1 /*
   2  * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "compiler/compileLog.hpp"
  27 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
  28 #include "gc/g1/heapRegion.hpp"
  29 #include "gc/shared/barrierSet.hpp"
  30 #include "gc/shared/cardTableModRefBS.hpp"
  31 #include "gc/shared/collectedHeap.hpp"
  32 #include "memory/resourceArea.hpp"
  33 #include "opto/addnode.hpp"
  34 #include "opto/castnode.hpp"
  35 #include "opto/convertnode.hpp"
  36 #include "opto/graphKit.hpp"
  37 #include "opto/idealKit.hpp"
  38 #include "opto/intrinsicnode.hpp"
  39 #include "opto/locknode.hpp"
  40 #include "opto/machnode.hpp"
  41 #include "opto/opaquenode.hpp"
  42 #include "opto/parse.hpp"
  43 #include "opto/rootnode.hpp"
  44 #include "opto/runtime.hpp"
  45 #include "runtime/deoptimization.hpp"
  46 #include "runtime/sharedRuntime.hpp"
  47 
  48 //----------------------------GraphKit-----------------------------------------
  49 // Main utility constructor.
  50 GraphKit::GraphKit(JVMState* jvms)
  51   : Phase(Phase::Parser),
  52     _env(C->env()),
  53     _gvn(*C->initial_gvn())
  54 {
  55   _exceptions = jvms->map()->next_exception();
  56   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  57   set_jvms(jvms);
  58 }
  59 
  60 // Private constructor for parser.
  61 GraphKit::GraphKit()
  62   : Phase(Phase::Parser),
  63     _env(C->env()),
  64     _gvn(*C->initial_gvn())
  65 {
  66   _exceptions = NULL;
  67   set_map(NULL);
  68   debug_only(_sp = -99);
  69   debug_only(set_bci(-99));
  70 }
  71 
  72 
  73 
  74 //---------------------------clean_stack---------------------------------------
  75 // Clear away rubbish from the stack area of the JVM state.
  76 // This destroys any arguments that may be waiting on the stack.
  77 void GraphKit::clean_stack(int from_sp) {
  78   SafePointNode* map      = this->map();
  79   JVMState*      jvms     = this->jvms();
  80   int            stk_size = jvms->stk_size();
  81   int            stkoff   = jvms->stkoff();
  82   Node*          top      = this->top();
  83   for (int i = from_sp; i < stk_size; i++) {
  84     if (map->in(stkoff + i) != top) {
  85       map->set_req(stkoff + i, top);
  86     }
  87   }
  88 }
  89 
  90 
  91 //--------------------------------sync_jvms-----------------------------------
  92 // Make sure our current jvms agrees with our parse state.
  93 JVMState* GraphKit::sync_jvms() const {
  94   JVMState* jvms = this->jvms();
  95   jvms->set_bci(bci());       // Record the new bci in the JVMState
  96   jvms->set_sp(sp());         // Record the new sp in the JVMState
  97   assert(jvms_in_sync(), "jvms is now in sync");
  98   return jvms;
  99 }
 100 
 101 //--------------------------------sync_jvms_for_reexecute---------------------
 102 // Make sure our current jvms agrees with our parse state.  This version
 103 // uses the reexecute_sp for reexecuting bytecodes.
 104 JVMState* GraphKit::sync_jvms_for_reexecute() {
 105   JVMState* jvms = this->jvms();
 106   jvms->set_bci(bci());          // Record the new bci in the JVMState
 107   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 108   return jvms;
 109 }
 110 
 111 #ifdef ASSERT
 112 bool GraphKit::jvms_in_sync() const {
 113   Parse* parse = is_Parse();
 114   if (parse == NULL) {
 115     if (bci() !=      jvms()->bci())          return false;
 116     if (sp()  != (int)jvms()->sp())           return false;
 117     return true;
 118   }
 119   if (jvms()->method() != parse->method())    return false;
 120   if (jvms()->bci()    != parse->bci())       return false;
 121   int jvms_sp = jvms()->sp();
 122   if (jvms_sp          != parse->sp())        return false;
 123   int jvms_depth = jvms()->depth();
 124   if (jvms_depth       != parse->depth())     return false;
 125   return true;
 126 }
 127 
 128 // Local helper checks for special internal merge points
 129 // used to accumulate and merge exception states.
 130 // They are marked by the region's in(0) edge being the map itself.
 131 // Such merge points must never "escape" into the parser at large,
 132 // until they have been handed to gvn.transform.
 133 static bool is_hidden_merge(Node* reg) {
 134   if (reg == NULL)  return false;
 135   if (reg->is_Phi()) {
 136     reg = reg->in(0);
 137     if (reg == NULL)  return false;
 138   }
 139   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 140 }
 141 
 142 void GraphKit::verify_map() const {
 143   if (map() == NULL)  return;  // null map is OK
 144   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 145   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 146   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 147 }
 148 
 149 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 150   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 151   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 152 }
 153 #endif
 154 
 155 //---------------------------stop_and_kill_map---------------------------------
 156 // Set _map to NULL, signalling a stop to further bytecode execution.
 157 // First smash the current map's control to a constant, to mark it dead.
 158 void GraphKit::stop_and_kill_map() {
 159   SafePointNode* dead_map = stop();
 160   if (dead_map != NULL) {
 161     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
 162     assert(dead_map->is_killed(), "must be so marked");
 163   }
 164 }
 165 
 166 
 167 //--------------------------------stopped--------------------------------------
 168 // Tell if _map is NULL, or control is top.
 169 bool GraphKit::stopped() {
 170   if (map() == NULL)           return true;
 171   else if (control() == top()) return true;
 172   else                         return false;
 173 }
 174 
 175 
 176 //-----------------------------has_ex_handler----------------------------------
 177 // Tell if this method or any caller method has exception handlers.
 178 bool GraphKit::has_ex_handler() {
 179   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 180     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 181       return true;
 182     }
 183   }
 184   return false;
 185 }
 186 
 187 //------------------------------save_ex_oop------------------------------------
 188 // Save an exception without blowing stack contents or other JVM state.
 189 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 190   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 191   ex_map->add_req(ex_oop);
 192   debug_only(verify_exception_state(ex_map));
 193 }
 194 
 195 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 196   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 197   Node* ex_oop = ex_map->in(ex_map->req()-1);
 198   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 199   return ex_oop;
 200 }
 201 
 202 //-----------------------------saved_ex_oop------------------------------------
 203 // Recover a saved exception from its map.
 204 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 205   return common_saved_ex_oop(ex_map, false);
 206 }
 207 
 208 //--------------------------clear_saved_ex_oop---------------------------------
 209 // Erase a previously saved exception from its map.
 210 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 211   return common_saved_ex_oop(ex_map, true);
 212 }
 213 
 214 #ifdef ASSERT
 215 //---------------------------has_saved_ex_oop----------------------------------
 216 // Erase a previously saved exception from its map.
 217 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 218   return ex_map->req() == ex_map->jvms()->endoff()+1;
 219 }
 220 #endif
 221 
 222 //-------------------------make_exception_state--------------------------------
 223 // Turn the current JVM state into an exception state, appending the ex_oop.
 224 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 225   sync_jvms();
 226   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 227   set_saved_ex_oop(ex_map, ex_oop);
 228   return ex_map;
 229 }
 230 
 231 
 232 //--------------------------add_exception_state--------------------------------
 233 // Add an exception to my list of exceptions.
 234 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 235   if (ex_map == NULL || ex_map->control() == top()) {
 236     return;
 237   }
 238 #ifdef ASSERT
 239   verify_exception_state(ex_map);
 240   if (has_exceptions()) {
 241     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 242   }
 243 #endif
 244 
 245   // If there is already an exception of exactly this type, merge with it.
 246   // In particular, null-checks and other low-level exceptions common up here.
 247   Node*       ex_oop  = saved_ex_oop(ex_map);
 248   const Type* ex_type = _gvn.type(ex_oop);
 249   if (ex_oop == top()) {
 250     // No action needed.
 251     return;
 252   }
 253   assert(ex_type->isa_instptr(), "exception must be an instance");
 254   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 255     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 256     // We check sp also because call bytecodes can generate exceptions
 257     // both before and after arguments are popped!
 258     if (ex_type2 == ex_type
 259         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 260       combine_exception_states(ex_map, e2);
 261       return;
 262     }
 263   }
 264 
 265   // No pre-existing exception of the same type.  Chain it on the list.
 266   push_exception_state(ex_map);
 267 }
 268 
 269 //-----------------------add_exception_states_from-----------------------------
 270 void GraphKit::add_exception_states_from(JVMState* jvms) {
 271   SafePointNode* ex_map = jvms->map()->next_exception();
 272   if (ex_map != NULL) {
 273     jvms->map()->set_next_exception(NULL);
 274     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 275       next_map = ex_map->next_exception();
 276       ex_map->set_next_exception(NULL);
 277       add_exception_state(ex_map);
 278     }
 279   }
 280 }
 281 
 282 //-----------------------transfer_exceptions_into_jvms-------------------------
 283 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 284   if (map() == NULL) {
 285     // We need a JVMS to carry the exceptions, but the map has gone away.
 286     // Create a scratch JVMS, cloned from any of the exception states...
 287     if (has_exceptions()) {
 288       _map = _exceptions;
 289       _map = clone_map();
 290       _map->set_next_exception(NULL);
 291       clear_saved_ex_oop(_map);
 292       debug_only(verify_map());
 293     } else {
 294       // ...or created from scratch
 295       JVMState* jvms = new (C) JVMState(_method, NULL);
 296       jvms->set_bci(_bci);
 297       jvms->set_sp(_sp);
 298       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
 299       set_jvms(jvms);
 300       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 301       set_all_memory(top());
 302       while (map()->req() < jvms->endoff())  map()->add_req(top());
 303     }
 304     // (This is a kludge, in case you didn't notice.)
 305     set_control(top());
 306   }
 307   JVMState* jvms = sync_jvms();
 308   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 309   jvms->map()->set_next_exception(_exceptions);
 310   _exceptions = NULL;   // done with this set of exceptions
 311   return jvms;
 312 }
 313 
 314 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 315   assert(is_hidden_merge(dstphi), "must be a special merge node");
 316   assert(is_hidden_merge(srcphi), "must be a special merge node");
 317   uint limit = srcphi->req();
 318   for (uint i = PhiNode::Input; i < limit; i++) {
 319     dstphi->add_req(srcphi->in(i));
 320   }
 321 }
 322 static inline void add_one_req(Node* dstphi, Node* src) {
 323   assert(is_hidden_merge(dstphi), "must be a special merge node");
 324   assert(!is_hidden_merge(src), "must not be a special merge node");
 325   dstphi->add_req(src);
 326 }
 327 
 328 //-----------------------combine_exception_states------------------------------
 329 // This helper function combines exception states by building phis on a
 330 // specially marked state-merging region.  These regions and phis are
 331 // untransformed, and can build up gradually.  The region is marked by
 332 // having a control input of its exception map, rather than NULL.  Such
 333 // regions do not appear except in this function, and in use_exception_state.
 334 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 335   if (failing())  return;  // dying anyway...
 336   JVMState* ex_jvms = ex_map->_jvms;
 337   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 338   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 339   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 340   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 341   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 342   assert(ex_map->req() == phi_map->req(), "matching maps");
 343   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 344   Node*         hidden_merge_mark = root();
 345   Node*         region  = phi_map->control();
 346   MergeMemNode* phi_mem = phi_map->merged_memory();
 347   MergeMemNode* ex_mem  = ex_map->merged_memory();
 348   if (region->in(0) != hidden_merge_mark) {
 349     // The control input is not (yet) a specially-marked region in phi_map.
 350     // Make it so, and build some phis.
 351     region = new RegionNode(2);
 352     _gvn.set_type(region, Type::CONTROL);
 353     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 354     region->init_req(1, phi_map->control());
 355     phi_map->set_control(region);
 356     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 357     record_for_igvn(io_phi);
 358     _gvn.set_type(io_phi, Type::ABIO);
 359     phi_map->set_i_o(io_phi);
 360     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 361       Node* m = mms.memory();
 362       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 363       record_for_igvn(m_phi);
 364       _gvn.set_type(m_phi, Type::MEMORY);
 365       mms.set_memory(m_phi);
 366     }
 367   }
 368 
 369   // Either or both of phi_map and ex_map might already be converted into phis.
 370   Node* ex_control = ex_map->control();
 371   // if there is special marking on ex_map also, we add multiple edges from src
 372   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 373   // how wide was the destination phi_map, originally?
 374   uint orig_width = region->req();
 375 
 376   if (add_multiple) {
 377     add_n_reqs(region, ex_control);
 378     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 379   } else {
 380     // ex_map has no merges, so we just add single edges everywhere
 381     add_one_req(region, ex_control);
 382     add_one_req(phi_map->i_o(), ex_map->i_o());
 383   }
 384   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 385     if (mms.is_empty()) {
 386       // get a copy of the base memory, and patch some inputs into it
 387       const TypePtr* adr_type = mms.adr_type(C);
 388       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 389       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 390       mms.set_memory(phi);
 391       // Prepare to append interesting stuff onto the newly sliced phi:
 392       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 393     }
 394     // Append stuff from ex_map:
 395     if (add_multiple) {
 396       add_n_reqs(mms.memory(), mms.memory2());
 397     } else {
 398       add_one_req(mms.memory(), mms.memory2());
 399     }
 400   }
 401   uint limit = ex_map->req();
 402   for (uint i = TypeFunc::Parms; i < limit; i++) {
 403     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 404     if (i == tos)  i = ex_jvms->monoff();
 405     Node* src = ex_map->in(i);
 406     Node* dst = phi_map->in(i);
 407     if (src != dst) {
 408       PhiNode* phi;
 409       if (dst->in(0) != region) {
 410         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 411         record_for_igvn(phi);
 412         _gvn.set_type(phi, phi->type());
 413         phi_map->set_req(i, dst);
 414         // Prepare to append interesting stuff onto the new phi:
 415         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 416       } else {
 417         assert(dst->is_Phi(), "nobody else uses a hidden region");
 418         phi = dst->as_Phi();
 419       }
 420       if (add_multiple && src->in(0) == ex_control) {
 421         // Both are phis.
 422         add_n_reqs(dst, src);
 423       } else {
 424         while (dst->req() < region->req())  add_one_req(dst, src);
 425       }
 426       const Type* srctype = _gvn.type(src);
 427       if (phi->type() != srctype) {
 428         const Type* dsttype = phi->type()->meet_speculative(srctype);
 429         if (phi->type() != dsttype) {
 430           phi->set_type(dsttype);
 431           _gvn.set_type(phi, dsttype);
 432         }
 433       }
 434     }
 435   }
 436   phi_map->merge_replaced_nodes_with(ex_map);
 437 }
 438 
 439 //--------------------------use_exception_state--------------------------------
 440 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 441   if (failing()) { stop(); return top(); }
 442   Node* region = phi_map->control();
 443   Node* hidden_merge_mark = root();
 444   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 445   Node* ex_oop = clear_saved_ex_oop(phi_map);
 446   if (region->in(0) == hidden_merge_mark) {
 447     // Special marking for internal ex-states.  Process the phis now.
 448     region->set_req(0, region);  // now it's an ordinary region
 449     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 450     // Note: Setting the jvms also sets the bci and sp.
 451     set_control(_gvn.transform(region));
 452     uint tos = jvms()->stkoff() + sp();
 453     for (uint i = 1; i < tos; i++) {
 454       Node* x = phi_map->in(i);
 455       if (x->in(0) == region) {
 456         assert(x->is_Phi(), "expected a special phi");
 457         phi_map->set_req(i, _gvn.transform(x));
 458       }
 459     }
 460     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 461       Node* x = mms.memory();
 462       if (x->in(0) == region) {
 463         assert(x->is_Phi(), "nobody else uses a hidden region");
 464         mms.set_memory(_gvn.transform(x));
 465       }
 466     }
 467     if (ex_oop->in(0) == region) {
 468       assert(ex_oop->is_Phi(), "expected a special phi");
 469       ex_oop = _gvn.transform(ex_oop);
 470     }
 471   } else {
 472     set_jvms(phi_map->jvms());
 473   }
 474 
 475   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 476   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 477   return ex_oop;
 478 }
 479 
 480 //---------------------------------java_bc-------------------------------------
 481 Bytecodes::Code GraphKit::java_bc() const {
 482   ciMethod* method = this->method();
 483   int       bci    = this->bci();
 484   if (method != NULL && bci != InvocationEntryBci)
 485     return method->java_code_at_bci(bci);
 486   else
 487     return Bytecodes::_illegal;
 488 }
 489 
 490 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 491                                                           bool must_throw) {
 492     // if the exception capability is set, then we will generate code
 493     // to check the JavaThread.should_post_on_exceptions flag to see
 494     // if we actually need to report exception events (for this
 495     // thread).  If we don't need to report exception events, we will
 496     // take the normal fast path provided by add_exception_events.  If
 497     // exception event reporting is enabled for this thread, we will
 498     // take the uncommon_trap in the BuildCutout below.
 499 
 500     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 501     Node* jthread = _gvn.transform(new ThreadLocalNode());
 502     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 503     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
 504 
 505     // Test the should_post_on_exceptions_flag vs. 0
 506     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
 507     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
 508 
 509     // Branch to slow_path if should_post_on_exceptions_flag was true
 510     { BuildCutout unless(this, tst, PROB_MAX);
 511       // Do not try anything fancy if we're notifying the VM on every throw.
 512       // Cf. case Bytecodes::_athrow in parse2.cpp.
 513       uncommon_trap(reason, Deoptimization::Action_none,
 514                     (ciKlass*)NULL, (char*)NULL, must_throw);
 515     }
 516 
 517 }
 518 
 519 //------------------------------builtin_throw----------------------------------
 520 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 521   bool must_throw = true;
 522 
 523   if (env()->jvmti_can_post_on_exceptions()) {
 524     // check if we must post exception events, take uncommon trap if so
 525     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 526     // here if should_post_on_exceptions is false
 527     // continue on with the normal codegen
 528   }
 529 
 530   // If this particular condition has not yet happened at this
 531   // bytecode, then use the uncommon trap mechanism, and allow for
 532   // a future recompilation if several traps occur here.
 533   // If the throw is hot, try to use a more complicated inline mechanism
 534   // which keeps execution inside the compiled code.
 535   bool treat_throw_as_hot = false;
 536   ciMethodData* md = method()->method_data();
 537 
 538   if (ProfileTraps) {
 539     if (too_many_traps(reason)) {
 540       treat_throw_as_hot = true;
 541     }
 542     // (If there is no MDO at all, assume it is early in
 543     // execution, and that any deopts are part of the
 544     // startup transient, and don't need to be remembered.)
 545 
 546     // Also, if there is a local exception handler, treat all throws
 547     // as hot if there has been at least one in this method.
 548     if (C->trap_count(reason) != 0
 549         && method()->method_data()->trap_count(reason) != 0
 550         && has_ex_handler()) {
 551         treat_throw_as_hot = true;
 552     }
 553   }
 554 
 555   // If this throw happens frequently, an uncommon trap might cause
 556   // a performance pothole.  If there is a local exception handler,
 557   // and if this particular bytecode appears to be deoptimizing often,
 558   // let us handle the throw inline, with a preconstructed instance.
 559   // Note:   If the deopt count has blown up, the uncommon trap
 560   // runtime is going to flush this nmethod, not matter what.
 561   if (treat_throw_as_hot
 562       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 563     // If the throw is local, we use a pre-existing instance and
 564     // punt on the backtrace.  This would lead to a missing backtrace
 565     // (a repeat of 4292742) if the backtrace object is ever asked
 566     // for its backtrace.
 567     // Fixing this remaining case of 4292742 requires some flavor of
 568     // escape analysis.  Leave that for the future.
 569     ciInstance* ex_obj = NULL;
 570     switch (reason) {
 571     case Deoptimization::Reason_null_check:
 572       ex_obj = env()->NullPointerException_instance();
 573       break;
 574     case Deoptimization::Reason_div0_check:
 575       ex_obj = env()->ArithmeticException_instance();
 576       break;
 577     case Deoptimization::Reason_range_check:
 578       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 579       break;
 580     case Deoptimization::Reason_class_check:
 581       if (java_bc() == Bytecodes::_aastore) {
 582         ex_obj = env()->ArrayStoreException_instance();
 583       } else {
 584         ex_obj = env()->ClassCastException_instance();
 585       }
 586       break;
 587     }
 588     if (failing()) { stop(); return; }  // exception allocation might fail
 589     if (ex_obj != NULL) {
 590       // Cheat with a preallocated exception object.
 591       if (C->log() != NULL)
 592         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 593                        Deoptimization::trap_reason_name(reason));
 594       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 595       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
 596 
 597       // Clear the detail message of the preallocated exception object.
 598       // Weblogic sometimes mutates the detail message of exceptions
 599       // using reflection.
 600       int offset = java_lang_Throwable::get_detailMessage_offset();
 601       const TypePtr* adr_typ = ex_con->add_offset(offset);
 602 
 603       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 604       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 605       // Conservatively release stores of object references.
 606       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
 607 
 608       add_exception_state(make_exception_state(ex_node));
 609       return;
 610     }
 611   }
 612 
 613   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 614   // It won't be much cheaper than bailing to the interp., since we'll
 615   // have to pass up all the debug-info, and the runtime will have to
 616   // create the stack trace.
 617 
 618   // Usual case:  Bail to interpreter.
 619   // Reserve the right to recompile if we haven't seen anything yet.
 620 
 621   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
 622   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 623   if (treat_throw_as_hot
 624       && (method()->method_data()->trap_recompiled_at(bci(), m)
 625           || C->too_many_traps(reason))) {
 626     // We cannot afford to take more traps here.  Suffer in the interpreter.
 627     if (C->log() != NULL)
 628       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 629                      Deoptimization::trap_reason_name(reason),
 630                      C->trap_count(reason));
 631     action = Deoptimization::Action_none;
 632   }
 633 
 634   // "must_throw" prunes the JVM state to include only the stack, if there
 635   // are no local exception handlers.  This should cut down on register
 636   // allocation time and code size, by drastically reducing the number
 637   // of in-edges on the call to the uncommon trap.
 638 
 639   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 640 }
 641 
 642 
 643 //----------------------------PreserveJVMState---------------------------------
 644 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 645   debug_only(kit->verify_map());
 646   _kit    = kit;
 647   _map    = kit->map();   // preserve the map
 648   _sp     = kit->sp();
 649   kit->set_map(clone_map ? kit->clone_map() : NULL);
 650 #ifdef ASSERT
 651   _bci    = kit->bci();
 652   Parse* parser = kit->is_Parse();
 653   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 654   _block  = block;
 655 #endif
 656 }
 657 PreserveJVMState::~PreserveJVMState() {
 658   GraphKit* kit = _kit;
 659 #ifdef ASSERT
 660   assert(kit->bci() == _bci, "bci must not shift");
 661   Parse* parser = kit->is_Parse();
 662   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 663   assert(block == _block,    "block must not shift");
 664 #endif
 665   kit->set_map(_map);
 666   kit->set_sp(_sp);
 667 }
 668 
 669 
 670 //-----------------------------BuildCutout-------------------------------------
 671 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 672   : PreserveJVMState(kit)
 673 {
 674   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 675   SafePointNode* outer_map = _map;   // preserved map is caller's
 676   SafePointNode* inner_map = kit->map();
 677   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 678   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
 679   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
 680 }
 681 BuildCutout::~BuildCutout() {
 682   GraphKit* kit = _kit;
 683   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 684 }
 685 
 686 //---------------------------PreserveReexecuteState----------------------------
 687 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 688   assert(!kit->stopped(), "must call stopped() before");
 689   _kit    =    kit;
 690   _sp     =    kit->sp();
 691   _reexecute = kit->jvms()->_reexecute;
 692 }
 693 PreserveReexecuteState::~PreserveReexecuteState() {
 694   if (_kit->stopped()) return;
 695   _kit->jvms()->_reexecute = _reexecute;
 696   _kit->set_sp(_sp);
 697 }
 698 
 699 //------------------------------clone_map--------------------------------------
 700 // Implementation of PreserveJVMState
 701 //
 702 // Only clone_map(...) here. If this function is only used in the
 703 // PreserveJVMState class we may want to get rid of this extra
 704 // function eventually and do it all there.
 705 
 706 SafePointNode* GraphKit::clone_map() {
 707   if (map() == NULL)  return NULL;
 708 
 709   // Clone the memory edge first
 710   Node* mem = MergeMemNode::make(map()->memory());
 711   gvn().set_type_bottom(mem);
 712 
 713   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 714   JVMState* jvms = this->jvms();
 715   JVMState* clonejvms = jvms->clone_shallow(C);
 716   clonemap->set_memory(mem);
 717   clonemap->set_jvms(clonejvms);
 718   clonejvms->set_map(clonemap);
 719   record_for_igvn(clonemap);
 720   gvn().set_type_bottom(clonemap);
 721   return clonemap;
 722 }
 723 
 724 
 725 //-----------------------------set_map_clone-----------------------------------
 726 void GraphKit::set_map_clone(SafePointNode* m) {
 727   _map = m;
 728   _map = clone_map();
 729   _map->set_next_exception(NULL);
 730   debug_only(verify_map());
 731 }
 732 
 733 
 734 //----------------------------kill_dead_locals---------------------------------
 735 // Detect any locals which are known to be dead, and force them to top.
 736 void GraphKit::kill_dead_locals() {
 737   // Consult the liveness information for the locals.  If any
 738   // of them are unused, then they can be replaced by top().  This
 739   // should help register allocation time and cut down on the size
 740   // of the deoptimization information.
 741 
 742   // This call is made from many of the bytecode handling
 743   // subroutines called from the Big Switch in do_one_bytecode.
 744   // Every bytecode which might include a slow path is responsible
 745   // for killing its dead locals.  The more consistent we
 746   // are about killing deads, the fewer useless phis will be
 747   // constructed for them at various merge points.
 748 
 749   // bci can be -1 (InvocationEntryBci).  We return the entry
 750   // liveness for the method.
 751 
 752   if (method() == NULL || method()->code_size() == 0) {
 753     // We are building a graph for a call to a native method.
 754     // All locals are live.
 755     return;
 756   }
 757 
 758   ResourceMark rm;
 759 
 760   // Consult the liveness information for the locals.  If any
 761   // of them are unused, then they can be replaced by top().  This
 762   // should help register allocation time and cut down on the size
 763   // of the deoptimization information.
 764   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 765 
 766   int len = (int)live_locals.size();
 767   assert(len <= jvms()->loc_size(), "too many live locals");
 768   for (int local = 0; local < len; local++) {
 769     if (!live_locals.at(local)) {
 770       set_local(local, top());
 771     }
 772   }
 773 }
 774 
 775 #ifdef ASSERT
 776 //-------------------------dead_locals_are_killed------------------------------
 777 // Return true if all dead locals are set to top in the map.
 778 // Used to assert "clean" debug info at various points.
 779 bool GraphKit::dead_locals_are_killed() {
 780   if (method() == NULL || method()->code_size() == 0) {
 781     // No locals need to be dead, so all is as it should be.
 782     return true;
 783   }
 784 
 785   // Make sure somebody called kill_dead_locals upstream.
 786   ResourceMark rm;
 787   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 788     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 789     SafePointNode* map = jvms->map();
 790     ciMethod* method = jvms->method();
 791     int       bci    = jvms->bci();
 792     if (jvms == this->jvms()) {
 793       bci = this->bci();  // it might not yet be synched
 794     }
 795     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 796     int len = (int)live_locals.size();
 797     if (!live_locals.is_valid() || len == 0)
 798       // This method is trivial, or is poisoned by a breakpoint.
 799       return true;
 800     assert(len == jvms->loc_size(), "live map consistent with locals map");
 801     for (int local = 0; local < len; local++) {
 802       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 803         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 804           tty->print_cr("Zombie local %d: ", local);
 805           jvms->dump();
 806         }
 807         return false;
 808       }
 809     }
 810   }
 811   return true;
 812 }
 813 
 814 #endif //ASSERT
 815 
 816 // Helper function for enforcing certain bytecodes to reexecute if
 817 // deoptimization happens
 818 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 819   ciMethod* cur_method = jvms->method();
 820   int       cur_bci   = jvms->bci();
 821   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 822     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 823     return Interpreter::bytecode_should_reexecute(code) ||
 824            is_anewarray && code == Bytecodes::_multianewarray;
 825     // Reexecute _multianewarray bytecode which was replaced with
 826     // sequence of [a]newarray. See Parse::do_multianewarray().
 827     //
 828     // Note: interpreter should not have it set since this optimization
 829     // is limited by dimensions and guarded by flag so in some cases
 830     // multianewarray() runtime calls will be generated and
 831     // the bytecode should not be reexecutes (stack will not be reset).
 832   } else
 833     return false;
 834 }
 835 
 836 // Helper function for adding JVMState and debug information to node
 837 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 838   // Add the safepoint edges to the call (or other safepoint).
 839 
 840   // Make sure dead locals are set to top.  This
 841   // should help register allocation time and cut down on the size
 842   // of the deoptimization information.
 843   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 844 
 845   // Walk the inline list to fill in the correct set of JVMState's
 846   // Also fill in the associated edges for each JVMState.
 847 
 848   // If the bytecode needs to be reexecuted we need to put
 849   // the arguments back on the stack.
 850   const bool should_reexecute = jvms()->should_reexecute();
 851   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 852 
 853   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 854   // undefined if the bci is different.  This is normal for Parse but it
 855   // should not happen for LibraryCallKit because only one bci is processed.
 856   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 857          "in LibraryCallKit the reexecute bit should not change");
 858 
 859   // If we are guaranteed to throw, we can prune everything but the
 860   // input to the current bytecode.
 861   bool can_prune_locals = false;
 862   uint stack_slots_not_pruned = 0;
 863   int inputs = 0, depth = 0;
 864   if (must_throw) {
 865     assert(method() == youngest_jvms->method(), "sanity");
 866     if (compute_stack_effects(inputs, depth)) {
 867       can_prune_locals = true;
 868       stack_slots_not_pruned = inputs;
 869     }
 870   }
 871 
 872   if (env()->should_retain_local_variables()) {
 873     // At any safepoint, this method can get breakpointed, which would
 874     // then require an immediate deoptimization.
 875     can_prune_locals = false;  // do not prune locals
 876     stack_slots_not_pruned = 0;
 877   }
 878 
 879   // do not scribble on the input jvms
 880   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 881   call->set_jvms(out_jvms); // Start jvms list for call node
 882 
 883   // For a known set of bytecodes, the interpreter should reexecute them if
 884   // deoptimization happens. We set the reexecute state for them here
 885   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 886       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 887     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 888   }
 889 
 890   // Presize the call:
 891   DEBUG_ONLY(uint non_debug_edges = call->req());
 892   call->add_req_batch(top(), youngest_jvms->debug_depth());
 893   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 894 
 895   // Set up edges so that the call looks like this:
 896   //  Call [state:] ctl io mem fptr retadr
 897   //       [parms:] parm0 ... parmN
 898   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 899   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 900   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 901   // Note that caller debug info precedes callee debug info.
 902 
 903   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 904   uint debug_ptr = call->req();
 905 
 906   // Loop over the map input edges associated with jvms, add them
 907   // to the call node, & reset all offsets to match call node array.
 908   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 909     uint debug_end   = debug_ptr;
 910     uint debug_start = debug_ptr - in_jvms->debug_size();
 911     debug_ptr = debug_start;  // back up the ptr
 912 
 913     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 914     uint j, k, l;
 915     SafePointNode* in_map = in_jvms->map();
 916     out_jvms->set_map(call);
 917 
 918     if (can_prune_locals) {
 919       assert(in_jvms->method() == out_jvms->method(), "sanity");
 920       // If the current throw can reach an exception handler in this JVMS,
 921       // then we must keep everything live that can reach that handler.
 922       // As a quick and dirty approximation, we look for any handlers at all.
 923       if (in_jvms->method()->has_exception_handlers()) {
 924         can_prune_locals = false;
 925       }
 926     }
 927 
 928     // Add the Locals
 929     k = in_jvms->locoff();
 930     l = in_jvms->loc_size();
 931     out_jvms->set_locoff(p);
 932     if (!can_prune_locals) {
 933       for (j = 0; j < l; j++)
 934         call->set_req(p++, in_map->in(k+j));
 935     } else {
 936       p += l;  // already set to top above by add_req_batch
 937     }
 938 
 939     // Add the Expression Stack
 940     k = in_jvms->stkoff();
 941     l = in_jvms->sp();
 942     out_jvms->set_stkoff(p);
 943     if (!can_prune_locals) {
 944       for (j = 0; j < l; j++)
 945         call->set_req(p++, in_map->in(k+j));
 946     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 947       // Divide stack into {S0,...,S1}, where S0 is set to top.
 948       uint s1 = stack_slots_not_pruned;
 949       stack_slots_not_pruned = 0;  // for next iteration
 950       if (s1 > l)  s1 = l;
 951       uint s0 = l - s1;
 952       p += s0;  // skip the tops preinstalled by add_req_batch
 953       for (j = s0; j < l; j++)
 954         call->set_req(p++, in_map->in(k+j));
 955     } else {
 956       p += l;  // already set to top above by add_req_batch
 957     }
 958 
 959     // Add the Monitors
 960     k = in_jvms->monoff();
 961     l = in_jvms->mon_size();
 962     out_jvms->set_monoff(p);
 963     for (j = 0; j < l; j++)
 964       call->set_req(p++, in_map->in(k+j));
 965 
 966     // Copy any scalar object fields.
 967     k = in_jvms->scloff();
 968     l = in_jvms->scl_size();
 969     out_jvms->set_scloff(p);
 970     for (j = 0; j < l; j++)
 971       call->set_req(p++, in_map->in(k+j));
 972 
 973     // Finish the new jvms.
 974     out_jvms->set_endoff(p);
 975 
 976     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 977     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 978     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 979     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 980     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 981     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 982 
 983     // Update the two tail pointers in parallel.
 984     out_jvms = out_jvms->caller();
 985     in_jvms  = in_jvms->caller();
 986   }
 987 
 988   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
 989 
 990   // Test the correctness of JVMState::debug_xxx accessors:
 991   assert(call->jvms()->debug_start() == non_debug_edges, "");
 992   assert(call->jvms()->debug_end()   == call->req(), "");
 993   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
 994 }
 995 
 996 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
 997   Bytecodes::Code code = java_bc();
 998   if (code == Bytecodes::_wide) {
 999     code = method()->java_code_at_bci(bci() + 1);
1000   }
1001 
1002   BasicType rtype = T_ILLEGAL;
1003   int       rsize = 0;
1004 
1005   if (code != Bytecodes::_illegal) {
1006     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1007     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1008     if (rtype < T_CONFLICT)
1009       rsize = type2size[rtype];
1010   }
1011 
1012   switch (code) {
1013   case Bytecodes::_illegal:
1014     return false;
1015 
1016   case Bytecodes::_ldc:
1017   case Bytecodes::_ldc_w:
1018   case Bytecodes::_ldc2_w:
1019     inputs = 0;
1020     break;
1021 
1022   case Bytecodes::_dup:         inputs = 1;  break;
1023   case Bytecodes::_dup_x1:      inputs = 2;  break;
1024   case Bytecodes::_dup_x2:      inputs = 3;  break;
1025   case Bytecodes::_dup2:        inputs = 2;  break;
1026   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1027   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1028   case Bytecodes::_swap:        inputs = 2;  break;
1029   case Bytecodes::_arraylength: inputs = 1;  break;
1030 
1031   case Bytecodes::_getstatic:
1032   case Bytecodes::_putstatic:
1033   case Bytecodes::_getfield:
1034   case Bytecodes::_putfield:
1035     {
1036       bool ignored_will_link;
1037       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1038       int      size  = field->type()->size();
1039       bool is_get = (depth >= 0), is_static = (depth & 1);
1040       inputs = (is_static ? 0 : 1);
1041       if (is_get) {
1042         depth = size - inputs;
1043       } else {
1044         inputs += size;        // putxxx pops the value from the stack
1045         depth = - inputs;
1046       }
1047     }
1048     break;
1049 
1050   case Bytecodes::_invokevirtual:
1051   case Bytecodes::_invokespecial:
1052   case Bytecodes::_invokestatic:
1053   case Bytecodes::_invokedynamic:
1054   case Bytecodes::_invokeinterface:
1055     {
1056       bool ignored_will_link;
1057       ciSignature* declared_signature = NULL;
1058       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1059       assert(declared_signature != NULL, "cannot be null");
1060       inputs   = declared_signature->arg_size_for_bc(code);
1061       int size = declared_signature->return_type()->size();
1062       depth = size - inputs;
1063     }
1064     break;
1065 
1066   case Bytecodes::_multianewarray:
1067     {
1068       ciBytecodeStream iter(method());
1069       iter.reset_to_bci(bci());
1070       iter.next();
1071       inputs = iter.get_dimensions();
1072       assert(rsize == 1, "");
1073       depth = rsize - inputs;
1074     }
1075     break;
1076 
1077   case Bytecodes::_ireturn:
1078   case Bytecodes::_lreturn:
1079   case Bytecodes::_freturn:
1080   case Bytecodes::_dreturn:
1081   case Bytecodes::_areturn:
1082     assert(rsize == -depth, "");
1083     inputs = rsize;
1084     break;
1085 
1086   case Bytecodes::_jsr:
1087   case Bytecodes::_jsr_w:
1088     inputs = 0;
1089     depth  = 1;                  // S.B. depth=1, not zero
1090     break;
1091 
1092   default:
1093     // bytecode produces a typed result
1094     inputs = rsize - depth;
1095     assert(inputs >= 0, "");
1096     break;
1097   }
1098 
1099 #ifdef ASSERT
1100   // spot check
1101   int outputs = depth + inputs;
1102   assert(outputs >= 0, "sanity");
1103   switch (code) {
1104   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1105   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1106   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1107   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1108   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1109   }
1110 #endif //ASSERT
1111 
1112   return true;
1113 }
1114 
1115 
1116 
1117 //------------------------------basic_plus_adr---------------------------------
1118 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1119   // short-circuit a common case
1120   if (offset == intcon(0))  return ptr;
1121   return _gvn.transform( new AddPNode(base, ptr, offset) );
1122 }
1123 
1124 Node* GraphKit::ConvI2L(Node* offset) {
1125   // short-circuit a common case
1126   jint offset_con = find_int_con(offset, Type::OffsetBot);
1127   if (offset_con != Type::OffsetBot) {
1128     return longcon((jlong) offset_con);
1129   }
1130   return _gvn.transform( new ConvI2LNode(offset));
1131 }
1132 
1133 Node* GraphKit::ConvI2UL(Node* offset) {
1134   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1135   if (offset_con != (juint) Type::OffsetBot) {
1136     return longcon((julong) offset_con);
1137   }
1138   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1139   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1140   return _gvn.transform( new AndLNode(conv, mask) );
1141 }
1142 
1143 Node* GraphKit::ConvL2I(Node* offset) {
1144   // short-circuit a common case
1145   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1146   if (offset_con != (jlong)Type::OffsetBot) {
1147     return intcon((int) offset_con);
1148   }
1149   return _gvn.transform( new ConvL2INode(offset));
1150 }
1151 
1152 //-------------------------load_object_klass-----------------------------------
1153 Node* GraphKit::load_object_klass(Node* obj) {
1154   // Special-case a fresh allocation to avoid building nodes:
1155   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1156   if (akls != NULL)  return akls;
1157   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1158   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1159 }
1160 
1161 //-------------------------load_array_length-----------------------------------
1162 Node* GraphKit::load_array_length(Node* array) {
1163   // Special-case a fresh allocation to avoid building nodes:
1164   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1165   Node *alen;
1166   if (alloc == NULL) {
1167     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1168     alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1169   } else {
1170     alen = alloc->Ideal_length();
1171     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1172     if (ccast != alen) {
1173       alen = _gvn.transform(ccast);
1174     }
1175   }
1176   return alen;
1177 }
1178 
1179 //------------------------------do_null_check----------------------------------
1180 // Helper function to do a NULL pointer check.  Returned value is
1181 // the incoming address with NULL casted away.  You are allowed to use the
1182 // not-null value only if you are control dependent on the test.
1183 #ifndef PRODUCT
1184 extern int explicit_null_checks_inserted,
1185            explicit_null_checks_elided;
1186 #endif
1187 Node* GraphKit::null_check_common(Node* value, BasicType type,
1188                                   // optional arguments for variations:
1189                                   bool assert_null,
1190                                   Node* *null_control,
1191                                   bool speculative) {
1192   assert(!assert_null || null_control == NULL, "not both at once");
1193   if (stopped())  return top();
1194   NOT_PRODUCT(explicit_null_checks_inserted++);
1195 
1196   // Construct NULL check
1197   Node *chk = NULL;
1198   switch(type) {
1199     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1200     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1201     case T_ARRAY  : // fall through
1202       type = T_OBJECT;  // simplify further tests
1203     case T_OBJECT : {
1204       const Type *t = _gvn.type( value );
1205 
1206       const TypeOopPtr* tp = t->isa_oopptr();
1207       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1208           // Only for do_null_check, not any of its siblings:
1209           && !assert_null && null_control == NULL) {
1210         // Usually, any field access or invocation on an unloaded oop type
1211         // will simply fail to link, since the statically linked class is
1212         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1213         // the static class is loaded but the sharper oop type is not.
1214         // Rather than checking for this obscure case in lots of places,
1215         // we simply observe that a null check on an unloaded class
1216         // will always be followed by a nonsense operation, so we
1217         // can just issue the uncommon trap here.
1218         // Our access to the unloaded class will only be correct
1219         // after it has been loaded and initialized, which requires
1220         // a trip through the interpreter.
1221 #ifndef PRODUCT
1222         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1223 #endif
1224         uncommon_trap(Deoptimization::Reason_unloaded,
1225                       Deoptimization::Action_reinterpret,
1226                       tp->klass(), "!loaded");
1227         return top();
1228       }
1229 
1230       if (assert_null) {
1231         // See if the type is contained in NULL_PTR.
1232         // If so, then the value is already null.
1233         if (t->higher_equal(TypePtr::NULL_PTR)) {
1234           NOT_PRODUCT(explicit_null_checks_elided++);
1235           return value;           // Elided null assert quickly!
1236         }
1237       } else {
1238         // See if mixing in the NULL pointer changes type.
1239         // If so, then the NULL pointer was not allowed in the original
1240         // type.  In other words, "value" was not-null.
1241         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1242           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1243           NOT_PRODUCT(explicit_null_checks_elided++);
1244           return value;           // Elided null check quickly!
1245         }
1246       }
1247       chk = new CmpPNode( value, null() );
1248       break;
1249     }
1250 
1251     default:
1252       fatal("unexpected type: %s", type2name(type));
1253   }
1254   assert(chk != NULL, "sanity check");
1255   chk = _gvn.transform(chk);
1256 
1257   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1258   BoolNode *btst = new BoolNode( chk, btest);
1259   Node   *tst = _gvn.transform( btst );
1260 
1261   //-----------
1262   // if peephole optimizations occurred, a prior test existed.
1263   // If a prior test existed, maybe it dominates as we can avoid this test.
1264   if (tst != btst && type == T_OBJECT) {
1265     // At this point we want to scan up the CFG to see if we can
1266     // find an identical test (and so avoid this test altogether).
1267     Node *cfg = control();
1268     int depth = 0;
1269     while( depth < 16 ) {       // Limit search depth for speed
1270       if( cfg->Opcode() == Op_IfTrue &&
1271           cfg->in(0)->in(1) == tst ) {
1272         // Found prior test.  Use "cast_not_null" to construct an identical
1273         // CastPP (and hence hash to) as already exists for the prior test.
1274         // Return that casted value.
1275         if (assert_null) {
1276           replace_in_map(value, null());
1277           return null();  // do not issue the redundant test
1278         }
1279         Node *oldcontrol = control();
1280         set_control(cfg);
1281         Node *res = cast_not_null(value);
1282         set_control(oldcontrol);
1283         NOT_PRODUCT(explicit_null_checks_elided++);
1284         return res;
1285       }
1286       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1287       if (cfg == NULL)  break;  // Quit at region nodes
1288       depth++;
1289     }
1290   }
1291 
1292   //-----------
1293   // Branch to failure if null
1294   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1295   Deoptimization::DeoptReason reason;
1296   if (assert_null) {
1297     reason = Deoptimization::Reason_null_assert;
1298   } else if (type == T_OBJECT) {
1299     reason = Deoptimization::reason_null_check(speculative);
1300   } else {
1301     reason = Deoptimization::Reason_div0_check;
1302   }
1303   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1304   // ciMethodData::has_trap_at will return a conservative -1 if any
1305   // must-be-null assertion has failed.  This could cause performance
1306   // problems for a method after its first do_null_assert failure.
1307   // Consider using 'Reason_class_check' instead?
1308 
1309   // To cause an implicit null check, we set the not-null probability
1310   // to the maximum (PROB_MAX).  For an explicit check the probability
1311   // is set to a smaller value.
1312   if (null_control != NULL || too_many_traps(reason)) {
1313     // probability is less likely
1314     ok_prob =  PROB_LIKELY_MAG(3);
1315   } else if (!assert_null &&
1316              (ImplicitNullCheckThreshold > 0) &&
1317              method() != NULL &&
1318              (method()->method_data()->trap_count(reason)
1319               >= (uint)ImplicitNullCheckThreshold)) {
1320     ok_prob =  PROB_LIKELY_MAG(3);
1321   }
1322 
1323   if (null_control != NULL) {
1324     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1325     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1326     set_control(      _gvn.transform( new IfTrueNode(iff)));
1327 #ifndef PRODUCT
1328     if (null_true == top()) {
1329       explicit_null_checks_elided++;
1330     }
1331 #endif
1332     (*null_control) = null_true;
1333   } else {
1334     BuildCutout unless(this, tst, ok_prob);
1335     // Check for optimizer eliding test at parse time
1336     if (stopped()) {
1337       // Failure not possible; do not bother making uncommon trap.
1338       NOT_PRODUCT(explicit_null_checks_elided++);
1339     } else if (assert_null) {
1340       uncommon_trap(reason,
1341                     Deoptimization::Action_make_not_entrant,
1342                     NULL, "assert_null");
1343     } else {
1344       replace_in_map(value, zerocon(type));
1345       builtin_throw(reason);
1346     }
1347   }
1348 
1349   // Must throw exception, fall-thru not possible?
1350   if (stopped()) {
1351     return top();               // No result
1352   }
1353 
1354   if (assert_null) {
1355     // Cast obj to null on this path.
1356     replace_in_map(value, zerocon(type));
1357     return zerocon(type);
1358   }
1359 
1360   // Cast obj to not-null on this path, if there is no null_control.
1361   // (If there is a null_control, a non-null value may come back to haunt us.)
1362   if (type == T_OBJECT) {
1363     Node* cast = cast_not_null(value, false);
1364     if (null_control == NULL || (*null_control) == top())
1365       replace_in_map(value, cast);
1366     value = cast;
1367   }
1368 
1369   return value;
1370 }
1371 
1372 
1373 //------------------------------cast_not_null----------------------------------
1374 // Cast obj to not-null on this path
1375 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1376   const Type *t = _gvn.type(obj);
1377   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1378   // Object is already not-null?
1379   if( t == t_not_null ) return obj;
1380 
1381   Node *cast = new CastPPNode(obj,t_not_null);
1382   cast->init_req(0, control());
1383   cast = _gvn.transform( cast );
1384 
1385   // Scan for instances of 'obj' in the current JVM mapping.
1386   // These instances are known to be not-null after the test.
1387   if (do_replace_in_map)
1388     replace_in_map(obj, cast);
1389 
1390   return cast;                  // Return casted value
1391 }
1392 
1393 
1394 //--------------------------replace_in_map-------------------------------------
1395 void GraphKit::replace_in_map(Node* old, Node* neww) {
1396   if (old == neww) {
1397     return;
1398   }
1399 
1400   map()->replace_edge(old, neww);
1401 
1402   // Note: This operation potentially replaces any edge
1403   // on the map.  This includes locals, stack, and monitors
1404   // of the current (innermost) JVM state.
1405 
1406   // don't let inconsistent types from profiling escape this
1407   // method
1408 
1409   const Type* told = _gvn.type(old);
1410   const Type* tnew = _gvn.type(neww);
1411 
1412   if (!tnew->higher_equal(told)) {
1413     return;
1414   }
1415 
1416   map()->record_replaced_node(old, neww);
1417 }
1418 
1419 
1420 //=============================================================================
1421 //--------------------------------memory---------------------------------------
1422 Node* GraphKit::memory(uint alias_idx) {
1423   MergeMemNode* mem = merged_memory();
1424   Node* p = mem->memory_at(alias_idx);
1425   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1426   return p;
1427 }
1428 
1429 //-----------------------------reset_memory------------------------------------
1430 Node* GraphKit::reset_memory() {
1431   Node* mem = map()->memory();
1432   // do not use this node for any more parsing!
1433   debug_only( map()->set_memory((Node*)NULL) );
1434   return _gvn.transform( mem );
1435 }
1436 
1437 //------------------------------set_all_memory---------------------------------
1438 void GraphKit::set_all_memory(Node* newmem) {
1439   Node* mergemem = MergeMemNode::make(newmem);
1440   gvn().set_type_bottom(mergemem);
1441   map()->set_memory(mergemem);
1442 }
1443 
1444 //------------------------------set_all_memory_call----------------------------
1445 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1446   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1447   set_all_memory(newmem);
1448 }
1449 
1450 //=============================================================================
1451 //
1452 // parser factory methods for MemNodes
1453 //
1454 // These are layered on top of the factory methods in LoadNode and StoreNode,
1455 // and integrate with the parser's memory state and _gvn engine.
1456 //
1457 
1458 // factory methods in "int adr_idx"
1459 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1460                           int adr_idx,
1461                           MemNode::MemOrd mo,
1462                           LoadNode::ControlDependency control_dependency,
1463                           bool require_atomic_access,
1464                           bool unaligned,
1465                           bool mismatched) {
1466   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1467   const TypePtr* adr_type = NULL; // debug-mode-only argument
1468   debug_only(adr_type = C->get_adr_type(adr_idx));
1469   Node* mem = memory(adr_idx);
1470   Node* ld;
1471   if (require_atomic_access && bt == T_LONG) {
1472     ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1473   } else if (require_atomic_access && bt == T_DOUBLE) {
1474     ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1475   } else {
1476     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1477   }
1478   ld = _gvn.transform(ld);
1479   if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1480     // Improve graph before escape analysis and boxing elimination.
1481     record_for_igvn(ld);
1482   }
1483   return ld;
1484 }
1485 
1486 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1487                                 int adr_idx,
1488                                 MemNode::MemOrd mo,
1489                                 bool require_atomic_access,
1490                                 bool unaligned,
1491                                 bool mismatched) {
1492   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1493   const TypePtr* adr_type = NULL;
1494   debug_only(adr_type = C->get_adr_type(adr_idx));
1495   Node *mem = memory(adr_idx);
1496   Node* st;
1497   if (require_atomic_access && bt == T_LONG) {
1498     st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1499   } else if (require_atomic_access && bt == T_DOUBLE) {
1500     st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1501   } else {
1502     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1503   }
1504   if (unaligned) {
1505     st->as_Store()->set_unaligned_access();
1506   }
1507   if (mismatched) {
1508     st->as_Store()->set_mismatched_access();
1509   }
1510   st = _gvn.transform(st);
1511   set_memory(st, adr_idx);
1512   // Back-to-back stores can only remove intermediate store with DU info
1513   // so push on worklist for optimizer.
1514   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1515     record_for_igvn(st);
1516 
1517   return st;
1518 }
1519 
1520 
1521 void GraphKit::pre_barrier(bool do_load,
1522                            Node* ctl,
1523                            Node* obj,
1524                            Node* adr,
1525                            uint  adr_idx,
1526                            Node* val,
1527                            const TypeOopPtr* val_type,
1528                            Node* pre_val,
1529                            BasicType bt) {
1530 
1531   BarrierSet* bs = Universe::heap()->barrier_set();
1532   set_control(ctl);
1533   switch (bs->kind()) {
1534     case BarrierSet::G1SATBCTLogging:
1535       g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1536       break;
1537 
1538     case BarrierSet::CardTableForRS:
1539     case BarrierSet::CardTableExtension:
1540     case BarrierSet::ModRef:
1541       break;
1542 
1543     default      :
1544       ShouldNotReachHere();
1545 
1546   }
1547 }
1548 
1549 bool GraphKit::can_move_pre_barrier() const {
1550   BarrierSet* bs = Universe::heap()->barrier_set();
1551   switch (bs->kind()) {
1552     case BarrierSet::G1SATBCTLogging:
1553       return true; // Can move it if no safepoint
1554 
1555     case BarrierSet::CardTableForRS:
1556     case BarrierSet::CardTableExtension:
1557     case BarrierSet::ModRef:
1558       return true; // There is no pre-barrier
1559 
1560     default      :
1561       ShouldNotReachHere();
1562   }
1563   return false;
1564 }
1565 
1566 void GraphKit::post_barrier(Node* ctl,
1567                             Node* store,
1568                             Node* obj,
1569                             Node* adr,
1570                             uint  adr_idx,
1571                             Node* val,
1572                             BasicType bt,
1573                             bool use_precise) {
1574   BarrierSet* bs = Universe::heap()->barrier_set();
1575   set_control(ctl);
1576   switch (bs->kind()) {
1577     case BarrierSet::G1SATBCTLogging:
1578       g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1579       break;
1580 
1581     case BarrierSet::CardTableForRS:
1582     case BarrierSet::CardTableExtension:
1583       write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1584       break;
1585 
1586     case BarrierSet::ModRef:
1587       break;
1588 
1589     default      :
1590       ShouldNotReachHere();
1591 
1592   }
1593 }
1594 
1595 Node* GraphKit::store_oop(Node* ctl,
1596                           Node* obj,
1597                           Node* adr,
1598                           const TypePtr* adr_type,
1599                           Node* val,
1600                           const TypeOopPtr* val_type,
1601                           BasicType bt,
1602                           bool use_precise,
1603                           MemNode::MemOrd mo,
1604                           bool mismatched) {
1605   // Transformation of a value which could be NULL pointer (CastPP #NULL)
1606   // could be delayed during Parse (for example, in adjust_map_after_if()).
1607   // Execute transformation here to avoid barrier generation in such case.
1608   if (_gvn.type(val) == TypePtr::NULL_PTR)
1609     val = _gvn.makecon(TypePtr::NULL_PTR);
1610 
1611   set_control(ctl);
1612   if (stopped()) return top(); // Dead path ?
1613 
1614   assert(bt == T_OBJECT, "sanity");
1615   assert(val != NULL, "not dead path");
1616   uint adr_idx = C->get_alias_index(adr_type);
1617   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1618 
1619   pre_barrier(true /* do_load */,
1620               control(), obj, adr, adr_idx, val, val_type,
1621               NULL /* pre_val */,
1622               bt);
1623 
1624   Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1625   post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1626   return store;
1627 }
1628 
1629 // Could be an array or object we don't know at compile time (unsafe ref.)
1630 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1631                              Node* obj,   // containing obj
1632                              Node* adr,  // actual adress to store val at
1633                              const TypePtr* adr_type,
1634                              Node* val,
1635                              BasicType bt,
1636                              MemNode::MemOrd mo,
1637                              bool mismatched) {
1638   Compile::AliasType* at = C->alias_type(adr_type);
1639   const TypeOopPtr* val_type = NULL;
1640   if (adr_type->isa_instptr()) {
1641     if (at->field() != NULL) {
1642       // known field.  This code is a copy of the do_put_xxx logic.
1643       ciField* field = at->field();
1644       if (!field->type()->is_loaded()) {
1645         val_type = TypeInstPtr::BOTTOM;
1646       } else {
1647         val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1648       }
1649     }
1650   } else if (adr_type->isa_aryptr()) {
1651     val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1652   }
1653   if (val_type == NULL) {
1654     val_type = TypeInstPtr::BOTTOM;
1655   }
1656   return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1657 }
1658 
1659 
1660 //-------------------------array_element_address-------------------------
1661 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1662                                       const TypeInt* sizetype, Node* ctrl) {
1663   uint shift  = exact_log2(type2aelembytes(elembt));
1664   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1665 
1666   // short-circuit a common case (saves lots of confusing waste motion)
1667   jint idx_con = find_int_con(idx, -1);
1668   if (idx_con >= 0) {
1669     intptr_t offset = header + ((intptr_t)idx_con << shift);
1670     return basic_plus_adr(ary, offset);
1671   }
1672 
1673   // must be correct type for alignment purposes
1674   Node* base  = basic_plus_adr(ary, header);
1675   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1676   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1677   return basic_plus_adr(ary, base, scale);
1678 }
1679 
1680 //-------------------------load_array_element-------------------------
1681 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1682   const Type* elemtype = arytype->elem();
1683   BasicType elembt = elemtype->array_element_basic_type();
1684   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1685   if (elembt == T_NARROWOOP) {
1686     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1687   }
1688   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1689   return ld;
1690 }
1691 
1692 //-------------------------set_arguments_for_java_call-------------------------
1693 // Arguments (pre-popped from the stack) are taken from the JVMS.
1694 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1695   // Add the call arguments:
1696   uint nargs = call->method()->arg_size();
1697   for (uint i = 0; i < nargs; i++) {
1698     Node* arg = argument(i);
1699     call->init_req(i + TypeFunc::Parms, arg);
1700   }
1701 }
1702 
1703 //---------------------------set_edges_for_java_call---------------------------
1704 // Connect a newly created call into the current JVMS.
1705 // A return value node (if any) is returned from set_edges_for_java_call.
1706 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1707 
1708   // Add the predefined inputs:
1709   call->init_req( TypeFunc::Control, control() );
1710   call->init_req( TypeFunc::I_O    , i_o() );
1711   call->init_req( TypeFunc::Memory , reset_memory() );
1712   call->init_req( TypeFunc::FramePtr, frameptr() );
1713   call->init_req( TypeFunc::ReturnAdr, top() );
1714 
1715   add_safepoint_edges(call, must_throw);
1716 
1717   Node* xcall = _gvn.transform(call);
1718 
1719   if (xcall == top()) {
1720     set_control(top());
1721     return;
1722   }
1723   assert(xcall == call, "call identity is stable");
1724 
1725   // Re-use the current map to produce the result.
1726 
1727   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1728   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1729   set_all_memory_call(xcall, separate_io_proj);
1730 
1731   //return xcall;   // no need, caller already has it
1732 }
1733 
1734 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1735   if (stopped())  return top();  // maybe the call folded up?
1736 
1737   // Capture the return value, if any.
1738   Node* ret;
1739   if (call->method() == NULL ||
1740       call->method()->return_type()->basic_type() == T_VOID)
1741         ret = top();
1742   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1743 
1744   // Note:  Since any out-of-line call can produce an exception,
1745   // we always insert an I_O projection from the call into the result.
1746 
1747   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1748 
1749   if (separate_io_proj) {
1750     // The caller requested separate projections be used by the fall
1751     // through and exceptional paths, so replace the projections for
1752     // the fall through path.
1753     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1754     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1755   }
1756   return ret;
1757 }
1758 
1759 //--------------------set_predefined_input_for_runtime_call--------------------
1760 // Reading and setting the memory state is way conservative here.
1761 // The real problem is that I am not doing real Type analysis on memory,
1762 // so I cannot distinguish card mark stores from other stores.  Across a GC
1763 // point the Store Barrier and the card mark memory has to agree.  I cannot
1764 // have a card mark store and its barrier split across the GC point from
1765 // either above or below.  Here I get that to happen by reading ALL of memory.
1766 // A better answer would be to separate out card marks from other memory.
1767 // For now, return the input memory state, so that it can be reused
1768 // after the call, if this call has restricted memory effects.
1769 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1770   // Set fixed predefined input arguments
1771   Node* memory = reset_memory();
1772   call->init_req( TypeFunc::Control,   control()  );
1773   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1774   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1775   call->init_req( TypeFunc::FramePtr,  frameptr() );
1776   call->init_req( TypeFunc::ReturnAdr, top()      );
1777   return memory;
1778 }
1779 
1780 //-------------------set_predefined_output_for_runtime_call--------------------
1781 // Set control and memory (not i_o) from the call.
1782 // If keep_mem is not NULL, use it for the output state,
1783 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1784 // If hook_mem is NULL, this call produces no memory effects at all.
1785 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1786 // then only that memory slice is taken from the call.
1787 // In the last case, we must put an appropriate memory barrier before
1788 // the call, so as to create the correct anti-dependencies on loads
1789 // preceding the call.
1790 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1791                                                       Node* keep_mem,
1792                                                       const TypePtr* hook_mem) {
1793   // no i/o
1794   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1795   if (keep_mem) {
1796     // First clone the existing memory state
1797     set_all_memory(keep_mem);
1798     if (hook_mem != NULL) {
1799       // Make memory for the call
1800       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1801       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1802       // We also use hook_mem to extract specific effects from arraycopy stubs.
1803       set_memory(mem, hook_mem);
1804     }
1805     // ...else the call has NO memory effects.
1806 
1807     // Make sure the call advertises its memory effects precisely.
1808     // This lets us build accurate anti-dependences in gcm.cpp.
1809     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1810            "call node must be constructed correctly");
1811   } else {
1812     assert(hook_mem == NULL, "");
1813     // This is not a "slow path" call; all memory comes from the call.
1814     set_all_memory_call(call);
1815   }
1816 }
1817 
1818 
1819 // Replace the call with the current state of the kit.
1820 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1821   JVMState* ejvms = NULL;
1822   if (has_exceptions()) {
1823     ejvms = transfer_exceptions_into_jvms();
1824   }
1825 
1826   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1827   ReplacedNodes replaced_nodes_exception;
1828   Node* ex_ctl = top();
1829 
1830   SafePointNode* final_state = stop();
1831 
1832   // Find all the needed outputs of this call
1833   CallProjections callprojs;
1834   call->extract_projections(&callprojs, true);
1835 
1836   Node* init_mem = call->in(TypeFunc::Memory);
1837   Node* final_mem = final_state->in(TypeFunc::Memory);
1838   Node* final_ctl = final_state->in(TypeFunc::Control);
1839   Node* final_io = final_state->in(TypeFunc::I_O);
1840 
1841   // Replace all the old call edges with the edges from the inlining result
1842   if (callprojs.fallthrough_catchproj != NULL) {
1843     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1844   }
1845   if (callprojs.fallthrough_memproj != NULL) {
1846     if (final_mem->is_MergeMem()) {
1847       // Parser's exits MergeMem was not transformed but may be optimized
1848       final_mem = _gvn.transform(final_mem);
1849     }
1850     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
1851   }
1852   if (callprojs.fallthrough_ioproj != NULL) {
1853     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
1854   }
1855 
1856   // Replace the result with the new result if it exists and is used
1857   if (callprojs.resproj != NULL && result != NULL) {
1858     C->gvn_replace_by(callprojs.resproj, result);
1859   }
1860 
1861   if (ejvms == NULL) {
1862     // No exception edges to simply kill off those paths
1863     if (callprojs.catchall_catchproj != NULL) {
1864       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1865     }
1866     if (callprojs.catchall_memproj != NULL) {
1867       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1868     }
1869     if (callprojs.catchall_ioproj != NULL) {
1870       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1871     }
1872     // Replace the old exception object with top
1873     if (callprojs.exobj != NULL) {
1874       C->gvn_replace_by(callprojs.exobj, C->top());
1875     }
1876   } else {
1877     GraphKit ekit(ejvms);
1878 
1879     // Load my combined exception state into the kit, with all phis transformed:
1880     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1881     replaced_nodes_exception = ex_map->replaced_nodes();
1882 
1883     Node* ex_oop = ekit.use_exception_state(ex_map);
1884 
1885     if (callprojs.catchall_catchproj != NULL) {
1886       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1887       ex_ctl = ekit.control();
1888     }
1889     if (callprojs.catchall_memproj != NULL) {
1890       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1891     }
1892     if (callprojs.catchall_ioproj != NULL) {
1893       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1894     }
1895 
1896     // Replace the old exception object with the newly created one
1897     if (callprojs.exobj != NULL) {
1898       C->gvn_replace_by(callprojs.exobj, ex_oop);
1899     }
1900   }
1901 
1902   // Disconnect the call from the graph
1903   call->disconnect_inputs(NULL, C);
1904   C->gvn_replace_by(call, C->top());
1905 
1906   // Clean up any MergeMems that feed other MergeMems since the
1907   // optimizer doesn't like that.
1908   if (final_mem->is_MergeMem()) {
1909     Node_List wl;
1910     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1911       Node* m = i.get();
1912       if (m->is_MergeMem() && !wl.contains(m)) {
1913         wl.push(m);
1914       }
1915     }
1916     while (wl.size()  > 0) {
1917       _gvn.transform(wl.pop());
1918     }
1919   }
1920 
1921   if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1922     replaced_nodes.apply(C, final_ctl);
1923   }
1924   if (!ex_ctl->is_top() && do_replaced_nodes) {
1925     replaced_nodes_exception.apply(C, ex_ctl);
1926   }
1927 }
1928 
1929 
1930 //------------------------------increment_counter------------------------------
1931 // for statistics: increment a VM counter by 1
1932 
1933 void GraphKit::increment_counter(address counter_addr) {
1934   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1935   increment_counter(adr1);
1936 }
1937 
1938 void GraphKit::increment_counter(Node* counter_addr) {
1939   int adr_type = Compile::AliasIdxRaw;
1940   Node* ctrl = control();
1941   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1942   Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1943   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1944 }
1945 
1946 
1947 //------------------------------uncommon_trap----------------------------------
1948 // Bail out to the interpreter in mid-method.  Implemented by calling the
1949 // uncommon_trap blob.  This helper function inserts a runtime call with the
1950 // right debug info.
1951 void GraphKit::uncommon_trap(int trap_request,
1952                              ciKlass* klass, const char* comment,
1953                              bool must_throw,
1954                              bool keep_exact_action) {
1955   if (failing())  stop();
1956   if (stopped())  return; // trap reachable?
1957 
1958   // Note:  If ProfileTraps is true, and if a deopt. actually
1959   // occurs here, the runtime will make sure an MDO exists.  There is
1960   // no need to call method()->ensure_method_data() at this point.
1961 
1962   // Set the stack pointer to the right value for reexecution:
1963   set_sp(reexecute_sp());
1964 
1965 #ifdef ASSERT
1966   if (!must_throw) {
1967     // Make sure the stack has at least enough depth to execute
1968     // the current bytecode.
1969     int inputs, ignored_depth;
1970     if (compute_stack_effects(inputs, ignored_depth)) {
1971       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1972              Bytecodes::name(java_bc()), sp(), inputs);
1973     }
1974   }
1975 #endif
1976 
1977   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1978   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1979 
1980   switch (action) {
1981   case Deoptimization::Action_maybe_recompile:
1982   case Deoptimization::Action_reinterpret:
1983     // Temporary fix for 6529811 to allow virtual calls to be sure they
1984     // get the chance to go from mono->bi->mega
1985     if (!keep_exact_action &&
1986         Deoptimization::trap_request_index(trap_request) < 0 &&
1987         too_many_recompiles(reason)) {
1988       // This BCI is causing too many recompilations.
1989       if (C->log() != NULL) {
1990         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
1991                 Deoptimization::trap_reason_name(reason),
1992                 Deoptimization::trap_action_name(action));
1993       }
1994       action = Deoptimization::Action_none;
1995       trap_request = Deoptimization::make_trap_request(reason, action);
1996     } else {
1997       C->set_trap_can_recompile(true);
1998     }
1999     break;
2000   case Deoptimization::Action_make_not_entrant:
2001     C->set_trap_can_recompile(true);
2002     break;
2003 #ifdef ASSERT
2004   case Deoptimization::Action_none:
2005   case Deoptimization::Action_make_not_compilable:
2006     break;
2007   default:
2008     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2009     break;
2010 #endif
2011   }
2012 
2013   if (TraceOptoParse) {
2014     char buf[100];
2015     tty->print_cr("Uncommon trap %s at bci:%d",
2016                   Deoptimization::format_trap_request(buf, sizeof(buf),
2017                                                       trap_request), bci());
2018   }
2019 
2020   CompileLog* log = C->log();
2021   if (log != NULL) {
2022     int kid = (klass == NULL)? -1: log->identify(klass);
2023     log->begin_elem("uncommon_trap bci='%d'", bci());
2024     char buf[100];
2025     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2026                                                           trap_request));
2027     if (kid >= 0)         log->print(" klass='%d'", kid);
2028     if (comment != NULL)  log->print(" comment='%s'", comment);
2029     log->end_elem();
2030   }
2031 
2032   // Make sure any guarding test views this path as very unlikely
2033   Node *i0 = control()->in(0);
2034   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2035     IfNode *iff = i0->as_If();
2036     float f = iff->_prob;   // Get prob
2037     if (control()->Opcode() == Op_IfTrue) {
2038       if (f > PROB_UNLIKELY_MAG(4))
2039         iff->_prob = PROB_MIN;
2040     } else {
2041       if (f < PROB_LIKELY_MAG(4))
2042         iff->_prob = PROB_MAX;
2043     }
2044   }
2045 
2046   // Clear out dead values from the debug info.
2047   kill_dead_locals();
2048 
2049   // Now insert the uncommon trap subroutine call
2050   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2051   const TypePtr* no_memory_effects = NULL;
2052   // Pass the index of the class to be loaded
2053   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2054                                  (must_throw ? RC_MUST_THROW : 0),
2055                                  OptoRuntime::uncommon_trap_Type(),
2056                                  call_addr, "uncommon_trap", no_memory_effects,
2057                                  intcon(trap_request));
2058   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2059          "must extract request correctly from the graph");
2060   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2061 
2062   call->set_req(TypeFunc::ReturnAdr, returnadr());
2063   // The debug info is the only real input to this call.
2064 
2065   // Halt-and-catch fire here.  The above call should never return!
2066   HaltNode* halt = new HaltNode(control(), frameptr());
2067   _gvn.set_type_bottom(halt);
2068   root()->add_req(halt);
2069 
2070   stop_and_kill_map();
2071 }
2072 
2073 
2074 //--------------------------just_allocated_object------------------------------
2075 // Report the object that was just allocated.
2076 // It must be the case that there are no intervening safepoints.
2077 // We use this to determine if an object is so "fresh" that
2078 // it does not require card marks.
2079 Node* GraphKit::just_allocated_object(Node* current_control) {
2080   if (C->recent_alloc_ctl() == current_control)
2081     return C->recent_alloc_obj();
2082   return NULL;
2083 }
2084 
2085 
2086 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2087   // (Note:  TypeFunc::make has a cache that makes this fast.)
2088   const TypeFunc* tf    = TypeFunc::make(dest_method);
2089   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2090   for (int j = 0; j < nargs; j++) {
2091     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2092     if( targ->basic_type() == T_DOUBLE ) {
2093       // If any parameters are doubles, they must be rounded before
2094       // the call, dstore_rounding does gvn.transform
2095       Node *arg = argument(j);
2096       arg = dstore_rounding(arg);
2097       set_argument(j, arg);
2098     }
2099   }
2100 }
2101 
2102 /**
2103  * Record profiling data exact_kls for Node n with the type system so
2104  * that it can propagate it (speculation)
2105  *
2106  * @param n          node that the type applies to
2107  * @param exact_kls  type from profiling
2108  * @param maybe_null did profiling see null?
2109  *
2110  * @return           node with improved type
2111  */
2112 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) {
2113   const Type* current_type = _gvn.type(n);
2114   assert(UseTypeSpeculation, "type speculation must be on");
2115 
2116   const TypePtr* speculative = current_type->speculative();
2117 
2118   // Should the klass from the profile be recorded in the speculative type?
2119   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2120     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2121     const TypeOopPtr* xtype = tklass->as_instance_type();
2122     assert(xtype->klass_is_exact(), "Should be exact");
2123     // Any reason to believe n is not null (from this profiling or a previous one)?
2124     const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2125     // record the new speculative type's depth
2126     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2127     speculative = speculative->with_inline_depth(jvms()->depth());
2128   } else if (current_type->would_improve_ptr(maybe_null)) {
2129     // Profiling report that null was never seen so we can change the
2130     // speculative type to non null ptr.
2131     assert(!maybe_null, "nothing to improve");
2132     if (speculative == NULL) {
2133       speculative = TypePtr::NOTNULL;
2134     } else {
2135       const TypePtr* ptr = TypePtr::NOTNULL;
2136       speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2137     }
2138   }
2139 
2140   if (speculative != current_type->speculative()) {
2141     // Build a type with a speculative type (what we think we know
2142     // about the type but will need a guard when we use it)
2143     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2144     // We're changing the type, we need a new CheckCast node to carry
2145     // the new type. The new type depends on the control: what
2146     // profiling tells us is only valid from here as far as we can
2147     // tell.
2148     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2149     cast = _gvn.transform(cast);
2150     replace_in_map(n, cast);
2151     n = cast;
2152   }
2153 
2154   return n;
2155 }
2156 
2157 /**
2158  * Record profiling data from receiver profiling at an invoke with the
2159  * type system so that it can propagate it (speculation)
2160  *
2161  * @param n  receiver node
2162  *
2163  * @return   node with improved type
2164  */
2165 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2166   if (!UseTypeSpeculation) {
2167     return n;
2168   }
2169   ciKlass* exact_kls = profile_has_unique_klass();
2170   bool maybe_null = true;
2171   if (java_bc() == Bytecodes::_checkcast ||
2172       java_bc() == Bytecodes::_instanceof ||
2173       java_bc() == Bytecodes::_aastore) {
2174     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2175     maybe_null = data == NULL ? true : data->as_BitData()->null_seen();
2176   }
2177   return record_profile_for_speculation(n, exact_kls, maybe_null);
2178 }
2179 
2180 /**
2181  * Record profiling data from argument profiling at an invoke with the
2182  * type system so that it can propagate it (speculation)
2183  *
2184  * @param dest_method  target method for the call
2185  * @param bc           what invoke bytecode is this?
2186  */
2187 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2188   if (!UseTypeSpeculation) {
2189     return;
2190   }
2191   const TypeFunc* tf    = TypeFunc::make(dest_method);
2192   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2193   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2194   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2195     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2196     if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2197       bool maybe_null = true;
2198       ciKlass* better_type = NULL;
2199       if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) {
2200         record_profile_for_speculation(argument(j), better_type, maybe_null);
2201       }
2202       i++;
2203     }
2204   }
2205 }
2206 
2207 /**
2208  * Record profiling data from parameter profiling at an invoke with
2209  * the type system so that it can propagate it (speculation)
2210  */
2211 void GraphKit::record_profiled_parameters_for_speculation() {
2212   if (!UseTypeSpeculation) {
2213     return;
2214   }
2215   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2216     if (_gvn.type(local(i))->isa_oopptr()) {
2217       bool maybe_null = true;
2218       ciKlass* better_type = NULL;
2219       if (method()->parameter_profiled_type(j, better_type, maybe_null)) {
2220         record_profile_for_speculation(local(i), better_type, maybe_null);
2221       }
2222       j++;
2223     }
2224   }
2225 }
2226 
2227 /**
2228  * Record profiling data from return value profiling at an invoke with
2229  * the type system so that it can propagate it (speculation)
2230  */
2231 void GraphKit::record_profiled_return_for_speculation() {
2232   if (!UseTypeSpeculation) {
2233     return;
2234   }
2235   bool maybe_null = true;
2236   ciKlass* better_type = NULL;
2237   if (method()->return_profiled_type(bci(), better_type, maybe_null)) {
2238     // If profiling reports a single type for the return value,
2239     // feed it to the type system so it can propagate it as a
2240     // speculative type
2241     record_profile_for_speculation(stack(sp()-1), better_type, maybe_null);
2242   }
2243 }
2244 
2245 void GraphKit::round_double_result(ciMethod* dest_method) {
2246   // A non-strict method may return a double value which has an extended
2247   // exponent, but this must not be visible in a caller which is 'strict'
2248   // If a strict caller invokes a non-strict callee, round a double result
2249 
2250   BasicType result_type = dest_method->return_type()->basic_type();
2251   assert( method() != NULL, "must have caller context");
2252   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2253     // Destination method's return value is on top of stack
2254     // dstore_rounding() does gvn.transform
2255     Node *result = pop_pair();
2256     result = dstore_rounding(result);
2257     push_pair(result);
2258   }
2259 }
2260 
2261 // rounding for strict float precision conformance
2262 Node* GraphKit::precision_rounding(Node* n) {
2263   return UseStrictFP && _method->flags().is_strict()
2264     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2265     ? _gvn.transform( new RoundFloatNode(0, n) )
2266     : n;
2267 }
2268 
2269 // rounding for strict double precision conformance
2270 Node* GraphKit::dprecision_rounding(Node *n) {
2271   return UseStrictFP && _method->flags().is_strict()
2272     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2273     ? _gvn.transform( new RoundDoubleNode(0, n) )
2274     : n;
2275 }
2276 
2277 // rounding for non-strict double stores
2278 Node* GraphKit::dstore_rounding(Node* n) {
2279   return Matcher::strict_fp_requires_explicit_rounding
2280     && UseSSE <= 1
2281     ? _gvn.transform( new RoundDoubleNode(0, n) )
2282     : n;
2283 }
2284 
2285 //=============================================================================
2286 // Generate a fast path/slow path idiom.  Graph looks like:
2287 // [foo] indicates that 'foo' is a parameter
2288 //
2289 //              [in]     NULL
2290 //                 \    /
2291 //                  CmpP
2292 //                  Bool ne
2293 //                   If
2294 //                  /  \
2295 //              True    False-<2>
2296 //              / |
2297 //             /  cast_not_null
2298 //           Load  |    |   ^
2299 //        [fast_test]   |   |
2300 // gvn to   opt_test    |   |
2301 //          /    \      |  <1>
2302 //      True     False  |
2303 //        |         \\  |
2304 //   [slow_call]     \[fast_result]
2305 //    Ctl   Val       \      \
2306 //     |               \      \
2307 //    Catch       <1>   \      \
2308 //   /    \        ^     \      \
2309 //  Ex    No_Ex    |      \      \
2310 //  |       \   \  |       \ <2>  \
2311 //  ...      \  [slow_res] |  |    \   [null_result]
2312 //            \         \--+--+---  |  |
2313 //             \           | /    \ | /
2314 //              --------Region     Phi
2315 //
2316 //=============================================================================
2317 // Code is structured as a series of driver functions all called 'do_XXX' that
2318 // call a set of helper functions.  Helper functions first, then drivers.
2319 
2320 //------------------------------null_check_oop---------------------------------
2321 // Null check oop.  Set null-path control into Region in slot 3.
2322 // Make a cast-not-nullness use the other not-null control.  Return cast.
2323 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2324                                bool never_see_null,
2325                                bool safe_for_replace,
2326                                bool speculative) {
2327   // Initial NULL check taken path
2328   (*null_control) = top();
2329   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2330 
2331   // Generate uncommon_trap:
2332   if (never_see_null && (*null_control) != top()) {
2333     // If we see an unexpected null at a check-cast we record it and force a
2334     // recompile; the offending check-cast will be compiled to handle NULLs.
2335     // If we see more than one offending BCI, then all checkcasts in the
2336     // method will be compiled to handle NULLs.
2337     PreserveJVMState pjvms(this);
2338     set_control(*null_control);
2339     replace_in_map(value, null());
2340     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2341     uncommon_trap(reason,
2342                   Deoptimization::Action_make_not_entrant);
2343     (*null_control) = top();    // NULL path is dead
2344   }
2345   if ((*null_control) == top() && safe_for_replace) {
2346     replace_in_map(value, cast);
2347   }
2348 
2349   // Cast away null-ness on the result
2350   return cast;
2351 }
2352 
2353 //------------------------------opt_iff----------------------------------------
2354 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2355 // Return slow-path control.
2356 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2357   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2358 
2359   // Fast path taken; set region slot 2
2360   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2361   region->init_req(2,fast_taken); // Capture fast-control
2362 
2363   // Fast path not-taken, i.e. slow path
2364   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2365   return slow_taken;
2366 }
2367 
2368 //-----------------------------make_runtime_call-------------------------------
2369 Node* GraphKit::make_runtime_call(int flags,
2370                                   const TypeFunc* call_type, address call_addr,
2371                                   const char* call_name,
2372                                   const TypePtr* adr_type,
2373                                   // The following parms are all optional.
2374                                   // The first NULL ends the list.
2375                                   Node* parm0, Node* parm1,
2376                                   Node* parm2, Node* parm3,
2377                                   Node* parm4, Node* parm5,
2378                                   Node* parm6, Node* parm7) {
2379   // Slow-path call
2380   bool is_leaf = !(flags & RC_NO_LEAF);
2381   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2382   if (call_name == NULL) {
2383     assert(!is_leaf, "must supply name for leaf");
2384     call_name = OptoRuntime::stub_name(call_addr);
2385   }
2386   CallNode* call;
2387   if (!is_leaf) {
2388     call = new CallStaticJavaNode(call_type, call_addr, call_name,
2389                                            bci(), adr_type);
2390   } else if (flags & RC_NO_FP) {
2391     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2392   } else {
2393     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2394   }
2395 
2396   // The following is similar to set_edges_for_java_call,
2397   // except that the memory effects of the call are restricted to AliasIdxRaw.
2398 
2399   // Slow path call has no side-effects, uses few values
2400   bool wide_in  = !(flags & RC_NARROW_MEM);
2401   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2402 
2403   Node* prev_mem = NULL;
2404   if (wide_in) {
2405     prev_mem = set_predefined_input_for_runtime_call(call);
2406   } else {
2407     assert(!wide_out, "narrow in => narrow out");
2408     Node* narrow_mem = memory(adr_type);
2409     prev_mem = reset_memory();
2410     map()->set_memory(narrow_mem);
2411     set_predefined_input_for_runtime_call(call);
2412   }
2413 
2414   // Hook each parm in order.  Stop looking at the first NULL.
2415   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2416   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2417   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2418   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2419   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2420   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2421   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2422   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2423     /* close each nested if ===> */  } } } } } } } }
2424   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2425 
2426   if (!is_leaf) {
2427     // Non-leaves can block and take safepoints:
2428     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2429   }
2430   // Non-leaves can throw exceptions:
2431   if (has_io) {
2432     call->set_req(TypeFunc::I_O, i_o());
2433   }
2434 
2435   if (flags & RC_UNCOMMON) {
2436     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2437     // (An "if" probability corresponds roughly to an unconditional count.
2438     // Sort of.)
2439     call->set_cnt(PROB_UNLIKELY_MAG(4));
2440   }
2441 
2442   Node* c = _gvn.transform(call);
2443   assert(c == call, "cannot disappear");
2444 
2445   if (wide_out) {
2446     // Slow path call has full side-effects.
2447     set_predefined_output_for_runtime_call(call);
2448   } else {
2449     // Slow path call has few side-effects, and/or sets few values.
2450     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2451   }
2452 
2453   if (has_io) {
2454     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2455   }
2456   return call;
2457 
2458 }
2459 
2460 //------------------------------merge_memory-----------------------------------
2461 // Merge memory from one path into the current memory state.
2462 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2463   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2464     Node* old_slice = mms.force_memory();
2465     Node* new_slice = mms.memory2();
2466     if (old_slice != new_slice) {
2467       PhiNode* phi;
2468       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2469         if (mms.is_empty()) {
2470           // clone base memory Phi's inputs for this memory slice
2471           assert(old_slice == mms.base_memory(), "sanity");
2472           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2473           _gvn.set_type(phi, Type::MEMORY);
2474           for (uint i = 1; i < phi->req(); i++) {
2475             phi->init_req(i, old_slice->in(i));
2476           }
2477         } else {
2478           phi = old_slice->as_Phi(); // Phi was generated already
2479         }
2480       } else {
2481         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2482         _gvn.set_type(phi, Type::MEMORY);
2483       }
2484       phi->set_req(new_path, new_slice);
2485       mms.set_memory(phi);
2486     }
2487   }
2488 }
2489 
2490 //------------------------------make_slow_call_ex------------------------------
2491 // Make the exception handler hookups for the slow call
2492 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2493   if (stopped())  return;
2494 
2495   // Make a catch node with just two handlers:  fall-through and catch-all
2496   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2497   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2498   Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2499   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2500 
2501   { PreserveJVMState pjvms(this);
2502     set_control(excp);
2503     set_i_o(i_o);
2504 
2505     if (excp != top()) {
2506       if (deoptimize) {
2507         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2508         uncommon_trap(Deoptimization::Reason_unhandled,
2509                       Deoptimization::Action_none);
2510       } else {
2511         // Create an exception state also.
2512         // Use an exact type if the caller has specified a specific exception.
2513         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2514         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2515         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2516       }
2517     }
2518   }
2519 
2520   // Get the no-exception control from the CatchNode.
2521   set_control(norm);
2522 }
2523 
2524 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2525   Node* cmp = NULL;
2526   switch(bt) {
2527   case T_INT: cmp = new CmpINode(in1, in2); break;
2528   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2529   default: fatal("unexpected comparison type %s", type2name(bt));
2530   }
2531   gvn->transform(cmp);
2532   Node* bol = gvn->transform(new BoolNode(cmp, test));
2533   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2534   gvn->transform(iff);
2535   if (!bol->is_Con()) gvn->record_for_igvn(iff);
2536   return iff;
2537 }
2538 
2539 
2540 //-------------------------------gen_subtype_check-----------------------------
2541 // Generate a subtyping check.  Takes as input the subtype and supertype.
2542 // Returns 2 values: sets the default control() to the true path and returns
2543 // the false path.  Only reads invariant memory; sets no (visible) memory.
2544 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2545 // but that's not exposed to the optimizer.  This call also doesn't take in an
2546 // Object; if you wish to check an Object you need to load the Object's class
2547 // prior to coming here.
2548 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2549   Compile* C = gvn->C;
2550 
2551   if ((*ctrl)->is_top()) {
2552     return C->top();
2553   }
2554 
2555   // Fast check for identical types, perhaps identical constants.
2556   // The types can even be identical non-constants, in cases
2557   // involving Array.newInstance, Object.clone, etc.
2558   if (subklass == superklass)
2559     return C->top();             // false path is dead; no test needed.
2560 
2561   if (gvn->type(superklass)->singleton()) {
2562     ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2563     ciKlass* subk   = gvn->type(subklass)->is_klassptr()->klass();
2564 
2565     // In the common case of an exact superklass, try to fold up the
2566     // test before generating code.  You may ask, why not just generate
2567     // the code and then let it fold up?  The answer is that the generated
2568     // code will necessarily include null checks, which do not always
2569     // completely fold away.  If they are also needless, then they turn
2570     // into a performance loss.  Example:
2571     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2572     // Here, the type of 'fa' is often exact, so the store check
2573     // of fa[1]=x will fold up, without testing the nullness of x.
2574     switch (C->static_subtype_check(superk, subk)) {
2575     case Compile::SSC_always_false:
2576       {
2577         Node* always_fail = *ctrl;
2578         *ctrl = gvn->C->top();
2579         return always_fail;
2580       }
2581     case Compile::SSC_always_true:
2582       return C->top();
2583     case Compile::SSC_easy_test:
2584       {
2585         // Just do a direct pointer compare and be done.
2586         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2587         *ctrl = gvn->transform(new IfTrueNode(iff));
2588         return gvn->transform(new IfFalseNode(iff));
2589       }
2590     case Compile::SSC_full_test:
2591       break;
2592     default:
2593       ShouldNotReachHere();
2594     }
2595   }
2596 
2597   // %%% Possible further optimization:  Even if the superklass is not exact,
2598   // if the subklass is the unique subtype of the superklass, the check
2599   // will always succeed.  We could leave a dependency behind to ensure this.
2600 
2601   // First load the super-klass's check-offset
2602   Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2603   Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2604   Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2605   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2606   bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2607 
2608   // Load from the sub-klass's super-class display list, or a 1-word cache of
2609   // the secondary superclass list, or a failing value with a sentinel offset
2610   // if the super-klass is an interface or exceptionally deep in the Java
2611   // hierarchy and we have to scan the secondary superclass list the hard way.
2612   // Worst-case type is a little odd: NULL is allowed as a result (usually
2613   // klass loads can never produce a NULL).
2614   Node *chk_off_X = chk_off;
2615 #ifdef _LP64
2616   chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2617 #endif
2618   Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2619   // For some types like interfaces the following loadKlass is from a 1-word
2620   // cache which is mutable so can't use immutable memory.  Other
2621   // types load from the super-class display table which is immutable.
2622   m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2623   Node *kmem = might_be_cache ? m : C->immutable_memory();
2624   Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2625 
2626   // Compile speed common case: ARE a subtype and we canNOT fail
2627   if( superklass == nkls )
2628     return C->top();             // false path is dead; no test needed.
2629 
2630   // See if we get an immediate positive hit.  Happens roughly 83% of the
2631   // time.  Test to see if the value loaded just previously from the subklass
2632   // is exactly the superklass.
2633   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2634   Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2635   *ctrl = gvn->transform(new IfFalseNode(iff1));
2636 
2637   // Compile speed common case: Check for being deterministic right now.  If
2638   // chk_off is a constant and not equal to cacheoff then we are NOT a
2639   // subklass.  In this case we need exactly the 1 test above and we can
2640   // return those results immediately.
2641   if (!might_be_cache) {
2642     Node* not_subtype_ctrl = *ctrl;
2643     *ctrl = iftrue1; // We need exactly the 1 test above
2644     return not_subtype_ctrl;
2645   }
2646 
2647   // Gather the various success & failures here
2648   RegionNode *r_ok_subtype = new RegionNode(4);
2649   gvn->record_for_igvn(r_ok_subtype);
2650   RegionNode *r_not_subtype = new RegionNode(3);
2651   gvn->record_for_igvn(r_not_subtype);
2652 
2653   r_ok_subtype->init_req(1, iftrue1);
2654 
2655   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2656   // is roughly 63% of the remaining cases).  Test to see if the loaded
2657   // check-offset points into the subklass display list or the 1-element
2658   // cache.  If it points to the display (and NOT the cache) and the display
2659   // missed then it's not a subtype.
2660   Node *cacheoff = gvn->intcon(cacheoff_con);
2661   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2662   r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2663   *ctrl = gvn->transform(new IfFalseNode(iff2));
2664 
2665   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2666   // No performance impact (too rare) but allows sharing of secondary arrays
2667   // which has some footprint reduction.
2668   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2669   r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2670   *ctrl = gvn->transform(new IfFalseNode(iff3));
2671 
2672   // -- Roads not taken here: --
2673   // We could also have chosen to perform the self-check at the beginning
2674   // of this code sequence, as the assembler does.  This would not pay off
2675   // the same way, since the optimizer, unlike the assembler, can perform
2676   // static type analysis to fold away many successful self-checks.
2677   // Non-foldable self checks work better here in second position, because
2678   // the initial primary superclass check subsumes a self-check for most
2679   // types.  An exception would be a secondary type like array-of-interface,
2680   // which does not appear in its own primary supertype display.
2681   // Finally, we could have chosen to move the self-check into the
2682   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2683   // dependent manner.  But it is worthwhile to have the check here,
2684   // where it can be perhaps be optimized.  The cost in code space is
2685   // small (register compare, branch).
2686 
2687   // Now do a linear scan of the secondary super-klass array.  Again, no real
2688   // performance impact (too rare) but it's gotta be done.
2689   // Since the code is rarely used, there is no penalty for moving it
2690   // out of line, and it can only improve I-cache density.
2691   // The decision to inline or out-of-line this final check is platform
2692   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2693   Node* psc = gvn->transform(
2694     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2695 
2696   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2697   r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2698   r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2699 
2700   // Return false path; set default control to true path.
2701   *ctrl = gvn->transform(r_ok_subtype);
2702   return gvn->transform(r_not_subtype);
2703 }
2704 
2705 // Profile-driven exact type check:
2706 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2707                                     float prob,
2708                                     Node* *casted_receiver) {
2709   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2710   Node* recv_klass = load_object_klass(receiver);
2711   Node* want_klass = makecon(tklass);
2712   Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2713   Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2714   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2715   set_control( _gvn.transform( new IfTrueNode (iff) ));
2716   Node* fail = _gvn.transform( new IfFalseNode(iff) );
2717 
2718   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2719   assert(recv_xtype->klass_is_exact(), "");
2720 
2721   // Subsume downstream occurrences of receiver with a cast to
2722   // recv_xtype, since now we know what the type will be.
2723   Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2724   (*casted_receiver) = _gvn.transform(cast);
2725   // (User must make the replace_in_map call.)
2726 
2727   return fail;
2728 }
2729 
2730 
2731 //------------------------------seems_never_null-------------------------------
2732 // Use null_seen information if it is available from the profile.
2733 // If we see an unexpected null at a type check we record it and force a
2734 // recompile; the offending check will be recompiled to handle NULLs.
2735 // If we see several offending BCIs, then all checks in the
2736 // method will be recompiled.
2737 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2738   speculating = !_gvn.type(obj)->speculative_maybe_null();
2739   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2740   if (UncommonNullCast               // Cutout for this technique
2741       && obj != null()               // And not the -Xcomp stupid case?
2742       && !too_many_traps(reason)
2743       ) {
2744     if (speculating) {
2745       return true;
2746     }
2747     if (data == NULL)
2748       // Edge case:  no mature data.  Be optimistic here.
2749       return true;
2750     // If the profile has not seen a null, assume it won't happen.
2751     assert(java_bc() == Bytecodes::_checkcast ||
2752            java_bc() == Bytecodes::_instanceof ||
2753            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2754     return !data->as_BitData()->null_seen();
2755   }
2756   speculating = false;
2757   return false;
2758 }
2759 
2760 //------------------------maybe_cast_profiled_receiver-------------------------
2761 // If the profile has seen exactly one type, narrow to exactly that type.
2762 // Subsequent type checks will always fold up.
2763 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2764                                              ciKlass* require_klass,
2765                                              ciKlass* spec_klass,
2766                                              bool safe_for_replace) {
2767   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2768 
2769   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2770 
2771   // Make sure we haven't already deoptimized from this tactic.
2772   if (too_many_traps(reason) || too_many_recompiles(reason))
2773     return NULL;
2774 
2775   // (No, this isn't a call, but it's enough like a virtual call
2776   // to use the same ciMethod accessor to get the profile info...)
2777   // If we have a speculative type use it instead of profiling (which
2778   // may not help us)
2779   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2780   if (exact_kls != NULL) {// no cast failures here
2781     if (require_klass == NULL ||
2782         C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2783       // If we narrow the type to match what the type profile sees or
2784       // the speculative type, we can then remove the rest of the
2785       // cast.
2786       // This is a win, even if the exact_kls is very specific,
2787       // because downstream operations, such as method calls,
2788       // will often benefit from the sharper type.
2789       Node* exact_obj = not_null_obj; // will get updated in place...
2790       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2791                                             &exact_obj);
2792       { PreserveJVMState pjvms(this);
2793         set_control(slow_ctl);
2794         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2795       }
2796       if (safe_for_replace) {
2797         replace_in_map(not_null_obj, exact_obj);
2798       }
2799       return exact_obj;
2800     }
2801     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2802   }
2803 
2804   return NULL;
2805 }
2806 
2807 /**
2808  * Cast obj to type and emit guard unless we had too many traps here
2809  * already
2810  *
2811  * @param obj       node being casted
2812  * @param type      type to cast the node to
2813  * @param not_null  true if we know node cannot be null
2814  */
2815 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2816                                         ciKlass* type,
2817                                         bool not_null) {
2818   if (stopped()) {
2819     return obj;
2820   }
2821 
2822   // type == NULL if profiling tells us this object is always null
2823   if (type != NULL) {
2824     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2825     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2826 
2827     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2828         !too_many_traps(class_reason) &&
2829         !too_many_recompiles(class_reason)) {
2830       Node* not_null_obj = NULL;
2831       // not_null is true if we know the object is not null and
2832       // there's no need for a null check
2833       if (!not_null) {
2834         Node* null_ctl = top();
2835         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2836         assert(null_ctl->is_top(), "no null control here");
2837       } else {
2838         not_null_obj = obj;
2839       }
2840 
2841       Node* exact_obj = not_null_obj;
2842       ciKlass* exact_kls = type;
2843       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2844                                             &exact_obj);
2845       {
2846         PreserveJVMState pjvms(this);
2847         set_control(slow_ctl);
2848         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2849       }
2850       replace_in_map(not_null_obj, exact_obj);
2851       obj = exact_obj;
2852     }
2853   } else {
2854     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2855         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2856       Node* exact_obj = null_assert(obj);
2857       replace_in_map(obj, exact_obj);
2858       obj = exact_obj;
2859     }
2860   }
2861   return obj;
2862 }
2863 
2864 //-------------------------------gen_instanceof--------------------------------
2865 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2866 // and the reflective instance-of call.
2867 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2868   kill_dead_locals();           // Benefit all the uncommon traps
2869   assert( !stopped(), "dead parse path should be checked in callers" );
2870   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2871          "must check for not-null not-dead klass in callers");
2872 
2873   // Make the merge point
2874   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2875   RegionNode* region = new RegionNode(PATH_LIMIT);
2876   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
2877   C->set_has_split_ifs(true); // Has chance for split-if optimization
2878 
2879   ciProfileData* data = NULL;
2880   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2881     data = method()->method_data()->bci_to_data(bci());
2882   }
2883   bool speculative_not_null = false;
2884   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2885                          && seems_never_null(obj, data, speculative_not_null));
2886 
2887   // Null check; get casted pointer; set region slot 3
2888   Node* null_ctl = top();
2889   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2890 
2891   // If not_null_obj is dead, only null-path is taken
2892   if (stopped()) {              // Doing instance-of on a NULL?
2893     set_control(null_ctl);
2894     return intcon(0);
2895   }
2896   region->init_req(_null_path, null_ctl);
2897   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2898   if (null_ctl == top()) {
2899     // Do this eagerly, so that pattern matches like is_diamond_phi
2900     // will work even during parsing.
2901     assert(_null_path == PATH_LIMIT-1, "delete last");
2902     region->del_req(_null_path);
2903     phi   ->del_req(_null_path);
2904   }
2905 
2906   // Do we know the type check always succeed?
2907   bool known_statically = false;
2908   if (_gvn.type(superklass)->singleton()) {
2909     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2910     ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2911     if (subk != NULL && subk->is_loaded()) {
2912       int static_res = C->static_subtype_check(superk, subk);
2913       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2914     }
2915   }
2916 
2917   if (known_statically && UseTypeSpeculation) {
2918     // If we know the type check always succeeds then we don't use the
2919     // profiling data at this bytecode. Don't lose it, feed it to the
2920     // type system as a speculative type.
2921     not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2922   } else {
2923     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2924     // We may not have profiling here or it may not help us. If we
2925     // have a speculative type use it to perform an exact cast.
2926     ciKlass* spec_obj_type = obj_type->speculative_type();
2927     if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2928       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2929       if (stopped()) {            // Profile disagrees with this path.
2930         set_control(null_ctl);    // Null is the only remaining possibility.
2931         return intcon(0);
2932       }
2933       if (cast_obj != NULL) {
2934         not_null_obj = cast_obj;
2935       }
2936     }
2937   }
2938 
2939   // Load the object's klass
2940   Node* obj_klass = load_object_klass(not_null_obj);
2941 
2942   // Generate the subtype check
2943   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2944 
2945   // Plug in the success path to the general merge in slot 1.
2946   region->init_req(_obj_path, control());
2947   phi   ->init_req(_obj_path, intcon(1));
2948 
2949   // Plug in the failing path to the general merge in slot 2.
2950   region->init_req(_fail_path, not_subtype_ctrl);
2951   phi   ->init_req(_fail_path, intcon(0));
2952 
2953   // Return final merged results
2954   set_control( _gvn.transform(region) );
2955   record_for_igvn(region);
2956   return _gvn.transform(phi);
2957 }
2958 
2959 //-------------------------------gen_checkcast---------------------------------
2960 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
2961 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
2962 // uncommon-trap paths work.  Adjust stack after this call.
2963 // If failure_control is supplied and not null, it is filled in with
2964 // the control edge for the cast failure.  Otherwise, an appropriate
2965 // uncommon trap or exception is thrown.
2966 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2967                               Node* *failure_control) {
2968   kill_dead_locals();           // Benefit all the uncommon traps
2969   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2970   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2971 
2972   // Fast cutout:  Check the case that the cast is vacuously true.
2973   // This detects the common cases where the test will short-circuit
2974   // away completely.  We do this before we perform the null check,
2975   // because if the test is going to turn into zero code, we don't
2976   // want a residual null check left around.  (Causes a slowdown,
2977   // for example, in some objArray manipulations, such as a[i]=a[j].)
2978   if (tk->singleton()) {
2979     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2980     if (objtp != NULL && objtp->klass() != NULL) {
2981       switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
2982       case Compile::SSC_always_true:
2983         // If we know the type check always succeed then we don't use
2984         // the profiling data at this bytecode. Don't lose it, feed it
2985         // to the type system as a speculative type.
2986         return record_profiled_receiver_for_speculation(obj);
2987       case Compile::SSC_always_false:
2988         // It needs a null check because a null will *pass* the cast check.
2989         // A non-null value will always produce an exception.
2990         return null_assert(obj);
2991       }
2992     }
2993   }
2994 
2995   ciProfileData* data = NULL;
2996   bool safe_for_replace = false;
2997   if (failure_control == NULL) {        // use MDO in regular case only
2998     assert(java_bc() == Bytecodes::_aastore ||
2999            java_bc() == Bytecodes::_checkcast,
3000            "interpreter profiles type checks only for these BCs");
3001     data = method()->method_data()->bci_to_data(bci());
3002     safe_for_replace = true;
3003   }
3004 
3005   // Make the merge point
3006   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3007   RegionNode* region = new RegionNode(PATH_LIMIT);
3008   Node*       phi    = new PhiNode(region, toop);
3009   C->set_has_split_ifs(true); // Has chance for split-if optimization
3010 
3011   // Use null-cast information if it is available
3012   bool speculative_not_null = false;
3013   bool never_see_null = ((failure_control == NULL)  // regular case only
3014                          && seems_never_null(obj, data, speculative_not_null));
3015 
3016   // Null check; get casted pointer; set region slot 3
3017   Node* null_ctl = top();
3018   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3019 
3020   // If not_null_obj is dead, only null-path is taken
3021   if (stopped()) {              // Doing instance-of on a NULL?
3022     set_control(null_ctl);
3023     return null();
3024   }
3025   region->init_req(_null_path, null_ctl);
3026   phi   ->init_req(_null_path, null());  // Set null path value
3027   if (null_ctl == top()) {
3028     // Do this eagerly, so that pattern matches like is_diamond_phi
3029     // will work even during parsing.
3030     assert(_null_path == PATH_LIMIT-1, "delete last");
3031     region->del_req(_null_path);
3032     phi   ->del_req(_null_path);
3033   }
3034 
3035   Node* cast_obj = NULL;
3036   if (tk->klass_is_exact()) {
3037     // The following optimization tries to statically cast the speculative type of the object
3038     // (for example obtained during profiling) to the type of the superklass and then do a
3039     // dynamic check that the type of the object is what we expect. To work correctly
3040     // for checkcast and aastore the type of superklass should be exact.
3041     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3042     // We may not have profiling here or it may not help us. If we have
3043     // a speculative type use it to perform an exact cast.
3044     ciKlass* spec_obj_type = obj_type->speculative_type();
3045     if (spec_obj_type != NULL || data != NULL) {
3046       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3047       if (cast_obj != NULL) {
3048         if (failure_control != NULL) // failure is now impossible
3049           (*failure_control) = top();
3050         // adjust the type of the phi to the exact klass:
3051         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3052       }
3053     }
3054   }
3055 
3056   if (cast_obj == NULL) {
3057     // Load the object's klass
3058     Node* obj_klass = load_object_klass(not_null_obj);
3059 
3060     // Generate the subtype check
3061     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3062 
3063     // Plug in success path into the merge
3064     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3065     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3066     if (failure_control == NULL) {
3067       if (not_subtype_ctrl != top()) { // If failure is possible
3068         PreserveJVMState pjvms(this);
3069         set_control(not_subtype_ctrl);
3070         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3071       }
3072     } else {
3073       (*failure_control) = not_subtype_ctrl;
3074     }
3075   }
3076 
3077   region->init_req(_obj_path, control());
3078   phi   ->init_req(_obj_path, cast_obj);
3079 
3080   // A merge of NULL or Casted-NotNull obj
3081   Node* res = _gvn.transform(phi);
3082 
3083   // Note I do NOT always 'replace_in_map(obj,result)' here.
3084   //  if( tk->klass()->can_be_primary_super()  )
3085     // This means that if I successfully store an Object into an array-of-String
3086     // I 'forget' that the Object is really now known to be a String.  I have to
3087     // do this because we don't have true union types for interfaces - if I store
3088     // a Baz into an array-of-Interface and then tell the optimizer it's an
3089     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3090     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3091   //  replace_in_map( obj, res );
3092 
3093   // Return final merged results
3094   set_control( _gvn.transform(region) );
3095   record_for_igvn(region);
3096   return res;
3097 }
3098 
3099 //------------------------------next_monitor-----------------------------------
3100 // What number should be given to the next monitor?
3101 int GraphKit::next_monitor() {
3102   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3103   int next = current + C->sync_stack_slots();
3104   // Keep the toplevel high water mark current:
3105   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3106   return current;
3107 }
3108 
3109 //------------------------------insert_mem_bar---------------------------------
3110 // Memory barrier to avoid floating things around
3111 // The membar serves as a pinch point between both control and all memory slices.
3112 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3113   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3114   mb->init_req(TypeFunc::Control, control());
3115   mb->init_req(TypeFunc::Memory,  reset_memory());
3116   Node* membar = _gvn.transform(mb);
3117   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3118   set_all_memory_call(membar);
3119   return membar;
3120 }
3121 
3122 //-------------------------insert_mem_bar_volatile----------------------------
3123 // Memory barrier to avoid floating things around
3124 // The membar serves as a pinch point between both control and memory(alias_idx).
3125 // If you want to make a pinch point on all memory slices, do not use this
3126 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3127 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3128   // When Parse::do_put_xxx updates a volatile field, it appends a series
3129   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3130   // The first membar is on the same memory slice as the field store opcode.
3131   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3132   // All the other membars (for other volatile slices, including AliasIdxBot,
3133   // which stands for all unknown volatile slices) are control-dependent
3134   // on the first membar.  This prevents later volatile loads or stores
3135   // from sliding up past the just-emitted store.
3136 
3137   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3138   mb->set_req(TypeFunc::Control,control());
3139   if (alias_idx == Compile::AliasIdxBot) {
3140     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3141   } else {
3142     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3143     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3144   }
3145   Node* membar = _gvn.transform(mb);
3146   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3147   if (alias_idx == Compile::AliasIdxBot) {
3148     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3149   } else {
3150     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3151   }
3152   return membar;
3153 }
3154 
3155 //------------------------------shared_lock------------------------------------
3156 // Emit locking code.
3157 FastLockNode* GraphKit::shared_lock(Node* obj) {
3158   // bci is either a monitorenter bc or InvocationEntryBci
3159   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3160   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3161 
3162   if( !GenerateSynchronizationCode )
3163     return NULL;                // Not locking things?
3164   if (stopped())                // Dead monitor?
3165     return NULL;
3166 
3167   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3168 
3169   // Box the stack location
3170   Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3171   Node* mem = reset_memory();
3172 
3173   FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3174   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3175     // Create the counters for this fast lock.
3176     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3177   }
3178 
3179   // Create the rtm counters for this fast lock if needed.
3180   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3181 
3182   // Add monitor to debug info for the slow path.  If we block inside the
3183   // slow path and de-opt, we need the monitor hanging around
3184   map()->push_monitor( flock );
3185 
3186   const TypeFunc *tf = LockNode::lock_type();
3187   LockNode *lock = new LockNode(C, tf);
3188 
3189   lock->init_req( TypeFunc::Control, control() );
3190   lock->init_req( TypeFunc::Memory , mem );
3191   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3192   lock->init_req( TypeFunc::FramePtr, frameptr() );
3193   lock->init_req( TypeFunc::ReturnAdr, top() );
3194 
3195   lock->init_req(TypeFunc::Parms + 0, obj);
3196   lock->init_req(TypeFunc::Parms + 1, box);
3197   lock->init_req(TypeFunc::Parms + 2, flock);
3198   add_safepoint_edges(lock);
3199 
3200   lock = _gvn.transform( lock )->as_Lock();
3201 
3202   // lock has no side-effects, sets few values
3203   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3204 
3205   insert_mem_bar(Op_MemBarAcquireLock);
3206 
3207   // Add this to the worklist so that the lock can be eliminated
3208   record_for_igvn(lock);
3209 
3210 #ifndef PRODUCT
3211   if (PrintLockStatistics) {
3212     // Update the counter for this lock.  Don't bother using an atomic
3213     // operation since we don't require absolute accuracy.
3214     lock->create_lock_counter(map()->jvms());
3215     increment_counter(lock->counter()->addr());
3216   }
3217 #endif
3218 
3219   return flock;
3220 }
3221 
3222 
3223 //------------------------------shared_unlock----------------------------------
3224 // Emit unlocking code.
3225 void GraphKit::shared_unlock(Node* box, Node* obj) {
3226   // bci is either a monitorenter bc or InvocationEntryBci
3227   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3228   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3229 
3230   if( !GenerateSynchronizationCode )
3231     return;
3232   if (stopped()) {               // Dead monitor?
3233     map()->pop_monitor();        // Kill monitor from debug info
3234     return;
3235   }
3236 
3237   // Memory barrier to avoid floating things down past the locked region
3238   insert_mem_bar(Op_MemBarReleaseLock);
3239 
3240   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3241   UnlockNode *unlock = new UnlockNode(C, tf);
3242 #ifdef ASSERT
3243   unlock->set_dbg_jvms(sync_jvms());
3244 #endif
3245   uint raw_idx = Compile::AliasIdxRaw;
3246   unlock->init_req( TypeFunc::Control, control() );
3247   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3248   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3249   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3250   unlock->init_req( TypeFunc::ReturnAdr, top() );
3251 
3252   unlock->init_req(TypeFunc::Parms + 0, obj);
3253   unlock->init_req(TypeFunc::Parms + 1, box);
3254   unlock = _gvn.transform(unlock)->as_Unlock();
3255 
3256   Node* mem = reset_memory();
3257 
3258   // unlock has no side-effects, sets few values
3259   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3260 
3261   // Kill monitor from debug info
3262   map()->pop_monitor( );
3263 }
3264 
3265 //-------------------------------get_layout_helper-----------------------------
3266 // If the given klass is a constant or known to be an array,
3267 // fetch the constant layout helper value into constant_value
3268 // and return (Node*)NULL.  Otherwise, load the non-constant
3269 // layout helper value, and return the node which represents it.
3270 // This two-faced routine is useful because allocation sites
3271 // almost always feature constant types.
3272 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3273   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3274   if (!StressReflectiveCode && inst_klass != NULL) {
3275     ciKlass* klass = inst_klass->klass();
3276     bool    xklass = inst_klass->klass_is_exact();
3277     if (xklass || klass->is_array_klass()) {
3278       jint lhelper = klass->layout_helper();
3279       if (lhelper != Klass::_lh_neutral_value) {
3280         constant_value = lhelper;
3281         return (Node*) NULL;
3282       }
3283     }
3284   }
3285   constant_value = Klass::_lh_neutral_value;  // put in a known value
3286   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3287   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3288 }
3289 
3290 // We just put in an allocate/initialize with a big raw-memory effect.
3291 // Hook selected additional alias categories on the initialization.
3292 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3293                                 MergeMemNode* init_in_merge,
3294                                 Node* init_out_raw) {
3295   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3296   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3297 
3298   Node* prevmem = kit.memory(alias_idx);
3299   init_in_merge->set_memory_at(alias_idx, prevmem);
3300   kit.set_memory(init_out_raw, alias_idx);
3301 }
3302 
3303 //---------------------------set_output_for_allocation-------------------------
3304 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3305                                           const TypeOopPtr* oop_type,
3306                                           bool deoptimize_on_exception) {
3307   int rawidx = Compile::AliasIdxRaw;
3308   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3309   add_safepoint_edges(alloc);
3310   Node* allocx = _gvn.transform(alloc);
3311   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3312   // create memory projection for i_o
3313   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3314   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3315 
3316   // create a memory projection as for the normal control path
3317   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3318   set_memory(malloc, rawidx);
3319 
3320   // a normal slow-call doesn't change i_o, but an allocation does
3321   // we create a separate i_o projection for the normal control path
3322   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3323   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3324 
3325   // put in an initialization barrier
3326   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3327                                                  rawoop)->as_Initialize();
3328   assert(alloc->initialization() == init,  "2-way macro link must work");
3329   assert(init ->allocation()     == alloc, "2-way macro link must work");
3330   {
3331     // Extract memory strands which may participate in the new object's
3332     // initialization, and source them from the new InitializeNode.
3333     // This will allow us to observe initializations when they occur,
3334     // and link them properly (as a group) to the InitializeNode.
3335     assert(init->in(InitializeNode::Memory) == malloc, "");
3336     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3337     init->set_req(InitializeNode::Memory, minit_in);
3338     record_for_igvn(minit_in); // fold it up later, if possible
3339     Node* minit_out = memory(rawidx);
3340     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3341     if (oop_type->isa_aryptr()) {
3342       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3343       int            elemidx  = C->get_alias_index(telemref);
3344       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3345     } else if (oop_type->isa_instptr()) {
3346       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3347       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3348         ciField* field = ik->nonstatic_field_at(i);
3349         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3350           continue;  // do not bother to track really large numbers of fields
3351         // Find (or create) the alias category for this field:
3352         int fieldidx = C->alias_type(field)->index();
3353         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3354       }
3355     }
3356   }
3357 
3358   // Cast raw oop to the real thing...
3359   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3360   javaoop = _gvn.transform(javaoop);
3361   C->set_recent_alloc(control(), javaoop);
3362   assert(just_allocated_object(control()) == javaoop, "just allocated");
3363 
3364 #ifdef ASSERT
3365   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3366     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3367            "Ideal_allocation works");
3368     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3369            "Ideal_allocation works");
3370     if (alloc->is_AllocateArray()) {
3371       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3372              "Ideal_allocation works");
3373       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3374              "Ideal_allocation works");
3375     } else {
3376       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3377     }
3378   }
3379 #endif //ASSERT
3380 
3381   return javaoop;
3382 }
3383 
3384 //---------------------------new_instance--------------------------------------
3385 // This routine takes a klass_node which may be constant (for a static type)
3386 // or may be non-constant (for reflective code).  It will work equally well
3387 // for either, and the graph will fold nicely if the optimizer later reduces
3388 // the type to a constant.
3389 // The optional arguments are for specialized use by intrinsics:
3390 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3391 //  - If 'return_size_val', report the the total object size to the caller.
3392 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3393 Node* GraphKit::new_instance(Node* klass_node,
3394                              Node* extra_slow_test,
3395                              Node* *return_size_val,
3396                              bool deoptimize_on_exception) {
3397   // Compute size in doublewords
3398   // The size is always an integral number of doublewords, represented
3399   // as a positive bytewise size stored in the klass's layout_helper.
3400   // The layout_helper also encodes (in a low bit) the need for a slow path.
3401   jint  layout_con = Klass::_lh_neutral_value;
3402   Node* layout_val = get_layout_helper(klass_node, layout_con);
3403   int   layout_is_con = (layout_val == NULL);
3404 
3405   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3406   // Generate the initial go-slow test.  It's either ALWAYS (return a
3407   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3408   // case) a computed value derived from the layout_helper.
3409   Node* initial_slow_test = NULL;
3410   if (layout_is_con) {
3411     assert(!StressReflectiveCode, "stress mode does not use these paths");
3412     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3413     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3414   } else {   // reflective case
3415     // This reflective path is used by Unsafe.allocateInstance.
3416     // (It may be stress-tested by specifying StressReflectiveCode.)
3417     // Basically, we want to get into the VM is there's an illegal argument.
3418     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3419     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3420     if (extra_slow_test != intcon(0)) {
3421       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3422     }
3423     // (Macro-expander will further convert this to a Bool, if necessary.)
3424   }
3425 
3426   // Find the size in bytes.  This is easy; it's the layout_helper.
3427   // The size value must be valid even if the slow path is taken.
3428   Node* size = NULL;
3429   if (layout_is_con) {
3430     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3431   } else {   // reflective case
3432     // This reflective path is used by clone and Unsafe.allocateInstance.
3433     size = ConvI2X(layout_val);
3434 
3435     // Clear the low bits to extract layout_helper_size_in_bytes:
3436     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3437     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3438     size = _gvn.transform( new AndXNode(size, mask) );
3439   }
3440   if (return_size_val != NULL) {
3441     (*return_size_val) = size;
3442   }
3443 
3444   // This is a precise notnull oop of the klass.
3445   // (Actually, it need not be precise if this is a reflective allocation.)
3446   // It's what we cast the result to.
3447   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3448   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3449   const TypeOopPtr* oop_type = tklass->as_instance_type();
3450 
3451   // Now generate allocation code
3452 
3453   // The entire memory state is needed for slow path of the allocation
3454   // since GC and deoptimization can happened.
3455   Node *mem = reset_memory();
3456   set_all_memory(mem); // Create new memory state
3457 
3458   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3459                                          control(), mem, i_o(),
3460                                          size, klass_node,
3461                                          initial_slow_test);
3462 
3463   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3464 }
3465 
3466 //-------------------------------new_array-------------------------------------
3467 // helper for both newarray and anewarray
3468 // The 'length' parameter is (obviously) the length of the array.
3469 // See comments on new_instance for the meaning of the other arguments.
3470 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3471                           Node* length,         // number of array elements
3472                           int   nargs,          // number of arguments to push back for uncommon trap
3473                           Node* *return_size_val,
3474                           bool deoptimize_on_exception) {
3475   jint  layout_con = Klass::_lh_neutral_value;
3476   Node* layout_val = get_layout_helper(klass_node, layout_con);
3477   int   layout_is_con = (layout_val == NULL);
3478 
3479   if (!layout_is_con && !StressReflectiveCode &&
3480       !too_many_traps(Deoptimization::Reason_class_check)) {
3481     // This is a reflective array creation site.
3482     // Optimistically assume that it is a subtype of Object[],
3483     // so that we can fold up all the address arithmetic.
3484     layout_con = Klass::array_layout_helper(T_OBJECT);
3485     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3486     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3487     { BuildCutout unless(this, bol_lh, PROB_MAX);
3488       inc_sp(nargs);
3489       uncommon_trap(Deoptimization::Reason_class_check,
3490                     Deoptimization::Action_maybe_recompile);
3491     }
3492     layout_val = NULL;
3493     layout_is_con = true;
3494   }
3495 
3496   // Generate the initial go-slow test.  Make sure we do not overflow
3497   // if length is huge (near 2Gig) or negative!  We do not need
3498   // exact double-words here, just a close approximation of needed
3499   // double-words.  We can't add any offset or rounding bits, lest we
3500   // take a size -1 of bytes and make it positive.  Use an unsigned
3501   // compare, so negative sizes look hugely positive.
3502   int fast_size_limit = FastAllocateSizeLimit;
3503   if (layout_is_con) {
3504     assert(!StressReflectiveCode, "stress mode does not use these paths");
3505     // Increase the size limit if we have exact knowledge of array type.
3506     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3507     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3508   }
3509 
3510   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3511   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3512 
3513   // --- Size Computation ---
3514   // array_size = round_to_heap(array_header + (length << elem_shift));
3515   // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3516   // and round_to(x, y) == ((x + y-1) & ~(y-1))
3517   // The rounding mask is strength-reduced, if possible.
3518   int round_mask = MinObjAlignmentInBytes - 1;
3519   Node* header_size = NULL;
3520   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3521   // (T_BYTE has the weakest alignment and size restrictions...)
3522   if (layout_is_con) {
3523     int       hsize  = Klass::layout_helper_header_size(layout_con);
3524     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3525     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3526     if ((round_mask & ~right_n_bits(eshift)) == 0)
3527       round_mask = 0;  // strength-reduce it if it goes away completely
3528     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3529     assert(header_size_min <= hsize, "generic minimum is smallest");
3530     header_size_min = hsize;
3531     header_size = intcon(hsize + round_mask);
3532   } else {
3533     Node* hss   = intcon(Klass::_lh_header_size_shift);
3534     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3535     Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3536     hsize       = _gvn.transform( new AndINode(hsize, hsm) );
3537     Node* mask  = intcon(round_mask);
3538     header_size = _gvn.transform( new AddINode(hsize, mask) );
3539   }
3540 
3541   Node* elem_shift = NULL;
3542   if (layout_is_con) {
3543     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3544     if (eshift != 0)
3545       elem_shift = intcon(eshift);
3546   } else {
3547     // There is no need to mask or shift this value.
3548     // The semantics of LShiftINode include an implicit mask to 0x1F.
3549     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3550     elem_shift = layout_val;
3551   }
3552 
3553   // Transition to native address size for all offset calculations:
3554   Node* lengthx = ConvI2X(length);
3555   Node* headerx = ConvI2X(header_size);
3556 #ifdef _LP64
3557   { const TypeInt* tilen = _gvn.find_int_type(length);
3558     if (tilen != NULL && tilen->_lo < 0) {
3559       // Add a manual constraint to a positive range.  Cf. array_element_address.
3560       jint size_max = fast_size_limit;
3561       if (size_max > tilen->_hi)  size_max = tilen->_hi;
3562       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3563 
3564       // Only do a narrow I2L conversion if the range check passed.
3565       IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3566       _gvn.transform(iff);
3567       RegionNode* region = new RegionNode(3);
3568       _gvn.set_type(region, Type::CONTROL);
3569       lengthx = new PhiNode(region, TypeLong::LONG);
3570       _gvn.set_type(lengthx, TypeLong::LONG);
3571 
3572       // Range check passed. Use ConvI2L node with narrow type.
3573       Node* passed = IfFalse(iff);
3574       region->init_req(1, passed);
3575       // Make I2L conversion control dependent to prevent it from
3576       // floating above the range check during loop optimizations.
3577       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3578 
3579       // Range check failed. Use ConvI2L with wide type because length may be invalid.
3580       region->init_req(2, IfTrue(iff));
3581       lengthx->init_req(2, ConvI2X(length));
3582 
3583       set_control(region);
3584       record_for_igvn(region);
3585       record_for_igvn(lengthx);
3586     }
3587   }
3588 #endif
3589 
3590   // Combine header size (plus rounding) and body size.  Then round down.
3591   // This computation cannot overflow, because it is used only in two
3592   // places, one where the length is sharply limited, and the other
3593   // after a successful allocation.
3594   Node* abody = lengthx;
3595   if (elem_shift != NULL)
3596     abody     = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3597   Node* size  = _gvn.transform( new AddXNode(headerx, abody) );
3598   if (round_mask != 0) {
3599     Node* mask = MakeConX(~round_mask);
3600     size       = _gvn.transform( new AndXNode(size, mask) );
3601   }
3602   // else if round_mask == 0, the size computation is self-rounding
3603 
3604   if (return_size_val != NULL) {
3605     // This is the size
3606     (*return_size_val) = size;
3607   }
3608 
3609   // Now generate allocation code
3610 
3611   // The entire memory state is needed for slow path of the allocation
3612   // since GC and deoptimization can happened.
3613   Node *mem = reset_memory();
3614   set_all_memory(mem); // Create new memory state
3615 
3616   if (initial_slow_test->is_Bool()) {
3617     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3618     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3619   }
3620 
3621   // Create the AllocateArrayNode and its result projections
3622   AllocateArrayNode* alloc
3623     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3624                             control(), mem, i_o(),
3625                             size, klass_node,
3626                             initial_slow_test,
3627                             length);
3628 
3629   // Cast to correct type.  Note that the klass_node may be constant or not,
3630   // and in the latter case the actual array type will be inexact also.
3631   // (This happens via a non-constant argument to inline_native_newArray.)
3632   // In any case, the value of klass_node provides the desired array type.
3633   const TypeInt* length_type = _gvn.find_int_type(length);
3634   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3635   if (ary_type->isa_aryptr() && length_type != NULL) {
3636     // Try to get a better type than POS for the size
3637     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3638   }
3639 
3640   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3641 
3642   // Cast length on remaining path to be as narrow as possible
3643   if (map()->find_edge(length) >= 0) {
3644     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3645     if (ccast != length) {
3646       _gvn.set_type_bottom(ccast);
3647       record_for_igvn(ccast);
3648       replace_in_map(length, ccast);
3649     }
3650   }
3651 
3652   return javaoop;
3653 }
3654 
3655 // The following "Ideal_foo" functions are placed here because they recognize
3656 // the graph shapes created by the functions immediately above.
3657 
3658 //---------------------------Ideal_allocation----------------------------------
3659 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3660 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3661   if (ptr == NULL) {     // reduce dumb test in callers
3662     return NULL;
3663   }
3664   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3665     ptr = ptr->in(1);
3666     if (ptr == NULL) return NULL;
3667   }
3668   // Return NULL for allocations with several casts:
3669   //   j.l.reflect.Array.newInstance(jobject, jint)
3670   //   Object.clone()
3671   // to keep more precise type from last cast.
3672   if (ptr->is_Proj()) {
3673     Node* allo = ptr->in(0);
3674     if (allo != NULL && allo->is_Allocate()) {
3675       return allo->as_Allocate();
3676     }
3677   }
3678   // Report failure to match.
3679   return NULL;
3680 }
3681 
3682 // Fancy version which also strips off an offset (and reports it to caller).
3683 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3684                                              intptr_t& offset) {
3685   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3686   if (base == NULL)  return NULL;
3687   return Ideal_allocation(base, phase);
3688 }
3689 
3690 // Trace Initialize <- Proj[Parm] <- Allocate
3691 AllocateNode* InitializeNode::allocation() {
3692   Node* rawoop = in(InitializeNode::RawAddress);
3693   if (rawoop->is_Proj()) {
3694     Node* alloc = rawoop->in(0);
3695     if (alloc->is_Allocate()) {
3696       return alloc->as_Allocate();
3697     }
3698   }
3699   return NULL;
3700 }
3701 
3702 // Trace Allocate -> Proj[Parm] -> Initialize
3703 InitializeNode* AllocateNode::initialization() {
3704   ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3705   if (rawoop == NULL)  return NULL;
3706   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3707     Node* init = rawoop->fast_out(i);
3708     if (init->is_Initialize()) {
3709       assert(init->as_Initialize()->allocation() == this, "2-way link");
3710       return init->as_Initialize();
3711     }
3712   }
3713   return NULL;
3714 }
3715 
3716 //----------------------------- loop predicates ---------------------------
3717 
3718 //------------------------------add_predicate_impl----------------------------
3719 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3720   // Too many traps seen?
3721   if (too_many_traps(reason)) {
3722 #ifdef ASSERT
3723     if (TraceLoopPredicate) {
3724       int tc = C->trap_count(reason);
3725       tty->print("too many traps=%s tcount=%d in ",
3726                     Deoptimization::trap_reason_name(reason), tc);
3727       method()->print(); // which method has too many predicate traps
3728       tty->cr();
3729     }
3730 #endif
3731     // We cannot afford to take more traps here,
3732     // do not generate predicate.
3733     return;
3734   }
3735 
3736   Node *cont    = _gvn.intcon(1);
3737   Node* opq     = _gvn.transform(new Opaque1Node(C, cont));
3738   Node *bol     = _gvn.transform(new Conv2BNode(opq));
3739   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3740   Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3741   C->add_predicate_opaq(opq);
3742   {
3743     PreserveJVMState pjvms(this);
3744     set_control(iffalse);
3745     inc_sp(nargs);
3746     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3747   }
3748   Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3749   set_control(iftrue);
3750 }
3751 
3752 //------------------------------add_predicate---------------------------------
3753 void GraphKit::add_predicate(int nargs) {
3754   if (UseLoopPredicate) {
3755     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3756   }
3757   // loop's limit check predicate should be near the loop.
3758   add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3759 }
3760 
3761 //----------------------------- store barriers ----------------------------
3762 #define __ ideal.
3763 
3764 void GraphKit::sync_kit(IdealKit& ideal) {
3765   set_all_memory(__ merged_memory());
3766   set_i_o(__ i_o());
3767   set_control(__ ctrl());
3768 }
3769 
3770 void GraphKit::final_sync(IdealKit& ideal) {
3771   // Final sync IdealKit and graphKit.
3772   sync_kit(ideal);
3773 }
3774 
3775 Node* GraphKit::byte_map_base_node() {
3776   // Get base of card map
3777   CardTableModRefBS* ct =
3778     barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3779   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
3780   if (ct->byte_map_base != NULL) {
3781     return makecon(TypeRawPtr::make((address)ct->byte_map_base));
3782   } else {
3783     return null();
3784   }
3785 }
3786 
3787 // vanilla/CMS post barrier
3788 // Insert a write-barrier store.  This is to let generational GC work; we have
3789 // to flag all oop-stores before the next GC point.
3790 void GraphKit::write_barrier_post(Node* oop_store,
3791                                   Node* obj,
3792                                   Node* adr,
3793                                   uint  adr_idx,
3794                                   Node* val,
3795                                   bool use_precise) {
3796   // No store check needed if we're storing a NULL or an old object
3797   // (latter case is probably a string constant). The concurrent
3798   // mark sweep garbage collector, however, needs to have all nonNull
3799   // oop updates flagged via card-marks.
3800   if (val != NULL && val->is_Con()) {
3801     // must be either an oop or NULL
3802     const Type* t = val->bottom_type();
3803     if (t == TypePtr::NULL_PTR || t == Type::TOP)
3804       // stores of null never (?) need barriers
3805       return;
3806   }
3807 
3808   if (use_ReduceInitialCardMarks()
3809       && obj == just_allocated_object(control())) {
3810     // We can skip marks on a freshly-allocated object in Eden.
3811     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3812     // That routine informs GC to take appropriate compensating steps,
3813     // upon a slow-path allocation, so as to make this card-mark
3814     // elision safe.
3815     return;
3816   }
3817 
3818   if (!use_precise) {
3819     // All card marks for a (non-array) instance are in one place:
3820     adr = obj;
3821   }
3822   // (Else it's an array (or unknown), and we want more precise card marks.)
3823   assert(adr != NULL, "");
3824 
3825   IdealKit ideal(this, true);
3826 
3827   // Convert the pointer to an int prior to doing math on it
3828   Node* cast = __ CastPX(__ ctrl(), adr);
3829 
3830   // Divide by card size
3831   assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3832          "Only one we handle so far.");
3833   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3834 
3835   // Combine card table base and card offset
3836   Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3837 
3838   // Get the alias_index for raw card-mark memory
3839   int adr_type = Compile::AliasIdxRaw;
3840   Node*   zero = __ ConI(0); // Dirty card value
3841   BasicType bt = T_BYTE;
3842 
3843   if (UseConcMarkSweepGC && UseCondCardMark) {
3844     insert_mem_bar(Op_MemBarVolatile);   // StoreLoad barrier
3845     __ sync_kit(this);
3846   }
3847 
3848   if (UseCondCardMark) {
3849     // The classic GC reference write barrier is typically implemented
3850     // as a store into the global card mark table.  Unfortunately
3851     // unconditional stores can result in false sharing and excessive
3852     // coherence traffic as well as false transactional aborts.
3853     // UseCondCardMark enables MP "polite" conditional card mark
3854     // stores.  In theory we could relax the load from ctrl() to
3855     // no_ctrl, but that doesn't buy much latitude.
3856     Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3857     __ if_then(card_val, BoolTest::ne, zero);
3858   }
3859 
3860   // Smash zero into card
3861   if( !UseConcMarkSweepGC ) {
3862     __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3863   } else {
3864     // Specialized path for CM store barrier
3865     __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3866   }
3867 
3868   if (UseCondCardMark) {
3869     __ end_if();
3870   }
3871 
3872   // Final sync IdealKit and GraphKit.
3873   final_sync(ideal);
3874 }
3875 /*
3876  * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3877  * required by SATB to make sure all objects live at the start of the
3878  * marking are kept alive, all reference updates need to any previous
3879  * reference stored before writing.
3880  *
3881  * If the previous value is NULL there is no need to save the old value.
3882  * References that are NULL are filtered during runtime by the barrier
3883  * code to avoid unnecessary queuing.
3884  *
3885  * However in the case of newly allocated objects it might be possible to
3886  * prove that the reference about to be overwritten is NULL during compile
3887  * time and avoid adding the barrier code completely.
3888  *
3889  * The compiler needs to determine that the object in which a field is about
3890  * to be written is newly allocated, and that no prior store to the same field
3891  * has happened since the allocation.
3892  *
3893  * Returns true if the pre-barrier can be removed
3894  */
3895 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3896                                          BasicType bt, uint adr_idx) {
3897   intptr_t offset = 0;
3898   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3899   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3900 
3901   if (offset == Type::OffsetBot) {
3902     return false; // cannot unalias unless there are precise offsets
3903   }
3904 
3905   if (alloc == NULL) {
3906     return false; // No allocation found
3907   }
3908 
3909   intptr_t size_in_bytes = type2aelembytes(bt);
3910 
3911   Node* mem = memory(adr_idx); // start searching here...
3912 
3913   for (int cnt = 0; cnt < 50; cnt++) {
3914 
3915     if (mem->is_Store()) {
3916 
3917       Node* st_adr = mem->in(MemNode::Address);
3918       intptr_t st_offset = 0;
3919       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
3920 
3921       if (st_base == NULL) {
3922         break; // inscrutable pointer
3923       }
3924 
3925       // Break we have found a store with same base and offset as ours so break
3926       if (st_base == base && st_offset == offset) {
3927         break;
3928       }
3929 
3930       if (st_offset != offset && st_offset != Type::OffsetBot) {
3931         const int MAX_STORE = BytesPerLong;
3932         if (st_offset >= offset + size_in_bytes ||
3933             st_offset <= offset - MAX_STORE ||
3934             st_offset <= offset - mem->as_Store()->memory_size()) {
3935           // Success:  The offsets are provably independent.
3936           // (You may ask, why not just test st_offset != offset and be done?
3937           // The answer is that stores of different sizes can co-exist
3938           // in the same sequence of RawMem effects.  We sometimes initialize
3939           // a whole 'tile' of array elements with a single jint or jlong.)
3940           mem = mem->in(MemNode::Memory);
3941           continue; // advance through independent store memory
3942         }
3943       }
3944 
3945       if (st_base != base
3946           && MemNode::detect_ptr_independence(base, alloc, st_base,
3947                                               AllocateNode::Ideal_allocation(st_base, phase),
3948                                               phase)) {
3949         // Success:  The bases are provably independent.
3950         mem = mem->in(MemNode::Memory);
3951         continue; // advance through independent store memory
3952       }
3953     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
3954 
3955       InitializeNode* st_init = mem->in(0)->as_Initialize();
3956       AllocateNode* st_alloc = st_init->allocation();
3957 
3958       // Make sure that we are looking at the same allocation site.
3959       // The alloc variable is guaranteed to not be null here from earlier check.
3960       if (alloc == st_alloc) {
3961         // Check that the initialization is storing NULL so that no previous store
3962         // has been moved up and directly write a reference
3963         Node* captured_store = st_init->find_captured_store(offset,
3964                                                             type2aelembytes(T_OBJECT),
3965                                                             phase);
3966         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
3967           return true;
3968         }
3969       }
3970     }
3971 
3972     // Unless there is an explicit 'continue', we must bail out here,
3973     // because 'mem' is an inscrutable memory state (e.g., a call).
3974     break;
3975   }
3976 
3977   return false;
3978 }
3979 
3980 // G1 pre/post barriers
3981 void GraphKit::g1_write_barrier_pre(bool do_load,
3982                                     Node* obj,
3983                                     Node* adr,
3984                                     uint alias_idx,
3985                                     Node* val,
3986                                     const TypeOopPtr* val_type,
3987                                     Node* pre_val,
3988                                     BasicType bt) {
3989 
3990   // Some sanity checks
3991   // Note: val is unused in this routine.
3992 
3993   if (do_load) {
3994     // We need to generate the load of the previous value
3995     assert(obj != NULL, "must have a base");
3996     assert(adr != NULL, "where are loading from?");
3997     assert(pre_val == NULL, "loaded already?");
3998     assert(val_type != NULL, "need a type");
3999 
4000     if (use_ReduceInitialCardMarks()
4001         && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4002       return;
4003     }
4004 
4005   } else {
4006     // In this case both val_type and alias_idx are unused.
4007     assert(pre_val != NULL, "must be loaded already");
4008     // Nothing to be done if pre_val is null.
4009     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4010     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4011   }
4012   assert(bt == T_OBJECT, "or we shouldn't be here");
4013 
4014   IdealKit ideal(this, true);
4015 
4016   Node* tls = __ thread(); // ThreadLocalStorage
4017 
4018   Node* no_ctrl = NULL;
4019   Node* no_base = __ top();
4020   Node* zero  = __ ConI(0);
4021   Node* zeroX = __ ConX(0);
4022 
4023   float likely  = PROB_LIKELY(0.999);
4024   float unlikely  = PROB_UNLIKELY(0.999);
4025 
4026   BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4027   assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4028 
4029   // Offsets into the thread
4030   const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +  // 648
4031                                           SATBMarkQueue::byte_offset_of_active());
4032   const int index_offset   = in_bytes(JavaThread::satb_mark_queue_offset() +  // 656
4033                                           SATBMarkQueue::byte_offset_of_index());
4034   const int buffer_offset  = in_bytes(JavaThread::satb_mark_queue_offset() +  // 652
4035                                           SATBMarkQueue::byte_offset_of_buf());
4036 
4037   // Now the actual pointers into the thread
4038   Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4039   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
4040   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
4041 
4042   // Now some of the values
4043   Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4044 
4045   // if (!marking)
4046   __ if_then(marking, BoolTest::ne, zero, unlikely); {
4047     BasicType index_bt = TypeX_X->basic_type();
4048     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4049     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4050 
4051     if (do_load) {
4052       // load original value
4053       // alias_idx correct??
4054       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4055     }
4056 
4057     // if (pre_val != NULL)
4058     __ if_then(pre_val, BoolTest::ne, null()); {
4059       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4060 
4061       // is the queue for this thread full?
4062       __ if_then(index, BoolTest::ne, zeroX, likely); {
4063 
4064         // decrement the index
4065         Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4066 
4067         // Now get the buffer location we will log the previous value into and store it
4068         Node *log_addr = __ AddP(no_base, buffer, next_index);
4069         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4070         // update the index
4071         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4072 
4073       } __ else_(); {
4074 
4075         // logging buffer is full, call the runtime
4076         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4077         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4078       } __ end_if();  // (!index)
4079     } __ end_if();  // (pre_val != NULL)
4080   } __ end_if();  // (!marking)
4081 
4082   // Final sync IdealKit and GraphKit.
4083   final_sync(ideal);
4084 }
4085 
4086 /*
4087  * G1 similar to any GC with a Young Generation requires a way to keep track of
4088  * references from Old Generation to Young Generation to make sure all live
4089  * objects are found. G1 also requires to keep track of object references
4090  * between different regions to enable evacuation of old regions, which is done
4091  * as part of mixed collections. References are tracked in remembered sets and
4092  * is continuously updated as reference are written to with the help of the
4093  * post-barrier.
4094  *
4095  * To reduce the number of updates to the remembered set the post-barrier
4096  * filters updates to fields in objects located in the Young Generation,
4097  * the same region as the reference, when the NULL is being written or
4098  * if the card is already marked as dirty by an earlier write.
4099  *
4100  * Under certain circumstances it is possible to avoid generating the
4101  * post-barrier completely if it is possible during compile time to prove
4102  * the object is newly allocated and that no safepoint exists between the
4103  * allocation and the store.
4104  *
4105  * In the case of slow allocation the allocation code must handle the barrier
4106  * as part of the allocation in the case the allocated object is not located
4107  * in the nursery, this would happen for humongous objects. This is similar to
4108  * how CMS is required to handle this case, see the comments for the method
4109  * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
4110  * A deferred card mark is required for these objects and handled in the above
4111  * mentioned methods.
4112  *
4113  * Returns true if the post barrier can be removed
4114  */
4115 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4116                                           Node* adr) {
4117   intptr_t      offset = 0;
4118   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4119   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
4120 
4121   if (offset == Type::OffsetBot) {
4122     return false; // cannot unalias unless there are precise offsets
4123   }
4124 
4125   if (alloc == NULL) {
4126      return false; // No allocation found
4127   }
4128 
4129   // Start search from Store node
4130   Node* mem = store->in(MemNode::Control);
4131   if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4132 
4133     InitializeNode* st_init = mem->in(0)->as_Initialize();
4134     AllocateNode*  st_alloc = st_init->allocation();
4135 
4136     // Make sure we are looking at the same allocation
4137     if (alloc == st_alloc) {
4138       return true;
4139     }
4140   }
4141 
4142   return false;
4143 }
4144 
4145 //
4146 // Update the card table and add card address to the queue
4147 //
4148 void GraphKit::g1_mark_card(IdealKit& ideal,
4149                             Node* card_adr,
4150                             Node* oop_store,
4151                             uint oop_alias_idx,
4152                             Node* index,
4153                             Node* index_adr,
4154                             Node* buffer,
4155                             const TypeFunc* tf) {
4156 
4157   Node* zero  = __ ConI(0);
4158   Node* zeroX = __ ConX(0);
4159   Node* no_base = __ top();
4160   BasicType card_bt = T_BYTE;
4161   // Smash zero into card. MUST BE ORDERED WRT TO STORE
4162   __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4163 
4164   //  Now do the queue work
4165   __ if_then(index, BoolTest::ne, zeroX); {
4166 
4167     Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4168     Node* log_addr = __ AddP(no_base, buffer, next_index);
4169 
4170     // Order, see storeCM.
4171     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4172     __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4173 
4174   } __ else_(); {
4175     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4176   } __ end_if();
4177 
4178 }
4179 
4180 void GraphKit::g1_write_barrier_post(Node* oop_store,
4181                                      Node* obj,
4182                                      Node* adr,
4183                                      uint alias_idx,
4184                                      Node* val,
4185                                      BasicType bt,
4186                                      bool use_precise) {
4187   // If we are writing a NULL then we need no post barrier
4188 
4189   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4190     // Must be NULL
4191     const Type* t = val->bottom_type();
4192     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4193     // No post barrier if writing NULLx
4194     return;
4195   }
4196 
4197   if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4198     // We can skip marks on a freshly-allocated object in Eden.
4199     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
4200     // That routine informs GC to take appropriate compensating steps,
4201     // upon a slow-path allocation, so as to make this card-mark
4202     // elision safe.
4203     return;
4204   }
4205 
4206   if (use_ReduceInitialCardMarks()
4207       && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4208     return;
4209   }
4210 
4211   if (!use_precise) {
4212     // All card marks for a (non-array) instance are in one place:
4213     adr = obj;
4214   }
4215   // (Else it's an array (or unknown), and we want more precise card marks.)
4216   assert(adr != NULL, "");
4217 
4218   IdealKit ideal(this, true);
4219 
4220   Node* tls = __ thread(); // ThreadLocalStorage
4221 
4222   Node* no_base = __ top();
4223   float likely  = PROB_LIKELY(0.999);
4224   float unlikely  = PROB_UNLIKELY(0.999);
4225   Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4226   Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4227   Node* zeroX = __ ConX(0);
4228 
4229   // Get the alias_index for raw card-mark memory
4230   const TypePtr* card_type = TypeRawPtr::BOTTOM;
4231 
4232   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4233 
4234   // Offsets into the thread
4235   const int index_offset  = in_bytes(JavaThread::dirty_card_queue_offset() +
4236                                      DirtyCardQueue::byte_offset_of_index());
4237   const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4238                                      DirtyCardQueue::byte_offset_of_buf());
4239 
4240   // Pointers into the thread
4241 
4242   Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4243   Node* index_adr =  __ AddP(no_base, tls, __ ConX(index_offset));
4244 
4245   // Now some values
4246   // Use ctrl to avoid hoisting these values past a safepoint, which could
4247   // potentially reset these fields in the JavaThread.
4248   Node* index  = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4249   Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4250 
4251   // Convert the store obj pointer to an int prior to doing math on it
4252   // Must use ctrl to prevent "integerized oop" existing across safepoint
4253   Node* cast =  __ CastPX(__ ctrl(), adr);
4254 
4255   // Divide pointer by card size
4256   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4257 
4258   // Combine card table base and card offset
4259   Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4260 
4261   // If we know the value being stored does it cross regions?
4262 
4263   if (val != NULL) {
4264     // Does the store cause us to cross regions?
4265 
4266     // Should be able to do an unsigned compare of region_size instead of
4267     // and extra shift. Do we have an unsigned compare??
4268     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4269     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4270 
4271     // if (xor_res == 0) same region so skip
4272     __ if_then(xor_res, BoolTest::ne, zeroX); {
4273 
4274       // No barrier if we are storing a NULL
4275       __ if_then(val, BoolTest::ne, null(), unlikely); {
4276 
4277         // Ok must mark the card if not already dirty
4278 
4279         // load the original value of the card
4280         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4281 
4282         __ if_then(card_val, BoolTest::ne, young_card); {
4283           sync_kit(ideal);
4284           // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4285           insert_mem_bar(Op_MemBarVolatile, oop_store);
4286           __ sync_kit(this);
4287 
4288           Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4289           __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4290             g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4291           } __ end_if();
4292         } __ end_if();
4293       } __ end_if();
4294     } __ end_if();
4295   } else {
4296     // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4297     // We don't need a barrier here if the destination is a newly allocated object
4298     // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4299     // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()).
4300     assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4301     Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4302     __ if_then(card_val, BoolTest::ne, young_card); {
4303       g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4304     } __ end_if();
4305   }
4306 
4307   // Final sync IdealKit and GraphKit.
4308   final_sync(ideal);
4309 }
4310 #undef __
4311 
4312 
4313 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4314   Node* len = load_array_length(load_String_value(ctrl, str));
4315   Node* coder = load_String_coder(ctrl, str);
4316   // Divide length by 2 if coder is UTF16
4317   return _gvn.transform(new RShiftINode(len, coder));
4318 }
4319 
4320 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4321   int value_offset = java_lang_String::value_offset_in_bytes();
4322   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4323                                                      false, NULL, 0);
4324   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4325   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4326                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4327                                                   ciTypeArrayKlass::make(T_BYTE), true, 0);
4328   int value_field_idx = C->get_alias_index(value_field_type);
4329   Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4330                          value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4331   // String.value field is known to be @Stable.
4332   if (UseImplicitStableValues) {
4333     load = cast_array_to_stable(load, value_type);
4334   }
4335   return load;
4336 }
4337 
4338 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4339   if (!CompactStrings) {
4340     return intcon(java_lang_String::CODER_UTF16);
4341   }
4342   int coder_offset = java_lang_String::coder_offset_in_bytes();
4343   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4344                                                      false, NULL, 0);
4345   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4346   int coder_field_idx = C->get_alias_index(coder_field_type);
4347   return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4348                    TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4349 }
4350 
4351 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4352   int value_offset = java_lang_String::value_offset_in_bytes();
4353   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4354                                                      false, NULL, 0);
4355   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4356   store_oop_to_object(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
4357       value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4358 }
4359 
4360 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4361   int coder_offset = java_lang_String::coder_offset_in_bytes();
4362   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4363                                                      false, NULL, 0);
4364   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4365   int coder_field_idx = C->get_alias_index(coder_field_type);
4366   store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4367                   value, T_BYTE, coder_field_idx, MemNode::unordered);
4368 }
4369 
4370 // Capture src and dst memory state with a MergeMemNode
4371 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4372   if (src_type == dst_type) {
4373     // Types are equal, we don't need a MergeMemNode
4374     return memory(src_type);
4375   }
4376   MergeMemNode* merge = MergeMemNode::make(map()->memory());
4377   record_for_igvn(merge); // fold it up later, if possible
4378   int src_idx = C->get_alias_index(src_type);
4379   int dst_idx = C->get_alias_index(dst_type);
4380   merge->set_memory_at(src_idx, memory(src_idx));
4381   merge->set_memory_at(dst_idx, memory(dst_idx));
4382   return merge;
4383 }
4384 
4385 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4386   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4387   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4388   // If input and output memory types differ, capture both states to preserve
4389   // the dependency between preceding and subsequent loads/stores.
4390   // For example, the following program:
4391   //  StoreB
4392   //  compress_string
4393   //  LoadB
4394   // has this memory graph (use->def):
4395   //  LoadB -> compress_string -> CharMem
4396   //             ... -> StoreB -> ByteMem
4397   // The intrinsic hides the dependency between LoadB and StoreB, causing
4398   // the load to read from memory not containing the result of the StoreB.
4399   // The correct memory graph should look like this:
4400   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4401   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4402   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4403   Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4404   set_memory(res_mem, TypeAryPtr::BYTES);
4405   return str;
4406 }
4407 
4408 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4409   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4410   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4411   // Capture src and dst memory (see comment in 'compress_string').
4412   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4413   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4414   set_memory(_gvn.transform(str), dst_type);
4415 }
4416 
4417 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4418   /**
4419    * int i_char = start;
4420    * for (int i_byte = 0; i_byte < count; i_byte++) {
4421    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
4422    * }
4423    */
4424   add_predicate();
4425   RegionNode* head = new RegionNode(3);
4426   head->init_req(1, control());
4427   gvn().set_type(head, Type::CONTROL);
4428   record_for_igvn(head);
4429 
4430   Node* i_byte = new PhiNode(head, TypeInt::INT);
4431   i_byte->init_req(1, intcon(0));
4432   gvn().set_type(i_byte, TypeInt::INT);
4433   record_for_igvn(i_byte);
4434 
4435   Node* i_char = new PhiNode(head, TypeInt::INT);
4436   i_char->init_req(1, start);
4437   gvn().set_type(i_char, TypeInt::INT);
4438   record_for_igvn(i_char);
4439 
4440   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4441   gvn().set_type(mem, Type::MEMORY);
4442   record_for_igvn(mem);
4443   set_control(head);
4444   set_memory(mem, TypeAryPtr::BYTES);
4445   Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4446   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4447                              AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4448                              false, false, true /* mismatched */);
4449 
4450   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4451   head->init_req(2, IfTrue(iff));
4452   mem->init_req(2, st);
4453   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4454   i_char->init_req(2, AddI(i_char, intcon(2)));
4455 
4456   set_control(IfFalse(iff));
4457   set_memory(st, TypeAryPtr::BYTES);
4458 }
4459 
4460 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4461   if (!field->is_constant()) {
4462     return NULL; // Field not marked as constant.
4463   }
4464   ciInstance* holder = NULL;
4465   if (!field->is_static()) {
4466     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4467     if (const_oop != NULL && const_oop->is_instance()) {
4468       holder = const_oop->as_instance();
4469     }
4470   }
4471   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4472                                                         /*is_unsigned_load=*/false);
4473   if (con_type != NULL) {
4474     return makecon(con_type);
4475   }
4476   return NULL;
4477 }
4478 
4479 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4480   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4481   // assumption of CCP analysis.
4482   return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4483 }