1 /*
   2  * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "compiler/compileLog.hpp"
  27 #include "gc/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 "opto/addnode.hpp"
  33 #include "opto/castnode.hpp"
  34 #include "opto/convertnode.hpp"
  35 #include "opto/graphKit.hpp"
  36 #include "opto/idealKit.hpp"
  37 #include "opto/intrinsicnode.hpp"
  38 #include "opto/locknode.hpp"
  39 #include "opto/machnode.hpp"
  40 #include "opto/opaquenode.hpp"
  41 #include "opto/parse.hpp"
  42 #include "opto/rootnode.hpp"
  43 #include "opto/runtime.hpp"
  44 #include "runtime/deoptimization.hpp"
  45 #include "runtime/sharedRuntime.hpp"
  46 
  47 //----------------------------GraphKit-----------------------------------------
  48 // Main utility constructor.
  49 GraphKit::GraphKit(JVMState* jvms)
  50   : Phase(Phase::Parser),
  51     _env(C->env()),
  52     _gvn(*C->initial_gvn())
  53 {
  54   _exceptions = jvms->map()->next_exception();
  55   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  56   set_jvms(jvms);
  57 }
  58 
  59 // Private constructor for parser.
  60 GraphKit::GraphKit()
  61   : Phase(Phase::Parser),
  62     _env(C->env()),
  63     _gvn(*C->initial_gvn())
  64 {
  65   _exceptions = NULL;
  66   set_map(NULL);
  67   debug_only(_sp = -99);
  68   debug_only(set_bci(-99));
  69 }
  70 
  71 
  72 
  73 //---------------------------clean_stack---------------------------------------
  74 // Clear away rubbish from the stack area of the JVM state.
  75 // This destroys any arguments that may be waiting on the stack.
  76 void GraphKit::clean_stack(int from_sp) {
  77   SafePointNode* map      = this->map();
  78   JVMState*      jvms     = this->jvms();
  79   int            stk_size = jvms->stk_size();
  80   int            stkoff   = jvms->stkoff();
  81   Node*          top      = this->top();
  82   for (int i = from_sp; i < stk_size; i++) {
  83     if (map->in(stkoff + i) != top) {
  84       map->set_req(stkoff + i, top);
  85     }
  86   }
  87 }
  88 
  89 
  90 //--------------------------------sync_jvms-----------------------------------
  91 // Make sure our current jvms agrees with our parse state.
  92 JVMState* GraphKit::sync_jvms() const {
  93   JVMState* jvms = this->jvms();
  94   jvms->set_bci(bci());       // Record the new bci in the JVMState
  95   jvms->set_sp(sp());         // Record the new sp in the JVMState
  96   assert(jvms_in_sync(), "jvms is now in sync");
  97   return jvms;
  98 }
  99 
 100 //--------------------------------sync_jvms_for_reexecute---------------------
 101 // Make sure our current jvms agrees with our parse state.  This version
 102 // uses the reexecute_sp for reexecuting bytecodes.
 103 JVMState* GraphKit::sync_jvms_for_reexecute() {
 104   JVMState* jvms = this->jvms();
 105   jvms->set_bci(bci());          // Record the new bci in the JVMState
 106   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
 107   return jvms;
 108 }
 109 
 110 #ifdef ASSERT
 111 bool GraphKit::jvms_in_sync() const {
 112   Parse* parse = is_Parse();
 113   if (parse == NULL) {
 114     if (bci() !=      jvms()->bci())          return false;
 115     if (sp()  != (int)jvms()->sp())           return false;
 116     return true;
 117   }
 118   if (jvms()->method() != parse->method())    return false;
 119   if (jvms()->bci()    != parse->bci())       return false;
 120   int jvms_sp = jvms()->sp();
 121   if (jvms_sp          != parse->sp())        return false;
 122   int jvms_depth = jvms()->depth();
 123   if (jvms_depth       != parse->depth())     return false;
 124   return true;
 125 }
 126 
 127 // Local helper checks for special internal merge points
 128 // used to accumulate and merge exception states.
 129 // They are marked by the region's in(0) edge being the map itself.
 130 // Such merge points must never "escape" into the parser at large,
 131 // until they have been handed to gvn.transform.
 132 static bool is_hidden_merge(Node* reg) {
 133   if (reg == NULL)  return false;
 134   if (reg->is_Phi()) {
 135     reg = reg->in(0);
 136     if (reg == NULL)  return false;
 137   }
 138   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 139 }
 140 
 141 void GraphKit::verify_map() const {
 142   if (map() == NULL)  return;  // null map is OK
 143   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 144   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 145   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 146 }
 147 
 148 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 149   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 150   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 151 }
 152 #endif
 153 
 154 //---------------------------stop_and_kill_map---------------------------------
 155 // Set _map to NULL, signalling a stop to further bytecode execution.
 156 // First smash the current map's control to a constant, to mark it dead.
 157 void GraphKit::stop_and_kill_map() {
 158   SafePointNode* dead_map = stop();
 159   if (dead_map != NULL) {
 160     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
 161     assert(dead_map->is_killed(), "must be so marked");
 162   }
 163 }
 164 
 165 
 166 //--------------------------------stopped--------------------------------------
 167 // Tell if _map is NULL, or control is top.
 168 bool GraphKit::stopped() {
 169   if (map() == NULL)           return true;
 170   else if (control() == top()) return true;
 171   else                         return false;
 172 }
 173 
 174 
 175 //-----------------------------has_ex_handler----------------------------------
 176 // Tell if this method or any caller method has exception handlers.
 177 bool GraphKit::has_ex_handler() {
 178   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 179     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 180       return true;
 181     }
 182   }
 183   return false;
 184 }
 185 
 186 //------------------------------save_ex_oop------------------------------------
 187 // Save an exception without blowing stack contents or other JVM state.
 188 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 189   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 190   ex_map->add_req(ex_oop);
 191   debug_only(verify_exception_state(ex_map));
 192 }
 193 
 194 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 195   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 196   Node* ex_oop = ex_map->in(ex_map->req()-1);
 197   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 198   return ex_oop;
 199 }
 200 
 201 //-----------------------------saved_ex_oop------------------------------------
 202 // Recover a saved exception from its map.
 203 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 204   return common_saved_ex_oop(ex_map, false);
 205 }
 206 
 207 //--------------------------clear_saved_ex_oop---------------------------------
 208 // Erase a previously saved exception from its map.
 209 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 210   return common_saved_ex_oop(ex_map, true);
 211 }
 212 
 213 #ifdef ASSERT
 214 //---------------------------has_saved_ex_oop----------------------------------
 215 // Erase a previously saved exception from its map.
 216 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 217   return ex_map->req() == ex_map->jvms()->endoff()+1;
 218 }
 219 #endif
 220 
 221 //-------------------------make_exception_state--------------------------------
 222 // Turn the current JVM state into an exception state, appending the ex_oop.
 223 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 224   sync_jvms();
 225   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 226   set_saved_ex_oop(ex_map, ex_oop);
 227   return ex_map;
 228 }
 229 
 230 
 231 //--------------------------add_exception_state--------------------------------
 232 // Add an exception to my list of exceptions.
 233 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 234   if (ex_map == NULL || ex_map->control() == top()) {
 235     return;
 236   }
 237 #ifdef ASSERT
 238   verify_exception_state(ex_map);
 239   if (has_exceptions()) {
 240     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 241   }
 242 #endif
 243 
 244   // If there is already an exception of exactly this type, merge with it.
 245   // In particular, null-checks and other low-level exceptions common up here.
 246   Node*       ex_oop  = saved_ex_oop(ex_map);
 247   const Type* ex_type = _gvn.type(ex_oop);
 248   if (ex_oop == top()) {
 249     // No action needed.
 250     return;
 251   }
 252   assert(ex_type->isa_instptr(), "exception must be an instance");
 253   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 254     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 255     // We check sp also because call bytecodes can generate exceptions
 256     // both before and after arguments are popped!
 257     if (ex_type2 == ex_type
 258         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 259       combine_exception_states(ex_map, e2);
 260       return;
 261     }
 262   }
 263 
 264   // No pre-existing exception of the same type.  Chain it on the list.
 265   push_exception_state(ex_map);
 266 }
 267 
 268 //-----------------------add_exception_states_from-----------------------------
 269 void GraphKit::add_exception_states_from(JVMState* jvms) {
 270   SafePointNode* ex_map = jvms->map()->next_exception();
 271   if (ex_map != NULL) {
 272     jvms->map()->set_next_exception(NULL);
 273     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 274       next_map = ex_map->next_exception();
 275       ex_map->set_next_exception(NULL);
 276       add_exception_state(ex_map);
 277     }
 278   }
 279 }
 280 
 281 //-----------------------transfer_exceptions_into_jvms-------------------------
 282 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 283   if (map() == NULL) {
 284     // We need a JVMS to carry the exceptions, but the map has gone away.
 285     // Create a scratch JVMS, cloned from any of the exception states...
 286     if (has_exceptions()) {
 287       _map = _exceptions;
 288       _map = clone_map();
 289       _map->set_next_exception(NULL);
 290       clear_saved_ex_oop(_map);
 291       debug_only(verify_map());
 292     } else {
 293       // ...or created from scratch
 294       JVMState* jvms = new (C) JVMState(_method, NULL);
 295       jvms->set_bci(_bci);
 296       jvms->set_sp(_sp);
 297       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
 298       set_jvms(jvms);
 299       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 300       set_all_memory(top());
 301       while (map()->req() < jvms->endoff())  map()->add_req(top());
 302     }
 303     // (This is a kludge, in case you didn't notice.)
 304     set_control(top());
 305   }
 306   JVMState* jvms = sync_jvms();
 307   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 308   jvms->map()->set_next_exception(_exceptions);
 309   _exceptions = NULL;   // done with this set of exceptions
 310   return jvms;
 311 }
 312 
 313 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 314   assert(is_hidden_merge(dstphi), "must be a special merge node");
 315   assert(is_hidden_merge(srcphi), "must be a special merge node");
 316   uint limit = srcphi->req();
 317   for (uint i = PhiNode::Input; i < limit; i++) {
 318     dstphi->add_req(srcphi->in(i));
 319   }
 320 }
 321 static inline void add_one_req(Node* dstphi, Node* src) {
 322   assert(is_hidden_merge(dstphi), "must be a special merge node");
 323   assert(!is_hidden_merge(src), "must not be a special merge node");
 324   dstphi->add_req(src);
 325 }
 326 
 327 //-----------------------combine_exception_states------------------------------
 328 // This helper function combines exception states by building phis on a
 329 // specially marked state-merging region.  These regions and phis are
 330 // untransformed, and can build up gradually.  The region is marked by
 331 // having a control input of its exception map, rather than NULL.  Such
 332 // regions do not appear except in this function, and in use_exception_state.
 333 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 334   if (failing())  return;  // dying anyway...
 335   JVMState* ex_jvms = ex_map->_jvms;
 336   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 337   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 338   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 339   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 340   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
 341   assert(ex_map->req() == phi_map->req(), "matching maps");
 342   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 343   Node*         hidden_merge_mark = root();
 344   Node*         region  = phi_map->control();
 345   MergeMemNode* phi_mem = phi_map->merged_memory();
 346   MergeMemNode* ex_mem  = ex_map->merged_memory();
 347   if (region->in(0) != hidden_merge_mark) {
 348     // The control input is not (yet) a specially-marked region in phi_map.
 349     // Make it so, and build some phis.
 350     region = new RegionNode(2);
 351     _gvn.set_type(region, Type::CONTROL);
 352     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 353     region->init_req(1, phi_map->control());
 354     phi_map->set_control(region);
 355     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 356     record_for_igvn(io_phi);
 357     _gvn.set_type(io_phi, Type::ABIO);
 358     phi_map->set_i_o(io_phi);
 359     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 360       Node* m = mms.memory();
 361       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 362       record_for_igvn(m_phi);
 363       _gvn.set_type(m_phi, Type::MEMORY);
 364       mms.set_memory(m_phi);
 365     }
 366   }
 367 
 368   // Either or both of phi_map and ex_map might already be converted into phis.
 369   Node* ex_control = ex_map->control();
 370   // if there is special marking on ex_map also, we add multiple edges from src
 371   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 372   // how wide was the destination phi_map, originally?
 373   uint orig_width = region->req();
 374 
 375   if (add_multiple) {
 376     add_n_reqs(region, ex_control);
 377     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 378   } else {
 379     // ex_map has no merges, so we just add single edges everywhere
 380     add_one_req(region, ex_control);
 381     add_one_req(phi_map->i_o(), ex_map->i_o());
 382   }
 383   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 384     if (mms.is_empty()) {
 385       // get a copy of the base memory, and patch some inputs into it
 386       const TypePtr* adr_type = mms.adr_type(C);
 387       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 388       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 389       mms.set_memory(phi);
 390       // Prepare to append interesting stuff onto the newly sliced phi:
 391       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 392     }
 393     // Append stuff from ex_map:
 394     if (add_multiple) {
 395       add_n_reqs(mms.memory(), mms.memory2());
 396     } else {
 397       add_one_req(mms.memory(), mms.memory2());
 398     }
 399   }
 400   uint limit = ex_map->req();
 401   for (uint i = TypeFunc::Parms; i < limit; i++) {
 402     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 403     if (i == tos)  i = ex_jvms->monoff();
 404     Node* src = ex_map->in(i);
 405     Node* dst = phi_map->in(i);
 406     if (src != dst) {
 407       PhiNode* phi;
 408       if (dst->in(0) != region) {
 409         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 410         record_for_igvn(phi);
 411         _gvn.set_type(phi, phi->type());
 412         phi_map->set_req(i, dst);
 413         // Prepare to append interesting stuff onto the new phi:
 414         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 415       } else {
 416         assert(dst->is_Phi(), "nobody else uses a hidden region");
 417         phi = dst->as_Phi();
 418       }
 419       if (add_multiple && src->in(0) == ex_control) {
 420         // Both are phis.
 421         add_n_reqs(dst, src);
 422       } else {
 423         while (dst->req() < region->req())  add_one_req(dst, src);
 424       }
 425       const Type* srctype = _gvn.type(src);
 426       if (phi->type() != srctype) {
 427         const Type* dsttype = phi->type()->meet_speculative(srctype);
 428         if (phi->type() != dsttype) {
 429           phi->set_type(dsttype);
 430           _gvn.set_type(phi, dsttype);
 431         }
 432       }
 433     }
 434   }
 435   phi_map->merge_replaced_nodes_with(ex_map);
 436 }
 437 
 438 //--------------------------use_exception_state--------------------------------
 439 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 440   if (failing()) { stop(); return top(); }
 441   Node* region = phi_map->control();
 442   Node* hidden_merge_mark = root();
 443   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 444   Node* ex_oop = clear_saved_ex_oop(phi_map);
 445   if (region->in(0) == hidden_merge_mark) {
 446     // Special marking for internal ex-states.  Process the phis now.
 447     region->set_req(0, region);  // now it's an ordinary region
 448     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 449     // Note: Setting the jvms also sets the bci and sp.
 450     set_control(_gvn.transform(region));
 451     uint tos = jvms()->stkoff() + sp();
 452     for (uint i = 1; i < tos; i++) {
 453       Node* x = phi_map->in(i);
 454       if (x->in(0) == region) {
 455         assert(x->is_Phi(), "expected a special phi");
 456         phi_map->set_req(i, _gvn.transform(x));
 457       }
 458     }
 459     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 460       Node* x = mms.memory();
 461       if (x->in(0) == region) {
 462         assert(x->is_Phi(), "nobody else uses a hidden region");
 463         mms.set_memory(_gvn.transform(x));
 464       }
 465     }
 466     if (ex_oop->in(0) == region) {
 467       assert(ex_oop->is_Phi(), "expected a special phi");
 468       ex_oop = _gvn.transform(ex_oop);
 469     }
 470   } else {
 471     set_jvms(phi_map->jvms());
 472   }
 473 
 474   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 475   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 476   return ex_oop;
 477 }
 478 
 479 //---------------------------------java_bc-------------------------------------
 480 Bytecodes::Code GraphKit::java_bc() const {
 481   ciMethod* method = this->method();
 482   int       bci    = this->bci();
 483   if (method != NULL && bci != InvocationEntryBci)
 484     return method->java_code_at_bci(bci);
 485   else
 486     return Bytecodes::_illegal;
 487 }
 488 
 489 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 490                                                           bool must_throw) {
 491     // if the exception capability is set, then we will generate code
 492     // to check the JavaThread.should_post_on_exceptions flag to see
 493     // if we actually need to report exception events (for this
 494     // thread).  If we don't need to report exception events, we will
 495     // take the normal fast path provided by add_exception_events.  If
 496     // exception event reporting is enabled for this thread, we will
 497     // take the uncommon_trap in the BuildCutout below.
 498 
 499     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 500     Node* jthread = _gvn.transform(new ThreadLocalNode());
 501     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 502     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
 503 
 504     // Test the should_post_on_exceptions_flag vs. 0
 505     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
 506     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
 507 
 508     // Branch to slow_path if should_post_on_exceptions_flag was true
 509     { BuildCutout unless(this, tst, PROB_MAX);
 510       // Do not try anything fancy if we're notifying the VM on every throw.
 511       // Cf. case Bytecodes::_athrow in parse2.cpp.
 512       uncommon_trap(reason, Deoptimization::Action_none,
 513                     (ciKlass*)NULL, (char*)NULL, must_throw);
 514     }
 515 
 516 }
 517 
 518 //------------------------------builtin_throw----------------------------------
 519 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 520   bool must_throw = true;
 521 
 522   if (env()->jvmti_can_post_on_exceptions()) {
 523     // check if we must post exception events, take uncommon trap if so
 524     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 525     // here if should_post_on_exceptions is false
 526     // continue on with the normal codegen
 527   }
 528 
 529   // If this particular condition has not yet happened at this
 530   // bytecode, then use the uncommon trap mechanism, and allow for
 531   // a future recompilation if several traps occur here.
 532   // If the throw is hot, try to use a more complicated inline mechanism
 533   // which keeps execution inside the compiled code.
 534   bool treat_throw_as_hot = false;
 535   ciMethodData* md = method()->method_data();
 536 
 537   if (ProfileTraps) {
 538     if (too_many_traps(reason)) {
 539       treat_throw_as_hot = true;
 540     }
 541     // (If there is no MDO at all, assume it is early in
 542     // execution, and that any deopts are part of the
 543     // startup transient, and don't need to be remembered.)
 544 
 545     // Also, if there is a local exception handler, treat all throws
 546     // as hot if there has been at least one in this method.
 547     if (C->trap_count(reason) != 0
 548         && method()->method_data()->trap_count(reason) != 0
 549         && has_ex_handler()) {
 550         treat_throw_as_hot = true;
 551     }
 552   }
 553 
 554   // If this throw happens frequently, an uncommon trap might cause
 555   // a performance pothole.  If there is a local exception handler,
 556   // and if this particular bytecode appears to be deoptimizing often,
 557   // let us handle the throw inline, with a preconstructed instance.
 558   // Note:   If the deopt count has blown up, the uncommon trap
 559   // runtime is going to flush this nmethod, not matter what.
 560   if (treat_throw_as_hot
 561       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 562     // If the throw is local, we use a pre-existing instance and
 563     // punt on the backtrace.  This would lead to a missing backtrace
 564     // (a repeat of 4292742) if the backtrace object is ever asked
 565     // for its backtrace.
 566     // Fixing this remaining case of 4292742 requires some flavor of
 567     // escape analysis.  Leave that for the future.
 568     ciInstance* ex_obj = NULL;
 569     switch (reason) {
 570     case Deoptimization::Reason_null_check:
 571       ex_obj = env()->NullPointerException_instance();
 572       break;
 573     case Deoptimization::Reason_div0_check:
 574       ex_obj = env()->ArithmeticException_instance();
 575       break;
 576     case Deoptimization::Reason_range_check:
 577       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 578       break;
 579     case Deoptimization::Reason_class_check:
 580       if (java_bc() == Bytecodes::_aastore) {
 581         ex_obj = env()->ArrayStoreException_instance();
 582       } else {
 583         ex_obj = env()->ClassCastException_instance();
 584       }
 585       break;
 586     }
 587     if (failing()) { stop(); return; }  // exception allocation might fail
 588     if (ex_obj != NULL) {
 589       // Cheat with a preallocated exception object.
 590       if (C->log() != NULL)
 591         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 592                        Deoptimization::trap_reason_name(reason));
 593       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 594       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
 595 
 596       // Clear the detail message of the preallocated exception object.
 597       // Weblogic sometimes mutates the detail message of exceptions
 598       // using reflection.
 599       int offset = java_lang_Throwable::get_detailMessage_offset();
 600       const TypePtr* adr_typ = ex_con->add_offset(offset);
 601 
 602       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 603       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 604       // Conservatively release stores of object references.
 605       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
 606 
 607       add_exception_state(make_exception_state(ex_node));
 608       return;
 609     }
 610   }
 611 
 612   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 613   // It won't be much cheaper than bailing to the interp., since we'll
 614   // have to pass up all the debug-info, and the runtime will have to
 615   // create the stack trace.
 616 
 617   // Usual case:  Bail to interpreter.
 618   // Reserve the right to recompile if we haven't seen anything yet.
 619 
 620   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
 621   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 622   if (treat_throw_as_hot
 623       && (method()->method_data()->trap_recompiled_at(bci(), m)
 624           || C->too_many_traps(reason))) {
 625     // We cannot afford to take more traps here.  Suffer in the interpreter.
 626     if (C->log() != NULL)
 627       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 628                      Deoptimization::trap_reason_name(reason),
 629                      C->trap_count(reason));
 630     action = Deoptimization::Action_none;
 631   }
 632 
 633   // "must_throw" prunes the JVM state to include only the stack, if there
 634   // are no local exception handlers.  This should cut down on register
 635   // allocation time and code size, by drastically reducing the number
 636   // of in-edges on the call to the uncommon trap.
 637 
 638   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 639 }
 640 
 641 
 642 //----------------------------PreserveJVMState---------------------------------
 643 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 644   debug_only(kit->verify_map());
 645   _kit    = kit;
 646   _map    = kit->map();   // preserve the map
 647   _sp     = kit->sp();
 648   kit->set_map(clone_map ? kit->clone_map() : NULL);
 649 #ifdef ASSERT
 650   _bci    = kit->bci();
 651   Parse* parser = kit->is_Parse();
 652   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 653   _block  = block;
 654 #endif
 655 }
 656 PreserveJVMState::~PreserveJVMState() {
 657   GraphKit* kit = _kit;
 658 #ifdef ASSERT
 659   assert(kit->bci() == _bci, "bci must not shift");
 660   Parse* parser = kit->is_Parse();
 661   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 662   assert(block == _block,    "block must not shift");
 663 #endif
 664   kit->set_map(_map);
 665   kit->set_sp(_sp);
 666 }
 667 
 668 
 669 //-----------------------------BuildCutout-------------------------------------
 670 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 671   : PreserveJVMState(kit)
 672 {
 673   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 674   SafePointNode* outer_map = _map;   // preserved map is caller's
 675   SafePointNode* inner_map = kit->map();
 676   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 677   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
 678   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
 679 }
 680 BuildCutout::~BuildCutout() {
 681   GraphKit* kit = _kit;
 682   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 683 }
 684 
 685 //---------------------------PreserveReexecuteState----------------------------
 686 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 687   assert(!kit->stopped(), "must call stopped() before");
 688   _kit    =    kit;
 689   _sp     =    kit->sp();
 690   _reexecute = kit->jvms()->_reexecute;
 691 }
 692 PreserveReexecuteState::~PreserveReexecuteState() {
 693   if (_kit->stopped()) return;
 694   _kit->jvms()->_reexecute = _reexecute;
 695   _kit->set_sp(_sp);
 696 }
 697 
 698 //------------------------------clone_map--------------------------------------
 699 // Implementation of PreserveJVMState
 700 //
 701 // Only clone_map(...) here. If this function is only used in the
 702 // PreserveJVMState class we may want to get rid of this extra
 703 // function eventually and do it all there.
 704 
 705 SafePointNode* GraphKit::clone_map() {
 706   if (map() == NULL)  return NULL;
 707 
 708   // Clone the memory edge first
 709   Node* mem = MergeMemNode::make(map()->memory());
 710   gvn().set_type_bottom(mem);
 711 
 712   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 713   JVMState* jvms = this->jvms();
 714   JVMState* clonejvms = jvms->clone_shallow(C);
 715   clonemap->set_memory(mem);
 716   clonemap->set_jvms(clonejvms);
 717   clonejvms->set_map(clonemap);
 718   record_for_igvn(clonemap);
 719   gvn().set_type_bottom(clonemap);
 720   return clonemap;
 721 }
 722 
 723 
 724 //-----------------------------set_map_clone-----------------------------------
 725 void GraphKit::set_map_clone(SafePointNode* m) {
 726   _map = m;
 727   _map = clone_map();
 728   _map->set_next_exception(NULL);
 729   debug_only(verify_map());
 730 }
 731 
 732 
 733 //----------------------------kill_dead_locals---------------------------------
 734 // Detect any locals which are known to be dead, and force them to top.
 735 void GraphKit::kill_dead_locals() {
 736   // Consult the liveness information for the locals.  If any
 737   // of them are unused, then they can be replaced by top().  This
 738   // should help register allocation time and cut down on the size
 739   // of the deoptimization information.
 740 
 741   // This call is made from many of the bytecode handling
 742   // subroutines called from the Big Switch in do_one_bytecode.
 743   // Every bytecode which might include a slow path is responsible
 744   // for killing its dead locals.  The more consistent we
 745   // are about killing deads, the fewer useless phis will be
 746   // constructed for them at various merge points.
 747 
 748   // bci can be -1 (InvocationEntryBci).  We return the entry
 749   // liveness for the method.
 750 
 751   if (method() == NULL || method()->code_size() == 0) {
 752     // We are building a graph for a call to a native method.
 753     // All locals are live.
 754     return;
 755   }
 756 
 757   ResourceMark rm;
 758 
 759   // Consult the liveness information for the locals.  If any
 760   // of them are unused, then they can be replaced by top().  This
 761   // should help register allocation time and cut down on the size
 762   // of the deoptimization information.
 763   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 764 
 765   int len = (int)live_locals.size();
 766   assert(len <= jvms()->loc_size(), "too many live locals");
 767   for (int local = 0; local < len; local++) {
 768     if (!live_locals.at(local)) {
 769       set_local(local, top());
 770     }
 771   }
 772 }
 773 
 774 #ifdef ASSERT
 775 //-------------------------dead_locals_are_killed------------------------------
 776 // Return true if all dead locals are set to top in the map.
 777 // Used to assert "clean" debug info at various points.
 778 bool GraphKit::dead_locals_are_killed() {
 779   if (method() == NULL || method()->code_size() == 0) {
 780     // No locals need to be dead, so all is as it should be.
 781     return true;
 782   }
 783 
 784   // Make sure somebody called kill_dead_locals upstream.
 785   ResourceMark rm;
 786   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 787     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 788     SafePointNode* map = jvms->map();
 789     ciMethod* method = jvms->method();
 790     int       bci    = jvms->bci();
 791     if (jvms == this->jvms()) {
 792       bci = this->bci();  // it might not yet be synched
 793     }
 794     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 795     int len = (int)live_locals.size();
 796     if (!live_locals.is_valid() || len == 0)
 797       // This method is trivial, or is poisoned by a breakpoint.
 798       return true;
 799     assert(len == jvms->loc_size(), "live map consistent with locals map");
 800     for (int local = 0; local < len; local++) {
 801       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 802         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 803           tty->print_cr("Zombie local %d: ", local);
 804           jvms->dump();
 805         }
 806         return false;
 807       }
 808     }
 809   }
 810   return true;
 811 }
 812 
 813 #endif //ASSERT
 814 
 815 // Helper function for enforcing certain bytecodes to reexecute if
 816 // deoptimization happens
 817 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 818   ciMethod* cur_method = jvms->method();
 819   int       cur_bci   = jvms->bci();
 820   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 821     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 822     return Interpreter::bytecode_should_reexecute(code) ||
 823            is_anewarray && code == Bytecodes::_multianewarray;
 824     // Reexecute _multianewarray bytecode which was replaced with
 825     // sequence of [a]newarray. See Parse::do_multianewarray().
 826     //
 827     // Note: interpreter should not have it set since this optimization
 828     // is limited by dimensions and guarded by flag so in some cases
 829     // multianewarray() runtime calls will be generated and
 830     // the bytecode should not be reexecutes (stack will not be reset).
 831   } else
 832     return false;
 833 }
 834 
 835 // Helper function for adding JVMState and debug information to node
 836 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 837   // Add the safepoint edges to the call (or other safepoint).
 838 
 839   // Make sure dead locals are set to top.  This
 840   // should help register allocation time and cut down on the size
 841   // of the deoptimization information.
 842   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 843 
 844   // Walk the inline list to fill in the correct set of JVMState's
 845   // Also fill in the associated edges for each JVMState.
 846 
 847   // If the bytecode needs to be reexecuted we need to put
 848   // the arguments back on the stack.
 849   const bool should_reexecute = jvms()->should_reexecute();
 850   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
 851 
 852   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
 853   // undefined if the bci is different.  This is normal for Parse but it
 854   // should not happen for LibraryCallKit because only one bci is processed.
 855   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
 856          "in LibraryCallKit the reexecute bit should not change");
 857 
 858   // If we are guaranteed to throw, we can prune everything but the
 859   // input to the current bytecode.
 860   bool can_prune_locals = false;
 861   uint stack_slots_not_pruned = 0;
 862   int inputs = 0, depth = 0;
 863   if (must_throw) {
 864     assert(method() == youngest_jvms->method(), "sanity");
 865     if (compute_stack_effects(inputs, depth)) {
 866       can_prune_locals = true;
 867       stack_slots_not_pruned = inputs;
 868     }
 869   }
 870 
 871   if (env()->should_retain_local_variables()) {
 872     // At any safepoint, this method can get breakpointed, which would
 873     // then require an immediate deoptimization.
 874     can_prune_locals = false;  // do not prune locals
 875     stack_slots_not_pruned = 0;
 876   }
 877 
 878   // do not scribble on the input jvms
 879   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 880   call->set_jvms(out_jvms); // Start jvms list for call node
 881 
 882   // For a known set of bytecodes, the interpreter should reexecute them if
 883   // deoptimization happens. We set the reexecute state for them here
 884   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 885       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 886     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 887   }
 888 
 889   // Presize the call:
 890   DEBUG_ONLY(uint non_debug_edges = call->req());
 891   call->add_req_batch(top(), youngest_jvms->debug_depth());
 892   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 893 
 894   // Set up edges so that the call looks like this:
 895   //  Call [state:] ctl io mem fptr retadr
 896   //       [parms:] parm0 ... parmN
 897   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 898   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 899   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 900   // Note that caller debug info precedes callee debug info.
 901 
 902   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 903   uint debug_ptr = call->req();
 904 
 905   // Loop over the map input edges associated with jvms, add them
 906   // to the call node, & reset all offsets to match call node array.
 907   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 908     uint debug_end   = debug_ptr;
 909     uint debug_start = debug_ptr - in_jvms->debug_size();
 910     debug_ptr = debug_start;  // back up the ptr
 911 
 912     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 913     uint j, k, l;
 914     SafePointNode* in_map = in_jvms->map();
 915     out_jvms->set_map(call);
 916 
 917     if (can_prune_locals) {
 918       assert(in_jvms->method() == out_jvms->method(), "sanity");
 919       // If the current throw can reach an exception handler in this JVMS,
 920       // then we must keep everything live that can reach that handler.
 921       // As a quick and dirty approximation, we look for any handlers at all.
 922       if (in_jvms->method()->has_exception_handlers()) {
 923         can_prune_locals = false;
 924       }
 925     }
 926 
 927     // Add the Locals
 928     k = in_jvms->locoff();
 929     l = in_jvms->loc_size();
 930     out_jvms->set_locoff(p);
 931     if (!can_prune_locals) {
 932       for (j = 0; j < l; j++)
 933         call->set_req(p++, in_map->in(k+j));
 934     } else {
 935       p += l;  // already set to top above by add_req_batch
 936     }
 937 
 938     // Add the Expression Stack
 939     k = in_jvms->stkoff();
 940     l = in_jvms->sp();
 941     out_jvms->set_stkoff(p);
 942     if (!can_prune_locals) {
 943       for (j = 0; j < l; j++)
 944         call->set_req(p++, in_map->in(k+j));
 945     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 946       // Divide stack into {S0,...,S1}, where S0 is set to top.
 947       uint s1 = stack_slots_not_pruned;
 948       stack_slots_not_pruned = 0;  // for next iteration
 949       if (s1 > l)  s1 = l;
 950       uint s0 = l - s1;
 951       p += s0;  // skip the tops preinstalled by add_req_batch
 952       for (j = s0; j < l; j++)
 953         call->set_req(p++, in_map->in(k+j));
 954     } else {
 955       p += l;  // already set to top above by add_req_batch
 956     }
 957 
 958     // Add the Monitors
 959     k = in_jvms->monoff();
 960     l = in_jvms->mon_size();
 961     out_jvms->set_monoff(p);
 962     for (j = 0; j < l; j++)
 963       call->set_req(p++, in_map->in(k+j));
 964 
 965     // Copy any scalar object fields.
 966     k = in_jvms->scloff();
 967     l = in_jvms->scl_size();
 968     out_jvms->set_scloff(p);
 969     for (j = 0; j < l; j++)
 970       call->set_req(p++, in_map->in(k+j));
 971 
 972     // Finish the new jvms.
 973     out_jvms->set_endoff(p);
 974 
 975     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 976     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 977     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 978     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 979     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 980     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 981 
 982     // Update the two tail pointers in parallel.
 983     out_jvms = out_jvms->caller();
 984     in_jvms  = in_jvms->caller();
 985   }
 986 
 987   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
 988 
 989   // Test the correctness of JVMState::debug_xxx accessors:
 990   assert(call->jvms()->debug_start() == non_debug_edges, "");
 991   assert(call->jvms()->debug_end()   == call->req(), "");
 992   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
 993 }
 994 
 995 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
 996   Bytecodes::Code code = java_bc();
 997   if (code == Bytecodes::_wide) {
 998     code = method()->java_code_at_bci(bci() + 1);
 999   }
1000 
1001   BasicType rtype = T_ILLEGAL;
1002   int       rsize = 0;
1003 
1004   if (code != Bytecodes::_illegal) {
1005     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1006     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1007     if (rtype < T_CONFLICT)
1008       rsize = type2size[rtype];
1009   }
1010 
1011   switch (code) {
1012   case Bytecodes::_illegal:
1013     return false;
1014 
1015   case Bytecodes::_ldc:
1016   case Bytecodes::_ldc_w:
1017   case Bytecodes::_ldc2_w:
1018     inputs = 0;
1019     break;
1020 
1021   case Bytecodes::_dup:         inputs = 1;  break;
1022   case Bytecodes::_dup_x1:      inputs = 2;  break;
1023   case Bytecodes::_dup_x2:      inputs = 3;  break;
1024   case Bytecodes::_dup2:        inputs = 2;  break;
1025   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1026   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1027   case Bytecodes::_swap:        inputs = 2;  break;
1028   case Bytecodes::_arraylength: inputs = 1;  break;
1029 
1030   case Bytecodes::_getstatic:
1031   case Bytecodes::_putstatic:
1032   case Bytecodes::_getfield:
1033   case Bytecodes::_putfield:
1034     {
1035       bool ignored_will_link;
1036       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1037       int      size  = field->type()->size();
1038       bool is_get = (depth >= 0), is_static = (depth & 1);
1039       inputs = (is_static ? 0 : 1);
1040       if (is_get) {
1041         depth = size - inputs;
1042       } else {
1043         inputs += size;        // putxxx pops the value from the stack
1044         depth = - inputs;
1045       }
1046     }
1047     break;
1048 
1049   case Bytecodes::_invokevirtual:
1050   case Bytecodes::_invokespecial:
1051   case Bytecodes::_invokestatic:
1052   case Bytecodes::_invokedynamic:
1053   case Bytecodes::_invokeinterface:
1054     {
1055       bool ignored_will_link;
1056       ciSignature* declared_signature = NULL;
1057       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1058       assert(declared_signature != NULL, "cannot be null");
1059       inputs   = declared_signature->arg_size_for_bc(code);
1060       int size = declared_signature->return_type()->size();
1061       depth = size - inputs;
1062     }
1063     break;
1064 
1065   case Bytecodes::_multianewarray:
1066     {
1067       ciBytecodeStream iter(method());
1068       iter.reset_to_bci(bci());
1069       iter.next();
1070       inputs = iter.get_dimensions();
1071       assert(rsize == 1, "");
1072       depth = rsize - inputs;
1073     }
1074     break;
1075 
1076   case Bytecodes::_ireturn:
1077   case Bytecodes::_lreturn:
1078   case Bytecodes::_freturn:
1079   case Bytecodes::_dreturn:
1080   case Bytecodes::_areturn:
1081     assert(rsize = -depth, "");
1082     inputs = rsize;
1083     break;
1084 
1085   case Bytecodes::_jsr:
1086   case Bytecodes::_jsr_w:
1087     inputs = 0;
1088     depth  = 1;                  // S.B. depth=1, not zero
1089     break;
1090 
1091   default:
1092     // bytecode produces a typed result
1093     inputs = rsize - depth;
1094     assert(inputs >= 0, "");
1095     break;
1096   }
1097 
1098 #ifdef ASSERT
1099   // spot check
1100   int outputs = depth + inputs;
1101   assert(outputs >= 0, "sanity");
1102   switch (code) {
1103   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1104   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1105   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1106   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1107   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1108   }
1109 #endif //ASSERT
1110 
1111   return true;
1112 }
1113 
1114 
1115 
1116 //------------------------------basic_plus_adr---------------------------------
1117 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1118   // short-circuit a common case
1119   if (offset == intcon(0))  return ptr;
1120   return _gvn.transform( new AddPNode(base, ptr, offset) );
1121 }
1122 
1123 Node* GraphKit::ConvI2L(Node* offset) {
1124   // short-circuit a common case
1125   jint offset_con = find_int_con(offset, Type::OffsetBot);
1126   if (offset_con != Type::OffsetBot) {
1127     return longcon((jlong) offset_con);
1128   }
1129   return _gvn.transform( new ConvI2LNode(offset));
1130 }
1131 
1132 Node* GraphKit::ConvI2UL(Node* offset) {
1133   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1134   if (offset_con != (juint) Type::OffsetBot) {
1135     return longcon((julong) offset_con);
1136   }
1137   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1138   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1139   return _gvn.transform( new AndLNode(conv, mask) );
1140 }
1141 
1142 Node* GraphKit::ConvL2I(Node* offset) {
1143   // short-circuit a common case
1144   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1145   if (offset_con != (jlong)Type::OffsetBot) {
1146     return intcon((int) offset_con);
1147   }
1148   return _gvn.transform( new ConvL2INode(offset));
1149 }
1150 
1151 //-------------------------load_object_klass-----------------------------------
1152 Node* GraphKit::load_object_klass(Node* obj) {
1153   // Special-case a fresh allocation to avoid building nodes:
1154   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1155   if (akls != NULL)  return akls;
1156   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1157   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1158 }
1159 
1160 //-------------------------load_array_length-----------------------------------
1161 Node* GraphKit::load_array_length(Node* array) {
1162   // Special-case a fresh allocation to avoid building nodes:
1163   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1164   Node *alen;
1165   if (alloc == NULL) {
1166     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1167     alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1168   } else {
1169     alen = alloc->Ideal_length();
1170     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1171     if (ccast != alen) {
1172       alen = _gvn.transform(ccast);
1173     }
1174   }
1175   return alen;
1176 }
1177 
1178 //------------------------------do_null_check----------------------------------
1179 // Helper function to do a NULL pointer check.  Returned value is
1180 // the incoming address with NULL casted away.  You are allowed to use the
1181 // not-null value only if you are control dependent on the test.
1182 extern int explicit_null_checks_inserted,
1183            explicit_null_checks_elided;
1184 Node* GraphKit::null_check_common(Node* value, BasicType type,
1185                                   // optional arguments for variations:
1186                                   bool assert_null,
1187                                   Node* *null_control,
1188                                   bool speculative) {
1189   assert(!assert_null || null_control == NULL, "not both at once");
1190   if (stopped())  return top();
1191   if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1192     // For some performance testing, we may wish to suppress null checking.
1193     value = cast_not_null(value);   // Make it appear to be non-null (4962416).
1194     return value;
1195   }
1196   explicit_null_checks_inserted++;
1197 
1198   // Construct NULL check
1199   Node *chk = NULL;
1200   switch(type) {
1201     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1202     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1203     case T_ARRAY  : // fall through
1204       type = T_OBJECT;  // simplify further tests
1205     case T_OBJECT : {
1206       const Type *t = _gvn.type( value );
1207 
1208       const TypeOopPtr* tp = t->isa_oopptr();
1209       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1210           // Only for do_null_check, not any of its siblings:
1211           && !assert_null && null_control == NULL) {
1212         // Usually, any field access or invocation on an unloaded oop type
1213         // will simply fail to link, since the statically linked class is
1214         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1215         // the static class is loaded but the sharper oop type is not.
1216         // Rather than checking for this obscure case in lots of places,
1217         // we simply observe that a null check on an unloaded class
1218         // will always be followed by a nonsense operation, so we
1219         // can just issue the uncommon trap here.
1220         // Our access to the unloaded class will only be correct
1221         // after it has been loaded and initialized, which requires
1222         // a trip through the interpreter.
1223 #ifndef PRODUCT
1224         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1225 #endif
1226         uncommon_trap(Deoptimization::Reason_unloaded,
1227                       Deoptimization::Action_reinterpret,
1228                       tp->klass(), "!loaded");
1229         return top();
1230       }
1231 
1232       if (assert_null) {
1233         // See if the type is contained in NULL_PTR.
1234         // If so, then the value is already null.
1235         if (t->higher_equal(TypePtr::NULL_PTR)) {
1236           explicit_null_checks_elided++;
1237           return value;           // Elided null assert quickly!
1238         }
1239       } else {
1240         // See if mixing in the NULL pointer changes type.
1241         // If so, then the NULL pointer was not allowed in the original
1242         // type.  In other words, "value" was not-null.
1243         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1244           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1245           explicit_null_checks_elided++;
1246           return value;           // Elided null check quickly!
1247         }
1248       }
1249       chk = new CmpPNode( value, null() );
1250       break;
1251     }
1252 
1253     default:
1254       fatal("unexpected type: %s", type2name(type));
1255   }
1256   assert(chk != NULL, "sanity check");
1257   chk = _gvn.transform(chk);
1258 
1259   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1260   BoolNode *btst = new BoolNode( chk, btest);
1261   Node   *tst = _gvn.transform( btst );
1262 
1263   //-----------
1264   // if peephole optimizations occurred, a prior test existed.
1265   // If a prior test existed, maybe it dominates as we can avoid this test.
1266   if (tst != btst && type == T_OBJECT) {
1267     // At this point we want to scan up the CFG to see if we can
1268     // find an identical test (and so avoid this test altogether).
1269     Node *cfg = control();
1270     int depth = 0;
1271     while( depth < 16 ) {       // Limit search depth for speed
1272       if( cfg->Opcode() == Op_IfTrue &&
1273           cfg->in(0)->in(1) == tst ) {
1274         // Found prior test.  Use "cast_not_null" to construct an identical
1275         // CastPP (and hence hash to) as already exists for the prior test.
1276         // Return that casted value.
1277         if (assert_null) {
1278           replace_in_map(value, null());
1279           return null();  // do not issue the redundant test
1280         }
1281         Node *oldcontrol = control();
1282         set_control(cfg);
1283         Node *res = cast_not_null(value);
1284         set_control(oldcontrol);
1285         explicit_null_checks_elided++;
1286         return res;
1287       }
1288       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1289       if (cfg == NULL)  break;  // Quit at region nodes
1290       depth++;
1291     }
1292   }
1293 
1294   //-----------
1295   // Branch to failure if null
1296   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1297   Deoptimization::DeoptReason reason;
1298   if (assert_null) {
1299     reason = Deoptimization::Reason_null_assert;
1300   } else if (type == T_OBJECT) {
1301     reason = Deoptimization::reason_null_check(speculative);
1302   } else {
1303     reason = Deoptimization::Reason_div0_check;
1304   }
1305   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1306   // ciMethodData::has_trap_at will return a conservative -1 if any
1307   // must-be-null assertion has failed.  This could cause performance
1308   // problems for a method after its first do_null_assert failure.
1309   // Consider using 'Reason_class_check' instead?
1310 
1311   // To cause an implicit null check, we set the not-null probability
1312   // to the maximum (PROB_MAX).  For an explicit check the probability
1313   // is set to a smaller value.
1314   if (null_control != NULL || too_many_traps(reason)) {
1315     // probability is less likely
1316     ok_prob =  PROB_LIKELY_MAG(3);
1317   } else if (!assert_null &&
1318              (ImplicitNullCheckThreshold > 0) &&
1319              method() != NULL &&
1320              (method()->method_data()->trap_count(reason)
1321               >= (uint)ImplicitNullCheckThreshold)) {
1322     ok_prob =  PROB_LIKELY_MAG(3);
1323   }
1324 
1325   if (null_control != NULL) {
1326     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1327     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1328     set_control(      _gvn.transform( new IfTrueNode(iff)));
1329     if (null_true == top())
1330       explicit_null_checks_elided++;
1331     (*null_control) = null_true;
1332   } else {
1333     BuildCutout unless(this, tst, ok_prob);
1334     // Check for optimizer eliding test at parse time
1335     if (stopped()) {
1336       // Failure not possible; do not bother making uncommon trap.
1337       explicit_null_checks_elided++;
1338     } else if (assert_null) {
1339       uncommon_trap(reason,
1340                     Deoptimization::Action_make_not_entrant,
1341                     NULL, "assert_null");
1342     } else {
1343       replace_in_map(value, zerocon(type));
1344       builtin_throw(reason);
1345     }
1346   }
1347 
1348   // Must throw exception, fall-thru not possible?
1349   if (stopped()) {
1350     return top();               // No result
1351   }
1352 
1353   if (assert_null) {
1354     // Cast obj to null on this path.
1355     replace_in_map(value, zerocon(type));
1356     return zerocon(type);
1357   }
1358 
1359   // Cast obj to not-null on this path, if there is no null_control.
1360   // (If there is a null_control, a non-null value may come back to haunt us.)
1361   if (type == T_OBJECT) {
1362     Node* cast = cast_not_null(value, false);
1363     if (null_control == NULL || (*null_control) == top())
1364       replace_in_map(value, cast);
1365     value = cast;
1366   }
1367 
1368   return value;
1369 }
1370 
1371 
1372 //------------------------------cast_not_null----------------------------------
1373 // Cast obj to not-null on this path
1374 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1375   const Type *t = _gvn.type(obj);
1376   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1377   // Object is already not-null?
1378   if( t == t_not_null ) return obj;
1379 
1380   Node *cast = new CastPPNode(obj,t_not_null);
1381   cast->init_req(0, control());
1382   cast = _gvn.transform( cast );
1383 
1384   // Scan for instances of 'obj' in the current JVM mapping.
1385   // These instances are known to be not-null after the test.
1386   if (do_replace_in_map)
1387     replace_in_map(obj, cast);
1388 
1389   return cast;                  // Return casted value
1390 }
1391 
1392 
1393 //--------------------------replace_in_map-------------------------------------
1394 void GraphKit::replace_in_map(Node* old, Node* neww) {
1395   if (old == neww) {
1396     return;
1397   }
1398 
1399   map()->replace_edge(old, neww);
1400 
1401   // Note: This operation potentially replaces any edge
1402   // on the map.  This includes locals, stack, and monitors
1403   // of the current (innermost) JVM state.
1404 
1405   // don't let inconsistent types from profiling escape this
1406   // method
1407 
1408   const Type* told = _gvn.type(old);
1409   const Type* tnew = _gvn.type(neww);
1410 
1411   if (!tnew->higher_equal(told)) {
1412     return;
1413   }
1414 
1415   map()->record_replaced_node(old, neww);
1416 }
1417 
1418 
1419 //=============================================================================
1420 //--------------------------------memory---------------------------------------
1421 Node* GraphKit::memory(uint alias_idx) {
1422   MergeMemNode* mem = merged_memory();
1423   Node* p = mem->memory_at(alias_idx);
1424   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1425   return p;
1426 }
1427 
1428 //-----------------------------reset_memory------------------------------------
1429 Node* GraphKit::reset_memory() {
1430   Node* mem = map()->memory();
1431   // do not use this node for any more parsing!
1432   debug_only( map()->set_memory((Node*)NULL) );
1433   return _gvn.transform( mem );
1434 }
1435 
1436 //------------------------------set_all_memory---------------------------------
1437 void GraphKit::set_all_memory(Node* newmem) {
1438   Node* mergemem = MergeMemNode::make(newmem);
1439   gvn().set_type_bottom(mergemem);
1440   map()->set_memory(mergemem);
1441 }
1442 
1443 //------------------------------set_all_memory_call----------------------------
1444 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1445   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1446   set_all_memory(newmem);
1447 }
1448 
1449 //=============================================================================
1450 //
1451 // parser factory methods for MemNodes
1452 //
1453 // These are layered on top of the factory methods in LoadNode and StoreNode,
1454 // and integrate with the parser's memory state and _gvn engine.
1455 //
1456 
1457 // factory methods in "int adr_idx"
1458 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1459                           int adr_idx,
1460                           MemNode::MemOrd mo,
1461                           LoadNode::ControlDependency control_dependency,
1462                           bool require_atomic_access,
1463                           bool unaligned,
1464                           bool mismatched) {
1465   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1466   const TypePtr* adr_type = NULL; // debug-mode-only argument
1467   debug_only(adr_type = C->get_adr_type(adr_idx));
1468   Node* mem = memory(adr_idx);
1469   Node* ld;
1470   if (require_atomic_access && bt == T_LONG) {
1471     ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1472   } else if (require_atomic_access && bt == T_DOUBLE) {
1473     ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1474   } else {
1475     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1476   }
1477   ld = _gvn.transform(ld);
1478   if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1479     // Improve graph before escape analysis and boxing elimination.
1480     record_for_igvn(ld);
1481   }
1482   return ld;
1483 }
1484 
1485 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1486                                 int adr_idx,
1487                                 MemNode::MemOrd mo,
1488                                 bool require_atomic_access,
1489                                 bool unaligned,
1490                                 bool mismatched) {
1491   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1492   const TypePtr* adr_type = NULL;
1493   debug_only(adr_type = C->get_adr_type(adr_idx));
1494   Node *mem = memory(adr_idx);
1495   Node* st;
1496   if (require_atomic_access && bt == T_LONG) {
1497     st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1498   } else if (require_atomic_access && bt == T_DOUBLE) {
1499     st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1500   } else {
1501     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1502   }
1503   if (unaligned) {
1504     st->as_Store()->set_unaligned_access();
1505   }
1506   if (mismatched) {
1507     st->as_Store()->set_mismatched_access();
1508   }
1509   st = _gvn.transform(st);
1510   set_memory(st, adr_idx);
1511   // Back-to-back stores can only remove intermediate store with DU info
1512   // so push on worklist for optimizer.
1513   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1514     record_for_igvn(st);
1515 
1516   return st;
1517 }
1518 
1519 
1520 void GraphKit::pre_barrier(bool do_load,
1521                            Node* ctl,
1522                            Node* obj,
1523                            Node* adr,
1524                            uint  adr_idx,
1525                            Node* val,
1526                            const TypeOopPtr* val_type,
1527                            Node* pre_val,
1528                            BasicType bt) {
1529 
1530   BarrierSet* bs = Universe::heap()->barrier_set();
1531   set_control(ctl);
1532   switch (bs->kind()) {
1533     case BarrierSet::G1SATBCTLogging:
1534       g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1535       break;
1536 
1537     case BarrierSet::CardTableForRS:
1538     case BarrierSet::CardTableExtension:
1539     case BarrierSet::ModRef:
1540       break;
1541 
1542     default      :
1543       ShouldNotReachHere();
1544 
1545   }
1546 }
1547 
1548 bool GraphKit::can_move_pre_barrier() const {
1549   BarrierSet* bs = Universe::heap()->barrier_set();
1550   switch (bs->kind()) {
1551     case BarrierSet::G1SATBCTLogging:
1552       return true; // Can move it if no safepoint
1553 
1554     case BarrierSet::CardTableForRS:
1555     case BarrierSet::CardTableExtension:
1556     case BarrierSet::ModRef:
1557       return true; // There is no pre-barrier
1558 
1559     default      :
1560       ShouldNotReachHere();
1561   }
1562   return false;
1563 }
1564 
1565 void GraphKit::post_barrier(Node* ctl,
1566                             Node* store,
1567                             Node* obj,
1568                             Node* adr,
1569                             uint  adr_idx,
1570                             Node* val,
1571                             BasicType bt,
1572                             bool use_precise) {
1573   BarrierSet* bs = Universe::heap()->barrier_set();
1574   set_control(ctl);
1575   switch (bs->kind()) {
1576     case BarrierSet::G1SATBCTLogging:
1577       g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1578       break;
1579 
1580     case BarrierSet::CardTableForRS:
1581     case BarrierSet::CardTableExtension:
1582       write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1583       break;
1584 
1585     case BarrierSet::ModRef:
1586       break;
1587 
1588     default      :
1589       ShouldNotReachHere();
1590 
1591   }
1592 }
1593 
1594 Node* GraphKit::store_oop(Node* ctl,
1595                           Node* obj,
1596                           Node* adr,
1597                           const TypePtr* adr_type,
1598                           Node* val,
1599                           const TypeOopPtr* val_type,
1600                           BasicType bt,
1601                           bool use_precise,
1602                           MemNode::MemOrd mo,
1603                           bool mismatched) {
1604   // Transformation of a value which could be NULL pointer (CastPP #NULL)
1605   // could be delayed during Parse (for example, in adjust_map_after_if()).
1606   // Execute transformation here to avoid barrier generation in such case.
1607   if (_gvn.type(val) == TypePtr::NULL_PTR)
1608     val = _gvn.makecon(TypePtr::NULL_PTR);
1609 
1610   set_control(ctl);
1611   if (stopped()) return top(); // Dead path ?
1612 
1613   assert(bt == T_OBJECT, "sanity");
1614   assert(val != NULL, "not dead path");
1615   uint adr_idx = C->get_alias_index(adr_type);
1616   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1617 
1618   pre_barrier(true /* do_load */,
1619               control(), obj, adr, adr_idx, val, val_type,
1620               NULL /* pre_val */,
1621               bt);
1622 
1623   Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1624   post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1625   return store;
1626 }
1627 
1628 // Could be an array or object we don't know at compile time (unsafe ref.)
1629 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1630                              Node* obj,   // containing obj
1631                              Node* adr,  // actual adress to store val at
1632                              const TypePtr* adr_type,
1633                              Node* val,
1634                              BasicType bt,
1635                              MemNode::MemOrd mo,
1636                              bool mismatched) {
1637   Compile::AliasType* at = C->alias_type(adr_type);
1638   const TypeOopPtr* val_type = NULL;
1639   if (adr_type->isa_instptr()) {
1640     if (at->field() != NULL) {
1641       // known field.  This code is a copy of the do_put_xxx logic.
1642       ciField* field = at->field();
1643       if (!field->type()->is_loaded()) {
1644         val_type = TypeInstPtr::BOTTOM;
1645       } else {
1646         val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1647       }
1648     }
1649   } else if (adr_type->isa_aryptr()) {
1650     val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1651   }
1652   if (val_type == NULL) {
1653     val_type = TypeInstPtr::BOTTOM;
1654   }
1655   return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1656 }
1657 
1658 
1659 //-------------------------array_element_address-------------------------
1660 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1661                                       const TypeInt* sizetype) {
1662   uint shift  = exact_log2(type2aelembytes(elembt));
1663   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1664 
1665   // short-circuit a common case (saves lots of confusing waste motion)
1666   jint idx_con = find_int_con(idx, -1);
1667   if (idx_con >= 0) {
1668     intptr_t offset = header + ((intptr_t)idx_con << shift);
1669     return basic_plus_adr(ary, offset);
1670   }
1671 
1672   // must be correct type for alignment purposes
1673   Node* base  = basic_plus_adr(ary, header);
1674   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype);
1675   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1676   return basic_plus_adr(ary, base, scale);
1677 }
1678 
1679 //-------------------------load_array_element-------------------------
1680 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1681   const Type* elemtype = arytype->elem();
1682   BasicType elembt = elemtype->array_element_basic_type();
1683   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1684   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1685   return ld;
1686 }
1687 
1688 //-------------------------set_arguments_for_java_call-------------------------
1689 // Arguments (pre-popped from the stack) are taken from the JVMS.
1690 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1691   // Add the call arguments:
1692   uint nargs = call->method()->arg_size();
1693   for (uint i = 0; i < nargs; i++) {
1694     Node* arg = argument(i);
1695     call->init_req(i + TypeFunc::Parms, arg);
1696   }
1697 }
1698 
1699 //---------------------------set_edges_for_java_call---------------------------
1700 // Connect a newly created call into the current JVMS.
1701 // A return value node (if any) is returned from set_edges_for_java_call.
1702 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1703 
1704   // Add the predefined inputs:
1705   call->init_req( TypeFunc::Control, control() );
1706   call->init_req( TypeFunc::I_O    , i_o() );
1707   call->init_req( TypeFunc::Memory , reset_memory() );
1708   call->init_req( TypeFunc::FramePtr, frameptr() );
1709   call->init_req( TypeFunc::ReturnAdr, top() );
1710 
1711   add_safepoint_edges(call, must_throw);
1712 
1713   Node* xcall = _gvn.transform(call);
1714 
1715   if (xcall == top()) {
1716     set_control(top());
1717     return;
1718   }
1719   assert(xcall == call, "call identity is stable");
1720 
1721   // Re-use the current map to produce the result.
1722 
1723   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1724   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1725   set_all_memory_call(xcall, separate_io_proj);
1726 
1727   //return xcall;   // no need, caller already has it
1728 }
1729 
1730 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1731   if (stopped())  return top();  // maybe the call folded up?
1732 
1733   // Capture the return value, if any.
1734   Node* ret;
1735   if (call->method() == NULL ||
1736       call->method()->return_type()->basic_type() == T_VOID)
1737         ret = top();
1738   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1739 
1740   // Note:  Since any out-of-line call can produce an exception,
1741   // we always insert an I_O projection from the call into the result.
1742 
1743   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1744 
1745   if (separate_io_proj) {
1746     // The caller requested separate projections be used by the fall
1747     // through and exceptional paths, so replace the projections for
1748     // the fall through path.
1749     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1750     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1751   }
1752   return ret;
1753 }
1754 
1755 //--------------------set_predefined_input_for_runtime_call--------------------
1756 // Reading and setting the memory state is way conservative here.
1757 // The real problem is that I am not doing real Type analysis on memory,
1758 // so I cannot distinguish card mark stores from other stores.  Across a GC
1759 // point the Store Barrier and the card mark memory has to agree.  I cannot
1760 // have a card mark store and its barrier split across the GC point from
1761 // either above or below.  Here I get that to happen by reading ALL of memory.
1762 // A better answer would be to separate out card marks from other memory.
1763 // For now, return the input memory state, so that it can be reused
1764 // after the call, if this call has restricted memory effects.
1765 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1766   // Set fixed predefined input arguments
1767   Node* memory = reset_memory();
1768   call->init_req( TypeFunc::Control,   control()  );
1769   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1770   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1771   call->init_req( TypeFunc::FramePtr,  frameptr() );
1772   call->init_req( TypeFunc::ReturnAdr, top()      );
1773   return memory;
1774 }
1775 
1776 //-------------------set_predefined_output_for_runtime_call--------------------
1777 // Set control and memory (not i_o) from the call.
1778 // If keep_mem is not NULL, use it for the output state,
1779 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1780 // If hook_mem is NULL, this call produces no memory effects at all.
1781 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1782 // then only that memory slice is taken from the call.
1783 // In the last case, we must put an appropriate memory barrier before
1784 // the call, so as to create the correct anti-dependencies on loads
1785 // preceding the call.
1786 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1787                                                       Node* keep_mem,
1788                                                       const TypePtr* hook_mem) {
1789   // no i/o
1790   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1791   if (keep_mem) {
1792     // First clone the existing memory state
1793     set_all_memory(keep_mem);
1794     if (hook_mem != NULL) {
1795       // Make memory for the call
1796       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1797       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1798       // We also use hook_mem to extract specific effects from arraycopy stubs.
1799       set_memory(mem, hook_mem);
1800     }
1801     // ...else the call has NO memory effects.
1802 
1803     // Make sure the call advertises its memory effects precisely.
1804     // This lets us build accurate anti-dependences in gcm.cpp.
1805     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1806            "call node must be constructed correctly");
1807   } else {
1808     assert(hook_mem == NULL, "");
1809     // This is not a "slow path" call; all memory comes from the call.
1810     set_all_memory_call(call);
1811   }
1812 }
1813 
1814 
1815 // Replace the call with the current state of the kit.
1816 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1817   JVMState* ejvms = NULL;
1818   if (has_exceptions()) {
1819     ejvms = transfer_exceptions_into_jvms();
1820   }
1821 
1822   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1823   ReplacedNodes replaced_nodes_exception;
1824   Node* ex_ctl = top();
1825 
1826   SafePointNode* final_state = stop();
1827 
1828   // Find all the needed outputs of this call
1829   CallProjections callprojs;
1830   call->extract_projections(&callprojs, true);
1831 
1832   Node* init_mem = call->in(TypeFunc::Memory);
1833   Node* final_mem = final_state->in(TypeFunc::Memory);
1834   Node* final_ctl = final_state->in(TypeFunc::Control);
1835   Node* final_io = final_state->in(TypeFunc::I_O);
1836 
1837   // Replace all the old call edges with the edges from the inlining result
1838   if (callprojs.fallthrough_catchproj != NULL) {
1839     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1840   }
1841   if (callprojs.fallthrough_memproj != NULL) {
1842     if (final_mem->is_MergeMem()) {
1843       // Parser's exits MergeMem was not transformed but may be optimized
1844       final_mem = _gvn.transform(final_mem);
1845     }
1846     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
1847   }
1848   if (callprojs.fallthrough_ioproj != NULL) {
1849     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
1850   }
1851 
1852   // Replace the result with the new result if it exists and is used
1853   if (callprojs.resproj != NULL && result != NULL) {
1854     C->gvn_replace_by(callprojs.resproj, result);
1855   }
1856 
1857   if (ejvms == NULL) {
1858     // No exception edges to simply kill off those paths
1859     if (callprojs.catchall_catchproj != NULL) {
1860       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1861     }
1862     if (callprojs.catchall_memproj != NULL) {
1863       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1864     }
1865     if (callprojs.catchall_ioproj != NULL) {
1866       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1867     }
1868     // Replace the old exception object with top
1869     if (callprojs.exobj != NULL) {
1870       C->gvn_replace_by(callprojs.exobj, C->top());
1871     }
1872   } else {
1873     GraphKit ekit(ejvms);
1874 
1875     // Load my combined exception state into the kit, with all phis transformed:
1876     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1877     replaced_nodes_exception = ex_map->replaced_nodes();
1878 
1879     Node* ex_oop = ekit.use_exception_state(ex_map);
1880 
1881     if (callprojs.catchall_catchproj != NULL) {
1882       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1883       ex_ctl = ekit.control();
1884     }
1885     if (callprojs.catchall_memproj != NULL) {
1886       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1887     }
1888     if (callprojs.catchall_ioproj != NULL) {
1889       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1890     }
1891 
1892     // Replace the old exception object with the newly created one
1893     if (callprojs.exobj != NULL) {
1894       C->gvn_replace_by(callprojs.exobj, ex_oop);
1895     }
1896   }
1897 
1898   // Disconnect the call from the graph
1899   call->disconnect_inputs(NULL, C);
1900   C->gvn_replace_by(call, C->top());
1901 
1902   // Clean up any MergeMems that feed other MergeMems since the
1903   // optimizer doesn't like that.
1904   if (final_mem->is_MergeMem()) {
1905     Node_List wl;
1906     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1907       Node* m = i.get();
1908       if (m->is_MergeMem() && !wl.contains(m)) {
1909         wl.push(m);
1910       }
1911     }
1912     while (wl.size()  > 0) {
1913       _gvn.transform(wl.pop());
1914     }
1915   }
1916 
1917   if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1918     replaced_nodes.apply(C, final_ctl);
1919   }
1920   if (!ex_ctl->is_top() && do_replaced_nodes) {
1921     replaced_nodes_exception.apply(C, ex_ctl);
1922   }
1923 }
1924 
1925 
1926 //------------------------------increment_counter------------------------------
1927 // for statistics: increment a VM counter by 1
1928 
1929 void GraphKit::increment_counter(address counter_addr) {
1930   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1931   increment_counter(adr1);
1932 }
1933 
1934 void GraphKit::increment_counter(Node* counter_addr) {
1935   int adr_type = Compile::AliasIdxRaw;
1936   Node* ctrl = control();
1937   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1938   Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1939   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1940 }
1941 
1942 
1943 //------------------------------uncommon_trap----------------------------------
1944 // Bail out to the interpreter in mid-method.  Implemented by calling the
1945 // uncommon_trap blob.  This helper function inserts a runtime call with the
1946 // right debug info.
1947 void GraphKit::uncommon_trap(int trap_request,
1948                              ciKlass* klass, const char* comment,
1949                              bool must_throw,
1950                              bool keep_exact_action) {
1951   if (failing())  stop();
1952   if (stopped())  return; // trap reachable?
1953 
1954   // Note:  If ProfileTraps is true, and if a deopt. actually
1955   // occurs here, the runtime will make sure an MDO exists.  There is
1956   // no need to call method()->ensure_method_data() at this point.
1957 
1958   // Set the stack pointer to the right value for reexecution:
1959   set_sp(reexecute_sp());
1960 
1961 #ifdef ASSERT
1962   if (!must_throw) {
1963     // Make sure the stack has at least enough depth to execute
1964     // the current bytecode.
1965     int inputs, ignored_depth;
1966     if (compute_stack_effects(inputs, ignored_depth)) {
1967       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1968              Bytecodes::name(java_bc()), sp(), inputs);
1969     }
1970   }
1971 #endif
1972 
1973   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1974   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1975 
1976   switch (action) {
1977   case Deoptimization::Action_maybe_recompile:
1978   case Deoptimization::Action_reinterpret:
1979     // Temporary fix for 6529811 to allow virtual calls to be sure they
1980     // get the chance to go from mono->bi->mega
1981     if (!keep_exact_action &&
1982         Deoptimization::trap_request_index(trap_request) < 0 &&
1983         too_many_recompiles(reason)) {
1984       // This BCI is causing too many recompilations.
1985       if (C->log() != NULL) {
1986         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
1987                 Deoptimization::trap_reason_name(reason),
1988                 Deoptimization::trap_action_name(action));
1989       }
1990       action = Deoptimization::Action_none;
1991       trap_request = Deoptimization::make_trap_request(reason, action);
1992     } else {
1993       C->set_trap_can_recompile(true);
1994     }
1995     break;
1996   case Deoptimization::Action_make_not_entrant:
1997     C->set_trap_can_recompile(true);
1998     break;
1999 #ifdef ASSERT
2000   case Deoptimization::Action_none:
2001   case Deoptimization::Action_make_not_compilable:
2002     break;
2003   default:
2004     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2005     break;
2006 #endif
2007   }
2008 
2009   if (TraceOptoParse) {
2010     char buf[100];
2011     tty->print_cr("Uncommon trap %s at bci:%d",
2012                   Deoptimization::format_trap_request(buf, sizeof(buf),
2013                                                       trap_request), bci());
2014   }
2015 
2016   CompileLog* log = C->log();
2017   if (log != NULL) {
2018     int kid = (klass == NULL)? -1: log->identify(klass);
2019     log->begin_elem("uncommon_trap bci='%d'", bci());
2020     char buf[100];
2021     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2022                                                           trap_request));
2023     if (kid >= 0)         log->print(" klass='%d'", kid);
2024     if (comment != NULL)  log->print(" comment='%s'", comment);
2025     log->end_elem();
2026   }
2027 
2028   // Make sure any guarding test views this path as very unlikely
2029   Node *i0 = control()->in(0);
2030   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2031     IfNode *iff = i0->as_If();
2032     float f = iff->_prob;   // Get prob
2033     if (control()->Opcode() == Op_IfTrue) {
2034       if (f > PROB_UNLIKELY_MAG(4))
2035         iff->_prob = PROB_MIN;
2036     } else {
2037       if (f < PROB_LIKELY_MAG(4))
2038         iff->_prob = PROB_MAX;
2039     }
2040   }
2041 
2042   // Clear out dead values from the debug info.
2043   kill_dead_locals();
2044 
2045   // Now insert the uncommon trap subroutine call
2046   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2047   const TypePtr* no_memory_effects = NULL;
2048   // Pass the index of the class to be loaded
2049   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2050                                  (must_throw ? RC_MUST_THROW : 0),
2051                                  OptoRuntime::uncommon_trap_Type(),
2052                                  call_addr, "uncommon_trap", no_memory_effects,
2053                                  intcon(trap_request));
2054   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2055          "must extract request correctly from the graph");
2056   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2057 
2058   call->set_req(TypeFunc::ReturnAdr, returnadr());
2059   // The debug info is the only real input to this call.
2060 
2061   // Halt-and-catch fire here.  The above call should never return!
2062   HaltNode* halt = new HaltNode(control(), frameptr());
2063   _gvn.set_type_bottom(halt);
2064   root()->add_req(halt);
2065 
2066   stop_and_kill_map();
2067 }
2068 
2069 
2070 //--------------------------just_allocated_object------------------------------
2071 // Report the object that was just allocated.
2072 // It must be the case that there are no intervening safepoints.
2073 // We use this to determine if an object is so "fresh" that
2074 // it does not require card marks.
2075 Node* GraphKit::just_allocated_object(Node* current_control) {
2076   if (C->recent_alloc_ctl() == current_control)
2077     return C->recent_alloc_obj();
2078   return NULL;
2079 }
2080 
2081 
2082 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2083   // (Note:  TypeFunc::make has a cache that makes this fast.)
2084   const TypeFunc* tf    = TypeFunc::make(dest_method);
2085   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2086   for (int j = 0; j < nargs; j++) {
2087     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2088     if( targ->basic_type() == T_DOUBLE ) {
2089       // If any parameters are doubles, they must be rounded before
2090       // the call, dstore_rounding does gvn.transform
2091       Node *arg = argument(j);
2092       arg = dstore_rounding(arg);
2093       set_argument(j, arg);
2094     }
2095   }
2096 }
2097 
2098 /**
2099  * Record profiling data exact_kls for Node n with the type system so
2100  * that it can propagate it (speculation)
2101  *
2102  * @param n          node that the type applies to
2103  * @param exact_kls  type from profiling
2104  * @param maybe_null did profiling see null?
2105  *
2106  * @return           node with improved type
2107  */
2108 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) {
2109   const Type* current_type = _gvn.type(n);
2110   assert(UseTypeSpeculation, "type speculation must be on");
2111 
2112   const TypePtr* speculative = current_type->speculative();
2113 
2114   // Should the klass from the profile be recorded in the speculative type?
2115   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2116     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2117     const TypeOopPtr* xtype = tklass->as_instance_type();
2118     assert(xtype->klass_is_exact(), "Should be exact");
2119     // Any reason to believe n is not null (from this profiling or a previous one)?
2120     const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2121     // record the new speculative type's depth
2122     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2123     speculative = speculative->with_inline_depth(jvms()->depth());
2124   } else if (current_type->would_improve_ptr(maybe_null)) {
2125     // Profiling report that null was never seen so we can change the
2126     // speculative type to non null ptr.
2127     assert(!maybe_null, "nothing to improve");
2128     if (speculative == NULL) {
2129       speculative = TypePtr::NOTNULL;
2130     } else {
2131       const TypePtr* ptr = TypePtr::NOTNULL;
2132       speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2133     }
2134   }
2135 
2136   if (speculative != current_type->speculative()) {
2137     // Build a type with a speculative type (what we think we know
2138     // about the type but will need a guard when we use it)
2139     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2140     // We're changing the type, we need a new CheckCast node to carry
2141     // the new type. The new type depends on the control: what
2142     // profiling tells us is only valid from here as far as we can
2143     // tell.
2144     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2145     cast = _gvn.transform(cast);
2146     replace_in_map(n, cast);
2147     n = cast;
2148   }
2149 
2150   return n;
2151 }
2152 
2153 /**
2154  * Record profiling data from receiver profiling at an invoke with the
2155  * type system so that it can propagate it (speculation)
2156  *
2157  * @param n  receiver node
2158  *
2159  * @return   node with improved type
2160  */
2161 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2162   if (!UseTypeSpeculation) {
2163     return n;
2164   }
2165   ciKlass* exact_kls = profile_has_unique_klass();
2166   bool maybe_null = true;
2167   if (java_bc() == Bytecodes::_checkcast ||
2168       java_bc() == Bytecodes::_instanceof ||
2169       java_bc() == Bytecodes::_aastore) {
2170     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2171     bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen();
2172   }
2173   return record_profile_for_speculation(n, exact_kls, maybe_null);
2174   return n;
2175 }
2176 
2177 /**
2178  * Record profiling data from argument profiling at an invoke with the
2179  * type system so that it can propagate it (speculation)
2180  *
2181  * @param dest_method  target method for the call
2182  * @param bc           what invoke bytecode is this?
2183  */
2184 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2185   if (!UseTypeSpeculation) {
2186     return;
2187   }
2188   const TypeFunc* tf    = TypeFunc::make(dest_method);
2189   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2190   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2191   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2192     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2193     if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2194       bool maybe_null = true;
2195       ciKlass* better_type = NULL;
2196       if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) {
2197         record_profile_for_speculation(argument(j), better_type, maybe_null);
2198       }
2199       i++;
2200     }
2201   }
2202 }
2203 
2204 /**
2205  * Record profiling data from parameter profiling at an invoke with
2206  * the type system so that it can propagate it (speculation)
2207  */
2208 void GraphKit::record_profiled_parameters_for_speculation() {
2209   if (!UseTypeSpeculation) {
2210     return;
2211   }
2212   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2213     if (_gvn.type(local(i))->isa_oopptr()) {
2214       bool maybe_null = true;
2215       ciKlass* better_type = NULL;
2216       if (method()->parameter_profiled_type(j, better_type, maybe_null)) {
2217         record_profile_for_speculation(local(i), better_type, maybe_null);
2218       }
2219       j++;
2220     }
2221   }
2222 }
2223 
2224 /**
2225  * Record profiling data from return value profiling at an invoke with
2226  * the type system so that it can propagate it (speculation)
2227  */
2228 void GraphKit::record_profiled_return_for_speculation() {
2229   if (!UseTypeSpeculation) {
2230     return;
2231   }
2232   bool maybe_null = true;
2233   ciKlass* better_type = NULL;
2234   if (method()->return_profiled_type(bci(), better_type, maybe_null)) {
2235     // If profiling reports a single type for the return value,
2236     // feed it to the type system so it can propagate it as a
2237     // speculative type
2238     record_profile_for_speculation(stack(sp()-1), better_type, maybe_null);
2239   }
2240 }
2241 
2242 void GraphKit::round_double_result(ciMethod* dest_method) {
2243   // A non-strict method may return a double value which has an extended
2244   // exponent, but this must not be visible in a caller which is 'strict'
2245   // If a strict caller invokes a non-strict callee, round a double result
2246 
2247   BasicType result_type = dest_method->return_type()->basic_type();
2248   assert( method() != NULL, "must have caller context");
2249   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2250     // Destination method's return value is on top of stack
2251     // dstore_rounding() does gvn.transform
2252     Node *result = pop_pair();
2253     result = dstore_rounding(result);
2254     push_pair(result);
2255   }
2256 }
2257 
2258 // rounding for strict float precision conformance
2259 Node* GraphKit::precision_rounding(Node* n) {
2260   return UseStrictFP && _method->flags().is_strict()
2261     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2262     ? _gvn.transform( new RoundFloatNode(0, n) )
2263     : n;
2264 }
2265 
2266 // rounding for strict double precision conformance
2267 Node* GraphKit::dprecision_rounding(Node *n) {
2268   return UseStrictFP && _method->flags().is_strict()
2269     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2270     ? _gvn.transform( new RoundDoubleNode(0, n) )
2271     : n;
2272 }
2273 
2274 // rounding for non-strict double stores
2275 Node* GraphKit::dstore_rounding(Node* n) {
2276   return Matcher::strict_fp_requires_explicit_rounding
2277     && UseSSE <= 1
2278     ? _gvn.transform( new RoundDoubleNode(0, n) )
2279     : n;
2280 }
2281 
2282 //=============================================================================
2283 // Generate a fast path/slow path idiom.  Graph looks like:
2284 // [foo] indicates that 'foo' is a parameter
2285 //
2286 //              [in]     NULL
2287 //                 \    /
2288 //                  CmpP
2289 //                  Bool ne
2290 //                   If
2291 //                  /  \
2292 //              True    False-<2>
2293 //              / |
2294 //             /  cast_not_null
2295 //           Load  |    |   ^
2296 //        [fast_test]   |   |
2297 // gvn to   opt_test    |   |
2298 //          /    \      |  <1>
2299 //      True     False  |
2300 //        |         \\  |
2301 //   [slow_call]     \[fast_result]
2302 //    Ctl   Val       \      \
2303 //     |               \      \
2304 //    Catch       <1>   \      \
2305 //   /    \        ^     \      \
2306 //  Ex    No_Ex    |      \      \
2307 //  |       \   \  |       \ <2>  \
2308 //  ...      \  [slow_res] |  |    \   [null_result]
2309 //            \         \--+--+---  |  |
2310 //             \           | /    \ | /
2311 //              --------Region     Phi
2312 //
2313 //=============================================================================
2314 // Code is structured as a series of driver functions all called 'do_XXX' that
2315 // call a set of helper functions.  Helper functions first, then drivers.
2316 
2317 //------------------------------null_check_oop---------------------------------
2318 // Null check oop.  Set null-path control into Region in slot 3.
2319 // Make a cast-not-nullness use the other not-null control.  Return cast.
2320 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2321                                bool never_see_null,
2322                                bool safe_for_replace,
2323                                bool speculative) {
2324   // Initial NULL check taken path
2325   (*null_control) = top();
2326   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2327 
2328   // Generate uncommon_trap:
2329   if (never_see_null && (*null_control) != top()) {
2330     // If we see an unexpected null at a check-cast we record it and force a
2331     // recompile; the offending check-cast will be compiled to handle NULLs.
2332     // If we see more than one offending BCI, then all checkcasts in the
2333     // method will be compiled to handle NULLs.
2334     PreserveJVMState pjvms(this);
2335     set_control(*null_control);
2336     replace_in_map(value, null());
2337     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2338     uncommon_trap(reason,
2339                   Deoptimization::Action_make_not_entrant);
2340     (*null_control) = top();    // NULL path is dead
2341   }
2342   if ((*null_control) == top() && safe_for_replace) {
2343     replace_in_map(value, cast);
2344   }
2345 
2346   // Cast away null-ness on the result
2347   return cast;
2348 }
2349 
2350 //------------------------------opt_iff----------------------------------------
2351 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2352 // Return slow-path control.
2353 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2354   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2355 
2356   // Fast path taken; set region slot 2
2357   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2358   region->init_req(2,fast_taken); // Capture fast-control
2359 
2360   // Fast path not-taken, i.e. slow path
2361   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2362   return slow_taken;
2363 }
2364 
2365 //-----------------------------make_runtime_call-------------------------------
2366 Node* GraphKit::make_runtime_call(int flags,
2367                                   const TypeFunc* call_type, address call_addr,
2368                                   const char* call_name,
2369                                   const TypePtr* adr_type,
2370                                   // The following parms are all optional.
2371                                   // The first NULL ends the list.
2372                                   Node* parm0, Node* parm1,
2373                                   Node* parm2, Node* parm3,
2374                                   Node* parm4, Node* parm5,
2375                                   Node* parm6, Node* parm7) {
2376   // Slow-path call
2377   bool is_leaf = !(flags & RC_NO_LEAF);
2378   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2379   if (call_name == NULL) {
2380     assert(!is_leaf, "must supply name for leaf");
2381     call_name = OptoRuntime::stub_name(call_addr);
2382   }
2383   CallNode* call;
2384   if (!is_leaf) {
2385     call = new CallStaticJavaNode(call_type, call_addr, call_name,
2386                                            bci(), adr_type);
2387   } else if (flags & RC_NO_FP) {
2388     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2389   } else {
2390     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2391   }
2392 
2393   // The following is similar to set_edges_for_java_call,
2394   // except that the memory effects of the call are restricted to AliasIdxRaw.
2395 
2396   // Slow path call has no side-effects, uses few values
2397   bool wide_in  = !(flags & RC_NARROW_MEM);
2398   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2399 
2400   Node* prev_mem = NULL;
2401   if (wide_in) {
2402     prev_mem = set_predefined_input_for_runtime_call(call);
2403   } else {
2404     assert(!wide_out, "narrow in => narrow out");
2405     Node* narrow_mem = memory(adr_type);
2406     prev_mem = reset_memory();
2407     map()->set_memory(narrow_mem);
2408     set_predefined_input_for_runtime_call(call);
2409   }
2410 
2411   // Hook each parm in order.  Stop looking at the first NULL.
2412   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2413   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2414   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2415   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2416   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2417   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2418   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2419   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2420     /* close each nested if ===> */  } } } } } } } }
2421   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2422 
2423   if (!is_leaf) {
2424     // Non-leaves can block and take safepoints:
2425     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2426   }
2427   // Non-leaves can throw exceptions:
2428   if (has_io) {
2429     call->set_req(TypeFunc::I_O, i_o());
2430   }
2431 
2432   if (flags & RC_UNCOMMON) {
2433     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2434     // (An "if" probability corresponds roughly to an unconditional count.
2435     // Sort of.)
2436     call->set_cnt(PROB_UNLIKELY_MAG(4));
2437   }
2438 
2439   Node* c = _gvn.transform(call);
2440   assert(c == call, "cannot disappear");
2441 
2442   if (wide_out) {
2443     // Slow path call has full side-effects.
2444     set_predefined_output_for_runtime_call(call);
2445   } else {
2446     // Slow path call has few side-effects, and/or sets few values.
2447     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2448   }
2449 
2450   if (has_io) {
2451     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2452   }
2453   return call;
2454 
2455 }
2456 
2457 //------------------------------merge_memory-----------------------------------
2458 // Merge memory from one path into the current memory state.
2459 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2460   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2461     Node* old_slice = mms.force_memory();
2462     Node* new_slice = mms.memory2();
2463     if (old_slice != new_slice) {
2464       PhiNode* phi;
2465       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2466         if (mms.is_empty()) {
2467           // clone base memory Phi's inputs for this memory slice
2468           assert(old_slice == mms.base_memory(), "sanity");
2469           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2470           _gvn.set_type(phi, Type::MEMORY);
2471           for (uint i = 1; i < phi->req(); i++) {
2472             phi->init_req(i, old_slice->in(i));
2473           }
2474         } else {
2475           phi = old_slice->as_Phi(); // Phi was generated already
2476         }
2477       } else {
2478         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2479         _gvn.set_type(phi, Type::MEMORY);
2480       }
2481       phi->set_req(new_path, new_slice);
2482       mms.set_memory(phi);
2483     }
2484   }
2485 }
2486 
2487 //------------------------------make_slow_call_ex------------------------------
2488 // Make the exception handler hookups for the slow call
2489 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2490   if (stopped())  return;
2491 
2492   // Make a catch node with just two handlers:  fall-through and catch-all
2493   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2494   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2495   Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2496   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2497 
2498   { PreserveJVMState pjvms(this);
2499     set_control(excp);
2500     set_i_o(i_o);
2501 
2502     if (excp != top()) {
2503       if (deoptimize) {
2504         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2505         uncommon_trap(Deoptimization::Reason_unhandled,
2506                       Deoptimization::Action_none);
2507       } else {
2508         // Create an exception state also.
2509         // Use an exact type if the caller has specified a specific exception.
2510         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2511         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2512         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2513       }
2514     }
2515   }
2516 
2517   // Get the no-exception control from the CatchNode.
2518   set_control(norm);
2519 }
2520 
2521 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2522   Node* cmp = NULL;
2523   switch(bt) {
2524   case T_INT: cmp = new CmpINode(in1, in2); break;
2525   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2526   default: fatal("unexpected comparison type %s", type2name(bt));
2527   }
2528   gvn->transform(cmp);
2529   Node* bol = gvn->transform(new BoolNode(cmp, test));
2530   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2531   gvn->transform(iff);
2532   if (!bol->is_Con()) gvn->record_for_igvn(iff);
2533   return iff;
2534 }
2535 
2536 
2537 //-------------------------------gen_subtype_check-----------------------------
2538 // Generate a subtyping check.  Takes as input the subtype and supertype.
2539 // Returns 2 values: sets the default control() to the true path and returns
2540 // the false path.  Only reads invariant memory; sets no (visible) memory.
2541 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2542 // but that's not exposed to the optimizer.  This call also doesn't take in an
2543 // Object; if you wish to check an Object you need to load the Object's class
2544 // prior to coming here.
2545 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2546   Compile* C = gvn->C;
2547 
2548   if ((*ctrl)->is_top()) {
2549     return C->top();
2550   }
2551 
2552   // Fast check for identical types, perhaps identical constants.
2553   // The types can even be identical non-constants, in cases
2554   // involving Array.newInstance, Object.clone, etc.
2555   if (subklass == superklass)
2556     return C->top();             // false path is dead; no test needed.
2557 
2558   if (gvn->type(superklass)->singleton()) {
2559     ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2560     ciKlass* subk   = gvn->type(subklass)->is_klassptr()->klass();
2561 
2562     // In the common case of an exact superklass, try to fold up the
2563     // test before generating code.  You may ask, why not just generate
2564     // the code and then let it fold up?  The answer is that the generated
2565     // code will necessarily include null checks, which do not always
2566     // completely fold away.  If they are also needless, then they turn
2567     // into a performance loss.  Example:
2568     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2569     // Here, the type of 'fa' is often exact, so the store check
2570     // of fa[1]=x will fold up, without testing the nullness of x.
2571     switch (C->static_subtype_check(superk, subk)) {
2572     case Compile::SSC_always_false:
2573       {
2574         Node* always_fail = *ctrl;
2575         *ctrl = gvn->C->top();
2576         return always_fail;
2577       }
2578     case Compile::SSC_always_true:
2579       return C->top();
2580     case Compile::SSC_easy_test:
2581       {
2582         // Just do a direct pointer compare and be done.
2583         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2584         *ctrl = gvn->transform(new IfTrueNode(iff));
2585         return gvn->transform(new IfFalseNode(iff));
2586       }
2587     case Compile::SSC_full_test:
2588       break;
2589     default:
2590       ShouldNotReachHere();
2591     }
2592   }
2593 
2594   // %%% Possible further optimization:  Even if the superklass is not exact,
2595   // if the subklass is the unique subtype of the superklass, the check
2596   // will always succeed.  We could leave a dependency behind to ensure this.
2597 
2598   // First load the super-klass's check-offset
2599   Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2600   Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2601   Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2602   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2603   bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2604 
2605   // Load from the sub-klass's super-class display list, or a 1-word cache of
2606   // the secondary superclass list, or a failing value with a sentinel offset
2607   // if the super-klass is an interface or exceptionally deep in the Java
2608   // hierarchy and we have to scan the secondary superclass list the hard way.
2609   // Worst-case type is a little odd: NULL is allowed as a result (usually
2610   // klass loads can never produce a NULL).
2611   Node *chk_off_X = chk_off;
2612 #ifdef _LP64
2613   chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2614 #endif
2615   Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2616   // For some types like interfaces the following loadKlass is from a 1-word
2617   // cache which is mutable so can't use immutable memory.  Other
2618   // types load from the super-class display table which is immutable.
2619   m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2620   Node *kmem = might_be_cache ? m : C->immutable_memory();
2621   Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2622 
2623   // Compile speed common case: ARE a subtype and we canNOT fail
2624   if( superklass == nkls )
2625     return C->top();             // false path is dead; no test needed.
2626 
2627   // See if we get an immediate positive hit.  Happens roughly 83% of the
2628   // time.  Test to see if the value loaded just previously from the subklass
2629   // is exactly the superklass.
2630   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2631   Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2632   *ctrl = gvn->transform(new IfFalseNode(iff1));
2633 
2634   // Compile speed common case: Check for being deterministic right now.  If
2635   // chk_off is a constant and not equal to cacheoff then we are NOT a
2636   // subklass.  In this case we need exactly the 1 test above and we can
2637   // return those results immediately.
2638   if (!might_be_cache) {
2639     Node* not_subtype_ctrl = *ctrl;
2640     *ctrl = iftrue1; // We need exactly the 1 test above
2641     return not_subtype_ctrl;
2642   }
2643 
2644   // Gather the various success & failures here
2645   RegionNode *r_ok_subtype = new RegionNode(4);
2646   gvn->record_for_igvn(r_ok_subtype);
2647   RegionNode *r_not_subtype = new RegionNode(3);
2648   gvn->record_for_igvn(r_not_subtype);
2649 
2650   r_ok_subtype->init_req(1, iftrue1);
2651 
2652   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2653   // is roughly 63% of the remaining cases).  Test to see if the loaded
2654   // check-offset points into the subklass display list or the 1-element
2655   // cache.  If it points to the display (and NOT the cache) and the display
2656   // missed then it's not a subtype.
2657   Node *cacheoff = gvn->intcon(cacheoff_con);
2658   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2659   r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2660   *ctrl = gvn->transform(new IfFalseNode(iff2));
2661 
2662   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2663   // No performance impact (too rare) but allows sharing of secondary arrays
2664   // which has some footprint reduction.
2665   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2666   r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2667   *ctrl = gvn->transform(new IfFalseNode(iff3));
2668 
2669   // -- Roads not taken here: --
2670   // We could also have chosen to perform the self-check at the beginning
2671   // of this code sequence, as the assembler does.  This would not pay off
2672   // the same way, since the optimizer, unlike the assembler, can perform
2673   // static type analysis to fold away many successful self-checks.
2674   // Non-foldable self checks work better here in second position, because
2675   // the initial primary superclass check subsumes a self-check for most
2676   // types.  An exception would be a secondary type like array-of-interface,
2677   // which does not appear in its own primary supertype display.
2678   // Finally, we could have chosen to move the self-check into the
2679   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2680   // dependent manner.  But it is worthwhile to have the check here,
2681   // where it can be perhaps be optimized.  The cost in code space is
2682   // small (register compare, branch).
2683 
2684   // Now do a linear scan of the secondary super-klass array.  Again, no real
2685   // performance impact (too rare) but it's gotta be done.
2686   // Since the code is rarely used, there is no penalty for moving it
2687   // out of line, and it can only improve I-cache density.
2688   // The decision to inline or out-of-line this final check is platform
2689   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2690   Node* psc = gvn->transform(
2691     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2692 
2693   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2694   r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2695   r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2696 
2697   // Return false path; set default control to true path.
2698   *ctrl = gvn->transform(r_ok_subtype);
2699   return gvn->transform(r_not_subtype);
2700 }
2701 
2702 // Profile-driven exact type check:
2703 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2704                                     float prob,
2705                                     Node* *casted_receiver) {
2706   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2707   Node* recv_klass = load_object_klass(receiver);
2708   Node* want_klass = makecon(tklass);
2709   Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2710   Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2711   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2712   set_control( _gvn.transform( new IfTrueNode (iff) ));
2713   Node* fail = _gvn.transform( new IfFalseNode(iff) );
2714 
2715   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2716   assert(recv_xtype->klass_is_exact(), "");
2717 
2718   // Subsume downstream occurrences of receiver with a cast to
2719   // recv_xtype, since now we know what the type will be.
2720   Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2721   (*casted_receiver) = _gvn.transform(cast);
2722   // (User must make the replace_in_map call.)
2723 
2724   return fail;
2725 }
2726 
2727 
2728 //------------------------------seems_never_null-------------------------------
2729 // Use null_seen information if it is available from the profile.
2730 // If we see an unexpected null at a type check we record it and force a
2731 // recompile; the offending check will be recompiled to handle NULLs.
2732 // If we see several offending BCIs, then all checks in the
2733 // method will be recompiled.
2734 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2735   speculating = !_gvn.type(obj)->speculative_maybe_null();
2736   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2737   if (UncommonNullCast               // Cutout for this technique
2738       && obj != null()               // And not the -Xcomp stupid case?
2739       && !too_many_traps(reason)
2740       ) {
2741     if (speculating) {
2742       return true;
2743     }
2744     if (data == NULL)
2745       // Edge case:  no mature data.  Be optimistic here.
2746       return true;
2747     // If the profile has not seen a null, assume it won't happen.
2748     assert(java_bc() == Bytecodes::_checkcast ||
2749            java_bc() == Bytecodes::_instanceof ||
2750            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2751     return !data->as_BitData()->null_seen();
2752   }
2753   speculating = false;
2754   return false;
2755 }
2756 
2757 //------------------------maybe_cast_profiled_receiver-------------------------
2758 // If the profile has seen exactly one type, narrow to exactly that type.
2759 // Subsequent type checks will always fold up.
2760 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2761                                              ciKlass* require_klass,
2762                                              ciKlass* spec_klass,
2763                                              bool safe_for_replace) {
2764   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2765 
2766   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2767 
2768   // Make sure we haven't already deoptimized from this tactic.
2769   if (too_many_traps(reason) || too_many_recompiles(reason))
2770     return NULL;
2771 
2772   // (No, this isn't a call, but it's enough like a virtual call
2773   // to use the same ciMethod accessor to get the profile info...)
2774   // If we have a speculative type use it instead of profiling (which
2775   // may not help us)
2776   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2777   if (exact_kls != NULL) {// no cast failures here
2778     if (require_klass == NULL ||
2779         C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2780       // If we narrow the type to match what the type profile sees or
2781       // the speculative type, we can then remove the rest of the
2782       // cast.
2783       // This is a win, even if the exact_kls is very specific,
2784       // because downstream operations, such as method calls,
2785       // will often benefit from the sharper type.
2786       Node* exact_obj = not_null_obj; // will get updated in place...
2787       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2788                                             &exact_obj);
2789       { PreserveJVMState pjvms(this);
2790         set_control(slow_ctl);
2791         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2792       }
2793       if (safe_for_replace) {
2794         replace_in_map(not_null_obj, exact_obj);
2795       }
2796       return exact_obj;
2797     }
2798     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2799   }
2800 
2801   return NULL;
2802 }
2803 
2804 /**
2805  * Cast obj to type and emit guard unless we had too many traps here
2806  * already
2807  *
2808  * @param obj       node being casted
2809  * @param type      type to cast the node to
2810  * @param not_null  true if we know node cannot be null
2811  */
2812 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2813                                         ciKlass* type,
2814                                         bool not_null) {
2815   if (stopped()) {
2816     return obj;
2817   }
2818 
2819   // type == NULL if profiling tells us this object is always null
2820   if (type != NULL) {
2821     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2822     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2823 
2824     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2825         !too_many_traps(class_reason) &&
2826         !too_many_recompiles(class_reason)) {
2827       Node* not_null_obj = NULL;
2828       // not_null is true if we know the object is not null and
2829       // there's no need for a null check
2830       if (!not_null) {
2831         Node* null_ctl = top();
2832         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2833         assert(null_ctl->is_top(), "no null control here");
2834       } else {
2835         not_null_obj = obj;
2836       }
2837 
2838       Node* exact_obj = not_null_obj;
2839       ciKlass* exact_kls = type;
2840       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2841                                             &exact_obj);
2842       {
2843         PreserveJVMState pjvms(this);
2844         set_control(slow_ctl);
2845         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2846       }
2847       replace_in_map(not_null_obj, exact_obj);
2848       obj = exact_obj;
2849     }
2850   } else {
2851     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2852         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2853       Node* exact_obj = null_assert(obj);
2854       replace_in_map(obj, exact_obj);
2855       obj = exact_obj;
2856     }
2857   }
2858   return obj;
2859 }
2860 
2861 //-------------------------------gen_instanceof--------------------------------
2862 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2863 // and the reflective instance-of call.
2864 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2865   kill_dead_locals();           // Benefit all the uncommon traps
2866   assert( !stopped(), "dead parse path should be checked in callers" );
2867   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2868          "must check for not-null not-dead klass in callers");
2869 
2870   // Make the merge point
2871   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2872   RegionNode* region = new RegionNode(PATH_LIMIT);
2873   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
2874   C->set_has_split_ifs(true); // Has chance for split-if optimization
2875 
2876   ciProfileData* data = NULL;
2877   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2878     data = method()->method_data()->bci_to_data(bci());
2879   }
2880   bool speculative_not_null = false;
2881   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2882                          && seems_never_null(obj, data, speculative_not_null));
2883 
2884   // Null check; get casted pointer; set region slot 3
2885   Node* null_ctl = top();
2886   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2887 
2888   // If not_null_obj is dead, only null-path is taken
2889   if (stopped()) {              // Doing instance-of on a NULL?
2890     set_control(null_ctl);
2891     return intcon(0);
2892   }
2893   region->init_req(_null_path, null_ctl);
2894   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2895   if (null_ctl == top()) {
2896     // Do this eagerly, so that pattern matches like is_diamond_phi
2897     // will work even during parsing.
2898     assert(_null_path == PATH_LIMIT-1, "delete last");
2899     region->del_req(_null_path);
2900     phi   ->del_req(_null_path);
2901   }
2902 
2903   // Do we know the type check always succeed?
2904   bool known_statically = false;
2905   if (_gvn.type(superklass)->singleton()) {
2906     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2907     ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2908     if (subk != NULL && subk->is_loaded()) {
2909       int static_res = C->static_subtype_check(superk, subk);
2910       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2911     }
2912   }
2913 
2914   if (known_statically && UseTypeSpeculation) {
2915     // If we know the type check always succeeds then we don't use the
2916     // profiling data at this bytecode. Don't lose it, feed it to the
2917     // type system as a speculative type.
2918     not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2919   } else {
2920     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2921     // We may not have profiling here or it may not help us. If we
2922     // have a speculative type use it to perform an exact cast.
2923     ciKlass* spec_obj_type = obj_type->speculative_type();
2924     if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2925       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2926       if (stopped()) {            // Profile disagrees with this path.
2927         set_control(null_ctl);    // Null is the only remaining possibility.
2928         return intcon(0);
2929       }
2930       if (cast_obj != NULL) {
2931         not_null_obj = cast_obj;
2932       }
2933     }
2934   }
2935 
2936   // Load the object's klass
2937   Node* obj_klass = load_object_klass(not_null_obj);
2938 
2939   // Generate the subtype check
2940   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2941 
2942   // Plug in the success path to the general merge in slot 1.
2943   region->init_req(_obj_path, control());
2944   phi   ->init_req(_obj_path, intcon(1));
2945 
2946   // Plug in the failing path to the general merge in slot 2.
2947   region->init_req(_fail_path, not_subtype_ctrl);
2948   phi   ->init_req(_fail_path, intcon(0));
2949 
2950   // Return final merged results
2951   set_control( _gvn.transform(region) );
2952   record_for_igvn(region);
2953   return _gvn.transform(phi);
2954 }
2955 
2956 //-------------------------------gen_checkcast---------------------------------
2957 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
2958 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
2959 // uncommon-trap paths work.  Adjust stack after this call.
2960 // If failure_control is supplied and not null, it is filled in with
2961 // the control edge for the cast failure.  Otherwise, an appropriate
2962 // uncommon trap or exception is thrown.
2963 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2964                               Node* *failure_control) {
2965   kill_dead_locals();           // Benefit all the uncommon traps
2966   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2967   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2968 
2969   // Fast cutout:  Check the case that the cast is vacuously true.
2970   // This detects the common cases where the test will short-circuit
2971   // away completely.  We do this before we perform the null check,
2972   // because if the test is going to turn into zero code, we don't
2973   // want a residual null check left around.  (Causes a slowdown,
2974   // for example, in some objArray manipulations, such as a[i]=a[j].)
2975   if (tk->singleton()) {
2976     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2977     if (objtp != NULL && objtp->klass() != NULL) {
2978       switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
2979       case Compile::SSC_always_true:
2980         // If we know the type check always succeed then we don't use
2981         // the profiling data at this bytecode. Don't lose it, feed it
2982         // to the type system as a speculative type.
2983         return record_profiled_receiver_for_speculation(obj);
2984       case Compile::SSC_always_false:
2985         // It needs a null check because a null will *pass* the cast check.
2986         // A non-null value will always produce an exception.
2987         return null_assert(obj);
2988       }
2989     }
2990   }
2991 
2992   ciProfileData* data = NULL;
2993   bool safe_for_replace = false;
2994   if (failure_control == NULL) {        // use MDO in regular case only
2995     assert(java_bc() == Bytecodes::_aastore ||
2996            java_bc() == Bytecodes::_checkcast,
2997            "interpreter profiles type checks only for these BCs");
2998     data = method()->method_data()->bci_to_data(bci());
2999     safe_for_replace = true;
3000   }
3001 
3002   // Make the merge point
3003   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3004   RegionNode* region = new RegionNode(PATH_LIMIT);
3005   Node*       phi    = new PhiNode(region, toop);
3006   C->set_has_split_ifs(true); // Has chance for split-if optimization
3007 
3008   // Use null-cast information if it is available
3009   bool speculative_not_null = false;
3010   bool never_see_null = ((failure_control == NULL)  // regular case only
3011                          && seems_never_null(obj, data, speculative_not_null));
3012 
3013   // Null check; get casted pointer; set region slot 3
3014   Node* null_ctl = top();
3015   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3016 
3017   // If not_null_obj is dead, only null-path is taken
3018   if (stopped()) {              // Doing instance-of on a NULL?
3019     set_control(null_ctl);
3020     return null();
3021   }
3022   region->init_req(_null_path, null_ctl);
3023   phi   ->init_req(_null_path, null());  // Set null path value
3024   if (null_ctl == top()) {
3025     // Do this eagerly, so that pattern matches like is_diamond_phi
3026     // will work even during parsing.
3027     assert(_null_path == PATH_LIMIT-1, "delete last");
3028     region->del_req(_null_path);
3029     phi   ->del_req(_null_path);
3030   }
3031 
3032   Node* cast_obj = NULL;
3033   if (tk->klass_is_exact()) {
3034     // The following optimization tries to statically cast the speculative type of the object
3035     // (for example obtained during profiling) to the type of the superklass and then do a
3036     // dynamic check that the type of the object is what we expect. To work correctly
3037     // for checkcast and aastore the type of superklass should be exact.
3038     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3039     // We may not have profiling here or it may not help us. If we have
3040     // a speculative type use it to perform an exact cast.
3041     ciKlass* spec_obj_type = obj_type->speculative_type();
3042     if (spec_obj_type != NULL || data != NULL) {
3043       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3044       if (cast_obj != NULL) {
3045         if (failure_control != NULL) // failure is now impossible
3046           (*failure_control) = top();
3047         // adjust the type of the phi to the exact klass:
3048         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3049       }
3050     }
3051   }
3052 
3053   if (cast_obj == NULL) {
3054     // Load the object's klass
3055     Node* obj_klass = load_object_klass(not_null_obj);
3056 
3057     // Generate the subtype check
3058     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3059 
3060     // Plug in success path into the merge
3061     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3062     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3063     if (failure_control == NULL) {
3064       if (not_subtype_ctrl != top()) { // If failure is possible
3065         PreserveJVMState pjvms(this);
3066         set_control(not_subtype_ctrl);
3067         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3068       }
3069     } else {
3070       (*failure_control) = not_subtype_ctrl;
3071     }
3072   }
3073 
3074   region->init_req(_obj_path, control());
3075   phi   ->init_req(_obj_path, cast_obj);
3076 
3077   // A merge of NULL or Casted-NotNull obj
3078   Node* res = _gvn.transform(phi);
3079 
3080   // Note I do NOT always 'replace_in_map(obj,result)' here.
3081   //  if( tk->klass()->can_be_primary_super()  )
3082     // This means that if I successfully store an Object into an array-of-String
3083     // I 'forget' that the Object is really now known to be a String.  I have to
3084     // do this because we don't have true union types for interfaces - if I store
3085     // a Baz into an array-of-Interface and then tell the optimizer it's an
3086     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3087     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3088   //  replace_in_map( obj, res );
3089 
3090   // Return final merged results
3091   set_control( _gvn.transform(region) );
3092   record_for_igvn(region);
3093   return res;
3094 }
3095 
3096 //------------------------------next_monitor-----------------------------------
3097 // What number should be given to the next monitor?
3098 int GraphKit::next_monitor() {
3099   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3100   int next = current + C->sync_stack_slots();
3101   // Keep the toplevel high water mark current:
3102   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3103   return current;
3104 }
3105 
3106 //------------------------------insert_mem_bar---------------------------------
3107 // Memory barrier to avoid floating things around
3108 // The membar serves as a pinch point between both control and all memory slices.
3109 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3110   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3111   mb->init_req(TypeFunc::Control, control());
3112   mb->init_req(TypeFunc::Memory,  reset_memory());
3113   Node* membar = _gvn.transform(mb);
3114   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3115   set_all_memory_call(membar);
3116   return membar;
3117 }
3118 
3119 //-------------------------insert_mem_bar_volatile----------------------------
3120 // Memory barrier to avoid floating things around
3121 // The membar serves as a pinch point between both control and memory(alias_idx).
3122 // If you want to make a pinch point on all memory slices, do not use this
3123 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3124 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3125   // When Parse::do_put_xxx updates a volatile field, it appends a series
3126   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3127   // The first membar is on the same memory slice as the field store opcode.
3128   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3129   // All the other membars (for other volatile slices, including AliasIdxBot,
3130   // which stands for all unknown volatile slices) are control-dependent
3131   // on the first membar.  This prevents later volatile loads or stores
3132   // from sliding up past the just-emitted store.
3133 
3134   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3135   mb->set_req(TypeFunc::Control,control());
3136   if (alias_idx == Compile::AliasIdxBot) {
3137     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3138   } else {
3139     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3140     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3141   }
3142   Node* membar = _gvn.transform(mb);
3143   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3144   if (alias_idx == Compile::AliasIdxBot) {
3145     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3146   } else {
3147     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3148   }
3149   return membar;
3150 }
3151 
3152 //------------------------------shared_lock------------------------------------
3153 // Emit locking code.
3154 FastLockNode* GraphKit::shared_lock(Node* obj) {
3155   // bci is either a monitorenter bc or InvocationEntryBci
3156   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3157   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3158 
3159   if( !GenerateSynchronizationCode )
3160     return NULL;                // Not locking things?
3161   if (stopped())                // Dead monitor?
3162     return NULL;
3163 
3164   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3165 
3166   // Box the stack location
3167   Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3168   Node* mem = reset_memory();
3169 
3170   FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3171   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3172     // Create the counters for this fast lock.
3173     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3174   }
3175 
3176   // Create the rtm counters for this fast lock if needed.
3177   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3178 
3179   // Add monitor to debug info for the slow path.  If we block inside the
3180   // slow path and de-opt, we need the monitor hanging around
3181   map()->push_monitor( flock );
3182 
3183   const TypeFunc *tf = LockNode::lock_type();
3184   LockNode *lock = new LockNode(C, tf);
3185 
3186   lock->init_req( TypeFunc::Control, control() );
3187   lock->init_req( TypeFunc::Memory , mem );
3188   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3189   lock->init_req( TypeFunc::FramePtr, frameptr() );
3190   lock->init_req( TypeFunc::ReturnAdr, top() );
3191 
3192   lock->init_req(TypeFunc::Parms + 0, obj);
3193   lock->init_req(TypeFunc::Parms + 1, box);
3194   lock->init_req(TypeFunc::Parms + 2, flock);
3195   add_safepoint_edges(lock);
3196 
3197   lock = _gvn.transform( lock )->as_Lock();
3198 
3199   // lock has no side-effects, sets few values
3200   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3201 
3202   insert_mem_bar(Op_MemBarAcquireLock);
3203 
3204   // Add this to the worklist so that the lock can be eliminated
3205   record_for_igvn(lock);
3206 
3207 #ifndef PRODUCT
3208   if (PrintLockStatistics) {
3209     // Update the counter for this lock.  Don't bother using an atomic
3210     // operation since we don't require absolute accuracy.
3211     lock->create_lock_counter(map()->jvms());
3212     increment_counter(lock->counter()->addr());
3213   }
3214 #endif
3215 
3216   return flock;
3217 }
3218 
3219 
3220 //------------------------------shared_unlock----------------------------------
3221 // Emit unlocking code.
3222 void GraphKit::shared_unlock(Node* box, Node* obj) {
3223   // bci is either a monitorenter bc or InvocationEntryBci
3224   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3225   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3226 
3227   if( !GenerateSynchronizationCode )
3228     return;
3229   if (stopped()) {               // Dead monitor?
3230     map()->pop_monitor();        // Kill monitor from debug info
3231     return;
3232   }
3233 
3234   // Memory barrier to avoid floating things down past the locked region
3235   insert_mem_bar(Op_MemBarReleaseLock);
3236 
3237   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3238   UnlockNode *unlock = new UnlockNode(C, tf);
3239 #ifdef ASSERT
3240   unlock->set_dbg_jvms(sync_jvms());
3241 #endif
3242   uint raw_idx = Compile::AliasIdxRaw;
3243   unlock->init_req( TypeFunc::Control, control() );
3244   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3245   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3246   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3247   unlock->init_req( TypeFunc::ReturnAdr, top() );
3248 
3249   unlock->init_req(TypeFunc::Parms + 0, obj);
3250   unlock->init_req(TypeFunc::Parms + 1, box);
3251   unlock = _gvn.transform(unlock)->as_Unlock();
3252 
3253   Node* mem = reset_memory();
3254 
3255   // unlock has no side-effects, sets few values
3256   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3257 
3258   // Kill monitor from debug info
3259   map()->pop_monitor( );
3260 }
3261 
3262 //-------------------------------get_layout_helper-----------------------------
3263 // If the given klass is a constant or known to be an array,
3264 // fetch the constant layout helper value into constant_value
3265 // and return (Node*)NULL.  Otherwise, load the non-constant
3266 // layout helper value, and return the node which represents it.
3267 // This two-faced routine is useful because allocation sites
3268 // almost always feature constant types.
3269 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3270   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3271   if (!StressReflectiveCode && inst_klass != NULL) {
3272     ciKlass* klass = inst_klass->klass();
3273     bool    xklass = inst_klass->klass_is_exact();
3274     if (xklass || klass->is_array_klass()) {
3275       jint lhelper = klass->layout_helper();
3276       if (lhelper != Klass::_lh_neutral_value) {
3277         constant_value = lhelper;
3278         return (Node*) NULL;
3279       }
3280     }
3281   }
3282   constant_value = Klass::_lh_neutral_value;  // put in a known value
3283   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3284   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3285 }
3286 
3287 // We just put in an allocate/initialize with a big raw-memory effect.
3288 // Hook selected additional alias categories on the initialization.
3289 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3290                                 MergeMemNode* init_in_merge,
3291                                 Node* init_out_raw) {
3292   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3293   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3294 
3295   Node* prevmem = kit.memory(alias_idx);
3296   init_in_merge->set_memory_at(alias_idx, prevmem);
3297   kit.set_memory(init_out_raw, alias_idx);
3298 }
3299 
3300 //---------------------------set_output_for_allocation-------------------------
3301 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3302                                           const TypeOopPtr* oop_type,
3303                                           bool deoptimize_on_exception) {
3304   int rawidx = Compile::AliasIdxRaw;
3305   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3306   add_safepoint_edges(alloc);
3307   Node* allocx = _gvn.transform(alloc);
3308   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3309   // create memory projection for i_o
3310   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3311   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3312 
3313   // create a memory projection as for the normal control path
3314   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3315   set_memory(malloc, rawidx);
3316 
3317   // a normal slow-call doesn't change i_o, but an allocation does
3318   // we create a separate i_o projection for the normal control path
3319   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3320   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3321 
3322   // put in an initialization barrier
3323   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3324                                                  rawoop)->as_Initialize();
3325   assert(alloc->initialization() == init,  "2-way macro link must work");
3326   assert(init ->allocation()     == alloc, "2-way macro link must work");
3327   {
3328     // Extract memory strands which may participate in the new object's
3329     // initialization, and source them from the new InitializeNode.
3330     // This will allow us to observe initializations when they occur,
3331     // and link them properly (as a group) to the InitializeNode.
3332     assert(init->in(InitializeNode::Memory) == malloc, "");
3333     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3334     init->set_req(InitializeNode::Memory, minit_in);
3335     record_for_igvn(minit_in); // fold it up later, if possible
3336     Node* minit_out = memory(rawidx);
3337     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3338     if (oop_type->isa_aryptr()) {
3339       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3340       int            elemidx  = C->get_alias_index(telemref);
3341       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3342     } else if (oop_type->isa_instptr()) {
3343       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3344       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3345         ciField* field = ik->nonstatic_field_at(i);
3346         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3347           continue;  // do not bother to track really large numbers of fields
3348         // Find (or create) the alias category for this field:
3349         int fieldidx = C->alias_type(field)->index();
3350         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3351       }
3352     }
3353   }
3354 
3355   // Cast raw oop to the real thing...
3356   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3357   javaoop = _gvn.transform(javaoop);
3358   C->set_recent_alloc(control(), javaoop);
3359   assert(just_allocated_object(control()) == javaoop, "just allocated");
3360 
3361 #ifdef ASSERT
3362   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3363     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3364            "Ideal_allocation works");
3365     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3366            "Ideal_allocation works");
3367     if (alloc->is_AllocateArray()) {
3368       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3369              "Ideal_allocation works");
3370       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3371              "Ideal_allocation works");
3372     } else {
3373       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3374     }
3375   }
3376 #endif //ASSERT
3377 
3378   return javaoop;
3379 }
3380 
3381 //---------------------------new_instance--------------------------------------
3382 // This routine takes a klass_node which may be constant (for a static type)
3383 // or may be non-constant (for reflective code).  It will work equally well
3384 // for either, and the graph will fold nicely if the optimizer later reduces
3385 // the type to a constant.
3386 // The optional arguments are for specialized use by intrinsics:
3387 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3388 //  - If 'return_size_val', report the the total object size to the caller.
3389 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3390 Node* GraphKit::new_instance(Node* klass_node,
3391                              Node* extra_slow_test,
3392                              Node* *return_size_val,
3393                              bool deoptimize_on_exception) {
3394   // Compute size in doublewords
3395   // The size is always an integral number of doublewords, represented
3396   // as a positive bytewise size stored in the klass's layout_helper.
3397   // The layout_helper also encodes (in a low bit) the need for a slow path.
3398   jint  layout_con = Klass::_lh_neutral_value;
3399   Node* layout_val = get_layout_helper(klass_node, layout_con);
3400   int   layout_is_con = (layout_val == NULL);
3401 
3402   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3403   // Generate the initial go-slow test.  It's either ALWAYS (return a
3404   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3405   // case) a computed value derived from the layout_helper.
3406   Node* initial_slow_test = NULL;
3407   if (layout_is_con) {
3408     assert(!StressReflectiveCode, "stress mode does not use these paths");
3409     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3410     initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3411 
3412   } else {   // reflective case
3413     // This reflective path is used by Unsafe.allocateInstance.
3414     // (It may be stress-tested by specifying StressReflectiveCode.)
3415     // Basically, we want to get into the VM is there's an illegal argument.
3416     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3417     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3418     if (extra_slow_test != intcon(0)) {
3419       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3420     }
3421     // (Macro-expander will further convert this to a Bool, if necessary.)
3422   }
3423 
3424   // Find the size in bytes.  This is easy; it's the layout_helper.
3425   // The size value must be valid even if the slow path is taken.
3426   Node* size = NULL;
3427   if (layout_is_con) {
3428     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3429   } else {   // reflective case
3430     // This reflective path is used by clone and Unsafe.allocateInstance.
3431     size = ConvI2X(layout_val);
3432 
3433     // Clear the low bits to extract layout_helper_size_in_bytes:
3434     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3435     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3436     size = _gvn.transform( new AndXNode(size, mask) );
3437   }
3438   if (return_size_val != NULL) {
3439     (*return_size_val) = size;
3440   }
3441 
3442   // This is a precise notnull oop of the klass.
3443   // (Actually, it need not be precise if this is a reflective allocation.)
3444   // It's what we cast the result to.
3445   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3446   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3447   const TypeOopPtr* oop_type = tklass->as_instance_type();
3448 
3449   // Now generate allocation code
3450 
3451   // The entire memory state is needed for slow path of the allocation
3452   // since GC and deoptimization can happened.
3453   Node *mem = reset_memory();
3454   set_all_memory(mem); // Create new memory state
3455 
3456   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3457                                          control(), mem, i_o(),
3458                                          size, klass_node,
3459                                          initial_slow_test);
3460 
3461   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3462 }
3463 
3464 //-------------------------------new_array-------------------------------------
3465 // helper for both newarray and anewarray
3466 // The 'length' parameter is (obviously) the length of the array.
3467 // See comments on new_instance for the meaning of the other arguments.
3468 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3469                           Node* length,         // number of array elements
3470                           int   nargs,          // number of arguments to push back for uncommon trap
3471                           Node* *return_size_val,
3472                           bool deoptimize_on_exception) {
3473   jint  layout_con = Klass::_lh_neutral_value;
3474   Node* layout_val = get_layout_helper(klass_node, layout_con);
3475   int   layout_is_con = (layout_val == NULL);
3476 
3477   if (!layout_is_con && !StressReflectiveCode &&
3478       !too_many_traps(Deoptimization::Reason_class_check)) {
3479     // This is a reflective array creation site.
3480     // Optimistically assume that it is a subtype of Object[],
3481     // so that we can fold up all the address arithmetic.
3482     layout_con = Klass::array_layout_helper(T_OBJECT);
3483     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3484     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3485     { BuildCutout unless(this, bol_lh, PROB_MAX);
3486       inc_sp(nargs);
3487       uncommon_trap(Deoptimization::Reason_class_check,
3488                     Deoptimization::Action_maybe_recompile);
3489     }
3490     layout_val = NULL;
3491     layout_is_con = true;
3492   }
3493 
3494   // Generate the initial go-slow test.  Make sure we do not overflow
3495   // if length is huge (near 2Gig) or negative!  We do not need
3496   // exact double-words here, just a close approximation of needed
3497   // double-words.  We can't add any offset or rounding bits, lest we
3498   // take a size -1 of bytes and make it positive.  Use an unsigned
3499   // compare, so negative sizes look hugely positive.
3500   int fast_size_limit = FastAllocateSizeLimit;
3501   if (layout_is_con) {
3502     assert(!StressReflectiveCode, "stress mode does not use these paths");
3503     // Increase the size limit if we have exact knowledge of array type.
3504     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3505     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3506   }
3507 
3508   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3509   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3510   if (initial_slow_test->is_Bool()) {
3511     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3512     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3513   }
3514 
3515   // --- Size Computation ---
3516   // array_size = round_to_heap(array_header + (length << elem_shift));
3517   // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3518   // and round_to(x, y) == ((x + y-1) & ~(y-1))
3519   // The rounding mask is strength-reduced, if possible.
3520   int round_mask = MinObjAlignmentInBytes - 1;
3521   Node* header_size = NULL;
3522   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3523   // (T_BYTE has the weakest alignment and size restrictions...)
3524   if (layout_is_con) {
3525     int       hsize  = Klass::layout_helper_header_size(layout_con);
3526     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3527     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3528     if ((round_mask & ~right_n_bits(eshift)) == 0)
3529       round_mask = 0;  // strength-reduce it if it goes away completely
3530     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3531     assert(header_size_min <= hsize, "generic minimum is smallest");
3532     header_size_min = hsize;
3533     header_size = intcon(hsize + round_mask);
3534   } else {
3535     Node* hss   = intcon(Klass::_lh_header_size_shift);
3536     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3537     Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3538     hsize       = _gvn.transform( new AndINode(hsize, hsm) );
3539     Node* mask  = intcon(round_mask);
3540     header_size = _gvn.transform( new AddINode(hsize, mask) );
3541   }
3542 
3543   Node* elem_shift = NULL;
3544   if (layout_is_con) {
3545     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3546     if (eshift != 0)
3547       elem_shift = intcon(eshift);
3548   } else {
3549     // There is no need to mask or shift this value.
3550     // The semantics of LShiftINode include an implicit mask to 0x1F.
3551     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3552     elem_shift = layout_val;
3553   }
3554 
3555   // Transition to native address size for all offset calculations:
3556   Node* lengthx = ConvI2X(length);
3557   Node* headerx = ConvI2X(header_size);
3558 #ifdef _LP64
3559   { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3560     if (tllen != NULL && tllen->_lo < 0) {
3561       // Add a manual constraint to a positive range.  Cf. array_element_address.
3562       jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3563       if (size_max > tllen->_hi)  size_max = tllen->_hi;
3564       const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3565       lengthx = _gvn.transform( new ConvI2LNode(length, tlcon));
3566     }
3567   }
3568 #endif
3569 
3570   // Combine header size (plus rounding) and body size.  Then round down.
3571   // This computation cannot overflow, because it is used only in two
3572   // places, one where the length is sharply limited, and the other
3573   // after a successful allocation.
3574   Node* abody = lengthx;
3575   if (elem_shift != NULL)
3576     abody     = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3577   Node* size  = _gvn.transform( new AddXNode(headerx, abody) );
3578   if (round_mask != 0) {
3579     Node* mask = MakeConX(~round_mask);
3580     size       = _gvn.transform( new AndXNode(size, mask) );
3581   }
3582   // else if round_mask == 0, the size computation is self-rounding
3583 
3584   if (return_size_val != NULL) {
3585     // This is the size
3586     (*return_size_val) = size;
3587   }
3588 
3589   // Now generate allocation code
3590 
3591   // The entire memory state is needed for slow path of the allocation
3592   // since GC and deoptimization can happened.
3593   Node *mem = reset_memory();
3594   set_all_memory(mem); // Create new memory state
3595 
3596   // Create the AllocateArrayNode and its result projections
3597   AllocateArrayNode* alloc
3598     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3599                             control(), mem, i_o(),
3600                             size, klass_node,
3601                             initial_slow_test,
3602                             length);
3603 
3604   // Cast to correct type.  Note that the klass_node may be constant or not,
3605   // and in the latter case the actual array type will be inexact also.
3606   // (This happens via a non-constant argument to inline_native_newArray.)
3607   // In any case, the value of klass_node provides the desired array type.
3608   const TypeInt* length_type = _gvn.find_int_type(length);
3609   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3610   if (ary_type->isa_aryptr() && length_type != NULL) {
3611     // Try to get a better type than POS for the size
3612     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3613   }
3614 
3615   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3616 
3617   // Cast length on remaining path to be as narrow as possible
3618   if (map()->find_edge(length) >= 0) {
3619     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3620     if (ccast != length) {
3621       _gvn.set_type_bottom(ccast);
3622       record_for_igvn(ccast);
3623       replace_in_map(length, ccast);
3624     }
3625   }
3626 
3627   return javaoop;
3628 }
3629 
3630 // The following "Ideal_foo" functions are placed here because they recognize
3631 // the graph shapes created by the functions immediately above.
3632 
3633 //---------------------------Ideal_allocation----------------------------------
3634 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3635 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3636   if (ptr == NULL) {     // reduce dumb test in callers
3637     return NULL;
3638   }
3639   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3640     ptr = ptr->in(1);
3641     if (ptr == NULL) return NULL;
3642   }
3643   // Return NULL for allocations with several casts:
3644   //   j.l.reflect.Array.newInstance(jobject, jint)
3645   //   Object.clone()
3646   // to keep more precise type from last cast.
3647   if (ptr->is_Proj()) {
3648     Node* allo = ptr->in(0);
3649     if (allo != NULL && allo->is_Allocate()) {
3650       return allo->as_Allocate();
3651     }
3652   }
3653   // Report failure to match.
3654   return NULL;
3655 }
3656 
3657 // Fancy version which also strips off an offset (and reports it to caller).
3658 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3659                                              intptr_t& offset) {
3660   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3661   if (base == NULL)  return NULL;
3662   return Ideal_allocation(base, phase);
3663 }
3664 
3665 // Trace Initialize <- Proj[Parm] <- Allocate
3666 AllocateNode* InitializeNode::allocation() {
3667   Node* rawoop = in(InitializeNode::RawAddress);
3668   if (rawoop->is_Proj()) {
3669     Node* alloc = rawoop->in(0);
3670     if (alloc->is_Allocate()) {
3671       return alloc->as_Allocate();
3672     }
3673   }
3674   return NULL;
3675 }
3676 
3677 // Trace Allocate -> Proj[Parm] -> Initialize
3678 InitializeNode* AllocateNode::initialization() {
3679   ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3680   if (rawoop == NULL)  return NULL;
3681   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3682     Node* init = rawoop->fast_out(i);
3683     if (init->is_Initialize()) {
3684       assert(init->as_Initialize()->allocation() == this, "2-way link");
3685       return init->as_Initialize();
3686     }
3687   }
3688   return NULL;
3689 }
3690 
3691 //----------------------------- loop predicates ---------------------------
3692 
3693 //------------------------------add_predicate_impl----------------------------
3694 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3695   // Too many traps seen?
3696   if (too_many_traps(reason)) {
3697 #ifdef ASSERT
3698     if (TraceLoopPredicate) {
3699       int tc = C->trap_count(reason);
3700       tty->print("too many traps=%s tcount=%d in ",
3701                     Deoptimization::trap_reason_name(reason), tc);
3702       method()->print(); // which method has too many predicate traps
3703       tty->cr();
3704     }
3705 #endif
3706     // We cannot afford to take more traps here,
3707     // do not generate predicate.
3708     return;
3709   }
3710 
3711   Node *cont    = _gvn.intcon(1);
3712   Node* opq     = _gvn.transform(new Opaque1Node(C, cont));
3713   Node *bol     = _gvn.transform(new Conv2BNode(opq));
3714   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3715   Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3716   C->add_predicate_opaq(opq);
3717   {
3718     PreserveJVMState pjvms(this);
3719     set_control(iffalse);
3720     inc_sp(nargs);
3721     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3722   }
3723   Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3724   set_control(iftrue);
3725 }
3726 
3727 //------------------------------add_predicate---------------------------------
3728 void GraphKit::add_predicate(int nargs) {
3729   if (UseLoopPredicate) {
3730     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3731   }
3732   // loop's limit check predicate should be near the loop.
3733   if (LoopLimitCheck) {
3734     add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3735   }
3736 }
3737 
3738 //----------------------------- store barriers ----------------------------
3739 #define __ ideal.
3740 
3741 void GraphKit::sync_kit(IdealKit& ideal) {
3742   set_all_memory(__ merged_memory());
3743   set_i_o(__ i_o());
3744   set_control(__ ctrl());
3745 }
3746 
3747 void GraphKit::final_sync(IdealKit& ideal) {
3748   // Final sync IdealKit and graphKit.
3749   sync_kit(ideal);
3750 }
3751 
3752 Node* GraphKit::byte_map_base_node() {
3753   // Get base of card map
3754   CardTableModRefBS* ct =
3755     barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3756   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
3757   if (ct->byte_map_base != NULL) {
3758     return makecon(TypeRawPtr::make((address)ct->byte_map_base));
3759   } else {
3760     return null();
3761   }
3762 }
3763 
3764 // vanilla/CMS post barrier
3765 // Insert a write-barrier store.  This is to let generational GC work; we have
3766 // to flag all oop-stores before the next GC point.
3767 void GraphKit::write_barrier_post(Node* oop_store,
3768                                   Node* obj,
3769                                   Node* adr,
3770                                   uint  adr_idx,
3771                                   Node* val,
3772                                   bool use_precise) {
3773   // No store check needed if we're storing a NULL or an old object
3774   // (latter case is probably a string constant). The concurrent
3775   // mark sweep garbage collector, however, needs to have all nonNull
3776   // oop updates flagged via card-marks.
3777   if (val != NULL && val->is_Con()) {
3778     // must be either an oop or NULL
3779     const Type* t = val->bottom_type();
3780     if (t == TypePtr::NULL_PTR || t == Type::TOP)
3781       // stores of null never (?) need barriers
3782       return;
3783   }
3784 
3785   if (use_ReduceInitialCardMarks()
3786       && obj == just_allocated_object(control())) {
3787     // We can skip marks on a freshly-allocated object in Eden.
3788     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3789     // That routine informs GC to take appropriate compensating steps,
3790     // upon a slow-path allocation, so as to make this card-mark
3791     // elision safe.
3792     return;
3793   }
3794 
3795   if (!use_precise) {
3796     // All card marks for a (non-array) instance are in one place:
3797     adr = obj;
3798   }
3799   // (Else it's an array (or unknown), and we want more precise card marks.)
3800   assert(adr != NULL, "");
3801 
3802   IdealKit ideal(this, true);
3803 
3804   // Convert the pointer to an int prior to doing math on it
3805   Node* cast = __ CastPX(__ ctrl(), adr);
3806 
3807   // Divide by card size
3808   assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3809          "Only one we handle so far.");
3810   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3811 
3812   // Combine card table base and card offset
3813   Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3814 
3815   // Get the alias_index for raw card-mark memory
3816   int adr_type = Compile::AliasIdxRaw;
3817   Node*   zero = __ ConI(0); // Dirty card value
3818   BasicType bt = T_BYTE;
3819 
3820   if (UseConcMarkSweepGC && UseCondCardMark) {
3821     insert_mem_bar(Op_MemBarVolatile);   // StoreLoad barrier
3822     __ sync_kit(this);
3823   }
3824 
3825   if (UseCondCardMark) {
3826     // The classic GC reference write barrier is typically implemented
3827     // as a store into the global card mark table.  Unfortunately
3828     // unconditional stores can result in false sharing and excessive
3829     // coherence traffic as well as false transactional aborts.
3830     // UseCondCardMark enables MP "polite" conditional card mark
3831     // stores.  In theory we could relax the load from ctrl() to
3832     // no_ctrl, but that doesn't buy much latitude.
3833     Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3834     __ if_then(card_val, BoolTest::ne, zero);
3835   }
3836 
3837   // Smash zero into card
3838   if( !UseConcMarkSweepGC ) {
3839     __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3840   } else {
3841     // Specialized path for CM store barrier
3842     __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3843   }
3844 
3845   if (UseCondCardMark) {
3846     __ end_if();
3847   }
3848 
3849   // Final sync IdealKit and GraphKit.
3850   final_sync(ideal);
3851 }
3852 /*
3853  * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3854  * required by SATB to make sure all objects live at the start of the
3855  * marking are kept alive, all reference updates need to any previous
3856  * reference stored before writing.
3857  *
3858  * If the previous value is NULL there is no need to save the old value.
3859  * References that are NULL are filtered during runtime by the barrier
3860  * code to avoid unnecessary queuing.
3861  *
3862  * However in the case of newly allocated objects it might be possible to
3863  * prove that the reference about to be overwritten is NULL during compile
3864  * time and avoid adding the barrier code completely.
3865  *
3866  * The compiler needs to determine that the object in which a field is about
3867  * to be written is newly allocated, and that no prior store to the same field
3868  * has happened since the allocation.
3869  *
3870  * Returns true if the pre-barrier can be removed
3871  */
3872 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3873                                          BasicType bt, uint adr_idx) {
3874   intptr_t offset = 0;
3875   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3876   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3877 
3878   if (offset == Type::OffsetBot) {
3879     return false; // cannot unalias unless there are precise offsets
3880   }
3881 
3882   if (alloc == NULL) {
3883     return false; // No allocation found
3884   }
3885 
3886   intptr_t size_in_bytes = type2aelembytes(bt);
3887 
3888   Node* mem = memory(adr_idx); // start searching here...
3889 
3890   for (int cnt = 0; cnt < 50; cnt++) {
3891 
3892     if (mem->is_Store()) {
3893 
3894       Node* st_adr = mem->in(MemNode::Address);
3895       intptr_t st_offset = 0;
3896       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
3897 
3898       if (st_base == NULL) {
3899         break; // inscrutable pointer
3900       }
3901 
3902       // Break we have found a store with same base and offset as ours so break
3903       if (st_base == base && st_offset == offset) {
3904         break;
3905       }
3906 
3907       if (st_offset != offset && st_offset != Type::OffsetBot) {
3908         const int MAX_STORE = BytesPerLong;
3909         if (st_offset >= offset + size_in_bytes ||
3910             st_offset <= offset - MAX_STORE ||
3911             st_offset <= offset - mem->as_Store()->memory_size()) {
3912           // Success:  The offsets are provably independent.
3913           // (You may ask, why not just test st_offset != offset and be done?
3914           // The answer is that stores of different sizes can co-exist
3915           // in the same sequence of RawMem effects.  We sometimes initialize
3916           // a whole 'tile' of array elements with a single jint or jlong.)
3917           mem = mem->in(MemNode::Memory);
3918           continue; // advance through independent store memory
3919         }
3920       }
3921 
3922       if (st_base != base
3923           && MemNode::detect_ptr_independence(base, alloc, st_base,
3924                                               AllocateNode::Ideal_allocation(st_base, phase),
3925                                               phase)) {
3926         // Success:  The bases are provably independent.
3927         mem = mem->in(MemNode::Memory);
3928         continue; // advance through independent store memory
3929       }
3930     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
3931 
3932       InitializeNode* st_init = mem->in(0)->as_Initialize();
3933       AllocateNode* st_alloc = st_init->allocation();
3934 
3935       // Make sure that we are looking at the same allocation site.
3936       // The alloc variable is guaranteed to not be null here from earlier check.
3937       if (alloc == st_alloc) {
3938         // Check that the initialization is storing NULL so that no previous store
3939         // has been moved up and directly write a reference
3940         Node* captured_store = st_init->find_captured_store(offset,
3941                                                             type2aelembytes(T_OBJECT),
3942                                                             phase);
3943         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
3944           return true;
3945         }
3946       }
3947     }
3948 
3949     // Unless there is an explicit 'continue', we must bail out here,
3950     // because 'mem' is an inscrutable memory state (e.g., a call).
3951     break;
3952   }
3953 
3954   return false;
3955 }
3956 
3957 // G1 pre/post barriers
3958 void GraphKit::g1_write_barrier_pre(bool do_load,
3959                                     Node* obj,
3960                                     Node* adr,
3961                                     uint alias_idx,
3962                                     Node* val,
3963                                     const TypeOopPtr* val_type,
3964                                     Node* pre_val,
3965                                     BasicType bt) {
3966 
3967   // Some sanity checks
3968   // Note: val is unused in this routine.
3969 
3970   if (do_load) {
3971     // We need to generate the load of the previous value
3972     assert(obj != NULL, "must have a base");
3973     assert(adr != NULL, "where are loading from?");
3974     assert(pre_val == NULL, "loaded already?");
3975     assert(val_type != NULL, "need a type");
3976 
3977     if (use_ReduceInitialCardMarks()
3978         && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
3979       return;
3980     }
3981 
3982   } else {
3983     // In this case both val_type and alias_idx are unused.
3984     assert(pre_val != NULL, "must be loaded already");
3985     // Nothing to be done if pre_val is null.
3986     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3987     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3988   }
3989   assert(bt == T_OBJECT, "or we shouldn't be here");
3990 
3991   IdealKit ideal(this, true);
3992 
3993   Node* tls = __ thread(); // ThreadLocalStorage
3994 
3995   Node* no_ctrl = NULL;
3996   Node* no_base = __ top();
3997   Node* zero  = __ ConI(0);
3998   Node* zeroX = __ ConX(0);
3999 
4000   float likely  = PROB_LIKELY(0.999);
4001   float unlikely  = PROB_UNLIKELY(0.999);
4002 
4003   BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4004   assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4005 
4006   // Offsets into the thread
4007   const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +  // 648
4008                                           SATBMarkQueue::byte_offset_of_active());
4009   const int index_offset   = in_bytes(JavaThread::satb_mark_queue_offset() +  // 656
4010                                           SATBMarkQueue::byte_offset_of_index());
4011   const int buffer_offset  = in_bytes(JavaThread::satb_mark_queue_offset() +  // 652
4012                                           SATBMarkQueue::byte_offset_of_buf());
4013 
4014   // Now the actual pointers into the thread
4015   Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4016   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
4017   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
4018 
4019   // Now some of the values
4020   Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4021 
4022   // if (!marking)
4023   __ if_then(marking, BoolTest::ne, zero, unlikely); {
4024     BasicType index_bt = TypeX_X->basic_type();
4025     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4026     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4027 
4028     if (do_load) {
4029       // load original value
4030       // alias_idx correct??
4031       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4032     }
4033 
4034     // if (pre_val != NULL)
4035     __ if_then(pre_val, BoolTest::ne, null()); {
4036       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4037 
4038       // is the queue for this thread full?
4039       __ if_then(index, BoolTest::ne, zeroX, likely); {
4040 
4041         // decrement the index
4042         Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4043 
4044         // Now get the buffer location we will log the previous value into and store it
4045         Node *log_addr = __ AddP(no_base, buffer, next_index);
4046         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4047         // update the index
4048         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4049 
4050       } __ else_(); {
4051 
4052         // logging buffer is full, call the runtime
4053         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4054         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4055       } __ end_if();  // (!index)
4056     } __ end_if();  // (pre_val != NULL)
4057   } __ end_if();  // (!marking)
4058 
4059   // Final sync IdealKit and GraphKit.
4060   final_sync(ideal);
4061 }
4062 
4063 /*
4064  * G1 similar to any GC with a Young Generation requires a way to keep track of
4065  * references from Old Generation to Young Generation to make sure all live
4066  * objects are found. G1 also requires to keep track of object references
4067  * between different regions to enable evacuation of old regions, which is done
4068  * as part of mixed collections. References are tracked in remembered sets and
4069  * is continuously updated as reference are written to with the help of the
4070  * post-barrier.
4071  *
4072  * To reduce the number of updates to the remembered set the post-barrier
4073  * filters updates to fields in objects located in the Young Generation,
4074  * the same region as the reference, when the NULL is being written or
4075  * if the card is already marked as dirty by an earlier write.
4076  *
4077  * Under certain circumstances it is possible to avoid generating the
4078  * post-barrier completely if it is possible during compile time to prove
4079  * the object is newly allocated and that no safepoint exists between the
4080  * allocation and the store.
4081  *
4082  * In the case of slow allocation the allocation code must handle the barrier
4083  * as part of the allocation in the case the allocated object is not located
4084  * in the nursery, this would happen for humongous objects. This is similar to
4085  * how CMS is required to handle this case, see the comments for the method
4086  * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
4087  * A deferred card mark is required for these objects and handled in the above
4088  * mentioned methods.
4089  *
4090  * Returns true if the post barrier can be removed
4091  */
4092 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4093                                           Node* adr) {
4094   intptr_t      offset = 0;
4095   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4096   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
4097 
4098   if (offset == Type::OffsetBot) {
4099     return false; // cannot unalias unless there are precise offsets
4100   }
4101 
4102   if (alloc == NULL) {
4103      return false; // No allocation found
4104   }
4105 
4106   // Start search from Store node
4107   Node* mem = store->in(MemNode::Control);
4108   if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4109 
4110     InitializeNode* st_init = mem->in(0)->as_Initialize();
4111     AllocateNode*  st_alloc = st_init->allocation();
4112 
4113     // Make sure we are looking at the same allocation
4114     if (alloc == st_alloc) {
4115       return true;
4116     }
4117   }
4118 
4119   return false;
4120 }
4121 
4122 //
4123 // Update the card table and add card address to the queue
4124 //
4125 void GraphKit::g1_mark_card(IdealKit& ideal,
4126                             Node* card_adr,
4127                             Node* oop_store,
4128                             uint oop_alias_idx,
4129                             Node* index,
4130                             Node* index_adr,
4131                             Node* buffer,
4132                             const TypeFunc* tf) {
4133 
4134   Node* zero  = __ ConI(0);
4135   Node* zeroX = __ ConX(0);
4136   Node* no_base = __ top();
4137   BasicType card_bt = T_BYTE;
4138   // Smash zero into card. MUST BE ORDERED WRT TO STORE
4139   __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4140 
4141   //  Now do the queue work
4142   __ if_then(index, BoolTest::ne, zeroX); {
4143 
4144     Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4145     Node* log_addr = __ AddP(no_base, buffer, next_index);
4146 
4147     // Order, see storeCM.
4148     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4149     __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4150 
4151   } __ else_(); {
4152     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4153   } __ end_if();
4154 
4155 }
4156 
4157 void GraphKit::g1_write_barrier_post(Node* oop_store,
4158                                      Node* obj,
4159                                      Node* adr,
4160                                      uint alias_idx,
4161                                      Node* val,
4162                                      BasicType bt,
4163                                      bool use_precise) {
4164   // If we are writing a NULL then we need no post barrier
4165 
4166   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4167     // Must be NULL
4168     const Type* t = val->bottom_type();
4169     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4170     // No post barrier if writing NULLx
4171     return;
4172   }
4173 
4174   if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4175     // We can skip marks on a freshly-allocated object in Eden.
4176     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
4177     // That routine informs GC to take appropriate compensating steps,
4178     // upon a slow-path allocation, so as to make this card-mark
4179     // elision safe.
4180     return;
4181   }
4182 
4183   if (use_ReduceInitialCardMarks()
4184       && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4185     return;
4186   }
4187 
4188   if (!use_precise) {
4189     // All card marks for a (non-array) instance are in one place:
4190     adr = obj;
4191   }
4192   // (Else it's an array (or unknown), and we want more precise card marks.)
4193   assert(adr != NULL, "");
4194 
4195   IdealKit ideal(this, true);
4196 
4197   Node* tls = __ thread(); // ThreadLocalStorage
4198 
4199   Node* no_base = __ top();
4200   float likely  = PROB_LIKELY(0.999);
4201   float unlikely  = PROB_UNLIKELY(0.999);
4202   Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4203   Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4204   Node* zeroX = __ ConX(0);
4205 
4206   // Get the alias_index for raw card-mark memory
4207   const TypePtr* card_type = TypeRawPtr::BOTTOM;
4208 
4209   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4210 
4211   // Offsets into the thread
4212   const int index_offset  = in_bytes(JavaThread::dirty_card_queue_offset() +
4213                                      DirtyCardQueue::byte_offset_of_index());
4214   const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4215                                      DirtyCardQueue::byte_offset_of_buf());
4216 
4217   // Pointers into the thread
4218 
4219   Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4220   Node* index_adr =  __ AddP(no_base, tls, __ ConX(index_offset));
4221 
4222   // Now some values
4223   // Use ctrl to avoid hoisting these values past a safepoint, which could
4224   // potentially reset these fields in the JavaThread.
4225   Node* index  = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4226   Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4227 
4228   // Convert the store obj pointer to an int prior to doing math on it
4229   // Must use ctrl to prevent "integerized oop" existing across safepoint
4230   Node* cast =  __ CastPX(__ ctrl(), adr);
4231 
4232   // Divide pointer by card size
4233   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4234 
4235   // Combine card table base and card offset
4236   Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4237 
4238   // If we know the value being stored does it cross regions?
4239 
4240   if (val != NULL) {
4241     // Does the store cause us to cross regions?
4242 
4243     // Should be able to do an unsigned compare of region_size instead of
4244     // and extra shift. Do we have an unsigned compare??
4245     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4246     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4247 
4248     // if (xor_res == 0) same region so skip
4249     __ if_then(xor_res, BoolTest::ne, zeroX); {
4250 
4251       // No barrier if we are storing a NULL
4252       __ if_then(val, BoolTest::ne, null(), unlikely); {
4253 
4254         // Ok must mark the card if not already dirty
4255 
4256         // load the original value of the card
4257         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4258 
4259         __ if_then(card_val, BoolTest::ne, young_card); {
4260           sync_kit(ideal);
4261           // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4262           insert_mem_bar(Op_MemBarVolatile, oop_store);
4263           __ sync_kit(this);
4264 
4265           Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4266           __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4267             g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4268           } __ end_if();
4269         } __ end_if();
4270       } __ end_if();
4271     } __ end_if();
4272   } else {
4273     // Object.clone() instrinsic uses this path.
4274     g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4275   }
4276 
4277   // Final sync IdealKit and GraphKit.
4278   final_sync(ideal);
4279 }
4280 #undef __
4281 
4282 
4283 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4284   Node* len = load_array_length(load_String_value(ctrl, str));
4285   Node* coder = load_String_coder(ctrl, str);
4286   // Divide length by 2 if coder is UTF16
4287   return _gvn.transform(new RShiftINode(len, coder));
4288 }
4289 
4290 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4291   int value_offset = java_lang_String::value_offset_in_bytes();
4292   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4293                                                      false, NULL, 0);
4294   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4295   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4296                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4297                                                   ciTypeArrayKlass::make(T_BYTE), true, 0);
4298   int value_field_idx = C->get_alias_index(value_field_type);
4299   Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4300                          value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4301   // String.value field is known to be @Stable.
4302   if (UseImplicitStableValues) {
4303     load = cast_array_to_stable(load, value_type);
4304   }
4305   return load;
4306 }
4307 
4308 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4309   if (java_lang_String::has_coder_field()) {
4310     if (!CompactStrings) {
4311       return intcon(java_lang_String::CODER_UTF16);
4312     }
4313     int coder_offset = java_lang_String::coder_offset_in_bytes();
4314     const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4315                                                        false, NULL, 0);
4316     const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4317     int coder_field_idx = C->get_alias_index(coder_field_type);
4318     return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4319                      TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4320   } else {
4321     return intcon(0); // false
4322   }
4323 }
4324 
4325 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4326   int value_offset = java_lang_String::value_offset_in_bytes();
4327   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4328                                                      false, NULL, 0);
4329   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4330   store_oop_to_object(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
4331       value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4332 }
4333 
4334 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4335   int coder_offset = java_lang_String::coder_offset_in_bytes();
4336   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4337                                                      false, NULL, 0);
4338   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4339   int coder_field_idx = C->get_alias_index(coder_field_type);
4340   store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4341                   value, T_BYTE, coder_field_idx, MemNode::unordered);
4342 }
4343 
4344 // Capture src and dst memory state with a MergeMemNode
4345 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4346   if (src_type == dst_type) {
4347     // Types are equal, we don't need a MergeMemNode
4348     return memory(src_type);
4349   }
4350   MergeMemNode* merge = MergeMemNode::make(map()->memory());
4351   record_for_igvn(merge); // fold it up later, if possible
4352   int src_idx = C->get_alias_index(src_type);
4353   int dst_idx = C->get_alias_index(dst_type);
4354   merge->set_memory_at(src_idx, memory(src_idx));
4355   merge->set_memory_at(dst_idx, memory(dst_idx));
4356   return merge;
4357 }
4358 
4359 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4360   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4361   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4362   // If input and output memory types differ, capture both states to preserve
4363   // the dependency between preceding and subsequent loads/stores.
4364   // For example, the following program:
4365   //  StoreB
4366   //  compress_string
4367   //  LoadB
4368   // has this memory graph (use->def):
4369   //  LoadB -> compress_string -> CharMem
4370   //             ... -> StoreB -> ByteMem
4371   // The intrinsic hides the dependency between LoadB and StoreB, causing
4372   // the load to read from memory not containing the result of the StoreB.
4373   // The correct memory graph should look like this:
4374   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4375   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4376   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4377   Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4378   set_memory(res_mem, TypeAryPtr::BYTES);
4379   return str;
4380 }
4381 
4382 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4383   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4384   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4385   // Capture src and dst memory (see comment in 'compress_string').
4386   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4387   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4388   set_memory(_gvn.transform(str), dst_type);
4389 }
4390 
4391 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4392   /**
4393    * int i_char = start;
4394    * for (int i_byte = 0; i_byte < count; i_byte++) {
4395    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
4396    * }
4397    */
4398   add_predicate();
4399   RegionNode* head = new RegionNode(3);
4400   head->init_req(1, control());
4401   gvn().set_type(head, Type::CONTROL);
4402   record_for_igvn(head);
4403 
4404   Node* i_byte = new PhiNode(head, TypeInt::INT);
4405   i_byte->init_req(1, intcon(0));
4406   gvn().set_type(i_byte, TypeInt::INT);
4407   record_for_igvn(i_byte);
4408 
4409   Node* i_char = new PhiNode(head, TypeInt::INT);
4410   i_char->init_req(1, start);
4411   gvn().set_type(i_char, TypeInt::INT);
4412   record_for_igvn(i_char);
4413 
4414   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4415   gvn().set_type(mem, Type::MEMORY);
4416   record_for_igvn(mem);
4417   set_control(head);
4418   set_memory(mem, TypeAryPtr::BYTES);
4419   Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4420   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4421                              AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4422                              false, false, true /* mismatched */);
4423 
4424   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4425   head->init_req(2, IfTrue(iff));
4426   mem->init_req(2, st);
4427   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4428   i_char->init_req(2, AddI(i_char, intcon(2)));
4429 
4430   set_control(IfFalse(iff));
4431   set_memory(st, TypeAryPtr::BYTES);
4432 }
4433 
4434 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4435   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4436   // assumption of CCP analysis.
4437   return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4438 }