1 /*
   2  * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "compiler/compileLog.hpp"
  27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
  28 #include "gc_implementation/g1/heapRegion.hpp"
  29 #include "gc_interface/collectedHeap.hpp"
  30 #include "gc_implementation/shenandoah/shenandoahHeap.hpp"
  31 #include "memory/barrierSet.hpp"
  32 #include "memory/cardTableModRefBS.hpp"
  33 #include "opto/addnode.hpp"
  34 #include "opto/graphKit.hpp"
  35 #include "opto/idealKit.hpp"
  36 #include "opto/locknode.hpp"
  37 #include "opto/machnode.hpp"
  38 #include "opto/parse.hpp"
  39 #include "opto/rootnode.hpp"
  40 #include "opto/runtime.hpp"
  41 #include "runtime/deoptimization.hpp"
  42 #include "runtime/sharedRuntime.hpp"
  43 
  44 #if INCLUDE_ALL_GCS
  45 #include "gc_implementation/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  46 #include "gc_implementation/shenandoah/c2/shenandoahSupport.hpp"
  47 #endif
  48 
  49 //----------------------------GraphKit-----------------------------------------
  50 // Main utility constructor.
  51 GraphKit::GraphKit(JVMState* jvms)
  52   : Phase(Phase::Parser),
  53     _env(C->env()),
  54     _gvn(*C->initial_gvn())
  55 {
  56   _exceptions = jvms->map()->next_exception();
  57   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  58   set_jvms(jvms);
  59 }
  60 
  61 // Private constructor for parser.
  62 GraphKit::GraphKit()
  63   : Phase(Phase::Parser),
  64     _env(C->env()),
  65     _gvn(*C->initial_gvn())
  66 {
  67   _exceptions = NULL;
  68   set_map(NULL);
  69   debug_only(_sp = -99);
  70   debug_only(set_bci(-99));
  71 }
  72 
  73 
  74 
  75 //---------------------------clean_stack---------------------------------------
  76 // Clear away rubbish from the stack area of the JVM state.
  77 // This destroys any arguments that may be waiting on the stack.
  78 void GraphKit::clean_stack(int from_sp) {
  79   SafePointNode* map      = this->map();
  80   JVMState*      jvms     = this->jvms();
  81   int            stk_size = jvms->stk_size();
  82   int            stkoff   = jvms->stkoff();
  83   Node*          top      = this->top();
  84   for (int i = from_sp; i < stk_size; i++) {
  85     if (map->in(stkoff + i) != top) {
  86       map->set_req(stkoff + i, top);
  87     }
  88   }
  89 }
  90 
  91 
  92 //--------------------------------sync_jvms-----------------------------------
  93 // Make sure our current jvms agrees with our parse state.
  94 JVMState* GraphKit::sync_jvms() const {
  95   JVMState* jvms = this->jvms();
  96   jvms->set_bci(bci());       // Record the new bci in the JVMState
  97   jvms->set_sp(sp());         // Record the new sp in the JVMState
  98   assert(jvms_in_sync(), "jvms is now in sync");
  99   return jvms;
 100 }
 101 
 102 //--------------------------------sync_jvms_for_reexecute---------------------
 103 // Make sure our current jvms agrees with our parse state.  This version
 104 // uses the reexecute_sp for reexecuting bytecodes.
 105 JVMState* GraphKit::sync_jvms_for_reexecute() {
 106   JVMState* jvms = this->jvms();
 107   jvms->set_bci(bci());          // Record the new bci in the JVMState
 108   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 109   return jvms;
 110 }
 111 
 112 #ifdef ASSERT
 113 bool GraphKit::jvms_in_sync() const {
 114   Parse* parse = is_Parse();
 115   if (parse == NULL) {
 116     if (bci() !=      jvms()->bci())          return false;
 117     if (sp()  != (int)jvms()->sp())           return false;
 118     return true;
 119   }
 120   if (jvms()->method() != parse->method())    return false;
 121   if (jvms()->bci()    != parse->bci())       return false;
 122   int jvms_sp = jvms()->sp();
 123   if (jvms_sp          != parse->sp())        return false;
 124   int jvms_depth = jvms()->depth();
 125   if (jvms_depth       != parse->depth())     return false;
 126   return true;
 127 }
 128 
 129 // Local helper checks for special internal merge points
 130 // used to accumulate and merge exception states.
 131 // They are marked by the region's in(0) edge being the map itself.
 132 // Such merge points must never "escape" into the parser at large,
 133 // until they have been handed to gvn.transform.
 134 static bool is_hidden_merge(Node* reg) {
 135   if (reg == NULL)  return false;
 136   if (reg->is_Phi()) {
 137     reg = reg->in(0);
 138     if (reg == NULL)  return false;
 139   }
 140   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 141 }
 142 
 143 void GraphKit::verify_map() const {
 144   if (map() == NULL)  return;  // null map is OK
 145   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 146   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 147   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 148 }
 149 
 150 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 151   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 152   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 153 }
 154 #endif
 155 
 156 //---------------------------stop_and_kill_map---------------------------------
 157 // Set _map to NULL, signalling a stop to further bytecode execution.
 158 // First smash the current map's control to a constant, to mark it dead.
 159 void GraphKit::stop_and_kill_map() {
 160   SafePointNode* dead_map = stop();
 161   if (dead_map != NULL) {
 162     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
 163     assert(dead_map->is_killed(), "must be so marked");
 164   }
 165 }
 166 
 167 
 168 //--------------------------------stopped--------------------------------------
 169 // Tell if _map is NULL, or control is top.
 170 bool GraphKit::stopped() {
 171   if (map() == NULL)           return true;
 172   else if (control() == top()) return true;
 173   else                         return false;
 174 }
 175 
 176 
 177 //-----------------------------has_ex_handler----------------------------------
 178 // Tell if this method or any caller method has exception handlers.
 179 bool GraphKit::has_ex_handler() {
 180   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 181     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 182       return true;
 183     }
 184   }
 185   return false;
 186 }
 187 
 188 //------------------------------save_ex_oop------------------------------------
 189 // Save an exception without blowing stack contents or other JVM state.
 190 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 191   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 192   ex_map->add_req(ex_oop);
 193   debug_only(verify_exception_state(ex_map));
 194 }
 195 
 196 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 197   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 198   Node* ex_oop = ex_map->in(ex_map->req()-1);
 199   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 200   return ex_oop;
 201 }
 202 
 203 //-----------------------------saved_ex_oop------------------------------------
 204 // Recover a saved exception from its map.
 205 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 206   return common_saved_ex_oop(ex_map, false);
 207 }
 208 
 209 //--------------------------clear_saved_ex_oop---------------------------------
 210 // Erase a previously saved exception from its map.
 211 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 212   return common_saved_ex_oop(ex_map, true);
 213 }
 214 
 215 #ifdef ASSERT
 216 //---------------------------has_saved_ex_oop----------------------------------
 217 // Erase a previously saved exception from its map.
 218 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 219   return ex_map->req() == ex_map->jvms()->endoff()+1;
 220 }
 221 #endif
 222 
 223 //-------------------------make_exception_state--------------------------------
 224 // Turn the current JVM state into an exception state, appending the ex_oop.
 225 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 226   sync_jvms();
 227   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 228   set_saved_ex_oop(ex_map, ex_oop);
 229   return ex_map;
 230 }
 231 
 232 
 233 //--------------------------add_exception_state--------------------------------
 234 // Add an exception to my list of exceptions.
 235 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 236   if (ex_map == NULL || ex_map->control() == top()) {
 237     return;
 238   }
 239 #ifdef ASSERT
 240   verify_exception_state(ex_map);
 241   if (has_exceptions()) {
 242     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 243   }
 244 #endif
 245 
 246   // If there is already an exception of exactly this type, merge with it.
 247   // In particular, null-checks and other low-level exceptions common up here.
 248   Node*       ex_oop  = saved_ex_oop(ex_map);
 249   const Type* ex_type = _gvn.type(ex_oop);
 250   if (ex_oop == top()) {
 251     // No action needed.
 252     return;
 253   }
 254   assert(ex_type->isa_instptr(), "exception must be an instance");
 255   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 256     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 257     // We check sp also because call bytecodes can generate exceptions
 258     // both before and after arguments are popped!
 259     if (ex_type2 == ex_type
 260         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 261       combine_exception_states(ex_map, e2);
 262       return;
 263     }
 264   }
 265 
 266   // No pre-existing exception of the same type.  Chain it on the list.
 267   push_exception_state(ex_map);
 268 }
 269 
 270 //-----------------------add_exception_states_from-----------------------------
 271 void GraphKit::add_exception_states_from(JVMState* jvms) {
 272   SafePointNode* ex_map = jvms->map()->next_exception();
 273   if (ex_map != NULL) {
 274     jvms->map()->set_next_exception(NULL);
 275     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 276       next_map = ex_map->next_exception();
 277       ex_map->set_next_exception(NULL);
 278       add_exception_state(ex_map);
 279     }
 280   }
 281 }
 282 
 283 //-----------------------transfer_exceptions_into_jvms-------------------------
 284 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 285   if (map() == NULL) {
 286     // We need a JVMS to carry the exceptions, but the map has gone away.
 287     // Create a scratch JVMS, cloned from any of the exception states...
 288     if (has_exceptions()) {
 289       _map = _exceptions;
 290       _map = clone_map();
 291       _map->set_next_exception(NULL);
 292       clear_saved_ex_oop(_map);
 293       debug_only(verify_map());
 294     } else {
 295       // ...or created from scratch
 296       JVMState* jvms = new (C) JVMState(_method, NULL);
 297       jvms->set_bci(_bci);
 298       jvms->set_sp(_sp);
 299       jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
 300       set_jvms(jvms);
 301       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 302       set_all_memory(top());
 303       while (map()->req() < jvms->endoff())  map()->add_req(top());
 304     }
 305     // (This is a kludge, in case you didn't notice.)
 306     set_control(top());
 307   }
 308   JVMState* jvms = sync_jvms();
 309   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 310   jvms->map()->set_next_exception(_exceptions);
 311   _exceptions = NULL;   // done with this set of exceptions
 312   return jvms;
 313 }
 314 
 315 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 316   assert(is_hidden_merge(dstphi), "must be a special merge node");
 317   assert(is_hidden_merge(srcphi), "must be a special merge node");
 318   uint limit = srcphi->req();
 319   for (uint i = PhiNode::Input; i < limit; i++) {
 320     dstphi->add_req(srcphi->in(i));
 321   }
 322 }
 323 static inline void add_one_req(Node* dstphi, Node* src) {
 324   assert(is_hidden_merge(dstphi), "must be a special merge node");
 325   assert(!is_hidden_merge(src), "must not be a special merge node");
 326   dstphi->add_req(src);
 327 }
 328 
 329 //-----------------------combine_exception_states------------------------------
 330 // This helper function combines exception states by building phis on a
 331 // specially marked state-merging region.  These regions and phis are
 332 // untransformed, and can build up gradually.  The region is marked by
 333 // having a control input of its exception map, rather than NULL.  Such
 334 // regions do not appear except in this function, and in use_exception_state.
 335 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 336   if (failing())  return;  // dying anyway...
 337   JVMState* ex_jvms = ex_map->_jvms;
 338   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 339   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 340   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 341   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 342   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 343   assert(ex_map->req() == phi_map->req(), "matching maps");
 344   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 345   Node*         hidden_merge_mark = root();
 346   Node*         region  = phi_map->control();
 347   MergeMemNode* phi_mem = phi_map->merged_memory();
 348   MergeMemNode* ex_mem  = ex_map->merged_memory();
 349   if (region->in(0) != hidden_merge_mark) {
 350     // The control input is not (yet) a specially-marked region in phi_map.
 351     // Make it so, and build some phis.
 352     region = new (C) RegionNode(2);
 353     _gvn.set_type(region, Type::CONTROL);
 354     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 355     region->init_req(1, phi_map->control());
 356     phi_map->set_control(region);
 357     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 358     record_for_igvn(io_phi);
 359     _gvn.set_type(io_phi, Type::ABIO);
 360     phi_map->set_i_o(io_phi);
 361     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 362       Node* m = mms.memory();
 363       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 364       record_for_igvn(m_phi);
 365       _gvn.set_type(m_phi, Type::MEMORY);
 366       mms.set_memory(m_phi);
 367     }
 368   }
 369 
 370   // Either or both of phi_map and ex_map might already be converted into phis.
 371   Node* ex_control = ex_map->control();
 372   // if there is special marking on ex_map also, we add multiple edges from src
 373   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 374   // how wide was the destination phi_map, originally?
 375   uint orig_width = region->req();
 376 
 377   if (add_multiple) {
 378     add_n_reqs(region, ex_control);
 379     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 380   } else {
 381     // ex_map has no merges, so we just add single edges everywhere
 382     add_one_req(region, ex_control);
 383     add_one_req(phi_map->i_o(), ex_map->i_o());
 384   }
 385   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 386     if (mms.is_empty()) {
 387       // get a copy of the base memory, and patch some inputs into it
 388       const TypePtr* adr_type = mms.adr_type(C);
 389       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 390       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 391       mms.set_memory(phi);
 392       // Prepare to append interesting stuff onto the newly sliced phi:
 393       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 394     }
 395     // Append stuff from ex_map:
 396     if (add_multiple) {
 397       add_n_reqs(mms.memory(), mms.memory2());
 398     } else {
 399       add_one_req(mms.memory(), mms.memory2());
 400     }
 401   }
 402   uint limit = ex_map->req();
 403   for (uint i = TypeFunc::Parms; i < limit; i++) {
 404     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 405     if (i == tos)  i = ex_jvms->monoff();
 406     Node* src = ex_map->in(i);
 407     Node* dst = phi_map->in(i);
 408     if (src != dst) {
 409       PhiNode* phi;
 410       if (dst->in(0) != region) {
 411         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 412         record_for_igvn(phi);
 413         _gvn.set_type(phi, phi->type());
 414         phi_map->set_req(i, dst);
 415         // Prepare to append interesting stuff onto the new phi:
 416         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 417       } else {
 418         assert(dst->is_Phi(), "nobody else uses a hidden region");
 419         phi = dst->as_Phi();
 420       }
 421       if (add_multiple && src->in(0) == ex_control) {
 422         // Both are phis.
 423         add_n_reqs(dst, src);
 424       } else {
 425         while (dst->req() < region->req())  add_one_req(dst, src);
 426       }
 427       const Type* srctype = _gvn.type(src);
 428       if (phi->type() != srctype) {
 429         const Type* dsttype = phi->type()->meet_speculative(srctype);
 430         if (phi->type() != dsttype) {
 431           phi->set_type(dsttype);
 432           _gvn.set_type(phi, dsttype);
 433         }
 434       }
 435     }
 436   }
 437   phi_map->merge_replaced_nodes_with(ex_map);
 438 }
 439 
 440 //--------------------------use_exception_state--------------------------------
 441 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 442   if (failing()) { stop(); return top(); }
 443   Node* region = phi_map->control();
 444   Node* hidden_merge_mark = root();
 445   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 446   Node* ex_oop = clear_saved_ex_oop(phi_map);
 447   if (region->in(0) == hidden_merge_mark) {
 448     // Special marking for internal ex-states.  Process the phis now.
 449     region->set_req(0, region);  // now it's an ordinary region
 450     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 451     // Note: Setting the jvms also sets the bci and sp.
 452     set_control(_gvn.transform(region));
 453     uint tos = jvms()->stkoff() + sp();
 454     for (uint i = 1; i < tos; i++) {
 455       Node* x = phi_map->in(i);
 456       if (x->in(0) == region) {
 457         assert(x->is_Phi(), "expected a special phi");
 458         phi_map->set_req(i, _gvn.transform(x));
 459       }
 460     }
 461     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 462       Node* x = mms.memory();
 463       if (x->in(0) == region) {
 464         assert(x->is_Phi(), "nobody else uses a hidden region");
 465         mms.set_memory(_gvn.transform(x));
 466       }
 467     }
 468     if (ex_oop->in(0) == region) {
 469       assert(ex_oop->is_Phi(), "expected a special phi");
 470       ex_oop = _gvn.transform(ex_oop);
 471     }
 472   } else {
 473     set_jvms(phi_map->jvms());
 474   }
 475 
 476   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 477   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 478   return ex_oop;
 479 }
 480 
 481 //---------------------------------java_bc-------------------------------------
 482 Bytecodes::Code GraphKit::java_bc() const {
 483   ciMethod* method = this->method();
 484   int       bci    = this->bci();
 485   if (method != NULL && bci != InvocationEntryBci)
 486     return method->java_code_at_bci(bci);
 487   else
 488     return Bytecodes::_illegal;
 489 }
 490 
 491 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 492                                                           bool must_throw) {
 493     // if the exception capability is set, then we will generate code
 494     // to check the JavaThread.should_post_on_exceptions flag to see
 495     // if we actually need to report exception events (for this
 496     // thread).  If we don't need to report exception events, we will
 497     // take the normal fast path provided by add_exception_events.  If
 498     // exception event reporting is enabled for this thread, we will
 499     // take the uncommon_trap in the BuildCutout below.
 500 
 501     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 502     Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
 503     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 504     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
 505 
 506     // Test the should_post_on_exceptions_flag vs. 0
 507     Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
 508     Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
 509 
 510     // Branch to slow_path if should_post_on_exceptions_flag was true
 511     { BuildCutout unless(this, tst, PROB_MAX);
 512       // Do not try anything fancy if we're notifying the VM on every throw.
 513       // Cf. case Bytecodes::_athrow in parse2.cpp.
 514       uncommon_trap(reason, Deoptimization::Action_none,
 515                     (ciKlass*)NULL, (char*)NULL, must_throw);
 516     }
 517 
 518 }
 519 
 520 //------------------------------builtin_throw----------------------------------
 521 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 522   bool must_throw = true;
 523 
 524   if (env()->jvmti_can_post_on_exceptions()) {
 525     // check if we must post exception events, take uncommon trap if so
 526     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 527     // here if should_post_on_exceptions is false
 528     // continue on with the normal codegen
 529   }
 530 
 531   // If this particular condition has not yet happened at this
 532   // bytecode, then use the uncommon trap mechanism, and allow for
 533   // a future recompilation if several traps occur here.
 534   // If the throw is hot, try to use a more complicated inline mechanism
 535   // which keeps execution inside the compiled code.
 536   bool treat_throw_as_hot = false;
 537   ciMethodData* md = method()->method_data();
 538 
 539   if (ProfileTraps) {
 540     if (too_many_traps(reason)) {
 541       treat_throw_as_hot = true;
 542     }
 543     // (If there is no MDO at all, assume it is early in
 544     // execution, and that any deopts are part of the
 545     // startup transient, and don't need to be remembered.)
 546 
 547     // Also, if there is a local exception handler, treat all throws
 548     // as hot if there has been at least one in this method.
 549     if (C->trap_count(reason) != 0
 550         && method()->method_data()->trap_count(reason) != 0
 551         && has_ex_handler()) {
 552         treat_throw_as_hot = true;
 553     }
 554   }
 555 
 556   // If this throw happens frequently, an uncommon trap might cause
 557   // a performance pothole.  If there is a local exception handler,
 558   // and if this particular bytecode appears to be deoptimizing often,
 559   // let us handle the throw inline, with a preconstructed instance.
 560   // Note:   If the deopt count has blown up, the uncommon trap
 561   // runtime is going to flush this nmethod, not matter what.
 562   if (treat_throw_as_hot
 563       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 564     // If the throw is local, we use a pre-existing instance and
 565     // punt on the backtrace.  This would lead to a missing backtrace
 566     // (a repeat of 4292742) if the backtrace object is ever asked
 567     // for its backtrace.
 568     // Fixing this remaining case of 4292742 requires some flavor of
 569     // escape analysis.  Leave that for the future.
 570     ciInstance* ex_obj = NULL;
 571     switch (reason) {
 572     case Deoptimization::Reason_null_check:
 573       ex_obj = env()->NullPointerException_instance();
 574       break;
 575     case Deoptimization::Reason_div0_check:
 576       ex_obj = env()->ArithmeticException_instance();
 577       break;
 578     case Deoptimization::Reason_range_check:
 579       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 580       break;
 581     case Deoptimization::Reason_class_check:
 582       if (java_bc() == Bytecodes::_aastore) {
 583         ex_obj = env()->ArrayStoreException_instance();
 584       } else {
 585         ex_obj = env()->ClassCastException_instance();
 586       }
 587       break;
 588     }
 589     if (failing()) { stop(); return; }  // exception allocation might fail
 590     if (ex_obj != NULL) {
 591       // Cheat with a preallocated exception object.
 592       if (C->log() != NULL)
 593         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 594                        Deoptimization::trap_reason_name(reason));
 595       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 596       Node*              ex_node = _gvn.transform( ConNode::make(C, ex_con) );
 597 
 598       // Clear the detail message of the preallocated exception object.
 599       // Weblogic sometimes mutates the detail message of exceptions
 600       // using reflection.
 601       int offset = java_lang_Throwable::get_detailMessage_offset();
 602       const TypePtr* adr_typ = ex_con->add_offset(offset);
 603 
 604       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 605       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 606       // Conservatively release stores of object references.
 607       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
 608 
 609       add_exception_state(make_exception_state(ex_node));
 610       return;
 611     }
 612   }
 613 
 614   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 615   // It won't be much cheaper than bailing to the interp., since we'll
 616   // have to pass up all the debug-info, and the runtime will have to
 617   // create the stack trace.
 618 
 619   // Usual case:  Bail to interpreter.
 620   // Reserve the right to recompile if we haven't seen anything yet.
 621 
 622   assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
 623   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 624   if (treat_throw_as_hot
 625       && (method()->method_data()->trap_recompiled_at(bci(), NULL)
 626           || C->too_many_traps(reason))) {
 627     // We cannot afford to take more traps here.  Suffer in the interpreter.
 628     if (C->log() != NULL)
 629       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 630                      Deoptimization::trap_reason_name(reason),
 631                      C->trap_count(reason));
 632     action = Deoptimization::Action_none;
 633   }
 634 
 635   // "must_throw" prunes the JVM state to include only the stack, if there
 636   // are no local exception handlers.  This should cut down on register
 637   // allocation time and code size, by drastically reducing the number
 638   // of in-edges on the call to the uncommon trap.
 639 
 640   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 641 }
 642 
 643 
 644 //----------------------------PreserveJVMState---------------------------------
 645 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 646   debug_only(kit->verify_map());
 647   _kit    = kit;
 648   _map    = kit->map();   // preserve the map
 649   _sp     = kit->sp();
 650   kit->set_map(clone_map ? kit->clone_map() : NULL);
 651 #ifdef ASSERT
 652   _bci    = kit->bci();
 653   Parse* parser = kit->is_Parse();
 654   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 655   _block  = block;
 656 #endif
 657 }
 658 PreserveJVMState::~PreserveJVMState() {
 659   GraphKit* kit = _kit;
 660 #ifdef ASSERT
 661   assert(kit->bci() == _bci, "bci must not shift");
 662   Parse* parser = kit->is_Parse();
 663   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 664   assert(block == _block,    "block must not shift");
 665 #endif
 666   kit->set_map(_map);
 667   kit->set_sp(_sp);
 668 }
 669 
 670 
 671 //-----------------------------BuildCutout-------------------------------------
 672 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 673   : PreserveJVMState(kit)
 674 {
 675   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 676   SafePointNode* outer_map = _map;   // preserved map is caller's
 677   SafePointNode* inner_map = kit->map();
 678   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 679   outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
 680   inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
 681 }
 682 BuildCutout::~BuildCutout() {
 683   GraphKit* kit = _kit;
 684   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 685 }
 686 
 687 //---------------------------PreserveReexecuteState----------------------------
 688 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 689   assert(!kit->stopped(), "must call stopped() before");
 690   _kit    =    kit;
 691   _sp     =    kit->sp();
 692   _reexecute = kit->jvms()->_reexecute;
 693 }
 694 PreserveReexecuteState::~PreserveReexecuteState() {
 695   if (_kit->stopped()) return;
 696   _kit->jvms()->_reexecute = _reexecute;
 697   _kit->set_sp(_sp);
 698 }
 699 
 700 //------------------------------clone_map--------------------------------------
 701 // Implementation of PreserveJVMState
 702 //
 703 // Only clone_map(...) here. If this function is only used in the
 704 // PreserveJVMState class we may want to get rid of this extra
 705 // function eventually and do it all there.
 706 
 707 SafePointNode* GraphKit::clone_map() {
 708   if (map() == NULL)  return NULL;
 709 
 710   // Clone the memory edge first
 711   Node* mem = MergeMemNode::make(C, map()->memory());
 712   gvn().set_type_bottom(mem);
 713 
 714   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 715   JVMState* jvms = this->jvms();
 716   JVMState* clonejvms = jvms->clone_shallow(C);
 717   clonemap->set_memory(mem);
 718   clonemap->set_jvms(clonejvms);
 719   clonejvms->set_map(clonemap);
 720   record_for_igvn(clonemap);
 721   gvn().set_type_bottom(clonemap);
 722   return clonemap;
 723 }
 724 
 725 
 726 //-----------------------------set_map_clone-----------------------------------
 727 void GraphKit::set_map_clone(SafePointNode* m) {
 728   _map = m;
 729   _map = clone_map();
 730   _map->set_next_exception(NULL);
 731   debug_only(verify_map());
 732 }
 733 
 734 
 735 //----------------------------kill_dead_locals---------------------------------
 736 // Detect any locals which are known to be dead, and force them to top.
 737 void GraphKit::kill_dead_locals() {
 738   // Consult the liveness information for the locals.  If any
 739   // of them are unused, then they can be replaced by top().  This
 740   // should help register allocation time and cut down on the size
 741   // of the deoptimization information.
 742 
 743   // This call is made from many of the bytecode handling
 744   // subroutines called from the Big Switch in do_one_bytecode.
 745   // Every bytecode which might include a slow path is responsible
 746   // for killing its dead locals.  The more consistent we
 747   // are about killing deads, the fewer useless phis will be
 748   // constructed for them at various merge points.
 749 
 750   // bci can be -1 (InvocationEntryBci).  We return the entry
 751   // liveness for the method.
 752 
 753   if (method() == NULL || method()->code_size() == 0) {
 754     // We are building a graph for a call to a native method.
 755     // All locals are live.
 756     return;
 757   }
 758 
 759   ResourceMark rm;
 760 
 761   // Consult the liveness information for the locals.  If any
 762   // of them are unused, then they can be replaced by top().  This
 763   // should help register allocation time and cut down on the size
 764   // of the deoptimization information.
 765   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 766 
 767   int len = (int)live_locals.size();
 768   assert(len <= jvms()->loc_size(), "too many live locals");
 769   for (int local = 0; local < len; local++) {
 770     if (!live_locals.at(local)) {
 771       set_local(local, top());
 772     }
 773   }
 774 }
 775 
 776 #ifdef ASSERT
 777 //-------------------------dead_locals_are_killed------------------------------
 778 // Return true if all dead locals are set to top in the map.
 779 // Used to assert "clean" debug info at various points.
 780 bool GraphKit::dead_locals_are_killed() {
 781   if (method() == NULL || method()->code_size() == 0) {
 782     // No locals need to be dead, so all is as it should be.
 783     return true;
 784   }
 785 
 786   // Make sure somebody called kill_dead_locals upstream.
 787   ResourceMark rm;
 788   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 789     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 790     SafePointNode* map = jvms->map();
 791     ciMethod* method = jvms->method();
 792     int       bci    = jvms->bci();
 793     if (jvms == this->jvms()) {
 794       bci = this->bci();  // it might not yet be synched
 795     }
 796     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 797     int len = (int)live_locals.size();
 798     if (!live_locals.is_valid() || len == 0)
 799       // This method is trivial, or is poisoned by a breakpoint.
 800       return true;
 801     assert(len == jvms->loc_size(), "live map consistent with locals map");
 802     for (int local = 0; local < len; local++) {
 803       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 804         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 805           tty->print_cr("Zombie local %d: ", local);
 806           jvms->dump();
 807         }
 808         return false;
 809       }
 810     }
 811   }
 812   return true;
 813 }
 814 
 815 #endif //ASSERT
 816 
 817 // Helper function for enforcing certain bytecodes to reexecute if
 818 // deoptimization happens
 819 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 820   ciMethod* cur_method = jvms->method();
 821   int       cur_bci   = jvms->bci();
 822   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 823     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 824     return Interpreter::bytecode_should_reexecute(code) ||
 825            is_anewarray && code == Bytecodes::_multianewarray;
 826     // Reexecute _multianewarray bytecode which was replaced with
 827     // sequence of [a]newarray. See Parse::do_multianewarray().
 828     //
 829     // Note: interpreter should not have it set since this optimization
 830     // is limited by dimensions and guarded by flag so in some cases
 831     // multianewarray() runtime calls will be generated and
 832     // the bytecode should not be reexecutes (stack will not be reset).
 833   } else
 834     return false;
 835 }
 836 
 837 // Helper function for adding JVMState and debug information to node
 838 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 839   // Add the safepoint edges to the call (or other safepoint).
 840 
 841   // Make sure dead locals are set to top.  This
 842   // should help register allocation time and cut down on the size
 843   // of the deoptimization information.
 844   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 845 
 846   // Walk the inline list to fill in the correct set of JVMState's
 847   // Also fill in the associated edges for each JVMState.
 848 
 849   // If the bytecode needs to be reexecuted we need to put
 850   // the arguments back on the stack.
 851   const bool should_reexecute = jvms()->should_reexecute();
 852   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 853 
 854   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 855   // undefined if the bci is different.  This is normal for Parse but it
 856   // should not happen for LibraryCallKit because only one bci is processed.
 857   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 858          "in LibraryCallKit the reexecute bit should not change");
 859 
 860   // If we are guaranteed to throw, we can prune everything but the
 861   // input to the current bytecode.
 862   bool can_prune_locals = false;
 863   uint stack_slots_not_pruned = 0;
 864   int inputs = 0, depth = 0;
 865   if (must_throw) {
 866     assert(method() == youngest_jvms->method(), "sanity");
 867     if (compute_stack_effects(inputs, depth)) {
 868       can_prune_locals = true;
 869       stack_slots_not_pruned = inputs;
 870     }
 871   }
 872 
 873   if (env()->should_retain_local_variables()) {
 874     // At any safepoint, this method can get breakpointed, which would
 875     // then require an immediate deoptimization.
 876     can_prune_locals = false;  // do not prune locals
 877     stack_slots_not_pruned = 0;
 878   }
 879 
 880   // do not scribble on the input jvms
 881   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 882   call->set_jvms(out_jvms); // Start jvms list for call node
 883 
 884   // For a known set of bytecodes, the interpreter should reexecute them if
 885   // deoptimization happens. We set the reexecute state for them here
 886   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 887       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 888     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 889   }
 890 
 891   // Presize the call:
 892   DEBUG_ONLY(uint non_debug_edges = call->req());
 893   call->add_req_batch(top(), youngest_jvms->debug_depth());
 894   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 895 
 896   // Set up edges so that the call looks like this:
 897   //  Call [state:] ctl io mem fptr retadr
 898   //       [parms:] parm0 ... parmN
 899   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 900   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 901   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 902   // Note that caller debug info precedes callee debug info.
 903 
 904   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 905   uint debug_ptr = call->req();
 906 
 907   // Loop over the map input edges associated with jvms, add them
 908   // to the call node, & reset all offsets to match call node array.
 909   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 910     uint debug_end   = debug_ptr;
 911     uint debug_start = debug_ptr - in_jvms->debug_size();
 912     debug_ptr = debug_start;  // back up the ptr
 913 
 914     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 915     uint j, k, l;
 916     SafePointNode* in_map = in_jvms->map();
 917     out_jvms->set_map(call);
 918 
 919     if (can_prune_locals) {
 920       assert(in_jvms->method() == out_jvms->method(), "sanity");
 921       // If the current throw can reach an exception handler in this JVMS,
 922       // then we must keep everything live that can reach that handler.
 923       // As a quick and dirty approximation, we look for any handlers at all.
 924       if (in_jvms->method()->has_exception_handlers()) {
 925         can_prune_locals = false;
 926       }
 927     }
 928 
 929     // Add the Locals
 930     k = in_jvms->locoff();
 931     l = in_jvms->loc_size();
 932     out_jvms->set_locoff(p);
 933     if (!can_prune_locals) {
 934       for (j = 0; j < l; j++)
 935         call->set_req(p++, in_map->in(k+j));
 936     } else {
 937       p += l;  // already set to top above by add_req_batch
 938     }
 939 
 940     // Add the Expression Stack
 941     k = in_jvms->stkoff();
 942     l = in_jvms->sp();
 943     out_jvms->set_stkoff(p);
 944     if (!can_prune_locals) {
 945       for (j = 0; j < l; j++)
 946         call->set_req(p++, in_map->in(k+j));
 947     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 948       // Divide stack into {S0,...,S1}, where S0 is set to top.
 949       uint s1 = stack_slots_not_pruned;
 950       stack_slots_not_pruned = 0;  // for next iteration
 951       if (s1 > l)  s1 = l;
 952       uint s0 = l - s1;
 953       p += s0;  // skip the tops preinstalled by add_req_batch
 954       for (j = s0; j < l; j++)
 955         call->set_req(p++, in_map->in(k+j));
 956     } else {
 957       p += l;  // already set to top above by add_req_batch
 958     }
 959 
 960     // Add the Monitors
 961     k = in_jvms->monoff();
 962     l = in_jvms->mon_size();
 963     out_jvms->set_monoff(p);
 964     for (j = 0; j < l; j++)
 965       call->set_req(p++, in_map->in(k+j));
 966 
 967     // Copy any scalar object fields.
 968     k = in_jvms->scloff();
 969     l = in_jvms->scl_size();
 970     out_jvms->set_scloff(p);
 971     for (j = 0; j < l; j++)
 972       call->set_req(p++, in_map->in(k+j));
 973 
 974     // Finish the new jvms.
 975     out_jvms->set_endoff(p);
 976 
 977     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 978     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 979     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 980     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 981     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 982     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 983 
 984     // Update the two tail pointers in parallel.
 985     out_jvms = out_jvms->caller();
 986     in_jvms  = in_jvms->caller();
 987   }
 988 
 989   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
 990 
 991   // Test the correctness of JVMState::debug_xxx accessors:
 992   assert(call->jvms()->debug_start() == non_debug_edges, "");
 993   assert(call->jvms()->debug_end()   == call->req(), "");
 994   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
 995 }
 996 
 997 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
 998   Bytecodes::Code code = java_bc();
 999   if (code == Bytecodes::_wide) {
1000     code = method()->java_code_at_bci(bci() + 1);
1001   }
1002 
1003   BasicType rtype = T_ILLEGAL;
1004   int       rsize = 0;
1005 
1006   if (code != Bytecodes::_illegal) {
1007     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1008     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1009     if (rtype < T_CONFLICT)
1010       rsize = type2size[rtype];
1011   }
1012 
1013   switch (code) {
1014   case Bytecodes::_illegal:
1015     return false;
1016 
1017   case Bytecodes::_ldc:
1018   case Bytecodes::_ldc_w:
1019   case Bytecodes::_ldc2_w:
1020     inputs = 0;
1021     break;
1022 
1023   case Bytecodes::_dup:         inputs = 1;  break;
1024   case Bytecodes::_dup_x1:      inputs = 2;  break;
1025   case Bytecodes::_dup_x2:      inputs = 3;  break;
1026   case Bytecodes::_dup2:        inputs = 2;  break;
1027   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1028   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1029   case Bytecodes::_swap:        inputs = 2;  break;
1030   case Bytecodes::_arraylength: inputs = 1;  break;
1031 
1032   case Bytecodes::_getstatic:
1033   case Bytecodes::_putstatic:
1034   case Bytecodes::_getfield:
1035   case Bytecodes::_putfield:
1036     {
1037       bool ignored_will_link;
1038       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1039       int      size  = field->type()->size();
1040       bool is_get = (depth >= 0), is_static = (depth & 1);
1041       inputs = (is_static ? 0 : 1);
1042       if (is_get) {
1043         depth = size - inputs;
1044       } else {
1045         inputs += size;        // putxxx pops the value from the stack
1046         depth = - inputs;
1047       }
1048     }
1049     break;
1050 
1051   case Bytecodes::_invokevirtual:
1052   case Bytecodes::_invokespecial:
1053   case Bytecodes::_invokestatic:
1054   case Bytecodes::_invokedynamic:
1055   case Bytecodes::_invokeinterface:
1056     {
1057       bool ignored_will_link;
1058       ciSignature* declared_signature = NULL;
1059       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1060       assert(declared_signature != NULL, "cannot be null");
1061       inputs   = declared_signature->arg_size_for_bc(code);
1062       int size = declared_signature->return_type()->size();
1063       depth = size - inputs;
1064     }
1065     break;
1066 
1067   case Bytecodes::_multianewarray:
1068     {
1069       ciBytecodeStream iter(method());
1070       iter.reset_to_bci(bci());
1071       iter.next();
1072       inputs = iter.get_dimensions();
1073       assert(rsize == 1, "");
1074       depth = rsize - inputs;
1075     }
1076     break;
1077 
1078   case Bytecodes::_ireturn:
1079   case Bytecodes::_lreturn:
1080   case Bytecodes::_freturn:
1081   case Bytecodes::_dreturn:
1082   case Bytecodes::_areturn:
1083     assert(rsize = -depth, "");
1084     inputs = rsize;
1085     break;
1086 
1087   case Bytecodes::_jsr:
1088   case Bytecodes::_jsr_w:
1089     inputs = 0;
1090     depth  = 1;                  // S.B. depth=1, not zero
1091     break;
1092 
1093   default:
1094     // bytecode produces a typed result
1095     inputs = rsize - depth;
1096     assert(inputs >= 0, "");
1097     break;
1098   }
1099 
1100 #ifdef ASSERT
1101   // spot check
1102   int outputs = depth + inputs;
1103   assert(outputs >= 0, "sanity");
1104   switch (code) {
1105   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1106   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1107   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1108   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1109   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1110   }
1111 #endif //ASSERT
1112 
1113   return true;
1114 }
1115 
1116 
1117 
1118 //------------------------------basic_plus_adr---------------------------------
1119 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1120   // short-circuit a common case
1121   if (offset == intcon(0))  return ptr;
1122   return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1123 }
1124 
1125 Node* GraphKit::ConvI2L(Node* offset) {
1126   // short-circuit a common case
1127   jint offset_con = find_int_con(offset, Type::OffsetBot);
1128   if (offset_con != Type::OffsetBot) {
1129     return longcon((jlong) offset_con);
1130   }
1131   return _gvn.transform( new (C) ConvI2LNode(offset));
1132 }
1133 
1134 Node* GraphKit::ConvI2UL(Node* offset) {
1135   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1136   if (offset_con != (juint) Type::OffsetBot) {
1137     return longcon((julong) offset_con);
1138   }
1139   Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
1140   Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
1141   return _gvn.transform( new (C) AndLNode(conv, mask) );
1142 }
1143 
1144 Node* GraphKit::ConvL2I(Node* offset) {
1145   // short-circuit a common case
1146   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1147   if (offset_con != (jlong)Type::OffsetBot) {
1148     return intcon((int) offset_con);
1149   }
1150   return _gvn.transform( new (C) ConvL2INode(offset));
1151 }
1152 
1153 //-------------------------load_object_klass-----------------------------------
1154 Node* GraphKit::load_object_klass(Node* obj) {
1155   // Special-case a fresh allocation to avoid building nodes:
1156   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1157   if (akls != NULL)  return akls;
1158   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1159   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1160 }
1161 
1162 //-------------------------load_array_length-----------------------------------
1163 Node* GraphKit::load_array_length(Node* array) {
1164   // Special-case a fresh allocation to avoid building nodes:
1165   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1166   Node *alen;
1167   if (alloc == NULL) {
1168     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1169     alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1170   } else {
1171     alen = alloc->Ideal_length();
1172     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1173     if (ccast != alen) {
1174       alen = _gvn.transform(ccast);
1175     }
1176   }
1177   return alen;
1178 }
1179 
1180 //------------------------------do_null_check----------------------------------
1181 // Helper function to do a NULL pointer check.  Returned value is
1182 // the incoming address with NULL casted away.  You are allowed to use the
1183 // not-null value only if you are control dependent on the test.
1184 extern int explicit_null_checks_inserted,
1185            explicit_null_checks_elided;
1186 Node* GraphKit::null_check_common(Node* value, BasicType type,
1187                                   // optional arguments for variations:
1188                                   bool assert_null,
1189                                   Node* *null_control) {
1190   assert(!assert_null || null_control == NULL, "not both at once");
1191   if (stopped())  return top();
1192   if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1193     // For some performance testing, we may wish to suppress null checking.
1194     value = cast_not_null(value);   // Make it appear to be non-null (4962416).
1195     return value;
1196   }
1197   explicit_null_checks_inserted++;
1198 
1199   // Construct NULL check
1200   Node *chk = NULL;
1201   switch(type) {
1202     case T_LONG   : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1203     case T_INT    : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1204     case T_ARRAY  : // fall through
1205       type = T_OBJECT;  // simplify further tests
1206     case T_OBJECT : {
1207       const Type *t = _gvn.type( value );
1208 
1209       const TypeOopPtr* tp = t->isa_oopptr();
1210       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1211           // Only for do_null_check, not any of its siblings:
1212           && !assert_null && null_control == NULL) {
1213         // Usually, any field access or invocation on an unloaded oop type
1214         // will simply fail to link, since the statically linked class is
1215         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1216         // the static class is loaded but the sharper oop type is not.
1217         // Rather than checking for this obscure case in lots of places,
1218         // we simply observe that a null check on an unloaded class
1219         // will always be followed by a nonsense operation, so we
1220         // can just issue the uncommon trap here.
1221         // Our access to the unloaded class will only be correct
1222         // after it has been loaded and initialized, which requires
1223         // a trip through the interpreter.
1224 #ifndef PRODUCT
1225         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1226 #endif
1227         uncommon_trap(Deoptimization::Reason_unloaded,
1228                       Deoptimization::Action_reinterpret,
1229                       tp->klass(), "!loaded");
1230         return top();
1231       }
1232 
1233       if (assert_null) {
1234         // See if the type is contained in NULL_PTR.
1235         // If so, then the value is already null.
1236         if (t->higher_equal(TypePtr::NULL_PTR)) {
1237           explicit_null_checks_elided++;
1238           return value;           // Elided null assert quickly!
1239         }
1240       } else {
1241         // See if mixing in the NULL pointer changes type.
1242         // If so, then the NULL pointer was not allowed in the original
1243         // type.  In other words, "value" was not-null.
1244         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1245           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1246           explicit_null_checks_elided++;
1247           return value;           // Elided null check quickly!
1248         }
1249       }
1250       chk = new (C) CmpPNode( value, null() );
1251       break;
1252     }
1253 
1254     default:
1255       fatal(err_msg_res("unexpected type: %s", type2name(type)));
1256   }
1257   assert(chk != NULL, "sanity check");
1258   chk = _gvn.transform(chk);
1259 
1260   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1261   BoolNode *btst = new (C) BoolNode( chk, btest);
1262   Node   *tst = _gvn.transform( btst );
1263 
1264   //-----------
1265   // if peephole optimizations occurred, a prior test existed.
1266   // If a prior test existed, maybe it dominates as we can avoid this test.
1267   if (tst != btst && type == T_OBJECT) {
1268     // At this point we want to scan up the CFG to see if we can
1269     // find an identical test (and so avoid this test altogether).
1270     Node *cfg = control();
1271     int depth = 0;
1272     while( depth < 16 ) {       // Limit search depth for speed
1273       if( cfg->Opcode() == Op_IfTrue &&
1274           cfg->in(0)->in(1) == tst ) {
1275         // Found prior test.  Use "cast_not_null" to construct an identical
1276         // CastPP (and hence hash to) as already exists for the prior test.
1277         // Return that casted value.
1278         if (assert_null) {
1279           replace_in_map(value, null());
1280           return null();  // do not issue the redundant test
1281         }
1282         Node *oldcontrol = control();
1283         set_control(cfg);
1284         Node *res = cast_not_null(value);
1285         set_control(oldcontrol);
1286         explicit_null_checks_elided++;
1287         return res;
1288       }
1289       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1290       if (cfg == NULL)  break;  // Quit at region nodes
1291       depth++;
1292     }
1293   }
1294 
1295   //-----------
1296   // Branch to failure if null
1297   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1298   Deoptimization::DeoptReason reason;
1299   if (assert_null)
1300     reason = Deoptimization::Reason_null_assert;
1301   else if (type == T_OBJECT)
1302     reason = Deoptimization::Reason_null_check;
1303   else
1304     reason = Deoptimization::Reason_div0_check;
1305 
1306   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1307   // ciMethodData::has_trap_at will return a conservative -1 if any
1308   // must-be-null assertion has failed.  This could cause performance
1309   // problems for a method after its first do_null_assert failure.
1310   // Consider using 'Reason_class_check' instead?
1311 
1312   // To cause an implicit null check, we set the not-null probability
1313   // to the maximum (PROB_MAX).  For an explicit check the probability
1314   // is set to a smaller value.
1315   if (null_control != NULL || too_many_traps(reason)) {
1316     // probability is less likely
1317     ok_prob =  PROB_LIKELY_MAG(3);
1318   } else if (!assert_null &&
1319              (ImplicitNullCheckThreshold > 0) &&
1320              method() != NULL &&
1321              (method()->method_data()->trap_count(reason)
1322               >= (uint)ImplicitNullCheckThreshold)) {
1323     ok_prob =  PROB_LIKELY_MAG(3);
1324   }
1325 
1326   if (null_control != NULL) {
1327     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1328     Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1329     set_control(      _gvn.transform( new (C) IfTrueNode(iff)));
1330     if (null_true == top())
1331       explicit_null_checks_elided++;
1332     (*null_control) = null_true;
1333   } else {
1334     BuildCutout unless(this, tst, ok_prob);
1335     // Check for optimizer eliding test at parse time
1336     if (stopped()) {
1337       // Failure not possible; do not bother making uncommon trap.
1338       explicit_null_checks_elided++;
1339     } else if (assert_null) {
1340       uncommon_trap(reason,
1341                     Deoptimization::Action_make_not_entrant,
1342                     NULL, "assert_null");
1343     } else {
1344       replace_in_map(value, zerocon(type));
1345       builtin_throw(reason);
1346     }
1347   }
1348 
1349   // Must throw exception, fall-thru not possible?
1350   if (stopped()) {
1351     return top();               // No result
1352   }
1353 
1354   if (assert_null) {
1355     // Cast obj to null on this path.
1356     replace_in_map(value, zerocon(type));
1357     return zerocon(type);
1358   }
1359 
1360   // Cast obj to not-null on this path, if there is no null_control.
1361   // (If there is a null_control, a non-null value may come back to haunt us.)
1362   if (type == T_OBJECT) {
1363     Node* cast = cast_not_null(value, false);
1364     if (null_control == NULL || (*null_control) == top())
1365       replace_in_map(value, cast);
1366     value = cast;
1367   }
1368 
1369   return value;
1370 }
1371 
1372 
1373 //------------------------------cast_not_null----------------------------------
1374 // Cast obj to not-null on this path
1375 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1376   const Type *t = _gvn.type(obj);
1377   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1378   // Object is already not-null?
1379   if( t == t_not_null ) return obj;
1380 
1381   Node *cast = new (C) CastPPNode(obj,t_not_null);
1382   cast->init_req(0, control());
1383   cast = _gvn.transform( cast );
1384 
1385   // Scan for instances of 'obj' in the current JVM mapping.
1386   // These instances are known to be not-null after the test.
1387   if (do_replace_in_map)
1388     replace_in_map(obj, cast);
1389 
1390   return cast;                  // Return casted value
1391 }
1392 
1393 
1394 //--------------------------replace_in_map-------------------------------------
1395 void GraphKit::replace_in_map(Node* old, Node* neww) {
1396   if (old == neww) {
1397     return;
1398   }
1399 
1400   map()->replace_edge(old, neww);
1401 
1402   // Note: This operation potentially replaces any edge
1403   // on the map.  This includes locals, stack, and monitors
1404   // of the current (innermost) JVM state.
1405 
1406   // don't let inconsistent types from profiling escape this
1407   // method
1408 
1409   const Type* told = _gvn.type(old);
1410   const Type* tnew = _gvn.type(neww);
1411 
1412   if (!tnew->higher_equal(told)) {
1413     return;
1414   }
1415 
1416   map()->record_replaced_node(old, neww);
1417 }
1418 
1419 
1420 //=============================================================================
1421 //--------------------------------memory---------------------------------------
1422 Node* GraphKit::memory(uint alias_idx) {
1423   MergeMemNode* mem = merged_memory();
1424   Node* p = mem->memory_at(alias_idx);
1425   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1426   return p;
1427 }
1428 
1429 //-----------------------------reset_memory------------------------------------
1430 Node* GraphKit::reset_memory() {
1431   Node* mem = map()->memory();
1432   // do not use this node for any more parsing!
1433   debug_only( map()->set_memory((Node*)NULL) );
1434   return _gvn.transform( mem );
1435 }
1436 
1437 //------------------------------set_all_memory---------------------------------
1438 void GraphKit::set_all_memory(Node* newmem) {
1439   Node* mergemem = MergeMemNode::make(C, newmem);
1440   gvn().set_type_bottom(mergemem);
1441   map()->set_memory(mergemem);
1442 }
1443 
1444 //------------------------------set_all_memory_call----------------------------
1445 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1446   Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1447   set_all_memory(newmem);
1448 }
1449 
1450 //=============================================================================
1451 //
1452 // parser factory methods for MemNodes
1453 //
1454 // These are layered on top of the factory methods in LoadNode and StoreNode,
1455 // and integrate with the parser's memory state and _gvn engine.
1456 //
1457 
1458 // factory methods in "int adr_idx"
1459 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1460                           int adr_idx,
1461                           MemNode::MemOrd mo,
1462                           LoadNode::ControlDependency control_dependency,
1463                           bool require_atomic_access,
1464                           bool unaligned,
1465                           bool mismatched) {
1466   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1467   const TypePtr* adr_type = NULL; // debug-mode-only argument
1468   debug_only(adr_type = C->get_adr_type(adr_idx));
1469   Node* mem = memory(adr_idx);
1470   Node* ld;
1471   if (require_atomic_access && bt == T_LONG) {
1472     ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency);
1473   } else if (require_atomic_access && bt == T_DOUBLE) {
1474     ld = LoadDNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency);
1475   } else {
1476     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency);
1477   }
1478   if (unaligned) {
1479     ld->as_Load()->set_unaligned_access();
1480   }
1481   if (mismatched) {
1482     ld->as_Load()->set_mismatched_access();
1483   }
1484   ld = _gvn.transform(ld);
1485   if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1486     // Improve graph before escape analysis and boxing elimination.
1487     record_for_igvn(ld);
1488   }
1489   return ld;
1490 }
1491 
1492 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1493                                 int adr_idx,
1494                                 MemNode::MemOrd mo,
1495                                 bool require_atomic_access,
1496                                 bool unaligned,
1497                                 bool mismatched) {
1498   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1499   const TypePtr* adr_type = NULL;
1500   debug_only(adr_type = C->get_adr_type(adr_idx));
1501   Node *mem = memory(adr_idx);
1502   Node* st;
1503   if (require_atomic_access && bt == T_LONG) {
1504     st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1505   } else if (require_atomic_access && bt == T_DOUBLE) {
1506     st = StoreDNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1507   } else {
1508     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1509   }
1510   if (unaligned) {
1511     st->as_Store()->set_unaligned_access();
1512   }
1513   if (mismatched) {
1514     st->as_Store()->set_mismatched_access();
1515   }
1516   st = _gvn.transform(st);
1517   set_memory(st, adr_idx);
1518   // Back-to-back stores can only remove intermediate store with DU info
1519   // so push on worklist for optimizer.
1520   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1521     record_for_igvn(st);
1522 
1523   return st;
1524 }
1525 
1526 
1527 void GraphKit::pre_barrier(bool do_load,
1528                            Node* ctl,
1529                            Node* obj,
1530                            Node* adr,
1531                            uint  adr_idx,
1532                            Node* val,
1533                            const TypeOopPtr* val_type,
1534                            Node* pre_val,
1535                            BasicType bt) {
1536 
1537   BarrierSet* bs = Universe::heap()->barrier_set();
1538   set_control(ctl);
1539   switch (bs->kind()) {
1540     case BarrierSet::G1SATBCT:
1541     case BarrierSet::G1SATBCTLogging:
1542       g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1543       break;
1544     case BarrierSet::ShenandoahBarrierSet:
1545       if (ShenandoahSATBBarrier) {
1546         g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1547       }
1548       break;
1549     case BarrierSet::CardTableModRef:
1550     case BarrierSet::CardTableExtension:
1551     case BarrierSet::ModRef:
1552       break;
1553 
1554     case BarrierSet::Other:
1555     default      :
1556       ShouldNotReachHere();
1557 
1558   }
1559 }
1560 
1561 bool GraphKit::can_move_pre_barrier() const {
1562   BarrierSet* bs = Universe::heap()->barrier_set();
1563   switch (bs->kind()) {
1564     case BarrierSet::G1SATBCT:
1565     case BarrierSet::G1SATBCTLogging:
1566     case BarrierSet::ShenandoahBarrierSet:
1567       return true; // Can move it if no safepoint
1568 
1569     case BarrierSet::CardTableModRef:
1570     case BarrierSet::CardTableExtension:
1571     case BarrierSet::ModRef:
1572       return true; // There is no pre-barrier
1573 
1574     case BarrierSet::Other:
1575     default      :
1576       ShouldNotReachHere();
1577   }
1578   return false;
1579 }
1580 
1581 void GraphKit::post_barrier(Node* ctl,
1582                             Node* store,
1583                             Node* obj,
1584                             Node* adr,
1585                             uint  adr_idx,
1586                             Node* val,
1587                             BasicType bt,
1588                             bool use_precise) {
1589   BarrierSet* bs = Universe::heap()->barrier_set();
1590   set_control(ctl);
1591   switch (bs->kind()) {
1592     case BarrierSet::G1SATBCT:
1593     case BarrierSet::G1SATBCTLogging:
1594       g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1595       break;
1596     case BarrierSet::ShenandoahBarrierSet:
1597       if (ShenandoahStoreValEnqueueBarrier) {
1598         g1_write_barrier_pre(false, NULL, NULL, max_juint, NULL, NULL, val, bt);
1599       }
1600       break;
1601     case BarrierSet::CardTableModRef:
1602     case BarrierSet::CardTableExtension:
1603       write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1604       break;
1605 
1606     case BarrierSet::ModRef:
1607       break;
1608 
1609     case BarrierSet::Other:
1610     default      :
1611       ShouldNotReachHere();
1612 
1613   }
1614 }
1615 
1616 Node* GraphKit::store_oop(Node* ctl,
1617                           Node* obj,
1618                           Node* adr,
1619                           const TypePtr* adr_type,
1620                           Node* val,
1621                           const TypeOopPtr* val_type,
1622                           BasicType bt,
1623                           bool use_precise,
1624                           MemNode::MemOrd mo,
1625                           bool mismatched) {
1626   // Transformation of a value which could be NULL pointer (CastPP #NULL)
1627   // could be delayed during Parse (for example, in adjust_map_after_if()).
1628   // Execute transformation here to avoid barrier generation in such case.
1629   if (_gvn.type(val) == TypePtr::NULL_PTR)
1630     val = _gvn.makecon(TypePtr::NULL_PTR);
1631 
1632   set_control(ctl);
1633   if (stopped()) return top(); // Dead path ?
1634 
1635   assert(bt == T_OBJECT, "sanity");
1636   assert(val != NULL, "not dead path");
1637   uint adr_idx = C->get_alias_index(adr_type);
1638   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1639 
1640   pre_barrier(true /* do_load */,
1641               control(), obj, adr, adr_idx, val, val_type,
1642               NULL /* pre_val */,
1643               bt);
1644 
1645   Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1646   post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1647   return store;
1648 }
1649 
1650 // Could be an array or object we don't know at compile time (unsafe ref.)
1651 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1652                              Node* obj,   // containing obj
1653                              Node* adr,  // actual adress to store val at
1654                              const TypePtr* adr_type,
1655                              Node* val,
1656                              BasicType bt,
1657                              MemNode::MemOrd mo,
1658                              bool mismatched) {
1659   Compile::AliasType* at = C->alias_type(adr_type);
1660   const TypeOopPtr* val_type = NULL;
1661   if (adr_type->isa_instptr()) {
1662     if (at->field() != NULL) {
1663       // known field.  This code is a copy of the do_put_xxx logic.
1664       ciField* field = at->field();
1665       if (!field->type()->is_loaded()) {
1666         val_type = TypeInstPtr::BOTTOM;
1667       } else {
1668         val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1669       }
1670     }
1671   } else if (adr_type->isa_aryptr()) {
1672     val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1673   }
1674   if (val_type == NULL) {
1675     val_type = TypeInstPtr::BOTTOM;
1676   }
1677   return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1678 }
1679 
1680 
1681 //-------------------------array_element_address-------------------------
1682 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1683                                       const TypeInt* sizetype, Node* ctrl) {
1684   uint shift  = exact_log2(type2aelembytes(elembt));
1685   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1686 
1687   // short-circuit a common case (saves lots of confusing waste motion)
1688   jint idx_con = find_int_con(idx, -1);
1689   if (idx_con >= 0) {
1690     intptr_t offset = header + ((intptr_t)idx_con << shift);
1691     return basic_plus_adr(ary, offset);
1692   }
1693 
1694   // must be correct type for alignment purposes
1695   Node* base  = basic_plus_adr(ary, header);
1696 #ifdef _LP64
1697   // The scaled index operand to AddP must be a clean 64-bit value.
1698   // Java allows a 32-bit int to be incremented to a negative
1699   // value, which appears in a 64-bit register as a large
1700   // positive number.  Using that large positive number as an
1701   // operand in pointer arithmetic has bad consequences.
1702   // On the other hand, 32-bit overflow is rare, and the possibility
1703   // can often be excluded, if we annotate the ConvI2L node with
1704   // a type assertion that its value is known to be a small positive
1705   // number.  (The prior range check has ensured this.)
1706   // This assertion is used by ConvI2LNode::Ideal.
1707   int index_max = max_jint - 1;  // array size is max_jint, index is one less
1708   if (sizetype != NULL) index_max = sizetype->_hi - 1;
1709   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
1710   idx = C->constrained_convI2L(&_gvn, idx, iidxtype, ctrl);
1711 #endif
1712   Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1713   return basic_plus_adr(ary, base, scale);
1714 }
1715 
1716 //-------------------------load_array_element-------------------------
1717 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1718   const Type* elemtype = arytype->elem();
1719   BasicType elembt = elemtype->array_element_basic_type();
1720   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1721   if (elembt == T_NARROWOOP) {
1722     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1723   }
1724   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1725 #if INCLUDE_ALL_GCS
1726   if (UseShenandoahGC && (elembt == T_OBJECT || elembt == T_ARRAY)) {
1727     ld = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, ld);
1728   }
1729 #endif
1730   return ld;
1731 }
1732 
1733 //-------------------------set_arguments_for_java_call-------------------------
1734 // Arguments (pre-popped from the stack) are taken from the JVMS.
1735 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1736   // Add the call arguments:
1737   uint nargs = call->method()->arg_size();
1738   for (uint i = 0; i < nargs; i++) {
1739     Node* arg = argument(i);
1740     call->init_req(i + TypeFunc::Parms, arg);
1741   }
1742 }
1743 
1744 //---------------------------set_edges_for_java_call---------------------------
1745 // Connect a newly created call into the current JVMS.
1746 // A return value node (if any) is returned from set_edges_for_java_call.
1747 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1748 
1749   // Add the predefined inputs:
1750   call->init_req( TypeFunc::Control, control() );
1751   call->init_req( TypeFunc::I_O    , i_o() );
1752   call->init_req( TypeFunc::Memory , reset_memory() );
1753   call->init_req( TypeFunc::FramePtr, frameptr() );
1754   call->init_req( TypeFunc::ReturnAdr, top() );
1755 
1756   add_safepoint_edges(call, must_throw);
1757 
1758   Node* xcall = _gvn.transform(call);
1759 
1760   if (xcall == top()) {
1761     set_control(top());
1762     return;
1763   }
1764   assert(xcall == call, "call identity is stable");
1765 
1766   // Re-use the current map to produce the result.
1767 
1768   set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1769   set_i_o(    _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1770   set_all_memory_call(xcall, separate_io_proj);
1771 
1772   //return xcall;   // no need, caller already has it
1773 }
1774 
1775 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1776   if (stopped())  return top();  // maybe the call folded up?
1777 
1778   // Capture the return value, if any.
1779   Node* ret;
1780   if (call->method() == NULL ||
1781       call->method()->return_type()->basic_type() == T_VOID)
1782         ret = top();
1783   else  ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1784 
1785   // Note:  Since any out-of-line call can produce an exception,
1786   // we always insert an I_O projection from the call into the result.
1787 
1788   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1789 
1790   if (separate_io_proj) {
1791     // The caller requested separate projections be used by the fall
1792     // through and exceptional paths, so replace the projections for
1793     // the fall through path.
1794     set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1795     set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1796   }
1797   return ret;
1798 }
1799 
1800 //--------------------set_predefined_input_for_runtime_call--------------------
1801 // Reading and setting the memory state is way conservative here.
1802 // The real problem is that I am not doing real Type analysis on memory,
1803 // so I cannot distinguish card mark stores from other stores.  Across a GC
1804 // point the Store Barrier and the card mark memory has to agree.  I cannot
1805 // have a card mark store and its barrier split across the GC point from
1806 // either above or below.  Here I get that to happen by reading ALL of memory.
1807 // A better answer would be to separate out card marks from other memory.
1808 // For now, return the input memory state, so that it can be reused
1809 // after the call, if this call has restricted memory effects.
1810 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1811   // Set fixed predefined input arguments
1812   Node* memory = reset_memory();
1813   Node* m = narrow_mem == NULL ? memory : narrow_mem;
1814   call->init_req( TypeFunc::Control,   control()  );
1815   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1816   call->init_req( TypeFunc::Memory,    m          ); // may gc ptrs
1817   call->init_req( TypeFunc::FramePtr,  frameptr() );
1818   call->init_req( TypeFunc::ReturnAdr, top()      );
1819   return memory;
1820 }
1821 
1822 //-------------------set_predefined_output_for_runtime_call--------------------
1823 // Set control and memory (not i_o) from the call.
1824 // If keep_mem is not NULL, use it for the output state,
1825 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1826 // If hook_mem is NULL, this call produces no memory effects at all.
1827 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1828 // then only that memory slice is taken from the call.
1829 // In the last case, we must put an appropriate memory barrier before
1830 // the call, so as to create the correct anti-dependencies on loads
1831 // preceding the call.
1832 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1833                                                       Node* keep_mem,
1834                                                       const TypePtr* hook_mem) {
1835   // no i/o
1836   set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1837   if (keep_mem) {
1838     // First clone the existing memory state
1839     set_all_memory(keep_mem);
1840     if (hook_mem != NULL) {
1841       // Make memory for the call
1842       Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1843       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1844       // We also use hook_mem to extract specific effects from arraycopy stubs.
1845       set_memory(mem, hook_mem);
1846     }
1847     // ...else the call has NO memory effects.
1848 
1849     // Make sure the call advertises its memory effects precisely.
1850     // This lets us build accurate anti-dependences in gcm.cpp.
1851     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1852            "call node must be constructed correctly");
1853   } else {
1854     assert(hook_mem == NULL, "");
1855     // This is not a "slow path" call; all memory comes from the call.
1856     set_all_memory_call(call);
1857   }
1858 }
1859 
1860 
1861 // Replace the call with the current state of the kit.
1862 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1863   JVMState* ejvms = NULL;
1864   if (has_exceptions()) {
1865     ejvms = transfer_exceptions_into_jvms();
1866   }
1867 
1868   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1869   ReplacedNodes replaced_nodes_exception;
1870   Node* ex_ctl = top();
1871 
1872   SafePointNode* final_state = stop();
1873 
1874   // Find all the needed outputs of this call
1875   CallProjections callprojs;
1876   call->extract_projections(&callprojs, true);
1877 
1878   Node* init_mem = call->in(TypeFunc::Memory);
1879   Node* final_mem = final_state->in(TypeFunc::Memory);
1880   Node* final_ctl = final_state->in(TypeFunc::Control);
1881   Node* final_io = final_state->in(TypeFunc::I_O);
1882 
1883   // Replace all the old call edges with the edges from the inlining result
1884   if (callprojs.fallthrough_catchproj != NULL) {
1885     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1886   }
1887   if (callprojs.fallthrough_memproj != NULL) {
1888     if (final_mem->is_MergeMem()) {
1889       // Parser's exits MergeMem was not transformed but may be optimized
1890       final_mem = _gvn.transform(final_mem);
1891     }
1892     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
1893   }
1894   if (callprojs.fallthrough_ioproj != NULL) {
1895     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
1896   }
1897 
1898   // Replace the result with the new result if it exists and is used
1899   if (callprojs.resproj != NULL && result != NULL) {
1900     C->gvn_replace_by(callprojs.resproj, result);
1901   }
1902 
1903   if (ejvms == NULL) {
1904     // No exception edges to simply kill off those paths
1905     if (callprojs.catchall_catchproj != NULL) {
1906       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1907     }
1908     if (callprojs.catchall_memproj != NULL) {
1909       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1910     }
1911     if (callprojs.catchall_ioproj != NULL) {
1912       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1913     }
1914     // Replace the old exception object with top
1915     if (callprojs.exobj != NULL) {
1916       C->gvn_replace_by(callprojs.exobj, C->top());
1917     }
1918   } else {
1919     GraphKit ekit(ejvms);
1920 
1921     // Load my combined exception state into the kit, with all phis transformed:
1922     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1923     replaced_nodes_exception = ex_map->replaced_nodes();
1924 
1925     Node* ex_oop = ekit.use_exception_state(ex_map);
1926 
1927     if (callprojs.catchall_catchproj != NULL) {
1928       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1929       ex_ctl = ekit.control();
1930     }
1931     if (callprojs.catchall_memproj != NULL) {
1932       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1933     }
1934     if (callprojs.catchall_ioproj != NULL) {
1935       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1936     }
1937 
1938     // Replace the old exception object with the newly created one
1939     if (callprojs.exobj != NULL) {
1940       C->gvn_replace_by(callprojs.exobj, ex_oop);
1941     }
1942   }
1943 
1944   // Disconnect the call from the graph
1945   call->disconnect_inputs(NULL, C);
1946   C->gvn_replace_by(call, C->top());
1947 
1948   // Clean up any MergeMems that feed other MergeMems since the
1949   // optimizer doesn't like that.
1950   if (final_mem->is_MergeMem()) {
1951     Node_List wl;
1952     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1953       Node* m = i.get();
1954       if (m->is_MergeMem() && !wl.contains(m)) {
1955         wl.push(m);
1956       }
1957     }
1958     while (wl.size()  > 0) {
1959       _gvn.transform(wl.pop());
1960     }
1961   }
1962 
1963   if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1964     replaced_nodes.apply(C, final_ctl);
1965   }
1966   if (!ex_ctl->is_top() && do_replaced_nodes) {
1967     replaced_nodes_exception.apply(C, ex_ctl);
1968   }
1969 }
1970 
1971 
1972 //------------------------------increment_counter------------------------------
1973 // for statistics: increment a VM counter by 1
1974 
1975 void GraphKit::increment_counter(address counter_addr) {
1976   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1977   increment_counter(adr1);
1978 }
1979 
1980 void GraphKit::increment_counter(Node* counter_addr) {
1981   int adr_type = Compile::AliasIdxRaw;
1982   Node* ctrl = control();
1983   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1984   Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1985   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1986 }
1987 
1988 
1989 //------------------------------uncommon_trap----------------------------------
1990 // Bail out to the interpreter in mid-method.  Implemented by calling the
1991 // uncommon_trap blob.  This helper function inserts a runtime call with the
1992 // right debug info.
1993 void GraphKit::uncommon_trap(int trap_request,
1994                              ciKlass* klass, const char* comment,
1995                              bool must_throw,
1996                              bool keep_exact_action) {
1997   if (failing())  stop();
1998   if (stopped())  return; // trap reachable?
1999 
2000   // Note:  If ProfileTraps is true, and if a deopt. actually
2001   // occurs here, the runtime will make sure an MDO exists.  There is
2002   // no need to call method()->ensure_method_data() at this point.
2003 
2004   // Set the stack pointer to the right value for reexecution:
2005   set_sp(reexecute_sp());
2006 
2007 #ifdef ASSERT
2008   if (!must_throw) {
2009     // Make sure the stack has at least enough depth to execute
2010     // the current bytecode.
2011     int inputs, ignored_depth;
2012     if (compute_stack_effects(inputs, ignored_depth)) {
2013       assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2014              Bytecodes::name(java_bc()), sp(), inputs));
2015     }
2016   }
2017 #endif
2018 
2019   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2020   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2021 
2022   switch (action) {
2023   case Deoptimization::Action_maybe_recompile:
2024   case Deoptimization::Action_reinterpret:
2025     // Temporary fix for 6529811 to allow virtual calls to be sure they
2026     // get the chance to go from mono->bi->mega
2027     if (!keep_exact_action &&
2028         Deoptimization::trap_request_index(trap_request) < 0 &&
2029         too_many_recompiles(reason)) {
2030       // This BCI is causing too many recompilations.
2031       if (C->log() != NULL) {
2032         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2033                 Deoptimization::trap_reason_name(reason),
2034                 Deoptimization::trap_action_name(action));
2035       }
2036       action = Deoptimization::Action_none;
2037       trap_request = Deoptimization::make_trap_request(reason, action);
2038     } else {
2039       C->set_trap_can_recompile(true);
2040     }
2041     break;
2042   case Deoptimization::Action_make_not_entrant:
2043     C->set_trap_can_recompile(true);
2044     break;
2045 #ifdef ASSERT
2046   case Deoptimization::Action_none:
2047   case Deoptimization::Action_make_not_compilable:
2048     break;
2049   default:
2050     fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2051     break;
2052 #endif
2053   }
2054 
2055   if (TraceOptoParse) {
2056     char buf[100];
2057     tty->print_cr("Uncommon trap %s at bci:%d",
2058                   Deoptimization::format_trap_request(buf, sizeof(buf),
2059                                                       trap_request), bci());
2060   }
2061 
2062   CompileLog* log = C->log();
2063   if (log != NULL) {
2064     int kid = (klass == NULL)? -1: log->identify(klass);
2065     log->begin_elem("uncommon_trap bci='%d'", bci());
2066     char buf[100];
2067     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2068                                                           trap_request));
2069     if (kid >= 0)         log->print(" klass='%d'", kid);
2070     if (comment != NULL)  log->print(" comment='%s'", comment);
2071     log->end_elem();
2072   }
2073 
2074   // Make sure any guarding test views this path as very unlikely
2075   Node *i0 = control()->in(0);
2076   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2077     IfNode *iff = i0->as_If();
2078     float f = iff->_prob;   // Get prob
2079     if (control()->Opcode() == Op_IfTrue) {
2080       if (f > PROB_UNLIKELY_MAG(4))
2081         iff->_prob = PROB_MIN;
2082     } else {
2083       if (f < PROB_LIKELY_MAG(4))
2084         iff->_prob = PROB_MAX;
2085     }
2086   }
2087 
2088   // Clear out dead values from the debug info.
2089   kill_dead_locals();
2090 
2091   // Now insert the uncommon trap subroutine call
2092   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2093   const TypePtr* no_memory_effects = NULL;
2094   // Pass the index of the class to be loaded
2095   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2096                                  (must_throw ? RC_MUST_THROW : 0),
2097                                  OptoRuntime::uncommon_trap_Type(),
2098                                  call_addr, "uncommon_trap", no_memory_effects,
2099                                  intcon(trap_request));
2100   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2101          "must extract request correctly from the graph");
2102   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2103 
2104   call->set_req(TypeFunc::ReturnAdr, returnadr());
2105   // The debug info is the only real input to this call.
2106 
2107   // Halt-and-catch fire here.  The above call should never return!
2108   HaltNode* halt = new(C) HaltNode(control(), frameptr());
2109   _gvn.set_type_bottom(halt);
2110   root()->add_req(halt);
2111 
2112   stop_and_kill_map();
2113 }
2114 
2115 
2116 //--------------------------just_allocated_object------------------------------
2117 // Report the object that was just allocated.
2118 // It must be the case that there are no intervening safepoints.
2119 // We use this to determine if an object is so "fresh" that
2120 // it does not require card marks.
2121 Node* GraphKit::just_allocated_object(Node* current_control) {
2122   if (C->recent_alloc_ctl() == current_control)
2123     return C->recent_alloc_obj();
2124   return NULL;
2125 }
2126 
2127 
2128 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2129   // (Note:  TypeFunc::make has a cache that makes this fast.)
2130   const TypeFunc* tf    = TypeFunc::make(dest_method);
2131   int             nargs = tf->_domain->_cnt - TypeFunc::Parms;
2132   for (int j = 0; j < nargs; j++) {
2133     const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2134     if( targ->basic_type() == T_DOUBLE ) {
2135       // If any parameters are doubles, they must be rounded before
2136       // the call, dstore_rounding does gvn.transform
2137       Node *arg = argument(j);
2138       arg = dstore_rounding(arg);
2139       set_argument(j, arg);
2140     }
2141   }
2142 }
2143 
2144 /**
2145  * Record profiling data exact_kls for Node n with the type system so
2146  * that it can propagate it (speculation)
2147  *
2148  * @param n          node that the type applies to
2149  * @param exact_kls  type from profiling
2150  *
2151  * @return           node with improved type
2152  */
2153 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2154   const Type* current_type = _gvn.type(n);
2155   assert(UseTypeSpeculation, "type speculation must be on");
2156 
2157   const TypeOopPtr* speculative = current_type->speculative();
2158 
2159   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2160     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2161     const TypeOopPtr* xtype = tklass->as_instance_type();
2162     assert(xtype->klass_is_exact(), "Should be exact");
2163     // record the new speculative type's depth
2164     speculative = xtype->with_inline_depth(jvms()->depth());
2165   }
2166 
2167   if (speculative != current_type->speculative()) {
2168     // Build a type with a speculative type (what we think we know
2169     // about the type but will need a guard when we use it)
2170     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2171     // We're changing the type, we need a new CheckCast node to carry
2172     // the new type. The new type depends on the control: what
2173     // profiling tells us is only valid from here as far as we can
2174     // tell.
2175     Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2176     cast = _gvn.transform(cast);
2177     replace_in_map(n, cast);
2178     n = cast;
2179   }
2180 
2181   return n;
2182 }
2183 
2184 /**
2185  * Record profiling data from receiver profiling at an invoke with the
2186  * type system so that it can propagate it (speculation)
2187  *
2188  * @param n  receiver node
2189  *
2190  * @return   node with improved type
2191  */
2192 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2193   if (!UseTypeSpeculation) {
2194     return n;
2195   }
2196   ciKlass* exact_kls = profile_has_unique_klass();
2197   return record_profile_for_speculation(n, exact_kls);
2198 }
2199 
2200 /**
2201  * Record profiling data from argument profiling at an invoke with the
2202  * type system so that it can propagate it (speculation)
2203  *
2204  * @param dest_method  target method for the call
2205  * @param bc           what invoke bytecode is this?
2206  */
2207 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2208   if (!UseTypeSpeculation) {
2209     return;
2210   }
2211   const TypeFunc* tf    = TypeFunc::make(dest_method);
2212   int             nargs = tf->_domain->_cnt - TypeFunc::Parms;
2213   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2214   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2215     const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2216     if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2217       ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2218       if (better_type != NULL) {
2219         record_profile_for_speculation(argument(j), better_type);
2220       }
2221       i++;
2222     }
2223   }
2224 }
2225 
2226 /**
2227  * Record profiling data from parameter profiling at an invoke with
2228  * the type system so that it can propagate it (speculation)
2229  */
2230 void GraphKit::record_profiled_parameters_for_speculation() {
2231   if (!UseTypeSpeculation) {
2232     return;
2233   }
2234   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2235     if (_gvn.type(local(i))->isa_oopptr()) {
2236       ciKlass* better_type = method()->parameter_profiled_type(j);
2237       if (better_type != NULL) {
2238         record_profile_for_speculation(local(i), better_type);
2239       }
2240       j++;
2241     }
2242   }
2243 }
2244 
2245 void GraphKit::round_double_result(ciMethod* dest_method) {
2246   // A non-strict method may return a double value which has an extended
2247   // exponent, but this must not be visible in a caller which is 'strict'
2248   // If a strict caller invokes a non-strict callee, round a double result
2249 
2250   BasicType result_type = dest_method->return_type()->basic_type();
2251   assert( method() != NULL, "must have caller context");
2252   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2253     // Destination method's return value is on top of stack
2254     // dstore_rounding() does gvn.transform
2255     Node *result = pop_pair();
2256     result = dstore_rounding(result);
2257     push_pair(result);
2258   }
2259 }
2260 
2261 // rounding for strict float precision conformance
2262 Node* GraphKit::precision_rounding(Node* n) {
2263   return UseStrictFP && _method->flags().is_strict()
2264     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2265     ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2266     : n;
2267 }
2268 
2269 // rounding for strict double precision conformance
2270 Node* GraphKit::dprecision_rounding(Node *n) {
2271   return UseStrictFP && _method->flags().is_strict()
2272     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2273     ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2274     : n;
2275 }
2276 
2277 // rounding for non-strict double stores
2278 Node* GraphKit::dstore_rounding(Node* n) {
2279   return Matcher::strict_fp_requires_explicit_rounding
2280     && UseSSE <= 1
2281     ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2282     : n;
2283 }
2284 
2285 //=============================================================================
2286 // Generate a fast path/slow path idiom.  Graph looks like:
2287 // [foo] indicates that 'foo' is a parameter
2288 //
2289 //              [in]     NULL
2290 //                 \    /
2291 //                  CmpP
2292 //                  Bool ne
2293 //                   If
2294 //                  /  \
2295 //              True    False-<2>
2296 //              / |
2297 //             /  cast_not_null
2298 //           Load  |    |   ^
2299 //        [fast_test]   |   |
2300 // gvn to   opt_test    |   |
2301 //          /    \      |  <1>
2302 //      True     False  |
2303 //        |         \\  |
2304 //   [slow_call]     \[fast_result]
2305 //    Ctl   Val       \      \
2306 //     |               \      \
2307 //    Catch       <1>   \      \
2308 //   /    \        ^     \      \
2309 //  Ex    No_Ex    |      \      \
2310 //  |       \   \  |       \ <2>  \
2311 //  ...      \  [slow_res] |  |    \   [null_result]
2312 //            \         \--+--+---  |  |
2313 //             \           | /    \ | /
2314 //              --------Region     Phi
2315 //
2316 //=============================================================================
2317 // Code is structured as a series of driver functions all called 'do_XXX' that
2318 // call a set of helper functions.  Helper functions first, then drivers.
2319 
2320 //------------------------------null_check_oop---------------------------------
2321 // Null check oop.  Set null-path control into Region in slot 3.
2322 // Make a cast-not-nullness use the other not-null control.  Return cast.
2323 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2324                                bool never_see_null, bool safe_for_replace) {
2325   // Initial NULL check taken path
2326   (*null_control) = top();
2327   Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2328 
2329   // Generate uncommon_trap:
2330   if (never_see_null && (*null_control) != top()) {
2331     // If we see an unexpected null at a check-cast we record it and force a
2332     // recompile; the offending check-cast will be compiled to handle NULLs.
2333     // If we see more than one offending BCI, then all checkcasts in the
2334     // method will be compiled to handle NULLs.
2335     PreserveJVMState pjvms(this);
2336     set_control(*null_control);
2337     replace_in_map(value, null());
2338     uncommon_trap(Deoptimization::Reason_null_check,
2339                   Deoptimization::Action_make_not_entrant);
2340     (*null_control) = top();    // NULL path is dead
2341   }
2342   if ((*null_control) == top() && safe_for_replace) {
2343     replace_in_map(value, cast);
2344   }
2345 
2346   // Cast away null-ness on the result
2347   return cast;
2348 }
2349 
2350 //------------------------------opt_iff----------------------------------------
2351 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2352 // Return slow-path control.
2353 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2354   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2355 
2356   // Fast path taken; set region slot 2
2357   Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2358   region->init_req(2,fast_taken); // Capture fast-control
2359 
2360   // Fast path not-taken, i.e. slow path
2361   Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2362   return slow_taken;
2363 }
2364 
2365 //-----------------------------make_runtime_call-------------------------------
2366 Node* GraphKit::make_runtime_call(int flags,
2367                                   const TypeFunc* call_type, address call_addr,
2368                                   const char* call_name,
2369                                   const TypePtr* adr_type,
2370                                   // The following parms are all optional.
2371                                   // The first NULL ends the list.
2372                                   Node* parm0, Node* parm1,
2373                                   Node* parm2, Node* parm3,
2374                                   Node* parm4, Node* parm5,
2375                                   Node* parm6, Node* parm7) {
2376   // Slow-path call
2377   bool is_leaf = !(flags & RC_NO_LEAF);
2378   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2379   if (call_name == NULL) {
2380     assert(!is_leaf, "must supply name for leaf");
2381     call_name = OptoRuntime::stub_name(call_addr);
2382   }
2383   CallNode* call;
2384   if (!is_leaf) {
2385     call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2386                                            bci(), adr_type);
2387   } else if (flags & RC_NO_FP) {
2388     call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2389   } else {
2390     call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2391   }
2392 
2393   // The following is similar to set_edges_for_java_call,
2394   // except that the memory effects of the call are restricted to AliasIdxRaw.
2395 
2396   // Slow path call has no side-effects, uses few values
2397   bool wide_in  = !(flags & RC_NARROW_MEM);
2398   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2399 
2400   Node* prev_mem = NULL;
2401   if (wide_in) {
2402     prev_mem = set_predefined_input_for_runtime_call(call);
2403   } else {
2404     assert(!wide_out, "narrow in => narrow out");
2405     Node* narrow_mem = memory(adr_type);
2406     prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2407   }
2408 
2409   // Hook each parm in order.  Stop looking at the first NULL.
2410   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2411   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2412   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2413   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2414   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2415   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2416   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2417   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2418     /* close each nested if ===> */  } } } } } } } }
2419   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2420 
2421   if (!is_leaf) {
2422     // Non-leaves can block and take safepoints:
2423     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2424   }
2425   // Non-leaves can throw exceptions:
2426   if (has_io) {
2427     call->set_req(TypeFunc::I_O, i_o());
2428   }
2429 
2430   if (flags & RC_UNCOMMON) {
2431     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2432     // (An "if" probability corresponds roughly to an unconditional count.
2433     // Sort of.)
2434     call->set_cnt(PROB_UNLIKELY_MAG(4));
2435   }
2436 
2437   Node* c = _gvn.transform(call);
2438   assert(c == call, "cannot disappear");
2439 
2440   if (wide_out) {
2441     // Slow path call has full side-effects.
2442     set_predefined_output_for_runtime_call(call);
2443   } else {
2444     // Slow path call has few side-effects, and/or sets few values.
2445     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2446   }
2447 
2448   if (has_io) {
2449     set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2450   }
2451   return call;
2452 
2453 }
2454 
2455 //------------------------------merge_memory-----------------------------------
2456 // Merge memory from one path into the current memory state.
2457 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2458   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2459     Node* old_slice = mms.force_memory();
2460     Node* new_slice = mms.memory2();
2461     if (old_slice != new_slice) {
2462       PhiNode* phi;
2463       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2464         if (mms.is_empty()) {
2465           // clone base memory Phi's inputs for this memory slice
2466           assert(old_slice == mms.base_memory(), "sanity");
2467           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2468           _gvn.set_type(phi, Type::MEMORY);
2469           for (uint i = 1; i < phi->req(); i++) {
2470             phi->init_req(i, old_slice->in(i));
2471           }
2472         } else {
2473           phi = old_slice->as_Phi(); // Phi was generated already
2474         }
2475       } else {
2476         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2477         _gvn.set_type(phi, Type::MEMORY);
2478       }
2479       phi->set_req(new_path, new_slice);
2480       mms.set_memory(phi);
2481     }
2482   }
2483 }
2484 
2485 //------------------------------make_slow_call_ex------------------------------
2486 // Make the exception handler hookups for the slow call
2487 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2488   if (stopped())  return;
2489 
2490   // Make a catch node with just two handlers:  fall-through and catch-all
2491   Node* i_o  = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2492   Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2493   Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2494   Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2495 
2496   { PreserveJVMState pjvms(this);
2497     set_control(excp);
2498     set_i_o(i_o);
2499 
2500     if (excp != top()) {
2501       if (deoptimize) {
2502         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2503         uncommon_trap(Deoptimization::Reason_unhandled,
2504                       Deoptimization::Action_none);
2505       } else {
2506         // Create an exception state also.
2507         // Use an exact type if the caller has specified a specific exception.
2508         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2509         Node*       ex_oop  = new (C) CreateExNode(ex_type, control(), i_o);
2510         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2511       }
2512     }
2513   }
2514 
2515   // Get the no-exception control from the CatchNode.
2516   set_control(norm);
2517 }
2518 
2519 
2520 //-------------------------------gen_subtype_check-----------------------------
2521 // Generate a subtyping check.  Takes as input the subtype and supertype.
2522 // Returns 2 values: sets the default control() to the true path and returns
2523 // the false path.  Only reads invariant memory; sets no (visible) memory.
2524 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2525 // but that's not exposed to the optimizer.  This call also doesn't take in an
2526 // Object; if you wish to check an Object you need to load the Object's class
2527 // prior to coming here.
2528 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2529   // Fast check for identical types, perhaps identical constants.
2530   // The types can even be identical non-constants, in cases
2531   // involving Array.newInstance, Object.clone, etc.
2532   if (subklass == superklass)
2533     return top();             // false path is dead; no test needed.
2534 
2535   if (_gvn.type(superklass)->singleton()) {
2536     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2537     ciKlass* subk   = _gvn.type(subklass)->is_klassptr()->klass();
2538 
2539     // In the common case of an exact superklass, try to fold up the
2540     // test before generating code.  You may ask, why not just generate
2541     // the code and then let it fold up?  The answer is that the generated
2542     // code will necessarily include null checks, which do not always
2543     // completely fold away.  If they are also needless, then they turn
2544     // into a performance loss.  Example:
2545     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2546     // Here, the type of 'fa' is often exact, so the store check
2547     // of fa[1]=x will fold up, without testing the nullness of x.
2548     switch (static_subtype_check(superk, subk)) {
2549     case SSC_always_false:
2550       {
2551         Node* always_fail = control();
2552         set_control(top());
2553         return always_fail;
2554       }
2555     case SSC_always_true:
2556       return top();
2557     case SSC_easy_test:
2558       {
2559         // Just do a direct pointer compare and be done.
2560         Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2561         Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2562         IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2563         set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2564         return       _gvn.transform( new(C) IfFalseNode(iff) );
2565       }
2566     case SSC_full_test:
2567       break;
2568     default:
2569       ShouldNotReachHere();
2570     }
2571   }
2572 
2573   // %%% Possible further optimization:  Even if the superklass is not exact,
2574   // if the subklass is the unique subtype of the superklass, the check
2575   // will always succeed.  We could leave a dependency behind to ensure this.
2576 
2577   // First load the super-klass's check-offset
2578   Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2579   Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2580                                                    TypeInt::INT, MemNode::unordered));
2581   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2582   bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2583 
2584   // Load from the sub-klass's super-class display list, or a 1-word cache of
2585   // the secondary superclass list, or a failing value with a sentinel offset
2586   // if the super-klass is an interface or exceptionally deep in the Java
2587   // hierarchy and we have to scan the secondary superclass list the hard way.
2588   // Worst-case type is a little odd: NULL is allowed as a result (usually
2589   // klass loads can never produce a NULL).
2590   Node *chk_off_X = ConvI2X(chk_off);
2591   Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2592   // For some types like interfaces the following loadKlass is from a 1-word
2593   // cache which is mutable so can't use immutable memory.  Other
2594   // types load from the super-class display table which is immutable.
2595   Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2596   Node* nkls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2597 
2598   // Compile speed common case: ARE a subtype and we canNOT fail
2599   if( superklass == nkls )
2600     return top();             // false path is dead; no test needed.
2601 
2602   // See if we get an immediate positive hit.  Happens roughly 83% of the
2603   // time.  Test to see if the value loaded just previously from the subklass
2604   // is exactly the superklass.
2605   Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2606   Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2607   IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2608   Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2609   set_control(    _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2610 
2611   // Compile speed common case: Check for being deterministic right now.  If
2612   // chk_off is a constant and not equal to cacheoff then we are NOT a
2613   // subklass.  In this case we need exactly the 1 test above and we can
2614   // return those results immediately.
2615   if (!might_be_cache) {
2616     Node* not_subtype_ctrl = control();
2617     set_control(iftrue1); // We need exactly the 1 test above
2618     return not_subtype_ctrl;
2619   }
2620 
2621   // Gather the various success & failures here
2622   RegionNode *r_ok_subtype = new (C) RegionNode(4);
2623   record_for_igvn(r_ok_subtype);
2624   RegionNode *r_not_subtype = new (C) RegionNode(3);
2625   record_for_igvn(r_not_subtype);
2626 
2627   r_ok_subtype->init_req(1, iftrue1);
2628 
2629   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2630   // is roughly 63% of the remaining cases).  Test to see if the loaded
2631   // check-offset points into the subklass display list or the 1-element
2632   // cache.  If it points to the display (and NOT the cache) and the display
2633   // missed then it's not a subtype.
2634   Node *cacheoff = _gvn.intcon(cacheoff_con);
2635   Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2636   Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2637   IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2638   r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2639   set_control(                _gvn.transform( new (C) IfFalseNode(iff2) ) );
2640 
2641   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2642   // No performance impact (too rare) but allows sharing of secondary arrays
2643   // which has some footprint reduction.
2644   Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2645   Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2646   IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2647   r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2648   set_control(               _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2649 
2650   // -- Roads not taken here: --
2651   // We could also have chosen to perform the self-check at the beginning
2652   // of this code sequence, as the assembler does.  This would not pay off
2653   // the same way, since the optimizer, unlike the assembler, can perform
2654   // static type analysis to fold away many successful self-checks.
2655   // Non-foldable self checks work better here in second position, because
2656   // the initial primary superclass check subsumes a self-check for most
2657   // types.  An exception would be a secondary type like array-of-interface,
2658   // which does not appear in its own primary supertype display.
2659   // Finally, we could have chosen to move the self-check into the
2660   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2661   // dependent manner.  But it is worthwhile to have the check here,
2662   // where it can be perhaps be optimized.  The cost in code space is
2663   // small (register compare, branch).
2664 
2665   // Now do a linear scan of the secondary super-klass array.  Again, no real
2666   // performance impact (too rare) but it's gotta be done.
2667   // Since the code is rarely used, there is no penalty for moving it
2668   // out of line, and it can only improve I-cache density.
2669   // The decision to inline or out-of-line this final check is platform
2670   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2671   Node* psc = _gvn.transform(
2672     new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2673 
2674   Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2675   Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2676   IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2677   r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2678   r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2679 
2680   // Return false path; set default control to true path.
2681   set_control( _gvn.transform(r_ok_subtype) );
2682   return _gvn.transform(r_not_subtype);
2683 }
2684 
2685 //----------------------------static_subtype_check-----------------------------
2686 // Shortcut important common cases when superklass is exact:
2687 // (0) superklass is java.lang.Object (can occur in reflective code)
2688 // (1) subklass is already limited to a subtype of superklass => always ok
2689 // (2) subklass does not overlap with superklass => always fail
2690 // (3) superklass has NO subtypes and we can check with a simple compare.
2691 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2692   if (StressReflectiveCode) {
2693     return SSC_full_test;       // Let caller generate the general case.
2694   }
2695 
2696   if (superk == env()->Object_klass()) {
2697     return SSC_always_true;     // (0) this test cannot fail
2698   }
2699 
2700   ciType* superelem = superk;
2701   if (superelem->is_array_klass())
2702     superelem = superelem->as_array_klass()->base_element_type();
2703 
2704   if (!subk->is_interface()) {  // cannot trust static interface types yet
2705     if (subk->is_subtype_of(superk)) {
2706       return SSC_always_true;   // (1) false path dead; no dynamic test needed
2707     }
2708     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2709         !superk->is_subtype_of(subk)) {
2710       return SSC_always_false;
2711     }
2712   }
2713 
2714   // If casting to an instance klass, it must have no subtypes
2715   if (superk->is_interface()) {
2716     // Cannot trust interfaces yet.
2717     // %%% S.B. superk->nof_implementors() == 1
2718   } else if (superelem->is_instance_klass()) {
2719     ciInstanceKlass* ik = superelem->as_instance_klass();
2720     if (!ik->has_subklass() && !ik->is_interface()) {
2721       if (!ik->is_final()) {
2722         // Add a dependency if there is a chance of a later subclass.
2723         C->dependencies()->assert_leaf_type(ik);
2724       }
2725       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
2726     }
2727   } else {
2728     // A primitive array type has no subtypes.
2729     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
2730   }
2731 
2732   return SSC_full_test;
2733 }
2734 
2735 // Profile-driven exact type check:
2736 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2737                                     float prob,
2738                                     Node* *casted_receiver) {
2739   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2740   Node* recv_klass = load_object_klass(receiver);
2741   Node* want_klass = makecon(tklass);
2742   Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2743   Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2744   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2745   set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2746   Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2747 
2748   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2749   assert(recv_xtype->klass_is_exact(), "");
2750 
2751   // Subsume downstream occurrences of receiver with a cast to
2752   // recv_xtype, since now we know what the type will be.
2753   Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2754   (*casted_receiver) = _gvn.transform(cast);
2755   // (User must make the replace_in_map call.)
2756 
2757   return fail;
2758 }
2759 
2760 
2761 //------------------------------seems_never_null-------------------------------
2762 // Use null_seen information if it is available from the profile.
2763 // If we see an unexpected null at a type check we record it and force a
2764 // recompile; the offending check will be recompiled to handle NULLs.
2765 // If we see several offending BCIs, then all checks in the
2766 // method will be recompiled.
2767 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2768   if (UncommonNullCast               // Cutout for this technique
2769       && obj != null()               // And not the -Xcomp stupid case?
2770       && !too_many_traps(Deoptimization::Reason_null_check)
2771       ) {
2772     if (data == NULL)
2773       // Edge case:  no mature data.  Be optimistic here.
2774       return true;
2775     // If the profile has not seen a null, assume it won't happen.
2776     assert(java_bc() == Bytecodes::_checkcast ||
2777            java_bc() == Bytecodes::_instanceof ||
2778            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2779     return !data->as_BitData()->null_seen();
2780   }
2781   return false;
2782 }
2783 
2784 //------------------------maybe_cast_profiled_receiver-------------------------
2785 // If the profile has seen exactly one type, narrow to exactly that type.
2786 // Subsequent type checks will always fold up.
2787 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2788                                              ciKlass* require_klass,
2789                                              ciKlass* spec_klass,
2790                                              bool safe_for_replace) {
2791   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2792 
2793   Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2794 
2795   // Make sure we haven't already deoptimized from this tactic.
2796   if (too_many_traps(reason) || too_many_recompiles(reason))
2797     return NULL;
2798 
2799   // (No, this isn't a call, but it's enough like a virtual call
2800   // to use the same ciMethod accessor to get the profile info...)
2801   // If we have a speculative type use it instead of profiling (which
2802   // may not help us)
2803   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2804   if (exact_kls != NULL) {// no cast failures here
2805     if (require_klass == NULL ||
2806         static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2807       // If we narrow the type to match what the type profile sees or
2808       // the speculative type, we can then remove the rest of the
2809       // cast.
2810       // This is a win, even if the exact_kls is very specific,
2811       // because downstream operations, such as method calls,
2812       // will often benefit from the sharper type.
2813       Node* exact_obj = not_null_obj; // will get updated in place...
2814       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2815                                             &exact_obj);
2816       { PreserveJVMState pjvms(this);
2817         set_control(slow_ctl);
2818         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2819       }
2820       if (safe_for_replace) {
2821         replace_in_map(not_null_obj, exact_obj);
2822       }
2823       return exact_obj;
2824     }
2825     // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2826   }
2827 
2828   return NULL;
2829 }
2830 
2831 /**
2832  * Cast obj to type and emit guard unless we had too many traps here
2833  * already
2834  *
2835  * @param obj       node being casted
2836  * @param type      type to cast the node to
2837  * @param not_null  true if we know node cannot be null
2838  */
2839 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2840                                         ciKlass* type,
2841                                         bool not_null) {
2842   // type == NULL if profiling tells us this object is always null
2843   if (type != NULL) {
2844     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2845     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2846     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2847         !too_many_traps(class_reason) && !too_many_recompiles(class_reason)) {
2848       Node* not_null_obj = NULL;
2849       // not_null is true if we know the object is not null and
2850       // there's no need for a null check
2851       if (!not_null) {
2852         Node* null_ctl = top();
2853         not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2854         assert(null_ctl->is_top(), "no null control here");
2855       } else {
2856         not_null_obj = obj;
2857       }
2858 
2859       Node* exact_obj = not_null_obj;
2860       ciKlass* exact_kls = type;
2861       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2862                                             &exact_obj);
2863       {
2864         PreserveJVMState pjvms(this);
2865         set_control(slow_ctl);
2866         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2867       }
2868       replace_in_map(not_null_obj, exact_obj);
2869       obj = exact_obj;
2870     }
2871   } else {
2872     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2873         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2874       Node* exact_obj = null_assert(obj);
2875       replace_in_map(obj, exact_obj);
2876       obj = exact_obj;
2877     }
2878   }
2879   return obj;
2880 }
2881 
2882 //-------------------------------gen_instanceof--------------------------------
2883 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2884 // and the reflective instance-of call.
2885 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2886   kill_dead_locals();           // Benefit all the uncommon traps
2887   assert( !stopped(), "dead parse path should be checked in callers" );
2888   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2889          "must check for not-null not-dead klass in callers");
2890 
2891   // Make the merge point
2892   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2893   RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2894   Node*       phi    = new(C) PhiNode(region, TypeInt::BOOL);
2895   C->set_has_split_ifs(true); // Has chance for split-if optimization
2896 
2897   ciProfileData* data = NULL;
2898   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2899     data = method()->method_data()->bci_to_data(bci());
2900   }
2901   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2902                          && seems_never_null(obj, data));
2903 
2904   // Null check; get casted pointer; set region slot 3
2905   Node* null_ctl = top();
2906   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2907 
2908   // If not_null_obj is dead, only null-path is taken
2909   if (stopped()) {              // Doing instance-of on a NULL?
2910     set_control(null_ctl);
2911     return intcon(0);
2912   }
2913   region->init_req(_null_path, null_ctl);
2914   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2915   if (null_ctl == top()) {
2916     // Do this eagerly, so that pattern matches like is_diamond_phi
2917     // will work even during parsing.
2918     assert(_null_path == PATH_LIMIT-1, "delete last");
2919     region->del_req(_null_path);
2920     phi   ->del_req(_null_path);
2921   }
2922 
2923   // Do we know the type check always succeed?
2924   bool known_statically = false;
2925   if (_gvn.type(superklass)->singleton()) {
2926     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2927     ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2928     if (subk != NULL && subk->is_loaded()) {
2929       int static_res = static_subtype_check(superk, subk);
2930       known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2931     }
2932   }
2933 
2934   if (known_statically && UseTypeSpeculation) {
2935     // If we know the type check always succeeds then we don't use the
2936     // profiling data at this bytecode. Don't lose it, feed it to the
2937     // type system as a speculative type.
2938     not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2939   } else {
2940     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2941     // We may not have profiling here or it may not help us. If we
2942     // have a speculative type use it to perform an exact cast.
2943     ciKlass* spec_obj_type = obj_type->speculative_type();
2944     if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2945       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2946       if (stopped()) {            // Profile disagrees with this path.
2947         set_control(null_ctl);    // Null is the only remaining possibility.
2948         return intcon(0);
2949       }
2950       if (cast_obj != NULL) {
2951         not_null_obj = cast_obj;
2952       }
2953     }
2954   }
2955 
2956   // Load the object's klass
2957   Node* obj_klass = load_object_klass(not_null_obj);
2958 
2959   // Generate the subtype check
2960   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2961 
2962   // Plug in the success path to the general merge in slot 1.
2963   region->init_req(_obj_path, control());
2964   phi   ->init_req(_obj_path, intcon(1));
2965 
2966   // Plug in the failing path to the general merge in slot 2.
2967   region->init_req(_fail_path, not_subtype_ctrl);
2968   phi   ->init_req(_fail_path, intcon(0));
2969 
2970   // Return final merged results
2971   set_control( _gvn.transform(region) );
2972   record_for_igvn(region);
2973   return _gvn.transform(phi);
2974 }
2975 
2976 //-------------------------------gen_checkcast---------------------------------
2977 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
2978 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
2979 // uncommon-trap paths work.  Adjust stack after this call.
2980 // If failure_control is supplied and not null, it is filled in with
2981 // the control edge for the cast failure.  Otherwise, an appropriate
2982 // uncommon trap or exception is thrown.
2983 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2984                               Node* *failure_control) {
2985   kill_dead_locals();           // Benefit all the uncommon traps
2986   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2987   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2988 
2989   // Fast cutout:  Check the case that the cast is vacuously true.
2990   // This detects the common cases where the test will short-circuit
2991   // away completely.  We do this before we perform the null check,
2992   // because if the test is going to turn into zero code, we don't
2993   // want a residual null check left around.  (Causes a slowdown,
2994   // for example, in some objArray manipulations, such as a[i]=a[j].)
2995   if (tk->singleton()) {
2996     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2997     if (objtp != NULL && objtp->klass() != NULL) {
2998       switch (static_subtype_check(tk->klass(), objtp->klass())) {
2999       case SSC_always_true:
3000         // If we know the type check always succeed then we don't use
3001         // the profiling data at this bytecode. Don't lose it, feed it
3002         // to the type system as a speculative type.
3003         return record_profiled_receiver_for_speculation(obj);
3004       case SSC_always_false:
3005         // It needs a null check because a null will *pass* the cast check.
3006         // A non-null value will always produce an exception.
3007         return null_assert(obj);
3008       }
3009     }
3010   }
3011 
3012   ciProfileData* data = NULL;
3013   bool safe_for_replace = false;
3014   if (failure_control == NULL) {        // use MDO in regular case only
3015     assert(java_bc() == Bytecodes::_aastore ||
3016            java_bc() == Bytecodes::_checkcast,
3017            "interpreter profiles type checks only for these BCs");
3018     data = method()->method_data()->bci_to_data(bci());
3019     safe_for_replace = true;
3020   }
3021 
3022   // Make the merge point
3023   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3024   RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3025   Node*       phi    = new (C) PhiNode(region, toop);
3026   C->set_has_split_ifs(true); // Has chance for split-if optimization
3027 
3028   // Use null-cast information if it is available
3029   bool never_see_null = ((failure_control == NULL)  // regular case only
3030                          && seems_never_null(obj, data));
3031 
3032   // Null check; get casted pointer; set region slot 3
3033   Node* null_ctl = top();
3034   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
3035 
3036   // If not_null_obj is dead, only null-path is taken
3037   if (stopped()) {              // Doing instance-of on a NULL?
3038     set_control(null_ctl);
3039     return null();
3040   }
3041   region->init_req(_null_path, null_ctl);
3042   phi   ->init_req(_null_path, null());  // Set null path value
3043   if (null_ctl == top()) {
3044     // Do this eagerly, so that pattern matches like is_diamond_phi
3045     // will work even during parsing.
3046     assert(_null_path == PATH_LIMIT-1, "delete last");
3047     region->del_req(_null_path);
3048     phi   ->del_req(_null_path);
3049   }
3050 
3051   Node* cast_obj = NULL;
3052   if (tk->klass_is_exact()) {
3053     // The following optimization tries to statically cast the speculative type of the object
3054     // (for example obtained during profiling) to the type of the superklass and then do a
3055     // dynamic check that the type of the object is what we expect. To work correctly
3056     // for checkcast and aastore the type of superklass should be exact.
3057     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3058     // We may not have profiling here or it may not help us. If we have
3059     // a speculative type use it to perform an exact cast.
3060     ciKlass* spec_obj_type = obj_type->speculative_type();
3061     if (spec_obj_type != NULL ||
3062         (data != NULL &&
3063          // Counter has never been decremented (due to cast failure).
3064          // ...This is a reasonable thing to expect.  It is true of
3065          // all casts inserted by javac to implement generic types.
3066          data->as_CounterData()->count() >= 0)) {
3067       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3068       if (cast_obj != NULL) {
3069         if (failure_control != NULL) // failure is now impossible
3070           (*failure_control) = top();
3071         // adjust the type of the phi to the exact klass:
3072         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3073       }
3074     }
3075   }
3076 
3077   if (cast_obj == NULL) {
3078     // Load the object's klass
3079     Node* obj_klass = load_object_klass(not_null_obj);
3080 
3081     // Generate the subtype check
3082     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3083 
3084     // Plug in success path into the merge
3085     cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3086                                                          not_null_obj, toop));
3087     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3088     if (failure_control == NULL) {
3089       if (not_subtype_ctrl != top()) { // If failure is possible
3090         PreserveJVMState pjvms(this);
3091         set_control(not_subtype_ctrl);
3092         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3093       }
3094     } else {
3095       (*failure_control) = not_subtype_ctrl;
3096     }
3097   }
3098 
3099   region->init_req(_obj_path, control());
3100   phi   ->init_req(_obj_path, cast_obj);
3101 
3102   // A merge of NULL or Casted-NotNull obj
3103   Node* res = _gvn.transform(phi);
3104 
3105   // Note I do NOT always 'replace_in_map(obj,result)' here.
3106   //  if( tk->klass()->can_be_primary_super()  )
3107     // This means that if I successfully store an Object into an array-of-String
3108     // I 'forget' that the Object is really now known to be a String.  I have to
3109     // do this because we don't have true union types for interfaces - if I store
3110     // a Baz into an array-of-Interface and then tell the optimizer it's an
3111     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3112     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3113   //  replace_in_map( obj, res );
3114 
3115   // Return final merged results
3116   set_control( _gvn.transform(region) );
3117   record_for_igvn(region);
3118   return res;
3119 }
3120 
3121 //------------------------------next_monitor-----------------------------------
3122 // What number should be given to the next monitor?
3123 int GraphKit::next_monitor() {
3124   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3125   int next = current + C->sync_stack_slots();
3126   // Keep the toplevel high water mark current:
3127   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3128   return current;
3129 }
3130 
3131 //------------------------------insert_mem_bar---------------------------------
3132 // Memory barrier to avoid floating things around
3133 // The membar serves as a pinch point between both control and all memory slices.
3134 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3135   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3136   mb->init_req(TypeFunc::Control, control());
3137   mb->init_req(TypeFunc::Memory,  reset_memory());
3138   Node* membar = _gvn.transform(mb);
3139   set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3140   set_all_memory_call(membar);
3141   return membar;
3142 }
3143 
3144 //-------------------------insert_mem_bar_volatile----------------------------
3145 // Memory barrier to avoid floating things around
3146 // The membar serves as a pinch point between both control and memory(alias_idx).
3147 // If you want to make a pinch point on all memory slices, do not use this
3148 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3149 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3150   // When Parse::do_put_xxx updates a volatile field, it appends a series
3151   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3152   // The first membar is on the same memory slice as the field store opcode.
3153   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3154   // All the other membars (for other volatile slices, including AliasIdxBot,
3155   // which stands for all unknown volatile slices) are control-dependent
3156   // on the first membar.  This prevents later volatile loads or stores
3157   // from sliding up past the just-emitted store.
3158 
3159   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3160   mb->set_req(TypeFunc::Control,control());
3161   if (alias_idx == Compile::AliasIdxBot) {
3162     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3163   } else {
3164     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3165     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3166   }
3167   Node* membar = _gvn.transform(mb);
3168   set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3169   if (alias_idx == Compile::AliasIdxBot) {
3170     merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3171   } else {
3172     set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3173   }
3174   return membar;
3175 }
3176 
3177 //------------------------------shared_lock------------------------------------
3178 // Emit locking code.
3179 FastLockNode* GraphKit::shared_lock(Node* obj) {
3180   // bci is either a monitorenter bc or InvocationEntryBci
3181   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3182   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3183 
3184   if( !GenerateSynchronizationCode )
3185     return NULL;                // Not locking things?
3186   if (stopped())                // Dead monitor?
3187     return NULL;
3188 
3189   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3190 
3191   // Box the stack location
3192   Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3193   Node* mem = reset_memory();
3194 
3195   FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3196   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3197     // Create the counters for this fast lock.
3198     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3199   }
3200 
3201   // Create the rtm counters for this fast lock if needed.
3202   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3203 
3204   // Add monitor to debug info for the slow path.  If we block inside the
3205   // slow path and de-opt, we need the monitor hanging around
3206   map()->push_monitor( flock );
3207 
3208   const TypeFunc *tf = LockNode::lock_type();
3209   LockNode *lock = new (C) LockNode(C, tf);
3210 
3211   lock->init_req( TypeFunc::Control, control() );
3212   lock->init_req( TypeFunc::Memory , mem );
3213   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3214   lock->init_req( TypeFunc::FramePtr, frameptr() );
3215   lock->init_req( TypeFunc::ReturnAdr, top() );
3216 
3217   lock->init_req(TypeFunc::Parms + 0, obj);
3218   lock->init_req(TypeFunc::Parms + 1, box);
3219   lock->init_req(TypeFunc::Parms + 2, flock);
3220   add_safepoint_edges(lock);
3221 
3222   lock = _gvn.transform( lock )->as_Lock();
3223 
3224   // lock has no side-effects, sets few values
3225   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3226 
3227   insert_mem_bar(Op_MemBarAcquireLock);
3228 
3229   // Add this to the worklist so that the lock can be eliminated
3230   record_for_igvn(lock);
3231 
3232 #ifndef PRODUCT
3233   if (PrintLockStatistics) {
3234     // Update the counter for this lock.  Don't bother using an atomic
3235     // operation since we don't require absolute accuracy.
3236     lock->create_lock_counter(map()->jvms());
3237     increment_counter(lock->counter()->addr());
3238   }
3239 #endif
3240 
3241   return flock;
3242 }
3243 
3244 
3245 //------------------------------shared_unlock----------------------------------
3246 // Emit unlocking code.
3247 void GraphKit::shared_unlock(Node* box, Node* obj) {
3248   // bci is either a monitorenter bc or InvocationEntryBci
3249   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3250   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3251 
3252   if( !GenerateSynchronizationCode )
3253     return;
3254   if (stopped()) {               // Dead monitor?
3255     map()->pop_monitor();        // Kill monitor from debug info
3256     return;
3257   }
3258 
3259   // Memory barrier to avoid floating things down past the locked region
3260   insert_mem_bar(Op_MemBarReleaseLock);
3261 
3262   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3263   UnlockNode *unlock = new (C) UnlockNode(C, tf);
3264 #ifdef ASSERT
3265   unlock->set_dbg_jvms(sync_jvms());
3266 #endif
3267   uint raw_idx = Compile::AliasIdxRaw;
3268   unlock->init_req( TypeFunc::Control, control() );
3269   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3270   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3271   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3272   unlock->init_req( TypeFunc::ReturnAdr, top() );
3273 
3274   unlock->init_req(TypeFunc::Parms + 0, obj);
3275   unlock->init_req(TypeFunc::Parms + 1, box);
3276   unlock = _gvn.transform(unlock)->as_Unlock();
3277 
3278   Node* mem = reset_memory();
3279 
3280   // unlock has no side-effects, sets few values
3281   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3282 
3283   // Kill monitor from debug info
3284   map()->pop_monitor( );
3285 }
3286 
3287 //-------------------------------get_layout_helper-----------------------------
3288 // If the given klass is a constant or known to be an array,
3289 // fetch the constant layout helper value into constant_value
3290 // and return (Node*)NULL.  Otherwise, load the non-constant
3291 // layout helper value, and return the node which represents it.
3292 // This two-faced routine is useful because allocation sites
3293 // almost always feature constant types.
3294 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3295   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3296   if (!StressReflectiveCode && inst_klass != NULL) {
3297     ciKlass* klass = inst_klass->klass();
3298     bool    xklass = inst_klass->klass_is_exact();
3299     if (xklass || klass->is_array_klass()) {
3300       jint lhelper = klass->layout_helper();
3301       if (lhelper != Klass::_lh_neutral_value) {
3302         constant_value = lhelper;
3303         return (Node*) NULL;
3304       }
3305     }
3306   }
3307   constant_value = Klass::_lh_neutral_value;  // put in a known value
3308   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3309   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3310 }
3311 
3312 // We just put in an allocate/initialize with a big raw-memory effect.
3313 // Hook selected additional alias categories on the initialization.
3314 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3315                                 MergeMemNode* init_in_merge,
3316                                 Node* init_out_raw) {
3317   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3318   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3319 
3320   Node* prevmem = kit.memory(alias_idx);
3321   init_in_merge->set_memory_at(alias_idx, prevmem);
3322   kit.set_memory(init_out_raw, alias_idx);
3323 }
3324 
3325 //---------------------------set_output_for_allocation-------------------------
3326 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3327                                           const TypeOopPtr* oop_type,
3328                                           bool deoptimize_on_exception) {
3329   int rawidx = Compile::AliasIdxRaw;
3330   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3331   add_safepoint_edges(alloc);
3332   Node* allocx = _gvn.transform(alloc);
3333   set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3334   // create memory projection for i_o
3335   set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3336   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3337 
3338   // create a memory projection as for the normal control path
3339   Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3340   set_memory(malloc, rawidx);
3341 
3342   // a normal slow-call doesn't change i_o, but an allocation does
3343   // we create a separate i_o projection for the normal control path
3344   set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3345   Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3346 
3347   // put in an initialization barrier
3348   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3349                                                  rawoop)->as_Initialize();
3350   assert(alloc->initialization() == init,  "2-way macro link must work");
3351   assert(init ->allocation()     == alloc, "2-way macro link must work");
3352   {
3353     // Extract memory strands which may participate in the new object's
3354     // initialization, and source them from the new InitializeNode.
3355     // This will allow us to observe initializations when they occur,
3356     // and link them properly (as a group) to the InitializeNode.
3357     assert(init->in(InitializeNode::Memory) == malloc, "");
3358     MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3359     init->set_req(InitializeNode::Memory, minit_in);
3360     record_for_igvn(minit_in); // fold it up later, if possible
3361     Node* minit_out = memory(rawidx);
3362     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3363     if (oop_type->isa_aryptr()) {
3364       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3365       int            elemidx  = C->get_alias_index(telemref);
3366       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3367     } else if (oop_type->isa_instptr()) {
3368       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3369       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3370         ciField* field = ik->nonstatic_field_at(i);
3371         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3372           continue;  // do not bother to track really large numbers of fields
3373         // Find (or create) the alias category for this field:
3374         int fieldidx = C->alias_type(field)->index();
3375         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3376       }
3377     }
3378   }
3379 
3380   // Cast raw oop to the real thing...
3381   Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3382   javaoop = _gvn.transform(javaoop);
3383   C->set_recent_alloc(control(), javaoop);
3384   assert(just_allocated_object(control()) == javaoop, "just allocated");
3385 
3386 #ifdef ASSERT
3387   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3388     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3389            "Ideal_allocation works");
3390     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3391            "Ideal_allocation works");
3392     if (alloc->is_AllocateArray()) {
3393       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3394              "Ideal_allocation works");
3395       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3396              "Ideal_allocation works");
3397     } else {
3398       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3399     }
3400   }
3401 #endif //ASSERT
3402 
3403   return javaoop;
3404 }
3405 
3406 //---------------------------new_instance--------------------------------------
3407 // This routine takes a klass_node which may be constant (for a static type)
3408 // or may be non-constant (for reflective code).  It will work equally well
3409 // for either, and the graph will fold nicely if the optimizer later reduces
3410 // the type to a constant.
3411 // The optional arguments are for specialized use by intrinsics:
3412 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3413 //  - If 'return_size_val', report the the total object size to the caller.
3414 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3415 Node* GraphKit::new_instance(Node* klass_node,
3416                              Node* extra_slow_test,
3417                              Node* *return_size_val,
3418                              bool deoptimize_on_exception) {
3419   // Compute size in doublewords
3420   // The size is always an integral number of doublewords, represented
3421   // as a positive bytewise size stored in the klass's layout_helper.
3422   // The layout_helper also encodes (in a low bit) the need for a slow path.
3423   jint  layout_con = Klass::_lh_neutral_value;
3424   Node* layout_val = get_layout_helper(klass_node, layout_con);
3425   int   layout_is_con = (layout_val == NULL);
3426 
3427   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3428   // Generate the initial go-slow test.  It's either ALWAYS (return a
3429   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3430   // case) a computed value derived from the layout_helper.
3431   Node* initial_slow_test = NULL;
3432   if (layout_is_con) {
3433     assert(!StressReflectiveCode, "stress mode does not use these paths");
3434     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3435     initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3436 
3437   } else {   // reflective case
3438     // This reflective path is used by Unsafe.allocateInstance.
3439     // (It may be stress-tested by specifying StressReflectiveCode.)
3440     // Basically, we want to get into the VM is there's an illegal argument.
3441     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3442     initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3443     if (extra_slow_test != intcon(0)) {
3444       initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3445     }
3446     // (Macro-expander will further convert this to a Bool, if necessary.)
3447   }
3448 
3449   // Find the size in bytes.  This is easy; it's the layout_helper.
3450   // The size value must be valid even if the slow path is taken.
3451   Node* size = NULL;
3452   if (layout_is_con) {
3453     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3454   } else {   // reflective case
3455     // This reflective path is used by clone and Unsafe.allocateInstance.
3456     size = ConvI2X(layout_val);
3457 
3458     // Clear the low bits to extract layout_helper_size_in_bytes:
3459     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3460     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3461     size = _gvn.transform( new (C) AndXNode(size, mask) );
3462   }
3463   if (return_size_val != NULL) {
3464     (*return_size_val) = size;
3465   }
3466 
3467   // This is a precise notnull oop of the klass.
3468   // (Actually, it need not be precise if this is a reflective allocation.)
3469   // It's what we cast the result to.
3470   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3471   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3472   const TypeOopPtr* oop_type = tklass->as_instance_type();
3473 
3474   // Now generate allocation code
3475 
3476   // The entire memory state is needed for slow path of the allocation
3477   // since GC and deoptimization can happened.
3478   Node *mem = reset_memory();
3479   set_all_memory(mem); // Create new memory state
3480 
3481   AllocateNode* alloc
3482     = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3483                            control(), mem, i_o(),
3484                            size, klass_node,
3485                            initial_slow_test);
3486 
3487   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3488 }
3489 
3490 //-------------------------------new_array-------------------------------------
3491 // helper for both newarray and anewarray
3492 // The 'length' parameter is (obviously) the length of the array.
3493 // See comments on new_instance for the meaning of the other arguments.
3494 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3495                           Node* length,         // number of array elements
3496                           int   nargs,          // number of arguments to push back for uncommon trap
3497                           Node* *return_size_val,
3498                           bool deoptimize_on_exception) {
3499   jint  layout_con = Klass::_lh_neutral_value;
3500   Node* layout_val = get_layout_helper(klass_node, layout_con);
3501   int   layout_is_con = (layout_val == NULL);
3502 
3503   if (!layout_is_con && !StressReflectiveCode &&
3504       !too_many_traps(Deoptimization::Reason_class_check)) {
3505     // This is a reflective array creation site.
3506     // Optimistically assume that it is a subtype of Object[],
3507     // so that we can fold up all the address arithmetic.
3508     layout_con = Klass::array_layout_helper(T_OBJECT);
3509     Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3510     Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3511     { BuildCutout unless(this, bol_lh, PROB_MAX);
3512       inc_sp(nargs);
3513       uncommon_trap(Deoptimization::Reason_class_check,
3514                     Deoptimization::Action_maybe_recompile);
3515     }
3516     layout_val = NULL;
3517     layout_is_con = true;
3518   }
3519 
3520   // Generate the initial go-slow test.  Make sure we do not overflow
3521   // if length is huge (near 2Gig) or negative!  We do not need
3522   // exact double-words here, just a close approximation of needed
3523   // double-words.  We can't add any offset or rounding bits, lest we
3524   // take a size -1 of bytes and make it positive.  Use an unsigned
3525   // compare, so negative sizes look hugely positive.
3526   int fast_size_limit = FastAllocateSizeLimit;
3527   if (layout_is_con) {
3528     assert(!StressReflectiveCode, "stress mode does not use these paths");
3529     // Increase the size limit if we have exact knowledge of array type.
3530     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3531     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3532   }
3533 
3534   Node* initial_slow_cmp  = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3535   Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3536 
3537   // --- Size Computation ---
3538   // array_size = round_to_heap(array_header + (length << elem_shift));
3539   // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3540   // and round_to(x, y) == ((x + y-1) & ~(y-1))
3541   // The rounding mask is strength-reduced, if possible.
3542   int round_mask = MinObjAlignmentInBytes - 1;
3543   Node* header_size = NULL;
3544   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3545   // (T_BYTE has the weakest alignment and size restrictions...)
3546   if (layout_is_con) {
3547     int       hsize  = Klass::layout_helper_header_size(layout_con);
3548     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3549     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3550     if ((round_mask & ~right_n_bits(eshift)) == 0)
3551       round_mask = 0;  // strength-reduce it if it goes away completely
3552     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3553     assert(header_size_min <= hsize, "generic minimum is smallest");
3554     header_size_min = hsize;
3555     header_size = intcon(hsize + round_mask);
3556   } else {
3557     Node* hss   = intcon(Klass::_lh_header_size_shift);
3558     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3559     Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3560     hsize       = _gvn.transform( new(C) AndINode(hsize, hsm) );
3561     Node* mask  = intcon(round_mask);
3562     header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3563   }
3564 
3565   Node* elem_shift = NULL;
3566   if (layout_is_con) {
3567     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3568     if (eshift != 0)
3569       elem_shift = intcon(eshift);
3570   } else {
3571     // There is no need to mask or shift this value.
3572     // The semantics of LShiftINode include an implicit mask to 0x1F.
3573     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3574     elem_shift = layout_val;
3575   }
3576 
3577   // Transition to native address size for all offset calculations:
3578   Node* lengthx = ConvI2X(length);
3579   Node* headerx = ConvI2X(header_size);
3580 #ifdef _LP64
3581   { const TypeInt* tilen = _gvn.find_int_type(length);
3582     if (tilen != NULL && tilen->_lo < 0) {
3583       // Add a manual constraint to a positive range.  Cf. array_element_address.
3584       jlong size_max = fast_size_limit;
3585       if (size_max > tilen->_hi)  size_max = tilen->_hi;
3586       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3587 
3588       // Only do a narrow I2L conversion if the range check passed.
3589       IfNode* iff = new (C) IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3590       _gvn.transform(iff);
3591       RegionNode* region = new (C) RegionNode(3);
3592       _gvn.set_type(region, Type::CONTROL);
3593       lengthx = new (C) PhiNode(region, TypeLong::LONG);
3594       _gvn.set_type(lengthx, TypeLong::LONG);
3595 
3596       // Range check passed. Use ConvI2L node with narrow type.
3597       Node* passed = IfFalse(iff);
3598       region->init_req(1, passed);
3599       // Make I2L conversion control dependent to prevent it from
3600       // floating above the range check during loop optimizations.
3601       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3602 
3603       // Range check failed. Use ConvI2L with wide type because length may be invalid.
3604       region->init_req(2, IfTrue(iff));
3605       lengthx->init_req(2, ConvI2X(length));
3606 
3607       set_control(region);
3608       record_for_igvn(region);
3609       record_for_igvn(lengthx);
3610     }
3611   }
3612 #endif
3613 
3614   // Combine header size (plus rounding) and body size.  Then round down.
3615   // This computation cannot overflow, because it is used only in two
3616   // places, one where the length is sharply limited, and the other
3617   // after a successful allocation.
3618   Node* abody = lengthx;
3619   if (elem_shift != NULL)
3620     abody     = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3621   Node* size  = _gvn.transform( new(C) AddXNode(headerx, abody) );
3622   if (round_mask != 0) {
3623     Node* mask = MakeConX(~round_mask);
3624     size       = _gvn.transform( new(C) AndXNode(size, mask) );
3625   }
3626   // else if round_mask == 0, the size computation is self-rounding
3627 
3628   if (return_size_val != NULL) {
3629     // This is the size
3630     (*return_size_val) = size;
3631   }
3632 
3633   // Now generate allocation code
3634 
3635   // The entire memory state is needed for slow path of the allocation
3636   // since GC and deoptimization can happened.
3637   Node *mem = reset_memory();
3638   set_all_memory(mem); // Create new memory state
3639 
3640   if (initial_slow_test->is_Bool()) {
3641     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3642     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3643   }
3644 
3645   // Create the AllocateArrayNode and its result projections
3646   AllocateArrayNode* alloc
3647     = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3648                                 control(), mem, i_o(),
3649                                 size, klass_node,
3650                                 initial_slow_test,
3651                                 length);
3652 
3653   // Cast to correct type.  Note that the klass_node may be constant or not,
3654   // and in the latter case the actual array type will be inexact also.
3655   // (This happens via a non-constant argument to inline_native_newArray.)
3656   // In any case, the value of klass_node provides the desired array type.
3657   const TypeInt* length_type = _gvn.find_int_type(length);
3658   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3659   if (ary_type->isa_aryptr() && length_type != NULL) {
3660     // Try to get a better type than POS for the size
3661     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3662   }
3663 
3664   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3665 
3666   // Cast length on remaining path to be as narrow as possible
3667   if (map()->find_edge(length) >= 0) {
3668     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3669     if (ccast != length) {
3670       _gvn.set_type_bottom(ccast);
3671       record_for_igvn(ccast);
3672       replace_in_map(length, ccast);
3673     }
3674   }
3675 
3676   return javaoop;
3677 }
3678 
3679 // The following "Ideal_foo" functions are placed here because they recognize
3680 // the graph shapes created by the functions immediately above.
3681 
3682 //---------------------------Ideal_allocation----------------------------------
3683 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3684 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3685   if (ptr == NULL) {     // reduce dumb test in callers
3686     return NULL;
3687   }
3688 
3689 #if INCLUDE_ALL_GCS
3690   if (UseShenandoahGC) {
3691     ptr = ShenandoahBarrierSetC2::bsc2()->step_over_gc_barrier(ptr);
3692   }
3693 #endif
3694   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3695     ptr = ptr->in(1);
3696     if (ptr == NULL) return NULL;
3697   }
3698   // Return NULL for allocations with several casts:
3699   //   j.l.reflect.Array.newInstance(jobject, jint)
3700   //   Object.clone()
3701   // to keep more precise type from last cast.
3702   if (ptr->is_Proj()) {
3703     Node* allo = ptr->in(0);
3704     if (allo != NULL && allo->is_Allocate()) {
3705       return allo->as_Allocate();
3706     }
3707   }
3708   // Report failure to match.
3709   return NULL;
3710 }
3711 
3712 // Fancy version which also strips off an offset (and reports it to caller).
3713 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3714                                              intptr_t& offset) {
3715   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3716   if (base == NULL)  return NULL;
3717   return Ideal_allocation(base, phase);
3718 }
3719 
3720 // Trace Initialize <- Proj[Parm] <- Allocate
3721 AllocateNode* InitializeNode::allocation() {
3722   Node* rawoop = in(InitializeNode::RawAddress);
3723   if (rawoop->is_Proj()) {
3724     Node* alloc = rawoop->in(0);
3725     if (alloc->is_Allocate()) {
3726       return alloc->as_Allocate();
3727     }
3728   }
3729   return NULL;
3730 }
3731 
3732 // Trace Allocate -> Proj[Parm] -> Initialize
3733 InitializeNode* AllocateNode::initialization() {
3734   ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3735   if (rawoop == NULL)  return NULL;
3736   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3737     Node* init = rawoop->fast_out(i);
3738     if (init->is_Initialize()) {
3739       assert(init->as_Initialize()->allocation() == this, "2-way link");
3740       return init->as_Initialize();
3741     }
3742   }
3743   return NULL;
3744 }
3745 
3746 //----------------------------- loop predicates ---------------------------
3747 
3748 //------------------------------add_predicate_impl----------------------------
3749 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3750   // Too many traps seen?
3751   if (too_many_traps(reason)) {
3752 #ifdef ASSERT
3753     if (TraceLoopPredicate) {
3754       int tc = C->trap_count(reason);
3755       tty->print("too many traps=%s tcount=%d in ",
3756                     Deoptimization::trap_reason_name(reason), tc);
3757       method()->print(); // which method has too many predicate traps
3758       tty->cr();
3759     }
3760 #endif
3761     // We cannot afford to take more traps here,
3762     // do not generate predicate.
3763     return;
3764   }
3765 
3766   Node *cont    = _gvn.intcon(1);
3767   Node* opq     = _gvn.transform(new (C) Opaque1Node(C, cont));
3768   Node *bol     = _gvn.transform(new (C) Conv2BNode(opq));
3769   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3770   Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3771   C->add_predicate_opaq(opq);
3772   {
3773     PreserveJVMState pjvms(this);
3774     set_control(iffalse);
3775     inc_sp(nargs);
3776     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3777   }
3778   Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3779   set_control(iftrue);
3780 }
3781 
3782 //------------------------------add_predicate---------------------------------
3783 void GraphKit::add_predicate(int nargs) {
3784   if (UseLoopPredicate) {
3785     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3786   }
3787   // loop's limit check predicate should be near the loop.
3788   if (LoopLimitCheck) {
3789     add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3790   }
3791 }
3792 
3793 //----------------------------- store barriers ----------------------------
3794 #define __ ideal.
3795 
3796 void GraphKit::sync_kit(IdealKit& ideal) {
3797   set_all_memory(__ merged_memory());
3798   set_i_o(__ i_o());
3799   set_control(__ ctrl());
3800 }
3801 
3802 void GraphKit::final_sync(IdealKit& ideal) {
3803   // Final sync IdealKit and graphKit.
3804   sync_kit(ideal);
3805 }
3806 
3807 // vanilla/CMS post barrier
3808 // Insert a write-barrier store.  This is to let generational GC work; we have
3809 // to flag all oop-stores before the next GC point.
3810 void GraphKit::write_barrier_post(Node* oop_store,
3811                                   Node* obj,
3812                                   Node* adr,
3813                                   uint  adr_idx,
3814                                   Node* val,
3815                                   bool use_precise) {
3816   // No store check needed if we're storing a NULL or an old object
3817   // (latter case is probably a string constant). The concurrent
3818   // mark sweep garbage collector, however, needs to have all nonNull
3819   // oop updates flagged via card-marks.
3820   if (val != NULL && val->is_Con()) {
3821     // must be either an oop or NULL
3822     const Type* t = val->bottom_type();
3823     if (t == TypePtr::NULL_PTR || t == Type::TOP)
3824       // stores of null never (?) need barriers
3825       return;
3826   }
3827 
3828   if (use_ReduceInitialCardMarks()
3829       && obj == just_allocated_object(control())) {
3830     // We can skip marks on a freshly-allocated object in Eden.
3831     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3832     // That routine informs GC to take appropriate compensating steps,
3833     // upon a slow-path allocation, so as to make this card-mark
3834     // elision safe.
3835     return;
3836   }
3837 
3838   if (!use_precise) {
3839     // All card marks for a (non-array) instance are in one place:
3840     adr = obj;
3841   }
3842   // (Else it's an array (or unknown), and we want more precise card marks.)
3843   assert(adr != NULL, "");
3844 
3845   IdealKit ideal(this, true);
3846 
3847   // Convert the pointer to an int prior to doing math on it
3848   Node* cast = __ CastPX(__ ctrl(), adr);
3849 
3850   // Divide by card size
3851   assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3852          "Only one we handle so far.");
3853   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3854 
3855   // Combine card table base and card offset
3856   Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3857 
3858   // Get the alias_index for raw card-mark memory
3859   int adr_type = Compile::AliasIdxRaw;
3860   Node*   zero = __ ConI(0); // Dirty card value
3861   BasicType bt = T_BYTE;
3862 
3863   if (UseCondCardMark) {
3864     // The classic GC reference write barrier is typically implemented
3865     // as a store into the global card mark table.  Unfortunately
3866     // unconditional stores can result in false sharing and excessive
3867     // coherence traffic as well as false transactional aborts.
3868     // UseCondCardMark enables MP "polite" conditional card mark
3869     // stores.  In theory we could relax the load from ctrl() to
3870     // no_ctrl, but that doesn't buy much latitude.
3871     Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3872     __ if_then(card_val, BoolTest::ne, zero);
3873   }
3874 
3875   // Smash zero into card
3876   if( !UseConcMarkSweepGC ) {
3877 #if defined(AARCH64)
3878     __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3879 #else
3880     __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
3881 #endif
3882   } else {
3883     // Specialized path for CM store barrier
3884     __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3885   }
3886 
3887   if (UseCondCardMark) {
3888     __ end_if();
3889   }
3890 
3891   // Final sync IdealKit and GraphKit.
3892   final_sync(ideal);
3893 }
3894 
3895 // G1 pre/post barriers
3896 void GraphKit::g1_write_barrier_pre(bool do_load,
3897                                     Node* obj,
3898                                     Node* adr,
3899                                     uint alias_idx,
3900                                     Node* val,
3901                                     const TypeOopPtr* val_type,
3902                                     Node* pre_val,
3903                                     BasicType bt) {
3904 
3905   // Some sanity checks
3906   // Note: val is unused in this routine.
3907 
3908   if (do_load) {
3909     // We need to generate the load of the previous value
3910     assert(obj != NULL, "must have a base");
3911     assert(adr != NULL, "where are loading from?");
3912     assert(pre_val == NULL, "loaded already?");
3913     assert(val_type != NULL, "need a type");
3914   } else {
3915     // In this case both val_type and alias_idx are unused.
3916     assert(pre_val != NULL, "must be loaded already");
3917     // Nothing to be done if pre_val is null.
3918     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3919     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3920   }
3921   assert(bt == T_OBJECT, "or we shouldn't be here");
3922 
3923   IdealKit ideal(this, true);
3924 
3925   Node* tls = __ thread(); // ThreadLocalStorage
3926 
3927   Node* no_ctrl = NULL;
3928   Node* no_base = __ top();
3929   Node* zero  = __ ConI(0);
3930   Node* zeroX = __ ConX(0);
3931 
3932   float likely  = PROB_LIKELY(0.999);
3933   float unlikely  = PROB_UNLIKELY(0.999);
3934 
3935   BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3936   assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3937 
3938   // Offsets into the thread
3939   const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +  // 648
3940                                           PtrQueue::byte_offset_of_active());
3941   const int index_offset   = in_bytes(JavaThread::satb_mark_queue_offset() +  // 656
3942                                           PtrQueue::byte_offset_of_index());
3943   const int buffer_offset  = in_bytes(JavaThread::satb_mark_queue_offset() +  // 652
3944                                           PtrQueue::byte_offset_of_buf());
3945 
3946   // Now the actual pointers into the thread
3947   Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3948   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
3949   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
3950 
3951   // Now some of the values
3952   Node* marking;
3953   if (UseShenandoahGC) {
3954     Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(JavaThread::gc_state_offset())));
3955     Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
3956     marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
3957     assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
3958   } else {
3959     assert(UseG1GC, "should be");
3960     marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3961   }
3962 
3963   // if (!marking)
3964   __ if_then(marking, BoolTest::ne, zero, unlikely); {
3965     BasicType index_bt = TypeX_X->basic_type();
3966     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3967     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3968 
3969     if (do_load) {
3970       // load original value
3971       // alias_idx correct??
3972       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3973     }
3974 
3975     // if (pre_val != NULL)
3976     __ if_then(pre_val, BoolTest::ne, null()); {
3977       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3978 
3979       // is the queue for this thread full?
3980       __ if_then(index, BoolTest::ne, zeroX, likely); {
3981 
3982         // decrement the index
3983         Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3984 
3985         // Now get the buffer location we will log the previous value into and store it
3986         Node *log_addr = __ AddP(no_base, buffer, next_index);
3987         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3988         // update the index
3989         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3990 
3991       } __ else_(); {
3992 
3993         // logging buffer is full, call the runtime
3994         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3995         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3996       } __ end_if();  // (!index)
3997     } __ end_if();  // (pre_val != NULL)
3998   } __ end_if();  // (!marking)
3999 
4000   // Final sync IdealKit and GraphKit.
4001   final_sync(ideal);
4002 
4003 #if INCLUDE_ALL_GCS
4004   if (UseShenandoahGC && adr != NULL) {
4005     Node* c = control();
4006     Node* call = c->in(1)->in(1)->in(1)->in(0);
4007     assert(call->is_g1_wb_pre_call(), "g1_wb_pre call expected");
4008     call->add_req(adr);
4009   }
4010 #endif
4011 }
4012 
4013 //
4014 // Update the card table and add card address to the queue
4015 //
4016 void GraphKit::g1_mark_card(IdealKit& ideal,
4017                             Node* card_adr,
4018                             Node* oop_store,
4019                             uint oop_alias_idx,
4020                             Node* index,
4021                             Node* index_adr,
4022                             Node* buffer,
4023                             const TypeFunc* tf) {
4024 
4025   Node* zero  = __ ConI(0);
4026   Node* zeroX = __ ConX(0);
4027   Node* no_base = __ top();
4028   BasicType card_bt = T_BYTE;
4029   // Smash zero into card. MUST BE ORDERED WRT TO STORE
4030   __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4031 
4032   //  Now do the queue work
4033   __ if_then(index, BoolTest::ne, zeroX); {
4034 
4035     Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
4036     Node* log_addr = __ AddP(no_base, buffer, next_index);
4037 
4038     // Order, see storeCM.
4039     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4040     __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4041 
4042   } __ else_(); {
4043     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4044   } __ end_if();
4045 
4046 }
4047 
4048 void GraphKit::g1_write_barrier_post(Node* oop_store,
4049                                      Node* obj,
4050                                      Node* adr,
4051                                      uint alias_idx,
4052                                      Node* val,
4053                                      BasicType bt,
4054                                      bool use_precise) {
4055   // If we are writing a NULL then we need no post barrier
4056 
4057   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4058     // Must be NULL
4059     const Type* t = val->bottom_type();
4060     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4061     // No post barrier if writing NULLx
4062     return;
4063   }
4064 
4065   if (!use_precise) {
4066     // All card marks for a (non-array) instance are in one place:
4067     adr = obj;
4068   }
4069   // (Else it's an array (or unknown), and we want more precise card marks.)
4070   assert(adr != NULL, "");
4071 
4072   IdealKit ideal(this, true);
4073 
4074   Node* tls = __ thread(); // ThreadLocalStorage
4075 
4076   Node* no_base = __ top();
4077   float likely  = PROB_LIKELY(0.999);
4078   float unlikely  = PROB_UNLIKELY(0.999);
4079   Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4080   Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4081   Node* zeroX = __ ConX(0);
4082 
4083   // Get the alias_index for raw card-mark memory
4084   const TypePtr* card_type = TypeRawPtr::BOTTOM;
4085 
4086   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4087 
4088   // Offsets into the thread
4089   const int index_offset  = in_bytes(JavaThread::dirty_card_queue_offset() +
4090                                      PtrQueue::byte_offset_of_index());
4091   const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4092                                      PtrQueue::byte_offset_of_buf());
4093 
4094   // Pointers into the thread
4095 
4096   Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4097   Node* index_adr =  __ AddP(no_base, tls, __ ConX(index_offset));
4098 
4099   // Now some values
4100   // Use ctrl to avoid hoisting these values past a safepoint, which could
4101   // potentially reset these fields in the JavaThread.
4102   Node* index  = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4103   Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4104 
4105   // Convert the store obj pointer to an int prior to doing math on it
4106   // Must use ctrl to prevent "integerized oop" existing across safepoint
4107   Node* cast =  __ CastPX(__ ctrl(), adr);
4108 
4109   // Divide pointer by card size
4110   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4111 
4112   // Combine card table base and card offset
4113   Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4114 
4115   // If we know the value being stored does it cross regions?
4116 
4117   if (val != NULL) {
4118     // Does the store cause us to cross regions?
4119 
4120     // Should be able to do an unsigned compare of region_size instead of
4121     // and extra shift. Do we have an unsigned compare??
4122     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4123     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4124 
4125     // if (xor_res == 0) same region so skip
4126     __ if_then(xor_res, BoolTest::ne, zeroX); {
4127 
4128       // No barrier if we are storing a NULL
4129       __ if_then(val, BoolTest::ne, null(), unlikely); {
4130 
4131         // Ok must mark the card if not already dirty
4132 
4133         // load the original value of the card
4134         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4135 
4136         __ if_then(card_val, BoolTest::ne, young_card); {
4137           sync_kit(ideal);
4138           // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4139           insert_mem_bar(Op_MemBarVolatile, oop_store);
4140           __ sync_kit(this);
4141 
4142           Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4143           __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4144             g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4145           } __ end_if();
4146         } __ end_if();
4147       } __ end_if();
4148     } __ end_if();
4149   } else {
4150     // Object.clone() instrinsic uses this path.
4151     g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4152   }
4153 
4154   // Final sync IdealKit and GraphKit.
4155   final_sync(ideal);
4156 }
4157 #undef __
4158 
4159 
4160 
4161 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4162   if (java_lang_String::has_offset_field()) {
4163     int offset_offset = java_lang_String::offset_offset_in_bytes();
4164     const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4165                                                        false, NULL, 0);
4166     const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4167     int offset_field_idx = C->get_alias_index(offset_field_type);
4168     return make_load(ctrl,
4169                      basic_plus_adr(str, str, offset_offset),
4170                      TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4171   } else {
4172     return intcon(0);
4173   }
4174 }
4175 
4176 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4177   if (java_lang_String::has_count_field()) {
4178     int count_offset = java_lang_String::count_offset_in_bytes();
4179     const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4180                                                        false, NULL, 0);
4181     const TypePtr* count_field_type = string_type->add_offset(count_offset);
4182     int count_field_idx = C->get_alias_index(count_field_type);
4183     return make_load(ctrl,
4184                      basic_plus_adr(str, str, count_offset),
4185                      TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4186   } else {
4187     return load_array_length(load_String_value(ctrl, str));
4188   }
4189 }
4190 
4191 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4192   int value_offset = java_lang_String::value_offset_in_bytes();
4193   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4194                                                      false, NULL, 0);
4195   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4196   const TypeAryPtr*  value_type = TypeAryPtr::make(TypePtr::NotNull,
4197                                                    TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4198                                                    ciTypeArrayKlass::make(T_CHAR), true, 0);
4199   int value_field_idx = C->get_alias_index(value_field_type);
4200   Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4201                          value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4202 #if INCLUDE_ALL_GCS
4203   if (UseShenandoahGC) {
4204     load = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, load);
4205   }
4206 #endif
4207   // String.value field is known to be @Stable.
4208   if (UseImplicitStableValues) {
4209     load = cast_array_to_stable(load, value_type);
4210   }
4211   return load;
4212 }
4213 
4214 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4215   int offset_offset = java_lang_String::offset_offset_in_bytes();
4216   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4217                                                      false, NULL, 0);
4218   const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4219   int offset_field_idx = C->get_alias_index(offset_field_type);
4220   store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4221                   value, T_INT, offset_field_idx, MemNode::unordered);
4222 }
4223 
4224 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4225   int value_offset = java_lang_String::value_offset_in_bytes();
4226   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4227                                                      false, NULL, 0);
4228   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4229 
4230   store_oop_to_object(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
4231       value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4232 }
4233 
4234 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4235   int count_offset = java_lang_String::count_offset_in_bytes();
4236   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4237                                                      false, NULL, 0);
4238   const TypePtr* count_field_type = string_type->add_offset(count_offset);
4239   int count_field_idx = C->get_alias_index(count_field_type);
4240   store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4241                   value, T_INT, count_field_idx, MemNode::unordered);
4242 }
4243 
4244 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4245   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4246   // assumption of CCP analysis.
4247   return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4248 }
--- EOF ---