1 /*
   2  * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "compiler/compileLog.hpp"
  27 #include "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 "opto/addnode.hpp"
  34 #include "opto/castnode.hpp"
  35 #include "opto/convertnode.hpp"
  36 #include "opto/graphKit.hpp"
  37 #include "opto/idealKit.hpp"
  38 #include "opto/intrinsicnode.hpp"
  39 #include "opto/locknode.hpp"
  40 #include "opto/machnode.hpp"
  41 #include "opto/opaquenode.hpp"
  42 #include "opto/parse.hpp"
  43 #include "opto/rootnode.hpp"
  44 #include "opto/runtime.hpp"
  45 #include "opto/valuetypenode.hpp"
  46 #include "runtime/deoptimization.hpp"
  47 #include "runtime/sharedRuntime.hpp"
  48 
  49 //----------------------------GraphKit-----------------------------------------
  50 // Main utility constructor.
  51 GraphKit::GraphKit(JVMState* jvms)
  52   : Phase(Phase::Parser),
  53     _env(C->env()),
  54     _gvn(*C->initial_gvn())
  55 {
  56   _exceptions = jvms->map()->next_exception();
  57   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  58   set_jvms(jvms);
  59 }
  60 
  61 // Private constructor for parser.
  62 GraphKit::GraphKit()
  63   : Phase(Phase::Parser),
  64     _env(C->env()),
  65     _gvn(*C->initial_gvn())
  66 {
  67   _exceptions = NULL;
  68   set_map(NULL);
  69   debug_only(_sp = -99);
  70   debug_only(set_bci(-99));
  71 }
  72 
  73 
  74 
  75 //---------------------------clean_stack---------------------------------------
  76 // Clear away rubbish from the stack area of the JVM state.
  77 // This destroys any arguments that may be waiting on the stack.
  78 void GraphKit::clean_stack(int from_sp) {
  79   SafePointNode* map      = this->map();
  80   JVMState*      jvms     = this->jvms();
  81   int            stk_size = jvms->stk_size();
  82   int            stkoff   = jvms->stkoff();
  83   Node*          top      = this->top();
  84   for (int i = from_sp; i < stk_size; i++) {
  85     if (map->in(stkoff + i) != top) {
  86       map->set_req(stkoff + i, top);
  87     }
  88   }
  89 }
  90 
  91 
  92 //--------------------------------sync_jvms-----------------------------------
  93 // Make sure our current jvms agrees with our parse state.
  94 JVMState* GraphKit::sync_jvms() const {
  95   JVMState* jvms = this->jvms();
  96   jvms->set_bci(bci());       // Record the new bci in the JVMState
  97   jvms->set_sp(sp());         // Record the new sp in the JVMState
  98   assert(jvms_in_sync(), "jvms is now in sync");
  99   return jvms;
 100 }
 101 
 102 //--------------------------------sync_jvms_for_reexecute---------------------
 103 // Make sure our current jvms agrees with our parse state.  This version
 104 // uses the reexecute_sp for reexecuting bytecodes.
 105 JVMState* GraphKit::sync_jvms_for_reexecute() {
 106   JVMState* jvms = this->jvms();
 107   jvms->set_bci(bci());          // Record the new bci in the JVMState
 108   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 109   return jvms;
 110 }
 111 
 112 #ifdef ASSERT
 113 bool GraphKit::jvms_in_sync() const {
 114   Parse* parse = is_Parse();
 115   if (parse == NULL) {
 116     if (bci() !=      jvms()->bci())          return false;
 117     if (sp()  != (int)jvms()->sp())           return false;
 118     return true;
 119   }
 120   if (jvms()->method() != parse->method())    return false;
 121   if (jvms()->bci()    != parse->bci())       return false;
 122   int jvms_sp = jvms()->sp();
 123   if (jvms_sp          != parse->sp())        return false;
 124   int jvms_depth = jvms()->depth();
 125   if (jvms_depth       != parse->depth())     return false;
 126   return true;
 127 }
 128 
 129 // Local helper checks for special internal merge points
 130 // used to accumulate and merge exception states.
 131 // They are marked by the region's in(0) edge being the map itself.
 132 // Such merge points must never "escape" into the parser at large,
 133 // until they have been handed to gvn.transform.
 134 static bool is_hidden_merge(Node* reg) {
 135   if (reg == NULL)  return false;
 136   if (reg->is_Phi()) {
 137     reg = reg->in(0);
 138     if (reg == NULL)  return false;
 139   }
 140   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 141 }
 142 
 143 void GraphKit::verify_map() const {
 144   if (map() == NULL)  return;  // null map is OK
 145   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 146   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 147   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 148 }
 149 
 150 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 151   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 152   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 153 }
 154 #endif
 155 
 156 //---------------------------stop_and_kill_map---------------------------------
 157 // Set _map to NULL, signalling a stop to further bytecode execution.
 158 // First smash the current map's control to a constant, to mark it dead.
 159 void GraphKit::stop_and_kill_map() {
 160   SafePointNode* dead_map = stop();
 161   if (dead_map != NULL) {
 162     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
 163     assert(dead_map->is_killed(), "must be so marked");
 164   }
 165 }
 166 
 167 
 168 //--------------------------------stopped--------------------------------------
 169 // Tell if _map is NULL, or control is top.
 170 bool GraphKit::stopped() {
 171   if (map() == NULL)           return true;
 172   else if (control() == top()) return true;
 173   else                         return false;
 174 }
 175 
 176 
 177 //-----------------------------has_ex_handler----------------------------------
 178 // Tell if this method or any caller method has exception handlers.
 179 bool GraphKit::has_ex_handler() {
 180   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 181     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 182       return true;
 183     }
 184   }
 185   return false;
 186 }
 187 
 188 //------------------------------save_ex_oop------------------------------------
 189 // Save an exception without blowing stack contents or other JVM state.
 190 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 191   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 192   ex_map->add_req(ex_oop);
 193   debug_only(verify_exception_state(ex_map));
 194 }
 195 
 196 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 197   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 198   Node* ex_oop = ex_map->in(ex_map->req()-1);
 199   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 200   return ex_oop;
 201 }
 202 
 203 //-----------------------------saved_ex_oop------------------------------------
 204 // Recover a saved exception from its map.
 205 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 206   return common_saved_ex_oop(ex_map, false);
 207 }
 208 
 209 //--------------------------clear_saved_ex_oop---------------------------------
 210 // Erase a previously saved exception from its map.
 211 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 212   return common_saved_ex_oop(ex_map, true);
 213 }
 214 
 215 #ifdef ASSERT
 216 //---------------------------has_saved_ex_oop----------------------------------
 217 // Erase a previously saved exception from its map.
 218 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 219   return ex_map->req() == ex_map->jvms()->endoff()+1;
 220 }
 221 #endif
 222 
 223 //-------------------------make_exception_state--------------------------------
 224 // Turn the current JVM state into an exception state, appending the ex_oop.
 225 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 226   sync_jvms();
 227   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 228   set_saved_ex_oop(ex_map, ex_oop);
 229   return ex_map;
 230 }
 231 
 232 
 233 //--------------------------add_exception_state--------------------------------
 234 // Add an exception to my list of exceptions.
 235 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 236   if (ex_map == NULL || ex_map->control() == top()) {
 237     return;
 238   }
 239 #ifdef ASSERT
 240   verify_exception_state(ex_map);
 241   if (has_exceptions()) {
 242     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 243   }
 244 #endif
 245 
 246   // If there is already an exception of exactly this type, merge with it.
 247   // In particular, null-checks and other low-level exceptions common up here.
 248   Node*       ex_oop  = saved_ex_oop(ex_map);
 249   const Type* ex_type = _gvn.type(ex_oop);
 250   if (ex_oop == top()) {
 251     // No action needed.
 252     return;
 253   }
 254   assert(ex_type->isa_instptr(), "exception must be an instance");
 255   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 256     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 257     // We check sp also because call bytecodes can generate exceptions
 258     // both before and after arguments are popped!
 259     if (ex_type2 == ex_type
 260         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 261       combine_exception_states(ex_map, e2);
 262       return;
 263     }
 264   }
 265 
 266   // No pre-existing exception of the same type.  Chain it on the list.
 267   push_exception_state(ex_map);
 268 }
 269 
 270 //-----------------------add_exception_states_from-----------------------------
 271 void GraphKit::add_exception_states_from(JVMState* jvms) {
 272   SafePointNode* ex_map = jvms->map()->next_exception();
 273   if (ex_map != NULL) {
 274     jvms->map()->set_next_exception(NULL);
 275     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 276       next_map = ex_map->next_exception();
 277       ex_map->set_next_exception(NULL);
 278       add_exception_state(ex_map);
 279     }
 280   }
 281 }
 282 
 283 //-----------------------transfer_exceptions_into_jvms-------------------------
 284 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 285   if (map() == NULL) {
 286     // We need a JVMS to carry the exceptions, but the map has gone away.
 287     // Create a scratch JVMS, cloned from any of the exception states...
 288     if (has_exceptions()) {
 289       _map = _exceptions;
 290       _map = clone_map();
 291       _map->set_next_exception(NULL);
 292       clear_saved_ex_oop(_map);
 293       debug_only(verify_map());
 294     } else {
 295       // ...or created from scratch
 296       JVMState* jvms = new (C) JVMState(_method, NULL);
 297       jvms->set_bci(_bci);
 298       jvms->set_sp(_sp);
 299       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
 300       set_jvms(jvms);
 301       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 302       set_all_memory(top());
 303       while (map()->req() < jvms->endoff())  map()->add_req(top());
 304     }
 305     // (This is a kludge, in case you didn't notice.)
 306     set_control(top());
 307   }
 308   JVMState* jvms = sync_jvms();
 309   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 310   jvms->map()->set_next_exception(_exceptions);
 311   _exceptions = NULL;   // done with this set of exceptions
 312   return jvms;
 313 }
 314 
 315 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 316   assert(is_hidden_merge(dstphi), "must be a special merge node");
 317   assert(is_hidden_merge(srcphi), "must be a special merge node");
 318   uint limit = srcphi->req();
 319   for (uint i = PhiNode::Input; i < limit; i++) {
 320     dstphi->add_req(srcphi->in(i));
 321   }
 322 }
 323 static inline void add_one_req(Node* dstphi, Node* src) {
 324   assert(is_hidden_merge(dstphi), "must be a special merge node");
 325   assert(!is_hidden_merge(src), "must not be a special merge node");
 326   dstphi->add_req(src);
 327 }
 328 
 329 //-----------------------combine_exception_states------------------------------
 330 // This helper function combines exception states by building phis on a
 331 // specially marked state-merging region.  These regions and phis are
 332 // untransformed, and can build up gradually.  The region is marked by
 333 // having a control input of its exception map, rather than NULL.  Such
 334 // regions do not appear except in this function, and in use_exception_state.
 335 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 336   if (failing())  return;  // dying anyway...
 337   JVMState* ex_jvms = ex_map->_jvms;
 338   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 339   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 340   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 341   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 342   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 343   assert(ex_map->req() == phi_map->req(), "matching maps");
 344   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 345   Node*         hidden_merge_mark = root();
 346   Node*         region  = phi_map->control();
 347   MergeMemNode* phi_mem = phi_map->merged_memory();
 348   MergeMemNode* ex_mem  = ex_map->merged_memory();
 349   if (region->in(0) != hidden_merge_mark) {
 350     // The control input is not (yet) a specially-marked region in phi_map.
 351     // Make it so, and build some phis.
 352     region = new RegionNode(2);
 353     _gvn.set_type(region, Type::CONTROL);
 354     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 355     region->init_req(1, phi_map->control());
 356     phi_map->set_control(region);
 357     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 358     record_for_igvn(io_phi);
 359     _gvn.set_type(io_phi, Type::ABIO);
 360     phi_map->set_i_o(io_phi);
 361     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 362       Node* m = mms.memory();
 363       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 364       record_for_igvn(m_phi);
 365       _gvn.set_type(m_phi, Type::MEMORY);
 366       mms.set_memory(m_phi);
 367     }
 368   }
 369 
 370   // Either or both of phi_map and ex_map might already be converted into phis.
 371   Node* ex_control = ex_map->control();
 372   // if there is special marking on ex_map also, we add multiple edges from src
 373   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 374   // how wide was the destination phi_map, originally?
 375   uint orig_width = region->req();
 376 
 377   if (add_multiple) {
 378     add_n_reqs(region, ex_control);
 379     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 380   } else {
 381     // ex_map has no merges, so we just add single edges everywhere
 382     add_one_req(region, ex_control);
 383     add_one_req(phi_map->i_o(), ex_map->i_o());
 384   }
 385   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 386     if (mms.is_empty()) {
 387       // get a copy of the base memory, and patch some inputs into it
 388       const TypePtr* adr_type = mms.adr_type(C);
 389       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 390       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 391       mms.set_memory(phi);
 392       // Prepare to append interesting stuff onto the newly sliced phi:
 393       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 394     }
 395     // Append stuff from ex_map:
 396     if (add_multiple) {
 397       add_n_reqs(mms.memory(), mms.memory2());
 398     } else {
 399       add_one_req(mms.memory(), mms.memory2());
 400     }
 401   }
 402   uint limit = ex_map->req();
 403   for (uint i = TypeFunc::Parms; i < limit; i++) {
 404     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 405     if (i == tos)  i = ex_jvms->monoff();
 406     Node* src = ex_map->in(i);
 407     Node* dst = phi_map->in(i);
 408     if (src != dst) {
 409       PhiNode* phi;
 410       if (dst->in(0) != region) {
 411         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 412         record_for_igvn(phi);
 413         _gvn.set_type(phi, phi->type());
 414         phi_map->set_req(i, dst);
 415         // Prepare to append interesting stuff onto the new phi:
 416         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 417       } else {
 418         assert(dst->is_Phi(), "nobody else uses a hidden region");
 419         phi = dst->as_Phi();
 420       }
 421       if (add_multiple && src->in(0) == ex_control) {
 422         // Both are phis.
 423         add_n_reqs(dst, src);
 424       } else {
 425         while (dst->req() < region->req())  add_one_req(dst, src);
 426       }
 427       const Type* srctype = _gvn.type(src);
 428       if (phi->type() != srctype) {
 429         const Type* dsttype = phi->type()->meet_speculative(srctype);
 430         if (phi->type() != dsttype) {
 431           phi->set_type(dsttype);
 432           _gvn.set_type(phi, dsttype);
 433         }
 434       }
 435     }
 436   }
 437   phi_map->merge_replaced_nodes_with(ex_map);
 438 }
 439 
 440 //--------------------------use_exception_state--------------------------------
 441 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 442   if (failing()) { stop(); return top(); }
 443   Node* region = phi_map->control();
 444   Node* hidden_merge_mark = root();
 445   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 446   Node* ex_oop = clear_saved_ex_oop(phi_map);
 447   if (region->in(0) == hidden_merge_mark) {
 448     // Special marking for internal ex-states.  Process the phis now.
 449     region->set_req(0, region);  // now it's an ordinary region
 450     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 451     // Note: Setting the jvms also sets the bci and sp.
 452     set_control(_gvn.transform(region));
 453     uint tos = jvms()->stkoff() + sp();
 454     for (uint i = 1; i < tos; i++) {
 455       Node* x = phi_map->in(i);
 456       if (x->in(0) == region) {
 457         assert(x->is_Phi(), "expected a special phi");
 458         phi_map->set_req(i, _gvn.transform(x));
 459       }
 460     }
 461     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 462       Node* x = mms.memory();
 463       if (x->in(0) == region) {
 464         assert(x->is_Phi(), "nobody else uses a hidden region");
 465         mms.set_memory(_gvn.transform(x));
 466       }
 467     }
 468     if (ex_oop->in(0) == region) {
 469       assert(ex_oop->is_Phi(), "expected a special phi");
 470       ex_oop = _gvn.transform(ex_oop);
 471     }
 472   } else {
 473     set_jvms(phi_map->jvms());
 474   }
 475 
 476   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 477   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 478   return ex_oop;
 479 }
 480 
 481 //---------------------------------java_bc-------------------------------------
 482 Bytecodes::Code GraphKit::java_bc() const {
 483   ciMethod* method = this->method();
 484   int       bci    = this->bci();
 485   if (method != NULL && bci != InvocationEntryBci)
 486     return method->java_code_at_bci(bci);
 487   else
 488     return Bytecodes::_illegal;
 489 }
 490 
 491 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 492                                                           bool must_throw) {
 493     // if the exception capability is set, then we will generate code
 494     // to check the JavaThread.should_post_on_exceptions flag to see
 495     // if we actually need to report exception events (for this
 496     // thread).  If we don't need to report exception events, we will
 497     // take the normal fast path provided by add_exception_events.  If
 498     // exception event reporting is enabled for this thread, we will
 499     // take the uncommon_trap in the BuildCutout below.
 500 
 501     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 502     Node* jthread = _gvn.transform(new ThreadLocalNode());
 503     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 504     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
 505 
 506     // Test the should_post_on_exceptions_flag vs. 0
 507     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
 508     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
 509 
 510     // Branch to slow_path if should_post_on_exceptions_flag was true
 511     { BuildCutout unless(this, tst, PROB_MAX);
 512       // Do not try anything fancy if we're notifying the VM on every throw.
 513       // Cf. case Bytecodes::_athrow in parse2.cpp.
 514       uncommon_trap(reason, Deoptimization::Action_none,
 515                     (ciKlass*)NULL, (char*)NULL, must_throw);
 516     }
 517 
 518 }
 519 
 520 //------------------------------builtin_throw----------------------------------
 521 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 522   bool must_throw = true;
 523 
 524   if (env()->jvmti_can_post_on_exceptions()) {
 525     // check if we must post exception events, take uncommon trap if so
 526     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 527     // here if should_post_on_exceptions is false
 528     // continue on with the normal codegen
 529   }
 530 
 531   // If this particular condition has not yet happened at this
 532   // bytecode, then use the uncommon trap mechanism, and allow for
 533   // a future recompilation if several traps occur here.
 534   // If the throw is hot, try to use a more complicated inline mechanism
 535   // which keeps execution inside the compiled code.
 536   bool treat_throw_as_hot = false;
 537   ciMethodData* md = method()->method_data();
 538 
 539   if (ProfileTraps) {
 540     if (too_many_traps(reason)) {
 541       treat_throw_as_hot = true;
 542     }
 543     // (If there is no MDO at all, assume it is early in
 544     // execution, and that any deopts are part of the
 545     // startup transient, and don't need to be remembered.)
 546 
 547     // Also, if there is a local exception handler, treat all throws
 548     // as hot if there has been at least one in this method.
 549     if (C->trap_count(reason) != 0
 550         && method()->method_data()->trap_count(reason) != 0
 551         && has_ex_handler()) {
 552         treat_throw_as_hot = true;
 553     }
 554   }
 555 
 556   // If this throw happens frequently, an uncommon trap might cause
 557   // a performance pothole.  If there is a local exception handler,
 558   // and if this particular bytecode appears to be deoptimizing often,
 559   // let us handle the throw inline, with a preconstructed instance.
 560   // Note:   If the deopt count has blown up, the uncommon trap
 561   // runtime is going to flush this nmethod, not matter what.
 562   if (treat_throw_as_hot
 563       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 564     // If the throw is local, we use a pre-existing instance and
 565     // punt on the backtrace.  This would lead to a missing backtrace
 566     // (a repeat of 4292742) if the backtrace object is ever asked
 567     // for its backtrace.
 568     // Fixing this remaining case of 4292742 requires some flavor of
 569     // escape analysis.  Leave that for the future.
 570     ciInstance* ex_obj = NULL;
 571     switch (reason) {
 572     case Deoptimization::Reason_null_check:
 573       ex_obj = env()->NullPointerException_instance();
 574       break;
 575     case Deoptimization::Reason_div0_check:
 576       ex_obj = env()->ArithmeticException_instance();
 577       break;
 578     case Deoptimization::Reason_range_check:
 579       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 580       break;
 581     case Deoptimization::Reason_class_check:
 582       if (java_bc() == Bytecodes::_aastore) {
 583         ex_obj = env()->ArrayStoreException_instance();
 584       } else {
 585         ex_obj = env()->ClassCastException_instance();
 586       }
 587       break;
 588     }
 589     if (failing()) { stop(); return; }  // exception allocation might fail
 590     if (ex_obj != NULL) {
 591       // Cheat with a preallocated exception object.
 592       if (C->log() != NULL)
 593         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 594                        Deoptimization::trap_reason_name(reason));
 595       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 596       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
 597 
 598       // Clear the detail message of the preallocated exception object.
 599       // Weblogic sometimes mutates the detail message of exceptions
 600       // using reflection.
 601       int offset = java_lang_Throwable::get_detailMessage_offset();
 602       const TypePtr* adr_typ = ex_con->add_offset(offset);
 603 
 604       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 605       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 606       // Conservatively release stores of object references.
 607       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
 608 
 609       add_exception_state(make_exception_state(ex_node));
 610       return;
 611     }
 612   }
 613 
 614   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 615   // It won't be much cheaper than bailing to the interp., since we'll
 616   // have to pass up all the debug-info, and the runtime will have to
 617   // create the stack trace.
 618 
 619   // Usual case:  Bail to interpreter.
 620   // Reserve the right to recompile if we haven't seen anything yet.
 621 
 622   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
 623   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 624   if (treat_throw_as_hot
 625       && (method()->method_data()->trap_recompiled_at(bci(), m)
 626           || C->too_many_traps(reason))) {
 627     // We cannot afford to take more traps here.  Suffer in the interpreter.
 628     if (C->log() != NULL)
 629       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 630                      Deoptimization::trap_reason_name(reason),
 631                      C->trap_count(reason));
 632     action = Deoptimization::Action_none;
 633   }
 634 
 635   // "must_throw" prunes the JVM state to include only the stack, if there
 636   // are no local exception handlers.  This should cut down on register
 637   // allocation time and code size, by drastically reducing the number
 638   // of in-edges on the call to the uncommon trap.
 639 
 640   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 641 }
 642 
 643 
 644 //----------------------------PreserveJVMState---------------------------------
 645 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 646   debug_only(kit->verify_map());
 647   _kit    = kit;
 648   _map    = kit->map();   // preserve the map
 649   _sp     = kit->sp();
 650   kit->set_map(clone_map ? kit->clone_map() : NULL);
 651 #ifdef ASSERT
 652   _bci    = kit->bci();
 653   Parse* parser = kit->is_Parse();
 654   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 655   _block  = block;
 656 #endif
 657 }
 658 PreserveJVMState::~PreserveJVMState() {
 659   GraphKit* kit = _kit;
 660 #ifdef ASSERT
 661   assert(kit->bci() == _bci, "bci must not shift");
 662   Parse* parser = kit->is_Parse();
 663   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 664   assert(block == _block,    "block must not shift");
 665 #endif
 666   kit->set_map(_map);
 667   kit->set_sp(_sp);
 668 }
 669 
 670 
 671 //-----------------------------BuildCutout-------------------------------------
 672 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 673   : PreserveJVMState(kit)
 674 {
 675   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 676   SafePointNode* outer_map = _map;   // preserved map is caller's
 677   SafePointNode* inner_map = kit->map();
 678   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 679   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
 680   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
 681 }
 682 BuildCutout::~BuildCutout() {
 683   GraphKit* kit = _kit;
 684   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 685 }
 686 
 687 //---------------------------PreserveReexecuteState----------------------------
 688 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 689   assert(!kit->stopped(), "must call stopped() before");
 690   _kit    =    kit;
 691   _sp     =    kit->sp();
 692   _reexecute = kit->jvms()->_reexecute;
 693 }
 694 PreserveReexecuteState::~PreserveReexecuteState() {
 695   if (_kit->stopped()) return;
 696   _kit->jvms()->_reexecute = _reexecute;
 697   _kit->set_sp(_sp);
 698 }
 699 
 700 //------------------------------clone_map--------------------------------------
 701 // Implementation of PreserveJVMState
 702 //
 703 // Only clone_map(...) here. If this function is only used in the
 704 // PreserveJVMState class we may want to get rid of this extra
 705 // function eventually and do it all there.
 706 
 707 SafePointNode* GraphKit::clone_map() {
 708   if (map() == NULL)  return NULL;
 709 
 710   // Clone the memory edge first
 711   Node* mem = MergeMemNode::make(map()->memory());
 712   gvn().set_type_bottom(mem);
 713 
 714   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 715   JVMState* jvms = this->jvms();
 716   JVMState* clonejvms = jvms->clone_shallow(C);
 717   clonemap->set_memory(mem);
 718   clonemap->set_jvms(clonejvms);
 719   clonejvms->set_map(clonemap);
 720   record_for_igvn(clonemap);
 721   gvn().set_type_bottom(clonemap);
 722   return clonemap;
 723 }
 724 
 725 
 726 //-----------------------------set_map_clone-----------------------------------
 727 void GraphKit::set_map_clone(SafePointNode* m) {
 728   _map = m;
 729   _map = clone_map();
 730   _map->set_next_exception(NULL);
 731   debug_only(verify_map());
 732 }
 733 
 734 
 735 //----------------------------kill_dead_locals---------------------------------
 736 // Detect any locals which are known to be dead, and force them to top.
 737 void GraphKit::kill_dead_locals() {
 738   // Consult the liveness information for the locals.  If any
 739   // of them are unused, then they can be replaced by top().  This
 740   // should help register allocation time and cut down on the size
 741   // of the deoptimization information.
 742 
 743   // This call is made from many of the bytecode handling
 744   // subroutines called from the Big Switch in do_one_bytecode.
 745   // Every bytecode which might include a slow path is responsible
 746   // for killing its dead locals.  The more consistent we
 747   // are about killing deads, the fewer useless phis will be
 748   // constructed for them at various merge points.
 749 
 750   // bci can be -1 (InvocationEntryBci).  We return the entry
 751   // liveness for the method.
 752 
 753   if (method() == NULL || method()->code_size() == 0) {
 754     // We are building a graph for a call to a native method.
 755     // All locals are live.
 756     return;
 757   }
 758 
 759   ResourceMark rm;
 760 
 761   // Consult the liveness information for the locals.  If any
 762   // of them are unused, then they can be replaced by top().  This
 763   // should help register allocation time and cut down on the size
 764   // of the deoptimization information.
 765   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 766 
 767   int len = (int)live_locals.size();
 768   assert(len <= jvms()->loc_size(), "too many live locals");
 769   for (int local = 0; local < len; local++) {
 770     if (!live_locals.at(local)) {
 771       set_local(local, top());
 772     }
 773   }
 774 }
 775 
 776 #ifdef ASSERT
 777 //-------------------------dead_locals_are_killed------------------------------
 778 // Return true if all dead locals are set to top in the map.
 779 // Used to assert "clean" debug info at various points.
 780 bool GraphKit::dead_locals_are_killed() {
 781   if (method() == NULL || method()->code_size() == 0) {
 782     // No locals need to be dead, so all is as it should be.
 783     return true;
 784   }
 785 
 786   // Make sure somebody called kill_dead_locals upstream.
 787   ResourceMark rm;
 788   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 789     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 790     SafePointNode* map = jvms->map();
 791     ciMethod* method = jvms->method();
 792     int       bci    = jvms->bci();
 793     if (jvms == this->jvms()) {
 794       bci = this->bci();  // it might not yet be synched
 795     }
 796     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 797     int len = (int)live_locals.size();
 798     if (!live_locals.is_valid() || len == 0)
 799       // This method is trivial, or is poisoned by a breakpoint.
 800       return true;
 801     assert(len == jvms->loc_size(), "live map consistent with locals map");
 802     for (int local = 0; local < len; local++) {
 803       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 804         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 805           tty->print_cr("Zombie local %d: ", local);
 806           jvms->dump();
 807         }
 808         return false;
 809       }
 810     }
 811   }
 812   return true;
 813 }
 814 
 815 #endif //ASSERT
 816 
 817 // Helper function for enforcing certain bytecodes to reexecute if
 818 // deoptimization happens
 819 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 820   ciMethod* cur_method = jvms->method();
 821   int       cur_bci   = jvms->bci();
 822   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 823     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 824     return Interpreter::bytecode_should_reexecute(code) ||
 825            is_anewarray && code == Bytecodes::_multianewarray;
 826     // Reexecute _multianewarray bytecode which was replaced with
 827     // sequence of [a]newarray. See Parse::do_multianewarray().
 828     //
 829     // Note: interpreter should not have it set since this optimization
 830     // is limited by dimensions and guarded by flag so in some cases
 831     // multianewarray() runtime calls will be generated and
 832     // the bytecode should not be reexecutes (stack will not be reset).
 833   } else
 834     return false;
 835 }
 836 
 837 // Helper function for adding JVMState and debug information to node
 838 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 839   // Add the safepoint edges to the call (or other safepoint).
 840 
 841   // Make sure dead locals are set to top.  This
 842   // should help register allocation time and cut down on the size
 843   // of the deoptimization information.
 844   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 845 
 846   // Walk the inline list to fill in the correct set of JVMState's
 847   // Also fill in the associated edges for each JVMState.
 848 
 849   // If the bytecode needs to be reexecuted we need to put
 850   // the arguments back on the stack.
 851   const bool should_reexecute = jvms()->should_reexecute();
 852   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 853 
 854   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 855   // undefined if the bci is different.  This is normal for Parse but it
 856   // should not happen for LibraryCallKit because only one bci is processed.
 857   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 858          "in LibraryCallKit the reexecute bit should not change");
 859 
 860   // If we are guaranteed to throw, we can prune everything but the
 861   // input to the current bytecode.
 862   bool can_prune_locals = false;
 863   uint stack_slots_not_pruned = 0;
 864   int inputs = 0, depth = 0;
 865   if (must_throw) {
 866     assert(method() == youngest_jvms->method(), "sanity");
 867     if (compute_stack_effects(inputs, depth)) {
 868       can_prune_locals = true;
 869       stack_slots_not_pruned = inputs;
 870     }
 871   }
 872 
 873   if (env()->should_retain_local_variables()) {
 874     // At any safepoint, this method can get breakpointed, which would
 875     // then require an immediate deoptimization.
 876     can_prune_locals = false;  // do not prune locals
 877     stack_slots_not_pruned = 0;
 878   }
 879 
 880   // do not scribble on the input jvms
 881   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 882   call->set_jvms(out_jvms); // Start jvms list for call node
 883 
 884   // For a known set of bytecodes, the interpreter should reexecute them if
 885   // deoptimization happens. We set the reexecute state for them here
 886   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 887       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 888     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 889   }
 890 
 891   // Presize the call:
 892   DEBUG_ONLY(uint non_debug_edges = call->req());
 893   call->add_req_batch(top(), youngest_jvms->debug_depth());
 894   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 895 
 896   // Set up edges so that the call looks like this:
 897   //  Call [state:] ctl io mem fptr retadr
 898   //       [parms:] parm0 ... parmN
 899   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 900   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 901   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 902   // Note that caller debug info precedes callee debug info.
 903 
 904   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 905   uint debug_ptr = call->req();
 906 
 907   // Loop over the map input edges associated with jvms, add them
 908   // to the call node, & reset all offsets to match call node array.
 909   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 910     uint debug_end   = debug_ptr;
 911     uint debug_start = debug_ptr - in_jvms->debug_size();
 912     debug_ptr = debug_start;  // back up the ptr
 913 
 914     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 915     uint j, k, l;
 916     SafePointNode* in_map = in_jvms->map();
 917     out_jvms->set_map(call);
 918 
 919     if (can_prune_locals) {
 920       assert(in_jvms->method() == out_jvms->method(), "sanity");
 921       // If the current throw can reach an exception handler in this JVMS,
 922       // then we must keep everything live that can reach that handler.
 923       // As a quick and dirty approximation, we look for any handlers at all.
 924       if (in_jvms->method()->has_exception_handlers()) {
 925         can_prune_locals = false;
 926       }
 927     }
 928 
 929     // Add the Locals
 930     k = in_jvms->locoff();
 931     l = in_jvms->loc_size();
 932     out_jvms->set_locoff(p);
 933     if (!can_prune_locals) {
 934       for (j = 0; j < l; j++)
 935         call->set_req(p++, in_map->in(k+j));
 936     } else {
 937       p += l;  // already set to top above by add_req_batch
 938     }
 939 
 940     // Add the Expression Stack
 941     k = in_jvms->stkoff();
 942     l = in_jvms->sp();
 943     out_jvms->set_stkoff(p);
 944     if (!can_prune_locals) {
 945       for (j = 0; j < l; j++)
 946         call->set_req(p++, in_map->in(k+j));
 947     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 948       // Divide stack into {S0,...,S1}, where S0 is set to top.
 949       uint s1 = stack_slots_not_pruned;
 950       stack_slots_not_pruned = 0;  // for next iteration
 951       if (s1 > l)  s1 = l;
 952       uint s0 = l - s1;
 953       p += s0;  // skip the tops preinstalled by add_req_batch
 954       for (j = s0; j < l; j++)
 955         call->set_req(p++, in_map->in(k+j));
 956     } else {
 957       p += l;  // already set to top above by add_req_batch
 958     }
 959 
 960     // Add the Monitors
 961     k = in_jvms->monoff();
 962     l = in_jvms->mon_size();
 963     out_jvms->set_monoff(p);
 964     for (j = 0; j < l; j++)
 965       call->set_req(p++, in_map->in(k+j));
 966 
 967     // Copy any scalar object fields.
 968     k = in_jvms->scloff();
 969     l = in_jvms->scl_size();
 970     out_jvms->set_scloff(p);
 971     for (j = 0; j < l; j++)
 972       call->set_req(p++, in_map->in(k+j));
 973 
 974     // Finish the new jvms.
 975     out_jvms->set_endoff(p);
 976 
 977     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 978     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 979     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 980     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 981     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 982     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 983 
 984     // Update the two tail pointers in parallel.
 985     out_jvms = out_jvms->caller();
 986     in_jvms  = in_jvms->caller();
 987   }
 988 
 989   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
 990 
 991   // Test the correctness of JVMState::debug_xxx accessors:
 992   assert(call->jvms()->debug_start() == non_debug_edges, "");
 993   assert(call->jvms()->debug_end()   == call->req(), "");
 994   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
 995 }
 996 
 997 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
 998   Bytecodes::Code code = java_bc();
 999   if (code == Bytecodes::_wide) {
1000     code = method()->java_code_at_bci(bci() + 1);
1001   }
1002 
1003   BasicType rtype = T_ILLEGAL;
1004   int       rsize = 0;
1005 
1006   if (code != Bytecodes::_illegal) {
1007     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1008     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1009     if (rtype < T_CONFLICT)
1010       rsize = type2size[rtype];
1011   }
1012 
1013   switch (code) {
1014   case Bytecodes::_illegal:
1015     return false;
1016 
1017   case Bytecodes::_ldc:
1018   case Bytecodes::_ldc_w:
1019   case Bytecodes::_ldc2_w:
1020     inputs = 0;
1021     break;
1022 
1023   case Bytecodes::_dup:         inputs = 1;  break;
1024   case Bytecodes::_dup_x1:      inputs = 2;  break;
1025   case Bytecodes::_dup_x2:      inputs = 3;  break;
1026   case Bytecodes::_dup2:        inputs = 2;  break;
1027   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1028   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1029   case Bytecodes::_swap:        inputs = 2;  break;
1030   case Bytecodes::_arraylength: inputs = 1;  break;
1031 
1032   case Bytecodes::_getstatic:
1033   case Bytecodes::_putstatic:
1034   case Bytecodes::_getfield:
1035   case Bytecodes::_vgetfield:
1036   case Bytecodes::_putfield:
1037     {
1038       bool ignored_will_link;
1039       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1040       int      size  = field->type()->size();
1041       bool is_get = (depth >= 0), is_static = (depth & 1);
1042       inputs = (is_static ? 0 : 1);
1043       if (is_get) {
1044         depth = size - inputs;
1045       } else {
1046         inputs += size;        // putxxx pops the value from the stack
1047         depth = - inputs;
1048       }
1049     }
1050     break;
1051 
1052   case Bytecodes::_invokevirtual:
1053   case Bytecodes::_invokedirect:
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::_vnew: {
1081     // vnew pops the values from the stack
1082     ciValueKlass* vk = method()->holder()->as_value_klass();
1083     inputs = vk->param_size();
1084     depth = rsize - inputs;
1085     break;
1086   }
1087 
1088   case Bytecodes::_ireturn:
1089   case Bytecodes::_lreturn:
1090   case Bytecodes::_freturn:
1091   case Bytecodes::_dreturn:
1092   case Bytecodes::_areturn:
1093   case Bytecodes::_vreturn:
1094     assert(rsize = -depth, "");
1095     inputs = rsize;
1096     break;
1097 
1098   case Bytecodes::_jsr:
1099   case Bytecodes::_jsr_w:
1100     inputs = 0;
1101     depth  = 1;                  // S.B. depth=1, not zero
1102     break;
1103 
1104   default:
1105     // bytecode produces a typed result
1106     inputs = rsize - depth;
1107     assert(inputs >= 0, "");
1108     break;
1109   }
1110 
1111 #ifdef ASSERT
1112   // spot check
1113   int outputs = depth + inputs;
1114   assert(outputs >= 0, "sanity");
1115   switch (code) {
1116   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1117   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1118   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1119   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1120   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1121   }
1122 #endif //ASSERT
1123 
1124   return true;
1125 }
1126 
1127 
1128 
1129 //------------------------------basic_plus_adr---------------------------------
1130 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1131   // short-circuit a common case
1132   if (offset == intcon(0))  return ptr;
1133   return _gvn.transform( new AddPNode(base, ptr, offset) );
1134 }
1135 
1136 Node* GraphKit::ConvI2L(Node* offset) {
1137   // short-circuit a common case
1138   jint offset_con = find_int_con(offset, Type::OffsetBot);
1139   if (offset_con != Type::OffsetBot) {
1140     return longcon((jlong) offset_con);
1141   }
1142   return _gvn.transform( new ConvI2LNode(offset));
1143 }
1144 
1145 Node* GraphKit::ConvI2UL(Node* offset) {
1146   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1147   if (offset_con != (juint) Type::OffsetBot) {
1148     return longcon((julong) offset_con);
1149   }
1150   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1151   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1152   return _gvn.transform( new AndLNode(conv, mask) );
1153 }
1154 
1155 Node* GraphKit::ConvL2I(Node* offset) {
1156   // short-circuit a common case
1157   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1158   if (offset_con != (jlong)Type::OffsetBot) {
1159     return intcon((int) offset_con);
1160   }
1161   return _gvn.transform( new ConvL2INode(offset));
1162 }
1163 
1164 //-------------------------load_object_klass-----------------------------------
1165 Node* GraphKit::load_object_klass(Node* obj) {
1166   // Special-case a fresh allocation to avoid building nodes:
1167   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1168   if (akls != NULL)  return akls;
1169   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1170   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1171 }
1172 
1173 //-------------------------load_array_length-----------------------------------
1174 Node* GraphKit::load_array_length(Node* array) {
1175   // Special-case a fresh allocation to avoid building nodes:
1176   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1177   Node *alen;
1178   if (alloc == NULL) {
1179     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1180     alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1181   } else {
1182     alen = alloc->Ideal_length();
1183     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1184     if (ccast != alen) {
1185       alen = _gvn.transform(ccast);
1186     }
1187   }
1188   return alen;
1189 }
1190 
1191 //------------------------------do_null_check----------------------------------
1192 // Helper function to do a NULL pointer check.  Returned value is
1193 // the incoming address with NULL casted away.  You are allowed to use the
1194 // not-null value only if you are control dependent on the test.
1195 #ifndef PRODUCT
1196 extern int explicit_null_checks_inserted,
1197            explicit_null_checks_elided;
1198 #endif
1199 Node* GraphKit::null_check_common(Node* value, BasicType type,
1200                                   // optional arguments for variations:
1201                                   bool assert_null,
1202                                   Node* *null_control,
1203                                   bool speculative) {
1204   assert(!assert_null || null_control == NULL, "not both at once");
1205   if (stopped())  return top();
1206   if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1207     // For some performance testing, we may wish to suppress null checking.
1208     value = cast_not_null(value);   // Make it appear to be non-null (4962416).
1209     return value;
1210   }
1211   NOT_PRODUCT(explicit_null_checks_inserted++);
1212 
1213   // Construct NULL check
1214   Node *chk = NULL;
1215   switch(type) {
1216     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1217     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1218     case T_ARRAY  : // fall through
1219       type = T_OBJECT;  // simplify further tests
1220     case T_OBJECT : {
1221       const Type *t = _gvn.type( value );
1222 
1223       const TypeOopPtr* tp = t->isa_oopptr();
1224       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1225           // Only for do_null_check, not any of its siblings:
1226           && !assert_null && null_control == NULL) {
1227         // Usually, any field access or invocation on an unloaded oop type
1228         // will simply fail to link, since the statically linked class is
1229         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1230         // the static class is loaded but the sharper oop type is not.
1231         // Rather than checking for this obscure case in lots of places,
1232         // we simply observe that a null check on an unloaded class
1233         // will always be followed by a nonsense operation, so we
1234         // can just issue the uncommon trap here.
1235         // Our access to the unloaded class will only be correct
1236         // after it has been loaded and initialized, which requires
1237         // a trip through the interpreter.
1238 #ifndef PRODUCT
1239         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1240 #endif
1241         uncommon_trap(Deoptimization::Reason_unloaded,
1242                       Deoptimization::Action_reinterpret,
1243                       tp->klass(), "!loaded");
1244         return top();
1245       }
1246 
1247       if (assert_null) {
1248         // See if the type is contained in NULL_PTR.
1249         // If so, then the value is already null.
1250         if (t->higher_equal(TypePtr::NULL_PTR)) {
1251           NOT_PRODUCT(explicit_null_checks_elided++);
1252           return value;           // Elided null assert quickly!
1253         }
1254       } else {
1255         // See if mixing in the NULL pointer changes type.
1256         // If so, then the NULL pointer was not allowed in the original
1257         // type.  In other words, "value" was not-null.
1258         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1259           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1260           NOT_PRODUCT(explicit_null_checks_elided++);
1261           return value;           // Elided null check quickly!
1262         }
1263       }
1264       chk = new CmpPNode( value, null() );
1265       break;
1266     }
1267 
1268     default:
1269       fatal("unexpected type: %s", type2name(type));
1270   }
1271   assert(chk != NULL, "sanity check");
1272   chk = _gvn.transform(chk);
1273 
1274   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1275   BoolNode *btst = new BoolNode( chk, btest);
1276   Node   *tst = _gvn.transform( btst );
1277 
1278   //-----------
1279   // if peephole optimizations occurred, a prior test existed.
1280   // If a prior test existed, maybe it dominates as we can avoid this test.
1281   if (tst != btst && type == T_OBJECT) {
1282     // At this point we want to scan up the CFG to see if we can
1283     // find an identical test (and so avoid this test altogether).
1284     Node *cfg = control();
1285     int depth = 0;
1286     while( depth < 16 ) {       // Limit search depth for speed
1287       if( cfg->Opcode() == Op_IfTrue &&
1288           cfg->in(0)->in(1) == tst ) {
1289         // Found prior test.  Use "cast_not_null" to construct an identical
1290         // CastPP (and hence hash to) as already exists for the prior test.
1291         // Return that casted value.
1292         if (assert_null) {
1293           replace_in_map(value, null());
1294           return null();  // do not issue the redundant test
1295         }
1296         Node *oldcontrol = control();
1297         set_control(cfg);
1298         Node *res = cast_not_null(value);
1299         set_control(oldcontrol);
1300         NOT_PRODUCT(explicit_null_checks_elided++);
1301         return res;
1302       }
1303       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1304       if (cfg == NULL)  break;  // Quit at region nodes
1305       depth++;
1306     }
1307   }
1308 
1309   //-----------
1310   // Branch to failure if null
1311   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1312   Deoptimization::DeoptReason reason;
1313   if (assert_null) {
1314     reason = Deoptimization::Reason_null_assert;
1315   } else if (type == T_OBJECT) {
1316     reason = Deoptimization::reason_null_check(speculative);
1317   } else {
1318     reason = Deoptimization::Reason_div0_check;
1319   }
1320   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1321   // ciMethodData::has_trap_at will return a conservative -1 if any
1322   // must-be-null assertion has failed.  This could cause performance
1323   // problems for a method after its first do_null_assert failure.
1324   // Consider using 'Reason_class_check' instead?
1325 
1326   // To cause an implicit null check, we set the not-null probability
1327   // to the maximum (PROB_MAX).  For an explicit check the probability
1328   // is set to a smaller value.
1329   if (null_control != NULL || too_many_traps(reason)) {
1330     // probability is less likely
1331     ok_prob =  PROB_LIKELY_MAG(3);
1332   } else if (!assert_null &&
1333              (ImplicitNullCheckThreshold > 0) &&
1334              method() != NULL &&
1335              (method()->method_data()->trap_count(reason)
1336               >= (uint)ImplicitNullCheckThreshold)) {
1337     ok_prob =  PROB_LIKELY_MAG(3);
1338   }
1339 
1340   if (null_control != NULL) {
1341     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1342     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1343     set_control(      _gvn.transform( new IfTrueNode(iff)));
1344 #ifndef PRODUCT
1345     if (null_true == top()) {
1346       explicit_null_checks_elided++;
1347     }
1348 #endif
1349     (*null_control) = null_true;
1350   } else {
1351     BuildCutout unless(this, tst, ok_prob);
1352     // Check for optimizer eliding test at parse time
1353     if (stopped()) {
1354       // Failure not possible; do not bother making uncommon trap.
1355       NOT_PRODUCT(explicit_null_checks_elided++);
1356     } else if (assert_null) {
1357       uncommon_trap(reason,
1358                     Deoptimization::Action_make_not_entrant,
1359                     NULL, "assert_null");
1360     } else {
1361       replace_in_map(value, zerocon(type));
1362       builtin_throw(reason);
1363     }
1364   }
1365 
1366   // Must throw exception, fall-thru not possible?
1367   if (stopped()) {
1368     return top();               // No result
1369   }
1370 
1371   if (assert_null) {
1372     // Cast obj to null on this path.
1373     replace_in_map(value, zerocon(type));
1374     return zerocon(type);
1375   }
1376 
1377   // Cast obj to not-null on this path, if there is no null_control.
1378   // (If there is a null_control, a non-null value may come back to haunt us.)
1379   if (type == T_OBJECT) {
1380     Node* cast = cast_not_null(value, false);
1381     if (null_control == NULL || (*null_control) == top())
1382       replace_in_map(value, cast);
1383     value = cast;
1384   }
1385 
1386   return value;
1387 }
1388 
1389 
1390 //------------------------------cast_not_null----------------------------------
1391 // Cast obj to not-null on this path
1392 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1393   const Type *t = _gvn.type(obj);
1394   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1395   // Object is already not-null?
1396   if( t == t_not_null ) return obj;
1397 
1398   Node *cast = new CastPPNode(obj,t_not_null);
1399   cast->init_req(0, control());
1400   cast = _gvn.transform( cast );
1401 
1402   // Scan for instances of 'obj' in the current JVM mapping.
1403   // These instances are known to be not-null after the test.
1404   if (do_replace_in_map)
1405     replace_in_map(obj, cast);
1406 
1407   return cast;                  // Return casted value
1408 }
1409 
1410 
1411 //--------------------------replace_in_map-------------------------------------
1412 void GraphKit::replace_in_map(Node* old, Node* neww) {
1413   if (old == neww) {
1414     return;
1415   }
1416 
1417   map()->replace_edge(old, neww);
1418 
1419   // Note: This operation potentially replaces any edge
1420   // on the map.  This includes locals, stack, and monitors
1421   // of the current (innermost) JVM state.
1422 
1423   // don't let inconsistent types from profiling escape this
1424   // method
1425 
1426   const Type* told = _gvn.type(old);
1427   const Type* tnew = _gvn.type(neww);
1428 
1429   if (!tnew->higher_equal(told)) {
1430     return;
1431   }
1432 
1433   map()->record_replaced_node(old, neww);
1434 }
1435 
1436 
1437 //=============================================================================
1438 //--------------------------------memory---------------------------------------
1439 Node* GraphKit::memory(uint alias_idx) {
1440   MergeMemNode* mem = merged_memory();
1441   Node* p = mem->memory_at(alias_idx);
1442   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1443   return p;
1444 }
1445 
1446 //-----------------------------reset_memory------------------------------------
1447 Node* GraphKit::reset_memory() {
1448   Node* mem = map()->memory();
1449   // do not use this node for any more parsing!
1450   debug_only( map()->set_memory((Node*)NULL) );
1451   return _gvn.transform( mem );
1452 }
1453 
1454 //------------------------------set_all_memory---------------------------------
1455 void GraphKit::set_all_memory(Node* newmem) {
1456   Node* mergemem = MergeMemNode::make(newmem);
1457   gvn().set_type_bottom(mergemem);
1458   map()->set_memory(mergemem);
1459 }
1460 
1461 //------------------------------set_all_memory_call----------------------------
1462 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1463   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1464   set_all_memory(newmem);
1465 }
1466 
1467 //=============================================================================
1468 //
1469 // parser factory methods for MemNodes
1470 //
1471 // These are layered on top of the factory methods in LoadNode and StoreNode,
1472 // and integrate with the parser's memory state and _gvn engine.
1473 //
1474 
1475 // factory methods in "int adr_idx"
1476 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1477                           int adr_idx,
1478                           MemNode::MemOrd mo,
1479                           LoadNode::ControlDependency control_dependency,
1480                           bool require_atomic_access,
1481                           bool unaligned,
1482                           bool mismatched) {
1483   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1484   const TypePtr* adr_type = NULL; // debug-mode-only argument
1485   debug_only(adr_type = C->get_adr_type(adr_idx));
1486   Node* mem = memory(adr_idx);
1487   Node* ld;
1488   if (require_atomic_access && bt == T_LONG) {
1489     ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1490   } else if (require_atomic_access && bt == T_DOUBLE) {
1491     ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1492   } else {
1493     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1494   }
1495   ld = _gvn.transform(ld);
1496   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, "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   pre_barrier(true /* do_load */,
1637               control(), obj, adr, adr_idx, val, val_type,
1638               NULL /* pre_val */,
1639               bt);
1640 
1641   Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1642   post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1643   return store;
1644 }
1645 
1646 // Could be an array or object we don't know at compile time (unsafe ref.)
1647 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1648                              Node* obj,   // containing obj
1649                              Node* adr,  // actual adress to store val at
1650                              const TypePtr* adr_type,
1651                              Node* val,
1652                              BasicType bt,
1653                              MemNode::MemOrd mo,
1654                              bool mismatched) {
1655   Compile::AliasType* at = C->alias_type(adr_type);
1656   const TypeOopPtr* val_type = NULL;
1657   if (adr_type->isa_instptr()) {
1658     if (at->field() != NULL) {
1659       // known field.  This code is a copy of the do_put_xxx logic.
1660       ciField* field = at->field();
1661       if (!field->type()->is_loaded()) {
1662         val_type = TypeInstPtr::BOTTOM;
1663       } else {
1664         val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1665       }
1666     }
1667   } else if (adr_type->isa_aryptr()) {
1668     val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1669   }
1670   if (val_type == NULL) {
1671     val_type = TypeInstPtr::BOTTOM;
1672   }
1673   return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1674 }
1675 
1676 
1677 //-------------------------array_element_address-------------------------
1678 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1679                                       const TypeInt* sizetype, Node* ctrl) {
1680   uint shift  = exact_log2(type2aelembytes(elembt));
1681   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1682 
1683   // short-circuit a common case (saves lots of confusing waste motion)
1684   jint idx_con = find_int_con(idx, -1);
1685   if (idx_con >= 0) {
1686     intptr_t offset = header + ((intptr_t)idx_con << shift);
1687     return basic_plus_adr(ary, offset);
1688   }
1689 
1690   // must be correct type for alignment purposes
1691   Node* base  = basic_plus_adr(ary, header);
1692   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1693   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1694   return basic_plus_adr(ary, base, scale);
1695 }
1696 
1697 //-------------------------load_array_element-------------------------
1698 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1699   const Type* elemtype = arytype->elem();
1700   BasicType elembt = elemtype->array_element_basic_type();
1701   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1702   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1703   return ld;
1704 }
1705 
1706 //-------------------------set_arguments_for_java_call-------------------------
1707 // Arguments (pre-popped from the stack) are taken from the JVMS.
1708 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1709   // Add the call arguments:
1710   uint nargs = call->method()->arg_size();
1711   for (uint i = 0; i < nargs; i++) {
1712     Node* arg = argument(i);
1713     if (arg->is_ValueType()) {
1714       // Pass value type argument via oop to callee
1715       arg = arg->as_ValueType()->store_to_memory(this);
1716     }
1717     call->init_req(i + TypeFunc::Parms, arg);
1718   }
1719 }
1720 
1721 //---------------------------set_edges_for_java_call---------------------------
1722 // Connect a newly created call into the current JVMS.
1723 // A return value node (if any) is returned from set_edges_for_java_call.
1724 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1725 
1726   // Add the predefined inputs:
1727   call->init_req( TypeFunc::Control, control() );
1728   call->init_req( TypeFunc::I_O    , i_o() );
1729   call->init_req( TypeFunc::Memory , reset_memory() );
1730   call->init_req( TypeFunc::FramePtr, frameptr() );
1731   call->init_req( TypeFunc::ReturnAdr, top() );
1732 
1733   add_safepoint_edges(call, must_throw);
1734 
1735   Node* xcall = _gvn.transform(call);
1736 
1737   if (xcall == top()) {
1738     set_control(top());
1739     return;
1740   }
1741   assert(xcall == call, "call identity is stable");
1742 
1743   // Re-use the current map to produce the result.
1744 
1745   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1746   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1747   set_all_memory_call(xcall, separate_io_proj);
1748 
1749   //return xcall;   // no need, caller already has it
1750 }
1751 
1752 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1753   if (stopped())  return top();  // maybe the call folded up?
1754 
1755   // Capture the return value, if any.
1756   Node* ret;
1757   if (call->method() == NULL ||
1758       call->method()->return_type()->basic_type() == T_VOID)
1759         ret = top();
1760   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1761 
1762   // Note:  Since any out-of-line call can produce an exception,
1763   // we always insert an I_O projection from the call into the result.
1764 
1765   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1766 
1767   if (separate_io_proj) {
1768     // The caller requested separate projections be used by the fall
1769     // through and exceptional paths, so replace the projections for
1770     // the fall through path.
1771     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1772     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1773   }
1774   return ret;
1775 }
1776 
1777 //--------------------set_predefined_input_for_runtime_call--------------------
1778 // Reading and setting the memory state is way conservative here.
1779 // The real problem is that I am not doing real Type analysis on memory,
1780 // so I cannot distinguish card mark stores from other stores.  Across a GC
1781 // point the Store Barrier and the card mark memory has to agree.  I cannot
1782 // have a card mark store and its barrier split across the GC point from
1783 // either above or below.  Here I get that to happen by reading ALL of memory.
1784 // A better answer would be to separate out card marks from other memory.
1785 // For now, return the input memory state, so that it can be reused
1786 // after the call, if this call has restricted memory effects.
1787 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1788   // Set fixed predefined input arguments
1789   Node* memory = reset_memory();
1790   call->init_req( TypeFunc::Control,   control()  );
1791   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1792   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1793   call->init_req( TypeFunc::FramePtr,  frameptr() );
1794   call->init_req( TypeFunc::ReturnAdr, top()      );
1795   return memory;
1796 }
1797 
1798 //-------------------set_predefined_output_for_runtime_call--------------------
1799 // Set control and memory (not i_o) from the call.
1800 // If keep_mem is not NULL, use it for the output state,
1801 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1802 // If hook_mem is NULL, this call produces no memory effects at all.
1803 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1804 // then only that memory slice is taken from the call.
1805 // In the last case, we must put an appropriate memory barrier before
1806 // the call, so as to create the correct anti-dependencies on loads
1807 // preceding the call.
1808 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1809                                                       Node* keep_mem,
1810                                                       const TypePtr* hook_mem) {
1811   // no i/o
1812   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1813   if (keep_mem) {
1814     // First clone the existing memory state
1815     set_all_memory(keep_mem);
1816     if (hook_mem != NULL) {
1817       // Make memory for the call
1818       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1819       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1820       // We also use hook_mem to extract specific effects from arraycopy stubs.
1821       set_memory(mem, hook_mem);
1822     }
1823     // ...else the call has NO memory effects.
1824 
1825     // Make sure the call advertises its memory effects precisely.
1826     // This lets us build accurate anti-dependences in gcm.cpp.
1827     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1828            "call node must be constructed correctly");
1829   } else {
1830     assert(hook_mem == NULL, "");
1831     // This is not a "slow path" call; all memory comes from the call.
1832     set_all_memory_call(call);
1833   }
1834 }
1835 
1836 
1837 // Replace the call with the current state of the kit.
1838 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1839   JVMState* ejvms = NULL;
1840   if (has_exceptions()) {
1841     ejvms = transfer_exceptions_into_jvms();
1842   }
1843 
1844   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1845   ReplacedNodes replaced_nodes_exception;
1846   Node* ex_ctl = top();
1847 
1848   SafePointNode* final_state = stop();
1849 
1850   // Find all the needed outputs of this call
1851   CallProjections callprojs;
1852   call->extract_projections(&callprojs, true);
1853 
1854   Node* init_mem = call->in(TypeFunc::Memory);
1855   Node* final_mem = final_state->in(TypeFunc::Memory);
1856   Node* final_ctl = final_state->in(TypeFunc::Control);
1857   Node* final_io = final_state->in(TypeFunc::I_O);
1858 
1859   // Replace all the old call edges with the edges from the inlining result
1860   if (callprojs.fallthrough_catchproj != NULL) {
1861     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1862   }
1863   if (callprojs.fallthrough_memproj != NULL) {
1864     if (final_mem->is_MergeMem()) {
1865       // Parser's exits MergeMem was not transformed but may be optimized
1866       final_mem = _gvn.transform(final_mem);
1867     }
1868     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
1869   }
1870   if (callprojs.fallthrough_ioproj != NULL) {
1871     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
1872   }
1873 
1874   // Replace the result with the new result if it exists and is used
1875   if (callprojs.resproj != NULL && result != NULL) {
1876     C->gvn_replace_by(callprojs.resproj, result);
1877   }
1878 
1879   if (ejvms == NULL) {
1880     // No exception edges to simply kill off those paths
1881     if (callprojs.catchall_catchproj != NULL) {
1882       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1883     }
1884     if (callprojs.catchall_memproj != NULL) {
1885       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1886     }
1887     if (callprojs.catchall_ioproj != NULL) {
1888       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1889     }
1890     // Replace the old exception object with top
1891     if (callprojs.exobj != NULL) {
1892       C->gvn_replace_by(callprojs.exobj, C->top());
1893     }
1894   } else {
1895     GraphKit ekit(ejvms);
1896 
1897     // Load my combined exception state into the kit, with all phis transformed:
1898     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1899     replaced_nodes_exception = ex_map->replaced_nodes();
1900 
1901     Node* ex_oop = ekit.use_exception_state(ex_map);
1902 
1903     if (callprojs.catchall_catchproj != NULL) {
1904       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1905       ex_ctl = ekit.control();
1906     }
1907     if (callprojs.catchall_memproj != NULL) {
1908       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1909     }
1910     if (callprojs.catchall_ioproj != NULL) {
1911       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1912     }
1913 
1914     // Replace the old exception object with the newly created one
1915     if (callprojs.exobj != NULL) {
1916       C->gvn_replace_by(callprojs.exobj, ex_oop);
1917     }
1918   }
1919 
1920   // Disconnect the call from the graph
1921   call->disconnect_inputs(NULL, C);
1922   C->gvn_replace_by(call, C->top());
1923 
1924   // Clean up any MergeMems that feed other MergeMems since the
1925   // optimizer doesn't like that.
1926   if (final_mem->is_MergeMem()) {
1927     Node_List wl;
1928     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1929       Node* m = i.get();
1930       if (m->is_MergeMem() && !wl.contains(m)) {
1931         wl.push(m);
1932       }
1933     }
1934     while (wl.size()  > 0) {
1935       _gvn.transform(wl.pop());
1936     }
1937   }
1938 
1939   if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1940     replaced_nodes.apply(C, final_ctl);
1941   }
1942   if (!ex_ctl->is_top() && do_replaced_nodes) {
1943     replaced_nodes_exception.apply(C, ex_ctl);
1944   }
1945 }
1946 
1947 
1948 //------------------------------increment_counter------------------------------
1949 // for statistics: increment a VM counter by 1
1950 
1951 void GraphKit::increment_counter(address counter_addr) {
1952   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1953   increment_counter(adr1);
1954 }
1955 
1956 void GraphKit::increment_counter(Node* counter_addr) {
1957   int adr_type = Compile::AliasIdxRaw;
1958   Node* ctrl = control();
1959   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1960   Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1961   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1962 }
1963 
1964 
1965 //------------------------------uncommon_trap----------------------------------
1966 // Bail out to the interpreter in mid-method.  Implemented by calling the
1967 // uncommon_trap blob.  This helper function inserts a runtime call with the
1968 // right debug info.
1969 void GraphKit::uncommon_trap(int trap_request,
1970                              ciKlass* klass, const char* comment,
1971                              bool must_throw,
1972                              bool keep_exact_action) {
1973   if (failing())  stop();
1974   if (stopped())  return; // trap reachable?
1975 
1976   // Note:  If ProfileTraps is true, and if a deopt. actually
1977   // occurs here, the runtime will make sure an MDO exists.  There is
1978   // no need to call method()->ensure_method_data() at this point.
1979 
1980   // Set the stack pointer to the right value for reexecution:
1981   set_sp(reexecute_sp());
1982 
1983 #ifdef ASSERT
1984   if (!must_throw) {
1985     // Make sure the stack has at least enough depth to execute
1986     // the current bytecode.
1987     int inputs, ignored_depth;
1988     if (compute_stack_effects(inputs, ignored_depth)) {
1989       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1990              Bytecodes::name(java_bc()), sp(), inputs);
1991     }
1992   }
1993 #endif
1994 
1995   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1996   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1997 
1998   switch (action) {
1999   case Deoptimization::Action_maybe_recompile:
2000   case Deoptimization::Action_reinterpret:
2001     // Temporary fix for 6529811 to allow virtual calls to be sure they
2002     // get the chance to go from mono->bi->mega
2003     if (!keep_exact_action &&
2004         Deoptimization::trap_request_index(trap_request) < 0 &&
2005         too_many_recompiles(reason)) {
2006       // This BCI is causing too many recompilations.
2007       if (C->log() != NULL) {
2008         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2009                 Deoptimization::trap_reason_name(reason),
2010                 Deoptimization::trap_action_name(action));
2011       }
2012       action = Deoptimization::Action_none;
2013       trap_request = Deoptimization::make_trap_request(reason, action);
2014     } else {
2015       C->set_trap_can_recompile(true);
2016     }
2017     break;
2018   case Deoptimization::Action_make_not_entrant:
2019     C->set_trap_can_recompile(true);
2020     break;
2021 #ifdef ASSERT
2022   case Deoptimization::Action_none:
2023   case Deoptimization::Action_make_not_compilable:
2024     break;
2025   default:
2026     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2027     break;
2028 #endif
2029   }
2030 
2031   if (TraceOptoParse) {
2032     char buf[100];
2033     tty->print_cr("Uncommon trap %s at bci:%d",
2034                   Deoptimization::format_trap_request(buf, sizeof(buf),
2035                                                       trap_request), bci());
2036   }
2037 
2038   CompileLog* log = C->log();
2039   if (log != NULL) {
2040     int kid = (klass == NULL)? -1: log->identify(klass);
2041     log->begin_elem("uncommon_trap bci='%d'", bci());
2042     char buf[100];
2043     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2044                                                           trap_request));
2045     if (kid >= 0)         log->print(" klass='%d'", kid);
2046     if (comment != NULL)  log->print(" comment='%s'", comment);
2047     log->end_elem();
2048   }
2049 
2050   // Make sure any guarding test views this path as very unlikely
2051   Node *i0 = control()->in(0);
2052   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2053     IfNode *iff = i0->as_If();
2054     float f = iff->_prob;   // Get prob
2055     if (control()->Opcode() == Op_IfTrue) {
2056       if (f > PROB_UNLIKELY_MAG(4))
2057         iff->_prob = PROB_MIN;
2058     } else {
2059       if (f < PROB_LIKELY_MAG(4))
2060         iff->_prob = PROB_MAX;
2061     }
2062   }
2063 
2064   // Clear out dead values from the debug info.
2065   kill_dead_locals();
2066 
2067   // Make sure all value types are allocated before calling uncommon_trap
2068   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
2069     SafePointNode* map = jvms->map();
2070     // Search for value type locals
2071     for (int i = 0; i < jvms->loc_size(); ++i) {
2072       Node* local = map->local(jvms, i);
2073       if (local->is_ValueType()) {
2074         // Allocate value type
2075         local = local->as_ValueType()->store_to_memory(this);
2076         map->set_local(jvms, i, local);
2077       }
2078     }
2079   }
2080 
2081   // Now insert the uncommon trap subroutine call
2082   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2083   const TypePtr* no_memory_effects = NULL;
2084   // Pass the index of the class to be loaded
2085   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2086                                  (must_throw ? RC_MUST_THROW : 0),
2087                                  OptoRuntime::uncommon_trap_Type(),
2088                                  call_addr, "uncommon_trap", no_memory_effects,
2089                                  intcon(trap_request));
2090   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2091          "must extract request correctly from the graph");
2092   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2093 
2094   call->set_req(TypeFunc::ReturnAdr, returnadr());
2095   // The debug info is the only real input to this call.
2096 
2097   // Halt-and-catch fire here.  The above call should never return!
2098   HaltNode* halt = new HaltNode(control(), frameptr());
2099   _gvn.set_type_bottom(halt);
2100   root()->add_req(halt);
2101 
2102   stop_and_kill_map();
2103 }
2104 
2105 
2106 //--------------------------just_allocated_object------------------------------
2107 // Report the object that was just allocated.
2108 // It must be the case that there are no intervening safepoints.
2109 // We use this to determine if an object is so "fresh" that
2110 // it does not require card marks.
2111 Node* GraphKit::just_allocated_object(Node* current_control) {
2112   if (C->recent_alloc_ctl() == current_control)
2113     return C->recent_alloc_obj();
2114   return NULL;
2115 }
2116 
2117 
2118 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2119   // (Note:  TypeFunc::make has a cache that makes this fast.)
2120   const TypeFunc* tf    = TypeFunc::make(dest_method);
2121   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2122   for (int j = 0; j < nargs; j++) {
2123     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2124     if( targ->basic_type() == T_DOUBLE ) {
2125       // If any parameters are doubles, they must be rounded before
2126       // the call, dstore_rounding does gvn.transform
2127       Node *arg = argument(j);
2128       arg = dstore_rounding(arg);
2129       set_argument(j, arg);
2130     }
2131   }
2132 }
2133 
2134 /**
2135  * Record profiling data exact_kls for Node n with the type system so
2136  * that it can propagate it (speculation)
2137  *
2138  * @param n          node that the type applies to
2139  * @param exact_kls  type from profiling
2140  * @param maybe_null did profiling see null?
2141  *
2142  * @return           node with improved type
2143  */
2144 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) {
2145   const Type* current_type = _gvn.type(n);
2146   assert(UseTypeSpeculation, "type speculation must be on");
2147 
2148   const TypePtr* speculative = current_type->speculative();
2149 
2150   // Should the klass from the profile be recorded in the speculative type?
2151   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2152     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2153     const TypeOopPtr* xtype = tklass->as_instance_type();
2154     assert(xtype->klass_is_exact(), "Should be exact");
2155     // Any reason to believe n is not null (from this profiling or a previous one)?
2156     const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2157     // record the new speculative type's depth
2158     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2159     speculative = speculative->with_inline_depth(jvms()->depth());
2160   } else if (current_type->would_improve_ptr(maybe_null)) {
2161     // Profiling report that null was never seen so we can change the
2162     // speculative type to non null ptr.
2163     assert(!maybe_null, "nothing to improve");
2164     if (speculative == NULL) {
2165       speculative = TypePtr::NOTNULL;
2166     } else {
2167       const TypePtr* ptr = TypePtr::NOTNULL;
2168       speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2169     }
2170   }
2171 
2172   if (speculative != current_type->speculative()) {
2173     // Build a type with a speculative type (what we think we know
2174     // about the type but will need a guard when we use it)
2175     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2176     // We're changing the type, we need a new CheckCast node to carry
2177     // the new type. The new type depends on the control: what
2178     // profiling tells us is only valid from here as far as we can
2179     // tell.
2180     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2181     cast = _gvn.transform(cast);
2182     replace_in_map(n, cast);
2183     n = cast;
2184   }
2185 
2186   return n;
2187 }
2188 
2189 /**
2190  * Record profiling data from receiver profiling at an invoke with the
2191  * type system so that it can propagate it (speculation)
2192  *
2193  * @param n  receiver node
2194  *
2195  * @return   node with improved type
2196  */
2197 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2198   if (!UseTypeSpeculation) {
2199     return n;
2200   }
2201   ciKlass* exact_kls = profile_has_unique_klass();
2202   bool maybe_null = true;
2203   if (java_bc() == Bytecodes::_checkcast ||
2204       java_bc() == Bytecodes::_instanceof ||
2205       java_bc() == Bytecodes::_aastore) {
2206     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2207     bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen();
2208   }
2209   return record_profile_for_speculation(n, exact_kls, maybe_null);
2210   return n;
2211 }
2212 
2213 /**
2214  * Record profiling data from argument profiling at an invoke with the
2215  * type system so that it can propagate it (speculation)
2216  *
2217  * @param dest_method  target method for the call
2218  * @param bc           what invoke bytecode is this?
2219  */
2220 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2221   if (!UseTypeSpeculation) {
2222     return;
2223   }
2224   const TypeFunc* tf    = TypeFunc::make(dest_method);
2225   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2226   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2227   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2228     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2229     if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2230       bool maybe_null = true;
2231       ciKlass* better_type = NULL;
2232       if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) {
2233         record_profile_for_speculation(argument(j), better_type, maybe_null);
2234       }
2235       i++;
2236     }
2237   }
2238 }
2239 
2240 /**
2241  * Record profiling data from parameter profiling at an invoke with
2242  * the type system so that it can propagate it (speculation)
2243  */
2244 void GraphKit::record_profiled_parameters_for_speculation() {
2245   if (!UseTypeSpeculation) {
2246     return;
2247   }
2248   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2249     if (_gvn.type(local(i))->isa_oopptr()) {
2250       bool maybe_null = true;
2251       ciKlass* better_type = NULL;
2252       if (method()->parameter_profiled_type(j, better_type, maybe_null)) {
2253         record_profile_for_speculation(local(i), better_type, maybe_null);
2254       }
2255       j++;
2256     }
2257   }
2258 }
2259 
2260 /**
2261  * Record profiling data from return value profiling at an invoke with
2262  * the type system so that it can propagate it (speculation)
2263  */
2264 void GraphKit::record_profiled_return_for_speculation() {
2265   if (!UseTypeSpeculation) {
2266     return;
2267   }
2268   bool maybe_null = true;
2269   ciKlass* better_type = NULL;
2270   if (method()->return_profiled_type(bci(), better_type, maybe_null)) {
2271     // If profiling reports a single type for the return value,
2272     // feed it to the type system so it can propagate it as a
2273     // speculative type
2274     record_profile_for_speculation(stack(sp()-1), better_type, maybe_null);
2275   }
2276 }
2277 
2278 void GraphKit::round_double_result(ciMethod* dest_method) {
2279   // A non-strict method may return a double value which has an extended
2280   // exponent, but this must not be visible in a caller which is 'strict'
2281   // If a strict caller invokes a non-strict callee, round a double result
2282 
2283   BasicType result_type = dest_method->return_type()->basic_type();
2284   assert( method() != NULL, "must have caller context");
2285   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2286     // Destination method's return value is on top of stack
2287     // dstore_rounding() does gvn.transform
2288     Node *result = pop_pair();
2289     result = dstore_rounding(result);
2290     push_pair(result);
2291   }
2292 }
2293 
2294 // rounding for strict float precision conformance
2295 Node* GraphKit::precision_rounding(Node* n) {
2296   return UseStrictFP && _method->flags().is_strict()
2297     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2298     ? _gvn.transform( new RoundFloatNode(0, n) )
2299     : n;
2300 }
2301 
2302 // rounding for strict double precision conformance
2303 Node* GraphKit::dprecision_rounding(Node *n) {
2304   return UseStrictFP && _method->flags().is_strict()
2305     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2306     ? _gvn.transform( new RoundDoubleNode(0, n) )
2307     : n;
2308 }
2309 
2310 // rounding for non-strict double stores
2311 Node* GraphKit::dstore_rounding(Node* n) {
2312   return Matcher::strict_fp_requires_explicit_rounding
2313     && UseSSE <= 1
2314     ? _gvn.transform( new RoundDoubleNode(0, n) )
2315     : n;
2316 }
2317 
2318 //=============================================================================
2319 // Generate a fast path/slow path idiom.  Graph looks like:
2320 // [foo] indicates that 'foo' is a parameter
2321 //
2322 //              [in]     NULL
2323 //                 \    /
2324 //                  CmpP
2325 //                  Bool ne
2326 //                   If
2327 //                  /  \
2328 //              True    False-<2>
2329 //              / |
2330 //             /  cast_not_null
2331 //           Load  |    |   ^
2332 //        [fast_test]   |   |
2333 // gvn to   opt_test    |   |
2334 //          /    \      |  <1>
2335 //      True     False  |
2336 //        |         \\  |
2337 //   [slow_call]     \[fast_result]
2338 //    Ctl   Val       \      \
2339 //     |               \      \
2340 //    Catch       <1>   \      \
2341 //   /    \        ^     \      \
2342 //  Ex    No_Ex    |      \      \
2343 //  |       \   \  |       \ <2>  \
2344 //  ...      \  [slow_res] |  |    \   [null_result]
2345 //            \         \--+--+---  |  |
2346 //             \           | /    \ | /
2347 //              --------Region     Phi
2348 //
2349 //=============================================================================
2350 // Code is structured as a series of driver functions all called 'do_XXX' that
2351 // call a set of helper functions.  Helper functions first, then drivers.
2352 
2353 //------------------------------null_check_oop---------------------------------
2354 // Null check oop.  Set null-path control into Region in slot 3.
2355 // Make a cast-not-nullness use the other not-null control.  Return cast.
2356 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2357                                bool never_see_null,
2358                                bool safe_for_replace,
2359                                bool speculative) {
2360   // Initial NULL check taken path
2361   (*null_control) = top();
2362   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2363 
2364   // Generate uncommon_trap:
2365   if (never_see_null && (*null_control) != top()) {
2366     // If we see an unexpected null at a check-cast we record it and force a
2367     // recompile; the offending check-cast will be compiled to handle NULLs.
2368     // If we see more than one offending BCI, then all checkcasts in the
2369     // method will be compiled to handle NULLs.
2370     PreserveJVMState pjvms(this);
2371     set_control(*null_control);
2372     replace_in_map(value, null());
2373     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2374     uncommon_trap(reason,
2375                   Deoptimization::Action_make_not_entrant);
2376     (*null_control) = top();    // NULL path is dead
2377   }
2378   if ((*null_control) == top() && safe_for_replace) {
2379     replace_in_map(value, cast);
2380   }
2381 
2382   // Cast away null-ness on the result
2383   return cast;
2384 }
2385 
2386 //------------------------------opt_iff----------------------------------------
2387 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2388 // Return slow-path control.
2389 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2390   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2391 
2392   // Fast path taken; set region slot 2
2393   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2394   region->init_req(2,fast_taken); // Capture fast-control
2395 
2396   // Fast path not-taken, i.e. slow path
2397   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2398   return slow_taken;
2399 }
2400 
2401 //-----------------------------make_runtime_call-------------------------------
2402 Node* GraphKit::make_runtime_call(int flags,
2403                                   const TypeFunc* call_type, address call_addr,
2404                                   const char* call_name,
2405                                   const TypePtr* adr_type,
2406                                   // The following parms are all optional.
2407                                   // The first NULL ends the list.
2408                                   Node* parm0, Node* parm1,
2409                                   Node* parm2, Node* parm3,
2410                                   Node* parm4, Node* parm5,
2411                                   Node* parm6, Node* parm7) {
2412   // Slow-path call
2413   bool is_leaf = !(flags & RC_NO_LEAF);
2414   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2415   if (call_name == NULL) {
2416     assert(!is_leaf, "must supply name for leaf");
2417     call_name = OptoRuntime::stub_name(call_addr);
2418   }
2419   CallNode* call;
2420   if (!is_leaf) {
2421     call = new CallStaticJavaNode(call_type, call_addr, call_name,
2422                                            bci(), adr_type);
2423   } else if (flags & RC_NO_FP) {
2424     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2425   } else {
2426     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2427   }
2428 
2429   // The following is similar to set_edges_for_java_call,
2430   // except that the memory effects of the call are restricted to AliasIdxRaw.
2431 
2432   // Slow path call has no side-effects, uses few values
2433   bool wide_in  = !(flags & RC_NARROW_MEM);
2434   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2435 
2436   Node* prev_mem = NULL;
2437   if (wide_in) {
2438     prev_mem = set_predefined_input_for_runtime_call(call);
2439   } else {
2440     assert(!wide_out, "narrow in => narrow out");
2441     Node* narrow_mem = memory(adr_type);
2442     prev_mem = reset_memory();
2443     map()->set_memory(narrow_mem);
2444     set_predefined_input_for_runtime_call(call);
2445   }
2446 
2447   // Hook each parm in order.  Stop looking at the first NULL.
2448   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2449   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2450   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2451   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2452   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2453   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2454   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2455   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2456     /* close each nested if ===> */  } } } } } } } }
2457   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2458 
2459   if (!is_leaf) {
2460     // Non-leaves can block and take safepoints:
2461     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2462   }
2463   // Non-leaves can throw exceptions:
2464   if (has_io) {
2465     call->set_req(TypeFunc::I_O, i_o());
2466   }
2467 
2468   if (flags & RC_UNCOMMON) {
2469     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2470     // (An "if" probability corresponds roughly to an unconditional count.
2471     // Sort of.)
2472     call->set_cnt(PROB_UNLIKELY_MAG(4));
2473   }
2474 
2475   Node* c = _gvn.transform(call);
2476   assert(c == call, "cannot disappear");
2477 
2478   if (wide_out) {
2479     // Slow path call has full side-effects.
2480     set_predefined_output_for_runtime_call(call);
2481   } else {
2482     // Slow path call has few side-effects, and/or sets few values.
2483     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2484   }
2485 
2486   if (has_io) {
2487     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2488   }
2489   return call;
2490 
2491 }
2492 
2493 //------------------------------merge_memory-----------------------------------
2494 // Merge memory from one path into the current memory state.
2495 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2496   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2497     Node* old_slice = mms.force_memory();
2498     Node* new_slice = mms.memory2();
2499     if (old_slice != new_slice) {
2500       PhiNode* phi;
2501       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2502         if (mms.is_empty()) {
2503           // clone base memory Phi's inputs for this memory slice
2504           assert(old_slice == mms.base_memory(), "sanity");
2505           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2506           _gvn.set_type(phi, Type::MEMORY);
2507           for (uint i = 1; i < phi->req(); i++) {
2508             phi->init_req(i, old_slice->in(i));
2509           }
2510         } else {
2511           phi = old_slice->as_Phi(); // Phi was generated already
2512         }
2513       } else {
2514         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2515         _gvn.set_type(phi, Type::MEMORY);
2516       }
2517       phi->set_req(new_path, new_slice);
2518       mms.set_memory(phi);
2519     }
2520   }
2521 }
2522 
2523 //------------------------------make_slow_call_ex------------------------------
2524 // Make the exception handler hookups for the slow call
2525 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2526   if (stopped())  return;
2527 
2528   // Make a catch node with just two handlers:  fall-through and catch-all
2529   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2530   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2531   Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2532   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2533 
2534   { PreserveJVMState pjvms(this);
2535     set_control(excp);
2536     set_i_o(i_o);
2537 
2538     if (excp != top()) {
2539       if (deoptimize) {
2540         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2541         uncommon_trap(Deoptimization::Reason_unhandled,
2542                       Deoptimization::Action_none);
2543       } else {
2544         // Create an exception state also.
2545         // Use an exact type if the caller has specified a specific exception.
2546         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2547         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2548         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2549       }
2550     }
2551   }
2552 
2553   // Get the no-exception control from the CatchNode.
2554   set_control(norm);
2555 }
2556 
2557 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2558   Node* cmp = NULL;
2559   switch(bt) {
2560   case T_INT: cmp = new CmpINode(in1, in2); break;
2561   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2562   default: fatal("unexpected comparison type %s", type2name(bt));
2563   }
2564   gvn->transform(cmp);
2565   Node* bol = gvn->transform(new BoolNode(cmp, test));
2566   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2567   gvn->transform(iff);
2568   if (!bol->is_Con()) gvn->record_for_igvn(iff);
2569   return iff;
2570 }
2571 
2572 
2573 //-------------------------------gen_subtype_check-----------------------------
2574 // Generate a subtyping check.  Takes as input the subtype and supertype.
2575 // Returns 2 values: sets the default control() to the true path and returns
2576 // the false path.  Only reads invariant memory; sets no (visible) memory.
2577 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2578 // but that's not exposed to the optimizer.  This call also doesn't take in an
2579 // Object; if you wish to check an Object you need to load the Object's class
2580 // prior to coming here.
2581 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2582   Compile* C = gvn->C;
2583 
2584   if ((*ctrl)->is_top()) {
2585     return C->top();
2586   }
2587 
2588   // Fast check for identical types, perhaps identical constants.
2589   // The types can even be identical non-constants, in cases
2590   // involving Array.newInstance, Object.clone, etc.
2591   if (subklass == superklass)
2592     return C->top();             // false path is dead; no test needed.
2593 
2594   if (gvn->type(superklass)->singleton()) {
2595     ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2596     ciKlass* subk   = gvn->type(subklass)->is_klassptr()->klass();
2597 
2598     // In the common case of an exact superklass, try to fold up the
2599     // test before generating code.  You may ask, why not just generate
2600     // the code and then let it fold up?  The answer is that the generated
2601     // code will necessarily include null checks, which do not always
2602     // completely fold away.  If they are also needless, then they turn
2603     // into a performance loss.  Example:
2604     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2605     // Here, the type of 'fa' is often exact, so the store check
2606     // of fa[1]=x will fold up, without testing the nullness of x.
2607     switch (C->static_subtype_check(superk, subk)) {
2608     case Compile::SSC_always_false:
2609       {
2610         Node* always_fail = *ctrl;
2611         *ctrl = gvn->C->top();
2612         return always_fail;
2613       }
2614     case Compile::SSC_always_true:
2615       return C->top();
2616     case Compile::SSC_easy_test:
2617       {
2618         // Just do a direct pointer compare and be done.
2619         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2620         *ctrl = gvn->transform(new IfTrueNode(iff));
2621         return gvn->transform(new IfFalseNode(iff));
2622       }
2623     case Compile::SSC_full_test:
2624       break;
2625     default:
2626       ShouldNotReachHere();
2627     }
2628   }
2629 
2630   // %%% Possible further optimization:  Even if the superklass is not exact,
2631   // if the subklass is the unique subtype of the superklass, the check
2632   // will always succeed.  We could leave a dependency behind to ensure this.
2633 
2634   // First load the super-klass's check-offset
2635   Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2636   Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2637   Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2638   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2639   bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2640 
2641   // Load from the sub-klass's super-class display list, or a 1-word cache of
2642   // the secondary superclass list, or a failing value with a sentinel offset
2643   // if the super-klass is an interface or exceptionally deep in the Java
2644   // hierarchy and we have to scan the secondary superclass list the hard way.
2645   // Worst-case type is a little odd: NULL is allowed as a result (usually
2646   // klass loads can never produce a NULL).
2647   Node *chk_off_X = chk_off;
2648 #ifdef _LP64
2649   chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2650 #endif
2651   Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2652   // For some types like interfaces the following loadKlass is from a 1-word
2653   // cache which is mutable so can't use immutable memory.  Other
2654   // types load from the super-class display table which is immutable.
2655   m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2656   Node *kmem = might_be_cache ? m : C->immutable_memory();
2657   Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2658 
2659   // Compile speed common case: ARE a subtype and we canNOT fail
2660   if( superklass == nkls )
2661     return C->top();             // false path is dead; no test needed.
2662 
2663   // See if we get an immediate positive hit.  Happens roughly 83% of the
2664   // time.  Test to see if the value loaded just previously from the subklass
2665   // is exactly the superklass.
2666   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2667   Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2668   *ctrl = gvn->transform(new IfFalseNode(iff1));
2669 
2670   // Compile speed common case: Check for being deterministic right now.  If
2671   // chk_off is a constant and not equal to cacheoff then we are NOT a
2672   // subklass.  In this case we need exactly the 1 test above and we can
2673   // return those results immediately.
2674   if (!might_be_cache) {
2675     Node* not_subtype_ctrl = *ctrl;
2676     *ctrl = iftrue1; // We need exactly the 1 test above
2677     return not_subtype_ctrl;
2678   }
2679 
2680   // Gather the various success & failures here
2681   RegionNode *r_ok_subtype = new RegionNode(4);
2682   gvn->record_for_igvn(r_ok_subtype);
2683   RegionNode *r_not_subtype = new RegionNode(3);
2684   gvn->record_for_igvn(r_not_subtype);
2685 
2686   r_ok_subtype->init_req(1, iftrue1);
2687 
2688   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2689   // is roughly 63% of the remaining cases).  Test to see if the loaded
2690   // check-offset points into the subklass display list or the 1-element
2691   // cache.  If it points to the display (and NOT the cache) and the display
2692   // missed then it's not a subtype.
2693   Node *cacheoff = gvn->intcon(cacheoff_con);
2694   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2695   r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2696   *ctrl = gvn->transform(new IfFalseNode(iff2));
2697 
2698   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2699   // No performance impact (too rare) but allows sharing of secondary arrays
2700   // which has some footprint reduction.
2701   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2702   r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2703   *ctrl = gvn->transform(new IfFalseNode(iff3));
2704 
2705   // -- Roads not taken here: --
2706   // We could also have chosen to perform the self-check at the beginning
2707   // of this code sequence, as the assembler does.  This would not pay off
2708   // the same way, since the optimizer, unlike the assembler, can perform
2709   // static type analysis to fold away many successful self-checks.
2710   // Non-foldable self checks work better here in second position, because
2711   // the initial primary superclass check subsumes a self-check for most
2712   // types.  An exception would be a secondary type like array-of-interface,
2713   // which does not appear in its own primary supertype display.
2714   // Finally, we could have chosen to move the self-check into the
2715   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2716   // dependent manner.  But it is worthwhile to have the check here,
2717   // where it can be perhaps be optimized.  The cost in code space is
2718   // small (register compare, branch).
2719 
2720   // Now do a linear scan of the secondary super-klass array.  Again, no real
2721   // performance impact (too rare) but it's gotta be done.
2722   // Since the code is rarely used, there is no penalty for moving it
2723   // out of line, and it can only improve I-cache density.
2724   // The decision to inline or out-of-line this final check is platform
2725   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2726   Node* psc = gvn->transform(
2727     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2728 
2729   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2730   r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2731   r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2732 
2733   // Return false path; set default control to true path.
2734   *ctrl = gvn->transform(r_ok_subtype);
2735   return gvn->transform(r_not_subtype);
2736 }
2737 
2738 // Profile-driven exact type check:
2739 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2740                                     float prob,
2741                                     Node* *casted_receiver) {
2742   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2743   Node* recv_klass = load_object_klass(receiver);
2744   Node* want_klass = makecon(tklass);
2745   Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2746   Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2747   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2748   set_control( _gvn.transform( new IfTrueNode (iff) ));
2749   Node* fail = _gvn.transform( new IfFalseNode(iff) );
2750 
2751   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2752   assert(recv_xtype->klass_is_exact(), "");
2753 
2754   // Subsume downstream occurrences of receiver with a cast to
2755   // recv_xtype, since now we know what the type will be.
2756   Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2757   (*casted_receiver) = _gvn.transform(cast);
2758   // (User must make the replace_in_map call.)
2759 
2760   return fail;
2761 }
2762 
2763 
2764 //------------------------------seems_never_null-------------------------------
2765 // Use null_seen information if it is available from the profile.
2766 // If we see an unexpected null at a type check we record it and force a
2767 // recompile; the offending check will be recompiled to handle NULLs.
2768 // If we see several offending BCIs, then all checks in the
2769 // method will be recompiled.
2770 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2771   speculating = !_gvn.type(obj)->speculative_maybe_null();
2772   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2773   if (UncommonNullCast               // Cutout for this technique
2774       && obj != null()               // And not the -Xcomp stupid case?
2775       && !too_many_traps(reason)
2776       ) {
2777     if (speculating) {
2778       return true;
2779     }
2780     if (data == NULL)
2781       // Edge case:  no mature data.  Be optimistic here.
2782       return true;
2783     // If the profile has not seen a null, assume it won't happen.
2784     assert(java_bc() == Bytecodes::_checkcast ||
2785            java_bc() == Bytecodes::_instanceof ||
2786            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2787     return !data->as_BitData()->null_seen();
2788   }
2789   speculating = false;
2790   return false;
2791 }
2792 
2793 //------------------------maybe_cast_profiled_receiver-------------------------
2794 // If the profile has seen exactly one type, narrow to exactly that type.
2795 // Subsequent type checks will always fold up.
2796 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2797                                              ciKlass* require_klass,
2798                                              ciKlass* spec_klass,
2799                                              bool safe_for_replace) {
2800   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2801 
2802   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2803 
2804   // Make sure we haven't already deoptimized from this tactic.
2805   if (too_many_traps(reason) || too_many_recompiles(reason))
2806     return NULL;
2807 
2808   // (No, this isn't a call, but it's enough like a virtual call
2809   // to use the same ciMethod accessor to get the profile info...)
2810   // If we have a speculative type use it instead of profiling (which
2811   // may not help us)
2812   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2813   if (exact_kls != NULL) {// no cast failures here
2814     if (require_klass == NULL ||
2815         C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2816       // If we narrow the type to match what the type profile sees or
2817       // the speculative type, we can then remove the rest of the
2818       // cast.
2819       // This is a win, even if the exact_kls is very specific,
2820       // because downstream operations, such as method calls,
2821       // will often benefit from the sharper type.
2822       Node* exact_obj = not_null_obj; // will get updated in place...
2823       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2824                                             &exact_obj);
2825       { PreserveJVMState pjvms(this);
2826         set_control(slow_ctl);
2827         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2828       }
2829       if (safe_for_replace) {
2830         replace_in_map(not_null_obj, exact_obj);
2831       }
2832       return exact_obj;
2833     }
2834     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2835   }
2836 
2837   return NULL;
2838 }
2839 
2840 /**
2841  * Cast obj to type and emit guard unless we had too many traps here
2842  * already
2843  *
2844  * @param obj       node being casted
2845  * @param type      type to cast the node to
2846  * @param not_null  true if we know node cannot be null
2847  */
2848 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2849                                         ciKlass* type,
2850                                         bool not_null) {
2851   if (stopped()) {
2852     return obj;
2853   }
2854 
2855   // type == NULL if profiling tells us this object is always null
2856   if (type != NULL) {
2857     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2858     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2859 
2860     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2861         !too_many_traps(class_reason) &&
2862         !too_many_recompiles(class_reason)) {
2863       Node* not_null_obj = NULL;
2864       // not_null is true if we know the object is not null and
2865       // there's no need for a null check
2866       if (!not_null) {
2867         Node* null_ctl = top();
2868         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2869         assert(null_ctl->is_top(), "no null control here");
2870       } else {
2871         not_null_obj = obj;
2872       }
2873 
2874       Node* exact_obj = not_null_obj;
2875       ciKlass* exact_kls = type;
2876       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2877                                             &exact_obj);
2878       {
2879         PreserveJVMState pjvms(this);
2880         set_control(slow_ctl);
2881         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2882       }
2883       replace_in_map(not_null_obj, exact_obj);
2884       obj = exact_obj;
2885     }
2886   } else {
2887     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2888         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2889       Node* exact_obj = null_assert(obj);
2890       replace_in_map(obj, exact_obj);
2891       obj = exact_obj;
2892     }
2893   }
2894   return obj;
2895 }
2896 
2897 //-------------------------------gen_instanceof--------------------------------
2898 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2899 // and the reflective instance-of call.
2900 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2901   kill_dead_locals();           // Benefit all the uncommon traps
2902   assert( !stopped(), "dead parse path should be checked in callers" );
2903   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2904          "must check for not-null not-dead klass in callers");
2905 
2906   // Make the merge point
2907   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2908   RegionNode* region = new RegionNode(PATH_LIMIT);
2909   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
2910   C->set_has_split_ifs(true); // Has chance for split-if optimization
2911 
2912   ciProfileData* data = NULL;
2913   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2914     data = method()->method_data()->bci_to_data(bci());
2915   }
2916   bool speculative_not_null = false;
2917   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2918                          && seems_never_null(obj, data, speculative_not_null));
2919 
2920   // Null check; get casted pointer; set region slot 3
2921   Node* null_ctl = top();
2922   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2923 
2924   // If not_null_obj is dead, only null-path is taken
2925   if (stopped()) {              // Doing instance-of on a NULL?
2926     set_control(null_ctl);
2927     return intcon(0);
2928   }
2929   region->init_req(_null_path, null_ctl);
2930   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2931   if (null_ctl == top()) {
2932     // Do this eagerly, so that pattern matches like is_diamond_phi
2933     // will work even during parsing.
2934     assert(_null_path == PATH_LIMIT-1, "delete last");
2935     region->del_req(_null_path);
2936     phi   ->del_req(_null_path);
2937   }
2938 
2939   // Do we know the type check always succeed?
2940   bool known_statically = false;
2941   if (_gvn.type(superklass)->singleton()) {
2942     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2943     ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2944     if (subk != NULL && subk->is_loaded()) {
2945       int static_res = C->static_subtype_check(superk, subk);
2946       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2947     }
2948   }
2949 
2950   if (known_statically && UseTypeSpeculation) {
2951     // If we know the type check always succeeds then we don't use the
2952     // profiling data at this bytecode. Don't lose it, feed it to the
2953     // type system as a speculative type.
2954     not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2955   } else {
2956     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2957     // We may not have profiling here or it may not help us. If we
2958     // have a speculative type use it to perform an exact cast.
2959     ciKlass* spec_obj_type = obj_type->speculative_type();
2960     if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2961       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2962       if (stopped()) {            // Profile disagrees with this path.
2963         set_control(null_ctl);    // Null is the only remaining possibility.
2964         return intcon(0);
2965       }
2966       if (cast_obj != NULL) {
2967         not_null_obj = cast_obj;
2968       }
2969     }
2970   }
2971 
2972   // Load the object's klass
2973   Node* obj_klass = load_object_klass(not_null_obj);
2974 
2975   // Generate the subtype check
2976   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2977 
2978   // Plug in the success path to the general merge in slot 1.
2979   region->init_req(_obj_path, control());
2980   phi   ->init_req(_obj_path, intcon(1));
2981 
2982   // Plug in the failing path to the general merge in slot 2.
2983   region->init_req(_fail_path, not_subtype_ctrl);
2984   phi   ->init_req(_fail_path, intcon(0));
2985 
2986   // Return final merged results
2987   set_control( _gvn.transform(region) );
2988   record_for_igvn(region);
2989   return _gvn.transform(phi);
2990 }
2991 
2992 //-------------------------------gen_checkcast---------------------------------
2993 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
2994 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
2995 // uncommon-trap paths work.  Adjust stack after this call.
2996 // If failure_control is supplied and not null, it is filled in with
2997 // the control edge for the cast failure.  Otherwise, an appropriate
2998 // uncommon trap or exception is thrown.
2999 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3000                               Node* *failure_control) {
3001   kill_dead_locals();           // Benefit all the uncommon traps
3002   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3003   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3004 
3005   // Fast cutout:  Check the case that the cast is vacuously true.
3006   // This detects the common cases where the test will short-circuit
3007   // away completely.  We do this before we perform the null check,
3008   // because if the test is going to turn into zero code, we don't
3009   // want a residual null check left around.  (Causes a slowdown,
3010   // for example, in some objArray manipulations, such as a[i]=a[j].)
3011   if (tk->singleton()) {
3012     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3013     if (objtp != NULL && objtp->klass() != NULL) {
3014       switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3015       case Compile::SSC_always_true:
3016         // If we know the type check always succeed then we don't use
3017         // the profiling data at this bytecode. Don't lose it, feed it
3018         // to the type system as a speculative type.
3019         return record_profiled_receiver_for_speculation(obj);
3020       case Compile::SSC_always_false:
3021         // It needs a null check because a null will *pass* the cast check.
3022         // A non-null value will always produce an exception.
3023         return null_assert(obj);
3024       }
3025     }
3026   }
3027 
3028   ciProfileData* data = NULL;
3029   bool safe_for_replace = false;
3030   if (failure_control == NULL) {        // use MDO in regular case only
3031     assert(java_bc() == Bytecodes::_aastore ||
3032            java_bc() == Bytecodes::_checkcast,
3033            "interpreter profiles type checks only for these BCs");
3034     data = method()->method_data()->bci_to_data(bci());
3035     safe_for_replace = true;
3036   }
3037 
3038   // Make the merge point
3039   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3040   RegionNode* region = new RegionNode(PATH_LIMIT);
3041   Node*       phi    = new PhiNode(region, toop);
3042   C->set_has_split_ifs(true); // Has chance for split-if optimization
3043 
3044   // Use null-cast information if it is available
3045   bool speculative_not_null = false;
3046   bool never_see_null = ((failure_control == NULL)  // regular case only
3047                          && seems_never_null(obj, data, speculative_not_null));
3048 
3049   // Null check; get casted pointer; set region slot 3
3050   Node* null_ctl = top();
3051   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3052 
3053   // If not_null_obj is dead, only null-path is taken
3054   if (stopped()) {              // Doing instance-of on a NULL?
3055     set_control(null_ctl);
3056     return null();
3057   }
3058   region->init_req(_null_path, null_ctl);
3059   phi   ->init_req(_null_path, null());  // Set null path value
3060   if (null_ctl == top()) {
3061     // Do this eagerly, so that pattern matches like is_diamond_phi
3062     // will work even during parsing.
3063     assert(_null_path == PATH_LIMIT-1, "delete last");
3064     region->del_req(_null_path);
3065     phi   ->del_req(_null_path);
3066   }
3067 
3068   Node* cast_obj = NULL;
3069   if (tk->klass_is_exact()) {
3070     // The following optimization tries to statically cast the speculative type of the object
3071     // (for example obtained during profiling) to the type of the superklass and then do a
3072     // dynamic check that the type of the object is what we expect. To work correctly
3073     // for checkcast and aastore the type of superklass should be exact.
3074     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3075     // We may not have profiling here or it may not help us. If we have
3076     // a speculative type use it to perform an exact cast.
3077     ciKlass* spec_obj_type = obj_type->speculative_type();
3078     if (spec_obj_type != NULL || data != NULL) {
3079       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3080       if (cast_obj != NULL) {
3081         if (failure_control != NULL) // failure is now impossible
3082           (*failure_control) = top();
3083         // adjust the type of the phi to the exact klass:
3084         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3085       }
3086     }
3087   }
3088 
3089   if (cast_obj == NULL) {
3090     // Load the object's klass
3091     Node* obj_klass = load_object_klass(not_null_obj);
3092 
3093     // Generate the subtype check
3094     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3095 
3096     // Plug in success path into the merge
3097     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3098     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3099     if (failure_control == NULL) {
3100       if (not_subtype_ctrl != top()) { // If failure is possible
3101         PreserveJVMState pjvms(this);
3102         set_control(not_subtype_ctrl);
3103         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3104       }
3105     } else {
3106       (*failure_control) = not_subtype_ctrl;
3107     }
3108   }
3109 
3110   region->init_req(_obj_path, control());
3111   phi   ->init_req(_obj_path, cast_obj);
3112 
3113   // A merge of NULL or Casted-NotNull obj
3114   Node* res = _gvn.transform(phi);
3115 
3116   // Note I do NOT always 'replace_in_map(obj,result)' here.
3117   //  if( tk->klass()->can_be_primary_super()  )
3118     // This means that if I successfully store an Object into an array-of-String
3119     // I 'forget' that the Object is really now known to be a String.  I have to
3120     // do this because we don't have true union types for interfaces - if I store
3121     // a Baz into an array-of-Interface and then tell the optimizer it's an
3122     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3123     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3124   //  replace_in_map( obj, res );
3125 
3126   // Return final merged results
3127   set_control( _gvn.transform(region) );
3128   record_for_igvn(region);
3129   return res;
3130 }
3131 
3132 //------------------------------next_monitor-----------------------------------
3133 // What number should be given to the next monitor?
3134 int GraphKit::next_monitor() {
3135   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3136   int next = current + C->sync_stack_slots();
3137   // Keep the toplevel high water mark current:
3138   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3139   return current;
3140 }
3141 
3142 //------------------------------insert_mem_bar---------------------------------
3143 // Memory barrier to avoid floating things around
3144 // The membar serves as a pinch point between both control and all memory slices.
3145 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3146   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3147   mb->init_req(TypeFunc::Control, control());
3148   mb->init_req(TypeFunc::Memory,  reset_memory());
3149   Node* membar = _gvn.transform(mb);
3150   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3151   set_all_memory_call(membar);
3152   return membar;
3153 }
3154 
3155 //-------------------------insert_mem_bar_volatile----------------------------
3156 // Memory barrier to avoid floating things around
3157 // The membar serves as a pinch point between both control and memory(alias_idx).
3158 // If you want to make a pinch point on all memory slices, do not use this
3159 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3160 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3161   // When Parse::do_put_xxx updates a volatile field, it appends a series
3162   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3163   // The first membar is on the same memory slice as the field store opcode.
3164   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3165   // All the other membars (for other volatile slices, including AliasIdxBot,
3166   // which stands for all unknown volatile slices) are control-dependent
3167   // on the first membar.  This prevents later volatile loads or stores
3168   // from sliding up past the just-emitted store.
3169 
3170   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3171   mb->set_req(TypeFunc::Control,control());
3172   if (alias_idx == Compile::AliasIdxBot) {
3173     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3174   } else {
3175     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3176     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3177   }
3178   Node* membar = _gvn.transform(mb);
3179   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3180   if (alias_idx == Compile::AliasIdxBot) {
3181     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3182   } else {
3183     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3184   }
3185   return membar;
3186 }
3187 
3188 void GraphKit::insert_store_load_for_barrier() {
3189   Node* mem = reset_memory();
3190   MemBarNode* mb = MemBarNode::make(C, Op_MemBarVolatile, Compile::AliasIdxBot);
3191   mb->init_req(TypeFunc::Control, control());
3192   mb->init_req(TypeFunc::Memory, mem);
3193   Node* membar = _gvn.transform(mb);
3194   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3195   Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory));
3196   set_all_memory(mem);
3197   set_memory(newmem, Compile::AliasIdxRaw);
3198 }
3199 
3200 
3201 //------------------------------shared_lock------------------------------------
3202 // Emit locking code.
3203 FastLockNode* GraphKit::shared_lock(Node* obj) {
3204   // bci is either a monitorenter bc or InvocationEntryBci
3205   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3206   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3207 
3208   if( !GenerateSynchronizationCode )
3209     return NULL;                // Not locking things?
3210   if (stopped())                // Dead monitor?
3211     return NULL;
3212 
3213   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3214 
3215   // Box the stack location
3216   Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3217   Node* mem = reset_memory();
3218 
3219   FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3220   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3221     // Create the counters for this fast lock.
3222     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3223   }
3224 
3225   // Create the rtm counters for this fast lock if needed.
3226   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3227 
3228   // Add monitor to debug info for the slow path.  If we block inside the
3229   // slow path and de-opt, we need the monitor hanging around
3230   map()->push_monitor( flock );
3231 
3232   const TypeFunc *tf = LockNode::lock_type();
3233   LockNode *lock = new LockNode(C, tf);
3234 
3235   lock->init_req( TypeFunc::Control, control() );
3236   lock->init_req( TypeFunc::Memory , mem );
3237   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3238   lock->init_req( TypeFunc::FramePtr, frameptr() );
3239   lock->init_req( TypeFunc::ReturnAdr, top() );
3240 
3241   lock->init_req(TypeFunc::Parms + 0, obj);
3242   lock->init_req(TypeFunc::Parms + 1, box);
3243   lock->init_req(TypeFunc::Parms + 2, flock);
3244   add_safepoint_edges(lock);
3245 
3246   lock = _gvn.transform( lock )->as_Lock();
3247 
3248   // lock has no side-effects, sets few values
3249   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3250 
3251   insert_mem_bar(Op_MemBarAcquireLock);
3252 
3253   // Add this to the worklist so that the lock can be eliminated
3254   record_for_igvn(lock);
3255 
3256 #ifndef PRODUCT
3257   if (PrintLockStatistics) {
3258     // Update the counter for this lock.  Don't bother using an atomic
3259     // operation since we don't require absolute accuracy.
3260     lock->create_lock_counter(map()->jvms());
3261     increment_counter(lock->counter()->addr());
3262   }
3263 #endif
3264 
3265   return flock;
3266 }
3267 
3268 
3269 //------------------------------shared_unlock----------------------------------
3270 // Emit unlocking code.
3271 void GraphKit::shared_unlock(Node* box, Node* obj) {
3272   // bci is either a monitorenter bc or InvocationEntryBci
3273   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3274   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3275 
3276   if( !GenerateSynchronizationCode )
3277     return;
3278   if (stopped()) {               // Dead monitor?
3279     map()->pop_monitor();        // Kill monitor from debug info
3280     return;
3281   }
3282 
3283   // Memory barrier to avoid floating things down past the locked region
3284   insert_mem_bar(Op_MemBarReleaseLock);
3285 
3286   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3287   UnlockNode *unlock = new UnlockNode(C, tf);
3288 #ifdef ASSERT
3289   unlock->set_dbg_jvms(sync_jvms());
3290 #endif
3291   uint raw_idx = Compile::AliasIdxRaw;
3292   unlock->init_req( TypeFunc::Control, control() );
3293   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3294   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3295   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3296   unlock->init_req( TypeFunc::ReturnAdr, top() );
3297 
3298   unlock->init_req(TypeFunc::Parms + 0, obj);
3299   unlock->init_req(TypeFunc::Parms + 1, box);
3300   unlock = _gvn.transform(unlock)->as_Unlock();
3301 
3302   Node* mem = reset_memory();
3303 
3304   // unlock has no side-effects, sets few values
3305   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3306 
3307   // Kill monitor from debug info
3308   map()->pop_monitor( );
3309 }
3310 
3311 //-------------------------------get_layout_helper-----------------------------
3312 // If the given klass is a constant or known to be an array,
3313 // fetch the constant layout helper value into constant_value
3314 // and return (Node*)NULL.  Otherwise, load the non-constant
3315 // layout helper value, and return the node which represents it.
3316 // This two-faced routine is useful because allocation sites
3317 // almost always feature constant types.
3318 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3319   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3320   if (!StressReflectiveCode && inst_klass != NULL) {
3321     ciKlass* klass = inst_klass->klass();
3322     bool    xklass = inst_klass->klass_is_exact();
3323     if (xklass || klass->is_array_klass()) {
3324       jint lhelper = klass->layout_helper();
3325       if (lhelper != Klass::_lh_neutral_value) {
3326         constant_value = lhelper;
3327         return (Node*) NULL;
3328       }
3329     }
3330   }
3331   constant_value = Klass::_lh_neutral_value;  // put in a known value
3332   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3333   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3334 }
3335 
3336 // We just put in an allocate/initialize with a big raw-memory effect.
3337 // Hook selected additional alias categories on the initialization.
3338 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3339                                 MergeMemNode* init_in_merge,
3340                                 Node* init_out_raw) {
3341   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3342   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3343 
3344   Node* prevmem = kit.memory(alias_idx);
3345   init_in_merge->set_memory_at(alias_idx, prevmem);
3346   kit.set_memory(init_out_raw, alias_idx);
3347 }
3348 
3349 //---------------------------set_output_for_allocation-------------------------
3350 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3351                                           const TypeOopPtr* oop_type,
3352                                           bool deoptimize_on_exception) {
3353   int rawidx = Compile::AliasIdxRaw;
3354   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3355   add_safepoint_edges(alloc);
3356   Node* allocx = _gvn.transform(alloc);
3357   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3358   // create memory projection for i_o
3359   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3360   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3361 
3362   // create a memory projection as for the normal control path
3363   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3364   set_memory(malloc, rawidx);
3365 
3366   // a normal slow-call doesn't change i_o, but an allocation does
3367   // we create a separate i_o projection for the normal control path
3368   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3369   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3370 
3371   // put in an initialization barrier
3372   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3373                                                  rawoop)->as_Initialize();
3374   assert(alloc->initialization() == init,  "2-way macro link must work");
3375   assert(init ->allocation()     == alloc, "2-way macro link must work");
3376   {
3377     // Extract memory strands which may participate in the new object's
3378     // initialization, and source them from the new InitializeNode.
3379     // This will allow us to observe initializations when they occur,
3380     // and link them properly (as a group) to the InitializeNode.
3381     assert(init->in(InitializeNode::Memory) == malloc, "");
3382     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3383     init->set_req(InitializeNode::Memory, minit_in);
3384     record_for_igvn(minit_in); // fold it up later, if possible
3385     Node* minit_out = memory(rawidx);
3386     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3387     if (oop_type->isa_aryptr()) {
3388       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3389       int            elemidx  = C->get_alias_index(telemref);
3390       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3391     } else if (oop_type->isa_instptr()) {
3392       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3393       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3394         ciField* field = ik->nonstatic_field_at(i);
3395         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3396           continue;  // do not bother to track really large numbers of fields
3397         // Find (or create) the alias category for this field:
3398         int fieldidx = C->alias_type(field)->index();
3399         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3400       }
3401     }
3402   }
3403 
3404   // Cast raw oop to the real thing...
3405   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3406   javaoop = _gvn.transform(javaoop);
3407   C->set_recent_alloc(control(), javaoop);
3408   assert(just_allocated_object(control()) == javaoop, "just allocated");
3409 
3410 #ifdef ASSERT
3411   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3412     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3413            "Ideal_allocation works");
3414     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3415            "Ideal_allocation works");
3416     if (alloc->is_AllocateArray()) {
3417       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3418              "Ideal_allocation works");
3419       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3420              "Ideal_allocation works");
3421     } else {
3422       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3423     }
3424   }
3425 #endif //ASSERT
3426 
3427   return javaoop;
3428 }
3429 
3430 //---------------------------new_instance--------------------------------------
3431 // This routine takes a klass_node which may be constant (for a static type)
3432 // or may be non-constant (for reflective code).  It will work equally well
3433 // for either, and the graph will fold nicely if the optimizer later reduces
3434 // the type to a constant.
3435 // The optional arguments are for specialized use by intrinsics:
3436 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3437 //  - If 'return_size_val', report the the total object size to the caller.
3438 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3439 Node* GraphKit::new_instance(Node* klass_node,
3440                              Node* extra_slow_test,
3441                              Node* *return_size_val,
3442                              bool deoptimize_on_exception) {
3443   // Compute size in doublewords
3444   // The size is always an integral number of doublewords, represented
3445   // as a positive bytewise size stored in the klass's layout_helper.
3446   // The layout_helper also encodes (in a low bit) the need for a slow path.
3447   jint  layout_con = Klass::_lh_neutral_value;
3448   Node* layout_val = get_layout_helper(klass_node, layout_con);
3449   int   layout_is_con = (layout_val == NULL);
3450 
3451   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3452   // Generate the initial go-slow test.  It's either ALWAYS (return a
3453   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3454   // case) a computed value derived from the layout_helper.
3455   Node* initial_slow_test = NULL;
3456   if (layout_is_con) {
3457     assert(!StressReflectiveCode, "stress mode does not use these paths");
3458     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3459     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3460   } else {   // reflective case
3461     // This reflective path is used by Unsafe.allocateInstance.
3462     // (It may be stress-tested by specifying StressReflectiveCode.)
3463     // Basically, we want to get into the VM is there's an illegal argument.
3464     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3465     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3466     if (extra_slow_test != intcon(0)) {
3467       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3468     }
3469     // (Macro-expander will further convert this to a Bool, if necessary.)
3470   }
3471 
3472   // Find the size in bytes.  This is easy; it's the layout_helper.
3473   // The size value must be valid even if the slow path is taken.
3474   Node* size = NULL;
3475   if (layout_is_con) {
3476     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3477   } else {   // reflective case
3478     // This reflective path is used by clone and Unsafe.allocateInstance.
3479     size = ConvI2X(layout_val);
3480 
3481     // Clear the low bits to extract layout_helper_size_in_bytes:
3482     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3483     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3484     size = _gvn.transform( new AndXNode(size, mask) );
3485   }
3486   if (return_size_val != NULL) {
3487     (*return_size_val) = size;
3488   }
3489 
3490   // This is a precise notnull oop of the klass.
3491   // (Actually, it need not be precise if this is a reflective allocation.)
3492   // It's what we cast the result to.
3493   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3494   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3495   const TypeOopPtr* oop_type = tklass->as_instance_type();
3496 
3497   // Now generate allocation code
3498 
3499   // The entire memory state is needed for slow path of the allocation
3500   // since GC and deoptimization can happen.
3501   Node *mem = reset_memory();
3502   set_all_memory(mem); // Create new memory state
3503 
3504   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3505                                          control(), mem, i_o(),
3506                                          size, klass_node,
3507                                          initial_slow_test);
3508 
3509   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3510 }
3511 
3512 //-------------------------------new_array-------------------------------------
3513 // helper for both newarray and anewarray
3514 // The 'length' parameter is (obviously) the length of the array.
3515 // See comments on new_instance for the meaning of the other arguments.
3516 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3517                           Node* length,         // number of array elements
3518                           int   nargs,          // number of arguments to push back for uncommon trap
3519                           Node* *return_size_val,
3520                           bool deoptimize_on_exception) {
3521   jint  layout_con = Klass::_lh_neutral_value;
3522   Node* layout_val = get_layout_helper(klass_node, layout_con);
3523   int   layout_is_con = (layout_val == NULL);
3524 
3525   if (!layout_is_con && !StressReflectiveCode &&
3526       !too_many_traps(Deoptimization::Reason_class_check)) {
3527     // This is a reflective array creation site.
3528     // Optimistically assume that it is a subtype of Object[],
3529     // so that we can fold up all the address arithmetic.
3530     layout_con = Klass::array_layout_helper(T_OBJECT);
3531     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3532     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3533     { BuildCutout unless(this, bol_lh, PROB_MAX);
3534       inc_sp(nargs);
3535       uncommon_trap(Deoptimization::Reason_class_check,
3536                     Deoptimization::Action_maybe_recompile);
3537     }
3538     layout_val = NULL;
3539     layout_is_con = true;
3540   }
3541 
3542   // Generate the initial go-slow test.  Make sure we do not overflow
3543   // if length is huge (near 2Gig) or negative!  We do not need
3544   // exact double-words here, just a close approximation of needed
3545   // double-words.  We can't add any offset or rounding bits, lest we
3546   // take a size -1 of bytes and make it positive.  Use an unsigned
3547   // compare, so negative sizes look hugely positive.
3548   int fast_size_limit = FastAllocateSizeLimit;
3549   if (layout_is_con) {
3550     assert(!StressReflectiveCode, "stress mode does not use these paths");
3551     // Increase the size limit if we have exact knowledge of array type.
3552     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3553     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3554   }
3555 
3556   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3557   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3558 
3559   // --- Size Computation ---
3560   // array_size = round_to_heap(array_header + (length << elem_shift));
3561   // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3562   // and round_to(x, y) == ((x + y-1) & ~(y-1))
3563   // The rounding mask is strength-reduced, if possible.
3564   int round_mask = MinObjAlignmentInBytes - 1;
3565   Node* header_size = NULL;
3566   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3567   // (T_BYTE has the weakest alignment and size restrictions...)
3568   if (layout_is_con) {
3569     int       hsize  = Klass::layout_helper_header_size(layout_con);
3570     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3571     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3572     if ((round_mask & ~right_n_bits(eshift)) == 0)
3573       round_mask = 0;  // strength-reduce it if it goes away completely
3574     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3575     assert(header_size_min <= hsize, "generic minimum is smallest");
3576     header_size_min = hsize;
3577     header_size = intcon(hsize + round_mask);
3578   } else {
3579     Node* hss   = intcon(Klass::_lh_header_size_shift);
3580     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3581     Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3582     hsize       = _gvn.transform( new AndINode(hsize, hsm) );
3583     Node* mask  = intcon(round_mask);
3584     header_size = _gvn.transform( new AddINode(hsize, mask) );
3585   }
3586 
3587   Node* elem_shift = NULL;
3588   if (layout_is_con) {
3589     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3590     if (eshift != 0)
3591       elem_shift = intcon(eshift);
3592   } else {
3593     // There is no need to mask or shift this value.
3594     // The semantics of LShiftINode include an implicit mask to 0x1F.
3595     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3596     elem_shift = layout_val;
3597   }
3598 
3599   // Transition to native address size for all offset calculations:
3600   Node* lengthx = ConvI2X(length);
3601   Node* headerx = ConvI2X(header_size);
3602 #ifdef _LP64
3603   { const TypeInt* tilen = _gvn.find_int_type(length);
3604     if (tilen != NULL && tilen->_lo < 0) {
3605       // Add a manual constraint to a positive range.  Cf. array_element_address.
3606       jint size_max = fast_size_limit;
3607       if (size_max > tilen->_hi)  size_max = tilen->_hi;
3608       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3609 
3610       // Only do a narrow I2L conversion if the range check passed.
3611       IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3612       _gvn.transform(iff);
3613       RegionNode* region = new RegionNode(3);
3614       _gvn.set_type(region, Type::CONTROL);
3615       lengthx = new PhiNode(region, TypeLong::LONG);
3616       _gvn.set_type(lengthx, TypeLong::LONG);
3617 
3618       // Range check passed. Use ConvI2L node with narrow type.
3619       Node* passed = IfFalse(iff);
3620       region->init_req(1, passed);
3621       // Make I2L conversion control dependent to prevent it from
3622       // floating above the range check during loop optimizations.
3623       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3624 
3625       // Range check failed. Use ConvI2L with wide type because length may be invalid.
3626       region->init_req(2, IfTrue(iff));
3627       lengthx->init_req(2, ConvI2X(length));
3628 
3629       set_control(region);
3630       record_for_igvn(region);
3631       record_for_igvn(lengthx);
3632     }
3633   }
3634 #endif
3635 
3636   // Combine header size (plus rounding) and body size.  Then round down.
3637   // This computation cannot overflow, because it is used only in two
3638   // places, one where the length is sharply limited, and the other
3639   // after a successful allocation.
3640   Node* abody = lengthx;
3641   if (elem_shift != NULL)
3642     abody     = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3643   Node* size  = _gvn.transform( new AddXNode(headerx, abody) );
3644   if (round_mask != 0) {
3645     Node* mask = MakeConX(~round_mask);
3646     size       = _gvn.transform( new AndXNode(size, mask) );
3647   }
3648   // else if round_mask == 0, the size computation is self-rounding
3649 
3650   if (return_size_val != NULL) {
3651     // This is the size
3652     (*return_size_val) = size;
3653   }
3654 
3655   // Now generate allocation code
3656 
3657   // The entire memory state is needed for slow path of the allocation
3658   // since GC and deoptimization can happened.
3659   Node *mem = reset_memory();
3660   set_all_memory(mem); // Create new memory state
3661 
3662   if (initial_slow_test->is_Bool()) {
3663     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3664     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3665   }
3666 
3667   // Create the AllocateArrayNode and its result projections
3668   AllocateArrayNode* alloc
3669     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3670                             control(), mem, i_o(),
3671                             size, klass_node,
3672                             initial_slow_test,
3673                             length);
3674 
3675   // Cast to correct type.  Note that the klass_node may be constant or not,
3676   // and in the latter case the actual array type will be inexact also.
3677   // (This happens via a non-constant argument to inline_native_newArray.)
3678   // In any case, the value of klass_node provides the desired array type.
3679   const TypeInt* length_type = _gvn.find_int_type(length);
3680   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3681   if (ary_type->isa_aryptr() && length_type != NULL) {
3682     // Try to get a better type than POS for the size
3683     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3684   }
3685 
3686   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3687 
3688   // Cast length on remaining path to be as narrow as possible
3689   if (map()->find_edge(length) >= 0) {
3690     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3691     if (ccast != length) {
3692       _gvn.set_type_bottom(ccast);
3693       record_for_igvn(ccast);
3694       replace_in_map(length, ccast);
3695     }
3696   }
3697 
3698   return javaoop;
3699 }
3700 
3701 // The following "Ideal_foo" functions are placed here because they recognize
3702 // the graph shapes created by the functions immediately above.
3703 
3704 //---------------------------Ideal_allocation----------------------------------
3705 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3706 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3707   if (ptr == NULL) {     // reduce dumb test in callers
3708     return NULL;
3709   }
3710   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3711     ptr = ptr->in(1);
3712     if (ptr == NULL) return NULL;
3713   }
3714   // Return NULL for allocations with several casts:
3715   //   j.l.reflect.Array.newInstance(jobject, jint)
3716   //   Object.clone()
3717   // to keep more precise type from last cast.
3718   if (ptr->is_Proj()) {
3719     Node* allo = ptr->in(0);
3720     if (allo != NULL && allo->is_Allocate()) {
3721       return allo->as_Allocate();
3722     }
3723   }
3724   // Report failure to match.
3725   return NULL;
3726 }
3727 
3728 // Fancy version which also strips off an offset (and reports it to caller).
3729 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3730                                              intptr_t& offset) {
3731   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3732   if (base == NULL)  return NULL;
3733   return Ideal_allocation(base, phase);
3734 }
3735 
3736 // Trace Initialize <- Proj[Parm] <- Allocate
3737 AllocateNode* InitializeNode::allocation() {
3738   Node* rawoop = in(InitializeNode::RawAddress);
3739   if (rawoop->is_Proj()) {
3740     Node* alloc = rawoop->in(0);
3741     if (alloc->is_Allocate()) {
3742       return alloc->as_Allocate();
3743     }
3744   }
3745   return NULL;
3746 }
3747 
3748 // Trace Allocate -> Proj[Parm] -> Initialize
3749 InitializeNode* AllocateNode::initialization() {
3750   ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3751   if (rawoop == NULL)  return NULL;
3752   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3753     Node* init = rawoop->fast_out(i);
3754     if (init->is_Initialize()) {
3755       assert(init->as_Initialize()->allocation() == this, "2-way link");
3756       return init->as_Initialize();
3757     }
3758   }
3759   return NULL;
3760 }
3761 
3762 //----------------------------- loop predicates ---------------------------
3763 
3764 //------------------------------add_predicate_impl----------------------------
3765 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3766   // Too many traps seen?
3767   if (too_many_traps(reason)) {
3768 #ifdef ASSERT
3769     if (TraceLoopPredicate) {
3770       int tc = C->trap_count(reason);
3771       tty->print("too many traps=%s tcount=%d in ",
3772                     Deoptimization::trap_reason_name(reason), tc);
3773       method()->print(); // which method has too many predicate traps
3774       tty->cr();
3775     }
3776 #endif
3777     // We cannot afford to take more traps here,
3778     // do not generate predicate.
3779     return;
3780   }
3781 
3782   Node *cont    = _gvn.intcon(1);
3783   Node* opq     = _gvn.transform(new Opaque1Node(C, cont));
3784   Node *bol     = _gvn.transform(new Conv2BNode(opq));
3785   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3786   Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3787   C->add_predicate_opaq(opq);
3788   {
3789     PreserveJVMState pjvms(this);
3790     set_control(iffalse);
3791     inc_sp(nargs);
3792     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3793   }
3794   Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3795   set_control(iftrue);
3796 }
3797 
3798 //------------------------------add_predicate---------------------------------
3799 void GraphKit::add_predicate(int nargs) {
3800   if (UseLoopPredicate) {
3801     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3802   }
3803   // loop's limit check predicate should be near the loop.
3804   if (LoopLimitCheck) {
3805     add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3806   }
3807 }
3808 
3809 //----------------------------- store barriers ----------------------------
3810 #define __ ideal.
3811 
3812 void GraphKit::sync_kit(IdealKit& ideal) {
3813   set_all_memory(__ merged_memory());
3814   set_i_o(__ i_o());
3815   set_control(__ ctrl());
3816 }
3817 
3818 void GraphKit::final_sync(IdealKit& ideal) {
3819   // Final sync IdealKit and graphKit.
3820   sync_kit(ideal);
3821 }
3822 
3823 Node* GraphKit::byte_map_base_node() {
3824   // Get base of card map
3825   CardTableModRefBS* ct =
3826     barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3827   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
3828   if (ct->byte_map_base != NULL) {
3829     return makecon(TypeRawPtr::make((address)ct->byte_map_base));
3830   } else {
3831     return null();
3832   }
3833 }
3834 
3835 // vanilla/CMS post barrier
3836 // Insert a write-barrier store.  This is to let generational GC work; we have
3837 // to flag all oop-stores before the next GC point.
3838 void GraphKit::write_barrier_post(Node* oop_store,
3839                                   Node* obj,
3840                                   Node* adr,
3841                                   uint  adr_idx,
3842                                   Node* val,
3843                                   bool use_precise) {
3844   // No store check needed if we're storing a NULL or an old object
3845   // (latter case is probably a string constant). The concurrent
3846   // mark sweep garbage collector, however, needs to have all nonNull
3847   // oop updates flagged via card-marks.
3848   if (val != NULL && val->is_Con()) {
3849     // must be either an oop or NULL
3850     const Type* t = val->bottom_type();
3851     if (t == TypePtr::NULL_PTR || t == Type::TOP)
3852       // stores of null never (?) need barriers
3853       return;
3854   }
3855 
3856   if (use_ReduceInitialCardMarks()
3857       && obj == just_allocated_object(control())) {
3858     // We can skip marks on a freshly-allocated object in Eden.
3859     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3860     // That routine informs GC to take appropriate compensating steps,
3861     // upon a slow-path allocation, so as to make this card-mark
3862     // elision safe.
3863     return;
3864   }
3865 
3866   if (!use_precise) {
3867     // All card marks for a (non-array) instance are in one place:
3868     adr = obj;
3869   }
3870   // (Else it's an array (or unknown), and we want more precise card marks.)
3871   assert(adr != NULL, "");
3872 
3873   IdealKit ideal(this, true);
3874 
3875   // Convert the pointer to an int prior to doing math on it
3876   Node* cast = __ CastPX(__ ctrl(), adr);
3877 
3878   // Divide by card size
3879   assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3880          "Only one we handle so far.");
3881   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3882 
3883   // Combine card table base and card offset
3884   Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3885 
3886   // Get the alias_index for raw card-mark memory
3887   int adr_type = Compile::AliasIdxRaw;
3888   Node*   zero = __ ConI(0); // Dirty card value
3889   BasicType bt = T_BYTE;
3890 
3891   if (UseConcMarkSweepGC && UseCondCardMark) {
3892     insert_store_load_for_barrier();
3893     __ sync_kit(this);
3894   }
3895 
3896   if (UseCondCardMark) {
3897     // The classic GC reference write barrier is typically implemented
3898     // as a store into the global card mark table.  Unfortunately
3899     // unconditional stores can result in false sharing and excessive
3900     // coherence traffic as well as false transactional aborts.
3901     // UseCondCardMark enables MP "polite" conditional card mark
3902     // stores.  In theory we could relax the load from ctrl() to
3903     // no_ctrl, but that doesn't buy much latitude.
3904     Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3905     __ if_then(card_val, BoolTest::ne, zero);
3906   }
3907 
3908   // Smash zero into card
3909   if( !UseConcMarkSweepGC ) {
3910     __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3911   } else {
3912     // Specialized path for CM store barrier
3913     __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3914   }
3915 
3916   if (UseCondCardMark) {
3917     __ end_if();
3918   }
3919 
3920   // Final sync IdealKit and GraphKit.
3921   final_sync(ideal);
3922 }
3923 /*
3924  * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3925  * required by SATB to make sure all objects live at the start of the
3926  * marking are kept alive, all reference updates need to any previous
3927  * reference stored before writing.
3928  *
3929  * If the previous value is NULL there is no need to save the old value.
3930  * References that are NULL are filtered during runtime by the barrier
3931  * code to avoid unnecessary queuing.
3932  *
3933  * However in the case of newly allocated objects it might be possible to
3934  * prove that the reference about to be overwritten is NULL during compile
3935  * time and avoid adding the barrier code completely.
3936  *
3937  * The compiler needs to determine that the object in which a field is about
3938  * to be written is newly allocated, and that no prior store to the same field
3939  * has happened since the allocation.
3940  *
3941  * Returns true if the pre-barrier can be removed
3942  */
3943 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3944                                          BasicType bt, uint adr_idx) {
3945   intptr_t offset = 0;
3946   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3947   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3948 
3949   if (offset == Type::OffsetBot) {
3950     return false; // cannot unalias unless there are precise offsets
3951   }
3952 
3953   if (alloc == NULL) {
3954     return false; // No allocation found
3955   }
3956 
3957   intptr_t size_in_bytes = type2aelembytes(bt);
3958 
3959   Node* mem = memory(adr_idx); // start searching here...
3960 
3961   for (int cnt = 0; cnt < 50; cnt++) {
3962 
3963     if (mem->is_Store()) {
3964 
3965       Node* st_adr = mem->in(MemNode::Address);
3966       intptr_t st_offset = 0;
3967       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
3968 
3969       if (st_base == NULL) {
3970         break; // inscrutable pointer
3971       }
3972 
3973       // Break we have found a store with same base and offset as ours so break
3974       if (st_base == base && st_offset == offset) {
3975         break;
3976       }
3977 
3978       if (st_offset != offset && st_offset != Type::OffsetBot) {
3979         const int MAX_STORE = BytesPerLong;
3980         if (st_offset >= offset + size_in_bytes ||
3981             st_offset <= offset - MAX_STORE ||
3982             st_offset <= offset - mem->as_Store()->memory_size()) {
3983           // Success:  The offsets are provably independent.
3984           // (You may ask, why not just test st_offset != offset and be done?
3985           // The answer is that stores of different sizes can co-exist
3986           // in the same sequence of RawMem effects.  We sometimes initialize
3987           // a whole 'tile' of array elements with a single jint or jlong.)
3988           mem = mem->in(MemNode::Memory);
3989           continue; // advance through independent store memory
3990         }
3991       }
3992 
3993       if (st_base != base
3994           && MemNode::detect_ptr_independence(base, alloc, st_base,
3995                                               AllocateNode::Ideal_allocation(st_base, phase),
3996                                               phase)) {
3997         // Success:  The bases are provably independent.
3998         mem = mem->in(MemNode::Memory);
3999         continue; // advance through independent store memory
4000       }
4001     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4002 
4003       InitializeNode* st_init = mem->in(0)->as_Initialize();
4004       AllocateNode* st_alloc = st_init->allocation();
4005 
4006       // Make sure that we are looking at the same allocation site.
4007       // The alloc variable is guaranteed to not be null here from earlier check.
4008       if (alloc == st_alloc) {
4009         // Check that the initialization is storing NULL so that no previous store
4010         // has been moved up and directly write a reference
4011         Node* captured_store = st_init->find_captured_store(offset,
4012                                                             type2aelembytes(T_OBJECT),
4013                                                             phase);
4014         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4015           return true;
4016         }
4017       }
4018     }
4019 
4020     // Unless there is an explicit 'continue', we must bail out here,
4021     // because 'mem' is an inscrutable memory state (e.g., a call).
4022     break;
4023   }
4024 
4025   return false;
4026 }
4027 
4028 // G1 pre/post barriers
4029 void GraphKit::g1_write_barrier_pre(bool do_load,
4030                                     Node* obj,
4031                                     Node* adr,
4032                                     uint alias_idx,
4033                                     Node* val,
4034                                     const TypeOopPtr* val_type,
4035                                     Node* pre_val,
4036                                     BasicType bt) {
4037 
4038   // Some sanity checks
4039   // Note: val is unused in this routine.
4040 
4041   if (do_load) {
4042     // We need to generate the load of the previous value
4043     assert(obj != NULL, "must have a base");
4044     assert(adr != NULL, "where are loading from?");
4045     assert(pre_val == NULL, "loaded already?");
4046     assert(val_type != NULL, "need a type");
4047 
4048     if (use_ReduceInitialCardMarks()
4049         && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4050       return;
4051     }
4052 
4053   } else {
4054     // In this case both val_type and alias_idx are unused.
4055     assert(pre_val != NULL, "must be loaded already");
4056     // Nothing to be done if pre_val is null.
4057     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4058     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4059   }
4060   assert(bt == T_OBJECT, "or we shouldn't be here");
4061 
4062   IdealKit ideal(this, true);
4063 
4064   Node* tls = __ thread(); // ThreadLocalStorage
4065 
4066   Node* no_ctrl = NULL;
4067   Node* no_base = __ top();
4068   Node* zero  = __ ConI(0);
4069   Node* zeroX = __ ConX(0);
4070 
4071   float likely  = PROB_LIKELY(0.999);
4072   float unlikely  = PROB_UNLIKELY(0.999);
4073 
4074   BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4075   assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4076 
4077   // Offsets into the thread
4078   const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +  // 648
4079                                           SATBMarkQueue::byte_offset_of_active());
4080   const int index_offset   = in_bytes(JavaThread::satb_mark_queue_offset() +  // 656
4081                                           SATBMarkQueue::byte_offset_of_index());
4082   const int buffer_offset  = in_bytes(JavaThread::satb_mark_queue_offset() +  // 652
4083                                           SATBMarkQueue::byte_offset_of_buf());
4084 
4085   // Now the actual pointers into the thread
4086   Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4087   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
4088   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
4089 
4090   // Now some of the values
4091   Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4092 
4093   // if (!marking)
4094   __ if_then(marking, BoolTest::ne, zero, unlikely); {
4095     BasicType index_bt = TypeX_X->basic_type();
4096     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4097     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4098 
4099     if (do_load) {
4100       // load original value
4101       // alias_idx correct??
4102       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4103     }
4104 
4105     // if (pre_val != NULL)
4106     __ if_then(pre_val, BoolTest::ne, null()); {
4107       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4108 
4109       // is the queue for this thread full?
4110       __ if_then(index, BoolTest::ne, zeroX, likely); {
4111 
4112         // decrement the index
4113         Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4114 
4115         // Now get the buffer location we will log the previous value into and store it
4116         Node *log_addr = __ AddP(no_base, buffer, next_index);
4117         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4118         // update the index
4119         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4120 
4121       } __ else_(); {
4122 
4123         // logging buffer is full, call the runtime
4124         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4125         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4126       } __ end_if();  // (!index)
4127     } __ end_if();  // (pre_val != NULL)
4128   } __ end_if();  // (!marking)
4129 
4130   // Final sync IdealKit and GraphKit.
4131   final_sync(ideal);
4132 }
4133 
4134 /*
4135  * G1 similar to any GC with a Young Generation requires a way to keep track of
4136  * references from Old Generation to Young Generation to make sure all live
4137  * objects are found. G1 also requires to keep track of object references
4138  * between different regions to enable evacuation of old regions, which is done
4139  * as part of mixed collections. References are tracked in remembered sets and
4140  * is continuously updated as reference are written to with the help of the
4141  * post-barrier.
4142  *
4143  * To reduce the number of updates to the remembered set the post-barrier
4144  * filters updates to fields in objects located in the Young Generation,
4145  * the same region as the reference, when the NULL is being written or
4146  * if the card is already marked as dirty by an earlier write.
4147  *
4148  * Under certain circumstances it is possible to avoid generating the
4149  * post-barrier completely if it is possible during compile time to prove
4150  * the object is newly allocated and that no safepoint exists between the
4151  * allocation and the store.
4152  *
4153  * In the case of slow allocation the allocation code must handle the barrier
4154  * as part of the allocation in the case the allocated object is not located
4155  * in the nursery, this would happen for humongous objects. This is similar to
4156  * how CMS is required to handle this case, see the comments for the method
4157  * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
4158  * A deferred card mark is required for these objects and handled in the above
4159  * mentioned methods.
4160  *
4161  * Returns true if the post barrier can be removed
4162  */
4163 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4164                                           Node* adr) {
4165   intptr_t      offset = 0;
4166   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4167   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
4168 
4169   if (offset == Type::OffsetBot) {
4170     return false; // cannot unalias unless there are precise offsets
4171   }
4172 
4173   if (alloc == NULL) {
4174      return false; // No allocation found
4175   }
4176 
4177   // Start search from Store node
4178   Node* mem = store->in(MemNode::Control);
4179   if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4180 
4181     InitializeNode* st_init = mem->in(0)->as_Initialize();
4182     AllocateNode*  st_alloc = st_init->allocation();
4183 
4184     // Make sure we are looking at the same allocation
4185     if (alloc == st_alloc) {
4186       return true;
4187     }
4188   }
4189 
4190   return false;
4191 }
4192 
4193 //
4194 // Update the card table and add card address to the queue
4195 //
4196 void GraphKit::g1_mark_card(IdealKit& ideal,
4197                             Node* card_adr,
4198                             Node* oop_store,
4199                             uint oop_alias_idx,
4200                             Node* index,
4201                             Node* index_adr,
4202                             Node* buffer,
4203                             const TypeFunc* tf) {
4204 
4205   Node* zero  = __ ConI(0);
4206   Node* zeroX = __ ConX(0);
4207   Node* no_base = __ top();
4208   BasicType card_bt = T_BYTE;
4209   // Smash zero into card. MUST BE ORDERED WRT TO STORE
4210   __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4211 
4212   //  Now do the queue work
4213   __ if_then(index, BoolTest::ne, zeroX); {
4214 
4215     Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4216     Node* log_addr = __ AddP(no_base, buffer, next_index);
4217 
4218     // Order, see storeCM.
4219     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4220     __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4221 
4222   } __ else_(); {
4223     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4224   } __ end_if();
4225 
4226 }
4227 
4228 void GraphKit::g1_write_barrier_post(Node* oop_store,
4229                                      Node* obj,
4230                                      Node* adr,
4231                                      uint alias_idx,
4232                                      Node* val,
4233                                      BasicType bt,
4234                                      bool use_precise) {
4235   // If we are writing a NULL then we need no post barrier
4236 
4237   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4238     // Must be NULL
4239     const Type* t = val->bottom_type();
4240     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4241     // No post barrier if writing NULLx
4242     return;
4243   }
4244 
4245   if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4246     // We can skip marks on a freshly-allocated object in Eden.
4247     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
4248     // That routine informs GC to take appropriate compensating steps,
4249     // upon a slow-path allocation, so as to make this card-mark
4250     // elision safe.
4251     return;
4252   }
4253 
4254   if (use_ReduceInitialCardMarks()
4255       && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4256     return;
4257   }
4258 
4259   if (!use_precise) {
4260     // All card marks for a (non-array) instance are in one place:
4261     adr = obj;
4262   }
4263   // (Else it's an array (or unknown), and we want more precise card marks.)
4264   assert(adr != NULL, "");
4265 
4266   IdealKit ideal(this, true);
4267 
4268   Node* tls = __ thread(); // ThreadLocalStorage
4269 
4270   Node* no_base = __ top();
4271   float likely  = PROB_LIKELY(0.999);
4272   float unlikely  = PROB_UNLIKELY(0.999);
4273   Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4274   Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4275   Node* zeroX = __ ConX(0);
4276 
4277   // Get the alias_index for raw card-mark memory
4278   const TypePtr* card_type = TypeRawPtr::BOTTOM;
4279 
4280   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4281 
4282   // Offsets into the thread
4283   const int index_offset  = in_bytes(JavaThread::dirty_card_queue_offset() +
4284                                      DirtyCardQueue::byte_offset_of_index());
4285   const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4286                                      DirtyCardQueue::byte_offset_of_buf());
4287 
4288   // Pointers into the thread
4289 
4290   Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4291   Node* index_adr =  __ AddP(no_base, tls, __ ConX(index_offset));
4292 
4293   // Now some values
4294   // Use ctrl to avoid hoisting these values past a safepoint, which could
4295   // potentially reset these fields in the JavaThread.
4296   Node* index  = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4297   Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4298 
4299   // Convert the store obj pointer to an int prior to doing math on it
4300   // Must use ctrl to prevent "integerized oop" existing across safepoint
4301   Node* cast =  __ CastPX(__ ctrl(), adr);
4302 
4303   // Divide pointer by card size
4304   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4305 
4306   // Combine card table base and card offset
4307   Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4308 
4309   // If we know the value being stored does it cross regions?
4310 
4311   if (val != NULL) {
4312     // Does the store cause us to cross regions?
4313 
4314     // Should be able to do an unsigned compare of region_size instead of
4315     // and extra shift. Do we have an unsigned compare??
4316     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4317     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4318 
4319     // if (xor_res == 0) same region so skip
4320     __ if_then(xor_res, BoolTest::ne, zeroX); {
4321 
4322       // No barrier if we are storing a NULL
4323       __ if_then(val, BoolTest::ne, null(), unlikely); {
4324 
4325         // Ok must mark the card if not already dirty
4326 
4327         // load the original value of the card
4328         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4329 
4330         __ if_then(card_val, BoolTest::ne, young_card); {
4331           sync_kit(ideal);
4332           insert_store_load_for_barrier();
4333           __ sync_kit(this);
4334 
4335           Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4336           __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4337             g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4338           } __ end_if();
4339         } __ end_if();
4340       } __ end_if();
4341     } __ end_if();
4342   } else {
4343     // Object.clone() instrinsic uses this path.
4344     g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4345   }
4346 
4347   // Final sync IdealKit and GraphKit.
4348   final_sync(ideal);
4349 }
4350 #undef __
4351 
4352 
4353 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4354   Node* len = load_array_length(load_String_value(ctrl, str));
4355   Node* coder = load_String_coder(ctrl, str);
4356   // Divide length by 2 if coder is UTF16
4357   return _gvn.transform(new RShiftINode(len, coder));
4358 }
4359 
4360 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4361   int value_offset = java_lang_String::value_offset_in_bytes();
4362   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4363                                                      false, NULL, 0);
4364   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4365   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4366                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4367                                                   ciTypeArrayKlass::make(T_BYTE), true, 0);
4368   int value_field_idx = C->get_alias_index(value_field_type);
4369   Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4370                          value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4371   // String.value field is known to be @Stable.
4372   if (UseImplicitStableValues) {
4373     load = cast_array_to_stable(load, value_type);
4374   }
4375   return load;
4376 }
4377 
4378 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4379   if (java_lang_String::has_coder_field()) {
4380     if (!CompactStrings) {
4381       return intcon(java_lang_String::CODER_UTF16);
4382     }
4383     int coder_offset = java_lang_String::coder_offset_in_bytes();
4384     const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4385                                                        false, NULL, 0);
4386     const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4387     int coder_field_idx = C->get_alias_index(coder_field_type);
4388     return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4389                      TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4390   } else {
4391     return intcon(0); // false
4392   }
4393 }
4394 
4395 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4396   int value_offset = java_lang_String::value_offset_in_bytes();
4397   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4398                                                      false, NULL, 0);
4399   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4400   store_oop_to_object(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
4401       value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4402 }
4403 
4404 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4405   int coder_offset = java_lang_String::coder_offset_in_bytes();
4406   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4407                                                      false, NULL, 0);
4408   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4409   int coder_field_idx = C->get_alias_index(coder_field_type);
4410   store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4411                   value, T_BYTE, coder_field_idx, MemNode::unordered);
4412 }
4413 
4414 // Capture src and dst memory state with a MergeMemNode
4415 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4416   if (src_type == dst_type) {
4417     // Types are equal, we don't need a MergeMemNode
4418     return memory(src_type);
4419   }
4420   MergeMemNode* merge = MergeMemNode::make(map()->memory());
4421   record_for_igvn(merge); // fold it up later, if possible
4422   int src_idx = C->get_alias_index(src_type);
4423   int dst_idx = C->get_alias_index(dst_type);
4424   merge->set_memory_at(src_idx, memory(src_idx));
4425   merge->set_memory_at(dst_idx, memory(dst_idx));
4426   return merge;
4427 }
4428 
4429 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4430   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4431   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4432   // If input and output memory types differ, capture both states to preserve
4433   // the dependency between preceding and subsequent loads/stores.
4434   // For example, the following program:
4435   //  StoreB
4436   //  compress_string
4437   //  LoadB
4438   // has this memory graph (use->def):
4439   //  LoadB -> compress_string -> CharMem
4440   //             ... -> StoreB -> ByteMem
4441   // The intrinsic hides the dependency between LoadB and StoreB, causing
4442   // the load to read from memory not containing the result of the StoreB.
4443   // The correct memory graph should look like this:
4444   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4445   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4446   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4447   Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4448   set_memory(res_mem, TypeAryPtr::BYTES);
4449   return str;
4450 }
4451 
4452 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4453   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4454   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4455   // Capture src and dst memory (see comment in 'compress_string').
4456   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4457   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4458   set_memory(_gvn.transform(str), dst_type);
4459 }
4460 
4461 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4462   /**
4463    * int i_char = start;
4464    * for (int i_byte = 0; i_byte < count; i_byte++) {
4465    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
4466    * }
4467    */
4468   add_predicate();
4469   RegionNode* head = new RegionNode(3);
4470   head->init_req(1, control());
4471   gvn().set_type(head, Type::CONTROL);
4472   record_for_igvn(head);
4473 
4474   Node* i_byte = new PhiNode(head, TypeInt::INT);
4475   i_byte->init_req(1, intcon(0));
4476   gvn().set_type(i_byte, TypeInt::INT);
4477   record_for_igvn(i_byte);
4478 
4479   Node* i_char = new PhiNode(head, TypeInt::INT);
4480   i_char->init_req(1, start);
4481   gvn().set_type(i_char, TypeInt::INT);
4482   record_for_igvn(i_char);
4483 
4484   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4485   gvn().set_type(mem, Type::MEMORY);
4486   record_for_igvn(mem);
4487   set_control(head);
4488   set_memory(mem, TypeAryPtr::BYTES);
4489   Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4490   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4491                              AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4492                              false, false, true /* mismatched */);
4493 
4494   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4495   head->init_req(2, IfTrue(iff));
4496   mem->init_req(2, st);
4497   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4498   i_char->init_req(2, AddI(i_char, intcon(2)));
4499 
4500   set_control(IfFalse(iff));
4501   set_memory(st, TypeAryPtr::BYTES);
4502 }
4503 
4504 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4505   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4506   // assumption of CCP analysis.
4507   return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4508 }