1 /*
   2  * Copyright (c) 2001, 2016, 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 "ci/ciValueKlass.hpp"
  28 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
  29 #include "gc/g1/heapRegion.hpp"
  30 #include "gc/shared/barrierSet.hpp"
  31 #include "gc/shared/cardTableModRefBS.hpp"
  32 #include "gc/shared/collectedHeap.hpp"
  33 #include "memory/resourceArea.hpp"
  34 #include "opto/addnode.hpp"
  35 #include "opto/castnode.hpp"
  36 #include "opto/convertnode.hpp"
  37 #include "opto/graphKit.hpp"
  38 #include "opto/idealKit.hpp"
  39 #include "opto/intrinsicnode.hpp"
  40 #include "opto/locknode.hpp"
  41 #include "opto/machnode.hpp"
  42 #include "opto/opaquenode.hpp"
  43 #include "opto/parse.hpp"
  44 #include "opto/rootnode.hpp"
  45 #include "opto/runtime.hpp"
  46 #include "opto/valuetypenode.hpp"
  47 #include "runtime/deoptimization.hpp"
  48 #include "runtime/sharedRuntime.hpp"
  49 
  50 //----------------------------GraphKit-----------------------------------------
  51 // Main utility constructor.
  52 GraphKit::GraphKit(JVMState* jvms)
  53   : Phase(Phase::Parser),
  54     _env(C->env()),
  55     _gvn(*C->initial_gvn())
  56 {
  57   _exceptions = jvms->map()->next_exception();
  58   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  59   set_jvms(jvms);
  60 }
  61 
  62 // Private constructor for parser.
  63 GraphKit::GraphKit()
  64   : Phase(Phase::Parser),
  65     _env(C->env()),
  66     _gvn(*C->initial_gvn())
  67 {
  68   _exceptions = NULL;
  69   set_map(NULL);
  70   debug_only(_sp = -99);
  71   debug_only(set_bci(-99));
  72 }
  73 
  74 
  75 
  76 //---------------------------clean_stack---------------------------------------
  77 // Clear away rubbish from the stack area of the JVM state.
  78 // This destroys any arguments that may be waiting on the stack.
  79 void GraphKit::clean_stack(int from_sp) {
  80   SafePointNode* map      = this->map();
  81   JVMState*      jvms     = this->jvms();
  82   int            stk_size = jvms->stk_size();
  83   int            stkoff   = jvms->stkoff();
  84   Node*          top      = this->top();
  85   for (int i = from_sp; i < stk_size; i++) {
  86     if (map->in(stkoff + i) != top) {
  87       map->set_req(stkoff + i, top);
  88     }
  89   }
  90 }
  91 
  92 
  93 //--------------------------------sync_jvms-----------------------------------
  94 // Make sure our current jvms agrees with our parse state.
  95 JVMState* GraphKit::sync_jvms() const {
  96   JVMState* jvms = this->jvms();
  97   jvms->set_bci(bci());       // Record the new bci in the JVMState
  98   jvms->set_sp(sp());         // Record the new sp in the JVMState
  99   assert(jvms_in_sync(), "jvms is now in sync");
 100   return jvms;
 101 }
 102 
 103 //--------------------------------sync_jvms_for_reexecute---------------------
 104 // Make sure our current jvms agrees with our parse state.  This version
 105 // uses the reexecute_sp for reexecuting bytecodes.
 106 JVMState* GraphKit::sync_jvms_for_reexecute() {
 107   JVMState* jvms = this->jvms();
 108   jvms->set_bci(bci());          // Record the new bci in the JVMState
 109   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 110   return jvms;
 111 }
 112 
 113 #ifdef ASSERT
 114 bool GraphKit::jvms_in_sync() const {
 115   Parse* parse = is_Parse();
 116   if (parse == NULL) {
 117     if (bci() !=      jvms()->bci())          return false;
 118     if (sp()  != (int)jvms()->sp())           return false;
 119     return true;
 120   }
 121   if (jvms()->method() != parse->method())    return false;
 122   if (jvms()->bci()    != parse->bci())       return false;
 123   int jvms_sp = jvms()->sp();
 124   if (jvms_sp          != parse->sp())        return false;
 125   int jvms_depth = jvms()->depth();
 126   if (jvms_depth       != parse->depth())     return false;
 127   return true;
 128 }
 129 
 130 // Local helper checks for special internal merge points
 131 // used to accumulate and merge exception states.
 132 // They are marked by the region's in(0) edge being the map itself.
 133 // Such merge points must never "escape" into the parser at large,
 134 // until they have been handed to gvn.transform.
 135 static bool is_hidden_merge(Node* reg) {
 136   if (reg == NULL)  return false;
 137   if (reg->is_Phi()) {
 138     reg = reg->in(0);
 139     if (reg == NULL)  return false;
 140   }
 141   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 142 }
 143 
 144 void GraphKit::verify_map() const {
 145   if (map() == NULL)  return;  // null map is OK
 146   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 147   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 148   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 149 }
 150 
 151 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 152   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 153   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 154 }
 155 #endif
 156 
 157 //---------------------------stop_and_kill_map---------------------------------
 158 // Set _map to NULL, signalling a stop to further bytecode execution.
 159 // First smash the current map's control to a constant, to mark it dead.
 160 void GraphKit::stop_and_kill_map() {
 161   SafePointNode* dead_map = stop();
 162   if (dead_map != NULL) {
 163     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
 164     assert(dead_map->is_killed(), "must be so marked");
 165   }
 166 }
 167 
 168 
 169 //--------------------------------stopped--------------------------------------
 170 // Tell if _map is NULL, or control is top.
 171 bool GraphKit::stopped() {
 172   if (map() == NULL)           return true;
 173   else if (control() == top()) return true;
 174   else                         return false;
 175 }
 176 
 177 
 178 //-----------------------------has_ex_handler----------------------------------
 179 // Tell if this method or any caller method has exception handlers.
 180 bool GraphKit::has_ex_handler() {
 181   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 182     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 183       return true;
 184     }
 185   }
 186   return false;
 187 }
 188 
 189 //------------------------------save_ex_oop------------------------------------
 190 // Save an exception without blowing stack contents or other JVM state.
 191 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 192   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 193   ex_map->add_req(ex_oop);
 194   debug_only(verify_exception_state(ex_map));
 195 }
 196 
 197 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 198   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 199   Node* ex_oop = ex_map->in(ex_map->req()-1);
 200   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 201   return ex_oop;
 202 }
 203 
 204 //-----------------------------saved_ex_oop------------------------------------
 205 // Recover a saved exception from its map.
 206 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 207   return common_saved_ex_oop(ex_map, false);
 208 }
 209 
 210 //--------------------------clear_saved_ex_oop---------------------------------
 211 // Erase a previously saved exception from its map.
 212 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 213   return common_saved_ex_oop(ex_map, true);
 214 }
 215 
 216 #ifdef ASSERT
 217 //---------------------------has_saved_ex_oop----------------------------------
 218 // Erase a previously saved exception from its map.
 219 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 220   return ex_map->req() == ex_map->jvms()->endoff()+1;
 221 }
 222 #endif
 223 
 224 //-------------------------make_exception_state--------------------------------
 225 // Turn the current JVM state into an exception state, appending the ex_oop.
 226 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 227   sync_jvms();
 228   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 229   set_saved_ex_oop(ex_map, ex_oop);
 230   return ex_map;
 231 }
 232 
 233 
 234 //--------------------------add_exception_state--------------------------------
 235 // Add an exception to my list of exceptions.
 236 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 237   if (ex_map == NULL || ex_map->control() == top()) {
 238     return;
 239   }
 240 #ifdef ASSERT
 241   verify_exception_state(ex_map);
 242   if (has_exceptions()) {
 243     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 244   }
 245 #endif
 246 
 247   // If there is already an exception of exactly this type, merge with it.
 248   // In particular, null-checks and other low-level exceptions common up here.
 249   Node*       ex_oop  = saved_ex_oop(ex_map);
 250   const Type* ex_type = _gvn.type(ex_oop);
 251   if (ex_oop == top()) {
 252     // No action needed.
 253     return;
 254   }
 255   assert(ex_type->isa_instptr(), "exception must be an instance");
 256   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 257     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 258     // We check sp also because call bytecodes can generate exceptions
 259     // both before and after arguments are popped!
 260     if (ex_type2 == ex_type
 261         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 262       combine_exception_states(ex_map, e2);
 263       return;
 264     }
 265   }
 266 
 267   // No pre-existing exception of the same type.  Chain it on the list.
 268   push_exception_state(ex_map);
 269 }
 270 
 271 //-----------------------add_exception_states_from-----------------------------
 272 void GraphKit::add_exception_states_from(JVMState* jvms) {
 273   SafePointNode* ex_map = jvms->map()->next_exception();
 274   if (ex_map != NULL) {
 275     jvms->map()->set_next_exception(NULL);
 276     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 277       next_map = ex_map->next_exception();
 278       ex_map->set_next_exception(NULL);
 279       add_exception_state(ex_map);
 280     }
 281   }
 282 }
 283 
 284 //-----------------------transfer_exceptions_into_jvms-------------------------
 285 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 286   if (map() == NULL) {
 287     // We need a JVMS to carry the exceptions, but the map has gone away.
 288     // Create a scratch JVMS, cloned from any of the exception states...
 289     if (has_exceptions()) {
 290       _map = _exceptions;
 291       _map = clone_map();
 292       _map->set_next_exception(NULL);
 293       clear_saved_ex_oop(_map);
 294       debug_only(verify_map());
 295     } else {
 296       // ...or created from scratch
 297       JVMState* jvms = new (C) JVMState(_method, NULL);
 298       jvms->set_bci(_bci);
 299       jvms->set_sp(_sp);
 300       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
 301       set_jvms(jvms);
 302       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 303       set_all_memory(top());
 304       while (map()->req() < jvms->endoff())  map()->add_req(top());
 305     }
 306     // (This is a kludge, in case you didn't notice.)
 307     set_control(top());
 308   }
 309   JVMState* jvms = sync_jvms();
 310   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 311   jvms->map()->set_next_exception(_exceptions);
 312   _exceptions = NULL;   // done with this set of exceptions
 313   return jvms;
 314 }
 315 
 316 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 317   assert(is_hidden_merge(dstphi), "must be a special merge node");
 318   assert(is_hidden_merge(srcphi), "must be a special merge node");
 319   uint limit = srcphi->req();
 320   for (uint i = PhiNode::Input; i < limit; i++) {
 321     dstphi->add_req(srcphi->in(i));
 322   }
 323 }
 324 static inline void add_one_req(Node* dstphi, Node* src) {
 325   assert(is_hidden_merge(dstphi), "must be a special merge node");
 326   assert(!is_hidden_merge(src), "must not be a special merge node");
 327   dstphi->add_req(src);
 328 }
 329 
 330 //-----------------------combine_exception_states------------------------------
 331 // This helper function combines exception states by building phis on a
 332 // specially marked state-merging region.  These regions and phis are
 333 // untransformed, and can build up gradually.  The region is marked by
 334 // having a control input of its exception map, rather than NULL.  Such
 335 // regions do not appear except in this function, and in use_exception_state.
 336 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 337   if (failing())  return;  // dying anyway...
 338   JVMState* ex_jvms = ex_map->_jvms;
 339   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 340   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 341   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 342   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 343   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 344   assert(ex_map->req() == phi_map->req(), "matching maps");
 345   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 346   Node*         hidden_merge_mark = root();
 347   Node*         region  = phi_map->control();
 348   MergeMemNode* phi_mem = phi_map->merged_memory();
 349   MergeMemNode* ex_mem  = ex_map->merged_memory();
 350   if (region->in(0) != hidden_merge_mark) {
 351     // The control input is not (yet) a specially-marked region in phi_map.
 352     // Make it so, and build some phis.
 353     region = new RegionNode(2);
 354     _gvn.set_type(region, Type::CONTROL);
 355     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 356     region->init_req(1, phi_map->control());
 357     phi_map->set_control(region);
 358     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 359     record_for_igvn(io_phi);
 360     _gvn.set_type(io_phi, Type::ABIO);
 361     phi_map->set_i_o(io_phi);
 362     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 363       Node* m = mms.memory();
 364       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 365       record_for_igvn(m_phi);
 366       _gvn.set_type(m_phi, Type::MEMORY);
 367       mms.set_memory(m_phi);
 368     }
 369   }
 370 
 371   // Either or both of phi_map and ex_map might already be converted into phis.
 372   Node* ex_control = ex_map->control();
 373   // if there is special marking on ex_map also, we add multiple edges from src
 374   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 375   // how wide was the destination phi_map, originally?
 376   uint orig_width = region->req();
 377 
 378   if (add_multiple) {
 379     add_n_reqs(region, ex_control);
 380     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 381   } else {
 382     // ex_map has no merges, so we just add single edges everywhere
 383     add_one_req(region, ex_control);
 384     add_one_req(phi_map->i_o(), ex_map->i_o());
 385   }
 386   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 387     if (mms.is_empty()) {
 388       // get a copy of the base memory, and patch some inputs into it
 389       const TypePtr* adr_type = mms.adr_type(C);
 390       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 391       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 392       mms.set_memory(phi);
 393       // Prepare to append interesting stuff onto the newly sliced phi:
 394       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 395     }
 396     // Append stuff from ex_map:
 397     if (add_multiple) {
 398       add_n_reqs(mms.memory(), mms.memory2());
 399     } else {
 400       add_one_req(mms.memory(), mms.memory2());
 401     }
 402   }
 403   uint limit = ex_map->req();
 404   for (uint i = TypeFunc::Parms; i < limit; i++) {
 405     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 406     if (i == tos)  i = ex_jvms->monoff();
 407     Node* src = ex_map->in(i);
 408     Node* dst = phi_map->in(i);
 409     if (src != dst) {
 410       PhiNode* phi;
 411       if (dst->in(0) != region) {
 412         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 413         record_for_igvn(phi);
 414         _gvn.set_type(phi, phi->type());
 415         phi_map->set_req(i, dst);
 416         // Prepare to append interesting stuff onto the new phi:
 417         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 418       } else {
 419         assert(dst->is_Phi(), "nobody else uses a hidden region");
 420         phi = dst->as_Phi();
 421       }
 422       if (add_multiple && src->in(0) == ex_control) {
 423         // Both are phis.
 424         add_n_reqs(dst, src);
 425       } else {
 426         while (dst->req() < region->req())  add_one_req(dst, src);
 427       }
 428       const Type* srctype = _gvn.type(src);
 429       if (phi->type() != srctype) {
 430         const Type* dsttype = phi->type()->meet_speculative(srctype);
 431         if (phi->type() != dsttype) {
 432           phi->set_type(dsttype);
 433           _gvn.set_type(phi, dsttype);
 434         }
 435       }
 436     }
 437   }
 438   phi_map->merge_replaced_nodes_with(ex_map);
 439 }
 440 
 441 //--------------------------use_exception_state--------------------------------
 442 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 443   if (failing()) { stop(); return top(); }
 444   Node* region = phi_map->control();
 445   Node* hidden_merge_mark = root();
 446   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 447   Node* ex_oop = clear_saved_ex_oop(phi_map);
 448   if (region->in(0) == hidden_merge_mark) {
 449     // Special marking for internal ex-states.  Process the phis now.
 450     region->set_req(0, region);  // now it's an ordinary region
 451     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 452     // Note: Setting the jvms also sets the bci and sp.
 453     set_control(_gvn.transform(region));
 454     uint tos = jvms()->stkoff() + sp();
 455     for (uint i = 1; i < tos; i++) {
 456       Node* x = phi_map->in(i);
 457       if (x->in(0) == region) {
 458         assert(x->is_Phi(), "expected a special phi");
 459         phi_map->set_req(i, _gvn.transform(x));
 460       }
 461     }
 462     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 463       Node* x = mms.memory();
 464       if (x->in(0) == region) {
 465         assert(x->is_Phi(), "nobody else uses a hidden region");
 466         mms.set_memory(_gvn.transform(x));
 467       }
 468     }
 469     if (ex_oop->in(0) == region) {
 470       assert(ex_oop->is_Phi(), "expected a special phi");
 471       ex_oop = _gvn.transform(ex_oop);
 472     }
 473   } else {
 474     set_jvms(phi_map->jvms());
 475   }
 476 
 477   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 478   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 479   return ex_oop;
 480 }
 481 
 482 //---------------------------------java_bc-------------------------------------
 483 Bytecodes::Code GraphKit::java_bc() const {
 484   ciMethod* method = this->method();
 485   int       bci    = this->bci();
 486   if (method != NULL && bci != InvocationEntryBci)
 487     return method->java_code_at_bci(bci);
 488   else
 489     return Bytecodes::_illegal;
 490 }
 491 
 492 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 493                                                           bool must_throw) {
 494     // if the exception capability is set, then we will generate code
 495     // to check the JavaThread.should_post_on_exceptions flag to see
 496     // if we actually need to report exception events (for this
 497     // thread).  If we don't need to report exception events, we will
 498     // take the normal fast path provided by add_exception_events.  If
 499     // exception event reporting is enabled for this thread, we will
 500     // take the uncommon_trap in the BuildCutout below.
 501 
 502     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 503     Node* jthread = _gvn.transform(new ThreadLocalNode());
 504     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 505     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
 506 
 507     // Test the should_post_on_exceptions_flag vs. 0
 508     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
 509     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
 510 
 511     // Branch to slow_path if should_post_on_exceptions_flag was true
 512     { BuildCutout unless(this, tst, PROB_MAX);
 513       // Do not try anything fancy if we're notifying the VM on every throw.
 514       // Cf. case Bytecodes::_athrow in parse2.cpp.
 515       uncommon_trap(reason, Deoptimization::Action_none,
 516                     (ciKlass*)NULL, (char*)NULL, must_throw);
 517     }
 518 
 519 }
 520 
 521 //------------------------------builtin_throw----------------------------------
 522 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 523   bool must_throw = true;
 524 
 525   if (env()->jvmti_can_post_on_exceptions()) {
 526     // check if we must post exception events, take uncommon trap if so
 527     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 528     // here if should_post_on_exceptions is false
 529     // continue on with the normal codegen
 530   }
 531 
 532   // If this particular condition has not yet happened at this
 533   // bytecode, then use the uncommon trap mechanism, and allow for
 534   // a future recompilation if several traps occur here.
 535   // If the throw is hot, try to use a more complicated inline mechanism
 536   // which keeps execution inside the compiled code.
 537   bool treat_throw_as_hot = false;
 538   ciMethodData* md = method()->method_data();
 539 
 540   if (ProfileTraps) {
 541     if (too_many_traps(reason)) {
 542       treat_throw_as_hot = true;
 543     }
 544     // (If there is no MDO at all, assume it is early in
 545     // execution, and that any deopts are part of the
 546     // startup transient, and don't need to be remembered.)
 547 
 548     // Also, if there is a local exception handler, treat all throws
 549     // as hot if there has been at least one in this method.
 550     if (C->trap_count(reason) != 0
 551         && method()->method_data()->trap_count(reason) != 0
 552         && has_ex_handler()) {
 553         treat_throw_as_hot = true;
 554     }
 555   }
 556 
 557   // If this throw happens frequently, an uncommon trap might cause
 558   // a performance pothole.  If there is a local exception handler,
 559   // and if this particular bytecode appears to be deoptimizing often,
 560   // let us handle the throw inline, with a preconstructed instance.
 561   // Note:   If the deopt count has blown up, the uncommon trap
 562   // runtime is going to flush this nmethod, not matter what.
 563   if (treat_throw_as_hot
 564       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 565     // If the throw is local, we use a pre-existing instance and
 566     // punt on the backtrace.  This would lead to a missing backtrace
 567     // (a repeat of 4292742) if the backtrace object is ever asked
 568     // for its backtrace.
 569     // Fixing this remaining case of 4292742 requires some flavor of
 570     // escape analysis.  Leave that for the future.
 571     ciInstance* ex_obj = NULL;
 572     switch (reason) {
 573     case Deoptimization::Reason_null_check:
 574       ex_obj = env()->NullPointerException_instance();
 575       break;
 576     case Deoptimization::Reason_div0_check:
 577       ex_obj = env()->ArithmeticException_instance();
 578       break;
 579     case Deoptimization::Reason_range_check:
 580       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 581       break;
 582     case Deoptimization::Reason_class_check:
 583       if (java_bc() == Bytecodes::_aastore) {
 584         ex_obj = env()->ArrayStoreException_instance();
 585       } else {
 586         ex_obj = env()->ClassCastException_instance();
 587       }
 588       break;
 589     }
 590     if (failing()) { stop(); return; }  // exception allocation might fail
 591     if (ex_obj != NULL) {
 592       // Cheat with a preallocated exception object.
 593       if (C->log() != NULL)
 594         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 595                        Deoptimization::trap_reason_name(reason));
 596       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 597       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
 598 
 599       // Clear the detail message of the preallocated exception object.
 600       // Weblogic sometimes mutates the detail message of exceptions
 601       // using reflection.
 602       int offset = java_lang_Throwable::get_detailMessage_offset();
 603       const TypePtr* adr_typ = ex_con->add_offset(offset);
 604 
 605       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 606       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 607       // Conservatively release stores of object references.
 608       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
 609 
 610       add_exception_state(make_exception_state(ex_node));
 611       return;
 612     }
 613   }
 614 
 615   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 616   // It won't be much cheaper than bailing to the interp., since we'll
 617   // have to pass up all the debug-info, and the runtime will have to
 618   // create the stack trace.
 619 
 620   // Usual case:  Bail to interpreter.
 621   // Reserve the right to recompile if we haven't seen anything yet.
 622 
 623   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
 624   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 625   if (treat_throw_as_hot
 626       && (method()->method_data()->trap_recompiled_at(bci(), m)
 627           || C->too_many_traps(reason))) {
 628     // We cannot afford to take more traps here.  Suffer in the interpreter.
 629     if (C->log() != NULL)
 630       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 631                      Deoptimization::trap_reason_name(reason),
 632                      C->trap_count(reason));
 633     action = Deoptimization::Action_none;
 634   }
 635 
 636   // "must_throw" prunes the JVM state to include only the stack, if there
 637   // are no local exception handlers.  This should cut down on register
 638   // allocation time and code size, by drastically reducing the number
 639   // of in-edges on the call to the uncommon trap.
 640 
 641   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 642 }
 643 
 644 
 645 //----------------------------PreserveJVMState---------------------------------
 646 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 647   debug_only(kit->verify_map());
 648   _kit    = kit;
 649   _map    = kit->map();   // preserve the map
 650   _sp     = kit->sp();
 651   kit->set_map(clone_map ? kit->clone_map() : NULL);
 652 #ifdef ASSERT
 653   _bci    = kit->bci();
 654   Parse* parser = kit->is_Parse();
 655   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 656   _block  = block;
 657 #endif
 658 }
 659 PreserveJVMState::~PreserveJVMState() {
 660   GraphKit* kit = _kit;
 661 #ifdef ASSERT
 662   assert(kit->bci() == _bci, "bci must not shift");
 663   Parse* parser = kit->is_Parse();
 664   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 665   assert(block == _block,    "block must not shift");
 666 #endif
 667   kit->set_map(_map);
 668   kit->set_sp(_sp);
 669 }
 670 
 671 
 672 //-----------------------------BuildCutout-------------------------------------
 673 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 674   : PreserveJVMState(kit)
 675 {
 676   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 677   SafePointNode* outer_map = _map;   // preserved map is caller's
 678   SafePointNode* inner_map = kit->map();
 679   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 680   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
 681   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
 682 }
 683 BuildCutout::~BuildCutout() {
 684   GraphKit* kit = _kit;
 685   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 686 }
 687 
 688 //---------------------------PreserveReexecuteState----------------------------
 689 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 690   assert(!kit->stopped(), "must call stopped() before");
 691   _kit    =    kit;
 692   _sp     =    kit->sp();
 693   _reexecute = kit->jvms()->_reexecute;
 694 }
 695 PreserveReexecuteState::~PreserveReexecuteState() {
 696   if (_kit->stopped()) return;
 697   _kit->jvms()->_reexecute = _reexecute;
 698   _kit->set_sp(_sp);
 699 }
 700 
 701 //------------------------------clone_map--------------------------------------
 702 // Implementation of PreserveJVMState
 703 //
 704 // Only clone_map(...) here. If this function is only used in the
 705 // PreserveJVMState class we may want to get rid of this extra
 706 // function eventually and do it all there.
 707 
 708 SafePointNode* GraphKit::clone_map() {
 709   if (map() == NULL)  return NULL;
 710 
 711   // Clone the memory edge first
 712   Node* mem = MergeMemNode::make(map()->memory());
 713   gvn().set_type_bottom(mem);
 714 
 715   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 716   JVMState* jvms = this->jvms();
 717   JVMState* clonejvms = jvms->clone_shallow(C);
 718   clonemap->set_memory(mem);
 719   clonemap->set_jvms(clonejvms);
 720   clonejvms->set_map(clonemap);
 721   record_for_igvn(clonemap);
 722   gvn().set_type_bottom(clonemap);
 723   return clonemap;
 724 }
 725 
 726 
 727 //-----------------------------set_map_clone-----------------------------------
 728 void GraphKit::set_map_clone(SafePointNode* m) {
 729   _map = m;
 730   _map = clone_map();
 731   _map->set_next_exception(NULL);
 732   debug_only(verify_map());
 733 }
 734 
 735 
 736 //----------------------------kill_dead_locals---------------------------------
 737 // Detect any locals which are known to be dead, and force them to top.
 738 void GraphKit::kill_dead_locals() {
 739   // Consult the liveness information for the locals.  If any
 740   // of them are unused, then they can be replaced by top().  This
 741   // should help register allocation time and cut down on the size
 742   // of the deoptimization information.
 743 
 744   // This call is made from many of the bytecode handling
 745   // subroutines called from the Big Switch in do_one_bytecode.
 746   // Every bytecode which might include a slow path is responsible
 747   // for killing its dead locals.  The more consistent we
 748   // are about killing deads, the fewer useless phis will be
 749   // constructed for them at various merge points.
 750 
 751   // bci can be -1 (InvocationEntryBci).  We return the entry
 752   // liveness for the method.
 753 
 754   if (method() == NULL || method()->code_size() == 0) {
 755     // We are building a graph for a call to a native method.
 756     // All locals are live.
 757     return;
 758   }
 759 
 760   ResourceMark rm;
 761 
 762   // Consult the liveness information for the locals.  If any
 763   // of them are unused, then they can be replaced by top().  This
 764   // should help register allocation time and cut down on the size
 765   // of the deoptimization information.
 766   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 767 
 768   int len = (int)live_locals.size();
 769   assert(len <= jvms()->loc_size(), "too many live locals");
 770   for (int local = 0; local < len; local++) {
 771     if (!live_locals.at(local)) {
 772       set_local(local, top());
 773     }
 774   }
 775 }
 776 
 777 #ifdef ASSERT
 778 //-------------------------dead_locals_are_killed------------------------------
 779 // Return true if all dead locals are set to top in the map.
 780 // Used to assert "clean" debug info at various points.
 781 bool GraphKit::dead_locals_are_killed() {
 782   if (method() == NULL || method()->code_size() == 0) {
 783     // No locals need to be dead, so all is as it should be.
 784     return true;
 785   }
 786 
 787   // Make sure somebody called kill_dead_locals upstream.
 788   ResourceMark rm;
 789   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 790     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 791     SafePointNode* map = jvms->map();
 792     ciMethod* method = jvms->method();
 793     int       bci    = jvms->bci();
 794     if (jvms == this->jvms()) {
 795       bci = this->bci();  // it might not yet be synched
 796     }
 797     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 798     int len = (int)live_locals.size();
 799     if (!live_locals.is_valid() || len == 0)
 800       // This method is trivial, or is poisoned by a breakpoint.
 801       return true;
 802     assert(len == jvms->loc_size(), "live map consistent with locals map");
 803     for (int local = 0; local < len; local++) {
 804       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 805         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 806           tty->print_cr("Zombie local %d: ", local);
 807           jvms->dump();
 808         }
 809         return false;
 810       }
 811     }
 812   }
 813   return true;
 814 }
 815 
 816 #endif //ASSERT
 817 
 818 // Helper function for enforcing certain bytecodes to reexecute if
 819 // deoptimization happens
 820 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 821   ciMethod* cur_method = jvms->method();
 822   int       cur_bci   = jvms->bci();
 823   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 824     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 825     return Interpreter::bytecode_should_reexecute(code) ||
 826            is_anewarray && code == Bytecodes::_multianewarray;
 827     // Reexecute _multianewarray bytecode which was replaced with
 828     // sequence of [a]newarray. See Parse::do_multianewarray().
 829     //
 830     // Note: interpreter should not have it set since this optimization
 831     // is limited by dimensions and guarded by flag so in some cases
 832     // multianewarray() runtime calls will be generated and
 833     // the bytecode should not be reexecutes (stack will not be reset).
 834   } else
 835     return false;
 836 }
 837 
 838 // Helper function for adding JVMState and debug information to node
 839 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 840   // Add the safepoint edges to the call (or other safepoint).
 841 
 842   // Make sure dead locals are set to top.  This
 843   // should help register allocation time and cut down on the size
 844   // of the deoptimization information.
 845   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 846 
 847   // Walk the inline list to fill in the correct set of JVMState's
 848   // Also fill in the associated edges for each JVMState.
 849 
 850   // If the bytecode needs to be reexecuted we need to put
 851   // the arguments back on the stack.
 852   const bool should_reexecute = jvms()->should_reexecute();
 853   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 854 
 855   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 856   // undefined if the bci is different.  This is normal for Parse but it
 857   // should not happen for LibraryCallKit because only one bci is processed.
 858   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 859          "in LibraryCallKit the reexecute bit should not change");
 860 
 861   // If we are guaranteed to throw, we can prune everything but the
 862   // input to the current bytecode.
 863   bool can_prune_locals = false;
 864   uint stack_slots_not_pruned = 0;
 865   int inputs = 0, depth = 0;
 866   if (must_throw) {
 867     assert(method() == youngest_jvms->method(), "sanity");
 868     if (compute_stack_effects(inputs, depth)) {
 869       can_prune_locals = true;
 870       stack_slots_not_pruned = inputs;
 871     }
 872   }
 873 
 874   if (env()->should_retain_local_variables()) {
 875     // At any safepoint, this method can get breakpointed, which would
 876     // then require an immediate deoptimization.
 877     can_prune_locals = false;  // do not prune locals
 878     stack_slots_not_pruned = 0;
 879   }
 880 
 881   // do not scribble on the input jvms
 882   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 883   call->set_jvms(out_jvms); // Start jvms list for call node
 884 
 885   // For a known set of bytecodes, the interpreter should reexecute them if
 886   // deoptimization happens. We set the reexecute state for them here
 887   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 888       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 889     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 890   }
 891 
 892   // Presize the call:
 893   DEBUG_ONLY(uint non_debug_edges = call->req());
 894   call->add_req_batch(top(), youngest_jvms->debug_depth());
 895   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 896 
 897   // Set up edges so that the call looks like this:
 898   //  Call [state:] ctl io mem fptr retadr
 899   //       [parms:] parm0 ... parmN
 900   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 901   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 902   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 903   // Note that caller debug info precedes callee debug info.
 904 
 905   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 906   uint debug_ptr = call->req();
 907 
 908   // Loop over the map input edges associated with jvms, add them
 909   // to the call node, & reset all offsets to match call node array.
 910   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 911     uint debug_end   = debug_ptr;
 912     uint debug_start = debug_ptr - in_jvms->debug_size();
 913     debug_ptr = debug_start;  // back up the ptr
 914 
 915     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 916     uint j, k, l;
 917     SafePointNode* in_map = in_jvms->map();
 918     out_jvms->set_map(call);
 919 
 920     if (can_prune_locals) {
 921       assert(in_jvms->method() == out_jvms->method(), "sanity");
 922       // If the current throw can reach an exception handler in this JVMS,
 923       // then we must keep everything live that can reach that handler.
 924       // As a quick and dirty approximation, we look for any handlers at all.
 925       if (in_jvms->method()->has_exception_handlers()) {
 926         can_prune_locals = false;
 927       }
 928     }
 929 
 930     // Add the Locals
 931     k = in_jvms->locoff();
 932     l = in_jvms->loc_size();
 933     out_jvms->set_locoff(p);
 934     if (!can_prune_locals) {
 935       for (j = 0; j < l; j++)
 936         call->set_req(p++, in_map->in(k+j));
 937     } else {
 938       p += l;  // already set to top above by add_req_batch
 939     }
 940 
 941     // Add the Expression Stack
 942     k = in_jvms->stkoff();
 943     l = in_jvms->sp();
 944     out_jvms->set_stkoff(p);
 945     if (!can_prune_locals) {
 946       for (j = 0; j < l; j++)
 947         call->set_req(p++, in_map->in(k+j));
 948     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 949       // Divide stack into {S0,...,S1}, where S0 is set to top.
 950       uint s1 = stack_slots_not_pruned;
 951       stack_slots_not_pruned = 0;  // for next iteration
 952       if (s1 > l)  s1 = l;
 953       uint s0 = l - s1;
 954       p += s0;  // skip the tops preinstalled by add_req_batch
 955       for (j = s0; j < l; j++)
 956         call->set_req(p++, in_map->in(k+j));
 957     } else {
 958       p += l;  // already set to top above by add_req_batch
 959     }
 960 
 961     // Add the Monitors
 962     k = in_jvms->monoff();
 963     l = in_jvms->mon_size();
 964     out_jvms->set_monoff(p);
 965     for (j = 0; j < l; j++)
 966       call->set_req(p++, in_map->in(k+j));
 967 
 968     // Copy any scalar object fields.
 969     k = in_jvms->scloff();
 970     l = in_jvms->scl_size();
 971     out_jvms->set_scloff(p);
 972     for (j = 0; j < l; j++)
 973       call->set_req(p++, in_map->in(k+j));
 974 
 975     // Finish the new jvms.
 976     out_jvms->set_endoff(p);
 977 
 978     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 979     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 980     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 981     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 982     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 983     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 984 
 985     // Update the two tail pointers in parallel.
 986     out_jvms = out_jvms->caller();
 987     in_jvms  = in_jvms->caller();
 988   }
 989 
 990   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
 991 
 992   // Test the correctness of JVMState::debug_xxx accessors:
 993   assert(call->jvms()->debug_start() == non_debug_edges, "");
 994   assert(call->jvms()->debug_end()   == call->req(), "");
 995   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
 996 }
 997 
 998 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
 999   Bytecodes::Code code = java_bc();
1000   if (code == Bytecodes::_wide) {
1001     code = method()->java_code_at_bci(bci() + 1);
1002   }
1003 
1004   BasicType rtype = T_ILLEGAL;
1005   int       rsize = 0;
1006 
1007   if (code != Bytecodes::_illegal) {
1008     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1009     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1010     if (rtype < T_CONFLICT)
1011       rsize = type2size[rtype];
1012   }
1013 
1014   switch (code) {
1015   case Bytecodes::_illegal:
1016     return false;
1017 
1018   case Bytecodes::_ldc:
1019   case Bytecodes::_ldc_w:
1020   case Bytecodes::_ldc2_w:
1021     inputs = 0;
1022     break;
1023 
1024   case Bytecodes::_dup:         inputs = 1;  break;
1025   case Bytecodes::_dup_x1:      inputs = 2;  break;
1026   case Bytecodes::_dup_x2:      inputs = 3;  break;
1027   case Bytecodes::_dup2:        inputs = 2;  break;
1028   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1029   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1030   case Bytecodes::_swap:        inputs = 2;  break;
1031   case Bytecodes::_arraylength: inputs = 1;  break;
1032 
1033   case Bytecodes::_getstatic:
1034   case Bytecodes::_putstatic:
1035   case Bytecodes::_getfield:
1036   case Bytecodes::_vgetfield:
1037   case Bytecodes::_putfield:
1038     {
1039       bool ignored_will_link;
1040       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1041       int      size  = field->type()->size();
1042       bool is_get = (depth >= 0), is_static = (depth & 1);
1043       inputs = (is_static ? 0 : 1);
1044       if (is_get) {
1045         depth = size - inputs;
1046       } else {
1047         inputs += size;        // putxxx pops the value from the stack
1048         depth = - inputs;
1049       }
1050     }
1051     break;
1052 
1053   case Bytecodes::_invokevirtual:
1054   case Bytecodes::_invokespecial:
1055   case Bytecodes::_invokestatic:
1056   case Bytecodes::_invokedynamic:
1057   case Bytecodes::_invokeinterface:
1058     {
1059       bool ignored_will_link;
1060       ciSignature* declared_signature = NULL;
1061       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1062       assert(declared_signature != NULL, "cannot be null");
1063       inputs   = declared_signature->arg_size_for_bc(code);
1064       int size = declared_signature->return_type()->size();
1065       depth = size - inputs;
1066     }
1067     break;
1068 
1069   case Bytecodes::_multianewarray:
1070     {
1071       ciBytecodeStream iter(method());
1072       iter.reset_to_bci(bci());
1073       iter.next();
1074       inputs = iter.get_dimensions();
1075       assert(rsize == 1, "");
1076       depth = rsize - inputs;
1077     }
1078     break;
1079 
1080   case Bytecodes::_vwithfield: {
1081     bool ignored_will_link;
1082     ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1083     int      size  = field->type()->size();
1084     inputs = size+1;
1085     depth = rsize - inputs;
1086     break;
1087   }
1088 
1089   case Bytecodes::_ireturn:
1090   case Bytecodes::_lreturn:
1091   case Bytecodes::_freturn:
1092   case Bytecodes::_dreturn:
1093   case Bytecodes::_areturn:
1094   case Bytecodes::_vreturn:
1095     assert(rsize == -depth, "");
1096     inputs = rsize;
1097     break;
1098 
1099   case Bytecodes::_jsr:
1100   case Bytecodes::_jsr_w:
1101     inputs = 0;
1102     depth  = 1;                  // S.B. depth=1, not zero
1103     break;
1104 
1105   default:
1106     // bytecode produces a typed result
1107     inputs = rsize - depth;
1108     assert(inputs >= 0, "");
1109     break;
1110   }
1111 
1112 #ifdef ASSERT
1113   // spot check
1114   int outputs = depth + inputs;
1115   assert(outputs >= 0, "sanity");
1116   switch (code) {
1117   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1118   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1119   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1120   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1121   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1122   }
1123 #endif //ASSERT
1124 
1125   return true;
1126 }
1127 
1128 
1129 
1130 //------------------------------basic_plus_adr---------------------------------
1131 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1132   // short-circuit a common case
1133   if (offset == intcon(0))  return ptr;
1134   return _gvn.transform( new AddPNode(base, ptr, offset) );
1135 }
1136 
1137 Node* GraphKit::ConvI2L(Node* offset) {
1138   // short-circuit a common case
1139   jint offset_con = find_int_con(offset, Type::OffsetBot);
1140   if (offset_con != Type::OffsetBot) {
1141     return longcon((jlong) offset_con);
1142   }
1143   return _gvn.transform( new ConvI2LNode(offset));
1144 }
1145 
1146 Node* GraphKit::ConvI2UL(Node* offset) {
1147   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1148   if (offset_con != (juint) Type::OffsetBot) {
1149     return longcon((julong) offset_con);
1150   }
1151   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1152   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1153   return _gvn.transform( new AndLNode(conv, mask) );
1154 }
1155 
1156 Node* GraphKit::ConvL2I(Node* offset) {
1157   // short-circuit a common case
1158   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1159   if (offset_con != (jlong)Type::OffsetBot) {
1160     return intcon((int) offset_con);
1161   }
1162   return _gvn.transform( new ConvL2INode(offset));
1163 }
1164 
1165 //-------------------------load_object_klass-----------------------------------
1166 Node* GraphKit::load_object_klass(Node* obj) {
1167   // Special-case a fresh allocation to avoid building nodes:
1168   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1169   if (akls != NULL)  return akls;
1170   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1171   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1172 }
1173 
1174 //-------------------------load_array_length-----------------------------------
1175 Node* GraphKit::load_array_length(Node* array) {
1176   // Special-case a fresh allocation to avoid building nodes:
1177   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1178   Node *alen;
1179   if (alloc == NULL) {
1180     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1181     alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1182   } else {
1183     alen = alloc->Ideal_length();
1184     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1185     if (ccast != alen) {
1186       alen = _gvn.transform(ccast);
1187     }
1188   }
1189   return alen;
1190 }
1191 
1192 //------------------------------do_null_check----------------------------------
1193 // Helper function to do a NULL pointer check.  Returned value is
1194 // the incoming address with NULL casted away.  You are allowed to use the
1195 // not-null value only if you are control dependent on the test.
1196 #ifndef PRODUCT
1197 extern int explicit_null_checks_inserted,
1198            explicit_null_checks_elided;
1199 #endif
1200 Node* GraphKit::null_check_common(Node* value, BasicType type,
1201                                   // optional arguments for variations:
1202                                   bool assert_null,
1203                                   Node* *null_control,
1204                                   bool speculative) {
1205   assert(!assert_null || null_control == NULL, "not both at once");
1206   if (stopped())  return top();
1207   NOT_PRODUCT(explicit_null_checks_inserted++);
1208 
1209   // Construct NULL check
1210   Node *chk = NULL;
1211   switch(type) {
1212     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1213     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1214     case T_ARRAY  : // fall through
1215       type = T_OBJECT;  // simplify further tests
1216     case T_OBJECT : {
1217       const Type *t = _gvn.type( value );
1218 
1219       const TypeOopPtr* tp = t->isa_oopptr();
1220       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1221           // Only for do_null_check, not any of its siblings:
1222           && !assert_null && null_control == NULL) {
1223         // Usually, any field access or invocation on an unloaded oop type
1224         // will simply fail to link, since the statically linked class is
1225         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1226         // the static class is loaded but the sharper oop type is not.
1227         // Rather than checking for this obscure case in lots of places,
1228         // we simply observe that a null check on an unloaded class
1229         // will always be followed by a nonsense operation, so we
1230         // can just issue the uncommon trap here.
1231         // Our access to the unloaded class will only be correct
1232         // after it has been loaded and initialized, which requires
1233         // a trip through the interpreter.
1234 #ifndef PRODUCT
1235         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1236 #endif
1237         uncommon_trap(Deoptimization::Reason_unloaded,
1238                       Deoptimization::Action_reinterpret,
1239                       tp->klass(), "!loaded");
1240         return top();
1241       }
1242 
1243       if (assert_null) {
1244         // See if the type is contained in NULL_PTR.
1245         // If so, then the value is already null.
1246         if (t->higher_equal(TypePtr::NULL_PTR)) {
1247           NOT_PRODUCT(explicit_null_checks_elided++);
1248           return value;           // Elided null assert quickly!
1249         }
1250       } else {
1251         // See if mixing in the NULL pointer changes type.
1252         // If so, then the NULL pointer was not allowed in the original
1253         // type.  In other words, "value" was not-null.
1254         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1255           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1256           NOT_PRODUCT(explicit_null_checks_elided++);
1257           return value;           // Elided null check quickly!
1258         }
1259       }
1260       chk = new CmpPNode( value, null() );
1261       break;
1262     }
1263 
1264     default:
1265       fatal("unexpected type: %s", type2name(type));
1266   }
1267   assert(chk != NULL, "sanity check");
1268   chk = _gvn.transform(chk);
1269 
1270   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1271   BoolNode *btst = new BoolNode( chk, btest);
1272   Node   *tst = _gvn.transform( btst );
1273 
1274   //-----------
1275   // if peephole optimizations occurred, a prior test existed.
1276   // If a prior test existed, maybe it dominates as we can avoid this test.
1277   if (tst != btst && type == T_OBJECT) {
1278     // At this point we want to scan up the CFG to see if we can
1279     // find an identical test (and so avoid this test altogether).
1280     Node *cfg = control();
1281     int depth = 0;
1282     while( depth < 16 ) {       // Limit search depth for speed
1283       if( cfg->Opcode() == Op_IfTrue &&
1284           cfg->in(0)->in(1) == tst ) {
1285         // Found prior test.  Use "cast_not_null" to construct an identical
1286         // CastPP (and hence hash to) as already exists for the prior test.
1287         // Return that casted value.
1288         if (assert_null) {
1289           replace_in_map(value, null());
1290           return null();  // do not issue the redundant test
1291         }
1292         Node *oldcontrol = control();
1293         set_control(cfg);
1294         Node *res = cast_not_null(value);
1295         set_control(oldcontrol);
1296         NOT_PRODUCT(explicit_null_checks_elided++);
1297         return res;
1298       }
1299       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1300       if (cfg == NULL)  break;  // Quit at region nodes
1301       depth++;
1302     }
1303   }
1304 
1305   //-----------
1306   // Branch to failure if null
1307   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1308   Deoptimization::DeoptReason reason;
1309   if (assert_null) {
1310     reason = Deoptimization::Reason_null_assert;
1311   } else if (type == T_OBJECT) {
1312     reason = Deoptimization::reason_null_check(speculative);
1313   } else {
1314     reason = Deoptimization::Reason_div0_check;
1315   }
1316   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1317   // ciMethodData::has_trap_at will return a conservative -1 if any
1318   // must-be-null assertion has failed.  This could cause performance
1319   // problems for a method after its first do_null_assert failure.
1320   // Consider using 'Reason_class_check' instead?
1321 
1322   // To cause an implicit null check, we set the not-null probability
1323   // to the maximum (PROB_MAX).  For an explicit check the probability
1324   // is set to a smaller value.
1325   if (null_control != NULL || too_many_traps(reason)) {
1326     // probability is less likely
1327     ok_prob =  PROB_LIKELY_MAG(3);
1328   } else if (!assert_null &&
1329              (ImplicitNullCheckThreshold > 0) &&
1330              method() != NULL &&
1331              (method()->method_data()->trap_count(reason)
1332               >= (uint)ImplicitNullCheckThreshold)) {
1333     ok_prob =  PROB_LIKELY_MAG(3);
1334   }
1335 
1336   if (null_control != NULL) {
1337     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1338     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1339     set_control(      _gvn.transform( new IfTrueNode(iff)));
1340 #ifndef PRODUCT
1341     if (null_true == top()) {
1342       explicit_null_checks_elided++;
1343     }
1344 #endif
1345     (*null_control) = null_true;
1346   } else {
1347     BuildCutout unless(this, tst, ok_prob);
1348     // Check for optimizer eliding test at parse time
1349     if (stopped()) {
1350       // Failure not possible; do not bother making uncommon trap.
1351       NOT_PRODUCT(explicit_null_checks_elided++);
1352     } else if (assert_null) {
1353       uncommon_trap(reason,
1354                     Deoptimization::Action_make_not_entrant,
1355                     NULL, "assert_null");
1356     } else {
1357       replace_in_map(value, zerocon(type));
1358       builtin_throw(reason);
1359     }
1360   }
1361 
1362   // Must throw exception, fall-thru not possible?
1363   if (stopped()) {
1364     return top();               // No result
1365   }
1366 
1367   if (assert_null) {
1368     // Cast obj to null on this path.
1369     replace_in_map(value, zerocon(type));
1370     return zerocon(type);
1371   }
1372 
1373   // Cast obj to not-null on this path, if there is no null_control.
1374   // (If there is a null_control, a non-null value may come back to haunt us.)
1375   if (type == T_OBJECT) {
1376     Node* cast = cast_not_null(value, false);
1377     if (null_control == NULL || (*null_control) == top())
1378       replace_in_map(value, cast);
1379     value = cast;
1380   }
1381 
1382   return value;
1383 }
1384 
1385 
1386 //------------------------------cast_not_null----------------------------------
1387 // Cast obj to not-null on this path
1388 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1389   const Type *t = _gvn.type(obj);
1390   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1391   // Object is already not-null?
1392   if( t == t_not_null ) return obj;
1393 
1394   Node *cast = new CastPPNode(obj,t_not_null);
1395   cast->init_req(0, control());
1396   cast = _gvn.transform( cast );
1397 
1398   // Scan for instances of 'obj' in the current JVM mapping.
1399   // These instances are known to be not-null after the test.
1400   if (do_replace_in_map)
1401     replace_in_map(obj, cast);
1402 
1403   return cast;                  // Return casted value
1404 }
1405 
1406 
1407 //--------------------------replace_in_map-------------------------------------
1408 void GraphKit::replace_in_map(Node* old, Node* neww) {
1409   if (old == neww) {
1410     return;
1411   }
1412 
1413   map()->replace_edge(old, neww);
1414 
1415   // Note: This operation potentially replaces any edge
1416   // on the map.  This includes locals, stack, and monitors
1417   // of the current (innermost) JVM state.
1418 
1419   // don't let inconsistent types from profiling escape this
1420   // method
1421 
1422   const Type* told = _gvn.type(old);
1423   const Type* tnew = _gvn.type(neww);
1424 
1425   if (!tnew->higher_equal(told)) {
1426     return;
1427   }
1428 
1429   map()->record_replaced_node(old, neww);
1430 }
1431 
1432 
1433 //=============================================================================
1434 //--------------------------------memory---------------------------------------
1435 Node* GraphKit::memory(uint alias_idx) {
1436   MergeMemNode* mem = merged_memory();
1437   Node* p = mem->memory_at(alias_idx);
1438   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1439   return p;
1440 }
1441 
1442 //-----------------------------reset_memory------------------------------------
1443 Node* GraphKit::reset_memory() {
1444   Node* mem = map()->memory();
1445   // do not use this node for any more parsing!
1446   debug_only( map()->set_memory((Node*)NULL) );
1447   return _gvn.transform( mem );
1448 }
1449 
1450 //------------------------------set_all_memory---------------------------------
1451 void GraphKit::set_all_memory(Node* newmem) {
1452   Node* mergemem = MergeMemNode::make(newmem);
1453   gvn().set_type_bottom(mergemem);
1454   map()->set_memory(mergemem);
1455 }
1456 
1457 //------------------------------set_all_memory_call----------------------------
1458 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1459   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1460   set_all_memory(newmem);
1461 }
1462 
1463 //=============================================================================
1464 //
1465 // parser factory methods for MemNodes
1466 //
1467 // These are layered on top of the factory methods in LoadNode and StoreNode,
1468 // and integrate with the parser's memory state and _gvn engine.
1469 //
1470 
1471 // factory methods in "int adr_idx"
1472 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1473                           int adr_idx,
1474                           MemNode::MemOrd mo,
1475                           LoadNode::ControlDependency control_dependency,
1476                           bool require_atomic_access,
1477                           bool unaligned,
1478                           bool mismatched) {
1479   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1480   const TypePtr* adr_type = NULL; // debug-mode-only argument
1481   debug_only(adr_type = C->get_adr_type(adr_idx));
1482   Node* mem = memory(adr_idx);
1483   Node* ld;
1484   if (require_atomic_access && bt == T_LONG) {
1485     ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1486   } else if (require_atomic_access && bt == T_DOUBLE) {
1487     ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1488   } else {
1489     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1490   }
1491   ld = _gvn.transform(ld);
1492 
1493   if (bt == T_VALUETYPE) {
1494     // Load value type from oop
1495     ld = ValueTypeNode::make(gvn(), map()->memory(), ld);
1496   } else if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1497     // Improve graph before escape analysis and boxing elimination.
1498     record_for_igvn(ld);
1499   }
1500   return ld;
1501 }
1502 
1503 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1504                                 int adr_idx,
1505                                 MemNode::MemOrd mo,
1506                                 bool require_atomic_access,
1507                                 bool unaligned,
1508                                 bool mismatched) {
1509   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1510   const TypePtr* adr_type = NULL;
1511   debug_only(adr_type = C->get_adr_type(adr_idx));
1512   Node *mem = memory(adr_idx);
1513   Node* st;
1514   if (require_atomic_access && bt == T_LONG) {
1515     st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1516   } else if (require_atomic_access && bt == T_DOUBLE) {
1517     st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1518   } else {
1519     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1520   }
1521   if (unaligned) {
1522     st->as_Store()->set_unaligned_access();
1523   }
1524   if (mismatched) {
1525     st->as_Store()->set_mismatched_access();
1526   }
1527   st = _gvn.transform(st);
1528   set_memory(st, adr_idx);
1529   // Back-to-back stores can only remove intermediate store with DU info
1530   // so push on worklist for optimizer.
1531   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1532     record_for_igvn(st);
1533 
1534   return st;
1535 }
1536 
1537 
1538 void GraphKit::pre_barrier(bool do_load,
1539                            Node* ctl,
1540                            Node* obj,
1541                            Node* adr,
1542                            uint  adr_idx,
1543                            Node* val,
1544                            const TypeOopPtr* val_type,
1545                            Node* pre_val,
1546                            BasicType bt) {
1547 
1548   BarrierSet* bs = Universe::heap()->barrier_set();
1549   set_control(ctl);
1550   switch (bs->kind()) {
1551     case BarrierSet::G1SATBCTLogging:
1552       g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1553       break;
1554 
1555     case BarrierSet::CardTableForRS:
1556     case BarrierSet::CardTableExtension:
1557     case BarrierSet::ModRef:
1558       break;
1559 
1560     default      :
1561       ShouldNotReachHere();
1562 
1563   }
1564 }
1565 
1566 bool GraphKit::can_move_pre_barrier() const {
1567   BarrierSet* bs = Universe::heap()->barrier_set();
1568   switch (bs->kind()) {
1569     case BarrierSet::G1SATBCTLogging:
1570       return true; // Can move it if no safepoint
1571 
1572     case BarrierSet::CardTableForRS:
1573     case BarrierSet::CardTableExtension:
1574     case BarrierSet::ModRef:
1575       return true; // There is no pre-barrier
1576 
1577     default      :
1578       ShouldNotReachHere();
1579   }
1580   return false;
1581 }
1582 
1583 void GraphKit::post_barrier(Node* ctl,
1584                             Node* store,
1585                             Node* obj,
1586                             Node* adr,
1587                             uint  adr_idx,
1588                             Node* val,
1589                             BasicType bt,
1590                             bool use_precise) {
1591   BarrierSet* bs = Universe::heap()->barrier_set();
1592   set_control(ctl);
1593   switch (bs->kind()) {
1594     case BarrierSet::G1SATBCTLogging:
1595       g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1596       break;
1597 
1598     case BarrierSet::CardTableForRS:
1599     case BarrierSet::CardTableExtension:
1600       write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1601       break;
1602 
1603     case BarrierSet::ModRef:
1604       break;
1605 
1606     default      :
1607       ShouldNotReachHere();
1608 
1609   }
1610 }
1611 
1612 Node* GraphKit::store_oop(Node* ctl,
1613                           Node* obj,
1614                           Node* adr,
1615                           const TypePtr* adr_type,
1616                           Node* val,
1617                           const TypeOopPtr* val_type,
1618                           BasicType bt,
1619                           bool use_precise,
1620                           MemNode::MemOrd mo,
1621                           bool mismatched) {
1622   // Transformation of a value which could be NULL pointer (CastPP #NULL)
1623   // could be delayed during Parse (for example, in adjust_map_after_if()).
1624   // Execute transformation here to avoid barrier generation in such case.
1625   if (_gvn.type(val) == TypePtr::NULL_PTR)
1626     val = _gvn.makecon(TypePtr::NULL_PTR);
1627 
1628   set_control(ctl);
1629   if (stopped()) return top(); // Dead path ?
1630 
1631   assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity");
1632   assert(val != NULL, "not dead path");
1633   uint adr_idx = C->get_alias_index(adr_type);
1634   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1635 
1636   if (bt == T_VALUETYPE) {
1637     // Allocate value type and store oop
1638     val = val->as_ValueType()->store_to_memory(this);
1639   }
1640 
1641   pre_barrier(true /* do_load */,
1642               control(), obj, adr, adr_idx, val, val_type,
1643               NULL /* pre_val */,
1644               bt);
1645 
1646   Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1647   post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1648   return store;
1649 }
1650 
1651 // Could be an array or object we don't know at compile time (unsafe ref.)
1652 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1653                              Node* obj,   // containing obj
1654                              Node* adr,  // actual adress to store val at
1655                              const TypePtr* adr_type,
1656                              Node* val,
1657                              BasicType bt,
1658                              MemNode::MemOrd mo,
1659                              bool mismatched) {
1660   Compile::AliasType* at = C->alias_type(adr_type);
1661   const TypeOopPtr* val_type = NULL;
1662   if (adr_type->isa_instptr()) {
1663     if (at->field() != NULL) {
1664       // known field.  This code is a copy of the do_put_xxx logic.
1665       ciField* field = at->field();
1666       if (!field->type()->is_loaded()) {
1667         val_type = TypeInstPtr::BOTTOM;
1668       } else {
1669         val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1670       }
1671     }
1672   } else if (adr_type->isa_aryptr()) {
1673     val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1674   }
1675   if (val_type == NULL) {
1676     val_type = TypeInstPtr::BOTTOM;
1677   }
1678   return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1679 }
1680 
1681 
1682 //-------------------------array_element_address-------------------------
1683 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1684                                       const TypeInt* sizetype, Node* ctrl) {
1685   uint shift  = exact_log2(type2aelembytes(elembt));
1686   ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1687   if (arytype_klass->is_value_array_klass()) {
1688     ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1689     shift = vak->log2_element_size();
1690   }
1691   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1692 
1693   // short-circuit a common case (saves lots of confusing waste motion)
1694   jint idx_con = find_int_con(idx, -1);
1695   if (idx_con >= 0) {
1696     intptr_t offset = header + ((intptr_t)idx_con << shift);
1697     return basic_plus_adr(ary, offset);
1698   }
1699 
1700   // must be correct type for alignment purposes
1701   Node* base  = basic_plus_adr(ary, header);
1702   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1703   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1704   return basic_plus_adr(ary, base, scale);
1705 }
1706 
1707 //-------------------------load_array_element-------------------------
1708 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1709   const Type* elemtype = arytype->elem();
1710   BasicType elembt = elemtype->array_element_basic_type();
1711   assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1712   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1713   if (elembt == T_NARROWOOP) {
1714     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1715   }
1716   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1717   return ld;
1718 }
1719 
1720 //-------------------------set_arguments_for_java_call-------------------------
1721 // Arguments (pre-popped from the stack) are taken from the JVMS.
1722 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1723   // Add the call arguments:
1724   const TypeTuple* domain = call->tf()->domain_sig();
1725   uint nargs = domain->cnt();
1726   for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1727     Node* arg = argument(i-TypeFunc::Parms);
1728     if (ValueTypePassFieldsAsArgs) {
1729       if (arg->is_ValueType()) {
1730         ValueTypeNode* vt = arg->as_ValueType();
1731         if (domain->field_at(i)->is_valuetypeptr()->klass() != C->env()->___Value_klass()) {
1732           // We don't pass value type arguments by reference but instead
1733           // pass each field of the value type
1734           idx += vt->set_arguments_for_java_call(call, idx, *this);
1735         } else {
1736           arg = arg->as_ValueType()->store_to_memory(this);
1737           call->init_req(idx, arg);
1738           idx++;
1739         }
1740       } else {
1741         call->init_req(idx, arg);
1742         idx++;
1743       }
1744     } else {
1745       if (arg->is_ValueType()) {
1746         // Pass value type argument via oop to callee
1747         arg = arg->as_ValueType()->store_to_memory(this);
1748       }
1749       call->init_req(i, arg);
1750     }
1751   }
1752 }
1753 
1754 //---------------------------set_edges_for_java_call---------------------------
1755 // Connect a newly created call into the current JVMS.
1756 // A return value node (if any) is returned from set_edges_for_java_call.
1757 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1758 
1759   // Add the predefined inputs:
1760   call->init_req( TypeFunc::Control, control() );
1761   call->init_req( TypeFunc::I_O    , i_o() );
1762   call->init_req( TypeFunc::Memory , reset_memory() );
1763   call->init_req( TypeFunc::FramePtr, frameptr() );
1764   call->init_req( TypeFunc::ReturnAdr, top() );
1765 
1766   add_safepoint_edges(call, must_throw);
1767 
1768   Node* xcall = _gvn.transform(call);
1769 
1770   if (xcall == top()) {
1771     set_control(top());
1772     return;
1773   }
1774   assert(xcall == call, "call identity is stable");
1775 
1776   // Re-use the current map to produce the result.
1777 
1778   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1779   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1780   set_all_memory_call(xcall, separate_io_proj);
1781 
1782   //return xcall;   // no need, caller already has it
1783 }
1784 
1785 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1786   if (stopped())  return top();  // maybe the call folded up?
1787 
1788   // Capture the return value, if any.
1789   Node* ret;
1790   if (call->method() == NULL ||
1791       call->method()->return_type()->basic_type() == T_VOID)
1792         ret = top();
1793   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1794 
1795   // Note:  Since any out-of-line call can produce an exception,
1796   // we always insert an I_O projection from the call into the result.
1797 
1798   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1799 
1800   if (separate_io_proj) {
1801     // The caller requested separate projections be used by the fall
1802     // through and exceptional paths, so replace the projections for
1803     // the fall through path.
1804     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1805     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1806   }
1807   return ret;
1808 }
1809 
1810 //--------------------set_predefined_input_for_runtime_call--------------------
1811 // Reading and setting the memory state is way conservative here.
1812 // The real problem is that I am not doing real Type analysis on memory,
1813 // so I cannot distinguish card mark stores from other stores.  Across a GC
1814 // point the Store Barrier and the card mark memory has to agree.  I cannot
1815 // have a card mark store and its barrier split across the GC point from
1816 // either above or below.  Here I get that to happen by reading ALL of memory.
1817 // A better answer would be to separate out card marks from other memory.
1818 // For now, return the input memory state, so that it can be reused
1819 // after the call, if this call has restricted memory effects.
1820 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1821   // Set fixed predefined input arguments
1822   Node* memory = reset_memory();
1823   call->init_req( TypeFunc::Control,   control()  );
1824   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1825   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1826   call->init_req( TypeFunc::FramePtr,  frameptr() );
1827   call->init_req( TypeFunc::ReturnAdr, top()      );
1828   return memory;
1829 }
1830 
1831 //-------------------set_predefined_output_for_runtime_call--------------------
1832 // Set control and memory (not i_o) from the call.
1833 // If keep_mem is not NULL, use it for the output state,
1834 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1835 // If hook_mem is NULL, this call produces no memory effects at all.
1836 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1837 // then only that memory slice is taken from the call.
1838 // In the last case, we must put an appropriate memory barrier before
1839 // the call, so as to create the correct anti-dependencies on loads
1840 // preceding the call.
1841 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1842                                                       Node* keep_mem,
1843                                                       const TypePtr* hook_mem) {
1844   // no i/o
1845   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1846   if (keep_mem) {
1847     // First clone the existing memory state
1848     set_all_memory(keep_mem);
1849     if (hook_mem != NULL) {
1850       // Make memory for the call
1851       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1852       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1853       // We also use hook_mem to extract specific effects from arraycopy stubs.
1854       set_memory(mem, hook_mem);
1855     }
1856     // ...else the call has NO memory effects.
1857 
1858     // Make sure the call advertises its memory effects precisely.
1859     // This lets us build accurate anti-dependences in gcm.cpp.
1860     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1861            "call node must be constructed correctly");
1862   } else {
1863     assert(hook_mem == NULL, "");
1864     // This is not a "slow path" call; all memory comes from the call.
1865     set_all_memory_call(call);
1866   }
1867 }
1868 
1869 
1870 // Replace the call with the current state of the kit.
1871 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1872   JVMState* ejvms = NULL;
1873   if (has_exceptions()) {
1874     ejvms = transfer_exceptions_into_jvms();
1875   }
1876 
1877   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1878   ReplacedNodes replaced_nodes_exception;
1879   Node* ex_ctl = top();
1880 
1881   SafePointNode* final_state = stop();
1882 
1883   // Find all the needed outputs of this call
1884   CallProjections callprojs;
1885   call->extract_projections(&callprojs, true);
1886 
1887   Node* init_mem = call->in(TypeFunc::Memory);
1888   Node* final_mem = final_state->in(TypeFunc::Memory);
1889   Node* final_ctl = final_state->in(TypeFunc::Control);
1890   Node* final_io = final_state->in(TypeFunc::I_O);
1891 
1892   // Replace all the old call edges with the edges from the inlining result
1893   if (callprojs.fallthrough_catchproj != NULL) {
1894     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1895   }
1896   if (callprojs.fallthrough_memproj != NULL) {
1897     if (final_mem->is_MergeMem()) {
1898       // Parser's exits MergeMem was not transformed but may be optimized
1899       final_mem = _gvn.transform(final_mem);
1900     }
1901     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
1902   }
1903   if (callprojs.fallthrough_ioproj != NULL) {
1904     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
1905   }
1906 
1907   // Replace the result with the new result if it exists and is used
1908   if (callprojs.resproj != NULL && result != NULL) {
1909     C->gvn_replace_by(callprojs.resproj, result);
1910   }
1911 
1912   if (ejvms == NULL) {
1913     // No exception edges to simply kill off those paths
1914     if (callprojs.catchall_catchproj != NULL) {
1915       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1916     }
1917     if (callprojs.catchall_memproj != NULL) {
1918       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1919     }
1920     if (callprojs.catchall_ioproj != NULL) {
1921       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1922     }
1923     // Replace the old exception object with top
1924     if (callprojs.exobj != NULL) {
1925       C->gvn_replace_by(callprojs.exobj, C->top());
1926     }
1927   } else {
1928     GraphKit ekit(ejvms);
1929 
1930     // Load my combined exception state into the kit, with all phis transformed:
1931     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1932     replaced_nodes_exception = ex_map->replaced_nodes();
1933 
1934     Node* ex_oop = ekit.use_exception_state(ex_map);
1935 
1936     if (callprojs.catchall_catchproj != NULL) {
1937       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1938       ex_ctl = ekit.control();
1939     }
1940     if (callprojs.catchall_memproj != NULL) {
1941       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1942     }
1943     if (callprojs.catchall_ioproj != NULL) {
1944       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1945     }
1946 
1947     // Replace the old exception object with the newly created one
1948     if (callprojs.exobj != NULL) {
1949       C->gvn_replace_by(callprojs.exobj, ex_oop);
1950     }
1951   }
1952 
1953   // Disconnect the call from the graph
1954   call->disconnect_inputs(NULL, C);
1955   C->gvn_replace_by(call, C->top());
1956 
1957   // Clean up any MergeMems that feed other MergeMems since the
1958   // optimizer doesn't like that.
1959   if (final_mem->is_MergeMem()) {
1960     Node_List wl;
1961     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1962       Node* m = i.get();
1963       if (m->is_MergeMem() && !wl.contains(m)) {
1964         wl.push(m);
1965       }
1966     }
1967     while (wl.size()  > 0) {
1968       _gvn.transform(wl.pop());
1969     }
1970   }
1971 
1972   if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1973     replaced_nodes.apply(C, final_ctl);
1974   }
1975   if (!ex_ctl->is_top() && do_replaced_nodes) {
1976     replaced_nodes_exception.apply(C, ex_ctl);
1977   }
1978 }
1979 
1980 
1981 //------------------------------increment_counter------------------------------
1982 // for statistics: increment a VM counter by 1
1983 
1984 void GraphKit::increment_counter(address counter_addr) {
1985   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1986   increment_counter(adr1);
1987 }
1988 
1989 void GraphKit::increment_counter(Node* counter_addr) {
1990   int adr_type = Compile::AliasIdxRaw;
1991   Node* ctrl = control();
1992   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1993   Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1994   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1995 }
1996 
1997 
1998 //------------------------------uncommon_trap----------------------------------
1999 // Bail out to the interpreter in mid-method.  Implemented by calling the
2000 // uncommon_trap blob.  This helper function inserts a runtime call with the
2001 // right debug info.
2002 void GraphKit::uncommon_trap(int trap_request,
2003                              ciKlass* klass, const char* comment,
2004                              bool must_throw,
2005                              bool keep_exact_action) {
2006   if (failing())  stop();
2007   if (stopped())  return; // trap reachable?
2008 
2009   // Note:  If ProfileTraps is true, and if a deopt. actually
2010   // occurs here, the runtime will make sure an MDO exists.  There is
2011   // no need to call method()->ensure_method_data() at this point.
2012 
2013   // Set the stack pointer to the right value for reexecution:
2014   set_sp(reexecute_sp());
2015 
2016 #ifdef ASSERT
2017   if (!must_throw) {
2018     // Make sure the stack has at least enough depth to execute
2019     // the current bytecode.
2020     int inputs, ignored_depth;
2021     if (compute_stack_effects(inputs, ignored_depth)) {
2022       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2023              Bytecodes::name(java_bc()), sp(), inputs);
2024     }
2025   }
2026 #endif
2027 
2028   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2029   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2030 
2031   switch (action) {
2032   case Deoptimization::Action_maybe_recompile:
2033   case Deoptimization::Action_reinterpret:
2034     // Temporary fix for 6529811 to allow virtual calls to be sure they
2035     // get the chance to go from mono->bi->mega
2036     if (!keep_exact_action &&
2037         Deoptimization::trap_request_index(trap_request) < 0 &&
2038         too_many_recompiles(reason)) {
2039       // This BCI is causing too many recompilations.
2040       if (C->log() != NULL) {
2041         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2042                 Deoptimization::trap_reason_name(reason),
2043                 Deoptimization::trap_action_name(action));
2044       }
2045       action = Deoptimization::Action_none;
2046       trap_request = Deoptimization::make_trap_request(reason, action);
2047     } else {
2048       C->set_trap_can_recompile(true);
2049     }
2050     break;
2051   case Deoptimization::Action_make_not_entrant:
2052     C->set_trap_can_recompile(true);
2053     break;
2054 #ifdef ASSERT
2055   case Deoptimization::Action_none:
2056   case Deoptimization::Action_make_not_compilable:
2057     break;
2058   default:
2059     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2060     break;
2061 #endif
2062   }
2063 
2064   if (TraceOptoParse) {
2065     char buf[100];
2066     tty->print_cr("Uncommon trap %s at bci:%d",
2067                   Deoptimization::format_trap_request(buf, sizeof(buf),
2068                                                       trap_request), bci());
2069   }
2070 
2071   CompileLog* log = C->log();
2072   if (log != NULL) {
2073     int kid = (klass == NULL)? -1: log->identify(klass);
2074     log->begin_elem("uncommon_trap bci='%d'", bci());
2075     char buf[100];
2076     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2077                                                           trap_request));
2078     if (kid >= 0)         log->print(" klass='%d'", kid);
2079     if (comment != NULL)  log->print(" comment='%s'", comment);
2080     log->end_elem();
2081   }
2082 
2083   // Make sure any guarding test views this path as very unlikely
2084   Node *i0 = control()->in(0);
2085   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2086     IfNode *iff = i0->as_If();
2087     float f = iff->_prob;   // Get prob
2088     if (control()->Opcode() == Op_IfTrue) {
2089       if (f > PROB_UNLIKELY_MAG(4))
2090         iff->_prob = PROB_MIN;
2091     } else {
2092       if (f < PROB_LIKELY_MAG(4))
2093         iff->_prob = PROB_MAX;
2094     }
2095   }
2096 
2097   // Clear out dead values from the debug info.
2098   kill_dead_locals();
2099 
2100   // Now insert the uncommon trap subroutine call
2101   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2102   const TypePtr* no_memory_effects = NULL;
2103   // Pass the index of the class to be loaded
2104   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2105                                  (must_throw ? RC_MUST_THROW : 0),
2106                                  OptoRuntime::uncommon_trap_Type(),
2107                                  call_addr, "uncommon_trap", no_memory_effects,
2108                                  intcon(trap_request));
2109   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2110          "must extract request correctly from the graph");
2111   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2112 
2113   call->set_req(TypeFunc::ReturnAdr, returnadr());
2114   // The debug info is the only real input to this call.
2115 
2116   // Halt-and-catch fire here.  The above call should never return!
2117   HaltNode* halt = new HaltNode(control(), frameptr());
2118   _gvn.set_type_bottom(halt);
2119   root()->add_req(halt);
2120 
2121   stop_and_kill_map();
2122 }
2123 
2124 
2125 //--------------------------just_allocated_object------------------------------
2126 // Report the object that was just allocated.
2127 // It must be the case that there are no intervening safepoints.
2128 // We use this to determine if an object is so "fresh" that
2129 // it does not require card marks.
2130 Node* GraphKit::just_allocated_object(Node* current_control) {
2131   if (C->recent_alloc_ctl() == current_control)
2132     return C->recent_alloc_obj();
2133   return NULL;
2134 }
2135 
2136 
2137 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2138   // (Note:  TypeFunc::make has a cache that makes this fast.)
2139   const TypeFunc* tf    = TypeFunc::make(dest_method);
2140   int             nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2141   for (int j = 0; j < nargs; j++) {
2142     const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2143     if( targ->basic_type() == T_DOUBLE ) {
2144       // If any parameters are doubles, they must be rounded before
2145       // the call, dstore_rounding does gvn.transform
2146       Node *arg = argument(j);
2147       arg = dstore_rounding(arg);
2148       set_argument(j, arg);
2149     }
2150   }
2151 }
2152 
2153 /**
2154  * Record profiling data exact_kls for Node n with the type system so
2155  * that it can propagate it (speculation)
2156  *
2157  * @param n          node that the type applies to
2158  * @param exact_kls  type from profiling
2159  * @param maybe_null did profiling see null?
2160  *
2161  * @return           node with improved type
2162  */
2163 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) {
2164   const Type* current_type = _gvn.type(n);
2165   assert(UseTypeSpeculation, "type speculation must be on");
2166 
2167   const TypePtr* speculative = current_type->speculative();
2168 
2169   // Should the klass from the profile be recorded in the speculative type?
2170   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2171     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2172     const TypeOopPtr* xtype = tklass->as_instance_type();
2173     assert(xtype->klass_is_exact(), "Should be exact");
2174     // Any reason to believe n is not null (from this profiling or a previous one)?
2175     const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2176     // record the new speculative type's depth
2177     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2178     speculative = speculative->with_inline_depth(jvms()->depth());
2179   } else if (current_type->would_improve_ptr(maybe_null)) {
2180     // Profiling report that null was never seen so we can change the
2181     // speculative type to non null ptr.
2182     assert(!maybe_null, "nothing to improve");
2183     if (speculative == NULL) {
2184       speculative = TypePtr::NOTNULL;
2185     } else {
2186       const TypePtr* ptr = TypePtr::NOTNULL;
2187       speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2188     }
2189   }
2190 
2191   if (speculative != current_type->speculative()) {
2192     // Build a type with a speculative type (what we think we know
2193     // about the type but will need a guard when we use it)
2194     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2195     // We're changing the type, we need a new CheckCast node to carry
2196     // the new type. The new type depends on the control: what
2197     // profiling tells us is only valid from here as far as we can
2198     // tell.
2199     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2200     cast = _gvn.transform(cast);
2201     replace_in_map(n, cast);
2202     n = cast;
2203   }
2204 
2205   return n;
2206 }
2207 
2208 /**
2209  * Record profiling data from receiver profiling at an invoke with the
2210  * type system so that it can propagate it (speculation)
2211  *
2212  * @param n  receiver node
2213  *
2214  * @return   node with improved type
2215  */
2216 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2217   if (!UseTypeSpeculation) {
2218     return n;
2219   }
2220   ciKlass* exact_kls = profile_has_unique_klass();
2221   bool maybe_null = true;
2222   if (java_bc() == Bytecodes::_checkcast ||
2223       java_bc() == Bytecodes::_instanceof ||
2224       java_bc() == Bytecodes::_aastore) {
2225     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2226     maybe_null = data == NULL ? true : data->as_BitData()->null_seen();
2227   }
2228   return record_profile_for_speculation(n, exact_kls, maybe_null);
2229 }
2230 
2231 /**
2232  * Record profiling data from argument profiling at an invoke with the
2233  * type system so that it can propagate it (speculation)
2234  *
2235  * @param dest_method  target method for the call
2236  * @param bc           what invoke bytecode is this?
2237  */
2238 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2239   if (!UseTypeSpeculation) {
2240     return;
2241   }
2242   const TypeFunc* tf    = TypeFunc::make(dest_method);
2243   int             nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2244   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2245   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2246     const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2247     if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2248       bool maybe_null = true;
2249       ciKlass* better_type = NULL;
2250       if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) {
2251         record_profile_for_speculation(argument(j), better_type, maybe_null);
2252       }
2253       i++;
2254     }
2255   }
2256 }
2257 
2258 /**
2259  * Record profiling data from parameter profiling at an invoke with
2260  * the type system so that it can propagate it (speculation)
2261  */
2262 void GraphKit::record_profiled_parameters_for_speculation() {
2263   if (!UseTypeSpeculation) {
2264     return;
2265   }
2266   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2267     if (_gvn.type(local(i))->isa_oopptr()) {
2268       bool maybe_null = true;
2269       ciKlass* better_type = NULL;
2270       if (method()->parameter_profiled_type(j, better_type, maybe_null)) {
2271         record_profile_for_speculation(local(i), better_type, maybe_null);
2272       }
2273       j++;
2274     }
2275   }
2276 }
2277 
2278 /**
2279  * Record profiling data from return value profiling at an invoke with
2280  * the type system so that it can propagate it (speculation)
2281  */
2282 void GraphKit::record_profiled_return_for_speculation() {
2283   if (!UseTypeSpeculation) {
2284     return;
2285   }
2286   bool maybe_null = true;
2287   ciKlass* better_type = NULL;
2288   if (method()->return_profiled_type(bci(), better_type, maybe_null)) {
2289     // If profiling reports a single type for the return value,
2290     // feed it to the type system so it can propagate it as a
2291     // speculative type
2292     record_profile_for_speculation(stack(sp()-1), better_type, maybe_null);
2293   }
2294 }
2295 
2296 void GraphKit::round_double_result(ciMethod* dest_method) {
2297   // A non-strict method may return a double value which has an extended
2298   // exponent, but this must not be visible in a caller which is 'strict'
2299   // If a strict caller invokes a non-strict callee, round a double result
2300 
2301   BasicType result_type = dest_method->return_type()->basic_type();
2302   assert( method() != NULL, "must have caller context");
2303   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2304     // Destination method's return value is on top of stack
2305     // dstore_rounding() does gvn.transform
2306     Node *result = pop_pair();
2307     result = dstore_rounding(result);
2308     push_pair(result);
2309   }
2310 }
2311 
2312 // rounding for strict float precision conformance
2313 Node* GraphKit::precision_rounding(Node* n) {
2314   return UseStrictFP && _method->flags().is_strict()
2315     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2316     ? _gvn.transform( new RoundFloatNode(0, n) )
2317     : n;
2318 }
2319 
2320 // rounding for strict double precision conformance
2321 Node* GraphKit::dprecision_rounding(Node *n) {
2322   return UseStrictFP && _method->flags().is_strict()
2323     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2324     ? _gvn.transform( new RoundDoubleNode(0, n) )
2325     : n;
2326 }
2327 
2328 // rounding for non-strict double stores
2329 Node* GraphKit::dstore_rounding(Node* n) {
2330   return Matcher::strict_fp_requires_explicit_rounding
2331     && UseSSE <= 1
2332     ? _gvn.transform( new RoundDoubleNode(0, n) )
2333     : n;
2334 }
2335 
2336 //=============================================================================
2337 // Generate a fast path/slow path idiom.  Graph looks like:
2338 // [foo] indicates that 'foo' is a parameter
2339 //
2340 //              [in]     NULL
2341 //                 \    /
2342 //                  CmpP
2343 //                  Bool ne
2344 //                   If
2345 //                  /  \
2346 //              True    False-<2>
2347 //              / |
2348 //             /  cast_not_null
2349 //           Load  |    |   ^
2350 //        [fast_test]   |   |
2351 // gvn to   opt_test    |   |
2352 //          /    \      |  <1>
2353 //      True     False  |
2354 //        |         \\  |
2355 //   [slow_call]     \[fast_result]
2356 //    Ctl   Val       \      \
2357 //     |               \      \
2358 //    Catch       <1>   \      \
2359 //   /    \        ^     \      \
2360 //  Ex    No_Ex    |      \      \
2361 //  |       \   \  |       \ <2>  \
2362 //  ...      \  [slow_res] |  |    \   [null_result]
2363 //            \         \--+--+---  |  |
2364 //             \           | /    \ | /
2365 //              --------Region     Phi
2366 //
2367 //=============================================================================
2368 // Code is structured as a series of driver functions all called 'do_XXX' that
2369 // call a set of helper functions.  Helper functions first, then drivers.
2370 
2371 //------------------------------null_check_oop---------------------------------
2372 // Null check oop.  Set null-path control into Region in slot 3.
2373 // Make a cast-not-nullness use the other not-null control.  Return cast.
2374 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2375                                bool never_see_null,
2376                                bool safe_for_replace,
2377                                bool speculative) {
2378   // Initial NULL check taken path
2379   (*null_control) = top();
2380   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2381 
2382   // Generate uncommon_trap:
2383   if (never_see_null && (*null_control) != top()) {
2384     // If we see an unexpected null at a check-cast we record it and force a
2385     // recompile; the offending check-cast will be compiled to handle NULLs.
2386     // If we see more than one offending BCI, then all checkcasts in the
2387     // method will be compiled to handle NULLs.
2388     PreserveJVMState pjvms(this);
2389     set_control(*null_control);
2390     replace_in_map(value, null());
2391     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2392     uncommon_trap(reason,
2393                   Deoptimization::Action_make_not_entrant);
2394     (*null_control) = top();    // NULL path is dead
2395   }
2396   if ((*null_control) == top() && safe_for_replace) {
2397     replace_in_map(value, cast);
2398   }
2399 
2400   // Cast away null-ness on the result
2401   return cast;
2402 }
2403 
2404 //------------------------------opt_iff----------------------------------------
2405 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2406 // Return slow-path control.
2407 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2408   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2409 
2410   // Fast path taken; set region slot 2
2411   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2412   region->init_req(2,fast_taken); // Capture fast-control
2413 
2414   // Fast path not-taken, i.e. slow path
2415   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2416   return slow_taken;
2417 }
2418 
2419 //-----------------------------make_runtime_call-------------------------------
2420 Node* GraphKit::make_runtime_call(int flags,
2421                                   const TypeFunc* call_type, address call_addr,
2422                                   const char* call_name,
2423                                   const TypePtr* adr_type,
2424                                   // The following parms are all optional.
2425                                   // The first NULL ends the list.
2426                                   Node* parm0, Node* parm1,
2427                                   Node* parm2, Node* parm3,
2428                                   Node* parm4, Node* parm5,
2429                                   Node* parm6, Node* parm7) {
2430   // Slow-path call
2431   bool is_leaf = !(flags & RC_NO_LEAF);
2432   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2433   if (call_name == NULL) {
2434     assert(!is_leaf, "must supply name for leaf");
2435     call_name = OptoRuntime::stub_name(call_addr);
2436   }
2437   CallNode* call;
2438   if (!is_leaf) {
2439     call = new CallStaticJavaNode(call_type, call_addr, call_name,
2440                                            bci(), adr_type);
2441   } else if (flags & RC_NO_FP) {
2442     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2443   } else {
2444     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2445   }
2446 
2447   // The following is similar to set_edges_for_java_call,
2448   // except that the memory effects of the call are restricted to AliasIdxRaw.
2449 
2450   // Slow path call has no side-effects, uses few values
2451   bool wide_in  = !(flags & RC_NARROW_MEM);
2452   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2453 
2454   Node* prev_mem = NULL;
2455   if (wide_in) {
2456     prev_mem = set_predefined_input_for_runtime_call(call);
2457   } else {
2458     assert(!wide_out, "narrow in => narrow out");
2459     Node* narrow_mem = memory(adr_type);
2460     prev_mem = reset_memory();
2461     map()->set_memory(narrow_mem);
2462     set_predefined_input_for_runtime_call(call);
2463   }
2464 
2465   // Hook each parm in order.  Stop looking at the first NULL.
2466   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2467   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2468   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2469   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2470   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2471   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2472   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2473   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2474     /* close each nested if ===> */  } } } } } } } }
2475   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2476 
2477   if (!is_leaf) {
2478     // Non-leaves can block and take safepoints:
2479     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2480   }
2481   // Non-leaves can throw exceptions:
2482   if (has_io) {
2483     call->set_req(TypeFunc::I_O, i_o());
2484   }
2485 
2486   if (flags & RC_UNCOMMON) {
2487     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2488     // (An "if" probability corresponds roughly to an unconditional count.
2489     // Sort of.)
2490     call->set_cnt(PROB_UNLIKELY_MAG(4));
2491   }
2492 
2493   Node* c = _gvn.transform(call);
2494   assert(c == call, "cannot disappear");
2495 
2496   if (wide_out) {
2497     // Slow path call has full side-effects.
2498     set_predefined_output_for_runtime_call(call);
2499   } else {
2500     // Slow path call has few side-effects, and/or sets few values.
2501     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2502   }
2503 
2504   if (has_io) {
2505     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2506   }
2507   return call;
2508 
2509 }
2510 
2511 //------------------------------merge_memory-----------------------------------
2512 // Merge memory from one path into the current memory state.
2513 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2514   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2515     Node* old_slice = mms.force_memory();
2516     Node* new_slice = mms.memory2();
2517     if (old_slice != new_slice) {
2518       PhiNode* phi;
2519       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2520         if (mms.is_empty()) {
2521           // clone base memory Phi's inputs for this memory slice
2522           assert(old_slice == mms.base_memory(), "sanity");
2523           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2524           _gvn.set_type(phi, Type::MEMORY);
2525           for (uint i = 1; i < phi->req(); i++) {
2526             phi->init_req(i, old_slice->in(i));
2527           }
2528         } else {
2529           phi = old_slice->as_Phi(); // Phi was generated already
2530         }
2531       } else {
2532         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2533         _gvn.set_type(phi, Type::MEMORY);
2534       }
2535       phi->set_req(new_path, new_slice);
2536       mms.set_memory(phi);
2537     }
2538   }
2539 }
2540 
2541 //------------------------------make_slow_call_ex------------------------------
2542 // Make the exception handler hookups for the slow call
2543 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2544   if (stopped())  return;
2545 
2546   // Make a catch node with just two handlers:  fall-through and catch-all
2547   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2548   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2549   Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2550   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2551 
2552   { PreserveJVMState pjvms(this);
2553     set_control(excp);
2554     set_i_o(i_o);
2555 
2556     if (excp != top()) {
2557       if (deoptimize) {
2558         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2559         uncommon_trap(Deoptimization::Reason_unhandled,
2560                       Deoptimization::Action_none);
2561       } else {
2562         // Create an exception state also.
2563         // Use an exact type if the caller has specified a specific exception.
2564         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2565         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2566         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2567       }
2568     }
2569   }
2570 
2571   // Get the no-exception control from the CatchNode.
2572   set_control(norm);
2573 }
2574 
2575 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2576   Node* cmp = NULL;
2577   switch(bt) {
2578   case T_INT: cmp = new CmpINode(in1, in2); break;
2579   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2580   default: fatal("unexpected comparison type %s", type2name(bt));
2581   }
2582   gvn->transform(cmp);
2583   Node* bol = gvn->transform(new BoolNode(cmp, test));
2584   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2585   gvn->transform(iff);
2586   if (!bol->is_Con()) gvn->record_for_igvn(iff);
2587   return iff;
2588 }
2589 
2590 
2591 //-------------------------------gen_subtype_check-----------------------------
2592 // Generate a subtyping check.  Takes as input the subtype and supertype.
2593 // Returns 2 values: sets the default control() to the true path and returns
2594 // the false path.  Only reads invariant memory; sets no (visible) memory.
2595 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2596 // but that's not exposed to the optimizer.  This call also doesn't take in an
2597 // Object; if you wish to check an Object you need to load the Object's class
2598 // prior to coming here.
2599 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2600   Compile* C = gvn->C;
2601 
2602   if ((*ctrl)->is_top()) {
2603     return C->top();
2604   }
2605 
2606   // Fast check for identical types, perhaps identical constants.
2607   // The types can even be identical non-constants, in cases
2608   // involving Array.newInstance, Object.clone, etc.
2609   if (subklass == superklass)
2610     return C->top();             // false path is dead; no test needed.
2611 
2612   if (gvn->type(superklass)->singleton()) {
2613     ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2614     ciKlass* subk   = gvn->type(subklass)->is_klassptr()->klass();
2615 
2616     // In the common case of an exact superklass, try to fold up the
2617     // test before generating code.  You may ask, why not just generate
2618     // the code and then let it fold up?  The answer is that the generated
2619     // code will necessarily include null checks, which do not always
2620     // completely fold away.  If they are also needless, then they turn
2621     // into a performance loss.  Example:
2622     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2623     // Here, the type of 'fa' is often exact, so the store check
2624     // of fa[1]=x will fold up, without testing the nullness of x.
2625     switch (C->static_subtype_check(superk, subk)) {
2626     case Compile::SSC_always_false:
2627       {
2628         Node* always_fail = *ctrl;
2629         *ctrl = gvn->C->top();
2630         return always_fail;
2631       }
2632     case Compile::SSC_always_true:
2633       return C->top();
2634     case Compile::SSC_easy_test:
2635       {
2636         // Just do a direct pointer compare and be done.
2637         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2638         *ctrl = gvn->transform(new IfTrueNode(iff));
2639         return gvn->transform(new IfFalseNode(iff));
2640       }
2641     case Compile::SSC_full_test:
2642       break;
2643     default:
2644       ShouldNotReachHere();
2645     }
2646   }
2647 
2648   // %%% Possible further optimization:  Even if the superklass is not exact,
2649   // if the subklass is the unique subtype of the superklass, the check
2650   // will always succeed.  We could leave a dependency behind to ensure this.
2651 
2652   // First load the super-klass's check-offset
2653   Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2654   Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2655   Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2656   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2657   bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2658 
2659   // Load from the sub-klass's super-class display list, or a 1-word cache of
2660   // the secondary superclass list, or a failing value with a sentinel offset
2661   // if the super-klass is an interface or exceptionally deep in the Java
2662   // hierarchy and we have to scan the secondary superclass list the hard way.
2663   // Worst-case type is a little odd: NULL is allowed as a result (usually
2664   // klass loads can never produce a NULL).
2665   Node *chk_off_X = chk_off;
2666 #ifdef _LP64
2667   chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2668 #endif
2669   Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2670   // For some types like interfaces the following loadKlass is from a 1-word
2671   // cache which is mutable so can't use immutable memory.  Other
2672   // types load from the super-class display table which is immutable.
2673   m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2674   Node *kmem = might_be_cache ? m : C->immutable_memory();
2675   Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2676 
2677   // Compile speed common case: ARE a subtype and we canNOT fail
2678   if( superklass == nkls )
2679     return C->top();             // false path is dead; no test needed.
2680 
2681   // See if we get an immediate positive hit.  Happens roughly 83% of the
2682   // time.  Test to see if the value loaded just previously from the subklass
2683   // is exactly the superklass.
2684   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2685   Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2686   *ctrl = gvn->transform(new IfFalseNode(iff1));
2687 
2688   // Compile speed common case: Check for being deterministic right now.  If
2689   // chk_off is a constant and not equal to cacheoff then we are NOT a
2690   // subklass.  In this case we need exactly the 1 test above and we can
2691   // return those results immediately.
2692   if (!might_be_cache) {
2693     Node* not_subtype_ctrl = *ctrl;
2694     *ctrl = iftrue1; // We need exactly the 1 test above
2695     return not_subtype_ctrl;
2696   }
2697 
2698   // Gather the various success & failures here
2699   RegionNode *r_ok_subtype = new RegionNode(4);
2700   gvn->record_for_igvn(r_ok_subtype);
2701   RegionNode *r_not_subtype = new RegionNode(3);
2702   gvn->record_for_igvn(r_not_subtype);
2703 
2704   r_ok_subtype->init_req(1, iftrue1);
2705 
2706   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2707   // is roughly 63% of the remaining cases).  Test to see if the loaded
2708   // check-offset points into the subklass display list or the 1-element
2709   // cache.  If it points to the display (and NOT the cache) and the display
2710   // missed then it's not a subtype.
2711   Node *cacheoff = gvn->intcon(cacheoff_con);
2712   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2713   r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2714   *ctrl = gvn->transform(new IfFalseNode(iff2));
2715 
2716   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2717   // No performance impact (too rare) but allows sharing of secondary arrays
2718   // which has some footprint reduction.
2719   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2720   r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2721   *ctrl = gvn->transform(new IfFalseNode(iff3));
2722 
2723   // -- Roads not taken here: --
2724   // We could also have chosen to perform the self-check at the beginning
2725   // of this code sequence, as the assembler does.  This would not pay off
2726   // the same way, since the optimizer, unlike the assembler, can perform
2727   // static type analysis to fold away many successful self-checks.
2728   // Non-foldable self checks work better here in second position, because
2729   // the initial primary superclass check subsumes a self-check for most
2730   // types.  An exception would be a secondary type like array-of-interface,
2731   // which does not appear in its own primary supertype display.
2732   // Finally, we could have chosen to move the self-check into the
2733   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2734   // dependent manner.  But it is worthwhile to have the check here,
2735   // where it can be perhaps be optimized.  The cost in code space is
2736   // small (register compare, branch).
2737 
2738   // Now do a linear scan of the secondary super-klass array.  Again, no real
2739   // performance impact (too rare) but it's gotta be done.
2740   // Since the code is rarely used, there is no penalty for moving it
2741   // out of line, and it can only improve I-cache density.
2742   // The decision to inline or out-of-line this final check is platform
2743   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2744   Node* psc = gvn->transform(
2745     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2746 
2747   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2748   r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2749   r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2750 
2751   // Return false path; set default control to true path.
2752   *ctrl = gvn->transform(r_ok_subtype);
2753   return gvn->transform(r_not_subtype);
2754 }
2755 
2756 // Profile-driven exact type check:
2757 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2758                                     float prob,
2759                                     Node* *casted_receiver) {
2760   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2761   Node* recv_klass = load_object_klass(receiver);
2762   Node* want_klass = makecon(tklass);
2763   Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2764   Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2765   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2766   set_control( _gvn.transform( new IfTrueNode (iff) ));
2767   Node* fail = _gvn.transform( new IfFalseNode(iff) );
2768 
2769   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2770   assert(recv_xtype->klass_is_exact(), "");
2771 
2772   // Subsume downstream occurrences of receiver with a cast to
2773   // recv_xtype, since now we know what the type will be.
2774   Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2775   (*casted_receiver) = _gvn.transform(cast);
2776   // (User must make the replace_in_map call.)
2777 
2778   return fail;
2779 }
2780 
2781 
2782 //------------------------------seems_never_null-------------------------------
2783 // Use null_seen information if it is available from the profile.
2784 // If we see an unexpected null at a type check we record it and force a
2785 // recompile; the offending check will be recompiled to handle NULLs.
2786 // If we see several offending BCIs, then all checks in the
2787 // method will be recompiled.
2788 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2789   speculating = !_gvn.type(obj)->speculative_maybe_null();
2790   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2791   if (UncommonNullCast               // Cutout for this technique
2792       && obj != null()               // And not the -Xcomp stupid case?
2793       && !too_many_traps(reason)
2794       ) {
2795     if (speculating) {
2796       return true;
2797     }
2798     if (data == NULL)
2799       // Edge case:  no mature data.  Be optimistic here.
2800       return true;
2801     // If the profile has not seen a null, assume it won't happen.
2802     assert(java_bc() == Bytecodes::_checkcast ||
2803            java_bc() == Bytecodes::_instanceof ||
2804            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2805     return !data->as_BitData()->null_seen();
2806   }
2807   speculating = false;
2808   return false;
2809 }
2810 
2811 //------------------------maybe_cast_profiled_receiver-------------------------
2812 // If the profile has seen exactly one type, narrow to exactly that type.
2813 // Subsequent type checks will always fold up.
2814 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2815                                              ciKlass* require_klass,
2816                                              ciKlass* spec_klass,
2817                                              bool safe_for_replace) {
2818   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2819 
2820   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2821 
2822   // Make sure we haven't already deoptimized from this tactic.
2823   if (too_many_traps(reason) || too_many_recompiles(reason))
2824     return NULL;
2825 
2826   // (No, this isn't a call, but it's enough like a virtual call
2827   // to use the same ciMethod accessor to get the profile info...)
2828   // If we have a speculative type use it instead of profiling (which
2829   // may not help us)
2830   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2831   if (exact_kls != NULL) {// no cast failures here
2832     if (require_klass == NULL ||
2833         C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2834       // If we narrow the type to match what the type profile sees or
2835       // the speculative type, we can then remove the rest of the
2836       // cast.
2837       // This is a win, even if the exact_kls is very specific,
2838       // because downstream operations, such as method calls,
2839       // will often benefit from the sharper type.
2840       Node* exact_obj = not_null_obj; // will get updated in place...
2841       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2842                                             &exact_obj);
2843       { PreserveJVMState pjvms(this);
2844         set_control(slow_ctl);
2845         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2846       }
2847       if (safe_for_replace) {
2848         replace_in_map(not_null_obj, exact_obj);
2849       }
2850       return exact_obj;
2851     }
2852     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2853   }
2854 
2855   return NULL;
2856 }
2857 
2858 /**
2859  * Cast obj to type and emit guard unless we had too many traps here
2860  * already
2861  *
2862  * @param obj       node being casted
2863  * @param type      type to cast the node to
2864  * @param not_null  true if we know node cannot be null
2865  */
2866 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2867                                         ciKlass* type,
2868                                         bool not_null) {
2869   if (stopped()) {
2870     return obj;
2871   }
2872 
2873   // type == NULL if profiling tells us this object is always null
2874   if (type != NULL) {
2875     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2876     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2877 
2878     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2879         !too_many_traps(class_reason) &&
2880         !too_many_recompiles(class_reason)) {
2881       Node* not_null_obj = NULL;
2882       // not_null is true if we know the object is not null and
2883       // there's no need for a null check
2884       if (!not_null) {
2885         Node* null_ctl = top();
2886         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2887         assert(null_ctl->is_top(), "no null control here");
2888       } else {
2889         not_null_obj = obj;
2890       }
2891 
2892       Node* exact_obj = not_null_obj;
2893       ciKlass* exact_kls = type;
2894       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2895                                             &exact_obj);
2896       {
2897         PreserveJVMState pjvms(this);
2898         set_control(slow_ctl);
2899         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2900       }
2901       replace_in_map(not_null_obj, exact_obj);
2902       obj = exact_obj;
2903     }
2904   } else {
2905     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2906         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2907       Node* exact_obj = null_assert(obj);
2908       replace_in_map(obj, exact_obj);
2909       obj = exact_obj;
2910     }
2911   }
2912   return obj;
2913 }
2914 
2915 //-------------------------------gen_instanceof--------------------------------
2916 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2917 // and the reflective instance-of call.
2918 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2919   kill_dead_locals();           // Benefit all the uncommon traps
2920   assert( !stopped(), "dead parse path should be checked in callers" );
2921   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2922          "must check for not-null not-dead klass in callers");
2923 
2924   // Make the merge point
2925   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2926   RegionNode* region = new RegionNode(PATH_LIMIT);
2927   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
2928   C->set_has_split_ifs(true); // Has chance for split-if optimization
2929 
2930   ciProfileData* data = NULL;
2931   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2932     data = method()->method_data()->bci_to_data(bci());
2933   }
2934   bool speculative_not_null = false;
2935   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2936                          && seems_never_null(obj, data, speculative_not_null));
2937 
2938   // Null check; get casted pointer; set region slot 3
2939   Node* null_ctl = top();
2940   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2941 
2942   // If not_null_obj is dead, only null-path is taken
2943   if (stopped()) {              // Doing instance-of on a NULL?
2944     set_control(null_ctl);
2945     return intcon(0);
2946   }
2947   region->init_req(_null_path, null_ctl);
2948   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2949   if (null_ctl == top()) {
2950     // Do this eagerly, so that pattern matches like is_diamond_phi
2951     // will work even during parsing.
2952     assert(_null_path == PATH_LIMIT-1, "delete last");
2953     region->del_req(_null_path);
2954     phi   ->del_req(_null_path);
2955   }
2956 
2957   // Do we know the type check always succeed?
2958   bool known_statically = false;
2959   if (_gvn.type(superklass)->singleton()) {
2960     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2961     ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2962     if (subk != NULL && subk->is_loaded()) {
2963       int static_res = C->static_subtype_check(superk, subk);
2964       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2965     }
2966   }
2967 
2968   if (known_statically && UseTypeSpeculation) {
2969     // If we know the type check always succeeds then we don't use the
2970     // profiling data at this bytecode. Don't lose it, feed it to the
2971     // type system as a speculative type.
2972     not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2973   } else {
2974     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2975     // We may not have profiling here or it may not help us. If we
2976     // have a speculative type use it to perform an exact cast.
2977     ciKlass* spec_obj_type = obj_type->speculative_type();
2978     if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2979       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2980       if (stopped()) {            // Profile disagrees with this path.
2981         set_control(null_ctl);    // Null is the only remaining possibility.
2982         return intcon(0);
2983       }
2984       if (cast_obj != NULL) {
2985         not_null_obj = cast_obj;
2986       }
2987     }
2988   }
2989 
2990   // Load the object's klass
2991   Node* obj_klass = load_object_klass(not_null_obj);
2992 
2993   // Generate the subtype check
2994   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2995 
2996   // Plug in the success path to the general merge in slot 1.
2997   region->init_req(_obj_path, control());
2998   phi   ->init_req(_obj_path, intcon(1));
2999 
3000   // Plug in the failing path to the general merge in slot 2.
3001   region->init_req(_fail_path, not_subtype_ctrl);
3002   phi   ->init_req(_fail_path, intcon(0));
3003 
3004   // Return final merged results
3005   set_control( _gvn.transform(region) );
3006   record_for_igvn(region);
3007   return _gvn.transform(phi);
3008 }
3009 
3010 //-------------------------------gen_checkcast---------------------------------
3011 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
3012 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
3013 // uncommon-trap paths work.  Adjust stack after this call.
3014 // If failure_control is supplied and not null, it is filled in with
3015 // the control edge for the cast failure.  Otherwise, an appropriate
3016 // uncommon trap or exception is thrown.
3017 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3018                               Node* *failure_control) {
3019   kill_dead_locals();           // Benefit all the uncommon traps
3020   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3021   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3022 
3023   // Fast cutout:  Check the case that the cast is vacuously true.
3024   // This detects the common cases where the test will short-circuit
3025   // away completely.  We do this before we perform the null check,
3026   // because if the test is going to turn into zero code, we don't
3027   // want a residual null check left around.  (Causes a slowdown,
3028   // for example, in some objArray manipulations, such as a[i]=a[j].)
3029   if (tk->singleton()) {
3030     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3031     if (objtp != NULL && objtp->klass() != NULL) {
3032       switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3033       case Compile::SSC_always_true:
3034         // If we know the type check always succeed then we don't use
3035         // the profiling data at this bytecode. Don't lose it, feed it
3036         // to the type system as a speculative type.
3037         return record_profiled_receiver_for_speculation(obj);
3038       case Compile::SSC_always_false:
3039         // It needs a null check because a null will *pass* the cast check.
3040         // A non-null value will always produce an exception.
3041         return null_assert(obj);
3042       }
3043     }
3044   }
3045 
3046   ciProfileData* data = NULL;
3047   bool safe_for_replace = false;
3048   if (failure_control == NULL) {        // use MDO in regular case only
3049     assert(java_bc() == Bytecodes::_aastore ||
3050            java_bc() == Bytecodes::_checkcast,
3051            "interpreter profiles type checks only for these BCs");
3052     data = method()->method_data()->bci_to_data(bci());
3053     safe_for_replace = true;
3054   }
3055 
3056   // Make the merge point
3057   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3058   RegionNode* region = new RegionNode(PATH_LIMIT);
3059   Node*       phi    = new PhiNode(region, toop);
3060   C->set_has_split_ifs(true); // Has chance for split-if optimization
3061 
3062   // Use null-cast information if it is available
3063   bool speculative_not_null = false;
3064   bool never_see_null = ((failure_control == NULL)  // regular case only
3065                          && seems_never_null(obj, data, speculative_not_null));
3066 
3067   // Null check; get casted pointer; set region slot 3
3068   Node* null_ctl = top();
3069   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3070 
3071   // If not_null_obj is dead, only null-path is taken
3072   if (stopped()) {              // Doing instance-of on a NULL?
3073     set_control(null_ctl);
3074     return null();
3075   }
3076   region->init_req(_null_path, null_ctl);
3077   phi   ->init_req(_null_path, null());  // Set null path value
3078   if (null_ctl == top()) {
3079     // Do this eagerly, so that pattern matches like is_diamond_phi
3080     // will work even during parsing.
3081     assert(_null_path == PATH_LIMIT-1, "delete last");
3082     region->del_req(_null_path);
3083     phi   ->del_req(_null_path);
3084   }
3085 
3086   Node* cast_obj = NULL;
3087   if (tk->klass_is_exact()) {
3088     // The following optimization tries to statically cast the speculative type of the object
3089     // (for example obtained during profiling) to the type of the superklass and then do a
3090     // dynamic check that the type of the object is what we expect. To work correctly
3091     // for checkcast and aastore the type of superklass should be exact.
3092     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3093     // We may not have profiling here or it may not help us. If we have
3094     // a speculative type use it to perform an exact cast.
3095     ciKlass* spec_obj_type = obj_type->speculative_type();
3096     if (spec_obj_type != NULL || data != NULL) {
3097       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3098       if (cast_obj != NULL) {
3099         if (failure_control != NULL) // failure is now impossible
3100           (*failure_control) = top();
3101         // adjust the type of the phi to the exact klass:
3102         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3103       }
3104     }
3105   }
3106 
3107   if (cast_obj == NULL) {
3108     // Load the object's klass
3109     Node* obj_klass = load_object_klass(not_null_obj);
3110 
3111     // Generate the subtype check
3112     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3113 
3114     // Plug in success path into the merge
3115     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3116     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3117     if (failure_control == NULL) {
3118       if (not_subtype_ctrl != top()) { // If failure is possible
3119         PreserveJVMState pjvms(this);
3120         set_control(not_subtype_ctrl);
3121         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3122       }
3123     } else {
3124       (*failure_control) = not_subtype_ctrl;
3125     }
3126   }
3127 
3128   region->init_req(_obj_path, control());
3129   phi   ->init_req(_obj_path, cast_obj);
3130 
3131   // A merge of NULL or Casted-NotNull obj
3132   Node* res = _gvn.transform(phi);
3133 
3134   // Note I do NOT always 'replace_in_map(obj,result)' here.
3135   //  if( tk->klass()->can_be_primary_super()  )
3136     // This means that if I successfully store an Object into an array-of-String
3137     // I 'forget' that the Object is really now known to be a String.  I have to
3138     // do this because we don't have true union types for interfaces - if I store
3139     // a Baz into an array-of-Interface and then tell the optimizer it's an
3140     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3141     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3142   //  replace_in_map( obj, res );
3143 
3144   // Return final merged results
3145   set_control( _gvn.transform(region) );
3146   record_for_igvn(region);
3147   return res;
3148 }
3149 
3150 //------------------------------next_monitor-----------------------------------
3151 // What number should be given to the next monitor?
3152 int GraphKit::next_monitor() {
3153   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3154   int next = current + C->sync_stack_slots();
3155   // Keep the toplevel high water mark current:
3156   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3157   return current;
3158 }
3159 
3160 //------------------------------insert_mem_bar---------------------------------
3161 // Memory barrier to avoid floating things around
3162 // The membar serves as a pinch point between both control and all memory slices.
3163 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3164   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3165   mb->init_req(TypeFunc::Control, control());
3166   mb->init_req(TypeFunc::Memory,  reset_memory());
3167   Node* membar = _gvn.transform(mb);
3168   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3169   set_all_memory_call(membar);
3170   return membar;
3171 }
3172 
3173 //-------------------------insert_mem_bar_volatile----------------------------
3174 // Memory barrier to avoid floating things around
3175 // The membar serves as a pinch point between both control and memory(alias_idx).
3176 // If you want to make a pinch point on all memory slices, do not use this
3177 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3178 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3179   // When Parse::do_put_xxx updates a volatile field, it appends a series
3180   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3181   // The first membar is on the same memory slice as the field store opcode.
3182   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3183   // All the other membars (for other volatile slices, including AliasIdxBot,
3184   // which stands for all unknown volatile slices) are control-dependent
3185   // on the first membar.  This prevents later volatile loads or stores
3186   // from sliding up past the just-emitted store.
3187 
3188   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3189   mb->set_req(TypeFunc::Control,control());
3190   if (alias_idx == Compile::AliasIdxBot) {
3191     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3192   } else {
3193     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3194     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3195   }
3196   Node* membar = _gvn.transform(mb);
3197   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3198   if (alias_idx == Compile::AliasIdxBot) {
3199     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3200   } else {
3201     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3202   }
3203   return membar;
3204 }
3205 
3206 //------------------------------shared_lock------------------------------------
3207 // Emit locking code.
3208 FastLockNode* GraphKit::shared_lock(Node* obj) {
3209   // bci is either a monitorenter bc or InvocationEntryBci
3210   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3211   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3212 
3213   if( !GenerateSynchronizationCode )
3214     return NULL;                // Not locking things?
3215   if (stopped())                // Dead monitor?
3216     return NULL;
3217 
3218   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3219 
3220   // Box the stack location
3221   Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3222   Node* mem = reset_memory();
3223 
3224   FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3225   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3226     // Create the counters for this fast lock.
3227     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3228   }
3229 
3230   // Create the rtm counters for this fast lock if needed.
3231   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3232 
3233   // Add monitor to debug info for the slow path.  If we block inside the
3234   // slow path and de-opt, we need the monitor hanging around
3235   map()->push_monitor( flock );
3236 
3237   const TypeFunc *tf = LockNode::lock_type();
3238   LockNode *lock = new LockNode(C, tf);
3239 
3240   lock->init_req( TypeFunc::Control, control() );
3241   lock->init_req( TypeFunc::Memory , mem );
3242   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3243   lock->init_req( TypeFunc::FramePtr, frameptr() );
3244   lock->init_req( TypeFunc::ReturnAdr, top() );
3245 
3246   lock->init_req(TypeFunc::Parms + 0, obj);
3247   lock->init_req(TypeFunc::Parms + 1, box);
3248   lock->init_req(TypeFunc::Parms + 2, flock);
3249   add_safepoint_edges(lock);
3250 
3251   lock = _gvn.transform( lock )->as_Lock();
3252 
3253   // lock has no side-effects, sets few values
3254   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3255 
3256   insert_mem_bar(Op_MemBarAcquireLock);
3257 
3258   // Add this to the worklist so that the lock can be eliminated
3259   record_for_igvn(lock);
3260 
3261 #ifndef PRODUCT
3262   if (PrintLockStatistics) {
3263     // Update the counter for this lock.  Don't bother using an atomic
3264     // operation since we don't require absolute accuracy.
3265     lock->create_lock_counter(map()->jvms());
3266     increment_counter(lock->counter()->addr());
3267   }
3268 #endif
3269 
3270   return flock;
3271 }
3272 
3273 
3274 //------------------------------shared_unlock----------------------------------
3275 // Emit unlocking code.
3276 void GraphKit::shared_unlock(Node* box, Node* obj) {
3277   // bci is either a monitorenter bc or InvocationEntryBci
3278   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3279   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3280 
3281   if( !GenerateSynchronizationCode )
3282     return;
3283   if (stopped()) {               // Dead monitor?
3284     map()->pop_monitor();        // Kill monitor from debug info
3285     return;
3286   }
3287 
3288   // Memory barrier to avoid floating things down past the locked region
3289   insert_mem_bar(Op_MemBarReleaseLock);
3290 
3291   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3292   UnlockNode *unlock = new UnlockNode(C, tf);
3293 #ifdef ASSERT
3294   unlock->set_dbg_jvms(sync_jvms());
3295 #endif
3296   uint raw_idx = Compile::AliasIdxRaw;
3297   unlock->init_req( TypeFunc::Control, control() );
3298   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3299   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3300   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3301   unlock->init_req( TypeFunc::ReturnAdr, top() );
3302 
3303   unlock->init_req(TypeFunc::Parms + 0, obj);
3304   unlock->init_req(TypeFunc::Parms + 1, box);
3305   unlock = _gvn.transform(unlock)->as_Unlock();
3306 
3307   Node* mem = reset_memory();
3308 
3309   // unlock has no side-effects, sets few values
3310   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3311 
3312   // Kill monitor from debug info
3313   map()->pop_monitor( );
3314 }
3315 
3316 //-------------------------------get_layout_helper-----------------------------
3317 // If the given klass is a constant or known to be an array,
3318 // fetch the constant layout helper value into constant_value
3319 // and return (Node*)NULL.  Otherwise, load the non-constant
3320 // layout helper value, and return the node which represents it.
3321 // This two-faced routine is useful because allocation sites
3322 // almost always feature constant types.
3323 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3324   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3325   if (!StressReflectiveCode && inst_klass != NULL) {
3326     ciKlass* klass = inst_klass->klass();
3327     bool    xklass = inst_klass->klass_is_exact();
3328     if (xklass || klass->is_array_klass()) {
3329       jint lhelper = klass->layout_helper();
3330       if (lhelper != Klass::_lh_neutral_value) {
3331         constant_value = lhelper;
3332         return (Node*) NULL;
3333       }
3334     }
3335   }
3336   constant_value = Klass::_lh_neutral_value;  // put in a known value
3337   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3338   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3339 }
3340 
3341 // We just put in an allocate/initialize with a big raw-memory effect.
3342 // Hook selected additional alias categories on the initialization.
3343 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3344                                 MergeMemNode* init_in_merge,
3345                                 Node* init_out_raw) {
3346   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3347   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3348 
3349   Node* prevmem = kit.memory(alias_idx);
3350   init_in_merge->set_memory_at(alias_idx, prevmem);
3351   kit.set_memory(init_out_raw, alias_idx);
3352 }
3353 
3354 //---------------------------set_output_for_allocation-------------------------
3355 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3356                                           const TypeOopPtr* oop_type,
3357                                           bool deoptimize_on_exception) {
3358   int rawidx = Compile::AliasIdxRaw;
3359   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3360   add_safepoint_edges(alloc);
3361   Node* allocx = _gvn.transform(alloc);
3362   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3363   // create memory projection for i_o
3364   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3365   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3366 
3367   // create a memory projection as for the normal control path
3368   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3369   set_memory(malloc, rawidx);
3370 
3371   // a normal slow-call doesn't change i_o, but an allocation does
3372   // we create a separate i_o projection for the normal control path
3373   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3374   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3375 
3376   // put in an initialization barrier
3377   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3378                                                  rawoop)->as_Initialize();
3379   assert(alloc->initialization() == init,  "2-way macro link must work");
3380   assert(init ->allocation()     == alloc, "2-way macro link must work");
3381   {
3382     // Extract memory strands which may participate in the new object's
3383     // initialization, and source them from the new InitializeNode.
3384     // This will allow us to observe initializations when they occur,
3385     // and link them properly (as a group) to the InitializeNode.
3386     assert(init->in(InitializeNode::Memory) == malloc, "");
3387     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3388     init->set_req(InitializeNode::Memory, minit_in);
3389     record_for_igvn(minit_in); // fold it up later, if possible
3390     Node* minit_out = memory(rawidx);
3391     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3392     if (oop_type->isa_aryptr()) {
3393       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3394       int            elemidx  = C->get_alias_index(telemref);
3395       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3396     } else if (oop_type->isa_instptr() || oop_type->isa_valuetypeptr()) {
3397       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3398       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3399         ciField* field = ik->nonstatic_field_at(i);
3400         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3401           continue;  // do not bother to track really large numbers of fields
3402         // Find (or create) the alias category for this field:
3403         int fieldidx = C->alias_type(field)->index();
3404         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3405       }
3406     }
3407   }
3408 
3409   // Cast raw oop to the real thing...
3410   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3411   javaoop = _gvn.transform(javaoop);
3412   C->set_recent_alloc(control(), javaoop);
3413   assert(just_allocated_object(control()) == javaoop, "just allocated");
3414 
3415 #ifdef ASSERT
3416   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3417     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3418            "Ideal_allocation works");
3419     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3420            "Ideal_allocation works");
3421     if (alloc->is_AllocateArray()) {
3422       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3423              "Ideal_allocation works");
3424       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3425              "Ideal_allocation works");
3426     } else {
3427       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3428     }
3429   }
3430 #endif //ASSERT
3431 
3432   return javaoop;
3433 }
3434 
3435 //---------------------------new_instance--------------------------------------
3436 // This routine takes a klass_node which may be constant (for a static type)
3437 // or may be non-constant (for reflective code).  It will work equally well
3438 // for either, and the graph will fold nicely if the optimizer later reduces
3439 // the type to a constant.
3440 // The optional arguments are for specialized use by intrinsics:
3441 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3442 //  - If 'return_size_val', report the the total object size to the caller.
3443 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3444 Node* GraphKit::new_instance(Node* klass_node,
3445                              Node* extra_slow_test,
3446                              Node* *return_size_val,
3447                              bool deoptimize_on_exception) {
3448   // Compute size in doublewords
3449   // The size is always an integral number of doublewords, represented
3450   // as a positive bytewise size stored in the klass's layout_helper.
3451   // The layout_helper also encodes (in a low bit) the need for a slow path.
3452   jint  layout_con = Klass::_lh_neutral_value;
3453   Node* layout_val = get_layout_helper(klass_node, layout_con);
3454   int   layout_is_con = (layout_val == NULL);
3455 
3456   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3457   // Generate the initial go-slow test.  It's either ALWAYS (return a
3458   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3459   // case) a computed value derived from the layout_helper.
3460   Node* initial_slow_test = NULL;
3461   if (layout_is_con) {
3462     assert(!StressReflectiveCode, "stress mode does not use these paths");
3463     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3464     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3465   } else {   // reflective case
3466     // This reflective path is used by Unsafe.allocateInstance.
3467     // (It may be stress-tested by specifying StressReflectiveCode.)
3468     // Basically, we want to get into the VM is there's an illegal argument.
3469     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3470     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3471     if (extra_slow_test != intcon(0)) {
3472       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3473     }
3474     // (Macro-expander will further convert this to a Bool, if necessary.)
3475   }
3476 
3477   // Find the size in bytes.  This is easy; it's the layout_helper.
3478   // The size value must be valid even if the slow path is taken.
3479   Node* size = NULL;
3480   if (layout_is_con) {
3481     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3482   } else {   // reflective case
3483     // This reflective path is used by clone and Unsafe.allocateInstance.
3484     size = ConvI2X(layout_val);
3485 
3486     // Clear the low bits to extract layout_helper_size_in_bytes:
3487     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3488     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3489     size = _gvn.transform( new AndXNode(size, mask) );
3490   }
3491   if (return_size_val != NULL) {
3492     (*return_size_val) = size;
3493   }
3494 
3495   // This is a precise notnull oop of the klass.
3496   // (Actually, it need not be precise if this is a reflective allocation.)
3497   // It's what we cast the result to.
3498   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3499   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3500   const TypeOopPtr* oop_type = tklass->as_instance_type();
3501 
3502   // Now generate allocation code
3503 
3504   // The entire memory state is needed for slow path of the allocation
3505   // since GC and deoptimization can happen.
3506   Node *mem = reset_memory();
3507   set_all_memory(mem); // Create new memory state
3508 
3509   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3510                                          control(), mem, i_o(),
3511                                          size, klass_node,
3512                                          initial_slow_test);
3513 
3514   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3515 }
3516 
3517 //-------------------------------new_array-------------------------------------
3518 // helper for newarray, anewarray and vnewarray
3519 // The 'length' parameter is (obviously) the length of the array.
3520 // See comments on new_instance for the meaning of the other arguments.
3521 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3522                           Node* length,         // number of array elements
3523                           int   nargs,          // number of arguments to push back for uncommon trap
3524                           Node* *return_size_val,
3525                           bool deoptimize_on_exception) {
3526   jint  layout_con = Klass::_lh_neutral_value;
3527   Node* layout_val = get_layout_helper(klass_node, layout_con);
3528   int   layout_is_con = (layout_val == NULL);
3529 
3530   if (!layout_is_con && !StressReflectiveCode &&
3531       !too_many_traps(Deoptimization::Reason_class_check)) {
3532     // This is a reflective array creation site.
3533     // Optimistically assume that it is a subtype of Object[],
3534     // so that we can fold up all the address arithmetic.
3535     layout_con = Klass::array_layout_helper(T_OBJECT);
3536     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3537     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3538     { BuildCutout unless(this, bol_lh, PROB_MAX);
3539       inc_sp(nargs);
3540       uncommon_trap(Deoptimization::Reason_class_check,
3541                     Deoptimization::Action_maybe_recompile);
3542     }
3543     layout_val = NULL;
3544     layout_is_con = true;
3545   }
3546 
3547   // Generate the initial go-slow test.  Make sure we do not overflow
3548   // if length is huge (near 2Gig) or negative!  We do not need
3549   // exact double-words here, just a close approximation of needed
3550   // double-words.  We can't add any offset or rounding bits, lest we
3551   // take a size -1 of bytes and make it positive.  Use an unsigned
3552   // compare, so negative sizes look hugely positive.
3553   int fast_size_limit = FastAllocateSizeLimit;
3554   if (layout_is_con) {
3555     assert(!StressReflectiveCode, "stress mode does not use these paths");
3556     // Increase the size limit if we have exact knowledge of array type.
3557     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3558     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3559   }
3560 
3561   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3562   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3563 
3564   // --- Size Computation ---
3565   // array_size = round_to_heap(array_header + (length << elem_shift));
3566   // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3567   // and round_to(x, y) == ((x + y-1) & ~(y-1))
3568   // The rounding mask is strength-reduced, if possible.
3569   int round_mask = MinObjAlignmentInBytes - 1;
3570   Node* header_size = NULL;
3571   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3572   // (T_BYTE has the weakest alignment and size restrictions...)
3573   if (layout_is_con) {
3574     int       hsize  = Klass::layout_helper_header_size(layout_con);
3575     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3576     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3577     bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
3578     if ((round_mask & ~right_n_bits(eshift)) == 0)
3579       round_mask = 0;  // strength-reduce it if it goes away completely
3580     assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3581     assert(header_size_min <= hsize, "generic minimum is smallest");
3582     header_size_min = hsize;
3583     header_size = intcon(hsize + round_mask);
3584   } else {
3585     Node* hss   = intcon(Klass::_lh_header_size_shift);
3586     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3587     Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3588     hsize       = _gvn.transform( new AndINode(hsize, hsm) );
3589     Node* mask  = intcon(round_mask);
3590     header_size = _gvn.transform( new AddINode(hsize, mask) );
3591   }
3592 
3593   Node* elem_shift = NULL;
3594   if (layout_is_con) {
3595     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3596     if (eshift != 0)
3597       elem_shift = intcon(eshift);
3598   } else {
3599     // There is no need to mask or shift this value.
3600     // The semantics of LShiftINode include an implicit mask to 0x1F.
3601     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3602     elem_shift = layout_val;
3603   }
3604 
3605   // Transition to native address size for all offset calculations:
3606   Node* lengthx = ConvI2X(length);
3607   Node* headerx = ConvI2X(header_size);
3608 #ifdef _LP64
3609   { const TypeInt* tilen = _gvn.find_int_type(length);
3610     if (tilen != NULL && tilen->_lo < 0) {
3611       // Add a manual constraint to a positive range.  Cf. array_element_address.
3612       jint size_max = fast_size_limit;
3613       if (size_max > tilen->_hi)  size_max = tilen->_hi;
3614       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3615 
3616       // Only do a narrow I2L conversion if the range check passed.
3617       IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3618       _gvn.transform(iff);
3619       RegionNode* region = new RegionNode(3);
3620       _gvn.set_type(region, Type::CONTROL);
3621       lengthx = new PhiNode(region, TypeLong::LONG);
3622       _gvn.set_type(lengthx, TypeLong::LONG);
3623 
3624       // Range check passed. Use ConvI2L node with narrow type.
3625       Node* passed = IfFalse(iff);
3626       region->init_req(1, passed);
3627       // Make I2L conversion control dependent to prevent it from
3628       // floating above the range check during loop optimizations.
3629       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3630 
3631       // Range check failed. Use ConvI2L with wide type because length may be invalid.
3632       region->init_req(2, IfTrue(iff));
3633       lengthx->init_req(2, ConvI2X(length));
3634 
3635       set_control(region);
3636       record_for_igvn(region);
3637       record_for_igvn(lengthx);
3638     }
3639   }
3640 #endif
3641 
3642   // Combine header size (plus rounding) and body size.  Then round down.
3643   // This computation cannot overflow, because it is used only in two
3644   // places, one where the length is sharply limited, and the other
3645   // after a successful allocation.
3646   Node* abody = lengthx;
3647   if (elem_shift != NULL)
3648     abody     = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3649   Node* size  = _gvn.transform( new AddXNode(headerx, abody) );
3650   if (round_mask != 0) {
3651     Node* mask = MakeConX(~round_mask);
3652     size       = _gvn.transform( new AndXNode(size, mask) );
3653   }
3654   // else if round_mask == 0, the size computation is self-rounding
3655 
3656   if (return_size_val != NULL) {
3657     // This is the size
3658     (*return_size_val) = size;
3659   }
3660 
3661   // Now generate allocation code
3662 
3663   // The entire memory state is needed for slow path of the allocation
3664   // since GC and deoptimization can happen.
3665   Node *mem = reset_memory();
3666   set_all_memory(mem); // Create new memory state
3667 
3668   if (initial_slow_test->is_Bool()) {
3669     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3670     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3671   }
3672 
3673   // Create the AllocateArrayNode and its result projections
3674   AllocateArrayNode* alloc
3675     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3676                             control(), mem, i_o(),
3677                             size, klass_node,
3678                             initial_slow_test,
3679                             length);
3680 
3681   // Cast to correct type.  Note that the klass_node may be constant or not,
3682   // and in the latter case the actual array type will be inexact also.
3683   // (This happens via a non-constant argument to inline_native_newArray.)
3684   // In any case, the value of klass_node provides the desired array type.
3685   const TypeInt* length_type = _gvn.find_int_type(length);
3686   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3687   if (ary_type->isa_aryptr() && length_type != NULL) {
3688     // Try to get a better type than POS for the size
3689     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3690   }
3691 
3692   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3693 
3694   // Cast length on remaining path to be as narrow as possible
3695   if (map()->find_edge(length) >= 0) {
3696     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3697     if (ccast != length) {
3698       _gvn.set_type_bottom(ccast);
3699       record_for_igvn(ccast);
3700       replace_in_map(length, ccast);
3701     }
3702   }
3703 
3704   const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
3705   ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL;
3706   if (elem_klass != NULL && elem_klass->is_valuetype()) {
3707     ciValueKlass* vk = elem_klass->as_value_klass();
3708     if (!vk->flatten_array()) {
3709       // Non-flattened value type arrays need to be initialized with default value type oops
3710       initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs);
3711       InitializeNode* init = alloc->initialization();
3712       init->set_complete_with_arraycopy();
3713     }
3714   }
3715 
3716   return javaoop;
3717 }
3718 
3719 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) {
3720   // Check for zero length
3721   Node* null_ctl = top();
3722   null_check_common(length, T_INT, false, &null_ctl, false);
3723   if (stopped()) {
3724     set_control(null_ctl); // Always zero
3725     return;
3726   }
3727 
3728   // Prepare for merging control and IO
3729   RegionNode* res_ctl = new RegionNode(3);
3730   res_ctl->init_req(1, null_ctl);
3731   gvn().set_type(res_ctl, Type::CONTROL);
3732   record_for_igvn(res_ctl);
3733   Node* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO);
3734   gvn().set_type(res_io, Type::ABIO);
3735   record_for_igvn(res_io);
3736 
3737   // TODO comment
3738   SafePointNode* loop_map = NULL;
3739   {
3740     PreserveJVMState pjvms(this);
3741     // Create default value type and store it to memory
3742     Node* oop = ValueTypeNode::make_default(gvn(), vk);
3743     oop = oop->as_ValueType()->store_to_memory(this);
3744 
3745     length = SubI(length, intcon(1));
3746     add_predicate(nargs);
3747     RegionNode* loop = new RegionNode(3);
3748     loop->init_req(1, control());
3749     gvn().set_type(loop, Type::CONTROL);
3750     record_for_igvn(loop);
3751 
3752     Node* index = new PhiNode(loop, TypeInt::INT);
3753     index->init_req(1, intcon(0));
3754     gvn().set_type(index, TypeInt::INT);
3755     record_for_igvn(index);
3756 
3757     // TODO explain why we need to capture all memory
3758     PhiNode* mem = new PhiNode(loop, Type::MEMORY, TypePtr::BOTTOM);
3759     mem->init_req(1, reset_memory());
3760     gvn().set_type(mem, Type::MEMORY);
3761     record_for_igvn(mem);
3762     set_control(loop);
3763     set_all_memory(mem);
3764     // Initialize array element
3765     Node* adr = array_element_address(array, index, T_OBJECT);
3766     const TypeOopPtr* elemtype = TypeValueTypePtr::make(TypePtr::NotNull, vk);
3767     Node* store = store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_OBJECT, MemNode::release);
3768 
3769     IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
3770     loop->init_req(2, IfTrue(iff));
3771     mem->init_req(2, merged_memory());
3772     index->init_req(2, AddI(index, intcon(1)));
3773 
3774     res_ctl->init_req(2, IfFalse(iff));
3775     res_io->set_req(2, i_o());
3776     loop_map = stop();
3777   }
3778   // Set merged control, IO and memory
3779   set_control(res_ctl);
3780   set_i_o(res_io);
3781   merge_memory(loop_map->merged_memory(), res_ctl, 2);
3782 
3783   // Transform new memory Phis.
3784   for (MergeMemStream mms(merged_memory()); mms.next_non_empty();) {
3785     Node* phi = mms.memory();
3786     if (phi->is_Phi() && phi->in(0) == res_ctl) {
3787       mms.set_memory(gvn().transform(phi));
3788     }
3789   }
3790 }
3791 
3792 // The following "Ideal_foo" functions are placed here because they recognize
3793 // the graph shapes created by the functions immediately above.
3794 
3795 //---------------------------Ideal_allocation----------------------------------
3796 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3797 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3798   if (ptr == NULL) {     // reduce dumb test in callers
3799     return NULL;
3800   }
3801   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3802     ptr = ptr->in(1);
3803     if (ptr == NULL) return NULL;
3804   }
3805   // Return NULL for allocations with several casts:
3806   //   j.l.reflect.Array.newInstance(jobject, jint)
3807   //   Object.clone()
3808   // to keep more precise type from last cast.
3809   if (ptr->is_Proj()) {
3810     Node* allo = ptr->in(0);
3811     if (allo != NULL && allo->is_Allocate()) {
3812       return allo->as_Allocate();
3813     }
3814   }
3815   // Report failure to match.
3816   return NULL;
3817 }
3818 
3819 // Fancy version which also strips off an offset (and reports it to caller).
3820 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3821                                              intptr_t& offset) {
3822   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3823   if (base == NULL)  return NULL;
3824   return Ideal_allocation(base, phase);
3825 }
3826 
3827 // Trace Initialize <- Proj[Parm] <- Allocate
3828 AllocateNode* InitializeNode::allocation() {
3829   Node* rawoop = in(InitializeNode::RawAddress);
3830   if (rawoop->is_Proj()) {
3831     Node* alloc = rawoop->in(0);
3832     if (alloc->is_Allocate()) {
3833       return alloc->as_Allocate();
3834     }
3835   }
3836   return NULL;
3837 }
3838 
3839 // Trace Allocate -> Proj[Parm] -> Initialize
3840 InitializeNode* AllocateNode::initialization() {
3841   ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3842   if (rawoop == NULL)  return NULL;
3843   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3844     Node* init = rawoop->fast_out(i);
3845     if (init->is_Initialize()) {
3846       assert(init->as_Initialize()->allocation() == this, "2-way link");
3847       return init->as_Initialize();
3848     }
3849   }
3850   return NULL;
3851 }
3852 
3853 //----------------------------- loop predicates ---------------------------
3854 
3855 //------------------------------add_predicate_impl----------------------------
3856 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3857   // Too many traps seen?
3858   if (too_many_traps(reason)) {
3859 #ifdef ASSERT
3860     if (TraceLoopPredicate) {
3861       int tc = C->trap_count(reason);
3862       tty->print("too many traps=%s tcount=%d in ",
3863                     Deoptimization::trap_reason_name(reason), tc);
3864       method()->print(); // which method has too many predicate traps
3865       tty->cr();
3866     }
3867 #endif
3868     // We cannot afford to take more traps here,
3869     // do not generate predicate.
3870     return;
3871   }
3872 
3873   Node *cont    = _gvn.intcon(1);
3874   Node* opq     = _gvn.transform(new Opaque1Node(C, cont));
3875   Node *bol     = _gvn.transform(new Conv2BNode(opq));
3876   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3877   Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3878   C->add_predicate_opaq(opq);
3879   {
3880     PreserveJVMState pjvms(this);
3881     set_control(iffalse);
3882     inc_sp(nargs);
3883     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3884   }
3885   Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3886   set_control(iftrue);
3887 }
3888 
3889 //------------------------------add_predicate---------------------------------
3890 void GraphKit::add_predicate(int nargs) {
3891   if (UseLoopPredicate) {
3892     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3893   }
3894   // loop's limit check predicate should be near the loop.
3895   add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3896 }
3897 
3898 //----------------------------- store barriers ----------------------------
3899 #define __ ideal.
3900 
3901 void GraphKit::sync_kit(IdealKit& ideal) {
3902   set_all_memory(__ merged_memory());
3903   set_i_o(__ i_o());
3904   set_control(__ ctrl());
3905 }
3906 
3907 void GraphKit::final_sync(IdealKit& ideal) {
3908   // Final sync IdealKit and graphKit.
3909   sync_kit(ideal);
3910 }
3911 
3912 Node* GraphKit::byte_map_base_node() {
3913   // Get base of card map
3914   CardTableModRefBS* ct =
3915     barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3916   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
3917   if (ct->byte_map_base != NULL) {
3918     return makecon(TypeRawPtr::make((address)ct->byte_map_base));
3919   } else {
3920     return null();
3921   }
3922 }
3923 
3924 // vanilla/CMS post barrier
3925 // Insert a write-barrier store.  This is to let generational GC work; we have
3926 // to flag all oop-stores before the next GC point.
3927 void GraphKit::write_barrier_post(Node* oop_store,
3928                                   Node* obj,
3929                                   Node* adr,
3930                                   uint  adr_idx,
3931                                   Node* val,
3932                                   bool use_precise) {
3933   // No store check needed if we're storing a NULL or an old object
3934   // (latter case is probably a string constant). The concurrent
3935   // mark sweep garbage collector, however, needs to have all nonNull
3936   // oop updates flagged via card-marks.
3937   if (val != NULL && val->is_Con()) {
3938     // must be either an oop or NULL
3939     const Type* t = val->bottom_type();
3940     if (t == TypePtr::NULL_PTR || t == Type::TOP)
3941       // stores of null never (?) need barriers
3942       return;
3943   }
3944 
3945   if (use_ReduceInitialCardMarks()
3946       && obj == just_allocated_object(control())) {
3947     // We can skip marks on a freshly-allocated object in Eden.
3948     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3949     // That routine informs GC to take appropriate compensating steps,
3950     // upon a slow-path allocation, so as to make this card-mark
3951     // elision safe.
3952     return;
3953   }
3954 
3955   if (!use_precise) {
3956     // All card marks for a (non-array) instance are in one place:
3957     adr = obj;
3958   }
3959   // (Else it's an array (or unknown), and we want more precise card marks.)
3960   assert(adr != NULL, "");
3961 
3962   IdealKit ideal(this, true);
3963 
3964   // Convert the pointer to an int prior to doing math on it
3965   Node* cast = __ CastPX(__ ctrl(), adr);
3966 
3967   // Divide by card size
3968   assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3969          "Only one we handle so far.");
3970   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3971 
3972   // Combine card table base and card offset
3973   Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3974 
3975   // Get the alias_index for raw card-mark memory
3976   int adr_type = Compile::AliasIdxRaw;
3977   Node*   zero = __ ConI(0); // Dirty card value
3978   BasicType bt = T_BYTE;
3979 
3980   if (UseConcMarkSweepGC && UseCondCardMark) {
3981     insert_mem_bar(Op_MemBarVolatile);   // StoreLoad barrier
3982     __ sync_kit(this);
3983   }
3984 
3985   if (UseCondCardMark) {
3986     // The classic GC reference write barrier is typically implemented
3987     // as a store into the global card mark table.  Unfortunately
3988     // unconditional stores can result in false sharing and excessive
3989     // coherence traffic as well as false transactional aborts.
3990     // UseCondCardMark enables MP "polite" conditional card mark
3991     // stores.  In theory we could relax the load from ctrl() to
3992     // no_ctrl, but that doesn't buy much latitude.
3993     Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3994     __ if_then(card_val, BoolTest::ne, zero);
3995   }
3996 
3997   // Smash zero into card
3998   if( !UseConcMarkSweepGC ) {
3999     __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
4000   } else {
4001     // Specialized path for CM store barrier
4002     __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
4003   }
4004 
4005   if (UseCondCardMark) {
4006     __ end_if();
4007   }
4008 
4009   // Final sync IdealKit and GraphKit.
4010   final_sync(ideal);
4011 }
4012 /*
4013  * Determine if the G1 pre-barrier can be removed. The pre-barrier is
4014  * required by SATB to make sure all objects live at the start of the
4015  * marking are kept alive, all reference updates need to any previous
4016  * reference stored before writing.
4017  *
4018  * If the previous value is NULL there is no need to save the old value.
4019  * References that are NULL are filtered during runtime by the barrier
4020  * code to avoid unnecessary queuing.
4021  *
4022  * However in the case of newly allocated objects it might be possible to
4023  * prove that the reference about to be overwritten is NULL during compile
4024  * time and avoid adding the barrier code completely.
4025  *
4026  * The compiler needs to determine that the object in which a field is about
4027  * to be written is newly allocated, and that no prior store to the same field
4028  * has happened since the allocation.
4029  *
4030  * Returns true if the pre-barrier can be removed
4031  */
4032 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
4033                                          BasicType bt, uint adr_idx) {
4034   intptr_t offset = 0;
4035   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4036   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4037 
4038   if (offset == Type::OffsetBot) {
4039     return false; // cannot unalias unless there are precise offsets
4040   }
4041 
4042   if (alloc == NULL) {
4043     return false; // No allocation found
4044   }
4045 
4046   intptr_t size_in_bytes = type2aelembytes(bt);
4047 
4048   Node* mem = memory(adr_idx); // start searching here...
4049 
4050   for (int cnt = 0; cnt < 50; cnt++) {
4051 
4052     if (mem->is_Store()) {
4053 
4054       Node* st_adr = mem->in(MemNode::Address);
4055       intptr_t st_offset = 0;
4056       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4057 
4058       if (st_base == NULL) {
4059         break; // inscrutable pointer
4060       }
4061 
4062       // Break we have found a store with same base and offset as ours so break
4063       if (st_base == base && st_offset == offset) {
4064         break;
4065       }
4066 
4067       if (st_offset != offset && st_offset != Type::OffsetBot) {
4068         const int MAX_STORE = BytesPerLong;
4069         if (st_offset >= offset + size_in_bytes ||
4070             st_offset <= offset - MAX_STORE ||
4071             st_offset <= offset - mem->as_Store()->memory_size()) {
4072           // Success:  The offsets are provably independent.
4073           // (You may ask, why not just test st_offset != offset and be done?
4074           // The answer is that stores of different sizes can co-exist
4075           // in the same sequence of RawMem effects.  We sometimes initialize
4076           // a whole 'tile' of array elements with a single jint or jlong.)
4077           mem = mem->in(MemNode::Memory);
4078           continue; // advance through independent store memory
4079         }
4080       }
4081 
4082       if (st_base != base
4083           && MemNode::detect_ptr_independence(base, alloc, st_base,
4084                                               AllocateNode::Ideal_allocation(st_base, phase),
4085                                               phase)) {
4086         // Success:  The bases are provably independent.
4087         mem = mem->in(MemNode::Memory);
4088         continue; // advance through independent store memory
4089       }
4090     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4091 
4092       InitializeNode* st_init = mem->in(0)->as_Initialize();
4093       AllocateNode* st_alloc = st_init->allocation();
4094 
4095       // Make sure that we are looking at the same allocation site.
4096       // The alloc variable is guaranteed to not be null here from earlier check.
4097       if (alloc == st_alloc) {
4098         // Check that the initialization is storing NULL so that no previous store
4099         // has been moved up and directly write a reference
4100         Node* captured_store = st_init->find_captured_store(offset,
4101                                                             type2aelembytes(T_OBJECT),
4102                                                             phase);
4103         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4104           return true;
4105         }
4106       }
4107     }
4108 
4109     // Unless there is an explicit 'continue', we must bail out here,
4110     // because 'mem' is an inscrutable memory state (e.g., a call).
4111     break;
4112   }
4113 
4114   return false;
4115 }
4116 
4117 // G1 pre/post barriers
4118 void GraphKit::g1_write_barrier_pre(bool do_load,
4119                                     Node* obj,
4120                                     Node* adr,
4121                                     uint alias_idx,
4122                                     Node* val,
4123                                     const TypeOopPtr* val_type,
4124                                     Node* pre_val,
4125                                     BasicType bt) {
4126 
4127   // Some sanity checks
4128   // Note: val is unused in this routine.
4129 
4130   if (do_load) {
4131     // We need to generate the load of the previous value
4132     assert(obj != NULL, "must have a base");
4133     assert(adr != NULL, "where are loading from?");
4134     assert(pre_val == NULL, "loaded already?");
4135     assert(val_type != NULL, "need a type");
4136 
4137     if (use_ReduceInitialCardMarks()
4138         && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4139       return;
4140     }
4141 
4142   } else {
4143     // In this case both val_type and alias_idx are unused.
4144     assert(pre_val != NULL, "must be loaded already");
4145     // Nothing to be done if pre_val is null.
4146     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4147     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4148   }
4149   assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here");
4150 
4151   IdealKit ideal(this, true);
4152 
4153   Node* tls = __ thread(); // ThreadLocalStorage
4154 
4155   Node* no_ctrl = NULL;
4156   Node* no_base = __ top();
4157   Node* zero  = __ ConI(0);
4158   Node* zeroX = __ ConX(0);
4159 
4160   float likely  = PROB_LIKELY(0.999);
4161   float unlikely  = PROB_UNLIKELY(0.999);
4162 
4163   BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4164   assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4165 
4166   // Offsets into the thread
4167   const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +  // 648
4168                                           SATBMarkQueue::byte_offset_of_active());
4169   const int index_offset   = in_bytes(JavaThread::satb_mark_queue_offset() +  // 656
4170                                           SATBMarkQueue::byte_offset_of_index());
4171   const int buffer_offset  = in_bytes(JavaThread::satb_mark_queue_offset() +  // 652
4172                                           SATBMarkQueue::byte_offset_of_buf());
4173 
4174   // Now the actual pointers into the thread
4175   Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4176   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
4177   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
4178 
4179   // Now some of the values
4180   Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4181 
4182   // if (!marking)
4183   __ if_then(marking, BoolTest::ne, zero, unlikely); {
4184     BasicType index_bt = TypeX_X->basic_type();
4185     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4186     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4187 
4188     if (do_load) {
4189       // load original value
4190       // alias_idx correct??
4191       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4192     }
4193 
4194     // if (pre_val != NULL)
4195     __ if_then(pre_val, BoolTest::ne, null()); {
4196       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4197 
4198       // is the queue for this thread full?
4199       __ if_then(index, BoolTest::ne, zeroX, likely); {
4200 
4201         // decrement the index
4202         Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4203 
4204         // Now get the buffer location we will log the previous value into and store it
4205         Node *log_addr = __ AddP(no_base, buffer, next_index);
4206         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4207         // update the index
4208         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4209 
4210       } __ else_(); {
4211 
4212         // logging buffer is full, call the runtime
4213         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4214         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4215       } __ end_if();  // (!index)
4216     } __ end_if();  // (pre_val != NULL)
4217   } __ end_if();  // (!marking)
4218 
4219   // Final sync IdealKit and GraphKit.
4220   final_sync(ideal);
4221 }
4222 
4223 /*
4224  * G1 similar to any GC with a Young Generation requires a way to keep track of
4225  * references from Old Generation to Young Generation to make sure all live
4226  * objects are found. G1 also requires to keep track of object references
4227  * between different regions to enable evacuation of old regions, which is done
4228  * as part of mixed collections. References are tracked in remembered sets and
4229  * is continuously updated as reference are written to with the help of the
4230  * post-barrier.
4231  *
4232  * To reduce the number of updates to the remembered set the post-barrier
4233  * filters updates to fields in objects located in the Young Generation,
4234  * the same region as the reference, when the NULL is being written or
4235  * if the card is already marked as dirty by an earlier write.
4236  *
4237  * Under certain circumstances it is possible to avoid generating the
4238  * post-barrier completely if it is possible during compile time to prove
4239  * the object is newly allocated and that no safepoint exists between the
4240  * allocation and the store.
4241  *
4242  * In the case of slow allocation the allocation code must handle the barrier
4243  * as part of the allocation in the case the allocated object is not located
4244  * in the nursery, this would happen for humongous objects. This is similar to
4245  * how CMS is required to handle this case, see the comments for the method
4246  * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
4247  * A deferred card mark is required for these objects and handled in the above
4248  * mentioned methods.
4249  *
4250  * Returns true if the post barrier can be removed
4251  */
4252 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4253                                           Node* adr) {
4254   intptr_t      offset = 0;
4255   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4256   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
4257 
4258   if (offset == Type::OffsetBot) {
4259     return false; // cannot unalias unless there are precise offsets
4260   }
4261 
4262   if (alloc == NULL) {
4263      return false; // No allocation found
4264   }
4265 
4266   // Start search from Store node
4267   Node* mem = store->in(MemNode::Control);
4268   if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4269 
4270     InitializeNode* st_init = mem->in(0)->as_Initialize();
4271     AllocateNode*  st_alloc = st_init->allocation();
4272 
4273     // Make sure we are looking at the same allocation
4274     if (alloc == st_alloc) {
4275       return true;
4276     }
4277   }
4278 
4279   return false;
4280 }
4281 
4282 //
4283 // Update the card table and add card address to the queue
4284 //
4285 void GraphKit::g1_mark_card(IdealKit& ideal,
4286                             Node* card_adr,
4287                             Node* oop_store,
4288                             uint oop_alias_idx,
4289                             Node* index,
4290                             Node* index_adr,
4291                             Node* buffer,
4292                             const TypeFunc* tf) {
4293 
4294   Node* zero  = __ ConI(0);
4295   Node* zeroX = __ ConX(0);
4296   Node* no_base = __ top();
4297   BasicType card_bt = T_BYTE;
4298   // Smash zero into card. MUST BE ORDERED WRT TO STORE
4299   __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4300 
4301   //  Now do the queue work
4302   __ if_then(index, BoolTest::ne, zeroX); {
4303 
4304     Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4305     Node* log_addr = __ AddP(no_base, buffer, next_index);
4306 
4307     // Order, see storeCM.
4308     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4309     __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4310 
4311   } __ else_(); {
4312     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4313   } __ end_if();
4314 
4315 }
4316 
4317 void GraphKit::g1_write_barrier_post(Node* oop_store,
4318                                      Node* obj,
4319                                      Node* adr,
4320                                      uint alias_idx,
4321                                      Node* val,
4322                                      BasicType bt,
4323                                      bool use_precise) {
4324   // If we are writing a NULL then we need no post barrier
4325 
4326   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4327     // Must be NULL
4328     const Type* t = val->bottom_type();
4329     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4330     // No post barrier if writing NULLx
4331     return;
4332   }
4333 
4334   if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4335     // We can skip marks on a freshly-allocated object in Eden.
4336     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
4337     // That routine informs GC to take appropriate compensating steps,
4338     // upon a slow-path allocation, so as to make this card-mark
4339     // elision safe.
4340     return;
4341   }
4342 
4343   if (use_ReduceInitialCardMarks()
4344       && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4345     return;
4346   }
4347 
4348   if (!use_precise) {
4349     // All card marks for a (non-array) instance are in one place:
4350     adr = obj;
4351   }
4352   // (Else it's an array (or unknown), and we want more precise card marks.)
4353   assert(adr != NULL, "");
4354 
4355   IdealKit ideal(this, true);
4356 
4357   Node* tls = __ thread(); // ThreadLocalStorage
4358 
4359   Node* no_base = __ top();
4360   float likely  = PROB_LIKELY(0.999);
4361   float unlikely  = PROB_UNLIKELY(0.999);
4362   Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4363   Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4364   Node* zeroX = __ ConX(0);
4365 
4366   // Get the alias_index for raw card-mark memory
4367   const TypePtr* card_type = TypeRawPtr::BOTTOM;
4368 
4369   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4370 
4371   // Offsets into the thread
4372   const int index_offset  = in_bytes(JavaThread::dirty_card_queue_offset() +
4373                                      DirtyCardQueue::byte_offset_of_index());
4374   const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4375                                      DirtyCardQueue::byte_offset_of_buf());
4376 
4377   // Pointers into the thread
4378 
4379   Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4380   Node* index_adr =  __ AddP(no_base, tls, __ ConX(index_offset));
4381 
4382   // Now some values
4383   // Use ctrl to avoid hoisting these values past a safepoint, which could
4384   // potentially reset these fields in the JavaThread.
4385   Node* index  = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4386   Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4387 
4388   // Convert the store obj pointer to an int prior to doing math on it
4389   // Must use ctrl to prevent "integerized oop" existing across safepoint
4390   Node* cast =  __ CastPX(__ ctrl(), adr);
4391 
4392   // Divide pointer by card size
4393   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4394 
4395   // Combine card table base and card offset
4396   Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4397 
4398   // If we know the value being stored does it cross regions?
4399 
4400   if (val != NULL) {
4401     // Does the store cause us to cross regions?
4402 
4403     // Should be able to do an unsigned compare of region_size instead of
4404     // and extra shift. Do we have an unsigned compare??
4405     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4406     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4407 
4408     // if (xor_res == 0) same region so skip
4409     __ if_then(xor_res, BoolTest::ne, zeroX); {
4410 
4411       // No barrier if we are storing a NULL
4412       __ if_then(val, BoolTest::ne, null(), unlikely); {
4413 
4414         // Ok must mark the card if not already dirty
4415 
4416         // load the original value of the card
4417         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4418 
4419         __ if_then(card_val, BoolTest::ne, young_card); {
4420           sync_kit(ideal);
4421           // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4422           insert_mem_bar(Op_MemBarVolatile, oop_store);
4423           __ sync_kit(this);
4424 
4425           Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4426           __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4427             g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4428           } __ end_if();
4429         } __ end_if();
4430       } __ end_if();
4431     } __ end_if();
4432   } else {
4433     // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4434     // We don't need a barrier here if the destination is a newly allocated object
4435     // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4436     // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()).
4437     assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4438     Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4439     __ if_then(card_val, BoolTest::ne, young_card); {
4440       g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4441     } __ end_if();
4442   }
4443 
4444   // Final sync IdealKit and GraphKit.
4445   final_sync(ideal);
4446 }
4447 #undef __
4448 
4449 
4450 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4451   Node* len = load_array_length(load_String_value(ctrl, str));
4452   Node* coder = load_String_coder(ctrl, str);
4453   // Divide length by 2 if coder is UTF16
4454   return _gvn.transform(new RShiftINode(len, coder));
4455 }
4456 
4457 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4458   int value_offset = java_lang_String::value_offset_in_bytes();
4459   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4460                                                      false, NULL, Type::Offset(0));
4461   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4462   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4463                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4464                                                   ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4465   int value_field_idx = C->get_alias_index(value_field_type);
4466   Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4467                          value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4468   // String.value field is known to be @Stable.
4469   if (UseImplicitStableValues) {
4470     load = cast_array_to_stable(load, value_type);
4471   }
4472   return load;
4473 }
4474 
4475 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4476   if (!CompactStrings) {
4477     return intcon(java_lang_String::CODER_UTF16);
4478   }
4479   int coder_offset = java_lang_String::coder_offset_in_bytes();
4480   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4481                                                      false, NULL, Type::Offset(0));
4482   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4483   int coder_field_idx = C->get_alias_index(coder_field_type);
4484   return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4485                    TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4486 }
4487 
4488 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4489   int value_offset = java_lang_String::value_offset_in_bytes();
4490   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4491                                                      false, NULL, Type::Offset(0));
4492   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4493   store_oop_to_object(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
4494       value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4495 }
4496 
4497 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4498   int coder_offset = java_lang_String::coder_offset_in_bytes();
4499   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4500                                                      false, NULL, Type::Offset(0));
4501   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4502   int coder_field_idx = C->get_alias_index(coder_field_type);
4503   store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4504                   value, T_BYTE, coder_field_idx, MemNode::unordered);
4505 }
4506 
4507 // Capture src and dst memory state with a MergeMemNode
4508 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4509   if (src_type == dst_type) {
4510     // Types are equal, we don't need a MergeMemNode
4511     return memory(src_type);
4512   }
4513   MergeMemNode* merge = MergeMemNode::make(map()->memory());
4514   record_for_igvn(merge); // fold it up later, if possible
4515   int src_idx = C->get_alias_index(src_type);
4516   int dst_idx = C->get_alias_index(dst_type);
4517   merge->set_memory_at(src_idx, memory(src_idx));
4518   merge->set_memory_at(dst_idx, memory(dst_idx));
4519   return merge;
4520 }
4521 
4522 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4523   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4524   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4525   // If input and output memory types differ, capture both states to preserve
4526   // the dependency between preceding and subsequent loads/stores.
4527   // For example, the following program:
4528   //  StoreB
4529   //  compress_string
4530   //  LoadB
4531   // has this memory graph (use->def):
4532   //  LoadB -> compress_string -> CharMem
4533   //             ... -> StoreB -> ByteMem
4534   // The intrinsic hides the dependency between LoadB and StoreB, causing
4535   // the load to read from memory not containing the result of the StoreB.
4536   // The correct memory graph should look like this:
4537   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4538   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4539   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4540   Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4541   set_memory(res_mem, TypeAryPtr::BYTES);
4542   return str;
4543 }
4544 
4545 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4546   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4547   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4548   // Capture src and dst memory (see comment in 'compress_string').
4549   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4550   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4551   set_memory(_gvn.transform(str), dst_type);
4552 }
4553 
4554 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4555   /**
4556    * int i_char = start;
4557    * for (int i_byte = 0; i_byte < count; i_byte++) {
4558    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
4559    * }
4560    */
4561   add_predicate();
4562   RegionNode* head = new RegionNode(3);
4563   head->init_req(1, control());
4564   gvn().set_type(head, Type::CONTROL);
4565   record_for_igvn(head);
4566 
4567   Node* i_byte = new PhiNode(head, TypeInt::INT);
4568   i_byte->init_req(1, intcon(0));
4569   gvn().set_type(i_byte, TypeInt::INT);
4570   record_for_igvn(i_byte);
4571 
4572   Node* i_char = new PhiNode(head, TypeInt::INT);
4573   i_char->init_req(1, start);
4574   gvn().set_type(i_char, TypeInt::INT);
4575   record_for_igvn(i_char);
4576 
4577   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4578   gvn().set_type(mem, Type::MEMORY);
4579   record_for_igvn(mem);
4580   set_control(head);
4581   set_memory(mem, TypeAryPtr::BYTES);
4582   Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4583   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4584                              AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4585                              false, false, true /* mismatched */);
4586 
4587   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4588   head->init_req(2, IfTrue(iff));
4589   mem->init_req(2, st);
4590   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4591   i_char->init_req(2, AddI(i_char, intcon(2)));
4592 
4593   set_control(IfFalse(iff));
4594   set_memory(st, TypeAryPtr::BYTES);
4595 }
4596 
4597 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4598   if (!field->is_constant()) {
4599     return NULL; // Field not marked as constant.
4600   }
4601   ciInstance* holder = NULL;
4602   if (!field->is_static()) {
4603     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4604     if (const_oop != NULL && const_oop->is_instance()) {
4605       holder = const_oop->as_instance();
4606     }
4607   }
4608   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4609                                                         /*is_unsigned_load=*/false);
4610   if (con_type != NULL) {
4611     Node* con = makecon(con_type);
4612     if (field->layout_type() == T_VALUETYPE) {
4613       // Load value type from constant oop
4614       con = ValueTypeNode::make(gvn(), map()->memory(), con);
4615     }
4616     return con;
4617   }
4618   return NULL;
4619 }
4620 
4621 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4622   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4623   // assumption of CCP analysis.
4624   return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4625 }