1 /*
   2  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "ci/ciUtilities.hpp"
  27 #include "compiler/compileLog.hpp"
  28 #include "gc/shared/barrierSet.hpp"
  29 #include "gc/shared/c2/barrierSetC2.hpp"
  30 #include "interpreter/interpreter.hpp"
  31 #include "memory/resourceArea.hpp"
  32 #include "opto/addnode.hpp"
  33 #include "opto/castnode.hpp"
  34 #include "opto/convertnode.hpp"
  35 #include "opto/graphKit.hpp"
  36 #include "opto/idealKit.hpp"
  37 #include "opto/intrinsicnode.hpp"
  38 #include "opto/locknode.hpp"
  39 #include "opto/machnode.hpp"
  40 #include "opto/opaquenode.hpp"
  41 #include "opto/parse.hpp"
  42 #include "opto/rootnode.hpp"
  43 #include "opto/runtime.hpp"
  44 #include "runtime/deoptimization.hpp"
  45 #include "runtime/sharedRuntime.hpp"
  46 #include "utilities/macros.hpp"
  47 #if INCLUDE_SHENANDOAHGC
  48 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  49 #endif
  50 
  51 //----------------------------GraphKit-----------------------------------------
  52 // Main utility constructor.
  53 GraphKit::GraphKit(JVMState* jvms)
  54   : Phase(Phase::Parser),
  55     _env(C->env()),
  56     _gvn(*C->initial_gvn()),
  57     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  58 {
  59   _exceptions = jvms->map()->next_exception();
  60   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  61   set_jvms(jvms);
  62 }
  63 
  64 // Private constructor for parser.
  65 GraphKit::GraphKit()
  66   : Phase(Phase::Parser),
  67     _env(C->env()),
  68     _gvn(*C->initial_gvn()),
  69     _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  70 {
  71   _exceptions = NULL;
  72   set_map(NULL);
  73   debug_only(_sp = -99);
  74   debug_only(set_bci(-99));
  75 }
  76 
  77 
  78 
  79 //---------------------------clean_stack---------------------------------------
  80 // Clear away rubbish from the stack area of the JVM state.
  81 // This destroys any arguments that may be waiting on the stack.
  82 void GraphKit::clean_stack(int from_sp) {
  83   SafePointNode* map      = this->map();
  84   JVMState*      jvms     = this->jvms();
  85   int            stk_size = jvms->stk_size();
  86   int            stkoff   = jvms->stkoff();
  87   Node*          top      = this->top();
  88   for (int i = from_sp; i < stk_size; i++) {
  89     if (map->in(stkoff + i) != top) {
  90       map->set_req(stkoff + i, top);
  91     }
  92   }
  93 }
  94 
  95 
  96 //--------------------------------sync_jvms-----------------------------------
  97 // Make sure our current jvms agrees with our parse state.
  98 JVMState* GraphKit::sync_jvms() const {
  99   JVMState* jvms = this->jvms();
 100   jvms->set_bci(bci());       // Record the new bci in the JVMState
 101   jvms->set_sp(sp());         // Record the new sp in the JVMState
 102   assert(jvms_in_sync(), "jvms is now in sync");
 103   return jvms;
 104 }
 105 
 106 //--------------------------------sync_jvms_for_reexecute---------------------
 107 // Make sure our current jvms agrees with our parse state.  This version
 108 // uses the reexecute_sp for reexecuting bytecodes.
 109 JVMState* GraphKit::sync_jvms_for_reexecute() {
 110   JVMState* jvms = this->jvms();
 111   jvms->set_bci(bci());          // Record the new bci in the JVMState
 112   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 113   return jvms;
 114 }
 115 
 116 #ifdef ASSERT
 117 bool GraphKit::jvms_in_sync() const {
 118   Parse* parse = is_Parse();
 119   if (parse == NULL) {
 120     if (bci() !=      jvms()->bci())          return false;
 121     if (sp()  != (int)jvms()->sp())           return false;
 122     return true;
 123   }
 124   if (jvms()->method() != parse->method())    return false;
 125   if (jvms()->bci()    != parse->bci())       return false;
 126   int jvms_sp = jvms()->sp();
 127   if (jvms_sp          != parse->sp())        return false;
 128   int jvms_depth = jvms()->depth();
 129   if (jvms_depth       != parse->depth())     return false;
 130   return true;
 131 }
 132 
 133 // Local helper checks for special internal merge points
 134 // used to accumulate and merge exception states.
 135 // They are marked by the region's in(0) edge being the map itself.
 136 // Such merge points must never "escape" into the parser at large,
 137 // until they have been handed to gvn.transform.
 138 static bool is_hidden_merge(Node* reg) {
 139   if (reg == NULL)  return false;
 140   if (reg->is_Phi()) {
 141     reg = reg->in(0);
 142     if (reg == NULL)  return false;
 143   }
 144   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 145 }
 146 
 147 void GraphKit::verify_map() const {
 148   if (map() == NULL)  return;  // null map is OK
 149   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 150   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 151   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 152 }
 153 
 154 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 155   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 156   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 157 }
 158 #endif
 159 
 160 //---------------------------stop_and_kill_map---------------------------------
 161 // Set _map to NULL, signalling a stop to further bytecode execution.
 162 // First smash the current map's control to a constant, to mark it dead.
 163 void GraphKit::stop_and_kill_map() {
 164   SafePointNode* dead_map = stop();
 165   if (dead_map != NULL) {
 166     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
 167     assert(dead_map->is_killed(), "must be so marked");
 168   }
 169 }
 170 
 171 
 172 //--------------------------------stopped--------------------------------------
 173 // Tell if _map is NULL, or control is top.
 174 bool GraphKit::stopped() {
 175   if (map() == NULL)           return true;
 176   else if (control() == top()) return true;
 177   else                         return false;
 178 }
 179 
 180 
 181 //-----------------------------has_ex_handler----------------------------------
 182 // Tell if this method or any caller method has exception handlers.
 183 bool GraphKit::has_ex_handler() {
 184   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 185     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 186       return true;
 187     }
 188   }
 189   return false;
 190 }
 191 
 192 //------------------------------save_ex_oop------------------------------------
 193 // Save an exception without blowing stack contents or other JVM state.
 194 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 195   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 196   ex_map->add_req(ex_oop);
 197   debug_only(verify_exception_state(ex_map));
 198 }
 199 
 200 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 201   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 202   Node* ex_oop = ex_map->in(ex_map->req()-1);
 203   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 204   return ex_oop;
 205 }
 206 
 207 //-----------------------------saved_ex_oop------------------------------------
 208 // Recover a saved exception from its map.
 209 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 210   return common_saved_ex_oop(ex_map, false);
 211 }
 212 
 213 //--------------------------clear_saved_ex_oop---------------------------------
 214 // Erase a previously saved exception from its map.
 215 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 216   return common_saved_ex_oop(ex_map, true);
 217 }
 218 
 219 #ifdef ASSERT
 220 //---------------------------has_saved_ex_oop----------------------------------
 221 // Erase a previously saved exception from its map.
 222 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 223   return ex_map->req() == ex_map->jvms()->endoff()+1;
 224 }
 225 #endif
 226 
 227 //-------------------------make_exception_state--------------------------------
 228 // Turn the current JVM state into an exception state, appending the ex_oop.
 229 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 230   sync_jvms();
 231   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 232   set_saved_ex_oop(ex_map, ex_oop);
 233   return ex_map;
 234 }
 235 
 236 
 237 //--------------------------add_exception_state--------------------------------
 238 // Add an exception to my list of exceptions.
 239 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 240   if (ex_map == NULL || ex_map->control() == top()) {
 241     return;
 242   }
 243 #ifdef ASSERT
 244   verify_exception_state(ex_map);
 245   if (has_exceptions()) {
 246     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 247   }
 248 #endif
 249 
 250   // If there is already an exception of exactly this type, merge with it.
 251   // In particular, null-checks and other low-level exceptions common up here.
 252   Node*       ex_oop  = saved_ex_oop(ex_map);
 253   const Type* ex_type = _gvn.type(ex_oop);
 254   if (ex_oop == top()) {
 255     // No action needed.
 256     return;
 257   }
 258   assert(ex_type->isa_instptr(), "exception must be an instance");
 259   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 260     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 261     // We check sp also because call bytecodes can generate exceptions
 262     // both before and after arguments are popped!
 263     if (ex_type2 == ex_type
 264         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 265       combine_exception_states(ex_map, e2);
 266       return;
 267     }
 268   }
 269 
 270   // No pre-existing exception of the same type.  Chain it on the list.
 271   push_exception_state(ex_map);
 272 }
 273 
 274 //-----------------------add_exception_states_from-----------------------------
 275 void GraphKit::add_exception_states_from(JVMState* jvms) {
 276   SafePointNode* ex_map = jvms->map()->next_exception();
 277   if (ex_map != NULL) {
 278     jvms->map()->set_next_exception(NULL);
 279     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 280       next_map = ex_map->next_exception();
 281       ex_map->set_next_exception(NULL);
 282       add_exception_state(ex_map);
 283     }
 284   }
 285 }
 286 
 287 //-----------------------transfer_exceptions_into_jvms-------------------------
 288 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 289   if (map() == NULL) {
 290     // We need a JVMS to carry the exceptions, but the map has gone away.
 291     // Create a scratch JVMS, cloned from any of the exception states...
 292     if (has_exceptions()) {
 293       _map = _exceptions;
 294       _map = clone_map();
 295       _map->set_next_exception(NULL);
 296       clear_saved_ex_oop(_map);
 297       debug_only(verify_map());
 298     } else {
 299       // ...or created from scratch
 300       JVMState* jvms = new (C) JVMState(_method, NULL);
 301       jvms->set_bci(_bci);
 302       jvms->set_sp(_sp);
 303       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
 304       set_jvms(jvms);
 305       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 306       set_all_memory(top());
 307       while (map()->req() < jvms->endoff())  map()->add_req(top());
 308     }
 309     // (This is a kludge, in case you didn't notice.)
 310     set_control(top());
 311   }
 312   JVMState* jvms = sync_jvms();
 313   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 314   jvms->map()->set_next_exception(_exceptions);
 315   _exceptions = NULL;   // done with this set of exceptions
 316   return jvms;
 317 }
 318 
 319 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 320   assert(is_hidden_merge(dstphi), "must be a special merge node");
 321   assert(is_hidden_merge(srcphi), "must be a special merge node");
 322   uint limit = srcphi->req();
 323   for (uint i = PhiNode::Input; i < limit; i++) {
 324     dstphi->add_req(srcphi->in(i));
 325   }
 326 }
 327 static inline void add_one_req(Node* dstphi, Node* src) {
 328   assert(is_hidden_merge(dstphi), "must be a special merge node");
 329   assert(!is_hidden_merge(src), "must not be a special merge node");
 330   dstphi->add_req(src);
 331 }
 332 
 333 //-----------------------combine_exception_states------------------------------
 334 // This helper function combines exception states by building phis on a
 335 // specially marked state-merging region.  These regions and phis are
 336 // untransformed, and can build up gradually.  The region is marked by
 337 // having a control input of its exception map, rather than NULL.  Such
 338 // regions do not appear except in this function, and in use_exception_state.
 339 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 340   if (failing())  return;  // dying anyway...
 341   JVMState* ex_jvms = ex_map->_jvms;
 342   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 343   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 344   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 345   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 346   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 347   assert(ex_map->req() == phi_map->req(), "matching maps");
 348   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 349   Node*         hidden_merge_mark = root();
 350   Node*         region  = phi_map->control();
 351   MergeMemNode* phi_mem = phi_map->merged_memory();
 352   MergeMemNode* ex_mem  = ex_map->merged_memory();
 353   if (region->in(0) != hidden_merge_mark) {
 354     // The control input is not (yet) a specially-marked region in phi_map.
 355     // Make it so, and build some phis.
 356     region = new RegionNode(2);
 357     _gvn.set_type(region, Type::CONTROL);
 358     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 359     region->init_req(1, phi_map->control());
 360     phi_map->set_control(region);
 361     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 362     record_for_igvn(io_phi);
 363     _gvn.set_type(io_phi, Type::ABIO);
 364     phi_map->set_i_o(io_phi);
 365     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 366       Node* m = mms.memory();
 367       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 368       record_for_igvn(m_phi);
 369       _gvn.set_type(m_phi, Type::MEMORY);
 370       mms.set_memory(m_phi);
 371     }
 372   }
 373 
 374   // Either or both of phi_map and ex_map might already be converted into phis.
 375   Node* ex_control = ex_map->control();
 376   // if there is special marking on ex_map also, we add multiple edges from src
 377   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 378   // how wide was the destination phi_map, originally?
 379   uint orig_width = region->req();
 380 
 381   if (add_multiple) {
 382     add_n_reqs(region, ex_control);
 383     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 384   } else {
 385     // ex_map has no merges, so we just add single edges everywhere
 386     add_one_req(region, ex_control);
 387     add_one_req(phi_map->i_o(), ex_map->i_o());
 388   }
 389   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 390     if (mms.is_empty()) {
 391       // get a copy of the base memory, and patch some inputs into it
 392       const TypePtr* adr_type = mms.adr_type(C);
 393       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 394       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 395       mms.set_memory(phi);
 396       // Prepare to append interesting stuff onto the newly sliced phi:
 397       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 398     }
 399     // Append stuff from ex_map:
 400     if (add_multiple) {
 401       add_n_reqs(mms.memory(), mms.memory2());
 402     } else {
 403       add_one_req(mms.memory(), mms.memory2());
 404     }
 405   }
 406   uint limit = ex_map->req();
 407   for (uint i = TypeFunc::Parms; i < limit; i++) {
 408     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 409     if (i == tos)  i = ex_jvms->monoff();
 410     Node* src = ex_map->in(i);
 411     Node* dst = phi_map->in(i);
 412     if (src != dst) {
 413       PhiNode* phi;
 414       if (dst->in(0) != region) {
 415         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 416         record_for_igvn(phi);
 417         _gvn.set_type(phi, phi->type());
 418         phi_map->set_req(i, dst);
 419         // Prepare to append interesting stuff onto the new phi:
 420         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 421       } else {
 422         assert(dst->is_Phi(), "nobody else uses a hidden region");
 423         phi = dst->as_Phi();
 424       }
 425       if (add_multiple && src->in(0) == ex_control) {
 426         // Both are phis.
 427         add_n_reqs(dst, src);
 428       } else {
 429         while (dst->req() < region->req())  add_one_req(dst, src);
 430       }
 431       const Type* srctype = _gvn.type(src);
 432       if (phi->type() != srctype) {
 433         const Type* dsttype = phi->type()->meet_speculative(srctype);
 434         if (phi->type() != dsttype) {
 435           phi->set_type(dsttype);
 436           _gvn.set_type(phi, dsttype);
 437         }
 438       }
 439     }
 440   }
 441   phi_map->merge_replaced_nodes_with(ex_map);
 442 }
 443 
 444 //--------------------------use_exception_state--------------------------------
 445 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 446   if (failing()) { stop(); return top(); }
 447   Node* region = phi_map->control();
 448   Node* hidden_merge_mark = root();
 449   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 450   Node* ex_oop = clear_saved_ex_oop(phi_map);
 451   if (region->in(0) == hidden_merge_mark) {
 452     // Special marking for internal ex-states.  Process the phis now.
 453     region->set_req(0, region);  // now it's an ordinary region
 454     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 455     // Note: Setting the jvms also sets the bci and sp.
 456     set_control(_gvn.transform(region));
 457     uint tos = jvms()->stkoff() + sp();
 458     for (uint i = 1; i < tos; i++) {
 459       Node* x = phi_map->in(i);
 460       if (x->in(0) == region) {
 461         assert(x->is_Phi(), "expected a special phi");
 462         phi_map->set_req(i, _gvn.transform(x));
 463       }
 464     }
 465     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 466       Node* x = mms.memory();
 467       if (x->in(0) == region) {
 468         assert(x->is_Phi(), "nobody else uses a hidden region");
 469         mms.set_memory(_gvn.transform(x));
 470       }
 471     }
 472     if (ex_oop->in(0) == region) {
 473       assert(ex_oop->is_Phi(), "expected a special phi");
 474       ex_oop = _gvn.transform(ex_oop);
 475     }
 476   } else {
 477     set_jvms(phi_map->jvms());
 478   }
 479 
 480   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 481   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 482   return ex_oop;
 483 }
 484 
 485 //---------------------------------java_bc-------------------------------------
 486 Bytecodes::Code GraphKit::java_bc() const {
 487   ciMethod* method = this->method();
 488   int       bci    = this->bci();
 489   if (method != NULL && bci != InvocationEntryBci)
 490     return method->java_code_at_bci(bci);
 491   else
 492     return Bytecodes::_illegal;
 493 }
 494 
 495 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 496                                                           bool must_throw) {
 497     // if the exception capability is set, then we will generate code
 498     // to check the JavaThread.should_post_on_exceptions flag to see
 499     // if we actually need to report exception events (for this
 500     // thread).  If we don't need to report exception events, we will
 501     // take the normal fast path provided by add_exception_events.  If
 502     // exception event reporting is enabled for this thread, we will
 503     // take the uncommon_trap in the BuildCutout below.
 504 
 505     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 506     Node* jthread = _gvn.transform(new ThreadLocalNode());
 507     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 508     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
 509 
 510     // Test the should_post_on_exceptions_flag vs. 0
 511     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
 512     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
 513 
 514     // Branch to slow_path if should_post_on_exceptions_flag was true
 515     { BuildCutout unless(this, tst, PROB_MAX);
 516       // Do not try anything fancy if we're notifying the VM on every throw.
 517       // Cf. case Bytecodes::_athrow in parse2.cpp.
 518       uncommon_trap(reason, Deoptimization::Action_none,
 519                     (ciKlass*)NULL, (char*)NULL, must_throw);
 520     }
 521 
 522 }
 523 
 524 //------------------------------builtin_throw----------------------------------
 525 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 526   bool must_throw = true;
 527 
 528   if (env()->jvmti_can_post_on_exceptions()) {
 529     // check if we must post exception events, take uncommon trap if so
 530     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 531     // here if should_post_on_exceptions is false
 532     // continue on with the normal codegen
 533   }
 534 
 535   // If this particular condition has not yet happened at this
 536   // bytecode, then use the uncommon trap mechanism, and allow for
 537   // a future recompilation if several traps occur here.
 538   // If the throw is hot, try to use a more complicated inline mechanism
 539   // which keeps execution inside the compiled code.
 540   bool treat_throw_as_hot = false;
 541   ciMethodData* md = method()->method_data();
 542 
 543   if (ProfileTraps) {
 544     if (too_many_traps(reason)) {
 545       treat_throw_as_hot = true;
 546     }
 547     // (If there is no MDO at all, assume it is early in
 548     // execution, and that any deopts are part of the
 549     // startup transient, and don't need to be remembered.)
 550 
 551     // Also, if there is a local exception handler, treat all throws
 552     // as hot if there has been at least one in this method.
 553     if (C->trap_count(reason) != 0
 554         && method()->method_data()->trap_count(reason) != 0
 555         && has_ex_handler()) {
 556         treat_throw_as_hot = true;
 557     }
 558   }
 559 
 560   // If this throw happens frequently, an uncommon trap might cause
 561   // a performance pothole.  If there is a local exception handler,
 562   // and if this particular bytecode appears to be deoptimizing often,
 563   // let us handle the throw inline, with a preconstructed instance.
 564   // Note:   If the deopt count has blown up, the uncommon trap
 565   // runtime is going to flush this nmethod, not matter what.
 566   if (treat_throw_as_hot
 567       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 568     // If the throw is local, we use a pre-existing instance and
 569     // punt on the backtrace.  This would lead to a missing backtrace
 570     // (a repeat of 4292742) if the backtrace object is ever asked
 571     // for its backtrace.
 572     // Fixing this remaining case of 4292742 requires some flavor of
 573     // escape analysis.  Leave that for the future.
 574     ciInstance* ex_obj = NULL;
 575     switch (reason) {
 576     case Deoptimization::Reason_null_check:
 577       ex_obj = env()->NullPointerException_instance();
 578       break;
 579     case Deoptimization::Reason_div0_check:
 580       ex_obj = env()->ArithmeticException_instance();
 581       break;
 582     case Deoptimization::Reason_range_check:
 583       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 584       break;
 585     case Deoptimization::Reason_class_check:
 586       if (java_bc() == Bytecodes::_aastore) {
 587         ex_obj = env()->ArrayStoreException_instance();
 588       } else {
 589         ex_obj = env()->ClassCastException_instance();
 590       }
 591       break;
 592     default:
 593       break;
 594     }
 595     if (failing()) { stop(); return; }  // exception allocation might fail
 596     if (ex_obj != NULL) {
 597       // Cheat with a preallocated exception object.
 598       if (C->log() != NULL)
 599         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 600                        Deoptimization::trap_reason_name(reason));
 601       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 602       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
 603 
 604       ex_node = access_resolve_for_write(ex_node);
 605 
 606       // Clear the detail message of the preallocated exception object.
 607       // Weblogic sometimes mutates the detail message of exceptions
 608       // using reflection.
 609       int offset = java_lang_Throwable::get_detailMessage_offset();
 610       const TypePtr* adr_typ = ex_con->add_offset(offset);
 611 
 612       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 613       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 614       Node *store = access_store_at(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
 615 
 616       add_exception_state(make_exception_state(ex_node));
 617       return;
 618     }
 619   }
 620 
 621   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 622   // It won't be much cheaper than bailing to the interp., since we'll
 623   // have to pass up all the debug-info, and the runtime will have to
 624   // create the stack trace.
 625 
 626   // Usual case:  Bail to interpreter.
 627   // Reserve the right to recompile if we haven't seen anything yet.
 628 
 629   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
 630   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 631   if (treat_throw_as_hot
 632       && (method()->method_data()->trap_recompiled_at(bci(), m)
 633           || C->too_many_traps(reason))) {
 634     // We cannot afford to take more traps here.  Suffer in the interpreter.
 635     if (C->log() != NULL)
 636       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 637                      Deoptimization::trap_reason_name(reason),
 638                      C->trap_count(reason));
 639     action = Deoptimization::Action_none;
 640   }
 641 
 642   // "must_throw" prunes the JVM state to include only the stack, if there
 643   // are no local exception handlers.  This should cut down on register
 644   // allocation time and code size, by drastically reducing the number
 645   // of in-edges on the call to the uncommon trap.
 646 
 647   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 648 }
 649 
 650 
 651 //----------------------------PreserveJVMState---------------------------------
 652 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 653   debug_only(kit->verify_map());
 654   _kit    = kit;
 655   _map    = kit->map();   // preserve the map
 656   _sp     = kit->sp();
 657   kit->set_map(clone_map ? kit->clone_map() : NULL);
 658 #ifdef ASSERT
 659   _bci    = kit->bci();
 660   Parse* parser = kit->is_Parse();
 661   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 662   _block  = block;
 663 #endif
 664 }
 665 PreserveJVMState::~PreserveJVMState() {
 666   GraphKit* kit = _kit;
 667 #ifdef ASSERT
 668   assert(kit->bci() == _bci, "bci must not shift");
 669   Parse* parser = kit->is_Parse();
 670   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 671   assert(block == _block,    "block must not shift");
 672 #endif
 673   kit->set_map(_map);
 674   kit->set_sp(_sp);
 675 }
 676 
 677 
 678 //-----------------------------BuildCutout-------------------------------------
 679 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 680   : PreserveJVMState(kit)
 681 {
 682   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 683   SafePointNode* outer_map = _map;   // preserved map is caller's
 684   SafePointNode* inner_map = kit->map();
 685   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 686   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
 687   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
 688 }
 689 BuildCutout::~BuildCutout() {
 690   GraphKit* kit = _kit;
 691   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 692 }
 693 
 694 //---------------------------PreserveReexecuteState----------------------------
 695 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 696   assert(!kit->stopped(), "must call stopped() before");
 697   _kit    =    kit;
 698   _sp     =    kit->sp();
 699   _reexecute = kit->jvms()->_reexecute;
 700 }
 701 PreserveReexecuteState::~PreserveReexecuteState() {
 702   if (_kit->stopped()) return;
 703   _kit->jvms()->_reexecute = _reexecute;
 704   _kit->set_sp(_sp);
 705 }
 706 
 707 //------------------------------clone_map--------------------------------------
 708 // Implementation of PreserveJVMState
 709 //
 710 // Only clone_map(...) here. If this function is only used in the
 711 // PreserveJVMState class we may want to get rid of this extra
 712 // function eventually and do it all there.
 713 
 714 SafePointNode* GraphKit::clone_map() {
 715   if (map() == NULL)  return NULL;
 716 
 717   // Clone the memory edge first
 718   Node* mem = MergeMemNode::make(map()->memory());
 719   gvn().set_type_bottom(mem);
 720 
 721   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 722   JVMState* jvms = this->jvms();
 723   JVMState* clonejvms = jvms->clone_shallow(C);
 724   clonemap->set_memory(mem);
 725   clonemap->set_jvms(clonejvms);
 726   clonejvms->set_map(clonemap);
 727   record_for_igvn(clonemap);
 728   gvn().set_type_bottom(clonemap);
 729   return clonemap;
 730 }
 731 
 732 
 733 //-----------------------------set_map_clone-----------------------------------
 734 void GraphKit::set_map_clone(SafePointNode* m) {
 735   _map = m;
 736   _map = clone_map();
 737   _map->set_next_exception(NULL);
 738   debug_only(verify_map());
 739 }
 740 
 741 
 742 //----------------------------kill_dead_locals---------------------------------
 743 // Detect any locals which are known to be dead, and force them to top.
 744 void GraphKit::kill_dead_locals() {
 745   // Consult the liveness information for the locals.  If any
 746   // of them are unused, then they can be replaced by top().  This
 747   // should help register allocation time and cut down on the size
 748   // of the deoptimization information.
 749 
 750   // This call is made from many of the bytecode handling
 751   // subroutines called from the Big Switch in do_one_bytecode.
 752   // Every bytecode which might include a slow path is responsible
 753   // for killing its dead locals.  The more consistent we
 754   // are about killing deads, the fewer useless phis will be
 755   // constructed for them at various merge points.
 756 
 757   // bci can be -1 (InvocationEntryBci).  We return the entry
 758   // liveness for the method.
 759 
 760   if (method() == NULL || method()->code_size() == 0) {
 761     // We are building a graph for a call to a native method.
 762     // All locals are live.
 763     return;
 764   }
 765 
 766   ResourceMark rm;
 767 
 768   // Consult the liveness information for the locals.  If any
 769   // of them are unused, then they can be replaced by top().  This
 770   // should help register allocation time and cut down on the size
 771   // of the deoptimization information.
 772   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 773 
 774   int len = (int)live_locals.size();
 775   assert(len <= jvms()->loc_size(), "too many live locals");
 776   for (int local = 0; local < len; local++) {
 777     if (!live_locals.at(local)) {
 778       set_local(local, top());
 779     }
 780   }
 781 }
 782 
 783 #ifdef ASSERT
 784 //-------------------------dead_locals_are_killed------------------------------
 785 // Return true if all dead locals are set to top in the map.
 786 // Used to assert "clean" debug info at various points.
 787 bool GraphKit::dead_locals_are_killed() {
 788   if (method() == NULL || method()->code_size() == 0) {
 789     // No locals need to be dead, so all is as it should be.
 790     return true;
 791   }
 792 
 793   // Make sure somebody called kill_dead_locals upstream.
 794   ResourceMark rm;
 795   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 796     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 797     SafePointNode* map = jvms->map();
 798     ciMethod* method = jvms->method();
 799     int       bci    = jvms->bci();
 800     if (jvms == this->jvms()) {
 801       bci = this->bci();  // it might not yet be synched
 802     }
 803     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 804     int len = (int)live_locals.size();
 805     if (!live_locals.is_valid() || len == 0)
 806       // This method is trivial, or is poisoned by a breakpoint.
 807       return true;
 808     assert(len == jvms->loc_size(), "live map consistent with locals map");
 809     for (int local = 0; local < len; local++) {
 810       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 811         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 812           tty->print_cr("Zombie local %d: ", local);
 813           jvms->dump();
 814         }
 815         return false;
 816       }
 817     }
 818   }
 819   return true;
 820 }
 821 
 822 #endif //ASSERT
 823 
 824 // Helper function for enforcing certain bytecodes to reexecute if
 825 // deoptimization happens
 826 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 827   ciMethod* cur_method = jvms->method();
 828   int       cur_bci   = jvms->bci();
 829   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 830     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 831     return Interpreter::bytecode_should_reexecute(code) ||
 832            (is_anewarray && code == Bytecodes::_multianewarray);
 833     // Reexecute _multianewarray bytecode which was replaced with
 834     // sequence of [a]newarray. See Parse::do_multianewarray().
 835     //
 836     // Note: interpreter should not have it set since this optimization
 837     // is limited by dimensions and guarded by flag so in some cases
 838     // multianewarray() runtime calls will be generated and
 839     // the bytecode should not be reexecutes (stack will not be reset).
 840   } else
 841     return false;
 842 }
 843 
 844 // Helper function for adding JVMState and debug information to node
 845 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 846   // Add the safepoint edges to the call (or other safepoint).
 847 
 848   // Make sure dead locals are set to top.  This
 849   // should help register allocation time and cut down on the size
 850   // of the deoptimization information.
 851   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 852 
 853   // Walk the inline list to fill in the correct set of JVMState's
 854   // Also fill in the associated edges for each JVMState.
 855 
 856   // If the bytecode needs to be reexecuted we need to put
 857   // the arguments back on the stack.
 858   const bool should_reexecute = jvms()->should_reexecute();
 859   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 860 
 861   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 862   // undefined if the bci is different.  This is normal for Parse but it
 863   // should not happen for LibraryCallKit because only one bci is processed.
 864   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 865          "in LibraryCallKit the reexecute bit should not change");
 866 
 867   // If we are guaranteed to throw, we can prune everything but the
 868   // input to the current bytecode.
 869   bool can_prune_locals = false;
 870   uint stack_slots_not_pruned = 0;
 871   int inputs = 0, depth = 0;
 872   if (must_throw) {
 873     assert(method() == youngest_jvms->method(), "sanity");
 874     if (compute_stack_effects(inputs, depth)) {
 875       can_prune_locals = true;
 876       stack_slots_not_pruned = inputs;
 877     }
 878   }
 879 
 880   if (env()->should_retain_local_variables()) {
 881     // At any safepoint, this method can get breakpointed, which would
 882     // then require an immediate deoptimization.
 883     can_prune_locals = false;  // do not prune locals
 884     stack_slots_not_pruned = 0;
 885   }
 886 
 887   // do not scribble on the input jvms
 888   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 889   call->set_jvms(out_jvms); // Start jvms list for call node
 890 
 891   // For a known set of bytecodes, the interpreter should reexecute them if
 892   // deoptimization happens. We set the reexecute state for them here
 893   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 894       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 895     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 896   }
 897 
 898   // Presize the call:
 899   DEBUG_ONLY(uint non_debug_edges = call->req());
 900   call->add_req_batch(top(), youngest_jvms->debug_depth());
 901   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 902 
 903   // Set up edges so that the call looks like this:
 904   //  Call [state:] ctl io mem fptr retadr
 905   //       [parms:] parm0 ... parmN
 906   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 907   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 908   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 909   // Note that caller debug info precedes callee debug info.
 910 
 911   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 912   uint debug_ptr = call->req();
 913 
 914   // Loop over the map input edges associated with jvms, add them
 915   // to the call node, & reset all offsets to match call node array.
 916   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 917     uint debug_end   = debug_ptr;
 918     uint debug_start = debug_ptr - in_jvms->debug_size();
 919     debug_ptr = debug_start;  // back up the ptr
 920 
 921     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 922     uint j, k, l;
 923     SafePointNode* in_map = in_jvms->map();
 924     out_jvms->set_map(call);
 925 
 926     if (can_prune_locals) {
 927       assert(in_jvms->method() == out_jvms->method(), "sanity");
 928       // If the current throw can reach an exception handler in this JVMS,
 929       // then we must keep everything live that can reach that handler.
 930       // As a quick and dirty approximation, we look for any handlers at all.
 931       if (in_jvms->method()->has_exception_handlers()) {
 932         can_prune_locals = false;
 933       }
 934     }
 935 
 936     // Add the Locals
 937     k = in_jvms->locoff();
 938     l = in_jvms->loc_size();
 939     out_jvms->set_locoff(p);
 940     if (!can_prune_locals) {
 941       for (j = 0; j < l; j++)
 942         call->set_req(p++, in_map->in(k+j));
 943     } else {
 944       p += l;  // already set to top above by add_req_batch
 945     }
 946 
 947     // Add the Expression Stack
 948     k = in_jvms->stkoff();
 949     l = in_jvms->sp();
 950     out_jvms->set_stkoff(p);
 951     if (!can_prune_locals) {
 952       for (j = 0; j < l; j++)
 953         call->set_req(p++, in_map->in(k+j));
 954     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 955       // Divide stack into {S0,...,S1}, where S0 is set to top.
 956       uint s1 = stack_slots_not_pruned;
 957       stack_slots_not_pruned = 0;  // for next iteration
 958       if (s1 > l)  s1 = l;
 959       uint s0 = l - s1;
 960       p += s0;  // skip the tops preinstalled by add_req_batch
 961       for (j = s0; j < l; j++)
 962         call->set_req(p++, in_map->in(k+j));
 963     } else {
 964       p += l;  // already set to top above by add_req_batch
 965     }
 966 
 967     // Add the Monitors
 968     k = in_jvms->monoff();
 969     l = in_jvms->mon_size();
 970     out_jvms->set_monoff(p);
 971     for (j = 0; j < l; j++)
 972       call->set_req(p++, in_map->in(k+j));
 973 
 974     // Copy any scalar object fields.
 975     k = in_jvms->scloff();
 976     l = in_jvms->scl_size();
 977     out_jvms->set_scloff(p);
 978     for (j = 0; j < l; j++)
 979       call->set_req(p++, in_map->in(k+j));
 980 
 981     // Finish the new jvms.
 982     out_jvms->set_endoff(p);
 983 
 984     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 985     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 986     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 987     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 988     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 989     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 990 
 991     // Update the two tail pointers in parallel.
 992     out_jvms = out_jvms->caller();
 993     in_jvms  = in_jvms->caller();
 994   }
 995 
 996   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
 997 
 998   // Test the correctness of JVMState::debug_xxx accessors:
 999   assert(call->jvms()->debug_start() == non_debug_edges, "");
1000   assert(call->jvms()->debug_end()   == call->req(), "");
1001   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1002 }
1003 
1004 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1005   Bytecodes::Code code = java_bc();
1006   if (code == Bytecodes::_wide) {
1007     code = method()->java_code_at_bci(bci() + 1);
1008   }
1009 
1010   BasicType rtype = T_ILLEGAL;
1011   int       rsize = 0;
1012 
1013   if (code != Bytecodes::_illegal) {
1014     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1015     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1016     if (rtype < T_CONFLICT)
1017       rsize = type2size[rtype];
1018   }
1019 
1020   switch (code) {
1021   case Bytecodes::_illegal:
1022     return false;
1023 
1024   case Bytecodes::_ldc:
1025   case Bytecodes::_ldc_w:
1026   case Bytecodes::_ldc2_w:
1027     inputs = 0;
1028     break;
1029 
1030   case Bytecodes::_dup:         inputs = 1;  break;
1031   case Bytecodes::_dup_x1:      inputs = 2;  break;
1032   case Bytecodes::_dup_x2:      inputs = 3;  break;
1033   case Bytecodes::_dup2:        inputs = 2;  break;
1034   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1035   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1036   case Bytecodes::_swap:        inputs = 2;  break;
1037   case Bytecodes::_arraylength: inputs = 1;  break;
1038 
1039   case Bytecodes::_getstatic:
1040   case Bytecodes::_putstatic:
1041   case Bytecodes::_getfield:
1042   case Bytecodes::_putfield:
1043     {
1044       bool ignored_will_link;
1045       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1046       int      size  = field->type()->size();
1047       bool is_get = (depth >= 0), is_static = (depth & 1);
1048       inputs = (is_static ? 0 : 1);
1049       if (is_get) {
1050         depth = size - inputs;
1051       } else {
1052         inputs += size;        // putxxx pops the value from the stack
1053         depth = - inputs;
1054       }
1055     }
1056     break;
1057 
1058   case Bytecodes::_invokevirtual:
1059   case Bytecodes::_invokespecial:
1060   case Bytecodes::_invokestatic:
1061   case Bytecodes::_invokedynamic:
1062   case Bytecodes::_invokeinterface:
1063     {
1064       bool ignored_will_link;
1065       ciSignature* declared_signature = NULL;
1066       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1067       assert(declared_signature != NULL, "cannot be null");
1068       inputs   = declared_signature->arg_size_for_bc(code);
1069       int size = declared_signature->return_type()->size();
1070       depth = size - inputs;
1071     }
1072     break;
1073 
1074   case Bytecodes::_multianewarray:
1075     {
1076       ciBytecodeStream iter(method());
1077       iter.reset_to_bci(bci());
1078       iter.next();
1079       inputs = iter.get_dimensions();
1080       assert(rsize == 1, "");
1081       depth = rsize - inputs;
1082     }
1083     break;
1084 
1085   case Bytecodes::_ireturn:
1086   case Bytecodes::_lreturn:
1087   case Bytecodes::_freturn:
1088   case Bytecodes::_dreturn:
1089   case Bytecodes::_areturn:
1090     assert(rsize == -depth, "");
1091     inputs = rsize;
1092     break;
1093 
1094   case Bytecodes::_jsr:
1095   case Bytecodes::_jsr_w:
1096     inputs = 0;
1097     depth  = 1;                  // S.B. depth=1, not zero
1098     break;
1099 
1100   default:
1101     // bytecode produces a typed result
1102     inputs = rsize - depth;
1103     assert(inputs >= 0, "");
1104     break;
1105   }
1106 
1107 #ifdef ASSERT
1108   // spot check
1109   int outputs = depth + inputs;
1110   assert(outputs >= 0, "sanity");
1111   switch (code) {
1112   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1113   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1114   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1115   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1116   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1117   default:                    break;
1118   }
1119 #endif //ASSERT
1120 
1121   return true;
1122 }
1123 
1124 
1125 
1126 //------------------------------basic_plus_adr---------------------------------
1127 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1128   // short-circuit a common case
1129   if (offset == intcon(0))  return ptr;
1130   return _gvn.transform( new AddPNode(base, ptr, offset) );
1131 }
1132 
1133 Node* GraphKit::ConvI2L(Node* offset) {
1134   // short-circuit a common case
1135   jint offset_con = find_int_con(offset, Type::OffsetBot);
1136   if (offset_con != Type::OffsetBot) {
1137     return longcon((jlong) offset_con);
1138   }
1139   return _gvn.transform( new ConvI2LNode(offset));
1140 }
1141 
1142 Node* GraphKit::ConvI2UL(Node* offset) {
1143   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1144   if (offset_con != (juint) Type::OffsetBot) {
1145     return longcon((julong) offset_con);
1146   }
1147   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1148   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1149   return _gvn.transform( new AndLNode(conv, mask) );
1150 }
1151 
1152 Node* GraphKit::ConvL2I(Node* offset) {
1153   // short-circuit a common case
1154   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1155   if (offset_con != (jlong)Type::OffsetBot) {
1156     return intcon((int) offset_con);
1157   }
1158   return _gvn.transform( new ConvL2INode(offset));
1159 }
1160 
1161 //-------------------------load_object_klass-----------------------------------
1162 Node* GraphKit::load_object_klass(Node* obj) {
1163   // Special-case a fresh allocation to avoid building nodes:
1164   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1165   if (akls != NULL)  return akls;
1166   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1167   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1168 }
1169 
1170 //-------------------------load_array_length-----------------------------------
1171 Node* GraphKit::load_array_length(Node* array) {
1172   // Special-case a fresh allocation to avoid building nodes:
1173   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1174   Node *alen;
1175   if (alloc == NULL) {
1176     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1177     alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1178   } else {
1179     alen = alloc->Ideal_length();
1180     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1181     if (ccast != alen) {
1182       alen = _gvn.transform(ccast);
1183     }
1184   }
1185   return alen;
1186 }
1187 
1188 //------------------------------do_null_check----------------------------------
1189 // Helper function to do a NULL pointer check.  Returned value is
1190 // the incoming address with NULL casted away.  You are allowed to use the
1191 // not-null value only if you are control dependent on the test.
1192 #ifndef PRODUCT
1193 extern int explicit_null_checks_inserted,
1194            explicit_null_checks_elided;
1195 #endif
1196 Node* GraphKit::null_check_common(Node* value, BasicType type,
1197                                   // optional arguments for variations:
1198                                   bool assert_null,
1199                                   Node* *null_control,
1200                                   bool speculative) {
1201   assert(!assert_null || null_control == NULL, "not both at once");
1202   if (stopped())  return top();
1203   NOT_PRODUCT(explicit_null_checks_inserted++);
1204 
1205   // Construct NULL check
1206   Node *chk = NULL;
1207   switch(type) {
1208     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1209     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1210     case T_ARRAY  : // fall through
1211       type = T_OBJECT;  // simplify further tests
1212     case T_OBJECT : {
1213       const Type *t = _gvn.type( value );
1214 
1215       const TypeOopPtr* tp = t->isa_oopptr();
1216       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1217           // Only for do_null_check, not any of its siblings:
1218           && !assert_null && null_control == NULL) {
1219         // Usually, any field access or invocation on an unloaded oop type
1220         // will simply fail to link, since the statically linked class is
1221         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1222         // the static class is loaded but the sharper oop type is not.
1223         // Rather than checking for this obscure case in lots of places,
1224         // we simply observe that a null check on an unloaded class
1225         // will always be followed by a nonsense operation, so we
1226         // can just issue the uncommon trap here.
1227         // Our access to the unloaded class will only be correct
1228         // after it has been loaded and initialized, which requires
1229         // a trip through the interpreter.
1230 #ifndef PRODUCT
1231         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1232 #endif
1233         uncommon_trap(Deoptimization::Reason_unloaded,
1234                       Deoptimization::Action_reinterpret,
1235                       tp->klass(), "!loaded");
1236         return top();
1237       }
1238 
1239       if (assert_null) {
1240         // See if the type is contained in NULL_PTR.
1241         // If so, then the value is already null.
1242         if (t->higher_equal(TypePtr::NULL_PTR)) {
1243           NOT_PRODUCT(explicit_null_checks_elided++);
1244           return value;           // Elided null assert quickly!
1245         }
1246       } else {
1247         // See if mixing in the NULL pointer changes type.
1248         // If so, then the NULL pointer was not allowed in the original
1249         // type.  In other words, "value" was not-null.
1250         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1251           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1252           NOT_PRODUCT(explicit_null_checks_elided++);
1253           return value;           // Elided null check quickly!
1254         }
1255       }
1256       chk = new CmpPNode( value, null() );
1257       break;
1258     }
1259 
1260     default:
1261       fatal("unexpected type: %s", type2name(type));
1262   }
1263   assert(chk != NULL, "sanity check");
1264   chk = _gvn.transform(chk);
1265 
1266   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1267   BoolNode *btst = new BoolNode( chk, btest);
1268   Node   *tst = _gvn.transform( btst );
1269 
1270   //-----------
1271   // if peephole optimizations occurred, a prior test existed.
1272   // If a prior test existed, maybe it dominates as we can avoid this test.
1273   if (tst != btst && type == T_OBJECT) {
1274     // At this point we want to scan up the CFG to see if we can
1275     // find an identical test (and so avoid this test altogether).
1276     Node *cfg = control();
1277     int depth = 0;
1278     while( depth < 16 ) {       // Limit search depth for speed
1279       if( cfg->Opcode() == Op_IfTrue &&
1280           cfg->in(0)->in(1) == tst ) {
1281         // Found prior test.  Use "cast_not_null" to construct an identical
1282         // CastPP (and hence hash to) as already exists for the prior test.
1283         // Return that casted value.
1284         if (assert_null) {
1285           replace_in_map(value, null());
1286           return null();  // do not issue the redundant test
1287         }
1288         Node *oldcontrol = control();
1289         set_control(cfg);
1290         Node *res = cast_not_null(value);
1291         set_control(oldcontrol);
1292         NOT_PRODUCT(explicit_null_checks_elided++);
1293         return res;
1294       }
1295       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1296       if (cfg == NULL)  break;  // Quit at region nodes
1297       depth++;
1298     }
1299   }
1300 
1301   //-----------
1302   // Branch to failure if null
1303   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1304   Deoptimization::DeoptReason reason;
1305   if (assert_null) {
1306     reason = Deoptimization::reason_null_assert(speculative);
1307   } else if (type == T_OBJECT) {
1308     reason = Deoptimization::reason_null_check(speculative);
1309   } else {
1310     reason = Deoptimization::Reason_div0_check;
1311   }
1312   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1313   // ciMethodData::has_trap_at will return a conservative -1 if any
1314   // must-be-null assertion has failed.  This could cause performance
1315   // problems for a method after its first do_null_assert failure.
1316   // Consider using 'Reason_class_check' instead?
1317 
1318   // To cause an implicit null check, we set the not-null probability
1319   // to the maximum (PROB_MAX).  For an explicit check the probability
1320   // is set to a smaller value.
1321   if (null_control != NULL || too_many_traps(reason)) {
1322     // probability is less likely
1323     ok_prob =  PROB_LIKELY_MAG(3);
1324   } else if (!assert_null &&
1325              (ImplicitNullCheckThreshold > 0) &&
1326              method() != NULL &&
1327              (method()->method_data()->trap_count(reason)
1328               >= (uint)ImplicitNullCheckThreshold)) {
1329     ok_prob =  PROB_LIKELY_MAG(3);
1330   }
1331 
1332   if (null_control != NULL) {
1333     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1334     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1335     set_control(      _gvn.transform( new IfTrueNode(iff)));
1336 #ifndef PRODUCT
1337     if (null_true == top()) {
1338       explicit_null_checks_elided++;
1339     }
1340 #endif
1341     (*null_control) = null_true;
1342   } else {
1343     BuildCutout unless(this, tst, ok_prob);
1344     // Check for optimizer eliding test at parse time
1345     if (stopped()) {
1346       // Failure not possible; do not bother making uncommon trap.
1347       NOT_PRODUCT(explicit_null_checks_elided++);
1348     } else if (assert_null) {
1349       uncommon_trap(reason,
1350                     Deoptimization::Action_make_not_entrant,
1351                     NULL, "assert_null");
1352     } else {
1353       replace_in_map(value, zerocon(type));
1354       builtin_throw(reason);
1355     }
1356   }
1357 
1358   // Must throw exception, fall-thru not possible?
1359   if (stopped()) {
1360     return top();               // No result
1361   }
1362 
1363   if (assert_null) {
1364     // Cast obj to null on this path.
1365     replace_in_map(value, zerocon(type));
1366     return zerocon(type);
1367   }
1368 
1369   // Cast obj to not-null on this path, if there is no null_control.
1370   // (If there is a null_control, a non-null value may come back to haunt us.)
1371   if (type == T_OBJECT) {
1372     Node* cast = cast_not_null(value, false);
1373     if (null_control == NULL || (*null_control) == top())
1374       replace_in_map(value, cast);
1375     value = cast;
1376   }
1377 
1378   return value;
1379 }
1380 
1381 
1382 //------------------------------cast_not_null----------------------------------
1383 // Cast obj to not-null on this path
1384 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1385   const Type *t = _gvn.type(obj);
1386   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1387   // Object is already not-null?
1388   if( t == t_not_null ) return obj;
1389 
1390   Node *cast = new CastPPNode(obj,t_not_null);
1391   cast->init_req(0, control());
1392   cast = _gvn.transform( cast );
1393 
1394   // Scan for instances of 'obj' in the current JVM mapping.
1395   // These instances are known to be not-null after the test.
1396   if (do_replace_in_map)
1397     replace_in_map(obj, cast);
1398 
1399   return cast;                  // Return casted value
1400 }
1401 
1402 // Sometimes in intrinsics, we implicitly know an object is not null
1403 // (there's no actual null check) so we can cast it to not null. In
1404 // the course of optimizations, the input to the cast can become null.
1405 // In that case that data path will die and we need the control path
1406 // to become dead as well to keep the graph consistent. So we have to
1407 // add a check for null for which one branch can't be taken. It uses
1408 // an Opaque4 node that will cause the check to be removed after loop
1409 // opts so the test goes away and the compiled code doesn't execute a
1410 // useless check.
1411 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1412   Node* chk = _gvn.transform(new CmpPNode(value, null()));
1413   Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1414   Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1415   IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1416   _gvn.set_type(iff, iff->Value(&_gvn));
1417   Node *if_f = _gvn.transform(new IfFalseNode(iff));
1418   Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1419   Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1420   C->root()->add_req(halt);
1421   Node *if_t = _gvn.transform(new IfTrueNode(iff));
1422   set_control(if_t);
1423   return cast_not_null(value, do_replace_in_map);
1424 }
1425 
1426 
1427 //--------------------------replace_in_map-------------------------------------
1428 void GraphKit::replace_in_map(Node* old, Node* neww) {
1429   if (old == neww) {
1430     return;
1431   }
1432 
1433   map()->replace_edge(old, neww);
1434 
1435   // Note: This operation potentially replaces any edge
1436   // on the map.  This includes locals, stack, and monitors
1437   // of the current (innermost) JVM state.
1438 
1439   // don't let inconsistent types from profiling escape this
1440   // method
1441 
1442   const Type* told = _gvn.type(old);
1443   const Type* tnew = _gvn.type(neww);
1444 
1445   if (!tnew->higher_equal(told)) {
1446     return;
1447   }
1448 
1449   map()->record_replaced_node(old, neww);
1450 }
1451 
1452 
1453 //=============================================================================
1454 //--------------------------------memory---------------------------------------
1455 Node* GraphKit::memory(uint alias_idx) {
1456   MergeMemNode* mem = merged_memory();
1457   Node* p = mem->memory_at(alias_idx);
1458   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1459   return p;
1460 }
1461 
1462 //-----------------------------reset_memory------------------------------------
1463 Node* GraphKit::reset_memory() {
1464   Node* mem = map()->memory();
1465   // do not use this node for any more parsing!
1466   debug_only( map()->set_memory((Node*)NULL) );
1467   return _gvn.transform( mem );
1468 }
1469 
1470 //------------------------------set_all_memory---------------------------------
1471 void GraphKit::set_all_memory(Node* newmem) {
1472   Node* mergemem = MergeMemNode::make(newmem);
1473   gvn().set_type_bottom(mergemem);
1474   map()->set_memory(mergemem);
1475 }
1476 
1477 //------------------------------set_all_memory_call----------------------------
1478 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1479   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1480   set_all_memory(newmem);
1481 }
1482 
1483 //=============================================================================
1484 //
1485 // parser factory methods for MemNodes
1486 //
1487 // These are layered on top of the factory methods in LoadNode and StoreNode,
1488 // and integrate with the parser's memory state and _gvn engine.
1489 //
1490 
1491 // factory methods in "int adr_idx"
1492 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1493                           int adr_idx,
1494                           MemNode::MemOrd mo,
1495                           LoadNode::ControlDependency control_dependency,
1496                           bool require_atomic_access,
1497                           bool unaligned,
1498                           bool mismatched) {
1499   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1500   const TypePtr* adr_type = NULL; // debug-mode-only argument
1501   debug_only(adr_type = C->get_adr_type(adr_idx));
1502   Node* mem = memory(adr_idx);
1503   Node* ld;
1504   if (require_atomic_access && bt == T_LONG) {
1505     ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1506   } else if (require_atomic_access && bt == T_DOUBLE) {
1507     ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1508   } else {
1509     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1510   }
1511   ld = _gvn.transform(ld);
1512   if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1513     // Improve graph before escape analysis and boxing elimination.
1514     record_for_igvn(ld);
1515   }
1516   return ld;
1517 }
1518 
1519 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1520                                 int adr_idx,
1521                                 MemNode::MemOrd mo,
1522                                 bool require_atomic_access,
1523                                 bool unaligned,
1524                                 bool mismatched) {
1525   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1526   const TypePtr* adr_type = NULL;
1527   debug_only(adr_type = C->get_adr_type(adr_idx));
1528   Node *mem = memory(adr_idx);
1529   Node* st;
1530   if (require_atomic_access && bt == T_LONG) {
1531     st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1532   } else if (require_atomic_access && bt == T_DOUBLE) {
1533     st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1534   } else {
1535     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1536   }
1537   if (unaligned) {
1538     st->as_Store()->set_unaligned_access();
1539   }
1540   if (mismatched) {
1541     st->as_Store()->set_mismatched_access();
1542   }
1543   st = _gvn.transform(st);
1544   set_memory(st, adr_idx);
1545   // Back-to-back stores can only remove intermediate store with DU info
1546   // so push on worklist for optimizer.
1547   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1548     record_for_igvn(st);
1549 
1550   return st;
1551 }
1552 
1553 Node* GraphKit::access_store_at(Node* ctl,
1554                                 Node* obj,
1555                                 Node* adr,
1556                                 const TypePtr* adr_type,
1557                                 Node* val,
1558                                 const Type* val_type,
1559                                 BasicType bt,
1560                                 DecoratorSet decorators) {
1561   // Transformation of a value which could be NULL pointer (CastPP #NULL)
1562   // could be delayed during Parse (for example, in adjust_map_after_if()).
1563   // Execute transformation here to avoid barrier generation in such case.
1564   if (_gvn.type(val) == TypePtr::NULL_PTR) {
1565     val = _gvn.makecon(TypePtr::NULL_PTR);
1566   }
1567 
1568   set_control(ctl);
1569   if (stopped()) {
1570     return top(); // Dead path ?
1571   }
1572 
1573   assert(val != NULL, "not dead path");
1574 
1575   C2AccessValuePtr addr(adr, adr_type);
1576   C2AccessValue value(val, val_type);
1577   C2Access access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1578   if (access.is_raw()) {
1579     return _barrier_set->BarrierSetC2::store_at(access, value);
1580   } else {
1581     return _barrier_set->store_at(access, value);
1582   }
1583 }
1584 
1585 Node* GraphKit::access_load_at(Node* obj,   // containing obj
1586                                Node* adr,   // actual adress to store val at
1587                                const TypePtr* adr_type,
1588                                const Type* val_type,
1589                                BasicType bt,
1590                                DecoratorSet decorators) {
1591   if (stopped()) {
1592     return top(); // Dead path ?
1593   }
1594 
1595   C2AccessValuePtr addr(adr, adr_type);
1596   C2Access access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1597   if (access.is_raw()) {
1598     return _barrier_set->BarrierSetC2::load_at(access, val_type);
1599   } else {
1600     return _barrier_set->load_at(access, val_type);
1601   }
1602 }
1603 
1604 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* ctl,
1605                                              Node* obj,
1606                                              Node* adr,
1607                                              const TypePtr* adr_type,
1608                                              int alias_idx,
1609                                              Node* expected_val,
1610                                              Node* new_val,
1611                                              const Type* value_type,
1612                                              BasicType bt,
1613                                              DecoratorSet decorators) {
1614   set_control(ctl);
1615   C2AccessValuePtr addr(adr, adr_type);
1616   C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1617                         bt, obj, addr, alias_idx);
1618   if (access.is_raw()) {
1619     return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1620   } else {
1621     return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1622   }
1623 }
1624 
1625 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* ctl,
1626                                               Node* obj,
1627                                               Node* adr,
1628                                               const TypePtr* adr_type,
1629                                               int alias_idx,
1630                                               Node* expected_val,
1631                                               Node* new_val,
1632                                               const Type* value_type,
1633                                               BasicType bt,
1634                                               DecoratorSet decorators) {
1635   set_control(ctl);
1636   C2AccessValuePtr addr(adr, adr_type);
1637   C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1638                         bt, obj, addr, alias_idx);
1639   if (access.is_raw()) {
1640     return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1641   } else {
1642     return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1643   }
1644 }
1645 
1646 Node* GraphKit::access_atomic_xchg_at(Node* ctl,
1647                                       Node* obj,
1648                                       Node* adr,
1649                                       const TypePtr* adr_type,
1650                                       int alias_idx,
1651                                       Node* new_val,
1652                                       const Type* value_type,
1653                                       BasicType bt,
1654                                       DecoratorSet decorators) {
1655   set_control(ctl);
1656   C2AccessValuePtr addr(adr, adr_type);
1657   C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1658                         bt, obj, addr, alias_idx);
1659   if (access.is_raw()) {
1660     return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1661   } else {
1662     return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1663   }
1664 }
1665 
1666 Node* GraphKit::access_atomic_add_at(Node* ctl,
1667                                      Node* obj,
1668                                      Node* adr,
1669                                      const TypePtr* adr_type,
1670                                      int alias_idx,
1671                                      Node* new_val,
1672                                      const Type* value_type,
1673                                      BasicType bt,
1674                                      DecoratorSet decorators) {
1675   set_control(ctl);
1676   C2AccessValuePtr addr(adr, adr_type);
1677   C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1678   if (access.is_raw()) {
1679     return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1680   } else {
1681     return _barrier_set->atomic_add_at(access, new_val, value_type);
1682   }
1683 }
1684 
1685 void GraphKit::access_clone(Node* ctl, Node* src, Node* dst, Node* size, bool is_array) {
1686   set_control(ctl);
1687   return _barrier_set->clone(this, src, dst, size, is_array);
1688 }
1689 
1690 Node* GraphKit::access_resolve_for_read(Node* n) {
1691   return _barrier_set->resolve_for_read(this, n);
1692 }
1693 
1694 Node* GraphKit::access_resolve_for_write(Node* n) {
1695   return _barrier_set->resolve_for_write(this, n);
1696 }
1697 
1698 
1699 //-------------------------array_element_address-------------------------
1700 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1701                                       const TypeInt* sizetype, Node* ctrl) {
1702   uint shift  = exact_log2(type2aelembytes(elembt));
1703   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1704 
1705   // short-circuit a common case (saves lots of confusing waste motion)
1706   jint idx_con = find_int_con(idx, -1);
1707   if (idx_con >= 0) {
1708     intptr_t offset = header + ((intptr_t)idx_con << shift);
1709     return basic_plus_adr(ary, offset);
1710   }
1711 
1712   // must be correct type for alignment purposes
1713   Node* base  = basic_plus_adr(ary, header);
1714   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1715   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1716   return basic_plus_adr(ary, base, scale);
1717 }
1718 
1719 //-------------------------load_array_element-------------------------
1720 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1721   const Type* elemtype = arytype->elem();
1722   BasicType elembt = elemtype->array_element_basic_type();
1723   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1724   if (elembt == T_NARROWOOP) {
1725     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1726   }
1727   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1728   return ld;
1729 }
1730 
1731 //-------------------------set_arguments_for_java_call-------------------------
1732 // Arguments (pre-popped from the stack) are taken from the JVMS.
1733 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1734   // Add the call arguments:
1735   uint nargs = call->method()->arg_size();
1736   for (uint i = 0; i < nargs; i++) {
1737     Node* arg = argument(i);
1738     call->init_req(i + TypeFunc::Parms, arg);
1739   }
1740 }
1741 
1742 //---------------------------set_edges_for_java_call---------------------------
1743 // Connect a newly created call into the current JVMS.
1744 // A return value node (if any) is returned from set_edges_for_java_call.
1745 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1746 
1747   // Add the predefined inputs:
1748   call->init_req( TypeFunc::Control, control() );
1749   call->init_req( TypeFunc::I_O    , i_o() );
1750   call->init_req( TypeFunc::Memory , reset_memory() );
1751   call->init_req( TypeFunc::FramePtr, frameptr() );
1752   call->init_req( TypeFunc::ReturnAdr, top() );
1753 
1754   add_safepoint_edges(call, must_throw);
1755 
1756   Node* xcall = _gvn.transform(call);
1757 
1758   if (xcall == top()) {
1759     set_control(top());
1760     return;
1761   }
1762   assert(xcall == call, "call identity is stable");
1763 
1764   // Re-use the current map to produce the result.
1765 
1766   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1767   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1768   set_all_memory_call(xcall, separate_io_proj);
1769 
1770   //return xcall;   // no need, caller already has it
1771 }
1772 
1773 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1774   if (stopped())  return top();  // maybe the call folded up?
1775 
1776   // Capture the return value, if any.
1777   Node* ret;
1778   if (call->method() == NULL ||
1779       call->method()->return_type()->basic_type() == T_VOID)
1780         ret = top();
1781   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1782 
1783   // Note:  Since any out-of-line call can produce an exception,
1784   // we always insert an I_O projection from the call into the result.
1785 
1786   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1787 
1788   if (separate_io_proj) {
1789     // The caller requested separate projections be used by the fall
1790     // through and exceptional paths, so replace the projections for
1791     // the fall through path.
1792     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1793     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1794   }
1795   return ret;
1796 }
1797 
1798 //--------------------set_predefined_input_for_runtime_call--------------------
1799 // Reading and setting the memory state is way conservative here.
1800 // The real problem is that I am not doing real Type analysis on memory,
1801 // so I cannot distinguish card mark stores from other stores.  Across a GC
1802 // point the Store Barrier and the card mark memory has to agree.  I cannot
1803 // have a card mark store and its barrier split across the GC point from
1804 // either above or below.  Here I get that to happen by reading ALL of memory.
1805 // A better answer would be to separate out card marks from other memory.
1806 // For now, return the input memory state, so that it can be reused
1807 // after the call, if this call has restricted memory effects.
1808 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1809   // Set fixed predefined input arguments
1810   Node* memory = reset_memory();
1811   call->init_req( TypeFunc::Control,   control()  );
1812   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1813   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1814   call->init_req( TypeFunc::FramePtr,  frameptr() );
1815   call->init_req( TypeFunc::ReturnAdr, top()      );
1816   return memory;
1817 }
1818 
1819 //-------------------set_predefined_output_for_runtime_call--------------------
1820 // Set control and memory (not i_o) from the call.
1821 // If keep_mem is not NULL, use it for the output state,
1822 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1823 // If hook_mem is NULL, this call produces no memory effects at all.
1824 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1825 // then only that memory slice is taken from the call.
1826 // In the last case, we must put an appropriate memory barrier before
1827 // the call, so as to create the correct anti-dependencies on loads
1828 // preceding the call.
1829 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1830                                                       Node* keep_mem,
1831                                                       const TypePtr* hook_mem) {
1832   // no i/o
1833   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1834   if (keep_mem) {
1835     // First clone the existing memory state
1836     set_all_memory(keep_mem);
1837     if (hook_mem != NULL) {
1838       // Make memory for the call
1839       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1840       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1841       // We also use hook_mem to extract specific effects from arraycopy stubs.
1842       set_memory(mem, hook_mem);
1843     }
1844     // ...else the call has NO memory effects.
1845 
1846     // Make sure the call advertises its memory effects precisely.
1847     // This lets us build accurate anti-dependences in gcm.cpp.
1848     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1849            "call node must be constructed correctly");
1850   } else {
1851     assert(hook_mem == NULL, "");
1852     // This is not a "slow path" call; all memory comes from the call.
1853     set_all_memory_call(call);
1854   }
1855 }
1856 
1857 
1858 // Replace the call with the current state of the kit.
1859 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1860   JVMState* ejvms = NULL;
1861   if (has_exceptions()) {
1862     ejvms = transfer_exceptions_into_jvms();
1863   }
1864 
1865   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1866   ReplacedNodes replaced_nodes_exception;
1867   Node* ex_ctl = top();
1868 
1869   SafePointNode* final_state = stop();
1870 
1871   // Find all the needed outputs of this call
1872   CallProjections callprojs;
1873   call->extract_projections(&callprojs, true);
1874 
1875   Node* init_mem = call->in(TypeFunc::Memory);
1876   Node* final_mem = final_state->in(TypeFunc::Memory);
1877   Node* final_ctl = final_state->in(TypeFunc::Control);
1878   Node* final_io = final_state->in(TypeFunc::I_O);
1879 
1880   // Replace all the old call edges with the edges from the inlining result
1881   if (callprojs.fallthrough_catchproj != NULL) {
1882     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1883   }
1884   if (callprojs.fallthrough_memproj != NULL) {
1885     if (final_mem->is_MergeMem()) {
1886       // Parser's exits MergeMem was not transformed but may be optimized
1887       final_mem = _gvn.transform(final_mem);
1888     }
1889     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
1890   }
1891   if (callprojs.fallthrough_ioproj != NULL) {
1892     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
1893   }
1894 
1895   // Replace the result with the new result if it exists and is used
1896   if (callprojs.resproj != NULL && result != NULL) {
1897     C->gvn_replace_by(callprojs.resproj, result);
1898   }
1899 
1900   if (ejvms == NULL) {
1901     // No exception edges to simply kill off those paths
1902     if (callprojs.catchall_catchproj != NULL) {
1903       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1904     }
1905     if (callprojs.catchall_memproj != NULL) {
1906       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1907     }
1908     if (callprojs.catchall_ioproj != NULL) {
1909       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1910     }
1911     // Replace the old exception object with top
1912     if (callprojs.exobj != NULL) {
1913       C->gvn_replace_by(callprojs.exobj, C->top());
1914     }
1915   } else {
1916     GraphKit ekit(ejvms);
1917 
1918     // Load my combined exception state into the kit, with all phis transformed:
1919     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1920     replaced_nodes_exception = ex_map->replaced_nodes();
1921 
1922     Node* ex_oop = ekit.use_exception_state(ex_map);
1923 
1924     if (callprojs.catchall_catchproj != NULL) {
1925       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1926       ex_ctl = ekit.control();
1927     }
1928     if (callprojs.catchall_memproj != NULL) {
1929       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1930     }
1931     if (callprojs.catchall_ioproj != NULL) {
1932       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1933     }
1934 
1935     // Replace the old exception object with the newly created one
1936     if (callprojs.exobj != NULL) {
1937       C->gvn_replace_by(callprojs.exobj, ex_oop);
1938     }
1939   }
1940 
1941   // Disconnect the call from the graph
1942   call->disconnect_inputs(NULL, C);
1943   C->gvn_replace_by(call, C->top());
1944 
1945   // Clean up any MergeMems that feed other MergeMems since the
1946   // optimizer doesn't like that.
1947   if (final_mem->is_MergeMem()) {
1948     Node_List wl;
1949     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1950       Node* m = i.get();
1951       if (m->is_MergeMem() && !wl.contains(m)) {
1952         wl.push(m);
1953       }
1954     }
1955     while (wl.size()  > 0) {
1956       _gvn.transform(wl.pop());
1957     }
1958   }
1959 
1960   if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1961     replaced_nodes.apply(C, final_ctl);
1962   }
1963   if (!ex_ctl->is_top() && do_replaced_nodes) {
1964     replaced_nodes_exception.apply(C, ex_ctl);
1965   }
1966 }
1967 
1968 
1969 //------------------------------increment_counter------------------------------
1970 // for statistics: increment a VM counter by 1
1971 
1972 void GraphKit::increment_counter(address counter_addr) {
1973   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1974   increment_counter(adr1);
1975 }
1976 
1977 void GraphKit::increment_counter(Node* counter_addr) {
1978   int adr_type = Compile::AliasIdxRaw;
1979   Node* ctrl = control();
1980   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1981   Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1982   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1983 }
1984 
1985 
1986 //------------------------------uncommon_trap----------------------------------
1987 // Bail out to the interpreter in mid-method.  Implemented by calling the
1988 // uncommon_trap blob.  This helper function inserts a runtime call with the
1989 // right debug info.
1990 void GraphKit::uncommon_trap(int trap_request,
1991                              ciKlass* klass, const char* comment,
1992                              bool must_throw,
1993                              bool keep_exact_action) {
1994   if (failing())  stop();
1995   if (stopped())  return; // trap reachable?
1996 
1997   // Note:  If ProfileTraps is true, and if a deopt. actually
1998   // occurs here, the runtime will make sure an MDO exists.  There is
1999   // no need to call method()->ensure_method_data() at this point.
2000 
2001   // Set the stack pointer to the right value for reexecution:
2002   set_sp(reexecute_sp());
2003 
2004 #ifdef ASSERT
2005   if (!must_throw) {
2006     // Make sure the stack has at least enough depth to execute
2007     // the current bytecode.
2008     int inputs, ignored_depth;
2009     if (compute_stack_effects(inputs, ignored_depth)) {
2010       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2011              Bytecodes::name(java_bc()), sp(), inputs);
2012     }
2013   }
2014 #endif
2015 
2016   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2017   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2018 
2019   switch (action) {
2020   case Deoptimization::Action_maybe_recompile:
2021   case Deoptimization::Action_reinterpret:
2022     // Temporary fix for 6529811 to allow virtual calls to be sure they
2023     // get the chance to go from mono->bi->mega
2024     if (!keep_exact_action &&
2025         Deoptimization::trap_request_index(trap_request) < 0 &&
2026         too_many_recompiles(reason)) {
2027       // This BCI is causing too many recompilations.
2028       if (C->log() != NULL) {
2029         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2030                 Deoptimization::trap_reason_name(reason),
2031                 Deoptimization::trap_action_name(action));
2032       }
2033       action = Deoptimization::Action_none;
2034       trap_request = Deoptimization::make_trap_request(reason, action);
2035     } else {
2036       C->set_trap_can_recompile(true);
2037     }
2038     break;
2039   case Deoptimization::Action_make_not_entrant:
2040     C->set_trap_can_recompile(true);
2041     break;
2042   case Deoptimization::Action_none:
2043   case Deoptimization::Action_make_not_compilable:
2044     break;
2045   default:
2046 #ifdef ASSERT
2047     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2048 #endif
2049     break;
2050   }
2051 
2052   if (TraceOptoParse) {
2053     char buf[100];
2054     tty->print_cr("Uncommon trap %s at bci:%d",
2055                   Deoptimization::format_trap_request(buf, sizeof(buf),
2056                                                       trap_request), bci());
2057   }
2058 
2059   CompileLog* log = C->log();
2060   if (log != NULL) {
2061     int kid = (klass == NULL)? -1: log->identify(klass);
2062     log->begin_elem("uncommon_trap bci='%d'", bci());
2063     char buf[100];
2064     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2065                                                           trap_request));
2066     if (kid >= 0)         log->print(" klass='%d'", kid);
2067     if (comment != NULL)  log->print(" comment='%s'", comment);
2068     log->end_elem();
2069   }
2070 
2071   // Make sure any guarding test views this path as very unlikely
2072   Node *i0 = control()->in(0);
2073   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2074     IfNode *iff = i0->as_If();
2075     float f = iff->_prob;   // Get prob
2076     if (control()->Opcode() == Op_IfTrue) {
2077       if (f > PROB_UNLIKELY_MAG(4))
2078         iff->_prob = PROB_MIN;
2079     } else {
2080       if (f < PROB_LIKELY_MAG(4))
2081         iff->_prob = PROB_MAX;
2082     }
2083   }
2084 
2085   // Clear out dead values from the debug info.
2086   kill_dead_locals();
2087 
2088   // Now insert the uncommon trap subroutine call
2089   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2090   const TypePtr* no_memory_effects = NULL;
2091   // Pass the index of the class to be loaded
2092   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2093                                  (must_throw ? RC_MUST_THROW : 0),
2094                                  OptoRuntime::uncommon_trap_Type(),
2095                                  call_addr, "uncommon_trap", no_memory_effects,
2096                                  intcon(trap_request));
2097   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2098          "must extract request correctly from the graph");
2099   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2100 
2101   call->set_req(TypeFunc::ReturnAdr, returnadr());
2102   // The debug info is the only real input to this call.
2103 
2104   // Halt-and-catch fire here.  The above call should never return!
2105   HaltNode* halt = new HaltNode(control(), frameptr());
2106   _gvn.set_type_bottom(halt);
2107   root()->add_req(halt);
2108 
2109   stop_and_kill_map();
2110 }
2111 
2112 
2113 //--------------------------just_allocated_object------------------------------
2114 // Report the object that was just allocated.
2115 // It must be the case that there are no intervening safepoints.
2116 // We use this to determine if an object is so "fresh" that
2117 // it does not require card marks.
2118 Node* GraphKit::just_allocated_object(Node* current_control) {
2119   if (C->recent_alloc_ctl() == current_control)
2120     return C->recent_alloc_obj();
2121   return NULL;
2122 }
2123 
2124 
2125 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2126   // (Note:  TypeFunc::make has a cache that makes this fast.)
2127   const TypeFunc* tf    = TypeFunc::make(dest_method);
2128   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2129   for (int j = 0; j < nargs; j++) {
2130     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2131     if( targ->basic_type() == T_DOUBLE ) {
2132       // If any parameters are doubles, they must be rounded before
2133       // the call, dstore_rounding does gvn.transform
2134       Node *arg = argument(j);
2135       arg = dstore_rounding(arg);
2136       set_argument(j, arg);
2137     }
2138   }
2139 }
2140 
2141 /**
2142  * Record profiling data exact_kls for Node n with the type system so
2143  * that it can propagate it (speculation)
2144  *
2145  * @param n          node that the type applies to
2146  * @param exact_kls  type from profiling
2147  * @param maybe_null did profiling see null?
2148  *
2149  * @return           node with improved type
2150  */
2151 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2152   const Type* current_type = _gvn.type(n);
2153   assert(UseTypeSpeculation, "type speculation must be on");
2154 
2155   const TypePtr* speculative = current_type->speculative();
2156 
2157   // Should the klass from the profile be recorded in the speculative type?
2158   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2159     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2160     const TypeOopPtr* xtype = tklass->as_instance_type();
2161     assert(xtype->klass_is_exact(), "Should be exact");
2162     // Any reason to believe n is not null (from this profiling or a previous one)?
2163     assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2164     const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2165     // record the new speculative type's depth
2166     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2167     speculative = speculative->with_inline_depth(jvms()->depth());
2168   } else if (current_type->would_improve_ptr(ptr_kind)) {
2169     // Profiling report that null was never seen so we can change the
2170     // speculative type to non null ptr.
2171     if (ptr_kind == ProfileAlwaysNull) {
2172       speculative = TypePtr::NULL_PTR;
2173     } else {
2174       assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2175       const TypePtr* ptr = TypePtr::NOTNULL;
2176       if (speculative != NULL) {
2177         speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2178       } else {
2179         speculative = ptr;
2180       }
2181     }
2182   }
2183 
2184   if (speculative != current_type->speculative()) {
2185     // Build a type with a speculative type (what we think we know
2186     // about the type but will need a guard when we use it)
2187     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2188     // We're changing the type, we need a new CheckCast node to carry
2189     // the new type. The new type depends on the control: what
2190     // profiling tells us is only valid from here as far as we can
2191     // tell.
2192     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2193     cast = _gvn.transform(cast);
2194     replace_in_map(n, cast);
2195     n = cast;
2196   }
2197 
2198   return n;
2199 }
2200 
2201 /**
2202  * Record profiling data from receiver profiling at an invoke with the
2203  * type system so that it can propagate it (speculation)
2204  *
2205  * @param n  receiver node
2206  *
2207  * @return   node with improved type
2208  */
2209 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2210   if (!UseTypeSpeculation) {
2211     return n;
2212   }
2213   ciKlass* exact_kls = profile_has_unique_klass();
2214   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2215   if ((java_bc() == Bytecodes::_checkcast ||
2216        java_bc() == Bytecodes::_instanceof ||
2217        java_bc() == Bytecodes::_aastore) &&
2218       method()->method_data()->is_mature()) {
2219     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2220     if (data != NULL) {
2221       if (!data->as_BitData()->null_seen()) {
2222         ptr_kind = ProfileNeverNull;
2223       } else {
2224         assert(data->is_ReceiverTypeData(), "bad profile data type");
2225         ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2226         uint i = 0;
2227         for (; i < call->row_limit(); i++) {
2228           ciKlass* receiver = call->receiver(i);
2229           if (receiver != NULL) {
2230             break;
2231           }
2232         }
2233         ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2234       }
2235     }
2236   }
2237   return record_profile_for_speculation(n, exact_kls, ptr_kind);
2238 }
2239 
2240 /**
2241  * Record profiling data from argument profiling at an invoke with the
2242  * type system so that it can propagate it (speculation)
2243  *
2244  * @param dest_method  target method for the call
2245  * @param bc           what invoke bytecode is this?
2246  */
2247 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2248   if (!UseTypeSpeculation) {
2249     return;
2250   }
2251   const TypeFunc* tf    = TypeFunc::make(dest_method);
2252   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2253   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2254   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2255     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2256     if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2257       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2258       ciKlass* better_type = NULL;
2259       if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2260         record_profile_for_speculation(argument(j), better_type, ptr_kind);
2261       }
2262       i++;
2263     }
2264   }
2265 }
2266 
2267 /**
2268  * Record profiling data from parameter profiling at an invoke with
2269  * the type system so that it can propagate it (speculation)
2270  */
2271 void GraphKit::record_profiled_parameters_for_speculation() {
2272   if (!UseTypeSpeculation) {
2273     return;
2274   }
2275   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2276     if (_gvn.type(local(i))->isa_oopptr()) {
2277       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2278       ciKlass* better_type = NULL;
2279       if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2280         record_profile_for_speculation(local(i), better_type, ptr_kind);
2281       }
2282       j++;
2283     }
2284   }
2285 }
2286 
2287 /**
2288  * Record profiling data from return value profiling at an invoke with
2289  * the type system so that it can propagate it (speculation)
2290  */
2291 void GraphKit::record_profiled_return_for_speculation() {
2292   if (!UseTypeSpeculation) {
2293     return;
2294   }
2295   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2296   ciKlass* better_type = NULL;
2297   if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2298     // If profiling reports a single type for the return value,
2299     // feed it to the type system so it can propagate it as a
2300     // speculative type
2301     record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2302   }
2303 }
2304 
2305 void GraphKit::round_double_result(ciMethod* dest_method) {
2306   // A non-strict method may return a double value which has an extended
2307   // exponent, but this must not be visible in a caller which is 'strict'
2308   // If a strict caller invokes a non-strict callee, round a double result
2309 
2310   BasicType result_type = dest_method->return_type()->basic_type();
2311   assert( method() != NULL, "must have caller context");
2312   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2313     // Destination method's return value is on top of stack
2314     // dstore_rounding() does gvn.transform
2315     Node *result = pop_pair();
2316     result = dstore_rounding(result);
2317     push_pair(result);
2318   }
2319 }
2320 
2321 // rounding for strict float precision conformance
2322 Node* GraphKit::precision_rounding(Node* n) {
2323   return UseStrictFP && _method->flags().is_strict()
2324     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2325     ? _gvn.transform( new RoundFloatNode(0, n) )
2326     : n;
2327 }
2328 
2329 // rounding for strict double precision conformance
2330 Node* GraphKit::dprecision_rounding(Node *n) {
2331   return UseStrictFP && _method->flags().is_strict()
2332     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2333     ? _gvn.transform( new RoundDoubleNode(0, n) )
2334     : n;
2335 }
2336 
2337 // rounding for non-strict double stores
2338 Node* GraphKit::dstore_rounding(Node* n) {
2339   return Matcher::strict_fp_requires_explicit_rounding
2340     && UseSSE <= 1
2341     ? _gvn.transform( new RoundDoubleNode(0, n) )
2342     : n;
2343 }
2344 
2345 //=============================================================================
2346 // Generate a fast path/slow path idiom.  Graph looks like:
2347 // [foo] indicates that 'foo' is a parameter
2348 //
2349 //              [in]     NULL
2350 //                 \    /
2351 //                  CmpP
2352 //                  Bool ne
2353 //                   If
2354 //                  /  \
2355 //              True    False-<2>
2356 //              / |
2357 //             /  cast_not_null
2358 //           Load  |    |   ^
2359 //        [fast_test]   |   |
2360 // gvn to   opt_test    |   |
2361 //          /    \      |  <1>
2362 //      True     False  |
2363 //        |         \\  |
2364 //   [slow_call]     \[fast_result]
2365 //    Ctl   Val       \      \
2366 //     |               \      \
2367 //    Catch       <1>   \      \
2368 //   /    \        ^     \      \
2369 //  Ex    No_Ex    |      \      \
2370 //  |       \   \  |       \ <2>  \
2371 //  ...      \  [slow_res] |  |    \   [null_result]
2372 //            \         \--+--+---  |  |
2373 //             \           | /    \ | /
2374 //              --------Region     Phi
2375 //
2376 //=============================================================================
2377 // Code is structured as a series of driver functions all called 'do_XXX' that
2378 // call a set of helper functions.  Helper functions first, then drivers.
2379 
2380 //------------------------------null_check_oop---------------------------------
2381 // Null check oop.  Set null-path control into Region in slot 3.
2382 // Make a cast-not-nullness use the other not-null control.  Return cast.
2383 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2384                                bool never_see_null,
2385                                bool safe_for_replace,
2386                                bool speculative) {
2387   // Initial NULL check taken path
2388   (*null_control) = top();
2389   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2390 
2391   // Generate uncommon_trap:
2392   if (never_see_null && (*null_control) != top()) {
2393     // If we see an unexpected null at a check-cast we record it and force a
2394     // recompile; the offending check-cast will be compiled to handle NULLs.
2395     // If we see more than one offending BCI, then all checkcasts in the
2396     // method will be compiled to handle NULLs.
2397     PreserveJVMState pjvms(this);
2398     set_control(*null_control);
2399     replace_in_map(value, null());
2400     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2401     uncommon_trap(reason,
2402                   Deoptimization::Action_make_not_entrant);
2403     (*null_control) = top();    // NULL path is dead
2404   }
2405   if ((*null_control) == top() && safe_for_replace) {
2406     replace_in_map(value, cast);
2407   }
2408 
2409   // Cast away null-ness on the result
2410   return cast;
2411 }
2412 
2413 //------------------------------opt_iff----------------------------------------
2414 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2415 // Return slow-path control.
2416 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2417   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2418 
2419   // Fast path taken; set region slot 2
2420   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2421   region->init_req(2,fast_taken); // Capture fast-control
2422 
2423   // Fast path not-taken, i.e. slow path
2424   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2425   return slow_taken;
2426 }
2427 
2428 //-----------------------------make_runtime_call-------------------------------
2429 Node* GraphKit::make_runtime_call(int flags,
2430                                   const TypeFunc* call_type, address call_addr,
2431                                   const char* call_name,
2432                                   const TypePtr* adr_type,
2433                                   // The following parms are all optional.
2434                                   // The first NULL ends the list.
2435                                   Node* parm0, Node* parm1,
2436                                   Node* parm2, Node* parm3,
2437                                   Node* parm4, Node* parm5,
2438                                   Node* parm6, Node* parm7) {
2439   assert(call_addr != NULL, "must not call NULL targets");
2440 
2441   // Slow-path call
2442   bool is_leaf = !(flags & RC_NO_LEAF);
2443   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2444   if (call_name == NULL) {
2445     assert(!is_leaf, "must supply name for leaf");
2446     call_name = OptoRuntime::stub_name(call_addr);
2447   }
2448   CallNode* call;
2449   if (!is_leaf) {
2450     call = new CallStaticJavaNode(call_type, call_addr, call_name,
2451                                            bci(), adr_type);
2452   } else if (flags & RC_NO_FP) {
2453     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2454   } else {
2455     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2456   }
2457 
2458   // The following is similar to set_edges_for_java_call,
2459   // except that the memory effects of the call are restricted to AliasIdxRaw.
2460 
2461   // Slow path call has no side-effects, uses few values
2462   bool wide_in  = !(flags & RC_NARROW_MEM);
2463   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2464 
2465   Node* prev_mem = NULL;
2466   if (wide_in) {
2467     prev_mem = set_predefined_input_for_runtime_call(call);
2468   } else {
2469     assert(!wide_out, "narrow in => narrow out");
2470     Node* narrow_mem = memory(adr_type);
2471     prev_mem = reset_memory();
2472     map()->set_memory(narrow_mem);
2473     set_predefined_input_for_runtime_call(call);
2474   }
2475 
2476   // Hook each parm in order.  Stop looking at the first NULL.
2477   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2478   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2479   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2480   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2481   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2482   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2483   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2484   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2485     /* close each nested if ===> */  } } } } } } } }
2486   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2487 
2488   if (!is_leaf) {
2489     // Non-leaves can block and take safepoints:
2490     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2491   }
2492   // Non-leaves can throw exceptions:
2493   if (has_io) {
2494     call->set_req(TypeFunc::I_O, i_o());
2495   }
2496 
2497   if (flags & RC_UNCOMMON) {
2498     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2499     // (An "if" probability corresponds roughly to an unconditional count.
2500     // Sort of.)
2501     call->set_cnt(PROB_UNLIKELY_MAG(4));
2502   }
2503 
2504   Node* c = _gvn.transform(call);
2505   assert(c == call, "cannot disappear");
2506 
2507   if (wide_out) {
2508     // Slow path call has full side-effects.
2509     set_predefined_output_for_runtime_call(call);
2510   } else {
2511     // Slow path call has few side-effects, and/or sets few values.
2512     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2513   }
2514 
2515   if (has_io) {
2516     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2517   }
2518   return call;
2519 
2520 }
2521 
2522 //------------------------------merge_memory-----------------------------------
2523 // Merge memory from one path into the current memory state.
2524 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2525   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2526     Node* old_slice = mms.force_memory();
2527     Node* new_slice = mms.memory2();
2528     if (old_slice != new_slice) {
2529       PhiNode* phi;
2530       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2531         if (mms.is_empty()) {
2532           // clone base memory Phi's inputs for this memory slice
2533           assert(old_slice == mms.base_memory(), "sanity");
2534           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2535           _gvn.set_type(phi, Type::MEMORY);
2536           for (uint i = 1; i < phi->req(); i++) {
2537             phi->init_req(i, old_slice->in(i));
2538           }
2539         } else {
2540           phi = old_slice->as_Phi(); // Phi was generated already
2541         }
2542       } else {
2543         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2544         _gvn.set_type(phi, Type::MEMORY);
2545       }
2546       phi->set_req(new_path, new_slice);
2547       mms.set_memory(phi);
2548     }
2549   }
2550 }
2551 
2552 //------------------------------make_slow_call_ex------------------------------
2553 // Make the exception handler hookups for the slow call
2554 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2555   if (stopped())  return;
2556 
2557   // Make a catch node with just two handlers:  fall-through and catch-all
2558   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2559   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2560   Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2561   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2562 
2563   { PreserveJVMState pjvms(this);
2564     set_control(excp);
2565     set_i_o(i_o);
2566 
2567     if (excp != top()) {
2568       if (deoptimize) {
2569         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2570         uncommon_trap(Deoptimization::Reason_unhandled,
2571                       Deoptimization::Action_none);
2572       } else {
2573         // Create an exception state also.
2574         // Use an exact type if the caller has specified a specific exception.
2575         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2576         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2577         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2578       }
2579     }
2580   }
2581 
2582   // Get the no-exception control from the CatchNode.
2583   set_control(norm);
2584 }
2585 
2586 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2587   Node* cmp = NULL;
2588   switch(bt) {
2589   case T_INT: cmp = new CmpINode(in1, in2); break;
2590   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2591   default: fatal("unexpected comparison type %s", type2name(bt));
2592   }
2593   gvn->transform(cmp);
2594   Node* bol = gvn->transform(new BoolNode(cmp, test));
2595   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2596   gvn->transform(iff);
2597   if (!bol->is_Con()) gvn->record_for_igvn(iff);
2598   return iff;
2599 }
2600 
2601 
2602 //-------------------------------gen_subtype_check-----------------------------
2603 // Generate a subtyping check.  Takes as input the subtype and supertype.
2604 // Returns 2 values: sets the default control() to the true path and returns
2605 // the false path.  Only reads invariant memory; sets no (visible) memory.
2606 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2607 // but that's not exposed to the optimizer.  This call also doesn't take in an
2608 // Object; if you wish to check an Object you need to load the Object's class
2609 // prior to coming here.
2610 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2611   Compile* C = gvn->C;
2612 
2613   if ((*ctrl)->is_top()) {
2614     return C->top();
2615   }
2616 
2617   // Fast check for identical types, perhaps identical constants.
2618   // The types can even be identical non-constants, in cases
2619   // involving Array.newInstance, Object.clone, etc.
2620   if (subklass == superklass)
2621     return C->top();             // false path is dead; no test needed.
2622 
2623   if (gvn->type(superklass)->singleton()) {
2624     ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2625     ciKlass* subk   = gvn->type(subklass)->is_klassptr()->klass();
2626 
2627     // In the common case of an exact superklass, try to fold up the
2628     // test before generating code.  You may ask, why not just generate
2629     // the code and then let it fold up?  The answer is that the generated
2630     // code will necessarily include null checks, which do not always
2631     // completely fold away.  If they are also needless, then they turn
2632     // into a performance loss.  Example:
2633     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2634     // Here, the type of 'fa' is often exact, so the store check
2635     // of fa[1]=x will fold up, without testing the nullness of x.
2636     switch (C->static_subtype_check(superk, subk)) {
2637     case Compile::SSC_always_false:
2638       {
2639         Node* always_fail = *ctrl;
2640         *ctrl = gvn->C->top();
2641         return always_fail;
2642       }
2643     case Compile::SSC_always_true:
2644       return C->top();
2645     case Compile::SSC_easy_test:
2646       {
2647         // Just do a direct pointer compare and be done.
2648         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2649         *ctrl = gvn->transform(new IfTrueNode(iff));
2650         return gvn->transform(new IfFalseNode(iff));
2651       }
2652     case Compile::SSC_full_test:
2653       break;
2654     default:
2655       ShouldNotReachHere();
2656     }
2657   }
2658 
2659   // %%% Possible further optimization:  Even if the superklass is not exact,
2660   // if the subklass is the unique subtype of the superklass, the check
2661   // will always succeed.  We could leave a dependency behind to ensure this.
2662 
2663   // First load the super-klass's check-offset
2664   Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2665   Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2666   Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2667   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2668   bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2669 
2670   // Load from the sub-klass's super-class display list, or a 1-word cache of
2671   // the secondary superclass list, or a failing value with a sentinel offset
2672   // if the super-klass is an interface or exceptionally deep in the Java
2673   // hierarchy and we have to scan the secondary superclass list the hard way.
2674   // Worst-case type is a little odd: NULL is allowed as a result (usually
2675   // klass loads can never produce a NULL).
2676   Node *chk_off_X = chk_off;
2677 #ifdef _LP64
2678   chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2679 #endif
2680   Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2681   // For some types like interfaces the following loadKlass is from a 1-word
2682   // cache which is mutable so can't use immutable memory.  Other
2683   // types load from the super-class display table which is immutable.
2684   m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2685   Node *kmem = might_be_cache ? m : C->immutable_memory();
2686   Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2687 
2688   // Compile speed common case: ARE a subtype and we canNOT fail
2689   if( superklass == nkls )
2690     return C->top();             // false path is dead; no test needed.
2691 
2692   // See if we get an immediate positive hit.  Happens roughly 83% of the
2693   // time.  Test to see if the value loaded just previously from the subklass
2694   // is exactly the superklass.
2695   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2696   Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2697   *ctrl = gvn->transform(new IfFalseNode(iff1));
2698 
2699   // Compile speed common case: Check for being deterministic right now.  If
2700   // chk_off is a constant and not equal to cacheoff then we are NOT a
2701   // subklass.  In this case we need exactly the 1 test above and we can
2702   // return those results immediately.
2703   if (!might_be_cache) {
2704     Node* not_subtype_ctrl = *ctrl;
2705     *ctrl = iftrue1; // We need exactly the 1 test above
2706     return not_subtype_ctrl;
2707   }
2708 
2709   // Gather the various success & failures here
2710   RegionNode *r_ok_subtype = new RegionNode(4);
2711   gvn->record_for_igvn(r_ok_subtype);
2712   RegionNode *r_not_subtype = new RegionNode(3);
2713   gvn->record_for_igvn(r_not_subtype);
2714 
2715   r_ok_subtype->init_req(1, iftrue1);
2716 
2717   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2718   // is roughly 63% of the remaining cases).  Test to see if the loaded
2719   // check-offset points into the subklass display list or the 1-element
2720   // cache.  If it points to the display (and NOT the cache) and the display
2721   // missed then it's not a subtype.
2722   Node *cacheoff = gvn->intcon(cacheoff_con);
2723   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2724   r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2725   *ctrl = gvn->transform(new IfFalseNode(iff2));
2726 
2727   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2728   // No performance impact (too rare) but allows sharing of secondary arrays
2729   // which has some footprint reduction.
2730   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2731   r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2732   *ctrl = gvn->transform(new IfFalseNode(iff3));
2733 
2734   // -- Roads not taken here: --
2735   // We could also have chosen to perform the self-check at the beginning
2736   // of this code sequence, as the assembler does.  This would not pay off
2737   // the same way, since the optimizer, unlike the assembler, can perform
2738   // static type analysis to fold away many successful self-checks.
2739   // Non-foldable self checks work better here in second position, because
2740   // the initial primary superclass check subsumes a self-check for most
2741   // types.  An exception would be a secondary type like array-of-interface,
2742   // which does not appear in its own primary supertype display.
2743   // Finally, we could have chosen to move the self-check into the
2744   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2745   // dependent manner.  But it is worthwhile to have the check here,
2746   // where it can be perhaps be optimized.  The cost in code space is
2747   // small (register compare, branch).
2748 
2749   // Now do a linear scan of the secondary super-klass array.  Again, no real
2750   // performance impact (too rare) but it's gotta be done.
2751   // Since the code is rarely used, there is no penalty for moving it
2752   // out of line, and it can only improve I-cache density.
2753   // The decision to inline or out-of-line this final check is platform
2754   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2755   Node* psc = gvn->transform(
2756     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2757 
2758   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2759   r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2760   r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2761 
2762   // Return false path; set default control to true path.
2763   *ctrl = gvn->transform(r_ok_subtype);
2764   return gvn->transform(r_not_subtype);
2765 }
2766 
2767 // Profile-driven exact type check:
2768 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2769                                     float prob,
2770                                     Node* *casted_receiver) {
2771   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2772   Node* recv_klass = load_object_klass(receiver);
2773   Node* want_klass = makecon(tklass);
2774   Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2775   Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2776   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2777   set_control( _gvn.transform( new IfTrueNode (iff) ));
2778   Node* fail = _gvn.transform( new IfFalseNode(iff) );
2779 
2780   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2781   assert(recv_xtype->klass_is_exact(), "");
2782 
2783   // Subsume downstream occurrences of receiver with a cast to
2784   // recv_xtype, since now we know what the type will be.
2785   Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2786   (*casted_receiver) = _gvn.transform(cast);
2787   // (User must make the replace_in_map call.)
2788 
2789   return fail;
2790 }
2791 
2792 
2793 //------------------------------seems_never_null-------------------------------
2794 // Use null_seen information if it is available from the profile.
2795 // If we see an unexpected null at a type check we record it and force a
2796 // recompile; the offending check will be recompiled to handle NULLs.
2797 // If we see several offending BCIs, then all checks in the
2798 // method will be recompiled.
2799 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2800   speculating = !_gvn.type(obj)->speculative_maybe_null();
2801   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2802   if (UncommonNullCast               // Cutout for this technique
2803       && obj != null()               // And not the -Xcomp stupid case?
2804       && !too_many_traps(reason)
2805       ) {
2806     if (speculating) {
2807       return true;
2808     }
2809     if (data == NULL)
2810       // Edge case:  no mature data.  Be optimistic here.
2811       return true;
2812     // If the profile has not seen a null, assume it won't happen.
2813     assert(java_bc() == Bytecodes::_checkcast ||
2814            java_bc() == Bytecodes::_instanceof ||
2815            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2816     return !data->as_BitData()->null_seen();
2817   }
2818   speculating = false;
2819   return false;
2820 }
2821 
2822 //------------------------maybe_cast_profiled_receiver-------------------------
2823 // If the profile has seen exactly one type, narrow to exactly that type.
2824 // Subsequent type checks will always fold up.
2825 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2826                                              ciKlass* require_klass,
2827                                              ciKlass* spec_klass,
2828                                              bool safe_for_replace) {
2829   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2830 
2831   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2832 
2833   // Make sure we haven't already deoptimized from this tactic.
2834   if (too_many_traps(reason) || too_many_recompiles(reason))
2835     return NULL;
2836 
2837   // (No, this isn't a call, but it's enough like a virtual call
2838   // to use the same ciMethod accessor to get the profile info...)
2839   // If we have a speculative type use it instead of profiling (which
2840   // may not help us)
2841   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2842   if (exact_kls != NULL) {// no cast failures here
2843     if (require_klass == NULL ||
2844         C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2845       // If we narrow the type to match what the type profile sees or
2846       // the speculative type, we can then remove the rest of the
2847       // cast.
2848       // This is a win, even if the exact_kls is very specific,
2849       // because downstream operations, such as method calls,
2850       // will often benefit from the sharper type.
2851       Node* exact_obj = not_null_obj; // will get updated in place...
2852       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2853                                             &exact_obj);
2854       { PreserveJVMState pjvms(this);
2855         set_control(slow_ctl);
2856         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2857       }
2858       if (safe_for_replace) {
2859         replace_in_map(not_null_obj, exact_obj);
2860       }
2861       return exact_obj;
2862     }
2863     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2864   }
2865 
2866   return NULL;
2867 }
2868 
2869 /**
2870  * Cast obj to type and emit guard unless we had too many traps here
2871  * already
2872  *
2873  * @param obj       node being casted
2874  * @param type      type to cast the node to
2875  * @param not_null  true if we know node cannot be null
2876  */
2877 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2878                                         ciKlass* type,
2879                                         bool not_null) {
2880   if (stopped()) {
2881     return obj;
2882   }
2883 
2884   // type == NULL if profiling tells us this object is always null
2885   if (type != NULL) {
2886     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2887     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2888 
2889     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2890         !too_many_traps(class_reason) &&
2891         !too_many_recompiles(class_reason)) {
2892       Node* not_null_obj = NULL;
2893       // not_null is true if we know the object is not null and
2894       // there's no need for a null check
2895       if (!not_null) {
2896         Node* null_ctl = top();
2897         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2898         assert(null_ctl->is_top(), "no null control here");
2899       } else {
2900         not_null_obj = obj;
2901       }
2902 
2903       Node* exact_obj = not_null_obj;
2904       ciKlass* exact_kls = type;
2905       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2906                                             &exact_obj);
2907       {
2908         PreserveJVMState pjvms(this);
2909         set_control(slow_ctl);
2910         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2911       }
2912       replace_in_map(not_null_obj, exact_obj);
2913       obj = exact_obj;
2914     }
2915   } else {
2916     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2917         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2918       Node* exact_obj = null_assert(obj);
2919       replace_in_map(obj, exact_obj);
2920       obj = exact_obj;
2921     }
2922   }
2923   return obj;
2924 }
2925 
2926 //-------------------------------gen_instanceof--------------------------------
2927 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2928 // and the reflective instance-of call.
2929 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2930   kill_dead_locals();           // Benefit all the uncommon traps
2931   assert( !stopped(), "dead parse path should be checked in callers" );
2932   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2933          "must check for not-null not-dead klass in callers");
2934 
2935   // Make the merge point
2936   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2937   RegionNode* region = new RegionNode(PATH_LIMIT);
2938   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
2939   C->set_has_split_ifs(true); // Has chance for split-if optimization
2940 
2941   ciProfileData* data = NULL;
2942   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2943     data = method()->method_data()->bci_to_data(bci());
2944   }
2945   bool speculative_not_null = false;
2946   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2947                          && seems_never_null(obj, data, speculative_not_null));
2948 
2949   // Null check; get casted pointer; set region slot 3
2950   Node* null_ctl = top();
2951   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2952 
2953   // If not_null_obj is dead, only null-path is taken
2954   if (stopped()) {              // Doing instance-of on a NULL?
2955     set_control(null_ctl);
2956     return intcon(0);
2957   }
2958   region->init_req(_null_path, null_ctl);
2959   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2960   if (null_ctl == top()) {
2961     // Do this eagerly, so that pattern matches like is_diamond_phi
2962     // will work even during parsing.
2963     assert(_null_path == PATH_LIMIT-1, "delete last");
2964     region->del_req(_null_path);
2965     phi   ->del_req(_null_path);
2966   }
2967 
2968   // Do we know the type check always succeed?
2969   bool known_statically = false;
2970   if (_gvn.type(superklass)->singleton()) {
2971     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2972     ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2973     if (subk != NULL && subk->is_loaded()) {
2974       int static_res = C->static_subtype_check(superk, subk);
2975       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2976     }
2977   }
2978 
2979   if (!known_statically) {
2980     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2981     // We may not have profiling here or it may not help us. If we
2982     // have a speculative type use it to perform an exact cast.
2983     ciKlass* spec_obj_type = obj_type->speculative_type();
2984     if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2985       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2986       if (stopped()) {            // Profile disagrees with this path.
2987         set_control(null_ctl);    // Null is the only remaining possibility.
2988         return intcon(0);
2989       }
2990       if (cast_obj != NULL) {
2991         not_null_obj = cast_obj;
2992       }
2993     }
2994   }
2995 
2996   // Load the object's klass
2997   Node* obj_klass = load_object_klass(not_null_obj);
2998 
2999   // Generate the subtype check
3000   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3001 
3002   // Plug in the success path to the general merge in slot 1.
3003   region->init_req(_obj_path, control());
3004   phi   ->init_req(_obj_path, intcon(1));
3005 
3006   // Plug in the failing path to the general merge in slot 2.
3007   region->init_req(_fail_path, not_subtype_ctrl);
3008   phi   ->init_req(_fail_path, intcon(0));
3009 
3010   // Return final merged results
3011   set_control( _gvn.transform(region) );
3012   record_for_igvn(region);
3013 
3014   // If we know the type check always succeeds then we don't use the
3015   // profiling data at this bytecode. Don't lose it, feed it to the
3016   // type system as a speculative type.
3017   if (safe_for_replace) {
3018     Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3019     replace_in_map(obj, casted_obj);
3020   }
3021 
3022   return _gvn.transform(phi);
3023 }
3024 
3025 //-------------------------------gen_checkcast---------------------------------
3026 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
3027 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
3028 // uncommon-trap paths work.  Adjust stack after this call.
3029 // If failure_control is supplied and not null, it is filled in with
3030 // the control edge for the cast failure.  Otherwise, an appropriate
3031 // uncommon trap or exception is thrown.
3032 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3033                               Node* *failure_control) {
3034   kill_dead_locals();           // Benefit all the uncommon traps
3035   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3036   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3037 
3038   // Fast cutout:  Check the case that the cast is vacuously true.
3039   // This detects the common cases where the test will short-circuit
3040   // away completely.  We do this before we perform the null check,
3041   // because if the test is going to turn into zero code, we don't
3042   // want a residual null check left around.  (Causes a slowdown,
3043   // for example, in some objArray manipulations, such as a[i]=a[j].)
3044   if (tk->singleton()) {
3045     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3046     if (objtp != NULL && objtp->klass() != NULL) {
3047       switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3048       case Compile::SSC_always_true:
3049         // If we know the type check always succeed then we don't use
3050         // the profiling data at this bytecode. Don't lose it, feed it
3051         // to the type system as a speculative type.
3052         return record_profiled_receiver_for_speculation(obj);
3053       case Compile::SSC_always_false:
3054         // It needs a null check because a null will *pass* the cast check.
3055         // A non-null value will always produce an exception.
3056         return null_assert(obj);
3057       }
3058     }
3059   }
3060 
3061   ciProfileData* data = NULL;
3062   bool safe_for_replace = false;
3063   if (failure_control == NULL) {        // use MDO in regular case only
3064     assert(java_bc() == Bytecodes::_aastore ||
3065            java_bc() == Bytecodes::_checkcast,
3066            "interpreter profiles type checks only for these BCs");
3067     data = method()->method_data()->bci_to_data(bci());
3068     safe_for_replace = true;
3069   }
3070 
3071   // Make the merge point
3072   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3073   RegionNode* region = new RegionNode(PATH_LIMIT);
3074   Node*       phi    = new PhiNode(region, toop);
3075   C->set_has_split_ifs(true); // Has chance for split-if optimization
3076 
3077   // Use null-cast information if it is available
3078   bool speculative_not_null = false;
3079   bool never_see_null = ((failure_control == NULL)  // regular case only
3080                          && seems_never_null(obj, data, speculative_not_null));
3081 
3082   // Null check; get casted pointer; set region slot 3
3083   Node* null_ctl = top();
3084   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3085 
3086   // If not_null_obj is dead, only null-path is taken
3087   if (stopped()) {              // Doing instance-of on a NULL?
3088     set_control(null_ctl);
3089     return null();
3090   }
3091   region->init_req(_null_path, null_ctl);
3092   phi   ->init_req(_null_path, null());  // Set null path value
3093   if (null_ctl == top()) {
3094     // Do this eagerly, so that pattern matches like is_diamond_phi
3095     // will work even during parsing.
3096     assert(_null_path == PATH_LIMIT-1, "delete last");
3097     region->del_req(_null_path);
3098     phi   ->del_req(_null_path);
3099   }
3100 
3101   Node* cast_obj = NULL;
3102   if (tk->klass_is_exact()) {
3103     // The following optimization tries to statically cast the speculative type of the object
3104     // (for example obtained during profiling) to the type of the superklass and then do a
3105     // dynamic check that the type of the object is what we expect. To work correctly
3106     // for checkcast and aastore the type of superklass should be exact.
3107     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3108     // We may not have profiling here or it may not help us. If we have
3109     // a speculative type use it to perform an exact cast.
3110     ciKlass* spec_obj_type = obj_type->speculative_type();
3111     if (spec_obj_type != NULL || data != NULL) {
3112       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3113       if (cast_obj != NULL) {
3114         if (failure_control != NULL) // failure is now impossible
3115           (*failure_control) = top();
3116         // adjust the type of the phi to the exact klass:
3117         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3118       }
3119     }
3120   }
3121 
3122   if (cast_obj == NULL) {
3123     // Load the object's klass
3124     Node* obj_klass = load_object_klass(not_null_obj);
3125 
3126     // Generate the subtype check
3127     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3128 
3129     // Plug in success path into the merge
3130     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3131     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3132     if (failure_control == NULL) {
3133       if (not_subtype_ctrl != top()) { // If failure is possible
3134         PreserveJVMState pjvms(this);
3135         set_control(not_subtype_ctrl);
3136         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3137       }
3138     } else {
3139       (*failure_control) = not_subtype_ctrl;
3140     }
3141   }
3142 
3143   region->init_req(_obj_path, control());
3144   phi   ->init_req(_obj_path, cast_obj);
3145 
3146   // A merge of NULL or Casted-NotNull obj
3147   Node* res = _gvn.transform(phi);
3148 
3149   // Note I do NOT always 'replace_in_map(obj,result)' here.
3150   //  if( tk->klass()->can_be_primary_super()  )
3151     // This means that if I successfully store an Object into an array-of-String
3152     // I 'forget' that the Object is really now known to be a String.  I have to
3153     // do this because we don't have true union types for interfaces - if I store
3154     // a Baz into an array-of-Interface and then tell the optimizer it's an
3155     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3156     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3157   //  replace_in_map( obj, res );
3158 
3159   // Return final merged results
3160   set_control( _gvn.transform(region) );
3161   record_for_igvn(region);
3162 
3163   return record_profiled_receiver_for_speculation(res);
3164 }
3165 
3166 //------------------------------next_monitor-----------------------------------
3167 // What number should be given to the next monitor?
3168 int GraphKit::next_monitor() {
3169   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3170   int next = current + C->sync_stack_slots();
3171   // Keep the toplevel high water mark current:
3172   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3173   return current;
3174 }
3175 
3176 //------------------------------insert_mem_bar---------------------------------
3177 // Memory barrier to avoid floating things around
3178 // The membar serves as a pinch point between both control and all memory slices.
3179 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3180   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3181   mb->init_req(TypeFunc::Control, control());
3182   mb->init_req(TypeFunc::Memory,  reset_memory());
3183   Node* membar = _gvn.transform(mb);
3184   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3185   set_all_memory_call(membar);
3186   return membar;
3187 }
3188 
3189 //-------------------------insert_mem_bar_volatile----------------------------
3190 // Memory barrier to avoid floating things around
3191 // The membar serves as a pinch point between both control and memory(alias_idx).
3192 // If you want to make a pinch point on all memory slices, do not use this
3193 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3194 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3195   // When Parse::do_put_xxx updates a volatile field, it appends a series
3196   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3197   // The first membar is on the same memory slice as the field store opcode.
3198   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3199   // All the other membars (for other volatile slices, including AliasIdxBot,
3200   // which stands for all unknown volatile slices) are control-dependent
3201   // on the first membar.  This prevents later volatile loads or stores
3202   // from sliding up past the just-emitted store.
3203 
3204   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3205   mb->set_req(TypeFunc::Control,control());
3206   if (alias_idx == Compile::AliasIdxBot) {
3207     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3208   } else {
3209     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3210     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3211   }
3212   Node* membar = _gvn.transform(mb);
3213   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3214   if (alias_idx == Compile::AliasIdxBot) {
3215     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3216   } else {
3217     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3218   }
3219   return membar;
3220 }
3221 
3222 //------------------------------shared_lock------------------------------------
3223 // Emit locking code.
3224 FastLockNode* GraphKit::shared_lock(Node* obj) {
3225   // bci is either a monitorenter bc or InvocationEntryBci
3226   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3227   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3228 
3229   if( !GenerateSynchronizationCode )
3230     return NULL;                // Not locking things?
3231   if (stopped())                // Dead monitor?
3232     return NULL;
3233 
3234   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3235 
3236   obj = access_resolve_for_write(obj);
3237 
3238   // Box the stack location
3239   Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3240   Node* mem = reset_memory();
3241 
3242   FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3243   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3244     // Create the counters for this fast lock.
3245     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3246   }
3247 
3248   // Create the rtm counters for this fast lock if needed.
3249   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3250 
3251   // Add monitor to debug info for the slow path.  If we block inside the
3252   // slow path and de-opt, we need the monitor hanging around
3253   map()->push_monitor( flock );
3254 
3255   const TypeFunc *tf = LockNode::lock_type();
3256   LockNode *lock = new LockNode(C, tf);
3257 
3258   lock->init_req( TypeFunc::Control, control() );
3259   lock->init_req( TypeFunc::Memory , mem );
3260   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3261   lock->init_req( TypeFunc::FramePtr, frameptr() );
3262   lock->init_req( TypeFunc::ReturnAdr, top() );
3263 
3264   lock->init_req(TypeFunc::Parms + 0, obj);
3265   lock->init_req(TypeFunc::Parms + 1, box);
3266   lock->init_req(TypeFunc::Parms + 2, flock);
3267   add_safepoint_edges(lock);
3268 
3269   lock = _gvn.transform( lock )->as_Lock();
3270 
3271   // lock has no side-effects, sets few values
3272   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3273 
3274   insert_mem_bar(Op_MemBarAcquireLock);
3275 
3276   // Add this to the worklist so that the lock can be eliminated
3277   record_for_igvn(lock);
3278 
3279 #ifndef PRODUCT
3280   if (PrintLockStatistics) {
3281     // Update the counter for this lock.  Don't bother using an atomic
3282     // operation since we don't require absolute accuracy.
3283     lock->create_lock_counter(map()->jvms());
3284     increment_counter(lock->counter()->addr());
3285   }
3286 #endif
3287 
3288   return flock;
3289 }
3290 
3291 
3292 //------------------------------shared_unlock----------------------------------
3293 // Emit unlocking code.
3294 void GraphKit::shared_unlock(Node* box, Node* obj) {
3295   // bci is either a monitorenter bc or InvocationEntryBci
3296   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3297   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3298 
3299   if( !GenerateSynchronizationCode )
3300     return;
3301   if (stopped()) {               // Dead monitor?
3302     map()->pop_monitor();        // Kill monitor from debug info
3303     return;
3304   }
3305 
3306   // Memory barrier to avoid floating things down past the locked region
3307   insert_mem_bar(Op_MemBarReleaseLock);
3308 
3309   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3310   UnlockNode *unlock = new UnlockNode(C, tf);
3311 #ifdef ASSERT
3312   unlock->set_dbg_jvms(sync_jvms());
3313 #endif
3314   uint raw_idx = Compile::AliasIdxRaw;
3315   unlock->init_req( TypeFunc::Control, control() );
3316   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3317   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3318   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3319   unlock->init_req( TypeFunc::ReturnAdr, top() );
3320 
3321   unlock->init_req(TypeFunc::Parms + 0, obj);
3322   unlock->init_req(TypeFunc::Parms + 1, box);
3323   unlock = _gvn.transform(unlock)->as_Unlock();
3324 
3325   Node* mem = reset_memory();
3326 
3327   // unlock has no side-effects, sets few values
3328   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3329 
3330   // Kill monitor from debug info
3331   map()->pop_monitor( );
3332 }
3333 
3334 //-------------------------------get_layout_helper-----------------------------
3335 // If the given klass is a constant or known to be an array,
3336 // fetch the constant layout helper value into constant_value
3337 // and return (Node*)NULL.  Otherwise, load the non-constant
3338 // layout helper value, and return the node which represents it.
3339 // This two-faced routine is useful because allocation sites
3340 // almost always feature constant types.
3341 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3342   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3343   if (!StressReflectiveCode && inst_klass != NULL) {
3344     ciKlass* klass = inst_klass->klass();
3345     bool    xklass = inst_klass->klass_is_exact();
3346     if (xklass || klass->is_array_klass()) {
3347       jint lhelper = klass->layout_helper();
3348       if (lhelper != Klass::_lh_neutral_value) {
3349         constant_value = lhelper;
3350         return (Node*) NULL;
3351       }
3352     }
3353   }
3354   constant_value = Klass::_lh_neutral_value;  // put in a known value
3355   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3356   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3357 }
3358 
3359 // We just put in an allocate/initialize with a big raw-memory effect.
3360 // Hook selected additional alias categories on the initialization.
3361 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3362                                 MergeMemNode* init_in_merge,
3363                                 Node* init_out_raw) {
3364   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3365   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3366 
3367   Node* prevmem = kit.memory(alias_idx);
3368   init_in_merge->set_memory_at(alias_idx, prevmem);
3369   kit.set_memory(init_out_raw, alias_idx);
3370 }
3371 
3372 //---------------------------set_output_for_allocation-------------------------
3373 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3374                                           const TypeOopPtr* oop_type,
3375                                           bool deoptimize_on_exception) {
3376   int rawidx = Compile::AliasIdxRaw;
3377   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3378   add_safepoint_edges(alloc);
3379   Node* allocx = _gvn.transform(alloc);
3380   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3381   // create memory projection for i_o
3382   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3383   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3384 
3385   // create a memory projection as for the normal control path
3386   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3387   set_memory(malloc, rawidx);
3388 
3389   // a normal slow-call doesn't change i_o, but an allocation does
3390   // we create a separate i_o projection for the normal control path
3391   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3392   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3393 
3394   // put in an initialization barrier
3395   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3396                                                  rawoop)->as_Initialize();
3397   assert(alloc->initialization() == init,  "2-way macro link must work");
3398   assert(init ->allocation()     == alloc, "2-way macro link must work");
3399   {
3400     // Extract memory strands which may participate in the new object's
3401     // initialization, and source them from the new InitializeNode.
3402     // This will allow us to observe initializations when they occur,
3403     // and link them properly (as a group) to the InitializeNode.
3404     assert(init->in(InitializeNode::Memory) == malloc, "");
3405     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3406     init->set_req(InitializeNode::Memory, minit_in);
3407     record_for_igvn(minit_in); // fold it up later, if possible
3408     Node* minit_out = memory(rawidx);
3409     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3410     if (oop_type->isa_aryptr()) {
3411       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3412       int            elemidx  = C->get_alias_index(telemref);
3413       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3414     } else if (oop_type->isa_instptr()) {
3415       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3416       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3417         ciField* field = ik->nonstatic_field_at(i);
3418         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3419           continue;  // do not bother to track really large numbers of fields
3420         // Find (or create) the alias category for this field:
3421         int fieldidx = C->alias_type(field)->index();
3422         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3423       }
3424     }
3425   }
3426 
3427   // Cast raw oop to the real thing...
3428   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3429   javaoop = _gvn.transform(javaoop);
3430   C->set_recent_alloc(control(), javaoop);
3431   assert(just_allocated_object(control()) == javaoop, "just allocated");
3432 
3433 #ifdef ASSERT
3434   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3435     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3436            "Ideal_allocation works");
3437     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3438            "Ideal_allocation works");
3439     if (alloc->is_AllocateArray()) {
3440       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3441              "Ideal_allocation works");
3442       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3443              "Ideal_allocation works");
3444     } else {
3445       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3446     }
3447   }
3448 #endif //ASSERT
3449 
3450   return javaoop;
3451 }
3452 
3453 //---------------------------new_instance--------------------------------------
3454 // This routine takes a klass_node which may be constant (for a static type)
3455 // or may be non-constant (for reflective code).  It will work equally well
3456 // for either, and the graph will fold nicely if the optimizer later reduces
3457 // the type to a constant.
3458 // The optional arguments are for specialized use by intrinsics:
3459 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3460 //  - If 'return_size_val', report the the total object size to the caller.
3461 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3462 Node* GraphKit::new_instance(Node* klass_node,
3463                              Node* extra_slow_test,
3464                              Node* *return_size_val,
3465                              bool deoptimize_on_exception) {
3466   // Compute size in doublewords
3467   // The size is always an integral number of doublewords, represented
3468   // as a positive bytewise size stored in the klass's layout_helper.
3469   // The layout_helper also encodes (in a low bit) the need for a slow path.
3470   jint  layout_con = Klass::_lh_neutral_value;
3471   Node* layout_val = get_layout_helper(klass_node, layout_con);
3472   int   layout_is_con = (layout_val == NULL);
3473 
3474   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3475   // Generate the initial go-slow test.  It's either ALWAYS (return a
3476   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3477   // case) a computed value derived from the layout_helper.
3478   Node* initial_slow_test = NULL;
3479   if (layout_is_con) {
3480     assert(!StressReflectiveCode, "stress mode does not use these paths");
3481     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3482     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3483   } else {   // reflective case
3484     // This reflective path is used by Unsafe.allocateInstance.
3485     // (It may be stress-tested by specifying StressReflectiveCode.)
3486     // Basically, we want to get into the VM is there's an illegal argument.
3487     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3488     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3489     if (extra_slow_test != intcon(0)) {
3490       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3491     }
3492     // (Macro-expander will further convert this to a Bool, if necessary.)
3493   }
3494 
3495   // Find the size in bytes.  This is easy; it's the layout_helper.
3496   // The size value must be valid even if the slow path is taken.
3497   Node* size = NULL;
3498   if (layout_is_con) {
3499     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3500   } else {   // reflective case
3501     // This reflective path is used by clone and Unsafe.allocateInstance.
3502     size = ConvI2X(layout_val);
3503 
3504     // Clear the low bits to extract layout_helper_size_in_bytes:
3505     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3506     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3507     size = _gvn.transform( new AndXNode(size, mask) );
3508   }
3509   if (return_size_val != NULL) {
3510     (*return_size_val) = size;
3511   }
3512 
3513   // This is a precise notnull oop of the klass.
3514   // (Actually, it need not be precise if this is a reflective allocation.)
3515   // It's what we cast the result to.
3516   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3517   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3518   const TypeOopPtr* oop_type = tklass->as_instance_type();
3519 
3520   // Now generate allocation code
3521 
3522   // The entire memory state is needed for slow path of the allocation
3523   // since GC and deoptimization can happened.
3524   Node *mem = reset_memory();
3525   set_all_memory(mem); // Create new memory state
3526 
3527   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3528                                          control(), mem, i_o(),
3529                                          size, klass_node,
3530                                          initial_slow_test);
3531 
3532   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3533 }
3534 
3535 //-------------------------------new_array-------------------------------------
3536 // helper for both newarray and anewarray
3537 // The 'length' parameter is (obviously) the length of the array.
3538 // See comments on new_instance for the meaning of the other arguments.
3539 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3540                           Node* length,         // number of array elements
3541                           int   nargs,          // number of arguments to push back for uncommon trap
3542                           Node* *return_size_val,
3543                           bool deoptimize_on_exception) {
3544   jint  layout_con = Klass::_lh_neutral_value;
3545   Node* layout_val = get_layout_helper(klass_node, layout_con);
3546   int   layout_is_con = (layout_val == NULL);
3547 
3548   if (!layout_is_con && !StressReflectiveCode &&
3549       !too_many_traps(Deoptimization::Reason_class_check)) {
3550     // This is a reflective array creation site.
3551     // Optimistically assume that it is a subtype of Object[],
3552     // so that we can fold up all the address arithmetic.
3553     layout_con = Klass::array_layout_helper(T_OBJECT);
3554     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3555     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3556     { BuildCutout unless(this, bol_lh, PROB_MAX);
3557       inc_sp(nargs);
3558       uncommon_trap(Deoptimization::Reason_class_check,
3559                     Deoptimization::Action_maybe_recompile);
3560     }
3561     layout_val = NULL;
3562     layout_is_con = true;
3563   }
3564 
3565   // Generate the initial go-slow test.  Make sure we do not overflow
3566   // if length is huge (near 2Gig) or negative!  We do not need
3567   // exact double-words here, just a close approximation of needed
3568   // double-words.  We can't add any offset or rounding bits, lest we
3569   // take a size -1 of bytes and make it positive.  Use an unsigned
3570   // compare, so negative sizes look hugely positive.
3571   int fast_size_limit = FastAllocateSizeLimit;
3572   if (layout_is_con) {
3573     assert(!StressReflectiveCode, "stress mode does not use these paths");
3574     // Increase the size limit if we have exact knowledge of array type.
3575     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3576     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3577   }
3578 
3579   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3580   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3581 
3582   // --- Size Computation ---
3583   // array_size = round_to_heap(array_header + (length << elem_shift));
3584   // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3585   // and align_to(x, y) == ((x + y-1) & ~(y-1))
3586   // The rounding mask is strength-reduced, if possible.
3587   int round_mask = MinObjAlignmentInBytes - 1;
3588   Node* header_size = NULL;
3589   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3590   // (T_BYTE has the weakest alignment and size restrictions...)
3591   if (layout_is_con) {
3592     int       hsize  = Klass::layout_helper_header_size(layout_con);
3593     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3594     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3595     if ((round_mask & ~right_n_bits(eshift)) == 0)
3596       round_mask = 0;  // strength-reduce it if it goes away completely
3597     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3598     assert(header_size_min <= hsize, "generic minimum is smallest");
3599     header_size_min = hsize;
3600     header_size = intcon(hsize + round_mask);
3601   } else {
3602     Node* hss   = intcon(Klass::_lh_header_size_shift);
3603     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3604     Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3605     hsize       = _gvn.transform( new AndINode(hsize, hsm) );
3606     Node* mask  = intcon(round_mask);
3607     header_size = _gvn.transform( new AddINode(hsize, mask) );
3608   }
3609 
3610   Node* elem_shift = NULL;
3611   if (layout_is_con) {
3612     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3613     if (eshift != 0)
3614       elem_shift = intcon(eshift);
3615   } else {
3616     // There is no need to mask or shift this value.
3617     // The semantics of LShiftINode include an implicit mask to 0x1F.
3618     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3619     elem_shift = layout_val;
3620   }
3621 
3622   // Transition to native address size for all offset calculations:
3623   Node* lengthx = ConvI2X(length);
3624   Node* headerx = ConvI2X(header_size);
3625 #ifdef _LP64
3626   { const TypeInt* tilen = _gvn.find_int_type(length);
3627     if (tilen != NULL && tilen->_lo < 0) {
3628       // Add a manual constraint to a positive range.  Cf. array_element_address.
3629       jint size_max = fast_size_limit;
3630       if (size_max > tilen->_hi)  size_max = tilen->_hi;
3631       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3632 
3633       // Only do a narrow I2L conversion if the range check passed.
3634       IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3635       _gvn.transform(iff);
3636       RegionNode* region = new RegionNode(3);
3637       _gvn.set_type(region, Type::CONTROL);
3638       lengthx = new PhiNode(region, TypeLong::LONG);
3639       _gvn.set_type(lengthx, TypeLong::LONG);
3640 
3641       // Range check passed. Use ConvI2L node with narrow type.
3642       Node* passed = IfFalse(iff);
3643       region->init_req(1, passed);
3644       // Make I2L conversion control dependent to prevent it from
3645       // floating above the range check during loop optimizations.
3646       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3647 
3648       // Range check failed. Use ConvI2L with wide type because length may be invalid.
3649       region->init_req(2, IfTrue(iff));
3650       lengthx->init_req(2, ConvI2X(length));
3651 
3652       set_control(region);
3653       record_for_igvn(region);
3654       record_for_igvn(lengthx);
3655     }
3656   }
3657 #endif
3658 
3659   // Combine header size (plus rounding) and body size.  Then round down.
3660   // This computation cannot overflow, because it is used only in two
3661   // places, one where the length is sharply limited, and the other
3662   // after a successful allocation.
3663   Node* abody = lengthx;
3664   if (elem_shift != NULL)
3665     abody     = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3666   Node* size  = _gvn.transform( new AddXNode(headerx, abody) );
3667   if (round_mask != 0) {
3668     Node* mask = MakeConX(~round_mask);
3669     size       = _gvn.transform( new AndXNode(size, mask) );
3670   }
3671   // else if round_mask == 0, the size computation is self-rounding
3672 
3673   if (return_size_val != NULL) {
3674     // This is the size
3675     (*return_size_val) = size;
3676   }
3677 
3678   // Now generate allocation code
3679 
3680   // The entire memory state is needed for slow path of the allocation
3681   // since GC and deoptimization can happened.
3682   Node *mem = reset_memory();
3683   set_all_memory(mem); // Create new memory state
3684 
3685   if (initial_slow_test->is_Bool()) {
3686     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3687     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3688   }
3689 
3690   // Create the AllocateArrayNode and its result projections
3691   AllocateArrayNode* alloc
3692     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3693                             control(), mem, i_o(),
3694                             size, klass_node,
3695                             initial_slow_test,
3696                             length);
3697 
3698   // Cast to correct type.  Note that the klass_node may be constant or not,
3699   // and in the latter case the actual array type will be inexact also.
3700   // (This happens via a non-constant argument to inline_native_newArray.)
3701   // In any case, the value of klass_node provides the desired array type.
3702   const TypeInt* length_type = _gvn.find_int_type(length);
3703   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3704   if (ary_type->isa_aryptr() && length_type != NULL) {
3705     // Try to get a better type than POS for the size
3706     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3707   }
3708 
3709   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3710 
3711   // Cast length on remaining path to be as narrow as possible
3712   if (map()->find_edge(length) >= 0) {
3713     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3714     if (ccast != length) {
3715       _gvn.set_type_bottom(ccast);
3716       record_for_igvn(ccast);
3717       replace_in_map(length, ccast);
3718     }
3719   }
3720 
3721   return javaoop;
3722 }
3723 
3724 // The following "Ideal_foo" functions are placed here because they recognize
3725 // the graph shapes created by the functions immediately above.
3726 
3727 //---------------------------Ideal_allocation----------------------------------
3728 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3729 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3730   if (ptr == NULL) {     // reduce dumb test in callers
3731     return NULL;
3732   }
3733 
3734 #if INCLUDE_SHENANDOAHGC
3735   // Attempt to see through Shenandoah barriers.
3736   ptr = ShenandoahBarrierNode::skip_through_barrier(ptr);
3737 #endif
3738 
3739   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3740     ptr = ptr->in(1);
3741     if (ptr == NULL) return NULL;
3742   }
3743   // Return NULL for allocations with several casts:
3744   //   j.l.reflect.Array.newInstance(jobject, jint)
3745   //   Object.clone()
3746   // to keep more precise type from last cast.
3747   if (ptr->is_Proj()) {
3748     Node* allo = ptr->in(0);
3749     if (allo != NULL && allo->is_Allocate()) {
3750       return allo->as_Allocate();
3751     }
3752   }
3753   // Report failure to match.
3754   return NULL;
3755 }
3756 
3757 // Fancy version which also strips off an offset (and reports it to caller).
3758 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3759                                              intptr_t& offset) {
3760   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3761   if (base == NULL)  return NULL;
3762   return Ideal_allocation(base, phase);
3763 }
3764 
3765 // Trace Initialize <- Proj[Parm] <- Allocate
3766 AllocateNode* InitializeNode::allocation() {
3767   Node* rawoop = in(InitializeNode::RawAddress);
3768   if (rawoop->is_Proj()) {
3769     Node* alloc = rawoop->in(0);
3770     if (alloc->is_Allocate()) {
3771       return alloc->as_Allocate();
3772     }
3773   }
3774   return NULL;
3775 }
3776 
3777 // Trace Allocate -> Proj[Parm] -> Initialize
3778 InitializeNode* AllocateNode::initialization() {
3779   ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3780   if (rawoop == NULL)  return NULL;
3781   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3782     Node* init = rawoop->fast_out(i);
3783     if (init->is_Initialize()) {
3784       assert(init->as_Initialize()->allocation() == this, "2-way link");
3785       return init->as_Initialize();
3786     }
3787   }
3788   return NULL;
3789 }
3790 
3791 //----------------------------- loop predicates ---------------------------
3792 
3793 //------------------------------add_predicate_impl----------------------------
3794 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3795   // Too many traps seen?
3796   if (too_many_traps(reason)) {
3797 #ifdef ASSERT
3798     if (TraceLoopPredicate) {
3799       int tc = C->trap_count(reason);
3800       tty->print("too many traps=%s tcount=%d in ",
3801                     Deoptimization::trap_reason_name(reason), tc);
3802       method()->print(); // which method has too many predicate traps
3803       tty->cr();
3804     }
3805 #endif
3806     // We cannot afford to take more traps here,
3807     // do not generate predicate.
3808     return;
3809   }
3810 
3811   Node *cont    = _gvn.intcon(1);
3812   Node* opq     = _gvn.transform(new Opaque1Node(C, cont));
3813   Node *bol     = _gvn.transform(new Conv2BNode(opq));
3814   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3815   Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3816   C->add_predicate_opaq(opq);
3817   {
3818     PreserveJVMState pjvms(this);
3819     set_control(iffalse);
3820     inc_sp(nargs);
3821     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3822   }
3823   Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3824   set_control(iftrue);
3825 }
3826 
3827 //------------------------------add_predicate---------------------------------
3828 void GraphKit::add_predicate(int nargs) {
3829   if (UseLoopPredicate) {
3830     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3831   }
3832   if (UseProfiledLoopPredicate) {
3833     add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
3834   }
3835   // loop's limit check predicate should be near the loop.
3836   add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3837 }
3838 
3839 void GraphKit::sync_kit(IdealKit& ideal) {
3840   set_all_memory(ideal.merged_memory());
3841   set_i_o(ideal.i_o());
3842   set_control(ideal.ctrl());
3843 }
3844 
3845 void GraphKit::final_sync(IdealKit& ideal) {
3846   // Final sync IdealKit and graphKit.
3847   sync_kit(ideal);
3848 }
3849 
3850 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
3851   Node* len = load_array_length(load_String_value(ctrl, str));
3852   Node* coder = load_String_coder(ctrl, str);
3853   // Divide length by 2 if coder is UTF16
3854   return _gvn.transform(new RShiftINode(len, coder));
3855 }
3856 
3857 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
3858   int value_offset = java_lang_String::value_offset_in_bytes();
3859   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3860                                                      false, NULL, 0);
3861   const TypePtr* value_field_type = string_type->add_offset(value_offset);
3862   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3863                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
3864                                                   ciTypeArrayKlass::make(T_BYTE), true, 0);
3865 
3866 #if INCLUDE_SHENANDOAHGC
3867   if (!ShenandoahOptimizeInstanceFinals) {
3868     str = access_resolve_for_read(str);
3869   }
3870 #endif
3871 
3872   Node* p = basic_plus_adr(str, str, value_offset);
3873   Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
3874                               IN_HEAP | C2_CONTROL_DEPENDENT_LOAD);
3875   // String.value field is known to be @Stable.
3876   if (UseImplicitStableValues) {
3877     load = cast_array_to_stable(load, value_type);
3878   }
3879   return load;
3880 }
3881 
3882 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
3883   if (!CompactStrings) {
3884     return intcon(java_lang_String::CODER_UTF16);
3885   }
3886   int coder_offset = java_lang_String::coder_offset_in_bytes();
3887   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3888                                                      false, NULL, 0);
3889   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
3890   int coder_field_idx = C->get_alias_index(coder_field_type);
3891 
3892 #if INCLUDE_SHENANDOAHGC
3893   if (!ShenandoahOptimizeInstanceFinals) {
3894     str = access_resolve_for_read(str);
3895   }
3896 #endif
3897 
3898   return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
3899                    TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
3900 }
3901 
3902 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
3903   int value_offset = java_lang_String::value_offset_in_bytes();
3904   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3905                                                      false, NULL, 0);
3906   const TypePtr* value_field_type = string_type->add_offset(value_offset);
3907 
3908   str = access_resolve_for_write(str);
3909 
3910   access_store_at(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
3911                   value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP);
3912 }
3913 
3914 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
3915   int coder_offset = java_lang_String::coder_offset_in_bytes();
3916   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3917                                                      false, NULL, 0);
3918 
3919   str = access_resolve_for_write(str);
3920 
3921   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
3922   int coder_field_idx = C->get_alias_index(coder_field_type);
3923   store_to_memory(control(), basic_plus_adr(str, coder_offset),
3924                   value, T_BYTE, coder_field_idx, MemNode::unordered);
3925 }
3926 
3927 // Capture src and dst memory state with a MergeMemNode
3928 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
3929   if (src_type == dst_type) {
3930     // Types are equal, we don't need a MergeMemNode
3931     return memory(src_type);
3932   }
3933   MergeMemNode* merge = MergeMemNode::make(map()->memory());
3934   record_for_igvn(merge); // fold it up later, if possible
3935   int src_idx = C->get_alias_index(src_type);
3936   int dst_idx = C->get_alias_index(dst_type);
3937   merge->set_memory_at(src_idx, memory(src_idx));
3938   merge->set_memory_at(dst_idx, memory(dst_idx));
3939   return merge;
3940 }
3941 
3942 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
3943   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
3944   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
3945   // If input and output memory types differ, capture both states to preserve
3946   // the dependency between preceding and subsequent loads/stores.
3947   // For example, the following program:
3948   //  StoreB
3949   //  compress_string
3950   //  LoadB
3951   // has this memory graph (use->def):
3952   //  LoadB -> compress_string -> CharMem
3953   //             ... -> StoreB -> ByteMem
3954   // The intrinsic hides the dependency between LoadB and StoreB, causing
3955   // the load to read from memory not containing the result of the StoreB.
3956   // The correct memory graph should look like this:
3957   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
3958   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
3959   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
3960   Node* res_mem = _gvn.transform(new SCMemProjNode(str));
3961   set_memory(res_mem, TypeAryPtr::BYTES);
3962   return str;
3963 }
3964 
3965 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
3966   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
3967   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
3968   // Capture src and dst memory (see comment in 'compress_string').
3969   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
3970   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
3971   set_memory(_gvn.transform(str), dst_type);
3972 }
3973 
3974 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
3975   src = access_resolve_for_read(src);
3976   dst = access_resolve_for_write(dst);
3977 
3978   /**
3979    * int i_char = start;
3980    * for (int i_byte = 0; i_byte < count; i_byte++) {
3981    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
3982    * }
3983    */
3984   add_predicate();
3985   RegionNode* head = new RegionNode(3);
3986   head->init_req(1, control());
3987   gvn().set_type(head, Type::CONTROL);
3988   record_for_igvn(head);
3989 
3990   Node* i_byte = new PhiNode(head, TypeInt::INT);
3991   i_byte->init_req(1, intcon(0));
3992   gvn().set_type(i_byte, TypeInt::INT);
3993   record_for_igvn(i_byte);
3994 
3995   Node* i_char = new PhiNode(head, TypeInt::INT);
3996   i_char->init_req(1, start);
3997   gvn().set_type(i_char, TypeInt::INT);
3998   record_for_igvn(i_char);
3999 
4000   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4001   gvn().set_type(mem, Type::MEMORY);
4002   record_for_igvn(mem);
4003   set_control(head);
4004   set_memory(mem, TypeAryPtr::BYTES);
4005   Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4006   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4007                              AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4008                              false, false, true /* mismatched */);
4009 
4010   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4011   head->init_req(2, IfTrue(iff));
4012   mem->init_req(2, st);
4013   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4014   i_char->init_req(2, AddI(i_char, intcon(2)));
4015 
4016   set_control(IfFalse(iff));
4017   set_memory(st, TypeAryPtr::BYTES);
4018 }
4019 
4020 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4021   if (!field->is_constant()) {
4022     return NULL; // Field not marked as constant.
4023   }
4024   ciInstance* holder = NULL;
4025   if (!field->is_static()) {
4026     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4027     if (const_oop != NULL && const_oop->is_instance()) {
4028       holder = const_oop->as_instance();
4029     }
4030   }
4031   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4032                                                         /*is_unsigned_load=*/false);
4033   if (con_type != NULL) {
4034     return makecon(con_type);
4035   }
4036   return NULL;
4037 }
4038 
4039 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4040   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4041   // assumption of CCP analysis.
4042   return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4043 }