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