          markOopDesc::biased_lock_pattern);
2519   } else {
2520     region  = new RegionNode(3);
2521     // create a Phi for the memory state
2522     mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2523   }
2524 
2525   FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2526   funlock = transform_later( funlock )->as_FastUnlock();
2527   // Optimize test; set region slot 2
2528   Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2529   Node *thread = transform_later(new ThreadLocalNode());
2530 
2531   CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2532                                   CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2533                                   "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2534 
2535   extract_call_projections(call);
2536 
2537   assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2538            _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2539 
2540   // No exceptions for unlocking
2541   // Capture slow path
2542   // disconnect fall-through projection from call and create a new one
2543   // hook up users of fall-through projection to region
2544   Node *slow_ctrl = _fallthroughproj->clone();
2545   transform_later(slow_ctrl);
2546   _igvn.hash_delete(_fallthroughproj);
2547   _fallthroughproj->disconnect_inputs(NULL, C);
2548   region->init_req(1, slow_ctrl);
2549   // region inputs are now complete
2550   transform_later(region);
2551   _igvn.replace_node(_fallthroughproj, region);
2552 
2553   Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2554   mem_phi->init_req(1, memproj );
2555   mem_phi->init_req(2, mem);
2556   transform_later(mem_phi);
2557   _igvn.replace_node(_memproj_fallthrough, mem_phi);
2558 }
2559 
2560 //---------------------------eliminate_macro_nodes----------------------
2561 // Eliminate scalar replaced allocations and associated locks.
2562 void PhaseMacroExpand::eliminate_macro_nodes() {
2563   if (C->macro_count() == 0)
2564     return;
2565 
2566   // First, attempt to eliminate locks
2567   int cnt = C->macro_count();
2568   for (int i=0; i < cnt; i++) {
2569     Node *n = C->macro_node(i);
2570     if (n->is_AbstractLock()) { // Lock and Unlock nodes
2571       // Before elimination mark all associated (same box and obj)
2572       // lock and unlock nodes.
2573       mark_eliminated_locking_nodes(n->as_AbstractLock());
2574     }
2575   }
2576   bool progress = true;
2577   while (progress) {
2578     progress = false;
2579     for (int i = C->macro_count(); i > 0; i--) {
2580       Node * n = C->macro_node(i-1);
2581       bool success = false;
2582       debug_only(int old_macro_count = C->macro_count(););
2583       if (n->is_AbstractLock()) {
2584         success = eliminate_locking_node(n->as_AbstractLock());
2585       }
2586       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2587       progress = progress || success;
2588     }
2589   }
2590   // Next, attempt to eliminate allocations
2591   _has_locks = false;
2592   progress = true;
2593   while (progress) {
2594     progress = false;
2595     for (int i = C->macro_count(); i > 0; i--) {
2596       Node * n = C->macro_node(i-1);
2597       bool success = false;
2598       debug_only(int old_macro_count = C->macro_count(););
2599       switch (n->class_id()) {
2600       case Node::Class_Allocate:
2601       case Node::Class_AllocateArray:
2602         success = eliminate_allocate_node(n->as_Allocate());
2603         break;
2604       case Node::Class_CallStaticJava:
2605         success = eliminate_boxing_node(n->as_CallStaticJava());
2606         break;
2607       case Node::Class_Lock:
2608       case Node::Class_Unlock:
2609         assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2610         _has_locks = true;
2611         break;
2612       case Node::Class_ArrayCopy:
2613         break;
2614       default:
2615         assert(n->Opcode() == Op_LoopLimit ||
2616                n->Opcode() == Op_Opaque1   ||
2617                n->Opcode() == Op_Opaque2   ||
2618                n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2619       }
2620       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2621       progress = progress || success;
2622     }
2623   }
2624 }
2625 
2626 //------------------------------expand_macro_nodes----------------------
2627 //  Returns true if a failure occurred.
2628 bool PhaseMacroExpand::expand_macro_nodes() {
2629   // Last attempt to eliminate macro nodes.
2630   eliminate_macro_nodes();
2631 
2632   // Make sure expansion will not cause node limit to be exceeded.
2633   // Worst case is a macro node gets expanded into about 50 nodes.
2634   // Allow 50% more for optimization.
2635   if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2636     return true;
2637 
2638   // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2639   bool progress = true;
2640   while (progress) {
2641     progress = false;
2642     for (int i = C->macro_count(); i > 0; i--) {
2643       Node * n = C->macro_node(i-1);
2644       bool success = false;
2645       debug_only(int old_macro_count = C->macro_count(););
2646       if (n->Opcode() == Op_LoopLimit) {
2647         // Remove it from macro list and put on IGVN worklist to optimize.
2648         C->remove_macro_node(n);
2649         _igvn._worklist.push(n);
2650         success = true;
2651       } else if (n->Opcode() == Op_CallStaticJava) {
2652         // Remove it from macro list and put on IGVN worklist to optimize.
2653         C->remove_macro_node(n);
2654         _igvn._worklist.push(n);
2655         success = true;
2656       } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2657         _igvn.replace_node(n, n->in(1));
2658         success = true;
2659 #if INCLUDE_RTM_OPT
2660       } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2661         assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2662         assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2663         Node* cmp = n->unique_out();
2664 #ifdef ASSERT
2665         // Validate graph.
2666         assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2667         BoolNode* bol = cmp->unique_out()->as_Bool();
2668         assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2669                (bol->_test._test == BoolTest::ne), "");
2670         IfNode* ifn = bol->unique_out()->as_If();
2671         assert((ifn->outcnt() == 2) &&
2672                ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2673 #endif
2674         Node* repl = n->in(1);
2675         if (!_has_locks) {
2676           // Remove RTM state check if there are no locks in the code.
2677           // Replace input to compare the same value.
2678           repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2679         }
2680         _igvn.replace_node(n, repl);
2681         success = true;
2682 #endif
2683       }
2684       assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2685       progress = progress || success;
2686     }
2687   }
2688 
2689   // expand arraycopy "macro" nodes first
2690   // For ReduceBulkZeroing, we must first process all arraycopy nodes
2691   // before the allocate nodes are expanded.
2692   int macro_idx = C->macro_count() - 1;
2693   while (macro_idx >= 0) {
2694     Node * n = C->macro_node(macro_idx);
2695     assert(n->is_macro(), "only macro nodes expected here");
2696     if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2697       // node is unreachable, so don't try to expand it
2698       C->remove_macro_node(n);
2699     } else if (n->is_ArrayCopy()){
2700       int macro_count = C->macro_count();
2701       expand_arraycopy_node(n->as_ArrayCopy());
2702       assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2703     }
2704     if (C->failing())  return true;
2705     macro_idx --;
2706   }
2707 
2708   // expand "macro" nodes
2709   // nodes are removed from the macro list as they are processed
2710   while (C->macro_count() > 0) {
2711     int macro_count = C->macro_count();
2712     Node * n = C->macro_node(macro_count-1);
2713     assert(n->is_macro(), "only macro nodes expected here");
2714     if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2715       // node is unreachable, so don't try to expand it
2716       C->remove_macro_node(n);
2717       continue;
2718     }
2719     switch (n->class_id()) {
2720     case Node::Class_Allocate:
2721       expand_allocate(n->as_Allocate());
2722       break;
2723     case Node::Class_AllocateArray:
2724       expand_allocate_array(n->as_AllocateArray());
2725       break;
2726     case Node::Class_Lock:
2727       expand_lock_node(n->as_Lock());
2728       break;
2729     case Node::Class_Unlock:
2730       expand_unlock_node(n->as_Unlock());
2731       break;
2732     default:
2733       assert(false, "unknown node type in macro list");
2734     }
2735     assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2736     if (C->failing())  return true;
2737   }
2738 
2739   _igvn.set_delay_transform(false);
2740   _igvn.optimize();
2741   if (C->failing())  return true;
2742   return false;
2743 }