1 /*
   2  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "ci/ciUtilities.hpp"
  27 #include "compiler/compileLog.hpp"
  28 #include "gc/shared/barrierSet.hpp"
  29 #include "gc/shared/cardTable.hpp"
  30 #include "gc/shared/cardTableBarrierSet.hpp"
  31 #include "gc/shared/collectedHeap.hpp"
  32 #include "interpreter/interpreter.hpp"
  33 #include "memory/resourceArea.hpp"
  34 #include "opto/addnode.hpp"
  35 #include "opto/castnode.hpp"
  36 #include "opto/convertnode.hpp"
  37 #include "opto/graphKit.hpp"
  38 #include "opto/idealKit.hpp"
  39 #include "opto/intrinsicnode.hpp"
  40 #include "opto/locknode.hpp"
  41 #include "opto/machnode.hpp"
  42 #include "opto/opaquenode.hpp"
  43 #include "opto/parse.hpp"
  44 #include "opto/rootnode.hpp"
  45 #include "opto/runtime.hpp"
  46 #include "runtime/deoptimization.hpp"
  47 #include "runtime/sharedRuntime.hpp"
  48 #if INCLUDE_G1GC
  49 #include "gc/g1/g1CardTable.hpp"
  50 #include "gc/g1/g1ThreadLocalData.hpp"
  51 #include "gc/g1/heapRegion.hpp"
  52 #endif // INCLUDE_ALL_GCS
  53 
  54 //----------------------------GraphKit-----------------------------------------
  55 // Main utility constructor.
  56 GraphKit::GraphKit(JVMState* jvms)
  57   : Phase(Phase::Parser),
  58     _env(C->env()),
  59     _gvn(*C->initial_gvn())
  60 {
  61   _exceptions = jvms->map()->next_exception();
  62   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  63   set_jvms(jvms);
  64 }
  65 
  66 // Private constructor for parser.
  67 GraphKit::GraphKit()
  68   : Phase(Phase::Parser),
  69     _env(C->env()),
  70     _gvn(*C->initial_gvn())
  71 {
  72   _exceptions = NULL;
  73   set_map(NULL);
  74   debug_only(_sp = -99);
  75   debug_only(set_bci(-99));
  76 }
  77 
  78 
  79 
  80 //---------------------------clean_stack---------------------------------------
  81 // Clear away rubbish from the stack area of the JVM state.
  82 // This destroys any arguments that may be waiting on the stack.
  83 void GraphKit::clean_stack(int from_sp) {
  84   SafePointNode* map      = this->map();
  85   JVMState*      jvms     = this->jvms();
  86   int            stk_size = jvms->stk_size();
  87   int            stkoff   = jvms->stkoff();
  88   Node*          top      = this->top();
  89   for (int i = from_sp; i < stk_size; i++) {
  90     if (map->in(stkoff + i) != top) {
  91       map->set_req(stkoff + i, top);
  92     }
  93   }
  94 }
  95 
  96 
  97 //--------------------------------sync_jvms-----------------------------------
  98 // Make sure our current jvms agrees with our parse state.
  99 JVMState* GraphKit::sync_jvms() const {
 100   JVMState* jvms = this->jvms();
 101   jvms->set_bci(bci());       // Record the new bci in the JVMState
 102   jvms->set_sp(sp());         // Record the new sp in the JVMState
 103   assert(jvms_in_sync(), "jvms is now in sync");
 104   return jvms;
 105 }
 106 
 107 //--------------------------------sync_jvms_for_reexecute---------------------
 108 // Make sure our current jvms agrees with our parse state.  This version
 109 // uses the reexecute_sp for reexecuting bytecodes.
 110 JVMState* GraphKit::sync_jvms_for_reexecute() {
 111   JVMState* jvms = this->jvms();
 112   jvms->set_bci(bci());          // Record the new bci in the JVMState
 113   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 114   return jvms;
 115 }
 116 
 117 #ifdef ASSERT
 118 bool GraphKit::jvms_in_sync() const {
 119   Parse* parse = is_Parse();
 120   if (parse == NULL) {
 121     if (bci() !=      jvms()->bci())          return false;
 122     if (sp()  != (int)jvms()->sp())           return false;
 123     return true;
 124   }
 125   if (jvms()->method() != parse->method())    return false;
 126   if (jvms()->bci()    != parse->bci())       return false;
 127   int jvms_sp = jvms()->sp();
 128   if (jvms_sp          != parse->sp())        return false;
 129   int jvms_depth = jvms()->depth();
 130   if (jvms_depth       != parse->depth())     return false;
 131   return true;
 132 }
 133 
 134 // Local helper checks for special internal merge points
 135 // used to accumulate and merge exception states.
 136 // They are marked by the region's in(0) edge being the map itself.
 137 // Such merge points must never "escape" into the parser at large,
 138 // until they have been handed to gvn.transform.
 139 static bool is_hidden_merge(Node* reg) {
 140   if (reg == NULL)  return false;
 141   if (reg->is_Phi()) {
 142     reg = reg->in(0);
 143     if (reg == NULL)  return false;
 144   }
 145   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 146 }
 147 
 148 void GraphKit::verify_map() const {
 149   if (map() == NULL)  return;  // null map is OK
 150   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 151   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 152   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 153 }
 154 
 155 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 156   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 157   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 158 }
 159 #endif
 160 
 161 //---------------------------stop_and_kill_map---------------------------------
 162 // Set _map to NULL, signalling a stop to further bytecode execution.
 163 // First smash the current map's control to a constant, to mark it dead.
 164 void GraphKit::stop_and_kill_map() {
 165   SafePointNode* dead_map = stop();
 166   if (dead_map != NULL) {
 167     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
 168     assert(dead_map->is_killed(), "must be so marked");
 169   }
 170 }
 171 
 172 
 173 //--------------------------------stopped--------------------------------------
 174 // Tell if _map is NULL, or control is top.
 175 bool GraphKit::stopped() {
 176   if (map() == NULL)           return true;
 177   else if (control() == top()) return true;
 178   else                         return false;
 179 }
 180 
 181 
 182 //-----------------------------has_ex_handler----------------------------------
 183 // Tell if this method or any caller method has exception handlers.
 184 bool GraphKit::has_ex_handler() {
 185   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 186     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 187       return true;
 188     }
 189   }
 190   return false;
 191 }
 192 
 193 //------------------------------save_ex_oop------------------------------------
 194 // Save an exception without blowing stack contents or other JVM state.
 195 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 196   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 197   ex_map->add_req(ex_oop);
 198   debug_only(verify_exception_state(ex_map));
 199 }
 200 
 201 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 202   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 203   Node* ex_oop = ex_map->in(ex_map->req()-1);
 204   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 205   return ex_oop;
 206 }
 207 
 208 //-----------------------------saved_ex_oop------------------------------------
 209 // Recover a saved exception from its map.
 210 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 211   return common_saved_ex_oop(ex_map, false);
 212 }
 213 
 214 //--------------------------clear_saved_ex_oop---------------------------------
 215 // Erase a previously saved exception from its map.
 216 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 217   return common_saved_ex_oop(ex_map, true);
 218 }
 219 
 220 #ifdef ASSERT
 221 //---------------------------has_saved_ex_oop----------------------------------
 222 // Erase a previously saved exception from its map.
 223 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 224   return ex_map->req() == ex_map->jvms()->endoff()+1;
 225 }
 226 #endif
 227 
 228 //-------------------------make_exception_state--------------------------------
 229 // Turn the current JVM state into an exception state, appending the ex_oop.
 230 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 231   sync_jvms();
 232   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 233   set_saved_ex_oop(ex_map, ex_oop);
 234   return ex_map;
 235 }
 236 
 237 
 238 //--------------------------add_exception_state--------------------------------
 239 // Add an exception to my list of exceptions.
 240 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 241   if (ex_map == NULL || ex_map->control() == top()) {
 242     return;
 243   }
 244 #ifdef ASSERT
 245   verify_exception_state(ex_map);
 246   if (has_exceptions()) {
 247     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 248   }
 249 #endif
 250 
 251   // If there is already an exception of exactly this type, merge with it.
 252   // In particular, null-checks and other low-level exceptions common up here.
 253   Node*       ex_oop  = saved_ex_oop(ex_map);
 254   const Type* ex_type = _gvn.type(ex_oop);
 255   if (ex_oop == top()) {
 256     // No action needed.
 257     return;
 258   }
 259   assert(ex_type->isa_instptr(), "exception must be an instance");
 260   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 261     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 262     // We check sp also because call bytecodes can generate exceptions
 263     // both before and after arguments are popped!
 264     if (ex_type2 == ex_type
 265         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 266       combine_exception_states(ex_map, e2);
 267       return;
 268     }
 269   }
 270 
 271   // No pre-existing exception of the same type.  Chain it on the list.
 272   push_exception_state(ex_map);
 273 }
 274 
 275 //-----------------------add_exception_states_from-----------------------------
 276 void GraphKit::add_exception_states_from(JVMState* jvms) {
 277   SafePointNode* ex_map = jvms->map()->next_exception();
 278   if (ex_map != NULL) {
 279     jvms->map()->set_next_exception(NULL);
 280     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 281       next_map = ex_map->next_exception();
 282       ex_map->set_next_exception(NULL);
 283       add_exception_state(ex_map);
 284     }
 285   }
 286 }
 287 
 288 //-----------------------transfer_exceptions_into_jvms-------------------------
 289 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 290   if (map() == NULL) {
 291     // We need a JVMS to carry the exceptions, but the map has gone away.
 292     // Create a scratch JVMS, cloned from any of the exception states...
 293     if (has_exceptions()) {
 294       _map = _exceptions;
 295       _map = clone_map();
 296       _map->set_next_exception(NULL);
 297       clear_saved_ex_oop(_map);
 298       debug_only(verify_map());
 299     } else {
 300       // ...or created from scratch
 301       JVMState* jvms = new (C) JVMState(_method, NULL);
 302       jvms->set_bci(_bci);
 303       jvms->set_sp(_sp);
 304       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
 305       set_jvms(jvms);
 306       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 307       set_all_memory(top());
 308       while (map()->req() < jvms->endoff())  map()->add_req(top());
 309     }
 310     // (This is a kludge, in case you didn't notice.)
 311     set_control(top());
 312   }
 313   JVMState* jvms = sync_jvms();
 314   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 315   jvms->map()->set_next_exception(_exceptions);
 316   _exceptions = NULL;   // done with this set of exceptions
 317   return jvms;
 318 }
 319 
 320 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 321   assert(is_hidden_merge(dstphi), "must be a special merge node");
 322   assert(is_hidden_merge(srcphi), "must be a special merge node");
 323   uint limit = srcphi->req();
 324   for (uint i = PhiNode::Input; i < limit; i++) {
 325     dstphi->add_req(srcphi->in(i));
 326   }
 327 }
 328 static inline void add_one_req(Node* dstphi, Node* src) {
 329   assert(is_hidden_merge(dstphi), "must be a special merge node");
 330   assert(!is_hidden_merge(src), "must not be a special merge node");
 331   dstphi->add_req(src);
 332 }
 333 
 334 //-----------------------combine_exception_states------------------------------
 335 // This helper function combines exception states by building phis on a
 336 // specially marked state-merging region.  These regions and phis are
 337 // untransformed, and can build up gradually.  The region is marked by
 338 // having a control input of its exception map, rather than NULL.  Such
 339 // regions do not appear except in this function, and in use_exception_state.
 340 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 341   if (failing())  return;  // dying anyway...
 342   JVMState* ex_jvms = ex_map->_jvms;
 343   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 344   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 345   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 346   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 347   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 348   assert(ex_map->req() == phi_map->req(), "matching maps");
 349   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 350   Node*         hidden_merge_mark = root();
 351   Node*         region  = phi_map->control();
 352   MergeMemNode* phi_mem = phi_map->merged_memory();
 353   MergeMemNode* ex_mem  = ex_map->merged_memory();
 354   if (region->in(0) != hidden_merge_mark) {
 355     // The control input is not (yet) a specially-marked region in phi_map.
 356     // Make it so, and build some phis.
 357     region = new RegionNode(2);
 358     _gvn.set_type(region, Type::CONTROL);
 359     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 360     region->init_req(1, phi_map->control());
 361     phi_map->set_control(region);
 362     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 363     record_for_igvn(io_phi);
 364     _gvn.set_type(io_phi, Type::ABIO);
 365     phi_map->set_i_o(io_phi);
 366     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 367       Node* m = mms.memory();
 368       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 369       record_for_igvn(m_phi);
 370       _gvn.set_type(m_phi, Type::MEMORY);
 371       mms.set_memory(m_phi);
 372     }
 373   }
 374 
 375   // Either or both of phi_map and ex_map might already be converted into phis.
 376   Node* ex_control = ex_map->control();
 377   // if there is special marking on ex_map also, we add multiple edges from src
 378   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 379   // how wide was the destination phi_map, originally?
 380   uint orig_width = region->req();
 381 
 382   if (add_multiple) {
 383     add_n_reqs(region, ex_control);
 384     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 385   } else {
 386     // ex_map has no merges, so we just add single edges everywhere
 387     add_one_req(region, ex_control);
 388     add_one_req(phi_map->i_o(), ex_map->i_o());
 389   }
 390   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 391     if (mms.is_empty()) {
 392       // get a copy of the base memory, and patch some inputs into it
 393       const TypePtr* adr_type = mms.adr_type(C);
 394       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 395       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 396       mms.set_memory(phi);
 397       // Prepare to append interesting stuff onto the newly sliced phi:
 398       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 399     }
 400     // Append stuff from ex_map:
 401     if (add_multiple) {
 402       add_n_reqs(mms.memory(), mms.memory2());
 403     } else {
 404       add_one_req(mms.memory(), mms.memory2());
 405     }
 406   }
 407   uint limit = ex_map->req();
 408   for (uint i = TypeFunc::Parms; i < limit; i++) {
 409     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 410     if (i == tos)  i = ex_jvms->monoff();
 411     Node* src = ex_map->in(i);
 412     Node* dst = phi_map->in(i);
 413     if (src != dst) {
 414       PhiNode* phi;
 415       if (dst->in(0) != region) {
 416         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 417         record_for_igvn(phi);
 418         _gvn.set_type(phi, phi->type());
 419         phi_map->set_req(i, dst);
 420         // Prepare to append interesting stuff onto the new phi:
 421         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 422       } else {
 423         assert(dst->is_Phi(), "nobody else uses a hidden region");
 424         phi = dst->as_Phi();
 425       }
 426       if (add_multiple && src->in(0) == ex_control) {
 427         // Both are phis.
 428         add_n_reqs(dst, src);
 429       } else {
 430         while (dst->req() < region->req())  add_one_req(dst, src);
 431       }
 432       const Type* srctype = _gvn.type(src);
 433       if (phi->type() != srctype) {
 434         const Type* dsttype = phi->type()->meet_speculative(srctype);
 435         if (phi->type() != dsttype) {
 436           phi->set_type(dsttype);
 437           _gvn.set_type(phi, dsttype);
 438         }
 439       }
 440     }
 441   }
 442   phi_map->merge_replaced_nodes_with(ex_map);
 443 }
 444 
 445 //--------------------------use_exception_state--------------------------------
 446 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 447   if (failing()) { stop(); return top(); }
 448   Node* region = phi_map->control();
 449   Node* hidden_merge_mark = root();
 450   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 451   Node* ex_oop = clear_saved_ex_oop(phi_map);
 452   if (region->in(0) == hidden_merge_mark) {
 453     // Special marking for internal ex-states.  Process the phis now.
 454     region->set_req(0, region);  // now it's an ordinary region
 455     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 456     // Note: Setting the jvms also sets the bci and sp.
 457     set_control(_gvn.transform(region));
 458     uint tos = jvms()->stkoff() + sp();
 459     for (uint i = 1; i < tos; i++) {
 460       Node* x = phi_map->in(i);
 461       if (x->in(0) == region) {
 462         assert(x->is_Phi(), "expected a special phi");
 463         phi_map->set_req(i, _gvn.transform(x));
 464       }
 465     }
 466     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 467       Node* x = mms.memory();
 468       if (x->in(0) == region) {
 469         assert(x->is_Phi(), "nobody else uses a hidden region");
 470         mms.set_memory(_gvn.transform(x));
 471       }
 472     }
 473     if (ex_oop->in(0) == region) {
 474       assert(ex_oop->is_Phi(), "expected a special phi");
 475       ex_oop = _gvn.transform(ex_oop);
 476     }
 477   } else {
 478     set_jvms(phi_map->jvms());
 479   }
 480 
 481   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 482   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 483   return ex_oop;
 484 }
 485 
 486 //---------------------------------java_bc-------------------------------------
 487 Bytecodes::Code GraphKit::java_bc() const {
 488   ciMethod* method = this->method();
 489   int       bci    = this->bci();
 490   if (method != NULL && bci != InvocationEntryBci)
 491     return method->java_code_at_bci(bci);
 492   else
 493     return Bytecodes::_illegal;
 494 }
 495 
 496 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 497                                                           bool must_throw) {
 498     // if the exception capability is set, then we will generate code
 499     // to check the JavaThread.should_post_on_exceptions flag to see
 500     // if we actually need to report exception events (for this
 501     // thread).  If we don't need to report exception events, we will
 502     // take the normal fast path provided by add_exception_events.  If
 503     // exception event reporting is enabled for this thread, we will
 504     // take the uncommon_trap in the BuildCutout below.
 505 
 506     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 507     Node* jthread = _gvn.transform(new ThreadLocalNode());
 508     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 509     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
 510 
 511     // Test the should_post_on_exceptions_flag vs. 0
 512     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
 513     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
 514 
 515     // Branch to slow_path if should_post_on_exceptions_flag was true
 516     { BuildCutout unless(this, tst, PROB_MAX);
 517       // Do not try anything fancy if we're notifying the VM on every throw.
 518       // Cf. case Bytecodes::_athrow in parse2.cpp.
 519       uncommon_trap(reason, Deoptimization::Action_none,
 520                     (ciKlass*)NULL, (char*)NULL, must_throw);
 521     }
 522 
 523 }
 524 
 525 //------------------------------builtin_throw----------------------------------
 526 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 527   bool must_throw = true;
 528 
 529   if (env()->jvmti_can_post_on_exceptions()) {
 530     // check if we must post exception events, take uncommon trap if so
 531     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 532     // here if should_post_on_exceptions is false
 533     // continue on with the normal codegen
 534   }
 535 
 536   // If this particular condition has not yet happened at this
 537   // bytecode, then use the uncommon trap mechanism, and allow for
 538   // a future recompilation if several traps occur here.
 539   // If the throw is hot, try to use a more complicated inline mechanism
 540   // which keeps execution inside the compiled code.
 541   bool treat_throw_as_hot = false;
 542   ciMethodData* md = method()->method_data();
 543 
 544   if (ProfileTraps) {
 545     if (too_many_traps(reason)) {
 546       treat_throw_as_hot = true;
 547     }
 548     // (If there is no MDO at all, assume it is early in
 549     // execution, and that any deopts are part of the
 550     // startup transient, and don't need to be remembered.)
 551 
 552     // Also, if there is a local exception handler, treat all throws
 553     // as hot if there has been at least one in this method.
 554     if (C->trap_count(reason) != 0
 555         && method()->method_data()->trap_count(reason) != 0
 556         && has_ex_handler()) {
 557         treat_throw_as_hot = true;
 558     }
 559   }
 560 
 561   // If this throw happens frequently, an uncommon trap might cause
 562   // a performance pothole.  If there is a local exception handler,
 563   // and if this particular bytecode appears to be deoptimizing often,
 564   // let us handle the throw inline, with a preconstructed instance.
 565   // Note:   If the deopt count has blown up, the uncommon trap
 566   // runtime is going to flush this nmethod, not matter what.
 567   if (treat_throw_as_hot
 568       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 569     // If the throw is local, we use a pre-existing instance and
 570     // punt on the backtrace.  This would lead to a missing backtrace
 571     // (a repeat of 4292742) if the backtrace object is ever asked
 572     // for its backtrace.
 573     // Fixing this remaining case of 4292742 requires some flavor of
 574     // escape analysis.  Leave that for the future.
 575     ciInstance* ex_obj = NULL;
 576     switch (reason) {
 577     case Deoptimization::Reason_null_check:
 578       ex_obj = env()->NullPointerException_instance();
 579       break;
 580     case Deoptimization::Reason_div0_check:
 581       ex_obj = env()->ArithmeticException_instance();
 582       break;
 583     case Deoptimization::Reason_range_check:
 584       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 585       break;
 586     case Deoptimization::Reason_class_check:
 587       if (java_bc() == Bytecodes::_aastore) {
 588         ex_obj = env()->ArrayStoreException_instance();
 589       } else {
 590         ex_obj = env()->ClassCastException_instance();
 591       }
 592       break;
 593     default:
 594       break;
 595     }
 596     if (failing()) { stop(); return; }  // exception allocation might fail
 597     if (ex_obj != NULL) {
 598       // Cheat with a preallocated exception object.
 599       if (C->log() != NULL)
 600         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 601                        Deoptimization::trap_reason_name(reason));
 602       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 603       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
 604 
 605       // Clear the detail message of the preallocated exception object.
 606       // Weblogic sometimes mutates the detail message of exceptions
 607       // using reflection.
 608       int offset = java_lang_Throwable::get_detailMessage_offset();
 609       const TypePtr* adr_typ = ex_con->add_offset(offset);
 610 
 611       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 612       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 613       // Conservatively release stores of object references.
 614       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
 615 
 616       add_exception_state(make_exception_state(ex_node));
 617       return;
 618     }
 619   }
 620 
 621   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 622   // It won't be much cheaper than bailing to the interp., since we'll
 623   // have to pass up all the debug-info, and the runtime will have to
 624   // create the stack trace.
 625 
 626   // Usual case:  Bail to interpreter.
 627   // Reserve the right to recompile if we haven't seen anything yet.
 628 
 629   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
 630   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 631   if (treat_throw_as_hot
 632       && (method()->method_data()->trap_recompiled_at(bci(), m)
 633           || C->too_many_traps(reason))) {
 634     // We cannot afford to take more traps here.  Suffer in the interpreter.
 635     if (C->log() != NULL)
 636       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 637                      Deoptimization::trap_reason_name(reason),
 638                      C->trap_count(reason));
 639     action = Deoptimization::Action_none;
 640   }
 641 
 642   // "must_throw" prunes the JVM state to include only the stack, if there
 643   // are no local exception handlers.  This should cut down on register
 644   // allocation time and code size, by drastically reducing the number
 645   // of in-edges on the call to the uncommon trap.
 646 
 647   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 648 }
 649 
 650 
 651 //----------------------------PreserveJVMState---------------------------------
 652 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 653   debug_only(kit->verify_map());
 654   _kit    = kit;
 655   _map    = kit->map();   // preserve the map
 656   _sp     = kit->sp();
 657   kit->set_map(clone_map ? kit->clone_map() : NULL);
 658 #ifdef ASSERT
 659   _bci    = kit->bci();
 660   Parse* parser = kit->is_Parse();
 661   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 662   _block  = block;
 663 #endif
 664 }
 665 PreserveJVMState::~PreserveJVMState() {
 666   GraphKit* kit = _kit;
 667 #ifdef ASSERT
 668   assert(kit->bci() == _bci, "bci must not shift");
 669   Parse* parser = kit->is_Parse();
 670   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 671   assert(block == _block,    "block must not shift");
 672 #endif
 673   kit->set_map(_map);
 674   kit->set_sp(_sp);
 675 }
 676 
 677 
 678 //-----------------------------BuildCutout-------------------------------------
 679 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 680   : PreserveJVMState(kit)
 681 {
 682   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 683   SafePointNode* outer_map = _map;   // preserved map is caller's
 684   SafePointNode* inner_map = kit->map();
 685   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 686   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
 687   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
 688 }
 689 BuildCutout::~BuildCutout() {
 690   GraphKit* kit = _kit;
 691   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 692 }
 693 
 694 //---------------------------PreserveReexecuteState----------------------------
 695 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 696   assert(!kit->stopped(), "must call stopped() before");
 697   _kit    =    kit;
 698   _sp     =    kit->sp();
 699   _reexecute = kit->jvms()->_reexecute;
 700 }
 701 PreserveReexecuteState::~PreserveReexecuteState() {
 702   if (_kit->stopped()) return;
 703   _kit->jvms()->_reexecute = _reexecute;
 704   _kit->set_sp(_sp);
 705 }
 706 
 707 //------------------------------clone_map--------------------------------------
 708 // Implementation of PreserveJVMState
 709 //
 710 // Only clone_map(...) here. If this function is only used in the
 711 // PreserveJVMState class we may want to get rid of this extra
 712 // function eventually and do it all there.
 713 
 714 SafePointNode* GraphKit::clone_map() {
 715   if (map() == NULL)  return NULL;
 716 
 717   // Clone the memory edge first
 718   Node* mem = MergeMemNode::make(map()->memory());
 719   gvn().set_type_bottom(mem);
 720 
 721   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 722   JVMState* jvms = this->jvms();
 723   JVMState* clonejvms = jvms->clone_shallow(C);
 724   clonemap->set_memory(mem);
 725   clonemap->set_jvms(clonejvms);
 726   clonejvms->set_map(clonemap);
 727   record_for_igvn(clonemap);
 728   gvn().set_type_bottom(clonemap);
 729   return clonemap;
 730 }
 731 
 732 
 733 //-----------------------------set_map_clone-----------------------------------
 734 void GraphKit::set_map_clone(SafePointNode* m) {
 735   _map = m;
 736   _map = clone_map();
 737   _map->set_next_exception(NULL);
 738   debug_only(verify_map());
 739 }
 740 
 741 
 742 //----------------------------kill_dead_locals---------------------------------
 743 // Detect any locals which are known to be dead, and force them to top.
 744 void GraphKit::kill_dead_locals() {
 745   // Consult the liveness information for the locals.  If any
 746   // of them are unused, then they can be replaced by top().  This
 747   // should help register allocation time and cut down on the size
 748   // of the deoptimization information.
 749 
 750   // This call is made from many of the bytecode handling
 751   // subroutines called from the Big Switch in do_one_bytecode.
 752   // Every bytecode which might include a slow path is responsible
 753   // for killing its dead locals.  The more consistent we
 754   // are about killing deads, the fewer useless phis will be
 755   // constructed for them at various merge points.
 756 
 757   // bci can be -1 (InvocationEntryBci).  We return the entry
 758   // liveness for the method.
 759 
 760   if (method() == NULL || method()->code_size() == 0) {
 761     // We are building a graph for a call to a native method.
 762     // All locals are live.
 763     return;
 764   }
 765 
 766   ResourceMark rm;
 767 
 768   // Consult the liveness information for the locals.  If any
 769   // of them are unused, then they can be replaced by top().  This
 770   // should help register allocation time and cut down on the size
 771   // of the deoptimization information.
 772   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 773 
 774   int len = (int)live_locals.size();
 775   assert(len <= jvms()->loc_size(), "too many live locals");
 776   for (int local = 0; local < len; local++) {
 777     if (!live_locals.at(local)) {
 778       set_local(local, top());
 779     }
 780   }
 781 }
 782 
 783 #ifdef ASSERT
 784 //-------------------------dead_locals_are_killed------------------------------
 785 // Return true if all dead locals are set to top in the map.
 786 // Used to assert "clean" debug info at various points.
 787 bool GraphKit::dead_locals_are_killed() {
 788   if (method() == NULL || method()->code_size() == 0) {
 789     // No locals need to be dead, so all is as it should be.
 790     return true;
 791   }
 792 
 793   // Make sure somebody called kill_dead_locals upstream.
 794   ResourceMark rm;
 795   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 796     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 797     SafePointNode* map = jvms->map();
 798     ciMethod* method = jvms->method();
 799     int       bci    = jvms->bci();
 800     if (jvms == this->jvms()) {
 801       bci = this->bci();  // it might not yet be synched
 802     }
 803     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 804     int len = (int)live_locals.size();
 805     if (!live_locals.is_valid() || len == 0)
 806       // This method is trivial, or is poisoned by a breakpoint.
 807       return true;
 808     assert(len == jvms->loc_size(), "live map consistent with locals map");
 809     for (int local = 0; local < len; local++) {
 810       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 811         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 812           tty->print_cr("Zombie local %d: ", local);
 813           jvms->dump();
 814         }
 815         return false;
 816       }
 817     }
 818   }
 819   return true;
 820 }
 821 
 822 #endif //ASSERT
 823 
 824 // Helper function for enforcing certain bytecodes to reexecute if
 825 // deoptimization happens
 826 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 827   ciMethod* cur_method = jvms->method();
 828   int       cur_bci   = jvms->bci();
 829   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 830     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 831     return Interpreter::bytecode_should_reexecute(code) ||
 832            (is_anewarray && code == Bytecodes::_multianewarray);
 833     // Reexecute _multianewarray bytecode which was replaced with
 834     // sequence of [a]newarray. See Parse::do_multianewarray().
 835     //
 836     // Note: interpreter should not have it set since this optimization
 837     // is limited by dimensions and guarded by flag so in some cases
 838     // multianewarray() runtime calls will be generated and
 839     // the bytecode should not be reexecutes (stack will not be reset).
 840   } else
 841     return false;
 842 }
 843 
 844 // Helper function for adding JVMState and debug information to node
 845 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 846   // Add the safepoint edges to the call (or other safepoint).
 847 
 848   // Make sure dead locals are set to top.  This
 849   // should help register allocation time and cut down on the size
 850   // of the deoptimization information.
 851   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 852 
 853   // Walk the inline list to fill in the correct set of JVMState's
 854   // Also fill in the associated edges for each JVMState.
 855 
 856   // If the bytecode needs to be reexecuted we need to put
 857   // the arguments back on the stack.
 858   const bool should_reexecute = jvms()->should_reexecute();
 859   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 860 
 861   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 862   // undefined if the bci is different.  This is normal for Parse but it
 863   // should not happen for LibraryCallKit because only one bci is processed.
 864   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 865          "in LibraryCallKit the reexecute bit should not change");
 866 
 867   // If we are guaranteed to throw, we can prune everything but the
 868   // input to the current bytecode.
 869   bool can_prune_locals = false;
 870   uint stack_slots_not_pruned = 0;
 871   int inputs = 0, depth = 0;
 872   if (must_throw) {
 873     assert(method() == youngest_jvms->method(), "sanity");
 874     if (compute_stack_effects(inputs, depth)) {
 875       can_prune_locals = true;
 876       stack_slots_not_pruned = inputs;
 877     }
 878   }
 879 
 880   if (env()->should_retain_local_variables()) {
 881     // At any safepoint, this method can get breakpointed, which would
 882     // then require an immediate deoptimization.
 883     can_prune_locals = false;  // do not prune locals
 884     stack_slots_not_pruned = 0;
 885   }
 886 
 887   // do not scribble on the input jvms
 888   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 889   call->set_jvms(out_jvms); // Start jvms list for call node
 890 
 891   // For a known set of bytecodes, the interpreter should reexecute them if
 892   // deoptimization happens. We set the reexecute state for them here
 893   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 894       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 895     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 896   }
 897 
 898   // Presize the call:
 899   DEBUG_ONLY(uint non_debug_edges = call->req());
 900   call->add_req_batch(top(), youngest_jvms->debug_depth());
 901   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 902 
 903   // Set up edges so that the call looks like this:
 904   //  Call [state:] ctl io mem fptr retadr
 905   //       [parms:] parm0 ... parmN
 906   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 907   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 908   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 909   // Note that caller debug info precedes callee debug info.
 910 
 911   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 912   uint debug_ptr = call->req();
 913 
 914   // Loop over the map input edges associated with jvms, add them
 915   // to the call node, & reset all offsets to match call node array.
 916   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 917     uint debug_end   = debug_ptr;
 918     uint debug_start = debug_ptr - in_jvms->debug_size();
 919     debug_ptr = debug_start;  // back up the ptr
 920 
 921     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 922     uint j, k, l;
 923     SafePointNode* in_map = in_jvms->map();
 924     out_jvms->set_map(call);
 925 
 926     if (can_prune_locals) {
 927       assert(in_jvms->method() == out_jvms->method(), "sanity");
 928       // If the current throw can reach an exception handler in this JVMS,
 929       // then we must keep everything live that can reach that handler.
 930       // As a quick and dirty approximation, we look for any handlers at all.
 931       if (in_jvms->method()->has_exception_handlers()) {
 932         can_prune_locals = false;
 933       }
 934     }
 935 
 936     // Add the Locals
 937     k = in_jvms->locoff();
 938     l = in_jvms->loc_size();
 939     out_jvms->set_locoff(p);
 940     if (!can_prune_locals) {
 941       for (j = 0; j < l; j++)
 942         call->set_req(p++, in_map->in(k+j));
 943     } else {
 944       p += l;  // already set to top above by add_req_batch
 945     }
 946 
 947     // Add the Expression Stack
 948     k = in_jvms->stkoff();
 949     l = in_jvms->sp();
 950     out_jvms->set_stkoff(p);
 951     if (!can_prune_locals) {
 952       for (j = 0; j < l; j++)
 953         call->set_req(p++, in_map->in(k+j));
 954     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 955       // Divide stack into {S0,...,S1}, where S0 is set to top.
 956       uint s1 = stack_slots_not_pruned;
 957       stack_slots_not_pruned = 0;  // for next iteration
 958       if (s1 > l)  s1 = l;
 959       uint s0 = l - s1;
 960       p += s0;  // skip the tops preinstalled by add_req_batch
 961       for (j = s0; j < l; j++)
 962         call->set_req(p++, in_map->in(k+j));
 963     } else {
 964       p += l;  // already set to top above by add_req_batch
 965     }
 966 
 967     // Add the Monitors
 968     k = in_jvms->monoff();
 969     l = in_jvms->mon_size();
 970     out_jvms->set_monoff(p);
 971     for (j = 0; j < l; j++)
 972       call->set_req(p++, in_map->in(k+j));
 973 
 974     // Copy any scalar object fields.
 975     k = in_jvms->scloff();
 976     l = in_jvms->scl_size();
 977     out_jvms->set_scloff(p);
 978     for (j = 0; j < l; j++)
 979       call->set_req(p++, in_map->in(k+j));
 980 
 981     // Finish the new jvms.
 982     out_jvms->set_endoff(p);
 983 
 984     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 985     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 986     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 987     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 988     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 989     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 990 
 991     // Update the two tail pointers in parallel.
 992     out_jvms = out_jvms->caller();
 993     in_jvms  = in_jvms->caller();
 994   }
 995 
 996   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
 997 
 998   // Test the correctness of JVMState::debug_xxx accessors:
 999   assert(call->jvms()->debug_start() == non_debug_edges, "");
1000   assert(call->jvms()->debug_end()   == call->req(), "");
1001   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1002 }
1003 
1004 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1005   Bytecodes::Code code = java_bc();
1006   if (code == Bytecodes::_wide) {
1007     code = method()->java_code_at_bci(bci() + 1);
1008   }
1009 
1010   BasicType rtype = T_ILLEGAL;
1011   int       rsize = 0;
1012 
1013   if (code != Bytecodes::_illegal) {
1014     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1015     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1016     if (rtype < T_CONFLICT)
1017       rsize = type2size[rtype];
1018   }
1019 
1020   switch (code) {
1021   case Bytecodes::_illegal:
1022     return false;
1023 
1024   case Bytecodes::_ldc:
1025   case Bytecodes::_ldc_w:
1026   case Bytecodes::_ldc2_w:
1027     inputs = 0;
1028     break;
1029 
1030   case Bytecodes::_dup:         inputs = 1;  break;
1031   case Bytecodes::_dup_x1:      inputs = 2;  break;
1032   case Bytecodes::_dup_x2:      inputs = 3;  break;
1033   case Bytecodes::_dup2:        inputs = 2;  break;
1034   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1035   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1036   case Bytecodes::_swap:        inputs = 2;  break;
1037   case Bytecodes::_arraylength: inputs = 1;  break;
1038 
1039   case Bytecodes::_getstatic:
1040   case Bytecodes::_putstatic:
1041   case Bytecodes::_getfield:
1042   case Bytecodes::_putfield:
1043     {
1044       bool ignored_will_link;
1045       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1046       int      size  = field->type()->size();
1047       bool is_get = (depth >= 0), is_static = (depth & 1);
1048       inputs = (is_static ? 0 : 1);
1049       if (is_get) {
1050         depth = size - inputs;
1051       } else {
1052         inputs += size;        // putxxx pops the value from the stack
1053         depth = - inputs;
1054       }
1055     }
1056     break;
1057 
1058   case Bytecodes::_invokevirtual:
1059   case Bytecodes::_invokespecial:
1060   case Bytecodes::_invokestatic:
1061   case Bytecodes::_invokedynamic:
1062   case Bytecodes::_invokeinterface:
1063     {
1064       bool ignored_will_link;
1065       ciSignature* declared_signature = NULL;
1066       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1067       assert(declared_signature != NULL, "cannot be null");
1068       inputs   = declared_signature->arg_size_for_bc(code);
1069       int size = declared_signature->return_type()->size();
1070       depth = size - inputs;
1071     }
1072     break;
1073 
1074   case Bytecodes::_multianewarray:
1075     {
1076       ciBytecodeStream iter(method());
1077       iter.reset_to_bci(bci());
1078       iter.next();
1079       inputs = iter.get_dimensions();
1080       assert(rsize == 1, "");
1081       depth = rsize - inputs;
1082     }
1083     break;
1084 
1085   case Bytecodes::_ireturn:
1086   case Bytecodes::_lreturn:
1087   case Bytecodes::_freturn:
1088   case Bytecodes::_dreturn:
1089   case Bytecodes::_areturn:
1090     assert(rsize == -depth, "");
1091     inputs = rsize;
1092     break;
1093 
1094   case Bytecodes::_jsr:
1095   case Bytecodes::_jsr_w:
1096     inputs = 0;
1097     depth  = 1;                  // S.B. depth=1, not zero
1098     break;
1099 
1100   default:
1101     // bytecode produces a typed result
1102     inputs = rsize - depth;
1103     assert(inputs >= 0, "");
1104     break;
1105   }
1106 
1107 #ifdef ASSERT
1108   // spot check
1109   int outputs = depth + inputs;
1110   assert(outputs >= 0, "sanity");
1111   switch (code) {
1112   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1113   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1114   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1115   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1116   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1117   default:                    break;
1118   }
1119 #endif //ASSERT
1120 
1121   return true;
1122 }
1123 
1124 
1125 
1126 //------------------------------basic_plus_adr---------------------------------
1127 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1128   // short-circuit a common case
1129   if (offset == intcon(0))  return ptr;
1130   return _gvn.transform( new AddPNode(base, ptr, offset) );
1131 }
1132 
1133 Node* GraphKit::ConvI2L(Node* offset) {
1134   // short-circuit a common case
1135   jint offset_con = find_int_con(offset, Type::OffsetBot);
1136   if (offset_con != Type::OffsetBot) {
1137     return longcon((jlong) offset_con);
1138   }
1139   return _gvn.transform( new ConvI2LNode(offset));
1140 }
1141 
1142 Node* GraphKit::ConvI2UL(Node* offset) {
1143   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1144   if (offset_con != (juint) Type::OffsetBot) {
1145     return longcon((julong) offset_con);
1146   }
1147   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1148   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1149   return _gvn.transform( new AndLNode(conv, mask) );
1150 }
1151 
1152 Node* GraphKit::ConvL2I(Node* offset) {
1153   // short-circuit a common case
1154   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1155   if (offset_con != (jlong)Type::OffsetBot) {
1156     return intcon((int) offset_con);
1157   }
1158   return _gvn.transform( new ConvL2INode(offset));
1159 }
1160 
1161 //-------------------------load_object_klass-----------------------------------
1162 Node* GraphKit::load_object_klass(Node* obj) {
1163   // Special-case a fresh allocation to avoid building nodes:
1164   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1165   if (akls != NULL)  return akls;
1166   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1167   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1168 }
1169 
1170 //-------------------------load_array_length-----------------------------------
1171 Node* GraphKit::load_array_length(Node* array) {
1172   // Special-case a fresh allocation to avoid building nodes:
1173   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1174   Node *alen;
1175   if (alloc == NULL) {
1176     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1177     alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1178   } else {
1179     alen = alloc->Ideal_length();
1180     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1181     if (ccast != alen) {
1182       alen = _gvn.transform(ccast);
1183     }
1184   }
1185   return alen;
1186 }
1187 
1188 //------------------------------do_null_check----------------------------------
1189 // Helper function to do a NULL pointer check.  Returned value is
1190 // the incoming address with NULL casted away.  You are allowed to use the
1191 // not-null value only if you are control dependent on the test.
1192 #ifndef PRODUCT
1193 extern int explicit_null_checks_inserted,
1194            explicit_null_checks_elided;
1195 #endif
1196 Node* GraphKit::null_check_common(Node* value, BasicType type,
1197                                   // optional arguments for variations:
1198                                   bool assert_null,
1199                                   Node* *null_control,
1200                                   bool speculative) {
1201   assert(!assert_null || null_control == NULL, "not both at once");
1202   if (stopped())  return top();
1203   NOT_PRODUCT(explicit_null_checks_inserted++);
1204 
1205   // Construct NULL check
1206   Node *chk = NULL;
1207   switch(type) {
1208     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1209     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1210     case T_ARRAY  : // fall through
1211       type = T_OBJECT;  // simplify further tests
1212     case T_OBJECT : {
1213       const Type *t = _gvn.type( value );
1214 
1215       const TypeOopPtr* tp = t->isa_oopptr();
1216       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1217           // Only for do_null_check, not any of its siblings:
1218           && !assert_null && null_control == NULL) {
1219         // Usually, any field access or invocation on an unloaded oop type
1220         // will simply fail to link, since the statically linked class is
1221         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1222         // the static class is loaded but the sharper oop type is not.
1223         // Rather than checking for this obscure case in lots of places,
1224         // we simply observe that a null check on an unloaded class
1225         // will always be followed by a nonsense operation, so we
1226         // can just issue the uncommon trap here.
1227         // Our access to the unloaded class will only be correct
1228         // after it has been loaded and initialized, which requires
1229         // a trip through the interpreter.
1230 #ifndef PRODUCT
1231         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1232 #endif
1233         uncommon_trap(Deoptimization::Reason_unloaded,
1234                       Deoptimization::Action_reinterpret,
1235                       tp->klass(), "!loaded");
1236         return top();
1237       }
1238 
1239       if (assert_null) {
1240         // See if the type is contained in NULL_PTR.
1241         // If so, then the value is already null.
1242         if (t->higher_equal(TypePtr::NULL_PTR)) {
1243           NOT_PRODUCT(explicit_null_checks_elided++);
1244           return value;           // Elided null assert quickly!
1245         }
1246       } else {
1247         // See if mixing in the NULL pointer changes type.
1248         // If so, then the NULL pointer was not allowed in the original
1249         // type.  In other words, "value" was not-null.
1250         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1251           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1252           NOT_PRODUCT(explicit_null_checks_elided++);
1253           return value;           // Elided null check quickly!
1254         }
1255       }
1256       chk = new CmpPNode( value, null() );
1257       break;
1258     }
1259 
1260     default:
1261       fatal("unexpected type: %s", type2name(type));
1262   }
1263   assert(chk != NULL, "sanity check");
1264   chk = _gvn.transform(chk);
1265 
1266   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1267   BoolNode *btst = new BoolNode( chk, btest);
1268   Node   *tst = _gvn.transform( btst );
1269 
1270   //-----------
1271   // if peephole optimizations occurred, a prior test existed.
1272   // If a prior test existed, maybe it dominates as we can avoid this test.
1273   if (tst != btst && type == T_OBJECT) {
1274     // At this point we want to scan up the CFG to see if we can
1275     // find an identical test (and so avoid this test altogether).
1276     Node *cfg = control();
1277     int depth = 0;
1278     while( depth < 16 ) {       // Limit search depth for speed
1279       if( cfg->Opcode() == Op_IfTrue &&
1280           cfg->in(0)->in(1) == tst ) {
1281         // Found prior test.  Use "cast_not_null" to construct an identical
1282         // CastPP (and hence hash to) as already exists for the prior test.
1283         // Return that casted value.
1284         if (assert_null) {
1285           replace_in_map(value, null());
1286           return null();  // do not issue the redundant test
1287         }
1288         Node *oldcontrol = control();
1289         set_control(cfg);
1290         Node *res = cast_not_null(value);
1291         set_control(oldcontrol);
1292         NOT_PRODUCT(explicit_null_checks_elided++);
1293         return res;
1294       }
1295       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1296       if (cfg == NULL)  break;  // Quit at region nodes
1297       depth++;
1298     }
1299   }
1300 
1301   //-----------
1302   // Branch to failure if null
1303   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1304   Deoptimization::DeoptReason reason;
1305   if (assert_null) {
1306     reason = Deoptimization::reason_null_assert(speculative);
1307   } else if (type == T_OBJECT) {
1308     reason = Deoptimization::reason_null_check(speculative);
1309   } else {
1310     reason = Deoptimization::Reason_div0_check;
1311   }
1312   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1313   // ciMethodData::has_trap_at will return a conservative -1 if any
1314   // must-be-null assertion has failed.  This could cause performance
1315   // problems for a method after its first do_null_assert failure.
1316   // Consider using 'Reason_class_check' instead?
1317 
1318   // To cause an implicit null check, we set the not-null probability
1319   // to the maximum (PROB_MAX).  For an explicit check the probability
1320   // is set to a smaller value.
1321   if (null_control != NULL || too_many_traps(reason)) {
1322     // probability is less likely
1323     ok_prob =  PROB_LIKELY_MAG(3);
1324   } else if (!assert_null &&
1325              (ImplicitNullCheckThreshold > 0) &&
1326              method() != NULL &&
1327              (method()->method_data()->trap_count(reason)
1328               >= (uint)ImplicitNullCheckThreshold)) {
1329     ok_prob =  PROB_LIKELY_MAG(3);
1330   }
1331 
1332   if (null_control != NULL) {
1333     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1334     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1335     set_control(      _gvn.transform( new IfTrueNode(iff)));
1336 #ifndef PRODUCT
1337     if (null_true == top()) {
1338       explicit_null_checks_elided++;
1339     }
1340 #endif
1341     (*null_control) = null_true;
1342   } else {
1343     BuildCutout unless(this, tst, ok_prob);
1344     // Check for optimizer eliding test at parse time
1345     if (stopped()) {
1346       // Failure not possible; do not bother making uncommon trap.
1347       NOT_PRODUCT(explicit_null_checks_elided++);
1348     } else if (assert_null) {
1349       uncommon_trap(reason,
1350                     Deoptimization::Action_make_not_entrant,
1351                     NULL, "assert_null");
1352     } else {
1353       replace_in_map(value, zerocon(type));
1354       builtin_throw(reason);
1355     }
1356   }
1357 
1358   // Must throw exception, fall-thru not possible?
1359   if (stopped()) {
1360     return top();               // No result
1361   }
1362 
1363   if (assert_null) {
1364     // Cast obj to null on this path.
1365     replace_in_map(value, zerocon(type));
1366     return zerocon(type);
1367   }
1368 
1369   // Cast obj to not-null on this path, if there is no null_control.
1370   // (If there is a null_control, a non-null value may come back to haunt us.)
1371   if (type == T_OBJECT) {
1372     Node* cast = cast_not_null(value, false);
1373     if (null_control == NULL || (*null_control) == top())
1374       replace_in_map(value, cast);
1375     value = cast;
1376   }
1377 
1378   return value;
1379 }
1380 
1381 
1382 //------------------------------cast_not_null----------------------------------
1383 // Cast obj to not-null on this path
1384 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1385   const Type *t = _gvn.type(obj);
1386   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1387   // Object is already not-null?
1388   if( t == t_not_null ) return obj;
1389 
1390   Node *cast = new CastPPNode(obj,t_not_null);
1391   cast->init_req(0, control());
1392   cast = _gvn.transform( cast );
1393 
1394   // Scan for instances of 'obj' in the current JVM mapping.
1395   // These instances are known to be not-null after the test.
1396   if (do_replace_in_map)
1397     replace_in_map(obj, cast);
1398 
1399   return cast;                  // Return casted value
1400 }
1401 
1402 // Sometimes in intrinsics, we implicitly know an object is not null
1403 // (there's no actual null check) so we can cast it to not null. In
1404 // the course of optimizations, the input to the cast can become null.
1405 // In that case that data path will die and we need the control path
1406 // to become dead as well to keep the graph consistent. So we have to
1407 // add a check for null for which one branch can't be taken. It uses
1408 // an Opaque4 node that will cause the check to be removed after loop
1409 // opts so the test goes away and the compiled code doesn't execute a
1410 // useless check.
1411 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1412   Node* chk = _gvn.transform(new CmpPNode(value, null()));
1413   Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1414   Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1415   IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1416   _gvn.set_type(iff, iff->Value(&_gvn));
1417   Node *if_f = _gvn.transform(new IfFalseNode(iff));
1418   Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1419   Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1420   C->root()->add_req(halt);
1421   Node *if_t = _gvn.transform(new IfTrueNode(iff));
1422   set_control(if_t);
1423   return cast_not_null(value, do_replace_in_map);
1424 }
1425 
1426 
1427 //--------------------------replace_in_map-------------------------------------
1428 void GraphKit::replace_in_map(Node* old, Node* neww) {
1429   if (old == neww) {
1430     return;
1431   }
1432 
1433   map()->replace_edge(old, neww);
1434 
1435   // Note: This operation potentially replaces any edge
1436   // on the map.  This includes locals, stack, and monitors
1437   // of the current (innermost) JVM state.
1438 
1439   // don't let inconsistent types from profiling escape this
1440   // method
1441 
1442   const Type* told = _gvn.type(old);
1443   const Type* tnew = _gvn.type(neww);
1444 
1445   if (!tnew->higher_equal(told)) {
1446     return;
1447   }
1448 
1449   map()->record_replaced_node(old, neww);
1450 }
1451 
1452 
1453 //=============================================================================
1454 //--------------------------------memory---------------------------------------
1455 Node* GraphKit::memory(uint alias_idx) {
1456   MergeMemNode* mem = merged_memory();
1457   Node* p = mem->memory_at(alias_idx);
1458   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1459   return p;
1460 }
1461 
1462 //-----------------------------reset_memory------------------------------------
1463 Node* GraphKit::reset_memory() {
1464   Node* mem = map()->memory();
1465   // do not use this node for any more parsing!
1466   debug_only( map()->set_memory((Node*)NULL) );
1467   return _gvn.transform( mem );
1468 }
1469 
1470 //------------------------------set_all_memory---------------------------------
1471 void GraphKit::set_all_memory(Node* newmem) {
1472   Node* mergemem = MergeMemNode::make(newmem);
1473   gvn().set_type_bottom(mergemem);
1474   map()->set_memory(mergemem);
1475 }
1476 
1477 //------------------------------set_all_memory_call----------------------------
1478 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1479   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1480   set_all_memory(newmem);
1481 }
1482 
1483 //=============================================================================
1484 //
1485 // parser factory methods for MemNodes
1486 //
1487 // These are layered on top of the factory methods in LoadNode and StoreNode,
1488 // and integrate with the parser's memory state and _gvn engine.
1489 //
1490 
1491 // factory methods in "int adr_idx"
1492 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1493                           int adr_idx,
1494                           MemNode::MemOrd mo,
1495                           LoadNode::ControlDependency control_dependency,
1496                           bool require_atomic_access,
1497                           bool unaligned,
1498                           bool mismatched) {
1499   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1500   const TypePtr* adr_type = NULL; // debug-mode-only argument
1501   debug_only(adr_type = C->get_adr_type(adr_idx));
1502   Node* mem = memory(adr_idx);
1503   Node* ld;
1504   if (require_atomic_access && bt == T_LONG) {
1505     ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1506   } else if (require_atomic_access && bt == T_DOUBLE) {
1507     ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1508   } else {
1509     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1510   }
1511   ld = _gvn.transform(ld);
1512   if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1513     // Improve graph before escape analysis and boxing elimination.
1514     record_for_igvn(ld);
1515   }
1516   return ld;
1517 }
1518 
1519 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1520                                 int adr_idx,
1521                                 MemNode::MemOrd mo,
1522                                 bool require_atomic_access,
1523                                 bool unaligned,
1524                                 bool mismatched) {
1525   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1526   const TypePtr* adr_type = NULL;
1527   debug_only(adr_type = C->get_adr_type(adr_idx));
1528   Node *mem = memory(adr_idx);
1529   Node* st;
1530   if (require_atomic_access && bt == T_LONG) {
1531     st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1532   } else if (require_atomic_access && bt == T_DOUBLE) {
1533     st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1534   } else {
1535     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1536   }
1537   if (unaligned) {
1538     st->as_Store()->set_unaligned_access();
1539   }
1540   if (mismatched) {
1541     st->as_Store()->set_mismatched_access();
1542   }
1543   st = _gvn.transform(st);
1544   set_memory(st, adr_idx);
1545   // Back-to-back stores can only remove intermediate store with DU info
1546   // so push on worklist for optimizer.
1547   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1548     record_for_igvn(st);
1549 
1550   return st;
1551 }
1552 
1553 
1554 void GraphKit::pre_barrier(bool do_load,
1555                            Node* ctl,
1556                            Node* obj,
1557                            Node* adr,
1558                            uint  adr_idx,
1559                            Node* val,
1560                            const TypeOopPtr* val_type,
1561                            Node* pre_val,
1562                            BasicType bt) {
1563 
1564   BarrierSet* bs = BarrierSet::barrier_set();
1565   set_control(ctl);
1566   switch (bs->kind()) {
1567 
1568 #if INCLUDE_G1GC
1569     case BarrierSet::G1BarrierSet:
1570       g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1571       break;
1572 #endif
1573 
1574     case BarrierSet::CardTableBarrierSet:
1575       break;
1576 
1577     default      :
1578       ShouldNotReachHere();
1579 
1580   }
1581 }
1582 
1583 bool GraphKit::can_move_pre_barrier() const {
1584   BarrierSet* bs = BarrierSet::barrier_set();
1585   switch (bs->kind()) {
1586 
1587 #if INCLUDE_G1GC
1588     case BarrierSet::G1BarrierSet:
1589       return true; // Can move it if no safepoint
1590 #endif
1591 
1592     case BarrierSet::CardTableBarrierSet:
1593       return true; // There is no pre-barrier
1594 
1595     default      :
1596       ShouldNotReachHere();
1597   }
1598   return false;
1599 }
1600 
1601 void GraphKit::post_barrier(Node* ctl,
1602                             Node* store,
1603                             Node* obj,
1604                             Node* adr,
1605                             uint  adr_idx,
1606                             Node* val,
1607                             BasicType bt,
1608                             bool use_precise) {
1609   BarrierSet* bs = BarrierSet::barrier_set();
1610   set_control(ctl);
1611   switch (bs->kind()) {
1612 #if INCLUDE_G1GC
1613     case BarrierSet::G1BarrierSet:
1614       g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1615       break;
1616 #endif
1617 
1618     case BarrierSet::CardTableBarrierSet:
1619       write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1620       break;
1621 
1622     default      :
1623       ShouldNotReachHere();
1624 
1625   }
1626 }
1627 
1628 Node* GraphKit::store_oop(Node* ctl,
1629                           Node* obj,
1630                           Node* adr,
1631                           const TypePtr* adr_type,
1632                           Node* val,
1633                           const TypeOopPtr* val_type,
1634                           BasicType bt,
1635                           bool use_precise,
1636                           MemNode::MemOrd mo,
1637                           bool mismatched) {
1638   // Transformation of a value which could be NULL pointer (CastPP #NULL)
1639   // could be delayed during Parse (for example, in adjust_map_after_if()).
1640   // Execute transformation here to avoid barrier generation in such case.
1641   if (_gvn.type(val) == TypePtr::NULL_PTR)
1642     val = _gvn.makecon(TypePtr::NULL_PTR);
1643 
1644   set_control(ctl);
1645   if (stopped()) return top(); // Dead path ?
1646 
1647   assert(bt == T_OBJECT, "sanity");
1648   assert(val != NULL, "not dead path");
1649   uint adr_idx = C->get_alias_index(adr_type);
1650   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1651 
1652   pre_barrier(true /* do_load */,
1653               control(), obj, adr, adr_idx, val, val_type,
1654               NULL /* pre_val */,
1655               bt);
1656 
1657   Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1658   post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1659   return store;
1660 }
1661 
1662 // Could be an array or object we don't know at compile time (unsafe ref.)
1663 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1664                              Node* obj,   // containing obj
1665                              Node* adr,  // actual adress to store val at
1666                              const TypePtr* adr_type,
1667                              Node* val,
1668                              BasicType bt,
1669                              MemNode::MemOrd mo,
1670                              bool mismatched) {
1671   Compile::AliasType* at = C->alias_type(adr_type);
1672   const TypeOopPtr* val_type = NULL;
1673   if (adr_type->isa_instptr()) {
1674     if (at->field() != NULL) {
1675       // known field.  This code is a copy of the do_put_xxx logic.
1676       ciField* field = at->field();
1677       if (!field->type()->is_loaded()) {
1678         val_type = TypeInstPtr::BOTTOM;
1679       } else {
1680         val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1681       }
1682     }
1683   } else if (adr_type->isa_aryptr()) {
1684     val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1685   }
1686   if (val_type == NULL) {
1687     val_type = TypeInstPtr::BOTTOM;
1688   }
1689   return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1690 }
1691 
1692 
1693 //-------------------------array_element_address-------------------------
1694 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1695                                       const TypeInt* sizetype, Node* ctrl) {
1696   uint shift  = exact_log2(type2aelembytes(elembt));
1697   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1698 
1699   // short-circuit a common case (saves lots of confusing waste motion)
1700   jint idx_con = find_int_con(idx, -1);
1701   if (idx_con >= 0) {
1702     intptr_t offset = header + ((intptr_t)idx_con << shift);
1703     return basic_plus_adr(ary, offset);
1704   }
1705 
1706   // must be correct type for alignment purposes
1707   Node* base  = basic_plus_adr(ary, header);
1708   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1709   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1710   return basic_plus_adr(ary, base, scale);
1711 }
1712 
1713 //-------------------------load_array_element-------------------------
1714 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1715   const Type* elemtype = arytype->elem();
1716   BasicType elembt = elemtype->array_element_basic_type();
1717   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1718   if (elembt == T_NARROWOOP) {
1719     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1720   }
1721   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1722   return ld;
1723 }
1724 
1725 //-------------------------set_arguments_for_java_call-------------------------
1726 // Arguments (pre-popped from the stack) are taken from the JVMS.
1727 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1728   // Add the call arguments:
1729   uint nargs = call->method()->arg_size();
1730   for (uint i = 0; i < nargs; i++) {
1731     Node* arg = argument(i);
1732     call->init_req(i + TypeFunc::Parms, arg);
1733   }
1734 }
1735 
1736 //---------------------------set_edges_for_java_call---------------------------
1737 // Connect a newly created call into the current JVMS.
1738 // A return value node (if any) is returned from set_edges_for_java_call.
1739 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1740 
1741   // Add the predefined inputs:
1742   call->init_req( TypeFunc::Control, control() );
1743   call->init_req( TypeFunc::I_O    , i_o() );
1744   call->init_req( TypeFunc::Memory , reset_memory() );
1745   call->init_req( TypeFunc::FramePtr, frameptr() );
1746   call->init_req( TypeFunc::ReturnAdr, top() );
1747 
1748   add_safepoint_edges(call, must_throw);
1749 
1750   Node* xcall = _gvn.transform(call);
1751 
1752   if (xcall == top()) {
1753     set_control(top());
1754     return;
1755   }
1756   assert(xcall == call, "call identity is stable");
1757 
1758   // Re-use the current map to produce the result.
1759 
1760   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1761   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1762   set_all_memory_call(xcall, separate_io_proj);
1763 
1764   //return xcall;   // no need, caller already has it
1765 }
1766 
1767 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1768   if (stopped())  return top();  // maybe the call folded up?
1769 
1770   // Capture the return value, if any.
1771   Node* ret;
1772   if (call->method() == NULL ||
1773       call->method()->return_type()->basic_type() == T_VOID)
1774         ret = top();
1775   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1776 
1777   // Note:  Since any out-of-line call can produce an exception,
1778   // we always insert an I_O projection from the call into the result.
1779 
1780   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1781 
1782   if (separate_io_proj) {
1783     // The caller requested separate projections be used by the fall
1784     // through and exceptional paths, so replace the projections for
1785     // the fall through path.
1786     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1787     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1788   }
1789   return ret;
1790 }
1791 
1792 //--------------------set_predefined_input_for_runtime_call--------------------
1793 // Reading and setting the memory state is way conservative here.
1794 // The real problem is that I am not doing real Type analysis on memory,
1795 // so I cannot distinguish card mark stores from other stores.  Across a GC
1796 // point the Store Barrier and the card mark memory has to agree.  I cannot
1797 // have a card mark store and its barrier split across the GC point from
1798 // either above or below.  Here I get that to happen by reading ALL of memory.
1799 // A better answer would be to separate out card marks from other memory.
1800 // For now, return the input memory state, so that it can be reused
1801 // after the call, if this call has restricted memory effects.
1802 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1803   // Set fixed predefined input arguments
1804   Node* memory = reset_memory();
1805   call->init_req( TypeFunc::Control,   control()  );
1806   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1807   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1808   call->init_req( TypeFunc::FramePtr,  frameptr() );
1809   call->init_req( TypeFunc::ReturnAdr, top()      );
1810   return memory;
1811 }
1812 
1813 //-------------------set_predefined_output_for_runtime_call--------------------
1814 // Set control and memory (not i_o) from the call.
1815 // If keep_mem is not NULL, use it for the output state,
1816 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1817 // If hook_mem is NULL, this call produces no memory effects at all.
1818 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1819 // then only that memory slice is taken from the call.
1820 // In the last case, we must put an appropriate memory barrier before
1821 // the call, so as to create the correct anti-dependencies on loads
1822 // preceding the call.
1823 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1824                                                       Node* keep_mem,
1825                                                       const TypePtr* hook_mem) {
1826   // no i/o
1827   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1828   if (keep_mem) {
1829     // First clone the existing memory state
1830     set_all_memory(keep_mem);
1831     if (hook_mem != NULL) {
1832       // Make memory for the call
1833       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1834       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1835       // We also use hook_mem to extract specific effects from arraycopy stubs.
1836       set_memory(mem, hook_mem);
1837     }
1838     // ...else the call has NO memory effects.
1839 
1840     // Make sure the call advertises its memory effects precisely.
1841     // This lets us build accurate anti-dependences in gcm.cpp.
1842     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1843            "call node must be constructed correctly");
1844   } else {
1845     assert(hook_mem == NULL, "");
1846     // This is not a "slow path" call; all memory comes from the call.
1847     set_all_memory_call(call);
1848   }
1849 }
1850 
1851 
1852 // Replace the call with the current state of the kit.
1853 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1854   JVMState* ejvms = NULL;
1855   if (has_exceptions()) {
1856     ejvms = transfer_exceptions_into_jvms();
1857   }
1858 
1859   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1860   ReplacedNodes replaced_nodes_exception;
1861   Node* ex_ctl = top();
1862 
1863   SafePointNode* final_state = stop();
1864 
1865   // Find all the needed outputs of this call
1866   CallProjections callprojs;
1867   call->extract_projections(&callprojs, true);
1868 
1869   Node* init_mem = call->in(TypeFunc::Memory);
1870   Node* final_mem = final_state->in(TypeFunc::Memory);
1871   Node* final_ctl = final_state->in(TypeFunc::Control);
1872   Node* final_io = final_state->in(TypeFunc::I_O);
1873 
1874   // Replace all the old call edges with the edges from the inlining result
1875   if (callprojs.fallthrough_catchproj != NULL) {
1876     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1877   }
1878   if (callprojs.fallthrough_memproj != NULL) {
1879     if (final_mem->is_MergeMem()) {
1880       // Parser's exits MergeMem was not transformed but may be optimized
1881       final_mem = _gvn.transform(final_mem);
1882     }
1883     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
1884   }
1885   if (callprojs.fallthrough_ioproj != NULL) {
1886     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
1887   }
1888 
1889   // Replace the result with the new result if it exists and is used
1890   if (callprojs.resproj != NULL && result != NULL) {
1891     C->gvn_replace_by(callprojs.resproj, result);
1892   }
1893 
1894   if (ejvms == NULL) {
1895     // No exception edges to simply kill off those paths
1896     if (callprojs.catchall_catchproj != NULL) {
1897       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1898     }
1899     if (callprojs.catchall_memproj != NULL) {
1900       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1901     }
1902     if (callprojs.catchall_ioproj != NULL) {
1903       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1904     }
1905     // Replace the old exception object with top
1906     if (callprojs.exobj != NULL) {
1907       C->gvn_replace_by(callprojs.exobj, C->top());
1908     }
1909   } else {
1910     GraphKit ekit(ejvms);
1911 
1912     // Load my combined exception state into the kit, with all phis transformed:
1913     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1914     replaced_nodes_exception = ex_map->replaced_nodes();
1915 
1916     Node* ex_oop = ekit.use_exception_state(ex_map);
1917 
1918     if (callprojs.catchall_catchproj != NULL) {
1919       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1920       ex_ctl = ekit.control();
1921     }
1922     if (callprojs.catchall_memproj != NULL) {
1923       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1924     }
1925     if (callprojs.catchall_ioproj != NULL) {
1926       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1927     }
1928 
1929     // Replace the old exception object with the newly created one
1930     if (callprojs.exobj != NULL) {
1931       C->gvn_replace_by(callprojs.exobj, ex_oop);
1932     }
1933   }
1934 
1935   // Disconnect the call from the graph
1936   call->disconnect_inputs(NULL, C);
1937   C->gvn_replace_by(call, C->top());
1938 
1939   // Clean up any MergeMems that feed other MergeMems since the
1940   // optimizer doesn't like that.
1941   if (final_mem->is_MergeMem()) {
1942     Node_List wl;
1943     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1944       Node* m = i.get();
1945       if (m->is_MergeMem() && !wl.contains(m)) {
1946         wl.push(m);
1947       }
1948     }
1949     while (wl.size()  > 0) {
1950       _gvn.transform(wl.pop());
1951     }
1952   }
1953 
1954   if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1955     replaced_nodes.apply(C, final_ctl);
1956   }
1957   if (!ex_ctl->is_top() && do_replaced_nodes) {
1958     replaced_nodes_exception.apply(C, ex_ctl);
1959   }
1960 }
1961 
1962 
1963 //------------------------------increment_counter------------------------------
1964 // for statistics: increment a VM counter by 1
1965 
1966 void GraphKit::increment_counter(address counter_addr) {
1967   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1968   increment_counter(adr1);
1969 }
1970 
1971 void GraphKit::increment_counter(Node* counter_addr) {
1972   int adr_type = Compile::AliasIdxRaw;
1973   Node* ctrl = control();
1974   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1975   Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1976   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1977 }
1978 
1979 
1980 //------------------------------uncommon_trap----------------------------------
1981 // Bail out to the interpreter in mid-method.  Implemented by calling the
1982 // uncommon_trap blob.  This helper function inserts a runtime call with the
1983 // right debug info.
1984 void GraphKit::uncommon_trap(int trap_request,
1985                              ciKlass* klass, const char* comment,
1986                              bool must_throw,
1987                              bool keep_exact_action) {
1988   if (failing())  stop();
1989   if (stopped())  return; // trap reachable?
1990 
1991   // Note:  If ProfileTraps is true, and if a deopt. actually
1992   // occurs here, the runtime will make sure an MDO exists.  There is
1993   // no need to call method()->ensure_method_data() at this point.
1994 
1995   // Set the stack pointer to the right value for reexecution:
1996   set_sp(reexecute_sp());
1997 
1998 #ifdef ASSERT
1999   if (!must_throw) {
2000     // Make sure the stack has at least enough depth to execute
2001     // the current bytecode.
2002     int inputs, ignored_depth;
2003     if (compute_stack_effects(inputs, ignored_depth)) {
2004       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2005              Bytecodes::name(java_bc()), sp(), inputs);
2006     }
2007   }
2008 #endif
2009 
2010   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2011   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2012 
2013   switch (action) {
2014   case Deoptimization::Action_maybe_recompile:
2015   case Deoptimization::Action_reinterpret:
2016     // Temporary fix for 6529811 to allow virtual calls to be sure they
2017     // get the chance to go from mono->bi->mega
2018     if (!keep_exact_action &&
2019         Deoptimization::trap_request_index(trap_request) < 0 &&
2020         too_many_recompiles(reason)) {
2021       // This BCI is causing too many recompilations.
2022       if (C->log() != NULL) {
2023         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2024                 Deoptimization::trap_reason_name(reason),
2025                 Deoptimization::trap_action_name(action));
2026       }
2027       action = Deoptimization::Action_none;
2028       trap_request = Deoptimization::make_trap_request(reason, action);
2029     } else {
2030       C->set_trap_can_recompile(true);
2031     }
2032     break;
2033   case Deoptimization::Action_make_not_entrant:
2034     C->set_trap_can_recompile(true);
2035     break;
2036   case Deoptimization::Action_none:
2037   case Deoptimization::Action_make_not_compilable:
2038     break;
2039   default:
2040 #ifdef ASSERT
2041     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2042 #endif
2043     break;
2044   }
2045 
2046   if (TraceOptoParse) {
2047     char buf[100];
2048     tty->print_cr("Uncommon trap %s at bci:%d",
2049                   Deoptimization::format_trap_request(buf, sizeof(buf),
2050                                                       trap_request), bci());
2051   }
2052 
2053   CompileLog* log = C->log();
2054   if (log != NULL) {
2055     int kid = (klass == NULL)? -1: log->identify(klass);
2056     log->begin_elem("uncommon_trap bci='%d'", bci());
2057     char buf[100];
2058     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2059                                                           trap_request));
2060     if (kid >= 0)         log->print(" klass='%d'", kid);
2061     if (comment != NULL)  log->print(" comment='%s'", comment);
2062     log->end_elem();
2063   }
2064 
2065   // Make sure any guarding test views this path as very unlikely
2066   Node *i0 = control()->in(0);
2067   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2068     IfNode *iff = i0->as_If();
2069     float f = iff->_prob;   // Get prob
2070     if (control()->Opcode() == Op_IfTrue) {
2071       if (f > PROB_UNLIKELY_MAG(4))
2072         iff->_prob = PROB_MIN;
2073     } else {
2074       if (f < PROB_LIKELY_MAG(4))
2075         iff->_prob = PROB_MAX;
2076     }
2077   }
2078 
2079   // Clear out dead values from the debug info.
2080   kill_dead_locals();
2081 
2082   // Now insert the uncommon trap subroutine call
2083   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2084   const TypePtr* no_memory_effects = NULL;
2085   // Pass the index of the class to be loaded
2086   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2087                                  (must_throw ? RC_MUST_THROW : 0),
2088                                  OptoRuntime::uncommon_trap_Type(),
2089                                  call_addr, "uncommon_trap", no_memory_effects,
2090                                  intcon(trap_request));
2091   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2092          "must extract request correctly from the graph");
2093   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2094 
2095   call->set_req(TypeFunc::ReturnAdr, returnadr());
2096   // The debug info is the only real input to this call.
2097 
2098   // Halt-and-catch fire here.  The above call should never return!
2099   HaltNode* halt = new HaltNode(control(), frameptr());
2100   _gvn.set_type_bottom(halt);
2101   root()->add_req(halt);
2102 
2103   stop_and_kill_map();
2104 }
2105 
2106 
2107 //--------------------------just_allocated_object------------------------------
2108 // Report the object that was just allocated.
2109 // It must be the case that there are no intervening safepoints.
2110 // We use this to determine if an object is so "fresh" that
2111 // it does not require card marks.
2112 Node* GraphKit::just_allocated_object(Node* current_control) {
2113   if (C->recent_alloc_ctl() == current_control)
2114     return C->recent_alloc_obj();
2115   return NULL;
2116 }
2117 
2118 
2119 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2120   // (Note:  TypeFunc::make has a cache that makes this fast.)
2121   const TypeFunc* tf    = TypeFunc::make(dest_method);
2122   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2123   for (int j = 0; j < nargs; j++) {
2124     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2125     if( targ->basic_type() == T_DOUBLE ) {
2126       // If any parameters are doubles, they must be rounded before
2127       // the call, dstore_rounding does gvn.transform
2128       Node *arg = argument(j);
2129       arg = dstore_rounding(arg);
2130       set_argument(j, arg);
2131     }
2132   }
2133 }
2134 
2135 /**
2136  * Record profiling data exact_kls for Node n with the type system so
2137  * that it can propagate it (speculation)
2138  *
2139  * @param n          node that the type applies to
2140  * @param exact_kls  type from profiling
2141  * @param maybe_null did profiling see null?
2142  *
2143  * @return           node with improved type
2144  */
2145 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2146   const Type* current_type = _gvn.type(n);
2147   assert(UseTypeSpeculation, "type speculation must be on");
2148 
2149   const TypePtr* speculative = current_type->speculative();
2150 
2151   // Should the klass from the profile be recorded in the speculative type?
2152   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2153     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2154     const TypeOopPtr* xtype = tklass->as_instance_type();
2155     assert(xtype->klass_is_exact(), "Should be exact");
2156     // Any reason to believe n is not null (from this profiling or a previous one)?
2157     assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2158     const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2159     // record the new speculative type's depth
2160     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2161     speculative = speculative->with_inline_depth(jvms()->depth());
2162   } else if (current_type->would_improve_ptr(ptr_kind)) {
2163     // Profiling report that null was never seen so we can change the
2164     // speculative type to non null ptr.
2165     if (ptr_kind == ProfileAlwaysNull) {
2166       speculative = TypePtr::NULL_PTR;
2167     } else {
2168       assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2169       const TypePtr* ptr = TypePtr::NOTNULL;
2170       if (speculative != NULL) {
2171         speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2172       } else {
2173         speculative = ptr;
2174       }
2175     }
2176   }
2177 
2178   if (speculative != current_type->speculative()) {
2179     // Build a type with a speculative type (what we think we know
2180     // about the type but will need a guard when we use it)
2181     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2182     // We're changing the type, we need a new CheckCast node to carry
2183     // the new type. The new type depends on the control: what
2184     // profiling tells us is only valid from here as far as we can
2185     // tell.
2186     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2187     cast = _gvn.transform(cast);
2188     replace_in_map(n, cast);
2189     n = cast;
2190   }
2191 
2192   return n;
2193 }
2194 
2195 /**
2196  * Record profiling data from receiver profiling at an invoke with the
2197  * type system so that it can propagate it (speculation)
2198  *
2199  * @param n  receiver node
2200  *
2201  * @return   node with improved type
2202  */
2203 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2204   if (!UseTypeSpeculation) {
2205     return n;
2206   }
2207   ciKlass* exact_kls = profile_has_unique_klass();
2208   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2209   if ((java_bc() == Bytecodes::_checkcast ||
2210        java_bc() == Bytecodes::_instanceof ||
2211        java_bc() == Bytecodes::_aastore) &&
2212       method()->method_data()->is_mature()) {
2213     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2214     if (data != NULL) {
2215       if (!data->as_BitData()->null_seen()) {
2216         ptr_kind = ProfileNeverNull;
2217       } else {
2218         assert(data->is_ReceiverTypeData(), "bad profile data type");
2219         ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2220         uint i = 0;
2221         for (; i < call->row_limit(); i++) {
2222           ciKlass* receiver = call->receiver(i);
2223           if (receiver != NULL) {
2224             break;
2225           }
2226         }
2227         ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2228       }
2229     }
2230   }
2231   return record_profile_for_speculation(n, exact_kls, ptr_kind);
2232 }
2233 
2234 /**
2235  * Record profiling data from argument profiling at an invoke with the
2236  * type system so that it can propagate it (speculation)
2237  *
2238  * @param dest_method  target method for the call
2239  * @param bc           what invoke bytecode is this?
2240  */
2241 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2242   if (!UseTypeSpeculation) {
2243     return;
2244   }
2245   const TypeFunc* tf    = TypeFunc::make(dest_method);
2246   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2247   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2248   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2249     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2250     if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2251       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2252       ciKlass* better_type = NULL;
2253       if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2254         record_profile_for_speculation(argument(j), better_type, ptr_kind);
2255       }
2256       i++;
2257     }
2258   }
2259 }
2260 
2261 /**
2262  * Record profiling data from parameter profiling at an invoke with
2263  * the type system so that it can propagate it (speculation)
2264  */
2265 void GraphKit::record_profiled_parameters_for_speculation() {
2266   if (!UseTypeSpeculation) {
2267     return;
2268   }
2269   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2270     if (_gvn.type(local(i))->isa_oopptr()) {
2271       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2272       ciKlass* better_type = NULL;
2273       if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2274         record_profile_for_speculation(local(i), better_type, ptr_kind);
2275       }
2276       j++;
2277     }
2278   }
2279 }
2280 
2281 /**
2282  * Record profiling data from return value profiling at an invoke with
2283  * the type system so that it can propagate it (speculation)
2284  */
2285 void GraphKit::record_profiled_return_for_speculation() {
2286   if (!UseTypeSpeculation) {
2287     return;
2288   }
2289   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2290   ciKlass* better_type = NULL;
2291   if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2292     // If profiling reports a single type for the return value,
2293     // feed it to the type system so it can propagate it as a
2294     // speculative type
2295     record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2296   }
2297 }
2298 
2299 void GraphKit::round_double_result(ciMethod* dest_method) {
2300   // A non-strict method may return a double value which has an extended
2301   // exponent, but this must not be visible in a caller which is 'strict'
2302   // If a strict caller invokes a non-strict callee, round a double result
2303 
2304   BasicType result_type = dest_method->return_type()->basic_type();
2305   assert( method() != NULL, "must have caller context");
2306   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2307     // Destination method's return value is on top of stack
2308     // dstore_rounding() does gvn.transform
2309     Node *result = pop_pair();
2310     result = dstore_rounding(result);
2311     push_pair(result);
2312   }
2313 }
2314 
2315 // rounding for strict float precision conformance
2316 Node* GraphKit::precision_rounding(Node* n) {
2317   return UseStrictFP && _method->flags().is_strict()
2318     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2319     ? _gvn.transform( new RoundFloatNode(0, n) )
2320     : n;
2321 }
2322 
2323 // rounding for strict double precision conformance
2324 Node* GraphKit::dprecision_rounding(Node *n) {
2325   return UseStrictFP && _method->flags().is_strict()
2326     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2327     ? _gvn.transform( new RoundDoubleNode(0, n) )
2328     : n;
2329 }
2330 
2331 // rounding for non-strict double stores
2332 Node* GraphKit::dstore_rounding(Node* n) {
2333   return Matcher::strict_fp_requires_explicit_rounding
2334     && UseSSE <= 1
2335     ? _gvn.transform( new RoundDoubleNode(0, n) )
2336     : n;
2337 }
2338 
2339 //=============================================================================
2340 // Generate a fast path/slow path idiom.  Graph looks like:
2341 // [foo] indicates that 'foo' is a parameter
2342 //
2343 //              [in]     NULL
2344 //                 \    /
2345 //                  CmpP
2346 //                  Bool ne
2347 //                   If
2348 //                  /  \
2349 //              True    False-<2>
2350 //              / |
2351 //             /  cast_not_null
2352 //           Load  |    |   ^
2353 //        [fast_test]   |   |
2354 // gvn to   opt_test    |   |
2355 //          /    \      |  <1>
2356 //      True     False  |
2357 //        |         \\  |
2358 //   [slow_call]     \[fast_result]
2359 //    Ctl   Val       \      \
2360 //     |               \      \
2361 //    Catch       <1>   \      \
2362 //   /    \        ^     \      \
2363 //  Ex    No_Ex    |      \      \
2364 //  |       \   \  |       \ <2>  \
2365 //  ...      \  [slow_res] |  |    \   [null_result]
2366 //            \         \--+--+---  |  |
2367 //             \           | /    \ | /
2368 //              --------Region     Phi
2369 //
2370 //=============================================================================
2371 // Code is structured as a series of driver functions all called 'do_XXX' that
2372 // call a set of helper functions.  Helper functions first, then drivers.
2373 
2374 //------------------------------null_check_oop---------------------------------
2375 // Null check oop.  Set null-path control into Region in slot 3.
2376 // Make a cast-not-nullness use the other not-null control.  Return cast.
2377 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2378                                bool never_see_null,
2379                                bool safe_for_replace,
2380                                bool speculative) {
2381   // Initial NULL check taken path
2382   (*null_control) = top();
2383   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2384 
2385   // Generate uncommon_trap:
2386   if (never_see_null && (*null_control) != top()) {
2387     // If we see an unexpected null at a check-cast we record it and force a
2388     // recompile; the offending check-cast will be compiled to handle NULLs.
2389     // If we see more than one offending BCI, then all checkcasts in the
2390     // method will be compiled to handle NULLs.
2391     PreserveJVMState pjvms(this);
2392     set_control(*null_control);
2393     replace_in_map(value, null());
2394     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2395     uncommon_trap(reason,
2396                   Deoptimization::Action_make_not_entrant);
2397     (*null_control) = top();    // NULL path is dead
2398   }
2399   if ((*null_control) == top() && safe_for_replace) {
2400     replace_in_map(value, cast);
2401   }
2402 
2403   // Cast away null-ness on the result
2404   return cast;
2405 }
2406 
2407 //------------------------------opt_iff----------------------------------------
2408 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2409 // Return slow-path control.
2410 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2411   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2412 
2413   // Fast path taken; set region slot 2
2414   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2415   region->init_req(2,fast_taken); // Capture fast-control
2416 
2417   // Fast path not-taken, i.e. slow path
2418   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2419   return slow_taken;
2420 }
2421 
2422 //-----------------------------make_runtime_call-------------------------------
2423 Node* GraphKit::make_runtime_call(int flags,
2424                                   const TypeFunc* call_type, address call_addr,
2425                                   const char* call_name,
2426                                   const TypePtr* adr_type,
2427                                   // The following parms are all optional.
2428                                   // The first NULL ends the list.
2429                                   Node* parm0, Node* parm1,
2430                                   Node* parm2, Node* parm3,
2431                                   Node* parm4, Node* parm5,
2432                                   Node* parm6, Node* parm7) {
2433   // Slow-path call
2434   bool is_leaf = !(flags & RC_NO_LEAF);
2435   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2436   if (call_name == NULL) {
2437     assert(!is_leaf, "must supply name for leaf");
2438     call_name = OptoRuntime::stub_name(call_addr);
2439   }
2440   CallNode* call;
2441   if (!is_leaf) {
2442     call = new CallStaticJavaNode(call_type, call_addr, call_name,
2443                                            bci(), adr_type);
2444   } else if (flags & RC_NO_FP) {
2445     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2446   } else {
2447     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2448   }
2449 
2450   // The following is similar to set_edges_for_java_call,
2451   // except that the memory effects of the call are restricted to AliasIdxRaw.
2452 
2453   // Slow path call has no side-effects, uses few values
2454   bool wide_in  = !(flags & RC_NARROW_MEM);
2455   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2456 
2457   Node* prev_mem = NULL;
2458   if (wide_in) {
2459     prev_mem = set_predefined_input_for_runtime_call(call);
2460   } else {
2461     assert(!wide_out, "narrow in => narrow out");
2462     Node* narrow_mem = memory(adr_type);
2463     prev_mem = reset_memory();
2464     map()->set_memory(narrow_mem);
2465     set_predefined_input_for_runtime_call(call);
2466   }
2467 
2468   // Hook each parm in order.  Stop looking at the first NULL.
2469   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2470   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2471   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2472   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2473   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2474   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2475   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2476   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2477     /* close each nested if ===> */  } } } } } } } }
2478   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2479 
2480   if (!is_leaf) {
2481     // Non-leaves can block and take safepoints:
2482     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2483   }
2484   // Non-leaves can throw exceptions:
2485   if (has_io) {
2486     call->set_req(TypeFunc::I_O, i_o());
2487   }
2488 
2489   if (flags & RC_UNCOMMON) {
2490     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2491     // (An "if" probability corresponds roughly to an unconditional count.
2492     // Sort of.)
2493     call->set_cnt(PROB_UNLIKELY_MAG(4));
2494   }
2495 
2496   Node* c = _gvn.transform(call);
2497   assert(c == call, "cannot disappear");
2498 
2499   if (wide_out) {
2500     // Slow path call has full side-effects.
2501     set_predefined_output_for_runtime_call(call);
2502   } else {
2503     // Slow path call has few side-effects, and/or sets few values.
2504     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2505   }
2506 
2507   if (has_io) {
2508     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2509   }
2510   return call;
2511 
2512 }
2513 
2514 //------------------------------merge_memory-----------------------------------
2515 // Merge memory from one path into the current memory state.
2516 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2517   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2518     Node* old_slice = mms.force_memory();
2519     Node* new_slice = mms.memory2();
2520     if (old_slice != new_slice) {
2521       PhiNode* phi;
2522       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2523         if (mms.is_empty()) {
2524           // clone base memory Phi's inputs for this memory slice
2525           assert(old_slice == mms.base_memory(), "sanity");
2526           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2527           _gvn.set_type(phi, Type::MEMORY);
2528           for (uint i = 1; i < phi->req(); i++) {
2529             phi->init_req(i, old_slice->in(i));
2530           }
2531         } else {
2532           phi = old_slice->as_Phi(); // Phi was generated already
2533         }
2534       } else {
2535         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2536         _gvn.set_type(phi, Type::MEMORY);
2537       }
2538       phi->set_req(new_path, new_slice);
2539       mms.set_memory(phi);
2540     }
2541   }
2542 }
2543 
2544 //------------------------------make_slow_call_ex------------------------------
2545 // Make the exception handler hookups for the slow call
2546 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2547   if (stopped())  return;
2548 
2549   // Make a catch node with just two handlers:  fall-through and catch-all
2550   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2551   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2552   Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2553   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2554 
2555   { PreserveJVMState pjvms(this);
2556     set_control(excp);
2557     set_i_o(i_o);
2558 
2559     if (excp != top()) {
2560       if (deoptimize) {
2561         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2562         uncommon_trap(Deoptimization::Reason_unhandled,
2563                       Deoptimization::Action_none);
2564       } else {
2565         // Create an exception state also.
2566         // Use an exact type if the caller has specified a specific exception.
2567         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2568         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2569         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2570       }
2571     }
2572   }
2573 
2574   // Get the no-exception control from the CatchNode.
2575   set_control(norm);
2576 }
2577 
2578 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2579   Node* cmp = NULL;
2580   switch(bt) {
2581   case T_INT: cmp = new CmpINode(in1, in2); break;
2582   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2583   default: fatal("unexpected comparison type %s", type2name(bt));
2584   }
2585   gvn->transform(cmp);
2586   Node* bol = gvn->transform(new BoolNode(cmp, test));
2587   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2588   gvn->transform(iff);
2589   if (!bol->is_Con()) gvn->record_for_igvn(iff);
2590   return iff;
2591 }
2592 
2593 
2594 //-------------------------------gen_subtype_check-----------------------------
2595 // Generate a subtyping check.  Takes as input the subtype and supertype.
2596 // Returns 2 values: sets the default control() to the true path and returns
2597 // the false path.  Only reads invariant memory; sets no (visible) memory.
2598 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2599 // but that's not exposed to the optimizer.  This call also doesn't take in an
2600 // Object; if you wish to check an Object you need to load the Object's class
2601 // prior to coming here.
2602 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2603   Compile* C = gvn->C;
2604 
2605   if ((*ctrl)->is_top()) {
2606     return C->top();
2607   }
2608 
2609   // Fast check for identical types, perhaps identical constants.
2610   // The types can even be identical non-constants, in cases
2611   // involving Array.newInstance, Object.clone, etc.
2612   if (subklass == superklass)
2613     return C->top();             // false path is dead; no test needed.
2614 
2615   if (gvn->type(superklass)->singleton()) {
2616     ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2617     ciKlass* subk   = gvn->type(subklass)->is_klassptr()->klass();
2618 
2619     // In the common case of an exact superklass, try to fold up the
2620     // test before generating code.  You may ask, why not just generate
2621     // the code and then let it fold up?  The answer is that the generated
2622     // code will necessarily include null checks, which do not always
2623     // completely fold away.  If they are also needless, then they turn
2624     // into a performance loss.  Example:
2625     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2626     // Here, the type of 'fa' is often exact, so the store check
2627     // of fa[1]=x will fold up, without testing the nullness of x.
2628     switch (C->static_subtype_check(superk, subk)) {
2629     case Compile::SSC_always_false:
2630       {
2631         Node* always_fail = *ctrl;
2632         *ctrl = gvn->C->top();
2633         return always_fail;
2634       }
2635     case Compile::SSC_always_true:
2636       return C->top();
2637     case Compile::SSC_easy_test:
2638       {
2639         // Just do a direct pointer compare and be done.
2640         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2641         *ctrl = gvn->transform(new IfTrueNode(iff));
2642         return gvn->transform(new IfFalseNode(iff));
2643       }
2644     case Compile::SSC_full_test:
2645       break;
2646     default:
2647       ShouldNotReachHere();
2648     }
2649   }
2650 
2651   // %%% Possible further optimization:  Even if the superklass is not exact,
2652   // if the subklass is the unique subtype of the superklass, the check
2653   // will always succeed.  We could leave a dependency behind to ensure this.
2654 
2655   // First load the super-klass's check-offset
2656   Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2657   Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2658   Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2659   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2660   bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2661 
2662   // Load from the sub-klass's super-class display list, or a 1-word cache of
2663   // the secondary superclass list, or a failing value with a sentinel offset
2664   // if the super-klass is an interface or exceptionally deep in the Java
2665   // hierarchy and we have to scan the secondary superclass list the hard way.
2666   // Worst-case type is a little odd: NULL is allowed as a result (usually
2667   // klass loads can never produce a NULL).
2668   Node *chk_off_X = chk_off;
2669 #ifdef _LP64
2670   chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2671 #endif
2672   Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2673   // For some types like interfaces the following loadKlass is from a 1-word
2674   // cache which is mutable so can't use immutable memory.  Other
2675   // types load from the super-class display table which is immutable.
2676   m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2677   Node *kmem = might_be_cache ? m : C->immutable_memory();
2678   Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2679 
2680   // Compile speed common case: ARE a subtype and we canNOT fail
2681   if( superklass == nkls )
2682     return C->top();             // false path is dead; no test needed.
2683 
2684   // See if we get an immediate positive hit.  Happens roughly 83% of the
2685   // time.  Test to see if the value loaded just previously from the subklass
2686   // is exactly the superklass.
2687   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2688   Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2689   *ctrl = gvn->transform(new IfFalseNode(iff1));
2690 
2691   // Compile speed common case: Check for being deterministic right now.  If
2692   // chk_off is a constant and not equal to cacheoff then we are NOT a
2693   // subklass.  In this case we need exactly the 1 test above and we can
2694   // return those results immediately.
2695   if (!might_be_cache) {
2696     Node* not_subtype_ctrl = *ctrl;
2697     *ctrl = iftrue1; // We need exactly the 1 test above
2698     return not_subtype_ctrl;
2699   }
2700 
2701   // Gather the various success & failures here
2702   RegionNode *r_ok_subtype = new RegionNode(4);
2703   gvn->record_for_igvn(r_ok_subtype);
2704   RegionNode *r_not_subtype = new RegionNode(3);
2705   gvn->record_for_igvn(r_not_subtype);
2706 
2707   r_ok_subtype->init_req(1, iftrue1);
2708 
2709   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2710   // is roughly 63% of the remaining cases).  Test to see if the loaded
2711   // check-offset points into the subklass display list or the 1-element
2712   // cache.  If it points to the display (and NOT the cache) and the display
2713   // missed then it's not a subtype.
2714   Node *cacheoff = gvn->intcon(cacheoff_con);
2715   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2716   r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2717   *ctrl = gvn->transform(new IfFalseNode(iff2));
2718 
2719   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2720   // No performance impact (too rare) but allows sharing of secondary arrays
2721   // which has some footprint reduction.
2722   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2723   r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2724   *ctrl = gvn->transform(new IfFalseNode(iff3));
2725 
2726   // -- Roads not taken here: --
2727   // We could also have chosen to perform the self-check at the beginning
2728   // of this code sequence, as the assembler does.  This would not pay off
2729   // the same way, since the optimizer, unlike the assembler, can perform
2730   // static type analysis to fold away many successful self-checks.
2731   // Non-foldable self checks work better here in second position, because
2732   // the initial primary superclass check subsumes a self-check for most
2733   // types.  An exception would be a secondary type like array-of-interface,
2734   // which does not appear in its own primary supertype display.
2735   // Finally, we could have chosen to move the self-check into the
2736   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2737   // dependent manner.  But it is worthwhile to have the check here,
2738   // where it can be perhaps be optimized.  The cost in code space is
2739   // small (register compare, branch).
2740 
2741   // Now do a linear scan of the secondary super-klass array.  Again, no real
2742   // performance impact (too rare) but it's gotta be done.
2743   // Since the code is rarely used, there is no penalty for moving it
2744   // out of line, and it can only improve I-cache density.
2745   // The decision to inline or out-of-line this final check is platform
2746   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2747   Node* psc = gvn->transform(
2748     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2749 
2750   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2751   r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2752   r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2753 
2754   // Return false path; set default control to true path.
2755   *ctrl = gvn->transform(r_ok_subtype);
2756   return gvn->transform(r_not_subtype);
2757 }
2758 
2759 // Profile-driven exact type check:
2760 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2761                                     float prob,
2762                                     Node* *casted_receiver) {
2763   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2764   Node* recv_klass = load_object_klass(receiver);
2765   Node* want_klass = makecon(tklass);
2766   Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2767   Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2768   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2769   set_control( _gvn.transform( new IfTrueNode (iff) ));
2770   Node* fail = _gvn.transform( new IfFalseNode(iff) );
2771 
2772   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2773   assert(recv_xtype->klass_is_exact(), "");
2774 
2775   // Subsume downstream occurrences of receiver with a cast to
2776   // recv_xtype, since now we know what the type will be.
2777   Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2778   (*casted_receiver) = _gvn.transform(cast);
2779   // (User must make the replace_in_map call.)
2780 
2781   return fail;
2782 }
2783 
2784 
2785 //------------------------------seems_never_null-------------------------------
2786 // Use null_seen information if it is available from the profile.
2787 // If we see an unexpected null at a type check we record it and force a
2788 // recompile; the offending check will be recompiled to handle NULLs.
2789 // If we see several offending BCIs, then all checks in the
2790 // method will be recompiled.
2791 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2792   speculating = !_gvn.type(obj)->speculative_maybe_null();
2793   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2794   if (UncommonNullCast               // Cutout for this technique
2795       && obj != null()               // And not the -Xcomp stupid case?
2796       && !too_many_traps(reason)
2797       ) {
2798     if (speculating) {
2799       return true;
2800     }
2801     if (data == NULL)
2802       // Edge case:  no mature data.  Be optimistic here.
2803       return true;
2804     // If the profile has not seen a null, assume it won't happen.
2805     assert(java_bc() == Bytecodes::_checkcast ||
2806            java_bc() == Bytecodes::_instanceof ||
2807            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2808     return !data->as_BitData()->null_seen();
2809   }
2810   speculating = false;
2811   return false;
2812 }
2813 
2814 //------------------------maybe_cast_profiled_receiver-------------------------
2815 // If the profile has seen exactly one type, narrow to exactly that type.
2816 // Subsequent type checks will always fold up.
2817 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2818                                              ciKlass* require_klass,
2819                                              ciKlass* spec_klass,
2820                                              bool safe_for_replace) {
2821   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2822 
2823   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2824 
2825   // Make sure we haven't already deoptimized from this tactic.
2826   if (too_many_traps(reason) || too_many_recompiles(reason))
2827     return NULL;
2828 
2829   // (No, this isn't a call, but it's enough like a virtual call
2830   // to use the same ciMethod accessor to get the profile info...)
2831   // If we have a speculative type use it instead of profiling (which
2832   // may not help us)
2833   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2834   if (exact_kls != NULL) {// no cast failures here
2835     if (require_klass == NULL ||
2836         C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2837       // If we narrow the type to match what the type profile sees or
2838       // the speculative type, we can then remove the rest of the
2839       // cast.
2840       // This is a win, even if the exact_kls is very specific,
2841       // because downstream operations, such as method calls,
2842       // will often benefit from the sharper type.
2843       Node* exact_obj = not_null_obj; // will get updated in place...
2844       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2845                                             &exact_obj);
2846       { PreserveJVMState pjvms(this);
2847         set_control(slow_ctl);
2848         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2849       }
2850       if (safe_for_replace) {
2851         replace_in_map(not_null_obj, exact_obj);
2852       }
2853       return exact_obj;
2854     }
2855     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2856   }
2857 
2858   return NULL;
2859 }
2860 
2861 /**
2862  * Cast obj to type and emit guard unless we had too many traps here
2863  * already
2864  *
2865  * @param obj       node being casted
2866  * @param type      type to cast the node to
2867  * @param not_null  true if we know node cannot be null
2868  */
2869 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2870                                         ciKlass* type,
2871                                         bool not_null) {
2872   if (stopped()) {
2873     return obj;
2874   }
2875 
2876   // type == NULL if profiling tells us this object is always null
2877   if (type != NULL) {
2878     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2879     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2880 
2881     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2882         !too_many_traps(class_reason) &&
2883         !too_many_recompiles(class_reason)) {
2884       Node* not_null_obj = NULL;
2885       // not_null is true if we know the object is not null and
2886       // there's no need for a null check
2887       if (!not_null) {
2888         Node* null_ctl = top();
2889         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2890         assert(null_ctl->is_top(), "no null control here");
2891       } else {
2892         not_null_obj = obj;
2893       }
2894 
2895       Node* exact_obj = not_null_obj;
2896       ciKlass* exact_kls = type;
2897       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2898                                             &exact_obj);
2899       {
2900         PreserveJVMState pjvms(this);
2901         set_control(slow_ctl);
2902         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2903       }
2904       replace_in_map(not_null_obj, exact_obj);
2905       obj = exact_obj;
2906     }
2907   } else {
2908     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2909         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2910       Node* exact_obj = null_assert(obj);
2911       replace_in_map(obj, exact_obj);
2912       obj = exact_obj;
2913     }
2914   }
2915   return obj;
2916 }
2917 
2918 //-------------------------------gen_instanceof--------------------------------
2919 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2920 // and the reflective instance-of call.
2921 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2922   kill_dead_locals();           // Benefit all the uncommon traps
2923   assert( !stopped(), "dead parse path should be checked in callers" );
2924   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2925          "must check for not-null not-dead klass in callers");
2926 
2927   // Make the merge point
2928   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2929   RegionNode* region = new RegionNode(PATH_LIMIT);
2930   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
2931   C->set_has_split_ifs(true); // Has chance for split-if optimization
2932 
2933   ciProfileData* data = NULL;
2934   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2935     data = method()->method_data()->bci_to_data(bci());
2936   }
2937   bool speculative_not_null = false;
2938   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2939                          && seems_never_null(obj, data, speculative_not_null));
2940 
2941   // Null check; get casted pointer; set region slot 3
2942   Node* null_ctl = top();
2943   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2944 
2945   // If not_null_obj is dead, only null-path is taken
2946   if (stopped()) {              // Doing instance-of on a NULL?
2947     set_control(null_ctl);
2948     return intcon(0);
2949   }
2950   region->init_req(_null_path, null_ctl);
2951   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2952   if (null_ctl == top()) {
2953     // Do this eagerly, so that pattern matches like is_diamond_phi
2954     // will work even during parsing.
2955     assert(_null_path == PATH_LIMIT-1, "delete last");
2956     region->del_req(_null_path);
2957     phi   ->del_req(_null_path);
2958   }
2959 
2960   // Do we know the type check always succeed?
2961   bool known_statically = false;
2962   if (_gvn.type(superklass)->singleton()) {
2963     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2964     ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2965     if (subk != NULL && subk->is_loaded()) {
2966       int static_res = C->static_subtype_check(superk, subk);
2967       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2968     }
2969   }
2970 
2971   if (!known_statically) {
2972     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2973     // We may not have profiling here or it may not help us. If we
2974     // have a speculative type use it to perform an exact cast.
2975     ciKlass* spec_obj_type = obj_type->speculative_type();
2976     if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2977       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2978       if (stopped()) {            // Profile disagrees with this path.
2979         set_control(null_ctl);    // Null is the only remaining possibility.
2980         return intcon(0);
2981       }
2982       if (cast_obj != NULL) {
2983         not_null_obj = cast_obj;
2984       }
2985     }
2986   }
2987 
2988   // Load the object's klass
2989   Node* obj_klass = load_object_klass(not_null_obj);
2990 
2991   // Generate the subtype check
2992   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2993 
2994   // Plug in the success path to the general merge in slot 1.
2995   region->init_req(_obj_path, control());
2996   phi   ->init_req(_obj_path, intcon(1));
2997 
2998   // Plug in the failing path to the general merge in slot 2.
2999   region->init_req(_fail_path, not_subtype_ctrl);
3000   phi   ->init_req(_fail_path, intcon(0));
3001 
3002   // Return final merged results
3003   set_control( _gvn.transform(region) );
3004   record_for_igvn(region);
3005 
3006   // If we know the type check always succeeds then we don't use the
3007   // profiling data at this bytecode. Don't lose it, feed it to the
3008   // type system as a speculative type.
3009   if (safe_for_replace) {
3010     Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3011     replace_in_map(obj, casted_obj);
3012   }
3013 
3014   return _gvn.transform(phi);
3015 }
3016 
3017 //-------------------------------gen_checkcast---------------------------------
3018 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
3019 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
3020 // uncommon-trap paths work.  Adjust stack after this call.
3021 // If failure_control is supplied and not null, it is filled in with
3022 // the control edge for the cast failure.  Otherwise, an appropriate
3023 // uncommon trap or exception is thrown.
3024 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3025                               Node* *failure_control) {
3026   kill_dead_locals();           // Benefit all the uncommon traps
3027   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3028   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3029 
3030   // Fast cutout:  Check the case that the cast is vacuously true.
3031   // This detects the common cases where the test will short-circuit
3032   // away completely.  We do this before we perform the null check,
3033   // because if the test is going to turn into zero code, we don't
3034   // want a residual null check left around.  (Causes a slowdown,
3035   // for example, in some objArray manipulations, such as a[i]=a[j].)
3036   if (tk->singleton()) {
3037     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3038     if (objtp != NULL && objtp->klass() != NULL) {
3039       switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3040       case Compile::SSC_always_true:
3041         // If we know the type check always succeed then we don't use
3042         // the profiling data at this bytecode. Don't lose it, feed it
3043         // to the type system as a speculative type.
3044         return record_profiled_receiver_for_speculation(obj);
3045       case Compile::SSC_always_false:
3046         // It needs a null check because a null will *pass* the cast check.
3047         // A non-null value will always produce an exception.
3048         return null_assert(obj);
3049       }
3050     }
3051   }
3052 
3053   ciProfileData* data = NULL;
3054   bool safe_for_replace = false;
3055   if (failure_control == NULL) {        // use MDO in regular case only
3056     assert(java_bc() == Bytecodes::_aastore ||
3057            java_bc() == Bytecodes::_checkcast,
3058            "interpreter profiles type checks only for these BCs");
3059     data = method()->method_data()->bci_to_data(bci());
3060     safe_for_replace = true;
3061   }
3062 
3063   // Make the merge point
3064   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3065   RegionNode* region = new RegionNode(PATH_LIMIT);
3066   Node*       phi    = new PhiNode(region, toop);
3067   C->set_has_split_ifs(true); // Has chance for split-if optimization
3068 
3069   // Use null-cast information if it is available
3070   bool speculative_not_null = false;
3071   bool never_see_null = ((failure_control == NULL)  // regular case only
3072                          && seems_never_null(obj, data, speculative_not_null));
3073 
3074   // Null check; get casted pointer; set region slot 3
3075   Node* null_ctl = top();
3076   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3077 
3078   // If not_null_obj is dead, only null-path is taken
3079   if (stopped()) {              // Doing instance-of on a NULL?
3080     set_control(null_ctl);
3081     return null();
3082   }
3083   region->init_req(_null_path, null_ctl);
3084   phi   ->init_req(_null_path, null());  // Set null path value
3085   if (null_ctl == top()) {
3086     // Do this eagerly, so that pattern matches like is_diamond_phi
3087     // will work even during parsing.
3088     assert(_null_path == PATH_LIMIT-1, "delete last");
3089     region->del_req(_null_path);
3090     phi   ->del_req(_null_path);
3091   }
3092 
3093   Node* cast_obj = NULL;
3094   if (tk->klass_is_exact()) {
3095     // The following optimization tries to statically cast the speculative type of the object
3096     // (for example obtained during profiling) to the type of the superklass and then do a
3097     // dynamic check that the type of the object is what we expect. To work correctly
3098     // for checkcast and aastore the type of superklass should be exact.
3099     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3100     // We may not have profiling here or it may not help us. If we have
3101     // a speculative type use it to perform an exact cast.
3102     ciKlass* spec_obj_type = obj_type->speculative_type();
3103     if (spec_obj_type != NULL || data != NULL) {
3104       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3105       if (cast_obj != NULL) {
3106         if (failure_control != NULL) // failure is now impossible
3107           (*failure_control) = top();
3108         // adjust the type of the phi to the exact klass:
3109         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3110       }
3111     }
3112   }
3113 
3114   if (cast_obj == NULL) {
3115     // Load the object's klass
3116     Node* obj_klass = load_object_klass(not_null_obj);
3117 
3118     // Generate the subtype check
3119     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3120 
3121     // Plug in success path into the merge
3122     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3123     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3124     if (failure_control == NULL) {
3125       if (not_subtype_ctrl != top()) { // If failure is possible
3126         PreserveJVMState pjvms(this);
3127         set_control(not_subtype_ctrl);
3128         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3129       }
3130     } else {
3131       (*failure_control) = not_subtype_ctrl;
3132     }
3133   }
3134 
3135   region->init_req(_obj_path, control());
3136   phi   ->init_req(_obj_path, cast_obj);
3137 
3138   // A merge of NULL or Casted-NotNull obj
3139   Node* res = _gvn.transform(phi);
3140 
3141   // Note I do NOT always 'replace_in_map(obj,result)' here.
3142   //  if( tk->klass()->can_be_primary_super()  )
3143     // This means that if I successfully store an Object into an array-of-String
3144     // I 'forget' that the Object is really now known to be a String.  I have to
3145     // do this because we don't have true union types for interfaces - if I store
3146     // a Baz into an array-of-Interface and then tell the optimizer it's an
3147     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3148     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3149   //  replace_in_map( obj, res );
3150 
3151   // Return final merged results
3152   set_control( _gvn.transform(region) );
3153   record_for_igvn(region);
3154 
3155   return record_profiled_receiver_for_speculation(res);
3156 }
3157 
3158 //------------------------------next_monitor-----------------------------------
3159 // What number should be given to the next monitor?
3160 int GraphKit::next_monitor() {
3161   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3162   int next = current + C->sync_stack_slots();
3163   // Keep the toplevel high water mark current:
3164   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3165   return current;
3166 }
3167 
3168 //------------------------------insert_mem_bar---------------------------------
3169 // Memory barrier to avoid floating things around
3170 // The membar serves as a pinch point between both control and all memory slices.
3171 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3172   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3173   mb->init_req(TypeFunc::Control, control());
3174   mb->init_req(TypeFunc::Memory,  reset_memory());
3175   Node* membar = _gvn.transform(mb);
3176   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3177   set_all_memory_call(membar);
3178   return membar;
3179 }
3180 
3181 //-------------------------insert_mem_bar_volatile----------------------------
3182 // Memory barrier to avoid floating things around
3183 // The membar serves as a pinch point between both control and memory(alias_idx).
3184 // If you want to make a pinch point on all memory slices, do not use this
3185 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3186 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3187   // When Parse::do_put_xxx updates a volatile field, it appends a series
3188   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3189   // The first membar is on the same memory slice as the field store opcode.
3190   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3191   // All the other membars (for other volatile slices, including AliasIdxBot,
3192   // which stands for all unknown volatile slices) are control-dependent
3193   // on the first membar.  This prevents later volatile loads or stores
3194   // from sliding up past the just-emitted store.
3195 
3196   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3197   mb->set_req(TypeFunc::Control,control());
3198   if (alias_idx == Compile::AliasIdxBot) {
3199     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3200   } else {
3201     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3202     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3203   }
3204   Node* membar = _gvn.transform(mb);
3205   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3206   if (alias_idx == Compile::AliasIdxBot) {
3207     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3208   } else {
3209     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3210   }
3211   return membar;
3212 }
3213 
3214 //------------------------------shared_lock------------------------------------
3215 // Emit locking code.
3216 FastLockNode* GraphKit::shared_lock(Node* obj) {
3217   // bci is either a monitorenter bc or InvocationEntryBci
3218   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3219   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3220 
3221   if( !GenerateSynchronizationCode )
3222     return NULL;                // Not locking things?
3223   if (stopped())                // Dead monitor?
3224     return NULL;
3225 
3226   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3227 
3228   // Box the stack location
3229   Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3230   Node* mem = reset_memory();
3231 
3232   FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3233   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3234     // Create the counters for this fast lock.
3235     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3236   }
3237 
3238   // Create the rtm counters for this fast lock if needed.
3239   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3240 
3241   // Add monitor to debug info for the slow path.  If we block inside the
3242   // slow path and de-opt, we need the monitor hanging around
3243   map()->push_monitor( flock );
3244 
3245   const TypeFunc *tf = LockNode::lock_type();
3246   LockNode *lock = new LockNode(C, tf);
3247 
3248   lock->init_req( TypeFunc::Control, control() );
3249   lock->init_req( TypeFunc::Memory , mem );
3250   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3251   lock->init_req( TypeFunc::FramePtr, frameptr() );
3252   lock->init_req( TypeFunc::ReturnAdr, top() );
3253 
3254   lock->init_req(TypeFunc::Parms + 0, obj);
3255   lock->init_req(TypeFunc::Parms + 1, box);
3256   lock->init_req(TypeFunc::Parms + 2, flock);
3257   add_safepoint_edges(lock);
3258 
3259   lock = _gvn.transform( lock )->as_Lock();
3260 
3261   // lock has no side-effects, sets few values
3262   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3263 
3264   insert_mem_bar(Op_MemBarAcquireLock);
3265 
3266   // Add this to the worklist so that the lock can be eliminated
3267   record_for_igvn(lock);
3268 
3269 #ifndef PRODUCT
3270   if (PrintLockStatistics) {
3271     // Update the counter for this lock.  Don't bother using an atomic
3272     // operation since we don't require absolute accuracy.
3273     lock->create_lock_counter(map()->jvms());
3274     increment_counter(lock->counter()->addr());
3275   }
3276 #endif
3277 
3278   return flock;
3279 }
3280 
3281 
3282 //------------------------------shared_unlock----------------------------------
3283 // Emit unlocking code.
3284 void GraphKit::shared_unlock(Node* box, Node* obj) {
3285   // bci is either a monitorenter bc or InvocationEntryBci
3286   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3287   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3288 
3289   if( !GenerateSynchronizationCode )
3290     return;
3291   if (stopped()) {               // Dead monitor?
3292     map()->pop_monitor();        // Kill monitor from debug info
3293     return;
3294   }
3295 
3296   // Memory barrier to avoid floating things down past the locked region
3297   insert_mem_bar(Op_MemBarReleaseLock);
3298 
3299   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3300   UnlockNode *unlock = new UnlockNode(C, tf);
3301 #ifdef ASSERT
3302   unlock->set_dbg_jvms(sync_jvms());
3303 #endif
3304   uint raw_idx = Compile::AliasIdxRaw;
3305   unlock->init_req( TypeFunc::Control, control() );
3306   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3307   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3308   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3309   unlock->init_req( TypeFunc::ReturnAdr, top() );
3310 
3311   unlock->init_req(TypeFunc::Parms + 0, obj);
3312   unlock->init_req(TypeFunc::Parms + 1, box);
3313   unlock = _gvn.transform(unlock)->as_Unlock();
3314 
3315   Node* mem = reset_memory();
3316 
3317   // unlock has no side-effects, sets few values
3318   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3319 
3320   // Kill monitor from debug info
3321   map()->pop_monitor( );
3322 }
3323 
3324 //-------------------------------get_layout_helper-----------------------------
3325 // If the given klass is a constant or known to be an array,
3326 // fetch the constant layout helper value into constant_value
3327 // and return (Node*)NULL.  Otherwise, load the non-constant
3328 // layout helper value, and return the node which represents it.
3329 // This two-faced routine is useful because allocation sites
3330 // almost always feature constant types.
3331 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3332   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3333   if (!StressReflectiveCode && inst_klass != NULL) {
3334     ciKlass* klass = inst_klass->klass();
3335     bool    xklass = inst_klass->klass_is_exact();
3336     if (xklass || klass->is_array_klass()) {
3337       jint lhelper = klass->layout_helper();
3338       if (lhelper != Klass::_lh_neutral_value) {
3339         constant_value = lhelper;
3340         return (Node*) NULL;
3341       }
3342     }
3343   }
3344   constant_value = Klass::_lh_neutral_value;  // put in a known value
3345   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3346   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3347 }
3348 
3349 // We just put in an allocate/initialize with a big raw-memory effect.
3350 // Hook selected additional alias categories on the initialization.
3351 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3352                                 MergeMemNode* init_in_merge,
3353                                 Node* init_out_raw) {
3354   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3355   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3356 
3357   Node* prevmem = kit.memory(alias_idx);
3358   init_in_merge->set_memory_at(alias_idx, prevmem);
3359   kit.set_memory(init_out_raw, alias_idx);
3360 }
3361 
3362 //---------------------------set_output_for_allocation-------------------------
3363 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3364                                           const TypeOopPtr* oop_type,
3365                                           bool deoptimize_on_exception) {
3366   int rawidx = Compile::AliasIdxRaw;
3367   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3368   add_safepoint_edges(alloc);
3369   Node* allocx = _gvn.transform(alloc);
3370   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3371   // create memory projection for i_o
3372   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3373   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3374 
3375   // create a memory projection as for the normal control path
3376   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3377   set_memory(malloc, rawidx);
3378 
3379   // a normal slow-call doesn't change i_o, but an allocation does
3380   // we create a separate i_o projection for the normal control path
3381   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3382   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3383 
3384   // put in an initialization barrier
3385   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3386                                                  rawoop)->as_Initialize();
3387   assert(alloc->initialization() == init,  "2-way macro link must work");
3388   assert(init ->allocation()     == alloc, "2-way macro link must work");
3389   {
3390     // Extract memory strands which may participate in the new object's
3391     // initialization, and source them from the new InitializeNode.
3392     // This will allow us to observe initializations when they occur,
3393     // and link them properly (as a group) to the InitializeNode.
3394     assert(init->in(InitializeNode::Memory) == malloc, "");
3395     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3396     init->set_req(InitializeNode::Memory, minit_in);
3397     record_for_igvn(minit_in); // fold it up later, if possible
3398     Node* minit_out = memory(rawidx);
3399     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3400     if (oop_type->isa_aryptr()) {
3401       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3402       int            elemidx  = C->get_alias_index(telemref);
3403       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3404     } else if (oop_type->isa_instptr()) {
3405       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3406       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3407         ciField* field = ik->nonstatic_field_at(i);
3408         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3409           continue;  // do not bother to track really large numbers of fields
3410         // Find (or create) the alias category for this field:
3411         int fieldidx = C->alias_type(field)->index();
3412         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3413       }
3414     }
3415   }
3416 
3417   // Cast raw oop to the real thing...
3418   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3419   javaoop = _gvn.transform(javaoop);
3420   C->set_recent_alloc(control(), javaoop);
3421   assert(just_allocated_object(control()) == javaoop, "just allocated");
3422 
3423 #ifdef ASSERT
3424   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3425     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3426            "Ideal_allocation works");
3427     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3428            "Ideal_allocation works");
3429     if (alloc->is_AllocateArray()) {
3430       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3431              "Ideal_allocation works");
3432       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3433              "Ideal_allocation works");
3434     } else {
3435       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3436     }
3437   }
3438 #endif //ASSERT
3439 
3440   return javaoop;
3441 }
3442 
3443 //---------------------------new_instance--------------------------------------
3444 // This routine takes a klass_node which may be constant (for a static type)
3445 // or may be non-constant (for reflective code).  It will work equally well
3446 // for either, and the graph will fold nicely if the optimizer later reduces
3447 // the type to a constant.
3448 // The optional arguments are for specialized use by intrinsics:
3449 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3450 //  - If 'return_size_val', report the the total object size to the caller.
3451 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3452 Node* GraphKit::new_instance(Node* klass_node,
3453                              Node* extra_slow_test,
3454                              Node* *return_size_val,
3455                              bool deoptimize_on_exception) {
3456   // Compute size in doublewords
3457   // The size is always an integral number of doublewords, represented
3458   // as a positive bytewise size stored in the klass's layout_helper.
3459   // The layout_helper also encodes (in a low bit) the need for a slow path.
3460   jint  layout_con = Klass::_lh_neutral_value;
3461   Node* layout_val = get_layout_helper(klass_node, layout_con);
3462   int   layout_is_con = (layout_val == NULL);
3463 
3464   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3465   // Generate the initial go-slow test.  It's either ALWAYS (return a
3466   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3467   // case) a computed value derived from the layout_helper.
3468   Node* initial_slow_test = NULL;
3469   if (layout_is_con) {
3470     assert(!StressReflectiveCode, "stress mode does not use these paths");
3471     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3472     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3473   } else {   // reflective case
3474     // This reflective path is used by Unsafe.allocateInstance.
3475     // (It may be stress-tested by specifying StressReflectiveCode.)
3476     // Basically, we want to get into the VM is there's an illegal argument.
3477     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3478     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3479     if (extra_slow_test != intcon(0)) {
3480       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3481     }
3482     // (Macro-expander will further convert this to a Bool, if necessary.)
3483   }
3484 
3485   // Find the size in bytes.  This is easy; it's the layout_helper.
3486   // The size value must be valid even if the slow path is taken.
3487   Node* size = NULL;
3488   if (layout_is_con) {
3489     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3490   } else {   // reflective case
3491     // This reflective path is used by clone and Unsafe.allocateInstance.
3492     size = ConvI2X(layout_val);
3493 
3494     // Clear the low bits to extract layout_helper_size_in_bytes:
3495     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3496     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3497     size = _gvn.transform( new AndXNode(size, mask) );
3498   }
3499   if (return_size_val != NULL) {
3500     (*return_size_val) = size;
3501   }
3502 
3503   // This is a precise notnull oop of the klass.
3504   // (Actually, it need not be precise if this is a reflective allocation.)
3505   // It's what we cast the result to.
3506   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3507   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3508   const TypeOopPtr* oop_type = tklass->as_instance_type();
3509 
3510   // Now generate allocation code
3511 
3512   // The entire memory state is needed for slow path of the allocation
3513   // since GC and deoptimization can happened.
3514   Node *mem = reset_memory();
3515   set_all_memory(mem); // Create new memory state
3516 
3517   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3518                                          control(), mem, i_o(),
3519                                          size, klass_node,
3520                                          initial_slow_test);
3521 
3522   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3523 }
3524 
3525 //-------------------------------new_array-------------------------------------
3526 // helper for both newarray and anewarray
3527 // The 'length' parameter is (obviously) the length of the array.
3528 // See comments on new_instance for the meaning of the other arguments.
3529 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3530                           Node* length,         // number of array elements
3531                           int   nargs,          // number of arguments to push back for uncommon trap
3532                           Node* *return_size_val,
3533                           bool deoptimize_on_exception) {
3534   jint  layout_con = Klass::_lh_neutral_value;
3535   Node* layout_val = get_layout_helper(klass_node, layout_con);
3536   int   layout_is_con = (layout_val == NULL);
3537 
3538   if (!layout_is_con && !StressReflectiveCode &&
3539       !too_many_traps(Deoptimization::Reason_class_check)) {
3540     // This is a reflective array creation site.
3541     // Optimistically assume that it is a subtype of Object[],
3542     // so that we can fold up all the address arithmetic.
3543     layout_con = Klass::array_layout_helper(T_OBJECT);
3544     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3545     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3546     { BuildCutout unless(this, bol_lh, PROB_MAX);
3547       inc_sp(nargs);
3548       uncommon_trap(Deoptimization::Reason_class_check,
3549                     Deoptimization::Action_maybe_recompile);
3550     }
3551     layout_val = NULL;
3552     layout_is_con = true;
3553   }
3554 
3555   // Generate the initial go-slow test.  Make sure we do not overflow
3556   // if length is huge (near 2Gig) or negative!  We do not need
3557   // exact double-words here, just a close approximation of needed
3558   // double-words.  We can't add any offset or rounding bits, lest we
3559   // take a size -1 of bytes and make it positive.  Use an unsigned
3560   // compare, so negative sizes look hugely positive.
3561   int fast_size_limit = FastAllocateSizeLimit;
3562   if (layout_is_con) {
3563     assert(!StressReflectiveCode, "stress mode does not use these paths");
3564     // Increase the size limit if we have exact knowledge of array type.
3565     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3566     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3567   }
3568 
3569   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3570   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3571 
3572   // --- Size Computation ---
3573   // array_size = round_to_heap(array_header + (length << elem_shift));
3574   // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3575   // and align_to(x, y) == ((x + y-1) & ~(y-1))
3576   // The rounding mask is strength-reduced, if possible.
3577   int round_mask = MinObjAlignmentInBytes - 1;
3578   Node* header_size = NULL;
3579   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3580   // (T_BYTE has the weakest alignment and size restrictions...)
3581   if (layout_is_con) {
3582     int       hsize  = Klass::layout_helper_header_size(layout_con);
3583     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3584     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3585     if ((round_mask & ~right_n_bits(eshift)) == 0)
3586       round_mask = 0;  // strength-reduce it if it goes away completely
3587     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3588     assert(header_size_min <= hsize, "generic minimum is smallest");
3589     header_size_min = hsize;
3590     header_size = intcon(hsize + round_mask);
3591   } else {
3592     Node* hss   = intcon(Klass::_lh_header_size_shift);
3593     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3594     Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3595     hsize       = _gvn.transform( new AndINode(hsize, hsm) );
3596     Node* mask  = intcon(round_mask);
3597     header_size = _gvn.transform( new AddINode(hsize, mask) );
3598   }
3599 
3600   Node* elem_shift = NULL;
3601   if (layout_is_con) {
3602     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3603     if (eshift != 0)
3604       elem_shift = intcon(eshift);
3605   } else {
3606     // There is no need to mask or shift this value.
3607     // The semantics of LShiftINode include an implicit mask to 0x1F.
3608     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3609     elem_shift = layout_val;
3610   }
3611 
3612   // Transition to native address size for all offset calculations:
3613   Node* lengthx = ConvI2X(length);
3614   Node* headerx = ConvI2X(header_size);
3615 #ifdef _LP64
3616   { const TypeInt* tilen = _gvn.find_int_type(length);
3617     if (tilen != NULL && tilen->_lo < 0) {
3618       // Add a manual constraint to a positive range.  Cf. array_element_address.
3619       jint size_max = fast_size_limit;
3620       if (size_max > tilen->_hi)  size_max = tilen->_hi;
3621       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3622 
3623       // Only do a narrow I2L conversion if the range check passed.
3624       IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3625       _gvn.transform(iff);
3626       RegionNode* region = new RegionNode(3);
3627       _gvn.set_type(region, Type::CONTROL);
3628       lengthx = new PhiNode(region, TypeLong::LONG);
3629       _gvn.set_type(lengthx, TypeLong::LONG);
3630 
3631       // Range check passed. Use ConvI2L node with narrow type.
3632       Node* passed = IfFalse(iff);
3633       region->init_req(1, passed);
3634       // Make I2L conversion control dependent to prevent it from
3635       // floating above the range check during loop optimizations.
3636       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3637 
3638       // Range check failed. Use ConvI2L with wide type because length may be invalid.
3639       region->init_req(2, IfTrue(iff));
3640       lengthx->init_req(2, ConvI2X(length));
3641 
3642       set_control(region);
3643       record_for_igvn(region);
3644       record_for_igvn(lengthx);
3645     }
3646   }
3647 #endif
3648 
3649   // Combine header size (plus rounding) and body size.  Then round down.
3650   // This computation cannot overflow, because it is used only in two
3651   // places, one where the length is sharply limited, and the other
3652   // after a successful allocation.
3653   Node* abody = lengthx;
3654   if (elem_shift != NULL)
3655     abody     = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3656   Node* size  = _gvn.transform( new AddXNode(headerx, abody) );
3657   if (round_mask != 0) {
3658     Node* mask = MakeConX(~round_mask);
3659     size       = _gvn.transform( new AndXNode(size, mask) );
3660   }
3661   // else if round_mask == 0, the size computation is self-rounding
3662 
3663   if (return_size_val != NULL) {
3664     // This is the size
3665     (*return_size_val) = size;
3666   }
3667 
3668   // Now generate allocation code
3669 
3670   // The entire memory state is needed for slow path of the allocation
3671   // since GC and deoptimization can happened.
3672   Node *mem = reset_memory();
3673   set_all_memory(mem); // Create new memory state
3674 
3675   if (initial_slow_test->is_Bool()) {
3676     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3677     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3678   }
3679 
3680   // Create the AllocateArrayNode and its result projections
3681   AllocateArrayNode* alloc
3682     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3683                             control(), mem, i_o(),
3684                             size, klass_node,
3685                             initial_slow_test,
3686                             length);
3687 
3688   // Cast to correct type.  Note that the klass_node may be constant or not,
3689   // and in the latter case the actual array type will be inexact also.
3690   // (This happens via a non-constant argument to inline_native_newArray.)
3691   // In any case, the value of klass_node provides the desired array type.
3692   const TypeInt* length_type = _gvn.find_int_type(length);
3693   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3694   if (ary_type->isa_aryptr() && length_type != NULL) {
3695     // Try to get a better type than POS for the size
3696     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3697   }
3698 
3699   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3700 
3701   // Cast length on remaining path to be as narrow as possible
3702   if (map()->find_edge(length) >= 0) {
3703     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3704     if (ccast != length) {
3705       _gvn.set_type_bottom(ccast);
3706       record_for_igvn(ccast);
3707       replace_in_map(length, ccast);
3708     }
3709   }
3710 
3711   return javaoop;
3712 }
3713 
3714 // The following "Ideal_foo" functions are placed here because they recognize
3715 // the graph shapes created by the functions immediately above.
3716 
3717 //---------------------------Ideal_allocation----------------------------------
3718 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3719 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3720   if (ptr == NULL) {     // reduce dumb test in callers
3721     return NULL;
3722   }
3723   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3724     ptr = ptr->in(1);
3725     if (ptr == NULL) return NULL;
3726   }
3727   // Return NULL for allocations with several casts:
3728   //   j.l.reflect.Array.newInstance(jobject, jint)
3729   //   Object.clone()
3730   // to keep more precise type from last cast.
3731   if (ptr->is_Proj()) {
3732     Node* allo = ptr->in(0);
3733     if (allo != NULL && allo->is_Allocate()) {
3734       return allo->as_Allocate();
3735     }
3736   }
3737   // Report failure to match.
3738   return NULL;
3739 }
3740 
3741 // Fancy version which also strips off an offset (and reports it to caller).
3742 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3743                                              intptr_t& offset) {
3744   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3745   if (base == NULL)  return NULL;
3746   return Ideal_allocation(base, phase);
3747 }
3748 
3749 // Trace Initialize <- Proj[Parm] <- Allocate
3750 AllocateNode* InitializeNode::allocation() {
3751   Node* rawoop = in(InitializeNode::RawAddress);
3752   if (rawoop->is_Proj()) {
3753     Node* alloc = rawoop->in(0);
3754     if (alloc->is_Allocate()) {
3755       return alloc->as_Allocate();
3756     }
3757   }
3758   return NULL;
3759 }
3760 
3761 // Trace Allocate -> Proj[Parm] -> Initialize
3762 InitializeNode* AllocateNode::initialization() {
3763   ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3764   if (rawoop == NULL)  return NULL;
3765   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3766     Node* init = rawoop->fast_out(i);
3767     if (init->is_Initialize()) {
3768       assert(init->as_Initialize()->allocation() == this, "2-way link");
3769       return init->as_Initialize();
3770     }
3771   }
3772   return NULL;
3773 }
3774 
3775 //----------------------------- loop predicates ---------------------------
3776 
3777 //------------------------------add_predicate_impl----------------------------
3778 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3779   // Too many traps seen?
3780   if (too_many_traps(reason)) {
3781 #ifdef ASSERT
3782     if (TraceLoopPredicate) {
3783       int tc = C->trap_count(reason);
3784       tty->print("too many traps=%s tcount=%d in ",
3785                     Deoptimization::trap_reason_name(reason), tc);
3786       method()->print(); // which method has too many predicate traps
3787       tty->cr();
3788     }
3789 #endif
3790     // We cannot afford to take more traps here,
3791     // do not generate predicate.
3792     return;
3793   }
3794 
3795   Node *cont    = _gvn.intcon(1);
3796   Node* opq     = _gvn.transform(new Opaque1Node(C, cont));
3797   Node *bol     = _gvn.transform(new Conv2BNode(opq));
3798   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3799   Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3800   C->add_predicate_opaq(opq);
3801   {
3802     PreserveJVMState pjvms(this);
3803     set_control(iffalse);
3804     inc_sp(nargs);
3805     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3806   }
3807   Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3808   set_control(iftrue);
3809 }
3810 
3811 //------------------------------add_predicate---------------------------------
3812 void GraphKit::add_predicate(int nargs) {
3813   if (UseLoopPredicate) {
3814     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3815   }
3816   // loop's limit check predicate should be near the loop.
3817   add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3818 }
3819 
3820 //----------------------------- store barriers ----------------------------
3821 #define __ ideal.
3822 
3823 bool GraphKit::use_ReduceInitialCardMarks() {
3824   BarrierSet *bs = BarrierSet::barrier_set();
3825   return bs->is_a(BarrierSet::CardTableBarrierSet)
3826          && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers()
3827          && ReduceInitialCardMarks;
3828 }
3829 
3830 void GraphKit::sync_kit(IdealKit& ideal) {
3831   set_all_memory(__ merged_memory());
3832   set_i_o(__ i_o());
3833   set_control(__ ctrl());
3834 }
3835 
3836 void GraphKit::final_sync(IdealKit& ideal) {
3837   // Final sync IdealKit and graphKit.
3838   sync_kit(ideal);
3839 }
3840 
3841 Node* GraphKit::byte_map_base_node() {
3842   // Get base of card map
3843   jbyte* card_table_base = ci_card_table_address();
3844   if (card_table_base != NULL) {
3845     return makecon(TypeRawPtr::make((address)card_table_base));
3846   } else {
3847     return null();
3848   }
3849 }
3850 
3851 // vanilla/CMS post barrier
3852 // Insert a write-barrier store.  This is to let generational GC work; we have
3853 // to flag all oop-stores before the next GC point.
3854 void GraphKit::write_barrier_post(Node* oop_store,
3855                                   Node* obj,
3856                                   Node* adr,
3857                                   uint  adr_idx,
3858                                   Node* val,
3859                                   bool use_precise) {
3860   // No store check needed if we're storing a NULL or an old object
3861   // (latter case is probably a string constant). The concurrent
3862   // mark sweep garbage collector, however, needs to have all nonNull
3863   // oop updates flagged via card-marks.
3864   if (val != NULL && val->is_Con()) {
3865     // must be either an oop or NULL
3866     const Type* t = val->bottom_type();
3867     if (t == TypePtr::NULL_PTR || t == Type::TOP)
3868       // stores of null never (?) need barriers
3869       return;
3870   }
3871 
3872   if (use_ReduceInitialCardMarks()
3873       && obj == just_allocated_object(control())) {
3874     // We can skip marks on a freshly-allocated object in Eden.
3875     // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
3876     // That routine informs GC to take appropriate compensating steps,
3877     // upon a slow-path allocation, so as to make this card-mark
3878     // elision safe.
3879     return;
3880   }
3881 
3882   if (!use_precise) {
3883     // All card marks for a (non-array) instance are in one place:
3884     adr = obj;
3885   }
3886   // (Else it's an array (or unknown), and we want more precise card marks.)
3887   assert(adr != NULL, "");
3888 
3889   IdealKit ideal(this, true);
3890 
3891   // Convert the pointer to an int prior to doing math on it
3892   Node* cast = __ CastPX(__ ctrl(), adr);
3893 
3894   // Divide by card size
3895   assert(BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet),
3896          "Only one we handle so far.");
3897   Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
3898 
3899   // Combine card table base and card offset
3900   Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3901 
3902   // Get the alias_index for raw card-mark memory
3903   int adr_type = Compile::AliasIdxRaw;
3904   Node*   zero = __ ConI(0); // Dirty card value
3905   BasicType bt = T_BYTE;
3906 
3907   if (UseConcMarkSweepGC && UseCondCardMark) {
3908     insert_mem_bar(Op_MemBarVolatile);   // StoreLoad barrier
3909     __ sync_kit(this);
3910   }
3911 
3912   if (UseCondCardMark) {
3913     // The classic GC reference write barrier is typically implemented
3914     // as a store into the global card mark table.  Unfortunately
3915     // unconditional stores can result in false sharing and excessive
3916     // coherence traffic as well as false transactional aborts.
3917     // UseCondCardMark enables MP "polite" conditional card mark
3918     // stores.  In theory we could relax the load from ctrl() to
3919     // no_ctrl, but that doesn't buy much latitude.
3920     Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3921     __ if_then(card_val, BoolTest::ne, zero);
3922   }
3923 
3924   // Smash zero into card
3925   if( !UseConcMarkSweepGC ) {
3926     __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3927   } else {
3928     // Specialized path for CM store barrier
3929     __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3930   }
3931 
3932   if (UseCondCardMark) {
3933     __ end_if();
3934   }
3935 
3936   // Final sync IdealKit and GraphKit.
3937   final_sync(ideal);
3938 }
3939 
3940 #if INCLUDE_G1GC
3941 
3942 /*
3943  * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3944  * required by SATB to make sure all objects live at the start of the
3945  * marking are kept alive, all reference updates need to any previous
3946  * reference stored before writing.
3947  *
3948  * If the previous value is NULL there is no need to save the old value.
3949  * References that are NULL are filtered during runtime by the barrier
3950  * code to avoid unnecessary queuing.
3951  *
3952  * However in the case of newly allocated objects it might be possible to
3953  * prove that the reference about to be overwritten is NULL during compile
3954  * time and avoid adding the barrier code completely.
3955  *
3956  * The compiler needs to determine that the object in which a field is about
3957  * to be written is newly allocated, and that no prior store to the same field
3958  * has happened since the allocation.
3959  *
3960  * Returns true if the pre-barrier can be removed
3961  */
3962 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3963                                          BasicType bt, uint adr_idx) {
3964   intptr_t offset = 0;
3965   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3966   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3967 
3968   if (offset == Type::OffsetBot) {
3969     return false; // cannot unalias unless there are precise offsets
3970   }
3971 
3972   if (alloc == NULL) {
3973     return false; // No allocation found
3974   }
3975 
3976   intptr_t size_in_bytes = type2aelembytes(bt);
3977 
3978   Node* mem = memory(adr_idx); // start searching here...
3979 
3980   for (int cnt = 0; cnt < 50; cnt++) {
3981 
3982     if (mem->is_Store()) {
3983 
3984       Node* st_adr = mem->in(MemNode::Address);
3985       intptr_t st_offset = 0;
3986       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
3987 
3988       if (st_base == NULL) {
3989         break; // inscrutable pointer
3990       }
3991 
3992       // Break we have found a store with same base and offset as ours so break
3993       if (st_base == base && st_offset == offset) {
3994         break;
3995       }
3996 
3997       if (st_offset != offset && st_offset != Type::OffsetBot) {
3998         const int MAX_STORE = BytesPerLong;
3999         if (st_offset >= offset + size_in_bytes ||
4000             st_offset <= offset - MAX_STORE ||
4001             st_offset <= offset - mem->as_Store()->memory_size()) {
4002           // Success:  The offsets are provably independent.
4003           // (You may ask, why not just test st_offset != offset and be done?
4004           // The answer is that stores of different sizes can co-exist
4005           // in the same sequence of RawMem effects.  We sometimes initialize
4006           // a whole 'tile' of array elements with a single jint or jlong.)
4007           mem = mem->in(MemNode::Memory);
4008           continue; // advance through independent store memory
4009         }
4010       }
4011 
4012       if (st_base != base
4013           && MemNode::detect_ptr_independence(base, alloc, st_base,
4014                                               AllocateNode::Ideal_allocation(st_base, phase),
4015                                               phase)) {
4016         // Success:  The bases are provably independent.
4017         mem = mem->in(MemNode::Memory);
4018         continue; // advance through independent store memory
4019       }
4020     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4021 
4022       InitializeNode* st_init = mem->in(0)->as_Initialize();
4023       AllocateNode* st_alloc = st_init->allocation();
4024 
4025       // Make sure that we are looking at the same allocation site.
4026       // The alloc variable is guaranteed to not be null here from earlier check.
4027       if (alloc == st_alloc) {
4028         // Check that the initialization is storing NULL so that no previous store
4029         // has been moved up and directly write a reference
4030         Node* captured_store = st_init->find_captured_store(offset,
4031                                                             type2aelembytes(T_OBJECT),
4032                                                             phase);
4033         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4034           return true;
4035         }
4036       }
4037     }
4038 
4039     // Unless there is an explicit 'continue', we must bail out here,
4040     // because 'mem' is an inscrutable memory state (e.g., a call).
4041     break;
4042   }
4043 
4044   return false;
4045 }
4046 
4047 // G1 pre/post barriers
4048 void GraphKit::g1_write_barrier_pre(bool do_load,
4049                                     Node* obj,
4050                                     Node* adr,
4051                                     uint alias_idx,
4052                                     Node* val,
4053                                     const TypeOopPtr* val_type,
4054                                     Node* pre_val,
4055                                     BasicType bt) {
4056 
4057   // Some sanity checks
4058   // Note: val is unused in this routine.
4059 
4060   if (do_load) {
4061     // We need to generate the load of the previous value
4062     assert(obj != NULL, "must have a base");
4063     assert(adr != NULL, "where are loading from?");
4064     assert(pre_val == NULL, "loaded already?");
4065     assert(val_type != NULL, "need a type");
4066 
4067     if (use_ReduceInitialCardMarks()
4068         && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4069       return;
4070     }
4071 
4072   } else {
4073     // In this case both val_type and alias_idx are unused.
4074     assert(pre_val != NULL, "must be loaded already");
4075     // Nothing to be done if pre_val is null.
4076     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4077     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4078   }
4079   assert(bt == T_OBJECT, "or we shouldn't be here");
4080 
4081   IdealKit ideal(this, true);
4082 
4083   Node* tls = __ thread(); // ThreadLocalStorage
4084 
4085   Node* no_ctrl = NULL;
4086   Node* no_base = __ top();
4087   Node* zero  = __ ConI(0);
4088   Node* zeroX = __ ConX(0);
4089 
4090   float likely  = PROB_LIKELY(0.999);
4091   float unlikely  = PROB_UNLIKELY(0.999);
4092 
4093   BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4094   assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4095 
4096   // Offsets into the thread
4097   const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
4098   const int index_offset   = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
4099   const int buffer_offset  = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
4100 
4101   // Now the actual pointers into the thread
4102   Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4103   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
4104   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
4105 
4106   // Now some of the values
4107   Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4108 
4109   // if (!marking)
4110   __ if_then(marking, BoolTest::ne, zero, unlikely); {
4111     BasicType index_bt = TypeX_X->basic_type();
4112     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4113     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4114 
4115     if (do_load) {
4116       // load original value
4117       // alias_idx correct??
4118       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4119     }
4120 
4121     // if (pre_val != NULL)
4122     __ if_then(pre_val, BoolTest::ne, null()); {
4123       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4124 
4125       // is the queue for this thread full?
4126       __ if_then(index, BoolTest::ne, zeroX, likely); {
4127 
4128         // decrement the index
4129         Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4130 
4131         // Now get the buffer location we will log the previous value into and store it
4132         Node *log_addr = __ AddP(no_base, buffer, next_index);
4133         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4134         // update the index
4135         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4136 
4137       } __ else_(); {
4138 
4139         // logging buffer is full, call the runtime
4140         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4141         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4142       } __ end_if();  // (!index)
4143     } __ end_if();  // (pre_val != NULL)
4144   } __ end_if();  // (!marking)
4145 
4146   // Final sync IdealKit and GraphKit.
4147   final_sync(ideal);
4148 }
4149 
4150 /*
4151  * G1 similar to any GC with a Young Generation requires a way to keep track of
4152  * references from Old Generation to Young Generation to make sure all live
4153  * objects are found. G1 also requires to keep track of object references
4154  * between different regions to enable evacuation of old regions, which is done
4155  * as part of mixed collections. References are tracked in remembered sets and
4156  * is continuously updated as reference are written to with the help of the
4157  * post-barrier.
4158  *
4159  * To reduce the number of updates to the remembered set the post-barrier
4160  * filters updates to fields in objects located in the Young Generation,
4161  * the same region as the reference, when the NULL is being written or
4162  * if the card is already marked as dirty by an earlier write.
4163  *
4164  * Under certain circumstances it is possible to avoid generating the
4165  * post-barrier completely if it is possible during compile time to prove
4166  * the object is newly allocated and that no safepoint exists between the
4167  * allocation and the store.
4168  *
4169  * In the case of slow allocation the allocation code must handle the barrier
4170  * as part of the allocation in the case the allocated object is not located
4171  * in the nursery, this would happen for humongous objects. This is similar to
4172  * how CMS is required to handle this case, see the comments for the method
4173  * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier.
4174  * A deferred card mark is required for these objects and handled in the above
4175  * mentioned methods.
4176  *
4177  * Returns true if the post barrier can be removed
4178  */
4179 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4180                                           Node* adr) {
4181   intptr_t      offset = 0;
4182   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4183   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
4184 
4185   if (offset == Type::OffsetBot) {
4186     return false; // cannot unalias unless there are precise offsets
4187   }
4188 
4189   if (alloc == NULL) {
4190      return false; // No allocation found
4191   }
4192 
4193   // Start search from Store node
4194   Node* mem = store->in(MemNode::Control);
4195   if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4196 
4197     InitializeNode* st_init = mem->in(0)->as_Initialize();
4198     AllocateNode*  st_alloc = st_init->allocation();
4199 
4200     // Make sure we are looking at the same allocation
4201     if (alloc == st_alloc) {
4202       return true;
4203     }
4204   }
4205 
4206   return false;
4207 }
4208 
4209 //
4210 // Update the card table and add card address to the queue
4211 //
4212 void GraphKit::g1_mark_card(IdealKit& ideal,
4213                             Node* card_adr,
4214                             Node* oop_store,
4215                             uint oop_alias_idx,
4216                             Node* index,
4217                             Node* index_adr,
4218                             Node* buffer,
4219                             const TypeFunc* tf) {
4220 
4221   Node* zero  = __ ConI(0);
4222   Node* zeroX = __ ConX(0);
4223   Node* no_base = __ top();
4224   BasicType card_bt = T_BYTE;
4225   // Smash zero into card. MUST BE ORDERED WRT TO STORE
4226   __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4227 
4228   //  Now do the queue work
4229   __ if_then(index, BoolTest::ne, zeroX); {
4230 
4231     Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4232     Node* log_addr = __ AddP(no_base, buffer, next_index);
4233 
4234     // Order, see storeCM.
4235     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4236     __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4237 
4238   } __ else_(); {
4239     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4240   } __ end_if();
4241 
4242 }
4243 
4244 void GraphKit::g1_write_barrier_post(Node* oop_store,
4245                                      Node* obj,
4246                                      Node* adr,
4247                                      uint alias_idx,
4248                                      Node* val,
4249                                      BasicType bt,
4250                                      bool use_precise) {
4251   // If we are writing a NULL then we need no post barrier
4252 
4253   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4254     // Must be NULL
4255     const Type* t = val->bottom_type();
4256     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4257     // No post barrier if writing NULLx
4258     return;
4259   }
4260 
4261   if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4262     // We can skip marks on a freshly-allocated object in Eden.
4263     // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
4264     // That routine informs GC to take appropriate compensating steps,
4265     // upon a slow-path allocation, so as to make this card-mark
4266     // elision safe.
4267     return;
4268   }
4269 
4270   if (use_ReduceInitialCardMarks()
4271       && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4272     return;
4273   }
4274 
4275   if (!use_precise) {
4276     // All card marks for a (non-array) instance are in one place:
4277     adr = obj;
4278   }
4279   // (Else it's an array (or unknown), and we want more precise card marks.)
4280   assert(adr != NULL, "");
4281 
4282   IdealKit ideal(this, true);
4283 
4284   Node* tls = __ thread(); // ThreadLocalStorage
4285 
4286   Node* no_base = __ top();
4287   float likely  = PROB_LIKELY(0.999);
4288   float unlikely  = PROB_UNLIKELY(0.999);
4289   Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
4290   Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val());
4291   Node* zeroX = __ ConX(0);
4292 
4293   // Get the alias_index for raw card-mark memory
4294   const TypePtr* card_type = TypeRawPtr::BOTTOM;
4295 
4296   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4297 
4298   // Offsets into the thread
4299   const int index_offset  = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
4300   const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
4301 
4302   // Pointers into the thread
4303 
4304   Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4305   Node* index_adr =  __ AddP(no_base, tls, __ ConX(index_offset));
4306 
4307   // Now some values
4308   // Use ctrl to avoid hoisting these values past a safepoint, which could
4309   // potentially reset these fields in the JavaThread.
4310   Node* index  = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4311   Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4312 
4313   // Convert the store obj pointer to an int prior to doing math on it
4314   // Must use ctrl to prevent "integerized oop" existing across safepoint
4315   Node* cast =  __ CastPX(__ ctrl(), adr);
4316 
4317   // Divide pointer by card size
4318   Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
4319 
4320   // Combine card table base and card offset
4321   Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4322 
4323   // If we know the value being stored does it cross regions?
4324 
4325   if (val != NULL) {
4326     // Does the store cause us to cross regions?
4327 
4328     // Should be able to do an unsigned compare of region_size instead of
4329     // and extra shift. Do we have an unsigned compare??
4330     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4331     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4332 
4333     // if (xor_res == 0) same region so skip
4334     __ if_then(xor_res, BoolTest::ne, zeroX); {
4335 
4336       // No barrier if we are storing a NULL
4337       __ if_then(val, BoolTest::ne, null(), unlikely); {
4338 
4339         // Ok must mark the card if not already dirty
4340 
4341         // load the original value of the card
4342         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4343 
4344         __ if_then(card_val, BoolTest::ne, young_card); {
4345           sync_kit(ideal);
4346           // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4347           insert_mem_bar(Op_MemBarVolatile, oop_store);
4348           __ sync_kit(this);
4349 
4350           Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4351           __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4352             g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4353           } __ end_if();
4354         } __ end_if();
4355       } __ end_if();
4356     } __ end_if();
4357   } else {
4358     // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4359     // We don't need a barrier here if the destination is a newly allocated object
4360     // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4361     // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
4362     assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4363     Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4364     __ if_then(card_val, BoolTest::ne, young_card); {
4365       g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4366     } __ end_if();
4367   }
4368 
4369   // Final sync IdealKit and GraphKit.
4370   final_sync(ideal);
4371 }
4372 #undef __
4373 
4374 #endif // INCLUDE_G1GC
4375 
4376 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4377   Node* len = load_array_length(load_String_value(ctrl, str));
4378   Node* coder = load_String_coder(ctrl, str);
4379   // Divide length by 2 if coder is UTF16
4380   return _gvn.transform(new RShiftINode(len, coder));
4381 }
4382 
4383 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4384   int value_offset = java_lang_String::value_offset_in_bytes();
4385   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4386                                                      false, NULL, 0);
4387   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4388   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4389                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4390                                                   ciTypeArrayKlass::make(T_BYTE), true, 0);
4391   int value_field_idx = C->get_alias_index(value_field_type);
4392   Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4393                          value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4394   // String.value field is known to be @Stable.
4395   if (UseImplicitStableValues) {
4396     load = cast_array_to_stable(load, value_type);
4397   }
4398   return load;
4399 }
4400 
4401 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4402   if (!CompactStrings) {
4403     return intcon(java_lang_String::CODER_UTF16);
4404   }
4405   int coder_offset = java_lang_String::coder_offset_in_bytes();
4406   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4407                                                      false, NULL, 0);
4408   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4409   int coder_field_idx = C->get_alias_index(coder_field_type);
4410   return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4411                    TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4412 }
4413 
4414 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4415   int value_offset = java_lang_String::value_offset_in_bytes();
4416   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4417                                                      false, NULL, 0);
4418   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4419   store_oop_to_object(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
4420       value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4421 }
4422 
4423 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4424   int coder_offset = java_lang_String::coder_offset_in_bytes();
4425   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4426                                                      false, NULL, 0);
4427   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4428   int coder_field_idx = C->get_alias_index(coder_field_type);
4429   store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4430                   value, T_BYTE, coder_field_idx, MemNode::unordered);
4431 }
4432 
4433 // Capture src and dst memory state with a MergeMemNode
4434 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4435   if (src_type == dst_type) {
4436     // Types are equal, we don't need a MergeMemNode
4437     return memory(src_type);
4438   }
4439   MergeMemNode* merge = MergeMemNode::make(map()->memory());
4440   record_for_igvn(merge); // fold it up later, if possible
4441   int src_idx = C->get_alias_index(src_type);
4442   int dst_idx = C->get_alias_index(dst_type);
4443   merge->set_memory_at(src_idx, memory(src_idx));
4444   merge->set_memory_at(dst_idx, memory(dst_idx));
4445   return merge;
4446 }
4447 
4448 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4449   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4450   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4451   // If input and output memory types differ, capture both states to preserve
4452   // the dependency between preceding and subsequent loads/stores.
4453   // For example, the following program:
4454   //  StoreB
4455   //  compress_string
4456   //  LoadB
4457   // has this memory graph (use->def):
4458   //  LoadB -> compress_string -> CharMem
4459   //             ... -> StoreB -> ByteMem
4460   // The intrinsic hides the dependency between LoadB and StoreB, causing
4461   // the load to read from memory not containing the result of the StoreB.
4462   // The correct memory graph should look like this:
4463   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4464   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4465   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4466   Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4467   set_memory(res_mem, TypeAryPtr::BYTES);
4468   return str;
4469 }
4470 
4471 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4472   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4473   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4474   // Capture src and dst memory (see comment in 'compress_string').
4475   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4476   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4477   set_memory(_gvn.transform(str), dst_type);
4478 }
4479 
4480 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4481   /**
4482    * int i_char = start;
4483    * for (int i_byte = 0; i_byte < count; i_byte++) {
4484    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
4485    * }
4486    */
4487   add_predicate();
4488   RegionNode* head = new RegionNode(3);
4489   head->init_req(1, control());
4490   gvn().set_type(head, Type::CONTROL);
4491   record_for_igvn(head);
4492 
4493   Node* i_byte = new PhiNode(head, TypeInt::INT);
4494   i_byte->init_req(1, intcon(0));
4495   gvn().set_type(i_byte, TypeInt::INT);
4496   record_for_igvn(i_byte);
4497 
4498   Node* i_char = new PhiNode(head, TypeInt::INT);
4499   i_char->init_req(1, start);
4500   gvn().set_type(i_char, TypeInt::INT);
4501   record_for_igvn(i_char);
4502 
4503   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4504   gvn().set_type(mem, Type::MEMORY);
4505   record_for_igvn(mem);
4506   set_control(head);
4507   set_memory(mem, TypeAryPtr::BYTES);
4508   Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4509   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4510                              AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4511                              false, false, true /* mismatched */);
4512 
4513   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4514   head->init_req(2, IfTrue(iff));
4515   mem->init_req(2, st);
4516   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4517   i_char->init_req(2, AddI(i_char, intcon(2)));
4518 
4519   set_control(IfFalse(iff));
4520   set_memory(st, TypeAryPtr::BYTES);
4521 }
4522 
4523 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4524   if (!field->is_constant()) {
4525     return NULL; // Field not marked as constant.
4526   }
4527   ciInstance* holder = NULL;
4528   if (!field->is_static()) {
4529     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4530     if (const_oop != NULL && const_oop->is_instance()) {
4531       holder = const_oop->as_instance();
4532     }
4533   }
4534   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4535                                                         /*is_unsigned_load=*/false);
4536   if (con_type != NULL) {
4537     return makecon(con_type);
4538   }
4539   return NULL;
4540 }
4541 
4542 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4543   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4544   // assumption of CCP analysis.
4545   return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4546 }